| 1 | /* |
| 2 | * IO priority handling helper functions common to the libaio and io_uring |
| 3 | * engines. |
| 4 | */ |
| 5 | |
| 6 | #include "cmdprio.h" |
| 7 | |
| 8 | /* |
| 9 | * Temporary array used during parsing. Will be freed after the corresponding |
| 10 | * struct bsprio_desc has been generated and saved in cmdprio->bsprio_desc. |
| 11 | */ |
| 12 | struct cmdprio_parse_result { |
| 13 | struct split_prio *entries; |
| 14 | int nr_entries; |
| 15 | }; |
| 16 | |
| 17 | /* |
| 18 | * Temporary array used during init. Will be freed after the corresponding |
| 19 | * struct clat_prio_stat array has been saved in td->ts.clat_prio and the |
| 20 | * matching clat_prio_indexes have been saved in each struct cmdprio_prio. |
| 21 | */ |
| 22 | struct cmdprio_values { |
| 23 | unsigned int *prios; |
| 24 | int nr_prios; |
| 25 | }; |
| 26 | |
| 27 | static int find_clat_prio_index(unsigned int *all_prios, int nr_prios, |
| 28 | int32_t prio) |
| 29 | { |
| 30 | int i; |
| 31 | |
| 32 | for (i = 0; i < nr_prios; i++) { |
| 33 | if (all_prios[i] == prio) |
| 34 | return i; |
| 35 | } |
| 36 | |
| 37 | return -1; |
| 38 | } |
| 39 | |
| 40 | /** |
| 41 | * assign_clat_prio_index - In order to avoid stat.c the need to loop through |
| 42 | * all possible priorities each time add_clat_sample() / add_lat_sample() is |
| 43 | * called, save which index to use in each cmdprio_prio. This will later be |
| 44 | * propagated to the io_u, if the specific io_u was determined to use a cmdprio |
| 45 | * priority value. |
| 46 | */ |
| 47 | static void assign_clat_prio_index(struct cmdprio_prio *prio, |
| 48 | struct cmdprio_values *values) |
| 49 | { |
| 50 | int clat_prio_index = find_clat_prio_index(values->prios, |
| 51 | values->nr_prios, |
| 52 | prio->prio); |
| 53 | if (clat_prio_index == -1) { |
| 54 | clat_prio_index = values->nr_prios; |
| 55 | values->prios[clat_prio_index] = prio->prio; |
| 56 | values->nr_prios++; |
| 57 | } |
| 58 | prio->clat_prio_index = clat_prio_index; |
| 59 | } |
| 60 | |
| 61 | /** |
| 62 | * init_cmdprio_values - Allocate a temporary array that can hold all unique |
| 63 | * priorities (per ddir), so that we can assign_clat_prio_index() for each |
| 64 | * cmdprio_prio during setup. This temporary array is freed after setup. |
| 65 | */ |
| 66 | static int init_cmdprio_values(struct cmdprio_values *values, |
| 67 | int max_unique_prios, struct thread_stat *ts) |
| 68 | { |
| 69 | values->prios = calloc(max_unique_prios + 1, |
| 70 | sizeof(*values->prios)); |
| 71 | if (!values->prios) |
| 72 | return 1; |
| 73 | |
| 74 | /* td->ioprio/ts->ioprio is always stored at index 0. */ |
| 75 | values->prios[0] = ts->ioprio; |
| 76 | values->nr_prios++; |
| 77 | |
| 78 | return 0; |
| 79 | } |
| 80 | |
| 81 | /** |
| 82 | * init_ts_clat_prio - Allocates and fills a clat_prio_stat array which holds |
| 83 | * all unique priorities (per ddir). |
| 84 | */ |
| 85 | static int init_ts_clat_prio(struct thread_stat *ts, enum fio_ddir ddir, |
| 86 | struct cmdprio_values *values) |
| 87 | { |
| 88 | int i; |
| 89 | |
| 90 | if (alloc_clat_prio_stat_ddir(ts, ddir, values->nr_prios)) |
| 91 | return 1; |
| 92 | |
| 93 | for (i = 0; i < values->nr_prios; i++) |
| 94 | ts->clat_prio[ddir][i].ioprio = values->prios[i]; |
| 95 | |
| 96 | return 0; |
| 97 | } |
| 98 | |
| 99 | static int fio_cmdprio_fill_bsprio(struct cmdprio_bsprio *bsprio, |
| 100 | struct split_prio *entries, |
| 101 | struct cmdprio_values *values, |
| 102 | int implicit_cmdprio, int start, int end) |
| 103 | { |
| 104 | struct cmdprio_prio *prio; |
| 105 | int i = end - start + 1; |
| 106 | |
| 107 | bsprio->prios = calloc(i, sizeof(*bsprio->prios)); |
| 108 | if (!bsprio->prios) |
| 109 | return 1; |
| 110 | |
| 111 | bsprio->bs = entries[start].bs; |
| 112 | bsprio->nr_prios = 0; |
| 113 | for (i = start; i <= end; i++) { |
| 114 | prio = &bsprio->prios[bsprio->nr_prios]; |
| 115 | prio->perc = entries[i].perc; |
| 116 | if (entries[i].prio == -1) |
| 117 | prio->prio = implicit_cmdprio; |
| 118 | else |
| 119 | prio->prio = entries[i].prio; |
| 120 | assign_clat_prio_index(prio, values); |
| 121 | bsprio->tot_perc += entries[i].perc; |
| 122 | if (bsprio->tot_perc > 100) { |
| 123 | log_err("fio: cmdprio_bssplit total percentage " |
| 124 | "for bs: %"PRIu64" exceeds 100\n", |
| 125 | bsprio->bs); |
| 126 | free(bsprio->prios); |
| 127 | return 1; |
| 128 | } |
| 129 | bsprio->nr_prios++; |
| 130 | } |
| 131 | |
| 132 | return 0; |
| 133 | } |
| 134 | |
| 135 | static int |
| 136 | fio_cmdprio_generate_bsprio_desc(struct cmdprio_bsprio_desc *bsprio_desc, |
| 137 | struct cmdprio_parse_result *parse_res, |
| 138 | struct cmdprio_values *values, |
| 139 | int implicit_cmdprio) |
| 140 | { |
| 141 | struct split_prio *entries = parse_res->entries; |
| 142 | int nr_entries = parse_res->nr_entries; |
| 143 | struct cmdprio_bsprio *bsprio; |
| 144 | int i, start, count = 0; |
| 145 | |
| 146 | /* |
| 147 | * The parsed result is sorted by blocksize, so count only the number |
| 148 | * of different blocksizes, to know how many cmdprio_bsprio we need. |
| 149 | */ |
| 150 | for (i = 0; i < nr_entries; i++) { |
| 151 | while (i + 1 < nr_entries && entries[i].bs == entries[i + 1].bs) |
| 152 | i++; |
| 153 | count++; |
| 154 | } |
| 155 | |
| 156 | /* |
| 157 | * This allocation is not freed on error. Instead, the calling function |
| 158 | * is responsible for calling fio_cmdprio_cleanup() on error. |
| 159 | */ |
| 160 | bsprio_desc->bsprios = calloc(count, sizeof(*bsprio_desc->bsprios)); |
| 161 | if (!bsprio_desc->bsprios) |
| 162 | return 1; |
| 163 | |
| 164 | start = 0; |
| 165 | bsprio_desc->nr_bsprios = 0; |
| 166 | for (i = 0; i < nr_entries; i++) { |
| 167 | while (i + 1 < nr_entries && entries[i].bs == entries[i + 1].bs) |
| 168 | i++; |
| 169 | bsprio = &bsprio_desc->bsprios[bsprio_desc->nr_bsprios]; |
| 170 | /* |
| 171 | * All parsed entries with the same blocksize get saved in the |
| 172 | * same cmdprio_bsprio, to expedite the search in the hot path. |
| 173 | */ |
| 174 | if (fio_cmdprio_fill_bsprio(bsprio, entries, values, |
| 175 | implicit_cmdprio, start, i)) |
| 176 | return 1; |
| 177 | |
| 178 | start = i + 1; |
| 179 | bsprio_desc->nr_bsprios++; |
| 180 | } |
| 181 | |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | static int fio_cmdprio_bssplit_ddir(struct thread_options *to, void *cb_arg, |
| 186 | enum fio_ddir ddir, char *str, bool data) |
| 187 | { |
| 188 | struct cmdprio_parse_result *parse_res_arr = cb_arg; |
| 189 | struct cmdprio_parse_result *parse_res = &parse_res_arr[ddir]; |
| 190 | |
| 191 | if (ddir == DDIR_TRIM) |
| 192 | return 0; |
| 193 | |
| 194 | if (split_parse_prio_ddir(to, &parse_res->entries, |
| 195 | &parse_res->nr_entries, str)) |
| 196 | return 1; |
| 197 | |
| 198 | return 0; |
| 199 | } |
| 200 | |
| 201 | static int fio_cmdprio_bssplit_parse(struct thread_data *td, const char *input, |
| 202 | struct cmdprio_parse_result *parse_res) |
| 203 | { |
| 204 | char *str, *p; |
| 205 | int ret = 0; |
| 206 | |
| 207 | p = str = strdup(input); |
| 208 | |
| 209 | strip_blank_front(&str); |
| 210 | strip_blank_end(str); |
| 211 | |
| 212 | ret = str_split_parse(td, str, fio_cmdprio_bssplit_ddir, parse_res, |
| 213 | false); |
| 214 | |
| 215 | free(p); |
| 216 | return ret; |
| 217 | } |
| 218 | |
| 219 | /** |
| 220 | * fio_cmdprio_percentage - Returns the percentage of I/Os that should |
| 221 | * use a cmdprio priority value (rather than the default context priority). |
| 222 | * |
| 223 | * For CMDPRIO_MODE_BSSPLIT, if the percentage is non-zero, we will also |
| 224 | * return the matching bsprio, to avoid the same linear search elsewhere. |
| 225 | * For CMDPRIO_MODE_PERC, we will never return a bsprio. |
| 226 | */ |
| 227 | static int fio_cmdprio_percentage(struct cmdprio *cmdprio, struct io_u *io_u, |
| 228 | struct cmdprio_bsprio **bsprio) |
| 229 | { |
| 230 | struct cmdprio_bsprio *bsprio_entry; |
| 231 | enum fio_ddir ddir = io_u->ddir; |
| 232 | int i; |
| 233 | |
| 234 | switch (cmdprio->mode) { |
| 235 | case CMDPRIO_MODE_PERC: |
| 236 | *bsprio = NULL; |
| 237 | return cmdprio->perc_entry[ddir].perc; |
| 238 | case CMDPRIO_MODE_BSSPLIT: |
| 239 | for (i = 0; i < cmdprio->bsprio_desc[ddir].nr_bsprios; i++) { |
| 240 | bsprio_entry = &cmdprio->bsprio_desc[ddir].bsprios[i]; |
| 241 | if (bsprio_entry->bs == io_u->buflen) { |
| 242 | *bsprio = bsprio_entry; |
| 243 | return bsprio_entry->tot_perc; |
| 244 | } |
| 245 | } |
| 246 | break; |
| 247 | default: |
| 248 | /* |
| 249 | * An I/O engine should never call this function if cmdprio |
| 250 | * is not is use. |
| 251 | */ |
| 252 | assert(0); |
| 253 | } |
| 254 | |
| 255 | /* |
| 256 | * This is totally fine, the given blocksize simply does not |
| 257 | * have any (non-zero) cmdprio_bssplit entries defined. |
| 258 | */ |
| 259 | *bsprio = NULL; |
| 260 | return 0; |
| 261 | } |
| 262 | |
| 263 | /** |
| 264 | * fio_cmdprio_set_ioprio - Set an io_u ioprio according to cmdprio options |
| 265 | * |
| 266 | * Generates a random percentage value to determine if an io_u ioprio needs |
| 267 | * to be set. If the random percentage value is within the user specified |
| 268 | * percentage of I/Os that should use a cmdprio priority value (rather than |
| 269 | * the default priority), then this function updates the io_u with an ioprio |
| 270 | * value as defined by the cmdprio/cmdprio_hint/cmdprio_class or |
| 271 | * cmdprio_bssplit options. |
| 272 | * |
| 273 | * Return true if the io_u ioprio was changed and false otherwise. |
| 274 | */ |
| 275 | bool fio_cmdprio_set_ioprio(struct thread_data *td, struct cmdprio *cmdprio, |
| 276 | struct io_u *io_u) |
| 277 | { |
| 278 | struct cmdprio_bsprio *bsprio; |
| 279 | unsigned int p, rand; |
| 280 | uint32_t perc = 0; |
| 281 | int i; |
| 282 | |
| 283 | p = fio_cmdprio_percentage(cmdprio, io_u, &bsprio); |
| 284 | if (!p) |
| 285 | return false; |
| 286 | |
| 287 | rand = rand_between(&td->prio_state, 0, 99); |
| 288 | if (rand >= p) |
| 289 | return false; |
| 290 | |
| 291 | switch (cmdprio->mode) { |
| 292 | case CMDPRIO_MODE_PERC: |
| 293 | io_u->ioprio = cmdprio->perc_entry[io_u->ddir].prio; |
| 294 | io_u->clat_prio_index = |
| 295 | cmdprio->perc_entry[io_u->ddir].clat_prio_index; |
| 296 | return true; |
| 297 | case CMDPRIO_MODE_BSSPLIT: |
| 298 | assert(bsprio); |
| 299 | for (i = 0; i < bsprio->nr_prios; i++) { |
| 300 | struct cmdprio_prio *prio = &bsprio->prios[i]; |
| 301 | |
| 302 | perc += prio->perc; |
| 303 | if (rand < perc) { |
| 304 | io_u->ioprio = prio->prio; |
| 305 | io_u->clat_prio_index = prio->clat_prio_index; |
| 306 | return true; |
| 307 | } |
| 308 | } |
| 309 | break; |
| 310 | default: |
| 311 | assert(0); |
| 312 | } |
| 313 | |
| 314 | /* When rand < p (total perc), we should always find a cmdprio_prio. */ |
| 315 | assert(0); |
| 316 | return false; |
| 317 | } |
| 318 | |
| 319 | static int fio_cmdprio_gen_perc(struct thread_data *td, struct cmdprio *cmdprio) |
| 320 | { |
| 321 | struct cmdprio_options *options = cmdprio->options; |
| 322 | struct cmdprio_prio *prio; |
| 323 | struct cmdprio_values values[CMDPRIO_RWDIR_CNT] = {}; |
| 324 | struct thread_stat *ts = &td->ts; |
| 325 | enum fio_ddir ddir; |
| 326 | int ret; |
| 327 | |
| 328 | for (ddir = 0; ddir < CMDPRIO_RWDIR_CNT; ddir++) { |
| 329 | /* |
| 330 | * Do not allocate a clat_prio array nor set the cmdprio struct |
| 331 | * if zero percent of the I/Os (for the ddir) should use a |
| 332 | * cmdprio priority value, or when the ddir is not enabled. |
| 333 | */ |
| 334 | if (!options->percentage[ddir] || |
| 335 | (ddir == DDIR_READ && !td_read(td)) || |
| 336 | (ddir == DDIR_WRITE && !td_write(td))) |
| 337 | continue; |
| 338 | |
| 339 | ret = init_cmdprio_values(&values[ddir], 1, ts); |
| 340 | if (ret) |
| 341 | goto err; |
| 342 | |
| 343 | prio = &cmdprio->perc_entry[ddir]; |
| 344 | prio->perc = options->percentage[ddir]; |
| 345 | prio->prio = ioprio_value(options->class[ddir], |
| 346 | options->level[ddir], |
| 347 | options->hint[ddir]); |
| 348 | assign_clat_prio_index(prio, &values[ddir]); |
| 349 | |
| 350 | ret = init_ts_clat_prio(ts, ddir, &values[ddir]); |
| 351 | if (ret) |
| 352 | goto err; |
| 353 | |
| 354 | free(values[ddir].prios); |
| 355 | values[ddir].prios = NULL; |
| 356 | values[ddir].nr_prios = 0; |
| 357 | } |
| 358 | |
| 359 | return 0; |
| 360 | |
| 361 | err: |
| 362 | for (ddir = 0; ddir < CMDPRIO_RWDIR_CNT; ddir++) |
| 363 | free(values[ddir].prios); |
| 364 | free_clat_prio_stats(ts); |
| 365 | |
| 366 | return ret; |
| 367 | } |
| 368 | |
| 369 | static int fio_cmdprio_parse_and_gen_bssplit(struct thread_data *td, |
| 370 | struct cmdprio *cmdprio) |
| 371 | { |
| 372 | struct cmdprio_options *options = cmdprio->options; |
| 373 | struct cmdprio_parse_result parse_res[CMDPRIO_RWDIR_CNT] = {}; |
| 374 | struct cmdprio_values values[CMDPRIO_RWDIR_CNT] = {}; |
| 375 | struct thread_stat *ts = &td->ts; |
| 376 | int ret, implicit_cmdprio; |
| 377 | enum fio_ddir ddir; |
| 378 | |
| 379 | ret = fio_cmdprio_bssplit_parse(td, options->bssplit_str, |
| 380 | &parse_res[0]); |
| 381 | if (ret) |
| 382 | goto err; |
| 383 | |
| 384 | for (ddir = 0; ddir < CMDPRIO_RWDIR_CNT; ddir++) { |
| 385 | /* |
| 386 | * Do not allocate a clat_prio array nor set the cmdprio structs |
| 387 | * if there are no non-zero entries (for the ddir), or when the |
| 388 | * ddir is not enabled. |
| 389 | */ |
| 390 | if (!parse_res[ddir].nr_entries || |
| 391 | (ddir == DDIR_READ && !td_read(td)) || |
| 392 | (ddir == DDIR_WRITE && !td_write(td))) { |
| 393 | free(parse_res[ddir].entries); |
| 394 | parse_res[ddir].entries = NULL; |
| 395 | parse_res[ddir].nr_entries = 0; |
| 396 | continue; |
| 397 | } |
| 398 | |
| 399 | ret = init_cmdprio_values(&values[ddir], |
| 400 | parse_res[ddir].nr_entries, ts); |
| 401 | if (ret) |
| 402 | goto err; |
| 403 | |
| 404 | implicit_cmdprio = ioprio_value(options->class[ddir], |
| 405 | options->level[ddir], |
| 406 | options->hint[ddir]); |
| 407 | |
| 408 | ret = fio_cmdprio_generate_bsprio_desc(&cmdprio->bsprio_desc[ddir], |
| 409 | &parse_res[ddir], |
| 410 | &values[ddir], |
| 411 | implicit_cmdprio); |
| 412 | if (ret) |
| 413 | goto err; |
| 414 | |
| 415 | free(parse_res[ddir].entries); |
| 416 | parse_res[ddir].entries = NULL; |
| 417 | parse_res[ddir].nr_entries = 0; |
| 418 | |
| 419 | ret = init_ts_clat_prio(ts, ddir, &values[ddir]); |
| 420 | if (ret) |
| 421 | goto err; |
| 422 | |
| 423 | free(values[ddir].prios); |
| 424 | values[ddir].prios = NULL; |
| 425 | values[ddir].nr_prios = 0; |
| 426 | } |
| 427 | |
| 428 | return 0; |
| 429 | |
| 430 | err: |
| 431 | for (ddir = 0; ddir < CMDPRIO_RWDIR_CNT; ddir++) { |
| 432 | free(parse_res[ddir].entries); |
| 433 | free(values[ddir].prios); |
| 434 | } |
| 435 | free_clat_prio_stats(ts); |
| 436 | fio_cmdprio_cleanup(cmdprio); |
| 437 | |
| 438 | return ret; |
| 439 | } |
| 440 | |
| 441 | static int fio_cmdprio_parse_and_gen(struct thread_data *td, |
| 442 | struct cmdprio *cmdprio) |
| 443 | { |
| 444 | struct cmdprio_options *options = cmdprio->options; |
| 445 | int i, ret; |
| 446 | |
| 447 | /* |
| 448 | * If cmdprio_percentage/cmdprio_bssplit is set and cmdprio_class |
| 449 | * is not set, default to RT priority class. |
| 450 | */ |
| 451 | for (i = 0; i < CMDPRIO_RWDIR_CNT; i++) { |
| 452 | /* |
| 453 | * A cmdprio value is only used when fio_cmdprio_percentage() |
| 454 | * returns non-zero, so it is safe to set a class even for a |
| 455 | * DDIR that will never use it. |
| 456 | */ |
| 457 | if (!options->class[i]) |
| 458 | options->class[i] = IOPRIO_CLASS_RT; |
| 459 | } |
| 460 | |
| 461 | switch (cmdprio->mode) { |
| 462 | case CMDPRIO_MODE_BSSPLIT: |
| 463 | ret = fio_cmdprio_parse_and_gen_bssplit(td, cmdprio); |
| 464 | break; |
| 465 | case CMDPRIO_MODE_PERC: |
| 466 | ret = fio_cmdprio_gen_perc(td, cmdprio); |
| 467 | break; |
| 468 | default: |
| 469 | assert(0); |
| 470 | return 1; |
| 471 | } |
| 472 | |
| 473 | return ret; |
| 474 | } |
| 475 | |
| 476 | void fio_cmdprio_cleanup(struct cmdprio *cmdprio) |
| 477 | { |
| 478 | enum fio_ddir ddir; |
| 479 | int i; |
| 480 | |
| 481 | for (ddir = 0; ddir < CMDPRIO_RWDIR_CNT; ddir++) { |
| 482 | for (i = 0; i < cmdprio->bsprio_desc[ddir].nr_bsprios; i++) |
| 483 | free(cmdprio->bsprio_desc[ddir].bsprios[i].prios); |
| 484 | free(cmdprio->bsprio_desc[ddir].bsprios); |
| 485 | cmdprio->bsprio_desc[ddir].bsprios = NULL; |
| 486 | cmdprio->bsprio_desc[ddir].nr_bsprios = 0; |
| 487 | } |
| 488 | |
| 489 | /* |
| 490 | * options points to a cmdprio_options struct that is part of td->eo. |
| 491 | * td->eo itself will be freed by free_ioengine(). |
| 492 | */ |
| 493 | cmdprio->options = NULL; |
| 494 | } |
| 495 | |
| 496 | int fio_cmdprio_init(struct thread_data *td, struct cmdprio *cmdprio, |
| 497 | struct cmdprio_options *options) |
| 498 | { |
| 499 | struct thread_options *to = &td->o; |
| 500 | bool has_cmdprio_percentage = false; |
| 501 | bool has_cmdprio_bssplit = false; |
| 502 | int i; |
| 503 | |
| 504 | cmdprio->options = options; |
| 505 | |
| 506 | if (options->bssplit_str && strlen(options->bssplit_str)) |
| 507 | has_cmdprio_bssplit = true; |
| 508 | |
| 509 | for (i = 0; i < CMDPRIO_RWDIR_CNT; i++) { |
| 510 | if (options->percentage[i]) |
| 511 | has_cmdprio_percentage = true; |
| 512 | } |
| 513 | |
| 514 | /* |
| 515 | * Check for option conflicts |
| 516 | */ |
| 517 | if (has_cmdprio_percentage && has_cmdprio_bssplit) { |
| 518 | log_err("%s: cmdprio_percentage and cmdprio_bssplit options " |
| 519 | "are mutually exclusive\n", |
| 520 | to->name); |
| 521 | return 1; |
| 522 | } |
| 523 | |
| 524 | if (has_cmdprio_bssplit) |
| 525 | cmdprio->mode = CMDPRIO_MODE_BSSPLIT; |
| 526 | else if (has_cmdprio_percentage) |
| 527 | cmdprio->mode = CMDPRIO_MODE_PERC; |
| 528 | else |
| 529 | cmdprio->mode = CMDPRIO_MODE_NONE; |
| 530 | |
| 531 | /* Nothing left to do if cmdprio is not used */ |
| 532 | if (cmdprio->mode == CMDPRIO_MODE_NONE) |
| 533 | return 0; |
| 534 | |
| 535 | return fio_cmdprio_parse_and_gen(td, cmdprio); |
| 536 | } |