| 1 | /* |
| 2 | * Copyright (C) 2018 Western Digital Corporation or its affiliates. |
| 3 | * |
| 4 | * This file is released under the GPL. |
| 5 | */ |
| 6 | |
| 7 | #include <errno.h> |
| 8 | #include <string.h> |
| 9 | #include <stdlib.h> |
| 10 | #include <fcntl.h> |
| 11 | #include <sys/stat.h> |
| 12 | #include <unistd.h> |
| 13 | |
| 14 | #include "os/os.h" |
| 15 | #include "file.h" |
| 16 | #include "fio.h" |
| 17 | #include "lib/pow2.h" |
| 18 | #include "log.h" |
| 19 | #include "oslib/asprintf.h" |
| 20 | #include "smalloc.h" |
| 21 | #include "verify.h" |
| 22 | #include "pshared.h" |
| 23 | #include "zbd.h" |
| 24 | |
| 25 | static bool is_valid_offset(const struct fio_file *f, uint64_t offset) |
| 26 | { |
| 27 | return (uint64_t)(offset - f->file_offset) < f->io_size; |
| 28 | } |
| 29 | |
| 30 | static inline unsigned int zbd_zone_idx(const struct fio_file *f, |
| 31 | struct fio_zone_info *zone) |
| 32 | { |
| 33 | return zone - f->zbd_info->zone_info; |
| 34 | } |
| 35 | |
| 36 | /** |
| 37 | * zbd_offset_to_zone_idx - convert an offset into a zone number |
| 38 | * @f: file pointer. |
| 39 | * @offset: offset in bytes. If this offset is in the first zone_size bytes |
| 40 | * past the disk size then the index of the sentinel is returned. |
| 41 | */ |
| 42 | static unsigned int zbd_offset_to_zone_idx(const struct fio_file *f, |
| 43 | uint64_t offset) |
| 44 | { |
| 45 | uint32_t zone_idx; |
| 46 | |
| 47 | if (f->zbd_info->zone_size_log2 > 0) |
| 48 | zone_idx = offset >> f->zbd_info->zone_size_log2; |
| 49 | else |
| 50 | zone_idx = offset / f->zbd_info->zone_size; |
| 51 | |
| 52 | return min(zone_idx, f->zbd_info->nr_zones); |
| 53 | } |
| 54 | |
| 55 | /** |
| 56 | * zbd_zone_end - Return zone end location |
| 57 | * @z: zone info pointer. |
| 58 | */ |
| 59 | static inline uint64_t zbd_zone_end(const struct fio_zone_info *z) |
| 60 | { |
| 61 | return (z+1)->start; |
| 62 | } |
| 63 | |
| 64 | /** |
| 65 | * zbd_zone_capacity_end - Return zone capacity limit end location |
| 66 | * @z: zone info pointer. |
| 67 | */ |
| 68 | static inline uint64_t zbd_zone_capacity_end(const struct fio_zone_info *z) |
| 69 | { |
| 70 | return z->start + z->capacity; |
| 71 | } |
| 72 | |
| 73 | /** |
| 74 | * zbd_zone_full - verify whether a minimum number of bytes remain in a zone |
| 75 | * @f: file pointer. |
| 76 | * @z: zone info pointer. |
| 77 | * @required: minimum number of bytes that must remain in a zone. |
| 78 | * |
| 79 | * The caller must hold z->mutex. |
| 80 | */ |
| 81 | static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z, |
| 82 | uint64_t required) |
| 83 | { |
| 84 | assert((required & 511) == 0); |
| 85 | |
| 86 | return z->has_wp && |
| 87 | z->wp + required > zbd_zone_capacity_end(z); |
| 88 | } |
| 89 | |
| 90 | static void zone_lock(struct thread_data *td, const struct fio_file *f, |
| 91 | struct fio_zone_info *z) |
| 92 | { |
| 93 | struct zoned_block_device_info *zbd = f->zbd_info; |
| 94 | uint32_t nz = z - zbd->zone_info; |
| 95 | |
| 96 | /* A thread should never lock zones outside its working area. */ |
| 97 | assert(f->min_zone <= nz && nz < f->max_zone); |
| 98 | |
| 99 | assert(z->has_wp); |
| 100 | |
| 101 | /* |
| 102 | * Lock the io_u target zone. The zone will be unlocked if io_u offset |
| 103 | * is changed or when io_u completes and zbd_put_io() executed. |
| 104 | * To avoid multiple jobs doing asynchronous I/Os from deadlocking each |
| 105 | * other waiting for zone locks when building an io_u batch, first |
| 106 | * only trylock the zone. If the zone is already locked by another job, |
| 107 | * process the currently queued I/Os so that I/O progress is made and |
| 108 | * zones unlocked. |
| 109 | */ |
| 110 | if (pthread_mutex_trylock(&z->mutex) != 0) { |
| 111 | if (!td_ioengine_flagged(td, FIO_SYNCIO)) |
| 112 | io_u_quiesce(td); |
| 113 | pthread_mutex_lock(&z->mutex); |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | static inline void zone_unlock(struct fio_zone_info *z) |
| 118 | { |
| 119 | int ret; |
| 120 | |
| 121 | assert(z->has_wp); |
| 122 | ret = pthread_mutex_unlock(&z->mutex); |
| 123 | assert(!ret); |
| 124 | } |
| 125 | |
| 126 | static inline struct fio_zone_info *zbd_get_zone(const struct fio_file *f, |
| 127 | unsigned int zone_idx) |
| 128 | { |
| 129 | return &f->zbd_info->zone_info[zone_idx]; |
| 130 | } |
| 131 | |
| 132 | static inline struct fio_zone_info * |
| 133 | zbd_offset_to_zone(const struct fio_file *f, uint64_t offset) |
| 134 | { |
| 135 | return zbd_get_zone(f, zbd_offset_to_zone_idx(f, offset)); |
| 136 | } |
| 137 | |
| 138 | /** |
| 139 | * zbd_get_zoned_model - Get a device zoned model |
| 140 | * @td: FIO thread data |
| 141 | * @f: FIO file for which to get model information |
| 142 | */ |
| 143 | static int zbd_get_zoned_model(struct thread_data *td, struct fio_file *f, |
| 144 | enum zbd_zoned_model *model) |
| 145 | { |
| 146 | int ret; |
| 147 | |
| 148 | if (f->filetype == FIO_TYPE_PIPE) { |
| 149 | log_err("zonemode=zbd does not support pipes\n"); |
| 150 | return -EINVAL; |
| 151 | } |
| 152 | |
| 153 | /* If regular file, always emulate zones inside the file. */ |
| 154 | if (f->filetype == FIO_TYPE_FILE) { |
| 155 | *model = ZBD_NONE; |
| 156 | return 0; |
| 157 | } |
| 158 | |
| 159 | if (td->io_ops && td->io_ops->get_zoned_model) |
| 160 | ret = td->io_ops->get_zoned_model(td, f, model); |
| 161 | else |
| 162 | ret = blkzoned_get_zoned_model(td, f, model); |
| 163 | if (ret < 0) { |
| 164 | td_verror(td, errno, "get zoned model failed"); |
| 165 | log_err("%s: get zoned model failed (%d).\n", |
| 166 | f->file_name, errno); |
| 167 | } |
| 168 | |
| 169 | return ret; |
| 170 | } |
| 171 | |
| 172 | /** |
| 173 | * zbd_report_zones - Get zone information |
| 174 | * @td: FIO thread data. |
| 175 | * @f: FIO file for which to get zone information |
| 176 | * @offset: offset from which to report zones |
| 177 | * @zones: Array of struct zbd_zone |
| 178 | * @nr_zones: Size of @zones array |
| 179 | * |
| 180 | * Get zone information into @zones starting from the zone at offset @offset |
| 181 | * for the device specified by @f. |
| 182 | * |
| 183 | * Returns the number of zones reported upon success and a negative error code |
| 184 | * upon failure. If the zone report is empty, always assume an error (device |
| 185 | * problem) and return -EIO. |
| 186 | */ |
| 187 | static int zbd_report_zones(struct thread_data *td, struct fio_file *f, |
| 188 | uint64_t offset, struct zbd_zone *zones, |
| 189 | unsigned int nr_zones) |
| 190 | { |
| 191 | int ret; |
| 192 | |
| 193 | if (td->io_ops && td->io_ops->report_zones) |
| 194 | ret = td->io_ops->report_zones(td, f, offset, zones, nr_zones); |
| 195 | else |
| 196 | ret = blkzoned_report_zones(td, f, offset, zones, nr_zones); |
| 197 | if (ret < 0) { |
| 198 | td_verror(td, errno, "report zones failed"); |
| 199 | log_err("%s: report zones from sector %"PRIu64" failed (%d).\n", |
| 200 | f->file_name, offset >> 9, errno); |
| 201 | } else if (ret == 0) { |
| 202 | td_verror(td, errno, "Empty zone report"); |
| 203 | log_err("%s: report zones from sector %"PRIu64" is empty.\n", |
| 204 | f->file_name, offset >> 9); |
| 205 | ret = -EIO; |
| 206 | } |
| 207 | |
| 208 | return ret; |
| 209 | } |
| 210 | |
| 211 | /** |
| 212 | * zbd_reset_wp - reset the write pointer of a range of zones |
| 213 | * @td: FIO thread data. |
| 214 | * @f: FIO file for which to reset zones |
| 215 | * @offset: Starting offset of the first zone to reset |
| 216 | * @length: Length of the range of zones to reset |
| 217 | * |
| 218 | * Reset the write pointer of all zones in the range @offset...@offset+@length. |
| 219 | * Returns 0 upon success and a negative error code upon failure. |
| 220 | */ |
| 221 | static int zbd_reset_wp(struct thread_data *td, struct fio_file *f, |
| 222 | uint64_t offset, uint64_t length) |
| 223 | { |
| 224 | int ret; |
| 225 | |
| 226 | if (td->io_ops && td->io_ops->reset_wp) |
| 227 | ret = td->io_ops->reset_wp(td, f, offset, length); |
| 228 | else |
| 229 | ret = blkzoned_reset_wp(td, f, offset, length); |
| 230 | if (ret < 0) { |
| 231 | td_verror(td, errno, "resetting wp failed"); |
| 232 | log_err("%s: resetting wp for %"PRIu64" sectors at sector %"PRIu64" failed (%d).\n", |
| 233 | f->file_name, length >> 9, offset >> 9, errno); |
| 234 | } |
| 235 | |
| 236 | return ret; |
| 237 | } |
| 238 | |
| 239 | /** |
| 240 | * zbd_reset_zone - reset the write pointer of a single zone |
| 241 | * @td: FIO thread data. |
| 242 | * @f: FIO file associated with the disk for which to reset a write pointer. |
| 243 | * @z: Zone to reset. |
| 244 | * |
| 245 | * Returns 0 upon success and a negative error code upon failure. |
| 246 | * |
| 247 | * The caller must hold z->mutex. |
| 248 | */ |
| 249 | static int zbd_reset_zone(struct thread_data *td, struct fio_file *f, |
| 250 | struct fio_zone_info *z) |
| 251 | { |
| 252 | uint64_t offset = z->start; |
| 253 | uint64_t length = (z+1)->start - offset; |
| 254 | uint64_t data_in_zone = z->wp - z->start; |
| 255 | int ret = 0; |
| 256 | |
| 257 | if (!data_in_zone) |
| 258 | return 0; |
| 259 | |
| 260 | assert(is_valid_offset(f, offset + length - 1)); |
| 261 | |
| 262 | dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", |
| 263 | f->file_name, zbd_zone_idx(f, z)); |
| 264 | |
| 265 | switch (f->zbd_info->model) { |
| 266 | case ZBD_HOST_AWARE: |
| 267 | case ZBD_HOST_MANAGED: |
| 268 | ret = zbd_reset_wp(td, f, offset, length); |
| 269 | if (ret < 0) |
| 270 | return ret; |
| 271 | break; |
| 272 | default: |
| 273 | break; |
| 274 | } |
| 275 | |
| 276 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 277 | f->zbd_info->sectors_with_data -= data_in_zone; |
| 278 | f->zbd_info->wp_sectors_with_data -= data_in_zone; |
| 279 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 280 | |
| 281 | z->wp = z->start; |
| 282 | z->verify_block = 0; |
| 283 | |
| 284 | td->ts.nr_zone_resets++; |
| 285 | |
| 286 | return ret; |
| 287 | } |
| 288 | |
| 289 | /** |
| 290 | * zbd_close_zone - Remove a zone from the open zones array. |
| 291 | * @td: FIO thread data. |
| 292 | * @f: FIO file associated with the disk for which to reset a write pointer. |
| 293 | * @zone_idx: Index of the zone to remove. |
| 294 | * |
| 295 | * The caller must hold f->zbd_info->mutex. |
| 296 | */ |
| 297 | static void zbd_close_zone(struct thread_data *td, const struct fio_file *f, |
| 298 | struct fio_zone_info *z) |
| 299 | { |
| 300 | uint32_t ozi; |
| 301 | |
| 302 | if (!z->open) |
| 303 | return; |
| 304 | |
| 305 | for (ozi = 0; ozi < f->zbd_info->num_open_zones; ozi++) { |
| 306 | if (zbd_get_zone(f, f->zbd_info->open_zones[ozi]) == z) |
| 307 | break; |
| 308 | } |
| 309 | if (ozi == f->zbd_info->num_open_zones) |
| 310 | return; |
| 311 | |
| 312 | dprint(FD_ZBD, "%s: closing zone %u\n", |
| 313 | f->file_name, zbd_zone_idx(f, z)); |
| 314 | |
| 315 | memmove(f->zbd_info->open_zones + ozi, |
| 316 | f->zbd_info->open_zones + ozi + 1, |
| 317 | (ZBD_MAX_OPEN_ZONES - (ozi + 1)) * |
| 318 | sizeof(f->zbd_info->open_zones[0])); |
| 319 | |
| 320 | f->zbd_info->num_open_zones--; |
| 321 | td->num_open_zones--; |
| 322 | z->open = 0; |
| 323 | } |
| 324 | |
| 325 | /** |
| 326 | * zbd_reset_zones - Reset a range of zones. |
| 327 | * @td: fio thread data. |
| 328 | * @f: fio file for which to reset zones |
| 329 | * @zb: first zone to reset. |
| 330 | * @ze: first zone not to reset. |
| 331 | * |
| 332 | * Returns 0 upon success and 1 upon failure. |
| 333 | */ |
| 334 | static int zbd_reset_zones(struct thread_data *td, struct fio_file *f, |
| 335 | struct fio_zone_info *const zb, |
| 336 | struct fio_zone_info *const ze) |
| 337 | { |
| 338 | struct fio_zone_info *z; |
| 339 | const uint64_t min_bs = td->o.min_bs[DDIR_WRITE]; |
| 340 | int res = 0; |
| 341 | |
| 342 | assert(min_bs); |
| 343 | |
| 344 | dprint(FD_ZBD, "%s: examining zones %u .. %u\n", |
| 345 | f->file_name, zbd_zone_idx(f, zb), zbd_zone_idx(f, ze)); |
| 346 | |
| 347 | for (z = zb; z < ze; z++) { |
| 348 | if (!z->has_wp) |
| 349 | continue; |
| 350 | |
| 351 | zone_lock(td, f, z); |
| 352 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 353 | zbd_close_zone(td, f, z); |
| 354 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 355 | |
| 356 | if (z->wp != z->start) { |
| 357 | dprint(FD_ZBD, "%s: resetting zone %u\n", |
| 358 | f->file_name, zbd_zone_idx(f, z)); |
| 359 | if (zbd_reset_zone(td, f, z) < 0) |
| 360 | res = 1; |
| 361 | } |
| 362 | |
| 363 | zone_unlock(z); |
| 364 | } |
| 365 | |
| 366 | return res; |
| 367 | } |
| 368 | |
| 369 | /** |
| 370 | * zbd_get_max_open_zones - Get the maximum number of open zones |
| 371 | * @td: FIO thread data |
| 372 | * @f: FIO file for which to get max open zones |
| 373 | * @max_open_zones: Upon success, result will be stored here. |
| 374 | * |
| 375 | * A @max_open_zones value set to zero means no limit. |
| 376 | * |
| 377 | * Returns 0 upon success and a negative error code upon failure. |
| 378 | */ |
| 379 | static int zbd_get_max_open_zones(struct thread_data *td, struct fio_file *f, |
| 380 | unsigned int *max_open_zones) |
| 381 | { |
| 382 | int ret; |
| 383 | |
| 384 | if (td->io_ops && td->io_ops->get_max_open_zones) |
| 385 | ret = td->io_ops->get_max_open_zones(td, f, max_open_zones); |
| 386 | else |
| 387 | ret = blkzoned_get_max_open_zones(td, f, max_open_zones); |
| 388 | if (ret < 0) { |
| 389 | td_verror(td, errno, "get max open zones failed"); |
| 390 | log_err("%s: get max open zones failed (%d).\n", |
| 391 | f->file_name, errno); |
| 392 | } |
| 393 | |
| 394 | return ret; |
| 395 | } |
| 396 | |
| 397 | /** |
| 398 | * zbd_open_zone - Add a zone to the array of open zones. |
| 399 | * @td: fio thread data. |
| 400 | * @f: fio file that has the open zones to add. |
| 401 | * @zone_idx: Index of the zone to add. |
| 402 | * |
| 403 | * Open a ZBD zone if it is not already open. Returns true if either the zone |
| 404 | * was already open or if the zone was successfully added to the array of open |
| 405 | * zones without exceeding the maximum number of open zones. Returns false if |
| 406 | * the zone was not already open and opening the zone would cause the zone limit |
| 407 | * to be exceeded. |
| 408 | */ |
| 409 | static bool zbd_open_zone(struct thread_data *td, const struct fio_file *f, |
| 410 | struct fio_zone_info *z) |
| 411 | { |
| 412 | const uint64_t min_bs = td->o.min_bs[DDIR_WRITE]; |
| 413 | struct zoned_block_device_info *zbdi = f->zbd_info; |
| 414 | uint32_t zone_idx = zbd_zone_idx(f, z); |
| 415 | bool res = true; |
| 416 | |
| 417 | if (z->cond == ZBD_ZONE_COND_OFFLINE) |
| 418 | return false; |
| 419 | |
| 420 | /* |
| 421 | * Skip full zones with data verification enabled because resetting a |
| 422 | * zone causes data loss and hence causes verification to fail. |
| 423 | */ |
| 424 | if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs)) |
| 425 | return false; |
| 426 | |
| 427 | /* |
| 428 | * zbdi->max_open_zones == 0 means that there is no limit on the maximum |
| 429 | * number of open zones. In this case, do no track open zones in |
| 430 | * zbdi->open_zones array. |
| 431 | */ |
| 432 | if (!zbdi->max_open_zones) |
| 433 | return true; |
| 434 | |
| 435 | pthread_mutex_lock(&zbdi->mutex); |
| 436 | |
| 437 | if (z->open) { |
| 438 | /* |
| 439 | * If the zone is going to be completely filled by writes |
| 440 | * already in-flight, handle it as a full zone instead of an |
| 441 | * open zone. |
| 442 | */ |
| 443 | if (z->wp >= zbd_zone_capacity_end(z)) |
| 444 | res = false; |
| 445 | goto out; |
| 446 | } |
| 447 | |
| 448 | res = false; |
| 449 | /* Zero means no limit */ |
| 450 | if (td->o.job_max_open_zones > 0 && |
| 451 | td->num_open_zones >= td->o.job_max_open_zones) |
| 452 | goto out; |
| 453 | if (zbdi->num_open_zones >= zbdi->max_open_zones) |
| 454 | goto out; |
| 455 | |
| 456 | dprint(FD_ZBD, "%s: opening zone %u\n", |
| 457 | f->file_name, zone_idx); |
| 458 | |
| 459 | zbdi->open_zones[zbdi->num_open_zones++] = zone_idx; |
| 460 | td->num_open_zones++; |
| 461 | z->open = 1; |
| 462 | res = true; |
| 463 | |
| 464 | out: |
| 465 | pthread_mutex_unlock(&zbdi->mutex); |
| 466 | return res; |
| 467 | } |
| 468 | |
| 469 | /* Verify whether direct I/O is used for all host-managed zoned block drives. */ |
| 470 | static bool zbd_using_direct_io(void) |
| 471 | { |
| 472 | struct thread_data *td; |
| 473 | struct fio_file *f; |
| 474 | int i, j; |
| 475 | |
| 476 | for_each_td(td, i) { |
| 477 | if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE)) |
| 478 | continue; |
| 479 | for_each_file(td, f, j) { |
| 480 | if (f->zbd_info && f->filetype == FIO_TYPE_BLOCK && |
| 481 | f->zbd_info->model == ZBD_HOST_MANAGED) |
| 482 | return false; |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | return true; |
| 487 | } |
| 488 | |
| 489 | /* Whether or not the I/O range for f includes one or more sequential zones */ |
| 490 | static bool zbd_is_seq_job(struct fio_file *f) |
| 491 | { |
| 492 | uint32_t zone_idx, zone_idx_b, zone_idx_e; |
| 493 | |
| 494 | assert(f->zbd_info); |
| 495 | |
| 496 | if (f->io_size == 0) |
| 497 | return false; |
| 498 | |
| 499 | zone_idx_b = zbd_offset_to_zone_idx(f, f->file_offset); |
| 500 | zone_idx_e = |
| 501 | zbd_offset_to_zone_idx(f, f->file_offset + f->io_size - 1); |
| 502 | for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++) |
| 503 | if (zbd_get_zone(f, zone_idx)->has_wp) |
| 504 | return true; |
| 505 | |
| 506 | return false; |
| 507 | } |
| 508 | |
| 509 | /* |
| 510 | * Verify whether the file offset and size parameters are aligned with zone |
| 511 | * boundaries. If the file offset is not aligned, align it down to the start of |
| 512 | * the zone containing the start offset and align up the file io_size parameter. |
| 513 | */ |
| 514 | static bool zbd_zone_align_file_sizes(struct thread_data *td, |
| 515 | struct fio_file *f) |
| 516 | { |
| 517 | const struct fio_zone_info *z; |
| 518 | uint64_t new_offset, new_end; |
| 519 | |
| 520 | if (!f->zbd_info) |
| 521 | return true; |
| 522 | if (f->file_offset >= f->real_file_size) |
| 523 | return true; |
| 524 | if (!zbd_is_seq_job(f)) |
| 525 | return true; |
| 526 | |
| 527 | if (!td->o.zone_size) { |
| 528 | td->o.zone_size = f->zbd_info->zone_size; |
| 529 | if (!td->o.zone_size) { |
| 530 | log_err("%s: invalid 0 zone size\n", |
| 531 | f->file_name); |
| 532 | return false; |
| 533 | } |
| 534 | } else if (td->o.zone_size != f->zbd_info->zone_size) { |
| 535 | log_err("%s: zonesize %llu does not match the device zone size %"PRIu64".\n", |
| 536 | f->file_name, td->o.zone_size, |
| 537 | f->zbd_info->zone_size); |
| 538 | return false; |
| 539 | } |
| 540 | |
| 541 | if (td->o.zone_skip % td->o.zone_size) { |
| 542 | log_err("%s: zoneskip %llu is not a multiple of the device zone size %llu.\n", |
| 543 | f->file_name, td->o.zone_skip, |
| 544 | td->o.zone_size); |
| 545 | return false; |
| 546 | } |
| 547 | |
| 548 | z = zbd_offset_to_zone(f, f->file_offset); |
| 549 | if ((f->file_offset != z->start) && |
| 550 | (td->o.td_ddir != TD_DDIR_READ)) { |
| 551 | new_offset = zbd_zone_end(z); |
| 552 | if (new_offset >= f->file_offset + f->io_size) { |
| 553 | log_info("%s: io_size must be at least one zone\n", |
| 554 | f->file_name); |
| 555 | return false; |
| 556 | } |
| 557 | log_info("%s: rounded up offset from %"PRIu64" to %"PRIu64"\n", |
| 558 | f->file_name, f->file_offset, |
| 559 | new_offset); |
| 560 | f->io_size -= (new_offset - f->file_offset); |
| 561 | f->file_offset = new_offset; |
| 562 | } |
| 563 | |
| 564 | z = zbd_offset_to_zone(f, f->file_offset + f->io_size); |
| 565 | new_end = z->start; |
| 566 | if ((td->o.td_ddir != TD_DDIR_READ) && |
| 567 | (f->file_offset + f->io_size != new_end)) { |
| 568 | if (new_end <= f->file_offset) { |
| 569 | log_info("%s: io_size must be at least one zone\n", |
| 570 | f->file_name); |
| 571 | return false; |
| 572 | } |
| 573 | log_info("%s: rounded down io_size from %"PRIu64" to %"PRIu64"\n", |
| 574 | f->file_name, f->io_size, |
| 575 | new_end - f->file_offset); |
| 576 | f->io_size = new_end - f->file_offset; |
| 577 | } |
| 578 | |
| 579 | return true; |
| 580 | } |
| 581 | |
| 582 | /* |
| 583 | * Verify whether offset and size parameters are aligned with zone boundaries. |
| 584 | */ |
| 585 | static bool zbd_verify_sizes(void) |
| 586 | { |
| 587 | struct thread_data *td; |
| 588 | struct fio_file *f; |
| 589 | int i, j; |
| 590 | |
| 591 | for_each_td(td, i) { |
| 592 | for_each_file(td, f, j) { |
| 593 | if (!zbd_zone_align_file_sizes(td, f)) |
| 594 | return false; |
| 595 | } |
| 596 | } |
| 597 | |
| 598 | return true; |
| 599 | } |
| 600 | |
| 601 | static bool zbd_verify_bs(void) |
| 602 | { |
| 603 | struct thread_data *td; |
| 604 | struct fio_file *f; |
| 605 | int i, j, k; |
| 606 | |
| 607 | for_each_td(td, i) { |
| 608 | if (td_trim(td) && |
| 609 | (td->o.min_bs[DDIR_TRIM] != td->o.max_bs[DDIR_TRIM] || |
| 610 | td->o.bssplit_nr[DDIR_TRIM])) { |
| 611 | log_info("bsrange and bssplit are not allowed for trim with zonemode=zbd\n"); |
| 612 | return false; |
| 613 | } |
| 614 | for_each_file(td, f, j) { |
| 615 | uint64_t zone_size; |
| 616 | |
| 617 | if (!f->zbd_info) |
| 618 | continue; |
| 619 | |
| 620 | zone_size = f->zbd_info->zone_size; |
| 621 | if (td_trim(td) && td->o.bs[DDIR_TRIM] != zone_size) { |
| 622 | log_info("%s: trim block size %llu is not the zone size %"PRIu64"\n", |
| 623 | f->file_name, td->o.bs[DDIR_TRIM], |
| 624 | zone_size); |
| 625 | return false; |
| 626 | } |
| 627 | for (k = 0; k < FIO_ARRAY_SIZE(td->o.bs); k++) { |
| 628 | if (td->o.verify != VERIFY_NONE && |
| 629 | zone_size % td->o.bs[k] != 0) { |
| 630 | log_info("%s: block size %llu is not a divisor of the zone size %"PRIu64"\n", |
| 631 | f->file_name, td->o.bs[k], |
| 632 | zone_size); |
| 633 | return false; |
| 634 | } |
| 635 | } |
| 636 | } |
| 637 | } |
| 638 | return true; |
| 639 | } |
| 640 | |
| 641 | static int ilog2(uint64_t i) |
| 642 | { |
| 643 | int log = -1; |
| 644 | |
| 645 | while (i) { |
| 646 | i >>= 1; |
| 647 | log++; |
| 648 | } |
| 649 | return log; |
| 650 | } |
| 651 | |
| 652 | /* |
| 653 | * Initialize f->zbd_info for devices that are not zoned block devices. This |
| 654 | * allows to execute a ZBD workload against a non-ZBD device. |
| 655 | */ |
| 656 | static int init_zone_info(struct thread_data *td, struct fio_file *f) |
| 657 | { |
| 658 | uint32_t nr_zones; |
| 659 | struct fio_zone_info *p; |
| 660 | uint64_t zone_size = td->o.zone_size; |
| 661 | uint64_t zone_capacity = td->o.zone_capacity; |
| 662 | struct zoned_block_device_info *zbd_info = NULL; |
| 663 | int i; |
| 664 | |
| 665 | if (zone_size == 0) { |
| 666 | log_err("%s: Specifying the zone size is mandatory for regular file/block device with --zonemode=zbd\n\n", |
| 667 | f->file_name); |
| 668 | return 1; |
| 669 | } |
| 670 | |
| 671 | if (zone_size < 512) { |
| 672 | log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n", |
| 673 | f->file_name); |
| 674 | return 1; |
| 675 | } |
| 676 | |
| 677 | if (zone_capacity == 0) |
| 678 | zone_capacity = zone_size; |
| 679 | |
| 680 | if (zone_capacity > zone_size) { |
| 681 | log_err("%s: job parameter zonecapacity %llu is larger than zone size %llu\n", |
| 682 | f->file_name, td->o.zone_capacity, td->o.zone_size); |
| 683 | return 1; |
| 684 | } |
| 685 | |
| 686 | if (f->real_file_size < zone_size) { |
| 687 | log_err("%s: file/device size %"PRIu64" is smaller than zone size %"PRIu64"\n", |
| 688 | f->file_name, f->real_file_size, zone_size); |
| 689 | return -EINVAL; |
| 690 | } |
| 691 | |
| 692 | nr_zones = (f->real_file_size + zone_size - 1) / zone_size; |
| 693 | zbd_info = scalloc(1, sizeof(*zbd_info) + |
| 694 | (nr_zones + 1) * sizeof(zbd_info->zone_info[0])); |
| 695 | if (!zbd_info) |
| 696 | return -ENOMEM; |
| 697 | |
| 698 | mutex_init_pshared(&zbd_info->mutex); |
| 699 | zbd_info->refcount = 1; |
| 700 | p = &zbd_info->zone_info[0]; |
| 701 | for (i = 0; i < nr_zones; i++, p++) { |
| 702 | mutex_init_pshared_with_type(&p->mutex, |
| 703 | PTHREAD_MUTEX_RECURSIVE); |
| 704 | p->start = i * zone_size; |
| 705 | p->wp = p->start; |
| 706 | p->type = ZBD_ZONE_TYPE_SWR; |
| 707 | p->cond = ZBD_ZONE_COND_EMPTY; |
| 708 | p->capacity = zone_capacity; |
| 709 | p->has_wp = 1; |
| 710 | } |
| 711 | /* a sentinel */ |
| 712 | p->start = nr_zones * zone_size; |
| 713 | |
| 714 | f->zbd_info = zbd_info; |
| 715 | f->zbd_info->zone_size = zone_size; |
| 716 | f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ? |
| 717 | ilog2(zone_size) : 0; |
| 718 | f->zbd_info->nr_zones = nr_zones; |
| 719 | return 0; |
| 720 | } |
| 721 | |
| 722 | /* |
| 723 | * Maximum number of zones to report in one operation. |
| 724 | */ |
| 725 | #define ZBD_REPORT_MAX_ZONES 8192U |
| 726 | |
| 727 | /* |
| 728 | * Parse the device zone report and store it in f->zbd_info. Must be called |
| 729 | * only for devices that are zoned, namely those with a model != ZBD_NONE. |
| 730 | */ |
| 731 | static int parse_zone_info(struct thread_data *td, struct fio_file *f) |
| 732 | { |
| 733 | int nr_zones, nrz; |
| 734 | struct zbd_zone *zones, *z; |
| 735 | struct fio_zone_info *p; |
| 736 | uint64_t zone_size, offset; |
| 737 | struct zoned_block_device_info *zbd_info = NULL; |
| 738 | int i, j, ret = -ENOMEM; |
| 739 | |
| 740 | zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone)); |
| 741 | if (!zones) |
| 742 | goto out; |
| 743 | |
| 744 | nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES); |
| 745 | if (nrz < 0) { |
| 746 | ret = nrz; |
| 747 | log_info("fio: report zones (offset 0) failed for %s (%d).\n", |
| 748 | f->file_name, -ret); |
| 749 | goto out; |
| 750 | } |
| 751 | |
| 752 | zone_size = zones[0].len; |
| 753 | nr_zones = (f->real_file_size + zone_size - 1) / zone_size; |
| 754 | |
| 755 | if (td->o.zone_size == 0) { |
| 756 | td->o.zone_size = zone_size; |
| 757 | } else if (td->o.zone_size != zone_size) { |
| 758 | log_err("fio: %s job parameter zonesize %llu does not match disk zone size %"PRIu64".\n", |
| 759 | f->file_name, td->o.zone_size, zone_size); |
| 760 | ret = -EINVAL; |
| 761 | goto out; |
| 762 | } |
| 763 | |
| 764 | dprint(FD_ZBD, "Device %s has %d zones of size %"PRIu64" KB\n", |
| 765 | f->file_name, nr_zones, zone_size / 1024); |
| 766 | |
| 767 | zbd_info = scalloc(1, sizeof(*zbd_info) + |
| 768 | (nr_zones + 1) * sizeof(zbd_info->zone_info[0])); |
| 769 | if (!zbd_info) |
| 770 | goto out; |
| 771 | mutex_init_pshared(&zbd_info->mutex); |
| 772 | zbd_info->refcount = 1; |
| 773 | p = &zbd_info->zone_info[0]; |
| 774 | for (offset = 0, j = 0; j < nr_zones;) { |
| 775 | z = &zones[0]; |
| 776 | for (i = 0; i < nrz; i++, j++, z++, p++) { |
| 777 | mutex_init_pshared_with_type(&p->mutex, |
| 778 | PTHREAD_MUTEX_RECURSIVE); |
| 779 | p->start = z->start; |
| 780 | p->capacity = z->capacity; |
| 781 | |
| 782 | switch (z->cond) { |
| 783 | case ZBD_ZONE_COND_NOT_WP: |
| 784 | case ZBD_ZONE_COND_FULL: |
| 785 | p->wp = p->start + p->capacity; |
| 786 | break; |
| 787 | default: |
| 788 | assert(z->start <= z->wp); |
| 789 | assert(z->wp <= z->start + zone_size); |
| 790 | p->wp = z->wp; |
| 791 | break; |
| 792 | } |
| 793 | |
| 794 | switch (z->type) { |
| 795 | case ZBD_ZONE_TYPE_SWR: |
| 796 | p->has_wp = 1; |
| 797 | break; |
| 798 | default: |
| 799 | p->has_wp = 0; |
| 800 | } |
| 801 | p->type = z->type; |
| 802 | p->cond = z->cond; |
| 803 | |
| 804 | if (j > 0 && p->start != p[-1].start + zone_size) { |
| 805 | log_info("%s: invalid zone data\n", |
| 806 | f->file_name); |
| 807 | ret = -EINVAL; |
| 808 | goto out; |
| 809 | } |
| 810 | } |
| 811 | z--; |
| 812 | offset = z->start + z->len; |
| 813 | if (j >= nr_zones) |
| 814 | break; |
| 815 | |
| 816 | nrz = zbd_report_zones(td, f, offset, zones, |
| 817 | min((uint32_t)(nr_zones - j), |
| 818 | ZBD_REPORT_MAX_ZONES)); |
| 819 | if (nrz < 0) { |
| 820 | ret = nrz; |
| 821 | log_info("fio: report zones (offset %"PRIu64") failed for %s (%d).\n", |
| 822 | offset, f->file_name, -ret); |
| 823 | goto out; |
| 824 | } |
| 825 | } |
| 826 | |
| 827 | /* a sentinel */ |
| 828 | zbd_info->zone_info[nr_zones].start = offset; |
| 829 | |
| 830 | f->zbd_info = zbd_info; |
| 831 | f->zbd_info->zone_size = zone_size; |
| 832 | f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ? |
| 833 | ilog2(zone_size) : 0; |
| 834 | f->zbd_info->nr_zones = nr_zones; |
| 835 | zbd_info = NULL; |
| 836 | ret = 0; |
| 837 | |
| 838 | out: |
| 839 | sfree(zbd_info); |
| 840 | free(zones); |
| 841 | return ret; |
| 842 | } |
| 843 | |
| 844 | static int zbd_set_max_open_zones(struct thread_data *td, struct fio_file *f) |
| 845 | { |
| 846 | struct zoned_block_device_info *zbd = f->zbd_info; |
| 847 | unsigned int max_open_zones; |
| 848 | int ret; |
| 849 | |
| 850 | if (zbd->model != ZBD_HOST_MANAGED || td->o.ignore_zone_limits) { |
| 851 | /* Only host-managed devices have a max open limit */ |
| 852 | zbd->max_open_zones = td->o.max_open_zones; |
| 853 | goto out; |
| 854 | } |
| 855 | |
| 856 | /* If host-managed, get the max open limit */ |
| 857 | ret = zbd_get_max_open_zones(td, f, &max_open_zones); |
| 858 | if (ret) |
| 859 | return ret; |
| 860 | |
| 861 | if (!max_open_zones) { |
| 862 | /* No device limit */ |
| 863 | zbd->max_open_zones = td->o.max_open_zones; |
| 864 | } else if (!td->o.max_open_zones) { |
| 865 | /* No user limit. Set limit to device limit */ |
| 866 | zbd->max_open_zones = max_open_zones; |
| 867 | } else if (td->o.max_open_zones <= max_open_zones) { |
| 868 | /* Both user limit and dev limit. User limit not too large */ |
| 869 | zbd->max_open_zones = td->o.max_open_zones; |
| 870 | } else { |
| 871 | /* Both user limit and dev limit. User limit too large */ |
| 872 | td_verror(td, EINVAL, |
| 873 | "Specified --max_open_zones is too large"); |
| 874 | log_err("Specified --max_open_zones (%d) is larger than max (%u)\n", |
| 875 | td->o.max_open_zones, max_open_zones); |
| 876 | return -EINVAL; |
| 877 | } |
| 878 | |
| 879 | out: |
| 880 | /* Ensure that the limit is not larger than FIO's internal limit */ |
| 881 | if (zbd->max_open_zones > ZBD_MAX_OPEN_ZONES) { |
| 882 | td_verror(td, EINVAL, "'max_open_zones' value is too large"); |
| 883 | log_err("'max_open_zones' value is larger than %u\n", |
| 884 | ZBD_MAX_OPEN_ZONES); |
| 885 | return -EINVAL; |
| 886 | } |
| 887 | |
| 888 | dprint(FD_ZBD, "%s: using max open zones limit: %"PRIu32"\n", |
| 889 | f->file_name, zbd->max_open_zones); |
| 890 | |
| 891 | return 0; |
| 892 | } |
| 893 | |
| 894 | /* |
| 895 | * Allocate zone information and store it into f->zbd_info if zonemode=zbd. |
| 896 | * |
| 897 | * Returns 0 upon success and a negative error code upon failure. |
| 898 | */ |
| 899 | static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f) |
| 900 | { |
| 901 | enum zbd_zoned_model zbd_model; |
| 902 | int ret; |
| 903 | |
| 904 | assert(td->o.zone_mode == ZONE_MODE_ZBD); |
| 905 | |
| 906 | ret = zbd_get_zoned_model(td, f, &zbd_model); |
| 907 | if (ret) |
| 908 | return ret; |
| 909 | |
| 910 | switch (zbd_model) { |
| 911 | case ZBD_HOST_AWARE: |
| 912 | case ZBD_HOST_MANAGED: |
| 913 | ret = parse_zone_info(td, f); |
| 914 | if (ret) |
| 915 | return ret; |
| 916 | break; |
| 917 | case ZBD_NONE: |
| 918 | ret = init_zone_info(td, f); |
| 919 | if (ret) |
| 920 | return ret; |
| 921 | break; |
| 922 | default: |
| 923 | td_verror(td, EINVAL, "Unsupported zoned model"); |
| 924 | log_err("Unsupported zoned model\n"); |
| 925 | return -EINVAL; |
| 926 | } |
| 927 | |
| 928 | assert(f->zbd_info); |
| 929 | f->zbd_info->model = zbd_model; |
| 930 | |
| 931 | ret = zbd_set_max_open_zones(td, f); |
| 932 | if (ret) { |
| 933 | zbd_free_zone_info(f); |
| 934 | return ret; |
| 935 | } |
| 936 | |
| 937 | return 0; |
| 938 | } |
| 939 | |
| 940 | void zbd_free_zone_info(struct fio_file *f) |
| 941 | { |
| 942 | uint32_t refcount; |
| 943 | |
| 944 | assert(f->zbd_info); |
| 945 | |
| 946 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 947 | refcount = --f->zbd_info->refcount; |
| 948 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 949 | |
| 950 | assert((int32_t)refcount >= 0); |
| 951 | if (refcount == 0) |
| 952 | sfree(f->zbd_info); |
| 953 | f->zbd_info = NULL; |
| 954 | } |
| 955 | |
| 956 | /* |
| 957 | * Initialize f->zbd_info. |
| 958 | * |
| 959 | * Returns 0 upon success and a negative error code upon failure. |
| 960 | * |
| 961 | * Note: this function can only work correctly if it is called before the first |
| 962 | * fio fork() call. |
| 963 | */ |
| 964 | static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file) |
| 965 | { |
| 966 | struct thread_data *td2; |
| 967 | struct fio_file *f2; |
| 968 | int i, j, ret; |
| 969 | |
| 970 | for_each_td(td2, i) { |
| 971 | for_each_file(td2, f2, j) { |
| 972 | if (td2 == td && f2 == file) |
| 973 | continue; |
| 974 | if (!f2->zbd_info || |
| 975 | strcmp(f2->file_name, file->file_name) != 0) |
| 976 | continue; |
| 977 | file->zbd_info = f2->zbd_info; |
| 978 | file->zbd_info->refcount++; |
| 979 | return 0; |
| 980 | } |
| 981 | } |
| 982 | |
| 983 | ret = zbd_create_zone_info(td, file); |
| 984 | if (ret < 0) |
| 985 | td_verror(td, -ret, "zbd_create_zone_info() failed"); |
| 986 | |
| 987 | return ret; |
| 988 | } |
| 989 | |
| 990 | int zbd_init_files(struct thread_data *td) |
| 991 | { |
| 992 | struct fio_file *f; |
| 993 | int i; |
| 994 | |
| 995 | for_each_file(td, f, i) { |
| 996 | if (zbd_init_zone_info(td, f)) |
| 997 | return 1; |
| 998 | } |
| 999 | |
| 1000 | return 0; |
| 1001 | } |
| 1002 | |
| 1003 | void zbd_recalc_options_with_zone_granularity(struct thread_data *td) |
| 1004 | { |
| 1005 | struct fio_file *f; |
| 1006 | int i; |
| 1007 | |
| 1008 | for_each_file(td, f, i) { |
| 1009 | struct zoned_block_device_info *zbd = f->zbd_info; |
| 1010 | uint64_t zone_size; |
| 1011 | |
| 1012 | /* zonemode=strided doesn't get per-file zone size. */ |
| 1013 | zone_size = zbd ? zbd->zone_size : td->o.zone_size; |
| 1014 | if (zone_size == 0) |
| 1015 | continue; |
| 1016 | |
| 1017 | if (td->o.size_nz > 0) |
| 1018 | td->o.size = td->o.size_nz * zone_size; |
| 1019 | if (td->o.io_size_nz > 0) |
| 1020 | td->o.io_size = td->o.io_size_nz * zone_size; |
| 1021 | if (td->o.start_offset_nz > 0) |
| 1022 | td->o.start_offset = td->o.start_offset_nz * zone_size; |
| 1023 | if (td->o.offset_increment_nz > 0) |
| 1024 | td->o.offset_increment = |
| 1025 | td->o.offset_increment_nz * zone_size; |
| 1026 | if (td->o.zone_skip_nz > 0) |
| 1027 | td->o.zone_skip = td->o.zone_skip_nz * zone_size; |
| 1028 | } |
| 1029 | } |
| 1030 | |
| 1031 | int zbd_setup_files(struct thread_data *td) |
| 1032 | { |
| 1033 | struct fio_file *f; |
| 1034 | int i; |
| 1035 | |
| 1036 | if (!zbd_using_direct_io()) { |
| 1037 | log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n"); |
| 1038 | return 1; |
| 1039 | } |
| 1040 | |
| 1041 | if (!zbd_verify_sizes()) |
| 1042 | return 1; |
| 1043 | |
| 1044 | if (!zbd_verify_bs()) |
| 1045 | return 1; |
| 1046 | |
| 1047 | for_each_file(td, f, i) { |
| 1048 | struct zoned_block_device_info *zbd = f->zbd_info; |
| 1049 | struct fio_zone_info *z; |
| 1050 | int zi; |
| 1051 | |
| 1052 | assert(zbd); |
| 1053 | |
| 1054 | f->min_zone = zbd_offset_to_zone_idx(f, f->file_offset); |
| 1055 | f->max_zone = |
| 1056 | zbd_offset_to_zone_idx(f, f->file_offset + f->io_size); |
| 1057 | |
| 1058 | /* |
| 1059 | * When all zones in the I/O range are conventional, io_size |
| 1060 | * can be smaller than zone size, making min_zone the same |
| 1061 | * as max_zone. This is why the assert below needs to be made |
| 1062 | * conditional. |
| 1063 | */ |
| 1064 | if (zbd_is_seq_job(f)) |
| 1065 | assert(f->min_zone < f->max_zone); |
| 1066 | |
| 1067 | if (td->o.max_open_zones > 0 && |
| 1068 | zbd->max_open_zones != td->o.max_open_zones) { |
| 1069 | log_err("Different 'max_open_zones' values\n"); |
| 1070 | return 1; |
| 1071 | } |
| 1072 | |
| 1073 | /* |
| 1074 | * The per job max open zones limit cannot be used without a |
| 1075 | * global max open zones limit. (As the tracking of open zones |
| 1076 | * is disabled when there is no global max open zones limit.) |
| 1077 | */ |
| 1078 | if (td->o.job_max_open_zones && !zbd->max_open_zones) { |
| 1079 | log_err("'job_max_open_zones' cannot be used without a global open zones limit\n"); |
| 1080 | return 1; |
| 1081 | } |
| 1082 | |
| 1083 | /* |
| 1084 | * zbd->max_open_zones is the global limit shared for all jobs |
| 1085 | * that target the same zoned block device. Force sync the per |
| 1086 | * thread global limit with the actual global limit. (The real |
| 1087 | * per thread/job limit is stored in td->o.job_max_open_zones). |
| 1088 | */ |
| 1089 | td->o.max_open_zones = zbd->max_open_zones; |
| 1090 | |
| 1091 | for (zi = f->min_zone; zi < f->max_zone; zi++) { |
| 1092 | z = &zbd->zone_info[zi]; |
| 1093 | if (z->cond != ZBD_ZONE_COND_IMP_OPEN && |
| 1094 | z->cond != ZBD_ZONE_COND_EXP_OPEN) |
| 1095 | continue; |
| 1096 | if (zbd_open_zone(td, f, z)) |
| 1097 | continue; |
| 1098 | /* |
| 1099 | * If the number of open zones exceeds specified limits, |
| 1100 | * reset all extra open zones. |
| 1101 | */ |
| 1102 | if (zbd_reset_zone(td, f, z) < 0) { |
| 1103 | log_err("Failed to reest zone %d\n", zi); |
| 1104 | return 1; |
| 1105 | } |
| 1106 | } |
| 1107 | } |
| 1108 | |
| 1109 | return 0; |
| 1110 | } |
| 1111 | |
| 1112 | /* |
| 1113 | * Reset zbd_info.write_cnt, the counter that counts down towards the next |
| 1114 | * zone reset. |
| 1115 | */ |
| 1116 | static void _zbd_reset_write_cnt(const struct thread_data *td, |
| 1117 | const struct fio_file *f) |
| 1118 | { |
| 1119 | assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1); |
| 1120 | |
| 1121 | f->zbd_info->write_cnt = td->o.zrf.u.f ? |
| 1122 | min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX; |
| 1123 | } |
| 1124 | |
| 1125 | static void zbd_reset_write_cnt(const struct thread_data *td, |
| 1126 | const struct fio_file *f) |
| 1127 | { |
| 1128 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 1129 | _zbd_reset_write_cnt(td, f); |
| 1130 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 1131 | } |
| 1132 | |
| 1133 | static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td, |
| 1134 | const struct fio_file *f) |
| 1135 | { |
| 1136 | uint32_t write_cnt = 0; |
| 1137 | |
| 1138 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 1139 | assert(f->zbd_info->write_cnt); |
| 1140 | if (f->zbd_info->write_cnt) |
| 1141 | write_cnt = --f->zbd_info->write_cnt; |
| 1142 | if (write_cnt == 0) |
| 1143 | _zbd_reset_write_cnt(td, f); |
| 1144 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 1145 | |
| 1146 | return write_cnt == 0; |
| 1147 | } |
| 1148 | |
| 1149 | enum swd_action { |
| 1150 | CHECK_SWD, |
| 1151 | SET_SWD, |
| 1152 | }; |
| 1153 | |
| 1154 | /* Calculate the number of sectors with data (swd) and perform action 'a' */ |
| 1155 | static uint64_t zbd_process_swd(struct thread_data *td, |
| 1156 | const struct fio_file *f, enum swd_action a) |
| 1157 | { |
| 1158 | struct fio_zone_info *zb, *ze, *z; |
| 1159 | uint64_t swd = 0; |
| 1160 | uint64_t wp_swd = 0; |
| 1161 | |
| 1162 | zb = zbd_get_zone(f, f->min_zone); |
| 1163 | ze = zbd_get_zone(f, f->max_zone); |
| 1164 | for (z = zb; z < ze; z++) { |
| 1165 | if (z->has_wp) { |
| 1166 | zone_lock(td, f, z); |
| 1167 | wp_swd += z->wp - z->start; |
| 1168 | } |
| 1169 | swd += z->wp - z->start; |
| 1170 | } |
| 1171 | |
| 1172 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 1173 | switch (a) { |
| 1174 | case CHECK_SWD: |
| 1175 | assert(f->zbd_info->sectors_with_data == swd); |
| 1176 | assert(f->zbd_info->wp_sectors_with_data == wp_swd); |
| 1177 | break; |
| 1178 | case SET_SWD: |
| 1179 | f->zbd_info->sectors_with_data = swd; |
| 1180 | f->zbd_info->wp_sectors_with_data = wp_swd; |
| 1181 | break; |
| 1182 | } |
| 1183 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 1184 | |
| 1185 | for (z = zb; z < ze; z++) |
| 1186 | if (z->has_wp) |
| 1187 | zone_unlock(z); |
| 1188 | |
| 1189 | return swd; |
| 1190 | } |
| 1191 | |
| 1192 | /* |
| 1193 | * The swd check is useful for debugging but takes too much time to leave |
| 1194 | * it enabled all the time. Hence it is disabled by default. |
| 1195 | */ |
| 1196 | static const bool enable_check_swd = false; |
| 1197 | |
| 1198 | /* Check whether the values of zbd_info.*sectors_with_data are correct. */ |
| 1199 | static void zbd_check_swd(struct thread_data *td, const struct fio_file *f) |
| 1200 | { |
| 1201 | if (!enable_check_swd) |
| 1202 | return; |
| 1203 | |
| 1204 | zbd_process_swd(td, f, CHECK_SWD); |
| 1205 | } |
| 1206 | |
| 1207 | void zbd_file_reset(struct thread_data *td, struct fio_file *f) |
| 1208 | { |
| 1209 | struct fio_zone_info *zb, *ze; |
| 1210 | uint64_t swd; |
| 1211 | |
| 1212 | if (!f->zbd_info || !td_write(td)) |
| 1213 | return; |
| 1214 | |
| 1215 | zb = zbd_get_zone(f, f->min_zone); |
| 1216 | ze = zbd_get_zone(f, f->max_zone); |
| 1217 | swd = zbd_process_swd(td, f, SET_SWD); |
| 1218 | |
| 1219 | dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", |
| 1220 | __func__, f->file_name, swd); |
| 1221 | |
| 1222 | /* |
| 1223 | * If data verification is enabled reset the affected zones before |
| 1224 | * writing any data to avoid that a zone reset has to be issued while |
| 1225 | * writing data, which causes data loss. |
| 1226 | */ |
| 1227 | if (td->o.verify != VERIFY_NONE && td->runstate != TD_VERIFYING) |
| 1228 | zbd_reset_zones(td, f, zb, ze); |
| 1229 | zbd_reset_write_cnt(td, f); |
| 1230 | } |
| 1231 | |
| 1232 | /* Return random zone index for one of the open zones. */ |
| 1233 | static uint32_t pick_random_zone_idx(const struct fio_file *f, |
| 1234 | const struct io_u *io_u) |
| 1235 | { |
| 1236 | return (io_u->offset - f->file_offset) * |
| 1237 | f->zbd_info->num_open_zones / f->io_size; |
| 1238 | } |
| 1239 | |
| 1240 | static bool any_io_in_flight(void) |
| 1241 | { |
| 1242 | struct thread_data *td; |
| 1243 | int i; |
| 1244 | |
| 1245 | for_each_td(td, i) { |
| 1246 | if (td->io_u_in_flight) |
| 1247 | return true; |
| 1248 | } |
| 1249 | |
| 1250 | return false; |
| 1251 | } |
| 1252 | |
| 1253 | /* |
| 1254 | * Modify the offset of an I/O unit that does not refer to an open zone such |
| 1255 | * that it refers to an open zone. Close an open zone and open a new zone if |
| 1256 | * necessary. The open zone is searched across sequential zones. |
| 1257 | * This algorithm can only work correctly if all write pointers are |
| 1258 | * a multiple of the fio block size. The caller must neither hold z->mutex |
| 1259 | * nor f->zbd_info->mutex. Returns with z->mutex held upon success. |
| 1260 | */ |
| 1261 | static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td, |
| 1262 | struct io_u *io_u) |
| 1263 | { |
| 1264 | const uint64_t min_bs = td->o.min_bs[io_u->ddir]; |
| 1265 | struct fio_file *f = io_u->file; |
| 1266 | struct zoned_block_device_info *zbdi = f->zbd_info; |
| 1267 | struct fio_zone_info *z; |
| 1268 | unsigned int open_zone_idx = -1; |
| 1269 | uint32_t zone_idx, new_zone_idx; |
| 1270 | int i; |
| 1271 | bool wait_zone_close; |
| 1272 | bool in_flight; |
| 1273 | bool should_retry = true; |
| 1274 | |
| 1275 | assert(is_valid_offset(f, io_u->offset)); |
| 1276 | |
| 1277 | if (zbdi->max_open_zones || td->o.job_max_open_zones) { |
| 1278 | /* |
| 1279 | * This statement accesses zbdi->open_zones[] on purpose |
| 1280 | * without locking. |
| 1281 | */ |
| 1282 | zone_idx = zbdi->open_zones[pick_random_zone_idx(f, io_u)]; |
| 1283 | } else { |
| 1284 | zone_idx = zbd_offset_to_zone_idx(f, io_u->offset); |
| 1285 | } |
| 1286 | if (zone_idx < f->min_zone) |
| 1287 | zone_idx = f->min_zone; |
| 1288 | else if (zone_idx >= f->max_zone) |
| 1289 | zone_idx = f->max_zone - 1; |
| 1290 | |
| 1291 | dprint(FD_ZBD, |
| 1292 | "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n", |
| 1293 | __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen); |
| 1294 | |
| 1295 | /* |
| 1296 | * Since z->mutex is the outer lock and zbdi->mutex the inner |
| 1297 | * lock it can happen that the state of the zone with index zone_idx |
| 1298 | * has changed after 'z' has been assigned and before zbdi->mutex |
| 1299 | * has been obtained. Hence the loop. |
| 1300 | */ |
| 1301 | for (;;) { |
| 1302 | uint32_t tmp_idx; |
| 1303 | |
| 1304 | z = zbd_get_zone(f, zone_idx); |
| 1305 | if (z->has_wp) |
| 1306 | zone_lock(td, f, z); |
| 1307 | |
| 1308 | pthread_mutex_lock(&zbdi->mutex); |
| 1309 | |
| 1310 | if (z->has_wp) { |
| 1311 | if (z->cond != ZBD_ZONE_COND_OFFLINE && |
| 1312 | zbdi->max_open_zones == 0 && |
| 1313 | td->o.job_max_open_zones == 0) |
| 1314 | goto examine_zone; |
| 1315 | if (zbdi->num_open_zones == 0) { |
| 1316 | dprint(FD_ZBD, "%s(%s): no zones are open\n", |
| 1317 | __func__, f->file_name); |
| 1318 | goto open_other_zone; |
| 1319 | } |
| 1320 | } |
| 1321 | |
| 1322 | /* |
| 1323 | * List of opened zones is per-device, shared across all |
| 1324 | * threads. Start with quasi-random candidate zone. Ignore |
| 1325 | * zones which don't belong to thread's offset/size area. |
| 1326 | */ |
| 1327 | open_zone_idx = pick_random_zone_idx(f, io_u); |
| 1328 | assert(!open_zone_idx || |
| 1329 | open_zone_idx < zbdi->num_open_zones); |
| 1330 | tmp_idx = open_zone_idx; |
| 1331 | |
| 1332 | for (i = 0; i < zbdi->num_open_zones; i++) { |
| 1333 | uint32_t tmpz; |
| 1334 | |
| 1335 | if (tmp_idx >= zbdi->num_open_zones) |
| 1336 | tmp_idx = 0; |
| 1337 | tmpz = zbdi->open_zones[tmp_idx]; |
| 1338 | if (f->min_zone <= tmpz && tmpz < f->max_zone) { |
| 1339 | open_zone_idx = tmp_idx; |
| 1340 | goto found_candidate_zone; |
| 1341 | } |
| 1342 | |
| 1343 | tmp_idx++; |
| 1344 | } |
| 1345 | |
| 1346 | dprint(FD_ZBD, "%s(%s): no candidate zone\n", |
| 1347 | __func__, f->file_name); |
| 1348 | |
| 1349 | pthread_mutex_unlock(&zbdi->mutex); |
| 1350 | |
| 1351 | if (z->has_wp) |
| 1352 | zone_unlock(z); |
| 1353 | |
| 1354 | return NULL; |
| 1355 | |
| 1356 | found_candidate_zone: |
| 1357 | new_zone_idx = zbdi->open_zones[open_zone_idx]; |
| 1358 | if (new_zone_idx == zone_idx) |
| 1359 | break; |
| 1360 | zone_idx = new_zone_idx; |
| 1361 | |
| 1362 | pthread_mutex_unlock(&zbdi->mutex); |
| 1363 | |
| 1364 | if (z->has_wp) |
| 1365 | zone_unlock(z); |
| 1366 | } |
| 1367 | |
| 1368 | /* Both z->mutex and zbdi->mutex are held. */ |
| 1369 | |
| 1370 | examine_zone: |
| 1371 | if (z->wp + min_bs <= zbd_zone_capacity_end(z)) { |
| 1372 | pthread_mutex_unlock(&zbdi->mutex); |
| 1373 | goto out; |
| 1374 | } |
| 1375 | |
| 1376 | open_other_zone: |
| 1377 | /* Check if number of open zones reaches one of limits. */ |
| 1378 | wait_zone_close = |
| 1379 | zbdi->num_open_zones == f->max_zone - f->min_zone || |
| 1380 | (zbdi->max_open_zones && |
| 1381 | zbdi->num_open_zones == zbdi->max_open_zones) || |
| 1382 | (td->o.job_max_open_zones && |
| 1383 | td->num_open_zones == td->o.job_max_open_zones); |
| 1384 | |
| 1385 | pthread_mutex_unlock(&zbdi->mutex); |
| 1386 | |
| 1387 | /* Only z->mutex is held. */ |
| 1388 | |
| 1389 | /* |
| 1390 | * When number of open zones reaches to one of limits, wait for |
| 1391 | * zone close before opening a new zone. |
| 1392 | */ |
| 1393 | if (wait_zone_close) { |
| 1394 | dprint(FD_ZBD, |
| 1395 | "%s(%s): quiesce to allow open zones to close\n", |
| 1396 | __func__, f->file_name); |
| 1397 | io_u_quiesce(td); |
| 1398 | } |
| 1399 | |
| 1400 | retry: |
| 1401 | /* Zone 'z' is full, so try to open a new zone. */ |
| 1402 | for (i = f->io_size / zbdi->zone_size; i > 0; i--) { |
| 1403 | zone_idx++; |
| 1404 | if (z->has_wp) |
| 1405 | zone_unlock(z); |
| 1406 | z++; |
| 1407 | if (!is_valid_offset(f, z->start)) { |
| 1408 | /* Wrap-around. */ |
| 1409 | zone_idx = f->min_zone; |
| 1410 | z = zbd_get_zone(f, zone_idx); |
| 1411 | } |
| 1412 | assert(is_valid_offset(f, z->start)); |
| 1413 | if (!z->has_wp) |
| 1414 | continue; |
| 1415 | zone_lock(td, f, z); |
| 1416 | if (z->open) |
| 1417 | continue; |
| 1418 | if (zbd_open_zone(td, f, z)) |
| 1419 | goto out; |
| 1420 | } |
| 1421 | |
| 1422 | /* Only z->mutex is held. */ |
| 1423 | |
| 1424 | /* Check whether the write fits in any of the already opened zones. */ |
| 1425 | pthread_mutex_lock(&zbdi->mutex); |
| 1426 | for (i = 0; i < zbdi->num_open_zones; i++) { |
| 1427 | zone_idx = zbdi->open_zones[i]; |
| 1428 | if (zone_idx < f->min_zone || zone_idx >= f->max_zone) |
| 1429 | continue; |
| 1430 | pthread_mutex_unlock(&zbdi->mutex); |
| 1431 | zone_unlock(z); |
| 1432 | |
| 1433 | z = zbd_get_zone(f, zone_idx); |
| 1434 | |
| 1435 | zone_lock(td, f, z); |
| 1436 | if (z->wp + min_bs <= zbd_zone_capacity_end(z)) |
| 1437 | goto out; |
| 1438 | pthread_mutex_lock(&zbdi->mutex); |
| 1439 | } |
| 1440 | |
| 1441 | /* |
| 1442 | * When any I/O is in-flight or when all I/Os in-flight get completed, |
| 1443 | * the I/Os might have closed zones then retry the steps to open a zone. |
| 1444 | * Before retry, call io_u_quiesce() to complete in-flight writes. |
| 1445 | */ |
| 1446 | in_flight = any_io_in_flight(); |
| 1447 | if (in_flight || should_retry) { |
| 1448 | dprint(FD_ZBD, |
| 1449 | "%s(%s): wait zone close and retry open zones\n", |
| 1450 | __func__, f->file_name); |
| 1451 | pthread_mutex_unlock(&zbdi->mutex); |
| 1452 | zone_unlock(z); |
| 1453 | io_u_quiesce(td); |
| 1454 | zone_lock(td, f, z); |
| 1455 | should_retry = in_flight; |
| 1456 | goto retry; |
| 1457 | } |
| 1458 | |
| 1459 | pthread_mutex_unlock(&zbdi->mutex); |
| 1460 | |
| 1461 | zone_unlock(z); |
| 1462 | |
| 1463 | dprint(FD_ZBD, "%s(%s): did not open another zone\n", |
| 1464 | __func__, f->file_name); |
| 1465 | |
| 1466 | return NULL; |
| 1467 | |
| 1468 | out: |
| 1469 | dprint(FD_ZBD, "%s(%s): returning zone %d\n", |
| 1470 | __func__, f->file_name, zone_idx); |
| 1471 | |
| 1472 | io_u->offset = z->start; |
| 1473 | assert(z->has_wp); |
| 1474 | assert(z->cond != ZBD_ZONE_COND_OFFLINE); |
| 1475 | |
| 1476 | return z; |
| 1477 | } |
| 1478 | |
| 1479 | /* The caller must hold z->mutex. */ |
| 1480 | static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td, |
| 1481 | struct io_u *io_u, |
| 1482 | struct fio_zone_info *z) |
| 1483 | { |
| 1484 | const struct fio_file *f = io_u->file; |
| 1485 | const uint64_t min_bs = td->o.min_bs[DDIR_WRITE]; |
| 1486 | |
| 1487 | if (!zbd_open_zone(td, f, z)) { |
| 1488 | zone_unlock(z); |
| 1489 | z = zbd_convert_to_open_zone(td, io_u); |
| 1490 | assert(z); |
| 1491 | } |
| 1492 | |
| 1493 | if (z->verify_block * min_bs >= z->capacity) { |
| 1494 | log_err("%s: %d * %"PRIu64" >= %"PRIu64"\n", |
| 1495 | f->file_name, z->verify_block, min_bs, z->capacity); |
| 1496 | /* |
| 1497 | * If the assertion below fails during a test run, adding |
| 1498 | * "--experimental_verify=1" to the command line may help. |
| 1499 | */ |
| 1500 | assert(false); |
| 1501 | } |
| 1502 | |
| 1503 | io_u->offset = z->start + z->verify_block * min_bs; |
| 1504 | if (io_u->offset + io_u->buflen >= zbd_zone_capacity_end(z)) { |
| 1505 | log_err("%s: %llu + %llu >= %"PRIu64"\n", |
| 1506 | f->file_name, io_u->offset, io_u->buflen, |
| 1507 | zbd_zone_capacity_end(z)); |
| 1508 | assert(false); |
| 1509 | } |
| 1510 | z->verify_block += io_u->buflen / min_bs; |
| 1511 | |
| 1512 | return z; |
| 1513 | } |
| 1514 | |
| 1515 | /* |
| 1516 | * Find another zone which has @min_bytes of readable data. Search in zones |
| 1517 | * @zb + 1 .. @zl. For random workload, also search in zones @zb - 1 .. @zf. |
| 1518 | * |
| 1519 | * Either returns NULL or returns a zone pointer. When the zone has write |
| 1520 | * pointer, hold the mutex for the zone. |
| 1521 | */ |
| 1522 | static struct fio_zone_info * |
| 1523 | zbd_find_zone(struct thread_data *td, struct io_u *io_u, uint64_t min_bytes, |
| 1524 | struct fio_zone_info *zb, struct fio_zone_info *zl) |
| 1525 | { |
| 1526 | struct fio_file *f = io_u->file; |
| 1527 | struct fio_zone_info *z1, *z2; |
| 1528 | const struct fio_zone_info *const zf = zbd_get_zone(f, f->min_zone); |
| 1529 | |
| 1530 | /* |
| 1531 | * Skip to the next non-empty zone in case of sequential I/O and to |
| 1532 | * the nearest non-empty zone in case of random I/O. |
| 1533 | */ |
| 1534 | for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) { |
| 1535 | if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) { |
| 1536 | if (z1->has_wp) |
| 1537 | zone_lock(td, f, z1); |
| 1538 | if (z1->start + min_bytes <= z1->wp) |
| 1539 | return z1; |
| 1540 | if (z1->has_wp) |
| 1541 | zone_unlock(z1); |
| 1542 | } else if (!td_random(td)) { |
| 1543 | break; |
| 1544 | } |
| 1545 | |
| 1546 | if (td_random(td) && z2 >= zf && |
| 1547 | z2->cond != ZBD_ZONE_COND_OFFLINE) { |
| 1548 | if (z2->has_wp) |
| 1549 | zone_lock(td, f, z2); |
| 1550 | if (z2->start + min_bytes <= z2->wp) |
| 1551 | return z2; |
| 1552 | if (z2->has_wp) |
| 1553 | zone_unlock(z2); |
| 1554 | } |
| 1555 | } |
| 1556 | |
| 1557 | dprint(FD_ZBD, |
| 1558 | "%s: no zone has %"PRIu64" bytes of readable data\n", |
| 1559 | f->file_name, min_bytes); |
| 1560 | |
| 1561 | return NULL; |
| 1562 | } |
| 1563 | |
| 1564 | /** |
| 1565 | * zbd_end_zone_io - update zone status at command completion |
| 1566 | * @io_u: I/O unit |
| 1567 | * @z: zone info pointer |
| 1568 | * |
| 1569 | * If the write command made the zone full, close it. |
| 1570 | * |
| 1571 | * The caller must hold z->mutex. |
| 1572 | */ |
| 1573 | static void zbd_end_zone_io(struct thread_data *td, const struct io_u *io_u, |
| 1574 | struct fio_zone_info *z) |
| 1575 | { |
| 1576 | const struct fio_file *f = io_u->file; |
| 1577 | |
| 1578 | if (io_u->ddir == DDIR_WRITE && |
| 1579 | io_u->offset + io_u->buflen >= zbd_zone_capacity_end(z)) { |
| 1580 | pthread_mutex_lock(&f->zbd_info->mutex); |
| 1581 | zbd_close_zone(td, f, z); |
| 1582 | pthread_mutex_unlock(&f->zbd_info->mutex); |
| 1583 | } |
| 1584 | } |
| 1585 | |
| 1586 | /** |
| 1587 | * zbd_queue_io - update the write pointer of a sequential zone |
| 1588 | * @io_u: I/O unit |
| 1589 | * @success: Whether or not the I/O unit has been queued successfully |
| 1590 | * @q: queueing status (busy, completed or queued). |
| 1591 | * |
| 1592 | * For write and trim operations, update the write pointer of the I/O unit |
| 1593 | * target zone. |
| 1594 | */ |
| 1595 | static void zbd_queue_io(struct thread_data *td, struct io_u *io_u, int q, |
| 1596 | bool success) |
| 1597 | { |
| 1598 | const struct fio_file *f = io_u->file; |
| 1599 | struct zoned_block_device_info *zbd_info = f->zbd_info; |
| 1600 | struct fio_zone_info *z; |
| 1601 | uint64_t zone_end; |
| 1602 | |
| 1603 | assert(zbd_info); |
| 1604 | |
| 1605 | z = zbd_offset_to_zone(f, io_u->offset); |
| 1606 | assert(z->has_wp); |
| 1607 | |
| 1608 | if (!success) |
| 1609 | goto unlock; |
| 1610 | |
| 1611 | dprint(FD_ZBD, |
| 1612 | "%s: queued I/O (%lld, %llu) for zone %u\n", |
| 1613 | f->file_name, io_u->offset, io_u->buflen, zbd_zone_idx(f, z)); |
| 1614 | |
| 1615 | switch (io_u->ddir) { |
| 1616 | case DDIR_WRITE: |
| 1617 | zone_end = min((uint64_t)(io_u->offset + io_u->buflen), |
| 1618 | zbd_zone_capacity_end(z)); |
| 1619 | |
| 1620 | /* |
| 1621 | * z->wp > zone_end means that one or more I/O errors |
| 1622 | * have occurred. |
| 1623 | */ |
| 1624 | pthread_mutex_lock(&zbd_info->mutex); |
| 1625 | if (z->wp <= zone_end) { |
| 1626 | zbd_info->sectors_with_data += zone_end - z->wp; |
| 1627 | zbd_info->wp_sectors_with_data += zone_end - z->wp; |
| 1628 | } |
| 1629 | pthread_mutex_unlock(&zbd_info->mutex); |
| 1630 | z->wp = zone_end; |
| 1631 | break; |
| 1632 | default: |
| 1633 | break; |
| 1634 | } |
| 1635 | |
| 1636 | if (q == FIO_Q_COMPLETED && !io_u->error) |
| 1637 | zbd_end_zone_io(td, io_u, z); |
| 1638 | |
| 1639 | unlock: |
| 1640 | if (!success || q != FIO_Q_QUEUED) { |
| 1641 | /* BUSY or COMPLETED: unlock the zone */ |
| 1642 | zone_unlock(z); |
| 1643 | io_u->zbd_put_io = NULL; |
| 1644 | } |
| 1645 | } |
| 1646 | |
| 1647 | /** |
| 1648 | * zbd_put_io - Unlock an I/O unit target zone lock |
| 1649 | * @io_u: I/O unit |
| 1650 | */ |
| 1651 | static void zbd_put_io(struct thread_data *td, const struct io_u *io_u) |
| 1652 | { |
| 1653 | const struct fio_file *f = io_u->file; |
| 1654 | struct zoned_block_device_info *zbd_info = f->zbd_info; |
| 1655 | struct fio_zone_info *z; |
| 1656 | |
| 1657 | assert(zbd_info); |
| 1658 | |
| 1659 | z = zbd_offset_to_zone(f, io_u->offset); |
| 1660 | assert(z->has_wp); |
| 1661 | |
| 1662 | dprint(FD_ZBD, |
| 1663 | "%s: terminate I/O (%lld, %llu) for zone %u\n", |
| 1664 | f->file_name, io_u->offset, io_u->buflen, zbd_zone_idx(f, z)); |
| 1665 | |
| 1666 | zbd_end_zone_io(td, io_u, z); |
| 1667 | |
| 1668 | zone_unlock(z); |
| 1669 | zbd_check_swd(td, f); |
| 1670 | } |
| 1671 | |
| 1672 | /* |
| 1673 | * Windows and MacOS do not define this. |
| 1674 | */ |
| 1675 | #ifndef EREMOTEIO |
| 1676 | #define EREMOTEIO 121 /* POSIX value */ |
| 1677 | #endif |
| 1678 | |
| 1679 | bool zbd_unaligned_write(int error_code) |
| 1680 | { |
| 1681 | switch (error_code) { |
| 1682 | case EIO: |
| 1683 | case EREMOTEIO: |
| 1684 | return true; |
| 1685 | } |
| 1686 | return false; |
| 1687 | } |
| 1688 | |
| 1689 | /** |
| 1690 | * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives |
| 1691 | * @td: FIO thread data. |
| 1692 | * @io_u: FIO I/O unit. |
| 1693 | * |
| 1694 | * For sequential workloads, change the file offset to skip zoneskip bytes when |
| 1695 | * no more IO can be performed in the current zone. |
| 1696 | * - For read workloads, zoneskip is applied when the io has reached the end of |
| 1697 | * the zone or the zone write position (when td->o.read_beyond_wp is false). |
| 1698 | * - For write workloads, zoneskip is applied when the zone is full. |
| 1699 | * This applies only to read and write operations. |
| 1700 | */ |
| 1701 | void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u) |
| 1702 | { |
| 1703 | struct fio_file *f = io_u->file; |
| 1704 | enum fio_ddir ddir = io_u->ddir; |
| 1705 | struct fio_zone_info *z; |
| 1706 | |
| 1707 | assert(td->o.zone_mode == ZONE_MODE_ZBD); |
| 1708 | assert(td->o.zone_size); |
| 1709 | assert(f->zbd_info); |
| 1710 | |
| 1711 | z = zbd_offset_to_zone(f, f->last_pos[ddir]); |
| 1712 | |
| 1713 | /* |
| 1714 | * When the zone capacity is smaller than the zone size and the I/O is |
| 1715 | * sequential write, skip to zone end if the latest position is at the |
| 1716 | * zone capacity limit. |
| 1717 | */ |
| 1718 | if (z->capacity < f->zbd_info->zone_size && |
| 1719 | !td_random(td) && ddir == DDIR_WRITE && |
| 1720 | f->last_pos[ddir] >= zbd_zone_capacity_end(z)) { |
| 1721 | dprint(FD_ZBD, |
| 1722 | "%s: Jump from zone capacity limit to zone end:" |
| 1723 | " (%"PRIu64" -> %"PRIu64") for zone %u (%"PRIu64")\n", |
| 1724 | f->file_name, f->last_pos[ddir], |
| 1725 | zbd_zone_end(z), zbd_zone_idx(f, z), z->capacity); |
| 1726 | td->io_skip_bytes += zbd_zone_end(z) - f->last_pos[ddir]; |
| 1727 | f->last_pos[ddir] = zbd_zone_end(z); |
| 1728 | } |
| 1729 | |
| 1730 | /* |
| 1731 | * zone_skip is valid only for sequential workloads. |
| 1732 | */ |
| 1733 | if (td_random(td) || !td->o.zone_skip) |
| 1734 | return; |
| 1735 | |
| 1736 | /* |
| 1737 | * It is time to switch to a new zone if: |
| 1738 | * - zone_bytes == zone_size bytes have already been accessed |
| 1739 | * - The last position reached the end of the current zone. |
| 1740 | * - For reads with td->o.read_beyond_wp == false, the last position |
| 1741 | * reached the zone write pointer. |
| 1742 | */ |
| 1743 | if (td->zone_bytes >= td->o.zone_size || |
| 1744 | f->last_pos[ddir] >= zbd_zone_end(z) || |
| 1745 | (ddir == DDIR_READ && |
| 1746 | (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) { |
| 1747 | /* |
| 1748 | * Skip zones. |
| 1749 | */ |
| 1750 | td->zone_bytes = 0; |
| 1751 | f->file_offset += td->o.zone_size + td->o.zone_skip; |
| 1752 | |
| 1753 | /* |
| 1754 | * Wrap from the beginning, if we exceed the file size |
| 1755 | */ |
| 1756 | if (f->file_offset >= f->real_file_size) |
| 1757 | f->file_offset = get_start_offset(td, f); |
| 1758 | |
| 1759 | f->last_pos[ddir] = f->file_offset; |
| 1760 | td->io_skip_bytes += td->o.zone_skip; |
| 1761 | } |
| 1762 | } |
| 1763 | |
| 1764 | /** |
| 1765 | * zbd_adjust_ddir - Adjust an I/O direction for zonemode=zbd. |
| 1766 | * |
| 1767 | * @td: FIO thread data. |
| 1768 | * @io_u: FIO I/O unit. |
| 1769 | * @ddir: I/O direction before adjustment. |
| 1770 | * |
| 1771 | * Return adjusted I/O direction. |
| 1772 | */ |
| 1773 | enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u, |
| 1774 | enum fio_ddir ddir) |
| 1775 | { |
| 1776 | /* |
| 1777 | * In case read direction is chosen for the first random I/O, fio with |
| 1778 | * zonemode=zbd stops because no data can be read from zoned block |
| 1779 | * devices with all empty zones. Overwrite the first I/O direction as |
| 1780 | * write to make sure data to read exists. |
| 1781 | */ |
| 1782 | assert(io_u->file->zbd_info); |
| 1783 | if (ddir != DDIR_READ || !td_rw(td)) |
| 1784 | return ddir; |
| 1785 | |
| 1786 | if (io_u->file->zbd_info->sectors_with_data || |
| 1787 | td->o.read_beyond_wp) |
| 1788 | return DDIR_READ; |
| 1789 | |
| 1790 | return DDIR_WRITE; |
| 1791 | } |
| 1792 | |
| 1793 | /** |
| 1794 | * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives |
| 1795 | * @td: FIO thread data. |
| 1796 | * @io_u: FIO I/O unit. |
| 1797 | * |
| 1798 | * Locking strategy: returns with z->mutex locked if and only if z refers |
| 1799 | * to a sequential zone and if io_u_accept is returned. z is the zone that |
| 1800 | * corresponds to io_u->offset at the end of this function. |
| 1801 | */ |
| 1802 | enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u) |
| 1803 | { |
| 1804 | struct fio_file *f = io_u->file; |
| 1805 | struct zoned_block_device_info *zbdi = f->zbd_info; |
| 1806 | struct fio_zone_info *zb, *zl, *orig_zb; |
| 1807 | uint32_t orig_len = io_u->buflen; |
| 1808 | uint64_t min_bs = td->o.min_bs[io_u->ddir]; |
| 1809 | uint64_t new_len; |
| 1810 | int64_t range; |
| 1811 | |
| 1812 | assert(zbdi); |
| 1813 | assert(min_bs); |
| 1814 | assert(is_valid_offset(f, io_u->offset)); |
| 1815 | assert(io_u->buflen); |
| 1816 | |
| 1817 | zb = zbd_offset_to_zone(f, io_u->offset); |
| 1818 | orig_zb = zb; |
| 1819 | |
| 1820 | if (!zb->has_wp) { |
| 1821 | /* Accept non-write I/Os for conventional zones. */ |
| 1822 | if (io_u->ddir != DDIR_WRITE) |
| 1823 | return io_u_accept; |
| 1824 | |
| 1825 | /* |
| 1826 | * Make sure that writes to conventional zones |
| 1827 | * don't cross over to any sequential zones. |
| 1828 | */ |
| 1829 | if (!(zb + 1)->has_wp || |
| 1830 | io_u->offset + io_u->buflen <= (zb + 1)->start) |
| 1831 | return io_u_accept; |
| 1832 | |
| 1833 | if (io_u->offset + min_bs > (zb + 1)->start) { |
| 1834 | dprint(FD_IO, |
| 1835 | "%s: off=%llu + min_bs=%"PRIu64" > next zone %"PRIu64"\n", |
| 1836 | f->file_name, io_u->offset, |
| 1837 | min_bs, (zb + 1)->start); |
| 1838 | io_u->offset = |
| 1839 | zb->start + (zb + 1)->start - io_u->offset; |
| 1840 | new_len = min(io_u->buflen, |
| 1841 | (zb + 1)->start - io_u->offset); |
| 1842 | } else { |
| 1843 | new_len = (zb + 1)->start - io_u->offset; |
| 1844 | } |
| 1845 | |
| 1846 | io_u->buflen = new_len / min_bs * min_bs; |
| 1847 | |
| 1848 | return io_u_accept; |
| 1849 | } |
| 1850 | |
| 1851 | /* |
| 1852 | * Accept the I/O offset for reads if reading beyond the write pointer |
| 1853 | * is enabled. |
| 1854 | */ |
| 1855 | if (zb->cond != ZBD_ZONE_COND_OFFLINE && |
| 1856 | io_u->ddir == DDIR_READ && td->o.read_beyond_wp) |
| 1857 | return io_u_accept; |
| 1858 | |
| 1859 | zbd_check_swd(td, f); |
| 1860 | |
| 1861 | zone_lock(td, f, zb); |
| 1862 | |
| 1863 | switch (io_u->ddir) { |
| 1864 | case DDIR_READ: |
| 1865 | if (td->runstate == TD_VERIFYING && td_write(td)) { |
| 1866 | zb = zbd_replay_write_order(td, io_u, zb); |
| 1867 | goto accept; |
| 1868 | } |
| 1869 | |
| 1870 | /* |
| 1871 | * Check that there is enough written data in the zone to do an |
| 1872 | * I/O of at least min_bs B. If there isn't, find a new zone for |
| 1873 | * the I/O. |
| 1874 | */ |
| 1875 | range = zb->cond != ZBD_ZONE_COND_OFFLINE ? |
| 1876 | zb->wp - zb->start : 0; |
| 1877 | if (range < min_bs || |
| 1878 | ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) { |
| 1879 | zone_unlock(zb); |
| 1880 | zl = zbd_get_zone(f, f->max_zone); |
| 1881 | zb = zbd_find_zone(td, io_u, min_bs, zb, zl); |
| 1882 | if (!zb) { |
| 1883 | dprint(FD_ZBD, |
| 1884 | "%s: zbd_find_zone(%lld, %llu) failed\n", |
| 1885 | f->file_name, io_u->offset, |
| 1886 | io_u->buflen); |
| 1887 | goto eof; |
| 1888 | } |
| 1889 | /* |
| 1890 | * zbd_find_zone() returned a zone with a range of at |
| 1891 | * least min_bs. |
| 1892 | */ |
| 1893 | range = zb->wp - zb->start; |
| 1894 | assert(range >= min_bs); |
| 1895 | |
| 1896 | if (!td_random(td)) |
| 1897 | io_u->offset = zb->start; |
| 1898 | } |
| 1899 | |
| 1900 | /* |
| 1901 | * Make sure the I/O is within the zone valid data range while |
| 1902 | * maximizing the I/O size and preserving randomness. |
| 1903 | */ |
| 1904 | if (range <= io_u->buflen) |
| 1905 | io_u->offset = zb->start; |
| 1906 | else if (td_random(td)) |
| 1907 | io_u->offset = zb->start + |
| 1908 | ((io_u->offset - orig_zb->start) % |
| 1909 | (range - io_u->buflen)) / min_bs * min_bs; |
| 1910 | |
| 1911 | /* |
| 1912 | * When zbd_find_zone() returns a conventional zone, |
| 1913 | * we can simply accept the new i/o offset here. |
| 1914 | */ |
| 1915 | if (!zb->has_wp) |
| 1916 | return io_u_accept; |
| 1917 | |
| 1918 | /* |
| 1919 | * Make sure the I/O does not cross over the zone wp position. |
| 1920 | */ |
| 1921 | new_len = min((unsigned long long)io_u->buflen, |
| 1922 | (unsigned long long)(zb->wp - io_u->offset)); |
| 1923 | new_len = new_len / min_bs * min_bs; |
| 1924 | if (new_len < io_u->buflen) { |
| 1925 | io_u->buflen = new_len; |
| 1926 | dprint(FD_IO, "Changed length from %u into %llu\n", |
| 1927 | orig_len, io_u->buflen); |
| 1928 | } |
| 1929 | |
| 1930 | assert(zb->start <= io_u->offset); |
| 1931 | assert(io_u->offset + io_u->buflen <= zb->wp); |
| 1932 | |
| 1933 | goto accept; |
| 1934 | |
| 1935 | case DDIR_WRITE: |
| 1936 | if (io_u->buflen > zbdi->zone_size) { |
| 1937 | td_verror(td, EINVAL, "I/O buflen exceeds zone size"); |
| 1938 | dprint(FD_IO, |
| 1939 | "%s: I/O buflen %llu exceeds zone size %"PRIu64"\n", |
| 1940 | f->file_name, io_u->buflen, zbdi->zone_size); |
| 1941 | goto eof; |
| 1942 | } |
| 1943 | |
| 1944 | if (!zbd_open_zone(td, f, zb)) { |
| 1945 | zone_unlock(zb); |
| 1946 | zb = zbd_convert_to_open_zone(td, io_u); |
| 1947 | if (!zb) { |
| 1948 | dprint(FD_IO, "%s: can't convert to open zone", |
| 1949 | f->file_name); |
| 1950 | goto eof; |
| 1951 | } |
| 1952 | } |
| 1953 | |
| 1954 | /* Check whether the zone reset threshold has been exceeded */ |
| 1955 | if (td->o.zrf.u.f) { |
| 1956 | if (zbdi->wp_sectors_with_data >= f->io_size * td->o.zrt.u.f && |
| 1957 | zbd_dec_and_reset_write_cnt(td, f)) |
| 1958 | zb->reset_zone = 1; |
| 1959 | } |
| 1960 | |
| 1961 | /* Reset the zone pointer if necessary */ |
| 1962 | if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) { |
| 1963 | assert(td->o.verify == VERIFY_NONE); |
| 1964 | /* |
| 1965 | * Since previous write requests may have been submitted |
| 1966 | * asynchronously and since we will submit the zone |
| 1967 | * reset synchronously, wait until previously submitted |
| 1968 | * write requests have completed before issuing a |
| 1969 | * zone reset. |
| 1970 | */ |
| 1971 | io_u_quiesce(td); |
| 1972 | zb->reset_zone = 0; |
| 1973 | if (zbd_reset_zone(td, f, zb) < 0) |
| 1974 | goto eof; |
| 1975 | |
| 1976 | if (zb->capacity < min_bs) { |
| 1977 | td_verror(td, EINVAL, "ZCAP is less min_bs"); |
| 1978 | log_err("zone capacity %"PRIu64" smaller than minimum block size %"PRIu64"\n", |
| 1979 | zb->capacity, min_bs); |
| 1980 | goto eof; |
| 1981 | } |
| 1982 | } |
| 1983 | |
| 1984 | /* Make writes occur at the write pointer */ |
| 1985 | assert(!zbd_zone_full(f, zb, min_bs)); |
| 1986 | io_u->offset = zb->wp; |
| 1987 | if (!is_valid_offset(f, io_u->offset)) { |
| 1988 | td_verror(td, EINVAL, "invalid WP value"); |
| 1989 | dprint(FD_ZBD, "%s: dropped request with offset %llu\n", |
| 1990 | f->file_name, io_u->offset); |
| 1991 | goto eof; |
| 1992 | } |
| 1993 | |
| 1994 | /* |
| 1995 | * Make sure that the buflen is a multiple of the minimal |
| 1996 | * block size. Give up if shrinking would make the request too |
| 1997 | * small. |
| 1998 | */ |
| 1999 | new_len = min((unsigned long long)io_u->buflen, |
| 2000 | zbd_zone_capacity_end(zb) - io_u->offset); |
| 2001 | new_len = new_len / min_bs * min_bs; |
| 2002 | if (new_len == io_u->buflen) |
| 2003 | goto accept; |
| 2004 | if (new_len >= min_bs) { |
| 2005 | io_u->buflen = new_len; |
| 2006 | dprint(FD_IO, "Changed length from %u into %llu\n", |
| 2007 | orig_len, io_u->buflen); |
| 2008 | goto accept; |
| 2009 | } |
| 2010 | |
| 2011 | td_verror(td, EIO, "zone remainder too small"); |
| 2012 | log_err("zone remainder %lld smaller than min block size %"PRIu64"\n", |
| 2013 | (zbd_zone_capacity_end(zb) - io_u->offset), min_bs); |
| 2014 | |
| 2015 | goto eof; |
| 2016 | |
| 2017 | case DDIR_TRIM: |
| 2018 | /* Check random trim targets a non-empty zone */ |
| 2019 | if (!td_random(td) || zb->wp > zb->start) |
| 2020 | goto accept; |
| 2021 | |
| 2022 | /* Find out a non-empty zone to trim */ |
| 2023 | zone_unlock(zb); |
| 2024 | zl = zbd_get_zone(f, f->max_zone); |
| 2025 | zb = zbd_find_zone(td, io_u, 1, zb, zl); |
| 2026 | if (zb) { |
| 2027 | io_u->offset = zb->start; |
| 2028 | dprint(FD_ZBD, "%s: found new zone(%lld) for trim\n", |
| 2029 | f->file_name, io_u->offset); |
| 2030 | goto accept; |
| 2031 | } |
| 2032 | |
| 2033 | goto eof; |
| 2034 | |
| 2035 | case DDIR_SYNC: |
| 2036 | /* fall-through */ |
| 2037 | case DDIR_DATASYNC: |
| 2038 | case DDIR_SYNC_FILE_RANGE: |
| 2039 | case DDIR_WAIT: |
| 2040 | case DDIR_LAST: |
| 2041 | case DDIR_INVAL: |
| 2042 | goto accept; |
| 2043 | } |
| 2044 | |
| 2045 | assert(false); |
| 2046 | |
| 2047 | accept: |
| 2048 | assert(zb->has_wp); |
| 2049 | assert(zb->cond != ZBD_ZONE_COND_OFFLINE); |
| 2050 | assert(!io_u->zbd_queue_io); |
| 2051 | assert(!io_u->zbd_put_io); |
| 2052 | |
| 2053 | io_u->zbd_queue_io = zbd_queue_io; |
| 2054 | io_u->zbd_put_io = zbd_put_io; |
| 2055 | |
| 2056 | /* |
| 2057 | * Since we return with the zone lock still held, |
| 2058 | * add an annotation to let Coverity know that it |
| 2059 | * is intentional. |
| 2060 | */ |
| 2061 | /* coverity[missing_unlock] */ |
| 2062 | |
| 2063 | return io_u_accept; |
| 2064 | |
| 2065 | eof: |
| 2066 | if (zb && zb->has_wp) |
| 2067 | zone_unlock(zb); |
| 2068 | |
| 2069 | return io_u_eof; |
| 2070 | } |
| 2071 | |
| 2072 | /* Return a string with ZBD statistics */ |
| 2073 | char *zbd_write_status(const struct thread_stat *ts) |
| 2074 | { |
| 2075 | char *res; |
| 2076 | |
| 2077 | if (asprintf(&res, "; %"PRIu64" zone resets", ts->nr_zone_resets) < 0) |
| 2078 | return NULL; |
| 2079 | return res; |
| 2080 | } |
| 2081 | |
| 2082 | /** |
| 2083 | * zbd_do_io_u_trim - If reset zone is applicable, do reset zone instead of trim |
| 2084 | * |
| 2085 | * @td: FIO thread data. |
| 2086 | * @io_u: FIO I/O unit. |
| 2087 | * |
| 2088 | * It is assumed that z->mutex is already locked. |
| 2089 | * Return io_u_completed when reset zone succeeds. Return 0 when the target zone |
| 2090 | * does not have write pointer. On error, return negative errno. |
| 2091 | */ |
| 2092 | int zbd_do_io_u_trim(const struct thread_data *td, struct io_u *io_u) |
| 2093 | { |
| 2094 | struct fio_file *f = io_u->file; |
| 2095 | struct fio_zone_info *z; |
| 2096 | int ret; |
| 2097 | |
| 2098 | z = zbd_offset_to_zone(f, io_u->offset); |
| 2099 | if (!z->has_wp) |
| 2100 | return 0; |
| 2101 | |
| 2102 | if (io_u->offset != z->start) { |
| 2103 | log_err("Trim offset not at zone start (%lld)\n", |
| 2104 | io_u->offset); |
| 2105 | return -EINVAL; |
| 2106 | } |
| 2107 | |
| 2108 | ret = zbd_reset_zone((struct thread_data *)td, f, z); |
| 2109 | if (ret < 0) |
| 2110 | return ret; |
| 2111 | |
| 2112 | return io_u_completed; |
| 2113 | } |