1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2018 Intel Corporation. 3 * All rights reserved. 4 */ 5 6 #ifndef SPDK_BDEV_RAID_INTERNAL_H 7 #define SPDK_BDEV_RAID_INTERNAL_H 8 9 #include "spdk/bdev_module.h" 10 #include "spdk/uuid.h" 11 12 #define RAID_BDEV_MIN_DATA_OFFSET_SIZE (1024*1024) /* 1 MiB */ 13 14 enum raid_level { 15 INVALID_RAID_LEVEL = -1, 16 RAID0 = 0, 17 RAID1 = 1, 18 RAID5F = 95, /* 0x5f */ 19 CONCAT = 99, 20 }; 21 22 /* 23 * Raid state describes the state of the raid. This raid bdev can be either in 24 * configured list or configuring list 25 */ 26 enum raid_bdev_state { 27 /* raid bdev is ready and is seen by upper layers */ 28 RAID_BDEV_STATE_ONLINE, 29 30 /* 31 * raid bdev is configuring, not all underlying bdevs are present. 32 * And can't be seen by upper layers. 33 */ 34 RAID_BDEV_STATE_CONFIGURING, 35 36 /* 37 * In offline state, raid bdev layer will complete all incoming commands without 38 * submitting to underlying base nvme bdevs 39 */ 40 RAID_BDEV_STATE_OFFLINE, 41 42 /* raid bdev state max, new states should be added before this */ 43 RAID_BDEV_STATE_MAX 44 }; 45 46 enum raid_process_type { 47 RAID_PROCESS_NONE, 48 RAID_PROCESS_REBUILD, 49 RAID_PROCESS_MAX 50 }; 51 52 typedef void (*raid_base_bdev_cb)(void *ctx, int status); 53 54 /* 55 * raid_base_bdev_info contains information for the base bdevs which are part of some 56 * raid. This structure contains the per base bdev information. Whatever is 57 * required per base device for raid bdev will be kept here 58 */ 59 struct raid_base_bdev_info { 60 /* The raid bdev that this base bdev belongs to */ 61 struct raid_bdev *raid_bdev; 62 63 /* name of the bdev */ 64 char *name; 65 66 /* uuid of the bdev */ 67 struct spdk_uuid uuid; 68 69 /* 70 * Pointer to base bdev descriptor opened by raid bdev. This is NULL when the bdev for 71 * this slot is missing. 72 */ 73 struct spdk_bdev_desc *desc; 74 75 /* offset in blocks from the start of the base bdev to the start of the data region */ 76 uint64_t data_offset; 77 78 /* size in blocks of the base bdev's data region */ 79 uint64_t data_size; 80 81 /* 82 * When underlying base device calls the hot plug function on drive removal, 83 * this flag will be set and later after doing some processing, base device 84 * descriptor will be closed 85 */ 86 bool remove_scheduled; 87 88 /* callback for base bdev removal */ 89 raid_base_bdev_cb remove_cb; 90 91 /* context of the callback */ 92 void *remove_cb_ctx; 93 94 /* Hold the number of blocks to know how large the base bdev is resized. */ 95 uint64_t blockcnt; 96 97 /* io channel for the app thread */ 98 struct spdk_io_channel *app_thread_ch; 99 100 /* Set to true when base bdev has completed the configuration process */ 101 bool is_configured; 102 103 /* callback for base bdev configuration */ 104 raid_base_bdev_cb configure_cb; 105 106 /* context of the callback */ 107 void *configure_cb_ctx; 108 }; 109 110 struct raid_bdev_io; 111 typedef void (*raid_bdev_io_completion_cb)(struct raid_bdev_io *raid_io, 112 enum spdk_bdev_io_status status); 113 114 /* 115 * raid_bdev_io is the context part of bdev_io. It contains the information 116 * related to bdev_io for a raid bdev 117 */ 118 struct raid_bdev_io { 119 /* The raid bdev associated with this IO */ 120 struct raid_bdev *raid_bdev; 121 122 uint64_t offset_blocks; 123 uint64_t num_blocks; 124 struct iovec *iovs; 125 int iovcnt; 126 enum spdk_bdev_io_type type; 127 struct spdk_memory_domain *memory_domain; 128 void *memory_domain_ctx; 129 void *md_buf; 130 131 /* WaitQ entry, used only in waitq logic */ 132 struct spdk_bdev_io_wait_entry waitq_entry; 133 134 /* Context of the original channel for this IO */ 135 struct raid_bdev_io_channel *raid_ch; 136 137 /* Used for tracking progress on io requests sent to member disks. */ 138 uint64_t base_bdev_io_remaining; 139 uint8_t base_bdev_io_submitted; 140 enum spdk_bdev_io_status base_bdev_io_status; 141 142 /* Private data for the raid module */ 143 void *module_private; 144 145 /* Custom completion callback. Overrides bdev_io completion if set. */ 146 raid_bdev_io_completion_cb completion_cb; 147 148 struct { 149 uint64_t offset; 150 struct iovec *iov; 151 struct iovec iov_copy; 152 } split; 153 }; 154 155 struct raid_bdev_process_request { 156 struct raid_bdev_process *process; 157 struct raid_base_bdev_info *target; 158 struct spdk_io_channel *target_ch; 159 uint64_t offset_blocks; 160 uint32_t num_blocks; 161 struct iovec iov; 162 void *md_buf; 163 /* bdev_io is raid_io's driver_ctx - don't reorder them! 164 * These are needed for re-using raid module I/O functions for process I/O. */ 165 struct spdk_bdev_io bdev_io; 166 struct raid_bdev_io raid_io; 167 TAILQ_ENTRY(raid_bdev_process_request) link; 168 }; 169 170 /* 171 * raid_bdev is the single entity structure which contains SPDK block device 172 * and the information related to any raid bdev either configured or 173 * in configuring list. io device is created on this. 174 */ 175 struct raid_bdev { 176 /* raid bdev device, this will get registered in bdev layer */ 177 struct spdk_bdev bdev; 178 179 /* the raid bdev descriptor, opened for internal use */ 180 struct spdk_bdev_desc *self_desc; 181 182 /* link of raid bdev to link it to global raid bdev list */ 183 TAILQ_ENTRY(raid_bdev) global_link; 184 185 /* array of base bdev info */ 186 struct raid_base_bdev_info *base_bdev_info; 187 188 /* lock to protect the base bdev array */ 189 struct spdk_spinlock base_bdev_lock; 190 191 /* strip size of raid bdev in blocks */ 192 uint32_t strip_size; 193 194 /* strip size of raid bdev in KB */ 195 uint32_t strip_size_kb; 196 197 /* strip size bit shift for optimized calculation */ 198 uint32_t strip_size_shift; 199 200 /* block length bit shift for optimized calculation */ 201 uint32_t blocklen_shift; 202 203 /* state of raid bdev */ 204 enum raid_bdev_state state; 205 206 /* number of base bdevs comprising raid bdev */ 207 uint8_t num_base_bdevs; 208 209 /* number of base bdevs discovered */ 210 uint8_t num_base_bdevs_discovered; 211 212 /* 213 * Number of operational base bdevs, i.e. how many we know/expect to be working. This 214 * will be less than num_base_bdevs when starting a degraded array. 215 */ 216 uint8_t num_base_bdevs_operational; 217 218 /* minimum number of viable base bdevs that are required by array to operate */ 219 uint8_t min_base_bdevs_operational; 220 221 /* Raid Level of this raid bdev */ 222 enum raid_level level; 223 224 /* Set to true if destroy of this raid bdev is started. */ 225 bool destroy_started; 226 227 /* Module for RAID-level specific operations */ 228 struct raid_bdev_module *module; 229 230 /* Private data for the raid module */ 231 void *module_private; 232 233 /* Superblock */ 234 struct raid_bdev_superblock *sb; 235 236 /* Raid bdev background process, e.g. rebuild */ 237 struct raid_bdev_process *process; 238 }; 239 240 #define RAID_FOR_EACH_BASE_BDEV(r, i) \ 241 for (i = r->base_bdev_info; i < r->base_bdev_info + r->num_base_bdevs; i++) 242 243 struct raid_bdev_io_channel; 244 245 /* TAIL head for raid bdev list */ 246 TAILQ_HEAD(raid_all_tailq, raid_bdev); 247 248 extern struct raid_all_tailq g_raid_bdev_list; 249 250 typedef void (*raid_bdev_destruct_cb)(void *cb_ctx, int rc); 251 252 int raid_bdev_create(const char *name, uint32_t strip_size, uint8_t num_base_bdevs, 253 enum raid_level level, bool superblock, const struct spdk_uuid *uuid, 254 struct raid_bdev **raid_bdev_out); 255 void raid_bdev_delete(struct raid_bdev *raid_bdev, raid_bdev_destruct_cb cb_fn, void *cb_ctx); 256 int raid_bdev_add_base_device(struct raid_bdev *raid_bdev, const char *name, uint8_t slot, 257 raid_base_bdev_cb cb_fn, void *cb_ctx); 258 struct raid_bdev *raid_bdev_find_by_name(const char *name); 259 enum raid_level raid_bdev_str_to_level(const char *str); 260 const char *raid_bdev_level_to_str(enum raid_level level); 261 enum raid_bdev_state raid_bdev_str_to_state(const char *str); 262 const char *raid_bdev_state_to_str(enum raid_bdev_state state); 263 const char *raid_bdev_process_to_str(enum raid_process_type value); 264 void raid_bdev_write_info_json(struct raid_bdev *raid_bdev, struct spdk_json_write_ctx *w); 265 int raid_bdev_remove_base_bdev(struct spdk_bdev *base_bdev, raid_base_bdev_cb cb_fn, void *cb_ctx); 266 int raid_bdev_attach_base_bdev(struct raid_bdev *raid_bdev, struct spdk_bdev *base_bdev, 267 raid_base_bdev_cb cb_fn, void *cb_ctx); 268 269 /* 270 * RAID module descriptor 271 */ 272 struct raid_bdev_module { 273 /* RAID level implemented by this module */ 274 enum raid_level level; 275 276 /* Minimum required number of base bdevs. Must be > 0. */ 277 uint8_t base_bdevs_min; 278 279 /* 280 * RAID constraint. Determines number of base bdevs that can be removed 281 * without failing the array. 282 */ 283 struct { 284 enum { 285 CONSTRAINT_UNSET = 0, 286 CONSTRAINT_MAX_BASE_BDEVS_REMOVED, 287 CONSTRAINT_MIN_BASE_BDEVS_OPERATIONAL, 288 } type; 289 uint8_t value; 290 } base_bdevs_constraint; 291 292 /* Set to true if this module supports memory domains. */ 293 bool memory_domains_supported; 294 295 /* 296 * Called when the raid is starting, right before changing the state to 297 * online and registering the bdev. Parameters of the bdev like blockcnt 298 * should be set here. 299 * 300 * Non-zero return value will abort the startup process. 301 */ 302 int (*start)(struct raid_bdev *raid_bdev); 303 304 /* 305 * Called when the raid is stopping, right before changing the state to 306 * offline and unregistering the bdev. Optional. 307 * 308 * The function should return false if it is asynchronous. Then, after 309 * the async operation has completed and the module is fully stopped 310 * raid_bdev_module_stop_done() must be called. 311 */ 312 bool (*stop)(struct raid_bdev *raid_bdev); 313 314 /* Handler for R/W requests */ 315 void (*submit_rw_request)(struct raid_bdev_io *raid_io); 316 317 /* Handler for requests without payload (flush, unmap). Optional. */ 318 void (*submit_null_payload_request)(struct raid_bdev_io *raid_io); 319 320 /* 321 * Called when the bdev's IO channel is created to get the module's private IO channel. 322 * Optional. 323 */ 324 struct spdk_io_channel *(*get_io_channel)(struct raid_bdev *raid_bdev); 325 326 /* 327 * Called when a base_bdev is resized to resize the raid if the condition 328 * is satisfied. 329 */ 330 void (*resize)(struct raid_bdev *raid_bdev); 331 332 /* Handler for raid process requests. Required for raid modules with redundancy. */ 333 int (*submit_process_request)(struct raid_bdev_process_request *process_req, 334 struct raid_bdev_io_channel *raid_ch); 335 336 TAILQ_ENTRY(raid_bdev_module) link; 337 }; 338 339 void raid_bdev_module_list_add(struct raid_bdev_module *raid_module); 340 341 #define __RAID_MODULE_REGISTER(line) __RAID_MODULE_REGISTER_(line) 342 #define __RAID_MODULE_REGISTER_(line) raid_module_register_##line 343 344 #define RAID_MODULE_REGISTER(_module) \ 345 __attribute__((constructor)) static void \ 346 __RAID_MODULE_REGISTER(__LINE__)(void) \ 347 { \ 348 raid_bdev_module_list_add(_module); \ 349 } 350 351 bool raid_bdev_io_complete_part(struct raid_bdev_io *raid_io, uint64_t completed, 352 enum spdk_bdev_io_status status); 353 void raid_bdev_queue_io_wait(struct raid_bdev_io *raid_io, struct spdk_bdev *bdev, 354 struct spdk_io_channel *ch, spdk_bdev_io_wait_cb cb_fn); 355 void raid_bdev_io_complete(struct raid_bdev_io *raid_io, enum spdk_bdev_io_status status); 356 void raid_bdev_module_stop_done(struct raid_bdev *raid_bdev); 357 struct spdk_io_channel *raid_bdev_channel_get_base_channel(struct raid_bdev_io_channel *raid_ch, 358 uint8_t idx); 359 void *raid_bdev_channel_get_module_ctx(struct raid_bdev_io_channel *raid_ch); 360 void raid_bdev_process_request_complete(struct raid_bdev_process_request *process_req, int status); 361 void raid_bdev_io_init(struct raid_bdev_io *raid_io, struct raid_bdev_io_channel *raid_ch, 362 enum spdk_bdev_io_type type, uint64_t offset_blocks, 363 uint64_t num_blocks, struct iovec *iovs, int iovcnt, void *md_buf, 364 struct spdk_memory_domain *memory_domain, void *memory_domain_ctx); 365 366 static inline uint8_t 367 raid_bdev_base_bdev_slot(struct raid_base_bdev_info *base_info) 368 { 369 return base_info - base_info->raid_bdev->base_bdev_info; 370 } 371 372 /** 373 * Raid bdev I/O read/write wrapper for spdk_bdev_readv_blocks_ext function. 374 */ 375 int raid_bdev_readv_blocks_ext(struct raid_base_bdev_info *base_info, struct spdk_io_channel *ch, 376 struct iovec *iov, int iovcnt, uint64_t offset_blocks, 377 uint64_t num_blocks, spdk_bdev_io_completion_cb cb, void *cb_arg, 378 struct spdk_bdev_ext_io_opts *opts); 379 380 /** 381 * Raid bdev I/O read/write wrapper for spdk_bdev_writev_blocks_ext function. 382 */ 383 int raid_bdev_writev_blocks_ext(struct raid_base_bdev_info *base_info, struct spdk_io_channel *ch, 384 struct iovec *iov, int iovcnt, uint64_t offset_blocks, 385 uint64_t num_blocks, spdk_bdev_io_completion_cb cb, void *cb_arg, 386 struct spdk_bdev_ext_io_opts *opts); 387 388 /** 389 * Raid bdev I/O read/write wrapper for spdk_bdev_unmap_blocks function. 390 */ 391 static inline int 392 raid_bdev_unmap_blocks(struct raid_base_bdev_info *base_info, struct spdk_io_channel *ch, 393 uint64_t offset_blocks, uint64_t num_blocks, 394 spdk_bdev_io_completion_cb cb, void *cb_arg) 395 { 396 return spdk_bdev_unmap_blocks(base_info->desc, ch, base_info->data_offset + offset_blocks, 397 num_blocks, cb, cb_arg); 398 } 399 400 /** 401 * Raid bdev I/O read/write wrapper for spdk_bdev_flush_blocks function. 402 */ 403 static inline int 404 raid_bdev_flush_blocks(struct raid_base_bdev_info *base_info, struct spdk_io_channel *ch, 405 uint64_t offset_blocks, uint64_t num_blocks, 406 spdk_bdev_io_completion_cb cb, void *cb_arg) 407 { 408 return spdk_bdev_flush_blocks(base_info->desc, ch, base_info->data_offset + offset_blocks, 409 num_blocks, cb, cb_arg); 410 } 411 412 /* 413 * Definitions related to raid bdev superblock 414 */ 415 416 #define RAID_BDEV_SB_VERSION_MAJOR 1 417 #define RAID_BDEV_SB_VERSION_MINOR 0 418 419 #define RAID_BDEV_SB_NAME_SIZE 64 420 421 enum raid_bdev_sb_base_bdev_state { 422 RAID_SB_BASE_BDEV_MISSING = 0, 423 RAID_SB_BASE_BDEV_CONFIGURED = 1, 424 RAID_SB_BASE_BDEV_FAILED = 2, 425 RAID_SB_BASE_BDEV_SPARE = 3, 426 }; 427 428 struct raid_bdev_sb_base_bdev { 429 /* uuid of the base bdev */ 430 struct spdk_uuid uuid; 431 /* offset in blocks from base device start to the start of raid data area */ 432 uint64_t data_offset; 433 /* size in blocks of the base device raid data area */ 434 uint64_t data_size; 435 /* state of the base bdev */ 436 uint32_t state; 437 /* feature/status flags */ 438 uint32_t flags; 439 /* slot number of this base bdev in the raid */ 440 uint8_t slot; 441 442 uint8_t reserved[23]; 443 }; 444 SPDK_STATIC_ASSERT(sizeof(struct raid_bdev_sb_base_bdev) == 64, "incorrect size"); 445 446 struct raid_bdev_superblock { 447 #define RAID_BDEV_SB_SIG "SPDKRAID" 448 uint8_t signature[8]; 449 struct { 450 /* incremented when a breaking change in the superblock structure is made */ 451 uint16_t major; 452 /* incremented for changes in the superblock that are backward compatible */ 453 uint16_t minor; 454 } version; 455 /* length in bytes of the entire superblock */ 456 uint32_t length; 457 /* crc32c checksum of the entire superblock */ 458 uint32_t crc; 459 /* feature/status flags */ 460 uint32_t flags; 461 /* unique id of the raid bdev */ 462 struct spdk_uuid uuid; 463 /* name of the raid bdev */ 464 uint8_t name[RAID_BDEV_SB_NAME_SIZE]; 465 /* size of the raid bdev in blocks */ 466 uint64_t raid_size; 467 /* the raid bdev block size - must be the same for all base bdevs */ 468 uint32_t block_size; 469 /* the raid level */ 470 uint32_t level; 471 /* strip (chunk) size in blocks */ 472 uint32_t strip_size; 473 /* state of the raid */ 474 uint32_t state; 475 /* sequence number, incremented on every superblock update */ 476 uint64_t seq_number; 477 /* number of raid base devices */ 478 uint8_t num_base_bdevs; 479 480 uint8_t reserved[118]; 481 482 /* size of the base bdevs array */ 483 uint8_t base_bdevs_size; 484 /* array of base bdev descriptors */ 485 struct raid_bdev_sb_base_bdev base_bdevs[]; 486 }; 487 SPDK_STATIC_ASSERT(sizeof(struct raid_bdev_superblock) == 256, "incorrect size"); 488 489 #define RAID_BDEV_SB_MAX_LENGTH \ 490 SPDK_ALIGN_CEIL((sizeof(struct raid_bdev_superblock) + UINT8_MAX * sizeof(struct raid_bdev_sb_base_bdev)), 0x1000) 491 492 SPDK_STATIC_ASSERT(RAID_BDEV_SB_MAX_LENGTH < RAID_BDEV_MIN_DATA_OFFSET_SIZE, 493 "Incorrect min data offset"); 494 495 typedef void (*raid_bdev_write_sb_cb)(int status, struct raid_bdev *raid_bdev, void *ctx); 496 typedef void (*raid_bdev_load_sb_cb)(const struct raid_bdev_superblock *sb, int status, void *ctx); 497 498 void raid_bdev_init_superblock(struct raid_bdev *raid_bdev); 499 void raid_bdev_write_superblock(struct raid_bdev *raid_bdev, raid_bdev_write_sb_cb cb, 500 void *cb_ctx); 501 int raid_bdev_load_base_bdev_superblock(struct spdk_bdev_desc *desc, struct spdk_io_channel *ch, 502 raid_bdev_load_sb_cb cb, void *cb_ctx); 503 504 struct spdk_raid_bdev_opts { 505 /* Size of the background process window in KiB */ 506 uint32_t process_window_size_kb; 507 }; 508 509 void raid_bdev_get_opts(struct spdk_raid_bdev_opts *opts); 510 int raid_bdev_set_opts(const struct spdk_raid_bdev_opts *opts); 511 512 #endif /* SPDK_BDEV_RAID_INTERNAL_H */ 513