1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (C) 2019 Intel Corporation. 3 * All rights reserved. 4 * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved. 5 */ 6 7 #include "bdev_raid.h" 8 9 #include "spdk/env.h" 10 #include "spdk/thread.h" 11 #include "spdk/string.h" 12 #include "spdk/util.h" 13 14 #include "spdk/log.h" 15 16 /* 17 * brief: 18 * raid0_bdev_io_completion function is called by lower layers to notify raid 19 * module that particular bdev_io is completed. 20 * params: 21 * bdev_io - pointer to bdev io submitted to lower layers, like child io 22 * success - bdev_io status 23 * cb_arg - function callback context (parent raid_bdev_io) 24 * returns: 25 * none 26 */ 27 static void 28 raid0_bdev_io_completion(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 29 { 30 struct raid_bdev_io *raid_io = cb_arg; 31 32 spdk_bdev_free_io(bdev_io); 33 34 if (success) { 35 raid_bdev_io_complete(raid_io, SPDK_BDEV_IO_STATUS_SUCCESS); 36 } else { 37 raid_bdev_io_complete(raid_io, SPDK_BDEV_IO_STATUS_FAILED); 38 } 39 } 40 41 static void raid0_submit_rw_request(struct raid_bdev_io *raid_io); 42 43 static void 44 _raid0_submit_rw_request(void *_raid_io) 45 { 46 struct raid_bdev_io *raid_io = _raid_io; 47 48 raid0_submit_rw_request(raid_io); 49 } 50 51 /* 52 * brief: 53 * raid0_submit_rw_request function is used to submit I/O to the correct 54 * member disk for raid0 bdevs. 55 * params: 56 * raid_io 57 * returns: 58 * none 59 */ 60 static void 61 raid0_submit_rw_request(struct raid_bdev_io *raid_io) 62 { 63 struct spdk_bdev_io *bdev_io = spdk_bdev_io_from_ctx(raid_io); 64 struct spdk_bdev_ext_io_opts io_opts = {}; 65 struct raid_bdev_io_channel *raid_ch = raid_io->raid_ch; 66 struct raid_bdev *raid_bdev = raid_io->raid_bdev; 67 uint64_t pd_strip; 68 uint32_t offset_in_strip; 69 uint64_t pd_lba; 70 uint64_t pd_blocks; 71 uint8_t pd_idx; 72 int ret = 0; 73 uint64_t start_strip; 74 uint64_t end_strip; 75 struct raid_base_bdev_info *base_info; 76 struct spdk_io_channel *base_ch; 77 78 start_strip = bdev_io->u.bdev.offset_blocks >> raid_bdev->strip_size_shift; 79 end_strip = (bdev_io->u.bdev.offset_blocks + bdev_io->u.bdev.num_blocks - 1) >> 80 raid_bdev->strip_size_shift; 81 if (start_strip != end_strip && raid_bdev->num_base_bdevs > 1) { 82 assert(false); 83 SPDK_ERRLOG("I/O spans strip boundary!\n"); 84 raid_bdev_io_complete(raid_io, SPDK_BDEV_IO_STATUS_FAILED); 85 return; 86 } 87 88 pd_strip = start_strip / raid_bdev->num_base_bdevs; 89 pd_idx = start_strip % raid_bdev->num_base_bdevs; 90 offset_in_strip = bdev_io->u.bdev.offset_blocks & (raid_bdev->strip_size - 1); 91 pd_lba = (pd_strip << raid_bdev->strip_size_shift) + offset_in_strip; 92 pd_blocks = bdev_io->u.bdev.num_blocks; 93 base_info = &raid_bdev->base_bdev_info[pd_idx]; 94 if (base_info->desc == NULL) { 95 SPDK_ERRLOG("base bdev desc null for pd_idx %u\n", pd_idx); 96 assert(0); 97 } 98 99 /* 100 * Submit child io to bdev layer with using base bdev descriptors, base 101 * bdev lba, base bdev child io length in blocks, buffer, completion 102 * function and function callback context 103 */ 104 assert(raid_ch != NULL); 105 assert(raid_ch->base_channel); 106 base_ch = raid_ch->base_channel[pd_idx]; 107 108 io_opts.size = sizeof(io_opts); 109 io_opts.memory_domain = bdev_io->u.bdev.memory_domain; 110 io_opts.memory_domain_ctx = bdev_io->u.bdev.memory_domain_ctx; 111 io_opts.metadata = bdev_io->u.bdev.md_buf; 112 113 if (bdev_io->type == SPDK_BDEV_IO_TYPE_READ) { 114 ret = spdk_bdev_readv_blocks_ext(base_info->desc, base_ch, 115 bdev_io->u.bdev.iovs, bdev_io->u.bdev.iovcnt, 116 pd_lba, pd_blocks, raid0_bdev_io_completion, 117 raid_io, &io_opts); 118 } else if (bdev_io->type == SPDK_BDEV_IO_TYPE_WRITE) { 119 ret = spdk_bdev_writev_blocks_ext(base_info->desc, base_ch, 120 bdev_io->u.bdev.iovs, bdev_io->u.bdev.iovcnt, 121 pd_lba, pd_blocks, raid0_bdev_io_completion, 122 raid_io, &io_opts); 123 } else { 124 SPDK_ERRLOG("Recvd not supported io type %u\n", bdev_io->type); 125 assert(0); 126 } 127 128 if (ret == -ENOMEM) { 129 raid_bdev_queue_io_wait(raid_io, spdk_bdev_desc_get_bdev(base_info->desc), 130 base_ch, _raid0_submit_rw_request); 131 } else if (ret != 0) { 132 SPDK_ERRLOG("bdev io submit error not due to ENOMEM, it should not happen\n"); 133 assert(false); 134 raid_bdev_io_complete(raid_io, SPDK_BDEV_IO_STATUS_FAILED); 135 } 136 } 137 138 /* raid0 IO range */ 139 struct raid_bdev_io_range { 140 uint64_t strip_size; 141 uint64_t start_strip_in_disk; 142 uint64_t end_strip_in_disk; 143 uint64_t start_offset_in_strip; 144 uint64_t end_offset_in_strip; 145 uint8_t start_disk; 146 uint8_t end_disk; 147 uint8_t n_disks_involved; 148 }; 149 150 static inline void 151 _raid0_get_io_range(struct raid_bdev_io_range *io_range, 152 uint8_t num_base_bdevs, uint64_t strip_size, uint64_t strip_size_shift, 153 uint64_t offset_blocks, uint64_t num_blocks) 154 { 155 uint64_t start_strip; 156 uint64_t end_strip; 157 uint64_t total_blocks; 158 159 io_range->strip_size = strip_size; 160 total_blocks = offset_blocks + num_blocks - (num_blocks > 0); 161 162 /* The start and end strip index in raid0 bdev scope */ 163 start_strip = offset_blocks >> strip_size_shift; 164 end_strip = total_blocks >> strip_size_shift; 165 io_range->start_strip_in_disk = start_strip / num_base_bdevs; 166 io_range->end_strip_in_disk = end_strip / num_base_bdevs; 167 168 /* The first strip may have unaligned start LBA offset. 169 * The end strip may have unaligned end LBA offset. 170 * Strips between them certainly have aligned offset and length to boundaries. 171 */ 172 io_range->start_offset_in_strip = offset_blocks % strip_size; 173 io_range->end_offset_in_strip = total_blocks % strip_size; 174 175 /* The base bdev indexes in which start and end strips are located */ 176 io_range->start_disk = start_strip % num_base_bdevs; 177 io_range->end_disk = end_strip % num_base_bdevs; 178 179 /* Calculate how many base_bdevs are involved in io operation. 180 * Number of base bdevs involved is between 1 and num_base_bdevs. 181 * It will be 1 if the first strip and last strip are the same one. 182 */ 183 io_range->n_disks_involved = spdk_min((end_strip - start_strip + 1), num_base_bdevs); 184 } 185 186 static inline void 187 _raid0_split_io_range(struct raid_bdev_io_range *io_range, uint8_t disk_idx, 188 uint64_t *_offset_in_disk, uint64_t *_nblocks_in_disk) 189 { 190 uint64_t n_strips_in_disk; 191 uint64_t start_offset_in_disk; 192 uint64_t end_offset_in_disk; 193 uint64_t offset_in_disk; 194 uint64_t nblocks_in_disk; 195 uint64_t start_strip_in_disk; 196 uint64_t end_strip_in_disk; 197 198 start_strip_in_disk = io_range->start_strip_in_disk; 199 if (disk_idx < io_range->start_disk) { 200 start_strip_in_disk += 1; 201 } 202 203 end_strip_in_disk = io_range->end_strip_in_disk; 204 if (disk_idx > io_range->end_disk) { 205 end_strip_in_disk -= 1; 206 } 207 208 assert(end_strip_in_disk >= start_strip_in_disk); 209 n_strips_in_disk = end_strip_in_disk - start_strip_in_disk + 1; 210 211 if (disk_idx == io_range->start_disk) { 212 start_offset_in_disk = io_range->start_offset_in_strip; 213 } else { 214 start_offset_in_disk = 0; 215 } 216 217 if (disk_idx == io_range->end_disk) { 218 end_offset_in_disk = io_range->end_offset_in_strip; 219 } else { 220 end_offset_in_disk = io_range->strip_size - 1; 221 } 222 223 offset_in_disk = start_offset_in_disk + start_strip_in_disk * io_range->strip_size; 224 nblocks_in_disk = (n_strips_in_disk - 1) * io_range->strip_size 225 + end_offset_in_disk - start_offset_in_disk + 1; 226 227 SPDK_DEBUGLOG(bdev_raid0, 228 "raid_bdev (strip_size 0x%" PRIx64 ") splits IO to base_bdev (%u) at (0x%" PRIx64 ", 0x%" PRIx64 229 ").\n", 230 io_range->strip_size, disk_idx, offset_in_disk, nblocks_in_disk); 231 232 *_offset_in_disk = offset_in_disk; 233 *_nblocks_in_disk = nblocks_in_disk; 234 } 235 236 static void raid0_submit_null_payload_request(struct raid_bdev_io *raid_io); 237 238 static void 239 _raid0_submit_null_payload_request(void *_raid_io) 240 { 241 struct raid_bdev_io *raid_io = _raid_io; 242 243 raid0_submit_null_payload_request(raid_io); 244 } 245 246 static void 247 raid0_base_io_complete(struct spdk_bdev_io *bdev_io, bool success, void *cb_arg) 248 { 249 struct raid_bdev_io *raid_io = cb_arg; 250 251 raid_bdev_io_complete_part(raid_io, 1, success ? 252 SPDK_BDEV_IO_STATUS_SUCCESS : 253 SPDK_BDEV_IO_STATUS_FAILED); 254 255 spdk_bdev_free_io(bdev_io); 256 } 257 258 /* 259 * brief: 260 * raid0_submit_null_payload_request function submits the next batch of 261 * io requests with range but without payload, like FLUSH and UNMAP, to member disks; 262 * it will submit as many as possible unless one base io request fails with -ENOMEM, 263 * in which case it will queue itself for later submission. 264 * params: 265 * bdev_io - pointer to parent bdev_io on raid bdev device 266 * returns: 267 * none 268 */ 269 static void 270 raid0_submit_null_payload_request(struct raid_bdev_io *raid_io) 271 { 272 struct spdk_bdev_io *bdev_io; 273 struct raid_bdev *raid_bdev; 274 struct raid_bdev_io_range io_range; 275 int ret; 276 struct raid_base_bdev_info *base_info; 277 struct spdk_io_channel *base_ch; 278 279 bdev_io = spdk_bdev_io_from_ctx(raid_io); 280 raid_bdev = raid_io->raid_bdev; 281 282 _raid0_get_io_range(&io_range, raid_bdev->num_base_bdevs, 283 raid_bdev->strip_size, raid_bdev->strip_size_shift, 284 bdev_io->u.bdev.offset_blocks, bdev_io->u.bdev.num_blocks); 285 286 if (raid_io->base_bdev_io_remaining == 0) { 287 raid_io->base_bdev_io_remaining = io_range.n_disks_involved; 288 } 289 290 while (raid_io->base_bdev_io_submitted < io_range.n_disks_involved) { 291 uint8_t disk_idx; 292 uint64_t offset_in_disk; 293 uint64_t nblocks_in_disk; 294 295 /* base_bdev is started from start_disk to end_disk. 296 * It is possible that index of start_disk is larger than end_disk's. 297 */ 298 disk_idx = (io_range.start_disk + raid_io->base_bdev_io_submitted) % raid_bdev->num_base_bdevs; 299 base_info = &raid_bdev->base_bdev_info[disk_idx]; 300 base_ch = raid_io->raid_ch->base_channel[disk_idx]; 301 302 _raid0_split_io_range(&io_range, disk_idx, &offset_in_disk, &nblocks_in_disk); 303 304 switch (bdev_io->type) { 305 case SPDK_BDEV_IO_TYPE_UNMAP: 306 ret = spdk_bdev_unmap_blocks(base_info->desc, base_ch, 307 offset_in_disk, nblocks_in_disk, 308 raid0_base_io_complete, raid_io); 309 break; 310 311 case SPDK_BDEV_IO_TYPE_FLUSH: 312 ret = spdk_bdev_flush_blocks(base_info->desc, base_ch, 313 offset_in_disk, nblocks_in_disk, 314 raid0_base_io_complete, raid_io); 315 break; 316 317 default: 318 SPDK_ERRLOG("submit request, invalid io type with null payload %u\n", bdev_io->type); 319 assert(false); 320 ret = -EIO; 321 } 322 323 if (ret == 0) { 324 raid_io->base_bdev_io_submitted++; 325 } else if (ret == -ENOMEM) { 326 raid_bdev_queue_io_wait(raid_io, spdk_bdev_desc_get_bdev(base_info->desc), 327 base_ch, _raid0_submit_null_payload_request); 328 return; 329 } else { 330 SPDK_ERRLOG("bdev io submit error not due to ENOMEM, it should not happen\n"); 331 assert(false); 332 raid_bdev_io_complete(raid_io, SPDK_BDEV_IO_STATUS_FAILED); 333 return; 334 } 335 } 336 } 337 338 static uint64_t 339 raid0_calculate_blockcnt(struct raid_bdev *raid_bdev) 340 { 341 uint64_t min_blockcnt = UINT64_MAX; 342 struct raid_base_bdev_info *base_info; 343 344 RAID_FOR_EACH_BASE_BDEV(raid_bdev, base_info) { 345 /* Calculate minimum block count from all base bdevs */ 346 min_blockcnt = spdk_min(min_blockcnt, spdk_bdev_desc_get_bdev(base_info->desc)->blockcnt); 347 } 348 349 /* 350 * Take the minimum block count based approach where total block count 351 * of raid bdev is the number of base bdev times the minimum block count 352 * of any base bdev. 353 */ 354 SPDK_DEBUGLOG(bdev_raid0, "min blockcount %" PRIu64 ", numbasedev %u, strip size shift %u\n", 355 min_blockcnt, raid_bdev->num_base_bdevs, raid_bdev->strip_size_shift); 356 357 return ((min_blockcnt >> raid_bdev->strip_size_shift) << 358 raid_bdev->strip_size_shift) * raid_bdev->num_base_bdevs; 359 } 360 361 static int 362 raid0_start(struct raid_bdev *raid_bdev) 363 { 364 raid_bdev->bdev.blockcnt = raid0_calculate_blockcnt(raid_bdev); 365 366 if (raid_bdev->num_base_bdevs > 1) { 367 raid_bdev->bdev.optimal_io_boundary = raid_bdev->strip_size; 368 raid_bdev->bdev.split_on_optimal_io_boundary = true; 369 } else { 370 /* Do not need to split reads/writes on single bdev RAID modules. */ 371 raid_bdev->bdev.optimal_io_boundary = 0; 372 raid_bdev->bdev.split_on_optimal_io_boundary = false; 373 } 374 375 return 0; 376 } 377 378 static void 379 raid0_resize(struct raid_bdev *raid_bdev) 380 { 381 uint64_t blockcnt; 382 int rc; 383 384 blockcnt = raid0_calculate_blockcnt(raid_bdev); 385 386 if (blockcnt == raid_bdev->bdev.blockcnt) { 387 return; 388 } 389 390 SPDK_NOTICELOG("raid0 '%s': min blockcount was changed from %" PRIu64 " to %" PRIu64 "\n", 391 raid_bdev->bdev.name, 392 raid_bdev->bdev.blockcnt, 393 blockcnt); 394 395 rc = spdk_bdev_notify_blockcnt_change(&raid_bdev->bdev, blockcnt); 396 if (rc != 0) { 397 SPDK_ERRLOG("Failed to notify blockcount change\n"); 398 } 399 } 400 401 static struct raid_bdev_module g_raid0_module = { 402 .level = RAID0, 403 .base_bdevs_min = 1, 404 .memory_domains_supported = true, 405 .start = raid0_start, 406 .submit_rw_request = raid0_submit_rw_request, 407 .submit_null_payload_request = raid0_submit_null_payload_request, 408 .resize = raid0_resize, 409 }; 410 RAID_MODULE_REGISTER(&g_raid0_module) 411 412 SPDK_LOG_REGISTER_COMPONENT(bdev_raid0) 413