fs/btrfs/zoned.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/bitops.h>
   4 #include <linux/slab.h>
   5 #include <linux/blkdev.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/atomic.h>
   8 #include <linux/vmalloc.h>
   9 #include "ctree.h"
  10 #include "volumes.h"
  11 #include "zoned.h"
  12 #include "rcu-string.h"
  13 #include "disk-io.h"
  14 #include "block-group.h"
  15 #include "transaction.h"
  16 #include "dev-replace.h"
  17 #include "space-info.h"
  18 #include "fs.h"
  19 #include "accessors.h"
  20 #include "bio.h"
  21
  22 /* Maximum number of zones to report per blkdev_report_zones() call */
  23 #define BTRFS_REPORT_NR_ZONES   4096
  24 /* Invalid allocation pointer value for missing devices */
  25 #define WP_MISSING_DEV ((u64)-1)
  26 /* Pseudo write pointer value for conventional zone */
  27 #define WP_CONVENTIONAL ((u64)-2)
  28
  29 /*
  30  * Location of the first zone of superblock logging zone pairs.
  31  *
  32  * - primary superblock:    0B (zone 0)
  33  * - first copy:          512G (zone starting at that offset)
  34  * - second copy:           4T (zone starting at that offset)
  35  */
  36 #define BTRFS_SB_LOG_PRIMARY_OFFSET     (0ULL)
  37 #define BTRFS_SB_LOG_FIRST_OFFSET       (512ULL * SZ_1G)
  38 #define BTRFS_SB_LOG_SECOND_OFFSET      (4096ULL * SZ_1G)
  39
  40 #define BTRFS_SB_LOG_FIRST_SHIFT        const_ilog2(BTRFS_SB_LOG_FIRST_OFFSET)
  41 #define BTRFS_SB_LOG_SECOND_SHIFT       const_ilog2(BTRFS_SB_LOG_SECOND_OFFSET)
  42
  43 /* Number of superblock log zones */
  44 #define BTRFS_NR_SB_LOG_ZONES 2
  45
  46 /*
  47  * Minimum of active zones we need:
  48  *
  49  * - BTRFS_SUPER_MIRROR_MAX zones for superblock mirrors
  50  * - 3 zones to ensure at least one zone per SYSTEM, META and DATA block group
  51  * - 1 zone for tree-log dedicated block group
  52  * - 1 zone for relocation
  53  */
  54 #define BTRFS_MIN_ACTIVE_ZONES          (BTRFS_SUPER_MIRROR_MAX + 5)
  55
  56 /*
  57  * Minimum / maximum supported zone size. Currently, SMR disks have a zone
  58  * size of 256MiB, and we are expecting ZNS drives to be in the 1-4GiB range.
  59  * We do not expect the zone size to become larger than 8GiB or smaller than
  60  * 4MiB in the near future.
  61  */
  62 #define BTRFS_MAX_ZONE_SIZE             SZ_8G
  63 #define BTRFS_MIN_ZONE_SIZE             SZ_4M
  64
  65 #define SUPER_INFO_SECTORS      ((u64)BTRFS_SUPER_INFO_SIZE >> SECTOR_SHIFT)
  66
  67 static inline bool sb_zone_is_full(const struct blk_zone *zone)
  68 {
  69         return (zone->cond == BLK_ZONE_COND_FULL) ||
  70                 (zone->wp + SUPER_INFO_SECTORS > zone->start + zone->capacity);
  71 }
  72
  73 static int copy_zone_info_cb(struct blk_zone *zone, unsigned int idx, void *data)
  74 {
  75         struct blk_zone *zones = data;
  76
  77         memcpy(&zones[idx], zone, sizeof(*zone));
  78
  79         return 0;
  80 }
  81
  82 static int sb_write_pointer(struct block_device *bdev, struct blk_zone *zones,
  83                             u64 *wp_ret)
  84 {
  85         bool empty[BTRFS_NR_SB_LOG_ZONES];
  86         bool full[BTRFS_NR_SB_LOG_ZONES];
  87         sector_t sector;
  88         int i;
  89
  90         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
  91                 ASSERT(zones[i].type != BLK_ZONE_TYPE_CONVENTIONAL);
  92                 empty[i] = (zones[i].cond == BLK_ZONE_COND_EMPTY);
  93                 full[i] = sb_zone_is_full(&zones[i]);
  94         }
  95
  96         /*
  97          * Possible states of log buffer zones
  98          *
  99          *           Empty[0]  In use[0]  Full[0]
 100          * Empty[1]         *          0        1
 101          * In use[1]        x          x        1
 102          * Full[1]          0          0        C
 103          *
 104          * Log position:
 105          *   *: Special case, no superblock is written
 106          *   0: Use write pointer of zones[0]
 107          *   1: Use write pointer of zones[1]
 108          *   C: Compare super blocks from zones[0] and zones[1], use the latest
 109          *      one determined by generation
 110          *   x: Invalid state
 111          */
 112
 113         if (empty[0] && empty[1]) {
 114                 /* Special case to distinguish no superblock to read */
 115                 *wp_ret = zones[0].start << SECTOR_SHIFT;
 116                 return -ENOENT;
 117         } else if (full[0] && full[1]) {
 118                 /* Compare two super blocks */
 119                 struct address_space *mapping = bdev->bd_inode->i_mapping;
 120                 struct page *page[BTRFS_NR_SB_LOG_ZONES];
 121                 struct btrfs_super_block *super[BTRFS_NR_SB_LOG_ZONES];
 122                 int i;
 123
 124                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 125                         u64 zone_end = (zones[i].start + zones[i].capacity) << SECTOR_SHIFT;
 126                         u64 bytenr = ALIGN_DOWN(zone_end, BTRFS_SUPER_INFO_SIZE) -
 127                                                 BTRFS_SUPER_INFO_SIZE;
 128
 129                         page[i] = read_cache_page_gfp(mapping,
 130                                         bytenr >> PAGE_SHIFT, GFP_NOFS);
 131                         if (IS_ERR(page[i])) {
 132                                 if (i == 1)
 133                                         btrfs_release_disk_super(super[0]);
 134                                 return PTR_ERR(page[i]);
 135                         }
 136                         super[i] = page_address(page[i]);
 137                 }
 138
 139                 if (btrfs_super_generation(super[0]) >
 140                     btrfs_super_generation(super[1]))
 141                         sector = zones[1].start;
 142                 else
 143                         sector = zones[0].start;
 144
 145                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++)
 146                         btrfs_release_disk_super(super[i]);
 147         } else if (!full[0] && (empty[1] || full[1])) {
 148                 sector = zones[0].wp;
 149         } else if (full[0]) {
 150                 sector = zones[1].wp;
 151         } else {
 152                 return -EUCLEAN;
 153         }
 154         *wp_ret = sector << SECTOR_SHIFT;
 155         return 0;
 156 }
 157
 158 /*
 159  * Get the first zone number of the superblock mirror
 160  */
 161 static inline u32 sb_zone_number(int shift, int mirror)
 162 {
 163         u64 zone = U64_MAX;
 164
 165         ASSERT(mirror < BTRFS_SUPER_MIRROR_MAX);
 166         switch (mirror) {
 167         case 0: zone = 0; break;
 168         case 1: zone = 1ULL << (BTRFS_SB_LOG_FIRST_SHIFT - shift); break;
 169         case 2: zone = 1ULL << (BTRFS_SB_LOG_SECOND_SHIFT - shift); break;
 170         }
 171
 172         ASSERT(zone <= U32_MAX);
 173
 174         return (u32)zone;
 175 }
 176
 177 static inline sector_t zone_start_sector(u32 zone_number,
 178                                          struct block_device *bdev)
 179 {
 180         return (sector_t)zone_number << ilog2(bdev_zone_sectors(bdev));
 181 }
 182
 183 static inline u64 zone_start_physical(u32 zone_number,
 184                                       struct btrfs_zoned_device_info *zone_info)
 185 {
 186         return (u64)zone_number << zone_info->zone_size_shift;
 187 }
 188
 189 /*
 190  * Emulate blkdev_report_zones() for a non-zoned device. It slices up the block
 191  * device into static sized chunks and fake a conventional zone on each of
 192  * them.
 193  */
 194 static int emulate_report_zones(struct btrfs_device *device, u64 pos,
 195                                 struct blk_zone *zones, unsigned int nr_zones)
 196 {
 197         const sector_t zone_sectors = device->fs_info->zone_size >> SECTOR_SHIFT;
 198         sector_t bdev_size = bdev_nr_sectors(device->bdev);
 199         unsigned int i;
 200
 201         pos >>= SECTOR_SHIFT;
 202         for (i = 0; i < nr_zones; i++) {
 203                 zones[i].start = i * zone_sectors + pos;
 204                 zones[i].len = zone_sectors;
 205                 zones[i].capacity = zone_sectors;
 206                 zones[i].wp = zones[i].start + zone_sectors;
 207                 zones[i].type = BLK_ZONE_TYPE_CONVENTIONAL;
 208                 zones[i].cond = BLK_ZONE_COND_NOT_WP;
 209
 210                 if (zones[i].wp >= bdev_size) {
 211                         i++;
 212                         break;
 213                 }
 214         }
 215
 216         return i;
 217 }
 218
 219 static int btrfs_get_dev_zones(struct btrfs_device *device, u64 pos,
 220                                struct blk_zone *zones, unsigned int *nr_zones)
 221 {
 222         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 223         int ret;
 224
 225         if (!*nr_zones)
 226                 return 0;
 227
 228         if (!bdev_is_zoned(device->bdev)) {
 229                 ret = emulate_report_zones(device, pos, zones, *nr_zones);
 230                 *nr_zones = ret;
 231                 return 0;
 232         }
 233
 234         /* Check cache */
 235         if (zinfo->zone_cache) {
 236                 unsigned int i;
 237                 u32 zno;
 238
 239                 ASSERT(IS_ALIGNED(pos, zinfo->zone_size));
 240                 zno = pos >> zinfo->zone_size_shift;
 241                 /*
 242                  * We cannot report zones beyond the zone end. So, it is OK to
 243                  * cap *nr_zones to at the end.
 244                  */
 245                 *nr_zones = min_t(u32, *nr_zones, zinfo->nr_zones - zno);
 246
 247                 for (i = 0; i < *nr_zones; i++) {
 248                         struct blk_zone *zone_info;
 249
 250                         zone_info = &zinfo->zone_cache[zno + i];
 251                         if (!zone_info->len)
 252                                 break;
 253                 }
 254
 255                 if (i == *nr_zones) {
 256                         /* Cache hit on all the zones */
 257                         memcpy(zones, zinfo->zone_cache + zno,
 258                                sizeof(*zinfo->zone_cache) * *nr_zones);
 259                         return 0;
 260                 }
 261         }
 262
 263         ret = blkdev_report_zones(device->bdev, pos >> SECTOR_SHIFT, *nr_zones,
 264                                   copy_zone_info_cb, zones);
 265         if (ret < 0) {
 266                 btrfs_err_in_rcu(device->fs_info,
 267                                  "zoned: failed to read zone %llu on %s (devid %llu)",
 268                                  pos, rcu_str_deref(device->name),
 269                                  device->devid);
 270                 return ret;
 271         }
 272         *nr_zones = ret;
 273         if (!ret)
 274                 return -EIO;
 275
 276         /* Populate cache */
 277         if (zinfo->zone_cache) {
 278                 u32 zno = pos >> zinfo->zone_size_shift;
 279
 280                 memcpy(zinfo->zone_cache + zno, zones,
 281                        sizeof(*zinfo->zone_cache) * *nr_zones);
 282         }
 283
 284         return 0;
 285 }
 286
 287 /* The emulated zone size is determined from the size of device extent */
 288 static int calculate_emulated_zone_size(struct btrfs_fs_info *fs_info)
 289 {
 290         struct btrfs_path *path;
 291         struct btrfs_root *root = fs_info->dev_root;
 292         struct btrfs_key key;
 293         struct extent_buffer *leaf;
 294         struct btrfs_dev_extent *dext;
 295         int ret = 0;
 296
 297         key.objectid = 1;
 298         key.type = BTRFS_DEV_EXTENT_KEY;
 299         key.offset = 0;
 300
 301         path = btrfs_alloc_path();
 302         if (!path)
 303                 return -ENOMEM;
 304
 305         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 306         if (ret < 0)
 307                 goto out;
 308
 309         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 310                 ret = btrfs_next_leaf(root, path);
 311                 if (ret < 0)
 312                         goto out;
 313                 /* No dev extents at all? Not good */
 314                 if (ret > 0) {
 315                         ret = -EUCLEAN;
 316                         goto out;
 317                 }
 318         }
 319
 320         leaf = path->nodes[0];
 321         dext = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
 322         fs_info->zone_size = btrfs_dev_extent_length(leaf, dext);
 323         ret = 0;
 324
 325 out:
 326         btrfs_free_path(path);
 327
 328         return ret;
 329 }
 330
 331 int btrfs_get_dev_zone_info_all_devices(struct btrfs_fs_info *fs_info)
 332 {
 333         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 334         struct btrfs_device *device;
 335         int ret = 0;
 336
 337         /* fs_info->zone_size might not set yet. Use the incomapt flag here. */
 338         if (!btrfs_fs_incompat(fs_info, ZONED))
 339                 return 0;
 340
 341         mutex_lock(&fs_devices->device_list_mutex);
 342         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 343                 /* We can skip reading of zone info for missing devices */
 344                 if (!device->bdev)
 345                         continue;
 346
 347                 ret = btrfs_get_dev_zone_info(device, true);
 348                 if (ret)
 349                         break;
 350         }
 351         mutex_unlock(&fs_devices->device_list_mutex);
 352
 353         return ret;
 354 }
 355
 356 int btrfs_get_dev_zone_info(struct btrfs_device *device, bool populate_cache)
 357 {
 358         struct btrfs_fs_info *fs_info = device->fs_info;
 359         struct btrfs_zoned_device_info *zone_info = NULL;
 360         struct block_device *bdev = device->bdev;
 361         unsigned int max_active_zones;
 362         unsigned int nactive;
 363         sector_t nr_sectors;
 364         sector_t sector = 0;
 365         struct blk_zone *zones = NULL;
 366         unsigned int i, nreported = 0, nr_zones;
 367         sector_t zone_sectors;
 368         char *model, *emulated;
 369         int ret;
 370
 371         /*
 372          * Cannot use btrfs_is_zoned here, since fs_info::zone_size might not
 373          * yet be set.
 374          */
 375         if (!btrfs_fs_incompat(fs_info, ZONED))
 376                 return 0;
 377
 378         if (device->zone_info)
 379                 return 0;
 380
 381         zone_info = kzalloc(sizeof(*zone_info), GFP_KERNEL);
 382         if (!zone_info)
 383                 return -ENOMEM;
 384
 385         device->zone_info = zone_info;
 386
 387         if (!bdev_is_zoned(bdev)) {
 388                 if (!fs_info->zone_size) {
 389                         ret = calculate_emulated_zone_size(fs_info);
 390                         if (ret)
 391                                 goto out;
 392                 }
 393
 394                 ASSERT(fs_info->zone_size);
 395                 zone_sectors = fs_info->zone_size >> SECTOR_SHIFT;
 396         } else {
 397                 zone_sectors = bdev_zone_sectors(bdev);
 398         }
 399
 400         ASSERT(is_power_of_two_u64(zone_sectors));
 401         zone_info->zone_size = zone_sectors << SECTOR_SHIFT;
 402
 403         /* We reject devices with a zone size larger than 8GB */
 404         if (zone_info->zone_size > BTRFS_MAX_ZONE_SIZE) {
 405                 btrfs_err_in_rcu(fs_info,
 406                 "zoned: %s: zone size %llu larger than supported maximum %llu",
 407                                  rcu_str_deref(device->name),
 408                                  zone_info->zone_size, BTRFS_MAX_ZONE_SIZE);
 409                 ret = -EINVAL;
 410                 goto out;
 411         } else if (zone_info->zone_size < BTRFS_MIN_ZONE_SIZE) {
 412                 btrfs_err_in_rcu(fs_info,
 413                 "zoned: %s: zone size %llu smaller than supported minimum %u",
 414                                  rcu_str_deref(device->name),
 415                                  zone_info->zone_size, BTRFS_MIN_ZONE_SIZE);
 416                 ret = -EINVAL;
 417                 goto out;
 418         }
 419
 420         nr_sectors = bdev_nr_sectors(bdev);
 421         zone_info->zone_size_shift = ilog2(zone_info->zone_size);
 422         zone_info->nr_zones = nr_sectors >> ilog2(zone_sectors);
 423         if (!IS_ALIGNED(nr_sectors, zone_sectors))
 424                 zone_info->nr_zones++;
 425
 426         max_active_zones = bdev_max_active_zones(bdev);
 427         if (max_active_zones && max_active_zones < BTRFS_MIN_ACTIVE_ZONES) {
 428                 btrfs_err_in_rcu(fs_info,
 429 "zoned: %s: max active zones %u is too small, need at least %u active zones",
 430                                  rcu_str_deref(device->name), max_active_zones,
 431                                  BTRFS_MIN_ACTIVE_ZONES);
 432                 ret = -EINVAL;
 433                 goto out;
 434         }
 435         zone_info->max_active_zones = max_active_zones;
 436
 437         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 438         if (!zone_info->seq_zones) {
 439                 ret = -ENOMEM;
 440                 goto out;
 441         }
 442
 443         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 444         if (!zone_info->empty_zones) {
 445                 ret = -ENOMEM;
 446                 goto out;
 447         }
 448
 449         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 450         if (!zone_info->active_zones) {
 451                 ret = -ENOMEM;
 452                 goto out;
 453         }
 454
 455         zones = kvcalloc(BTRFS_REPORT_NR_ZONES, sizeof(struct blk_zone), GFP_KERNEL);
 456         if (!zones) {
 457                 ret = -ENOMEM;
 458                 goto out;
 459         }
 460
 461         /*
 462          * Enable zone cache only for a zoned device. On a non-zoned device, we
 463          * fill the zone info with emulated CONVENTIONAL zones, so no need to
 464          * use the cache.
 465          */
 466         if (populate_cache && bdev_is_zoned(device->bdev)) {
 467                 zone_info->zone_cache = vzalloc(sizeof(struct blk_zone) *
 468                                                 zone_info->nr_zones);
 469                 if (!zone_info->zone_cache) {
 470                         btrfs_err_in_rcu(device->fs_info,
 471                                 "zoned: failed to allocate zone cache for %s",
 472                                 rcu_str_deref(device->name));
 473                         ret = -ENOMEM;
 474                         goto out;
 475                 }
 476         }
 477
 478         /* Get zones type */
 479         nactive = 0;
 480         while (sector < nr_sectors) {
 481                 nr_zones = BTRFS_REPORT_NR_ZONES;
 482                 ret = btrfs_get_dev_zones(device, sector << SECTOR_SHIFT, zones,
 483                                           &nr_zones);
 484                 if (ret)
 485                         goto out;
 486
 487                 for (i = 0; i < nr_zones; i++) {
 488                         if (zones[i].type == BLK_ZONE_TYPE_SEQWRITE_REQ)
 489                                 __set_bit(nreported, zone_info->seq_zones);
 490                         switch (zones[i].cond) {
 491                         case BLK_ZONE_COND_EMPTY:
 492                                 __set_bit(nreported, zone_info->empty_zones);
 493                                 break;
 494                         case BLK_ZONE_COND_IMP_OPEN:
 495                         case BLK_ZONE_COND_EXP_OPEN:
 496                         case BLK_ZONE_COND_CLOSED:
 497                                 __set_bit(nreported, zone_info->active_zones);
 498                                 nactive++;
 499                                 break;
 500                         }
 501                         nreported++;
 502                 }
 503                 sector = zones[nr_zones - 1].start + zones[nr_zones - 1].len;
 504         }
 505
 506         if (nreported != zone_info->nr_zones) {
 507                 btrfs_err_in_rcu(device->fs_info,
 508                                  "inconsistent number of zones on %s (%u/%u)",
 509                                  rcu_str_deref(device->name), nreported,
 510                                  zone_info->nr_zones);
 511                 ret = -EIO;
 512                 goto out;
 513         }
 514
 515         if (max_active_zones) {
 516                 if (nactive > max_active_zones) {
 517                         btrfs_err_in_rcu(device->fs_info,
 518                         "zoned: %u active zones on %s exceeds max_active_zones %u",
 519                                          nactive, rcu_str_deref(device->name),
 520                                          max_active_zones);
 521                         ret = -EIO;
 522                         goto out;
 523                 }
 524                 atomic_set(&zone_info->active_zones_left,
 525                            max_active_zones - nactive);
 526                 set_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags);
 527         }
 528
 529         /* Validate superblock log */
 530         nr_zones = BTRFS_NR_SB_LOG_ZONES;
 531         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 532                 u32 sb_zone;
 533                 u64 sb_wp;
 534                 int sb_pos = BTRFS_NR_SB_LOG_ZONES * i;
 535
 536                 sb_zone = sb_zone_number(zone_info->zone_size_shift, i);
 537                 if (sb_zone + 1 >= zone_info->nr_zones)
 538                         continue;
 539
 540                 ret = btrfs_get_dev_zones(device,
 541                                           zone_start_physical(sb_zone, zone_info),
 542                                           &zone_info->sb_zones[sb_pos],
 543                                           &nr_zones);
 544                 if (ret)
 545                         goto out;
 546
 547                 if (nr_zones != BTRFS_NR_SB_LOG_ZONES) {
 548                         btrfs_err_in_rcu(device->fs_info,
 549         "zoned: failed to read super block log zone info at devid %llu zone %u",
 550                                          device->devid, sb_zone);
 551                         ret = -EUCLEAN;
 552                         goto out;
 553                 }
 554
 555                 /*
 556                  * If zones[0] is conventional, always use the beginning of the
 557                  * zone to record superblock. No need to validate in that case.
 558                  */
 559                 if (zone_info->sb_zones[BTRFS_NR_SB_LOG_ZONES * i].type ==
 560                     BLK_ZONE_TYPE_CONVENTIONAL)
 561                         continue;
 562
 563                 ret = sb_write_pointer(device->bdev,
 564                                        &zone_info->sb_zones[sb_pos], &sb_wp);
 565                 if (ret != -ENOENT && ret) {
 566                         btrfs_err_in_rcu(device->fs_info,
 567                         "zoned: super block log zone corrupted devid %llu zone %u",
 568                                          device->devid, sb_zone);
 569                         ret = -EUCLEAN;
 570                         goto out;
 571                 }
 572         }
 573
 574
 575         kvfree(zones);
 576
 577         switch (bdev_zoned_model(bdev)) {
 578         case BLK_ZONED_HM:
 579                 model = "host-managed zoned";
 580                 emulated = "";
 581                 break;
 582         case BLK_ZONED_HA:
 583                 model = "host-aware zoned";
 584                 emulated = "";
 585                 break;
 586         case BLK_ZONED_NONE:
 587                 model = "regular";
 588                 emulated = "emulated ";
 589                 break;
 590         default:
 591                 /* Just in case */
 592                 btrfs_err_in_rcu(fs_info, "zoned: unsupported model %d on %s",
 593                                  bdev_zoned_model(bdev),
 594                                  rcu_str_deref(device->name));
 595                 ret = -EOPNOTSUPP;
 596                 goto out_free_zone_info;
 597         }
 598
 599         btrfs_info_in_rcu(fs_info,
 600                 "%s block device %s, %u %szones of %llu bytes",
 601                 model, rcu_str_deref(device->name), zone_info->nr_zones,
 602                 emulated, zone_info->zone_size);
 603
 604         return 0;
 605
 606 out:
 607         kvfree(zones);
 608 out_free_zone_info:
 609         btrfs_destroy_dev_zone_info(device);
 610
 611         return ret;
 612 }
 613
 614 void btrfs_destroy_dev_zone_info(struct btrfs_device *device)
 615 {
 616         struct btrfs_zoned_device_info *zone_info = device->zone_info;
 617
 618         if (!zone_info)
 619                 return;
 620
 621         bitmap_free(zone_info->active_zones);
 622         bitmap_free(zone_info->seq_zones);
 623         bitmap_free(zone_info->empty_zones);
 624         vfree(zone_info->zone_cache);
 625         kfree(zone_info);
 626         device->zone_info = NULL;
 627 }
 628
 629 struct btrfs_zoned_device_info *btrfs_clone_dev_zone_info(struct btrfs_device *orig_dev)
 630 {
 631         struct btrfs_zoned_device_info *zone_info;
 632
 633         zone_info = kmemdup(orig_dev->zone_info, sizeof(*zone_info), GFP_KERNEL);
 634         if (!zone_info)
 635                 return NULL;
 636
 637         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 638         if (!zone_info->seq_zones)
 639                 goto out;
 640
 641         bitmap_copy(zone_info->seq_zones, orig_dev->zone_info->seq_zones,
 642                     zone_info->nr_zones);
 643
 644         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 645         if (!zone_info->empty_zones)
 646                 goto out;
 647
 648         bitmap_copy(zone_info->empty_zones, orig_dev->zone_info->empty_zones,
 649                     zone_info->nr_zones);
 650
 651         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 652         if (!zone_info->active_zones)
 653                 goto out;
 654
 655         bitmap_copy(zone_info->active_zones, orig_dev->zone_info->active_zones,
 656                     zone_info->nr_zones);
 657         zone_info->zone_cache = NULL;
 658
 659         return zone_info;
 660
 661 out:
 662         bitmap_free(zone_info->seq_zones);
 663         bitmap_free(zone_info->empty_zones);
 664         bitmap_free(zone_info->active_zones);
 665         kfree(zone_info);
 666         return NULL;
 667 }
 668
 669 int btrfs_get_dev_zone(struct btrfs_device *device, u64 pos,
 670                        struct blk_zone *zone)
 671 {
 672         unsigned int nr_zones = 1;
 673         int ret;
 674
 675         ret = btrfs_get_dev_zones(device, pos, zone, &nr_zones);
 676         if (ret != 0 || !nr_zones)
 677                 return ret ? ret : -EIO;
 678
 679         return 0;
 680 }
 681
 682 static int btrfs_check_for_zoned_device(struct btrfs_fs_info *fs_info)
 683 {
 684         struct btrfs_device *device;
 685
 686         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 687                 if (device->bdev &&
 688                     bdev_zoned_model(device->bdev) == BLK_ZONED_HM) {
 689                         btrfs_err(fs_info,
 690                                 "zoned: mode not enabled but zoned device found: %pg",
 691                                 device->bdev);
 692                         return -EINVAL;
 693                 }
 694         }
 695
 696         return 0;
 697 }
 698
 699 int btrfs_check_zoned_mode(struct btrfs_fs_info *fs_info)
 700 {
 701         struct queue_limits *lim = &fs_info->limits;
 702         struct btrfs_device *device;
 703         u64 zone_size = 0;
 704         int ret;
 705
 706         /*
 707          * Host-Managed devices can't be used without the ZONED flag.  With the
 708          * ZONED all devices can be used, using zone emulation if required.
 709          */
 710         if (!btrfs_fs_incompat(fs_info, ZONED))
 711                 return btrfs_check_for_zoned_device(fs_info);
 712
 713         blk_set_stacking_limits(lim);
 714
 715         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 716                 struct btrfs_zoned_device_info *zone_info = device->zone_info;
 717
 718                 if (!device->bdev)
 719                         continue;
 720
 721                 if (!zone_size) {
 722                         zone_size = zone_info->zone_size;
 723                 } else if (zone_info->zone_size != zone_size) {
 724                         btrfs_err(fs_info,
 725                 "zoned: unequal block device zone sizes: have %llu found %llu",
 726                                   zone_info->zone_size, zone_size);
 727                         return -EINVAL;
 728                 }
 729
 730                 /*
 731                  * With the zoned emulation, we can have non-zoned device on the
 732                  * zoned mode. In this case, we don't have a valid max zone
 733                  * append size.
 734                  */
 735                 if (bdev_is_zoned(device->bdev)) {
 736                         blk_stack_limits(lim,
 737                                          &bdev_get_queue(device->bdev)->limits,
 738                                          0);
 739                 }
 740         }
 741
 742         /*
 743          * stripe_size is always aligned to BTRFS_STRIPE_LEN in
 744          * btrfs_create_chunk(). Since we want stripe_len == zone_size,
 745          * check the alignment here.
 746          */
 747         if (!IS_ALIGNED(zone_size, BTRFS_STRIPE_LEN)) {
 748                 btrfs_err(fs_info,
 749                           "zoned: zone size %llu not aligned to stripe %u",
 750                           zone_size, BTRFS_STRIPE_LEN);
 751                 return -EINVAL;
 752         }
 753
 754         if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
 755                 btrfs_err(fs_info, "zoned: mixed block groups not supported");
 756                 return -EINVAL;
 757         }
 758
 759         fs_info->zone_size = zone_size;
 760         /*
 761          * Also limit max_zone_append_size by max_segments * PAGE_SIZE.
 762          * Technically, we can have multiple pages per segment. But, since
 763          * we add the pages one by one to a bio, and cannot increase the
 764          * metadata reservation even if it increases the number of extents, it
 765          * is safe to stick with the limit.
 766          */
 767         fs_info->max_zone_append_size = ALIGN_DOWN(
 768                 min3((u64)lim->max_zone_append_sectors << SECTOR_SHIFT,
 769                      (u64)lim->max_sectors << SECTOR_SHIFT,
 770                      (u64)lim->max_segments << PAGE_SHIFT),
 771                 fs_info->sectorsize);
 772         fs_info->fs_devices->chunk_alloc_policy = BTRFS_CHUNK_ALLOC_ZONED;
 773         if (fs_info->max_zone_append_size < fs_info->max_extent_size)
 774                 fs_info->max_extent_size = fs_info->max_zone_append_size;
 775
 776         /*
 777          * Check mount options here, because we might change fs_info->zoned
 778          * from fs_info->zone_size.
 779          */
 780         ret = btrfs_check_mountopts_zoned(fs_info);
 781         if (ret)
 782                 return ret;
 783
 784         btrfs_info(fs_info, "zoned mode enabled with zone size %llu", zone_size);
 785         return 0;
 786 }
 787
 788 int btrfs_check_mountopts_zoned(struct btrfs_fs_info *info)
 789 {
 790         if (!btrfs_is_zoned(info))
 791                 return 0;
 792
 793         /*
 794          * Space cache writing is not COWed. Disable that to avoid write errors
 795          * in sequential zones.
 796          */
 797         if (btrfs_test_opt(info, SPACE_CACHE)) {
 798                 btrfs_err(info, "zoned: space cache v1 is not supported");
 799                 return -EINVAL;
 800         }
 801
 802         if (btrfs_test_opt(info, NODATACOW)) {
 803                 btrfs_err(info, "zoned: NODATACOW not supported");
 804                 return -EINVAL;
 805         }
 806
 807         return 0;
 808 }
 809
 810 static int sb_log_location(struct block_device *bdev, struct blk_zone *zones,
 811                            int rw, u64 *bytenr_ret)
 812 {
 813         u64 wp;
 814         int ret;
 815
 816         if (zones[0].type == BLK_ZONE_TYPE_CONVENTIONAL) {
 817                 *bytenr_ret = zones[0].start << SECTOR_SHIFT;
 818                 return 0;
 819         }
 820
 821         ret = sb_write_pointer(bdev, zones, &wp);
 822         if (ret != -ENOENT && ret < 0)
 823                 return ret;
 824
 825         if (rw == WRITE) {
 826                 struct blk_zone *reset = NULL;
 827
 828                 if (wp == zones[0].start << SECTOR_SHIFT)
 829                         reset = &zones[0];
 830                 else if (wp == zones[1].start << SECTOR_SHIFT)
 831                         reset = &zones[1];
 832
 833                 if (reset && reset->cond != BLK_ZONE_COND_EMPTY) {
 834                         ASSERT(sb_zone_is_full(reset));
 835
 836                         ret = blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
 837                                                reset->start, reset->len,
 838                                                GFP_NOFS);
 839                         if (ret)
 840                                 return ret;
 841
 842                         reset->cond = BLK_ZONE_COND_EMPTY;
 843                         reset->wp = reset->start;
 844                 }
 845         } else if (ret != -ENOENT) {
 846                 /*
 847                  * For READ, we want the previous one. Move write pointer to
 848                  * the end of a zone, if it is at the head of a zone.
 849                  */
 850                 u64 zone_end = 0;
 851
 852                 if (wp == zones[0].start << SECTOR_SHIFT)
 853                         zone_end = zones[1].start + zones[1].capacity;
 854                 else if (wp == zones[1].start << SECTOR_SHIFT)
 855                         zone_end = zones[0].start + zones[0].capacity;
 856                 if (zone_end)
 857                         wp = ALIGN_DOWN(zone_end << SECTOR_SHIFT,
 858                                         BTRFS_SUPER_INFO_SIZE);
 859
 860                 wp -= BTRFS_SUPER_INFO_SIZE;
 861         }
 862
 863         *bytenr_ret = wp;
 864         return 0;
 865
 866 }
 867
 868 int btrfs_sb_log_location_bdev(struct block_device *bdev, int mirror, int rw,
 869                                u64 *bytenr_ret)
 870 {
 871         struct blk_zone zones[BTRFS_NR_SB_LOG_ZONES];
 872         sector_t zone_sectors;
 873         u32 sb_zone;
 874         int ret;
 875         u8 zone_sectors_shift;
 876         sector_t nr_sectors;
 877         u32 nr_zones;
 878
 879         if (!bdev_is_zoned(bdev)) {
 880                 *bytenr_ret = btrfs_sb_offset(mirror);
 881                 return 0;
 882         }
 883
 884         ASSERT(rw == READ || rw == WRITE);
 885
 886         zone_sectors = bdev_zone_sectors(bdev);
 887         if (!is_power_of_2(zone_sectors))
 888                 return -EINVAL;
 889         zone_sectors_shift = ilog2(zone_sectors);
 890         nr_sectors = bdev_nr_sectors(bdev);
 891         nr_zones = nr_sectors >> zone_sectors_shift;
 892
 893         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
 894         if (sb_zone + 1 >= nr_zones)
 895                 return -ENOENT;
 896
 897         ret = blkdev_report_zones(bdev, zone_start_sector(sb_zone, bdev),
 898                                   BTRFS_NR_SB_LOG_ZONES, copy_zone_info_cb,
 899                                   zones);
 900         if (ret < 0)
 901                 return ret;
 902         if (ret != BTRFS_NR_SB_LOG_ZONES)
 903                 return -EIO;
 904
 905         return sb_log_location(bdev, zones, rw, bytenr_ret);
 906 }
 907
 908 int btrfs_sb_log_location(struct btrfs_device *device, int mirror, int rw,
 909                           u64 *bytenr_ret)
 910 {
 911         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 912         u32 zone_num;
 913
 914         /*
 915          * For a zoned filesystem on a non-zoned block device, use the same
 916          * super block locations as regular filesystem. Doing so, the super
 917          * block can always be retrieved and the zoned flag of the volume
 918          * detected from the super block information.
 919          */
 920         if (!bdev_is_zoned(device->bdev)) {
 921                 *bytenr_ret = btrfs_sb_offset(mirror);
 922                 return 0;
 923         }
 924
 925         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 926         if (zone_num + 1 >= zinfo->nr_zones)
 927                 return -ENOENT;
 928
 929         return sb_log_location(device->bdev,
 930                                &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror],
 931                                rw, bytenr_ret);
 932 }
 933
 934 static inline bool is_sb_log_zone(struct btrfs_zoned_device_info *zinfo,
 935                                   int mirror)
 936 {
 937         u32 zone_num;
 938
 939         if (!zinfo)
 940                 return false;
 941
 942         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 943         if (zone_num + 1 >= zinfo->nr_zones)
 944                 return false;
 945
 946         if (!test_bit(zone_num, zinfo->seq_zones))
 947                 return false;
 948
 949         return true;
 950 }
 951
 952 int btrfs_advance_sb_log(struct btrfs_device *device, int mirror)
 953 {
 954         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 955         struct blk_zone *zone;
 956         int i;
 957
 958         if (!is_sb_log_zone(zinfo, mirror))
 959                 return 0;
 960
 961         zone = &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror];
 962         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 963                 /* Advance the next zone */
 964                 if (zone->cond == BLK_ZONE_COND_FULL) {
 965                         zone++;
 966                         continue;
 967                 }
 968
 969                 if (zone->cond == BLK_ZONE_COND_EMPTY)
 970                         zone->cond = BLK_ZONE_COND_IMP_OPEN;
 971
 972                 zone->wp += SUPER_INFO_SECTORS;
 973
 974                 if (sb_zone_is_full(zone)) {
 975                         /*
 976                          * No room left to write new superblock. Since
 977                          * superblock is written with REQ_SYNC, it is safe to
 978                          * finish the zone now.
 979                          *
 980                          * If the write pointer is exactly at the capacity,
 981                          * explicit ZONE_FINISH is not necessary.
 982                          */
 983                         if (zone->wp != zone->start + zone->capacity) {
 984                                 int ret;
 985
 986                                 ret = blkdev_zone_mgmt(device->bdev,
 987                                                 REQ_OP_ZONE_FINISH, zone->start,
 988                                                 zone->len, GFP_NOFS);
 989                                 if (ret)
 990                                         return ret;
 991                         }
 992
 993                         zone->wp = zone->start + zone->len;
 994                         zone->cond = BLK_ZONE_COND_FULL;
 995                 }
 996                 return 0;
 997         }
 998
 999         /* All the zones are FULL. Should not reach here. */
1000         ASSERT(0);
1001         return -EIO;
1002 }
1003
1004 int btrfs_reset_sb_log_zones(struct block_device *bdev, int mirror)
1005 {
1006         sector_t zone_sectors;
1007         sector_t nr_sectors;
1008         u8 zone_sectors_shift;
1009         u32 sb_zone;
1010         u32 nr_zones;
1011
1012         zone_sectors = bdev_zone_sectors(bdev);
1013         zone_sectors_shift = ilog2(zone_sectors);
1014         nr_sectors = bdev_nr_sectors(bdev);
1015         nr_zones = nr_sectors >> zone_sectors_shift;
1016
1017         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
1018         if (sb_zone + 1 >= nr_zones)
1019                 return -ENOENT;
1020
1021         return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
1022                                 zone_start_sector(sb_zone, bdev),
1023                                 zone_sectors * BTRFS_NR_SB_LOG_ZONES, GFP_NOFS);
1024 }
1025
1026 /*
1027  * Find allocatable zones within a given region.
1028  *
1029  * @device:     the device to allocate a region on
1030  * @hole_start: the position of the hole to allocate the region
1031  * @num_bytes:  size of wanted region
1032  * @hole_end:   the end of the hole
1033  * @return:     position of allocatable zones
1034  *
1035  * Allocatable region should not contain any superblock locations.
1036  */
1037 u64 btrfs_find_allocatable_zones(struct btrfs_device *device, u64 hole_start,
1038                                  u64 hole_end, u64 num_bytes)
1039 {
1040         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1041         const u8 shift = zinfo->zone_size_shift;
1042         u64 nzones = num_bytes >> shift;
1043         u64 pos = hole_start;
1044         u64 begin, end;
1045         bool have_sb;
1046         int i;
1047
1048         ASSERT(IS_ALIGNED(hole_start, zinfo->zone_size));
1049         ASSERT(IS_ALIGNED(num_bytes, zinfo->zone_size));
1050
1051         while (pos < hole_end) {
1052                 begin = pos >> shift;
1053                 end = begin + nzones;
1054
1055                 if (end > zinfo->nr_zones)
1056                         return hole_end;
1057
1058                 /* Check if zones in the region are all empty */
1059                 if (btrfs_dev_is_sequential(device, pos) &&
1060                     find_next_zero_bit(zinfo->empty_zones, end, begin) != end) {
1061                         pos += zinfo->zone_size;
1062                         continue;
1063                 }
1064
1065                 have_sb = false;
1066                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1067                         u32 sb_zone;
1068                         u64 sb_pos;
1069
1070                         sb_zone = sb_zone_number(shift, i);
1071                         if (!(end <= sb_zone ||
1072                               sb_zone + BTRFS_NR_SB_LOG_ZONES <= begin)) {
1073                                 have_sb = true;
1074                                 pos = zone_start_physical(
1075                                         sb_zone + BTRFS_NR_SB_LOG_ZONES, zinfo);
1076                                 break;
1077                         }
1078
1079                         /* We also need to exclude regular superblock positions */
1080                         sb_pos = btrfs_sb_offset(i);
1081                         if (!(pos + num_bytes <= sb_pos ||
1082                               sb_pos + BTRFS_SUPER_INFO_SIZE <= pos)) {
1083                                 have_sb = true;
1084                                 pos = ALIGN(sb_pos + BTRFS_SUPER_INFO_SIZE,
1085                                             zinfo->zone_size);
1086                                 break;
1087                         }
1088                 }
1089                 if (!have_sb)
1090                         break;
1091         }
1092
1093         return pos;
1094 }
1095
1096 static bool btrfs_dev_set_active_zone(struct btrfs_device *device, u64 pos)
1097 {
1098         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1099         unsigned int zno = (pos >> zone_info->zone_size_shift);
1100
1101         /* We can use any number of zones */
1102         if (zone_info->max_active_zones == 0)
1103                 return true;
1104
1105         if (!test_bit(zno, zone_info->active_zones)) {
1106                 /* Active zone left? */
1107                 if (atomic_dec_if_positive(&zone_info->active_zones_left) < 0)
1108                         return false;
1109                 if (test_and_set_bit(zno, zone_info->active_zones)) {
1110                         /* Someone already set the bit */
1111                         atomic_inc(&zone_info->active_zones_left);
1112                 }
1113         }
1114
1115         return true;
1116 }
1117
1118 static void btrfs_dev_clear_active_zone(struct btrfs_device *device, u64 pos)
1119 {
1120         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1121         unsigned int zno = (pos >> zone_info->zone_size_shift);
1122
1123         /* We can use any number of zones */
1124         if (zone_info->max_active_zones == 0)
1125                 return;
1126
1127         if (test_and_clear_bit(zno, zone_info->active_zones))
1128                 atomic_inc(&zone_info->active_zones_left);
1129 }
1130
1131 int btrfs_reset_device_zone(struct btrfs_device *device, u64 physical,
1132                             u64 length, u64 *bytes)
1133 {
1134         int ret;
1135
1136         *bytes = 0;
1137         ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_RESET,
1138                                physical >> SECTOR_SHIFT, length >> SECTOR_SHIFT,
1139                                GFP_NOFS);
1140         if (ret)
1141                 return ret;
1142
1143         *bytes = length;
1144         while (length) {
1145                 btrfs_dev_set_zone_empty(device, physical);
1146                 btrfs_dev_clear_active_zone(device, physical);
1147                 physical += device->zone_info->zone_size;
1148                 length -= device->zone_info->zone_size;
1149         }
1150
1151         return 0;
1152 }
1153
1154 int btrfs_ensure_empty_zones(struct btrfs_device *device, u64 start, u64 size)
1155 {
1156         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1157         const u8 shift = zinfo->zone_size_shift;
1158         unsigned long begin = start >> shift;
1159         unsigned long end = (start + size) >> shift;
1160         u64 pos;
1161         int ret;
1162
1163         ASSERT(IS_ALIGNED(start, zinfo->zone_size));
1164         ASSERT(IS_ALIGNED(size, zinfo->zone_size));
1165
1166         if (end > zinfo->nr_zones)
1167                 return -ERANGE;
1168
1169         /* All the zones are conventional */
1170         if (find_next_bit(zinfo->seq_zones, end, begin) == end)
1171                 return 0;
1172
1173         /* All the zones are sequential and empty */
1174         if (find_next_zero_bit(zinfo->seq_zones, end, begin) == end &&
1175             find_next_zero_bit(zinfo->empty_zones, end, begin) == end)
1176                 return 0;
1177
1178         for (pos = start; pos < start + size; pos += zinfo->zone_size) {
1179                 u64 reset_bytes;
1180
1181                 if (!btrfs_dev_is_sequential(device, pos) ||
1182                     btrfs_dev_is_empty_zone(device, pos))
1183                         continue;
1184
1185                 /* Free regions should be empty */
1186                 btrfs_warn_in_rcu(
1187                         device->fs_info,
1188                 "zoned: resetting device %s (devid %llu) zone %llu for allocation",
1189                         rcu_str_deref(device->name), device->devid, pos >> shift);
1190                 WARN_ON_ONCE(1);
1191
1192                 ret = btrfs_reset_device_zone(device, pos, zinfo->zone_size,
1193                                               &reset_bytes);
1194                 if (ret)
1195                         return ret;
1196         }
1197
1198         return 0;
1199 }
1200
1201 /*
1202  * Calculate an allocation pointer from the extent allocation information
1203  * for a block group consist of conventional zones. It is pointed to the
1204  * end of the highest addressed extent in the block group as an allocation
1205  * offset.
1206  */
1207 static int calculate_alloc_pointer(struct btrfs_block_group *cache,
1208                                    u64 *offset_ret, bool new)
1209 {
1210         struct btrfs_fs_info *fs_info = cache->fs_info;
1211         struct btrfs_root *root;
1212         struct btrfs_path *path;
1213         struct btrfs_key key;
1214         struct btrfs_key found_key;
1215         int ret;
1216         u64 length;
1217
1218         /*
1219          * Avoid  tree lookups for a new block group, there's no use for it.
1220          * It must always be 0.
1221          *
1222          * Also, we have a lock chain of extent buffer lock -> chunk mutex.
1223          * For new a block group, this function is called from
1224          * btrfs_make_block_group() which is already taking the chunk mutex.
1225          * Thus, we cannot call calculate_alloc_pointer() which takes extent
1226          * buffer locks to avoid deadlock.
1227          */
1228         if (new) {
1229                 *offset_ret = 0;
1230                 return 0;
1231         }
1232
1233         path = btrfs_alloc_path();
1234         if (!path)
1235                 return -ENOMEM;
1236
1237         key.objectid = cache->start + cache->length;
1238         key.type = 0;
1239         key.offset = 0;
1240
1241         root = btrfs_extent_root(fs_info, key.objectid);
1242         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1243         /* We should not find the exact match */
1244         if (!ret)
1245                 ret = -EUCLEAN;
1246         if (ret < 0)
1247                 goto out;
1248
1249         ret = btrfs_previous_extent_item(root, path, cache->start);
1250         if (ret) {
1251                 if (ret == 1) {
1252                         ret = 0;
1253                         *offset_ret = 0;
1254                 }
1255                 goto out;
1256         }
1257
1258         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
1259
1260         if (found_key.type == BTRFS_EXTENT_ITEM_KEY)
1261                 length = found_key.offset;
1262         else
1263                 length = fs_info->nodesize;
1264
1265         if (!(found_key.objectid >= cache->start &&
1266                found_key.objectid + length <= cache->start + cache->length)) {
1267                 ret = -EUCLEAN;
1268                 goto out;
1269         }
1270         *offset_ret = found_key.objectid + length - cache->start;
1271         ret = 0;
1272
1273 out:
1274         btrfs_free_path(path);
1275         return ret;
1276 }
1277
1278 int btrfs_load_block_group_zone_info(struct btrfs_block_group *cache, bool new)
1279 {
1280         struct btrfs_fs_info *fs_info = cache->fs_info;
1281         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1282         struct extent_map *em;
1283         struct map_lookup *map;
1284         struct btrfs_device *device;
1285         u64 logical = cache->start;
1286         u64 length = cache->length;
1287         int ret;
1288         int i;
1289         unsigned int nofs_flag;
1290         u64 *alloc_offsets = NULL;
1291         u64 *caps = NULL;
1292         u64 *physical = NULL;
1293         unsigned long *active = NULL;
1294         u64 last_alloc = 0;
1295         u32 num_sequential = 0, num_conventional = 0;
1296
1297         if (!btrfs_is_zoned(fs_info))
1298                 return 0;
1299
1300         /* Sanity check */
1301         if (!IS_ALIGNED(length, fs_info->zone_size)) {
1302                 btrfs_err(fs_info,
1303                 "zoned: block group %llu len %llu unaligned to zone size %llu",
1304                           logical, length, fs_info->zone_size);
1305                 return -EIO;
1306         }
1307
1308         /* Get the chunk mapping */
1309         read_lock(&em_tree->lock);
1310         em = lookup_extent_mapping(em_tree, logical, length);
1311         read_unlock(&em_tree->lock);
1312
1313         if (!em)
1314                 return -EINVAL;
1315
1316         map = em->map_lookup;
1317
1318         cache->physical_map = kmemdup(map, map_lookup_size(map->num_stripes), GFP_NOFS);
1319         if (!cache->physical_map) {
1320                 ret = -ENOMEM;
1321                 goto out;
1322         }
1323
1324         alloc_offsets = kcalloc(map->num_stripes, sizeof(*alloc_offsets), GFP_NOFS);
1325         if (!alloc_offsets) {
1326                 ret = -ENOMEM;
1327                 goto out;
1328         }
1329
1330         caps = kcalloc(map->num_stripes, sizeof(*caps), GFP_NOFS);
1331         if (!caps) {
1332                 ret = -ENOMEM;
1333                 goto out;
1334         }
1335
1336         physical = kcalloc(map->num_stripes, sizeof(*physical), GFP_NOFS);
1337         if (!physical) {
1338                 ret = -ENOMEM;
1339                 goto out;
1340         }
1341
1342         active = bitmap_zalloc(map->num_stripes, GFP_NOFS);
1343         if (!active) {
1344                 ret = -ENOMEM;
1345                 goto out;
1346         }
1347
1348         for (i = 0; i < map->num_stripes; i++) {
1349                 bool is_sequential;
1350                 struct blk_zone zone;
1351                 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1352                 int dev_replace_is_ongoing = 0;
1353
1354                 device = map->stripes[i].dev;
1355                 physical[i] = map->stripes[i].physical;
1356
1357                 if (device->bdev == NULL) {
1358                         alloc_offsets[i] = WP_MISSING_DEV;
1359                         continue;
1360                 }
1361
1362                 is_sequential = btrfs_dev_is_sequential(device, physical[i]);
1363                 if (is_sequential)
1364                         num_sequential++;
1365                 else
1366                         num_conventional++;
1367
1368                 /*
1369                  * Consider a zone as active if we can allow any number of
1370                  * active zones.
1371                  */
1372                 if (!device->zone_info->max_active_zones)
1373                         __set_bit(i, active);
1374
1375                 if (!is_sequential) {
1376                         alloc_offsets[i] = WP_CONVENTIONAL;
1377                         continue;
1378                 }
1379
1380                 /*
1381                  * This zone will be used for allocation, so mark this zone
1382                  * non-empty.
1383                  */
1384                 btrfs_dev_clear_zone_empty(device, physical[i]);
1385
1386                 down_read(&dev_replace->rwsem);
1387                 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
1388                 if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL)
1389                         btrfs_dev_clear_zone_empty(dev_replace->tgtdev, physical[i]);
1390                 up_read(&dev_replace->rwsem);
1391
1392                 /*
1393                  * The group is mapped to a sequential zone. Get the zone write
1394                  * pointer to determine the allocation offset within the zone.
1395                  */
1396                 WARN_ON(!IS_ALIGNED(physical[i], fs_info->zone_size));
1397                 nofs_flag = memalloc_nofs_save();
1398                 ret = btrfs_get_dev_zone(device, physical[i], &zone);
1399                 memalloc_nofs_restore(nofs_flag);
1400                 if (ret == -EIO || ret == -EOPNOTSUPP) {
1401                         ret = 0;
1402                         alloc_offsets[i] = WP_MISSING_DEV;
1403                         continue;
1404                 } else if (ret) {
1405                         goto out;
1406                 }
1407
1408                 if (zone.type == BLK_ZONE_TYPE_CONVENTIONAL) {
1409                         btrfs_err_in_rcu(fs_info,
1410         "zoned: unexpected conventional zone %llu on device %s (devid %llu)",
1411                                 zone.start << SECTOR_SHIFT,
1412                                 rcu_str_deref(device->name), device->devid);
1413                         ret = -EIO;
1414                         goto out;
1415                 }
1416
1417                 caps[i] = (zone.capacity << SECTOR_SHIFT);
1418
1419                 switch (zone.cond) {
1420                 case BLK_ZONE_COND_OFFLINE:
1421                 case BLK_ZONE_COND_READONLY:
1422                         btrfs_err(fs_info,
1423                 "zoned: offline/readonly zone %llu on device %s (devid %llu)",
1424                                   physical[i] >> device->zone_info->zone_size_shift,
1425                                   rcu_str_deref(device->name), device->devid);
1426                         alloc_offsets[i] = WP_MISSING_DEV;
1427                         break;
1428                 case BLK_ZONE_COND_EMPTY:
1429                         alloc_offsets[i] = 0;
1430                         break;
1431                 case BLK_ZONE_COND_FULL:
1432                         alloc_offsets[i] = caps[i];
1433                         break;
1434                 default:
1435                         /* Partially used zone */
1436                         alloc_offsets[i] =
1437                                         ((zone.wp - zone.start) << SECTOR_SHIFT);
1438                         __set_bit(i, active);
1439                         break;
1440                 }
1441         }
1442
1443         if (num_sequential > 0)
1444                 set_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1445
1446         if (num_conventional > 0) {
1447                 /* Zone capacity is always zone size in emulation */
1448                 cache->zone_capacity = cache->length;
1449                 ret = calculate_alloc_pointer(cache, &last_alloc, new);
1450                 if (ret) {
1451                         btrfs_err(fs_info,
1452                         "zoned: failed to determine allocation offset of bg %llu",
1453                                   cache->start);
1454                         goto out;
1455                 } else if (map->num_stripes == num_conventional) {
1456                         cache->alloc_offset = last_alloc;
1457                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1458                         goto out;
1459                 }
1460         }
1461
1462         switch (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
1463         case 0: /* single */
1464                 if (alloc_offsets[0] == WP_MISSING_DEV) {
1465                         btrfs_err(fs_info,
1466                         "zoned: cannot recover write pointer for zone %llu",
1467                                 physical[0]);
1468                         ret = -EIO;
1469                         goto out;
1470                 }
1471                 cache->alloc_offset = alloc_offsets[0];
1472                 cache->zone_capacity = caps[0];
1473                 if (test_bit(0, active))
1474                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1475                 break;
1476         case BTRFS_BLOCK_GROUP_DUP:
1477                 if (map->type & BTRFS_BLOCK_GROUP_DATA) {
1478                         btrfs_err(fs_info, "zoned: profile DUP not yet supported on data bg");
1479                         ret = -EINVAL;
1480                         goto out;
1481                 }
1482                 if (alloc_offsets[0] == WP_MISSING_DEV) {
1483                         btrfs_err(fs_info,
1484                         "zoned: cannot recover write pointer for zone %llu",
1485                                 physical[0]);
1486                         ret = -EIO;
1487                         goto out;
1488                 }
1489                 if (alloc_offsets[1] == WP_MISSING_DEV) {
1490                         btrfs_err(fs_info,
1491                         "zoned: cannot recover write pointer for zone %llu",
1492                                 physical[1]);
1493                         ret = -EIO;
1494                         goto out;
1495                 }
1496                 if (alloc_offsets[0] != alloc_offsets[1]) {
1497                         btrfs_err(fs_info,
1498                         "zoned: write pointer offset mismatch of zones in DUP profile");
1499                         ret = -EIO;
1500                         goto out;
1501                 }
1502                 if (test_bit(0, active) != test_bit(1, active)) {
1503                         if (!btrfs_zone_activate(cache)) {
1504                                 ret = -EIO;
1505                                 goto out;
1506                         }
1507                 } else {
1508                         if (test_bit(0, active))
1509                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
1510                                         &cache->runtime_flags);
1511                 }
1512                 cache->alloc_offset = alloc_offsets[0];
1513                 cache->zone_capacity = min(caps[0], caps[1]);
1514                 break;
1515         case BTRFS_BLOCK_GROUP_RAID1:
1516         case BTRFS_BLOCK_GROUP_RAID0:
1517         case BTRFS_BLOCK_GROUP_RAID10:
1518         case BTRFS_BLOCK_GROUP_RAID5:
1519         case BTRFS_BLOCK_GROUP_RAID6:
1520                 /* non-single profiles are not supported yet */
1521         default:
1522                 btrfs_err(fs_info, "zoned: profile %s not yet supported",
1523                           btrfs_bg_type_to_raid_name(map->type));
1524                 ret = -EINVAL;
1525                 goto out;
1526         }
1527
1528 out:
1529         if (cache->alloc_offset > fs_info->zone_size) {
1530                 btrfs_err(fs_info,
1531                         "zoned: invalid write pointer %llu in block group %llu",
1532                         cache->alloc_offset, cache->start);
1533                 ret = -EIO;
1534         }
1535
1536         if (cache->alloc_offset > cache->zone_capacity) {
1537                 btrfs_err(fs_info,
1538 "zoned: invalid write pointer %llu (larger than zone capacity %llu) in block group %llu",
1539                           cache->alloc_offset, cache->zone_capacity,
1540                           cache->start);
1541                 ret = -EIO;
1542         }
1543
1544         /* An extent is allocated after the write pointer */
1545         if (!ret && num_conventional && last_alloc > cache->alloc_offset) {
1546                 btrfs_err(fs_info,
1547                           "zoned: got wrong write pointer in BG %llu: %llu > %llu",
1548                           logical, last_alloc, cache->alloc_offset);
1549                 ret = -EIO;
1550         }
1551
1552         if (!ret) {
1553                 cache->meta_write_pointer = cache->alloc_offset + cache->start;
1554                 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags)) {
1555                         btrfs_get_block_group(cache);
1556                         spin_lock(&fs_info->zone_active_bgs_lock);
1557                         list_add_tail(&cache->active_bg_list,
1558                                       &fs_info->zone_active_bgs);
1559                         spin_unlock(&fs_info->zone_active_bgs_lock);
1560                 }
1561         } else {
1562                 kfree(cache->physical_map);
1563                 cache->physical_map = NULL;
1564         }
1565         bitmap_free(active);
1566         kfree(physical);
1567         kfree(caps);
1568         kfree(alloc_offsets);
1569         free_extent_map(em);
1570
1571         return ret;
1572 }
1573
1574 void btrfs_calc_zone_unusable(struct btrfs_block_group *cache)
1575 {
1576         u64 unusable, free;
1577
1578         if (!btrfs_is_zoned(cache->fs_info))
1579                 return;
1580
1581         WARN_ON(cache->bytes_super != 0);
1582
1583         /* Check for block groups never get activated */
1584         if (test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &cache->fs_info->flags) &&
1585             cache->flags & (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM) &&
1586             !test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags) &&
1587             cache->alloc_offset == 0) {
1588                 unusable = cache->length;
1589                 free = 0;
1590         } else {
1591                 unusable = (cache->alloc_offset - cache->used) +
1592                            (cache->length - cache->zone_capacity);
1593                 free = cache->zone_capacity - cache->alloc_offset;
1594         }
1595
1596         /* We only need ->free_space in ALLOC_SEQ block groups */
1597         cache->cached = BTRFS_CACHE_FINISHED;
1598         cache->free_space_ctl->free_space = free;
1599         cache->zone_unusable = unusable;
1600 }
1601
1602 void btrfs_redirty_list_add(struct btrfs_transaction *trans,
1603                             struct extent_buffer *eb)
1604 {
1605         struct btrfs_fs_info *fs_info = eb->fs_info;
1606
1607         if (!btrfs_is_zoned(fs_info) ||
1608             btrfs_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN) ||
1609             !list_empty(&eb->release_list))
1610                 return;
1611
1612         memzero_extent_buffer(eb, 0, eb->len);
1613         set_bit(EXTENT_BUFFER_NO_CHECK, &eb->bflags);
1614         set_extent_buffer_dirty(eb);
1615         set_extent_bits_nowait(&trans->dirty_pages, eb->start,
1616                                eb->start + eb->len - 1, EXTENT_DIRTY);
1617
1618         spin_lock(&trans->releasing_ebs_lock);
1619         list_add_tail(&eb->release_list, &trans->releasing_ebs);
1620         spin_unlock(&trans->releasing_ebs_lock);
1621         atomic_inc(&eb->refs);
1622 }
1623
1624 void btrfs_free_redirty_list(struct btrfs_transaction *trans)
1625 {
1626         spin_lock(&trans->releasing_ebs_lock);
1627         while (!list_empty(&trans->releasing_ebs)) {
1628                 struct extent_buffer *eb;
1629
1630                 eb = list_first_entry(&trans->releasing_ebs,
1631                                       struct extent_buffer, release_list);
1632                 list_del_init(&eb->release_list);
1633                 free_extent_buffer(eb);
1634         }
1635         spin_unlock(&trans->releasing_ebs_lock);
1636 }
1637
1638 bool btrfs_use_zone_append(struct btrfs_bio *bbio)
1639 {
1640         u64 start = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT);
1641         struct btrfs_inode *inode = bbio->inode;
1642         struct btrfs_fs_info *fs_info = bbio->fs_info;
1643         struct btrfs_block_group *cache;
1644         bool ret = false;
1645
1646         if (!btrfs_is_zoned(fs_info))
1647                 return false;
1648
1649         if (!inode || !is_data_inode(&inode->vfs_inode))
1650                 return false;
1651
1652         if (btrfs_op(&bbio->bio) != BTRFS_MAP_WRITE)
1653                 return false;
1654
1655         /*
1656          * Using REQ_OP_ZONE_APPNED for relocation can break assumptions on the
1657          * extent layout the relocation code has.
1658          * Furthermore we have set aside own block-group from which only the
1659          * relocation "process" can allocate and make sure only one process at a
1660          * time can add pages to an extent that gets relocated, so it's safe to
1661          * use regular REQ_OP_WRITE for this special case.
1662          */
1663         if (btrfs_is_data_reloc_root(inode->root))
1664                 return false;
1665
1666         cache = btrfs_lookup_block_group(fs_info, start);
1667         ASSERT(cache);
1668         if (!cache)
1669                 return false;
1670
1671         ret = !!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1672         btrfs_put_block_group(cache);
1673
1674         return ret;
1675 }
1676
1677 void btrfs_record_physical_zoned(struct btrfs_bio *bbio)
1678 {
1679         const u64 physical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
1680         struct btrfs_ordered_extent *ordered;
1681
1682         ordered = btrfs_lookup_ordered_extent(bbio->inode, bbio->file_offset);
1683         if (WARN_ON(!ordered))
1684                 return;
1685
1686         ordered->physical = physical;
1687         btrfs_put_ordered_extent(ordered);
1688 }
1689
1690 void btrfs_rewrite_logical_zoned(struct btrfs_ordered_extent *ordered)
1691 {
1692         struct btrfs_inode *inode = BTRFS_I(ordered->inode);
1693         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1694         struct extent_map_tree *em_tree;
1695         struct extent_map *em;
1696         struct btrfs_ordered_sum *sum;
1697         u64 orig_logical = ordered->disk_bytenr;
1698         struct map_lookup *map;
1699         u64 physical = ordered->physical;
1700         u64 chunk_start_phys;
1701         u64 logical;
1702
1703         em = btrfs_get_chunk_map(fs_info, orig_logical, 1);
1704         if (IS_ERR(em))
1705                 return;
1706         map = em->map_lookup;
1707         chunk_start_phys = map->stripes[0].physical;
1708
1709         if (WARN_ON_ONCE(map->num_stripes > 1) ||
1710             WARN_ON_ONCE((map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) != 0) ||
1711             WARN_ON_ONCE(physical < chunk_start_phys) ||
1712             WARN_ON_ONCE(physical > chunk_start_phys + em->orig_block_len)) {
1713                 free_extent_map(em);
1714                 return;
1715         }
1716         logical = em->start + (physical - map->stripes[0].physical);
1717         free_extent_map(em);
1718
1719         if (orig_logical == logical)
1720                 return;
1721
1722         ordered->disk_bytenr = logical;
1723
1724         em_tree = &inode->extent_tree;
1725         write_lock(&em_tree->lock);
1726         em = search_extent_mapping(em_tree, ordered->file_offset,
1727                                    ordered->num_bytes);
1728         em->block_start = logical;
1729         free_extent_map(em);
1730         write_unlock(&em_tree->lock);
1731
1732         list_for_each_entry(sum, &ordered->list, list) {
1733                 if (logical < orig_logical)
1734                         sum->bytenr -= orig_logical - logical;
1735                 else
1736                         sum->bytenr += logical - orig_logical;
1737         }
1738 }
1739
1740 bool btrfs_check_meta_write_pointer(struct btrfs_fs_info *fs_info,
1741                                     struct extent_buffer *eb,
1742                                     struct btrfs_block_group **cache_ret)
1743 {
1744         struct btrfs_block_group *cache;
1745         bool ret = true;
1746
1747         if (!btrfs_is_zoned(fs_info))
1748                 return true;
1749
1750         cache = btrfs_lookup_block_group(fs_info, eb->start);
1751         if (!cache)
1752                 return true;
1753
1754         if (cache->meta_write_pointer != eb->start) {
1755                 btrfs_put_block_group(cache);
1756                 cache = NULL;
1757                 ret = false;
1758         } else {
1759                 cache->meta_write_pointer = eb->start + eb->len;
1760         }
1761
1762         *cache_ret = cache;
1763
1764         return ret;
1765 }
1766
1767 void btrfs_revert_meta_write_pointer(struct btrfs_block_group *cache,
1768                                      struct extent_buffer *eb)
1769 {
1770         if (!btrfs_is_zoned(eb->fs_info) || !cache)
1771                 return;
1772
1773         ASSERT(cache->meta_write_pointer == eb->start + eb->len);
1774         cache->meta_write_pointer = eb->start;
1775 }
1776
1777 int btrfs_zoned_issue_zeroout(struct btrfs_device *device, u64 physical, u64 length)
1778 {
1779         if (!btrfs_dev_is_sequential(device, physical))
1780                 return -EOPNOTSUPP;
1781
1782         return blkdev_issue_zeroout(device->bdev, physical >> SECTOR_SHIFT,
1783                                     length >> SECTOR_SHIFT, GFP_NOFS, 0);
1784 }
1785
1786 static int read_zone_info(struct btrfs_fs_info *fs_info, u64 logical,
1787                           struct blk_zone *zone)
1788 {
1789         struct btrfs_io_context *bioc = NULL;
1790         u64 mapped_length = PAGE_SIZE;
1791         unsigned int nofs_flag;
1792         int nmirrors;
1793         int i, ret;
1794
1795         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
1796                                &mapped_length, &bioc);
1797         if (ret || !bioc || mapped_length < PAGE_SIZE) {
1798                 ret = -EIO;
1799                 goto out_put_bioc;
1800         }
1801
1802         if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1803                 ret = -EINVAL;
1804                 goto out_put_bioc;
1805         }
1806
1807         nofs_flag = memalloc_nofs_save();
1808         nmirrors = (int)bioc->num_stripes;
1809         for (i = 0; i < nmirrors; i++) {
1810                 u64 physical = bioc->stripes[i].physical;
1811                 struct btrfs_device *dev = bioc->stripes[i].dev;
1812
1813                 /* Missing device */
1814                 if (!dev->bdev)
1815                         continue;
1816
1817                 ret = btrfs_get_dev_zone(dev, physical, zone);
1818                 /* Failing device */
1819                 if (ret == -EIO || ret == -EOPNOTSUPP)
1820                         continue;
1821                 break;
1822         }
1823         memalloc_nofs_restore(nofs_flag);
1824 out_put_bioc:
1825         btrfs_put_bioc(bioc);
1826         return ret;
1827 }
1828
1829 /*
1830  * Synchronize write pointer in a zone at @physical_start on @tgt_dev, by
1831  * filling zeros between @physical_pos to a write pointer of dev-replace
1832  * source device.
1833  */
1834 int btrfs_sync_zone_write_pointer(struct btrfs_device *tgt_dev, u64 logical,
1835                                     u64 physical_start, u64 physical_pos)
1836 {
1837         struct btrfs_fs_info *fs_info = tgt_dev->fs_info;
1838         struct blk_zone zone;
1839         u64 length;
1840         u64 wp;
1841         int ret;
1842
1843         if (!btrfs_dev_is_sequential(tgt_dev, physical_pos))
1844                 return 0;
1845
1846         ret = read_zone_info(fs_info, logical, &zone);
1847         if (ret)
1848                 return ret;
1849
1850         wp = physical_start + ((zone.wp - zone.start) << SECTOR_SHIFT);
1851
1852         if (physical_pos == wp)
1853                 return 0;
1854
1855         if (physical_pos > wp)
1856                 return -EUCLEAN;
1857
1858         length = wp - physical_pos;
1859         return btrfs_zoned_issue_zeroout(tgt_dev, physical_pos, length);
1860 }
1861
1862 /*
1863  * Activate block group and underlying device zones
1864  *
1865  * @block_group: the block group to activate
1866  *
1867  * Return: true on success, false otherwise
1868  */
1869 bool btrfs_zone_activate(struct btrfs_block_group *block_group)
1870 {
1871         struct btrfs_fs_info *fs_info = block_group->fs_info;
1872         struct btrfs_space_info *space_info = block_group->space_info;
1873         struct map_lookup *map;
1874         struct btrfs_device *device;
1875         u64 physical;
1876         bool ret;
1877         int i;
1878
1879         if (!btrfs_is_zoned(block_group->fs_info))
1880                 return true;
1881
1882         map = block_group->physical_map;
1883
1884         spin_lock(&space_info->lock);
1885         spin_lock(&block_group->lock);
1886         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
1887                 ret = true;
1888                 goto out_unlock;
1889         }
1890
1891         /* No space left */
1892         if (btrfs_zoned_bg_is_full(block_group)) {
1893                 ret = false;
1894                 goto out_unlock;
1895         }
1896
1897         for (i = 0; i < map->num_stripes; i++) {
1898                 device = map->stripes[i].dev;
1899                 physical = map->stripes[i].physical;
1900
1901                 if (device->zone_info->max_active_zones == 0)
1902                         continue;
1903
1904                 if (!btrfs_dev_set_active_zone(device, physical)) {
1905                         /* Cannot activate the zone */
1906                         ret = false;
1907                         goto out_unlock;
1908                 }
1909         }
1910
1911         /* Successfully activated all the zones */
1912         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
1913         WARN_ON(block_group->alloc_offset != 0);
1914         if (block_group->zone_unusable == block_group->length) {
1915                 block_group->zone_unusable = block_group->length - block_group->zone_capacity;
1916                 space_info->bytes_zone_unusable -= block_group->zone_capacity;
1917         }
1918         spin_unlock(&block_group->lock);
1919         btrfs_try_granting_tickets(fs_info, space_info);
1920         spin_unlock(&space_info->lock);
1921
1922         /* For the active block group list */
1923         btrfs_get_block_group(block_group);
1924
1925         spin_lock(&fs_info->zone_active_bgs_lock);
1926         list_add_tail(&block_group->active_bg_list, &fs_info->zone_active_bgs);
1927         spin_unlock(&fs_info->zone_active_bgs_lock);
1928
1929         return true;
1930
1931 out_unlock:
1932         spin_unlock(&block_group->lock);
1933         spin_unlock(&space_info->lock);
1934         return ret;
1935 }
1936
1937 static void wait_eb_writebacks(struct btrfs_block_group *block_group)
1938 {
1939         struct btrfs_fs_info *fs_info = block_group->fs_info;
1940         const u64 end = block_group->start + block_group->length;
1941         struct radix_tree_iter iter;
1942         struct extent_buffer *eb;
1943         void __rcu **slot;
1944
1945         rcu_read_lock();
1946         radix_tree_for_each_slot(slot, &fs_info->buffer_radix, &iter,
1947                                  block_group->start >> fs_info->sectorsize_bits) {
1948                 eb = radix_tree_deref_slot(slot);
1949                 if (!eb)
1950                         continue;
1951                 if (radix_tree_deref_retry(eb)) {
1952                         slot = radix_tree_iter_retry(&iter);
1953                         continue;
1954                 }
1955
1956                 if (eb->start < block_group->start)
1957                         continue;
1958                 if (eb->start >= end)
1959                         break;
1960
1961                 slot = radix_tree_iter_resume(slot, &iter);
1962                 rcu_read_unlock();
1963                 wait_on_extent_buffer_writeback(eb);
1964                 rcu_read_lock();
1965         }
1966         rcu_read_unlock();
1967 }
1968
1969 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written)
1970 {
1971         struct btrfs_fs_info *fs_info = block_group->fs_info;
1972         struct map_lookup *map;
1973         const bool is_metadata = (block_group->flags &
1974                         (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM));
1975         int ret = 0;
1976         int i;
1977
1978         spin_lock(&block_group->lock);
1979         if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
1980                 spin_unlock(&block_group->lock);
1981                 return 0;
1982         }
1983
1984         /* Check if we have unwritten allocated space */
1985         if (is_metadata &&
1986             block_group->start + block_group->alloc_offset > block_group->meta_write_pointer) {
1987                 spin_unlock(&block_group->lock);
1988                 return -EAGAIN;
1989         }
1990
1991         /*
1992          * If we are sure that the block group is full (= no more room left for
1993          * new allocation) and the IO for the last usable block is completed, we
1994          * don't need to wait for the other IOs. This holds because we ensure
1995          * the sequential IO submissions using the ZONE_APPEND command for data
1996          * and block_group->meta_write_pointer for metadata.
1997          */
1998         if (!fully_written) {
1999                 spin_unlock(&block_group->lock);
2000
2001                 ret = btrfs_inc_block_group_ro(block_group, false);
2002                 if (ret)
2003                         return ret;
2004
2005                 /* Ensure all writes in this block group finish */
2006                 btrfs_wait_block_group_reservations(block_group);
2007                 /* No need to wait for NOCOW writers. Zoned mode does not allow that */
2008                 btrfs_wait_ordered_roots(fs_info, U64_MAX, block_group->start,
2009                                          block_group->length);
2010                 /* Wait for extent buffers to be written. */
2011                 if (is_metadata)
2012                         wait_eb_writebacks(block_group);
2013
2014                 spin_lock(&block_group->lock);
2015
2016                 /*
2017                  * Bail out if someone already deactivated the block group, or
2018                  * allocated space is left in the block group.
2019                  */
2020                 if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2021                               &block_group->runtime_flags)) {
2022                         spin_unlock(&block_group->lock);
2023                         btrfs_dec_block_group_ro(block_group);
2024                         return 0;
2025                 }
2026
2027                 if (block_group->reserved) {
2028                         spin_unlock(&block_group->lock);
2029                         btrfs_dec_block_group_ro(block_group);
2030                         return -EAGAIN;
2031                 }
2032         }
2033
2034         clear_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2035         block_group->alloc_offset = block_group->zone_capacity;
2036         block_group->free_space_ctl->free_space = 0;
2037         btrfs_clear_treelog_bg(block_group);
2038         btrfs_clear_data_reloc_bg(block_group);
2039         spin_unlock(&block_group->lock);
2040
2041         map = block_group->physical_map;
2042         for (i = 0; i < map->num_stripes; i++) {
2043                 struct btrfs_device *device = map->stripes[i].dev;
2044                 const u64 physical = map->stripes[i].physical;
2045
2046                 if (device->zone_info->max_active_zones == 0)
2047                         continue;
2048
2049                 ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_FINISH,
2050                                        physical >> SECTOR_SHIFT,
2051                                        device->zone_info->zone_size >> SECTOR_SHIFT,
2052                                        GFP_NOFS);
2053
2054                 if (ret)
2055                         return ret;
2056
2057                 btrfs_dev_clear_active_zone(device, physical);
2058         }
2059
2060         if (!fully_written)
2061                 btrfs_dec_block_group_ro(block_group);
2062
2063         spin_lock(&fs_info->zone_active_bgs_lock);
2064         ASSERT(!list_empty(&block_group->active_bg_list));
2065         list_del_init(&block_group->active_bg_list);
2066         spin_unlock(&fs_info->zone_active_bgs_lock);
2067
2068         /* For active_bg_list */
2069         btrfs_put_block_group(block_group);
2070
2071         clear_and_wake_up_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2072
2073         return 0;
2074 }
2075
2076 int btrfs_zone_finish(struct btrfs_block_group *block_group)
2077 {
2078         if (!btrfs_is_zoned(block_group->fs_info))
2079                 return 0;
2080
2081         return do_zone_finish(block_group, false);
2082 }
2083
2084 bool btrfs_can_activate_zone(struct btrfs_fs_devices *fs_devices, u64 flags)
2085 {
2086         struct btrfs_fs_info *fs_info = fs_devices->fs_info;
2087         struct btrfs_device *device;
2088         bool ret = false;
2089
2090         if (!btrfs_is_zoned(fs_info))
2091                 return true;
2092
2093         /* Check if there is a device with active zones left */
2094         mutex_lock(&fs_info->chunk_mutex);
2095         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
2096                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2097
2098                 if (!device->bdev)
2099                         continue;
2100
2101                 if (!zinfo->max_active_zones) {
2102                         ret = true;
2103                         break;
2104                 }
2105
2106                 switch (flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
2107                 case 0: /* single */
2108                         ret = (atomic_read(&zinfo->active_zones_left) >= 1);
2109                         break;
2110                 case BTRFS_BLOCK_GROUP_DUP:
2111                         ret = (atomic_read(&zinfo->active_zones_left) >= 2);
2112                         break;
2113                 }
2114                 if (ret)
2115                         break;
2116         }
2117         mutex_unlock(&fs_info->chunk_mutex);
2118
2119         if (!ret)
2120                 set_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2121
2122         return ret;
2123 }
2124
2125 void btrfs_zone_finish_endio(struct btrfs_fs_info *fs_info, u64 logical, u64 length)
2126 {
2127         struct btrfs_block_group *block_group;
2128         u64 min_alloc_bytes;
2129
2130         if (!btrfs_is_zoned(fs_info))
2131                 return;
2132
2133         block_group = btrfs_lookup_block_group(fs_info, logical);
2134         ASSERT(block_group);
2135
2136         /* No MIXED_BG on zoned btrfs. */
2137         if (block_group->flags & BTRFS_BLOCK_GROUP_DATA)
2138                 min_alloc_bytes = fs_info->sectorsize;
2139         else
2140                 min_alloc_bytes = fs_info->nodesize;
2141
2142         /* Bail out if we can allocate more data from this block group. */
2143         if (logical + length + min_alloc_bytes <=
2144             block_group->start + block_group->zone_capacity)
2145                 goto out;
2146
2147         do_zone_finish(block_group, true);
2148
2149 out:
2150         btrfs_put_block_group(block_group);
2151 }
2152
2153 static void btrfs_zone_finish_endio_workfn(struct work_struct *work)
2154 {
2155         struct btrfs_block_group *bg =
2156                 container_of(work, struct btrfs_block_group, zone_finish_work);
2157
2158         wait_on_extent_buffer_writeback(bg->last_eb);
2159         free_extent_buffer(bg->last_eb);
2160         btrfs_zone_finish_endio(bg->fs_info, bg->start, bg->length);
2161         btrfs_put_block_group(bg);
2162 }
2163
2164 void btrfs_schedule_zone_finish_bg(struct btrfs_block_group *bg,
2165                                    struct extent_buffer *eb)
2166 {
2167         if (!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &bg->runtime_flags) ||
2168             eb->start + eb->len * 2 <= bg->start + bg->zone_capacity)
2169                 return;
2170
2171         if (WARN_ON(bg->zone_finish_work.func == btrfs_zone_finish_endio_workfn)) {
2172                 btrfs_err(bg->fs_info, "double scheduling of bg %llu zone finishing",
2173                           bg->start);
2174                 return;
2175         }
2176
2177         /* For the work */
2178         btrfs_get_block_group(bg);
2179         atomic_inc(&eb->refs);
2180         bg->last_eb = eb;
2181         INIT_WORK(&bg->zone_finish_work, btrfs_zone_finish_endio_workfn);
2182         queue_work(system_unbound_wq, &bg->zone_finish_work);
2183 }
2184
2185 void btrfs_clear_data_reloc_bg(struct btrfs_block_group *bg)
2186 {
2187         struct btrfs_fs_info *fs_info = bg->fs_info;
2188
2189         spin_lock(&fs_info->relocation_bg_lock);
2190         if (fs_info->data_reloc_bg == bg->start)
2191                 fs_info->data_reloc_bg = 0;
2192         spin_unlock(&fs_info->relocation_bg_lock);
2193 }
2194
2195 void btrfs_free_zone_cache(struct btrfs_fs_info *fs_info)
2196 {
2197         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2198         struct btrfs_device *device;
2199
2200         if (!btrfs_is_zoned(fs_info))
2201                 return;
2202
2203         mutex_lock(&fs_devices->device_list_mutex);
2204         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2205                 if (device->zone_info) {
2206                         vfree(device->zone_info->zone_cache);
2207                         device->zone_info->zone_cache = NULL;
2208                 }
2209         }
2210         mutex_unlock(&fs_devices->device_list_mutex);
2211 }
2212
2213 bool btrfs_zoned_should_reclaim(struct btrfs_fs_info *fs_info)
2214 {
2215         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2216         struct btrfs_device *device;
2217         u64 used = 0;
2218         u64 total = 0;
2219         u64 factor;
2220
2221         ASSERT(btrfs_is_zoned(fs_info));
2222
2223         if (fs_info->bg_reclaim_threshold == 0)
2224                 return false;
2225
2226         mutex_lock(&fs_devices->device_list_mutex);
2227         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2228                 if (!device->bdev)
2229                         continue;
2230
2231                 total += device->disk_total_bytes;
2232                 used += device->bytes_used;
2233         }
2234         mutex_unlock(&fs_devices->device_list_mutex);
2235
2236         factor = div64_u64(used * 100, total);
2237         return factor >= fs_info->bg_reclaim_threshold;
2238 }
2239
2240 void btrfs_zoned_release_data_reloc_bg(struct btrfs_fs_info *fs_info, u64 logical,
2241                                        u64 length)
2242 {
2243         struct btrfs_block_group *block_group;
2244
2245         if (!btrfs_is_zoned(fs_info))
2246                 return;
2247
2248         block_group = btrfs_lookup_block_group(fs_info, logical);
2249         /* It should be called on a previous data relocation block group. */
2250         ASSERT(block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA));
2251
2252         spin_lock(&block_group->lock);
2253         if (!test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))
2254                 goto out;
2255
2256         /* All relocation extents are written. */
2257         if (block_group->start + block_group->alloc_offset == logical + length) {
2258                 /* Now, release this block group for further allocations. */
2259                 clear_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2260                           &block_group->runtime_flags);
2261         }
2262
2263 out:
2264         spin_unlock(&block_group->lock);
2265         btrfs_put_block_group(block_group);
2266 }
2267
2268 int btrfs_zone_finish_one_bg(struct btrfs_fs_info *fs_info)
2269 {
2270         struct btrfs_block_group *block_group;
2271         struct btrfs_block_group *min_bg = NULL;
2272         u64 min_avail = U64_MAX;
2273         int ret;
2274
2275         spin_lock(&fs_info->zone_active_bgs_lock);
2276         list_for_each_entry(block_group, &fs_info->zone_active_bgs,
2277                             active_bg_list) {
2278                 u64 avail;
2279
2280                 spin_lock(&block_group->lock);
2281                 if (block_group->reserved || block_group->alloc_offset == 0 ||
2282                     (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM)) {
2283                         spin_unlock(&block_group->lock);
2284                         continue;
2285                 }
2286
2287                 avail = block_group->zone_capacity - block_group->alloc_offset;
2288                 if (min_avail > avail) {
2289                         if (min_bg)
2290                                 btrfs_put_block_group(min_bg);
2291                         min_bg = block_group;
2292                         min_avail = avail;
2293                         btrfs_get_block_group(min_bg);
2294                 }
2295                 spin_unlock(&block_group->lock);
2296         }
2297         spin_unlock(&fs_info->zone_active_bgs_lock);
2298
2299         if (!min_bg)
2300                 return 0;
2301
2302         ret = btrfs_zone_finish(min_bg);
2303         btrfs_put_block_group(min_bg);
2304
2305         return ret < 0 ? ret : 1;
2306 }
2307
2308 int btrfs_zoned_activate_one_bg(struct btrfs_fs_info *fs_info,
2309                                 struct btrfs_space_info *space_info,
2310                                 bool do_finish)
2311 {
2312         struct btrfs_block_group *bg;
2313         int index;
2314
2315         if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
2316                 return 0;
2317
2318         for (;;) {
2319                 int ret;
2320                 bool need_finish = false;
2321
2322                 down_read(&space_info->groups_sem);
2323                 for (index = 0; index < BTRFS_NR_RAID_TYPES; index++) {
2324                         list_for_each_entry(bg, &space_info->block_groups[index],
2325                                             list) {
2326                                 if (!spin_trylock(&bg->lock))
2327                                         continue;
2328                                 if (btrfs_zoned_bg_is_full(bg) ||
2329                                     test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2330                                              &bg->runtime_flags)) {
2331                                         spin_unlock(&bg->lock);
2332                                         continue;
2333                                 }
2334                                 spin_unlock(&bg->lock);
2335
2336                                 if (btrfs_zone_activate(bg)) {
2337                                         up_read(&space_info->groups_sem);
2338                                         return 1;
2339                                 }
2340
2341                                 need_finish = true;
2342                         }
2343                 }
2344                 up_read(&space_info->groups_sem);
2345
2346                 if (!do_finish || !need_finish)
2347                         break;
2348
2349                 ret = btrfs_zone_finish_one_bg(fs_info);
2350                 if (ret == 0)
2351                         break;
2352                 if (ret < 0)
2353                         return ret;
2354         }
2355
2356         return 0;
2357 }