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