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