fs/btrfs/zoned.c

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