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 "super.h"
  19 #include "fs.h"
  20 #include "accessors.h"
  21 #include "bio.h"
  22
  23 /* Maximum number of zones to report per blkdev_report_zones() call */
  24 #define BTRFS_REPORT_NR_ZONES   4096
  25 /* Invalid allocation pointer value for missing devices */
  26 #define WP_MISSING_DEV ((u64)-1)
  27 /* Pseudo write pointer value for conventional zone */
  28 #define WP_CONVENTIONAL ((u64)-2)
  29
  30 /*
  31  * Location of the first zone of superblock logging zone pairs.
  32  *
  33  * - primary superblock:    0B (zone 0)
  34  * - first copy:          512G (zone starting at that offset)
  35  * - second copy:           4T (zone starting at that offset)
  36  */
  37 #define BTRFS_SB_LOG_PRIMARY_OFFSET     (0ULL)
  38 #define BTRFS_SB_LOG_FIRST_OFFSET       (512ULL * SZ_1G)
  39 #define BTRFS_SB_LOG_SECOND_OFFSET      (4096ULL * SZ_1G)
  40
  41 #define BTRFS_SB_LOG_FIRST_SHIFT        const_ilog2(BTRFS_SB_LOG_FIRST_OFFSET)
  42 #define BTRFS_SB_LOG_SECOND_SHIFT       const_ilog2(BTRFS_SB_LOG_SECOND_OFFSET)
  43
  44 /* Number of superblock log zones */
  45 #define BTRFS_NR_SB_LOG_ZONES 2
  46
  47 /*
  48  * Minimum of active zones we need:
  49  *
  50  * - BTRFS_SUPER_MIRROR_MAX zones for superblock mirrors
  51  * - 3 zones to ensure at least one zone per SYSTEM, META and DATA block group
  52  * - 1 zone for tree-log dedicated block group
  53  * - 1 zone for relocation
  54  */
  55 #define BTRFS_MIN_ACTIVE_ZONES          (BTRFS_SUPER_MIRROR_MAX + 5)
  56
  57 /*
  58  * Minimum / maximum supported zone size. Currently, SMR disks have a zone
  59  * size of 256MiB, and we are expecting ZNS drives to be in the 1-4GiB range.
  60  * We do not expect the zone size to become larger than 8GiB or smaller than
  61  * 4MiB in the near future.
  62  */
  63 #define BTRFS_MAX_ZONE_SIZE             SZ_8G
  64 #define BTRFS_MIN_ZONE_SIZE             SZ_4M
  65
  66 #define SUPER_INFO_SECTORS      ((u64)BTRFS_SUPER_INFO_SIZE >> SECTOR_SHIFT)
  67
  68 static void wait_eb_writebacks(struct btrfs_block_group *block_group);
  69 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written);
  70
  71 static inline bool sb_zone_is_full(const struct blk_zone *zone)
  72 {
  73         return (zone->cond == BLK_ZONE_COND_FULL) ||
  74                 (zone->wp + SUPER_INFO_SECTORS > zone->start + zone->capacity);
  75 }
  76
  77 static int copy_zone_info_cb(struct blk_zone *zone, unsigned int idx, void *data)
  78 {
  79         struct blk_zone *zones = data;
  80
  81         memcpy(&zones[idx], zone, sizeof(*zone));
  82
  83         return 0;
  84 }
  85
  86 static int sb_write_pointer(struct block_device *bdev, struct blk_zone *zones,
  87                             u64 *wp_ret)
  88 {
  89         bool empty[BTRFS_NR_SB_LOG_ZONES];
  90         bool full[BTRFS_NR_SB_LOG_ZONES];
  91         sector_t sector;
  92         int i;
  93
  94         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
  95                 ASSERT(zones[i].type != BLK_ZONE_TYPE_CONVENTIONAL);
  96                 empty[i] = (zones[i].cond == BLK_ZONE_COND_EMPTY);
  97                 full[i] = sb_zone_is_full(&zones[i]);
  98         }
  99
 100         /*
 101          * Possible states of log buffer zones
 102          *
 103          *           Empty[0]  In use[0]  Full[0]
 104          * Empty[1]         *          0        1
 105          * In use[1]        x          x        1
 106          * Full[1]          0          0        C
 107          *
 108          * Log position:
 109          *   *: Special case, no superblock is written
 110          *   0: Use write pointer of zones[0]
 111          *   1: Use write pointer of zones[1]
 112          *   C: Compare super blocks from zones[0] and zones[1], use the latest
 113          *      one determined by generation
 114          *   x: Invalid state
 115          */
 116
 117         if (empty[0] && empty[1]) {
 118                 /* Special case to distinguish no superblock to read */
 119                 *wp_ret = zones[0].start << SECTOR_SHIFT;
 120                 return -ENOENT;
 121         } else if (full[0] && full[1]) {
 122                 /* Compare two super blocks */
 123                 struct address_space *mapping = bdev->bd_inode->i_mapping;
 124                 struct page *page[BTRFS_NR_SB_LOG_ZONES];
 125                 struct btrfs_super_block *super[BTRFS_NR_SB_LOG_ZONES];
 126                 int i;
 127
 128                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 129                         u64 zone_end = (zones[i].start + zones[i].capacity) << SECTOR_SHIFT;
 130                         u64 bytenr = ALIGN_DOWN(zone_end, BTRFS_SUPER_INFO_SIZE) -
 131                                                 BTRFS_SUPER_INFO_SIZE;
 132
 133                         page[i] = read_cache_page_gfp(mapping,
 134                                         bytenr >> PAGE_SHIFT, GFP_NOFS);
 135                         if (IS_ERR(page[i])) {
 136                                 if (i == 1)
 137                                         btrfs_release_disk_super(super[0]);
 138                                 return PTR_ERR(page[i]);
 139                         }
 140                         super[i] = page_address(page[i]);
 141                 }
 142
 143                 if (btrfs_super_generation(super[0]) >
 144                     btrfs_super_generation(super[1]))
 145                         sector = zones[1].start;
 146                 else
 147                         sector = zones[0].start;
 148
 149                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++)
 150                         btrfs_release_disk_super(super[i]);
 151         } else if (!full[0] && (empty[1] || full[1])) {
 152                 sector = zones[0].wp;
 153         } else if (full[0]) {
 154                 sector = zones[1].wp;
 155         } else {
 156                 return -EUCLEAN;
 157         }
 158         *wp_ret = sector << SECTOR_SHIFT;
 159         return 0;
 160 }
 161
 162 /*
 163  * Get the first zone number of the superblock mirror
 164  */
 165 static inline u32 sb_zone_number(int shift, int mirror)
 166 {
 167         u64 zone = U64_MAX;
 168
 169         ASSERT(mirror < BTRFS_SUPER_MIRROR_MAX);
 170         switch (mirror) {
 171         case 0: zone = 0; break;
 172         case 1: zone = 1ULL << (BTRFS_SB_LOG_FIRST_SHIFT - shift); break;
 173         case 2: zone = 1ULL << (BTRFS_SB_LOG_SECOND_SHIFT - shift); break;
 174         }
 175
 176         ASSERT(zone <= U32_MAX);
 177
 178         return (u32)zone;
 179 }
 180
 181 static inline sector_t zone_start_sector(u32 zone_number,
 182                                          struct block_device *bdev)
 183 {
 184         return (sector_t)zone_number << ilog2(bdev_zone_sectors(bdev));
 185 }
 186
 187 static inline u64 zone_start_physical(u32 zone_number,
 188                                       struct btrfs_zoned_device_info *zone_info)
 189 {
 190         return (u64)zone_number << zone_info->zone_size_shift;
 191 }
 192
 193 /*
 194  * Emulate blkdev_report_zones() for a non-zoned device. It slices up the block
 195  * device into static sized chunks and fake a conventional zone on each of
 196  * them.
 197  */
 198 static int emulate_report_zones(struct btrfs_device *device, u64 pos,
 199                                 struct blk_zone *zones, unsigned int nr_zones)
 200 {
 201         const sector_t zone_sectors = device->fs_info->zone_size >> SECTOR_SHIFT;
 202         sector_t bdev_size = bdev_nr_sectors(device->bdev);
 203         unsigned int i;
 204
 205         pos >>= SECTOR_SHIFT;
 206         for (i = 0; i < nr_zones; i++) {
 207                 zones[i].start = i * zone_sectors + pos;
 208                 zones[i].len = zone_sectors;
 209                 zones[i].capacity = zone_sectors;
 210                 zones[i].wp = zones[i].start + zone_sectors;
 211                 zones[i].type = BLK_ZONE_TYPE_CONVENTIONAL;
 212                 zones[i].cond = BLK_ZONE_COND_NOT_WP;
 213
 214                 if (zones[i].wp >= bdev_size) {
 215                         i++;
 216                         break;
 217                 }
 218         }
 219
 220         return i;
 221 }
 222
 223 static int btrfs_get_dev_zones(struct btrfs_device *device, u64 pos,
 224                                struct blk_zone *zones, unsigned int *nr_zones)
 225 {
 226         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 227         int ret;
 228
 229         if (!*nr_zones)
 230                 return 0;
 231
 232         if (!bdev_is_zoned(device->bdev)) {
 233                 ret = emulate_report_zones(device, pos, zones, *nr_zones);
 234                 *nr_zones = ret;
 235                 return 0;
 236         }
 237
 238         /* Check cache */
 239         if (zinfo->zone_cache) {
 240                 unsigned int i;
 241                 u32 zno;
 242
 243                 ASSERT(IS_ALIGNED(pos, zinfo->zone_size));
 244                 zno = pos >> zinfo->zone_size_shift;
 245                 /*
 246                  * We cannot report zones beyond the zone end. So, it is OK to
 247                  * cap *nr_zones to at the end.
 248                  */
 249                 *nr_zones = min_t(u32, *nr_zones, zinfo->nr_zones - zno);
 250
 251                 for (i = 0; i < *nr_zones; i++) {
 252                         struct blk_zone *zone_info;
 253
 254                         zone_info = &zinfo->zone_cache[zno + i];
 255                         if (!zone_info->len)
 256                                 break;
 257                 }
 258
 259                 if (i == *nr_zones) {
 260                         /* Cache hit on all the zones */
 261                         memcpy(zones, zinfo->zone_cache + zno,
 262                                sizeof(*zinfo->zone_cache) * *nr_zones);
 263                         return 0;
 264                 }
 265         }
 266
 267         ret = blkdev_report_zones(device->bdev, pos >> SECTOR_SHIFT, *nr_zones,
 268                                   copy_zone_info_cb, zones);
 269         if (ret < 0) {
 270                 btrfs_err_in_rcu(device->fs_info,
 271                                  "zoned: failed to read zone %llu on %s (devid %llu)",
 272                                  pos, rcu_str_deref(device->name),
 273                                  device->devid);
 274                 return ret;
 275         }
 276         *nr_zones = ret;
 277         if (!ret)
 278                 return -EIO;
 279
 280         /* Populate cache */
 281         if (zinfo->zone_cache) {
 282                 u32 zno = pos >> zinfo->zone_size_shift;
 283
 284                 memcpy(zinfo->zone_cache + zno, zones,
 285                        sizeof(*zinfo->zone_cache) * *nr_zones);
 286         }
 287
 288         return 0;
 289 }
 290
 291 /* The emulated zone size is determined from the size of device extent */
 292 static int calculate_emulated_zone_size(struct btrfs_fs_info *fs_info)
 293 {
 294         struct btrfs_path *path;
 295         struct btrfs_root *root = fs_info->dev_root;
 296         struct btrfs_key key;
 297         struct extent_buffer *leaf;
 298         struct btrfs_dev_extent *dext;
 299         int ret = 0;
 300
 301         key.objectid = 1;
 302         key.type = BTRFS_DEV_EXTENT_KEY;
 303         key.offset = 0;
 304
 305         path = btrfs_alloc_path();
 306         if (!path)
 307                 return -ENOMEM;
 308
 309         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 310         if (ret < 0)
 311                 goto out;
 312
 313         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 314                 ret = btrfs_next_leaf(root, path);
 315                 if (ret < 0)
 316                         goto out;
 317                 /* No dev extents at all? Not good */
 318                 if (ret > 0) {
 319                         ret = -EUCLEAN;
 320                         goto out;
 321                 }
 322         }
 323
 324         leaf = path->nodes[0];
 325         dext = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
 326         fs_info->zone_size = btrfs_dev_extent_length(leaf, dext);
 327         ret = 0;
 328
 329 out:
 330         btrfs_free_path(path);
 331
 332         return ret;
 333 }
 334
 335 int btrfs_get_dev_zone_info_all_devices(struct btrfs_fs_info *fs_info)
 336 {
 337         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 338         struct btrfs_device *device;
 339         int ret = 0;
 340
 341         /* fs_info->zone_size might not set yet. Use the incomapt flag here. */
 342         if (!btrfs_fs_incompat(fs_info, ZONED))
 343                 return 0;
 344
 345         mutex_lock(&fs_devices->device_list_mutex);
 346         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 347                 /* We can skip reading of zone info for missing devices */
 348                 if (!device->bdev)
 349                         continue;
 350
 351                 ret = btrfs_get_dev_zone_info(device, true);
 352                 if (ret)
 353                         break;
 354         }
 355         mutex_unlock(&fs_devices->device_list_mutex);
 356
 357         return ret;
 358 }
 359
 360 int btrfs_get_dev_zone_info(struct btrfs_device *device, bool populate_cache)
 361 {
 362         struct btrfs_fs_info *fs_info = device->fs_info;
 363         struct btrfs_zoned_device_info *zone_info = NULL;
 364         struct block_device *bdev = device->bdev;
 365         unsigned int max_active_zones;
 366         unsigned int nactive;
 367         sector_t nr_sectors;
 368         sector_t sector = 0;
 369         struct blk_zone *zones = NULL;
 370         unsigned int i, nreported = 0, nr_zones;
 371         sector_t zone_sectors;
 372         char *model, *emulated;
 373         int ret;
 374
 375         /*
 376          * Cannot use btrfs_is_zoned here, since fs_info::zone_size might not
 377          * yet be set.
 378          */
 379         if (!btrfs_fs_incompat(fs_info, ZONED))
 380                 return 0;
 381
 382         if (device->zone_info)
 383                 return 0;
 384
 385         zone_info = kzalloc(sizeof(*zone_info), GFP_KERNEL);
 386         if (!zone_info)
 387                 return -ENOMEM;
 388
 389         device->zone_info = zone_info;
 390
 391         if (!bdev_is_zoned(bdev)) {
 392                 if (!fs_info->zone_size) {
 393                         ret = calculate_emulated_zone_size(fs_info);
 394                         if (ret)
 395                                 goto out;
 396                 }
 397
 398                 ASSERT(fs_info->zone_size);
 399                 zone_sectors = fs_info->zone_size >> SECTOR_SHIFT;
 400         } else {
 401                 zone_sectors = bdev_zone_sectors(bdev);
 402         }
 403
 404         ASSERT(is_power_of_two_u64(zone_sectors));
 405         zone_info->zone_size = zone_sectors << SECTOR_SHIFT;
 406
 407         /* We reject devices with a zone size larger than 8GB */
 408         if (zone_info->zone_size > BTRFS_MAX_ZONE_SIZE) {
 409                 btrfs_err_in_rcu(fs_info,
 410                 "zoned: %s: zone size %llu larger than supported maximum %llu",
 411                                  rcu_str_deref(device->name),
 412                                  zone_info->zone_size, BTRFS_MAX_ZONE_SIZE);
 413                 ret = -EINVAL;
 414                 goto out;
 415         } else if (zone_info->zone_size < BTRFS_MIN_ZONE_SIZE) {
 416                 btrfs_err_in_rcu(fs_info,
 417                 "zoned: %s: zone size %llu smaller than supported minimum %u",
 418                                  rcu_str_deref(device->name),
 419                                  zone_info->zone_size, BTRFS_MIN_ZONE_SIZE);
 420                 ret = -EINVAL;
 421                 goto out;
 422         }
 423
 424         nr_sectors = bdev_nr_sectors(bdev);
 425         zone_info->zone_size_shift = ilog2(zone_info->zone_size);
 426         zone_info->nr_zones = nr_sectors >> ilog2(zone_sectors);
 427         if (!IS_ALIGNED(nr_sectors, zone_sectors))
 428                 zone_info->nr_zones++;
 429
 430         max_active_zones = bdev_max_active_zones(bdev);
 431         if (max_active_zones && max_active_zones < BTRFS_MIN_ACTIVE_ZONES) {
 432                 btrfs_err_in_rcu(fs_info,
 433 "zoned: %s: max active zones %u is too small, need at least %u active zones",
 434                                  rcu_str_deref(device->name), max_active_zones,
 435                                  BTRFS_MIN_ACTIVE_ZONES);
 436                 ret = -EINVAL;
 437                 goto out;
 438         }
 439         zone_info->max_active_zones = max_active_zones;
 440
 441         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 442         if (!zone_info->seq_zones) {
 443                 ret = -ENOMEM;
 444                 goto out;
 445         }
 446
 447         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 448         if (!zone_info->empty_zones) {
 449                 ret = -ENOMEM;
 450                 goto out;
 451         }
 452
 453         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 454         if (!zone_info->active_zones) {
 455                 ret = -ENOMEM;
 456                 goto out;
 457         }
 458
 459         zones = kvcalloc(BTRFS_REPORT_NR_ZONES, sizeof(struct blk_zone), GFP_KERNEL);
 460         if (!zones) {
 461                 ret = -ENOMEM;
 462                 goto out;
 463         }
 464
 465         /*
 466          * Enable zone cache only for a zoned device. On a non-zoned device, we
 467          * fill the zone info with emulated CONVENTIONAL zones, so no need to
 468          * use the cache.
 469          */
 470         if (populate_cache && bdev_is_zoned(device->bdev)) {
 471                 zone_info->zone_cache = vcalloc(zone_info->nr_zones,
 472                                                 sizeof(struct blk_zone));
 473                 if (!zone_info->zone_cache) {
 474                         btrfs_err_in_rcu(device->fs_info,
 475                                 "zoned: failed to allocate zone cache for %s",
 476                                 rcu_str_deref(device->name));
 477                         ret = -ENOMEM;
 478                         goto out;
 479                 }
 480         }
 481
 482         /* Get zones type */
 483         nactive = 0;
 484         while (sector < nr_sectors) {
 485                 nr_zones = BTRFS_REPORT_NR_ZONES;
 486                 ret = btrfs_get_dev_zones(device, sector << SECTOR_SHIFT, zones,
 487                                           &nr_zones);
 488                 if (ret)
 489                         goto out;
 490
 491                 for (i = 0; i < nr_zones; i++) {
 492                         if (zones[i].type == BLK_ZONE_TYPE_SEQWRITE_REQ)
 493                                 __set_bit(nreported, zone_info->seq_zones);
 494                         switch (zones[i].cond) {
 495                         case BLK_ZONE_COND_EMPTY:
 496                                 __set_bit(nreported, zone_info->empty_zones);
 497                                 break;
 498                         case BLK_ZONE_COND_IMP_OPEN:
 499                         case BLK_ZONE_COND_EXP_OPEN:
 500                         case BLK_ZONE_COND_CLOSED:
 501                                 __set_bit(nreported, zone_info->active_zones);
 502                                 nactive++;
 503                                 break;
 504                         }
 505                         nreported++;
 506                 }
 507                 sector = zones[nr_zones - 1].start + zones[nr_zones - 1].len;
 508         }
 509
 510         if (nreported != zone_info->nr_zones) {
 511                 btrfs_err_in_rcu(device->fs_info,
 512                                  "inconsistent number of zones on %s (%u/%u)",
 513                                  rcu_str_deref(device->name), nreported,
 514                                  zone_info->nr_zones);
 515                 ret = -EIO;
 516                 goto out;
 517         }
 518
 519         if (max_active_zones) {
 520                 if (nactive > max_active_zones) {
 521                         btrfs_err_in_rcu(device->fs_info,
 522                         "zoned: %u active zones on %s exceeds max_active_zones %u",
 523                                          nactive, rcu_str_deref(device->name),
 524                                          max_active_zones);
 525                         ret = -EIO;
 526                         goto out;
 527                 }
 528                 atomic_set(&zone_info->active_zones_left,
 529                            max_active_zones - nactive);
 530                 set_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags);
 531         }
 532
 533         /* Validate superblock log */
 534         nr_zones = BTRFS_NR_SB_LOG_ZONES;
 535         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 536                 u32 sb_zone;
 537                 u64 sb_wp;
 538                 int sb_pos = BTRFS_NR_SB_LOG_ZONES * i;
 539
 540                 sb_zone = sb_zone_number(zone_info->zone_size_shift, i);
 541                 if (sb_zone + 1 >= zone_info->nr_zones)
 542                         continue;
 543
 544                 ret = btrfs_get_dev_zones(device,
 545                                           zone_start_physical(sb_zone, zone_info),
 546                                           &zone_info->sb_zones[sb_pos],
 547                                           &nr_zones);
 548                 if (ret)
 549                         goto out;
 550
 551                 if (nr_zones != BTRFS_NR_SB_LOG_ZONES) {
 552                         btrfs_err_in_rcu(device->fs_info,
 553         "zoned: failed to read super block log zone info at devid %llu zone %u",
 554                                          device->devid, sb_zone);
 555                         ret = -EUCLEAN;
 556                         goto out;
 557                 }
 558
 559                 /*
 560                  * If zones[0] is conventional, always use the beginning of the
 561                  * zone to record superblock. No need to validate in that case.
 562                  */
 563                 if (zone_info->sb_zones[BTRFS_NR_SB_LOG_ZONES * i].type ==
 564                     BLK_ZONE_TYPE_CONVENTIONAL)
 565                         continue;
 566
 567                 ret = sb_write_pointer(device->bdev,
 568                                        &zone_info->sb_zones[sb_pos], &sb_wp);
 569                 if (ret != -ENOENT && ret) {
 570                         btrfs_err_in_rcu(device->fs_info,
 571                         "zoned: super block log zone corrupted devid %llu zone %u",
 572                                          device->devid, sb_zone);
 573                         ret = -EUCLEAN;
 574                         goto out;
 575                 }
 576         }
 577
 578
 579         kvfree(zones);
 580
 581         switch (bdev_zoned_model(bdev)) {
 582         case BLK_ZONED_HM:
 583                 model = "host-managed zoned";
 584                 emulated = "";
 585                 break;
 586         case BLK_ZONED_HA:
 587                 model = "host-aware zoned";
 588                 emulated = "";
 589                 break;
 590         case BLK_ZONED_NONE:
 591                 model = "regular";
 592                 emulated = "emulated ";
 593                 break;
 594         default:
 595                 /* Just in case */
 596                 btrfs_err_in_rcu(fs_info, "zoned: unsupported model %d on %s",
 597                                  bdev_zoned_model(bdev),
 598                                  rcu_str_deref(device->name));
 599                 ret = -EOPNOTSUPP;
 600                 goto out_free_zone_info;
 601         }
 602
 603         btrfs_info_in_rcu(fs_info,
 604                 "%s block device %s, %u %szones of %llu bytes",
 605                 model, rcu_str_deref(device->name), zone_info->nr_zones,
 606                 emulated, zone_info->zone_size);
 607
 608         return 0;
 609
 610 out:
 611         kvfree(zones);
 612 out_free_zone_info:
 613         btrfs_destroy_dev_zone_info(device);
 614
 615         return ret;
 616 }
 617
 618 void btrfs_destroy_dev_zone_info(struct btrfs_device *device)
 619 {
 620         struct btrfs_zoned_device_info *zone_info = device->zone_info;
 621
 622         if (!zone_info)
 623                 return;
 624
 625         bitmap_free(zone_info->active_zones);
 626         bitmap_free(zone_info->seq_zones);
 627         bitmap_free(zone_info->empty_zones);
 628         vfree(zone_info->zone_cache);
 629         kfree(zone_info);
 630         device->zone_info = NULL;
 631 }
 632
 633 struct btrfs_zoned_device_info *btrfs_clone_dev_zone_info(struct btrfs_device *orig_dev)
 634 {
 635         struct btrfs_zoned_device_info *zone_info;
 636
 637         zone_info = kmemdup(orig_dev->zone_info, sizeof(*zone_info), GFP_KERNEL);
 638         if (!zone_info)
 639                 return NULL;
 640
 641         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 642         if (!zone_info->seq_zones)
 643                 goto out;
 644
 645         bitmap_copy(zone_info->seq_zones, orig_dev->zone_info->seq_zones,
 646                     zone_info->nr_zones);
 647
 648         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 649         if (!zone_info->empty_zones)
 650                 goto out;
 651
 652         bitmap_copy(zone_info->empty_zones, orig_dev->zone_info->empty_zones,
 653                     zone_info->nr_zones);
 654
 655         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 656         if (!zone_info->active_zones)
 657                 goto out;
 658
 659         bitmap_copy(zone_info->active_zones, orig_dev->zone_info->active_zones,
 660                     zone_info->nr_zones);
 661         zone_info->zone_cache = NULL;
 662
 663         return zone_info;
 664
 665 out:
 666         bitmap_free(zone_info->seq_zones);
 667         bitmap_free(zone_info->empty_zones);
 668         bitmap_free(zone_info->active_zones);
 669         kfree(zone_info);
 670         return NULL;
 671 }
 672
 673 int btrfs_get_dev_zone(struct btrfs_device *device, u64 pos,
 674                        struct blk_zone *zone)
 675 {
 676         unsigned int nr_zones = 1;
 677         int ret;
 678
 679         ret = btrfs_get_dev_zones(device, pos, zone, &nr_zones);
 680         if (ret != 0 || !nr_zones)
 681                 return ret ? ret : -EIO;
 682
 683         return 0;
 684 }
 685
 686 static int btrfs_check_for_zoned_device(struct btrfs_fs_info *fs_info)
 687 {
 688         struct btrfs_device *device;
 689
 690         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 691                 if (device->bdev &&
 692                     bdev_zoned_model(device->bdev) == BLK_ZONED_HM) {
 693                         btrfs_err(fs_info,
 694                                 "zoned: mode not enabled but zoned device found: %pg",
 695                                 device->bdev);
 696                         return -EINVAL;
 697                 }
 698         }
 699
 700         return 0;
 701 }
 702
 703 int btrfs_check_zoned_mode(struct btrfs_fs_info *fs_info)
 704 {
 705         struct queue_limits *lim = &fs_info->limits;
 706         struct btrfs_device *device;
 707         u64 zone_size = 0;
 708         int ret;
 709
 710         /*
 711          * Host-Managed devices can't be used without the ZONED flag.  With the
 712          * ZONED all devices can be used, using zone emulation if required.
 713          */
 714         if (!btrfs_fs_incompat(fs_info, ZONED))
 715                 return btrfs_check_for_zoned_device(fs_info);
 716
 717         blk_set_stacking_limits(lim);
 718
 719         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 720                 struct btrfs_zoned_device_info *zone_info = device->zone_info;
 721
 722                 if (!device->bdev)
 723                         continue;
 724
 725                 if (!zone_size) {
 726                         zone_size = zone_info->zone_size;
 727                 } else if (zone_info->zone_size != zone_size) {
 728                         btrfs_err(fs_info,
 729                 "zoned: unequal block device zone sizes: have %llu found %llu",
 730                                   zone_info->zone_size, zone_size);
 731                         return -EINVAL;
 732                 }
 733
 734                 /*
 735                  * With the zoned emulation, we can have non-zoned device on the
 736                  * zoned mode. In this case, we don't have a valid max zone
 737                  * append size.
 738                  */
 739                 if (bdev_is_zoned(device->bdev)) {
 740                         blk_stack_limits(lim,
 741                                          &bdev_get_queue(device->bdev)->limits,
 742                                          0);
 743                 }
 744         }
 745
 746         /*
 747          * stripe_size is always aligned to BTRFS_STRIPE_LEN in
 748          * btrfs_create_chunk(). Since we want stripe_len == zone_size,
 749          * check the alignment here.
 750          */
 751         if (!IS_ALIGNED(zone_size, BTRFS_STRIPE_LEN)) {
 752                 btrfs_err(fs_info,
 753                           "zoned: zone size %llu not aligned to stripe %u",
 754                           zone_size, BTRFS_STRIPE_LEN);
 755                 return -EINVAL;
 756         }
 757
 758         if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
 759                 btrfs_err(fs_info, "zoned: mixed block groups not supported");
 760                 return -EINVAL;
 761         }
 762
 763         fs_info->zone_size = zone_size;
 764         /*
 765          * Also limit max_zone_append_size by max_segments * PAGE_SIZE.
 766          * Technically, we can have multiple pages per segment. But, since
 767          * we add the pages one by one to a bio, and cannot increase the
 768          * metadata reservation even if it increases the number of extents, it
 769          * is safe to stick with the limit.
 770          */
 771         fs_info->max_zone_append_size = ALIGN_DOWN(
 772                 min3((u64)lim->max_zone_append_sectors << SECTOR_SHIFT,
 773                      (u64)lim->max_sectors << SECTOR_SHIFT,
 774                      (u64)lim->max_segments << PAGE_SHIFT),
 775                 fs_info->sectorsize);
 776         fs_info->fs_devices->chunk_alloc_policy = BTRFS_CHUNK_ALLOC_ZONED;
 777         if (fs_info->max_zone_append_size < fs_info->max_extent_size)
 778                 fs_info->max_extent_size = fs_info->max_zone_append_size;
 779
 780         /*
 781          * Check mount options here, because we might change fs_info->zoned
 782          * from fs_info->zone_size.
 783          */
 784         ret = btrfs_check_mountopts_zoned(fs_info);
 785         if (ret)
 786                 return ret;
 787
 788         btrfs_info(fs_info, "zoned mode enabled with zone size %llu", zone_size);
 789         return 0;
 790 }
 791
 792 int btrfs_check_mountopts_zoned(struct btrfs_fs_info *info)
 793 {
 794         if (!btrfs_is_zoned(info))
 795                 return 0;
 796
 797         /*
 798          * Space cache writing is not COWed. Disable that to avoid write errors
 799          * in sequential zones.
 800          */
 801         if (btrfs_test_opt(info, SPACE_CACHE)) {
 802                 btrfs_err(info, "zoned: space cache v1 is not supported");
 803                 return -EINVAL;
 804         }
 805
 806         if (btrfs_test_opt(info, NODATACOW)) {
 807                 btrfs_err(info, "zoned: NODATACOW not supported");
 808                 return -EINVAL;
 809         }
 810
 811         btrfs_clear_and_info(info, DISCARD_ASYNC,
 812                         "zoned: async discard ignored and disabled for zoned mode");
 813
 814         return 0;
 815 }
 816
 817 static int sb_log_location(struct block_device *bdev, struct blk_zone *zones,
 818                            int rw, u64 *bytenr_ret)
 819 {
 820         u64 wp;
 821         int ret;
 822
 823         if (zones[0].type == BLK_ZONE_TYPE_CONVENTIONAL) {
 824                 *bytenr_ret = zones[0].start << SECTOR_SHIFT;
 825                 return 0;
 826         }
 827
 828         ret = sb_write_pointer(bdev, zones, &wp);
 829         if (ret != -ENOENT && ret < 0)
 830                 return ret;
 831
 832         if (rw == WRITE) {
 833                 struct blk_zone *reset = NULL;
 834
 835                 if (wp == zones[0].start << SECTOR_SHIFT)
 836                         reset = &zones[0];
 837                 else if (wp == zones[1].start << SECTOR_SHIFT)
 838                         reset = &zones[1];
 839
 840                 if (reset && reset->cond != BLK_ZONE_COND_EMPTY) {
 841                         ASSERT(sb_zone_is_full(reset));
 842
 843                         ret = blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
 844                                                reset->start, reset->len,
 845                                                GFP_NOFS);
 846                         if (ret)
 847                                 return ret;
 848
 849                         reset->cond = BLK_ZONE_COND_EMPTY;
 850                         reset->wp = reset->start;
 851                 }
 852         } else if (ret != -ENOENT) {
 853                 /*
 854                  * For READ, we want the previous one. Move write pointer to
 855                  * the end of a zone, if it is at the head of a zone.
 856                  */
 857                 u64 zone_end = 0;
 858
 859                 if (wp == zones[0].start << SECTOR_SHIFT)
 860                         zone_end = zones[1].start + zones[1].capacity;
 861                 else if (wp == zones[1].start << SECTOR_SHIFT)
 862                         zone_end = zones[0].start + zones[0].capacity;
 863                 if (zone_end)
 864                         wp = ALIGN_DOWN(zone_end << SECTOR_SHIFT,
 865                                         BTRFS_SUPER_INFO_SIZE);
 866
 867                 wp -= BTRFS_SUPER_INFO_SIZE;
 868         }
 869
 870         *bytenr_ret = wp;
 871         return 0;
 872
 873 }
 874
 875 int btrfs_sb_log_location_bdev(struct block_device *bdev, int mirror, int rw,
 876                                u64 *bytenr_ret)
 877 {
 878         struct blk_zone zones[BTRFS_NR_SB_LOG_ZONES];
 879         sector_t zone_sectors;
 880         u32 sb_zone;
 881         int ret;
 882         u8 zone_sectors_shift;
 883         sector_t nr_sectors;
 884         u32 nr_zones;
 885
 886         if (!bdev_is_zoned(bdev)) {
 887                 *bytenr_ret = btrfs_sb_offset(mirror);
 888                 return 0;
 889         }
 890
 891         ASSERT(rw == READ || rw == WRITE);
 892
 893         zone_sectors = bdev_zone_sectors(bdev);
 894         if (!is_power_of_2(zone_sectors))
 895                 return -EINVAL;
 896         zone_sectors_shift = ilog2(zone_sectors);
 897         nr_sectors = bdev_nr_sectors(bdev);
 898         nr_zones = nr_sectors >> zone_sectors_shift;
 899
 900         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
 901         if (sb_zone + 1 >= nr_zones)
 902                 return -ENOENT;
 903
 904         ret = blkdev_report_zones(bdev, zone_start_sector(sb_zone, bdev),
 905                                   BTRFS_NR_SB_LOG_ZONES, copy_zone_info_cb,
 906                                   zones);
 907         if (ret < 0)
 908                 return ret;
 909         if (ret != BTRFS_NR_SB_LOG_ZONES)
 910                 return -EIO;
 911
 912         return sb_log_location(bdev, zones, rw, bytenr_ret);
 913 }
 914
 915 int btrfs_sb_log_location(struct btrfs_device *device, int mirror, int rw,
 916                           u64 *bytenr_ret)
 917 {
 918         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 919         u32 zone_num;
 920
 921         /*
 922          * For a zoned filesystem on a non-zoned block device, use the same
 923          * super block locations as regular filesystem. Doing so, the super
 924          * block can always be retrieved and the zoned flag of the volume
 925          * detected from the super block information.
 926          */
 927         if (!bdev_is_zoned(device->bdev)) {
 928                 *bytenr_ret = btrfs_sb_offset(mirror);
 929                 return 0;
 930         }
 931
 932         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 933         if (zone_num + 1 >= zinfo->nr_zones)
 934                 return -ENOENT;
 935
 936         return sb_log_location(device->bdev,
 937                                &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror],
 938                                rw, bytenr_ret);
 939 }
 940
 941 static inline bool is_sb_log_zone(struct btrfs_zoned_device_info *zinfo,
 942                                   int mirror)
 943 {
 944         u32 zone_num;
 945
 946         if (!zinfo)
 947                 return false;
 948
 949         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 950         if (zone_num + 1 >= zinfo->nr_zones)
 951                 return false;
 952
 953         if (!test_bit(zone_num, zinfo->seq_zones))
 954                 return false;
 955
 956         return true;
 957 }
 958
 959 int btrfs_advance_sb_log(struct btrfs_device *device, int mirror)
 960 {
 961         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 962         struct blk_zone *zone;
 963         int i;
 964
 965         if (!is_sb_log_zone(zinfo, mirror))
 966                 return 0;
 967
 968         zone = &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror];
 969         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 970                 /* Advance the next zone */
 971                 if (zone->cond == BLK_ZONE_COND_FULL) {
 972                         zone++;
 973                         continue;
 974                 }
 975
 976                 if (zone->cond == BLK_ZONE_COND_EMPTY)
 977                         zone->cond = BLK_ZONE_COND_IMP_OPEN;
 978
 979                 zone->wp += SUPER_INFO_SECTORS;
 980
 981                 if (sb_zone_is_full(zone)) {
 982                         /*
 983                          * No room left to write new superblock. Since
 984                          * superblock is written with REQ_SYNC, it is safe to
 985                          * finish the zone now.
 986                          *
 987                          * If the write pointer is exactly at the capacity,
 988                          * explicit ZONE_FINISH is not necessary.
 989                          */
 990                         if (zone->wp != zone->start + zone->capacity) {
 991                                 int ret;
 992
 993                                 ret = blkdev_zone_mgmt(device->bdev,
 994                                                 REQ_OP_ZONE_FINISH, zone->start,
 995                                                 zone->len, GFP_NOFS);
 996                                 if (ret)
 997                                         return ret;
 998                         }
 999
1000                         zone->wp = zone->start + zone->len;
1001                         zone->cond = BLK_ZONE_COND_FULL;
1002                 }
1003                 return 0;
1004         }
1005
1006         /* All the zones are FULL. Should not reach here. */
1007         ASSERT(0);
1008         return -EIO;
1009 }
1010
1011 int btrfs_reset_sb_log_zones(struct block_device *bdev, int mirror)
1012 {
1013         sector_t zone_sectors;
1014         sector_t nr_sectors;
1015         u8 zone_sectors_shift;
1016         u32 sb_zone;
1017         u32 nr_zones;
1018
1019         zone_sectors = bdev_zone_sectors(bdev);
1020         zone_sectors_shift = ilog2(zone_sectors);
1021         nr_sectors = bdev_nr_sectors(bdev);
1022         nr_zones = nr_sectors >> zone_sectors_shift;
1023
1024         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
1025         if (sb_zone + 1 >= nr_zones)
1026                 return -ENOENT;
1027
1028         return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
1029                                 zone_start_sector(sb_zone, bdev),
1030                                 zone_sectors * BTRFS_NR_SB_LOG_ZONES, GFP_NOFS);
1031 }
1032
1033 /*
1034  * Find allocatable zones within a given region.
1035  *
1036  * @device:     the device to allocate a region on
1037  * @hole_start: the position of the hole to allocate the region
1038  * @num_bytes:  size of wanted region
1039  * @hole_end:   the end of the hole
1040  * @return:     position of allocatable zones
1041  *
1042  * Allocatable region should not contain any superblock locations.
1043  */
1044 u64 btrfs_find_allocatable_zones(struct btrfs_device *device, u64 hole_start,
1045                                  u64 hole_end, u64 num_bytes)
1046 {
1047         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1048         const u8 shift = zinfo->zone_size_shift;
1049         u64 nzones = num_bytes >> shift;
1050         u64 pos = hole_start;
1051         u64 begin, end;
1052         bool have_sb;
1053         int i;
1054
1055         ASSERT(IS_ALIGNED(hole_start, zinfo->zone_size));
1056         ASSERT(IS_ALIGNED(num_bytes, zinfo->zone_size));
1057
1058         while (pos < hole_end) {
1059                 begin = pos >> shift;
1060                 end = begin + nzones;
1061
1062                 if (end > zinfo->nr_zones)
1063                         return hole_end;
1064
1065                 /* Check if zones in the region are all empty */
1066                 if (btrfs_dev_is_sequential(device, pos) &&
1067                     !bitmap_test_range_all_set(zinfo->empty_zones, begin, nzones)) {
1068                         pos += zinfo->zone_size;
1069                         continue;
1070                 }
1071
1072                 have_sb = false;
1073                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1074                         u32 sb_zone;
1075                         u64 sb_pos;
1076
1077                         sb_zone = sb_zone_number(shift, i);
1078                         if (!(end <= sb_zone ||
1079                               sb_zone + BTRFS_NR_SB_LOG_ZONES <= begin)) {
1080                                 have_sb = true;
1081                                 pos = zone_start_physical(
1082                                         sb_zone + BTRFS_NR_SB_LOG_ZONES, zinfo);
1083                                 break;
1084                         }
1085
1086                         /* We also need to exclude regular superblock positions */
1087                         sb_pos = btrfs_sb_offset(i);
1088                         if (!(pos + num_bytes <= sb_pos ||
1089                               sb_pos + BTRFS_SUPER_INFO_SIZE <= pos)) {
1090                                 have_sb = true;
1091                                 pos = ALIGN(sb_pos + BTRFS_SUPER_INFO_SIZE,
1092                                             zinfo->zone_size);
1093                                 break;
1094                         }
1095                 }
1096                 if (!have_sb)
1097                         break;
1098         }
1099
1100         return pos;
1101 }
1102
1103 static bool btrfs_dev_set_active_zone(struct btrfs_device *device, u64 pos)
1104 {
1105         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1106         unsigned int zno = (pos >> zone_info->zone_size_shift);
1107
1108         /* We can use any number of zones */
1109         if (zone_info->max_active_zones == 0)
1110                 return true;
1111
1112         if (!test_bit(zno, zone_info->active_zones)) {
1113                 /* Active zone left? */
1114                 if (atomic_dec_if_positive(&zone_info->active_zones_left) < 0)
1115                         return false;
1116                 if (test_and_set_bit(zno, zone_info->active_zones)) {
1117                         /* Someone already set the bit */
1118                         atomic_inc(&zone_info->active_zones_left);
1119                 }
1120         }
1121
1122         return true;
1123 }
1124
1125 static void btrfs_dev_clear_active_zone(struct btrfs_device *device, u64 pos)
1126 {
1127         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1128         unsigned int zno = (pos >> zone_info->zone_size_shift);
1129
1130         /* We can use any number of zones */
1131         if (zone_info->max_active_zones == 0)
1132                 return;
1133
1134         if (test_and_clear_bit(zno, zone_info->active_zones))
1135                 atomic_inc(&zone_info->active_zones_left);
1136 }
1137
1138 int btrfs_reset_device_zone(struct btrfs_device *device, u64 physical,
1139                             u64 length, u64 *bytes)
1140 {
1141         int ret;
1142
1143         *bytes = 0;
1144         ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_RESET,
1145                                physical >> SECTOR_SHIFT, length >> SECTOR_SHIFT,
1146                                GFP_NOFS);
1147         if (ret)
1148                 return ret;
1149
1150         *bytes = length;
1151         while (length) {
1152                 btrfs_dev_set_zone_empty(device, physical);
1153                 btrfs_dev_clear_active_zone(device, physical);
1154                 physical += device->zone_info->zone_size;
1155                 length -= device->zone_info->zone_size;
1156         }
1157
1158         return 0;
1159 }
1160
1161 int btrfs_ensure_empty_zones(struct btrfs_device *device, u64 start, u64 size)
1162 {
1163         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1164         const u8 shift = zinfo->zone_size_shift;
1165         unsigned long begin = start >> shift;
1166         unsigned long nbits = size >> shift;
1167         u64 pos;
1168         int ret;
1169
1170         ASSERT(IS_ALIGNED(start, zinfo->zone_size));
1171         ASSERT(IS_ALIGNED(size, zinfo->zone_size));
1172
1173         if (begin + nbits > zinfo->nr_zones)
1174                 return -ERANGE;
1175
1176         /* All the zones are conventional */
1177         if (bitmap_test_range_all_zero(zinfo->seq_zones, begin, nbits))
1178                 return 0;
1179
1180         /* All the zones are sequential and empty */
1181         if (bitmap_test_range_all_set(zinfo->seq_zones, begin, nbits) &&
1182             bitmap_test_range_all_set(zinfo->empty_zones, begin, nbits))
1183                 return 0;
1184
1185         for (pos = start; pos < start + size; pos += zinfo->zone_size) {
1186                 u64 reset_bytes;
1187
1188                 if (!btrfs_dev_is_sequential(device, pos) ||
1189                     btrfs_dev_is_empty_zone(device, pos))
1190                         continue;
1191
1192                 /* Free regions should be empty */
1193                 btrfs_warn_in_rcu(
1194                         device->fs_info,
1195                 "zoned: resetting device %s (devid %llu) zone %llu for allocation",
1196                         rcu_str_deref(device->name), device->devid, pos >> shift);
1197                 WARN_ON_ONCE(1);
1198
1199                 ret = btrfs_reset_device_zone(device, pos, zinfo->zone_size,
1200                                               &reset_bytes);
1201                 if (ret)
1202                         return ret;
1203         }
1204
1205         return 0;
1206 }
1207
1208 /*
1209  * Calculate an allocation pointer from the extent allocation information
1210  * for a block group consist of conventional zones. It is pointed to the
1211  * end of the highest addressed extent in the block group as an allocation
1212  * offset.
1213  */
1214 static int calculate_alloc_pointer(struct btrfs_block_group *cache,
1215                                    u64 *offset_ret, bool new)
1216 {
1217         struct btrfs_fs_info *fs_info = cache->fs_info;
1218         struct btrfs_root *root;
1219         struct btrfs_path *path;
1220         struct btrfs_key key;
1221         struct btrfs_key found_key;
1222         int ret;
1223         u64 length;
1224
1225         /*
1226          * Avoid  tree lookups for a new block group, there's no use for it.
1227          * It must always be 0.
1228          *
1229          * Also, we have a lock chain of extent buffer lock -> chunk mutex.
1230          * For new a block group, this function is called from
1231          * btrfs_make_block_group() which is already taking the chunk mutex.
1232          * Thus, we cannot call calculate_alloc_pointer() which takes extent
1233          * buffer locks to avoid deadlock.
1234          */
1235         if (new) {
1236                 *offset_ret = 0;
1237                 return 0;
1238         }
1239
1240         path = btrfs_alloc_path();
1241         if (!path)
1242                 return -ENOMEM;
1243
1244         key.objectid = cache->start + cache->length;
1245         key.type = 0;
1246         key.offset = 0;
1247
1248         root = btrfs_extent_root(fs_info, key.objectid);
1249         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1250         /* We should not find the exact match */
1251         if (!ret)
1252                 ret = -EUCLEAN;
1253         if (ret < 0)
1254                 goto out;
1255
1256         ret = btrfs_previous_extent_item(root, path, cache->start);
1257         if (ret) {
1258                 if (ret == 1) {
1259                         ret = 0;
1260                         *offset_ret = 0;
1261                 }
1262                 goto out;
1263         }
1264
1265         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
1266
1267         if (found_key.type == BTRFS_EXTENT_ITEM_KEY)
1268                 length = found_key.offset;
1269         else
1270                 length = fs_info->nodesize;
1271
1272         if (!(found_key.objectid >= cache->start &&
1273                found_key.objectid + length <= cache->start + cache->length)) {
1274                 ret = -EUCLEAN;
1275                 goto out;
1276         }
1277         *offset_ret = found_key.objectid + length - cache->start;
1278         ret = 0;
1279
1280 out:
1281         btrfs_free_path(path);
1282         return ret;
1283 }
1284
1285 struct zone_info {
1286         u64 physical;
1287         u64 capacity;
1288         u64 alloc_offset;
1289 };
1290
1291 static int btrfs_load_zone_info(struct btrfs_fs_info *fs_info, int zone_idx,
1292                                 struct zone_info *info, unsigned long *active,
1293                                 struct map_lookup *map)
1294 {
1295         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1296         struct btrfs_device *device = map->stripes[zone_idx].dev;
1297         int dev_replace_is_ongoing = 0;
1298         unsigned int nofs_flag;
1299         struct blk_zone zone;
1300         int ret;
1301
1302         info->physical = map->stripes[zone_idx].physical;
1303
1304         if (!device->bdev) {
1305                 info->alloc_offset = WP_MISSING_DEV;
1306                 return 0;
1307         }
1308
1309         /* Consider a zone as active if we can allow any number of active zones. */
1310         if (!device->zone_info->max_active_zones)
1311                 __set_bit(zone_idx, active);
1312
1313         if (!btrfs_dev_is_sequential(device, info->physical)) {
1314                 info->alloc_offset = WP_CONVENTIONAL;
1315                 return 0;
1316         }
1317
1318         /* This zone will be used for allocation, so mark this zone non-empty. */
1319         btrfs_dev_clear_zone_empty(device, info->physical);
1320
1321         down_read(&dev_replace->rwsem);
1322         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
1323         if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL)
1324                 btrfs_dev_clear_zone_empty(dev_replace->tgtdev, info->physical);
1325         up_read(&dev_replace->rwsem);
1326
1327         /*
1328          * The group is mapped to a sequential zone. Get the zone write pointer
1329          * to determine the allocation offset within the zone.
1330          */
1331         WARN_ON(!IS_ALIGNED(info->physical, fs_info->zone_size));
1332         nofs_flag = memalloc_nofs_save();
1333         ret = btrfs_get_dev_zone(device, info->physical, &zone);
1334         memalloc_nofs_restore(nofs_flag);
1335         if (ret) {
1336                 if (ret != -EIO && ret != -EOPNOTSUPP)
1337                         return ret;
1338                 info->alloc_offset = WP_MISSING_DEV;
1339                 return 0;
1340         }
1341
1342         if (zone.type == BLK_ZONE_TYPE_CONVENTIONAL) {
1343                 btrfs_err_in_rcu(fs_info,
1344                 "zoned: unexpected conventional zone %llu on device %s (devid %llu)",
1345                         zone.start << SECTOR_SHIFT, rcu_str_deref(device->name),
1346                         device->devid);
1347                 return -EIO;
1348         }
1349
1350         info->capacity = (zone.capacity << SECTOR_SHIFT);
1351
1352         switch (zone.cond) {
1353         case BLK_ZONE_COND_OFFLINE:
1354         case BLK_ZONE_COND_READONLY:
1355                 btrfs_err(fs_info,
1356                 "zoned: offline/readonly zone %llu on device %s (devid %llu)",
1357                           (info->physical >> device->zone_info->zone_size_shift),
1358                           rcu_str_deref(device->name), device->devid);
1359                 info->alloc_offset = WP_MISSING_DEV;
1360                 break;
1361         case BLK_ZONE_COND_EMPTY:
1362                 info->alloc_offset = 0;
1363                 break;
1364         case BLK_ZONE_COND_FULL:
1365                 info->alloc_offset = info->capacity;
1366                 break;
1367         default:
1368                 /* Partially used zone. */
1369                 info->alloc_offset = ((zone.wp - zone.start) << SECTOR_SHIFT);
1370                 __set_bit(zone_idx, active);
1371                 break;
1372         }
1373
1374         return 0;
1375 }
1376
1377 static int btrfs_load_block_group_single(struct btrfs_block_group *bg,
1378                                          struct zone_info *info,
1379                                          unsigned long *active)
1380 {
1381         if (info->alloc_offset == WP_MISSING_DEV) {
1382                 btrfs_err(bg->fs_info,
1383                         "zoned: cannot recover write pointer for zone %llu",
1384                         info->physical);
1385                 return -EIO;
1386         }
1387
1388         bg->alloc_offset = info->alloc_offset;
1389         bg->zone_capacity = info->capacity;
1390         if (test_bit(0, active))
1391                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1392         return 0;
1393 }
1394
1395 static int btrfs_load_block_group_dup(struct btrfs_block_group *bg,
1396                                       struct map_lookup *map,
1397                                       struct zone_info *zone_info,
1398                                       unsigned long *active)
1399 {
1400         struct btrfs_fs_info *fs_info = bg->fs_info;
1401
1402         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1403                 btrfs_err(fs_info, "zoned: data DUP profile needs raid-stripe-tree");
1404                 return -EINVAL;
1405         }
1406
1407         if (zone_info[0].alloc_offset == WP_MISSING_DEV) {
1408                 btrfs_err(bg->fs_info,
1409                           "zoned: cannot recover write pointer for zone %llu",
1410                           zone_info[0].physical);
1411                 return -EIO;
1412         }
1413         if (zone_info[1].alloc_offset == WP_MISSING_DEV) {
1414                 btrfs_err(bg->fs_info,
1415                           "zoned: cannot recover write pointer for zone %llu",
1416                           zone_info[1].physical);
1417                 return -EIO;
1418         }
1419         if (zone_info[0].alloc_offset != zone_info[1].alloc_offset) {
1420                 btrfs_err(bg->fs_info,
1421                           "zoned: write pointer offset mismatch of zones in DUP profile");
1422                 return -EIO;
1423         }
1424
1425         if (test_bit(0, active) != test_bit(1, active)) {
1426                 if (!btrfs_zone_activate(bg))
1427                         return -EIO;
1428         } else if (test_bit(0, active)) {
1429                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1430         }
1431
1432         bg->alloc_offset = zone_info[0].alloc_offset;
1433         bg->zone_capacity = min(zone_info[0].capacity, zone_info[1].capacity);
1434         return 0;
1435 }
1436
1437 static int btrfs_load_block_group_raid1(struct btrfs_block_group *bg,
1438                                         struct map_lookup *map,
1439                                         struct zone_info *zone_info,
1440                                         unsigned long *active)
1441 {
1442         struct btrfs_fs_info *fs_info = bg->fs_info;
1443         int i;
1444
1445         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1446                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1447                           btrfs_bg_type_to_raid_name(map->type));
1448                 return -EINVAL;
1449         }
1450
1451         for (i = 0; i < map->num_stripes; i++) {
1452                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1453                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1454                         continue;
1455
1456                 if ((zone_info[0].alloc_offset != zone_info[i].alloc_offset) &&
1457                     !btrfs_test_opt(fs_info, DEGRADED)) {
1458                         btrfs_err(fs_info,
1459                         "zoned: write pointer offset mismatch of zones in %s profile",
1460                                   btrfs_bg_type_to_raid_name(map->type));
1461                         return -EIO;
1462                 }
1463                 if (test_bit(0, active) != test_bit(i, active)) {
1464                         if (!btrfs_test_opt(fs_info, DEGRADED) &&
1465                             !btrfs_zone_activate(bg)) {
1466                                 return -EIO;
1467                         }
1468                 } else {
1469                         if (test_bit(0, active))
1470                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1471                 }
1472                 /* In case a device is missing we have a cap of 0, so don't use it. */
1473                 bg->zone_capacity = min_not_zero(zone_info[0].capacity,
1474                                                  zone_info[1].capacity);
1475         }
1476
1477         if (zone_info[0].alloc_offset != WP_MISSING_DEV)
1478                 bg->alloc_offset = zone_info[0].alloc_offset;
1479         else
1480                 bg->alloc_offset = zone_info[i - 1].alloc_offset;
1481
1482         return 0;
1483 }
1484
1485 static int btrfs_load_block_group_raid0(struct btrfs_block_group *bg,
1486                                         struct map_lookup *map,
1487                                         struct zone_info *zone_info,
1488                                         unsigned long *active)
1489 {
1490         struct btrfs_fs_info *fs_info = bg->fs_info;
1491
1492         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1493                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1494                           btrfs_bg_type_to_raid_name(map->type));
1495                 return -EINVAL;
1496         }
1497
1498         for (int i = 0; i < map->num_stripes; i++) {
1499                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1500                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1501                         continue;
1502
1503                 if (test_bit(0, active) != test_bit(i, active)) {
1504                         if (!btrfs_zone_activate(bg))
1505                                 return -EIO;
1506                 } else {
1507                         if (test_bit(0, active))
1508                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1509                 }
1510                 bg->zone_capacity += zone_info[i].capacity;
1511                 bg->alloc_offset += zone_info[i].alloc_offset;
1512         }
1513
1514         return 0;
1515 }
1516
1517 static int btrfs_load_block_group_raid10(struct btrfs_block_group *bg,
1518                                          struct map_lookup *map,
1519                                          struct zone_info *zone_info,
1520                                          unsigned long *active)
1521 {
1522         struct btrfs_fs_info *fs_info = bg->fs_info;
1523
1524         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1525                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1526                           btrfs_bg_type_to_raid_name(map->type));
1527                 return -EINVAL;
1528         }
1529
1530         for (int i = 0; i < map->num_stripes; i++) {
1531                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1532                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1533                         continue;
1534
1535                 if (test_bit(0, active) != test_bit(i, active)) {
1536                         if (!btrfs_zone_activate(bg))
1537                                 return -EIO;
1538                 } else {
1539                         if (test_bit(0, active))
1540                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1541                 }
1542
1543                 if ((i % map->sub_stripes) == 0) {
1544                         bg->zone_capacity += zone_info[i].capacity;
1545                         bg->alloc_offset += zone_info[i].alloc_offset;
1546                 }
1547         }
1548
1549         return 0;
1550 }
1551
1552 int btrfs_load_block_group_zone_info(struct btrfs_block_group *cache, bool new)
1553 {
1554         struct btrfs_fs_info *fs_info = cache->fs_info;
1555         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1556         struct extent_map *em;
1557         struct map_lookup *map;
1558         u64 logical = cache->start;
1559         u64 length = cache->length;
1560         struct zone_info *zone_info = NULL;
1561         int ret;
1562         int i;
1563         unsigned long *active = NULL;
1564         u64 last_alloc = 0;
1565         u32 num_sequential = 0, num_conventional = 0;
1566
1567         if (!btrfs_is_zoned(fs_info))
1568                 return 0;
1569
1570         /* Sanity check */
1571         if (!IS_ALIGNED(length, fs_info->zone_size)) {
1572                 btrfs_err(fs_info,
1573                 "zoned: block group %llu len %llu unaligned to zone size %llu",
1574                           logical, length, fs_info->zone_size);
1575                 return -EIO;
1576         }
1577
1578         /* Get the chunk mapping */
1579         read_lock(&em_tree->lock);
1580         em = lookup_extent_mapping(em_tree, logical, length);
1581         read_unlock(&em_tree->lock);
1582
1583         if (!em)
1584                 return -EINVAL;
1585
1586         map = em->map_lookup;
1587
1588         cache->physical_map = kmemdup(map, map_lookup_size(map->num_stripes), GFP_NOFS);
1589         if (!cache->physical_map) {
1590                 ret = -ENOMEM;
1591                 goto out;
1592         }
1593
1594         zone_info = kcalloc(map->num_stripes, sizeof(*zone_info), GFP_NOFS);
1595         if (!zone_info) {
1596                 ret = -ENOMEM;
1597                 goto out;
1598         }
1599
1600         active = bitmap_zalloc(map->num_stripes, GFP_NOFS);
1601         if (!active) {
1602                 ret = -ENOMEM;
1603                 goto out;
1604         }
1605
1606         for (i = 0; i < map->num_stripes; i++) {
1607                 ret = btrfs_load_zone_info(fs_info, i, &zone_info[i], active, map);
1608                 if (ret)
1609                         goto out;
1610
1611                 if (zone_info[i].alloc_offset == WP_CONVENTIONAL)
1612                         num_conventional++;
1613                 else
1614                         num_sequential++;
1615         }
1616
1617         if (num_sequential > 0)
1618                 set_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1619
1620         if (num_conventional > 0) {
1621                 /* Zone capacity is always zone size in emulation */
1622                 cache->zone_capacity = cache->length;
1623                 ret = calculate_alloc_pointer(cache, &last_alloc, new);
1624                 if (ret) {
1625                         btrfs_err(fs_info,
1626                         "zoned: failed to determine allocation offset of bg %llu",
1627                                   cache->start);
1628                         goto out;
1629                 } else if (map->num_stripes == num_conventional) {
1630                         cache->alloc_offset = last_alloc;
1631                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1632                         goto out;
1633                 }
1634         }
1635
1636         switch (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
1637         case 0: /* single */
1638                 ret = btrfs_load_block_group_single(cache, &zone_info[0], active);
1639                 break;
1640         case BTRFS_BLOCK_GROUP_DUP:
1641                 ret = btrfs_load_block_group_dup(cache, map, zone_info, active);
1642                 break;
1643         case BTRFS_BLOCK_GROUP_RAID1:
1644         case BTRFS_BLOCK_GROUP_RAID1C3:
1645         case BTRFS_BLOCK_GROUP_RAID1C4:
1646                 ret = btrfs_load_block_group_raid1(cache, map, zone_info, active);
1647                 break;
1648         case BTRFS_BLOCK_GROUP_RAID0:
1649                 ret = btrfs_load_block_group_raid0(cache, map, zone_info, active);
1650                 break;
1651         case BTRFS_BLOCK_GROUP_RAID10:
1652                 ret = btrfs_load_block_group_raid10(cache, map, zone_info, active);
1653                 break;
1654         case BTRFS_BLOCK_GROUP_RAID5:
1655         case BTRFS_BLOCK_GROUP_RAID6:
1656         default:
1657                 btrfs_err(fs_info, "zoned: profile %s not yet supported",
1658                           btrfs_bg_type_to_raid_name(map->type));
1659                 ret = -EINVAL;
1660                 goto out;
1661         }
1662
1663 out:
1664         if (cache->alloc_offset > cache->zone_capacity) {
1665                 btrfs_err(fs_info,
1666 "zoned: invalid write pointer %llu (larger than zone capacity %llu) in block group %llu",
1667                           cache->alloc_offset, cache->zone_capacity,
1668                           cache->start);
1669                 ret = -EIO;
1670         }
1671
1672         /* An extent is allocated after the write pointer */
1673         if (!ret && num_conventional && last_alloc > cache->alloc_offset) {
1674                 btrfs_err(fs_info,
1675                           "zoned: got wrong write pointer in BG %llu: %llu > %llu",
1676                           logical, last_alloc, cache->alloc_offset);
1677                 ret = -EIO;
1678         }
1679
1680         if (!ret) {
1681                 cache->meta_write_pointer = cache->alloc_offset + cache->start;
1682                 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags)) {
1683                         btrfs_get_block_group(cache);
1684                         spin_lock(&fs_info->zone_active_bgs_lock);
1685                         list_add_tail(&cache->active_bg_list,
1686                                       &fs_info->zone_active_bgs);
1687                         spin_unlock(&fs_info->zone_active_bgs_lock);
1688                 }
1689         } else {
1690                 kfree(cache->physical_map);
1691                 cache->physical_map = NULL;
1692         }
1693         bitmap_free(active);
1694         kfree(zone_info);
1695         free_extent_map(em);
1696
1697         return ret;
1698 }
1699
1700 void btrfs_calc_zone_unusable(struct btrfs_block_group *cache)
1701 {
1702         u64 unusable, free;
1703
1704         if (!btrfs_is_zoned(cache->fs_info))
1705                 return;
1706
1707         WARN_ON(cache->bytes_super != 0);
1708         unusable = (cache->alloc_offset - cache->used) +
1709                    (cache->length - cache->zone_capacity);
1710         free = cache->zone_capacity - cache->alloc_offset;
1711
1712         /* We only need ->free_space in ALLOC_SEQ block groups */
1713         cache->cached = BTRFS_CACHE_FINISHED;
1714         cache->free_space_ctl->free_space = free;
1715         cache->zone_unusable = unusable;
1716 }
1717
1718 void btrfs_redirty_list_add(struct btrfs_transaction *trans,
1719                             struct extent_buffer *eb)
1720 {
1721         if (!btrfs_is_zoned(eb->fs_info) ||
1722             btrfs_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN))
1723                 return;
1724
1725         ASSERT(!test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
1726
1727         memzero_extent_buffer(eb, 0, eb->len);
1728         set_bit(EXTENT_BUFFER_NO_CHECK, &eb->bflags);
1729         set_extent_buffer_dirty(eb);
1730         set_extent_bit(&trans->dirty_pages, eb->start, eb->start + eb->len - 1,
1731                         EXTENT_DIRTY, NULL);
1732 }
1733
1734 bool btrfs_use_zone_append(struct btrfs_bio *bbio)
1735 {
1736         u64 start = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT);
1737         struct btrfs_inode *inode = bbio->inode;
1738         struct btrfs_fs_info *fs_info = bbio->fs_info;
1739         struct btrfs_block_group *cache;
1740         bool ret = false;
1741
1742         if (!btrfs_is_zoned(fs_info))
1743                 return false;
1744
1745         if (!inode || !is_data_inode(&inode->vfs_inode))
1746                 return false;
1747
1748         if (btrfs_op(&bbio->bio) != BTRFS_MAP_WRITE)
1749                 return false;
1750
1751         /*
1752          * Using REQ_OP_ZONE_APPNED for relocation can break assumptions on the
1753          * extent layout the relocation code has.
1754          * Furthermore we have set aside own block-group from which only the
1755          * relocation "process" can allocate and make sure only one process at a
1756          * time can add pages to an extent that gets relocated, so it's safe to
1757          * use regular REQ_OP_WRITE for this special case.
1758          */
1759         if (btrfs_is_data_reloc_root(inode->root))
1760                 return false;
1761
1762         cache = btrfs_lookup_block_group(fs_info, start);
1763         ASSERT(cache);
1764         if (!cache)
1765                 return false;
1766
1767         ret = !!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1768         btrfs_put_block_group(cache);
1769
1770         return ret;
1771 }
1772
1773 void btrfs_record_physical_zoned(struct btrfs_bio *bbio)
1774 {
1775         const u64 physical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
1776         struct btrfs_ordered_sum *sum = bbio->sums;
1777
1778         if (physical < bbio->orig_physical)
1779                 sum->logical -= bbio->orig_physical - physical;
1780         else
1781                 sum->logical += physical - bbio->orig_physical;
1782 }
1783
1784 static void btrfs_rewrite_logical_zoned(struct btrfs_ordered_extent *ordered,
1785                                         u64 logical)
1786 {
1787         struct extent_map_tree *em_tree = &BTRFS_I(ordered->inode)->extent_tree;
1788         struct extent_map *em;
1789
1790         ordered->disk_bytenr = logical;
1791
1792         write_lock(&em_tree->lock);
1793         em = search_extent_mapping(em_tree, ordered->file_offset,
1794                                    ordered->num_bytes);
1795         em->block_start = logical;
1796         free_extent_map(em);
1797         write_unlock(&em_tree->lock);
1798 }
1799
1800 static bool btrfs_zoned_split_ordered(struct btrfs_ordered_extent *ordered,
1801                                       u64 logical, u64 len)
1802 {
1803         struct btrfs_ordered_extent *new;
1804
1805         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags) &&
1806             split_extent_map(BTRFS_I(ordered->inode), ordered->file_offset,
1807                              ordered->num_bytes, len, logical))
1808                 return false;
1809
1810         new = btrfs_split_ordered_extent(ordered, len);
1811         if (IS_ERR(new))
1812                 return false;
1813         new->disk_bytenr = logical;
1814         btrfs_finish_one_ordered(new);
1815         return true;
1816 }
1817
1818 void btrfs_finish_ordered_zoned(struct btrfs_ordered_extent *ordered)
1819 {
1820         struct btrfs_inode *inode = BTRFS_I(ordered->inode);
1821         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1822         struct btrfs_ordered_sum *sum;
1823         u64 logical, len;
1824
1825         /*
1826          * Write to pre-allocated region is for the data relocation, and so
1827          * it should use WRITE operation. No split/rewrite are necessary.
1828          */
1829         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags))
1830                 return;
1831
1832         ASSERT(!list_empty(&ordered->list));
1833         /* The ordered->list can be empty in the above pre-alloc case. */
1834         sum = list_first_entry(&ordered->list, struct btrfs_ordered_sum, list);
1835         logical = sum->logical;
1836         len = sum->len;
1837
1838         while (len < ordered->disk_num_bytes) {
1839                 sum = list_next_entry(sum, list);
1840                 if (sum->logical == logical + len) {
1841                         len += sum->len;
1842                         continue;
1843                 }
1844                 if (!btrfs_zoned_split_ordered(ordered, logical, len)) {
1845                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
1846                         btrfs_err(fs_info, "failed to split ordered extent");
1847                         goto out;
1848                 }
1849                 logical = sum->logical;
1850                 len = sum->len;
1851         }
1852
1853         if (ordered->disk_bytenr != logical)
1854                 btrfs_rewrite_logical_zoned(ordered, logical);
1855
1856 out:
1857         /*
1858          * If we end up here for nodatasum I/O, the btrfs_ordered_sum structures
1859          * were allocated by btrfs_alloc_dummy_sum only to record the logical
1860          * addresses and don't contain actual checksums.  We thus must free them
1861          * here so that we don't attempt to log the csums later.
1862          */
1863         if ((inode->flags & BTRFS_INODE_NODATASUM) ||
1864             test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state)) {
1865                 while ((sum = list_first_entry_or_null(&ordered->list,
1866                                                        typeof(*sum), list))) {
1867                         list_del(&sum->list);
1868                         kfree(sum);
1869                 }
1870         }
1871 }
1872
1873 static bool check_bg_is_active(struct btrfs_eb_write_context *ctx,
1874                                struct btrfs_block_group **active_bg)
1875 {
1876         const struct writeback_control *wbc = ctx->wbc;
1877         struct btrfs_block_group *block_group = ctx->zoned_bg;
1878         struct btrfs_fs_info *fs_info = block_group->fs_info;
1879
1880         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
1881                 return true;
1882
1883         if (fs_info->treelog_bg == block_group->start) {
1884                 if (!btrfs_zone_activate(block_group)) {
1885                         int ret_fin = btrfs_zone_finish_one_bg(fs_info);
1886
1887                         if (ret_fin != 1 || !btrfs_zone_activate(block_group))
1888                                 return false;
1889                 }
1890         } else if (*active_bg != block_group) {
1891                 struct btrfs_block_group *tgt = *active_bg;
1892
1893                 /* zoned_meta_io_lock protects fs_info->active_{meta,system}_bg. */
1894                 lockdep_assert_held(&fs_info->zoned_meta_io_lock);
1895
1896                 if (tgt) {
1897                         /*
1898                          * If there is an unsent IO left in the allocated area,
1899                          * we cannot wait for them as it may cause a deadlock.
1900                          */
1901                         if (tgt->meta_write_pointer < tgt->start + tgt->alloc_offset) {
1902                                 if (wbc->sync_mode == WB_SYNC_NONE ||
1903                                     (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync))
1904                                         return false;
1905                         }
1906
1907                         /* Pivot active metadata/system block group. */
1908                         btrfs_zoned_meta_io_unlock(fs_info);
1909                         wait_eb_writebacks(tgt);
1910                         do_zone_finish(tgt, true);
1911                         btrfs_zoned_meta_io_lock(fs_info);
1912                         if (*active_bg == tgt) {
1913                                 btrfs_put_block_group(tgt);
1914                                 *active_bg = NULL;
1915                         }
1916                 }
1917                 if (!btrfs_zone_activate(block_group))
1918                         return false;
1919                 if (*active_bg != block_group) {
1920                         ASSERT(*active_bg == NULL);
1921                         *active_bg = block_group;
1922                         btrfs_get_block_group(block_group);
1923                 }
1924         }
1925
1926         return true;
1927 }
1928
1929 /*
1930  * Check if @ctx->eb is aligned to the write pointer.
1931  *
1932  * Return:
1933  *   0:        @ctx->eb is at the write pointer. You can write it.
1934  *   -EAGAIN:  There is a hole. The caller should handle the case.
1935  *   -EBUSY:   There is a hole, but the caller can just bail out.
1936  */
1937 int btrfs_check_meta_write_pointer(struct btrfs_fs_info *fs_info,
1938                                    struct btrfs_eb_write_context *ctx)
1939 {
1940         const struct writeback_control *wbc = ctx->wbc;
1941         const struct extent_buffer *eb = ctx->eb;
1942         struct btrfs_block_group *block_group = ctx->zoned_bg;
1943
1944         if (!btrfs_is_zoned(fs_info))
1945                 return 0;
1946
1947         if (block_group) {
1948                 if (block_group->start > eb->start ||
1949                     block_group->start + block_group->length <= eb->start) {
1950                         btrfs_put_block_group(block_group);
1951                         block_group = NULL;
1952                         ctx->zoned_bg = NULL;
1953                 }
1954         }
1955
1956         if (!block_group) {
1957                 block_group = btrfs_lookup_block_group(fs_info, eb->start);
1958                 if (!block_group)
1959                         return 0;
1960                 ctx->zoned_bg = block_group;
1961         }
1962
1963         if (block_group->meta_write_pointer == eb->start) {
1964                 struct btrfs_block_group **tgt;
1965
1966                 if (!test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags))
1967                         return 0;
1968
1969                 if (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM)
1970                         tgt = &fs_info->active_system_bg;
1971                 else
1972                         tgt = &fs_info->active_meta_bg;
1973                 if (check_bg_is_active(ctx, tgt))
1974                         return 0;
1975         }
1976
1977         /*
1978          * Since we may release fs_info->zoned_meta_io_lock, someone can already
1979          * start writing this eb. In that case, we can just bail out.
1980          */
1981         if (block_group->meta_write_pointer > eb->start)
1982                 return -EBUSY;
1983
1984         /* If for_sync, this hole will be filled with trasnsaction commit. */
1985         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
1986                 return -EAGAIN;
1987         return -EBUSY;
1988 }
1989
1990 int btrfs_zoned_issue_zeroout(struct btrfs_device *device, u64 physical, u64 length)
1991 {
1992         if (!btrfs_dev_is_sequential(device, physical))
1993                 return -EOPNOTSUPP;
1994
1995         return blkdev_issue_zeroout(device->bdev, physical >> SECTOR_SHIFT,
1996                                     length >> SECTOR_SHIFT, GFP_NOFS, 0);
1997 }
1998
1999 static int read_zone_info(struct btrfs_fs_info *fs_info, u64 logical,
2000                           struct blk_zone *zone)
2001 {
2002         struct btrfs_io_context *bioc = NULL;
2003         u64 mapped_length = PAGE_SIZE;
2004         unsigned int nofs_flag;
2005         int nmirrors;
2006         int i, ret;
2007
2008         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2009                               &mapped_length, &bioc, NULL, NULL);
2010         if (ret || !bioc || mapped_length < PAGE_SIZE) {
2011                 ret = -EIO;
2012                 goto out_put_bioc;
2013         }
2014
2015         if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
2016                 ret = -EINVAL;
2017                 goto out_put_bioc;
2018         }
2019
2020         nofs_flag = memalloc_nofs_save();
2021         nmirrors = (int)bioc->num_stripes;
2022         for (i = 0; i < nmirrors; i++) {
2023                 u64 physical = bioc->stripes[i].physical;
2024                 struct btrfs_device *dev = bioc->stripes[i].dev;
2025
2026                 /* Missing device */
2027                 if (!dev->bdev)
2028                         continue;
2029
2030                 ret = btrfs_get_dev_zone(dev, physical, zone);
2031                 /* Failing device */
2032                 if (ret == -EIO || ret == -EOPNOTSUPP)
2033                         continue;
2034                 break;
2035         }
2036         memalloc_nofs_restore(nofs_flag);
2037 out_put_bioc:
2038         btrfs_put_bioc(bioc);
2039         return ret;
2040 }
2041
2042 /*
2043  * Synchronize write pointer in a zone at @physical_start on @tgt_dev, by
2044  * filling zeros between @physical_pos to a write pointer of dev-replace
2045  * source device.
2046  */
2047 int btrfs_sync_zone_write_pointer(struct btrfs_device *tgt_dev, u64 logical,
2048                                     u64 physical_start, u64 physical_pos)
2049 {
2050         struct btrfs_fs_info *fs_info = tgt_dev->fs_info;
2051         struct blk_zone zone;
2052         u64 length;
2053         u64 wp;
2054         int ret;
2055
2056         if (!btrfs_dev_is_sequential(tgt_dev, physical_pos))
2057                 return 0;
2058
2059         ret = read_zone_info(fs_info, logical, &zone);
2060         if (ret)
2061                 return ret;
2062
2063         wp = physical_start + ((zone.wp - zone.start) << SECTOR_SHIFT);
2064
2065         if (physical_pos == wp)
2066                 return 0;
2067
2068         if (physical_pos > wp)
2069                 return -EUCLEAN;
2070
2071         length = wp - physical_pos;
2072         return btrfs_zoned_issue_zeroout(tgt_dev, physical_pos, length);
2073 }
2074
2075 /*
2076  * Activate block group and underlying device zones
2077  *
2078  * @block_group: the block group to activate
2079  *
2080  * Return: true on success, false otherwise
2081  */
2082 bool btrfs_zone_activate(struct btrfs_block_group *block_group)
2083 {
2084         struct btrfs_fs_info *fs_info = block_group->fs_info;
2085         struct map_lookup *map;
2086         struct btrfs_device *device;
2087         u64 physical;
2088         const bool is_data = (block_group->flags & BTRFS_BLOCK_GROUP_DATA);
2089         bool ret;
2090         int i;
2091
2092         if (!btrfs_is_zoned(block_group->fs_info))
2093                 return true;
2094
2095         map = block_group->physical_map;
2096
2097         spin_lock(&block_group->lock);
2098         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
2099                 ret = true;
2100                 goto out_unlock;
2101         }
2102
2103         /* No space left */
2104         if (btrfs_zoned_bg_is_full(block_group)) {
2105                 ret = false;
2106                 goto out_unlock;
2107         }
2108
2109         spin_lock(&fs_info->zone_active_bgs_lock);
2110         for (i = 0; i < map->num_stripes; i++) {
2111                 struct btrfs_zoned_device_info *zinfo;
2112                 int reserved = 0;
2113
2114                 device = map->stripes[i].dev;
2115                 physical = map->stripes[i].physical;
2116                 zinfo = device->zone_info;
2117
2118                 if (zinfo->max_active_zones == 0)
2119                         continue;
2120
2121                 if (is_data)
2122                         reserved = zinfo->reserved_active_zones;
2123                 /*
2124                  * For the data block group, leave active zones for one
2125                  * metadata block group and one system block group.
2126                  */
2127                 if (atomic_read(&zinfo->active_zones_left) <= reserved) {
2128                         ret = false;
2129                         spin_unlock(&fs_info->zone_active_bgs_lock);
2130                         goto out_unlock;
2131                 }
2132
2133                 if (!btrfs_dev_set_active_zone(device, physical)) {
2134                         /* Cannot activate the zone */
2135                         ret = false;
2136                         spin_unlock(&fs_info->zone_active_bgs_lock);
2137                         goto out_unlock;
2138                 }
2139                 if (!is_data)
2140                         zinfo->reserved_active_zones--;
2141         }
2142         spin_unlock(&fs_info->zone_active_bgs_lock);
2143
2144         /* Successfully activated all the zones */
2145         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2146         spin_unlock(&block_group->lock);
2147
2148         /* For the active block group list */
2149         btrfs_get_block_group(block_group);
2150
2151         spin_lock(&fs_info->zone_active_bgs_lock);
2152         list_add_tail(&block_group->active_bg_list, &fs_info->zone_active_bgs);
2153         spin_unlock(&fs_info->zone_active_bgs_lock);
2154
2155         return true;
2156
2157 out_unlock:
2158         spin_unlock(&block_group->lock);
2159         return ret;
2160 }
2161
2162 static void wait_eb_writebacks(struct btrfs_block_group *block_group)
2163 {
2164         struct btrfs_fs_info *fs_info = block_group->fs_info;
2165         const u64 end = block_group->start + block_group->length;
2166         struct radix_tree_iter iter;
2167         struct extent_buffer *eb;
2168         void __rcu **slot;
2169
2170         rcu_read_lock();
2171         radix_tree_for_each_slot(slot, &fs_info->buffer_radix, &iter,
2172                                  block_group->start >> fs_info->sectorsize_bits) {
2173                 eb = radix_tree_deref_slot(slot);
2174                 if (!eb)
2175                         continue;
2176                 if (radix_tree_deref_retry(eb)) {
2177                         slot = radix_tree_iter_retry(&iter);
2178                         continue;
2179                 }
2180
2181                 if (eb->start < block_group->start)
2182                         continue;
2183                 if (eb->start >= end)
2184                         break;
2185
2186                 slot = radix_tree_iter_resume(slot, &iter);
2187                 rcu_read_unlock();
2188                 wait_on_extent_buffer_writeback(eb);
2189                 rcu_read_lock();
2190         }
2191         rcu_read_unlock();
2192 }
2193
2194 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written)
2195 {
2196         struct btrfs_fs_info *fs_info = block_group->fs_info;
2197         struct map_lookup *map;
2198         const bool is_metadata = (block_group->flags &
2199                         (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM));
2200         int ret = 0;
2201         int i;
2202
2203         spin_lock(&block_group->lock);
2204         if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
2205                 spin_unlock(&block_group->lock);
2206                 return 0;
2207         }
2208
2209         /* Check if we have unwritten allocated space */
2210         if (is_metadata &&
2211             block_group->start + block_group->alloc_offset > block_group->meta_write_pointer) {
2212                 spin_unlock(&block_group->lock);
2213                 return -EAGAIN;
2214         }
2215
2216         /*
2217          * If we are sure that the block group is full (= no more room left for
2218          * new allocation) and the IO for the last usable block is completed, we
2219          * don't need to wait for the other IOs. This holds because we ensure
2220          * the sequential IO submissions using the ZONE_APPEND command for data
2221          * and block_group->meta_write_pointer for metadata.
2222          */
2223         if (!fully_written) {
2224                 if (test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
2225                         spin_unlock(&block_group->lock);
2226                         return -EAGAIN;
2227                 }
2228                 spin_unlock(&block_group->lock);
2229
2230                 ret = btrfs_inc_block_group_ro(block_group, false);
2231                 if (ret)
2232                         return ret;
2233
2234                 /* Ensure all writes in this block group finish */
2235                 btrfs_wait_block_group_reservations(block_group);
2236                 /* No need to wait for NOCOW writers. Zoned mode does not allow that */
2237                 btrfs_wait_ordered_roots(fs_info, U64_MAX, block_group->start,
2238                                          block_group->length);
2239                 /* Wait for extent buffers to be written. */
2240                 if (is_metadata)
2241                         wait_eb_writebacks(block_group);
2242
2243                 spin_lock(&block_group->lock);
2244
2245                 /*
2246                  * Bail out if someone already deactivated the block group, or
2247                  * allocated space is left in the block group.
2248                  */
2249                 if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2250                               &block_group->runtime_flags)) {
2251                         spin_unlock(&block_group->lock);
2252                         btrfs_dec_block_group_ro(block_group);
2253                         return 0;
2254                 }
2255
2256                 if (block_group->reserved ||
2257                     test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2258                              &block_group->runtime_flags)) {
2259                         spin_unlock(&block_group->lock);
2260                         btrfs_dec_block_group_ro(block_group);
2261                         return -EAGAIN;
2262                 }
2263         }
2264
2265         clear_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2266         block_group->alloc_offset = block_group->zone_capacity;
2267         if (block_group->flags & (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM))
2268                 block_group->meta_write_pointer = block_group->start +
2269                                                   block_group->zone_capacity;
2270         block_group->free_space_ctl->free_space = 0;
2271         btrfs_clear_treelog_bg(block_group);
2272         btrfs_clear_data_reloc_bg(block_group);
2273         spin_unlock(&block_group->lock);
2274
2275         map = block_group->physical_map;
2276         for (i = 0; i < map->num_stripes; i++) {
2277                 struct btrfs_device *device = map->stripes[i].dev;
2278                 const u64 physical = map->stripes[i].physical;
2279                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2280
2281                 if (zinfo->max_active_zones == 0)
2282                         continue;
2283
2284                 ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_FINISH,
2285                                        physical >> SECTOR_SHIFT,
2286                                        zinfo->zone_size >> SECTOR_SHIFT,
2287                                        GFP_NOFS);
2288
2289                 if (ret)
2290                         return ret;
2291
2292                 if (!(block_group->flags & BTRFS_BLOCK_GROUP_DATA))
2293                         zinfo->reserved_active_zones++;
2294                 btrfs_dev_clear_active_zone(device, physical);
2295         }
2296
2297         if (!fully_written)
2298                 btrfs_dec_block_group_ro(block_group);
2299
2300         spin_lock(&fs_info->zone_active_bgs_lock);
2301         ASSERT(!list_empty(&block_group->active_bg_list));
2302         list_del_init(&block_group->active_bg_list);
2303         spin_unlock(&fs_info->zone_active_bgs_lock);
2304
2305         /* For active_bg_list */
2306         btrfs_put_block_group(block_group);
2307
2308         clear_and_wake_up_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2309
2310         return 0;
2311 }
2312
2313 int btrfs_zone_finish(struct btrfs_block_group *block_group)
2314 {
2315         if (!btrfs_is_zoned(block_group->fs_info))
2316                 return 0;
2317
2318         return do_zone_finish(block_group, false);
2319 }
2320
2321 bool btrfs_can_activate_zone(struct btrfs_fs_devices *fs_devices, u64 flags)
2322 {
2323         struct btrfs_fs_info *fs_info = fs_devices->fs_info;
2324         struct btrfs_device *device;
2325         bool ret = false;
2326
2327         if (!btrfs_is_zoned(fs_info))
2328                 return true;
2329
2330         /* Check if there is a device with active zones left */
2331         mutex_lock(&fs_info->chunk_mutex);
2332         spin_lock(&fs_info->zone_active_bgs_lock);
2333         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
2334                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2335                 int reserved = 0;
2336
2337                 if (!device->bdev)
2338                         continue;
2339
2340                 if (!zinfo->max_active_zones) {
2341                         ret = true;
2342                         break;
2343                 }
2344
2345                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2346                         reserved = zinfo->reserved_active_zones;
2347
2348                 switch (flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
2349                 case 0: /* single */
2350                         ret = (atomic_read(&zinfo->active_zones_left) >= (1 + reserved));
2351                         break;
2352                 case BTRFS_BLOCK_GROUP_DUP:
2353                         ret = (atomic_read(&zinfo->active_zones_left) >= (2 + reserved));
2354                         break;
2355                 }
2356                 if (ret)
2357                         break;
2358         }
2359         spin_unlock(&fs_info->zone_active_bgs_lock);
2360         mutex_unlock(&fs_info->chunk_mutex);
2361
2362         if (!ret)
2363                 set_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2364
2365         return ret;
2366 }
2367
2368 void btrfs_zone_finish_endio(struct btrfs_fs_info *fs_info, u64 logical, u64 length)
2369 {
2370         struct btrfs_block_group *block_group;
2371         u64 min_alloc_bytes;
2372
2373         if (!btrfs_is_zoned(fs_info))
2374                 return;
2375
2376         block_group = btrfs_lookup_block_group(fs_info, logical);
2377         ASSERT(block_group);
2378
2379         /* No MIXED_BG on zoned btrfs. */
2380         if (block_group->flags & BTRFS_BLOCK_GROUP_DATA)
2381                 min_alloc_bytes = fs_info->sectorsize;
2382         else
2383                 min_alloc_bytes = fs_info->nodesize;
2384
2385         /* Bail out if we can allocate more data from this block group. */
2386         if (logical + length + min_alloc_bytes <=
2387             block_group->start + block_group->zone_capacity)
2388                 goto out;
2389
2390         do_zone_finish(block_group, true);
2391
2392 out:
2393         btrfs_put_block_group(block_group);
2394 }
2395
2396 static void btrfs_zone_finish_endio_workfn(struct work_struct *work)
2397 {
2398         struct btrfs_block_group *bg =
2399                 container_of(work, struct btrfs_block_group, zone_finish_work);
2400
2401         wait_on_extent_buffer_writeback(bg->last_eb);
2402         free_extent_buffer(bg->last_eb);
2403         btrfs_zone_finish_endio(bg->fs_info, bg->start, bg->length);
2404         btrfs_put_block_group(bg);
2405 }
2406
2407 void btrfs_schedule_zone_finish_bg(struct btrfs_block_group *bg,
2408                                    struct extent_buffer *eb)
2409 {
2410         if (!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &bg->runtime_flags) ||
2411             eb->start + eb->len * 2 <= bg->start + bg->zone_capacity)
2412                 return;
2413
2414         if (WARN_ON(bg->zone_finish_work.func == btrfs_zone_finish_endio_workfn)) {
2415                 btrfs_err(bg->fs_info, "double scheduling of bg %llu zone finishing",
2416                           bg->start);
2417                 return;
2418         }
2419
2420         /* For the work */
2421         btrfs_get_block_group(bg);
2422         atomic_inc(&eb->refs);
2423         bg->last_eb = eb;
2424         INIT_WORK(&bg->zone_finish_work, btrfs_zone_finish_endio_workfn);
2425         queue_work(system_unbound_wq, &bg->zone_finish_work);
2426 }
2427
2428 void btrfs_clear_data_reloc_bg(struct btrfs_block_group *bg)
2429 {
2430         struct btrfs_fs_info *fs_info = bg->fs_info;
2431
2432         spin_lock(&fs_info->relocation_bg_lock);
2433         if (fs_info->data_reloc_bg == bg->start)
2434                 fs_info->data_reloc_bg = 0;
2435         spin_unlock(&fs_info->relocation_bg_lock);
2436 }
2437
2438 void btrfs_free_zone_cache(struct btrfs_fs_info *fs_info)
2439 {
2440         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2441         struct btrfs_device *device;
2442
2443         if (!btrfs_is_zoned(fs_info))
2444                 return;
2445
2446         mutex_lock(&fs_devices->device_list_mutex);
2447         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2448                 if (device->zone_info) {
2449                         vfree(device->zone_info->zone_cache);
2450                         device->zone_info->zone_cache = NULL;
2451                 }
2452         }
2453         mutex_unlock(&fs_devices->device_list_mutex);
2454 }
2455
2456 bool btrfs_zoned_should_reclaim(struct btrfs_fs_info *fs_info)
2457 {
2458         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2459         struct btrfs_device *device;
2460         u64 used = 0;
2461         u64 total = 0;
2462         u64 factor;
2463
2464         ASSERT(btrfs_is_zoned(fs_info));
2465
2466         if (fs_info->bg_reclaim_threshold == 0)
2467                 return false;
2468
2469         mutex_lock(&fs_devices->device_list_mutex);
2470         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2471                 if (!device->bdev)
2472                         continue;
2473
2474                 total += device->disk_total_bytes;
2475                 used += device->bytes_used;
2476         }
2477         mutex_unlock(&fs_devices->device_list_mutex);
2478
2479         factor = div64_u64(used * 100, total);
2480         return factor >= fs_info->bg_reclaim_threshold;
2481 }
2482
2483 void btrfs_zoned_release_data_reloc_bg(struct btrfs_fs_info *fs_info, u64 logical,
2484                                        u64 length)
2485 {
2486         struct btrfs_block_group *block_group;
2487
2488         if (!btrfs_is_zoned(fs_info))
2489                 return;
2490
2491         block_group = btrfs_lookup_block_group(fs_info, logical);
2492         /* It should be called on a previous data relocation block group. */
2493         ASSERT(block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA));
2494
2495         spin_lock(&block_group->lock);
2496         if (!test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))
2497                 goto out;
2498
2499         /* All relocation extents are written. */
2500         if (block_group->start + block_group->alloc_offset == logical + length) {
2501                 /*
2502                  * Now, release this block group for further allocations and
2503                  * zone finish.
2504                  */
2505                 clear_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2506                           &block_group->runtime_flags);
2507         }
2508
2509 out:
2510         spin_unlock(&block_group->lock);
2511         btrfs_put_block_group(block_group);
2512 }
2513
2514 int btrfs_zone_finish_one_bg(struct btrfs_fs_info *fs_info)
2515 {
2516         struct btrfs_block_group *block_group;
2517         struct btrfs_block_group *min_bg = NULL;
2518         u64 min_avail = U64_MAX;
2519         int ret;
2520
2521         spin_lock(&fs_info->zone_active_bgs_lock);
2522         list_for_each_entry(block_group, &fs_info->zone_active_bgs,
2523                             active_bg_list) {
2524                 u64 avail;
2525
2526                 spin_lock(&block_group->lock);
2527                 if (block_group->reserved || block_group->alloc_offset == 0 ||
2528                     (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM) ||
2529                     test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
2530                         spin_unlock(&block_group->lock);
2531                         continue;
2532                 }
2533
2534                 avail = block_group->zone_capacity - block_group->alloc_offset;
2535                 if (min_avail > avail) {
2536                         if (min_bg)
2537                                 btrfs_put_block_group(min_bg);
2538                         min_bg = block_group;
2539                         min_avail = avail;
2540                         btrfs_get_block_group(min_bg);
2541                 }
2542                 spin_unlock(&block_group->lock);
2543         }
2544         spin_unlock(&fs_info->zone_active_bgs_lock);
2545
2546         if (!min_bg)
2547                 return 0;
2548
2549         ret = btrfs_zone_finish(min_bg);
2550         btrfs_put_block_group(min_bg);
2551
2552         return ret < 0 ? ret : 1;
2553 }
2554
2555 int btrfs_zoned_activate_one_bg(struct btrfs_fs_info *fs_info,
2556                                 struct btrfs_space_info *space_info,
2557                                 bool do_finish)
2558 {
2559         struct btrfs_block_group *bg;
2560         int index;
2561
2562         if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
2563                 return 0;
2564
2565         for (;;) {
2566                 int ret;
2567                 bool need_finish = false;
2568
2569                 down_read(&space_info->groups_sem);
2570                 for (index = 0; index < BTRFS_NR_RAID_TYPES; index++) {
2571                         list_for_each_entry(bg, &space_info->block_groups[index],
2572                                             list) {
2573                                 if (!spin_trylock(&bg->lock))
2574                                         continue;
2575                                 if (btrfs_zoned_bg_is_full(bg) ||
2576                                     test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2577                                              &bg->runtime_flags)) {
2578                                         spin_unlock(&bg->lock);
2579                                         continue;
2580                                 }
2581                                 spin_unlock(&bg->lock);
2582
2583                                 if (btrfs_zone_activate(bg)) {
2584                                         up_read(&space_info->groups_sem);
2585                                         return 1;
2586                                 }
2587
2588                                 need_finish = true;
2589                         }
2590                 }
2591                 up_read(&space_info->groups_sem);
2592
2593                 if (!do_finish || !need_finish)
2594                         break;
2595
2596                 ret = btrfs_zone_finish_one_bg(fs_info);
2597                 if (ret == 0)
2598                         break;
2599                 if (ret < 0)
2600                         return ret;
2601         }
2602
2603         return 0;
2604 }
2605
2606 /*
2607  * Reserve zones for one metadata block group, one tree-log block group, and one
2608  * system block group.
2609  */
2610 void btrfs_check_active_zone_reservation(struct btrfs_fs_info *fs_info)
2611 {
2612         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2613         struct btrfs_block_group *block_group;
2614         struct btrfs_device *device;
2615         /* Reserve zones for normal SINGLE metadata and tree-log block group. */
2616         unsigned int metadata_reserve = 2;
2617         /* Reserve a zone for SINGLE system block group. */
2618         unsigned int system_reserve = 1;
2619
2620         if (!test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags))
2621                 return;
2622
2623         /*
2624          * This function is called from the mount context. So, there is no
2625          * parallel process touching the bits. No need for read_seqretry().
2626          */
2627         if (fs_info->avail_metadata_alloc_bits & BTRFS_BLOCK_GROUP_DUP)
2628                 metadata_reserve = 4;
2629         if (fs_info->avail_system_alloc_bits & BTRFS_BLOCK_GROUP_DUP)
2630                 system_reserve = 2;
2631
2632         /* Apply the reservation on all the devices. */
2633         mutex_lock(&fs_devices->device_list_mutex);
2634         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2635                 if (!device->bdev)
2636                         continue;
2637
2638                 device->zone_info->reserved_active_zones =
2639                         metadata_reserve + system_reserve;
2640         }
2641         mutex_unlock(&fs_devices->device_list_mutex);
2642
2643         /* Release reservation for currently active block groups. */
2644         spin_lock(&fs_info->zone_active_bgs_lock);
2645         list_for_each_entry(block_group, &fs_info->zone_active_bgs, active_bg_list) {
2646                 struct map_lookup *map = block_group->physical_map;
2647
2648                 if (!(block_group->flags &
2649                       (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM)))
2650                         continue;
2651
2652                 for (int i = 0; i < map->num_stripes; i++)
2653                         map->stripes[i].dev->zone_info->reserved_active_zones--;
2654         }
2655         spin_unlock(&fs_info->zone_active_bgs_lock);
2656 }