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         if (bdev_is_zoned(bdev)) {
 582                 model = "host-managed zoned";
 583                 emulated = "";
 584         } else {
 585                 model = "regular";
 586                 emulated = "emulated ";
 587         }
 588
 589         btrfs_info_in_rcu(fs_info,
 590                 "%s block device %s, %u %szones of %llu bytes",
 591                 model, rcu_str_deref(device->name), zone_info->nr_zones,
 592                 emulated, zone_info->zone_size);
 593
 594         return 0;
 595
 596 out:
 597         kvfree(zones);
 598         btrfs_destroy_dev_zone_info(device);
 599         return ret;
 600 }
 601
 602 void btrfs_destroy_dev_zone_info(struct btrfs_device *device)
 603 {
 604         struct btrfs_zoned_device_info *zone_info = device->zone_info;
 605
 606         if (!zone_info)
 607                 return;
 608
 609         bitmap_free(zone_info->active_zones);
 610         bitmap_free(zone_info->seq_zones);
 611         bitmap_free(zone_info->empty_zones);
 612         vfree(zone_info->zone_cache);
 613         kfree(zone_info);
 614         device->zone_info = NULL;
 615 }
 616
 617 struct btrfs_zoned_device_info *btrfs_clone_dev_zone_info(struct btrfs_device *orig_dev)
 618 {
 619         struct btrfs_zoned_device_info *zone_info;
 620
 621         zone_info = kmemdup(orig_dev->zone_info, sizeof(*zone_info), GFP_KERNEL);
 622         if (!zone_info)
 623                 return NULL;
 624
 625         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 626         if (!zone_info->seq_zones)
 627                 goto out;
 628
 629         bitmap_copy(zone_info->seq_zones, orig_dev->zone_info->seq_zones,
 630                     zone_info->nr_zones);
 631
 632         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 633         if (!zone_info->empty_zones)
 634                 goto out;
 635
 636         bitmap_copy(zone_info->empty_zones, orig_dev->zone_info->empty_zones,
 637                     zone_info->nr_zones);
 638
 639         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 640         if (!zone_info->active_zones)
 641                 goto out;
 642
 643         bitmap_copy(zone_info->active_zones, orig_dev->zone_info->active_zones,
 644                     zone_info->nr_zones);
 645         zone_info->zone_cache = NULL;
 646
 647         return zone_info;
 648
 649 out:
 650         bitmap_free(zone_info->seq_zones);
 651         bitmap_free(zone_info->empty_zones);
 652         bitmap_free(zone_info->active_zones);
 653         kfree(zone_info);
 654         return NULL;
 655 }
 656
 657 int btrfs_get_dev_zone(struct btrfs_device *device, u64 pos,
 658                        struct blk_zone *zone)
 659 {
 660         unsigned int nr_zones = 1;
 661         int ret;
 662
 663         ret = btrfs_get_dev_zones(device, pos, zone, &nr_zones);
 664         if (ret != 0 || !nr_zones)
 665                 return ret ? ret : -EIO;
 666
 667         return 0;
 668 }
 669
 670 static int btrfs_check_for_zoned_device(struct btrfs_fs_info *fs_info)
 671 {
 672         struct btrfs_device *device;
 673
 674         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 675                 if (device->bdev && bdev_is_zoned(device->bdev)) {
 676                         btrfs_err(fs_info,
 677                                 "zoned: mode not enabled but zoned device found: %pg",
 678                                 device->bdev);
 679                         return -EINVAL;
 680                 }
 681         }
 682
 683         return 0;
 684 }
 685
 686 int btrfs_check_zoned_mode(struct btrfs_fs_info *fs_info)
 687 {
 688         struct queue_limits *lim = &fs_info->limits;
 689         struct btrfs_device *device;
 690         u64 zone_size = 0;
 691         int ret;
 692
 693         /*
 694          * Host-Managed devices can't be used without the ZONED flag.  With the
 695          * ZONED all devices can be used, using zone emulation if required.
 696          */
 697         if (!btrfs_fs_incompat(fs_info, ZONED))
 698                 return btrfs_check_for_zoned_device(fs_info);
 699
 700         blk_set_stacking_limits(lim);
 701
 702         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 703                 struct btrfs_zoned_device_info *zone_info = device->zone_info;
 704
 705                 if (!device->bdev)
 706                         continue;
 707
 708                 if (!zone_size) {
 709                         zone_size = zone_info->zone_size;
 710                 } else if (zone_info->zone_size != zone_size) {
 711                         btrfs_err(fs_info,
 712                 "zoned: unequal block device zone sizes: have %llu found %llu",
 713                                   zone_info->zone_size, zone_size);
 714                         return -EINVAL;
 715                 }
 716
 717                 /*
 718                  * With the zoned emulation, we can have non-zoned device on the
 719                  * zoned mode. In this case, we don't have a valid max zone
 720                  * append size.
 721                  */
 722                 if (bdev_is_zoned(device->bdev)) {
 723                         blk_stack_limits(lim,
 724                                          &bdev_get_queue(device->bdev)->limits,
 725                                          0);
 726                 }
 727         }
 728
 729         /*
 730          * stripe_size is always aligned to BTRFS_STRIPE_LEN in
 731          * btrfs_create_chunk(). Since we want stripe_len == zone_size,
 732          * check the alignment here.
 733          */
 734         if (!IS_ALIGNED(zone_size, BTRFS_STRIPE_LEN)) {
 735                 btrfs_err(fs_info,
 736                           "zoned: zone size %llu not aligned to stripe %u",
 737                           zone_size, BTRFS_STRIPE_LEN);
 738                 return -EINVAL;
 739         }
 740
 741         if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
 742                 btrfs_err(fs_info, "zoned: mixed block groups not supported");
 743                 return -EINVAL;
 744         }
 745
 746         fs_info->zone_size = zone_size;
 747         /*
 748          * Also limit max_zone_append_size by max_segments * PAGE_SIZE.
 749          * Technically, we can have multiple pages per segment. But, since
 750          * we add the pages one by one to a bio, and cannot increase the
 751          * metadata reservation even if it increases the number of extents, it
 752          * is safe to stick with the limit.
 753          */
 754         fs_info->max_zone_append_size = ALIGN_DOWN(
 755                 min3((u64)lim->max_zone_append_sectors << SECTOR_SHIFT,
 756                      (u64)lim->max_sectors << SECTOR_SHIFT,
 757                      (u64)lim->max_segments << PAGE_SHIFT),
 758                 fs_info->sectorsize);
 759         fs_info->fs_devices->chunk_alloc_policy = BTRFS_CHUNK_ALLOC_ZONED;
 760         if (fs_info->max_zone_append_size < fs_info->max_extent_size)
 761                 fs_info->max_extent_size = fs_info->max_zone_append_size;
 762
 763         /*
 764          * Check mount options here, because we might change fs_info->zoned
 765          * from fs_info->zone_size.
 766          */
 767         ret = btrfs_check_mountopts_zoned(fs_info, &fs_info->mount_opt);
 768         if (ret)
 769                 return ret;
 770
 771         btrfs_info(fs_info, "zoned mode enabled with zone size %llu", zone_size);
 772         return 0;
 773 }
 774
 775 int btrfs_check_mountopts_zoned(struct btrfs_fs_info *info, unsigned long *mount_opt)
 776 {
 777         if (!btrfs_is_zoned(info))
 778                 return 0;
 779
 780         /*
 781          * Space cache writing is not COWed. Disable that to avoid write errors
 782          * in sequential zones.
 783          */
 784         if (btrfs_raw_test_opt(*mount_opt, SPACE_CACHE)) {
 785                 btrfs_err(info, "zoned: space cache v1 is not supported");
 786                 return -EINVAL;
 787         }
 788
 789         if (btrfs_raw_test_opt(*mount_opt, NODATACOW)) {
 790                 btrfs_err(info, "zoned: NODATACOW not supported");
 791                 return -EINVAL;
 792         }
 793
 794         if (btrfs_raw_test_opt(*mount_opt, DISCARD_ASYNC)) {
 795                 btrfs_info(info,
 796                            "zoned: async discard ignored and disabled for zoned mode");
 797                 btrfs_clear_opt(*mount_opt, DISCARD_ASYNC);
 798         }
 799
 800         return 0;
 801 }
 802
 803 static int sb_log_location(struct block_device *bdev, struct blk_zone *zones,
 804                            int rw, u64 *bytenr_ret)
 805 {
 806         u64 wp;
 807         int ret;
 808
 809         if (zones[0].type == BLK_ZONE_TYPE_CONVENTIONAL) {
 810                 *bytenr_ret = zones[0].start << SECTOR_SHIFT;
 811                 return 0;
 812         }
 813
 814         ret = sb_write_pointer(bdev, zones, &wp);
 815         if (ret != -ENOENT && ret < 0)
 816                 return ret;
 817
 818         if (rw == WRITE) {
 819                 struct blk_zone *reset = NULL;
 820
 821                 if (wp == zones[0].start << SECTOR_SHIFT)
 822                         reset = &zones[0];
 823                 else if (wp == zones[1].start << SECTOR_SHIFT)
 824                         reset = &zones[1];
 825
 826                 if (reset && reset->cond != BLK_ZONE_COND_EMPTY) {
 827                         ASSERT(sb_zone_is_full(reset));
 828
 829                         ret = blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
 830                                                reset->start, reset->len,
 831                                                GFP_NOFS);
 832                         if (ret)
 833                                 return ret;
 834
 835                         reset->cond = BLK_ZONE_COND_EMPTY;
 836                         reset->wp = reset->start;
 837                 }
 838         } else if (ret != -ENOENT) {
 839                 /*
 840                  * For READ, we want the previous one. Move write pointer to
 841                  * the end of a zone, if it is at the head of a zone.
 842                  */
 843                 u64 zone_end = 0;
 844
 845                 if (wp == zones[0].start << SECTOR_SHIFT)
 846                         zone_end = zones[1].start + zones[1].capacity;
 847                 else if (wp == zones[1].start << SECTOR_SHIFT)
 848                         zone_end = zones[0].start + zones[0].capacity;
 849                 if (zone_end)
 850                         wp = ALIGN_DOWN(zone_end << SECTOR_SHIFT,
 851                                         BTRFS_SUPER_INFO_SIZE);
 852
 853                 wp -= BTRFS_SUPER_INFO_SIZE;
 854         }
 855
 856         *bytenr_ret = wp;
 857         return 0;
 858
 859 }
 860
 861 int btrfs_sb_log_location_bdev(struct block_device *bdev, int mirror, int rw,
 862                                u64 *bytenr_ret)
 863 {
 864         struct blk_zone zones[BTRFS_NR_SB_LOG_ZONES];
 865         sector_t zone_sectors;
 866         u32 sb_zone;
 867         int ret;
 868         u8 zone_sectors_shift;
 869         sector_t nr_sectors;
 870         u32 nr_zones;
 871
 872         if (!bdev_is_zoned(bdev)) {
 873                 *bytenr_ret = btrfs_sb_offset(mirror);
 874                 return 0;
 875         }
 876
 877         ASSERT(rw == READ || rw == WRITE);
 878
 879         zone_sectors = bdev_zone_sectors(bdev);
 880         if (!is_power_of_2(zone_sectors))
 881                 return -EINVAL;
 882         zone_sectors_shift = ilog2(zone_sectors);
 883         nr_sectors = bdev_nr_sectors(bdev);
 884         nr_zones = nr_sectors >> zone_sectors_shift;
 885
 886         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
 887         if (sb_zone + 1 >= nr_zones)
 888                 return -ENOENT;
 889
 890         ret = blkdev_report_zones(bdev, zone_start_sector(sb_zone, bdev),
 891                                   BTRFS_NR_SB_LOG_ZONES, copy_zone_info_cb,
 892                                   zones);
 893         if (ret < 0)
 894                 return ret;
 895         if (ret != BTRFS_NR_SB_LOG_ZONES)
 896                 return -EIO;
 897
 898         return sb_log_location(bdev, zones, rw, bytenr_ret);
 899 }
 900
 901 int btrfs_sb_log_location(struct btrfs_device *device, int mirror, int rw,
 902                           u64 *bytenr_ret)
 903 {
 904         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 905         u32 zone_num;
 906
 907         /*
 908          * For a zoned filesystem on a non-zoned block device, use the same
 909          * super block locations as regular filesystem. Doing so, the super
 910          * block can always be retrieved and the zoned flag of the volume
 911          * detected from the super block information.
 912          */
 913         if (!bdev_is_zoned(device->bdev)) {
 914                 *bytenr_ret = btrfs_sb_offset(mirror);
 915                 return 0;
 916         }
 917
 918         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 919         if (zone_num + 1 >= zinfo->nr_zones)
 920                 return -ENOENT;
 921
 922         return sb_log_location(device->bdev,
 923                                &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror],
 924                                rw, bytenr_ret);
 925 }
 926
 927 static inline bool is_sb_log_zone(struct btrfs_zoned_device_info *zinfo,
 928                                   int mirror)
 929 {
 930         u32 zone_num;
 931
 932         if (!zinfo)
 933                 return false;
 934
 935         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 936         if (zone_num + 1 >= zinfo->nr_zones)
 937                 return false;
 938
 939         if (!test_bit(zone_num, zinfo->seq_zones))
 940                 return false;
 941
 942         return true;
 943 }
 944
 945 int btrfs_advance_sb_log(struct btrfs_device *device, int mirror)
 946 {
 947         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 948         struct blk_zone *zone;
 949         int i;
 950
 951         if (!is_sb_log_zone(zinfo, mirror))
 952                 return 0;
 953
 954         zone = &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror];
 955         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 956                 /* Advance the next zone */
 957                 if (zone->cond == BLK_ZONE_COND_FULL) {
 958                         zone++;
 959                         continue;
 960                 }
 961
 962                 if (zone->cond == BLK_ZONE_COND_EMPTY)
 963                         zone->cond = BLK_ZONE_COND_IMP_OPEN;
 964
 965                 zone->wp += SUPER_INFO_SECTORS;
 966
 967                 if (sb_zone_is_full(zone)) {
 968                         /*
 969                          * No room left to write new superblock. Since
 970                          * superblock is written with REQ_SYNC, it is safe to
 971                          * finish the zone now.
 972                          *
 973                          * If the write pointer is exactly at the capacity,
 974                          * explicit ZONE_FINISH is not necessary.
 975                          */
 976                         if (zone->wp != zone->start + zone->capacity) {
 977                                 int ret;
 978
 979                                 ret = blkdev_zone_mgmt(device->bdev,
 980                                                 REQ_OP_ZONE_FINISH, zone->start,
 981                                                 zone->len, GFP_NOFS);
 982                                 if (ret)
 983                                         return ret;
 984                         }
 985
 986                         zone->wp = zone->start + zone->len;
 987                         zone->cond = BLK_ZONE_COND_FULL;
 988                 }
 989                 return 0;
 990         }
 991
 992         /* All the zones are FULL. Should not reach here. */
 993         ASSERT(0);
 994         return -EIO;
 995 }
 996
 997 int btrfs_reset_sb_log_zones(struct block_device *bdev, int mirror)
 998 {
 999         sector_t zone_sectors;
1000         sector_t nr_sectors;
1001         u8 zone_sectors_shift;
1002         u32 sb_zone;
1003         u32 nr_zones;
1004
1005         zone_sectors = bdev_zone_sectors(bdev);
1006         zone_sectors_shift = ilog2(zone_sectors);
1007         nr_sectors = bdev_nr_sectors(bdev);
1008         nr_zones = nr_sectors >> zone_sectors_shift;
1009
1010         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
1011         if (sb_zone + 1 >= nr_zones)
1012                 return -ENOENT;
1013
1014         return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
1015                                 zone_start_sector(sb_zone, bdev),
1016                                 zone_sectors * BTRFS_NR_SB_LOG_ZONES, GFP_NOFS);
1017 }
1018
1019 /*
1020  * Find allocatable zones within a given region.
1021  *
1022  * @device:     the device to allocate a region on
1023  * @hole_start: the position of the hole to allocate the region
1024  * @num_bytes:  size of wanted region
1025  * @hole_end:   the end of the hole
1026  * @return:     position of allocatable zones
1027  *
1028  * Allocatable region should not contain any superblock locations.
1029  */
1030 u64 btrfs_find_allocatable_zones(struct btrfs_device *device, u64 hole_start,
1031                                  u64 hole_end, u64 num_bytes)
1032 {
1033         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1034         const u8 shift = zinfo->zone_size_shift;
1035         u64 nzones = num_bytes >> shift;
1036         u64 pos = hole_start;
1037         u64 begin, end;
1038         bool have_sb;
1039         int i;
1040
1041         ASSERT(IS_ALIGNED(hole_start, zinfo->zone_size));
1042         ASSERT(IS_ALIGNED(num_bytes, zinfo->zone_size));
1043
1044         while (pos < hole_end) {
1045                 begin = pos >> shift;
1046                 end = begin + nzones;
1047
1048                 if (end > zinfo->nr_zones)
1049                         return hole_end;
1050
1051                 /* Check if zones in the region are all empty */
1052                 if (btrfs_dev_is_sequential(device, pos) &&
1053                     !bitmap_test_range_all_set(zinfo->empty_zones, begin, nzones)) {
1054                         pos += zinfo->zone_size;
1055                         continue;
1056                 }
1057
1058                 have_sb = false;
1059                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1060                         u32 sb_zone;
1061                         u64 sb_pos;
1062
1063                         sb_zone = sb_zone_number(shift, i);
1064                         if (!(end <= sb_zone ||
1065                               sb_zone + BTRFS_NR_SB_LOG_ZONES <= begin)) {
1066                                 have_sb = true;
1067                                 pos = zone_start_physical(
1068                                         sb_zone + BTRFS_NR_SB_LOG_ZONES, zinfo);
1069                                 break;
1070                         }
1071
1072                         /* We also need to exclude regular superblock positions */
1073                         sb_pos = btrfs_sb_offset(i);
1074                         if (!(pos + num_bytes <= sb_pos ||
1075                               sb_pos + BTRFS_SUPER_INFO_SIZE <= pos)) {
1076                                 have_sb = true;
1077                                 pos = ALIGN(sb_pos + BTRFS_SUPER_INFO_SIZE,
1078                                             zinfo->zone_size);
1079                                 break;
1080                         }
1081                 }
1082                 if (!have_sb)
1083                         break;
1084         }
1085
1086         return pos;
1087 }
1088
1089 static bool btrfs_dev_set_active_zone(struct btrfs_device *device, u64 pos)
1090 {
1091         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1092         unsigned int zno = (pos >> zone_info->zone_size_shift);
1093
1094         /* We can use any number of zones */
1095         if (zone_info->max_active_zones == 0)
1096                 return true;
1097
1098         if (!test_bit(zno, zone_info->active_zones)) {
1099                 /* Active zone left? */
1100                 if (atomic_dec_if_positive(&zone_info->active_zones_left) < 0)
1101                         return false;
1102                 if (test_and_set_bit(zno, zone_info->active_zones)) {
1103                         /* Someone already set the bit */
1104                         atomic_inc(&zone_info->active_zones_left);
1105                 }
1106         }
1107
1108         return true;
1109 }
1110
1111 static void btrfs_dev_clear_active_zone(struct btrfs_device *device, u64 pos)
1112 {
1113         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1114         unsigned int zno = (pos >> zone_info->zone_size_shift);
1115
1116         /* We can use any number of zones */
1117         if (zone_info->max_active_zones == 0)
1118                 return;
1119
1120         if (test_and_clear_bit(zno, zone_info->active_zones))
1121                 atomic_inc(&zone_info->active_zones_left);
1122 }
1123
1124 int btrfs_reset_device_zone(struct btrfs_device *device, u64 physical,
1125                             u64 length, u64 *bytes)
1126 {
1127         int ret;
1128
1129         *bytes = 0;
1130         ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_RESET,
1131                                physical >> SECTOR_SHIFT, length >> SECTOR_SHIFT,
1132                                GFP_NOFS);
1133         if (ret)
1134                 return ret;
1135
1136         *bytes = length;
1137         while (length) {
1138                 btrfs_dev_set_zone_empty(device, physical);
1139                 btrfs_dev_clear_active_zone(device, physical);
1140                 physical += device->zone_info->zone_size;
1141                 length -= device->zone_info->zone_size;
1142         }
1143
1144         return 0;
1145 }
1146
1147 int btrfs_ensure_empty_zones(struct btrfs_device *device, u64 start, u64 size)
1148 {
1149         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1150         const u8 shift = zinfo->zone_size_shift;
1151         unsigned long begin = start >> shift;
1152         unsigned long nbits = size >> shift;
1153         u64 pos;
1154         int ret;
1155
1156         ASSERT(IS_ALIGNED(start, zinfo->zone_size));
1157         ASSERT(IS_ALIGNED(size, zinfo->zone_size));
1158
1159         if (begin + nbits > zinfo->nr_zones)
1160                 return -ERANGE;
1161
1162         /* All the zones are conventional */
1163         if (bitmap_test_range_all_zero(zinfo->seq_zones, begin, nbits))
1164                 return 0;
1165
1166         /* All the zones are sequential and empty */
1167         if (bitmap_test_range_all_set(zinfo->seq_zones, begin, nbits) &&
1168             bitmap_test_range_all_set(zinfo->empty_zones, begin, nbits))
1169                 return 0;
1170
1171         for (pos = start; pos < start + size; pos += zinfo->zone_size) {
1172                 u64 reset_bytes;
1173
1174                 if (!btrfs_dev_is_sequential(device, pos) ||
1175                     btrfs_dev_is_empty_zone(device, pos))
1176                         continue;
1177
1178                 /* Free regions should be empty */
1179                 btrfs_warn_in_rcu(
1180                         device->fs_info,
1181                 "zoned: resetting device %s (devid %llu) zone %llu for allocation",
1182                         rcu_str_deref(device->name), device->devid, pos >> shift);
1183                 WARN_ON_ONCE(1);
1184
1185                 ret = btrfs_reset_device_zone(device, pos, zinfo->zone_size,
1186                                               &reset_bytes);
1187                 if (ret)
1188                         return ret;
1189         }
1190
1191         return 0;
1192 }
1193
1194 /*
1195  * Calculate an allocation pointer from the extent allocation information
1196  * for a block group consist of conventional zones. It is pointed to the
1197  * end of the highest addressed extent in the block group as an allocation
1198  * offset.
1199  */
1200 static int calculate_alloc_pointer(struct btrfs_block_group *cache,
1201                                    u64 *offset_ret, bool new)
1202 {
1203         struct btrfs_fs_info *fs_info = cache->fs_info;
1204         struct btrfs_root *root;
1205         struct btrfs_path *path;
1206         struct btrfs_key key;
1207         struct btrfs_key found_key;
1208         int ret;
1209         u64 length;
1210
1211         /*
1212          * Avoid  tree lookups for a new block group, there's no use for it.
1213          * It must always be 0.
1214          *
1215          * Also, we have a lock chain of extent buffer lock -> chunk mutex.
1216          * For new a block group, this function is called from
1217          * btrfs_make_block_group() which is already taking the chunk mutex.
1218          * Thus, we cannot call calculate_alloc_pointer() which takes extent
1219          * buffer locks to avoid deadlock.
1220          */
1221         if (new) {
1222                 *offset_ret = 0;
1223                 return 0;
1224         }
1225
1226         path = btrfs_alloc_path();
1227         if (!path)
1228                 return -ENOMEM;
1229
1230         key.objectid = cache->start + cache->length;
1231         key.type = 0;
1232         key.offset = 0;
1233
1234         root = btrfs_extent_root(fs_info, key.objectid);
1235         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1236         /* We should not find the exact match */
1237         if (!ret)
1238                 ret = -EUCLEAN;
1239         if (ret < 0)
1240                 goto out;
1241
1242         ret = btrfs_previous_extent_item(root, path, cache->start);
1243         if (ret) {
1244                 if (ret == 1) {
1245                         ret = 0;
1246                         *offset_ret = 0;
1247                 }
1248                 goto out;
1249         }
1250
1251         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
1252
1253         if (found_key.type == BTRFS_EXTENT_ITEM_KEY)
1254                 length = found_key.offset;
1255         else
1256                 length = fs_info->nodesize;
1257
1258         if (!(found_key.objectid >= cache->start &&
1259                found_key.objectid + length <= cache->start + cache->length)) {
1260                 ret = -EUCLEAN;
1261                 goto out;
1262         }
1263         *offset_ret = found_key.objectid + length - cache->start;
1264         ret = 0;
1265
1266 out:
1267         btrfs_free_path(path);
1268         return ret;
1269 }
1270
1271 struct zone_info {
1272         u64 physical;
1273         u64 capacity;
1274         u64 alloc_offset;
1275 };
1276
1277 static int btrfs_load_zone_info(struct btrfs_fs_info *fs_info, int zone_idx,
1278                                 struct zone_info *info, unsigned long *active,
1279                                 struct btrfs_chunk_map *map)
1280 {
1281         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1282         struct btrfs_device *device = map->stripes[zone_idx].dev;
1283         int dev_replace_is_ongoing = 0;
1284         unsigned int nofs_flag;
1285         struct blk_zone zone;
1286         int ret;
1287
1288         info->physical = map->stripes[zone_idx].physical;
1289
1290         if (!device->bdev) {
1291                 info->alloc_offset = WP_MISSING_DEV;
1292                 return 0;
1293         }
1294
1295         /* Consider a zone as active if we can allow any number of active zones. */
1296         if (!device->zone_info->max_active_zones)
1297                 __set_bit(zone_idx, active);
1298
1299         if (!btrfs_dev_is_sequential(device, info->physical)) {
1300                 info->alloc_offset = WP_CONVENTIONAL;
1301                 return 0;
1302         }
1303
1304         /* This zone will be used for allocation, so mark this zone non-empty. */
1305         btrfs_dev_clear_zone_empty(device, info->physical);
1306
1307         down_read(&dev_replace->rwsem);
1308         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
1309         if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL)
1310                 btrfs_dev_clear_zone_empty(dev_replace->tgtdev, info->physical);
1311         up_read(&dev_replace->rwsem);
1312
1313         /*
1314          * The group is mapped to a sequential zone. Get the zone write pointer
1315          * to determine the allocation offset within the zone.
1316          */
1317         WARN_ON(!IS_ALIGNED(info->physical, fs_info->zone_size));
1318         nofs_flag = memalloc_nofs_save();
1319         ret = btrfs_get_dev_zone(device, info->physical, &zone);
1320         memalloc_nofs_restore(nofs_flag);
1321         if (ret) {
1322                 if (ret != -EIO && ret != -EOPNOTSUPP)
1323                         return ret;
1324                 info->alloc_offset = WP_MISSING_DEV;
1325                 return 0;
1326         }
1327
1328         if (zone.type == BLK_ZONE_TYPE_CONVENTIONAL) {
1329                 btrfs_err_in_rcu(fs_info,
1330                 "zoned: unexpected conventional zone %llu on device %s (devid %llu)",
1331                         zone.start << SECTOR_SHIFT, rcu_str_deref(device->name),
1332                         device->devid);
1333                 return -EIO;
1334         }
1335
1336         info->capacity = (zone.capacity << SECTOR_SHIFT);
1337
1338         switch (zone.cond) {
1339         case BLK_ZONE_COND_OFFLINE:
1340         case BLK_ZONE_COND_READONLY:
1341                 btrfs_err(fs_info,
1342                 "zoned: offline/readonly zone %llu on device %s (devid %llu)",
1343                           (info->physical >> device->zone_info->zone_size_shift),
1344                           rcu_str_deref(device->name), device->devid);
1345                 info->alloc_offset = WP_MISSING_DEV;
1346                 break;
1347         case BLK_ZONE_COND_EMPTY:
1348                 info->alloc_offset = 0;
1349                 break;
1350         case BLK_ZONE_COND_FULL:
1351                 info->alloc_offset = info->capacity;
1352                 break;
1353         default:
1354                 /* Partially used zone. */
1355                 info->alloc_offset = ((zone.wp - zone.start) << SECTOR_SHIFT);
1356                 __set_bit(zone_idx, active);
1357                 break;
1358         }
1359
1360         return 0;
1361 }
1362
1363 static int btrfs_load_block_group_single(struct btrfs_block_group *bg,
1364                                          struct zone_info *info,
1365                                          unsigned long *active)
1366 {
1367         if (info->alloc_offset == WP_MISSING_DEV) {
1368                 btrfs_err(bg->fs_info,
1369                         "zoned: cannot recover write pointer for zone %llu",
1370                         info->physical);
1371                 return -EIO;
1372         }
1373
1374         bg->alloc_offset = info->alloc_offset;
1375         bg->zone_capacity = info->capacity;
1376         if (test_bit(0, active))
1377                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1378         return 0;
1379 }
1380
1381 static int btrfs_load_block_group_dup(struct btrfs_block_group *bg,
1382                                       struct btrfs_chunk_map *map,
1383                                       struct zone_info *zone_info,
1384                                       unsigned long *active)
1385 {
1386         struct btrfs_fs_info *fs_info = bg->fs_info;
1387
1388         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1389                 btrfs_err(fs_info, "zoned: data DUP profile needs raid-stripe-tree");
1390                 return -EINVAL;
1391         }
1392
1393         if (zone_info[0].alloc_offset == WP_MISSING_DEV) {
1394                 btrfs_err(bg->fs_info,
1395                           "zoned: cannot recover write pointer for zone %llu",
1396                           zone_info[0].physical);
1397                 return -EIO;
1398         }
1399         if (zone_info[1].alloc_offset == WP_MISSING_DEV) {
1400                 btrfs_err(bg->fs_info,
1401                           "zoned: cannot recover write pointer for zone %llu",
1402                           zone_info[1].physical);
1403                 return -EIO;
1404         }
1405         if (zone_info[0].alloc_offset != zone_info[1].alloc_offset) {
1406                 btrfs_err(bg->fs_info,
1407                           "zoned: write pointer offset mismatch of zones in DUP profile");
1408                 return -EIO;
1409         }
1410
1411         if (test_bit(0, active) != test_bit(1, active)) {
1412                 if (!btrfs_zone_activate(bg))
1413                         return -EIO;
1414         } else if (test_bit(0, active)) {
1415                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1416         }
1417
1418         bg->alloc_offset = zone_info[0].alloc_offset;
1419         bg->zone_capacity = min(zone_info[0].capacity, zone_info[1].capacity);
1420         return 0;
1421 }
1422
1423 static int btrfs_load_block_group_raid1(struct btrfs_block_group *bg,
1424                                         struct btrfs_chunk_map *map,
1425                                         struct zone_info *zone_info,
1426                                         unsigned long *active)
1427 {
1428         struct btrfs_fs_info *fs_info = bg->fs_info;
1429         int i;
1430
1431         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1432                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1433                           btrfs_bg_type_to_raid_name(map->type));
1434                 return -EINVAL;
1435         }
1436
1437         for (i = 0; i < map->num_stripes; i++) {
1438                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1439                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1440                         continue;
1441
1442                 if ((zone_info[0].alloc_offset != zone_info[i].alloc_offset) &&
1443                     !btrfs_test_opt(fs_info, DEGRADED)) {
1444                         btrfs_err(fs_info,
1445                         "zoned: write pointer offset mismatch of zones in %s profile",
1446                                   btrfs_bg_type_to_raid_name(map->type));
1447                         return -EIO;
1448                 }
1449                 if (test_bit(0, active) != test_bit(i, active)) {
1450                         if (!btrfs_test_opt(fs_info, DEGRADED) &&
1451                             !btrfs_zone_activate(bg)) {
1452                                 return -EIO;
1453                         }
1454                 } else {
1455                         if (test_bit(0, active))
1456                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1457                 }
1458                 /* In case a device is missing we have a cap of 0, so don't use it. */
1459                 bg->zone_capacity = min_not_zero(zone_info[0].capacity,
1460                                                  zone_info[1].capacity);
1461         }
1462
1463         if (zone_info[0].alloc_offset != WP_MISSING_DEV)
1464                 bg->alloc_offset = zone_info[0].alloc_offset;
1465         else
1466                 bg->alloc_offset = zone_info[i - 1].alloc_offset;
1467
1468         return 0;
1469 }
1470
1471 static int btrfs_load_block_group_raid0(struct btrfs_block_group *bg,
1472                                         struct btrfs_chunk_map *map,
1473                                         struct zone_info *zone_info,
1474                                         unsigned long *active)
1475 {
1476         struct btrfs_fs_info *fs_info = bg->fs_info;
1477
1478         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1479                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1480                           btrfs_bg_type_to_raid_name(map->type));
1481                 return -EINVAL;
1482         }
1483
1484         for (int i = 0; i < map->num_stripes; i++) {
1485                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1486                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1487                         continue;
1488
1489                 if (test_bit(0, active) != test_bit(i, active)) {
1490                         if (!btrfs_zone_activate(bg))
1491                                 return -EIO;
1492                 } else {
1493                         if (test_bit(0, active))
1494                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1495                 }
1496                 bg->zone_capacity += zone_info[i].capacity;
1497                 bg->alloc_offset += zone_info[i].alloc_offset;
1498         }
1499
1500         return 0;
1501 }
1502
1503 static int btrfs_load_block_group_raid10(struct btrfs_block_group *bg,
1504                                          struct btrfs_chunk_map *map,
1505                                          struct zone_info *zone_info,
1506                                          unsigned long *active)
1507 {
1508         struct btrfs_fs_info *fs_info = bg->fs_info;
1509
1510         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1511                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1512                           btrfs_bg_type_to_raid_name(map->type));
1513                 return -EINVAL;
1514         }
1515
1516         for (int i = 0; i < map->num_stripes; i++) {
1517                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1518                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1519                         continue;
1520
1521                 if (test_bit(0, active) != test_bit(i, active)) {
1522                         if (!btrfs_zone_activate(bg))
1523                                 return -EIO;
1524                 } else {
1525                         if (test_bit(0, active))
1526                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1527                 }
1528
1529                 if ((i % map->sub_stripes) == 0) {
1530                         bg->zone_capacity += zone_info[i].capacity;
1531                         bg->alloc_offset += zone_info[i].alloc_offset;
1532                 }
1533         }
1534
1535         return 0;
1536 }
1537
1538 int btrfs_load_block_group_zone_info(struct btrfs_block_group *cache, bool new)
1539 {
1540         struct btrfs_fs_info *fs_info = cache->fs_info;
1541         struct btrfs_chunk_map *map;
1542         u64 logical = cache->start;
1543         u64 length = cache->length;
1544         struct zone_info *zone_info = NULL;
1545         int ret;
1546         int i;
1547         unsigned long *active = NULL;
1548         u64 last_alloc = 0;
1549         u32 num_sequential = 0, num_conventional = 0;
1550
1551         if (!btrfs_is_zoned(fs_info))
1552                 return 0;
1553
1554         /* Sanity check */
1555         if (!IS_ALIGNED(length, fs_info->zone_size)) {
1556                 btrfs_err(fs_info,
1557                 "zoned: block group %llu len %llu unaligned to zone size %llu",
1558                           logical, length, fs_info->zone_size);
1559                 return -EIO;
1560         }
1561
1562         map = btrfs_find_chunk_map(fs_info, logical, length);
1563         if (!map)
1564                 return -EINVAL;
1565
1566         cache->physical_map = map;
1567
1568         zone_info = kcalloc(map->num_stripes, sizeof(*zone_info), GFP_NOFS);
1569         if (!zone_info) {
1570                 ret = -ENOMEM;
1571                 goto out;
1572         }
1573
1574         active = bitmap_zalloc(map->num_stripes, GFP_NOFS);
1575         if (!active) {
1576                 ret = -ENOMEM;
1577                 goto out;
1578         }
1579
1580         for (i = 0; i < map->num_stripes; i++) {
1581                 ret = btrfs_load_zone_info(fs_info, i, &zone_info[i], active, map);
1582                 if (ret)
1583                         goto out;
1584
1585                 if (zone_info[i].alloc_offset == WP_CONVENTIONAL)
1586                         num_conventional++;
1587                 else
1588                         num_sequential++;
1589         }
1590
1591         if (num_sequential > 0)
1592                 set_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1593
1594         if (num_conventional > 0) {
1595                 /* Zone capacity is always zone size in emulation */
1596                 cache->zone_capacity = cache->length;
1597                 ret = calculate_alloc_pointer(cache, &last_alloc, new);
1598                 if (ret) {
1599                         btrfs_err(fs_info,
1600                         "zoned: failed to determine allocation offset of bg %llu",
1601                                   cache->start);
1602                         goto out;
1603                 } else if (map->num_stripes == num_conventional) {
1604                         cache->alloc_offset = last_alloc;
1605                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1606                         goto out;
1607                 }
1608         }
1609
1610         switch (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
1611         case 0: /* single */
1612                 ret = btrfs_load_block_group_single(cache, &zone_info[0], active);
1613                 break;
1614         case BTRFS_BLOCK_GROUP_DUP:
1615                 ret = btrfs_load_block_group_dup(cache, map, zone_info, active);
1616                 break;
1617         case BTRFS_BLOCK_GROUP_RAID1:
1618         case BTRFS_BLOCK_GROUP_RAID1C3:
1619         case BTRFS_BLOCK_GROUP_RAID1C4:
1620                 ret = btrfs_load_block_group_raid1(cache, map, zone_info, active);
1621                 break;
1622         case BTRFS_BLOCK_GROUP_RAID0:
1623                 ret = btrfs_load_block_group_raid0(cache, map, zone_info, active);
1624                 break;
1625         case BTRFS_BLOCK_GROUP_RAID10:
1626                 ret = btrfs_load_block_group_raid10(cache, map, zone_info, active);
1627                 break;
1628         case BTRFS_BLOCK_GROUP_RAID5:
1629         case BTRFS_BLOCK_GROUP_RAID6:
1630         default:
1631                 btrfs_err(fs_info, "zoned: profile %s not yet supported",
1632                           btrfs_bg_type_to_raid_name(map->type));
1633                 ret = -EINVAL;
1634                 goto out;
1635         }
1636
1637 out:
1638         /* Reject non SINGLE data profiles without RST */
1639         if ((map->type & BTRFS_BLOCK_GROUP_DATA) &&
1640             (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) &&
1641             !fs_info->stripe_root) {
1642                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1643                           btrfs_bg_type_to_raid_name(map->type));
1644                 return -EINVAL;
1645         }
1646
1647         if (cache->alloc_offset > cache->zone_capacity) {
1648                 btrfs_err(fs_info,
1649 "zoned: invalid write pointer %llu (larger than zone capacity %llu) in block group %llu",
1650                           cache->alloc_offset, cache->zone_capacity,
1651                           cache->start);
1652                 ret = -EIO;
1653         }
1654
1655         /* An extent is allocated after the write pointer */
1656         if (!ret && num_conventional && last_alloc > cache->alloc_offset) {
1657                 btrfs_err(fs_info,
1658                           "zoned: got wrong write pointer in BG %llu: %llu > %llu",
1659                           logical, last_alloc, cache->alloc_offset);
1660                 ret = -EIO;
1661         }
1662
1663         if (!ret) {
1664                 cache->meta_write_pointer = cache->alloc_offset + cache->start;
1665                 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags)) {
1666                         btrfs_get_block_group(cache);
1667                         spin_lock(&fs_info->zone_active_bgs_lock);
1668                         list_add_tail(&cache->active_bg_list,
1669                                       &fs_info->zone_active_bgs);
1670                         spin_unlock(&fs_info->zone_active_bgs_lock);
1671                 }
1672         } else {
1673                 btrfs_free_chunk_map(cache->physical_map);
1674                 cache->physical_map = NULL;
1675         }
1676         bitmap_free(active);
1677         kfree(zone_info);
1678
1679         return ret;
1680 }
1681
1682 void btrfs_calc_zone_unusable(struct btrfs_block_group *cache)
1683 {
1684         u64 unusable, free;
1685
1686         if (!btrfs_is_zoned(cache->fs_info))
1687                 return;
1688
1689         WARN_ON(cache->bytes_super != 0);
1690         unusable = (cache->alloc_offset - cache->used) +
1691                    (cache->length - cache->zone_capacity);
1692         free = cache->zone_capacity - cache->alloc_offset;
1693
1694         /* We only need ->free_space in ALLOC_SEQ block groups */
1695         cache->cached = BTRFS_CACHE_FINISHED;
1696         cache->free_space_ctl->free_space = free;
1697         cache->zone_unusable = unusable;
1698 }
1699
1700 bool btrfs_use_zone_append(struct btrfs_bio *bbio)
1701 {
1702         u64 start = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT);
1703         struct btrfs_inode *inode = bbio->inode;
1704         struct btrfs_fs_info *fs_info = bbio->fs_info;
1705         struct btrfs_block_group *cache;
1706         bool ret = false;
1707
1708         if (!btrfs_is_zoned(fs_info))
1709                 return false;
1710
1711         if (!inode || !is_data_inode(&inode->vfs_inode))
1712                 return false;
1713
1714         if (btrfs_op(&bbio->bio) != BTRFS_MAP_WRITE)
1715                 return false;
1716
1717         /*
1718          * Using REQ_OP_ZONE_APPNED for relocation can break assumptions on the
1719          * extent layout the relocation code has.
1720          * Furthermore we have set aside own block-group from which only the
1721          * relocation "process" can allocate and make sure only one process at a
1722          * time can add pages to an extent that gets relocated, so it's safe to
1723          * use regular REQ_OP_WRITE for this special case.
1724          */
1725         if (btrfs_is_data_reloc_root(inode->root))
1726                 return false;
1727
1728         cache = btrfs_lookup_block_group(fs_info, start);
1729         ASSERT(cache);
1730         if (!cache)
1731                 return false;
1732
1733         ret = !!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1734         btrfs_put_block_group(cache);
1735
1736         return ret;
1737 }
1738
1739 void btrfs_record_physical_zoned(struct btrfs_bio *bbio)
1740 {
1741         const u64 physical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
1742         struct btrfs_ordered_sum *sum = bbio->sums;
1743
1744         if (physical < bbio->orig_physical)
1745                 sum->logical -= bbio->orig_physical - physical;
1746         else
1747                 sum->logical += physical - bbio->orig_physical;
1748 }
1749
1750 static void btrfs_rewrite_logical_zoned(struct btrfs_ordered_extent *ordered,
1751                                         u64 logical)
1752 {
1753         struct extent_map_tree *em_tree = &BTRFS_I(ordered->inode)->extent_tree;
1754         struct extent_map *em;
1755
1756         ordered->disk_bytenr = logical;
1757
1758         write_lock(&em_tree->lock);
1759         em = search_extent_mapping(em_tree, ordered->file_offset,
1760                                    ordered->num_bytes);
1761         em->block_start = logical;
1762         free_extent_map(em);
1763         write_unlock(&em_tree->lock);
1764 }
1765
1766 static bool btrfs_zoned_split_ordered(struct btrfs_ordered_extent *ordered,
1767                                       u64 logical, u64 len)
1768 {
1769         struct btrfs_ordered_extent *new;
1770
1771         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags) &&
1772             split_extent_map(BTRFS_I(ordered->inode), ordered->file_offset,
1773                              ordered->num_bytes, len, logical))
1774                 return false;
1775
1776         new = btrfs_split_ordered_extent(ordered, len);
1777         if (IS_ERR(new))
1778                 return false;
1779         new->disk_bytenr = logical;
1780         btrfs_finish_one_ordered(new);
1781         return true;
1782 }
1783
1784 void btrfs_finish_ordered_zoned(struct btrfs_ordered_extent *ordered)
1785 {
1786         struct btrfs_inode *inode = BTRFS_I(ordered->inode);
1787         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1788         struct btrfs_ordered_sum *sum;
1789         u64 logical, len;
1790
1791         /*
1792          * Write to pre-allocated region is for the data relocation, and so
1793          * it should use WRITE operation. No split/rewrite are necessary.
1794          */
1795         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags))
1796                 return;
1797
1798         ASSERT(!list_empty(&ordered->list));
1799         /* The ordered->list can be empty in the above pre-alloc case. */
1800         sum = list_first_entry(&ordered->list, struct btrfs_ordered_sum, list);
1801         logical = sum->logical;
1802         len = sum->len;
1803
1804         while (len < ordered->disk_num_bytes) {
1805                 sum = list_next_entry(sum, list);
1806                 if (sum->logical == logical + len) {
1807                         len += sum->len;
1808                         continue;
1809                 }
1810                 if (!btrfs_zoned_split_ordered(ordered, logical, len)) {
1811                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
1812                         btrfs_err(fs_info, "failed to split ordered extent");
1813                         goto out;
1814                 }
1815                 logical = sum->logical;
1816                 len = sum->len;
1817         }
1818
1819         if (ordered->disk_bytenr != logical)
1820                 btrfs_rewrite_logical_zoned(ordered, logical);
1821
1822 out:
1823         /*
1824          * If we end up here for nodatasum I/O, the btrfs_ordered_sum structures
1825          * were allocated by btrfs_alloc_dummy_sum only to record the logical
1826          * addresses and don't contain actual checksums.  We thus must free them
1827          * here so that we don't attempt to log the csums later.
1828          */
1829         if ((inode->flags & BTRFS_INODE_NODATASUM) ||
1830             test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state)) {
1831                 while ((sum = list_first_entry_or_null(&ordered->list,
1832                                                        typeof(*sum), list))) {
1833                         list_del(&sum->list);
1834                         kfree(sum);
1835                 }
1836         }
1837 }
1838
1839 static bool check_bg_is_active(struct btrfs_eb_write_context *ctx,
1840                                struct btrfs_block_group **active_bg)
1841 {
1842         const struct writeback_control *wbc = ctx->wbc;
1843         struct btrfs_block_group *block_group = ctx->zoned_bg;
1844         struct btrfs_fs_info *fs_info = block_group->fs_info;
1845
1846         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
1847                 return true;
1848
1849         if (fs_info->treelog_bg == block_group->start) {
1850                 if (!btrfs_zone_activate(block_group)) {
1851                         int ret_fin = btrfs_zone_finish_one_bg(fs_info);
1852
1853                         if (ret_fin != 1 || !btrfs_zone_activate(block_group))
1854                                 return false;
1855                 }
1856         } else if (*active_bg != block_group) {
1857                 struct btrfs_block_group *tgt = *active_bg;
1858
1859                 /* zoned_meta_io_lock protects fs_info->active_{meta,system}_bg. */
1860                 lockdep_assert_held(&fs_info->zoned_meta_io_lock);
1861
1862                 if (tgt) {
1863                         /*
1864                          * If there is an unsent IO left in the allocated area,
1865                          * we cannot wait for them as it may cause a deadlock.
1866                          */
1867                         if (tgt->meta_write_pointer < tgt->start + tgt->alloc_offset) {
1868                                 if (wbc->sync_mode == WB_SYNC_NONE ||
1869                                     (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync))
1870                                         return false;
1871                         }
1872
1873                         /* Pivot active metadata/system block group. */
1874                         btrfs_zoned_meta_io_unlock(fs_info);
1875                         wait_eb_writebacks(tgt);
1876                         do_zone_finish(tgt, true);
1877                         btrfs_zoned_meta_io_lock(fs_info);
1878                         if (*active_bg == tgt) {
1879                                 btrfs_put_block_group(tgt);
1880                                 *active_bg = NULL;
1881                         }
1882                 }
1883                 if (!btrfs_zone_activate(block_group))
1884                         return false;
1885                 if (*active_bg != block_group) {
1886                         ASSERT(*active_bg == NULL);
1887                         *active_bg = block_group;
1888                         btrfs_get_block_group(block_group);
1889                 }
1890         }
1891
1892         return true;
1893 }
1894
1895 /*
1896  * Check if @ctx->eb is aligned to the write pointer.
1897  *
1898  * Return:
1899  *   0:        @ctx->eb is at the write pointer. You can write it.
1900  *   -EAGAIN:  There is a hole. The caller should handle the case.
1901  *   -EBUSY:   There is a hole, but the caller can just bail out.
1902  */
1903 int btrfs_check_meta_write_pointer(struct btrfs_fs_info *fs_info,
1904                                    struct btrfs_eb_write_context *ctx)
1905 {
1906         const struct writeback_control *wbc = ctx->wbc;
1907         const struct extent_buffer *eb = ctx->eb;
1908         struct btrfs_block_group *block_group = ctx->zoned_bg;
1909
1910         if (!btrfs_is_zoned(fs_info))
1911                 return 0;
1912
1913         if (block_group) {
1914                 if (block_group->start > eb->start ||
1915                     block_group->start + block_group->length <= eb->start) {
1916                         btrfs_put_block_group(block_group);
1917                         block_group = NULL;
1918                         ctx->zoned_bg = NULL;
1919                 }
1920         }
1921
1922         if (!block_group) {
1923                 block_group = btrfs_lookup_block_group(fs_info, eb->start);
1924                 if (!block_group)
1925                         return 0;
1926                 ctx->zoned_bg = block_group;
1927         }
1928
1929         if (block_group->meta_write_pointer == eb->start) {
1930                 struct btrfs_block_group **tgt;
1931
1932                 if (!test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags))
1933                         return 0;
1934
1935                 if (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM)
1936                         tgt = &fs_info->active_system_bg;
1937                 else
1938                         tgt = &fs_info->active_meta_bg;
1939                 if (check_bg_is_active(ctx, tgt))
1940                         return 0;
1941         }
1942
1943         /*
1944          * Since we may release fs_info->zoned_meta_io_lock, someone can already
1945          * start writing this eb. In that case, we can just bail out.
1946          */
1947         if (block_group->meta_write_pointer > eb->start)
1948                 return -EBUSY;
1949
1950         /* If for_sync, this hole will be filled with trasnsaction commit. */
1951         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
1952                 return -EAGAIN;
1953         return -EBUSY;
1954 }
1955
1956 int btrfs_zoned_issue_zeroout(struct btrfs_device *device, u64 physical, u64 length)
1957 {
1958         if (!btrfs_dev_is_sequential(device, physical))
1959                 return -EOPNOTSUPP;
1960
1961         return blkdev_issue_zeroout(device->bdev, physical >> SECTOR_SHIFT,
1962                                     length >> SECTOR_SHIFT, GFP_NOFS, 0);
1963 }
1964
1965 static int read_zone_info(struct btrfs_fs_info *fs_info, u64 logical,
1966                           struct blk_zone *zone)
1967 {
1968         struct btrfs_io_context *bioc = NULL;
1969         u64 mapped_length = PAGE_SIZE;
1970         unsigned int nofs_flag;
1971         int nmirrors;
1972         int i, ret;
1973
1974         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
1975                               &mapped_length, &bioc, NULL, NULL);
1976         if (ret || !bioc || mapped_length < PAGE_SIZE) {
1977                 ret = -EIO;
1978                 goto out_put_bioc;
1979         }
1980
1981         if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1982                 ret = -EINVAL;
1983                 goto out_put_bioc;
1984         }
1985
1986         nofs_flag = memalloc_nofs_save();
1987         nmirrors = (int)bioc->num_stripes;
1988         for (i = 0; i < nmirrors; i++) {
1989                 u64 physical = bioc->stripes[i].physical;
1990                 struct btrfs_device *dev = bioc->stripes[i].dev;
1991
1992                 /* Missing device */
1993                 if (!dev->bdev)
1994                         continue;
1995
1996                 ret = btrfs_get_dev_zone(dev, physical, zone);
1997                 /* Failing device */
1998                 if (ret == -EIO || ret == -EOPNOTSUPP)
1999                         continue;
2000                 break;
2001         }
2002         memalloc_nofs_restore(nofs_flag);
2003 out_put_bioc:
2004         btrfs_put_bioc(bioc);
2005         return ret;
2006 }
2007
2008 /*
2009  * Synchronize write pointer in a zone at @physical_start on @tgt_dev, by
2010  * filling zeros between @physical_pos to a write pointer of dev-replace
2011  * source device.
2012  */
2013 int btrfs_sync_zone_write_pointer(struct btrfs_device *tgt_dev, u64 logical,
2014                                     u64 physical_start, u64 physical_pos)
2015 {
2016         struct btrfs_fs_info *fs_info = tgt_dev->fs_info;
2017         struct blk_zone zone;
2018         u64 length;
2019         u64 wp;
2020         int ret;
2021
2022         if (!btrfs_dev_is_sequential(tgt_dev, physical_pos))
2023                 return 0;
2024
2025         ret = read_zone_info(fs_info, logical, &zone);
2026         if (ret)
2027                 return ret;
2028
2029         wp = physical_start + ((zone.wp - zone.start) << SECTOR_SHIFT);
2030
2031         if (physical_pos == wp)
2032                 return 0;
2033
2034         if (physical_pos > wp)
2035                 return -EUCLEAN;
2036
2037         length = wp - physical_pos;
2038         return btrfs_zoned_issue_zeroout(tgt_dev, physical_pos, length);
2039 }
2040
2041 /*
2042  * Activate block group and underlying device zones
2043  *
2044  * @block_group: the block group to activate
2045  *
2046  * Return: true on success, false otherwise
2047  */
2048 bool btrfs_zone_activate(struct btrfs_block_group *block_group)
2049 {
2050         struct btrfs_fs_info *fs_info = block_group->fs_info;
2051         struct btrfs_chunk_map *map;
2052         struct btrfs_device *device;
2053         u64 physical;
2054         const bool is_data = (block_group->flags & BTRFS_BLOCK_GROUP_DATA);
2055         bool ret;
2056         int i;
2057
2058         if (!btrfs_is_zoned(block_group->fs_info))
2059                 return true;
2060
2061         map = block_group->physical_map;
2062
2063         spin_lock(&fs_info->zone_active_bgs_lock);
2064         spin_lock(&block_group->lock);
2065         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
2066                 ret = true;
2067                 goto out_unlock;
2068         }
2069
2070         /* No space left */
2071         if (btrfs_zoned_bg_is_full(block_group)) {
2072                 ret = false;
2073                 goto out_unlock;
2074         }
2075
2076         for (i = 0; i < map->num_stripes; i++) {
2077                 struct btrfs_zoned_device_info *zinfo;
2078                 int reserved = 0;
2079
2080                 device = map->stripes[i].dev;
2081                 physical = map->stripes[i].physical;
2082                 zinfo = device->zone_info;
2083
2084                 if (zinfo->max_active_zones == 0)
2085                         continue;
2086
2087                 if (is_data)
2088                         reserved = zinfo->reserved_active_zones;
2089                 /*
2090                  * For the data block group, leave active zones for one
2091                  * metadata block group and one system block group.
2092                  */
2093                 if (atomic_read(&zinfo->active_zones_left) <= reserved) {
2094                         ret = false;
2095                         goto out_unlock;
2096                 }
2097
2098                 if (!btrfs_dev_set_active_zone(device, physical)) {
2099                         /* Cannot activate the zone */
2100                         ret = false;
2101                         goto out_unlock;
2102                 }
2103                 if (!is_data)
2104                         zinfo->reserved_active_zones--;
2105         }
2106
2107         /* Successfully activated all the zones */
2108         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2109         spin_unlock(&block_group->lock);
2110
2111         /* For the active block group list */
2112         btrfs_get_block_group(block_group);
2113         list_add_tail(&block_group->active_bg_list, &fs_info->zone_active_bgs);
2114         spin_unlock(&fs_info->zone_active_bgs_lock);
2115
2116         return true;
2117
2118 out_unlock:
2119         spin_unlock(&block_group->lock);
2120         spin_unlock(&fs_info->zone_active_bgs_lock);
2121         return ret;
2122 }
2123
2124 static void wait_eb_writebacks(struct btrfs_block_group *block_group)
2125 {
2126         struct btrfs_fs_info *fs_info = block_group->fs_info;
2127         const u64 end = block_group->start + block_group->length;
2128         struct radix_tree_iter iter;
2129         struct extent_buffer *eb;
2130         void __rcu **slot;
2131
2132         rcu_read_lock();
2133         radix_tree_for_each_slot(slot, &fs_info->buffer_radix, &iter,
2134                                  block_group->start >> fs_info->sectorsize_bits) {
2135                 eb = radix_tree_deref_slot(slot);
2136                 if (!eb)
2137                         continue;
2138                 if (radix_tree_deref_retry(eb)) {
2139                         slot = radix_tree_iter_retry(&iter);
2140                         continue;
2141                 }
2142
2143                 if (eb->start < block_group->start)
2144                         continue;
2145                 if (eb->start >= end)
2146                         break;
2147
2148                 slot = radix_tree_iter_resume(slot, &iter);
2149                 rcu_read_unlock();
2150                 wait_on_extent_buffer_writeback(eb);
2151                 rcu_read_lock();
2152         }
2153         rcu_read_unlock();
2154 }
2155
2156 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written)
2157 {
2158         struct btrfs_fs_info *fs_info = block_group->fs_info;
2159         struct btrfs_chunk_map *map;
2160         const bool is_metadata = (block_group->flags &
2161                         (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM));
2162         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2163         int ret = 0;
2164         int i;
2165
2166         spin_lock(&block_group->lock);
2167         if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
2168                 spin_unlock(&block_group->lock);
2169                 return 0;
2170         }
2171
2172         /* Check if we have unwritten allocated space */
2173         if (is_metadata &&
2174             block_group->start + block_group->alloc_offset > block_group->meta_write_pointer) {
2175                 spin_unlock(&block_group->lock);
2176                 return -EAGAIN;
2177         }
2178
2179         /*
2180          * If we are sure that the block group is full (= no more room left for
2181          * new allocation) and the IO for the last usable block is completed, we
2182          * don't need to wait for the other IOs. This holds because we ensure
2183          * the sequential IO submissions using the ZONE_APPEND command for data
2184          * and block_group->meta_write_pointer for metadata.
2185          */
2186         if (!fully_written) {
2187                 if (test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
2188                         spin_unlock(&block_group->lock);
2189                         return -EAGAIN;
2190                 }
2191                 spin_unlock(&block_group->lock);
2192
2193                 ret = btrfs_inc_block_group_ro(block_group, false);
2194                 if (ret)
2195                         return ret;
2196
2197                 /* Ensure all writes in this block group finish */
2198                 btrfs_wait_block_group_reservations(block_group);
2199                 /* No need to wait for NOCOW writers. Zoned mode does not allow that */
2200                 btrfs_wait_ordered_roots(fs_info, U64_MAX, block_group->start,
2201                                          block_group->length);
2202                 /* Wait for extent buffers to be written. */
2203                 if (is_metadata)
2204                         wait_eb_writebacks(block_group);
2205
2206                 spin_lock(&block_group->lock);
2207
2208                 /*
2209                  * Bail out if someone already deactivated the block group, or
2210                  * allocated space is left in the block group.
2211                  */
2212                 if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2213                               &block_group->runtime_flags)) {
2214                         spin_unlock(&block_group->lock);
2215                         btrfs_dec_block_group_ro(block_group);
2216                         return 0;
2217                 }
2218
2219                 if (block_group->reserved ||
2220                     test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2221                              &block_group->runtime_flags)) {
2222                         spin_unlock(&block_group->lock);
2223                         btrfs_dec_block_group_ro(block_group);
2224                         return -EAGAIN;
2225                 }
2226         }
2227
2228         clear_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2229         block_group->alloc_offset = block_group->zone_capacity;
2230         if (block_group->flags & (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM))
2231                 block_group->meta_write_pointer = block_group->start +
2232                                                   block_group->zone_capacity;
2233         block_group->free_space_ctl->free_space = 0;
2234         btrfs_clear_treelog_bg(block_group);
2235         btrfs_clear_data_reloc_bg(block_group);
2236         spin_unlock(&block_group->lock);
2237
2238         down_read(&dev_replace->rwsem);
2239         map = block_group->physical_map;
2240         for (i = 0; i < map->num_stripes; i++) {
2241                 struct btrfs_device *device = map->stripes[i].dev;
2242                 const u64 physical = map->stripes[i].physical;
2243                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2244
2245                 if (zinfo->max_active_zones == 0)
2246                         continue;
2247
2248                 ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_FINISH,
2249                                        physical >> SECTOR_SHIFT,
2250                                        zinfo->zone_size >> SECTOR_SHIFT,
2251                                        GFP_NOFS);
2252
2253                 if (ret) {
2254                         up_read(&dev_replace->rwsem);
2255                         return ret;
2256                 }
2257
2258                 if (!(block_group->flags & BTRFS_BLOCK_GROUP_DATA))
2259                         zinfo->reserved_active_zones++;
2260                 btrfs_dev_clear_active_zone(device, physical);
2261         }
2262         up_read(&dev_replace->rwsem);
2263
2264         if (!fully_written)
2265                 btrfs_dec_block_group_ro(block_group);
2266
2267         spin_lock(&fs_info->zone_active_bgs_lock);
2268         ASSERT(!list_empty(&block_group->active_bg_list));
2269         list_del_init(&block_group->active_bg_list);
2270         spin_unlock(&fs_info->zone_active_bgs_lock);
2271
2272         /* For active_bg_list */
2273         btrfs_put_block_group(block_group);
2274
2275         clear_and_wake_up_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2276
2277         return 0;
2278 }
2279
2280 int btrfs_zone_finish(struct btrfs_block_group *block_group)
2281 {
2282         if (!btrfs_is_zoned(block_group->fs_info))
2283                 return 0;
2284
2285         return do_zone_finish(block_group, false);
2286 }
2287
2288 bool btrfs_can_activate_zone(struct btrfs_fs_devices *fs_devices, u64 flags)
2289 {
2290         struct btrfs_fs_info *fs_info = fs_devices->fs_info;
2291         struct btrfs_device *device;
2292         bool ret = false;
2293
2294         if (!btrfs_is_zoned(fs_info))
2295                 return true;
2296
2297         /* Check if there is a device with active zones left */
2298         mutex_lock(&fs_info->chunk_mutex);
2299         spin_lock(&fs_info->zone_active_bgs_lock);
2300         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
2301                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2302                 int reserved = 0;
2303
2304                 if (!device->bdev)
2305                         continue;
2306
2307                 if (!zinfo->max_active_zones) {
2308                         ret = true;
2309                         break;
2310                 }
2311
2312                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2313                         reserved = zinfo->reserved_active_zones;
2314
2315                 switch (flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
2316                 case 0: /* single */
2317                         ret = (atomic_read(&zinfo->active_zones_left) >= (1 + reserved));
2318                         break;
2319                 case BTRFS_BLOCK_GROUP_DUP:
2320                         ret = (atomic_read(&zinfo->active_zones_left) >= (2 + reserved));
2321                         break;
2322                 }
2323                 if (ret)
2324                         break;
2325         }
2326         spin_unlock(&fs_info->zone_active_bgs_lock);
2327         mutex_unlock(&fs_info->chunk_mutex);
2328
2329         if (!ret)
2330                 set_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2331
2332         return ret;
2333 }
2334
2335 void btrfs_zone_finish_endio(struct btrfs_fs_info *fs_info, u64 logical, u64 length)
2336 {
2337         struct btrfs_block_group *block_group;
2338         u64 min_alloc_bytes;
2339
2340         if (!btrfs_is_zoned(fs_info))
2341                 return;
2342
2343         block_group = btrfs_lookup_block_group(fs_info, logical);
2344         ASSERT(block_group);
2345
2346         /* No MIXED_BG on zoned btrfs. */
2347         if (block_group->flags & BTRFS_BLOCK_GROUP_DATA)
2348                 min_alloc_bytes = fs_info->sectorsize;
2349         else
2350                 min_alloc_bytes = fs_info->nodesize;
2351
2352         /* Bail out if we can allocate more data from this block group. */
2353         if (logical + length + min_alloc_bytes <=
2354             block_group->start + block_group->zone_capacity)
2355                 goto out;
2356
2357         do_zone_finish(block_group, true);
2358
2359 out:
2360         btrfs_put_block_group(block_group);
2361 }
2362
2363 static void btrfs_zone_finish_endio_workfn(struct work_struct *work)
2364 {
2365         struct btrfs_block_group *bg =
2366                 container_of(work, struct btrfs_block_group, zone_finish_work);
2367
2368         wait_on_extent_buffer_writeback(bg->last_eb);
2369         free_extent_buffer(bg->last_eb);
2370         btrfs_zone_finish_endio(bg->fs_info, bg->start, bg->length);
2371         btrfs_put_block_group(bg);
2372 }
2373
2374 void btrfs_schedule_zone_finish_bg(struct btrfs_block_group *bg,
2375                                    struct extent_buffer *eb)
2376 {
2377         if (!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &bg->runtime_flags) ||
2378             eb->start + eb->len * 2 <= bg->start + bg->zone_capacity)
2379                 return;
2380
2381         if (WARN_ON(bg->zone_finish_work.func == btrfs_zone_finish_endio_workfn)) {
2382                 btrfs_err(bg->fs_info, "double scheduling of bg %llu zone finishing",
2383                           bg->start);
2384                 return;
2385         }
2386
2387         /* For the work */
2388         btrfs_get_block_group(bg);
2389         atomic_inc(&eb->refs);
2390         bg->last_eb = eb;
2391         INIT_WORK(&bg->zone_finish_work, btrfs_zone_finish_endio_workfn);
2392         queue_work(system_unbound_wq, &bg->zone_finish_work);
2393 }
2394
2395 void btrfs_clear_data_reloc_bg(struct btrfs_block_group *bg)
2396 {
2397         struct btrfs_fs_info *fs_info = bg->fs_info;
2398
2399         spin_lock(&fs_info->relocation_bg_lock);
2400         if (fs_info->data_reloc_bg == bg->start)
2401                 fs_info->data_reloc_bg = 0;
2402         spin_unlock(&fs_info->relocation_bg_lock);
2403 }
2404
2405 void btrfs_free_zone_cache(struct btrfs_fs_info *fs_info)
2406 {
2407         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2408         struct btrfs_device *device;
2409
2410         if (!btrfs_is_zoned(fs_info))
2411                 return;
2412
2413         mutex_lock(&fs_devices->device_list_mutex);
2414         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2415                 if (device->zone_info) {
2416                         vfree(device->zone_info->zone_cache);
2417                         device->zone_info->zone_cache = NULL;
2418                 }
2419         }
2420         mutex_unlock(&fs_devices->device_list_mutex);
2421 }
2422
2423 bool btrfs_zoned_should_reclaim(struct btrfs_fs_info *fs_info)
2424 {
2425         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2426         struct btrfs_device *device;
2427         u64 used = 0;
2428         u64 total = 0;
2429         u64 factor;
2430
2431         ASSERT(btrfs_is_zoned(fs_info));
2432
2433         if (fs_info->bg_reclaim_threshold == 0)
2434                 return false;
2435
2436         mutex_lock(&fs_devices->device_list_mutex);
2437         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2438                 if (!device->bdev)
2439                         continue;
2440
2441                 total += device->disk_total_bytes;
2442                 used += device->bytes_used;
2443         }
2444         mutex_unlock(&fs_devices->device_list_mutex);
2445
2446         factor = div64_u64(used * 100, total);
2447         return factor >= fs_info->bg_reclaim_threshold;
2448 }
2449
2450 void btrfs_zoned_release_data_reloc_bg(struct btrfs_fs_info *fs_info, u64 logical,
2451                                        u64 length)
2452 {
2453         struct btrfs_block_group *block_group;
2454
2455         if (!btrfs_is_zoned(fs_info))
2456                 return;
2457
2458         block_group = btrfs_lookup_block_group(fs_info, logical);
2459         /* It should be called on a previous data relocation block group. */
2460         ASSERT(block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA));
2461
2462         spin_lock(&block_group->lock);
2463         if (!test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))
2464                 goto out;
2465
2466         /* All relocation extents are written. */
2467         if (block_group->start + block_group->alloc_offset == logical + length) {
2468                 /*
2469                  * Now, release this block group for further allocations and
2470                  * zone finish.
2471                  */
2472                 clear_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2473                           &block_group->runtime_flags);
2474         }
2475
2476 out:
2477         spin_unlock(&block_group->lock);
2478         btrfs_put_block_group(block_group);
2479 }
2480
2481 int btrfs_zone_finish_one_bg(struct btrfs_fs_info *fs_info)
2482 {
2483         struct btrfs_block_group *block_group;
2484         struct btrfs_block_group *min_bg = NULL;
2485         u64 min_avail = U64_MAX;
2486         int ret;
2487
2488         spin_lock(&fs_info->zone_active_bgs_lock);
2489         list_for_each_entry(block_group, &fs_info->zone_active_bgs,
2490                             active_bg_list) {
2491                 u64 avail;
2492
2493                 spin_lock(&block_group->lock);
2494                 if (block_group->reserved || block_group->alloc_offset == 0 ||
2495                     (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM) ||
2496                     test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
2497                         spin_unlock(&block_group->lock);
2498                         continue;
2499                 }
2500
2501                 avail = block_group->zone_capacity - block_group->alloc_offset;
2502                 if (min_avail > avail) {
2503                         if (min_bg)
2504                                 btrfs_put_block_group(min_bg);
2505                         min_bg = block_group;
2506                         min_avail = avail;
2507                         btrfs_get_block_group(min_bg);
2508                 }
2509                 spin_unlock(&block_group->lock);
2510         }
2511         spin_unlock(&fs_info->zone_active_bgs_lock);
2512
2513         if (!min_bg)
2514                 return 0;
2515
2516         ret = btrfs_zone_finish(min_bg);
2517         btrfs_put_block_group(min_bg);
2518
2519         return ret < 0 ? ret : 1;
2520 }
2521
2522 int btrfs_zoned_activate_one_bg(struct btrfs_fs_info *fs_info,
2523                                 struct btrfs_space_info *space_info,
2524                                 bool do_finish)
2525 {
2526         struct btrfs_block_group *bg;
2527         int index;
2528
2529         if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
2530                 return 0;
2531
2532         for (;;) {
2533                 int ret;
2534                 bool need_finish = false;
2535
2536                 down_read(&space_info->groups_sem);
2537                 for (index = 0; index < BTRFS_NR_RAID_TYPES; index++) {
2538                         list_for_each_entry(bg, &space_info->block_groups[index],
2539                                             list) {
2540                                 if (!spin_trylock(&bg->lock))
2541                                         continue;
2542                                 if (btrfs_zoned_bg_is_full(bg) ||
2543                                     test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2544                                              &bg->runtime_flags)) {
2545                                         spin_unlock(&bg->lock);
2546                                         continue;
2547                                 }
2548                                 spin_unlock(&bg->lock);
2549
2550                                 if (btrfs_zone_activate(bg)) {
2551                                         up_read(&space_info->groups_sem);
2552                                         return 1;
2553                                 }
2554
2555                                 need_finish = true;
2556                         }
2557                 }
2558                 up_read(&space_info->groups_sem);
2559
2560                 if (!do_finish || !need_finish)
2561                         break;
2562
2563                 ret = btrfs_zone_finish_one_bg(fs_info);
2564                 if (ret == 0)
2565                         break;
2566                 if (ret < 0)
2567                         return ret;
2568         }
2569
2570         return 0;
2571 }
2572
2573 /*
2574  * Reserve zones for one metadata block group, one tree-log block group, and one
2575  * system block group.
2576  */
2577 void btrfs_check_active_zone_reservation(struct btrfs_fs_info *fs_info)
2578 {
2579         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2580         struct btrfs_block_group *block_group;
2581         struct btrfs_device *device;
2582         /* Reserve zones for normal SINGLE metadata and tree-log block group. */
2583         unsigned int metadata_reserve = 2;
2584         /* Reserve a zone for SINGLE system block group. */
2585         unsigned int system_reserve = 1;
2586
2587         if (!test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags))
2588                 return;
2589
2590         /*
2591          * This function is called from the mount context. So, there is no
2592          * parallel process touching the bits. No need for read_seqretry().
2593          */
2594         if (fs_info->avail_metadata_alloc_bits & BTRFS_BLOCK_GROUP_DUP)
2595                 metadata_reserve = 4;
2596         if (fs_info->avail_system_alloc_bits & BTRFS_BLOCK_GROUP_DUP)
2597                 system_reserve = 2;
2598
2599         /* Apply the reservation on all the devices. */
2600         mutex_lock(&fs_devices->device_list_mutex);
2601         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2602                 if (!device->bdev)
2603                         continue;
2604
2605                 device->zone_info->reserved_active_zones =
2606                         metadata_reserve + system_reserve;
2607         }
2608         mutex_unlock(&fs_devices->device_list_mutex);
2609
2610         /* Release reservation for currently active block groups. */
2611         spin_lock(&fs_info->zone_active_bgs_lock);
2612         list_for_each_entry(block_group, &fs_info->zone_active_bgs, active_bg_list) {
2613                 struct btrfs_chunk_map *map = block_group->physical_map;
2614
2615                 if (!(block_group->flags &
2616                       (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM)))
2617                         continue;
2618
2619                 for (int i = 0; i < map->num_stripes; i++)
2620                         map->stripes[i].dev->zone_info->reserved_active_zones--;
2621         }
2622         spin_unlock(&fs_info->zone_active_bgs_lock);
2623 }