fs/btrfs/zoned.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/bitops.h>
   4 #include <linux/slab.h>
   5 #include <linux/blkdev.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/atomic.h>
   8 #include <linux/vmalloc.h>
   9 #include "ctree.h"
  10 #include "volumes.h"
  11 #include "zoned.h"
  12 #include "rcu-string.h"
  13 #include "disk-io.h"
  14 #include "block-group.h"
  15 #include "transaction.h"
  16 #include "dev-replace.h"
  17 #include "space-info.h"
  18 #include "super.h"
  19 #include "fs.h"
  20 #include "accessors.h"
  21 #include "bio.h"
  22
  23 /* Maximum number of zones to report per blkdev_report_zones() call */
  24 #define BTRFS_REPORT_NR_ZONES   4096
  25 /* Invalid allocation pointer value for missing devices */
  26 #define WP_MISSING_DEV ((u64)-1)
  27 /* Pseudo write pointer value for conventional zone */
  28 #define WP_CONVENTIONAL ((u64)-2)
  29
  30 /*
  31  * Location of the first zone of superblock logging zone pairs.
  32  *
  33  * - primary superblock:    0B (zone 0)
  34  * - first copy:          512G (zone starting at that offset)
  35  * - second copy:           4T (zone starting at that offset)
  36  */
  37 #define BTRFS_SB_LOG_PRIMARY_OFFSET     (0ULL)
  38 #define BTRFS_SB_LOG_FIRST_OFFSET       (512ULL * SZ_1G)
  39 #define BTRFS_SB_LOG_SECOND_OFFSET      (4096ULL * SZ_1G)
  40
  41 #define BTRFS_SB_LOG_FIRST_SHIFT        const_ilog2(BTRFS_SB_LOG_FIRST_OFFSET)
  42 #define BTRFS_SB_LOG_SECOND_SHIFT       const_ilog2(BTRFS_SB_LOG_SECOND_OFFSET)
  43
  44 /* Number of superblock log zones */
  45 #define BTRFS_NR_SB_LOG_ZONES 2
  46
  47 /*
  48  * Minimum of active zones we need:
  49  *
  50  * - BTRFS_SUPER_MIRROR_MAX zones for superblock mirrors
  51  * - 3 zones to ensure at least one zone per SYSTEM, META and DATA block group
  52  * - 1 zone for tree-log dedicated block group
  53  * - 1 zone for relocation
  54  */
  55 #define BTRFS_MIN_ACTIVE_ZONES          (BTRFS_SUPER_MIRROR_MAX + 5)
  56
  57 /*
  58  * Minimum / maximum supported zone size. Currently, SMR disks have a zone
  59  * size of 256MiB, and we are expecting ZNS drives to be in the 1-4GiB range.
  60  * We do not expect the zone size to become larger than 8GiB or smaller than
  61  * 4MiB in the near future.
  62  */
  63 #define BTRFS_MAX_ZONE_SIZE             SZ_8G
  64 #define BTRFS_MIN_ZONE_SIZE             SZ_4M
  65
  66 #define SUPER_INFO_SECTORS      ((u64)BTRFS_SUPER_INFO_SIZE >> SECTOR_SHIFT)
  67
  68 static void wait_eb_writebacks(struct btrfs_block_group *block_group);
  69 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written);
  70
  71 static inline bool sb_zone_is_full(const struct blk_zone *zone)
  72 {
  73         return (zone->cond == BLK_ZONE_COND_FULL) ||
  74                 (zone->wp + SUPER_INFO_SECTORS > zone->start + zone->capacity);
  75 }
  76
  77 static int copy_zone_info_cb(struct blk_zone *zone, unsigned int idx, void *data)
  78 {
  79         struct blk_zone *zones = data;
  80
  81         memcpy(&zones[idx], zone, sizeof(*zone));
  82
  83         return 0;
  84 }
  85
  86 static int sb_write_pointer(struct block_device *bdev, struct blk_zone *zones,
  87                             u64 *wp_ret)
  88 {
  89         bool empty[BTRFS_NR_SB_LOG_ZONES];
  90         bool full[BTRFS_NR_SB_LOG_ZONES];
  91         sector_t sector;
  92         int i;
  93
  94         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
  95                 ASSERT(zones[i].type != BLK_ZONE_TYPE_CONVENTIONAL);
  96                 empty[i] = (zones[i].cond == BLK_ZONE_COND_EMPTY);
  97                 full[i] = sb_zone_is_full(&zones[i]);
  98         }
  99
 100         /*
 101          * Possible states of log buffer zones
 102          *
 103          *           Empty[0]  In use[0]  Full[0]
 104          * Empty[1]         *          0        1
 105          * In use[1]        x          x        1
 106          * Full[1]          0          0        C
 107          *
 108          * Log position:
 109          *   *: Special case, no superblock is written
 110          *   0: Use write pointer of zones[0]
 111          *   1: Use write pointer of zones[1]
 112          *   C: Compare super blocks from zones[0] and zones[1], use the latest
 113          *      one determined by generation
 114          *   x: Invalid state
 115          */
 116
 117         if (empty[0] && empty[1]) {
 118                 /* Special case to distinguish no superblock to read */
 119                 *wp_ret = zones[0].start << SECTOR_SHIFT;
 120                 return -ENOENT;
 121         } else if (full[0] && full[1]) {
 122                 /* Compare two super blocks */
 123                 struct address_space *mapping = bdev->bd_inode->i_mapping;
 124                 struct page *page[BTRFS_NR_SB_LOG_ZONES];
 125                 struct btrfs_super_block *super[BTRFS_NR_SB_LOG_ZONES];
 126                 int i;
 127
 128                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 129                         u64 zone_end = (zones[i].start + zones[i].capacity) << SECTOR_SHIFT;
 130                         u64 bytenr = ALIGN_DOWN(zone_end, BTRFS_SUPER_INFO_SIZE) -
 131                                                 BTRFS_SUPER_INFO_SIZE;
 132
 133                         page[i] = read_cache_page_gfp(mapping,
 134                                         bytenr >> PAGE_SHIFT, GFP_NOFS);
 135                         if (IS_ERR(page[i])) {
 136                                 if (i == 1)
 137                                         btrfs_release_disk_super(super[0]);
 138                                 return PTR_ERR(page[i]);
 139                         }
 140                         super[i] = page_address(page[i]);
 141                 }
 142
 143                 if (btrfs_super_generation(super[0]) >
 144                     btrfs_super_generation(super[1]))
 145                         sector = zones[1].start;
 146                 else
 147                         sector = zones[0].start;
 148
 149                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++)
 150                         btrfs_release_disk_super(super[i]);
 151         } else if (!full[0] && (empty[1] || full[1])) {
 152                 sector = zones[0].wp;
 153         } else if (full[0]) {
 154                 sector = zones[1].wp;
 155         } else {
 156                 return -EUCLEAN;
 157         }
 158         *wp_ret = sector << SECTOR_SHIFT;
 159         return 0;
 160 }
 161
 162 /*
 163  * Get the first zone number of the superblock mirror
 164  */
 165 static inline u32 sb_zone_number(int shift, int mirror)
 166 {
 167         u64 zone = U64_MAX;
 168
 169         ASSERT(mirror < BTRFS_SUPER_MIRROR_MAX);
 170         switch (mirror) {
 171         case 0: zone = 0; break;
 172         case 1: zone = 1ULL << (BTRFS_SB_LOG_FIRST_SHIFT - shift); break;
 173         case 2: zone = 1ULL << (BTRFS_SB_LOG_SECOND_SHIFT - shift); break;
 174         }
 175
 176         ASSERT(zone <= U32_MAX);
 177
 178         return (u32)zone;
 179 }
 180
 181 static inline sector_t zone_start_sector(u32 zone_number,
 182                                          struct block_device *bdev)
 183 {
 184         return (sector_t)zone_number << ilog2(bdev_zone_sectors(bdev));
 185 }
 186
 187 static inline u64 zone_start_physical(u32 zone_number,
 188                                       struct btrfs_zoned_device_info *zone_info)
 189 {
 190         return (u64)zone_number << zone_info->zone_size_shift;
 191 }
 192
 193 /*
 194  * Emulate blkdev_report_zones() for a non-zoned device. It slices up the block
 195  * device into static sized chunks and fake a conventional zone on each of
 196  * them.
 197  */
 198 static int emulate_report_zones(struct btrfs_device *device, u64 pos,
 199                                 struct blk_zone *zones, unsigned int nr_zones)
 200 {
 201         const sector_t zone_sectors = device->fs_info->zone_size >> SECTOR_SHIFT;
 202         sector_t bdev_size = bdev_nr_sectors(device->bdev);
 203         unsigned int i;
 204
 205         pos >>= SECTOR_SHIFT;
 206         for (i = 0; i < nr_zones; i++) {
 207                 zones[i].start = i * zone_sectors + pos;
 208                 zones[i].len = zone_sectors;
 209                 zones[i].capacity = zone_sectors;
 210                 zones[i].wp = zones[i].start + zone_sectors;
 211                 zones[i].type = BLK_ZONE_TYPE_CONVENTIONAL;
 212                 zones[i].cond = BLK_ZONE_COND_NOT_WP;
 213
 214                 if (zones[i].wp >= bdev_size) {
 215                         i++;
 216                         break;
 217                 }
 218         }
 219
 220         return i;
 221 }
 222
 223 static int btrfs_get_dev_zones(struct btrfs_device *device, u64 pos,
 224                                struct blk_zone *zones, unsigned int *nr_zones)
 225 {
 226         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 227         int ret;
 228
 229         if (!*nr_zones)
 230                 return 0;
 231
 232         if (!bdev_is_zoned(device->bdev)) {
 233                 ret = emulate_report_zones(device, pos, zones, *nr_zones);
 234                 *nr_zones = ret;
 235                 return 0;
 236         }
 237
 238         /* Check cache */
 239         if (zinfo->zone_cache) {
 240                 unsigned int i;
 241                 u32 zno;
 242
 243                 ASSERT(IS_ALIGNED(pos, zinfo->zone_size));
 244                 zno = pos >> zinfo->zone_size_shift;
 245                 /*
 246                  * We cannot report zones beyond the zone end. So, it is OK to
 247                  * cap *nr_zones to at the end.
 248                  */
 249                 *nr_zones = min_t(u32, *nr_zones, zinfo->nr_zones - zno);
 250
 251                 for (i = 0; i < *nr_zones; i++) {
 252                         struct blk_zone *zone_info;
 253
 254                         zone_info = &zinfo->zone_cache[zno + i];
 255                         if (!zone_info->len)
 256                                 break;
 257                 }
 258
 259                 if (i == *nr_zones) {
 260                         /* Cache hit on all the zones */
 261                         memcpy(zones, zinfo->zone_cache + zno,
 262                                sizeof(*zinfo->zone_cache) * *nr_zones);
 263                         return 0;
 264                 }
 265         }
 266
 267         ret = blkdev_report_zones(device->bdev, pos >> SECTOR_SHIFT, *nr_zones,
 268                                   copy_zone_info_cb, zones);
 269         if (ret < 0) {
 270                 btrfs_err_in_rcu(device->fs_info,
 271                                  "zoned: failed to read zone %llu on %s (devid %llu)",
 272                                  pos, rcu_str_deref(device->name),
 273                                  device->devid);
 274                 return ret;
 275         }
 276         *nr_zones = ret;
 277         if (!ret)
 278                 return -EIO;
 279
 280         /* Populate cache */
 281         if (zinfo->zone_cache) {
 282                 u32 zno = pos >> zinfo->zone_size_shift;
 283
 284                 memcpy(zinfo->zone_cache + zno, zones,
 285                        sizeof(*zinfo->zone_cache) * *nr_zones);
 286         }
 287
 288         return 0;
 289 }
 290
 291 /* The emulated zone size is determined from the size of device extent */
 292 static int calculate_emulated_zone_size(struct btrfs_fs_info *fs_info)
 293 {
 294         struct btrfs_path *path;
 295         struct btrfs_root *root = fs_info->dev_root;
 296         struct btrfs_key key;
 297         struct extent_buffer *leaf;
 298         struct btrfs_dev_extent *dext;
 299         int ret = 0;
 300
 301         key.objectid = 1;
 302         key.type = BTRFS_DEV_EXTENT_KEY;
 303         key.offset = 0;
 304
 305         path = btrfs_alloc_path();
 306         if (!path)
 307                 return -ENOMEM;
 308
 309         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 310         if (ret < 0)
 311                 goto out;
 312
 313         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
 314                 ret = btrfs_next_leaf(root, path);
 315                 if (ret < 0)
 316                         goto out;
 317                 /* No dev extents at all? Not good */
 318                 if (ret > 0) {
 319                         ret = -EUCLEAN;
 320                         goto out;
 321                 }
 322         }
 323
 324         leaf = path->nodes[0];
 325         dext = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
 326         fs_info->zone_size = btrfs_dev_extent_length(leaf, dext);
 327         ret = 0;
 328
 329 out:
 330         btrfs_free_path(path);
 331
 332         return ret;
 333 }
 334
 335 int btrfs_get_dev_zone_info_all_devices(struct btrfs_fs_info *fs_info)
 336 {
 337         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 338         struct btrfs_device *device;
 339         int ret = 0;
 340
 341         /* fs_info->zone_size might not set yet. Use the incomapt flag here. */
 342         if (!btrfs_fs_incompat(fs_info, ZONED))
 343                 return 0;
 344
 345         mutex_lock(&fs_devices->device_list_mutex);
 346         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 347                 /* We can skip reading of zone info for missing devices */
 348                 if (!device->bdev)
 349                         continue;
 350
 351                 ret = btrfs_get_dev_zone_info(device, true);
 352                 if (ret)
 353                         break;
 354         }
 355         mutex_unlock(&fs_devices->device_list_mutex);
 356
 357         return ret;
 358 }
 359
 360 int btrfs_get_dev_zone_info(struct btrfs_device *device, bool populate_cache)
 361 {
 362         struct btrfs_fs_info *fs_info = device->fs_info;
 363         struct btrfs_zoned_device_info *zone_info = NULL;
 364         struct block_device *bdev = device->bdev;
 365         unsigned int max_active_zones;
 366         unsigned int nactive;
 367         sector_t nr_sectors;
 368         sector_t sector = 0;
 369         struct blk_zone *zones = NULL;
 370         unsigned int i, nreported = 0, nr_zones;
 371         sector_t zone_sectors;
 372         char *model, *emulated;
 373         int ret;
 374
 375         /*
 376          * Cannot use btrfs_is_zoned here, since fs_info::zone_size might not
 377          * yet be set.
 378          */
 379         if (!btrfs_fs_incompat(fs_info, ZONED))
 380                 return 0;
 381
 382         if (device->zone_info)
 383                 return 0;
 384
 385         zone_info = kzalloc(sizeof(*zone_info), GFP_KERNEL);
 386         if (!zone_info)
 387                 return -ENOMEM;
 388
 389         device->zone_info = zone_info;
 390
 391         if (!bdev_is_zoned(bdev)) {
 392                 if (!fs_info->zone_size) {
 393                         ret = calculate_emulated_zone_size(fs_info);
 394                         if (ret)
 395                                 goto out;
 396                 }
 397
 398                 ASSERT(fs_info->zone_size);
 399                 zone_sectors = fs_info->zone_size >> SECTOR_SHIFT;
 400         } else {
 401                 zone_sectors = bdev_zone_sectors(bdev);
 402         }
 403
 404         ASSERT(is_power_of_two_u64(zone_sectors));
 405         zone_info->zone_size = zone_sectors << SECTOR_SHIFT;
 406
 407         /* We reject devices with a zone size larger than 8GB */
 408         if (zone_info->zone_size > BTRFS_MAX_ZONE_SIZE) {
 409                 btrfs_err_in_rcu(fs_info,
 410                 "zoned: %s: zone size %llu larger than supported maximum %llu",
 411                                  rcu_str_deref(device->name),
 412                                  zone_info->zone_size, BTRFS_MAX_ZONE_SIZE);
 413                 ret = -EINVAL;
 414                 goto out;
 415         } else if (zone_info->zone_size < BTRFS_MIN_ZONE_SIZE) {
 416                 btrfs_err_in_rcu(fs_info,
 417                 "zoned: %s: zone size %llu smaller than supported minimum %u",
 418                                  rcu_str_deref(device->name),
 419                                  zone_info->zone_size, BTRFS_MIN_ZONE_SIZE);
 420                 ret = -EINVAL;
 421                 goto out;
 422         }
 423
 424         nr_sectors = bdev_nr_sectors(bdev);
 425         zone_info->zone_size_shift = ilog2(zone_info->zone_size);
 426         zone_info->nr_zones = nr_sectors >> ilog2(zone_sectors);
 427         if (!IS_ALIGNED(nr_sectors, zone_sectors))
 428                 zone_info->nr_zones++;
 429
 430         max_active_zones = bdev_max_active_zones(bdev);
 431         if (max_active_zones && max_active_zones < BTRFS_MIN_ACTIVE_ZONES) {
 432                 btrfs_err_in_rcu(fs_info,
 433 "zoned: %s: max active zones %u is too small, need at least %u active zones",
 434                                  rcu_str_deref(device->name), max_active_zones,
 435                                  BTRFS_MIN_ACTIVE_ZONES);
 436                 ret = -EINVAL;
 437                 goto out;
 438         }
 439         zone_info->max_active_zones = max_active_zones;
 440
 441         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 442         if (!zone_info->seq_zones) {
 443                 ret = -ENOMEM;
 444                 goto out;
 445         }
 446
 447         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 448         if (!zone_info->empty_zones) {
 449                 ret = -ENOMEM;
 450                 goto out;
 451         }
 452
 453         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 454         if (!zone_info->active_zones) {
 455                 ret = -ENOMEM;
 456                 goto out;
 457         }
 458
 459         zones = kvcalloc(BTRFS_REPORT_NR_ZONES, sizeof(struct blk_zone), GFP_KERNEL);
 460         if (!zones) {
 461                 ret = -ENOMEM;
 462                 goto out;
 463         }
 464
 465         /*
 466          * Enable zone cache only for a zoned device. On a non-zoned device, we
 467          * fill the zone info with emulated CONVENTIONAL zones, so no need to
 468          * use the cache.
 469          */
 470         if (populate_cache && bdev_is_zoned(device->bdev)) {
 471                 zone_info->zone_cache = vcalloc(zone_info->nr_zones,
 472                                                 sizeof(struct blk_zone));
 473                 if (!zone_info->zone_cache) {
 474                         btrfs_err_in_rcu(device->fs_info,
 475                                 "zoned: failed to allocate zone cache for %s",
 476                                 rcu_str_deref(device->name));
 477                         ret = -ENOMEM;
 478                         goto out;
 479                 }
 480         }
 481
 482         /* Get zones type */
 483         nactive = 0;
 484         while (sector < nr_sectors) {
 485                 nr_zones = BTRFS_REPORT_NR_ZONES;
 486                 ret = btrfs_get_dev_zones(device, sector << SECTOR_SHIFT, zones,
 487                                           &nr_zones);
 488                 if (ret)
 489                         goto out;
 490
 491                 for (i = 0; i < nr_zones; i++) {
 492                         if (zones[i].type == BLK_ZONE_TYPE_SEQWRITE_REQ)
 493                                 __set_bit(nreported, zone_info->seq_zones);
 494                         switch (zones[i].cond) {
 495                         case BLK_ZONE_COND_EMPTY:
 496                                 __set_bit(nreported, zone_info->empty_zones);
 497                                 break;
 498                         case BLK_ZONE_COND_IMP_OPEN:
 499                         case BLK_ZONE_COND_EXP_OPEN:
 500                         case BLK_ZONE_COND_CLOSED:
 501                                 __set_bit(nreported, zone_info->active_zones);
 502                                 nactive++;
 503                                 break;
 504                         }
 505                         nreported++;
 506                 }
 507                 sector = zones[nr_zones - 1].start + zones[nr_zones - 1].len;
 508         }
 509
 510         if (nreported != zone_info->nr_zones) {
 511                 btrfs_err_in_rcu(device->fs_info,
 512                                  "inconsistent number of zones on %s (%u/%u)",
 513                                  rcu_str_deref(device->name), nreported,
 514                                  zone_info->nr_zones);
 515                 ret = -EIO;
 516                 goto out;
 517         }
 518
 519         if (max_active_zones) {
 520                 if (nactive > max_active_zones) {
 521                         btrfs_err_in_rcu(device->fs_info,
 522                         "zoned: %u active zones on %s exceeds max_active_zones %u",
 523                                          nactive, rcu_str_deref(device->name),
 524                                          max_active_zones);
 525                         ret = -EIO;
 526                         goto out;
 527                 }
 528                 atomic_set(&zone_info->active_zones_left,
 529                            max_active_zones - nactive);
 530                 set_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags);
 531         }
 532
 533         /* Validate superblock log */
 534         nr_zones = BTRFS_NR_SB_LOG_ZONES;
 535         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 536                 u32 sb_zone;
 537                 u64 sb_wp;
 538                 int sb_pos = BTRFS_NR_SB_LOG_ZONES * i;
 539
 540                 sb_zone = sb_zone_number(zone_info->zone_size_shift, i);
 541                 if (sb_zone + 1 >= zone_info->nr_zones)
 542                         continue;
 543
 544                 ret = btrfs_get_dev_zones(device,
 545                                           zone_start_physical(sb_zone, zone_info),
 546                                           &zone_info->sb_zones[sb_pos],
 547                                           &nr_zones);
 548                 if (ret)
 549                         goto out;
 550
 551                 if (nr_zones != BTRFS_NR_SB_LOG_ZONES) {
 552                         btrfs_err_in_rcu(device->fs_info,
 553         "zoned: failed to read super block log zone info at devid %llu zone %u",
 554                                          device->devid, sb_zone);
 555                         ret = -EUCLEAN;
 556                         goto out;
 557                 }
 558
 559                 /*
 560                  * If zones[0] is conventional, always use the beginning of the
 561                  * zone to record superblock. No need to validate in that case.
 562                  */
 563                 if (zone_info->sb_zones[BTRFS_NR_SB_LOG_ZONES * i].type ==
 564                     BLK_ZONE_TYPE_CONVENTIONAL)
 565                         continue;
 566
 567                 ret = sb_write_pointer(device->bdev,
 568                                        &zone_info->sb_zones[sb_pos], &sb_wp);
 569                 if (ret != -ENOENT && ret) {
 570                         btrfs_err_in_rcu(device->fs_info,
 571                         "zoned: super block log zone corrupted devid %llu zone %u",
 572                                          device->devid, sb_zone);
 573                         ret = -EUCLEAN;
 574                         goto out;
 575                 }
 576         }
 577
 578
 579         kvfree(zones);
 580
 581         switch (bdev_zoned_model(bdev)) {
 582         case BLK_ZONED_HM:
 583                 model = "host-managed zoned";
 584                 emulated = "";
 585                 break;
 586         case BLK_ZONED_HA:
 587                 model = "host-aware zoned";
 588                 emulated = "";
 589                 break;
 590         case BLK_ZONED_NONE:
 591                 model = "regular";
 592                 emulated = "emulated ";
 593                 break;
 594         default:
 595                 /* Just in case */
 596                 btrfs_err_in_rcu(fs_info, "zoned: unsupported model %d on %s",
 597                                  bdev_zoned_model(bdev),
 598                                  rcu_str_deref(device->name));
 599                 ret = -EOPNOTSUPP;
 600                 goto out_free_zone_info;
 601         }
 602
 603         btrfs_info_in_rcu(fs_info,
 604                 "%s block device %s, %u %szones of %llu bytes",
 605                 model, rcu_str_deref(device->name), zone_info->nr_zones,
 606                 emulated, zone_info->zone_size);
 607
 608         return 0;
 609
 610 out:
 611         kvfree(zones);
 612 out_free_zone_info:
 613         btrfs_destroy_dev_zone_info(device);
 614
 615         return ret;
 616 }
 617
 618 void btrfs_destroy_dev_zone_info(struct btrfs_device *device)
 619 {
 620         struct btrfs_zoned_device_info *zone_info = device->zone_info;
 621
 622         if (!zone_info)
 623                 return;
 624
 625         bitmap_free(zone_info->active_zones);
 626         bitmap_free(zone_info->seq_zones);
 627         bitmap_free(zone_info->empty_zones);
 628         vfree(zone_info->zone_cache);
 629         kfree(zone_info);
 630         device->zone_info = NULL;
 631 }
 632
 633 struct btrfs_zoned_device_info *btrfs_clone_dev_zone_info(struct btrfs_device *orig_dev)
 634 {
 635         struct btrfs_zoned_device_info *zone_info;
 636
 637         zone_info = kmemdup(orig_dev->zone_info, sizeof(*zone_info), GFP_KERNEL);
 638         if (!zone_info)
 639                 return NULL;
 640
 641         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 642         if (!zone_info->seq_zones)
 643                 goto out;
 644
 645         bitmap_copy(zone_info->seq_zones, orig_dev->zone_info->seq_zones,
 646                     zone_info->nr_zones);
 647
 648         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 649         if (!zone_info->empty_zones)
 650                 goto out;
 651
 652         bitmap_copy(zone_info->empty_zones, orig_dev->zone_info->empty_zones,
 653                     zone_info->nr_zones);
 654
 655         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 656         if (!zone_info->active_zones)
 657                 goto out;
 658
 659         bitmap_copy(zone_info->active_zones, orig_dev->zone_info->active_zones,
 660                     zone_info->nr_zones);
 661         zone_info->zone_cache = NULL;
 662
 663         return zone_info;
 664
 665 out:
 666         bitmap_free(zone_info->seq_zones);
 667         bitmap_free(zone_info->empty_zones);
 668         bitmap_free(zone_info->active_zones);
 669         kfree(zone_info);
 670         return NULL;
 671 }
 672
 673 int btrfs_get_dev_zone(struct btrfs_device *device, u64 pos,
 674                        struct blk_zone *zone)
 675 {
 676         unsigned int nr_zones = 1;
 677         int ret;
 678
 679         ret = btrfs_get_dev_zones(device, pos, zone, &nr_zones);
 680         if (ret != 0 || !nr_zones)
 681                 return ret ? ret : -EIO;
 682
 683         return 0;
 684 }
 685
 686 static int btrfs_check_for_zoned_device(struct btrfs_fs_info *fs_info)
 687 {
 688         struct btrfs_device *device;
 689
 690         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 691                 if (device->bdev &&
 692                     bdev_zoned_model(device->bdev) == BLK_ZONED_HM) {
 693                         btrfs_err(fs_info,
 694                                 "zoned: mode not enabled but zoned device found: %pg",
 695                                 device->bdev);
 696                         return -EINVAL;
 697                 }
 698         }
 699
 700         return 0;
 701 }
 702
 703 int btrfs_check_zoned_mode(struct btrfs_fs_info *fs_info)
 704 {
 705         struct queue_limits *lim = &fs_info->limits;
 706         struct btrfs_device *device;
 707         u64 zone_size = 0;
 708         int ret;
 709
 710         /*
 711          * Host-Managed devices can't be used without the ZONED flag.  With the
 712          * ZONED all devices can be used, using zone emulation if required.
 713          */
 714         if (!btrfs_fs_incompat(fs_info, ZONED))
 715                 return btrfs_check_for_zoned_device(fs_info);
 716
 717         blk_set_stacking_limits(lim);
 718
 719         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 720                 struct btrfs_zoned_device_info *zone_info = device->zone_info;
 721
 722                 if (!device->bdev)
 723                         continue;
 724
 725                 if (!zone_size) {
 726                         zone_size = zone_info->zone_size;
 727                 } else if (zone_info->zone_size != zone_size) {
 728                         btrfs_err(fs_info,
 729                 "zoned: unequal block device zone sizes: have %llu found %llu",
 730                                   zone_info->zone_size, zone_size);
 731                         return -EINVAL;
 732                 }
 733
 734                 /*
 735                  * With the zoned emulation, we can have non-zoned device on the
 736                  * zoned mode. In this case, we don't have a valid max zone
 737                  * append size.
 738                  */
 739                 if (bdev_is_zoned(device->bdev)) {
 740                         blk_stack_limits(lim,
 741                                          &bdev_get_queue(device->bdev)->limits,
 742                                          0);
 743                 }
 744         }
 745
 746         /*
 747          * stripe_size is always aligned to BTRFS_STRIPE_LEN in
 748          * btrfs_create_chunk(). Since we want stripe_len == zone_size,
 749          * check the alignment here.
 750          */
 751         if (!IS_ALIGNED(zone_size, BTRFS_STRIPE_LEN)) {
 752                 btrfs_err(fs_info,
 753                           "zoned: zone size %llu not aligned to stripe %u",
 754                           zone_size, BTRFS_STRIPE_LEN);
 755                 return -EINVAL;
 756         }
 757
 758         if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
 759                 btrfs_err(fs_info, "zoned: mixed block groups not supported");
 760                 return -EINVAL;
 761         }
 762
 763         fs_info->zone_size = zone_size;
 764         /*
 765          * Also limit max_zone_append_size by max_segments * PAGE_SIZE.
 766          * Technically, we can have multiple pages per segment. But, since
 767          * we add the pages one by one to a bio, and cannot increase the
 768          * metadata reservation even if it increases the number of extents, it
 769          * is safe to stick with the limit.
 770          */
 771         fs_info->max_zone_append_size = ALIGN_DOWN(
 772                 min3((u64)lim->max_zone_append_sectors << SECTOR_SHIFT,
 773                      (u64)lim->max_sectors << SECTOR_SHIFT,
 774                      (u64)lim->max_segments << PAGE_SHIFT),
 775                 fs_info->sectorsize);
 776         fs_info->fs_devices->chunk_alloc_policy = BTRFS_CHUNK_ALLOC_ZONED;
 777         if (fs_info->max_zone_append_size < fs_info->max_extent_size)
 778                 fs_info->max_extent_size = fs_info->max_zone_append_size;
 779
 780         /*
 781          * Check mount options here, because we might change fs_info->zoned
 782          * from fs_info->zone_size.
 783          */
 784         ret = btrfs_check_mountopts_zoned(fs_info, &fs_info->mount_opt);
 785         if (ret)
 786                 return ret;
 787
 788         btrfs_info(fs_info, "zoned mode enabled with zone size %llu", zone_size);
 789         return 0;
 790 }
 791
 792 int btrfs_check_mountopts_zoned(struct btrfs_fs_info *info, unsigned long *mount_opt)
 793 {
 794         if (!btrfs_is_zoned(info))
 795                 return 0;
 796
 797         /*
 798          * Space cache writing is not COWed. Disable that to avoid write errors
 799          * in sequential zones.
 800          */
 801         if (btrfs_raw_test_opt(*mount_opt, SPACE_CACHE)) {
 802                 btrfs_err(info, "zoned: space cache v1 is not supported");
 803                 return -EINVAL;
 804         }
 805
 806         if (btrfs_raw_test_opt(*mount_opt, NODATACOW)) {
 807                 btrfs_err(info, "zoned: NODATACOW not supported");
 808                 return -EINVAL;
 809         }
 810
 811         if (btrfs_raw_test_opt(*mount_opt, DISCARD_ASYNC)) {
 812                 btrfs_info(info,
 813                            "zoned: async discard ignored and disabled for zoned mode");
 814                 btrfs_clear_opt(*mount_opt, DISCARD_ASYNC);
 815         }
 816
 817         return 0;
 818 }
 819
 820 static int sb_log_location(struct block_device *bdev, struct blk_zone *zones,
 821                            int rw, u64 *bytenr_ret)
 822 {
 823         u64 wp;
 824         int ret;
 825
 826         if (zones[0].type == BLK_ZONE_TYPE_CONVENTIONAL) {
 827                 *bytenr_ret = zones[0].start << SECTOR_SHIFT;
 828                 return 0;
 829         }
 830
 831         ret = sb_write_pointer(bdev, zones, &wp);
 832         if (ret != -ENOENT && ret < 0)
 833                 return ret;
 834
 835         if (rw == WRITE) {
 836                 struct blk_zone *reset = NULL;
 837
 838                 if (wp == zones[0].start << SECTOR_SHIFT)
 839                         reset = &zones[0];
 840                 else if (wp == zones[1].start << SECTOR_SHIFT)
 841                         reset = &zones[1];
 842
 843                 if (reset && reset->cond != BLK_ZONE_COND_EMPTY) {
 844                         ASSERT(sb_zone_is_full(reset));
 845
 846                         ret = blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
 847                                                reset->start, reset->len,
 848                                                GFP_NOFS);
 849                         if (ret)
 850                                 return ret;
 851
 852                         reset->cond = BLK_ZONE_COND_EMPTY;
 853                         reset->wp = reset->start;
 854                 }
 855         } else if (ret != -ENOENT) {
 856                 /*
 857                  * For READ, we want the previous one. Move write pointer to
 858                  * the end of a zone, if it is at the head of a zone.
 859                  */
 860                 u64 zone_end = 0;
 861
 862                 if (wp == zones[0].start << SECTOR_SHIFT)
 863                         zone_end = zones[1].start + zones[1].capacity;
 864                 else if (wp == zones[1].start << SECTOR_SHIFT)
 865                         zone_end = zones[0].start + zones[0].capacity;
 866                 if (zone_end)
 867                         wp = ALIGN_DOWN(zone_end << SECTOR_SHIFT,
 868                                         BTRFS_SUPER_INFO_SIZE);
 869
 870                 wp -= BTRFS_SUPER_INFO_SIZE;
 871         }
 872
 873         *bytenr_ret = wp;
 874         return 0;
 875
 876 }
 877
 878 int btrfs_sb_log_location_bdev(struct block_device *bdev, int mirror, int rw,
 879                                u64 *bytenr_ret)
 880 {
 881         struct blk_zone zones[BTRFS_NR_SB_LOG_ZONES];
 882         sector_t zone_sectors;
 883         u32 sb_zone;
 884         int ret;
 885         u8 zone_sectors_shift;
 886         sector_t nr_sectors;
 887         u32 nr_zones;
 888
 889         if (!bdev_is_zoned(bdev)) {
 890                 *bytenr_ret = btrfs_sb_offset(mirror);
 891                 return 0;
 892         }
 893
 894         ASSERT(rw == READ || rw == WRITE);
 895
 896         zone_sectors = bdev_zone_sectors(bdev);
 897         if (!is_power_of_2(zone_sectors))
 898                 return -EINVAL;
 899         zone_sectors_shift = ilog2(zone_sectors);
 900         nr_sectors = bdev_nr_sectors(bdev);
 901         nr_zones = nr_sectors >> zone_sectors_shift;
 902
 903         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
 904         if (sb_zone + 1 >= nr_zones)
 905                 return -ENOENT;
 906
 907         ret = blkdev_report_zones(bdev, zone_start_sector(sb_zone, bdev),
 908                                   BTRFS_NR_SB_LOG_ZONES, copy_zone_info_cb,
 909                                   zones);
 910         if (ret < 0)
 911                 return ret;
 912         if (ret != BTRFS_NR_SB_LOG_ZONES)
 913                 return -EIO;
 914
 915         return sb_log_location(bdev, zones, rw, bytenr_ret);
 916 }
 917
 918 int btrfs_sb_log_location(struct btrfs_device *device, int mirror, int rw,
 919                           u64 *bytenr_ret)
 920 {
 921         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 922         u32 zone_num;
 923
 924         /*
 925          * For a zoned filesystem on a non-zoned block device, use the same
 926          * super block locations as regular filesystem. Doing so, the super
 927          * block can always be retrieved and the zoned flag of the volume
 928          * detected from the super block information.
 929          */
 930         if (!bdev_is_zoned(device->bdev)) {
 931                 *bytenr_ret = btrfs_sb_offset(mirror);
 932                 return 0;
 933         }
 934
 935         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 936         if (zone_num + 1 >= zinfo->nr_zones)
 937                 return -ENOENT;
 938
 939         return sb_log_location(device->bdev,
 940                                &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror],
 941                                rw, bytenr_ret);
 942 }
 943
 944 static inline bool is_sb_log_zone(struct btrfs_zoned_device_info *zinfo,
 945                                   int mirror)
 946 {
 947         u32 zone_num;
 948
 949         if (!zinfo)
 950                 return false;
 951
 952         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 953         if (zone_num + 1 >= zinfo->nr_zones)
 954                 return false;
 955
 956         if (!test_bit(zone_num, zinfo->seq_zones))
 957                 return false;
 958
 959         return true;
 960 }
 961
 962 int btrfs_advance_sb_log(struct btrfs_device *device, int mirror)
 963 {
 964         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 965         struct blk_zone *zone;
 966         int i;
 967
 968         if (!is_sb_log_zone(zinfo, mirror))
 969                 return 0;
 970
 971         zone = &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror];
 972         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 973                 /* Advance the next zone */
 974                 if (zone->cond == BLK_ZONE_COND_FULL) {
 975                         zone++;
 976                         continue;
 977                 }
 978
 979                 if (zone->cond == BLK_ZONE_COND_EMPTY)
 980                         zone->cond = BLK_ZONE_COND_IMP_OPEN;
 981
 982                 zone->wp += SUPER_INFO_SECTORS;
 983
 984                 if (sb_zone_is_full(zone)) {
 985                         /*
 986                          * No room left to write new superblock. Since
 987                          * superblock is written with REQ_SYNC, it is safe to
 988                          * finish the zone now.
 989                          *
 990                          * If the write pointer is exactly at the capacity,
 991                          * explicit ZONE_FINISH is not necessary.
 992                          */
 993                         if (zone->wp != zone->start + zone->capacity) {
 994                                 int ret;
 995
 996                                 ret = blkdev_zone_mgmt(device->bdev,
 997                                                 REQ_OP_ZONE_FINISH, zone->start,
 998                                                 zone->len, GFP_NOFS);
 999                                 if (ret)
1000                                         return ret;
1001                         }
1002
1003                         zone->wp = zone->start + zone->len;
1004                         zone->cond = BLK_ZONE_COND_FULL;
1005                 }
1006                 return 0;
1007         }
1008
1009         /* All the zones are FULL. Should not reach here. */
1010         ASSERT(0);
1011         return -EIO;
1012 }
1013
1014 int btrfs_reset_sb_log_zones(struct block_device *bdev, int mirror)
1015 {
1016         sector_t zone_sectors;
1017         sector_t nr_sectors;
1018         u8 zone_sectors_shift;
1019         u32 sb_zone;
1020         u32 nr_zones;
1021
1022         zone_sectors = bdev_zone_sectors(bdev);
1023         zone_sectors_shift = ilog2(zone_sectors);
1024         nr_sectors = bdev_nr_sectors(bdev);
1025         nr_zones = nr_sectors >> zone_sectors_shift;
1026
1027         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
1028         if (sb_zone + 1 >= nr_zones)
1029                 return -ENOENT;
1030
1031         return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
1032                                 zone_start_sector(sb_zone, bdev),
1033                                 zone_sectors * BTRFS_NR_SB_LOG_ZONES, GFP_NOFS);
1034 }
1035
1036 /*
1037  * Find allocatable zones within a given region.
1038  *
1039  * @device:     the device to allocate a region on
1040  * @hole_start: the position of the hole to allocate the region
1041  * @num_bytes:  size of wanted region
1042  * @hole_end:   the end of the hole
1043  * @return:     position of allocatable zones
1044  *
1045  * Allocatable region should not contain any superblock locations.
1046  */
1047 u64 btrfs_find_allocatable_zones(struct btrfs_device *device, u64 hole_start,
1048                                  u64 hole_end, u64 num_bytes)
1049 {
1050         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1051         const u8 shift = zinfo->zone_size_shift;
1052         u64 nzones = num_bytes >> shift;
1053         u64 pos = hole_start;
1054         u64 begin, end;
1055         bool have_sb;
1056         int i;
1057
1058         ASSERT(IS_ALIGNED(hole_start, zinfo->zone_size));
1059         ASSERT(IS_ALIGNED(num_bytes, zinfo->zone_size));
1060
1061         while (pos < hole_end) {
1062                 begin = pos >> shift;
1063                 end = begin + nzones;
1064
1065                 if (end > zinfo->nr_zones)
1066                         return hole_end;
1067
1068                 /* Check if zones in the region are all empty */
1069                 if (btrfs_dev_is_sequential(device, pos) &&
1070                     !bitmap_test_range_all_set(zinfo->empty_zones, begin, nzones)) {
1071                         pos += zinfo->zone_size;
1072                         continue;
1073                 }
1074
1075                 have_sb = false;
1076                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1077                         u32 sb_zone;
1078                         u64 sb_pos;
1079
1080                         sb_zone = sb_zone_number(shift, i);
1081                         if (!(end <= sb_zone ||
1082                               sb_zone + BTRFS_NR_SB_LOG_ZONES <= begin)) {
1083                                 have_sb = true;
1084                                 pos = zone_start_physical(
1085                                         sb_zone + BTRFS_NR_SB_LOG_ZONES, zinfo);
1086                                 break;
1087                         }
1088
1089                         /* We also need to exclude regular superblock positions */
1090                         sb_pos = btrfs_sb_offset(i);
1091                         if (!(pos + num_bytes <= sb_pos ||
1092                               sb_pos + BTRFS_SUPER_INFO_SIZE <= pos)) {
1093                                 have_sb = true;
1094                                 pos = ALIGN(sb_pos + BTRFS_SUPER_INFO_SIZE,
1095                                             zinfo->zone_size);
1096                                 break;
1097                         }
1098                 }
1099                 if (!have_sb)
1100                         break;
1101         }
1102
1103         return pos;
1104 }
1105
1106 static bool btrfs_dev_set_active_zone(struct btrfs_device *device, u64 pos)
1107 {
1108         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1109         unsigned int zno = (pos >> zone_info->zone_size_shift);
1110
1111         /* We can use any number of zones */
1112         if (zone_info->max_active_zones == 0)
1113                 return true;
1114
1115         if (!test_bit(zno, zone_info->active_zones)) {
1116                 /* Active zone left? */
1117                 if (atomic_dec_if_positive(&zone_info->active_zones_left) < 0)
1118                         return false;
1119                 if (test_and_set_bit(zno, zone_info->active_zones)) {
1120                         /* Someone already set the bit */
1121                         atomic_inc(&zone_info->active_zones_left);
1122                 }
1123         }
1124
1125         return true;
1126 }
1127
1128 static void btrfs_dev_clear_active_zone(struct btrfs_device *device, u64 pos)
1129 {
1130         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1131         unsigned int zno = (pos >> zone_info->zone_size_shift);
1132
1133         /* We can use any number of zones */
1134         if (zone_info->max_active_zones == 0)
1135                 return;
1136
1137         if (test_and_clear_bit(zno, zone_info->active_zones))
1138                 atomic_inc(&zone_info->active_zones_left);
1139 }
1140
1141 int btrfs_reset_device_zone(struct btrfs_device *device, u64 physical,
1142                             u64 length, u64 *bytes)
1143 {
1144         int ret;
1145
1146         *bytes = 0;
1147         ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_RESET,
1148                                physical >> SECTOR_SHIFT, length >> SECTOR_SHIFT,
1149                                GFP_NOFS);
1150         if (ret)
1151                 return ret;
1152
1153         *bytes = length;
1154         while (length) {
1155                 btrfs_dev_set_zone_empty(device, physical);
1156                 btrfs_dev_clear_active_zone(device, physical);
1157                 physical += device->zone_info->zone_size;
1158                 length -= device->zone_info->zone_size;
1159         }
1160
1161         return 0;
1162 }
1163
1164 int btrfs_ensure_empty_zones(struct btrfs_device *device, u64 start, u64 size)
1165 {
1166         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1167         const u8 shift = zinfo->zone_size_shift;
1168         unsigned long begin = start >> shift;
1169         unsigned long nbits = size >> shift;
1170         u64 pos;
1171         int ret;
1172
1173         ASSERT(IS_ALIGNED(start, zinfo->zone_size));
1174         ASSERT(IS_ALIGNED(size, zinfo->zone_size));
1175
1176         if (begin + nbits > zinfo->nr_zones)
1177                 return -ERANGE;
1178
1179         /* All the zones are conventional */
1180         if (bitmap_test_range_all_zero(zinfo->seq_zones, begin, nbits))
1181                 return 0;
1182
1183         /* All the zones are sequential and empty */
1184         if (bitmap_test_range_all_set(zinfo->seq_zones, begin, nbits) &&
1185             bitmap_test_range_all_set(zinfo->empty_zones, begin, nbits))
1186                 return 0;
1187
1188         for (pos = start; pos < start + size; pos += zinfo->zone_size) {
1189                 u64 reset_bytes;
1190
1191                 if (!btrfs_dev_is_sequential(device, pos) ||
1192                     btrfs_dev_is_empty_zone(device, pos))
1193                         continue;
1194
1195                 /* Free regions should be empty */
1196                 btrfs_warn_in_rcu(
1197                         device->fs_info,
1198                 "zoned: resetting device %s (devid %llu) zone %llu for allocation",
1199                         rcu_str_deref(device->name), device->devid, pos >> shift);
1200                 WARN_ON_ONCE(1);
1201
1202                 ret = btrfs_reset_device_zone(device, pos, zinfo->zone_size,
1203                                               &reset_bytes);
1204                 if (ret)
1205                         return ret;
1206         }
1207
1208         return 0;
1209 }
1210
1211 /*
1212  * Calculate an allocation pointer from the extent allocation information
1213  * for a block group consist of conventional zones. It is pointed to the
1214  * end of the highest addressed extent in the block group as an allocation
1215  * offset.
1216  */
1217 static int calculate_alloc_pointer(struct btrfs_block_group *cache,
1218                                    u64 *offset_ret, bool new)
1219 {
1220         struct btrfs_fs_info *fs_info = cache->fs_info;
1221         struct btrfs_root *root;
1222         struct btrfs_path *path;
1223         struct btrfs_key key;
1224         struct btrfs_key found_key;
1225         int ret;
1226         u64 length;
1227
1228         /*
1229          * Avoid  tree lookups for a new block group, there's no use for it.
1230          * It must always be 0.
1231          *
1232          * Also, we have a lock chain of extent buffer lock -> chunk mutex.
1233          * For new a block group, this function is called from
1234          * btrfs_make_block_group() which is already taking the chunk mutex.
1235          * Thus, we cannot call calculate_alloc_pointer() which takes extent
1236          * buffer locks to avoid deadlock.
1237          */
1238         if (new) {
1239                 *offset_ret = 0;
1240                 return 0;
1241         }
1242
1243         path = btrfs_alloc_path();
1244         if (!path)
1245                 return -ENOMEM;
1246
1247         key.objectid = cache->start + cache->length;
1248         key.type = 0;
1249         key.offset = 0;
1250
1251         root = btrfs_extent_root(fs_info, key.objectid);
1252         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1253         /* We should not find the exact match */
1254         if (!ret)
1255                 ret = -EUCLEAN;
1256         if (ret < 0)
1257                 goto out;
1258
1259         ret = btrfs_previous_extent_item(root, path, cache->start);
1260         if (ret) {
1261                 if (ret == 1) {
1262                         ret = 0;
1263                         *offset_ret = 0;
1264                 }
1265                 goto out;
1266         }
1267
1268         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
1269
1270         if (found_key.type == BTRFS_EXTENT_ITEM_KEY)
1271                 length = found_key.offset;
1272         else
1273                 length = fs_info->nodesize;
1274
1275         if (!(found_key.objectid >= cache->start &&
1276                found_key.objectid + length <= cache->start + cache->length)) {
1277                 ret = -EUCLEAN;
1278                 goto out;
1279         }
1280         *offset_ret = found_key.objectid + length - cache->start;
1281         ret = 0;
1282
1283 out:
1284         btrfs_free_path(path);
1285         return ret;
1286 }
1287
1288 struct zone_info {
1289         u64 physical;
1290         u64 capacity;
1291         u64 alloc_offset;
1292 };
1293
1294 static int btrfs_load_zone_info(struct btrfs_fs_info *fs_info, int zone_idx,
1295                                 struct zone_info *info, unsigned long *active,
1296                                 struct btrfs_chunk_map *map)
1297 {
1298         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1299         struct btrfs_device *device = map->stripes[zone_idx].dev;
1300         int dev_replace_is_ongoing = 0;
1301         unsigned int nofs_flag;
1302         struct blk_zone zone;
1303         int ret;
1304
1305         info->physical = map->stripes[zone_idx].physical;
1306
1307         if (!device->bdev) {
1308                 info->alloc_offset = WP_MISSING_DEV;
1309                 return 0;
1310         }
1311
1312         /* Consider a zone as active if we can allow any number of active zones. */
1313         if (!device->zone_info->max_active_zones)
1314                 __set_bit(zone_idx, active);
1315
1316         if (!btrfs_dev_is_sequential(device, info->physical)) {
1317                 info->alloc_offset = WP_CONVENTIONAL;
1318                 return 0;
1319         }
1320
1321         /* This zone will be used for allocation, so mark this zone non-empty. */
1322         btrfs_dev_clear_zone_empty(device, info->physical);
1323
1324         down_read(&dev_replace->rwsem);
1325         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
1326         if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL)
1327                 btrfs_dev_clear_zone_empty(dev_replace->tgtdev, info->physical);
1328         up_read(&dev_replace->rwsem);
1329
1330         /*
1331          * The group is mapped to a sequential zone. Get the zone write pointer
1332          * to determine the allocation offset within the zone.
1333          */
1334         WARN_ON(!IS_ALIGNED(info->physical, fs_info->zone_size));
1335         nofs_flag = memalloc_nofs_save();
1336         ret = btrfs_get_dev_zone(device, info->physical, &zone);
1337         memalloc_nofs_restore(nofs_flag);
1338         if (ret) {
1339                 if (ret != -EIO && ret != -EOPNOTSUPP)
1340                         return ret;
1341                 info->alloc_offset = WP_MISSING_DEV;
1342                 return 0;
1343         }
1344
1345         if (zone.type == BLK_ZONE_TYPE_CONVENTIONAL) {
1346                 btrfs_err_in_rcu(fs_info,
1347                 "zoned: unexpected conventional zone %llu on device %s (devid %llu)",
1348                         zone.start << SECTOR_SHIFT, rcu_str_deref(device->name),
1349                         device->devid);
1350                 return -EIO;
1351         }
1352
1353         info->capacity = (zone.capacity << SECTOR_SHIFT);
1354
1355         switch (zone.cond) {
1356         case BLK_ZONE_COND_OFFLINE:
1357         case BLK_ZONE_COND_READONLY:
1358                 btrfs_err(fs_info,
1359                 "zoned: offline/readonly zone %llu on device %s (devid %llu)",
1360                           (info->physical >> device->zone_info->zone_size_shift),
1361                           rcu_str_deref(device->name), device->devid);
1362                 info->alloc_offset = WP_MISSING_DEV;
1363                 break;
1364         case BLK_ZONE_COND_EMPTY:
1365                 info->alloc_offset = 0;
1366                 break;
1367         case BLK_ZONE_COND_FULL:
1368                 info->alloc_offset = info->capacity;
1369                 break;
1370         default:
1371                 /* Partially used zone. */
1372                 info->alloc_offset = ((zone.wp - zone.start) << SECTOR_SHIFT);
1373                 __set_bit(zone_idx, active);
1374                 break;
1375         }
1376
1377         return 0;
1378 }
1379
1380 static int btrfs_load_block_group_single(struct btrfs_block_group *bg,
1381                                          struct zone_info *info,
1382                                          unsigned long *active)
1383 {
1384         if (info->alloc_offset == WP_MISSING_DEV) {
1385                 btrfs_err(bg->fs_info,
1386                         "zoned: cannot recover write pointer for zone %llu",
1387                         info->physical);
1388                 return -EIO;
1389         }
1390
1391         bg->alloc_offset = info->alloc_offset;
1392         bg->zone_capacity = info->capacity;
1393         if (test_bit(0, active))
1394                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1395         return 0;
1396 }
1397
1398 static int btrfs_load_block_group_dup(struct btrfs_block_group *bg,
1399                                       struct btrfs_chunk_map *map,
1400                                       struct zone_info *zone_info,
1401                                       unsigned long *active)
1402 {
1403         struct btrfs_fs_info *fs_info = bg->fs_info;
1404
1405         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1406                 btrfs_err(fs_info, "zoned: data DUP profile needs raid-stripe-tree");
1407                 return -EINVAL;
1408         }
1409
1410         if (zone_info[0].alloc_offset == WP_MISSING_DEV) {
1411                 btrfs_err(bg->fs_info,
1412                           "zoned: cannot recover write pointer for zone %llu",
1413                           zone_info[0].physical);
1414                 return -EIO;
1415         }
1416         if (zone_info[1].alloc_offset == WP_MISSING_DEV) {
1417                 btrfs_err(bg->fs_info,
1418                           "zoned: cannot recover write pointer for zone %llu",
1419                           zone_info[1].physical);
1420                 return -EIO;
1421         }
1422         if (zone_info[0].alloc_offset != zone_info[1].alloc_offset) {
1423                 btrfs_err(bg->fs_info,
1424                           "zoned: write pointer offset mismatch of zones in DUP profile");
1425                 return -EIO;
1426         }
1427
1428         if (test_bit(0, active) != test_bit(1, active)) {
1429                 if (!btrfs_zone_activate(bg))
1430                         return -EIO;
1431         } else if (test_bit(0, active)) {
1432                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1433         }
1434
1435         bg->alloc_offset = zone_info[0].alloc_offset;
1436         bg->zone_capacity = min(zone_info[0].capacity, zone_info[1].capacity);
1437         return 0;
1438 }
1439
1440 static int btrfs_load_block_group_raid1(struct btrfs_block_group *bg,
1441                                         struct btrfs_chunk_map *map,
1442                                         struct zone_info *zone_info,
1443                                         unsigned long *active)
1444 {
1445         struct btrfs_fs_info *fs_info = bg->fs_info;
1446         int i;
1447
1448         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1449                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1450                           btrfs_bg_type_to_raid_name(map->type));
1451                 return -EINVAL;
1452         }
1453
1454         for (i = 0; i < map->num_stripes; i++) {
1455                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1456                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1457                         continue;
1458
1459                 if ((zone_info[0].alloc_offset != zone_info[i].alloc_offset) &&
1460                     !btrfs_test_opt(fs_info, DEGRADED)) {
1461                         btrfs_err(fs_info,
1462                         "zoned: write pointer offset mismatch of zones in %s profile",
1463                                   btrfs_bg_type_to_raid_name(map->type));
1464                         return -EIO;
1465                 }
1466                 if (test_bit(0, active) != test_bit(i, active)) {
1467                         if (!btrfs_test_opt(fs_info, DEGRADED) &&
1468                             !btrfs_zone_activate(bg)) {
1469                                 return -EIO;
1470                         }
1471                 } else {
1472                         if (test_bit(0, active))
1473                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1474                 }
1475                 /* In case a device is missing we have a cap of 0, so don't use it. */
1476                 bg->zone_capacity = min_not_zero(zone_info[0].capacity,
1477                                                  zone_info[1].capacity);
1478         }
1479
1480         if (zone_info[0].alloc_offset != WP_MISSING_DEV)
1481                 bg->alloc_offset = zone_info[0].alloc_offset;
1482         else
1483                 bg->alloc_offset = zone_info[i - 1].alloc_offset;
1484
1485         return 0;
1486 }
1487
1488 static int btrfs_load_block_group_raid0(struct btrfs_block_group *bg,
1489                                         struct btrfs_chunk_map *map,
1490                                         struct zone_info *zone_info,
1491                                         unsigned long *active)
1492 {
1493         struct btrfs_fs_info *fs_info = bg->fs_info;
1494
1495         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1496                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1497                           btrfs_bg_type_to_raid_name(map->type));
1498                 return -EINVAL;
1499         }
1500
1501         for (int i = 0; i < map->num_stripes; i++) {
1502                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1503                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1504                         continue;
1505
1506                 if (test_bit(0, active) != test_bit(i, active)) {
1507                         if (!btrfs_zone_activate(bg))
1508                                 return -EIO;
1509                 } else {
1510                         if (test_bit(0, active))
1511                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1512                 }
1513                 bg->zone_capacity += zone_info[i].capacity;
1514                 bg->alloc_offset += zone_info[i].alloc_offset;
1515         }
1516
1517         return 0;
1518 }
1519
1520 static int btrfs_load_block_group_raid10(struct btrfs_block_group *bg,
1521                                          struct btrfs_chunk_map *map,
1522                                          struct zone_info *zone_info,
1523                                          unsigned long *active)
1524 {
1525         struct btrfs_fs_info *fs_info = bg->fs_info;
1526
1527         if ((map->type & BTRFS_BLOCK_GROUP_DATA) && !fs_info->stripe_root) {
1528                 btrfs_err(fs_info, "zoned: data %s needs raid-stripe-tree",
1529                           btrfs_bg_type_to_raid_name(map->type));
1530                 return -EINVAL;
1531         }
1532
1533         for (int i = 0; i < map->num_stripes; i++) {
1534                 if (zone_info[i].alloc_offset == WP_MISSING_DEV ||
1535                     zone_info[i].alloc_offset == WP_CONVENTIONAL)
1536                         continue;
1537
1538                 if (test_bit(0, active) != test_bit(i, active)) {
1539                         if (!btrfs_zone_activate(bg))
1540                                 return -EIO;
1541                 } else {
1542                         if (test_bit(0, active))
1543                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &bg->runtime_flags);
1544                 }
1545
1546                 if ((i % map->sub_stripes) == 0) {
1547                         bg->zone_capacity += zone_info[i].capacity;
1548                         bg->alloc_offset += zone_info[i].alloc_offset;
1549                 }
1550         }
1551
1552         return 0;
1553 }
1554
1555 int btrfs_load_block_group_zone_info(struct btrfs_block_group *cache, bool new)
1556 {
1557         struct btrfs_fs_info *fs_info = cache->fs_info;
1558         struct btrfs_chunk_map *map;
1559         u64 logical = cache->start;
1560         u64 length = cache->length;
1561         struct zone_info *zone_info = NULL;
1562         int ret;
1563         int i;
1564         unsigned long *active = NULL;
1565         u64 last_alloc = 0;
1566         u32 num_sequential = 0, num_conventional = 0;
1567
1568         if (!btrfs_is_zoned(fs_info))
1569                 return 0;
1570
1571         /* Sanity check */
1572         if (!IS_ALIGNED(length, fs_info->zone_size)) {
1573                 btrfs_err(fs_info,
1574                 "zoned: block group %llu len %llu unaligned to zone size %llu",
1575                           logical, length, fs_info->zone_size);
1576                 return -EIO;
1577         }
1578
1579         map = btrfs_find_chunk_map(fs_info, logical, length);
1580         if (!map)
1581                 return -EINVAL;
1582
1583         cache->physical_map = btrfs_clone_chunk_map(map, GFP_NOFS);
1584         if (!cache->physical_map) {
1585                 ret = -ENOMEM;
1586                 goto out;
1587         }
1588
1589         zone_info = kcalloc(map->num_stripes, sizeof(*zone_info), GFP_NOFS);
1590         if (!zone_info) {
1591                 ret = -ENOMEM;
1592                 goto out;
1593         }
1594
1595         active = bitmap_zalloc(map->num_stripes, GFP_NOFS);
1596         if (!active) {
1597                 ret = -ENOMEM;
1598                 goto out;
1599         }
1600
1601         for (i = 0; i < map->num_stripes; i++) {
1602                 ret = btrfs_load_zone_info(fs_info, i, &zone_info[i], active, map);
1603                 if (ret)
1604                         goto out;
1605
1606                 if (zone_info[i].alloc_offset == WP_CONVENTIONAL)
1607                         num_conventional++;
1608                 else
1609                         num_sequential++;
1610         }
1611
1612         if (num_sequential > 0)
1613                 set_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1614
1615         if (num_conventional > 0) {
1616                 /* Zone capacity is always zone size in emulation */
1617                 cache->zone_capacity = cache->length;
1618                 ret = calculate_alloc_pointer(cache, &last_alloc, new);
1619                 if (ret) {
1620                         btrfs_err(fs_info,
1621                         "zoned: failed to determine allocation offset of bg %llu",
1622                                   cache->start);
1623                         goto out;
1624                 } else if (map->num_stripes == num_conventional) {
1625                         cache->alloc_offset = last_alloc;
1626                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1627                         goto out;
1628                 }
1629         }
1630
1631         switch (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
1632         case 0: /* single */
1633                 ret = btrfs_load_block_group_single(cache, &zone_info[0], active);
1634                 break;
1635         case BTRFS_BLOCK_GROUP_DUP:
1636                 ret = btrfs_load_block_group_dup(cache, map, zone_info, active);
1637                 break;
1638         case BTRFS_BLOCK_GROUP_RAID1:
1639         case BTRFS_BLOCK_GROUP_RAID1C3:
1640         case BTRFS_BLOCK_GROUP_RAID1C4:
1641                 ret = btrfs_load_block_group_raid1(cache, map, zone_info, active);
1642                 break;
1643         case BTRFS_BLOCK_GROUP_RAID0:
1644                 ret = btrfs_load_block_group_raid0(cache, map, zone_info, active);
1645                 break;
1646         case BTRFS_BLOCK_GROUP_RAID10:
1647                 ret = btrfs_load_block_group_raid10(cache, map, zone_info, active);
1648                 break;
1649         case BTRFS_BLOCK_GROUP_RAID5:
1650         case BTRFS_BLOCK_GROUP_RAID6:
1651         default:
1652                 btrfs_err(fs_info, "zoned: profile %s not yet supported",
1653                           btrfs_bg_type_to_raid_name(map->type));
1654                 ret = -EINVAL;
1655                 goto out;
1656         }
1657
1658 out:
1659         if (cache->alloc_offset > cache->zone_capacity) {
1660                 btrfs_err(fs_info,
1661 "zoned: invalid write pointer %llu (larger than zone capacity %llu) in block group %llu",
1662                           cache->alloc_offset, cache->zone_capacity,
1663                           cache->start);
1664                 ret = -EIO;
1665         }
1666
1667         /* An extent is allocated after the write pointer */
1668         if (!ret && num_conventional && last_alloc > cache->alloc_offset) {
1669                 btrfs_err(fs_info,
1670                           "zoned: got wrong write pointer in BG %llu: %llu > %llu",
1671                           logical, last_alloc, cache->alloc_offset);
1672                 ret = -EIO;
1673         }
1674
1675         if (!ret) {
1676                 cache->meta_write_pointer = cache->alloc_offset + cache->start;
1677                 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags)) {
1678                         btrfs_get_block_group(cache);
1679                         spin_lock(&fs_info->zone_active_bgs_lock);
1680                         list_add_tail(&cache->active_bg_list,
1681                                       &fs_info->zone_active_bgs);
1682                         spin_unlock(&fs_info->zone_active_bgs_lock);
1683                 }
1684         } else {
1685                 btrfs_free_chunk_map(cache->physical_map);
1686                 cache->physical_map = NULL;
1687         }
1688         bitmap_free(active);
1689         kfree(zone_info);
1690
1691         return ret;
1692 }
1693
1694 void btrfs_calc_zone_unusable(struct btrfs_block_group *cache)
1695 {
1696         u64 unusable, free;
1697
1698         if (!btrfs_is_zoned(cache->fs_info))
1699                 return;
1700
1701         WARN_ON(cache->bytes_super != 0);
1702         unusable = (cache->alloc_offset - cache->used) +
1703                    (cache->length - cache->zone_capacity);
1704         free = cache->zone_capacity - cache->alloc_offset;
1705
1706         /* We only need ->free_space in ALLOC_SEQ block groups */
1707         cache->cached = BTRFS_CACHE_FINISHED;
1708         cache->free_space_ctl->free_space = free;
1709         cache->zone_unusable = unusable;
1710 }
1711
1712 bool btrfs_use_zone_append(struct btrfs_bio *bbio)
1713 {
1714         u64 start = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT);
1715         struct btrfs_inode *inode = bbio->inode;
1716         struct btrfs_fs_info *fs_info = bbio->fs_info;
1717         struct btrfs_block_group *cache;
1718         bool ret = false;
1719
1720         if (!btrfs_is_zoned(fs_info))
1721                 return false;
1722
1723         if (!inode || !is_data_inode(&inode->vfs_inode))
1724                 return false;
1725
1726         if (btrfs_op(&bbio->bio) != BTRFS_MAP_WRITE)
1727                 return false;
1728
1729         /*
1730          * Using REQ_OP_ZONE_APPNED for relocation can break assumptions on the
1731          * extent layout the relocation code has.
1732          * Furthermore we have set aside own block-group from which only the
1733          * relocation "process" can allocate and make sure only one process at a
1734          * time can add pages to an extent that gets relocated, so it's safe to
1735          * use regular REQ_OP_WRITE for this special case.
1736          */
1737         if (btrfs_is_data_reloc_root(inode->root))
1738                 return false;
1739
1740         cache = btrfs_lookup_block_group(fs_info, start);
1741         ASSERT(cache);
1742         if (!cache)
1743                 return false;
1744
1745         ret = !!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1746         btrfs_put_block_group(cache);
1747
1748         return ret;
1749 }
1750
1751 void btrfs_record_physical_zoned(struct btrfs_bio *bbio)
1752 {
1753         const u64 physical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
1754         struct btrfs_ordered_sum *sum = bbio->sums;
1755
1756         if (physical < bbio->orig_physical)
1757                 sum->logical -= bbio->orig_physical - physical;
1758         else
1759                 sum->logical += physical - bbio->orig_physical;
1760 }
1761
1762 static void btrfs_rewrite_logical_zoned(struct btrfs_ordered_extent *ordered,
1763                                         u64 logical)
1764 {
1765         struct extent_map_tree *em_tree = &BTRFS_I(ordered->inode)->extent_tree;
1766         struct extent_map *em;
1767
1768         ordered->disk_bytenr = logical;
1769
1770         write_lock(&em_tree->lock);
1771         em = search_extent_mapping(em_tree, ordered->file_offset,
1772                                    ordered->num_bytes);
1773         em->block_start = logical;
1774         free_extent_map(em);
1775         write_unlock(&em_tree->lock);
1776 }
1777
1778 static bool btrfs_zoned_split_ordered(struct btrfs_ordered_extent *ordered,
1779                                       u64 logical, u64 len)
1780 {
1781         struct btrfs_ordered_extent *new;
1782
1783         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags) &&
1784             split_extent_map(BTRFS_I(ordered->inode), ordered->file_offset,
1785                              ordered->num_bytes, len, logical))
1786                 return false;
1787
1788         new = btrfs_split_ordered_extent(ordered, len);
1789         if (IS_ERR(new))
1790                 return false;
1791         new->disk_bytenr = logical;
1792         btrfs_finish_one_ordered(new);
1793         return true;
1794 }
1795
1796 void btrfs_finish_ordered_zoned(struct btrfs_ordered_extent *ordered)
1797 {
1798         struct btrfs_inode *inode = BTRFS_I(ordered->inode);
1799         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1800         struct btrfs_ordered_sum *sum;
1801         u64 logical, len;
1802
1803         /*
1804          * Write to pre-allocated region is for the data relocation, and so
1805          * it should use WRITE operation. No split/rewrite are necessary.
1806          */
1807         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags))
1808                 return;
1809
1810         ASSERT(!list_empty(&ordered->list));
1811         /* The ordered->list can be empty in the above pre-alloc case. */
1812         sum = list_first_entry(&ordered->list, struct btrfs_ordered_sum, list);
1813         logical = sum->logical;
1814         len = sum->len;
1815
1816         while (len < ordered->disk_num_bytes) {
1817                 sum = list_next_entry(sum, list);
1818                 if (sum->logical == logical + len) {
1819                         len += sum->len;
1820                         continue;
1821                 }
1822                 if (!btrfs_zoned_split_ordered(ordered, logical, len)) {
1823                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
1824                         btrfs_err(fs_info, "failed to split ordered extent");
1825                         goto out;
1826                 }
1827                 logical = sum->logical;
1828                 len = sum->len;
1829         }
1830
1831         if (ordered->disk_bytenr != logical)
1832                 btrfs_rewrite_logical_zoned(ordered, logical);
1833
1834 out:
1835         /*
1836          * If we end up here for nodatasum I/O, the btrfs_ordered_sum structures
1837          * were allocated by btrfs_alloc_dummy_sum only to record the logical
1838          * addresses and don't contain actual checksums.  We thus must free them
1839          * here so that we don't attempt to log the csums later.
1840          */
1841         if ((inode->flags & BTRFS_INODE_NODATASUM) ||
1842             test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state)) {
1843                 while ((sum = list_first_entry_or_null(&ordered->list,
1844                                                        typeof(*sum), list))) {
1845                         list_del(&sum->list);
1846                         kfree(sum);
1847                 }
1848         }
1849 }
1850
1851 static bool check_bg_is_active(struct btrfs_eb_write_context *ctx,
1852                                struct btrfs_block_group **active_bg)
1853 {
1854         const struct writeback_control *wbc = ctx->wbc;
1855         struct btrfs_block_group *block_group = ctx->zoned_bg;
1856         struct btrfs_fs_info *fs_info = block_group->fs_info;
1857
1858         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
1859                 return true;
1860
1861         if (fs_info->treelog_bg == block_group->start) {
1862                 if (!btrfs_zone_activate(block_group)) {
1863                         int ret_fin = btrfs_zone_finish_one_bg(fs_info);
1864
1865                         if (ret_fin != 1 || !btrfs_zone_activate(block_group))
1866                                 return false;
1867                 }
1868         } else if (*active_bg != block_group) {
1869                 struct btrfs_block_group *tgt = *active_bg;
1870
1871                 /* zoned_meta_io_lock protects fs_info->active_{meta,system}_bg. */
1872                 lockdep_assert_held(&fs_info->zoned_meta_io_lock);
1873
1874                 if (tgt) {
1875                         /*
1876                          * If there is an unsent IO left in the allocated area,
1877                          * we cannot wait for them as it may cause a deadlock.
1878                          */
1879                         if (tgt->meta_write_pointer < tgt->start + tgt->alloc_offset) {
1880                                 if (wbc->sync_mode == WB_SYNC_NONE ||
1881                                     (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync))
1882                                         return false;
1883                         }
1884
1885                         /* Pivot active metadata/system block group. */
1886                         btrfs_zoned_meta_io_unlock(fs_info);
1887                         wait_eb_writebacks(tgt);
1888                         do_zone_finish(tgt, true);
1889                         btrfs_zoned_meta_io_lock(fs_info);
1890                         if (*active_bg == tgt) {
1891                                 btrfs_put_block_group(tgt);
1892                                 *active_bg = NULL;
1893                         }
1894                 }
1895                 if (!btrfs_zone_activate(block_group))
1896                         return false;
1897                 if (*active_bg != block_group) {
1898                         ASSERT(*active_bg == NULL);
1899                         *active_bg = block_group;
1900                         btrfs_get_block_group(block_group);
1901                 }
1902         }
1903
1904         return true;
1905 }
1906
1907 /*
1908  * Check if @ctx->eb is aligned to the write pointer.
1909  *
1910  * Return:
1911  *   0:        @ctx->eb is at the write pointer. You can write it.
1912  *   -EAGAIN:  There is a hole. The caller should handle the case.
1913  *   -EBUSY:   There is a hole, but the caller can just bail out.
1914  */
1915 int btrfs_check_meta_write_pointer(struct btrfs_fs_info *fs_info,
1916                                    struct btrfs_eb_write_context *ctx)
1917 {
1918         const struct writeback_control *wbc = ctx->wbc;
1919         const struct extent_buffer *eb = ctx->eb;
1920         struct btrfs_block_group *block_group = ctx->zoned_bg;
1921
1922         if (!btrfs_is_zoned(fs_info))
1923                 return 0;
1924
1925         if (block_group) {
1926                 if (block_group->start > eb->start ||
1927                     block_group->start + block_group->length <= eb->start) {
1928                         btrfs_put_block_group(block_group);
1929                         block_group = NULL;
1930                         ctx->zoned_bg = NULL;
1931                 }
1932         }
1933
1934         if (!block_group) {
1935                 block_group = btrfs_lookup_block_group(fs_info, eb->start);
1936                 if (!block_group)
1937                         return 0;
1938                 ctx->zoned_bg = block_group;
1939         }
1940
1941         if (block_group->meta_write_pointer == eb->start) {
1942                 struct btrfs_block_group **tgt;
1943
1944                 if (!test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags))
1945                         return 0;
1946
1947                 if (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM)
1948                         tgt = &fs_info->active_system_bg;
1949                 else
1950                         tgt = &fs_info->active_meta_bg;
1951                 if (check_bg_is_active(ctx, tgt))
1952                         return 0;
1953         }
1954
1955         /*
1956          * Since we may release fs_info->zoned_meta_io_lock, someone can already
1957          * start writing this eb. In that case, we can just bail out.
1958          */
1959         if (block_group->meta_write_pointer > eb->start)
1960                 return -EBUSY;
1961
1962         /* If for_sync, this hole will be filled with trasnsaction commit. */
1963         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
1964                 return -EAGAIN;
1965         return -EBUSY;
1966 }
1967
1968 int btrfs_zoned_issue_zeroout(struct btrfs_device *device, u64 physical, u64 length)
1969 {
1970         if (!btrfs_dev_is_sequential(device, physical))
1971                 return -EOPNOTSUPP;
1972
1973         return blkdev_issue_zeroout(device->bdev, physical >> SECTOR_SHIFT,
1974                                     length >> SECTOR_SHIFT, GFP_NOFS, 0);
1975 }
1976
1977 static int read_zone_info(struct btrfs_fs_info *fs_info, u64 logical,
1978                           struct blk_zone *zone)
1979 {
1980         struct btrfs_io_context *bioc = NULL;
1981         u64 mapped_length = PAGE_SIZE;
1982         unsigned int nofs_flag;
1983         int nmirrors;
1984         int i, ret;
1985
1986         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
1987                               &mapped_length, &bioc, NULL, NULL);
1988         if (ret || !bioc || mapped_length < PAGE_SIZE) {
1989                 ret = -EIO;
1990                 goto out_put_bioc;
1991         }
1992
1993         if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1994                 ret = -EINVAL;
1995                 goto out_put_bioc;
1996         }
1997
1998         nofs_flag = memalloc_nofs_save();
1999         nmirrors = (int)bioc->num_stripes;
2000         for (i = 0; i < nmirrors; i++) {
2001                 u64 physical = bioc->stripes[i].physical;
2002                 struct btrfs_device *dev = bioc->stripes[i].dev;
2003
2004                 /* Missing device */
2005                 if (!dev->bdev)
2006                         continue;
2007
2008                 ret = btrfs_get_dev_zone(dev, physical, zone);
2009                 /* Failing device */
2010                 if (ret == -EIO || ret == -EOPNOTSUPP)
2011                         continue;
2012                 break;
2013         }
2014         memalloc_nofs_restore(nofs_flag);
2015 out_put_bioc:
2016         btrfs_put_bioc(bioc);
2017         return ret;
2018 }
2019
2020 /*
2021  * Synchronize write pointer in a zone at @physical_start on @tgt_dev, by
2022  * filling zeros between @physical_pos to a write pointer of dev-replace
2023  * source device.
2024  */
2025 int btrfs_sync_zone_write_pointer(struct btrfs_device *tgt_dev, u64 logical,
2026                                     u64 physical_start, u64 physical_pos)
2027 {
2028         struct btrfs_fs_info *fs_info = tgt_dev->fs_info;
2029         struct blk_zone zone;
2030         u64 length;
2031         u64 wp;
2032         int ret;
2033
2034         if (!btrfs_dev_is_sequential(tgt_dev, physical_pos))
2035                 return 0;
2036
2037         ret = read_zone_info(fs_info, logical, &zone);
2038         if (ret)
2039                 return ret;
2040
2041         wp = physical_start + ((zone.wp - zone.start) << SECTOR_SHIFT);
2042
2043         if (physical_pos == wp)
2044                 return 0;
2045
2046         if (physical_pos > wp)
2047                 return -EUCLEAN;
2048
2049         length = wp - physical_pos;
2050         return btrfs_zoned_issue_zeroout(tgt_dev, physical_pos, length);
2051 }
2052
2053 /*
2054  * Activate block group and underlying device zones
2055  *
2056  * @block_group: the block group to activate
2057  *
2058  * Return: true on success, false otherwise
2059  */
2060 bool btrfs_zone_activate(struct btrfs_block_group *block_group)
2061 {
2062         struct btrfs_fs_info *fs_info = block_group->fs_info;
2063         struct btrfs_chunk_map *map;
2064         struct btrfs_device *device;
2065         u64 physical;
2066         const bool is_data = (block_group->flags & BTRFS_BLOCK_GROUP_DATA);
2067         bool ret;
2068         int i;
2069
2070         if (!btrfs_is_zoned(block_group->fs_info))
2071                 return true;
2072
2073         map = block_group->physical_map;
2074
2075         spin_lock(&block_group->lock);
2076         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
2077                 ret = true;
2078                 goto out_unlock;
2079         }
2080
2081         /* No space left */
2082         if (btrfs_zoned_bg_is_full(block_group)) {
2083                 ret = false;
2084                 goto out_unlock;
2085         }
2086
2087         spin_lock(&fs_info->zone_active_bgs_lock);
2088         for (i = 0; i < map->num_stripes; i++) {
2089                 struct btrfs_zoned_device_info *zinfo;
2090                 int reserved = 0;
2091
2092                 device = map->stripes[i].dev;
2093                 physical = map->stripes[i].physical;
2094                 zinfo = device->zone_info;
2095
2096                 if (zinfo->max_active_zones == 0)
2097                         continue;
2098
2099                 if (is_data)
2100                         reserved = zinfo->reserved_active_zones;
2101                 /*
2102                  * For the data block group, leave active zones for one
2103                  * metadata block group and one system block group.
2104                  */
2105                 if (atomic_read(&zinfo->active_zones_left) <= reserved) {
2106                         ret = false;
2107                         spin_unlock(&fs_info->zone_active_bgs_lock);
2108                         goto out_unlock;
2109                 }
2110
2111                 if (!btrfs_dev_set_active_zone(device, physical)) {
2112                         /* Cannot activate the zone */
2113                         ret = false;
2114                         spin_unlock(&fs_info->zone_active_bgs_lock);
2115                         goto out_unlock;
2116                 }
2117                 if (!is_data)
2118                         zinfo->reserved_active_zones--;
2119         }
2120         spin_unlock(&fs_info->zone_active_bgs_lock);
2121
2122         /* Successfully activated all the zones */
2123         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2124         spin_unlock(&block_group->lock);
2125
2126         /* For the active block group list */
2127         btrfs_get_block_group(block_group);
2128
2129         spin_lock(&fs_info->zone_active_bgs_lock);
2130         list_add_tail(&block_group->active_bg_list, &fs_info->zone_active_bgs);
2131         spin_unlock(&fs_info->zone_active_bgs_lock);
2132
2133         return true;
2134
2135 out_unlock:
2136         spin_unlock(&block_group->lock);
2137         return ret;
2138 }
2139
2140 static void wait_eb_writebacks(struct btrfs_block_group *block_group)
2141 {
2142         struct btrfs_fs_info *fs_info = block_group->fs_info;
2143         const u64 end = block_group->start + block_group->length;
2144         struct radix_tree_iter iter;
2145         struct extent_buffer *eb;
2146         void __rcu **slot;
2147
2148         rcu_read_lock();
2149         radix_tree_for_each_slot(slot, &fs_info->buffer_radix, &iter,
2150                                  block_group->start >> fs_info->sectorsize_bits) {
2151                 eb = radix_tree_deref_slot(slot);
2152                 if (!eb)
2153                         continue;
2154                 if (radix_tree_deref_retry(eb)) {
2155                         slot = radix_tree_iter_retry(&iter);
2156                         continue;
2157                 }
2158
2159                 if (eb->start < block_group->start)
2160                         continue;
2161                 if (eb->start >= end)
2162                         break;
2163
2164                 slot = radix_tree_iter_resume(slot, &iter);
2165                 rcu_read_unlock();
2166                 wait_on_extent_buffer_writeback(eb);
2167                 rcu_read_lock();
2168         }
2169         rcu_read_unlock();
2170 }
2171
2172 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written)
2173 {
2174         struct btrfs_fs_info *fs_info = block_group->fs_info;
2175         struct btrfs_chunk_map *map;
2176         const bool is_metadata = (block_group->flags &
2177                         (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM));
2178         int ret = 0;
2179         int i;
2180
2181         spin_lock(&block_group->lock);
2182         if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
2183                 spin_unlock(&block_group->lock);
2184                 return 0;
2185         }
2186
2187         /* Check if we have unwritten allocated space */
2188         if (is_metadata &&
2189             block_group->start + block_group->alloc_offset > block_group->meta_write_pointer) {
2190                 spin_unlock(&block_group->lock);
2191                 return -EAGAIN;
2192         }
2193
2194         /*
2195          * If we are sure that the block group is full (= no more room left for
2196          * new allocation) and the IO for the last usable block is completed, we
2197          * don't need to wait for the other IOs. This holds because we ensure
2198          * the sequential IO submissions using the ZONE_APPEND command for data
2199          * and block_group->meta_write_pointer for metadata.
2200          */
2201         if (!fully_written) {
2202                 if (test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
2203                         spin_unlock(&block_group->lock);
2204                         return -EAGAIN;
2205                 }
2206                 spin_unlock(&block_group->lock);
2207
2208                 ret = btrfs_inc_block_group_ro(block_group, false);
2209                 if (ret)
2210                         return ret;
2211
2212                 /* Ensure all writes in this block group finish */
2213                 btrfs_wait_block_group_reservations(block_group);
2214                 /* No need to wait for NOCOW writers. Zoned mode does not allow that */
2215                 btrfs_wait_ordered_roots(fs_info, U64_MAX, block_group->start,
2216                                          block_group->length);
2217                 /* Wait for extent buffers to be written. */
2218                 if (is_metadata)
2219                         wait_eb_writebacks(block_group);
2220
2221                 spin_lock(&block_group->lock);
2222
2223                 /*
2224                  * Bail out if someone already deactivated the block group, or
2225                  * allocated space is left in the block group.
2226                  */
2227                 if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2228                               &block_group->runtime_flags)) {
2229                         spin_unlock(&block_group->lock);
2230                         btrfs_dec_block_group_ro(block_group);
2231                         return 0;
2232                 }
2233
2234                 if (block_group->reserved ||
2235                     test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2236                              &block_group->runtime_flags)) {
2237                         spin_unlock(&block_group->lock);
2238                         btrfs_dec_block_group_ro(block_group);
2239                         return -EAGAIN;
2240                 }
2241         }
2242
2243         clear_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2244         block_group->alloc_offset = block_group->zone_capacity;
2245         if (block_group->flags & (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM))
2246                 block_group->meta_write_pointer = block_group->start +
2247                                                   block_group->zone_capacity;
2248         block_group->free_space_ctl->free_space = 0;
2249         btrfs_clear_treelog_bg(block_group);
2250         btrfs_clear_data_reloc_bg(block_group);
2251         spin_unlock(&block_group->lock);
2252
2253         map = block_group->physical_map;
2254         for (i = 0; i < map->num_stripes; i++) {
2255                 struct btrfs_device *device = map->stripes[i].dev;
2256                 const u64 physical = map->stripes[i].physical;
2257                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2258
2259                 if (zinfo->max_active_zones == 0)
2260                         continue;
2261
2262                 ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_FINISH,
2263                                        physical >> SECTOR_SHIFT,
2264                                        zinfo->zone_size >> SECTOR_SHIFT,
2265                                        GFP_NOFS);
2266
2267                 if (ret)
2268                         return ret;
2269
2270                 if (!(block_group->flags & BTRFS_BLOCK_GROUP_DATA))
2271                         zinfo->reserved_active_zones++;
2272                 btrfs_dev_clear_active_zone(device, physical);
2273         }
2274
2275         if (!fully_written)
2276                 btrfs_dec_block_group_ro(block_group);
2277
2278         spin_lock(&fs_info->zone_active_bgs_lock);
2279         ASSERT(!list_empty(&block_group->active_bg_list));
2280         list_del_init(&block_group->active_bg_list);
2281         spin_unlock(&fs_info->zone_active_bgs_lock);
2282
2283         /* For active_bg_list */
2284         btrfs_put_block_group(block_group);
2285
2286         clear_and_wake_up_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2287
2288         return 0;
2289 }
2290
2291 int btrfs_zone_finish(struct btrfs_block_group *block_group)
2292 {
2293         if (!btrfs_is_zoned(block_group->fs_info))
2294                 return 0;
2295
2296         return do_zone_finish(block_group, false);
2297 }
2298
2299 bool btrfs_can_activate_zone(struct btrfs_fs_devices *fs_devices, u64 flags)
2300 {
2301         struct btrfs_fs_info *fs_info = fs_devices->fs_info;
2302         struct btrfs_device *device;
2303         bool ret = false;
2304
2305         if (!btrfs_is_zoned(fs_info))
2306                 return true;
2307
2308         /* Check if there is a device with active zones left */
2309         mutex_lock(&fs_info->chunk_mutex);
2310         spin_lock(&fs_info->zone_active_bgs_lock);
2311         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
2312                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2313                 int reserved = 0;
2314
2315                 if (!device->bdev)
2316                         continue;
2317
2318                 if (!zinfo->max_active_zones) {
2319                         ret = true;
2320                         break;
2321                 }
2322
2323                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2324                         reserved = zinfo->reserved_active_zones;
2325
2326                 switch (flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
2327                 case 0: /* single */
2328                         ret = (atomic_read(&zinfo->active_zones_left) >= (1 + reserved));
2329                         break;
2330                 case BTRFS_BLOCK_GROUP_DUP:
2331                         ret = (atomic_read(&zinfo->active_zones_left) >= (2 + reserved));
2332                         break;
2333                 }
2334                 if (ret)
2335                         break;
2336         }
2337         spin_unlock(&fs_info->zone_active_bgs_lock);
2338         mutex_unlock(&fs_info->chunk_mutex);
2339
2340         if (!ret)
2341                 set_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2342
2343         return ret;
2344 }
2345
2346 void btrfs_zone_finish_endio(struct btrfs_fs_info *fs_info, u64 logical, u64 length)
2347 {
2348         struct btrfs_block_group *block_group;
2349         u64 min_alloc_bytes;
2350
2351         if (!btrfs_is_zoned(fs_info))
2352                 return;
2353
2354         block_group = btrfs_lookup_block_group(fs_info, logical);
2355         ASSERT(block_group);
2356
2357         /* No MIXED_BG on zoned btrfs. */
2358         if (block_group->flags & BTRFS_BLOCK_GROUP_DATA)
2359                 min_alloc_bytes = fs_info->sectorsize;
2360         else
2361                 min_alloc_bytes = fs_info->nodesize;
2362
2363         /* Bail out if we can allocate more data from this block group. */
2364         if (logical + length + min_alloc_bytes <=
2365             block_group->start + block_group->zone_capacity)
2366                 goto out;
2367
2368         do_zone_finish(block_group, true);
2369
2370 out:
2371         btrfs_put_block_group(block_group);
2372 }
2373
2374 static void btrfs_zone_finish_endio_workfn(struct work_struct *work)
2375 {
2376         struct btrfs_block_group *bg =
2377                 container_of(work, struct btrfs_block_group, zone_finish_work);
2378
2379         wait_on_extent_buffer_writeback(bg->last_eb);
2380         free_extent_buffer(bg->last_eb);
2381         btrfs_zone_finish_endio(bg->fs_info, bg->start, bg->length);
2382         btrfs_put_block_group(bg);
2383 }
2384
2385 void btrfs_schedule_zone_finish_bg(struct btrfs_block_group *bg,
2386                                    struct extent_buffer *eb)
2387 {
2388         if (!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &bg->runtime_flags) ||
2389             eb->start + eb->len * 2 <= bg->start + bg->zone_capacity)
2390                 return;
2391
2392         if (WARN_ON(bg->zone_finish_work.func == btrfs_zone_finish_endio_workfn)) {
2393                 btrfs_err(bg->fs_info, "double scheduling of bg %llu zone finishing",
2394                           bg->start);
2395                 return;
2396         }
2397
2398         /* For the work */
2399         btrfs_get_block_group(bg);
2400         atomic_inc(&eb->refs);
2401         bg->last_eb = eb;
2402         INIT_WORK(&bg->zone_finish_work, btrfs_zone_finish_endio_workfn);
2403         queue_work(system_unbound_wq, &bg->zone_finish_work);
2404 }
2405
2406 void btrfs_clear_data_reloc_bg(struct btrfs_block_group *bg)
2407 {
2408         struct btrfs_fs_info *fs_info = bg->fs_info;
2409
2410         spin_lock(&fs_info->relocation_bg_lock);
2411         if (fs_info->data_reloc_bg == bg->start)
2412                 fs_info->data_reloc_bg = 0;
2413         spin_unlock(&fs_info->relocation_bg_lock);
2414 }
2415
2416 void btrfs_free_zone_cache(struct btrfs_fs_info *fs_info)
2417 {
2418         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2419         struct btrfs_device *device;
2420
2421         if (!btrfs_is_zoned(fs_info))
2422                 return;
2423
2424         mutex_lock(&fs_devices->device_list_mutex);
2425         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2426                 if (device->zone_info) {
2427                         vfree(device->zone_info->zone_cache);
2428                         device->zone_info->zone_cache = NULL;
2429                 }
2430         }
2431         mutex_unlock(&fs_devices->device_list_mutex);
2432 }
2433
2434 bool btrfs_zoned_should_reclaim(struct btrfs_fs_info *fs_info)
2435 {
2436         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2437         struct btrfs_device *device;
2438         u64 used = 0;
2439         u64 total = 0;
2440         u64 factor;
2441
2442         ASSERT(btrfs_is_zoned(fs_info));
2443
2444         if (fs_info->bg_reclaim_threshold == 0)
2445                 return false;
2446
2447         mutex_lock(&fs_devices->device_list_mutex);
2448         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2449                 if (!device->bdev)
2450                         continue;
2451
2452                 total += device->disk_total_bytes;
2453                 used += device->bytes_used;
2454         }
2455         mutex_unlock(&fs_devices->device_list_mutex);
2456
2457         factor = div64_u64(used * 100, total);
2458         return factor >= fs_info->bg_reclaim_threshold;
2459 }
2460
2461 void btrfs_zoned_release_data_reloc_bg(struct btrfs_fs_info *fs_info, u64 logical,
2462                                        u64 length)
2463 {
2464         struct btrfs_block_group *block_group;
2465
2466         if (!btrfs_is_zoned(fs_info))
2467                 return;
2468
2469         block_group = btrfs_lookup_block_group(fs_info, logical);
2470         /* It should be called on a previous data relocation block group. */
2471         ASSERT(block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA));
2472
2473         spin_lock(&block_group->lock);
2474         if (!test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))
2475                 goto out;
2476
2477         /* All relocation extents are written. */
2478         if (block_group->start + block_group->alloc_offset == logical + length) {
2479                 /*
2480                  * Now, release this block group for further allocations and
2481                  * zone finish.
2482                  */
2483                 clear_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2484                           &block_group->runtime_flags);
2485         }
2486
2487 out:
2488         spin_unlock(&block_group->lock);
2489         btrfs_put_block_group(block_group);
2490 }
2491
2492 int btrfs_zone_finish_one_bg(struct btrfs_fs_info *fs_info)
2493 {
2494         struct btrfs_block_group *block_group;
2495         struct btrfs_block_group *min_bg = NULL;
2496         u64 min_avail = U64_MAX;
2497         int ret;
2498
2499         spin_lock(&fs_info->zone_active_bgs_lock);
2500         list_for_each_entry(block_group, &fs_info->zone_active_bgs,
2501                             active_bg_list) {
2502                 u64 avail;
2503
2504                 spin_lock(&block_group->lock);
2505                 if (block_group->reserved || block_group->alloc_offset == 0 ||
2506                     (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM) ||
2507                     test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags)) {
2508                         spin_unlock(&block_group->lock);
2509                         continue;
2510                 }
2511
2512                 avail = block_group->zone_capacity - block_group->alloc_offset;
2513                 if (min_avail > avail) {
2514                         if (min_bg)
2515                                 btrfs_put_block_group(min_bg);
2516                         min_bg = block_group;
2517                         min_avail = avail;
2518                         btrfs_get_block_group(min_bg);
2519                 }
2520                 spin_unlock(&block_group->lock);
2521         }
2522         spin_unlock(&fs_info->zone_active_bgs_lock);
2523
2524         if (!min_bg)
2525                 return 0;
2526
2527         ret = btrfs_zone_finish(min_bg);
2528         btrfs_put_block_group(min_bg);
2529
2530         return ret < 0 ? ret : 1;
2531 }
2532
2533 int btrfs_zoned_activate_one_bg(struct btrfs_fs_info *fs_info,
2534                                 struct btrfs_space_info *space_info,
2535                                 bool do_finish)
2536 {
2537         struct btrfs_block_group *bg;
2538         int index;
2539
2540         if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
2541                 return 0;
2542
2543         for (;;) {
2544                 int ret;
2545                 bool need_finish = false;
2546
2547                 down_read(&space_info->groups_sem);
2548                 for (index = 0; index < BTRFS_NR_RAID_TYPES; index++) {
2549                         list_for_each_entry(bg, &space_info->block_groups[index],
2550                                             list) {
2551                                 if (!spin_trylock(&bg->lock))
2552                                         continue;
2553                                 if (btrfs_zoned_bg_is_full(bg) ||
2554                                     test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2555                                              &bg->runtime_flags)) {
2556                                         spin_unlock(&bg->lock);
2557                                         continue;
2558                                 }
2559                                 spin_unlock(&bg->lock);
2560
2561                                 if (btrfs_zone_activate(bg)) {
2562                                         up_read(&space_info->groups_sem);
2563                                         return 1;
2564                                 }
2565
2566                                 need_finish = true;
2567                         }
2568                 }
2569                 up_read(&space_info->groups_sem);
2570
2571                 if (!do_finish || !need_finish)
2572                         break;
2573
2574                 ret = btrfs_zone_finish_one_bg(fs_info);
2575                 if (ret == 0)
2576                         break;
2577                 if (ret < 0)
2578                         return ret;
2579         }
2580
2581         return 0;
2582 }
2583
2584 /*
2585  * Reserve zones for one metadata block group, one tree-log block group, and one
2586  * system block group.
2587  */
2588 void btrfs_check_active_zone_reservation(struct btrfs_fs_info *fs_info)
2589 {
2590         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2591         struct btrfs_block_group *block_group;
2592         struct btrfs_device *device;
2593         /* Reserve zones for normal SINGLE metadata and tree-log block group. */
2594         unsigned int metadata_reserve = 2;
2595         /* Reserve a zone for SINGLE system block group. */
2596         unsigned int system_reserve = 1;
2597
2598         if (!test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags))
2599                 return;
2600
2601         /*
2602          * This function is called from the mount context. So, there is no
2603          * parallel process touching the bits. No need for read_seqretry().
2604          */
2605         if (fs_info->avail_metadata_alloc_bits & BTRFS_BLOCK_GROUP_DUP)
2606                 metadata_reserve = 4;
2607         if (fs_info->avail_system_alloc_bits & BTRFS_BLOCK_GROUP_DUP)
2608                 system_reserve = 2;
2609
2610         /* Apply the reservation on all the devices. */
2611         mutex_lock(&fs_devices->device_list_mutex);
2612         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2613                 if (!device->bdev)
2614                         continue;
2615
2616                 device->zone_info->reserved_active_zones =
2617                         metadata_reserve + system_reserve;
2618         }
2619         mutex_unlock(&fs_devices->device_list_mutex);
2620
2621         /* Release reservation for currently active block groups. */
2622         spin_lock(&fs_info->zone_active_bgs_lock);
2623         list_for_each_entry(block_group, &fs_info->zone_active_bgs, active_bg_list) {
2624                 struct btrfs_chunk_map *map = block_group->physical_map;
2625
2626                 if (!(block_group->flags &
2627                       (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM)))
2628                         continue;
2629
2630                 for (int i = 0; i < map->num_stripes; i++)
2631                         map->stripes[i].dev->zone_info->reserved_active_zones--;
2632         }
2633         spin_unlock(&fs_info->zone_active_bgs_lock);
2634 }