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