fs/btrfs/zoned.c

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