fs/btrfs/zoned.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/bitops.h>
   4 #include <linux/slab.h>
   5 #include <linux/blkdev.h>
   6 #include <linux/sched/mm.h>
   7 #include <linux/atomic.h>
   8 #include <linux/vmalloc.h>
   9 #include "ctree.h"
  10 #include "volumes.h"
  11 #include "zoned.h"
  12 #include "rcu-string.h"
  13 #include "disk-io.h"
  14 #include "block-group.h"
  15 #include "transaction.h"
  16 #include "dev-replace.h"
  17 #include "space-info.h"
  18 #include "super.h"
  19 #include "fs.h"
  20 #include "accessors.h"
  21 #include "bio.h"
  22
  23 /* Maximum number of zones to report per blkdev_report_zones() call */
  24 #define BTRFS_REPORT_NR_ZONES   4096
  25 /* Invalid allocation pointer value for missing devices */
  26 #define WP_MISSING_DEV ((u64)-1)
  27 /* Pseudo write pointer value for conventional zone */
  28 #define WP_CONVENTIONAL ((u64)-2)
  29
  30 /*
  31  * Location of the first zone of superblock logging zone pairs.
  32  *
  33  * - primary superblock:    0B (zone 0)
  34  * - first copy:          512G (zone starting at that offset)
  35  * - second copy:           4T (zone starting at that offset)
  36  */
  37 #define BTRFS_SB_LOG_PRIMARY_OFFSET     (0ULL)
  38 #define BTRFS_SB_LOG_FIRST_OFFSET       (512ULL * SZ_1G)
  39 #define BTRFS_SB_LOG_SECOND_OFFSET      (4096ULL * SZ_1G)
  40
  41 #define BTRFS_SB_LOG_FIRST_SHIFT        const_ilog2(BTRFS_SB_LOG_FIRST_OFFSET)
  42 #define BTRFS_SB_LOG_SECOND_SHIFT       const_ilog2(BTRFS_SB_LOG_SECOND_OFFSET)
  43
  44 /* Number of superblock log zones */
  45 #define BTRFS_NR_SB_LOG_ZONES 2
  46
  47 /*
  48  * Minimum of active zones we need:
  49  *
  50  * - BTRFS_SUPER_MIRROR_MAX zones for superblock mirrors
  51  * - 3 zones to ensure at least one zone per SYSTEM, META and DATA block group
  52  * - 1 zone for tree-log dedicated block group
  53  * - 1 zone for relocation
  54  */
  55 #define BTRFS_MIN_ACTIVE_ZONES          (BTRFS_SUPER_MIRROR_MAX + 5)
  56
  57 /*
  58  * Minimum / maximum supported zone size. Currently, SMR disks have a zone
  59  * size of 256MiB, and we are expecting ZNS drives to be in the 1-4GiB range.
  60  * We do not expect the zone size to become larger than 8GiB or smaller than
  61  * 4MiB in the near future.
  62  */
  63 #define BTRFS_MAX_ZONE_SIZE             SZ_8G
  64 #define BTRFS_MIN_ZONE_SIZE             SZ_4M
  65
  66 #define SUPER_INFO_SECTORS      ((u64)BTRFS_SUPER_INFO_SIZE >> SECTOR_SHIFT)
  67
  68 static inline bool sb_zone_is_full(const struct blk_zone *zone)
  69 {
  70         return (zone->cond == BLK_ZONE_COND_FULL) ||
  71                 (zone->wp + SUPER_INFO_SECTORS > zone->start + zone->capacity);
  72 }
  73
  74 static int copy_zone_info_cb(struct blk_zone *zone, unsigned int idx, void *data)
  75 {
  76         struct blk_zone *zones = data;
  77
  78         memcpy(&zones[idx], zone, sizeof(*zone));
  79
  80         return 0;
  81 }
  82
  83 static int sb_write_pointer(struct block_device *bdev, struct blk_zone *zones,
  84                             u64 *wp_ret)
  85 {
  86         bool empty[BTRFS_NR_SB_LOG_ZONES];
  87         bool full[BTRFS_NR_SB_LOG_ZONES];
  88         sector_t sector;
  89         int i;
  90
  91         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
  92                 ASSERT(zones[i].type != BLK_ZONE_TYPE_CONVENTIONAL);
  93                 empty[i] = (zones[i].cond == BLK_ZONE_COND_EMPTY);
  94                 full[i] = sb_zone_is_full(&zones[i]);
  95         }
  96
  97         /*
  98          * Possible states of log buffer zones
  99          *
 100          *           Empty[0]  In use[0]  Full[0]
 101          * Empty[1]         *          0        1
 102          * In use[1]        x          x        1
 103          * Full[1]          0          0        C
 104          *
 105          * Log position:
 106          *   *: Special case, no superblock is written
 107          *   0: Use write pointer of zones[0]
 108          *   1: Use write pointer of zones[1]
 109          *   C: Compare super blocks from zones[0] and zones[1], use the latest
 110          *      one determined by generation
 111          *   x: Invalid state
 112          */
 113
 114         if (empty[0] && empty[1]) {
 115                 /* Special case to distinguish no superblock to read */
 116                 *wp_ret = zones[0].start << SECTOR_SHIFT;
 117                 return -ENOENT;
 118         } else if (full[0] && full[1]) {
 119                 /* Compare two super blocks */
 120                 struct address_space *mapping = bdev->bd_inode->i_mapping;
 121                 struct page *page[BTRFS_NR_SB_LOG_ZONES];
 122                 struct btrfs_super_block *super[BTRFS_NR_SB_LOG_ZONES];
 123                 int i;
 124
 125                 for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 126                         u64 zone_end = (zones[i].start + zones[i].capacity) << SECTOR_SHIFT;
 127                         u64 bytenr = ALIGN_DOWN(zone_end, BTRFS_SUPER_INFO_SIZE) -
 128                                                 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                 set_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags);
 528         }
 529
 530         /* Validate superblock log */
 531         nr_zones = BTRFS_NR_SB_LOG_ZONES;
 532         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 533                 u32 sb_zone;
 534                 u64 sb_wp;
 535                 int sb_pos = BTRFS_NR_SB_LOG_ZONES * i;
 536
 537                 sb_zone = sb_zone_number(zone_info->zone_size_shift, i);
 538                 if (sb_zone + 1 >= zone_info->nr_zones)
 539                         continue;
 540
 541                 ret = btrfs_get_dev_zones(device,
 542                                           zone_start_physical(sb_zone, zone_info),
 543                                           &zone_info->sb_zones[sb_pos],
 544                                           &nr_zones);
 545                 if (ret)
 546                         goto out;
 547
 548                 if (nr_zones != BTRFS_NR_SB_LOG_ZONES) {
 549                         btrfs_err_in_rcu(device->fs_info,
 550         "zoned: failed to read super block log zone info at devid %llu zone %u",
 551                                          device->devid, sb_zone);
 552                         ret = -EUCLEAN;
 553                         goto out;
 554                 }
 555
 556                 /*
 557                  * If zones[0] is conventional, always use the beginning of the
 558                  * zone to record superblock. No need to validate in that case.
 559                  */
 560                 if (zone_info->sb_zones[BTRFS_NR_SB_LOG_ZONES * i].type ==
 561                     BLK_ZONE_TYPE_CONVENTIONAL)
 562                         continue;
 563
 564                 ret = sb_write_pointer(device->bdev,
 565                                        &zone_info->sb_zones[sb_pos], &sb_wp);
 566                 if (ret != -ENOENT && ret) {
 567                         btrfs_err_in_rcu(device->fs_info,
 568                         "zoned: super block log zone corrupted devid %llu zone %u",
 569                                          device->devid, sb_zone);
 570                         ret = -EUCLEAN;
 571                         goto out;
 572                 }
 573         }
 574
 575
 576         kvfree(zones);
 577
 578         switch (bdev_zoned_model(bdev)) {
 579         case BLK_ZONED_HM:
 580                 model = "host-managed zoned";
 581                 emulated = "";
 582                 break;
 583         case BLK_ZONED_HA:
 584                 model = "host-aware zoned";
 585                 emulated = "";
 586                 break;
 587         case BLK_ZONED_NONE:
 588                 model = "regular";
 589                 emulated = "emulated ";
 590                 break;
 591         default:
 592                 /* Just in case */
 593                 btrfs_err_in_rcu(fs_info, "zoned: unsupported model %d on %s",
 594                                  bdev_zoned_model(bdev),
 595                                  rcu_str_deref(device->name));
 596                 ret = -EOPNOTSUPP;
 597                 goto out_free_zone_info;
 598         }
 599
 600         btrfs_info_in_rcu(fs_info,
 601                 "%s block device %s, %u %szones of %llu bytes",
 602                 model, rcu_str_deref(device->name), zone_info->nr_zones,
 603                 emulated, zone_info->zone_size);
 604
 605         return 0;
 606
 607 out:
 608         kvfree(zones);
 609 out_free_zone_info:
 610         btrfs_destroy_dev_zone_info(device);
 611
 612         return ret;
 613 }
 614
 615 void btrfs_destroy_dev_zone_info(struct btrfs_device *device)
 616 {
 617         struct btrfs_zoned_device_info *zone_info = device->zone_info;
 618
 619         if (!zone_info)
 620                 return;
 621
 622         bitmap_free(zone_info->active_zones);
 623         bitmap_free(zone_info->seq_zones);
 624         bitmap_free(zone_info->empty_zones);
 625         vfree(zone_info->zone_cache);
 626         kfree(zone_info);
 627         device->zone_info = NULL;
 628 }
 629
 630 struct btrfs_zoned_device_info *btrfs_clone_dev_zone_info(struct btrfs_device *orig_dev)
 631 {
 632         struct btrfs_zoned_device_info *zone_info;
 633
 634         zone_info = kmemdup(orig_dev->zone_info, sizeof(*zone_info), GFP_KERNEL);
 635         if (!zone_info)
 636                 return NULL;
 637
 638         zone_info->seq_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 639         if (!zone_info->seq_zones)
 640                 goto out;
 641
 642         bitmap_copy(zone_info->seq_zones, orig_dev->zone_info->seq_zones,
 643                     zone_info->nr_zones);
 644
 645         zone_info->empty_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 646         if (!zone_info->empty_zones)
 647                 goto out;
 648
 649         bitmap_copy(zone_info->empty_zones, orig_dev->zone_info->empty_zones,
 650                     zone_info->nr_zones);
 651
 652         zone_info->active_zones = bitmap_zalloc(zone_info->nr_zones, GFP_KERNEL);
 653         if (!zone_info->active_zones)
 654                 goto out;
 655
 656         bitmap_copy(zone_info->active_zones, orig_dev->zone_info->active_zones,
 657                     zone_info->nr_zones);
 658         zone_info->zone_cache = NULL;
 659
 660         return zone_info;
 661
 662 out:
 663         bitmap_free(zone_info->seq_zones);
 664         bitmap_free(zone_info->empty_zones);
 665         bitmap_free(zone_info->active_zones);
 666         kfree(zone_info);
 667         return NULL;
 668 }
 669
 670 int btrfs_get_dev_zone(struct btrfs_device *device, u64 pos,
 671                        struct blk_zone *zone)
 672 {
 673         unsigned int nr_zones = 1;
 674         int ret;
 675
 676         ret = btrfs_get_dev_zones(device, pos, zone, &nr_zones);
 677         if (ret != 0 || !nr_zones)
 678                 return ret ? ret : -EIO;
 679
 680         return 0;
 681 }
 682
 683 static int btrfs_check_for_zoned_device(struct btrfs_fs_info *fs_info)
 684 {
 685         struct btrfs_device *device;
 686
 687         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 688                 if (device->bdev &&
 689                     bdev_zoned_model(device->bdev) == BLK_ZONED_HM) {
 690                         btrfs_err(fs_info,
 691                                 "zoned: mode not enabled but zoned device found: %pg",
 692                                 device->bdev);
 693                         return -EINVAL;
 694                 }
 695         }
 696
 697         return 0;
 698 }
 699
 700 int btrfs_check_zoned_mode(struct btrfs_fs_info *fs_info)
 701 {
 702         struct queue_limits *lim = &fs_info->limits;
 703         struct btrfs_device *device;
 704         u64 zone_size = 0;
 705         int ret;
 706
 707         /*
 708          * Host-Managed devices can't be used without the ZONED flag.  With the
 709          * ZONED all devices can be used, using zone emulation if required.
 710          */
 711         if (!btrfs_fs_incompat(fs_info, ZONED))
 712                 return btrfs_check_for_zoned_device(fs_info);
 713
 714         blk_set_stacking_limits(lim);
 715
 716         list_for_each_entry(device, &fs_info->fs_devices->devices, dev_list) {
 717                 struct btrfs_zoned_device_info *zone_info = device->zone_info;
 718
 719                 if (!device->bdev)
 720                         continue;
 721
 722                 if (!zone_size) {
 723                         zone_size = zone_info->zone_size;
 724                 } else if (zone_info->zone_size != zone_size) {
 725                         btrfs_err(fs_info,
 726                 "zoned: unequal block device zone sizes: have %llu found %llu",
 727                                   zone_info->zone_size, zone_size);
 728                         return -EINVAL;
 729                 }
 730
 731                 /*
 732                  * With the zoned emulation, we can have non-zoned device on the
 733                  * zoned mode. In this case, we don't have a valid max zone
 734                  * append size.
 735                  */
 736                 if (bdev_is_zoned(device->bdev)) {
 737                         blk_stack_limits(lim,
 738                                          &bdev_get_queue(device->bdev)->limits,
 739                                          0);
 740                 }
 741         }
 742
 743         /*
 744          * stripe_size is always aligned to BTRFS_STRIPE_LEN in
 745          * btrfs_create_chunk(). Since we want stripe_len == zone_size,
 746          * check the alignment here.
 747          */
 748         if (!IS_ALIGNED(zone_size, BTRFS_STRIPE_LEN)) {
 749                 btrfs_err(fs_info,
 750                           "zoned: zone size %llu not aligned to stripe %u",
 751                           zone_size, BTRFS_STRIPE_LEN);
 752                 return -EINVAL;
 753         }
 754
 755         if (btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
 756                 btrfs_err(fs_info, "zoned: mixed block groups not supported");
 757                 return -EINVAL;
 758         }
 759
 760         fs_info->zone_size = zone_size;
 761         /*
 762          * Also limit max_zone_append_size by max_segments * PAGE_SIZE.
 763          * Technically, we can have multiple pages per segment. But, since
 764          * we add the pages one by one to a bio, and cannot increase the
 765          * metadata reservation even if it increases the number of extents, it
 766          * is safe to stick with the limit.
 767          */
 768         fs_info->max_zone_append_size = ALIGN_DOWN(
 769                 min3((u64)lim->max_zone_append_sectors << SECTOR_SHIFT,
 770                      (u64)lim->max_sectors << SECTOR_SHIFT,
 771                      (u64)lim->max_segments << PAGE_SHIFT),
 772                 fs_info->sectorsize);
 773         fs_info->fs_devices->chunk_alloc_policy = BTRFS_CHUNK_ALLOC_ZONED;
 774         if (fs_info->max_zone_append_size < fs_info->max_extent_size)
 775                 fs_info->max_extent_size = fs_info->max_zone_append_size;
 776
 777         /*
 778          * Check mount options here, because we might change fs_info->zoned
 779          * from fs_info->zone_size.
 780          */
 781         ret = btrfs_check_mountopts_zoned(fs_info);
 782         if (ret)
 783                 return ret;
 784
 785         btrfs_info(fs_info, "zoned mode enabled with zone size %llu", zone_size);
 786         return 0;
 787 }
 788
 789 int btrfs_check_mountopts_zoned(struct btrfs_fs_info *info)
 790 {
 791         if (!btrfs_is_zoned(info))
 792                 return 0;
 793
 794         /*
 795          * Space cache writing is not COWed. Disable that to avoid write errors
 796          * in sequential zones.
 797          */
 798         if (btrfs_test_opt(info, SPACE_CACHE)) {
 799                 btrfs_err(info, "zoned: space cache v1 is not supported");
 800                 return -EINVAL;
 801         }
 802
 803         if (btrfs_test_opt(info, NODATACOW)) {
 804                 btrfs_err(info, "zoned: NODATACOW not supported");
 805                 return -EINVAL;
 806         }
 807
 808         btrfs_clear_and_info(info, DISCARD_ASYNC,
 809                         "zoned: async discard ignored and disabled for zoned mode");
 810
 811         return 0;
 812 }
 813
 814 static int sb_log_location(struct block_device *bdev, struct blk_zone *zones,
 815                            int rw, u64 *bytenr_ret)
 816 {
 817         u64 wp;
 818         int ret;
 819
 820         if (zones[0].type == BLK_ZONE_TYPE_CONVENTIONAL) {
 821                 *bytenr_ret = zones[0].start << SECTOR_SHIFT;
 822                 return 0;
 823         }
 824
 825         ret = sb_write_pointer(bdev, zones, &wp);
 826         if (ret != -ENOENT && ret < 0)
 827                 return ret;
 828
 829         if (rw == WRITE) {
 830                 struct blk_zone *reset = NULL;
 831
 832                 if (wp == zones[0].start << SECTOR_SHIFT)
 833                         reset = &zones[0];
 834                 else if (wp == zones[1].start << SECTOR_SHIFT)
 835                         reset = &zones[1];
 836
 837                 if (reset && reset->cond != BLK_ZONE_COND_EMPTY) {
 838                         ASSERT(sb_zone_is_full(reset));
 839
 840                         ret = blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
 841                                                reset->start, reset->len,
 842                                                GFP_NOFS);
 843                         if (ret)
 844                                 return ret;
 845
 846                         reset->cond = BLK_ZONE_COND_EMPTY;
 847                         reset->wp = reset->start;
 848                 }
 849         } else if (ret != -ENOENT) {
 850                 /*
 851                  * For READ, we want the previous one. Move write pointer to
 852                  * the end of a zone, if it is at the head of a zone.
 853                  */
 854                 u64 zone_end = 0;
 855
 856                 if (wp == zones[0].start << SECTOR_SHIFT)
 857                         zone_end = zones[1].start + zones[1].capacity;
 858                 else if (wp == zones[1].start << SECTOR_SHIFT)
 859                         zone_end = zones[0].start + zones[0].capacity;
 860                 if (zone_end)
 861                         wp = ALIGN_DOWN(zone_end << SECTOR_SHIFT,
 862                                         BTRFS_SUPER_INFO_SIZE);
 863
 864                 wp -= BTRFS_SUPER_INFO_SIZE;
 865         }
 866
 867         *bytenr_ret = wp;
 868         return 0;
 869
 870 }
 871
 872 int btrfs_sb_log_location_bdev(struct block_device *bdev, int mirror, int rw,
 873                                u64 *bytenr_ret)
 874 {
 875         struct blk_zone zones[BTRFS_NR_SB_LOG_ZONES];
 876         sector_t zone_sectors;
 877         u32 sb_zone;
 878         int ret;
 879         u8 zone_sectors_shift;
 880         sector_t nr_sectors;
 881         u32 nr_zones;
 882
 883         if (!bdev_is_zoned(bdev)) {
 884                 *bytenr_ret = btrfs_sb_offset(mirror);
 885                 return 0;
 886         }
 887
 888         ASSERT(rw == READ || rw == WRITE);
 889
 890         zone_sectors = bdev_zone_sectors(bdev);
 891         if (!is_power_of_2(zone_sectors))
 892                 return -EINVAL;
 893         zone_sectors_shift = ilog2(zone_sectors);
 894         nr_sectors = bdev_nr_sectors(bdev);
 895         nr_zones = nr_sectors >> zone_sectors_shift;
 896
 897         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
 898         if (sb_zone + 1 >= nr_zones)
 899                 return -ENOENT;
 900
 901         ret = blkdev_report_zones(bdev, zone_start_sector(sb_zone, bdev),
 902                                   BTRFS_NR_SB_LOG_ZONES, copy_zone_info_cb,
 903                                   zones);
 904         if (ret < 0)
 905                 return ret;
 906         if (ret != BTRFS_NR_SB_LOG_ZONES)
 907                 return -EIO;
 908
 909         return sb_log_location(bdev, zones, rw, bytenr_ret);
 910 }
 911
 912 int btrfs_sb_log_location(struct btrfs_device *device, int mirror, int rw,
 913                           u64 *bytenr_ret)
 914 {
 915         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 916         u32 zone_num;
 917
 918         /*
 919          * For a zoned filesystem on a non-zoned block device, use the same
 920          * super block locations as regular filesystem. Doing so, the super
 921          * block can always be retrieved and the zoned flag of the volume
 922          * detected from the super block information.
 923          */
 924         if (!bdev_is_zoned(device->bdev)) {
 925                 *bytenr_ret = btrfs_sb_offset(mirror);
 926                 return 0;
 927         }
 928
 929         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 930         if (zone_num + 1 >= zinfo->nr_zones)
 931                 return -ENOENT;
 932
 933         return sb_log_location(device->bdev,
 934                                &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror],
 935                                rw, bytenr_ret);
 936 }
 937
 938 static inline bool is_sb_log_zone(struct btrfs_zoned_device_info *zinfo,
 939                                   int mirror)
 940 {
 941         u32 zone_num;
 942
 943         if (!zinfo)
 944                 return false;
 945
 946         zone_num = sb_zone_number(zinfo->zone_size_shift, mirror);
 947         if (zone_num + 1 >= zinfo->nr_zones)
 948                 return false;
 949
 950         if (!test_bit(zone_num, zinfo->seq_zones))
 951                 return false;
 952
 953         return true;
 954 }
 955
 956 int btrfs_advance_sb_log(struct btrfs_device *device, int mirror)
 957 {
 958         struct btrfs_zoned_device_info *zinfo = device->zone_info;
 959         struct blk_zone *zone;
 960         int i;
 961
 962         if (!is_sb_log_zone(zinfo, mirror))
 963                 return 0;
 964
 965         zone = &zinfo->sb_zones[BTRFS_NR_SB_LOG_ZONES * mirror];
 966         for (i = 0; i < BTRFS_NR_SB_LOG_ZONES; i++) {
 967                 /* Advance the next zone */
 968                 if (zone->cond == BLK_ZONE_COND_FULL) {
 969                         zone++;
 970                         continue;
 971                 }
 972
 973                 if (zone->cond == BLK_ZONE_COND_EMPTY)
 974                         zone->cond = BLK_ZONE_COND_IMP_OPEN;
 975
 976                 zone->wp += SUPER_INFO_SECTORS;
 977
 978                 if (sb_zone_is_full(zone)) {
 979                         /*
 980                          * No room left to write new superblock. Since
 981                          * superblock is written with REQ_SYNC, it is safe to
 982                          * finish the zone now.
 983                          *
 984                          * If the write pointer is exactly at the capacity,
 985                          * explicit ZONE_FINISH is not necessary.
 986                          */
 987                         if (zone->wp != zone->start + zone->capacity) {
 988                                 int ret;
 989
 990                                 ret = blkdev_zone_mgmt(device->bdev,
 991                                                 REQ_OP_ZONE_FINISH, zone->start,
 992                                                 zone->len, GFP_NOFS);
 993                                 if (ret)
 994                                         return ret;
 995                         }
 996
 997                         zone->wp = zone->start + zone->len;
 998                         zone->cond = BLK_ZONE_COND_FULL;
 999                 }
1000                 return 0;
1001         }
1002
1003         /* All the zones are FULL. Should not reach here. */
1004         ASSERT(0);
1005         return -EIO;
1006 }
1007
1008 int btrfs_reset_sb_log_zones(struct block_device *bdev, int mirror)
1009 {
1010         sector_t zone_sectors;
1011         sector_t nr_sectors;
1012         u8 zone_sectors_shift;
1013         u32 sb_zone;
1014         u32 nr_zones;
1015
1016         zone_sectors = bdev_zone_sectors(bdev);
1017         zone_sectors_shift = ilog2(zone_sectors);
1018         nr_sectors = bdev_nr_sectors(bdev);
1019         nr_zones = nr_sectors >> zone_sectors_shift;
1020
1021         sb_zone = sb_zone_number(zone_sectors_shift + SECTOR_SHIFT, mirror);
1022         if (sb_zone + 1 >= nr_zones)
1023                 return -ENOENT;
1024
1025         return blkdev_zone_mgmt(bdev, REQ_OP_ZONE_RESET,
1026                                 zone_start_sector(sb_zone, bdev),
1027                                 zone_sectors * BTRFS_NR_SB_LOG_ZONES, GFP_NOFS);
1028 }
1029
1030 /*
1031  * Find allocatable zones within a given region.
1032  *
1033  * @device:     the device to allocate a region on
1034  * @hole_start: the position of the hole to allocate the region
1035  * @num_bytes:  size of wanted region
1036  * @hole_end:   the end of the hole
1037  * @return:     position of allocatable zones
1038  *
1039  * Allocatable region should not contain any superblock locations.
1040  */
1041 u64 btrfs_find_allocatable_zones(struct btrfs_device *device, u64 hole_start,
1042                                  u64 hole_end, u64 num_bytes)
1043 {
1044         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1045         const u8 shift = zinfo->zone_size_shift;
1046         u64 nzones = num_bytes >> shift;
1047         u64 pos = hole_start;
1048         u64 begin, end;
1049         bool have_sb;
1050         int i;
1051
1052         ASSERT(IS_ALIGNED(hole_start, zinfo->zone_size));
1053         ASSERT(IS_ALIGNED(num_bytes, zinfo->zone_size));
1054
1055         while (pos < hole_end) {
1056                 begin = pos >> shift;
1057                 end = begin + nzones;
1058
1059                 if (end > zinfo->nr_zones)
1060                         return hole_end;
1061
1062                 /* Check if zones in the region are all empty */
1063                 if (btrfs_dev_is_sequential(device, pos) &&
1064                     !bitmap_test_range_all_set(zinfo->empty_zones, begin, nzones)) {
1065                         pos += zinfo->zone_size;
1066                         continue;
1067                 }
1068
1069                 have_sb = false;
1070                 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1071                         u32 sb_zone;
1072                         u64 sb_pos;
1073
1074                         sb_zone = sb_zone_number(shift, i);
1075                         if (!(end <= sb_zone ||
1076                               sb_zone + BTRFS_NR_SB_LOG_ZONES <= begin)) {
1077                                 have_sb = true;
1078                                 pos = zone_start_physical(
1079                                         sb_zone + BTRFS_NR_SB_LOG_ZONES, zinfo);
1080                                 break;
1081                         }
1082
1083                         /* We also need to exclude regular superblock positions */
1084                         sb_pos = btrfs_sb_offset(i);
1085                         if (!(pos + num_bytes <= sb_pos ||
1086                               sb_pos + BTRFS_SUPER_INFO_SIZE <= pos)) {
1087                                 have_sb = true;
1088                                 pos = ALIGN(sb_pos + BTRFS_SUPER_INFO_SIZE,
1089                                             zinfo->zone_size);
1090                                 break;
1091                         }
1092                 }
1093                 if (!have_sb)
1094                         break;
1095         }
1096
1097         return pos;
1098 }
1099
1100 static bool btrfs_dev_set_active_zone(struct btrfs_device *device, u64 pos)
1101 {
1102         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1103         unsigned int zno = (pos >> zone_info->zone_size_shift);
1104
1105         /* We can use any number of zones */
1106         if (zone_info->max_active_zones == 0)
1107                 return true;
1108
1109         if (!test_bit(zno, zone_info->active_zones)) {
1110                 /* Active zone left? */
1111                 if (atomic_dec_if_positive(&zone_info->active_zones_left) < 0)
1112                         return false;
1113                 if (test_and_set_bit(zno, zone_info->active_zones)) {
1114                         /* Someone already set the bit */
1115                         atomic_inc(&zone_info->active_zones_left);
1116                 }
1117         }
1118
1119         return true;
1120 }
1121
1122 static void btrfs_dev_clear_active_zone(struct btrfs_device *device, u64 pos)
1123 {
1124         struct btrfs_zoned_device_info *zone_info = device->zone_info;
1125         unsigned int zno = (pos >> zone_info->zone_size_shift);
1126
1127         /* We can use any number of zones */
1128         if (zone_info->max_active_zones == 0)
1129                 return;
1130
1131         if (test_and_clear_bit(zno, zone_info->active_zones))
1132                 atomic_inc(&zone_info->active_zones_left);
1133 }
1134
1135 int btrfs_reset_device_zone(struct btrfs_device *device, u64 physical,
1136                             u64 length, u64 *bytes)
1137 {
1138         int ret;
1139
1140         *bytes = 0;
1141         ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_RESET,
1142                                physical >> SECTOR_SHIFT, length >> SECTOR_SHIFT,
1143                                GFP_NOFS);
1144         if (ret)
1145                 return ret;
1146
1147         *bytes = length;
1148         while (length) {
1149                 btrfs_dev_set_zone_empty(device, physical);
1150                 btrfs_dev_clear_active_zone(device, physical);
1151                 physical += device->zone_info->zone_size;
1152                 length -= device->zone_info->zone_size;
1153         }
1154
1155         return 0;
1156 }
1157
1158 int btrfs_ensure_empty_zones(struct btrfs_device *device, u64 start, u64 size)
1159 {
1160         struct btrfs_zoned_device_info *zinfo = device->zone_info;
1161         const u8 shift = zinfo->zone_size_shift;
1162         unsigned long begin = start >> shift;
1163         unsigned long nbits = size >> shift;
1164         u64 pos;
1165         int ret;
1166
1167         ASSERT(IS_ALIGNED(start, zinfo->zone_size));
1168         ASSERT(IS_ALIGNED(size, zinfo->zone_size));
1169
1170         if (begin + nbits > zinfo->nr_zones)
1171                 return -ERANGE;
1172
1173         /* All the zones are conventional */
1174         if (bitmap_test_range_all_zero(zinfo->seq_zones, begin, nbits))
1175                 return 0;
1176
1177         /* All the zones are sequential and empty */
1178         if (bitmap_test_range_all_set(zinfo->seq_zones, begin, nbits) &&
1179             bitmap_test_range_all_set(zinfo->empty_zones, begin, nbits))
1180                 return 0;
1181
1182         for (pos = start; pos < start + size; pos += zinfo->zone_size) {
1183                 u64 reset_bytes;
1184
1185                 if (!btrfs_dev_is_sequential(device, pos) ||
1186                     btrfs_dev_is_empty_zone(device, pos))
1187                         continue;
1188
1189                 /* Free regions should be empty */
1190                 btrfs_warn_in_rcu(
1191                         device->fs_info,
1192                 "zoned: resetting device %s (devid %llu) zone %llu for allocation",
1193                         rcu_str_deref(device->name), device->devid, pos >> shift);
1194                 WARN_ON_ONCE(1);
1195
1196                 ret = btrfs_reset_device_zone(device, pos, zinfo->zone_size,
1197                                               &reset_bytes);
1198                 if (ret)
1199                         return ret;
1200         }
1201
1202         return 0;
1203 }
1204
1205 /*
1206  * Calculate an allocation pointer from the extent allocation information
1207  * for a block group consist of conventional zones. It is pointed to the
1208  * end of the highest addressed extent in the block group as an allocation
1209  * offset.
1210  */
1211 static int calculate_alloc_pointer(struct btrfs_block_group *cache,
1212                                    u64 *offset_ret, bool new)
1213 {
1214         struct btrfs_fs_info *fs_info = cache->fs_info;
1215         struct btrfs_root *root;
1216         struct btrfs_path *path;
1217         struct btrfs_key key;
1218         struct btrfs_key found_key;
1219         int ret;
1220         u64 length;
1221
1222         /*
1223          * Avoid  tree lookups for a new block group, there's no use for it.
1224          * It must always be 0.
1225          *
1226          * Also, we have a lock chain of extent buffer lock -> chunk mutex.
1227          * For new a block group, this function is called from
1228          * btrfs_make_block_group() which is already taking the chunk mutex.
1229          * Thus, we cannot call calculate_alloc_pointer() which takes extent
1230          * buffer locks to avoid deadlock.
1231          */
1232         if (new) {
1233                 *offset_ret = 0;
1234                 return 0;
1235         }
1236
1237         path = btrfs_alloc_path();
1238         if (!path)
1239                 return -ENOMEM;
1240
1241         key.objectid = cache->start + cache->length;
1242         key.type = 0;
1243         key.offset = 0;
1244
1245         root = btrfs_extent_root(fs_info, key.objectid);
1246         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1247         /* We should not find the exact match */
1248         if (!ret)
1249                 ret = -EUCLEAN;
1250         if (ret < 0)
1251                 goto out;
1252
1253         ret = btrfs_previous_extent_item(root, path, cache->start);
1254         if (ret) {
1255                 if (ret == 1) {
1256                         ret = 0;
1257                         *offset_ret = 0;
1258                 }
1259                 goto out;
1260         }
1261
1262         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
1263
1264         if (found_key.type == BTRFS_EXTENT_ITEM_KEY)
1265                 length = found_key.offset;
1266         else
1267                 length = fs_info->nodesize;
1268
1269         if (!(found_key.objectid >= cache->start &&
1270                found_key.objectid + length <= cache->start + cache->length)) {
1271                 ret = -EUCLEAN;
1272                 goto out;
1273         }
1274         *offset_ret = found_key.objectid + length - cache->start;
1275         ret = 0;
1276
1277 out:
1278         btrfs_free_path(path);
1279         return ret;
1280 }
1281
1282 int btrfs_load_block_group_zone_info(struct btrfs_block_group *cache, bool new)
1283 {
1284         struct btrfs_fs_info *fs_info = cache->fs_info;
1285         struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1286         struct extent_map *em;
1287         struct map_lookup *map;
1288         struct btrfs_device *device;
1289         u64 logical = cache->start;
1290         u64 length = cache->length;
1291         int ret;
1292         int i;
1293         unsigned int nofs_flag;
1294         u64 *alloc_offsets = NULL;
1295         u64 *caps = NULL;
1296         u64 *physical = NULL;
1297         unsigned long *active = NULL;
1298         u64 last_alloc = 0;
1299         u32 num_sequential = 0, num_conventional = 0;
1300
1301         if (!btrfs_is_zoned(fs_info))
1302                 return 0;
1303
1304         /* Sanity check */
1305         if (!IS_ALIGNED(length, fs_info->zone_size)) {
1306                 btrfs_err(fs_info,
1307                 "zoned: block group %llu len %llu unaligned to zone size %llu",
1308                           logical, length, fs_info->zone_size);
1309                 return -EIO;
1310         }
1311
1312         /* Get the chunk mapping */
1313         read_lock(&em_tree->lock);
1314         em = lookup_extent_mapping(em_tree, logical, length);
1315         read_unlock(&em_tree->lock);
1316
1317         if (!em)
1318                 return -EINVAL;
1319
1320         map = em->map_lookup;
1321
1322         cache->physical_map = kmemdup(map, map_lookup_size(map->num_stripes), GFP_NOFS);
1323         if (!cache->physical_map) {
1324                 ret = -ENOMEM;
1325                 goto out;
1326         }
1327
1328         alloc_offsets = kcalloc(map->num_stripes, sizeof(*alloc_offsets), GFP_NOFS);
1329         if (!alloc_offsets) {
1330                 ret = -ENOMEM;
1331                 goto out;
1332         }
1333
1334         caps = kcalloc(map->num_stripes, sizeof(*caps), GFP_NOFS);
1335         if (!caps) {
1336                 ret = -ENOMEM;
1337                 goto out;
1338         }
1339
1340         physical = kcalloc(map->num_stripes, sizeof(*physical), GFP_NOFS);
1341         if (!physical) {
1342                 ret = -ENOMEM;
1343                 goto out;
1344         }
1345
1346         active = bitmap_zalloc(map->num_stripes, GFP_NOFS);
1347         if (!active) {
1348                 ret = -ENOMEM;
1349                 goto out;
1350         }
1351
1352         for (i = 0; i < map->num_stripes; i++) {
1353                 bool is_sequential;
1354                 struct blk_zone zone;
1355                 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
1356                 int dev_replace_is_ongoing = 0;
1357
1358                 device = map->stripes[i].dev;
1359                 physical[i] = map->stripes[i].physical;
1360
1361                 if (device->bdev == NULL) {
1362                         alloc_offsets[i] = WP_MISSING_DEV;
1363                         continue;
1364                 }
1365
1366                 is_sequential = btrfs_dev_is_sequential(device, physical[i]);
1367                 if (is_sequential)
1368                         num_sequential++;
1369                 else
1370                         num_conventional++;
1371
1372                 /*
1373                  * Consider a zone as active if we can allow any number of
1374                  * active zones.
1375                  */
1376                 if (!device->zone_info->max_active_zones)
1377                         __set_bit(i, active);
1378
1379                 if (!is_sequential) {
1380                         alloc_offsets[i] = WP_CONVENTIONAL;
1381                         continue;
1382                 }
1383
1384                 /*
1385                  * This zone will be used for allocation, so mark this zone
1386                  * non-empty.
1387                  */
1388                 btrfs_dev_clear_zone_empty(device, physical[i]);
1389
1390                 down_read(&dev_replace->rwsem);
1391                 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
1392                 if (dev_replace_is_ongoing && dev_replace->tgtdev != NULL)
1393                         btrfs_dev_clear_zone_empty(dev_replace->tgtdev, physical[i]);
1394                 up_read(&dev_replace->rwsem);
1395
1396                 /*
1397                  * The group is mapped to a sequential zone. Get the zone write
1398                  * pointer to determine the allocation offset within the zone.
1399                  */
1400                 WARN_ON(!IS_ALIGNED(physical[i], fs_info->zone_size));
1401                 nofs_flag = memalloc_nofs_save();
1402                 ret = btrfs_get_dev_zone(device, physical[i], &zone);
1403                 memalloc_nofs_restore(nofs_flag);
1404                 if (ret == -EIO || ret == -EOPNOTSUPP) {
1405                         ret = 0;
1406                         alloc_offsets[i] = WP_MISSING_DEV;
1407                         continue;
1408                 } else if (ret) {
1409                         goto out;
1410                 }
1411
1412                 if (zone.type == BLK_ZONE_TYPE_CONVENTIONAL) {
1413                         btrfs_err_in_rcu(fs_info,
1414         "zoned: unexpected conventional zone %llu on device %s (devid %llu)",
1415                                 zone.start << SECTOR_SHIFT,
1416                                 rcu_str_deref(device->name), device->devid);
1417                         ret = -EIO;
1418                         goto out;
1419                 }
1420
1421                 caps[i] = (zone.capacity << SECTOR_SHIFT);
1422
1423                 switch (zone.cond) {
1424                 case BLK_ZONE_COND_OFFLINE:
1425                 case BLK_ZONE_COND_READONLY:
1426                         btrfs_err(fs_info,
1427                 "zoned: offline/readonly zone %llu on device %s (devid %llu)",
1428                                   physical[i] >> device->zone_info->zone_size_shift,
1429                                   rcu_str_deref(device->name), device->devid);
1430                         alloc_offsets[i] = WP_MISSING_DEV;
1431                         break;
1432                 case BLK_ZONE_COND_EMPTY:
1433                         alloc_offsets[i] = 0;
1434                         break;
1435                 case BLK_ZONE_COND_FULL:
1436                         alloc_offsets[i] = caps[i];
1437                         break;
1438                 default:
1439                         /* Partially used zone */
1440                         alloc_offsets[i] =
1441                                         ((zone.wp - zone.start) << SECTOR_SHIFT);
1442                         __set_bit(i, active);
1443                         break;
1444                 }
1445         }
1446
1447         if (num_sequential > 0)
1448                 set_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1449
1450         if (num_conventional > 0) {
1451                 /* Zone capacity is always zone size in emulation */
1452                 cache->zone_capacity = cache->length;
1453                 ret = calculate_alloc_pointer(cache, &last_alloc, new);
1454                 if (ret) {
1455                         btrfs_err(fs_info,
1456                         "zoned: failed to determine allocation offset of bg %llu",
1457                                   cache->start);
1458                         goto out;
1459                 } else if (map->num_stripes == num_conventional) {
1460                         cache->alloc_offset = last_alloc;
1461                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1462                         goto out;
1463                 }
1464         }
1465
1466         switch (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
1467         case 0: /* single */
1468                 if (alloc_offsets[0] == WP_MISSING_DEV) {
1469                         btrfs_err(fs_info,
1470                         "zoned: cannot recover write pointer for zone %llu",
1471                                 physical[0]);
1472                         ret = -EIO;
1473                         goto out;
1474                 }
1475                 cache->alloc_offset = alloc_offsets[0];
1476                 cache->zone_capacity = caps[0];
1477                 if (test_bit(0, active))
1478                         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags);
1479                 break;
1480         case BTRFS_BLOCK_GROUP_DUP:
1481                 if (map->type & BTRFS_BLOCK_GROUP_DATA) {
1482                         btrfs_err(fs_info, "zoned: profile DUP not yet supported on data bg");
1483                         ret = -EINVAL;
1484                         goto out;
1485                 }
1486                 if (alloc_offsets[0] == WP_MISSING_DEV) {
1487                         btrfs_err(fs_info,
1488                         "zoned: cannot recover write pointer for zone %llu",
1489                                 physical[0]);
1490                         ret = -EIO;
1491                         goto out;
1492                 }
1493                 if (alloc_offsets[1] == WP_MISSING_DEV) {
1494                         btrfs_err(fs_info,
1495                         "zoned: cannot recover write pointer for zone %llu",
1496                                 physical[1]);
1497                         ret = -EIO;
1498                         goto out;
1499                 }
1500                 if (alloc_offsets[0] != alloc_offsets[1]) {
1501                         btrfs_err(fs_info,
1502                         "zoned: write pointer offset mismatch of zones in DUP profile");
1503                         ret = -EIO;
1504                         goto out;
1505                 }
1506                 if (test_bit(0, active) != test_bit(1, active)) {
1507                         if (!btrfs_zone_activate(cache)) {
1508                                 ret = -EIO;
1509                                 goto out;
1510                         }
1511                 } else {
1512                         if (test_bit(0, active))
1513                                 set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
1514                                         &cache->runtime_flags);
1515                 }
1516                 cache->alloc_offset = alloc_offsets[0];
1517                 cache->zone_capacity = min(caps[0], caps[1]);
1518                 break;
1519         case BTRFS_BLOCK_GROUP_RAID1:
1520         case BTRFS_BLOCK_GROUP_RAID0:
1521         case BTRFS_BLOCK_GROUP_RAID10:
1522         case BTRFS_BLOCK_GROUP_RAID5:
1523         case BTRFS_BLOCK_GROUP_RAID6:
1524                 /* non-single profiles are not supported yet */
1525         default:
1526                 btrfs_err(fs_info, "zoned: profile %s not yet supported",
1527                           btrfs_bg_type_to_raid_name(map->type));
1528                 ret = -EINVAL;
1529                 goto out;
1530         }
1531
1532 out:
1533         if (cache->alloc_offset > fs_info->zone_size) {
1534                 btrfs_err(fs_info,
1535                         "zoned: invalid write pointer %llu in block group %llu",
1536                         cache->alloc_offset, cache->start);
1537                 ret = -EIO;
1538         }
1539
1540         if (cache->alloc_offset > cache->zone_capacity) {
1541                 btrfs_err(fs_info,
1542 "zoned: invalid write pointer %llu (larger than zone capacity %llu) in block group %llu",
1543                           cache->alloc_offset, cache->zone_capacity,
1544                           cache->start);
1545                 ret = -EIO;
1546         }
1547
1548         /* An extent is allocated after the write pointer */
1549         if (!ret && num_conventional && last_alloc > cache->alloc_offset) {
1550                 btrfs_err(fs_info,
1551                           "zoned: got wrong write pointer in BG %llu: %llu > %llu",
1552                           logical, last_alloc, cache->alloc_offset);
1553                 ret = -EIO;
1554         }
1555
1556         if (!ret) {
1557                 cache->meta_write_pointer = cache->alloc_offset + cache->start;
1558                 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags)) {
1559                         btrfs_get_block_group(cache);
1560                         spin_lock(&fs_info->zone_active_bgs_lock);
1561                         list_add_tail(&cache->active_bg_list,
1562                                       &fs_info->zone_active_bgs);
1563                         spin_unlock(&fs_info->zone_active_bgs_lock);
1564                 }
1565         } else {
1566                 kfree(cache->physical_map);
1567                 cache->physical_map = NULL;
1568         }
1569         bitmap_free(active);
1570         kfree(physical);
1571         kfree(caps);
1572         kfree(alloc_offsets);
1573         free_extent_map(em);
1574
1575         return ret;
1576 }
1577
1578 void btrfs_calc_zone_unusable(struct btrfs_block_group *cache)
1579 {
1580         u64 unusable, free;
1581
1582         if (!btrfs_is_zoned(cache->fs_info))
1583                 return;
1584
1585         WARN_ON(cache->bytes_super != 0);
1586
1587         /* Check for block groups never get activated */
1588         if (test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &cache->fs_info->flags) &&
1589             cache->flags & (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM) &&
1590             !test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &cache->runtime_flags) &&
1591             cache->alloc_offset == 0) {
1592                 unusable = cache->length;
1593                 free = 0;
1594         } else {
1595                 unusable = (cache->alloc_offset - cache->used) +
1596                            (cache->length - cache->zone_capacity);
1597                 free = cache->zone_capacity - cache->alloc_offset;
1598         }
1599
1600         /* We only need ->free_space in ALLOC_SEQ block groups */
1601         cache->cached = BTRFS_CACHE_FINISHED;
1602         cache->free_space_ctl->free_space = free;
1603         cache->zone_unusable = unusable;
1604 }
1605
1606 void btrfs_redirty_list_add(struct btrfs_transaction *trans,
1607                             struct extent_buffer *eb)
1608 {
1609         if (!btrfs_is_zoned(eb->fs_info) ||
1610             btrfs_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN))
1611                 return;
1612
1613         ASSERT(!test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
1614
1615         memzero_extent_buffer(eb, 0, eb->len);
1616         set_bit(EXTENT_BUFFER_NO_CHECK, &eb->bflags);
1617         set_extent_buffer_dirty(eb);
1618         set_extent_bit(&trans->dirty_pages, eb->start, eb->start + eb->len - 1,
1619                         EXTENT_DIRTY | EXTENT_NOWAIT, NULL);
1620 }
1621
1622 bool btrfs_use_zone_append(struct btrfs_bio *bbio)
1623 {
1624         u64 start = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT);
1625         struct btrfs_inode *inode = bbio->inode;
1626         struct btrfs_fs_info *fs_info = bbio->fs_info;
1627         struct btrfs_block_group *cache;
1628         bool ret = false;
1629
1630         if (!btrfs_is_zoned(fs_info))
1631                 return false;
1632
1633         if (!inode || !is_data_inode(&inode->vfs_inode))
1634                 return false;
1635
1636         if (btrfs_op(&bbio->bio) != BTRFS_MAP_WRITE)
1637                 return false;
1638
1639         /*
1640          * Using REQ_OP_ZONE_APPNED for relocation can break assumptions on the
1641          * extent layout the relocation code has.
1642          * Furthermore we have set aside own block-group from which only the
1643          * relocation "process" can allocate and make sure only one process at a
1644          * time can add pages to an extent that gets relocated, so it's safe to
1645          * use regular REQ_OP_WRITE for this special case.
1646          */
1647         if (btrfs_is_data_reloc_root(inode->root))
1648                 return false;
1649
1650         cache = btrfs_lookup_block_group(fs_info, start);
1651         ASSERT(cache);
1652         if (!cache)
1653                 return false;
1654
1655         ret = !!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &cache->runtime_flags);
1656         btrfs_put_block_group(cache);
1657
1658         return ret;
1659 }
1660
1661 void btrfs_record_physical_zoned(struct btrfs_bio *bbio)
1662 {
1663         const u64 physical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
1664         struct btrfs_ordered_sum *sum = bbio->sums;
1665
1666         if (physical < bbio->orig_physical)
1667                 sum->logical -= bbio->orig_physical - physical;
1668         else
1669                 sum->logical += physical - bbio->orig_physical;
1670 }
1671
1672 static void btrfs_rewrite_logical_zoned(struct btrfs_ordered_extent *ordered,
1673                                         u64 logical)
1674 {
1675         struct extent_map_tree *em_tree = &BTRFS_I(ordered->inode)->extent_tree;
1676         struct extent_map *em;
1677
1678         ordered->disk_bytenr = logical;
1679
1680         write_lock(&em_tree->lock);
1681         em = search_extent_mapping(em_tree, ordered->file_offset,
1682                                    ordered->num_bytes);
1683         em->block_start = logical;
1684         free_extent_map(em);
1685         write_unlock(&em_tree->lock);
1686 }
1687
1688 static bool btrfs_zoned_split_ordered(struct btrfs_ordered_extent *ordered,
1689                                       u64 logical, u64 len)
1690 {
1691         struct btrfs_ordered_extent *new;
1692
1693         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags) &&
1694             split_extent_map(BTRFS_I(ordered->inode), ordered->file_offset,
1695                              ordered->num_bytes, len, logical))
1696                 return false;
1697
1698         new = btrfs_split_ordered_extent(ordered, len);
1699         if (IS_ERR(new))
1700                 return false;
1701         new->disk_bytenr = logical;
1702         btrfs_finish_one_ordered(new);
1703         return true;
1704 }
1705
1706 void btrfs_finish_ordered_zoned(struct btrfs_ordered_extent *ordered)
1707 {
1708         struct btrfs_inode *inode = BTRFS_I(ordered->inode);
1709         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1710         struct btrfs_ordered_sum *sum =
1711                 list_first_entry(&ordered->list, typeof(*sum), list);
1712         u64 logical = sum->logical;
1713         u64 len = sum->len;
1714
1715         while (len < ordered->disk_num_bytes) {
1716                 sum = list_next_entry(sum, list);
1717                 if (sum->logical == logical + len) {
1718                         len += sum->len;
1719                         continue;
1720                 }
1721                 if (!btrfs_zoned_split_ordered(ordered, logical, len)) {
1722                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
1723                         btrfs_err(fs_info, "failed to split ordered extent");
1724                         goto out;
1725                 }
1726                 logical = sum->logical;
1727                 len = sum->len;
1728         }
1729
1730         if (ordered->disk_bytenr != logical)
1731                 btrfs_rewrite_logical_zoned(ordered, logical);
1732
1733 out:
1734         /*
1735          * If we end up here for nodatasum I/O, the btrfs_ordered_sum structures
1736          * were allocated by btrfs_alloc_dummy_sum only to record the logical
1737          * addresses and don't contain actual checksums.  We thus must free them
1738          * here so that we don't attempt to log the csums later.
1739          */
1740         if ((inode->flags & BTRFS_INODE_NODATASUM) ||
1741             test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state)) {
1742                 while ((sum = list_first_entry_or_null(&ordered->list,
1743                                                        typeof(*sum), list))) {
1744                         list_del(&sum->list);
1745                         kfree(sum);
1746                 }
1747         }
1748 }
1749
1750 bool btrfs_check_meta_write_pointer(struct btrfs_fs_info *fs_info,
1751                                     struct extent_buffer *eb,
1752                                     struct btrfs_block_group **cache_ret)
1753 {
1754         struct btrfs_block_group *cache;
1755         bool ret = true;
1756
1757         if (!btrfs_is_zoned(fs_info))
1758                 return true;
1759
1760         cache = btrfs_lookup_block_group(fs_info, eb->start);
1761         if (!cache)
1762                 return true;
1763
1764         if (cache->meta_write_pointer != eb->start) {
1765                 btrfs_put_block_group(cache);
1766                 cache = NULL;
1767                 ret = false;
1768         } else {
1769                 cache->meta_write_pointer = eb->start + eb->len;
1770         }
1771
1772         *cache_ret = cache;
1773
1774         return ret;
1775 }
1776
1777 void btrfs_revert_meta_write_pointer(struct btrfs_block_group *cache,
1778                                      struct extent_buffer *eb)
1779 {
1780         if (!btrfs_is_zoned(eb->fs_info) || !cache)
1781                 return;
1782
1783         ASSERT(cache->meta_write_pointer == eb->start + eb->len);
1784         cache->meta_write_pointer = eb->start;
1785 }
1786
1787 int btrfs_zoned_issue_zeroout(struct btrfs_device *device, u64 physical, u64 length)
1788 {
1789         if (!btrfs_dev_is_sequential(device, physical))
1790                 return -EOPNOTSUPP;
1791
1792         return blkdev_issue_zeroout(device->bdev, physical >> SECTOR_SHIFT,
1793                                     length >> SECTOR_SHIFT, GFP_NOFS, 0);
1794 }
1795
1796 static int read_zone_info(struct btrfs_fs_info *fs_info, u64 logical,
1797                           struct blk_zone *zone)
1798 {
1799         struct btrfs_io_context *bioc = NULL;
1800         u64 mapped_length = PAGE_SIZE;
1801         unsigned int nofs_flag;
1802         int nmirrors;
1803         int i, ret;
1804
1805         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
1806                               &mapped_length, &bioc, NULL, NULL, 1);
1807         if (ret || !bioc || mapped_length < PAGE_SIZE) {
1808                 ret = -EIO;
1809                 goto out_put_bioc;
1810         }
1811
1812         if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1813                 ret = -EINVAL;
1814                 goto out_put_bioc;
1815         }
1816
1817         nofs_flag = memalloc_nofs_save();
1818         nmirrors = (int)bioc->num_stripes;
1819         for (i = 0; i < nmirrors; i++) {
1820                 u64 physical = bioc->stripes[i].physical;
1821                 struct btrfs_device *dev = bioc->stripes[i].dev;
1822
1823                 /* Missing device */
1824                 if (!dev->bdev)
1825                         continue;
1826
1827                 ret = btrfs_get_dev_zone(dev, physical, zone);
1828                 /* Failing device */
1829                 if (ret == -EIO || ret == -EOPNOTSUPP)
1830                         continue;
1831                 break;
1832         }
1833         memalloc_nofs_restore(nofs_flag);
1834 out_put_bioc:
1835         btrfs_put_bioc(bioc);
1836         return ret;
1837 }
1838
1839 /*
1840  * Synchronize write pointer in a zone at @physical_start on @tgt_dev, by
1841  * filling zeros between @physical_pos to a write pointer of dev-replace
1842  * source device.
1843  */
1844 int btrfs_sync_zone_write_pointer(struct btrfs_device *tgt_dev, u64 logical,
1845                                     u64 physical_start, u64 physical_pos)
1846 {
1847         struct btrfs_fs_info *fs_info = tgt_dev->fs_info;
1848         struct blk_zone zone;
1849         u64 length;
1850         u64 wp;
1851         int ret;
1852
1853         if (!btrfs_dev_is_sequential(tgt_dev, physical_pos))
1854                 return 0;
1855
1856         ret = read_zone_info(fs_info, logical, &zone);
1857         if (ret)
1858                 return ret;
1859
1860         wp = physical_start + ((zone.wp - zone.start) << SECTOR_SHIFT);
1861
1862         if (physical_pos == wp)
1863                 return 0;
1864
1865         if (physical_pos > wp)
1866                 return -EUCLEAN;
1867
1868         length = wp - physical_pos;
1869         return btrfs_zoned_issue_zeroout(tgt_dev, physical_pos, length);
1870 }
1871
1872 /*
1873  * Activate block group and underlying device zones
1874  *
1875  * @block_group: the block group to activate
1876  *
1877  * Return: true on success, false otherwise
1878  */
1879 bool btrfs_zone_activate(struct btrfs_block_group *block_group)
1880 {
1881         struct btrfs_fs_info *fs_info = block_group->fs_info;
1882         struct btrfs_space_info *space_info = block_group->space_info;
1883         struct map_lookup *map;
1884         struct btrfs_device *device;
1885         u64 physical;
1886         bool ret;
1887         int i;
1888
1889         if (!btrfs_is_zoned(block_group->fs_info))
1890                 return true;
1891
1892         map = block_group->physical_map;
1893
1894         spin_lock(&space_info->lock);
1895         spin_lock(&block_group->lock);
1896         if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
1897                 ret = true;
1898                 goto out_unlock;
1899         }
1900
1901         /* No space left */
1902         if (btrfs_zoned_bg_is_full(block_group)) {
1903                 ret = false;
1904                 goto out_unlock;
1905         }
1906
1907         for (i = 0; i < map->num_stripes; i++) {
1908                 device = map->stripes[i].dev;
1909                 physical = map->stripes[i].physical;
1910
1911                 if (device->zone_info->max_active_zones == 0)
1912                         continue;
1913
1914                 if (!btrfs_dev_set_active_zone(device, physical)) {
1915                         /* Cannot activate the zone */
1916                         ret = false;
1917                         goto out_unlock;
1918                 }
1919         }
1920
1921         /* Successfully activated all the zones */
1922         set_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
1923         WARN_ON(block_group->alloc_offset != 0);
1924         if (block_group->zone_unusable == block_group->length) {
1925                 block_group->zone_unusable = block_group->length - block_group->zone_capacity;
1926                 space_info->bytes_zone_unusable -= block_group->zone_capacity;
1927         }
1928         spin_unlock(&block_group->lock);
1929         btrfs_try_granting_tickets(fs_info, space_info);
1930         spin_unlock(&space_info->lock);
1931
1932         /* For the active block group list */
1933         btrfs_get_block_group(block_group);
1934
1935         spin_lock(&fs_info->zone_active_bgs_lock);
1936         list_add_tail(&block_group->active_bg_list, &fs_info->zone_active_bgs);
1937         spin_unlock(&fs_info->zone_active_bgs_lock);
1938
1939         return true;
1940
1941 out_unlock:
1942         spin_unlock(&block_group->lock);
1943         spin_unlock(&space_info->lock);
1944         return ret;
1945 }
1946
1947 static void wait_eb_writebacks(struct btrfs_block_group *block_group)
1948 {
1949         struct btrfs_fs_info *fs_info = block_group->fs_info;
1950         const u64 end = block_group->start + block_group->length;
1951         struct radix_tree_iter iter;
1952         struct extent_buffer *eb;
1953         void __rcu **slot;
1954
1955         rcu_read_lock();
1956         radix_tree_for_each_slot(slot, &fs_info->buffer_radix, &iter,
1957                                  block_group->start >> fs_info->sectorsize_bits) {
1958                 eb = radix_tree_deref_slot(slot);
1959                 if (!eb)
1960                         continue;
1961                 if (radix_tree_deref_retry(eb)) {
1962                         slot = radix_tree_iter_retry(&iter);
1963                         continue;
1964                 }
1965
1966                 if (eb->start < block_group->start)
1967                         continue;
1968                 if (eb->start >= end)
1969                         break;
1970
1971                 slot = radix_tree_iter_resume(slot, &iter);
1972                 rcu_read_unlock();
1973                 wait_on_extent_buffer_writeback(eb);
1974                 rcu_read_lock();
1975         }
1976         rcu_read_unlock();
1977 }
1978
1979 static int do_zone_finish(struct btrfs_block_group *block_group, bool fully_written)
1980 {
1981         struct btrfs_fs_info *fs_info = block_group->fs_info;
1982         struct map_lookup *map;
1983         const bool is_metadata = (block_group->flags &
1984                         (BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_SYSTEM));
1985         int ret = 0;
1986         int i;
1987
1988         spin_lock(&block_group->lock);
1989         if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) {
1990                 spin_unlock(&block_group->lock);
1991                 return 0;
1992         }
1993
1994         /* Check if we have unwritten allocated space */
1995         if (is_metadata &&
1996             block_group->start + block_group->alloc_offset > block_group->meta_write_pointer) {
1997                 spin_unlock(&block_group->lock);
1998                 return -EAGAIN;
1999         }
2000
2001         /*
2002          * If we are sure that the block group is full (= no more room left for
2003          * new allocation) and the IO for the last usable block is completed, we
2004          * don't need to wait for the other IOs. This holds because we ensure
2005          * the sequential IO submissions using the ZONE_APPEND command for data
2006          * and block_group->meta_write_pointer for metadata.
2007          */
2008         if (!fully_written) {
2009                 spin_unlock(&block_group->lock);
2010
2011                 ret = btrfs_inc_block_group_ro(block_group, false);
2012                 if (ret)
2013                         return ret;
2014
2015                 /* Ensure all writes in this block group finish */
2016                 btrfs_wait_block_group_reservations(block_group);
2017                 /* No need to wait for NOCOW writers. Zoned mode does not allow that */
2018                 btrfs_wait_ordered_roots(fs_info, U64_MAX, block_group->start,
2019                                          block_group->length);
2020                 /* Wait for extent buffers to be written. */
2021                 if (is_metadata)
2022                         wait_eb_writebacks(block_group);
2023
2024                 spin_lock(&block_group->lock);
2025
2026                 /*
2027                  * Bail out if someone already deactivated the block group, or
2028                  * allocated space is left in the block group.
2029                  */
2030                 if (!test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2031                               &block_group->runtime_flags)) {
2032                         spin_unlock(&block_group->lock);
2033                         btrfs_dec_block_group_ro(block_group);
2034                         return 0;
2035                 }
2036
2037                 if (block_group->reserved) {
2038                         spin_unlock(&block_group->lock);
2039                         btrfs_dec_block_group_ro(block_group);
2040                         return -EAGAIN;
2041                 }
2042         }
2043
2044         clear_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags);
2045         block_group->alloc_offset = block_group->zone_capacity;
2046         block_group->free_space_ctl->free_space = 0;
2047         btrfs_clear_treelog_bg(block_group);
2048         btrfs_clear_data_reloc_bg(block_group);
2049         spin_unlock(&block_group->lock);
2050
2051         map = block_group->physical_map;
2052         for (i = 0; i < map->num_stripes; i++) {
2053                 struct btrfs_device *device = map->stripes[i].dev;
2054                 const u64 physical = map->stripes[i].physical;
2055
2056                 if (device->zone_info->max_active_zones == 0)
2057                         continue;
2058
2059                 ret = blkdev_zone_mgmt(device->bdev, REQ_OP_ZONE_FINISH,
2060                                        physical >> SECTOR_SHIFT,
2061                                        device->zone_info->zone_size >> SECTOR_SHIFT,
2062                                        GFP_NOFS);
2063
2064                 if (ret)
2065                         return ret;
2066
2067                 btrfs_dev_clear_active_zone(device, physical);
2068         }
2069
2070         if (!fully_written)
2071                 btrfs_dec_block_group_ro(block_group);
2072
2073         spin_lock(&fs_info->zone_active_bgs_lock);
2074         ASSERT(!list_empty(&block_group->active_bg_list));
2075         list_del_init(&block_group->active_bg_list);
2076         spin_unlock(&fs_info->zone_active_bgs_lock);
2077
2078         /* For active_bg_list */
2079         btrfs_put_block_group(block_group);
2080
2081         clear_and_wake_up_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2082
2083         return 0;
2084 }
2085
2086 int btrfs_zone_finish(struct btrfs_block_group *block_group)
2087 {
2088         if (!btrfs_is_zoned(block_group->fs_info))
2089                 return 0;
2090
2091         return do_zone_finish(block_group, false);
2092 }
2093
2094 bool btrfs_can_activate_zone(struct btrfs_fs_devices *fs_devices, u64 flags)
2095 {
2096         struct btrfs_fs_info *fs_info = fs_devices->fs_info;
2097         struct btrfs_device *device;
2098         bool ret = false;
2099
2100         if (!btrfs_is_zoned(fs_info))
2101                 return true;
2102
2103         /* Check if there is a device with active zones left */
2104         mutex_lock(&fs_info->chunk_mutex);
2105         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
2106                 struct btrfs_zoned_device_info *zinfo = device->zone_info;
2107
2108                 if (!device->bdev)
2109                         continue;
2110
2111                 if (!zinfo->max_active_zones) {
2112                         ret = true;
2113                         break;
2114                 }
2115
2116                 switch (flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
2117                 case 0: /* single */
2118                         ret = (atomic_read(&zinfo->active_zones_left) >= 1);
2119                         break;
2120                 case BTRFS_BLOCK_GROUP_DUP:
2121                         ret = (atomic_read(&zinfo->active_zones_left) >= 2);
2122                         break;
2123                 }
2124                 if (ret)
2125                         break;
2126         }
2127         mutex_unlock(&fs_info->chunk_mutex);
2128
2129         if (!ret)
2130                 set_bit(BTRFS_FS_NEED_ZONE_FINISH, &fs_info->flags);
2131
2132         return ret;
2133 }
2134
2135 void btrfs_zone_finish_endio(struct btrfs_fs_info *fs_info, u64 logical, u64 length)
2136 {
2137         struct btrfs_block_group *block_group;
2138         u64 min_alloc_bytes;
2139
2140         if (!btrfs_is_zoned(fs_info))
2141                 return;
2142
2143         block_group = btrfs_lookup_block_group(fs_info, logical);
2144         ASSERT(block_group);
2145
2146         /* No MIXED_BG on zoned btrfs. */
2147         if (block_group->flags & BTRFS_BLOCK_GROUP_DATA)
2148                 min_alloc_bytes = fs_info->sectorsize;
2149         else
2150                 min_alloc_bytes = fs_info->nodesize;
2151
2152         /* Bail out if we can allocate more data from this block group. */
2153         if (logical + length + min_alloc_bytes <=
2154             block_group->start + block_group->zone_capacity)
2155                 goto out;
2156
2157         do_zone_finish(block_group, true);
2158
2159 out:
2160         btrfs_put_block_group(block_group);
2161 }
2162
2163 static void btrfs_zone_finish_endio_workfn(struct work_struct *work)
2164 {
2165         struct btrfs_block_group *bg =
2166                 container_of(work, struct btrfs_block_group, zone_finish_work);
2167
2168         wait_on_extent_buffer_writeback(bg->last_eb);
2169         free_extent_buffer(bg->last_eb);
2170         btrfs_zone_finish_endio(bg->fs_info, bg->start, bg->length);
2171         btrfs_put_block_group(bg);
2172 }
2173
2174 void btrfs_schedule_zone_finish_bg(struct btrfs_block_group *bg,
2175                                    struct extent_buffer *eb)
2176 {
2177         if (!test_bit(BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE, &bg->runtime_flags) ||
2178             eb->start + eb->len * 2 <= bg->start + bg->zone_capacity)
2179                 return;
2180
2181         if (WARN_ON(bg->zone_finish_work.func == btrfs_zone_finish_endio_workfn)) {
2182                 btrfs_err(bg->fs_info, "double scheduling of bg %llu zone finishing",
2183                           bg->start);
2184                 return;
2185         }
2186
2187         /* For the work */
2188         btrfs_get_block_group(bg);
2189         atomic_inc(&eb->refs);
2190         bg->last_eb = eb;
2191         INIT_WORK(&bg->zone_finish_work, btrfs_zone_finish_endio_workfn);
2192         queue_work(system_unbound_wq, &bg->zone_finish_work);
2193 }
2194
2195 void btrfs_clear_data_reloc_bg(struct btrfs_block_group *bg)
2196 {
2197         struct btrfs_fs_info *fs_info = bg->fs_info;
2198
2199         spin_lock(&fs_info->relocation_bg_lock);
2200         if (fs_info->data_reloc_bg == bg->start)
2201                 fs_info->data_reloc_bg = 0;
2202         spin_unlock(&fs_info->relocation_bg_lock);
2203 }
2204
2205 void btrfs_free_zone_cache(struct btrfs_fs_info *fs_info)
2206 {
2207         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2208         struct btrfs_device *device;
2209
2210         if (!btrfs_is_zoned(fs_info))
2211                 return;
2212
2213         mutex_lock(&fs_devices->device_list_mutex);
2214         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2215                 if (device->zone_info) {
2216                         vfree(device->zone_info->zone_cache);
2217                         device->zone_info->zone_cache = NULL;
2218                 }
2219         }
2220         mutex_unlock(&fs_devices->device_list_mutex);
2221 }
2222
2223 bool btrfs_zoned_should_reclaim(struct btrfs_fs_info *fs_info)
2224 {
2225         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
2226         struct btrfs_device *device;
2227         u64 used = 0;
2228         u64 total = 0;
2229         u64 factor;
2230
2231         ASSERT(btrfs_is_zoned(fs_info));
2232
2233         if (fs_info->bg_reclaim_threshold == 0)
2234                 return false;
2235
2236         mutex_lock(&fs_devices->device_list_mutex);
2237         list_for_each_entry(device, &fs_devices->devices, dev_list) {
2238                 if (!device->bdev)
2239                         continue;
2240
2241                 total += device->disk_total_bytes;
2242                 used += device->bytes_used;
2243         }
2244         mutex_unlock(&fs_devices->device_list_mutex);
2245
2246         factor = div64_u64(used * 100, total);
2247         return factor >= fs_info->bg_reclaim_threshold;
2248 }
2249
2250 void btrfs_zoned_release_data_reloc_bg(struct btrfs_fs_info *fs_info, u64 logical,
2251                                        u64 length)
2252 {
2253         struct btrfs_block_group *block_group;
2254
2255         if (!btrfs_is_zoned(fs_info))
2256                 return;
2257
2258         block_group = btrfs_lookup_block_group(fs_info, logical);
2259         /* It should be called on a previous data relocation block group. */
2260         ASSERT(block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA));
2261
2262         spin_lock(&block_group->lock);
2263         if (!test_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC, &block_group->runtime_flags))
2264                 goto out;
2265
2266         /* All relocation extents are written. */
2267         if (block_group->start + block_group->alloc_offset == logical + length) {
2268                 /* Now, release this block group for further allocations. */
2269                 clear_bit(BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
2270                           &block_group->runtime_flags);
2271         }
2272
2273 out:
2274         spin_unlock(&block_group->lock);
2275         btrfs_put_block_group(block_group);
2276 }
2277
2278 int btrfs_zone_finish_one_bg(struct btrfs_fs_info *fs_info)
2279 {
2280         struct btrfs_block_group *block_group;
2281         struct btrfs_block_group *min_bg = NULL;
2282         u64 min_avail = U64_MAX;
2283         int ret;
2284
2285         spin_lock(&fs_info->zone_active_bgs_lock);
2286         list_for_each_entry(block_group, &fs_info->zone_active_bgs,
2287                             active_bg_list) {
2288                 u64 avail;
2289
2290                 spin_lock(&block_group->lock);
2291                 if (block_group->reserved || block_group->alloc_offset == 0 ||
2292                     (block_group->flags & BTRFS_BLOCK_GROUP_SYSTEM)) {
2293                         spin_unlock(&block_group->lock);
2294                         continue;
2295                 }
2296
2297                 avail = block_group->zone_capacity - block_group->alloc_offset;
2298                 if (min_avail > avail) {
2299                         if (min_bg)
2300                                 btrfs_put_block_group(min_bg);
2301                         min_bg = block_group;
2302                         min_avail = avail;
2303                         btrfs_get_block_group(min_bg);
2304                 }
2305                 spin_unlock(&block_group->lock);
2306         }
2307         spin_unlock(&fs_info->zone_active_bgs_lock);
2308
2309         if (!min_bg)
2310                 return 0;
2311
2312         ret = btrfs_zone_finish(min_bg);
2313         btrfs_put_block_group(min_bg);
2314
2315         return ret < 0 ? ret : 1;
2316 }
2317
2318 int btrfs_zoned_activate_one_bg(struct btrfs_fs_info *fs_info,
2319                                 struct btrfs_space_info *space_info,
2320                                 bool do_finish)
2321 {
2322         struct btrfs_block_group *bg;
2323         int index;
2324
2325         if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
2326                 return 0;
2327
2328         for (;;) {
2329                 int ret;
2330                 bool need_finish = false;
2331
2332                 down_read(&space_info->groups_sem);
2333                 for (index = 0; index < BTRFS_NR_RAID_TYPES; index++) {
2334                         list_for_each_entry(bg, &space_info->block_groups[index],
2335                                             list) {
2336                                 if (!spin_trylock(&bg->lock))
2337                                         continue;
2338                                 if (btrfs_zoned_bg_is_full(bg) ||
2339                                     test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2340                                              &bg->runtime_flags)) {
2341                                         spin_unlock(&bg->lock);
2342                                         continue;
2343                                 }
2344                                 spin_unlock(&bg->lock);
2345
2346                                 if (btrfs_zone_activate(bg)) {
2347                                         up_read(&space_info->groups_sem);
2348                                         return 1;
2349                                 }
2350
2351                                 need_finish = true;
2352                         }
2353                 }
2354                 up_read(&space_info->groups_sem);
2355
2356                 if (!do_finish || !need_finish)
2357                         break;
2358
2359                 ret = btrfs_zone_finish_one_bg(fs_info);
2360                 if (ret == 0)
2361                         break;
2362                 if (ret < 0)
2363                         return ret;
2364         }
2365
2366         return 0;
2367 }