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