16cbd5570SChris Mason /* 26cbd5570SChris Mason * Copyright (C) 2007 Oracle. All rights reserved. 36cbd5570SChris Mason * 46cbd5570SChris Mason * This program is free software; you can redistribute it and/or 56cbd5570SChris Mason * modify it under the terms of the GNU General Public 66cbd5570SChris Mason * License v2 as published by the Free Software Foundation. 76cbd5570SChris Mason * 86cbd5570SChris Mason * This program is distributed in the hope that it will be useful, 96cbd5570SChris Mason * but WITHOUT ANY WARRANTY; without even the implied warranty of 106cbd5570SChris Mason * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 116cbd5570SChris Mason * General Public License for more details. 126cbd5570SChris Mason * 136cbd5570SChris Mason * You should have received a copy of the GNU General Public 146cbd5570SChris Mason * License along with this program; if not, write to the 156cbd5570SChris Mason * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 166cbd5570SChris Mason * Boston, MA 021110-1307, USA. 176cbd5570SChris Mason */ 186cbd5570SChris Mason 19e20d96d6SChris Mason #include <linux/fs.h> 20d98237b3SChris Mason #include <linux/blkdev.h> 2187cbda5cSChris Mason #include <linux/scatterlist.h> 2222b0ebdaSChris Mason #include <linux/swap.h> 230f7d52f4SChris Mason #include <linux/radix-tree.h> 2435b7e476SChris Mason #include <linux/writeback.h> 25d397712bSChris Mason #include <linux/buffer_head.h> 26ce9adaa5SChris Mason #include <linux/workqueue.h> 27a74a4b97SChris Mason #include <linux/kthread.h> 285a0e3ad6STejun Heo #include <linux/slab.h> 29784b4e29SChris Mason #include <linux/migrate.h> 307a36ddecSDavid Sterba #include <linux/ratelimit.h> 316463fe58SStefan Behrens #include <linux/uuid.h> 32803b2f54SStefan Behrens #include <linux/semaphore.h> 337e75bf3fSDavid Sterba #include <asm/unaligned.h> 34eb60ceacSChris Mason #include "ctree.h" 35eb60ceacSChris Mason #include "disk-io.h" 360b947affSFilipe David Borba Manana #include "hash.h" 37e089f05cSChris Mason #include "transaction.h" 380f7d52f4SChris Mason #include "btrfs_inode.h" 390b86a832SChris Mason #include "volumes.h" 40db94535dSChris Mason #include "print-tree.h" 41925baeddSChris Mason #include "locking.h" 42e02119d5SChris Mason #include "tree-log.h" 43fa9c0d79SChris Mason #include "free-space-cache.h" 4470f6d82eSOmar Sandoval #include "free-space-tree.h" 45581bb050SLi Zefan #include "inode-map.h" 4621adbd5cSStefan Behrens #include "check-integrity.h" 47606686eeSJosef Bacik #include "rcu-string.h" 488dabb742SStefan Behrens #include "dev-replace.h" 4953b381b3SDavid Woodhouse #include "raid56.h" 505ac1d209SJeff Mahoney #include "sysfs.h" 51fcebe456SJosef Bacik #include "qgroup.h" 52ebb8765bSAnand Jain #include "compression.h" 53eb60ceacSChris Mason 54de0022b9SJosef Bacik #ifdef CONFIG_X86 55de0022b9SJosef Bacik #include <asm/cpufeature.h> 56de0022b9SJosef Bacik #endif 57de0022b9SJosef Bacik 58319e4d06SQu Wenruo #define BTRFS_SUPER_FLAG_SUPP (BTRFS_HEADER_FLAG_WRITTEN |\ 59319e4d06SQu Wenruo BTRFS_HEADER_FLAG_RELOC |\ 60319e4d06SQu Wenruo BTRFS_SUPER_FLAG_ERROR |\ 61319e4d06SQu Wenruo BTRFS_SUPER_FLAG_SEEDING |\ 62319e4d06SQu Wenruo BTRFS_SUPER_FLAG_METADUMP) 63319e4d06SQu Wenruo 64e8c9f186SDavid Sterba static const struct extent_io_ops btree_extent_io_ops; 658b712842SChris Mason static void end_workqueue_fn(struct btrfs_work *work); 664df27c4dSYan, Zheng static void free_fs_root(struct btrfs_root *root); 67fcd1f065SDavid Sterba static int btrfs_check_super_valid(struct btrfs_fs_info *fs_info, 68acce952bSliubo int read_only); 69143bede5SJeff Mahoney static void btrfs_destroy_ordered_extents(struct btrfs_root *root); 70acce952bSliubo static int btrfs_destroy_delayed_refs(struct btrfs_transaction *trans, 71acce952bSliubo struct btrfs_root *root); 72143bede5SJeff Mahoney static void btrfs_destroy_delalloc_inodes(struct btrfs_root *root); 73acce952bSliubo static int btrfs_destroy_marked_extents(struct btrfs_root *root, 74acce952bSliubo struct extent_io_tree *dirty_pages, 75acce952bSliubo int mark); 76acce952bSliubo static int btrfs_destroy_pinned_extent(struct btrfs_root *root, 77acce952bSliubo struct extent_io_tree *pinned_extents); 7848a3b636SEric Sandeen static int btrfs_cleanup_transaction(struct btrfs_root *root); 7948a3b636SEric Sandeen static void btrfs_error_commit_super(struct btrfs_root *root); 80ce9adaa5SChris Mason 81d352ac68SChris Mason /* 8297eb6b69SDavid Sterba * btrfs_end_io_wq structs are used to do processing in task context when an IO 8397eb6b69SDavid Sterba * is complete. This is used during reads to verify checksums, and it is used 84d352ac68SChris Mason * by writes to insert metadata for new file extents after IO is complete. 85d352ac68SChris Mason */ 8697eb6b69SDavid Sterba struct btrfs_end_io_wq { 87ce9adaa5SChris Mason struct bio *bio; 88ce9adaa5SChris Mason bio_end_io_t *end_io; 89ce9adaa5SChris Mason void *private; 90ce9adaa5SChris Mason struct btrfs_fs_info *info; 91ce9adaa5SChris Mason int error; 92bfebd8b5SDavid Sterba enum btrfs_wq_endio_type metadata; 93ce9adaa5SChris Mason struct list_head list; 948b712842SChris Mason struct btrfs_work work; 95ce9adaa5SChris Mason }; 960da5468fSChris Mason 9797eb6b69SDavid Sterba static struct kmem_cache *btrfs_end_io_wq_cache; 9897eb6b69SDavid Sterba 9997eb6b69SDavid Sterba int __init btrfs_end_io_wq_init(void) 10097eb6b69SDavid Sterba { 10197eb6b69SDavid Sterba btrfs_end_io_wq_cache = kmem_cache_create("btrfs_end_io_wq", 10297eb6b69SDavid Sterba sizeof(struct btrfs_end_io_wq), 10397eb6b69SDavid Sterba 0, 104fba4b697SNikolay Borisov SLAB_MEM_SPREAD, 10597eb6b69SDavid Sterba NULL); 10697eb6b69SDavid Sterba if (!btrfs_end_io_wq_cache) 10797eb6b69SDavid Sterba return -ENOMEM; 10897eb6b69SDavid Sterba return 0; 10997eb6b69SDavid Sterba } 11097eb6b69SDavid Sterba 11197eb6b69SDavid Sterba void btrfs_end_io_wq_exit(void) 11297eb6b69SDavid Sterba { 11397eb6b69SDavid Sterba kmem_cache_destroy(btrfs_end_io_wq_cache); 11497eb6b69SDavid Sterba } 11597eb6b69SDavid Sterba 116d352ac68SChris Mason /* 117d352ac68SChris Mason * async submit bios are used to offload expensive checksumming 118d352ac68SChris Mason * onto the worker threads. They checksum file and metadata bios 119d352ac68SChris Mason * just before they are sent down the IO stack. 120d352ac68SChris Mason */ 12144b8bd7eSChris Mason struct async_submit_bio { 12244b8bd7eSChris Mason struct inode *inode; 12344b8bd7eSChris Mason struct bio *bio; 12444b8bd7eSChris Mason struct list_head list; 1254a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_start; 1264a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_done; 12744b8bd7eSChris Mason int mirror_num; 128c8b97818SChris Mason unsigned long bio_flags; 129eaf25d93SChris Mason /* 130eaf25d93SChris Mason * bio_offset is optional, can be used if the pages in the bio 131eaf25d93SChris Mason * can't tell us where in the file the bio should go 132eaf25d93SChris Mason */ 133eaf25d93SChris Mason u64 bio_offset; 1348b712842SChris Mason struct btrfs_work work; 13579787eaaSJeff Mahoney int error; 13644b8bd7eSChris Mason }; 13744b8bd7eSChris Mason 13885d4e461SChris Mason /* 13985d4e461SChris Mason * Lockdep class keys for extent_buffer->lock's in this root. For a given 14085d4e461SChris Mason * eb, the lockdep key is determined by the btrfs_root it belongs to and 14185d4e461SChris Mason * the level the eb occupies in the tree. 1424008c04aSChris Mason * 14385d4e461SChris Mason * Different roots are used for different purposes and may nest inside each 14485d4e461SChris Mason * other and they require separate keysets. As lockdep keys should be 14585d4e461SChris Mason * static, assign keysets according to the purpose of the root as indicated 14685d4e461SChris Mason * by btrfs_root->objectid. This ensures that all special purpose roots 14785d4e461SChris Mason * have separate keysets. 1484008c04aSChris Mason * 14985d4e461SChris Mason * Lock-nesting across peer nodes is always done with the immediate parent 15085d4e461SChris Mason * node locked thus preventing deadlock. As lockdep doesn't know this, use 15185d4e461SChris Mason * subclass to avoid triggering lockdep warning in such cases. 1524008c04aSChris Mason * 15385d4e461SChris Mason * The key is set by the readpage_end_io_hook after the buffer has passed 15485d4e461SChris Mason * csum validation but before the pages are unlocked. It is also set by 15585d4e461SChris Mason * btrfs_init_new_buffer on freshly allocated blocks. 15685d4e461SChris Mason * 15785d4e461SChris Mason * We also add a check to make sure the highest level of the tree is the 15885d4e461SChris Mason * same as our lockdep setup here. If BTRFS_MAX_LEVEL changes, this code 15985d4e461SChris Mason * needs update as well. 1604008c04aSChris Mason */ 1614008c04aSChris Mason #ifdef CONFIG_DEBUG_LOCK_ALLOC 1624008c04aSChris Mason # if BTRFS_MAX_LEVEL != 8 1634008c04aSChris Mason # error 1644008c04aSChris Mason # endif 16585d4e461SChris Mason 16685d4e461SChris Mason static struct btrfs_lockdep_keyset { 16785d4e461SChris Mason u64 id; /* root objectid */ 16885d4e461SChris Mason const char *name_stem; /* lock name stem */ 16985d4e461SChris Mason char names[BTRFS_MAX_LEVEL + 1][20]; 17085d4e461SChris Mason struct lock_class_key keys[BTRFS_MAX_LEVEL + 1]; 17185d4e461SChris Mason } btrfs_lockdep_keysets[] = { 17285d4e461SChris Mason { .id = BTRFS_ROOT_TREE_OBJECTID, .name_stem = "root" }, 17385d4e461SChris Mason { .id = BTRFS_EXTENT_TREE_OBJECTID, .name_stem = "extent" }, 17485d4e461SChris Mason { .id = BTRFS_CHUNK_TREE_OBJECTID, .name_stem = "chunk" }, 17585d4e461SChris Mason { .id = BTRFS_DEV_TREE_OBJECTID, .name_stem = "dev" }, 17685d4e461SChris Mason { .id = BTRFS_FS_TREE_OBJECTID, .name_stem = "fs" }, 17785d4e461SChris Mason { .id = BTRFS_CSUM_TREE_OBJECTID, .name_stem = "csum" }, 17860b62978SDavid Sterba { .id = BTRFS_QUOTA_TREE_OBJECTID, .name_stem = "quota" }, 17985d4e461SChris Mason { .id = BTRFS_TREE_LOG_OBJECTID, .name_stem = "log" }, 18085d4e461SChris Mason { .id = BTRFS_TREE_RELOC_OBJECTID, .name_stem = "treloc" }, 18185d4e461SChris Mason { .id = BTRFS_DATA_RELOC_TREE_OBJECTID, .name_stem = "dreloc" }, 18213fd8da9SDavid Sterba { .id = BTRFS_UUID_TREE_OBJECTID, .name_stem = "uuid" }, 1836b20e0adSDavid Sterba { .id = BTRFS_FREE_SPACE_TREE_OBJECTID, .name_stem = "free-space" }, 18485d4e461SChris Mason { .id = 0, .name_stem = "tree" }, 1854008c04aSChris Mason }; 18685d4e461SChris Mason 18785d4e461SChris Mason void __init btrfs_init_lockdep(void) 18885d4e461SChris Mason { 18985d4e461SChris Mason int i, j; 19085d4e461SChris Mason 19185d4e461SChris Mason /* initialize lockdep class names */ 19285d4e461SChris Mason for (i = 0; i < ARRAY_SIZE(btrfs_lockdep_keysets); i++) { 19385d4e461SChris Mason struct btrfs_lockdep_keyset *ks = &btrfs_lockdep_keysets[i]; 19485d4e461SChris Mason 19585d4e461SChris Mason for (j = 0; j < ARRAY_SIZE(ks->names); j++) 19685d4e461SChris Mason snprintf(ks->names[j], sizeof(ks->names[j]), 19785d4e461SChris Mason "btrfs-%s-%02d", ks->name_stem, j); 19885d4e461SChris Mason } 19985d4e461SChris Mason } 20085d4e461SChris Mason 20185d4e461SChris Mason void btrfs_set_buffer_lockdep_class(u64 objectid, struct extent_buffer *eb, 20285d4e461SChris Mason int level) 20385d4e461SChris Mason { 20485d4e461SChris Mason struct btrfs_lockdep_keyset *ks; 20585d4e461SChris Mason 20685d4e461SChris Mason BUG_ON(level >= ARRAY_SIZE(ks->keys)); 20785d4e461SChris Mason 20885d4e461SChris Mason /* find the matching keyset, id 0 is the default entry */ 20985d4e461SChris Mason for (ks = btrfs_lockdep_keysets; ks->id; ks++) 21085d4e461SChris Mason if (ks->id == objectid) 21185d4e461SChris Mason break; 21285d4e461SChris Mason 21385d4e461SChris Mason lockdep_set_class_and_name(&eb->lock, 21485d4e461SChris Mason &ks->keys[level], ks->names[level]); 21585d4e461SChris Mason } 21685d4e461SChris Mason 2174008c04aSChris Mason #endif 2184008c04aSChris Mason 219d352ac68SChris Mason /* 220d352ac68SChris Mason * extents on the btree inode are pretty simple, there's one extent 221d352ac68SChris Mason * that covers the entire device 222d352ac68SChris Mason */ 223b2950863SChristoph Hellwig static struct extent_map *btree_get_extent(struct inode *inode, 224306e16ceSDavid Sterba struct page *page, size_t pg_offset, u64 start, u64 len, 2255f39d397SChris Mason int create) 2265f39d397SChris Mason { 2270b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 2285f39d397SChris Mason struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 2295f39d397SChris Mason struct extent_map *em; 2305f39d397SChris Mason int ret; 2315f39d397SChris Mason 232890871beSChris Mason read_lock(&em_tree->lock); 233d1310b2eSChris Mason em = lookup_extent_mapping(em_tree, start, len); 234a061fc8dSChris Mason if (em) { 2350b246afaSJeff Mahoney em->bdev = fs_info->fs_devices->latest_bdev; 236890871beSChris Mason read_unlock(&em_tree->lock); 2375f39d397SChris Mason goto out; 238a061fc8dSChris Mason } 239890871beSChris Mason read_unlock(&em_tree->lock); 2407b13b7b1SChris Mason 241172ddd60SDavid Sterba em = alloc_extent_map(); 2425f39d397SChris Mason if (!em) { 2435f39d397SChris Mason em = ERR_PTR(-ENOMEM); 2445f39d397SChris Mason goto out; 2455f39d397SChris Mason } 2465f39d397SChris Mason em->start = 0; 2470afbaf8cSChris Mason em->len = (u64)-1; 248c8b97818SChris Mason em->block_len = (u64)-1; 2495f39d397SChris Mason em->block_start = 0; 2500b246afaSJeff Mahoney em->bdev = fs_info->fs_devices->latest_bdev; 251d1310b2eSChris Mason 252890871beSChris Mason write_lock(&em_tree->lock); 25309a2a8f9SJosef Bacik ret = add_extent_mapping(em_tree, em, 0); 2545f39d397SChris Mason if (ret == -EEXIST) { 2555f39d397SChris Mason free_extent_map(em); 2567b13b7b1SChris Mason em = lookup_extent_mapping(em_tree, start, len); 257b4f359abSTsutomu Itoh if (!em) 2580433f20dSTsutomu Itoh em = ERR_PTR(-EIO); 2595f39d397SChris Mason } else if (ret) { 2607b13b7b1SChris Mason free_extent_map(em); 2610433f20dSTsutomu Itoh em = ERR_PTR(ret); 2625f39d397SChris Mason } 263890871beSChris Mason write_unlock(&em_tree->lock); 2647b13b7b1SChris Mason 2655f39d397SChris Mason out: 2665f39d397SChris Mason return em; 2675f39d397SChris Mason } 2685f39d397SChris Mason 269b0496686SLiu Bo u32 btrfs_csum_data(char *data, u32 seed, size_t len) 27019c00ddcSChris Mason { 2710b947affSFilipe David Borba Manana return btrfs_crc32c(seed, data, len); 27219c00ddcSChris Mason } 27319c00ddcSChris Mason 2740b5e3dafSDomagoj Tršan void btrfs_csum_final(u32 crc, u8 *result) 27519c00ddcSChris Mason { 2767e75bf3fSDavid Sterba put_unaligned_le32(~crc, result); 27719c00ddcSChris Mason } 27819c00ddcSChris Mason 279d352ac68SChris Mason /* 280d352ac68SChris Mason * compute the csum for a btree block, and either verify it or write it 281d352ac68SChris Mason * into the csum field of the block. 282d352ac68SChris Mason */ 28301d58472SDaniel Dressler static int csum_tree_block(struct btrfs_fs_info *fs_info, 28401d58472SDaniel Dressler struct extent_buffer *buf, 28519c00ddcSChris Mason int verify) 28619c00ddcSChris Mason { 28701d58472SDaniel Dressler u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); 288607d432dSJosef Bacik char *result = NULL; 28919c00ddcSChris Mason unsigned long len; 29019c00ddcSChris Mason unsigned long cur_len; 29119c00ddcSChris Mason unsigned long offset = BTRFS_CSUM_SIZE; 29219c00ddcSChris Mason char *kaddr; 29319c00ddcSChris Mason unsigned long map_start; 29419c00ddcSChris Mason unsigned long map_len; 29519c00ddcSChris Mason int err; 29619c00ddcSChris Mason u32 crc = ~(u32)0; 297607d432dSJosef Bacik unsigned long inline_result; 29819c00ddcSChris Mason 29919c00ddcSChris Mason len = buf->len - offset; 30019c00ddcSChris Mason while (len > 0) { 30119c00ddcSChris Mason err = map_private_extent_buffer(buf, offset, 32, 302a6591715SChris Mason &kaddr, &map_start, &map_len); 303d397712bSChris Mason if (err) 3048bd98f0eSAlex Lyakas return err; 30519c00ddcSChris Mason cur_len = min(len, map_len - (offset - map_start)); 306b0496686SLiu Bo crc = btrfs_csum_data(kaddr + offset - map_start, 30719c00ddcSChris Mason crc, cur_len); 30819c00ddcSChris Mason len -= cur_len; 30919c00ddcSChris Mason offset += cur_len; 31019c00ddcSChris Mason } 311607d432dSJosef Bacik if (csum_size > sizeof(inline_result)) { 31231e818feSDavid Sterba result = kzalloc(csum_size, GFP_NOFS); 313607d432dSJosef Bacik if (!result) 3148bd98f0eSAlex Lyakas return -ENOMEM; 315607d432dSJosef Bacik } else { 316607d432dSJosef Bacik result = (char *)&inline_result; 317607d432dSJosef Bacik } 318607d432dSJosef Bacik 31919c00ddcSChris Mason btrfs_csum_final(crc, result); 32019c00ddcSChris Mason 32119c00ddcSChris Mason if (verify) { 322607d432dSJosef Bacik if (memcmp_extent_buffer(buf, result, 0, csum_size)) { 323e4204dedSChris Mason u32 val; 324e4204dedSChris Mason u32 found = 0; 325607d432dSJosef Bacik memcpy(&found, result, csum_size); 326e4204dedSChris Mason 327607d432dSJosef Bacik read_extent_buffer(buf, &val, 0, csum_size); 32894647322SDavid Sterba btrfs_warn_rl(fs_info, 3295d163e0eSJeff Mahoney "%s checksum verify failed on %llu wanted %X found %X level %d", 33001d58472SDaniel Dressler fs_info->sb->s_id, buf->start, 331c1c9ff7cSGeert Uytterhoeven val, found, btrfs_header_level(buf)); 332607d432dSJosef Bacik if (result != (char *)&inline_result) 333607d432dSJosef Bacik kfree(result); 3348bd98f0eSAlex Lyakas return -EUCLEAN; 33519c00ddcSChris Mason } 33619c00ddcSChris Mason } else { 337607d432dSJosef Bacik write_extent_buffer(buf, result, 0, csum_size); 33819c00ddcSChris Mason } 339607d432dSJosef Bacik if (result != (char *)&inline_result) 340607d432dSJosef Bacik kfree(result); 34119c00ddcSChris Mason return 0; 34219c00ddcSChris Mason } 34319c00ddcSChris Mason 344d352ac68SChris Mason /* 345d352ac68SChris Mason * we can't consider a given block up to date unless the transid of the 346d352ac68SChris Mason * block matches the transid in the parent node's pointer. This is how we 347d352ac68SChris Mason * detect blocks that either didn't get written at all or got written 348d352ac68SChris Mason * in the wrong place. 349d352ac68SChris Mason */ 3501259ab75SChris Mason static int verify_parent_transid(struct extent_io_tree *io_tree, 351b9fab919SChris Mason struct extent_buffer *eb, u64 parent_transid, 352b9fab919SChris Mason int atomic) 3531259ab75SChris Mason { 3542ac55d41SJosef Bacik struct extent_state *cached_state = NULL; 3551259ab75SChris Mason int ret; 3562755a0deSDavid Sterba bool need_lock = (current->journal_info == BTRFS_SEND_TRANS_STUB); 3571259ab75SChris Mason 3581259ab75SChris Mason if (!parent_transid || btrfs_header_generation(eb) == parent_transid) 3591259ab75SChris Mason return 0; 3601259ab75SChris Mason 361b9fab919SChris Mason if (atomic) 362b9fab919SChris Mason return -EAGAIN; 363b9fab919SChris Mason 364a26e8c9fSJosef Bacik if (need_lock) { 365a26e8c9fSJosef Bacik btrfs_tree_read_lock(eb); 366a26e8c9fSJosef Bacik btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK); 367a26e8c9fSJosef Bacik } 368a26e8c9fSJosef Bacik 3692ac55d41SJosef Bacik lock_extent_bits(io_tree, eb->start, eb->start + eb->len - 1, 370ff13db41SDavid Sterba &cached_state); 3710b32f4bbSJosef Bacik if (extent_buffer_uptodate(eb) && 3721259ab75SChris Mason btrfs_header_generation(eb) == parent_transid) { 3731259ab75SChris Mason ret = 0; 3741259ab75SChris Mason goto out; 3751259ab75SChris Mason } 37694647322SDavid Sterba btrfs_err_rl(eb->fs_info, 37794647322SDavid Sterba "parent transid verify failed on %llu wanted %llu found %llu", 37894647322SDavid Sterba eb->start, 37929549aecSWang Shilong parent_transid, btrfs_header_generation(eb)); 3801259ab75SChris Mason ret = 1; 381a26e8c9fSJosef Bacik 382a26e8c9fSJosef Bacik /* 383a26e8c9fSJosef Bacik * Things reading via commit roots that don't have normal protection, 384a26e8c9fSJosef Bacik * like send, can have a really old block in cache that may point at a 38501327610SNicholas D Steeves * block that has been freed and re-allocated. So don't clear uptodate 386a26e8c9fSJosef Bacik * if we find an eb that is under IO (dirty/writeback) because we could 387a26e8c9fSJosef Bacik * end up reading in the stale data and then writing it back out and 388a26e8c9fSJosef Bacik * making everybody very sad. 389a26e8c9fSJosef Bacik */ 390a26e8c9fSJosef Bacik if (!extent_buffer_under_io(eb)) 3910b32f4bbSJosef Bacik clear_extent_buffer_uptodate(eb); 39233958dc6SChris Mason out: 3932ac55d41SJosef Bacik unlock_extent_cached(io_tree, eb->start, eb->start + eb->len - 1, 3942ac55d41SJosef Bacik &cached_state, GFP_NOFS); 395472b909fSJosef Bacik if (need_lock) 396a26e8c9fSJosef Bacik btrfs_tree_read_unlock_blocking(eb); 3971259ab75SChris Mason return ret; 3981259ab75SChris Mason } 3991259ab75SChris Mason 400d352ac68SChris Mason /* 4011104a885SDavid Sterba * Return 0 if the superblock checksum type matches the checksum value of that 4021104a885SDavid Sterba * algorithm. Pass the raw disk superblock data. 4031104a885SDavid Sterba */ 404ab8d0fc4SJeff Mahoney static int btrfs_check_super_csum(struct btrfs_fs_info *fs_info, 405ab8d0fc4SJeff Mahoney char *raw_disk_sb) 4061104a885SDavid Sterba { 4071104a885SDavid Sterba struct btrfs_super_block *disk_sb = 4081104a885SDavid Sterba (struct btrfs_super_block *)raw_disk_sb; 4091104a885SDavid Sterba u16 csum_type = btrfs_super_csum_type(disk_sb); 4101104a885SDavid Sterba int ret = 0; 4111104a885SDavid Sterba 4121104a885SDavid Sterba if (csum_type == BTRFS_CSUM_TYPE_CRC32) { 4131104a885SDavid Sterba u32 crc = ~(u32)0; 4141104a885SDavid Sterba const int csum_size = sizeof(crc); 4151104a885SDavid Sterba char result[csum_size]; 4161104a885SDavid Sterba 4171104a885SDavid Sterba /* 4181104a885SDavid Sterba * The super_block structure does not span the whole 4191104a885SDavid Sterba * BTRFS_SUPER_INFO_SIZE range, we expect that the unused space 42001327610SNicholas D Steeves * is filled with zeros and is included in the checksum. 4211104a885SDavid Sterba */ 4221104a885SDavid Sterba crc = btrfs_csum_data(raw_disk_sb + BTRFS_CSUM_SIZE, 4231104a885SDavid Sterba crc, BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE); 4241104a885SDavid Sterba btrfs_csum_final(crc, result); 4251104a885SDavid Sterba 4261104a885SDavid Sterba if (memcmp(raw_disk_sb, result, csum_size)) 4271104a885SDavid Sterba ret = 1; 4281104a885SDavid Sterba } 4291104a885SDavid Sterba 4301104a885SDavid Sterba if (csum_type >= ARRAY_SIZE(btrfs_csum_sizes)) { 431ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "unsupported checksum algorithm %u", 4321104a885SDavid Sterba csum_type); 4331104a885SDavid Sterba ret = 1; 4341104a885SDavid Sterba } 4351104a885SDavid Sterba 4361104a885SDavid Sterba return ret; 4371104a885SDavid Sterba } 4381104a885SDavid Sterba 4391104a885SDavid Sterba /* 440d352ac68SChris Mason * helper to read a given tree block, doing retries as required when 441d352ac68SChris Mason * the checksums don't match and we have alternate mirrors to try. 442d352ac68SChris Mason */ 443f188591eSChris Mason static int btree_read_extent_buffer_pages(struct btrfs_root *root, 444f188591eSChris Mason struct extent_buffer *eb, 4458436ea91SJosef Bacik u64 parent_transid) 446f188591eSChris Mason { 4470b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 448f188591eSChris Mason struct extent_io_tree *io_tree; 449ea466794SJosef Bacik int failed = 0; 450f188591eSChris Mason int ret; 451f188591eSChris Mason int num_copies = 0; 452f188591eSChris Mason int mirror_num = 0; 453ea466794SJosef Bacik int failed_mirror = 0; 454f188591eSChris Mason 455a826d6dcSJosef Bacik clear_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags); 4560b246afaSJeff Mahoney io_tree = &BTRFS_I(fs_info->btree_inode)->io_tree; 457f188591eSChris Mason while (1) { 4588436ea91SJosef Bacik ret = read_extent_buffer_pages(io_tree, eb, WAIT_COMPLETE, 459f188591eSChris Mason btree_get_extent, mirror_num); 460256dd1bbSStefan Behrens if (!ret) { 461256dd1bbSStefan Behrens if (!verify_parent_transid(io_tree, eb, 462b9fab919SChris Mason parent_transid, 0)) 463ea466794SJosef Bacik break; 464256dd1bbSStefan Behrens else 465256dd1bbSStefan Behrens ret = -EIO; 466256dd1bbSStefan Behrens } 467d397712bSChris Mason 468a826d6dcSJosef Bacik /* 469a826d6dcSJosef Bacik * This buffer's crc is fine, but its contents are corrupted, so 470a826d6dcSJosef Bacik * there is no reason to read the other copies, they won't be 471a826d6dcSJosef Bacik * any less wrong. 472a826d6dcSJosef Bacik */ 473a826d6dcSJosef Bacik if (test_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags)) 474ea466794SJosef Bacik break; 475ea466794SJosef Bacik 4760b246afaSJeff Mahoney num_copies = btrfs_num_copies(fs_info, 477f188591eSChris Mason eb->start, eb->len); 4784235298eSChris Mason if (num_copies == 1) 479ea466794SJosef Bacik break; 4804235298eSChris Mason 4815cf1ab56SJosef Bacik if (!failed_mirror) { 4825cf1ab56SJosef Bacik failed = 1; 4835cf1ab56SJosef Bacik failed_mirror = eb->read_mirror; 4845cf1ab56SJosef Bacik } 4855cf1ab56SJosef Bacik 486f188591eSChris Mason mirror_num++; 487ea466794SJosef Bacik if (mirror_num == failed_mirror) 488ea466794SJosef Bacik mirror_num++; 489ea466794SJosef Bacik 4904235298eSChris Mason if (mirror_num > num_copies) 491ea466794SJosef Bacik break; 492f188591eSChris Mason } 493ea466794SJosef Bacik 494c0901581SStefan Behrens if (failed && !ret && failed_mirror) 495ea466794SJosef Bacik repair_eb_io_failure(root, eb, failed_mirror); 496ea466794SJosef Bacik 497ea466794SJosef Bacik return ret; 498f188591eSChris Mason } 49919c00ddcSChris Mason 500d352ac68SChris Mason /* 501d397712bSChris Mason * checksum a dirty tree block before IO. This has extra checks to make sure 502d397712bSChris Mason * we only fill in the checksum field in the first page of a multi-page block 503d352ac68SChris Mason */ 504d397712bSChris Mason 50501d58472SDaniel Dressler static int csum_dirty_buffer(struct btrfs_fs_info *fs_info, struct page *page) 50619c00ddcSChris Mason { 5074eee4fa4SMiao Xie u64 start = page_offset(page); 50819c00ddcSChris Mason u64 found_start; 50919c00ddcSChris Mason struct extent_buffer *eb; 510f188591eSChris Mason 5114f2de97aSJosef Bacik eb = (struct extent_buffer *)page->private; 5124f2de97aSJosef Bacik if (page != eb->pages[0]) 5134f2de97aSJosef Bacik return 0; 5140f805531SAlex Lyakas 51519c00ddcSChris Mason found_start = btrfs_header_bytenr(eb); 5160f805531SAlex Lyakas /* 5170f805531SAlex Lyakas * Please do not consolidate these warnings into a single if. 5180f805531SAlex Lyakas * It is useful to know what went wrong. 5190f805531SAlex Lyakas */ 5200f805531SAlex Lyakas if (WARN_ON(found_start != start)) 5210f805531SAlex Lyakas return -EUCLEAN; 5220f805531SAlex Lyakas if (WARN_ON(!PageUptodate(page))) 5230f805531SAlex Lyakas return -EUCLEAN; 5240f805531SAlex Lyakas 5250f805531SAlex Lyakas ASSERT(memcmp_extent_buffer(eb, fs_info->fsid, 5260f805531SAlex Lyakas btrfs_header_fsid(), BTRFS_FSID_SIZE) == 0); 5270f805531SAlex Lyakas 5288bd98f0eSAlex Lyakas return csum_tree_block(fs_info, eb, 0); 52919c00ddcSChris Mason } 53019c00ddcSChris Mason 53101d58472SDaniel Dressler static int check_tree_block_fsid(struct btrfs_fs_info *fs_info, 5322b82032cSYan Zheng struct extent_buffer *eb) 5332b82032cSYan Zheng { 53401d58472SDaniel Dressler struct btrfs_fs_devices *fs_devices = fs_info->fs_devices; 5352b82032cSYan Zheng u8 fsid[BTRFS_UUID_SIZE]; 5362b82032cSYan Zheng int ret = 1; 5372b82032cSYan Zheng 5380a4e5586SRoss Kirk read_extent_buffer(eb, fsid, btrfs_header_fsid(), BTRFS_FSID_SIZE); 5392b82032cSYan Zheng while (fs_devices) { 5402b82032cSYan Zheng if (!memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE)) { 5412b82032cSYan Zheng ret = 0; 5422b82032cSYan Zheng break; 5432b82032cSYan Zheng } 5442b82032cSYan Zheng fs_devices = fs_devices->seed; 5452b82032cSYan Zheng } 5462b82032cSYan Zheng return ret; 5472b82032cSYan Zheng } 5482b82032cSYan Zheng 549a826d6dcSJosef Bacik #define CORRUPT(reason, eb, root, slot) \ 5500b246afaSJeff Mahoney btrfs_crit(root->fs_info, \ 5510b246afaSJeff Mahoney "corrupt %s, %s: block=%llu, root=%llu, slot=%d", \ 5526b722c17SLiu Bo btrfs_header_level(eb) == 0 ? "leaf" : "node", \ 5536b722c17SLiu Bo reason, btrfs_header_bytenr(eb), root->objectid, slot) 554a826d6dcSJosef Bacik 555a826d6dcSJosef Bacik static noinline int check_leaf(struct btrfs_root *root, 556a826d6dcSJosef Bacik struct extent_buffer *leaf) 557a826d6dcSJosef Bacik { 5580b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 559a826d6dcSJosef Bacik struct btrfs_key key; 560a826d6dcSJosef Bacik struct btrfs_key leaf_key; 561a826d6dcSJosef Bacik u32 nritems = btrfs_header_nritems(leaf); 562a826d6dcSJosef Bacik int slot; 563a826d6dcSJosef Bacik 5641ba98d08SLiu Bo if (nritems == 0) { 5651ba98d08SLiu Bo struct btrfs_root *check_root; 5661ba98d08SLiu Bo 5671ba98d08SLiu Bo key.objectid = btrfs_header_owner(leaf); 5681ba98d08SLiu Bo key.type = BTRFS_ROOT_ITEM_KEY; 5691ba98d08SLiu Bo key.offset = (u64)-1; 5701ba98d08SLiu Bo 5710b246afaSJeff Mahoney check_root = btrfs_get_fs_root(fs_info, &key, false); 5721ba98d08SLiu Bo /* 5731ba98d08SLiu Bo * The only reason we also check NULL here is that during 5741ba98d08SLiu Bo * open_ctree() some roots has not yet been set up. 5751ba98d08SLiu Bo */ 5761ba98d08SLiu Bo if (!IS_ERR_OR_NULL(check_root)) { 5771ba98d08SLiu Bo /* if leaf is the root, then it's fine */ 5781ba98d08SLiu Bo if (leaf->start != 5791ba98d08SLiu Bo btrfs_root_bytenr(&check_root->root_item)) { 5801ba98d08SLiu Bo CORRUPT("non-root leaf's nritems is 0", 5811ba98d08SLiu Bo leaf, root, 0); 5821ba98d08SLiu Bo return -EIO; 5831ba98d08SLiu Bo } 5841ba98d08SLiu Bo } 585a826d6dcSJosef Bacik return 0; 5861ba98d08SLiu Bo } 587a826d6dcSJosef Bacik 588a826d6dcSJosef Bacik /* Check the 0 item */ 589a826d6dcSJosef Bacik if (btrfs_item_offset_nr(leaf, 0) + btrfs_item_size_nr(leaf, 0) != 5900b246afaSJeff Mahoney BTRFS_LEAF_DATA_SIZE(fs_info)) { 591a826d6dcSJosef Bacik CORRUPT("invalid item offset size pair", leaf, root, 0); 592a826d6dcSJosef Bacik return -EIO; 593a826d6dcSJosef Bacik } 594a826d6dcSJosef Bacik 595a826d6dcSJosef Bacik /* 596a826d6dcSJosef Bacik * Check to make sure each items keys are in the correct order and their 597a826d6dcSJosef Bacik * offsets make sense. We only have to loop through nritems-1 because 598a826d6dcSJosef Bacik * we check the current slot against the next slot, which verifies the 599a826d6dcSJosef Bacik * next slot's offset+size makes sense and that the current's slot 600a826d6dcSJosef Bacik * offset is correct. 601a826d6dcSJosef Bacik */ 602a826d6dcSJosef Bacik for (slot = 0; slot < nritems - 1; slot++) { 603a826d6dcSJosef Bacik btrfs_item_key_to_cpu(leaf, &leaf_key, slot); 604a826d6dcSJosef Bacik btrfs_item_key_to_cpu(leaf, &key, slot + 1); 605a826d6dcSJosef Bacik 606a826d6dcSJosef Bacik /* Make sure the keys are in the right order */ 607a826d6dcSJosef Bacik if (btrfs_comp_cpu_keys(&leaf_key, &key) >= 0) { 608a826d6dcSJosef Bacik CORRUPT("bad key order", leaf, root, slot); 609a826d6dcSJosef Bacik return -EIO; 610a826d6dcSJosef Bacik } 611a826d6dcSJosef Bacik 612a826d6dcSJosef Bacik /* 613a826d6dcSJosef Bacik * Make sure the offset and ends are right, remember that the 614a826d6dcSJosef Bacik * item data starts at the end of the leaf and grows towards the 615a826d6dcSJosef Bacik * front. 616a826d6dcSJosef Bacik */ 617a826d6dcSJosef Bacik if (btrfs_item_offset_nr(leaf, slot) != 618a826d6dcSJosef Bacik btrfs_item_end_nr(leaf, slot + 1)) { 619a826d6dcSJosef Bacik CORRUPT("slot offset bad", leaf, root, slot); 620a826d6dcSJosef Bacik return -EIO; 621a826d6dcSJosef Bacik } 622a826d6dcSJosef Bacik 623a826d6dcSJosef Bacik /* 624a826d6dcSJosef Bacik * Check to make sure that we don't point outside of the leaf, 625a826d6dcSJosef Bacik * just in case all the items are consistent to each other, but 626a826d6dcSJosef Bacik * all point outside of the leaf. 627a826d6dcSJosef Bacik */ 628a826d6dcSJosef Bacik if (btrfs_item_end_nr(leaf, slot) > 6290b246afaSJeff Mahoney BTRFS_LEAF_DATA_SIZE(fs_info)) { 630a826d6dcSJosef Bacik CORRUPT("slot end outside of leaf", leaf, root, slot); 631a826d6dcSJosef Bacik return -EIO; 632a826d6dcSJosef Bacik } 633a826d6dcSJosef Bacik } 634a826d6dcSJosef Bacik 635a826d6dcSJosef Bacik return 0; 636a826d6dcSJosef Bacik } 637a826d6dcSJosef Bacik 638053ab70fSLiu Bo static int check_node(struct btrfs_root *root, struct extent_buffer *node) 639053ab70fSLiu Bo { 640053ab70fSLiu Bo unsigned long nr = btrfs_header_nritems(node); 6416b722c17SLiu Bo struct btrfs_key key, next_key; 6426b722c17SLiu Bo int slot; 6436b722c17SLiu Bo u64 bytenr; 6446b722c17SLiu Bo int ret = 0; 645053ab70fSLiu Bo 646da17066cSJeff Mahoney if (nr == 0 || nr > BTRFS_NODEPTRS_PER_BLOCK(root->fs_info)) { 647053ab70fSLiu Bo btrfs_crit(root->fs_info, 648053ab70fSLiu Bo "corrupt node: block %llu root %llu nritems %lu", 649053ab70fSLiu Bo node->start, root->objectid, nr); 650053ab70fSLiu Bo return -EIO; 651053ab70fSLiu Bo } 6526b722c17SLiu Bo 6536b722c17SLiu Bo for (slot = 0; slot < nr - 1; slot++) { 6546b722c17SLiu Bo bytenr = btrfs_node_blockptr(node, slot); 6556b722c17SLiu Bo btrfs_node_key_to_cpu(node, &key, slot); 6566b722c17SLiu Bo btrfs_node_key_to_cpu(node, &next_key, slot + 1); 6576b722c17SLiu Bo 6586b722c17SLiu Bo if (!bytenr) { 6596b722c17SLiu Bo CORRUPT("invalid item slot", node, root, slot); 6606b722c17SLiu Bo ret = -EIO; 6616b722c17SLiu Bo goto out; 6626b722c17SLiu Bo } 6636b722c17SLiu Bo 6646b722c17SLiu Bo if (btrfs_comp_cpu_keys(&key, &next_key) >= 0) { 6656b722c17SLiu Bo CORRUPT("bad key order", node, root, slot); 6666b722c17SLiu Bo ret = -EIO; 6676b722c17SLiu Bo goto out; 6686b722c17SLiu Bo } 6696b722c17SLiu Bo } 6706b722c17SLiu Bo out: 6716b722c17SLiu Bo return ret; 672053ab70fSLiu Bo } 673053ab70fSLiu Bo 674facc8a22SMiao Xie static int btree_readpage_end_io_hook(struct btrfs_io_bio *io_bio, 675facc8a22SMiao Xie u64 phy_offset, struct page *page, 676facc8a22SMiao Xie u64 start, u64 end, int mirror) 677ce9adaa5SChris Mason { 678ce9adaa5SChris Mason u64 found_start; 679ce9adaa5SChris Mason int found_level; 680ce9adaa5SChris Mason struct extent_buffer *eb; 681ce9adaa5SChris Mason struct btrfs_root *root = BTRFS_I(page->mapping->host)->root; 68202873e43SZhao Lei struct btrfs_fs_info *fs_info = root->fs_info; 683f188591eSChris Mason int ret = 0; 684727011e0SChris Mason int reads_done; 685ce9adaa5SChris Mason 686ce9adaa5SChris Mason if (!page->private) 687ce9adaa5SChris Mason goto out; 688d397712bSChris Mason 6894f2de97aSJosef Bacik eb = (struct extent_buffer *)page->private; 690d397712bSChris Mason 6910b32f4bbSJosef Bacik /* the pending IO might have been the only thing that kept this buffer 6920b32f4bbSJosef Bacik * in memory. Make sure we have a ref for all this other checks 6930b32f4bbSJosef Bacik */ 6940b32f4bbSJosef Bacik extent_buffer_get(eb); 6950b32f4bbSJosef Bacik 6960b32f4bbSJosef Bacik reads_done = atomic_dec_and_test(&eb->io_pages); 697727011e0SChris Mason if (!reads_done) 698727011e0SChris Mason goto err; 699f188591eSChris Mason 7005cf1ab56SJosef Bacik eb->read_mirror = mirror; 701656f30dbSFilipe Manana if (test_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags)) { 702ea466794SJosef Bacik ret = -EIO; 703ea466794SJosef Bacik goto err; 704ea466794SJosef Bacik } 705ea466794SJosef Bacik 706ce9adaa5SChris Mason found_start = btrfs_header_bytenr(eb); 707727011e0SChris Mason if (found_start != eb->start) { 70802873e43SZhao Lei btrfs_err_rl(fs_info, "bad tree block start %llu %llu", 70994647322SDavid Sterba found_start, eb->start); 710f188591eSChris Mason ret = -EIO; 711ce9adaa5SChris Mason goto err; 712ce9adaa5SChris Mason } 71302873e43SZhao Lei if (check_tree_block_fsid(fs_info, eb)) { 71402873e43SZhao Lei btrfs_err_rl(fs_info, "bad fsid on block %llu", 71594647322SDavid Sterba eb->start); 7161259ab75SChris Mason ret = -EIO; 7171259ab75SChris Mason goto err; 7181259ab75SChris Mason } 719ce9adaa5SChris Mason found_level = btrfs_header_level(eb); 7201c24c3ceSJosef Bacik if (found_level >= BTRFS_MAX_LEVEL) { 72102873e43SZhao Lei btrfs_err(fs_info, "bad tree block level %d", 7221c24c3ceSJosef Bacik (int)btrfs_header_level(eb)); 7231c24c3ceSJosef Bacik ret = -EIO; 7241c24c3ceSJosef Bacik goto err; 7251c24c3ceSJosef Bacik } 726ce9adaa5SChris Mason 72785d4e461SChris Mason btrfs_set_buffer_lockdep_class(btrfs_header_owner(eb), 72885d4e461SChris Mason eb, found_level); 7294008c04aSChris Mason 73002873e43SZhao Lei ret = csum_tree_block(fs_info, eb, 1); 7318bd98f0eSAlex Lyakas if (ret) 732a826d6dcSJosef Bacik goto err; 733a826d6dcSJosef Bacik 734a826d6dcSJosef Bacik /* 735a826d6dcSJosef Bacik * If this is a leaf block and it is corrupt, set the corrupt bit so 736a826d6dcSJosef Bacik * that we don't try and read the other copies of this block, just 737a826d6dcSJosef Bacik * return -EIO. 738a826d6dcSJosef Bacik */ 739a826d6dcSJosef Bacik if (found_level == 0 && check_leaf(root, eb)) { 740a826d6dcSJosef Bacik set_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags); 741a826d6dcSJosef Bacik ret = -EIO; 742a826d6dcSJosef Bacik } 743ce9adaa5SChris Mason 744053ab70fSLiu Bo if (found_level > 0 && check_node(root, eb)) 745053ab70fSLiu Bo ret = -EIO; 746053ab70fSLiu Bo 7470b32f4bbSJosef Bacik if (!ret) 7480b32f4bbSJosef Bacik set_extent_buffer_uptodate(eb); 749ce9adaa5SChris Mason err: 75079fb65a1SJosef Bacik if (reads_done && 75179fb65a1SJosef Bacik test_and_clear_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags)) 752fc2e901fSDavid Sterba btree_readahead_hook(fs_info, eb, ret); 7534bb31e92SArne Jansen 75453b381b3SDavid Woodhouse if (ret) { 75553b381b3SDavid Woodhouse /* 75653b381b3SDavid Woodhouse * our io error hook is going to dec the io pages 75753b381b3SDavid Woodhouse * again, we have to make sure it has something 75853b381b3SDavid Woodhouse * to decrement 75953b381b3SDavid Woodhouse */ 76053b381b3SDavid Woodhouse atomic_inc(&eb->io_pages); 7610b32f4bbSJosef Bacik clear_extent_buffer_uptodate(eb); 76253b381b3SDavid Woodhouse } 7630b32f4bbSJosef Bacik free_extent_buffer(eb); 764ce9adaa5SChris Mason out: 765f188591eSChris Mason return ret; 766ce9adaa5SChris Mason } 767ce9adaa5SChris Mason 768ea466794SJosef Bacik static int btree_io_failed_hook(struct page *page, int failed_mirror) 7694bb31e92SArne Jansen { 7704bb31e92SArne Jansen struct extent_buffer *eb; 7714bb31e92SArne Jansen 7724f2de97aSJosef Bacik eb = (struct extent_buffer *)page->private; 773656f30dbSFilipe Manana set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags); 7745cf1ab56SJosef Bacik eb->read_mirror = failed_mirror; 77553b381b3SDavid Woodhouse atomic_dec(&eb->io_pages); 776ea466794SJosef Bacik if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags)) 777fc2e901fSDavid Sterba btree_readahead_hook(eb->fs_info, eb, -EIO); 7784bb31e92SArne Jansen return -EIO; /* we fixed nothing */ 7794bb31e92SArne Jansen } 7804bb31e92SArne Jansen 7814246a0b6SChristoph Hellwig static void end_workqueue_bio(struct bio *bio) 782ce9adaa5SChris Mason { 78397eb6b69SDavid Sterba struct btrfs_end_io_wq *end_io_wq = bio->bi_private; 784ce9adaa5SChris Mason struct btrfs_fs_info *fs_info; 7859e0af237SLiu Bo struct btrfs_workqueue *wq; 7869e0af237SLiu Bo btrfs_work_func_t func; 787ce9adaa5SChris Mason 788ce9adaa5SChris Mason fs_info = end_io_wq->info; 7894246a0b6SChristoph Hellwig end_io_wq->error = bio->bi_error; 790d20f7043SChris Mason 79137226b21SMike Christie if (bio_op(bio) == REQ_OP_WRITE) { 7929e0af237SLiu Bo if (end_io_wq->metadata == BTRFS_WQ_ENDIO_METADATA) { 7939e0af237SLiu Bo wq = fs_info->endio_meta_write_workers; 7949e0af237SLiu Bo func = btrfs_endio_meta_write_helper; 7959e0af237SLiu Bo } else if (end_io_wq->metadata == BTRFS_WQ_ENDIO_FREE_SPACE) { 7969e0af237SLiu Bo wq = fs_info->endio_freespace_worker; 7979e0af237SLiu Bo func = btrfs_freespace_write_helper; 7989e0af237SLiu Bo } else if (end_io_wq->metadata == BTRFS_WQ_ENDIO_RAID56) { 7999e0af237SLiu Bo wq = fs_info->endio_raid56_workers; 8009e0af237SLiu Bo func = btrfs_endio_raid56_helper; 801d20f7043SChris Mason } else { 8029e0af237SLiu Bo wq = fs_info->endio_write_workers; 8039e0af237SLiu Bo func = btrfs_endio_write_helper; 804d20f7043SChris Mason } 8059e0af237SLiu Bo } else { 8068b110e39SMiao Xie if (unlikely(end_io_wq->metadata == 8078b110e39SMiao Xie BTRFS_WQ_ENDIO_DIO_REPAIR)) { 8088b110e39SMiao Xie wq = fs_info->endio_repair_workers; 8098b110e39SMiao Xie func = btrfs_endio_repair_helper; 8108b110e39SMiao Xie } else if (end_io_wq->metadata == BTRFS_WQ_ENDIO_RAID56) { 8119e0af237SLiu Bo wq = fs_info->endio_raid56_workers; 8129e0af237SLiu Bo func = btrfs_endio_raid56_helper; 8139e0af237SLiu Bo } else if (end_io_wq->metadata) { 8149e0af237SLiu Bo wq = fs_info->endio_meta_workers; 8159e0af237SLiu Bo func = btrfs_endio_meta_helper; 8169e0af237SLiu Bo } else { 8179e0af237SLiu Bo wq = fs_info->endio_workers; 8189e0af237SLiu Bo func = btrfs_endio_helper; 8199e0af237SLiu Bo } 8209e0af237SLiu Bo } 8219e0af237SLiu Bo 8229e0af237SLiu Bo btrfs_init_work(&end_io_wq->work, func, end_workqueue_fn, NULL, NULL); 8239e0af237SLiu Bo btrfs_queue_work(wq, &end_io_wq->work); 824ce9adaa5SChris Mason } 825ce9adaa5SChris Mason 82622c59948SChris Mason int btrfs_bio_wq_end_io(struct btrfs_fs_info *info, struct bio *bio, 827bfebd8b5SDavid Sterba enum btrfs_wq_endio_type metadata) 8280b86a832SChris Mason { 82997eb6b69SDavid Sterba struct btrfs_end_io_wq *end_io_wq; 8308b110e39SMiao Xie 83197eb6b69SDavid Sterba end_io_wq = kmem_cache_alloc(btrfs_end_io_wq_cache, GFP_NOFS); 832ce9adaa5SChris Mason if (!end_io_wq) 833ce9adaa5SChris Mason return -ENOMEM; 834ce9adaa5SChris Mason 835ce9adaa5SChris Mason end_io_wq->private = bio->bi_private; 836ce9adaa5SChris Mason end_io_wq->end_io = bio->bi_end_io; 83722c59948SChris Mason end_io_wq->info = info; 838ce9adaa5SChris Mason end_io_wq->error = 0; 839ce9adaa5SChris Mason end_io_wq->bio = bio; 84022c59948SChris Mason end_io_wq->metadata = metadata; 841ce9adaa5SChris Mason 842ce9adaa5SChris Mason bio->bi_private = end_io_wq; 843ce9adaa5SChris Mason bio->bi_end_io = end_workqueue_bio; 84422c59948SChris Mason return 0; 84522c59948SChris Mason } 84622c59948SChris Mason 847b64a2851SChris Mason unsigned long btrfs_async_submit_limit(struct btrfs_fs_info *info) 8484854ddd0SChris Mason { 8494854ddd0SChris Mason unsigned long limit = min_t(unsigned long, 8505cdc7ad3SQu Wenruo info->thread_pool_size, 8514854ddd0SChris Mason info->fs_devices->open_devices); 8524854ddd0SChris Mason return 256 * limit; 8534854ddd0SChris Mason } 8544854ddd0SChris Mason 8554a69a410SChris Mason static void run_one_async_start(struct btrfs_work *work) 8564a69a410SChris Mason { 8574a69a410SChris Mason struct async_submit_bio *async; 85879787eaaSJeff Mahoney int ret; 8594a69a410SChris Mason 8604a69a410SChris Mason async = container_of(work, struct async_submit_bio, work); 86181a75f67SMike Christie ret = async->submit_bio_start(async->inode, async->bio, 862eaf25d93SChris Mason async->mirror_num, async->bio_flags, 863eaf25d93SChris Mason async->bio_offset); 86479787eaaSJeff Mahoney if (ret) 86579787eaaSJeff Mahoney async->error = ret; 8664a69a410SChris Mason } 8674a69a410SChris Mason 8684a69a410SChris Mason static void run_one_async_done(struct btrfs_work *work) 8698b712842SChris Mason { 8708b712842SChris Mason struct btrfs_fs_info *fs_info; 8718b712842SChris Mason struct async_submit_bio *async; 8724854ddd0SChris Mason int limit; 8738b712842SChris Mason 8748b712842SChris Mason async = container_of(work, struct async_submit_bio, work); 8758b712842SChris Mason fs_info = BTRFS_I(async->inode)->root->fs_info; 8764854ddd0SChris Mason 877b64a2851SChris Mason limit = btrfs_async_submit_limit(fs_info); 8784854ddd0SChris Mason limit = limit * 2 / 3; 8794854ddd0SChris Mason 880ee863954SDavid Sterba /* 881ee863954SDavid Sterba * atomic_dec_return implies a barrier for waitqueue_active 882ee863954SDavid Sterba */ 88366657b31SJosef Bacik if (atomic_dec_return(&fs_info->nr_async_submits) < limit && 884b64a2851SChris Mason waitqueue_active(&fs_info->async_submit_wait)) 8854854ddd0SChris Mason wake_up(&fs_info->async_submit_wait); 8864854ddd0SChris Mason 887bb7ab3b9SAdam Buchbinder /* If an error occurred we just want to clean up the bio and move on */ 88879787eaaSJeff Mahoney if (async->error) { 8894246a0b6SChristoph Hellwig async->bio->bi_error = async->error; 8904246a0b6SChristoph Hellwig bio_endio(async->bio); 89179787eaaSJeff Mahoney return; 89279787eaaSJeff Mahoney } 89379787eaaSJeff Mahoney 89481a75f67SMike Christie async->submit_bio_done(async->inode, async->bio, async->mirror_num, 89581a75f67SMike Christie async->bio_flags, async->bio_offset); 8964a69a410SChris Mason } 8974a69a410SChris Mason 8984a69a410SChris Mason static void run_one_async_free(struct btrfs_work *work) 8994a69a410SChris Mason { 9004a69a410SChris Mason struct async_submit_bio *async; 9014a69a410SChris Mason 9024a69a410SChris Mason async = container_of(work, struct async_submit_bio, work); 9038b712842SChris Mason kfree(async); 9048b712842SChris Mason } 9058b712842SChris Mason 90644b8bd7eSChris Mason int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode, 90781a75f67SMike Christie struct bio *bio, int mirror_num, 908c8b97818SChris Mason unsigned long bio_flags, 909eaf25d93SChris Mason u64 bio_offset, 9104a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_start, 9114a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_done) 91244b8bd7eSChris Mason { 91344b8bd7eSChris Mason struct async_submit_bio *async; 91444b8bd7eSChris Mason 91544b8bd7eSChris Mason async = kmalloc(sizeof(*async), GFP_NOFS); 91644b8bd7eSChris Mason if (!async) 91744b8bd7eSChris Mason return -ENOMEM; 91844b8bd7eSChris Mason 91944b8bd7eSChris Mason async->inode = inode; 92044b8bd7eSChris Mason async->bio = bio; 92144b8bd7eSChris Mason async->mirror_num = mirror_num; 9224a69a410SChris Mason async->submit_bio_start = submit_bio_start; 9234a69a410SChris Mason async->submit_bio_done = submit_bio_done; 9244a69a410SChris Mason 9259e0af237SLiu Bo btrfs_init_work(&async->work, btrfs_worker_helper, run_one_async_start, 9265cdc7ad3SQu Wenruo run_one_async_done, run_one_async_free); 9274a69a410SChris Mason 928c8b97818SChris Mason async->bio_flags = bio_flags; 929eaf25d93SChris Mason async->bio_offset = bio_offset; 9308c8bee1dSChris Mason 93179787eaaSJeff Mahoney async->error = 0; 93279787eaaSJeff Mahoney 933cb03c743SChris Mason atomic_inc(&fs_info->nr_async_submits); 934d313d7a3SChris Mason 9351eff9d32SJens Axboe if (bio->bi_opf & REQ_SYNC) 9365cdc7ad3SQu Wenruo btrfs_set_work_high_priority(&async->work); 937d313d7a3SChris Mason 9385cdc7ad3SQu Wenruo btrfs_queue_work(fs_info->workers, &async->work); 9399473f16cSChris Mason 940771ed689SChris Mason while (atomic_read(&fs_info->async_submit_draining) && 941771ed689SChris Mason atomic_read(&fs_info->nr_async_submits)) { 942771ed689SChris Mason wait_event(fs_info->async_submit_wait, 943771ed689SChris Mason (atomic_read(&fs_info->nr_async_submits) == 0)); 944771ed689SChris Mason } 945771ed689SChris Mason 94644b8bd7eSChris Mason return 0; 94744b8bd7eSChris Mason } 94844b8bd7eSChris Mason 949ce3ed71aSChris Mason static int btree_csum_one_bio(struct bio *bio) 950ce3ed71aSChris Mason { 9512c30c71bSKent Overstreet struct bio_vec *bvec; 952ce3ed71aSChris Mason struct btrfs_root *root; 9532c30c71bSKent Overstreet int i, ret = 0; 954ce3ed71aSChris Mason 9552c30c71bSKent Overstreet bio_for_each_segment_all(bvec, bio, i) { 956ce3ed71aSChris Mason root = BTRFS_I(bvec->bv_page->mapping->host)->root; 95701d58472SDaniel Dressler ret = csum_dirty_buffer(root->fs_info, bvec->bv_page); 95879787eaaSJeff Mahoney if (ret) 95979787eaaSJeff Mahoney break; 960ce3ed71aSChris Mason } 9612c30c71bSKent Overstreet 96279787eaaSJeff Mahoney return ret; 963ce3ed71aSChris Mason } 964ce3ed71aSChris Mason 96581a75f67SMike Christie static int __btree_submit_bio_start(struct inode *inode, struct bio *bio, 96681a75f67SMike Christie int mirror_num, unsigned long bio_flags, 967eaf25d93SChris Mason u64 bio_offset) 96822c59948SChris Mason { 9698b712842SChris Mason /* 9708b712842SChris Mason * when we're called for a write, we're already in the async 9715443be45SChris Mason * submission context. Just jump into btrfs_map_bio 9728b712842SChris Mason */ 97379787eaaSJeff Mahoney return btree_csum_one_bio(bio); 97422c59948SChris Mason } 97522c59948SChris Mason 97681a75f67SMike Christie static int __btree_submit_bio_done(struct inode *inode, struct bio *bio, 977eaf25d93SChris Mason int mirror_num, unsigned long bio_flags, 978eaf25d93SChris Mason u64 bio_offset) 9794a69a410SChris Mason { 98061891923SStefan Behrens int ret; 98161891923SStefan Behrens 9828b712842SChris Mason /* 9834a69a410SChris Mason * when we're called for a write, we're already in the async 9844a69a410SChris Mason * submission context. Just jump into btrfs_map_bio 9858b712842SChris Mason */ 98681a75f67SMike Christie ret = btrfs_map_bio(BTRFS_I(inode)->root, bio, mirror_num, 1); 9874246a0b6SChristoph Hellwig if (ret) { 9884246a0b6SChristoph Hellwig bio->bi_error = ret; 9894246a0b6SChristoph Hellwig bio_endio(bio); 9904246a0b6SChristoph Hellwig } 99161891923SStefan Behrens return ret; 9920b86a832SChris Mason } 9930b86a832SChris Mason 994de0022b9SJosef Bacik static int check_async_write(struct inode *inode, unsigned long bio_flags) 995de0022b9SJosef Bacik { 996de0022b9SJosef Bacik if (bio_flags & EXTENT_BIO_TREE_LOG) 997de0022b9SJosef Bacik return 0; 998de0022b9SJosef Bacik #ifdef CONFIG_X86 999bc696ca0SBorislav Petkov if (static_cpu_has(X86_FEATURE_XMM4_2)) 1000de0022b9SJosef Bacik return 0; 1001de0022b9SJosef Bacik #endif 1002de0022b9SJosef Bacik return 1; 1003de0022b9SJosef Bacik } 1004de0022b9SJosef Bacik 100581a75f67SMike Christie static int btree_submit_bio_hook(struct inode *inode, struct bio *bio, 1006eaf25d93SChris Mason int mirror_num, unsigned long bio_flags, 1007eaf25d93SChris Mason u64 bio_offset) 100844b8bd7eSChris Mason { 10090b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 1010de0022b9SJosef Bacik int async = check_async_write(inode, bio_flags); 10114a69a410SChris Mason int ret; 1012cad321adSChris Mason 101337226b21SMike Christie if (bio_op(bio) != REQ_OP_WRITE) { 1014cad321adSChris Mason /* 1015cad321adSChris Mason * called for a read, do the setup so that checksum validation 1016cad321adSChris Mason * can happen in the async kernel threads 1017cad321adSChris Mason */ 10180b246afaSJeff Mahoney ret = btrfs_bio_wq_end_io(fs_info, bio, 10190b246afaSJeff Mahoney BTRFS_WQ_ENDIO_METADATA); 10201d4284bdSChris Mason if (ret) 102161891923SStefan Behrens goto out_w_error; 102281a75f67SMike Christie ret = btrfs_map_bio(BTRFS_I(inode)->root, bio, mirror_num, 0); 1023de0022b9SJosef Bacik } else if (!async) { 1024de0022b9SJosef Bacik ret = btree_csum_one_bio(bio); 1025de0022b9SJosef Bacik if (ret) 102661891923SStefan Behrens goto out_w_error; 102781a75f67SMike Christie ret = btrfs_map_bio(BTRFS_I(inode)->root, bio, mirror_num, 0); 102861891923SStefan Behrens } else { 1029cad321adSChris Mason /* 103061891923SStefan Behrens * kthread helpers are used to submit writes so that 103161891923SStefan Behrens * checksumming can happen in parallel across all CPUs 1032cad321adSChris Mason */ 10330b246afaSJeff Mahoney ret = btrfs_wq_submit_bio(fs_info, inode, bio, mirror_num, 0, 1034eaf25d93SChris Mason bio_offset, 10354a69a410SChris Mason __btree_submit_bio_start, 10364a69a410SChris Mason __btree_submit_bio_done); 103744b8bd7eSChris Mason } 103844b8bd7eSChris Mason 10394246a0b6SChristoph Hellwig if (ret) 10404246a0b6SChristoph Hellwig goto out_w_error; 10414246a0b6SChristoph Hellwig return 0; 10424246a0b6SChristoph Hellwig 104361891923SStefan Behrens out_w_error: 10444246a0b6SChristoph Hellwig bio->bi_error = ret; 10454246a0b6SChristoph Hellwig bio_endio(bio); 104661891923SStefan Behrens return ret; 104761891923SStefan Behrens } 104861891923SStefan Behrens 10493dd1462eSJan Beulich #ifdef CONFIG_MIGRATION 1050784b4e29SChris Mason static int btree_migratepage(struct address_space *mapping, 1051a6bc32b8SMel Gorman struct page *newpage, struct page *page, 1052a6bc32b8SMel Gorman enum migrate_mode mode) 1053784b4e29SChris Mason { 1054784b4e29SChris Mason /* 1055784b4e29SChris Mason * we can't safely write a btree page from here, 1056784b4e29SChris Mason * we haven't done the locking hook 1057784b4e29SChris Mason */ 1058784b4e29SChris Mason if (PageDirty(page)) 1059784b4e29SChris Mason return -EAGAIN; 1060784b4e29SChris Mason /* 1061784b4e29SChris Mason * Buffers may be managed in a filesystem specific way. 1062784b4e29SChris Mason * We must have no buffers or drop them. 1063784b4e29SChris Mason */ 1064784b4e29SChris Mason if (page_has_private(page) && 1065784b4e29SChris Mason !try_to_release_page(page, GFP_KERNEL)) 1066784b4e29SChris Mason return -EAGAIN; 1067a6bc32b8SMel Gorman return migrate_page(mapping, newpage, page, mode); 1068784b4e29SChris Mason } 10693dd1462eSJan Beulich #endif 1070784b4e29SChris Mason 10710da5468fSChris Mason 10720da5468fSChris Mason static int btree_writepages(struct address_space *mapping, 10730da5468fSChris Mason struct writeback_control *wbc) 10740da5468fSChris Mason { 1075e2d84521SMiao Xie struct btrfs_fs_info *fs_info; 1076e2d84521SMiao Xie int ret; 1077e2d84521SMiao Xie 1078d8d5f3e1SChris Mason if (wbc->sync_mode == WB_SYNC_NONE) { 1079448d640bSChris Mason 1080448d640bSChris Mason if (wbc->for_kupdate) 1081448d640bSChris Mason return 0; 1082448d640bSChris Mason 1083e2d84521SMiao Xie fs_info = BTRFS_I(mapping->host)->root->fs_info; 1084b9473439SChris Mason /* this is a bit racy, but that's ok */ 1085e2d84521SMiao Xie ret = percpu_counter_compare(&fs_info->dirty_metadata_bytes, 1086e2d84521SMiao Xie BTRFS_DIRTY_METADATA_THRESH); 1087e2d84521SMiao Xie if (ret < 0) 1088793955bcSChris Mason return 0; 1089793955bcSChris Mason } 10900b32f4bbSJosef Bacik return btree_write_cache_pages(mapping, wbc); 10910da5468fSChris Mason } 10920da5468fSChris Mason 1093b2950863SChristoph Hellwig static int btree_readpage(struct file *file, struct page *page) 10945f39d397SChris Mason { 1095d1310b2eSChris Mason struct extent_io_tree *tree; 1096d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 10978ddc7d9cSJan Schmidt return extent_read_full_page(tree, page, btree_get_extent, 0); 10985f39d397SChris Mason } 10995f39d397SChris Mason 110070dec807SChris Mason static int btree_releasepage(struct page *page, gfp_t gfp_flags) 11015f39d397SChris Mason { 110298509cfcSChris Mason if (PageWriteback(page) || PageDirty(page)) 110398509cfcSChris Mason return 0; 11040c4e538bSDavid Sterba 1105f7a52a40SDavid Sterba return try_release_extent_buffer(page); 1106d98237b3SChris Mason } 1107d98237b3SChris Mason 1108d47992f8SLukas Czerner static void btree_invalidatepage(struct page *page, unsigned int offset, 1109d47992f8SLukas Czerner unsigned int length) 1110d98237b3SChris Mason { 1111d1310b2eSChris Mason struct extent_io_tree *tree; 1112d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 11135f39d397SChris Mason extent_invalidatepage(tree, page, offset); 11145f39d397SChris Mason btree_releasepage(page, GFP_NOFS); 11159ad6b7bcSChris Mason if (PagePrivate(page)) { 1116efe120a0SFrank Holton btrfs_warn(BTRFS_I(page->mapping->host)->root->fs_info, 1117efe120a0SFrank Holton "page private not zero on page %llu", 1118efe120a0SFrank Holton (unsigned long long)page_offset(page)); 11199ad6b7bcSChris Mason ClearPagePrivate(page); 11209ad6b7bcSChris Mason set_page_private(page, 0); 112109cbfeafSKirill A. Shutemov put_page(page); 11229ad6b7bcSChris Mason } 1123d98237b3SChris Mason } 1124d98237b3SChris Mason 11250b32f4bbSJosef Bacik static int btree_set_page_dirty(struct page *page) 11260b32f4bbSJosef Bacik { 1127bb146eb2SJosef Bacik #ifdef DEBUG 11280b32f4bbSJosef Bacik struct extent_buffer *eb; 11290b32f4bbSJosef Bacik 11300b32f4bbSJosef Bacik BUG_ON(!PagePrivate(page)); 11310b32f4bbSJosef Bacik eb = (struct extent_buffer *)page->private; 11320b32f4bbSJosef Bacik BUG_ON(!eb); 11330b32f4bbSJosef Bacik BUG_ON(!test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)); 11340b32f4bbSJosef Bacik BUG_ON(!atomic_read(&eb->refs)); 11350b32f4bbSJosef Bacik btrfs_assert_tree_locked(eb); 1136bb146eb2SJosef Bacik #endif 11370b32f4bbSJosef Bacik return __set_page_dirty_nobuffers(page); 11380b32f4bbSJosef Bacik } 11390b32f4bbSJosef Bacik 11407f09410bSAlexey Dobriyan static const struct address_space_operations btree_aops = { 1141d98237b3SChris Mason .readpage = btree_readpage, 11420da5468fSChris Mason .writepages = btree_writepages, 11435f39d397SChris Mason .releasepage = btree_releasepage, 11445f39d397SChris Mason .invalidatepage = btree_invalidatepage, 11455a92bc88SChris Mason #ifdef CONFIG_MIGRATION 1146784b4e29SChris Mason .migratepage = btree_migratepage, 11475a92bc88SChris Mason #endif 11480b32f4bbSJosef Bacik .set_page_dirty = btree_set_page_dirty, 1149d98237b3SChris Mason }; 1150123abc88SChris Mason 1151d3e46feaSDavid Sterba void readahead_tree_block(struct btrfs_root *root, u64 bytenr) 1152090d1875SChris Mason { 11535f39d397SChris Mason struct extent_buffer *buf = NULL; 11545f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 1155090d1875SChris Mason 1156a83fffb7SDavid Sterba buf = btrfs_find_create_tree_block(root, bytenr); 1157c871b0f2SLiu Bo if (IS_ERR(buf)) 11586197d86eSDavid Sterba return; 1159d1310b2eSChris Mason read_extent_buffer_pages(&BTRFS_I(btree_inode)->io_tree, 11608436ea91SJosef Bacik buf, WAIT_NONE, btree_get_extent, 0); 11615f39d397SChris Mason free_extent_buffer(buf); 1162090d1875SChris Mason } 1163090d1875SChris Mason 1164c0dcaa4dSDavid Sterba int reada_tree_block_flagged(struct btrfs_root *root, u64 bytenr, 1165ab0fff03SArne Jansen int mirror_num, struct extent_buffer **eb) 1166ab0fff03SArne Jansen { 1167ab0fff03SArne Jansen struct extent_buffer *buf = NULL; 1168ab0fff03SArne Jansen struct inode *btree_inode = root->fs_info->btree_inode; 1169ab0fff03SArne Jansen struct extent_io_tree *io_tree = &BTRFS_I(btree_inode)->io_tree; 1170ab0fff03SArne Jansen int ret; 1171ab0fff03SArne Jansen 1172a83fffb7SDavid Sterba buf = btrfs_find_create_tree_block(root, bytenr); 1173c871b0f2SLiu Bo if (IS_ERR(buf)) 1174ab0fff03SArne Jansen return 0; 1175ab0fff03SArne Jansen 1176ab0fff03SArne Jansen set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags); 1177ab0fff03SArne Jansen 11788436ea91SJosef Bacik ret = read_extent_buffer_pages(io_tree, buf, WAIT_PAGE_LOCK, 1179ab0fff03SArne Jansen btree_get_extent, mirror_num); 1180ab0fff03SArne Jansen if (ret) { 1181ab0fff03SArne Jansen free_extent_buffer(buf); 1182ab0fff03SArne Jansen return ret; 1183ab0fff03SArne Jansen } 1184ab0fff03SArne Jansen 1185ab0fff03SArne Jansen if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) { 1186ab0fff03SArne Jansen free_extent_buffer(buf); 1187ab0fff03SArne Jansen return -EIO; 11880b32f4bbSJosef Bacik } else if (extent_buffer_uptodate(buf)) { 1189ab0fff03SArne Jansen *eb = buf; 1190ab0fff03SArne Jansen } else { 1191ab0fff03SArne Jansen free_extent_buffer(buf); 1192ab0fff03SArne Jansen } 1193ab0fff03SArne Jansen return 0; 1194ab0fff03SArne Jansen } 1195ab0fff03SArne Jansen 11960999df54SChris Mason struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root, 1197a83fffb7SDavid Sterba u64 bytenr) 11980999df54SChris Mason { 11990b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 12000b246afaSJeff Mahoney 12010b246afaSJeff Mahoney if (btrfs_is_testing(fs_info)) 12020b246afaSJeff Mahoney return alloc_test_extent_buffer(fs_info, bytenr); 12030b246afaSJeff Mahoney return alloc_extent_buffer(fs_info, bytenr); 12040999df54SChris Mason } 12050999df54SChris Mason 12060999df54SChris Mason 1207e02119d5SChris Mason int btrfs_write_tree_block(struct extent_buffer *buf) 1208e02119d5SChris Mason { 1209727011e0SChris Mason return filemap_fdatawrite_range(buf->pages[0]->mapping, buf->start, 12108aa38c31SChristoph Hellwig buf->start + buf->len - 1); 1211e02119d5SChris Mason } 1212e02119d5SChris Mason 1213e02119d5SChris Mason int btrfs_wait_tree_block_writeback(struct extent_buffer *buf) 1214e02119d5SChris Mason { 1215727011e0SChris Mason return filemap_fdatawait_range(buf->pages[0]->mapping, 12168aa38c31SChristoph Hellwig buf->start, buf->start + buf->len - 1); 1217e02119d5SChris Mason } 1218e02119d5SChris Mason 1219db94535dSChris Mason struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr, 1220ce86cd59SDavid Sterba u64 parent_transid) 1221e20d96d6SChris Mason { 12225f39d397SChris Mason struct extent_buffer *buf = NULL; 122319c00ddcSChris Mason int ret; 122419c00ddcSChris Mason 1225a83fffb7SDavid Sterba buf = btrfs_find_create_tree_block(root, bytenr); 1226c871b0f2SLiu Bo if (IS_ERR(buf)) 1227c871b0f2SLiu Bo return buf; 1228e4204dedSChris Mason 12298436ea91SJosef Bacik ret = btree_read_extent_buffer_pages(root, buf, parent_transid); 12300f0fe8f7SFilipe David Borba Manana if (ret) { 12310f0fe8f7SFilipe David Borba Manana free_extent_buffer(buf); 123264c043deSLiu Bo return ERR_PTR(ret); 12330f0fe8f7SFilipe David Borba Manana } 12345f39d397SChris Mason return buf; 1235ce9adaa5SChris Mason 1236eb60ceacSChris Mason } 1237eb60ceacSChris Mason 123801d58472SDaniel Dressler void clean_tree_block(struct btrfs_trans_handle *trans, 123901d58472SDaniel Dressler struct btrfs_fs_info *fs_info, 12405f39d397SChris Mason struct extent_buffer *buf) 1241ed2ff2cbSChris Mason { 124255c69072SChris Mason if (btrfs_header_generation(buf) == 1243e2d84521SMiao Xie fs_info->running_transaction->transid) { 1244b9447ef8SChris Mason btrfs_assert_tree_locked(buf); 1245b4ce94deSChris Mason 1246b9473439SChris Mason if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &buf->bflags)) { 1247e2d84521SMiao Xie __percpu_counter_add(&fs_info->dirty_metadata_bytes, 1248e2d84521SMiao Xie -buf->len, 1249e2d84521SMiao Xie fs_info->dirty_metadata_batch); 1250b9473439SChris Mason /* ugh, clear_extent_buffer_dirty needs to lock the page */ 1251b9473439SChris Mason btrfs_set_lock_blocking(buf); 12520b32f4bbSJosef Bacik clear_extent_buffer_dirty(buf); 1253925baeddSChris Mason } 12545f39d397SChris Mason } 1255ed7b63ebSJosef Bacik } 12565f39d397SChris Mason 12578257b2dcSMiao Xie static struct btrfs_subvolume_writers *btrfs_alloc_subvolume_writers(void) 12588257b2dcSMiao Xie { 12598257b2dcSMiao Xie struct btrfs_subvolume_writers *writers; 12608257b2dcSMiao Xie int ret; 12618257b2dcSMiao Xie 12628257b2dcSMiao Xie writers = kmalloc(sizeof(*writers), GFP_NOFS); 12638257b2dcSMiao Xie if (!writers) 12648257b2dcSMiao Xie return ERR_PTR(-ENOMEM); 12658257b2dcSMiao Xie 1266908c7f19STejun Heo ret = percpu_counter_init(&writers->counter, 0, GFP_KERNEL); 12678257b2dcSMiao Xie if (ret < 0) { 12688257b2dcSMiao Xie kfree(writers); 12698257b2dcSMiao Xie return ERR_PTR(ret); 12708257b2dcSMiao Xie } 12718257b2dcSMiao Xie 12728257b2dcSMiao Xie init_waitqueue_head(&writers->wait); 12738257b2dcSMiao Xie return writers; 12748257b2dcSMiao Xie } 12758257b2dcSMiao Xie 12768257b2dcSMiao Xie static void 12778257b2dcSMiao Xie btrfs_free_subvolume_writers(struct btrfs_subvolume_writers *writers) 12788257b2dcSMiao Xie { 12798257b2dcSMiao Xie percpu_counter_destroy(&writers->counter); 12808257b2dcSMiao Xie kfree(writers); 12818257b2dcSMiao Xie } 12828257b2dcSMiao Xie 1283da17066cSJeff Mahoney static void __setup_root(struct btrfs_root *root, struct btrfs_fs_info *fs_info, 1284e20d96d6SChris Mason u64 objectid) 1285d97e63b6SChris Mason { 12867c0260eeSJeff Mahoney bool dummy = test_bit(BTRFS_FS_STATE_DUMMY_FS_INFO, &fs_info->fs_state); 1287cfaa7295SChris Mason root->node = NULL; 1288a28ec197SChris Mason root->commit_root = NULL; 128927cdeb70SMiao Xie root->state = 0; 1290d68fc57bSYan, Zheng root->orphan_cleanup_state = 0; 12910b86a832SChris Mason 12920f7d52f4SChris Mason root->objectid = objectid; 12930f7d52f4SChris Mason root->last_trans = 0; 129413a8a7c8SYan, Zheng root->highest_objectid = 0; 1295eb73c1b7SMiao Xie root->nr_delalloc_inodes = 0; 1296199c2a9cSMiao Xie root->nr_ordered_extents = 0; 129758176a96SJosef Bacik root->name = NULL; 12986bef4d31SEric Paris root->inode_tree = RB_ROOT; 129916cdcec7SMiao Xie INIT_RADIX_TREE(&root->delayed_nodes_tree, GFP_ATOMIC); 1300f0486c68SYan, Zheng root->block_rsv = NULL; 1301d68fc57bSYan, Zheng root->orphan_block_rsv = NULL; 13020b86a832SChris Mason 13030b86a832SChris Mason INIT_LIST_HEAD(&root->dirty_list); 13045d4f98a2SYan Zheng INIT_LIST_HEAD(&root->root_list); 1305eb73c1b7SMiao Xie INIT_LIST_HEAD(&root->delalloc_inodes); 1306eb73c1b7SMiao Xie INIT_LIST_HEAD(&root->delalloc_root); 1307199c2a9cSMiao Xie INIT_LIST_HEAD(&root->ordered_extents); 1308199c2a9cSMiao Xie INIT_LIST_HEAD(&root->ordered_root); 13092ab28f32SJosef Bacik INIT_LIST_HEAD(&root->logged_list[0]); 13102ab28f32SJosef Bacik INIT_LIST_HEAD(&root->logged_list[1]); 1311d68fc57bSYan, Zheng spin_lock_init(&root->orphan_lock); 13125d4f98a2SYan Zheng spin_lock_init(&root->inode_lock); 1313eb73c1b7SMiao Xie spin_lock_init(&root->delalloc_lock); 1314199c2a9cSMiao Xie spin_lock_init(&root->ordered_extent_lock); 1315f0486c68SYan, Zheng spin_lock_init(&root->accounting_lock); 13162ab28f32SJosef Bacik spin_lock_init(&root->log_extents_lock[0]); 13172ab28f32SJosef Bacik spin_lock_init(&root->log_extents_lock[1]); 1318a2135011SChris Mason mutex_init(&root->objectid_mutex); 1319e02119d5SChris Mason mutex_init(&root->log_mutex); 132031f3d255SMiao Xie mutex_init(&root->ordered_extent_mutex); 1321573bfb72SMiao Xie mutex_init(&root->delalloc_mutex); 13227237f183SYan Zheng init_waitqueue_head(&root->log_writer_wait); 13237237f183SYan Zheng init_waitqueue_head(&root->log_commit_wait[0]); 13247237f183SYan Zheng init_waitqueue_head(&root->log_commit_wait[1]); 13258b050d35SMiao Xie INIT_LIST_HEAD(&root->log_ctxs[0]); 13268b050d35SMiao Xie INIT_LIST_HEAD(&root->log_ctxs[1]); 13277237f183SYan Zheng atomic_set(&root->log_commit[0], 0); 13287237f183SYan Zheng atomic_set(&root->log_commit[1], 0); 13297237f183SYan Zheng atomic_set(&root->log_writers, 0); 13302ecb7923SMiao Xie atomic_set(&root->log_batch, 0); 13318a35d95fSJosef Bacik atomic_set(&root->orphan_inodes, 0); 1332b0feb9d9SMiao Xie atomic_set(&root->refs, 1); 13338257b2dcSMiao Xie atomic_set(&root->will_be_snapshoted, 0); 133455eeaf05SQu Wenruo atomic_set(&root->qgroup_meta_rsv, 0); 13357237f183SYan Zheng root->log_transid = 0; 1336d1433debSMiao Xie root->log_transid_committed = -1; 1337257c62e1SChris Mason root->last_log_commit = 0; 13387c0260eeSJeff Mahoney if (!dummy) 1339d0c803c4SChris Mason extent_io_tree_init(&root->dirty_log_pages, 1340f993c883SDavid Sterba fs_info->btree_inode->i_mapping); 1341017e5369SChris Mason 13423768f368SChris Mason memset(&root->root_key, 0, sizeof(root->root_key)); 13433768f368SChris Mason memset(&root->root_item, 0, sizeof(root->root_item)); 13446702ed49SChris Mason memset(&root->defrag_progress, 0, sizeof(root->defrag_progress)); 13457c0260eeSJeff Mahoney if (!dummy) 13463f157a2fSChris Mason root->defrag_trans_start = fs_info->generation; 134706ea65a3SJosef Bacik else 134806ea65a3SJosef Bacik root->defrag_trans_start = 0; 13494d775673SChris Mason root->root_key.objectid = objectid; 13500ee5dc67SAl Viro root->anon_dev = 0; 13518ea05e3aSAlexander Block 13525f3ab90aSAnand Jain spin_lock_init(&root->root_item_lock); 13533768f368SChris Mason } 13543768f368SChris Mason 135574e4d827SDavid Sterba static struct btrfs_root *btrfs_alloc_root(struct btrfs_fs_info *fs_info, 135674e4d827SDavid Sterba gfp_t flags) 13576f07e42eSAl Viro { 135874e4d827SDavid Sterba struct btrfs_root *root = kzalloc(sizeof(*root), flags); 13596f07e42eSAl Viro if (root) 13606f07e42eSAl Viro root->fs_info = fs_info; 13616f07e42eSAl Viro return root; 13626f07e42eSAl Viro } 13636f07e42eSAl Viro 136406ea65a3SJosef Bacik #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS 136506ea65a3SJosef Bacik /* Should only be used by the testing infrastructure */ 1366da17066cSJeff Mahoney struct btrfs_root *btrfs_alloc_dummy_root(struct btrfs_fs_info *fs_info) 136706ea65a3SJosef Bacik { 136806ea65a3SJosef Bacik struct btrfs_root *root; 136906ea65a3SJosef Bacik 13707c0260eeSJeff Mahoney if (!fs_info) 13717c0260eeSJeff Mahoney return ERR_PTR(-EINVAL); 13727c0260eeSJeff Mahoney 13737c0260eeSJeff Mahoney root = btrfs_alloc_root(fs_info, GFP_KERNEL); 137406ea65a3SJosef Bacik if (!root) 137506ea65a3SJosef Bacik return ERR_PTR(-ENOMEM); 1376da17066cSJeff Mahoney 1377b9ef22deSFeifei Xu /* We don't use the stripesize in selftest, set it as sectorsize */ 1378da17066cSJeff Mahoney __setup_root(root, fs_info, BTRFS_ROOT_TREE_OBJECTID); 1379faa2dbf0SJosef Bacik root->alloc_bytenr = 0; 138006ea65a3SJosef Bacik 138106ea65a3SJosef Bacik return root; 138206ea65a3SJosef Bacik } 138306ea65a3SJosef Bacik #endif 138406ea65a3SJosef Bacik 138520897f5cSArne Jansen struct btrfs_root *btrfs_create_tree(struct btrfs_trans_handle *trans, 138620897f5cSArne Jansen struct btrfs_fs_info *fs_info, 138720897f5cSArne Jansen u64 objectid) 138820897f5cSArne Jansen { 138920897f5cSArne Jansen struct extent_buffer *leaf; 139020897f5cSArne Jansen struct btrfs_root *tree_root = fs_info->tree_root; 139120897f5cSArne Jansen struct btrfs_root *root; 139220897f5cSArne Jansen struct btrfs_key key; 139320897f5cSArne Jansen int ret = 0; 13946463fe58SStefan Behrens uuid_le uuid; 139520897f5cSArne Jansen 139674e4d827SDavid Sterba root = btrfs_alloc_root(fs_info, GFP_KERNEL); 139720897f5cSArne Jansen if (!root) 139820897f5cSArne Jansen return ERR_PTR(-ENOMEM); 139920897f5cSArne Jansen 1400da17066cSJeff Mahoney __setup_root(root, fs_info, objectid); 140120897f5cSArne Jansen root->root_key.objectid = objectid; 140220897f5cSArne Jansen root->root_key.type = BTRFS_ROOT_ITEM_KEY; 140320897f5cSArne Jansen root->root_key.offset = 0; 140420897f5cSArne Jansen 14054d75f8a9SDavid Sterba leaf = btrfs_alloc_tree_block(trans, root, 0, objectid, NULL, 0, 0, 0); 140620897f5cSArne Jansen if (IS_ERR(leaf)) { 140720897f5cSArne Jansen ret = PTR_ERR(leaf); 14081dd05682STsutomu Itoh leaf = NULL; 140920897f5cSArne Jansen goto fail; 141020897f5cSArne Jansen } 141120897f5cSArne Jansen 1412b159fa28SDavid Sterba memzero_extent_buffer(leaf, 0, sizeof(struct btrfs_header)); 141320897f5cSArne Jansen btrfs_set_header_bytenr(leaf, leaf->start); 141420897f5cSArne Jansen btrfs_set_header_generation(leaf, trans->transid); 141520897f5cSArne Jansen btrfs_set_header_backref_rev(leaf, BTRFS_MIXED_BACKREF_REV); 141620897f5cSArne Jansen btrfs_set_header_owner(leaf, objectid); 141720897f5cSArne Jansen root->node = leaf; 141820897f5cSArne Jansen 1419d24ee97bSDavid Sterba write_extent_buffer_fsid(leaf, fs_info->fsid); 1420d24ee97bSDavid Sterba write_extent_buffer_chunk_tree_uuid(leaf, fs_info->chunk_tree_uuid); 142120897f5cSArne Jansen btrfs_mark_buffer_dirty(leaf); 142220897f5cSArne Jansen 142320897f5cSArne Jansen root->commit_root = btrfs_root_node(root); 142427cdeb70SMiao Xie set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 142520897f5cSArne Jansen 142620897f5cSArne Jansen root->root_item.flags = 0; 142720897f5cSArne Jansen root->root_item.byte_limit = 0; 142820897f5cSArne Jansen btrfs_set_root_bytenr(&root->root_item, leaf->start); 142920897f5cSArne Jansen btrfs_set_root_generation(&root->root_item, trans->transid); 143020897f5cSArne Jansen btrfs_set_root_level(&root->root_item, 0); 143120897f5cSArne Jansen btrfs_set_root_refs(&root->root_item, 1); 143220897f5cSArne Jansen btrfs_set_root_used(&root->root_item, leaf->len); 143320897f5cSArne Jansen btrfs_set_root_last_snapshot(&root->root_item, 0); 143420897f5cSArne Jansen btrfs_set_root_dirid(&root->root_item, 0); 14356463fe58SStefan Behrens uuid_le_gen(&uuid); 14366463fe58SStefan Behrens memcpy(root->root_item.uuid, uuid.b, BTRFS_UUID_SIZE); 143720897f5cSArne Jansen root->root_item.drop_level = 0; 143820897f5cSArne Jansen 143920897f5cSArne Jansen key.objectid = objectid; 144020897f5cSArne Jansen key.type = BTRFS_ROOT_ITEM_KEY; 144120897f5cSArne Jansen key.offset = 0; 144220897f5cSArne Jansen ret = btrfs_insert_root(trans, tree_root, &key, &root->root_item); 144320897f5cSArne Jansen if (ret) 144420897f5cSArne Jansen goto fail; 144520897f5cSArne Jansen 144620897f5cSArne Jansen btrfs_tree_unlock(leaf); 144720897f5cSArne Jansen 144820897f5cSArne Jansen return root; 14491dd05682STsutomu Itoh 14501dd05682STsutomu Itoh fail: 14511dd05682STsutomu Itoh if (leaf) { 14521dd05682STsutomu Itoh btrfs_tree_unlock(leaf); 145359885b39STsutomu Itoh free_extent_buffer(root->commit_root); 14541dd05682STsutomu Itoh free_extent_buffer(leaf); 14551dd05682STsutomu Itoh } 14561dd05682STsutomu Itoh kfree(root); 14571dd05682STsutomu Itoh 14581dd05682STsutomu Itoh return ERR_PTR(ret); 145920897f5cSArne Jansen } 146020897f5cSArne Jansen 14617237f183SYan Zheng static struct btrfs_root *alloc_log_tree(struct btrfs_trans_handle *trans, 1462e02119d5SChris Mason struct btrfs_fs_info *fs_info) 14630f7d52f4SChris Mason { 14640f7d52f4SChris Mason struct btrfs_root *root; 14657237f183SYan Zheng struct extent_buffer *leaf; 1466e02119d5SChris Mason 146774e4d827SDavid Sterba root = btrfs_alloc_root(fs_info, GFP_NOFS); 1468e02119d5SChris Mason if (!root) 14697237f183SYan Zheng return ERR_PTR(-ENOMEM); 1470e02119d5SChris Mason 1471da17066cSJeff Mahoney __setup_root(root, fs_info, BTRFS_TREE_LOG_OBJECTID); 1472e02119d5SChris Mason 1473e02119d5SChris Mason root->root_key.objectid = BTRFS_TREE_LOG_OBJECTID; 1474e02119d5SChris Mason root->root_key.type = BTRFS_ROOT_ITEM_KEY; 1475e02119d5SChris Mason root->root_key.offset = BTRFS_TREE_LOG_OBJECTID; 147627cdeb70SMiao Xie 14777237f183SYan Zheng /* 147827cdeb70SMiao Xie * DON'T set REF_COWS for log trees 147927cdeb70SMiao Xie * 14807237f183SYan Zheng * log trees do not get reference counted because they go away 14817237f183SYan Zheng * before a real commit is actually done. They do store pointers 14827237f183SYan Zheng * to file data extents, and those reference counts still get 14837237f183SYan Zheng * updated (along with back refs to the log tree). 14847237f183SYan Zheng */ 1485e02119d5SChris Mason 14864d75f8a9SDavid Sterba leaf = btrfs_alloc_tree_block(trans, root, 0, BTRFS_TREE_LOG_OBJECTID, 14874d75f8a9SDavid Sterba NULL, 0, 0, 0); 14887237f183SYan Zheng if (IS_ERR(leaf)) { 14897237f183SYan Zheng kfree(root); 14907237f183SYan Zheng return ERR_CAST(leaf); 14917237f183SYan Zheng } 1492e02119d5SChris Mason 1493b159fa28SDavid Sterba memzero_extent_buffer(leaf, 0, sizeof(struct btrfs_header)); 14945d4f98a2SYan Zheng btrfs_set_header_bytenr(leaf, leaf->start); 14955d4f98a2SYan Zheng btrfs_set_header_generation(leaf, trans->transid); 14965d4f98a2SYan Zheng btrfs_set_header_backref_rev(leaf, BTRFS_MIXED_BACKREF_REV); 14975d4f98a2SYan Zheng btrfs_set_header_owner(leaf, BTRFS_TREE_LOG_OBJECTID); 14987237f183SYan Zheng root->node = leaf; 1499e02119d5SChris Mason 15000b246afaSJeff Mahoney write_extent_buffer_fsid(root->node, fs_info->fsid); 1501e02119d5SChris Mason btrfs_mark_buffer_dirty(root->node); 1502e02119d5SChris Mason btrfs_tree_unlock(root->node); 15037237f183SYan Zheng return root; 15047237f183SYan Zheng } 15057237f183SYan Zheng 15067237f183SYan Zheng int btrfs_init_log_root_tree(struct btrfs_trans_handle *trans, 15077237f183SYan Zheng struct btrfs_fs_info *fs_info) 15087237f183SYan Zheng { 15097237f183SYan Zheng struct btrfs_root *log_root; 15107237f183SYan Zheng 15117237f183SYan Zheng log_root = alloc_log_tree(trans, fs_info); 15127237f183SYan Zheng if (IS_ERR(log_root)) 15137237f183SYan Zheng return PTR_ERR(log_root); 15147237f183SYan Zheng WARN_ON(fs_info->log_root_tree); 15157237f183SYan Zheng fs_info->log_root_tree = log_root; 15167237f183SYan Zheng return 0; 15177237f183SYan Zheng } 15187237f183SYan Zheng 15197237f183SYan Zheng int btrfs_add_log_tree(struct btrfs_trans_handle *trans, 15207237f183SYan Zheng struct btrfs_root *root) 15217237f183SYan Zheng { 15220b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 15237237f183SYan Zheng struct btrfs_root *log_root; 15247237f183SYan Zheng struct btrfs_inode_item *inode_item; 15257237f183SYan Zheng 15260b246afaSJeff Mahoney log_root = alloc_log_tree(trans, fs_info); 15277237f183SYan Zheng if (IS_ERR(log_root)) 15287237f183SYan Zheng return PTR_ERR(log_root); 15297237f183SYan Zheng 15307237f183SYan Zheng log_root->last_trans = trans->transid; 15317237f183SYan Zheng log_root->root_key.offset = root->root_key.objectid; 15327237f183SYan Zheng 15337237f183SYan Zheng inode_item = &log_root->root_item.inode; 15343cae210fSQu Wenruo btrfs_set_stack_inode_generation(inode_item, 1); 15353cae210fSQu Wenruo btrfs_set_stack_inode_size(inode_item, 3); 15363cae210fSQu Wenruo btrfs_set_stack_inode_nlink(inode_item, 1); 1537da17066cSJeff Mahoney btrfs_set_stack_inode_nbytes(inode_item, 15380b246afaSJeff Mahoney fs_info->nodesize); 15393cae210fSQu Wenruo btrfs_set_stack_inode_mode(inode_item, S_IFDIR | 0755); 15407237f183SYan Zheng 15415d4f98a2SYan Zheng btrfs_set_root_node(&log_root->root_item, log_root->node); 15427237f183SYan Zheng 15437237f183SYan Zheng WARN_ON(root->log_root); 15447237f183SYan Zheng root->log_root = log_root; 15457237f183SYan Zheng root->log_transid = 0; 1546d1433debSMiao Xie root->log_transid_committed = -1; 1547257c62e1SChris Mason root->last_log_commit = 0; 1548e02119d5SChris Mason return 0; 1549e02119d5SChris Mason } 1550e02119d5SChris Mason 155135a3621bSStefan Behrens static struct btrfs_root *btrfs_read_tree_root(struct btrfs_root *tree_root, 1552cb517eabSMiao Xie struct btrfs_key *key) 1553e02119d5SChris Mason { 1554e02119d5SChris Mason struct btrfs_root *root; 1555e02119d5SChris Mason struct btrfs_fs_info *fs_info = tree_root->fs_info; 15560f7d52f4SChris Mason struct btrfs_path *path; 155784234f3aSYan Zheng u64 generation; 1558cb517eabSMiao Xie int ret; 1559cb517eabSMiao Xie 1560cb517eabSMiao Xie path = btrfs_alloc_path(); 1561cb517eabSMiao Xie if (!path) 1562cb517eabSMiao Xie return ERR_PTR(-ENOMEM); 15630f7d52f4SChris Mason 156474e4d827SDavid Sterba root = btrfs_alloc_root(fs_info, GFP_NOFS); 1565cb517eabSMiao Xie if (!root) { 1566cb517eabSMiao Xie ret = -ENOMEM; 1567cb517eabSMiao Xie goto alloc_fail; 15680f7d52f4SChris Mason } 15690f7d52f4SChris Mason 1570da17066cSJeff Mahoney __setup_root(root, fs_info, key->objectid); 15710f7d52f4SChris Mason 1572cb517eabSMiao Xie ret = btrfs_find_root(tree_root, key, path, 1573cb517eabSMiao Xie &root->root_item, &root->root_key); 15740f7d52f4SChris Mason if (ret) { 157513a8a7c8SYan, Zheng if (ret > 0) 157613a8a7c8SYan, Zheng ret = -ENOENT; 1577cb517eabSMiao Xie goto find_fail; 15780f7d52f4SChris Mason } 157913a8a7c8SYan, Zheng 158084234f3aSYan Zheng generation = btrfs_root_generation(&root->root_item); 1581db94535dSChris Mason root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item), 1582ce86cd59SDavid Sterba generation); 158364c043deSLiu Bo if (IS_ERR(root->node)) { 158464c043deSLiu Bo ret = PTR_ERR(root->node); 1585cb517eabSMiao Xie goto find_fail; 1586cb517eabSMiao Xie } else if (!btrfs_buffer_uptodate(root->node, generation, 0)) { 1587cb517eabSMiao Xie ret = -EIO; 158864c043deSLiu Bo free_extent_buffer(root->node); 158964c043deSLiu Bo goto find_fail; 1590416bc658SJosef Bacik } 15915d4f98a2SYan Zheng root->commit_root = btrfs_root_node(root); 159213a8a7c8SYan, Zheng out: 1593cb517eabSMiao Xie btrfs_free_path(path); 1594cb517eabSMiao Xie return root; 1595cb517eabSMiao Xie 1596cb517eabSMiao Xie find_fail: 1597cb517eabSMiao Xie kfree(root); 1598cb517eabSMiao Xie alloc_fail: 1599cb517eabSMiao Xie root = ERR_PTR(ret); 1600cb517eabSMiao Xie goto out; 1601cb517eabSMiao Xie } 1602cb517eabSMiao Xie 1603cb517eabSMiao Xie struct btrfs_root *btrfs_read_fs_root(struct btrfs_root *tree_root, 1604cb517eabSMiao Xie struct btrfs_key *location) 1605cb517eabSMiao Xie { 1606cb517eabSMiao Xie struct btrfs_root *root; 1607cb517eabSMiao Xie 1608cb517eabSMiao Xie root = btrfs_read_tree_root(tree_root, location); 1609cb517eabSMiao Xie if (IS_ERR(root)) 1610cb517eabSMiao Xie return root; 1611cb517eabSMiao Xie 1612cb517eabSMiao Xie if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) { 161327cdeb70SMiao Xie set_bit(BTRFS_ROOT_REF_COWS, &root->state); 161408fe4db1SLi Zefan btrfs_check_and_init_root_item(&root->root_item); 161508fe4db1SLi Zefan } 161613a8a7c8SYan, Zheng 16175eda7b5eSChris Mason return root; 16185eda7b5eSChris Mason } 16195eda7b5eSChris Mason 1620cb517eabSMiao Xie int btrfs_init_fs_root(struct btrfs_root *root) 1621cb517eabSMiao Xie { 1622cb517eabSMiao Xie int ret; 16238257b2dcSMiao Xie struct btrfs_subvolume_writers *writers; 1624cb517eabSMiao Xie 1625cb517eabSMiao Xie root->free_ino_ctl = kzalloc(sizeof(*root->free_ino_ctl), GFP_NOFS); 1626cb517eabSMiao Xie root->free_ino_pinned = kzalloc(sizeof(*root->free_ino_pinned), 1627cb517eabSMiao Xie GFP_NOFS); 1628cb517eabSMiao Xie if (!root->free_ino_pinned || !root->free_ino_ctl) { 1629cb517eabSMiao Xie ret = -ENOMEM; 1630cb517eabSMiao Xie goto fail; 1631cb517eabSMiao Xie } 1632cb517eabSMiao Xie 16338257b2dcSMiao Xie writers = btrfs_alloc_subvolume_writers(); 16348257b2dcSMiao Xie if (IS_ERR(writers)) { 16358257b2dcSMiao Xie ret = PTR_ERR(writers); 16368257b2dcSMiao Xie goto fail; 16378257b2dcSMiao Xie } 16388257b2dcSMiao Xie root->subv_writers = writers; 16398257b2dcSMiao Xie 1640cb517eabSMiao Xie btrfs_init_free_ino_ctl(root); 164157cdc8dbSDavid Sterba spin_lock_init(&root->ino_cache_lock); 164257cdc8dbSDavid Sterba init_waitqueue_head(&root->ino_cache_wait); 1643cb517eabSMiao Xie 1644cb517eabSMiao Xie ret = get_anon_bdev(&root->anon_dev); 1645cb517eabSMiao Xie if (ret) 1646876d2cf1SLiu Bo goto fail; 1647f32e48e9SChandan Rajendra 1648f32e48e9SChandan Rajendra mutex_lock(&root->objectid_mutex); 1649f32e48e9SChandan Rajendra ret = btrfs_find_highest_objectid(root, 1650f32e48e9SChandan Rajendra &root->highest_objectid); 1651f32e48e9SChandan Rajendra if (ret) { 1652f32e48e9SChandan Rajendra mutex_unlock(&root->objectid_mutex); 1653876d2cf1SLiu Bo goto fail; 1654f32e48e9SChandan Rajendra } 1655f32e48e9SChandan Rajendra 1656f32e48e9SChandan Rajendra ASSERT(root->highest_objectid <= BTRFS_LAST_FREE_OBJECTID); 1657f32e48e9SChandan Rajendra 1658f32e48e9SChandan Rajendra mutex_unlock(&root->objectid_mutex); 1659f32e48e9SChandan Rajendra 1660cb517eabSMiao Xie return 0; 1661cb517eabSMiao Xie fail: 1662876d2cf1SLiu Bo /* the caller is responsible to call free_fs_root */ 1663cb517eabSMiao Xie return ret; 1664cb517eabSMiao Xie } 1665cb517eabSMiao Xie 166635bbb97fSJeff Mahoney struct btrfs_root *btrfs_lookup_fs_root(struct btrfs_fs_info *fs_info, 1667cb517eabSMiao Xie u64 root_id) 1668cb517eabSMiao Xie { 1669cb517eabSMiao Xie struct btrfs_root *root; 1670cb517eabSMiao Xie 1671cb517eabSMiao Xie spin_lock(&fs_info->fs_roots_radix_lock); 1672cb517eabSMiao Xie root = radix_tree_lookup(&fs_info->fs_roots_radix, 1673cb517eabSMiao Xie (unsigned long)root_id); 1674cb517eabSMiao Xie spin_unlock(&fs_info->fs_roots_radix_lock); 1675cb517eabSMiao Xie return root; 1676cb517eabSMiao Xie } 1677cb517eabSMiao Xie 1678cb517eabSMiao Xie int btrfs_insert_fs_root(struct btrfs_fs_info *fs_info, 1679cb517eabSMiao Xie struct btrfs_root *root) 1680cb517eabSMiao Xie { 1681cb517eabSMiao Xie int ret; 1682cb517eabSMiao Xie 1683e1860a77SDavid Sterba ret = radix_tree_preload(GFP_NOFS); 1684cb517eabSMiao Xie if (ret) 1685cb517eabSMiao Xie return ret; 1686cb517eabSMiao Xie 1687cb517eabSMiao Xie spin_lock(&fs_info->fs_roots_radix_lock); 1688cb517eabSMiao Xie ret = radix_tree_insert(&fs_info->fs_roots_radix, 1689cb517eabSMiao Xie (unsigned long)root->root_key.objectid, 1690cb517eabSMiao Xie root); 1691cb517eabSMiao Xie if (ret == 0) 169227cdeb70SMiao Xie set_bit(BTRFS_ROOT_IN_RADIX, &root->state); 1693cb517eabSMiao Xie spin_unlock(&fs_info->fs_roots_radix_lock); 1694cb517eabSMiao Xie radix_tree_preload_end(); 1695cb517eabSMiao Xie 1696cb517eabSMiao Xie return ret; 1697cb517eabSMiao Xie } 1698cb517eabSMiao Xie 1699c00869f1SMiao Xie struct btrfs_root *btrfs_get_fs_root(struct btrfs_fs_info *fs_info, 1700c00869f1SMiao Xie struct btrfs_key *location, 1701c00869f1SMiao Xie bool check_ref) 17025eda7b5eSChris Mason { 17035eda7b5eSChris Mason struct btrfs_root *root; 1704381cf658SDavid Sterba struct btrfs_path *path; 17051d4c08e0SDavid Sterba struct btrfs_key key; 17065eda7b5eSChris Mason int ret; 17075eda7b5eSChris Mason 1708edbd8d4eSChris Mason if (location->objectid == BTRFS_ROOT_TREE_OBJECTID) 1709edbd8d4eSChris Mason return fs_info->tree_root; 1710edbd8d4eSChris Mason if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID) 1711edbd8d4eSChris Mason return fs_info->extent_root; 17128f18cf13SChris Mason if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID) 17138f18cf13SChris Mason return fs_info->chunk_root; 17148f18cf13SChris Mason if (location->objectid == BTRFS_DEV_TREE_OBJECTID) 17158f18cf13SChris Mason return fs_info->dev_root; 17160403e47eSYan Zheng if (location->objectid == BTRFS_CSUM_TREE_OBJECTID) 17170403e47eSYan Zheng return fs_info->csum_root; 1718bcef60f2SArne Jansen if (location->objectid == BTRFS_QUOTA_TREE_OBJECTID) 1719bcef60f2SArne Jansen return fs_info->quota_root ? fs_info->quota_root : 1720bcef60f2SArne Jansen ERR_PTR(-ENOENT); 1721f7a81ea4SStefan Behrens if (location->objectid == BTRFS_UUID_TREE_OBJECTID) 1722f7a81ea4SStefan Behrens return fs_info->uuid_root ? fs_info->uuid_root : 1723f7a81ea4SStefan Behrens ERR_PTR(-ENOENT); 172470f6d82eSOmar Sandoval if (location->objectid == BTRFS_FREE_SPACE_TREE_OBJECTID) 172570f6d82eSOmar Sandoval return fs_info->free_space_root ? fs_info->free_space_root : 172670f6d82eSOmar Sandoval ERR_PTR(-ENOENT); 17274df27c4dSYan, Zheng again: 1728cb517eabSMiao Xie root = btrfs_lookup_fs_root(fs_info, location->objectid); 172948475471SStefan Behrens if (root) { 1730c00869f1SMiao Xie if (check_ref && btrfs_root_refs(&root->root_item) == 0) 173148475471SStefan Behrens return ERR_PTR(-ENOENT); 17325eda7b5eSChris Mason return root; 173348475471SStefan Behrens } 17345eda7b5eSChris Mason 1735cb517eabSMiao Xie root = btrfs_read_fs_root(fs_info->tree_root, location); 17365eda7b5eSChris Mason if (IS_ERR(root)) 17375eda7b5eSChris Mason return root; 17383394e160SChris Mason 1739c00869f1SMiao Xie if (check_ref && btrfs_root_refs(&root->root_item) == 0) { 1740d68fc57bSYan, Zheng ret = -ENOENT; 1741d68fc57bSYan, Zheng goto fail; 1742d68fc57bSYan, Zheng } 1743d68fc57bSYan, Zheng 1744cb517eabSMiao Xie ret = btrfs_init_fs_root(root); 1745cb517eabSMiao Xie if (ret) 1746cb517eabSMiao Xie goto fail; 1747cb517eabSMiao Xie 1748381cf658SDavid Sterba path = btrfs_alloc_path(); 1749381cf658SDavid Sterba if (!path) { 1750381cf658SDavid Sterba ret = -ENOMEM; 1751381cf658SDavid Sterba goto fail; 1752381cf658SDavid Sterba } 17531d4c08e0SDavid Sterba key.objectid = BTRFS_ORPHAN_OBJECTID; 17541d4c08e0SDavid Sterba key.type = BTRFS_ORPHAN_ITEM_KEY; 17551d4c08e0SDavid Sterba key.offset = location->objectid; 17561d4c08e0SDavid Sterba 17571d4c08e0SDavid Sterba ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0); 1758381cf658SDavid Sterba btrfs_free_path(path); 1759d68fc57bSYan, Zheng if (ret < 0) 1760d68fc57bSYan, Zheng goto fail; 1761d68fc57bSYan, Zheng if (ret == 0) 176227cdeb70SMiao Xie set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state); 1763d68fc57bSYan, Zheng 1764cb517eabSMiao Xie ret = btrfs_insert_fs_root(fs_info, root); 17650f7d52f4SChris Mason if (ret) { 17664df27c4dSYan, Zheng if (ret == -EEXIST) { 17674df27c4dSYan, Zheng free_fs_root(root); 17684df27c4dSYan, Zheng goto again; 17690f7d52f4SChris Mason } 17704df27c4dSYan, Zheng goto fail; 17714df27c4dSYan, Zheng } 1772edbd8d4eSChris Mason return root; 17734df27c4dSYan, Zheng fail: 17744df27c4dSYan, Zheng free_fs_root(root); 17754df27c4dSYan, Zheng return ERR_PTR(ret); 1776edbd8d4eSChris Mason } 1777edbd8d4eSChris Mason 177804160088SChris Mason static int btrfs_congested_fn(void *congested_data, int bdi_bits) 177904160088SChris Mason { 178004160088SChris Mason struct btrfs_fs_info *info = (struct btrfs_fs_info *)congested_data; 178104160088SChris Mason int ret = 0; 178204160088SChris Mason struct btrfs_device *device; 178304160088SChris Mason struct backing_dev_info *bdi; 1784b7967db7SChris Mason 17851f78160cSXiao Guangrong rcu_read_lock(); 17861f78160cSXiao Guangrong list_for_each_entry_rcu(device, &info->fs_devices->devices, dev_list) { 1787dfe25020SChris Mason if (!device->bdev) 1788dfe25020SChris Mason continue; 178904160088SChris Mason bdi = blk_get_backing_dev_info(device->bdev); 1790ff9ea323STejun Heo if (bdi_congested(bdi, bdi_bits)) { 179104160088SChris Mason ret = 1; 179204160088SChris Mason break; 179304160088SChris Mason } 179404160088SChris Mason } 17951f78160cSXiao Guangrong rcu_read_unlock(); 179604160088SChris Mason return ret; 179704160088SChris Mason } 179804160088SChris Mason 179904160088SChris Mason static int setup_bdi(struct btrfs_fs_info *info, struct backing_dev_info *bdi) 180004160088SChris Mason { 1801ad081f14SJens Axboe int err; 1802ad081f14SJens Axboe 1803b4caecd4SChristoph Hellwig err = bdi_setup_and_register(bdi, "btrfs"); 1804ad081f14SJens Axboe if (err) 1805ad081f14SJens Axboe return err; 1806ad081f14SJens Axboe 180709cbfeafSKirill A. Shutemov bdi->ra_pages = VM_MAX_READAHEAD * 1024 / PAGE_SIZE; 180804160088SChris Mason bdi->congested_fn = btrfs_congested_fn; 180904160088SChris Mason bdi->congested_data = info; 1810da2f0f74SChris Mason bdi->capabilities |= BDI_CAP_CGROUP_WRITEBACK; 181104160088SChris Mason return 0; 181204160088SChris Mason } 181304160088SChris Mason 18148b712842SChris Mason /* 18158b712842SChris Mason * called by the kthread helper functions to finally call the bio end_io 18168b712842SChris Mason * functions. This is where read checksum verification actually happens 18178b712842SChris Mason */ 18188b712842SChris Mason static void end_workqueue_fn(struct btrfs_work *work) 1819ce9adaa5SChris Mason { 1820ce9adaa5SChris Mason struct bio *bio; 182197eb6b69SDavid Sterba struct btrfs_end_io_wq *end_io_wq; 1822ce9adaa5SChris Mason 182397eb6b69SDavid Sterba end_io_wq = container_of(work, struct btrfs_end_io_wq, work); 1824ce9adaa5SChris Mason bio = end_io_wq->bio; 18258b712842SChris Mason 18264246a0b6SChristoph Hellwig bio->bi_error = end_io_wq->error; 1827ce9adaa5SChris Mason bio->bi_private = end_io_wq->private; 1828ce9adaa5SChris Mason bio->bi_end_io = end_io_wq->end_io; 182997eb6b69SDavid Sterba kmem_cache_free(btrfs_end_io_wq_cache, end_io_wq); 18304246a0b6SChristoph Hellwig bio_endio(bio); 1831ce9adaa5SChris Mason } 183244b8bd7eSChris Mason 1833a74a4b97SChris Mason static int cleaner_kthread(void *arg) 1834a74a4b97SChris Mason { 1835a74a4b97SChris Mason struct btrfs_root *root = arg; 18360b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1837d0278245SMiao Xie int again; 1838da288d28SFilipe Manana struct btrfs_trans_handle *trans; 1839a74a4b97SChris Mason 1840a74a4b97SChris Mason do { 1841d0278245SMiao Xie again = 0; 18429d1a2a3aSDavid Sterba 1843d0278245SMiao Xie /* Make the cleaner go to sleep early. */ 1844babbf170SMiao Xie if (btrfs_need_cleaner_sleep(root)) 1845d0278245SMiao Xie goto sleep; 1846d0278245SMiao Xie 184790c711abSZygo Blaxell /* 184890c711abSZygo Blaxell * Do not do anything if we might cause open_ctree() to block 184990c711abSZygo Blaxell * before we have finished mounting the filesystem. 185090c711abSZygo Blaxell */ 18510b246afaSJeff Mahoney if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags)) 185290c711abSZygo Blaxell goto sleep; 185390c711abSZygo Blaxell 18540b246afaSJeff Mahoney if (!mutex_trylock(&fs_info->cleaner_mutex)) 1855d0278245SMiao Xie goto sleep; 1856d0278245SMiao Xie 1857dc7f370cSMiao Xie /* 1858dc7f370cSMiao Xie * Avoid the problem that we change the status of the fs 1859dc7f370cSMiao Xie * during the above check and trylock. 1860dc7f370cSMiao Xie */ 1861babbf170SMiao Xie if (btrfs_need_cleaner_sleep(root)) { 18620b246afaSJeff Mahoney mutex_unlock(&fs_info->cleaner_mutex); 1863dc7f370cSMiao Xie goto sleep; 1864dc7f370cSMiao Xie } 1865dc7f370cSMiao Xie 18660b246afaSJeff Mahoney mutex_lock(&fs_info->cleaner_delayed_iput_mutex); 186724bbcf04SYan, Zheng btrfs_run_delayed_iputs(root); 18680b246afaSJeff Mahoney mutex_unlock(&fs_info->cleaner_delayed_iput_mutex); 1869c2d6cb16SFilipe Manana 18709d1a2a3aSDavid Sterba again = btrfs_clean_one_deleted_snapshot(root); 18710b246afaSJeff Mahoney mutex_unlock(&fs_info->cleaner_mutex); 1872a74a4b97SChris Mason 1873d0278245SMiao Xie /* 187405323cd1SMiao Xie * The defragger has dealt with the R/O remount and umount, 187505323cd1SMiao Xie * needn't do anything special here. 1876d0278245SMiao Xie */ 18770b246afaSJeff Mahoney btrfs_run_defrag_inodes(fs_info); 187867c5e7d4SFilipe Manana 187967c5e7d4SFilipe Manana /* 188067c5e7d4SFilipe Manana * Acquires fs_info->delete_unused_bgs_mutex to avoid racing 188167c5e7d4SFilipe Manana * with relocation (btrfs_relocate_chunk) and relocation 188267c5e7d4SFilipe Manana * acquires fs_info->cleaner_mutex (btrfs_relocate_block_group) 188367c5e7d4SFilipe Manana * after acquiring fs_info->delete_unused_bgs_mutex. So we 188467c5e7d4SFilipe Manana * can't hold, nor need to, fs_info->cleaner_mutex when deleting 188567c5e7d4SFilipe Manana * unused block groups. 188667c5e7d4SFilipe Manana */ 18870b246afaSJeff Mahoney btrfs_delete_unused_bgs(fs_info); 1888d0278245SMiao Xie sleep: 1889838fe188SJiri Kosina if (!again) { 1890a74a4b97SChris Mason set_current_state(TASK_INTERRUPTIBLE); 18918929ecfaSYan, Zheng if (!kthread_should_stop()) 1892a74a4b97SChris Mason schedule(); 1893a74a4b97SChris Mason __set_current_state(TASK_RUNNING); 1894a74a4b97SChris Mason } 1895a74a4b97SChris Mason } while (!kthread_should_stop()); 1896da288d28SFilipe Manana 1897da288d28SFilipe Manana /* 1898da288d28SFilipe Manana * Transaction kthread is stopped before us and wakes us up. 1899da288d28SFilipe Manana * However we might have started a new transaction and COWed some 1900da288d28SFilipe Manana * tree blocks when deleting unused block groups for example. So 1901da288d28SFilipe Manana * make sure we commit the transaction we started to have a clean 1902da288d28SFilipe Manana * shutdown when evicting the btree inode - if it has dirty pages 1903da288d28SFilipe Manana * when we do the final iput() on it, eviction will trigger a 1904da288d28SFilipe Manana * writeback for it which will fail with null pointer dereferences 1905da288d28SFilipe Manana * since work queues and other resources were already released and 1906da288d28SFilipe Manana * destroyed by the time the iput/eviction/writeback is made. 1907da288d28SFilipe Manana */ 1908da288d28SFilipe Manana trans = btrfs_attach_transaction(root); 1909da288d28SFilipe Manana if (IS_ERR(trans)) { 1910da288d28SFilipe Manana if (PTR_ERR(trans) != -ENOENT) 19110b246afaSJeff Mahoney btrfs_err(fs_info, 1912da288d28SFilipe Manana "cleaner transaction attach returned %ld", 1913da288d28SFilipe Manana PTR_ERR(trans)); 1914da288d28SFilipe Manana } else { 1915da288d28SFilipe Manana int ret; 1916da288d28SFilipe Manana 1917da288d28SFilipe Manana ret = btrfs_commit_transaction(trans, root); 1918da288d28SFilipe Manana if (ret) 19190b246afaSJeff Mahoney btrfs_err(fs_info, 1920da288d28SFilipe Manana "cleaner open transaction commit returned %d", 1921da288d28SFilipe Manana ret); 1922da288d28SFilipe Manana } 1923da288d28SFilipe Manana 1924a74a4b97SChris Mason return 0; 1925a74a4b97SChris Mason } 1926a74a4b97SChris Mason 1927a74a4b97SChris Mason static int transaction_kthread(void *arg) 1928a74a4b97SChris Mason { 1929a74a4b97SChris Mason struct btrfs_root *root = arg; 19300b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 1931a74a4b97SChris Mason struct btrfs_trans_handle *trans; 1932a74a4b97SChris Mason struct btrfs_transaction *cur; 19338929ecfaSYan, Zheng u64 transid; 1934a74a4b97SChris Mason unsigned long now; 1935a74a4b97SChris Mason unsigned long delay; 1936914b2007SJan Kara bool cannot_commit; 1937a74a4b97SChris Mason 1938a74a4b97SChris Mason do { 1939914b2007SJan Kara cannot_commit = false; 19400b246afaSJeff Mahoney delay = HZ * fs_info->commit_interval; 19410b246afaSJeff Mahoney mutex_lock(&fs_info->transaction_kthread_mutex); 1942a74a4b97SChris Mason 19430b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 19440b246afaSJeff Mahoney cur = fs_info->running_transaction; 1945a74a4b97SChris Mason if (!cur) { 19460b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1947a74a4b97SChris Mason goto sleep; 1948a74a4b97SChris Mason } 194931153d81SYan Zheng 1950a74a4b97SChris Mason now = get_seconds(); 19514a9d8bdeSMiao Xie if (cur->state < TRANS_STATE_BLOCKED && 19528b87dc17SDavid Sterba (now < cur->start_time || 19530b246afaSJeff Mahoney now - cur->start_time < fs_info->commit_interval)) { 19540b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 1955a74a4b97SChris Mason delay = HZ * 5; 1956a74a4b97SChris Mason goto sleep; 1957a74a4b97SChris Mason } 19588929ecfaSYan, Zheng transid = cur->transid; 19590b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 196056bec294SChris Mason 196179787eaaSJeff Mahoney /* If the file system is aborted, this will always fail. */ 1962354aa0fbSMiao Xie trans = btrfs_attach_transaction(root); 1963914b2007SJan Kara if (IS_ERR(trans)) { 1964354aa0fbSMiao Xie if (PTR_ERR(trans) != -ENOENT) 1965914b2007SJan Kara cannot_commit = true; 196679787eaaSJeff Mahoney goto sleep; 1967914b2007SJan Kara } 19688929ecfaSYan, Zheng if (transid == trans->transid) { 196979787eaaSJeff Mahoney btrfs_commit_transaction(trans, root); 19708929ecfaSYan, Zheng } else { 19718929ecfaSYan, Zheng btrfs_end_transaction(trans, root); 19728929ecfaSYan, Zheng } 1973a74a4b97SChris Mason sleep: 19740b246afaSJeff Mahoney wake_up_process(fs_info->cleaner_kthread); 19750b246afaSJeff Mahoney mutex_unlock(&fs_info->transaction_kthread_mutex); 1976a74a4b97SChris Mason 19774e121c06SJosef Bacik if (unlikely(test_bit(BTRFS_FS_STATE_ERROR, 19780b246afaSJeff Mahoney &fs_info->fs_state))) 19794e121c06SJosef Bacik btrfs_cleanup_transaction(root); 1980a74a4b97SChris Mason set_current_state(TASK_INTERRUPTIBLE); 19818929ecfaSYan, Zheng if (!kthread_should_stop() && 19820b246afaSJeff Mahoney (!btrfs_transaction_blocked(fs_info) || 1983914b2007SJan Kara cannot_commit)) 1984a74a4b97SChris Mason schedule_timeout(delay); 1985a74a4b97SChris Mason __set_current_state(TASK_RUNNING); 1986a74a4b97SChris Mason } while (!kthread_should_stop()); 1987a74a4b97SChris Mason return 0; 1988a74a4b97SChris Mason } 1989a74a4b97SChris Mason 1990af31f5e5SChris Mason /* 1991af31f5e5SChris Mason * this will find the highest generation in the array of 1992af31f5e5SChris Mason * root backups. The index of the highest array is returned, 1993af31f5e5SChris Mason * or -1 if we can't find anything. 1994af31f5e5SChris Mason * 1995af31f5e5SChris Mason * We check to make sure the array is valid by comparing the 1996af31f5e5SChris Mason * generation of the latest root in the array with the generation 1997af31f5e5SChris Mason * in the super block. If they don't match we pitch it. 1998af31f5e5SChris Mason */ 1999af31f5e5SChris Mason static int find_newest_super_backup(struct btrfs_fs_info *info, u64 newest_gen) 2000af31f5e5SChris Mason { 2001af31f5e5SChris Mason u64 cur; 2002af31f5e5SChris Mason int newest_index = -1; 2003af31f5e5SChris Mason struct btrfs_root_backup *root_backup; 2004af31f5e5SChris Mason int i; 2005af31f5e5SChris Mason 2006af31f5e5SChris Mason for (i = 0; i < BTRFS_NUM_BACKUP_ROOTS; i++) { 2007af31f5e5SChris Mason root_backup = info->super_copy->super_roots + i; 2008af31f5e5SChris Mason cur = btrfs_backup_tree_root_gen(root_backup); 2009af31f5e5SChris Mason if (cur == newest_gen) 2010af31f5e5SChris Mason newest_index = i; 2011af31f5e5SChris Mason } 2012af31f5e5SChris Mason 2013af31f5e5SChris Mason /* check to see if we actually wrapped around */ 2014af31f5e5SChris Mason if (newest_index == BTRFS_NUM_BACKUP_ROOTS - 1) { 2015af31f5e5SChris Mason root_backup = info->super_copy->super_roots; 2016af31f5e5SChris Mason cur = btrfs_backup_tree_root_gen(root_backup); 2017af31f5e5SChris Mason if (cur == newest_gen) 2018af31f5e5SChris Mason newest_index = 0; 2019af31f5e5SChris Mason } 2020af31f5e5SChris Mason return newest_index; 2021af31f5e5SChris Mason } 2022af31f5e5SChris Mason 2023af31f5e5SChris Mason 2024af31f5e5SChris Mason /* 2025af31f5e5SChris Mason * find the oldest backup so we know where to store new entries 2026af31f5e5SChris Mason * in the backup array. This will set the backup_root_index 2027af31f5e5SChris Mason * field in the fs_info struct 2028af31f5e5SChris Mason */ 2029af31f5e5SChris Mason static void find_oldest_super_backup(struct btrfs_fs_info *info, 2030af31f5e5SChris Mason u64 newest_gen) 2031af31f5e5SChris Mason { 2032af31f5e5SChris Mason int newest_index = -1; 2033af31f5e5SChris Mason 2034af31f5e5SChris Mason newest_index = find_newest_super_backup(info, newest_gen); 2035af31f5e5SChris Mason /* if there was garbage in there, just move along */ 2036af31f5e5SChris Mason if (newest_index == -1) { 2037af31f5e5SChris Mason info->backup_root_index = 0; 2038af31f5e5SChris Mason } else { 2039af31f5e5SChris Mason info->backup_root_index = (newest_index + 1) % BTRFS_NUM_BACKUP_ROOTS; 2040af31f5e5SChris Mason } 2041af31f5e5SChris Mason } 2042af31f5e5SChris Mason 2043af31f5e5SChris Mason /* 2044af31f5e5SChris Mason * copy all the root pointers into the super backup array. 2045af31f5e5SChris Mason * this will bump the backup pointer by one when it is 2046af31f5e5SChris Mason * done 2047af31f5e5SChris Mason */ 2048af31f5e5SChris Mason static void backup_super_roots(struct btrfs_fs_info *info) 2049af31f5e5SChris Mason { 2050af31f5e5SChris Mason int next_backup; 2051af31f5e5SChris Mason struct btrfs_root_backup *root_backup; 2052af31f5e5SChris Mason int last_backup; 2053af31f5e5SChris Mason 2054af31f5e5SChris Mason next_backup = info->backup_root_index; 2055af31f5e5SChris Mason last_backup = (next_backup + BTRFS_NUM_BACKUP_ROOTS - 1) % 2056af31f5e5SChris Mason BTRFS_NUM_BACKUP_ROOTS; 2057af31f5e5SChris Mason 2058af31f5e5SChris Mason /* 2059af31f5e5SChris Mason * just overwrite the last backup if we're at the same generation 2060af31f5e5SChris Mason * this happens only at umount 2061af31f5e5SChris Mason */ 2062af31f5e5SChris Mason root_backup = info->super_for_commit->super_roots + last_backup; 2063af31f5e5SChris Mason if (btrfs_backup_tree_root_gen(root_backup) == 2064af31f5e5SChris Mason btrfs_header_generation(info->tree_root->node)) 2065af31f5e5SChris Mason next_backup = last_backup; 2066af31f5e5SChris Mason 2067af31f5e5SChris Mason root_backup = info->super_for_commit->super_roots + next_backup; 2068af31f5e5SChris Mason 2069af31f5e5SChris Mason /* 2070af31f5e5SChris Mason * make sure all of our padding and empty slots get zero filled 2071af31f5e5SChris Mason * regardless of which ones we use today 2072af31f5e5SChris Mason */ 2073af31f5e5SChris Mason memset(root_backup, 0, sizeof(*root_backup)); 2074af31f5e5SChris Mason 2075af31f5e5SChris Mason info->backup_root_index = (next_backup + 1) % BTRFS_NUM_BACKUP_ROOTS; 2076af31f5e5SChris Mason 2077af31f5e5SChris Mason btrfs_set_backup_tree_root(root_backup, info->tree_root->node->start); 2078af31f5e5SChris Mason btrfs_set_backup_tree_root_gen(root_backup, 2079af31f5e5SChris Mason btrfs_header_generation(info->tree_root->node)); 2080af31f5e5SChris Mason 2081af31f5e5SChris Mason btrfs_set_backup_tree_root_level(root_backup, 2082af31f5e5SChris Mason btrfs_header_level(info->tree_root->node)); 2083af31f5e5SChris Mason 2084af31f5e5SChris Mason btrfs_set_backup_chunk_root(root_backup, info->chunk_root->node->start); 2085af31f5e5SChris Mason btrfs_set_backup_chunk_root_gen(root_backup, 2086af31f5e5SChris Mason btrfs_header_generation(info->chunk_root->node)); 2087af31f5e5SChris Mason btrfs_set_backup_chunk_root_level(root_backup, 2088af31f5e5SChris Mason btrfs_header_level(info->chunk_root->node)); 2089af31f5e5SChris Mason 2090af31f5e5SChris Mason btrfs_set_backup_extent_root(root_backup, info->extent_root->node->start); 2091af31f5e5SChris Mason btrfs_set_backup_extent_root_gen(root_backup, 2092af31f5e5SChris Mason btrfs_header_generation(info->extent_root->node)); 2093af31f5e5SChris Mason btrfs_set_backup_extent_root_level(root_backup, 2094af31f5e5SChris Mason btrfs_header_level(info->extent_root->node)); 2095af31f5e5SChris Mason 20967c7e82a7SChris Mason /* 20977c7e82a7SChris Mason * we might commit during log recovery, which happens before we set 20987c7e82a7SChris Mason * the fs_root. Make sure it is valid before we fill it in. 20997c7e82a7SChris Mason */ 21007c7e82a7SChris Mason if (info->fs_root && info->fs_root->node) { 21017c7e82a7SChris Mason btrfs_set_backup_fs_root(root_backup, 21027c7e82a7SChris Mason info->fs_root->node->start); 2103af31f5e5SChris Mason btrfs_set_backup_fs_root_gen(root_backup, 2104af31f5e5SChris Mason btrfs_header_generation(info->fs_root->node)); 2105af31f5e5SChris Mason btrfs_set_backup_fs_root_level(root_backup, 2106af31f5e5SChris Mason btrfs_header_level(info->fs_root->node)); 21077c7e82a7SChris Mason } 2108af31f5e5SChris Mason 2109af31f5e5SChris Mason btrfs_set_backup_dev_root(root_backup, info->dev_root->node->start); 2110af31f5e5SChris Mason btrfs_set_backup_dev_root_gen(root_backup, 2111af31f5e5SChris Mason btrfs_header_generation(info->dev_root->node)); 2112af31f5e5SChris Mason btrfs_set_backup_dev_root_level(root_backup, 2113af31f5e5SChris Mason btrfs_header_level(info->dev_root->node)); 2114af31f5e5SChris Mason 2115af31f5e5SChris Mason btrfs_set_backup_csum_root(root_backup, info->csum_root->node->start); 2116af31f5e5SChris Mason btrfs_set_backup_csum_root_gen(root_backup, 2117af31f5e5SChris Mason btrfs_header_generation(info->csum_root->node)); 2118af31f5e5SChris Mason btrfs_set_backup_csum_root_level(root_backup, 2119af31f5e5SChris Mason btrfs_header_level(info->csum_root->node)); 2120af31f5e5SChris Mason 2121af31f5e5SChris Mason btrfs_set_backup_total_bytes(root_backup, 2122af31f5e5SChris Mason btrfs_super_total_bytes(info->super_copy)); 2123af31f5e5SChris Mason btrfs_set_backup_bytes_used(root_backup, 2124af31f5e5SChris Mason btrfs_super_bytes_used(info->super_copy)); 2125af31f5e5SChris Mason btrfs_set_backup_num_devices(root_backup, 2126af31f5e5SChris Mason btrfs_super_num_devices(info->super_copy)); 2127af31f5e5SChris Mason 2128af31f5e5SChris Mason /* 2129af31f5e5SChris Mason * if we don't copy this out to the super_copy, it won't get remembered 2130af31f5e5SChris Mason * for the next commit 2131af31f5e5SChris Mason */ 2132af31f5e5SChris Mason memcpy(&info->super_copy->super_roots, 2133af31f5e5SChris Mason &info->super_for_commit->super_roots, 2134af31f5e5SChris Mason sizeof(*root_backup) * BTRFS_NUM_BACKUP_ROOTS); 2135af31f5e5SChris Mason } 2136af31f5e5SChris Mason 2137af31f5e5SChris Mason /* 2138af31f5e5SChris Mason * this copies info out of the root backup array and back into 2139af31f5e5SChris Mason * the in-memory super block. It is meant to help iterate through 2140af31f5e5SChris Mason * the array, so you send it the number of backups you've already 2141af31f5e5SChris Mason * tried and the last backup index you used. 2142af31f5e5SChris Mason * 2143af31f5e5SChris Mason * this returns -1 when it has tried all the backups 2144af31f5e5SChris Mason */ 2145af31f5e5SChris Mason static noinline int next_root_backup(struct btrfs_fs_info *info, 2146af31f5e5SChris Mason struct btrfs_super_block *super, 2147af31f5e5SChris Mason int *num_backups_tried, int *backup_index) 2148af31f5e5SChris Mason { 2149af31f5e5SChris Mason struct btrfs_root_backup *root_backup; 2150af31f5e5SChris Mason int newest = *backup_index; 2151af31f5e5SChris Mason 2152af31f5e5SChris Mason if (*num_backups_tried == 0) { 2153af31f5e5SChris Mason u64 gen = btrfs_super_generation(super); 2154af31f5e5SChris Mason 2155af31f5e5SChris Mason newest = find_newest_super_backup(info, gen); 2156af31f5e5SChris Mason if (newest == -1) 2157af31f5e5SChris Mason return -1; 2158af31f5e5SChris Mason 2159af31f5e5SChris Mason *backup_index = newest; 2160af31f5e5SChris Mason *num_backups_tried = 1; 2161af31f5e5SChris Mason } else if (*num_backups_tried == BTRFS_NUM_BACKUP_ROOTS) { 2162af31f5e5SChris Mason /* we've tried all the backups, all done */ 2163af31f5e5SChris Mason return -1; 2164af31f5e5SChris Mason } else { 2165af31f5e5SChris Mason /* jump to the next oldest backup */ 2166af31f5e5SChris Mason newest = (*backup_index + BTRFS_NUM_BACKUP_ROOTS - 1) % 2167af31f5e5SChris Mason BTRFS_NUM_BACKUP_ROOTS; 2168af31f5e5SChris Mason *backup_index = newest; 2169af31f5e5SChris Mason *num_backups_tried += 1; 2170af31f5e5SChris Mason } 2171af31f5e5SChris Mason root_backup = super->super_roots + newest; 2172af31f5e5SChris Mason 2173af31f5e5SChris Mason btrfs_set_super_generation(super, 2174af31f5e5SChris Mason btrfs_backup_tree_root_gen(root_backup)); 2175af31f5e5SChris Mason btrfs_set_super_root(super, btrfs_backup_tree_root(root_backup)); 2176af31f5e5SChris Mason btrfs_set_super_root_level(super, 2177af31f5e5SChris Mason btrfs_backup_tree_root_level(root_backup)); 2178af31f5e5SChris Mason btrfs_set_super_bytes_used(super, btrfs_backup_bytes_used(root_backup)); 2179af31f5e5SChris Mason 2180af31f5e5SChris Mason /* 2181af31f5e5SChris Mason * fixme: the total bytes and num_devices need to match or we should 2182af31f5e5SChris Mason * need a fsck 2183af31f5e5SChris Mason */ 2184af31f5e5SChris Mason btrfs_set_super_total_bytes(super, btrfs_backup_total_bytes(root_backup)); 2185af31f5e5SChris Mason btrfs_set_super_num_devices(super, btrfs_backup_num_devices(root_backup)); 2186af31f5e5SChris Mason return 0; 2187af31f5e5SChris Mason } 2188af31f5e5SChris Mason 21897abadb64SLiu Bo /* helper to cleanup workers */ 21907abadb64SLiu Bo static void btrfs_stop_all_workers(struct btrfs_fs_info *fs_info) 21917abadb64SLiu Bo { 2192dc6e3209SQu Wenruo btrfs_destroy_workqueue(fs_info->fixup_workers); 2193afe3d242SQu Wenruo btrfs_destroy_workqueue(fs_info->delalloc_workers); 21945cdc7ad3SQu Wenruo btrfs_destroy_workqueue(fs_info->workers); 2195fccb5d86SQu Wenruo btrfs_destroy_workqueue(fs_info->endio_workers); 2196fccb5d86SQu Wenruo btrfs_destroy_workqueue(fs_info->endio_meta_workers); 2197fccb5d86SQu Wenruo btrfs_destroy_workqueue(fs_info->endio_raid56_workers); 21988b110e39SMiao Xie btrfs_destroy_workqueue(fs_info->endio_repair_workers); 2199d05a33acSQu Wenruo btrfs_destroy_workqueue(fs_info->rmw_workers); 2200fccb5d86SQu Wenruo btrfs_destroy_workqueue(fs_info->endio_meta_write_workers); 2201fccb5d86SQu Wenruo btrfs_destroy_workqueue(fs_info->endio_write_workers); 2202fccb5d86SQu Wenruo btrfs_destroy_workqueue(fs_info->endio_freespace_worker); 2203a8c93d4eSQu Wenruo btrfs_destroy_workqueue(fs_info->submit_workers); 22045b3bc44eSQu Wenruo btrfs_destroy_workqueue(fs_info->delayed_workers); 2205e66f0bb1SQu Wenruo btrfs_destroy_workqueue(fs_info->caching_workers); 2206736cfa15SQu Wenruo btrfs_destroy_workqueue(fs_info->readahead_workers); 2207a44903abSQu Wenruo btrfs_destroy_workqueue(fs_info->flush_workers); 2208fc97fab0SQu Wenruo btrfs_destroy_workqueue(fs_info->qgroup_rescan_workers); 2209a79b7d4bSChris Mason btrfs_destroy_workqueue(fs_info->extent_workers); 22107abadb64SLiu Bo } 22117abadb64SLiu Bo 22122e9f5954SRashika static void free_root_extent_buffers(struct btrfs_root *root) 22132e9f5954SRashika { 22142e9f5954SRashika if (root) { 22152e9f5954SRashika free_extent_buffer(root->node); 22162e9f5954SRashika free_extent_buffer(root->commit_root); 22172e9f5954SRashika root->node = NULL; 22182e9f5954SRashika root->commit_root = NULL; 22192e9f5954SRashika } 22202e9f5954SRashika } 22212e9f5954SRashika 2222af31f5e5SChris Mason /* helper to cleanup tree roots */ 2223af31f5e5SChris Mason static void free_root_pointers(struct btrfs_fs_info *info, int chunk_root) 2224af31f5e5SChris Mason { 22252e9f5954SRashika free_root_extent_buffers(info->tree_root); 2226655b09feSJosef Bacik 22272e9f5954SRashika free_root_extent_buffers(info->dev_root); 22282e9f5954SRashika free_root_extent_buffers(info->extent_root); 22292e9f5954SRashika free_root_extent_buffers(info->csum_root); 22302e9f5954SRashika free_root_extent_buffers(info->quota_root); 22312e9f5954SRashika free_root_extent_buffers(info->uuid_root); 22322e9f5954SRashika if (chunk_root) 22332e9f5954SRashika free_root_extent_buffers(info->chunk_root); 223470f6d82eSOmar Sandoval free_root_extent_buffers(info->free_space_root); 2235af31f5e5SChris Mason } 2236af31f5e5SChris Mason 2237faa2dbf0SJosef Bacik void btrfs_free_fs_roots(struct btrfs_fs_info *fs_info) 2238171f6537SJosef Bacik { 2239171f6537SJosef Bacik int ret; 2240171f6537SJosef Bacik struct btrfs_root *gang[8]; 2241171f6537SJosef Bacik int i; 2242171f6537SJosef Bacik 2243171f6537SJosef Bacik while (!list_empty(&fs_info->dead_roots)) { 2244171f6537SJosef Bacik gang[0] = list_entry(fs_info->dead_roots.next, 2245171f6537SJosef Bacik struct btrfs_root, root_list); 2246171f6537SJosef Bacik list_del(&gang[0]->root_list); 2247171f6537SJosef Bacik 224827cdeb70SMiao Xie if (test_bit(BTRFS_ROOT_IN_RADIX, &gang[0]->state)) { 2249cb517eabSMiao Xie btrfs_drop_and_free_fs_root(fs_info, gang[0]); 2250171f6537SJosef Bacik } else { 2251171f6537SJosef Bacik free_extent_buffer(gang[0]->node); 2252171f6537SJosef Bacik free_extent_buffer(gang[0]->commit_root); 2253b0feb9d9SMiao Xie btrfs_put_fs_root(gang[0]); 2254171f6537SJosef Bacik } 2255171f6537SJosef Bacik } 2256171f6537SJosef Bacik 2257171f6537SJosef Bacik while (1) { 2258171f6537SJosef Bacik ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix, 2259171f6537SJosef Bacik (void **)gang, 0, 2260171f6537SJosef Bacik ARRAY_SIZE(gang)); 2261171f6537SJosef Bacik if (!ret) 2262171f6537SJosef Bacik break; 2263171f6537SJosef Bacik for (i = 0; i < ret; i++) 2264cb517eabSMiao Xie btrfs_drop_and_free_fs_root(fs_info, gang[i]); 2265171f6537SJosef Bacik } 22661a4319ccSLiu Bo 22671a4319ccSLiu Bo if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) { 22681a4319ccSLiu Bo btrfs_free_log_root_tree(NULL, fs_info); 22691a4319ccSLiu Bo btrfs_destroy_pinned_extent(fs_info->tree_root, 22701a4319ccSLiu Bo fs_info->pinned_extents); 22711a4319ccSLiu Bo } 2272171f6537SJosef Bacik } 2273af31f5e5SChris Mason 2274638aa7edSEric Sandeen static void btrfs_init_scrub(struct btrfs_fs_info *fs_info) 2275638aa7edSEric Sandeen { 2276638aa7edSEric Sandeen mutex_init(&fs_info->scrub_lock); 2277638aa7edSEric Sandeen atomic_set(&fs_info->scrubs_running, 0); 2278638aa7edSEric Sandeen atomic_set(&fs_info->scrub_pause_req, 0); 2279638aa7edSEric Sandeen atomic_set(&fs_info->scrubs_paused, 0); 2280638aa7edSEric Sandeen atomic_set(&fs_info->scrub_cancel_req, 0); 2281638aa7edSEric Sandeen init_waitqueue_head(&fs_info->scrub_pause_wait); 2282638aa7edSEric Sandeen fs_info->scrub_workers_refcnt = 0; 2283638aa7edSEric Sandeen } 2284638aa7edSEric Sandeen 2285779a65a4SEric Sandeen static void btrfs_init_balance(struct btrfs_fs_info *fs_info) 2286779a65a4SEric Sandeen { 2287779a65a4SEric Sandeen spin_lock_init(&fs_info->balance_lock); 2288779a65a4SEric Sandeen mutex_init(&fs_info->balance_mutex); 2289779a65a4SEric Sandeen atomic_set(&fs_info->balance_running, 0); 2290779a65a4SEric Sandeen atomic_set(&fs_info->balance_pause_req, 0); 2291779a65a4SEric Sandeen atomic_set(&fs_info->balance_cancel_req, 0); 2292779a65a4SEric Sandeen fs_info->balance_ctl = NULL; 2293779a65a4SEric Sandeen init_waitqueue_head(&fs_info->balance_wait_q); 2294779a65a4SEric Sandeen } 2295779a65a4SEric Sandeen 22966bccf3abSJeff Mahoney static void btrfs_init_btree_inode(struct btrfs_fs_info *fs_info) 2297f37938e0SEric Sandeen { 2298f37938e0SEric Sandeen fs_info->btree_inode->i_ino = BTRFS_BTREE_INODE_OBJECTID; 2299f37938e0SEric Sandeen set_nlink(fs_info->btree_inode, 1); 2300f37938e0SEric Sandeen /* 2301f37938e0SEric Sandeen * we set the i_size on the btree inode to the max possible int. 2302f37938e0SEric Sandeen * the real end of the address space is determined by all of 2303f37938e0SEric Sandeen * the devices in the system 2304f37938e0SEric Sandeen */ 2305f37938e0SEric Sandeen fs_info->btree_inode->i_size = OFFSET_MAX; 2306f37938e0SEric Sandeen fs_info->btree_inode->i_mapping->a_ops = &btree_aops; 2307f37938e0SEric Sandeen 2308f37938e0SEric Sandeen RB_CLEAR_NODE(&BTRFS_I(fs_info->btree_inode)->rb_node); 2309f37938e0SEric Sandeen extent_io_tree_init(&BTRFS_I(fs_info->btree_inode)->io_tree, 2310f37938e0SEric Sandeen fs_info->btree_inode->i_mapping); 2311f37938e0SEric Sandeen BTRFS_I(fs_info->btree_inode)->io_tree.track_uptodate = 0; 2312f37938e0SEric Sandeen extent_map_tree_init(&BTRFS_I(fs_info->btree_inode)->extent_tree); 2313f37938e0SEric Sandeen 2314f37938e0SEric Sandeen BTRFS_I(fs_info->btree_inode)->io_tree.ops = &btree_extent_io_ops; 2315f37938e0SEric Sandeen 23166bccf3abSJeff Mahoney BTRFS_I(fs_info->btree_inode)->root = fs_info->tree_root; 2317f37938e0SEric Sandeen memset(&BTRFS_I(fs_info->btree_inode)->location, 0, 2318f37938e0SEric Sandeen sizeof(struct btrfs_key)); 2319f37938e0SEric Sandeen set_bit(BTRFS_INODE_DUMMY, 2320f37938e0SEric Sandeen &BTRFS_I(fs_info->btree_inode)->runtime_flags); 2321f37938e0SEric Sandeen btrfs_insert_inode_hash(fs_info->btree_inode); 2322f37938e0SEric Sandeen } 2323f37938e0SEric Sandeen 2324ad618368SEric Sandeen static void btrfs_init_dev_replace_locks(struct btrfs_fs_info *fs_info) 2325ad618368SEric Sandeen { 2326ad618368SEric Sandeen fs_info->dev_replace.lock_owner = 0; 2327ad618368SEric Sandeen atomic_set(&fs_info->dev_replace.nesting_level, 0); 2328ad618368SEric Sandeen mutex_init(&fs_info->dev_replace.lock_finishing_cancel_unmount); 232973beece9SLiu Bo rwlock_init(&fs_info->dev_replace.lock); 233073beece9SLiu Bo atomic_set(&fs_info->dev_replace.read_locks, 0); 233173beece9SLiu Bo atomic_set(&fs_info->dev_replace.blocking_readers, 0); 2332ad618368SEric Sandeen init_waitqueue_head(&fs_info->replace_wait); 233373beece9SLiu Bo init_waitqueue_head(&fs_info->dev_replace.read_lock_wq); 2334ad618368SEric Sandeen } 2335ad618368SEric Sandeen 2336f9e92e40SEric Sandeen static void btrfs_init_qgroup(struct btrfs_fs_info *fs_info) 2337f9e92e40SEric Sandeen { 2338f9e92e40SEric Sandeen spin_lock_init(&fs_info->qgroup_lock); 2339f9e92e40SEric Sandeen mutex_init(&fs_info->qgroup_ioctl_lock); 2340f9e92e40SEric Sandeen fs_info->qgroup_tree = RB_ROOT; 2341f9e92e40SEric Sandeen fs_info->qgroup_op_tree = RB_ROOT; 2342f9e92e40SEric Sandeen INIT_LIST_HEAD(&fs_info->dirty_qgroups); 2343f9e92e40SEric Sandeen fs_info->qgroup_seq = 1; 2344f9e92e40SEric Sandeen fs_info->qgroup_ulist = NULL; 2345d2c609b8SJeff Mahoney fs_info->qgroup_rescan_running = false; 2346f9e92e40SEric Sandeen mutex_init(&fs_info->qgroup_rescan_lock); 2347f9e92e40SEric Sandeen } 2348f9e92e40SEric Sandeen 23492a458198SEric Sandeen static int btrfs_init_workqueues(struct btrfs_fs_info *fs_info, 23502a458198SEric Sandeen struct btrfs_fs_devices *fs_devices) 23512a458198SEric Sandeen { 23522a458198SEric Sandeen int max_active = fs_info->thread_pool_size; 23536f011058SDavid Sterba unsigned int flags = WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_UNBOUND; 23542a458198SEric Sandeen 23552a458198SEric Sandeen fs_info->workers = 2356cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "worker", 2357cb001095SJeff Mahoney flags | WQ_HIGHPRI, max_active, 16); 23582a458198SEric Sandeen 23592a458198SEric Sandeen fs_info->delalloc_workers = 2360cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "delalloc", 2361cb001095SJeff Mahoney flags, max_active, 2); 23622a458198SEric Sandeen 23632a458198SEric Sandeen fs_info->flush_workers = 2364cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "flush_delalloc", 2365cb001095SJeff Mahoney flags, max_active, 0); 23662a458198SEric Sandeen 23672a458198SEric Sandeen fs_info->caching_workers = 2368cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "cache", flags, max_active, 0); 23692a458198SEric Sandeen 23702a458198SEric Sandeen /* 23712a458198SEric Sandeen * a higher idle thresh on the submit workers makes it much more 23722a458198SEric Sandeen * likely that bios will be send down in a sane order to the 23732a458198SEric Sandeen * devices 23742a458198SEric Sandeen */ 23752a458198SEric Sandeen fs_info->submit_workers = 2376cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "submit", flags, 23772a458198SEric Sandeen min_t(u64, fs_devices->num_devices, 23782a458198SEric Sandeen max_active), 64); 23792a458198SEric Sandeen 23802a458198SEric Sandeen fs_info->fixup_workers = 2381cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "fixup", flags, 1, 0); 23822a458198SEric Sandeen 23832a458198SEric Sandeen /* 23842a458198SEric Sandeen * endios are largely parallel and should have a very 23852a458198SEric Sandeen * low idle thresh 23862a458198SEric Sandeen */ 23872a458198SEric Sandeen fs_info->endio_workers = 2388cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "endio", flags, max_active, 4); 23892a458198SEric Sandeen fs_info->endio_meta_workers = 2390cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "endio-meta", flags, 2391cb001095SJeff Mahoney max_active, 4); 23922a458198SEric Sandeen fs_info->endio_meta_write_workers = 2393cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "endio-meta-write", flags, 2394cb001095SJeff Mahoney max_active, 2); 23952a458198SEric Sandeen fs_info->endio_raid56_workers = 2396cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "endio-raid56", flags, 2397cb001095SJeff Mahoney max_active, 4); 23982a458198SEric Sandeen fs_info->endio_repair_workers = 2399cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "endio-repair", flags, 1, 0); 24002a458198SEric Sandeen fs_info->rmw_workers = 2401cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "rmw", flags, max_active, 2); 24022a458198SEric Sandeen fs_info->endio_write_workers = 2403cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "endio-write", flags, 2404cb001095SJeff Mahoney max_active, 2); 24052a458198SEric Sandeen fs_info->endio_freespace_worker = 2406cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "freespace-write", flags, 2407cb001095SJeff Mahoney max_active, 0); 24082a458198SEric Sandeen fs_info->delayed_workers = 2409cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "delayed-meta", flags, 2410cb001095SJeff Mahoney max_active, 0); 24112a458198SEric Sandeen fs_info->readahead_workers = 2412cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "readahead", flags, 2413cb001095SJeff Mahoney max_active, 2); 24142a458198SEric Sandeen fs_info->qgroup_rescan_workers = 2415cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "qgroup-rescan", flags, 1, 0); 24162a458198SEric Sandeen fs_info->extent_workers = 2417cb001095SJeff Mahoney btrfs_alloc_workqueue(fs_info, "extent-refs", flags, 24182a458198SEric Sandeen min_t(u64, fs_devices->num_devices, 24192a458198SEric Sandeen max_active), 8); 24202a458198SEric Sandeen 24212a458198SEric Sandeen if (!(fs_info->workers && fs_info->delalloc_workers && 24222a458198SEric Sandeen fs_info->submit_workers && fs_info->flush_workers && 24232a458198SEric Sandeen fs_info->endio_workers && fs_info->endio_meta_workers && 24242a458198SEric Sandeen fs_info->endio_meta_write_workers && 24252a458198SEric Sandeen fs_info->endio_repair_workers && 24262a458198SEric Sandeen fs_info->endio_write_workers && fs_info->endio_raid56_workers && 24272a458198SEric Sandeen fs_info->endio_freespace_worker && fs_info->rmw_workers && 24282a458198SEric Sandeen fs_info->caching_workers && fs_info->readahead_workers && 24292a458198SEric Sandeen fs_info->fixup_workers && fs_info->delayed_workers && 24302a458198SEric Sandeen fs_info->extent_workers && 24312a458198SEric Sandeen fs_info->qgroup_rescan_workers)) { 24322a458198SEric Sandeen return -ENOMEM; 24332a458198SEric Sandeen } 24342a458198SEric Sandeen 24352a458198SEric Sandeen return 0; 24362a458198SEric Sandeen } 24372a458198SEric Sandeen 243863443bf5SEric Sandeen static int btrfs_replay_log(struct btrfs_fs_info *fs_info, 243963443bf5SEric Sandeen struct btrfs_fs_devices *fs_devices) 244063443bf5SEric Sandeen { 244163443bf5SEric Sandeen int ret; 244263443bf5SEric Sandeen struct btrfs_root *tree_root = fs_info->tree_root; 244363443bf5SEric Sandeen struct btrfs_root *log_tree_root; 244463443bf5SEric Sandeen struct btrfs_super_block *disk_super = fs_info->super_copy; 244563443bf5SEric Sandeen u64 bytenr = btrfs_super_log_root(disk_super); 244663443bf5SEric Sandeen 244763443bf5SEric Sandeen if (fs_devices->rw_devices == 0) { 2448f14d104dSDavid Sterba btrfs_warn(fs_info, "log replay required on RO media"); 244963443bf5SEric Sandeen return -EIO; 245063443bf5SEric Sandeen } 245163443bf5SEric Sandeen 245274e4d827SDavid Sterba log_tree_root = btrfs_alloc_root(fs_info, GFP_KERNEL); 245363443bf5SEric Sandeen if (!log_tree_root) 245463443bf5SEric Sandeen return -ENOMEM; 245563443bf5SEric Sandeen 2456da17066cSJeff Mahoney __setup_root(log_tree_root, fs_info, BTRFS_TREE_LOG_OBJECTID); 245763443bf5SEric Sandeen 245863443bf5SEric Sandeen log_tree_root->node = read_tree_block(tree_root, bytenr, 245963443bf5SEric Sandeen fs_info->generation + 1); 246064c043deSLiu Bo if (IS_ERR(log_tree_root->node)) { 2461f14d104dSDavid Sterba btrfs_warn(fs_info, "failed to read log tree"); 24620eeff236SLiu Bo ret = PTR_ERR(log_tree_root->node); 246364c043deSLiu Bo kfree(log_tree_root); 24640eeff236SLiu Bo return ret; 246564c043deSLiu Bo } else if (!extent_buffer_uptodate(log_tree_root->node)) { 2466f14d104dSDavid Sterba btrfs_err(fs_info, "failed to read log tree"); 246763443bf5SEric Sandeen free_extent_buffer(log_tree_root->node); 246863443bf5SEric Sandeen kfree(log_tree_root); 246963443bf5SEric Sandeen return -EIO; 247063443bf5SEric Sandeen } 247163443bf5SEric Sandeen /* returns with log_tree_root freed on success */ 247263443bf5SEric Sandeen ret = btrfs_recover_log_trees(log_tree_root); 247363443bf5SEric Sandeen if (ret) { 24740b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, 247563443bf5SEric Sandeen "Failed to recover log tree"); 247663443bf5SEric Sandeen free_extent_buffer(log_tree_root->node); 247763443bf5SEric Sandeen kfree(log_tree_root); 247863443bf5SEric Sandeen return ret; 247963443bf5SEric Sandeen } 248063443bf5SEric Sandeen 248163443bf5SEric Sandeen if (fs_info->sb->s_flags & MS_RDONLY) { 24826bccf3abSJeff Mahoney ret = btrfs_commit_super(fs_info); 248363443bf5SEric Sandeen if (ret) 248463443bf5SEric Sandeen return ret; 248563443bf5SEric Sandeen } 248663443bf5SEric Sandeen 248763443bf5SEric Sandeen return 0; 248863443bf5SEric Sandeen } 248963443bf5SEric Sandeen 24906bccf3abSJeff Mahoney static int btrfs_read_roots(struct btrfs_fs_info *fs_info) 24914bbcaa64SEric Sandeen { 24926bccf3abSJeff Mahoney struct btrfs_root *tree_root = fs_info->tree_root; 2493a4f3d2c4SDavid Sterba struct btrfs_root *root; 24944bbcaa64SEric Sandeen struct btrfs_key location; 24954bbcaa64SEric Sandeen int ret; 24964bbcaa64SEric Sandeen 24976bccf3abSJeff Mahoney BUG_ON(!fs_info->tree_root); 24986bccf3abSJeff Mahoney 24994bbcaa64SEric Sandeen location.objectid = BTRFS_EXTENT_TREE_OBJECTID; 25004bbcaa64SEric Sandeen location.type = BTRFS_ROOT_ITEM_KEY; 25014bbcaa64SEric Sandeen location.offset = 0; 25024bbcaa64SEric Sandeen 2503a4f3d2c4SDavid Sterba root = btrfs_read_tree_root(tree_root, &location); 2504a4f3d2c4SDavid Sterba if (IS_ERR(root)) 2505a4f3d2c4SDavid Sterba return PTR_ERR(root); 2506a4f3d2c4SDavid Sterba set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 2507a4f3d2c4SDavid Sterba fs_info->extent_root = root; 25084bbcaa64SEric Sandeen 25094bbcaa64SEric Sandeen location.objectid = BTRFS_DEV_TREE_OBJECTID; 2510a4f3d2c4SDavid Sterba root = btrfs_read_tree_root(tree_root, &location); 2511a4f3d2c4SDavid Sterba if (IS_ERR(root)) 2512a4f3d2c4SDavid Sterba return PTR_ERR(root); 2513a4f3d2c4SDavid Sterba set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 2514a4f3d2c4SDavid Sterba fs_info->dev_root = root; 25154bbcaa64SEric Sandeen btrfs_init_devices_late(fs_info); 25164bbcaa64SEric Sandeen 25174bbcaa64SEric Sandeen location.objectid = BTRFS_CSUM_TREE_OBJECTID; 2518a4f3d2c4SDavid Sterba root = btrfs_read_tree_root(tree_root, &location); 2519a4f3d2c4SDavid Sterba if (IS_ERR(root)) 2520a4f3d2c4SDavid Sterba return PTR_ERR(root); 2521a4f3d2c4SDavid Sterba set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 2522a4f3d2c4SDavid Sterba fs_info->csum_root = root; 25234bbcaa64SEric Sandeen 25244bbcaa64SEric Sandeen location.objectid = BTRFS_QUOTA_TREE_OBJECTID; 2525a4f3d2c4SDavid Sterba root = btrfs_read_tree_root(tree_root, &location); 2526a4f3d2c4SDavid Sterba if (!IS_ERR(root)) { 2527a4f3d2c4SDavid Sterba set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 2528afcdd129SJosef Bacik set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags); 2529a4f3d2c4SDavid Sterba fs_info->quota_root = root; 25304bbcaa64SEric Sandeen } 25314bbcaa64SEric Sandeen 25324bbcaa64SEric Sandeen location.objectid = BTRFS_UUID_TREE_OBJECTID; 2533a4f3d2c4SDavid Sterba root = btrfs_read_tree_root(tree_root, &location); 2534a4f3d2c4SDavid Sterba if (IS_ERR(root)) { 2535a4f3d2c4SDavid Sterba ret = PTR_ERR(root); 25364bbcaa64SEric Sandeen if (ret != -ENOENT) 25374bbcaa64SEric Sandeen return ret; 25384bbcaa64SEric Sandeen } else { 2539a4f3d2c4SDavid Sterba set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 2540a4f3d2c4SDavid Sterba fs_info->uuid_root = root; 25414bbcaa64SEric Sandeen } 25424bbcaa64SEric Sandeen 254370f6d82eSOmar Sandoval if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE)) { 254470f6d82eSOmar Sandoval location.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID; 254570f6d82eSOmar Sandoval root = btrfs_read_tree_root(tree_root, &location); 254670f6d82eSOmar Sandoval if (IS_ERR(root)) 254770f6d82eSOmar Sandoval return PTR_ERR(root); 254870f6d82eSOmar Sandoval set_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state); 254970f6d82eSOmar Sandoval fs_info->free_space_root = root; 255070f6d82eSOmar Sandoval } 255170f6d82eSOmar Sandoval 25524bbcaa64SEric Sandeen return 0; 25534bbcaa64SEric Sandeen } 25544bbcaa64SEric Sandeen 2555ad2b2c80SAl Viro int open_ctree(struct super_block *sb, 2556dfe25020SChris Mason struct btrfs_fs_devices *fs_devices, 2557dfe25020SChris Mason char *options) 2558eb60ceacSChris Mason { 2559db94535dSChris Mason u32 sectorsize; 2560db94535dSChris Mason u32 nodesize; 256187ee04ebSChris Mason u32 stripesize; 256284234f3aSYan Zheng u64 generation; 2563f2b636e8SJosef Bacik u64 features; 25643de4586cSChris Mason struct btrfs_key location; 2565a061fc8dSChris Mason struct buffer_head *bh; 25664d34b278SIlya Dryomov struct btrfs_super_block *disk_super; 2567815745cfSAl Viro struct btrfs_fs_info *fs_info = btrfs_sb(sb); 2568f84a8bd6SAl Viro struct btrfs_root *tree_root; 25694d34b278SIlya Dryomov struct btrfs_root *chunk_root; 2570eb60ceacSChris Mason int ret; 2571e58ca020SYan int err = -EINVAL; 2572af31f5e5SChris Mason int num_backups_tried = 0; 2573af31f5e5SChris Mason int backup_index = 0; 25745cdc7ad3SQu Wenruo int max_active; 25756675df31SOmar Sandoval int clear_free_space_tree = 0; 25764543df7eSChris Mason 257774e4d827SDavid Sterba tree_root = fs_info->tree_root = btrfs_alloc_root(fs_info, GFP_KERNEL); 257874e4d827SDavid Sterba chunk_root = fs_info->chunk_root = btrfs_alloc_root(fs_info, GFP_KERNEL); 2579cb517eabSMiao Xie if (!tree_root || !chunk_root) { 258039279cc3SChris Mason err = -ENOMEM; 258139279cc3SChris Mason goto fail; 258239279cc3SChris Mason } 258376dda93cSYan, Zheng 258476dda93cSYan, Zheng ret = init_srcu_struct(&fs_info->subvol_srcu); 258576dda93cSYan, Zheng if (ret) { 258676dda93cSYan, Zheng err = ret; 258776dda93cSYan, Zheng goto fail; 258876dda93cSYan, Zheng } 258976dda93cSYan, Zheng 259076dda93cSYan, Zheng ret = setup_bdi(fs_info, &fs_info->bdi); 259176dda93cSYan, Zheng if (ret) { 259276dda93cSYan, Zheng err = ret; 259376dda93cSYan, Zheng goto fail_srcu; 259476dda93cSYan, Zheng } 259576dda93cSYan, Zheng 2596908c7f19STejun Heo ret = percpu_counter_init(&fs_info->dirty_metadata_bytes, 0, GFP_KERNEL); 2597e2d84521SMiao Xie if (ret) { 2598e2d84521SMiao Xie err = ret; 2599e2d84521SMiao Xie goto fail_bdi; 2600e2d84521SMiao Xie } 260109cbfeafSKirill A. Shutemov fs_info->dirty_metadata_batch = PAGE_SIZE * 2602e2d84521SMiao Xie (1 + ilog2(nr_cpu_ids)); 2603e2d84521SMiao Xie 2604908c7f19STejun Heo ret = percpu_counter_init(&fs_info->delalloc_bytes, 0, GFP_KERNEL); 2605963d678bSMiao Xie if (ret) { 2606963d678bSMiao Xie err = ret; 2607963d678bSMiao Xie goto fail_dirty_metadata_bytes; 2608963d678bSMiao Xie } 2609963d678bSMiao Xie 2610908c7f19STejun Heo ret = percpu_counter_init(&fs_info->bio_counter, 0, GFP_KERNEL); 2611c404e0dcSMiao Xie if (ret) { 2612c404e0dcSMiao Xie err = ret; 2613c404e0dcSMiao Xie goto fail_delalloc_bytes; 2614c404e0dcSMiao Xie } 2615c404e0dcSMiao Xie 261676dda93cSYan, Zheng fs_info->btree_inode = new_inode(sb); 261776dda93cSYan, Zheng if (!fs_info->btree_inode) { 261876dda93cSYan, Zheng err = -ENOMEM; 2619c404e0dcSMiao Xie goto fail_bio_counter; 262076dda93cSYan, Zheng } 262176dda93cSYan, Zheng 2622a6591715SChris Mason mapping_set_gfp_mask(fs_info->btree_inode->i_mapping, GFP_NOFS); 26231561dedaSMiao Xie 262476dda93cSYan, Zheng INIT_RADIX_TREE(&fs_info->fs_roots_radix, GFP_ATOMIC); 2625f28491e0SJosef Bacik INIT_RADIX_TREE(&fs_info->buffer_radix, GFP_ATOMIC); 26268fd17795SChris Mason INIT_LIST_HEAD(&fs_info->trans_list); 2627facda1e7SChris Mason INIT_LIST_HEAD(&fs_info->dead_roots); 262824bbcf04SYan, Zheng INIT_LIST_HEAD(&fs_info->delayed_iputs); 2629eb73c1b7SMiao Xie INIT_LIST_HEAD(&fs_info->delalloc_roots); 263011833d66SYan Zheng INIT_LIST_HEAD(&fs_info->caching_block_groups); 2631eb73c1b7SMiao Xie spin_lock_init(&fs_info->delalloc_root_lock); 2632a4abeea4SJosef Bacik spin_lock_init(&fs_info->trans_lock); 263376dda93cSYan, Zheng spin_lock_init(&fs_info->fs_roots_radix_lock); 263424bbcf04SYan, Zheng spin_lock_init(&fs_info->delayed_iput_lock); 26354cb5300bSChris Mason spin_lock_init(&fs_info->defrag_inodes_lock); 26362bf64758SJosef Bacik spin_lock_init(&fs_info->free_chunk_lock); 2637f29021b2SJan Schmidt spin_lock_init(&fs_info->tree_mod_seq_lock); 2638ceda0864SMiao Xie spin_lock_init(&fs_info->super_lock); 2639fcebe456SJosef Bacik spin_lock_init(&fs_info->qgroup_op_lock); 2640f28491e0SJosef Bacik spin_lock_init(&fs_info->buffer_lock); 264147ab2a6cSJosef Bacik spin_lock_init(&fs_info->unused_bgs_lock); 2642f29021b2SJan Schmidt rwlock_init(&fs_info->tree_mod_log_lock); 2643d7c15171SZhao Lei mutex_init(&fs_info->unused_bg_unpin_mutex); 264467c5e7d4SFilipe Manana mutex_init(&fs_info->delete_unused_bgs_mutex); 26457585717fSChris Mason mutex_init(&fs_info->reloc_mutex); 2646573bfb72SMiao Xie mutex_init(&fs_info->delalloc_root_mutex); 2647c2d6cb16SFilipe Manana mutex_init(&fs_info->cleaner_delayed_iput_mutex); 2648de98ced9SMiao Xie seqlock_init(&fs_info->profiles_lock); 264919c00ddcSChris Mason 26500b86a832SChris Mason INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots); 26516324fbf3SChris Mason INIT_LIST_HEAD(&fs_info->space_info); 2652f29021b2SJan Schmidt INIT_LIST_HEAD(&fs_info->tree_mod_seq_list); 265347ab2a6cSJosef Bacik INIT_LIST_HEAD(&fs_info->unused_bgs); 26540b86a832SChris Mason btrfs_mapping_init(&fs_info->mapping_tree); 265566d8f3ddSMiao Xie btrfs_init_block_rsv(&fs_info->global_block_rsv, 265666d8f3ddSMiao Xie BTRFS_BLOCK_RSV_GLOBAL); 265766d8f3ddSMiao Xie btrfs_init_block_rsv(&fs_info->delalloc_block_rsv, 265866d8f3ddSMiao Xie BTRFS_BLOCK_RSV_DELALLOC); 265966d8f3ddSMiao Xie btrfs_init_block_rsv(&fs_info->trans_block_rsv, BTRFS_BLOCK_RSV_TRANS); 266066d8f3ddSMiao Xie btrfs_init_block_rsv(&fs_info->chunk_block_rsv, BTRFS_BLOCK_RSV_CHUNK); 266166d8f3ddSMiao Xie btrfs_init_block_rsv(&fs_info->empty_block_rsv, BTRFS_BLOCK_RSV_EMPTY); 266266d8f3ddSMiao Xie btrfs_init_block_rsv(&fs_info->delayed_block_rsv, 266366d8f3ddSMiao Xie BTRFS_BLOCK_RSV_DELOPS); 2664cb03c743SChris Mason atomic_set(&fs_info->nr_async_submits, 0); 2665771ed689SChris Mason atomic_set(&fs_info->async_delalloc_pages, 0); 26668c8bee1dSChris Mason atomic_set(&fs_info->async_submit_draining, 0); 26670986fe9eSChris Mason atomic_set(&fs_info->nr_async_bios, 0); 26684cb5300bSChris Mason atomic_set(&fs_info->defrag_running, 0); 2669fcebe456SJosef Bacik atomic_set(&fs_info->qgroup_op_seq, 0); 26702fefd558SZhao Lei atomic_set(&fs_info->reada_works_cnt, 0); 2671fc36ed7eSJan Schmidt atomic64_set(&fs_info->tree_mod_seq, 0); 26729e7cc91aSWang Xiaoguang fs_info->fs_frozen = 0; 2673e20d96d6SChris Mason fs_info->sb = sb; 267495ac567aSFilipe David Borba Manana fs_info->max_inline = BTRFS_DEFAULT_MAX_INLINE; 26759ed74f2dSJosef Bacik fs_info->metadata_ratio = 0; 26764cb5300bSChris Mason fs_info->defrag_inodes = RB_ROOT; 26772bf64758SJosef Bacik fs_info->free_chunk_space = 0; 2678f29021b2SJan Schmidt fs_info->tree_mod_log = RB_ROOT; 26798b87dc17SDavid Sterba fs_info->commit_interval = BTRFS_DEFAULT_COMMIT_INTERVAL; 2680f8c269d7SDavid Sterba fs_info->avg_delayed_ref_runtime = NSEC_PER_SEC >> 6; /* div by 64 */ 268190519d66SArne Jansen /* readahead state */ 2682d0164adcSMel Gorman INIT_RADIX_TREE(&fs_info->reada_tree, GFP_NOFS & ~__GFP_DIRECT_RECLAIM); 268390519d66SArne Jansen spin_lock_init(&fs_info->reada_lock); 2684c8b97818SChris Mason 2685b34b086cSChris Mason fs_info->thread_pool_size = min_t(unsigned long, 2686b34b086cSChris Mason num_online_cpus() + 2, 8); 26870afbaf8cSChris Mason 2688199c2a9cSMiao Xie INIT_LIST_HEAD(&fs_info->ordered_roots); 2689199c2a9cSMiao Xie spin_lock_init(&fs_info->ordered_root_lock); 269016cdcec7SMiao Xie fs_info->delayed_root = kmalloc(sizeof(struct btrfs_delayed_root), 269174e4d827SDavid Sterba GFP_KERNEL); 269216cdcec7SMiao Xie if (!fs_info->delayed_root) { 269316cdcec7SMiao Xie err = -ENOMEM; 269416cdcec7SMiao Xie goto fail_iput; 269516cdcec7SMiao Xie } 269616cdcec7SMiao Xie btrfs_init_delayed_root(fs_info->delayed_root); 26973eaa2885SChris Mason 2698638aa7edSEric Sandeen btrfs_init_scrub(fs_info); 269921adbd5cSStefan Behrens #ifdef CONFIG_BTRFS_FS_CHECK_INTEGRITY 270021adbd5cSStefan Behrens fs_info->check_integrity_print_mask = 0; 270121adbd5cSStefan Behrens #endif 2702779a65a4SEric Sandeen btrfs_init_balance(fs_info); 270321c7e756SMiao Xie btrfs_init_async_reclaim_work(&fs_info->async_reclaim_work); 2704a2de733cSArne Jansen 2705a061fc8dSChris Mason sb->s_blocksize = 4096; 2706a061fc8dSChris Mason sb->s_blocksize_bits = blksize_bits(4096); 270732a88aa1SJens Axboe sb->s_bdi = &fs_info->bdi; 2708a061fc8dSChris Mason 27096bccf3abSJeff Mahoney btrfs_init_btree_inode(fs_info); 271039279cc3SChris Mason 2711e02119d5SChris Mason spin_lock_init(&fs_info->block_group_cache_lock); 27126bef4d31SEric Paris fs_info->block_group_cache_tree = RB_ROOT; 2713a1897fddSLiu Bo fs_info->first_logical_byte = (u64)-1; 2714925baeddSChris Mason 271511833d66SYan Zheng extent_io_tree_init(&fs_info->freed_extents[0], 2716f993c883SDavid Sterba fs_info->btree_inode->i_mapping); 271711833d66SYan Zheng extent_io_tree_init(&fs_info->freed_extents[1], 2718f993c883SDavid Sterba fs_info->btree_inode->i_mapping); 271911833d66SYan Zheng fs_info->pinned_extents = &fs_info->freed_extents[0]; 2720afcdd129SJosef Bacik set_bit(BTRFS_FS_BARRIER, &fs_info->flags); 2721509659cdSChris Mason 27225a3f23d5SChris Mason mutex_init(&fs_info->ordered_operations_mutex); 272379154b1bSChris Mason mutex_init(&fs_info->tree_log_mutex); 2724925baeddSChris Mason mutex_init(&fs_info->chunk_mutex); 2725a74a4b97SChris Mason mutex_init(&fs_info->transaction_kthread_mutex); 2726a74a4b97SChris Mason mutex_init(&fs_info->cleaner_mutex); 27277d9eb12cSChris Mason mutex_init(&fs_info->volume_mutex); 27281bbc621eSChris Mason mutex_init(&fs_info->ro_block_group_mutex); 27299e351cc8SJosef Bacik init_rwsem(&fs_info->commit_root_sem); 2730c71bf099SYan, Zheng init_rwsem(&fs_info->cleanup_work_sem); 273176dda93cSYan, Zheng init_rwsem(&fs_info->subvol_sem); 2732803b2f54SStefan Behrens sema_init(&fs_info->uuid_tree_rescan_sem, 1); 2733fa9c0d79SChris Mason 2734ad618368SEric Sandeen btrfs_init_dev_replace_locks(fs_info); 2735f9e92e40SEric Sandeen btrfs_init_qgroup(fs_info); 2736416ac51dSArne Jansen 2737fa9c0d79SChris Mason btrfs_init_free_cluster(&fs_info->meta_alloc_cluster); 2738fa9c0d79SChris Mason btrfs_init_free_cluster(&fs_info->data_alloc_cluster); 2739fa9c0d79SChris Mason 2740e6dcd2dcSChris Mason init_waitqueue_head(&fs_info->transaction_throttle); 2741f9295749SChris Mason init_waitqueue_head(&fs_info->transaction_wait); 2742bb9c12c9SSage Weil init_waitqueue_head(&fs_info->transaction_blocked_wait); 27434854ddd0SChris Mason init_waitqueue_head(&fs_info->async_submit_wait); 27443768f368SChris Mason 274504216820SFilipe Manana INIT_LIST_HEAD(&fs_info->pinned_chunks); 274604216820SFilipe Manana 2747da17066cSJeff Mahoney /* Usable values until the real ones are cached from the superblock */ 2748da17066cSJeff Mahoney fs_info->nodesize = 4096; 2749da17066cSJeff Mahoney fs_info->sectorsize = 4096; 2750da17066cSJeff Mahoney fs_info->stripesize = 4096; 2751da17066cSJeff Mahoney 275253b381b3SDavid Woodhouse ret = btrfs_alloc_stripe_hash_table(fs_info); 275353b381b3SDavid Woodhouse if (ret) { 275483c8266aSDavid Sterba err = ret; 275553b381b3SDavid Woodhouse goto fail_alloc; 275653b381b3SDavid Woodhouse } 275753b381b3SDavid Woodhouse 2758da17066cSJeff Mahoney __setup_root(tree_root, fs_info, BTRFS_ROOT_TREE_OBJECTID); 27597eccb903SChris Mason 27603c4bb26bSChris Mason invalidate_bdev(fs_devices->latest_bdev); 27611104a885SDavid Sterba 27621104a885SDavid Sterba /* 27631104a885SDavid Sterba * Read super block and check the signature bytes only 27641104a885SDavid Sterba */ 2765a512bbf8SYan Zheng bh = btrfs_read_dev_super(fs_devices->latest_bdev); 276692fc03fbSAnand Jain if (IS_ERR(bh)) { 276792fc03fbSAnand Jain err = PTR_ERR(bh); 276816cdcec7SMiao Xie goto fail_alloc; 276920b45077SDave Young } 277039279cc3SChris Mason 27711104a885SDavid Sterba /* 27721104a885SDavid Sterba * We want to check superblock checksum, the type is stored inside. 27731104a885SDavid Sterba * Pass the whole disk block of size BTRFS_SUPER_INFO_SIZE (4k). 27741104a885SDavid Sterba */ 2775ab8d0fc4SJeff Mahoney if (btrfs_check_super_csum(fs_info, bh->b_data)) { 277605135f59SDavid Sterba btrfs_err(fs_info, "superblock checksum mismatch"); 27771104a885SDavid Sterba err = -EINVAL; 2778b2acdddfSAnand Jain brelse(bh); 27791104a885SDavid Sterba goto fail_alloc; 27801104a885SDavid Sterba } 27811104a885SDavid Sterba 27821104a885SDavid Sterba /* 27831104a885SDavid Sterba * super_copy is zeroed at allocation time and we never touch the 27841104a885SDavid Sterba * following bytes up to INFO_SIZE, the checksum is calculated from 27851104a885SDavid Sterba * the whole block of INFO_SIZE 27861104a885SDavid Sterba */ 27876c41761fSDavid Sterba memcpy(fs_info->super_copy, bh->b_data, sizeof(*fs_info->super_copy)); 27886c41761fSDavid Sterba memcpy(fs_info->super_for_commit, fs_info->super_copy, 27896c41761fSDavid Sterba sizeof(*fs_info->super_for_commit)); 2790a061fc8dSChris Mason brelse(bh); 27915f39d397SChris Mason 27926c41761fSDavid Sterba memcpy(fs_info->fsid, fs_info->super_copy->fsid, BTRFS_FSID_SIZE); 27930b86a832SChris Mason 27941104a885SDavid Sterba ret = btrfs_check_super_valid(fs_info, sb->s_flags & MS_RDONLY); 27951104a885SDavid Sterba if (ret) { 279605135f59SDavid Sterba btrfs_err(fs_info, "superblock contains fatal errors"); 27971104a885SDavid Sterba err = -EINVAL; 27981104a885SDavid Sterba goto fail_alloc; 27991104a885SDavid Sterba } 28001104a885SDavid Sterba 28016c41761fSDavid Sterba disk_super = fs_info->super_copy; 28020f7d52f4SChris Mason if (!btrfs_super_root(disk_super)) 280316cdcec7SMiao Xie goto fail_alloc; 28040f7d52f4SChris Mason 2805acce952bSliubo /* check FS state, whether FS is broken. */ 280687533c47SMiao Xie if (btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_ERROR) 280787533c47SMiao Xie set_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state); 2808acce952bSliubo 280975e7cb7fSLiu Bo /* 2810af31f5e5SChris Mason * run through our array of backup supers and setup 2811af31f5e5SChris Mason * our ring pointer to the oldest one 2812af31f5e5SChris Mason */ 2813af31f5e5SChris Mason generation = btrfs_super_generation(disk_super); 2814af31f5e5SChris Mason find_oldest_super_backup(fs_info, generation); 2815af31f5e5SChris Mason 2816af31f5e5SChris Mason /* 281775e7cb7fSLiu Bo * In the long term, we'll store the compression type in the super 281875e7cb7fSLiu Bo * block, and it'll be used for per file compression control. 281975e7cb7fSLiu Bo */ 282075e7cb7fSLiu Bo fs_info->compress_type = BTRFS_COMPRESS_ZLIB; 282175e7cb7fSLiu Bo 282296da0919SQu Wenruo ret = btrfs_parse_options(tree_root, options, sb->s_flags); 28232b82032cSYan Zheng if (ret) { 28242b82032cSYan Zheng err = ret; 282516cdcec7SMiao Xie goto fail_alloc; 28262b82032cSYan Zheng } 2827dfe25020SChris Mason 2828f2b636e8SJosef Bacik features = btrfs_super_incompat_flags(disk_super) & 2829f2b636e8SJosef Bacik ~BTRFS_FEATURE_INCOMPAT_SUPP; 2830f2b636e8SJosef Bacik if (features) { 283105135f59SDavid Sterba btrfs_err(fs_info, 283205135f59SDavid Sterba "cannot mount because of unsupported optional features (%llx)", 2833c1c9ff7cSGeert Uytterhoeven features); 2834f2b636e8SJosef Bacik err = -EINVAL; 283516cdcec7SMiao Xie goto fail_alloc; 2836f2b636e8SJosef Bacik } 2837f2b636e8SJosef Bacik 28385d4f98a2SYan Zheng features = btrfs_super_incompat_flags(disk_super); 28395d4f98a2SYan Zheng features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF; 28400b246afaSJeff Mahoney if (fs_info->compress_type == BTRFS_COMPRESS_LZO) 2841a6fa6faeSLi Zefan features |= BTRFS_FEATURE_INCOMPAT_COMPRESS_LZO; 2842727011e0SChris Mason 28433173a18fSJosef Bacik if (features & BTRFS_FEATURE_INCOMPAT_SKINNY_METADATA) 284405135f59SDavid Sterba btrfs_info(fs_info, "has skinny extents"); 28453173a18fSJosef Bacik 2846727011e0SChris Mason /* 2847727011e0SChris Mason * flag our filesystem as having big metadata blocks if 2848727011e0SChris Mason * they are bigger than the page size 2849727011e0SChris Mason */ 285009cbfeafSKirill A. Shutemov if (btrfs_super_nodesize(disk_super) > PAGE_SIZE) { 2851727011e0SChris Mason if (!(features & BTRFS_FEATURE_INCOMPAT_BIG_METADATA)) 285205135f59SDavid Sterba btrfs_info(fs_info, 285305135f59SDavid Sterba "flagging fs with big metadata feature"); 2854727011e0SChris Mason features |= BTRFS_FEATURE_INCOMPAT_BIG_METADATA; 2855727011e0SChris Mason } 2856727011e0SChris Mason 2857bc3f116fSChris Mason nodesize = btrfs_super_nodesize(disk_super); 2858bc3f116fSChris Mason sectorsize = btrfs_super_sectorsize(disk_super); 2859b7f67055SChandan Rajendra stripesize = sectorsize; 2860707e8a07SDavid Sterba fs_info->dirty_metadata_batch = nodesize * (1 + ilog2(nr_cpu_ids)); 2861963d678bSMiao Xie fs_info->delalloc_batch = sectorsize * 512 * (1 + ilog2(nr_cpu_ids)); 2862bc3f116fSChris Mason 2863da17066cSJeff Mahoney /* Cache block sizes */ 2864da17066cSJeff Mahoney fs_info->nodesize = nodesize; 2865da17066cSJeff Mahoney fs_info->sectorsize = sectorsize; 2866da17066cSJeff Mahoney fs_info->stripesize = stripesize; 2867da17066cSJeff Mahoney 2868bc3f116fSChris Mason /* 2869bc3f116fSChris Mason * mixed block groups end up with duplicate but slightly offset 2870bc3f116fSChris Mason * extent buffers for the same range. It leads to corruptions 2871bc3f116fSChris Mason */ 2872bc3f116fSChris Mason if ((features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) && 2873707e8a07SDavid Sterba (sectorsize != nodesize)) { 287405135f59SDavid Sterba btrfs_err(fs_info, 287505135f59SDavid Sterba "unequal nodesize/sectorsize (%u != %u) are not allowed for mixed block groups", 287605135f59SDavid Sterba nodesize, sectorsize); 2877bc3f116fSChris Mason goto fail_alloc; 2878bc3f116fSChris Mason } 2879bc3f116fSChris Mason 2880ceda0864SMiao Xie /* 2881ceda0864SMiao Xie * Needn't use the lock because there is no other task which will 2882ceda0864SMiao Xie * update the flag. 2883ceda0864SMiao Xie */ 28845d4f98a2SYan Zheng btrfs_set_super_incompat_flags(disk_super, features); 28855d4f98a2SYan Zheng 2886f2b636e8SJosef Bacik features = btrfs_super_compat_ro_flags(disk_super) & 2887f2b636e8SJosef Bacik ~BTRFS_FEATURE_COMPAT_RO_SUPP; 2888f2b636e8SJosef Bacik if (!(sb->s_flags & MS_RDONLY) && features) { 288905135f59SDavid Sterba btrfs_err(fs_info, 289005135f59SDavid Sterba "cannot mount read-write because of unsupported optional features (%llx)", 2891c1c9ff7cSGeert Uytterhoeven features); 2892f2b636e8SJosef Bacik err = -EINVAL; 289316cdcec7SMiao Xie goto fail_alloc; 2894f2b636e8SJosef Bacik } 289561d92c32SChris Mason 28965cdc7ad3SQu Wenruo max_active = fs_info->thread_pool_size; 289761d92c32SChris Mason 28982a458198SEric Sandeen ret = btrfs_init_workqueues(fs_info, fs_devices); 28992a458198SEric Sandeen if (ret) { 29002a458198SEric Sandeen err = ret; 29010dc3b84aSJosef Bacik goto fail_sb_buffer; 29020dc3b84aSJosef Bacik } 29034543df7eSChris Mason 29044575c9ccSChris Mason fs_info->bdi.ra_pages *= btrfs_super_num_devices(disk_super); 2905c8b97818SChris Mason fs_info->bdi.ra_pages = max(fs_info->bdi.ra_pages, 290609cbfeafSKirill A. Shutemov SZ_4M / PAGE_SIZE); 29074575c9ccSChris Mason 2908a061fc8dSChris Mason sb->s_blocksize = sectorsize; 2909a061fc8dSChris Mason sb->s_blocksize_bits = blksize_bits(sectorsize); 2910db94535dSChris Mason 2911925baeddSChris Mason mutex_lock(&fs_info->chunk_mutex); 29126bccf3abSJeff Mahoney ret = btrfs_read_sys_array(fs_info); 2913925baeddSChris Mason mutex_unlock(&fs_info->chunk_mutex); 291484eed90fSChris Mason if (ret) { 291505135f59SDavid Sterba btrfs_err(fs_info, "failed to read the system array: %d", ret); 29165d4f98a2SYan Zheng goto fail_sb_buffer; 291784eed90fSChris Mason } 29180b86a832SChris Mason 291984234f3aSYan Zheng generation = btrfs_super_chunk_root_generation(disk_super); 29200b86a832SChris Mason 2921da17066cSJeff Mahoney __setup_root(chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID); 29220b86a832SChris Mason 29230b86a832SChris Mason chunk_root->node = read_tree_block(chunk_root, 29240b86a832SChris Mason btrfs_super_chunk_root(disk_super), 2925ce86cd59SDavid Sterba generation); 292664c043deSLiu Bo if (IS_ERR(chunk_root->node) || 292764c043deSLiu Bo !extent_buffer_uptodate(chunk_root->node)) { 292805135f59SDavid Sterba btrfs_err(fs_info, "failed to read chunk root"); 2929e5fffbacSchandan if (!IS_ERR(chunk_root->node)) 2930e5fffbacSchandan free_extent_buffer(chunk_root->node); 293195ab1f64SZhao Lei chunk_root->node = NULL; 2932af31f5e5SChris Mason goto fail_tree_roots; 293383121942SDavid Woodhouse } 29345d4f98a2SYan Zheng btrfs_set_root_node(&chunk_root->root_item, chunk_root->node); 29355d4f98a2SYan Zheng chunk_root->commit_root = btrfs_root_node(chunk_root); 29360b86a832SChris Mason 2937e17cade2SChris Mason read_extent_buffer(chunk_root->node, fs_info->chunk_tree_uuid, 2938b308bc2fSGeert Uytterhoeven btrfs_header_chunk_tree_uuid(chunk_root->node), BTRFS_UUID_SIZE); 2939e17cade2SChris Mason 29405b4aacefSJeff Mahoney ret = btrfs_read_chunk_tree(fs_info); 29412b82032cSYan Zheng if (ret) { 294205135f59SDavid Sterba btrfs_err(fs_info, "failed to read chunk tree: %d", ret); 2943af31f5e5SChris Mason goto fail_tree_roots; 29442b82032cSYan Zheng } 29450b86a832SChris Mason 29468dabb742SStefan Behrens /* 29478dabb742SStefan Behrens * keep the device that is marked to be the target device for the 29488dabb742SStefan Behrens * dev_replace procedure 29498dabb742SStefan Behrens */ 29509eaed21eSEric Sandeen btrfs_close_extra_devices(fs_devices, 0); 2951dfe25020SChris Mason 2952a6b0d5c8SChris Mason if (!fs_devices->latest_bdev) { 295305135f59SDavid Sterba btrfs_err(fs_info, "failed to read devices"); 2954a6b0d5c8SChris Mason goto fail_tree_roots; 2955a6b0d5c8SChris Mason } 2956a6b0d5c8SChris Mason 2957af31f5e5SChris Mason retry_root_backup: 295884234f3aSYan Zheng generation = btrfs_super_generation(disk_super); 29590b86a832SChris Mason 2960e20d96d6SChris Mason tree_root->node = read_tree_block(tree_root, 2961db94535dSChris Mason btrfs_super_root(disk_super), 2962ce86cd59SDavid Sterba generation); 296364c043deSLiu Bo if (IS_ERR(tree_root->node) || 296464c043deSLiu Bo !extent_buffer_uptodate(tree_root->node)) { 296505135f59SDavid Sterba btrfs_warn(fs_info, "failed to read tree root"); 2966e5fffbacSchandan if (!IS_ERR(tree_root->node)) 2967e5fffbacSchandan free_extent_buffer(tree_root->node); 296895ab1f64SZhao Lei tree_root->node = NULL; 2969af31f5e5SChris Mason goto recovery_tree_root; 297083121942SDavid Woodhouse } 2971af31f5e5SChris Mason 29725d4f98a2SYan Zheng btrfs_set_root_node(&tree_root->root_item, tree_root->node); 29735d4f98a2SYan Zheng tree_root->commit_root = btrfs_root_node(tree_root); 297469e9c6c6SStefan Behrens btrfs_set_root_refs(&tree_root->root_item, 1); 2975db94535dSChris Mason 2976f32e48e9SChandan Rajendra mutex_lock(&tree_root->objectid_mutex); 2977f32e48e9SChandan Rajendra ret = btrfs_find_highest_objectid(tree_root, 2978f32e48e9SChandan Rajendra &tree_root->highest_objectid); 2979f32e48e9SChandan Rajendra if (ret) { 2980f32e48e9SChandan Rajendra mutex_unlock(&tree_root->objectid_mutex); 2981f32e48e9SChandan Rajendra goto recovery_tree_root; 2982f32e48e9SChandan Rajendra } 2983f32e48e9SChandan Rajendra 2984f32e48e9SChandan Rajendra ASSERT(tree_root->highest_objectid <= BTRFS_LAST_FREE_OBJECTID); 2985f32e48e9SChandan Rajendra 2986f32e48e9SChandan Rajendra mutex_unlock(&tree_root->objectid_mutex); 2987f32e48e9SChandan Rajendra 29886bccf3abSJeff Mahoney ret = btrfs_read_roots(fs_info); 29894bbcaa64SEric Sandeen if (ret) 2990af31f5e5SChris Mason goto recovery_tree_root; 2991f7a81ea4SStefan Behrens 29928929ecfaSYan, Zheng fs_info->generation = generation; 29938929ecfaSYan, Zheng fs_info->last_trans_committed = generation; 29948929ecfaSYan, Zheng 299568310a5eSIlya Dryomov ret = btrfs_recover_balance(fs_info); 299668310a5eSIlya Dryomov if (ret) { 299705135f59SDavid Sterba btrfs_err(fs_info, "failed to recover balance: %d", ret); 299868310a5eSIlya Dryomov goto fail_block_groups; 299968310a5eSIlya Dryomov } 300068310a5eSIlya Dryomov 3001733f4fbbSStefan Behrens ret = btrfs_init_dev_stats(fs_info); 3002733f4fbbSStefan Behrens if (ret) { 300305135f59SDavid Sterba btrfs_err(fs_info, "failed to init dev_stats: %d", ret); 3004733f4fbbSStefan Behrens goto fail_block_groups; 3005733f4fbbSStefan Behrens } 3006733f4fbbSStefan Behrens 30078dabb742SStefan Behrens ret = btrfs_init_dev_replace(fs_info); 30088dabb742SStefan Behrens if (ret) { 300905135f59SDavid Sterba btrfs_err(fs_info, "failed to init dev_replace: %d", ret); 30108dabb742SStefan Behrens goto fail_block_groups; 30118dabb742SStefan Behrens } 30128dabb742SStefan Behrens 30139eaed21eSEric Sandeen btrfs_close_extra_devices(fs_devices, 1); 30148dabb742SStefan Behrens 3015b7c35e81SAnand Jain ret = btrfs_sysfs_add_fsid(fs_devices, NULL); 3016b7c35e81SAnand Jain if (ret) { 301705135f59SDavid Sterba btrfs_err(fs_info, "failed to init sysfs fsid interface: %d", 301805135f59SDavid Sterba ret); 3019b7c35e81SAnand Jain goto fail_block_groups; 3020b7c35e81SAnand Jain } 3021b7c35e81SAnand Jain 3022b7c35e81SAnand Jain ret = btrfs_sysfs_add_device(fs_devices); 3023b7c35e81SAnand Jain if (ret) { 302405135f59SDavid Sterba btrfs_err(fs_info, "failed to init sysfs device interface: %d", 302505135f59SDavid Sterba ret); 3026b7c35e81SAnand Jain goto fail_fsdev_sysfs; 3027b7c35e81SAnand Jain } 3028b7c35e81SAnand Jain 302996f3136eSAnand Jain ret = btrfs_sysfs_add_mounted(fs_info); 30305ac1d209SJeff Mahoney if (ret) { 303105135f59SDavid Sterba btrfs_err(fs_info, "failed to init sysfs interface: %d", ret); 3032b7c35e81SAnand Jain goto fail_fsdev_sysfs; 30335ac1d209SJeff Mahoney } 30345ac1d209SJeff Mahoney 3035c59021f8Sliubo ret = btrfs_init_space_info(fs_info); 3036c59021f8Sliubo if (ret) { 303705135f59SDavid Sterba btrfs_err(fs_info, "failed to initialize space info: %d", ret); 30382365dd3cSAnand Jain goto fail_sysfs; 3039c59021f8Sliubo } 3040c59021f8Sliubo 30415b4aacefSJeff Mahoney ret = btrfs_read_block_groups(fs_info); 30421b1d1f66SJosef Bacik if (ret) { 304305135f59SDavid Sterba btrfs_err(fs_info, "failed to read block groups: %d", ret); 30442365dd3cSAnand Jain goto fail_sysfs; 30451b1d1f66SJosef Bacik } 30465af3e8ccSStefan Behrens fs_info->num_tolerated_disk_barrier_failures = 30475af3e8ccSStefan Behrens btrfs_calc_num_tolerated_disk_barrier_failures(fs_info); 3048292fd7fcSStefan Behrens if (fs_info->fs_devices->missing_devices > 3049292fd7fcSStefan Behrens fs_info->num_tolerated_disk_barrier_failures && 3050292fd7fcSStefan Behrens !(sb->s_flags & MS_RDONLY)) { 305105135f59SDavid Sterba btrfs_warn(fs_info, 305205135f59SDavid Sterba "missing devices (%llu) exceeds the limit (%d), writeable mount is not allowed", 305378fa1770SZhao Lei fs_info->fs_devices->missing_devices, 305478fa1770SZhao Lei fs_info->num_tolerated_disk_barrier_failures); 30552365dd3cSAnand Jain goto fail_sysfs; 3056292fd7fcSStefan Behrens } 30579078a3e1SChris Mason 3058a74a4b97SChris Mason fs_info->cleaner_kthread = kthread_run(cleaner_kthread, tree_root, 3059a74a4b97SChris Mason "btrfs-cleaner"); 306057506d50SQinghuang Feng if (IS_ERR(fs_info->cleaner_kthread)) 30612365dd3cSAnand Jain goto fail_sysfs; 3062a74a4b97SChris Mason 3063a74a4b97SChris Mason fs_info->transaction_kthread = kthread_run(transaction_kthread, 3064a74a4b97SChris Mason tree_root, 3065a74a4b97SChris Mason "btrfs-transaction"); 306657506d50SQinghuang Feng if (IS_ERR(fs_info->transaction_kthread)) 30673f157a2fSChris Mason goto fail_cleaner; 3068a74a4b97SChris Mason 30690b246afaSJeff Mahoney if (!btrfs_test_opt(fs_info, SSD) && 30700b246afaSJeff Mahoney !btrfs_test_opt(fs_info, NOSSD) && 3071c289811cSChris Mason !fs_info->fs_devices->rotating) { 307205135f59SDavid Sterba btrfs_info(fs_info, "detected SSD devices, enabling SSD mode"); 3073c289811cSChris Mason btrfs_set_opt(fs_info->mount_opt, SSD); 3074c289811cSChris Mason } 3075c289811cSChris Mason 3076572d9ab7SDavid Sterba /* 307701327610SNicholas D Steeves * Mount does not set all options immediately, we can do it now and do 3078572d9ab7SDavid Sterba * not have to wait for transaction commit 3079572d9ab7SDavid Sterba */ 3080572d9ab7SDavid Sterba btrfs_apply_pending_changes(fs_info); 30813818aea2SQu Wenruo 308221adbd5cSStefan Behrens #ifdef CONFIG_BTRFS_FS_CHECK_INTEGRITY 30830b246afaSJeff Mahoney if (btrfs_test_opt(fs_info, CHECK_INTEGRITY)) { 308421adbd5cSStefan Behrens ret = btrfsic_mount(tree_root, fs_devices, 30850b246afaSJeff Mahoney btrfs_test_opt(fs_info, 308621adbd5cSStefan Behrens CHECK_INTEGRITY_INCLUDING_EXTENT_DATA) ? 308721adbd5cSStefan Behrens 1 : 0, 308821adbd5cSStefan Behrens fs_info->check_integrity_print_mask); 308921adbd5cSStefan Behrens if (ret) 309005135f59SDavid Sterba btrfs_warn(fs_info, 309105135f59SDavid Sterba "failed to initialize integrity check module: %d", 309205135f59SDavid Sterba ret); 309321adbd5cSStefan Behrens } 309421adbd5cSStefan Behrens #endif 3095bcef60f2SArne Jansen ret = btrfs_read_qgroup_config(fs_info); 3096bcef60f2SArne Jansen if (ret) 3097bcef60f2SArne Jansen goto fail_trans_kthread; 309821adbd5cSStefan Behrens 309996da0919SQu Wenruo /* do not make disk changes in broken FS or nologreplay is given */ 310096da0919SQu Wenruo if (btrfs_super_log_root(disk_super) != 0 && 31010b246afaSJeff Mahoney !btrfs_test_opt(fs_info, NOLOGREPLAY)) { 310263443bf5SEric Sandeen ret = btrfs_replay_log(fs_info, fs_devices); 310379787eaaSJeff Mahoney if (ret) { 310463443bf5SEric Sandeen err = ret; 310528c16cbbSWang Shilong goto fail_qgroup; 3106e556ce2cSYan Zheng } 3107e02119d5SChris Mason } 31081a40e23bSZheng Yan 31096bccf3abSJeff Mahoney ret = btrfs_find_orphan_roots(fs_info); 311079787eaaSJeff Mahoney if (ret) 311128c16cbbSWang Shilong goto fail_qgroup; 311276dda93cSYan, Zheng 31137c2ca468SChris Mason if (!(sb->s_flags & MS_RDONLY)) { 3114d68fc57bSYan, Zheng ret = btrfs_cleanup_fs_roots(fs_info); 311544c44af2SIlya Dryomov if (ret) 311628c16cbbSWang Shilong goto fail_qgroup; 311790c711abSZygo Blaxell 311890c711abSZygo Blaxell mutex_lock(&fs_info->cleaner_mutex); 31195d4f98a2SYan Zheng ret = btrfs_recover_relocation(tree_root); 312090c711abSZygo Blaxell mutex_unlock(&fs_info->cleaner_mutex); 3121d7ce5843SMiao Xie if (ret < 0) { 312205135f59SDavid Sterba btrfs_warn(fs_info, "failed to recover relocation: %d", 312305135f59SDavid Sterba ret); 3124d7ce5843SMiao Xie err = -EINVAL; 3125bcef60f2SArne Jansen goto fail_qgroup; 3126d7ce5843SMiao Xie } 31277c2ca468SChris Mason } 31281a40e23bSZheng Yan 31293de4586cSChris Mason location.objectid = BTRFS_FS_TREE_OBJECTID; 31303de4586cSChris Mason location.type = BTRFS_ROOT_ITEM_KEY; 3131cb517eabSMiao Xie location.offset = 0; 31323de4586cSChris Mason 31333de4586cSChris Mason fs_info->fs_root = btrfs_read_fs_root_no_name(fs_info, &location); 31343140c9a3SDan Carpenter if (IS_ERR(fs_info->fs_root)) { 31353140c9a3SDan Carpenter err = PTR_ERR(fs_info->fs_root); 3136bcef60f2SArne Jansen goto fail_qgroup; 31373140c9a3SDan Carpenter } 3138c289811cSChris Mason 31392b6ba629SIlya Dryomov if (sb->s_flags & MS_RDONLY) 31402b6ba629SIlya Dryomov return 0; 31412b6ba629SIlya Dryomov 3142f8d468a1SOmar Sandoval if (btrfs_test_opt(fs_info, CLEAR_CACHE) && 3143f8d468a1SOmar Sandoval btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE)) { 31446675df31SOmar Sandoval clear_free_space_tree = 1; 31456675df31SOmar Sandoval } else if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) && 31466675df31SOmar Sandoval !btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID)) { 31476675df31SOmar Sandoval btrfs_warn(fs_info, "free space tree is invalid"); 31486675df31SOmar Sandoval clear_free_space_tree = 1; 31496675df31SOmar Sandoval } 31506675df31SOmar Sandoval 31516675df31SOmar Sandoval if (clear_free_space_tree) { 3152f8d468a1SOmar Sandoval btrfs_info(fs_info, "clearing free space tree"); 3153f8d468a1SOmar Sandoval ret = btrfs_clear_free_space_tree(fs_info); 3154f8d468a1SOmar Sandoval if (ret) { 3155f8d468a1SOmar Sandoval btrfs_warn(fs_info, 3156f8d468a1SOmar Sandoval "failed to clear free space tree: %d", ret); 31576bccf3abSJeff Mahoney close_ctree(fs_info); 3158f8d468a1SOmar Sandoval return ret; 3159f8d468a1SOmar Sandoval } 3160f8d468a1SOmar Sandoval } 3161f8d468a1SOmar Sandoval 31620b246afaSJeff Mahoney if (btrfs_test_opt(fs_info, FREE_SPACE_TREE) && 3163511711afSChris Mason !btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE)) { 316405135f59SDavid Sterba btrfs_info(fs_info, "creating free space tree"); 3165511711afSChris Mason ret = btrfs_create_free_space_tree(fs_info); 3166511711afSChris Mason if (ret) { 316705135f59SDavid Sterba btrfs_warn(fs_info, 316805135f59SDavid Sterba "failed to create free space tree: %d", ret); 31696bccf3abSJeff Mahoney close_ctree(fs_info); 3170511711afSChris Mason return ret; 3171511711afSChris Mason } 3172511711afSChris Mason } 3173511711afSChris Mason 3174e3acc2a6SJosef Bacik down_read(&fs_info->cleanup_work_sem); 31752b6ba629SIlya Dryomov if ((ret = btrfs_orphan_cleanup(fs_info->fs_root)) || 31762b6ba629SIlya Dryomov (ret = btrfs_orphan_cleanup(fs_info->tree_root))) { 31772b6ba629SIlya Dryomov up_read(&fs_info->cleanup_work_sem); 31786bccf3abSJeff Mahoney close_ctree(fs_info); 31792b6ba629SIlya Dryomov return ret; 31802b6ba629SIlya Dryomov } 3181e3acc2a6SJosef Bacik up_read(&fs_info->cleanup_work_sem); 318259641015SIlya Dryomov 31832b6ba629SIlya Dryomov ret = btrfs_resume_balance_async(fs_info); 31842b6ba629SIlya Dryomov if (ret) { 318505135f59SDavid Sterba btrfs_warn(fs_info, "failed to resume balance: %d", ret); 31866bccf3abSJeff Mahoney close_ctree(fs_info); 31872b6ba629SIlya Dryomov return ret; 3188e3acc2a6SJosef Bacik } 3189e3acc2a6SJosef Bacik 31908dabb742SStefan Behrens ret = btrfs_resume_dev_replace_async(fs_info); 31918dabb742SStefan Behrens if (ret) { 319205135f59SDavid Sterba btrfs_warn(fs_info, "failed to resume device replace: %d", ret); 31936bccf3abSJeff Mahoney close_ctree(fs_info); 31948dabb742SStefan Behrens return ret; 31958dabb742SStefan Behrens } 31968dabb742SStefan Behrens 3197b382a324SJan Schmidt btrfs_qgroup_rescan_resume(fs_info); 3198b382a324SJan Schmidt 31994bbcaa64SEric Sandeen if (!fs_info->uuid_root) { 320005135f59SDavid Sterba btrfs_info(fs_info, "creating UUID tree"); 3201f7a81ea4SStefan Behrens ret = btrfs_create_uuid_tree(fs_info); 3202f7a81ea4SStefan Behrens if (ret) { 320305135f59SDavid Sterba btrfs_warn(fs_info, 320405135f59SDavid Sterba "failed to create the UUID tree: %d", ret); 32056bccf3abSJeff Mahoney close_ctree(fs_info); 3206f7a81ea4SStefan Behrens return ret; 3207f7a81ea4SStefan Behrens } 32080b246afaSJeff Mahoney } else if (btrfs_test_opt(fs_info, RESCAN_UUID_TREE) || 32094bbcaa64SEric Sandeen fs_info->generation != 32104bbcaa64SEric Sandeen btrfs_super_uuid_tree_generation(disk_super)) { 321105135f59SDavid Sterba btrfs_info(fs_info, "checking UUID tree"); 321270f80175SStefan Behrens ret = btrfs_check_uuid_tree(fs_info); 321370f80175SStefan Behrens if (ret) { 321405135f59SDavid Sterba btrfs_warn(fs_info, 321505135f59SDavid Sterba "failed to check the UUID tree: %d", ret); 32166bccf3abSJeff Mahoney close_ctree(fs_info); 321770f80175SStefan Behrens return ret; 321870f80175SStefan Behrens } 321970f80175SStefan Behrens } else { 3220afcdd129SJosef Bacik set_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags); 3221f7a81ea4SStefan Behrens } 3222afcdd129SJosef Bacik set_bit(BTRFS_FS_OPEN, &fs_info->flags); 322347ab2a6cSJosef Bacik 32248dcddfa0SQu Wenruo /* 32258dcddfa0SQu Wenruo * backuproot only affect mount behavior, and if open_ctree succeeded, 32268dcddfa0SQu Wenruo * no need to keep the flag 32278dcddfa0SQu Wenruo */ 32288dcddfa0SQu Wenruo btrfs_clear_opt(fs_info->mount_opt, USEBACKUPROOT); 32298dcddfa0SQu Wenruo 3230ad2b2c80SAl Viro return 0; 323139279cc3SChris Mason 3232bcef60f2SArne Jansen fail_qgroup: 3233bcef60f2SArne Jansen btrfs_free_qgroup_config(fs_info); 32347c2ca468SChris Mason fail_trans_kthread: 32357c2ca468SChris Mason kthread_stop(fs_info->transaction_kthread); 323654067ae9SJosef Bacik btrfs_cleanup_transaction(fs_info->tree_root); 3237faa2dbf0SJosef Bacik btrfs_free_fs_roots(fs_info); 32383f157a2fSChris Mason fail_cleaner: 3239a74a4b97SChris Mason kthread_stop(fs_info->cleaner_kthread); 32407c2ca468SChris Mason 32417c2ca468SChris Mason /* 32427c2ca468SChris Mason * make sure we're done with the btree inode before we stop our 32437c2ca468SChris Mason * kthreads 32447c2ca468SChris Mason */ 32457c2ca468SChris Mason filemap_write_and_wait(fs_info->btree_inode->i_mapping); 32467c2ca468SChris Mason 32472365dd3cSAnand Jain fail_sysfs: 32486618a59bSAnand Jain btrfs_sysfs_remove_mounted(fs_info); 32492365dd3cSAnand Jain 3250b7c35e81SAnand Jain fail_fsdev_sysfs: 3251b7c35e81SAnand Jain btrfs_sysfs_remove_fsid(fs_info->fs_devices); 3252b7c35e81SAnand Jain 32531b1d1f66SJosef Bacik fail_block_groups: 325454067ae9SJosef Bacik btrfs_put_block_group_cache(fs_info); 32551b1d1f66SJosef Bacik btrfs_free_block_groups(fs_info); 3256af31f5e5SChris Mason 3257af31f5e5SChris Mason fail_tree_roots: 3258af31f5e5SChris Mason free_root_pointers(fs_info, 1); 32592b8195bbSMiao Xie invalidate_inode_pages2(fs_info->btree_inode->i_mapping); 3260af31f5e5SChris Mason 326139279cc3SChris Mason fail_sb_buffer: 32627abadb64SLiu Bo btrfs_stop_all_workers(fs_info); 326316cdcec7SMiao Xie fail_alloc: 32644543df7eSChris Mason fail_iput: 3265586e46e2SIlya Dryomov btrfs_mapping_tree_free(&fs_info->mapping_tree); 3266586e46e2SIlya Dryomov 32674543df7eSChris Mason iput(fs_info->btree_inode); 3268c404e0dcSMiao Xie fail_bio_counter: 3269c404e0dcSMiao Xie percpu_counter_destroy(&fs_info->bio_counter); 3270963d678bSMiao Xie fail_delalloc_bytes: 3271963d678bSMiao Xie percpu_counter_destroy(&fs_info->delalloc_bytes); 3272e2d84521SMiao Xie fail_dirty_metadata_bytes: 3273e2d84521SMiao Xie percpu_counter_destroy(&fs_info->dirty_metadata_bytes); 3274ad081f14SJens Axboe fail_bdi: 32757e662854SQinghuang Feng bdi_destroy(&fs_info->bdi); 327676dda93cSYan, Zheng fail_srcu: 327776dda93cSYan, Zheng cleanup_srcu_struct(&fs_info->subvol_srcu); 32787e662854SQinghuang Feng fail: 327953b381b3SDavid Woodhouse btrfs_free_stripe_hash_table(fs_info); 3280586e46e2SIlya Dryomov btrfs_close_devices(fs_info->fs_devices); 3281ad2b2c80SAl Viro return err; 3282af31f5e5SChris Mason 3283af31f5e5SChris Mason recovery_tree_root: 32840b246afaSJeff Mahoney if (!btrfs_test_opt(fs_info, USEBACKUPROOT)) 3285af31f5e5SChris Mason goto fail_tree_roots; 3286af31f5e5SChris Mason 3287af31f5e5SChris Mason free_root_pointers(fs_info, 0); 3288af31f5e5SChris Mason 3289af31f5e5SChris Mason /* don't use the log in recovery mode, it won't be valid */ 3290af31f5e5SChris Mason btrfs_set_super_log_root(disk_super, 0); 3291af31f5e5SChris Mason 3292af31f5e5SChris Mason /* we can't trust the free space cache either */ 3293af31f5e5SChris Mason btrfs_set_opt(fs_info->mount_opt, CLEAR_CACHE); 3294af31f5e5SChris Mason 3295af31f5e5SChris Mason ret = next_root_backup(fs_info, fs_info->super_copy, 3296af31f5e5SChris Mason &num_backups_tried, &backup_index); 3297af31f5e5SChris Mason if (ret == -1) 3298af31f5e5SChris Mason goto fail_block_groups; 3299af31f5e5SChris Mason goto retry_root_backup; 3300eb60ceacSChris Mason } 3301eb60ceacSChris Mason 3302f2984462SChris Mason static void btrfs_end_buffer_write_sync(struct buffer_head *bh, int uptodate) 3303f2984462SChris Mason { 3304f2984462SChris Mason if (uptodate) { 3305f2984462SChris Mason set_buffer_uptodate(bh); 3306f2984462SChris Mason } else { 3307442a4f63SStefan Behrens struct btrfs_device *device = (struct btrfs_device *) 3308442a4f63SStefan Behrens bh->b_private; 3309442a4f63SStefan Behrens 3310fb456252SJeff Mahoney btrfs_warn_rl_in_rcu(device->fs_info, 3311b14af3b4SDavid Sterba "lost page write due to IO error on %s", 3312606686eeSJosef Bacik rcu_str_deref(device->name)); 331301327610SNicholas D Steeves /* note, we don't set_buffer_write_io_error because we have 33141259ab75SChris Mason * our own ways of dealing with the IO errors 33151259ab75SChris Mason */ 3316f2984462SChris Mason clear_buffer_uptodate(bh); 3317442a4f63SStefan Behrens btrfs_dev_stat_inc_and_print(device, BTRFS_DEV_STAT_WRITE_ERRS); 3318f2984462SChris Mason } 3319f2984462SChris Mason unlock_buffer(bh); 3320f2984462SChris Mason put_bh(bh); 3321f2984462SChris Mason } 3322f2984462SChris Mason 332329c36d72SAnand Jain int btrfs_read_dev_one_super(struct block_device *bdev, int copy_num, 332429c36d72SAnand Jain struct buffer_head **bh_ret) 332529c36d72SAnand Jain { 332629c36d72SAnand Jain struct buffer_head *bh; 332729c36d72SAnand Jain struct btrfs_super_block *super; 332829c36d72SAnand Jain u64 bytenr; 332929c36d72SAnand Jain 333029c36d72SAnand Jain bytenr = btrfs_sb_offset(copy_num); 333129c36d72SAnand Jain if (bytenr + BTRFS_SUPER_INFO_SIZE >= i_size_read(bdev->bd_inode)) 333229c36d72SAnand Jain return -EINVAL; 333329c36d72SAnand Jain 333429c36d72SAnand Jain bh = __bread(bdev, bytenr / 4096, BTRFS_SUPER_INFO_SIZE); 333529c36d72SAnand Jain /* 333629c36d72SAnand Jain * If we fail to read from the underlying devices, as of now 333729c36d72SAnand Jain * the best option we have is to mark it EIO. 333829c36d72SAnand Jain */ 333929c36d72SAnand Jain if (!bh) 334029c36d72SAnand Jain return -EIO; 334129c36d72SAnand Jain 334229c36d72SAnand Jain super = (struct btrfs_super_block *)bh->b_data; 334329c36d72SAnand Jain if (btrfs_super_bytenr(super) != bytenr || 334429c36d72SAnand Jain btrfs_super_magic(super) != BTRFS_MAGIC) { 334529c36d72SAnand Jain brelse(bh); 334629c36d72SAnand Jain return -EINVAL; 334729c36d72SAnand Jain } 334829c36d72SAnand Jain 334929c36d72SAnand Jain *bh_ret = bh; 335029c36d72SAnand Jain return 0; 335129c36d72SAnand Jain } 335229c36d72SAnand Jain 335329c36d72SAnand Jain 3354a512bbf8SYan Zheng struct buffer_head *btrfs_read_dev_super(struct block_device *bdev) 3355a512bbf8SYan Zheng { 3356a512bbf8SYan Zheng struct buffer_head *bh; 3357a512bbf8SYan Zheng struct buffer_head *latest = NULL; 3358a512bbf8SYan Zheng struct btrfs_super_block *super; 3359a512bbf8SYan Zheng int i; 3360a512bbf8SYan Zheng u64 transid = 0; 336192fc03fbSAnand Jain int ret = -EINVAL; 3362a512bbf8SYan Zheng 3363a512bbf8SYan Zheng /* we would like to check all the supers, but that would make 3364a512bbf8SYan Zheng * a btrfs mount succeed after a mkfs from a different FS. 3365a512bbf8SYan Zheng * So, we need to add a special mount option to scan for 3366a512bbf8SYan Zheng * later supers, using BTRFS_SUPER_MIRROR_MAX instead 3367a512bbf8SYan Zheng */ 3368a512bbf8SYan Zheng for (i = 0; i < 1; i++) { 336929c36d72SAnand Jain ret = btrfs_read_dev_one_super(bdev, i, &bh); 337029c36d72SAnand Jain if (ret) 3371a512bbf8SYan Zheng continue; 3372a512bbf8SYan Zheng 3373a512bbf8SYan Zheng super = (struct btrfs_super_block *)bh->b_data; 3374a512bbf8SYan Zheng 3375a512bbf8SYan Zheng if (!latest || btrfs_super_generation(super) > transid) { 3376a512bbf8SYan Zheng brelse(latest); 3377a512bbf8SYan Zheng latest = bh; 3378a512bbf8SYan Zheng transid = btrfs_super_generation(super); 3379a512bbf8SYan Zheng } else { 3380a512bbf8SYan Zheng brelse(bh); 3381a512bbf8SYan Zheng } 3382a512bbf8SYan Zheng } 338392fc03fbSAnand Jain 338492fc03fbSAnand Jain if (!latest) 338592fc03fbSAnand Jain return ERR_PTR(ret); 338692fc03fbSAnand Jain 3387a512bbf8SYan Zheng return latest; 3388a512bbf8SYan Zheng } 3389a512bbf8SYan Zheng 33904eedeb75SHisashi Hifumi /* 33914eedeb75SHisashi Hifumi * this should be called twice, once with wait == 0 and 33924eedeb75SHisashi Hifumi * once with wait == 1. When wait == 0 is done, all the buffer heads 33934eedeb75SHisashi Hifumi * we write are pinned. 33944eedeb75SHisashi Hifumi * 33954eedeb75SHisashi Hifumi * They are released when wait == 1 is done. 33964eedeb75SHisashi Hifumi * max_mirrors must be the same for both runs, and it indicates how 33974eedeb75SHisashi Hifumi * many supers on this one device should be written. 33984eedeb75SHisashi Hifumi * 33994eedeb75SHisashi Hifumi * max_mirrors == 0 means to write them all. 34004eedeb75SHisashi Hifumi */ 3401a512bbf8SYan Zheng static int write_dev_supers(struct btrfs_device *device, 3402a512bbf8SYan Zheng struct btrfs_super_block *sb, 3403a512bbf8SYan Zheng int do_barriers, int wait, int max_mirrors) 3404a512bbf8SYan Zheng { 3405a512bbf8SYan Zheng struct buffer_head *bh; 3406a512bbf8SYan Zheng int i; 3407a512bbf8SYan Zheng int ret; 3408a512bbf8SYan Zheng int errors = 0; 3409a512bbf8SYan Zheng u32 crc; 3410a512bbf8SYan Zheng u64 bytenr; 3411a512bbf8SYan Zheng 3412a512bbf8SYan Zheng if (max_mirrors == 0) 3413a512bbf8SYan Zheng max_mirrors = BTRFS_SUPER_MIRROR_MAX; 3414a512bbf8SYan Zheng 3415a512bbf8SYan Zheng for (i = 0; i < max_mirrors; i++) { 3416a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 3417935e5cc9SMiao Xie if (bytenr + BTRFS_SUPER_INFO_SIZE >= 3418935e5cc9SMiao Xie device->commit_total_bytes) 3419a512bbf8SYan Zheng break; 3420a512bbf8SYan Zheng 3421a512bbf8SYan Zheng if (wait) { 3422a512bbf8SYan Zheng bh = __find_get_block(device->bdev, bytenr / 4096, 3423a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE); 3424634554dcSJosef Bacik if (!bh) { 3425634554dcSJosef Bacik errors++; 3426634554dcSJosef Bacik continue; 3427634554dcSJosef Bacik } 3428a512bbf8SYan Zheng wait_on_buffer(bh); 34294eedeb75SHisashi Hifumi if (!buffer_uptodate(bh)) 34304eedeb75SHisashi Hifumi errors++; 34314eedeb75SHisashi Hifumi 34324eedeb75SHisashi Hifumi /* drop our reference */ 34334eedeb75SHisashi Hifumi brelse(bh); 34344eedeb75SHisashi Hifumi 34354eedeb75SHisashi Hifumi /* drop the reference from the wait == 0 run */ 3436a512bbf8SYan Zheng brelse(bh); 3437a512bbf8SYan Zheng continue; 3438a512bbf8SYan Zheng } else { 3439a512bbf8SYan Zheng btrfs_set_super_bytenr(sb, bytenr); 3440a512bbf8SYan Zheng 3441a512bbf8SYan Zheng crc = ~(u32)0; 3442b0496686SLiu Bo crc = btrfs_csum_data((char *)sb + 3443a512bbf8SYan Zheng BTRFS_CSUM_SIZE, crc, 3444a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE - 3445a512bbf8SYan Zheng BTRFS_CSUM_SIZE); 3446a512bbf8SYan Zheng btrfs_csum_final(crc, sb->csum); 3447a512bbf8SYan Zheng 34484eedeb75SHisashi Hifumi /* 34494eedeb75SHisashi Hifumi * one reference for us, and we leave it for the 34504eedeb75SHisashi Hifumi * caller 34514eedeb75SHisashi Hifumi */ 3452a512bbf8SYan Zheng bh = __getblk(device->bdev, bytenr / 4096, 3453a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE); 3454634554dcSJosef Bacik if (!bh) { 3455fb456252SJeff Mahoney btrfs_err(device->fs_info, 3456f14d104dSDavid Sterba "couldn't get super buffer head for bytenr %llu", 3457f14d104dSDavid Sterba bytenr); 3458634554dcSJosef Bacik errors++; 3459634554dcSJosef Bacik continue; 3460634554dcSJosef Bacik } 3461634554dcSJosef Bacik 3462a512bbf8SYan Zheng memcpy(bh->b_data, sb, BTRFS_SUPER_INFO_SIZE); 3463a512bbf8SYan Zheng 34644eedeb75SHisashi Hifumi /* one reference for submit_bh */ 3465a512bbf8SYan Zheng get_bh(bh); 34664eedeb75SHisashi Hifumi 34674eedeb75SHisashi Hifumi set_buffer_uptodate(bh); 3468a512bbf8SYan Zheng lock_buffer(bh); 3469a512bbf8SYan Zheng bh->b_end_io = btrfs_end_buffer_write_sync; 3470442a4f63SStefan Behrens bh->b_private = device; 3471a512bbf8SYan Zheng } 3472a512bbf8SYan Zheng 3473387125fcSChris Mason /* 3474387125fcSChris Mason * we fua the first super. The others we allow 3475387125fcSChris Mason * to go down lazy. 3476387125fcSChris Mason */ 3477e8117c26SWang Shilong if (i == 0) 34782a222ca9SMike Christie ret = btrfsic_submit_bh(REQ_OP_WRITE, WRITE_FUA, bh); 3479e8117c26SWang Shilong else 34802a222ca9SMike Christie ret = btrfsic_submit_bh(REQ_OP_WRITE, WRITE_SYNC, bh); 34814eedeb75SHisashi Hifumi if (ret) 3482a512bbf8SYan Zheng errors++; 3483a512bbf8SYan Zheng } 3484a512bbf8SYan Zheng return errors < i ? 0 : -1; 3485a512bbf8SYan Zheng } 3486a512bbf8SYan Zheng 3487387125fcSChris Mason /* 3488387125fcSChris Mason * endio for the write_dev_flush, this will wake anyone waiting 3489387125fcSChris Mason * for the barrier when it is done 3490387125fcSChris Mason */ 34914246a0b6SChristoph Hellwig static void btrfs_end_empty_barrier(struct bio *bio) 3492387125fcSChris Mason { 3493387125fcSChris Mason if (bio->bi_private) 3494387125fcSChris Mason complete(bio->bi_private); 3495387125fcSChris Mason bio_put(bio); 3496387125fcSChris Mason } 3497387125fcSChris Mason 3498387125fcSChris Mason /* 3499387125fcSChris Mason * trigger flushes for one the devices. If you pass wait == 0, the flushes are 3500387125fcSChris Mason * sent down. With wait == 1, it waits for the previous flush. 3501387125fcSChris Mason * 3502387125fcSChris Mason * any device where the flush fails with eopnotsupp are flagged as not-barrier 3503387125fcSChris Mason * capable 3504387125fcSChris Mason */ 3505387125fcSChris Mason static int write_dev_flush(struct btrfs_device *device, int wait) 3506387125fcSChris Mason { 3507387125fcSChris Mason struct bio *bio; 3508387125fcSChris Mason int ret = 0; 3509387125fcSChris Mason 3510387125fcSChris Mason if (device->nobarriers) 3511387125fcSChris Mason return 0; 3512387125fcSChris Mason 3513387125fcSChris Mason if (wait) { 3514387125fcSChris Mason bio = device->flush_bio; 3515387125fcSChris Mason if (!bio) 3516387125fcSChris Mason return 0; 3517387125fcSChris Mason 3518387125fcSChris Mason wait_for_completion(&device->flush_wait); 3519387125fcSChris Mason 35204246a0b6SChristoph Hellwig if (bio->bi_error) { 35214246a0b6SChristoph Hellwig ret = bio->bi_error; 3522442a4f63SStefan Behrens btrfs_dev_stat_inc_and_print(device, 3523442a4f63SStefan Behrens BTRFS_DEV_STAT_FLUSH_ERRS); 3524387125fcSChris Mason } 3525387125fcSChris Mason 3526387125fcSChris Mason /* drop the reference from the wait == 0 run */ 3527387125fcSChris Mason bio_put(bio); 3528387125fcSChris Mason device->flush_bio = NULL; 3529387125fcSChris Mason 3530387125fcSChris Mason return ret; 3531387125fcSChris Mason } 3532387125fcSChris Mason 3533387125fcSChris Mason /* 3534387125fcSChris Mason * one reference for us, and we leave it for the 3535387125fcSChris Mason * caller 3536387125fcSChris Mason */ 35379c017abcSJesper Juhl device->flush_bio = NULL; 35389be3395bSChris Mason bio = btrfs_io_bio_alloc(GFP_NOFS, 0); 3539387125fcSChris Mason if (!bio) 3540387125fcSChris Mason return -ENOMEM; 3541387125fcSChris Mason 3542387125fcSChris Mason bio->bi_end_io = btrfs_end_empty_barrier; 3543387125fcSChris Mason bio->bi_bdev = device->bdev; 354437226b21SMike Christie bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_FLUSH); 3545387125fcSChris Mason init_completion(&device->flush_wait); 3546387125fcSChris Mason bio->bi_private = &device->flush_wait; 3547387125fcSChris Mason device->flush_bio = bio; 3548387125fcSChris Mason 3549387125fcSChris Mason bio_get(bio); 35504e49ea4aSMike Christie btrfsic_submit_bio(bio); 3551387125fcSChris Mason 3552387125fcSChris Mason return 0; 3553387125fcSChris Mason } 3554387125fcSChris Mason 3555387125fcSChris Mason /* 3556387125fcSChris Mason * send an empty flush down to each device in parallel, 3557387125fcSChris Mason * then wait for them 3558387125fcSChris Mason */ 3559387125fcSChris Mason static int barrier_all_devices(struct btrfs_fs_info *info) 3560387125fcSChris Mason { 3561387125fcSChris Mason struct list_head *head; 3562387125fcSChris Mason struct btrfs_device *dev; 35635af3e8ccSStefan Behrens int errors_send = 0; 35645af3e8ccSStefan Behrens int errors_wait = 0; 3565387125fcSChris Mason int ret; 3566387125fcSChris Mason 3567387125fcSChris Mason /* send down all the barriers */ 3568387125fcSChris Mason head = &info->fs_devices->devices; 3569387125fcSChris Mason list_for_each_entry_rcu(dev, head, dev_list) { 3570f88ba6a2SHidetoshi Seto if (dev->missing) 3571f88ba6a2SHidetoshi Seto continue; 3572387125fcSChris Mason if (!dev->bdev) { 35735af3e8ccSStefan Behrens errors_send++; 3574387125fcSChris Mason continue; 3575387125fcSChris Mason } 3576387125fcSChris Mason if (!dev->in_fs_metadata || !dev->writeable) 3577387125fcSChris Mason continue; 3578387125fcSChris Mason 3579387125fcSChris Mason ret = write_dev_flush(dev, 0); 3580387125fcSChris Mason if (ret) 35815af3e8ccSStefan Behrens errors_send++; 3582387125fcSChris Mason } 3583387125fcSChris Mason 3584387125fcSChris Mason /* wait for all the barriers */ 3585387125fcSChris Mason list_for_each_entry_rcu(dev, head, dev_list) { 3586f88ba6a2SHidetoshi Seto if (dev->missing) 3587f88ba6a2SHidetoshi Seto continue; 3588387125fcSChris Mason if (!dev->bdev) { 35895af3e8ccSStefan Behrens errors_wait++; 3590387125fcSChris Mason continue; 3591387125fcSChris Mason } 3592387125fcSChris Mason if (!dev->in_fs_metadata || !dev->writeable) 3593387125fcSChris Mason continue; 3594387125fcSChris Mason 3595387125fcSChris Mason ret = write_dev_flush(dev, 1); 3596387125fcSChris Mason if (ret) 35975af3e8ccSStefan Behrens errors_wait++; 3598387125fcSChris Mason } 35995af3e8ccSStefan Behrens if (errors_send > info->num_tolerated_disk_barrier_failures || 36005af3e8ccSStefan Behrens errors_wait > info->num_tolerated_disk_barrier_failures) 3601387125fcSChris Mason return -EIO; 3602387125fcSChris Mason return 0; 3603387125fcSChris Mason } 3604387125fcSChris Mason 3605943c6e99SZhao Lei int btrfs_get_num_tolerated_disk_barrier_failures(u64 flags) 3606943c6e99SZhao Lei { 36078789f4feSZhao Lei int raid_type; 36088789f4feSZhao Lei int min_tolerated = INT_MAX; 3609943c6e99SZhao Lei 36108789f4feSZhao Lei if ((flags & BTRFS_BLOCK_GROUP_PROFILE_MASK) == 0 || 36118789f4feSZhao Lei (flags & BTRFS_AVAIL_ALLOC_BIT_SINGLE)) 36128789f4feSZhao Lei min_tolerated = min(min_tolerated, 36138789f4feSZhao Lei btrfs_raid_array[BTRFS_RAID_SINGLE]. 36148789f4feSZhao Lei tolerated_failures); 3615943c6e99SZhao Lei 36168789f4feSZhao Lei for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) { 36178789f4feSZhao Lei if (raid_type == BTRFS_RAID_SINGLE) 36188789f4feSZhao Lei continue; 36198789f4feSZhao Lei if (!(flags & btrfs_raid_group[raid_type])) 36208789f4feSZhao Lei continue; 36218789f4feSZhao Lei min_tolerated = min(min_tolerated, 36228789f4feSZhao Lei btrfs_raid_array[raid_type]. 36238789f4feSZhao Lei tolerated_failures); 36248789f4feSZhao Lei } 3625943c6e99SZhao Lei 36268789f4feSZhao Lei if (min_tolerated == INT_MAX) { 3627ab8d0fc4SJeff Mahoney pr_warn("BTRFS: unknown raid flag: %llu", flags); 36288789f4feSZhao Lei min_tolerated = 0; 36298789f4feSZhao Lei } 36308789f4feSZhao Lei 36318789f4feSZhao Lei return min_tolerated; 3632943c6e99SZhao Lei } 3633943c6e99SZhao Lei 36345af3e8ccSStefan Behrens int btrfs_calc_num_tolerated_disk_barrier_failures( 36355af3e8ccSStefan Behrens struct btrfs_fs_info *fs_info) 36365af3e8ccSStefan Behrens { 36375af3e8ccSStefan Behrens struct btrfs_ioctl_space_info space; 36385af3e8ccSStefan Behrens struct btrfs_space_info *sinfo; 36395af3e8ccSStefan Behrens u64 types[] = {BTRFS_BLOCK_GROUP_DATA, 36405af3e8ccSStefan Behrens BTRFS_BLOCK_GROUP_SYSTEM, 36415af3e8ccSStefan Behrens BTRFS_BLOCK_GROUP_METADATA, 36425af3e8ccSStefan Behrens BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA}; 36435af3e8ccSStefan Behrens int i; 36445af3e8ccSStefan Behrens int c; 36455af3e8ccSStefan Behrens int num_tolerated_disk_barrier_failures = 36465af3e8ccSStefan Behrens (int)fs_info->fs_devices->num_devices; 36475af3e8ccSStefan Behrens 36482c458045SZhao Lei for (i = 0; i < ARRAY_SIZE(types); i++) { 36495af3e8ccSStefan Behrens struct btrfs_space_info *tmp; 36505af3e8ccSStefan Behrens 36515af3e8ccSStefan Behrens sinfo = NULL; 36525af3e8ccSStefan Behrens rcu_read_lock(); 36535af3e8ccSStefan Behrens list_for_each_entry_rcu(tmp, &fs_info->space_info, list) { 36545af3e8ccSStefan Behrens if (tmp->flags == types[i]) { 36555af3e8ccSStefan Behrens sinfo = tmp; 36565af3e8ccSStefan Behrens break; 36575af3e8ccSStefan Behrens } 36585af3e8ccSStefan Behrens } 36595af3e8ccSStefan Behrens rcu_read_unlock(); 36605af3e8ccSStefan Behrens 36615af3e8ccSStefan Behrens if (!sinfo) 36625af3e8ccSStefan Behrens continue; 36635af3e8ccSStefan Behrens 36645af3e8ccSStefan Behrens down_read(&sinfo->groups_sem); 36655af3e8ccSStefan Behrens for (c = 0; c < BTRFS_NR_RAID_TYPES; c++) { 36665af3e8ccSStefan Behrens u64 flags; 36675af3e8ccSStefan Behrens 36682c458045SZhao Lei if (list_empty(&sinfo->block_groups[c])) 36692c458045SZhao Lei continue; 36702c458045SZhao Lei 36712c458045SZhao Lei btrfs_get_block_group_info(&sinfo->block_groups[c], 36722c458045SZhao Lei &space); 36732c458045SZhao Lei if (space.total_bytes == 0 || space.used_bytes == 0) 36745af3e8ccSStefan Behrens continue; 36755af3e8ccSStefan Behrens flags = space.flags; 3676943c6e99SZhao Lei 3677943c6e99SZhao Lei num_tolerated_disk_barrier_failures = min( 3678943c6e99SZhao Lei num_tolerated_disk_barrier_failures, 3679943c6e99SZhao Lei btrfs_get_num_tolerated_disk_barrier_failures( 3680943c6e99SZhao Lei flags)); 36815af3e8ccSStefan Behrens } 36825af3e8ccSStefan Behrens up_read(&sinfo->groups_sem); 36835af3e8ccSStefan Behrens } 36845af3e8ccSStefan Behrens 36855af3e8ccSStefan Behrens return num_tolerated_disk_barrier_failures; 36865af3e8ccSStefan Behrens } 36875af3e8ccSStefan Behrens 368848a3b636SEric Sandeen static int write_all_supers(struct btrfs_root *root, int max_mirrors) 3689f2984462SChris Mason { 36900b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 3691e5e9a520SChris Mason struct list_head *head; 3692f2984462SChris Mason struct btrfs_device *dev; 3693a061fc8dSChris Mason struct btrfs_super_block *sb; 3694f2984462SChris Mason struct btrfs_dev_item *dev_item; 3695f2984462SChris Mason int ret; 3696f2984462SChris Mason int do_barriers; 3697a236aed1SChris Mason int max_errors; 3698a236aed1SChris Mason int total_errors = 0; 3699a061fc8dSChris Mason u64 flags; 3700f2984462SChris Mason 37010b246afaSJeff Mahoney do_barriers = !btrfs_test_opt(fs_info, NOBARRIER); 37020b246afaSJeff Mahoney backup_super_roots(fs_info); 3703f2984462SChris Mason 37040b246afaSJeff Mahoney sb = fs_info->super_for_commit; 3705a061fc8dSChris Mason dev_item = &sb->dev_item; 3706e5e9a520SChris Mason 37070b246afaSJeff Mahoney mutex_lock(&fs_info->fs_devices->device_list_mutex); 37080b246afaSJeff Mahoney head = &fs_info->fs_devices->devices; 37090b246afaSJeff Mahoney max_errors = btrfs_super_num_devices(fs_info->super_copy) - 1; 3710387125fcSChris Mason 37115af3e8ccSStefan Behrens if (do_barriers) { 37120b246afaSJeff Mahoney ret = barrier_all_devices(fs_info); 37135af3e8ccSStefan Behrens if (ret) { 37145af3e8ccSStefan Behrens mutex_unlock( 37150b246afaSJeff Mahoney &fs_info->fs_devices->device_list_mutex); 37160b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, ret, 37175af3e8ccSStefan Behrens "errors while submitting device barriers."); 37185af3e8ccSStefan Behrens return ret; 37195af3e8ccSStefan Behrens } 37205af3e8ccSStefan Behrens } 3721387125fcSChris Mason 37221f78160cSXiao Guangrong list_for_each_entry_rcu(dev, head, dev_list) { 3723dfe25020SChris Mason if (!dev->bdev) { 3724dfe25020SChris Mason total_errors++; 3725dfe25020SChris Mason continue; 3726dfe25020SChris Mason } 37272b82032cSYan Zheng if (!dev->in_fs_metadata || !dev->writeable) 3728dfe25020SChris Mason continue; 3729dfe25020SChris Mason 37302b82032cSYan Zheng btrfs_set_stack_device_generation(dev_item, 0); 3731a061fc8dSChris Mason btrfs_set_stack_device_type(dev_item, dev->type); 3732a061fc8dSChris Mason btrfs_set_stack_device_id(dev_item, dev->devid); 37337df69d3eSMiao Xie btrfs_set_stack_device_total_bytes(dev_item, 3734935e5cc9SMiao Xie dev->commit_total_bytes); 3735ce7213c7SMiao Xie btrfs_set_stack_device_bytes_used(dev_item, 3736ce7213c7SMiao Xie dev->commit_bytes_used); 3737a061fc8dSChris Mason btrfs_set_stack_device_io_align(dev_item, dev->io_align); 3738a061fc8dSChris Mason btrfs_set_stack_device_io_width(dev_item, dev->io_width); 3739a061fc8dSChris Mason btrfs_set_stack_device_sector_size(dev_item, dev->sector_size); 3740a061fc8dSChris Mason memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE); 37412b82032cSYan Zheng memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE); 3742a512bbf8SYan Zheng 3743a061fc8dSChris Mason flags = btrfs_super_flags(sb); 3744a061fc8dSChris Mason btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN); 3745f2984462SChris Mason 3746a512bbf8SYan Zheng ret = write_dev_supers(dev, sb, do_barriers, 0, max_mirrors); 3747a236aed1SChris Mason if (ret) 3748a236aed1SChris Mason total_errors++; 3749f2984462SChris Mason } 3750a236aed1SChris Mason if (total_errors > max_errors) { 37510b246afaSJeff Mahoney btrfs_err(fs_info, "%d errors while writing supers", 3752d397712bSChris Mason total_errors); 37530b246afaSJeff Mahoney mutex_unlock(&fs_info->fs_devices->device_list_mutex); 375479787eaaSJeff Mahoney 37559d565ba4SStefan Behrens /* FUA is masked off if unsupported and can't be the reason */ 37560b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, -EIO, 37570b246afaSJeff Mahoney "%d errors while writing supers", 37580b246afaSJeff Mahoney total_errors); 37599d565ba4SStefan Behrens return -EIO; 3760a236aed1SChris Mason } 3761f2984462SChris Mason 3762a512bbf8SYan Zheng total_errors = 0; 37631f78160cSXiao Guangrong list_for_each_entry_rcu(dev, head, dev_list) { 3764dfe25020SChris Mason if (!dev->bdev) 3765dfe25020SChris Mason continue; 37662b82032cSYan Zheng if (!dev->in_fs_metadata || !dev->writeable) 3767dfe25020SChris Mason continue; 3768dfe25020SChris Mason 3769a512bbf8SYan Zheng ret = write_dev_supers(dev, sb, do_barriers, 1, max_mirrors); 3770a512bbf8SYan Zheng if (ret) 37711259ab75SChris Mason total_errors++; 3772f2984462SChris Mason } 37730b246afaSJeff Mahoney mutex_unlock(&fs_info->fs_devices->device_list_mutex); 3774a236aed1SChris Mason if (total_errors > max_errors) { 37750b246afaSJeff Mahoney btrfs_handle_fs_error(fs_info, -EIO, 37760b246afaSJeff Mahoney "%d errors while writing supers", 37770b246afaSJeff Mahoney total_errors); 377879787eaaSJeff Mahoney return -EIO; 3779a236aed1SChris Mason } 3780f2984462SChris Mason return 0; 3781f2984462SChris Mason } 3782f2984462SChris Mason 3783a512bbf8SYan Zheng int write_ctree_super(struct btrfs_trans_handle *trans, 3784a512bbf8SYan Zheng struct btrfs_root *root, int max_mirrors) 3785cfaa7295SChris Mason { 3786f570e757SRashika return write_all_supers(root, max_mirrors); 3787cfaa7295SChris Mason } 3788cfaa7295SChris Mason 3789cb517eabSMiao Xie /* Drop a fs root from the radix tree and free it. */ 3790cb517eabSMiao Xie void btrfs_drop_and_free_fs_root(struct btrfs_fs_info *fs_info, 3791cb517eabSMiao Xie struct btrfs_root *root) 37922619ba1fSChris Mason { 37934df27c4dSYan, Zheng spin_lock(&fs_info->fs_roots_radix_lock); 37942619ba1fSChris Mason radix_tree_delete(&fs_info->fs_roots_radix, 37952619ba1fSChris Mason (unsigned long)root->root_key.objectid); 37964df27c4dSYan, Zheng spin_unlock(&fs_info->fs_roots_radix_lock); 379776dda93cSYan, Zheng 379876dda93cSYan, Zheng if (btrfs_root_refs(&root->root_item) == 0) 379976dda93cSYan, Zheng synchronize_srcu(&fs_info->subvol_srcu); 380076dda93cSYan, Zheng 38011c1ea4f7SLiu Bo if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) { 38023321719eSLiu Bo btrfs_free_log(NULL, root); 38031c1ea4f7SLiu Bo if (root->reloc_root) { 38041c1ea4f7SLiu Bo free_extent_buffer(root->reloc_root->node); 38051c1ea4f7SLiu Bo free_extent_buffer(root->reloc_root->commit_root); 38061c1ea4f7SLiu Bo btrfs_put_fs_root(root->reloc_root); 38071c1ea4f7SLiu Bo root->reloc_root = NULL; 38081c1ea4f7SLiu Bo } 38091c1ea4f7SLiu Bo } 38103321719eSLiu Bo 3811faa2dbf0SJosef Bacik if (root->free_ino_pinned) 3812581bb050SLi Zefan __btrfs_remove_free_space_cache(root->free_ino_pinned); 3813faa2dbf0SJosef Bacik if (root->free_ino_ctl) 3814581bb050SLi Zefan __btrfs_remove_free_space_cache(root->free_ino_ctl); 38154df27c4dSYan, Zheng free_fs_root(root); 38164df27c4dSYan, Zheng } 38174df27c4dSYan, Zheng 38184df27c4dSYan, Zheng static void free_fs_root(struct btrfs_root *root) 38194df27c4dSYan, Zheng { 382057cdc8dbSDavid Sterba iput(root->ino_cache_inode); 38214df27c4dSYan, Zheng WARN_ON(!RB_EMPTY_ROOT(&root->inode_tree)); 38221cb048f5SFilipe David Borba Manana btrfs_free_block_rsv(root, root->orphan_block_rsv); 38231cb048f5SFilipe David Borba Manana root->orphan_block_rsv = NULL; 38240ee5dc67SAl Viro if (root->anon_dev) 38250ee5dc67SAl Viro free_anon_bdev(root->anon_dev); 38268257b2dcSMiao Xie if (root->subv_writers) 38278257b2dcSMiao Xie btrfs_free_subvolume_writers(root->subv_writers); 38285f39d397SChris Mason free_extent_buffer(root->node); 38295f39d397SChris Mason free_extent_buffer(root->commit_root); 3830581bb050SLi Zefan kfree(root->free_ino_ctl); 3831581bb050SLi Zefan kfree(root->free_ino_pinned); 383258176a96SJosef Bacik kfree(root->name); 3833b0feb9d9SMiao Xie btrfs_put_fs_root(root); 38342619ba1fSChris Mason } 38352619ba1fSChris Mason 3836cb517eabSMiao Xie void btrfs_free_fs_root(struct btrfs_root *root) 3837cb517eabSMiao Xie { 3838cb517eabSMiao Xie free_fs_root(root); 3839eb60ceacSChris Mason } 3840eb60ceacSChris Mason 3841c146afadSYan Zheng int btrfs_cleanup_fs_roots(struct btrfs_fs_info *fs_info) 3842c146afadSYan Zheng { 3843c146afadSYan Zheng u64 root_objectid = 0; 3844c146afadSYan Zheng struct btrfs_root *gang[8]; 384565d33fd7SQu Wenruo int i = 0; 384665d33fd7SQu Wenruo int err = 0; 384765d33fd7SQu Wenruo unsigned int ret = 0; 384865d33fd7SQu Wenruo int index; 3849c146afadSYan Zheng 3850c146afadSYan Zheng while (1) { 385165d33fd7SQu Wenruo index = srcu_read_lock(&fs_info->subvol_srcu); 3852c146afadSYan Zheng ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix, 3853c146afadSYan Zheng (void **)gang, root_objectid, 3854c146afadSYan Zheng ARRAY_SIZE(gang)); 385565d33fd7SQu Wenruo if (!ret) { 385665d33fd7SQu Wenruo srcu_read_unlock(&fs_info->subvol_srcu, index); 3857c146afadSYan Zheng break; 385865d33fd7SQu Wenruo } 38595d4f98a2SYan Zheng root_objectid = gang[ret - 1]->root_key.objectid + 1; 386066b4ffd1SJosef Bacik 386165d33fd7SQu Wenruo for (i = 0; i < ret; i++) { 386265d33fd7SQu Wenruo /* Avoid to grab roots in dead_roots */ 386365d33fd7SQu Wenruo if (btrfs_root_refs(&gang[i]->root_item) == 0) { 386465d33fd7SQu Wenruo gang[i] = NULL; 386565d33fd7SQu Wenruo continue; 386665d33fd7SQu Wenruo } 386765d33fd7SQu Wenruo /* grab all the search result for later use */ 386865d33fd7SQu Wenruo gang[i] = btrfs_grab_fs_root(gang[i]); 386965d33fd7SQu Wenruo } 387065d33fd7SQu Wenruo srcu_read_unlock(&fs_info->subvol_srcu, index); 387165d33fd7SQu Wenruo 387265d33fd7SQu Wenruo for (i = 0; i < ret; i++) { 387365d33fd7SQu Wenruo if (!gang[i]) 387465d33fd7SQu Wenruo continue; 3875c146afadSYan Zheng root_objectid = gang[i]->root_key.objectid; 387666b4ffd1SJosef Bacik err = btrfs_orphan_cleanup(gang[i]); 387766b4ffd1SJosef Bacik if (err) 387865d33fd7SQu Wenruo break; 387965d33fd7SQu Wenruo btrfs_put_fs_root(gang[i]); 3880c146afadSYan Zheng } 3881c146afadSYan Zheng root_objectid++; 3882c146afadSYan Zheng } 388365d33fd7SQu Wenruo 388465d33fd7SQu Wenruo /* release the uncleaned roots due to error */ 388565d33fd7SQu Wenruo for (; i < ret; i++) { 388665d33fd7SQu Wenruo if (gang[i]) 388765d33fd7SQu Wenruo btrfs_put_fs_root(gang[i]); 388865d33fd7SQu Wenruo } 388965d33fd7SQu Wenruo return err; 3890c146afadSYan Zheng } 3891c146afadSYan Zheng 38926bccf3abSJeff Mahoney int btrfs_commit_super(struct btrfs_fs_info *fs_info) 3893c146afadSYan Zheng { 38946bccf3abSJeff Mahoney struct btrfs_root *root = fs_info->tree_root; 3895c146afadSYan Zheng struct btrfs_trans_handle *trans; 3896c146afadSYan Zheng 38970b246afaSJeff Mahoney mutex_lock(&fs_info->cleaner_mutex); 389824bbcf04SYan, Zheng btrfs_run_delayed_iputs(root); 38990b246afaSJeff Mahoney mutex_unlock(&fs_info->cleaner_mutex); 39000b246afaSJeff Mahoney wake_up_process(fs_info->cleaner_kthread); 3901c71bf099SYan, Zheng 3902c71bf099SYan, Zheng /* wait until ongoing cleanup work done */ 39030b246afaSJeff Mahoney down_write(&fs_info->cleanup_work_sem); 39040b246afaSJeff Mahoney up_write(&fs_info->cleanup_work_sem); 3905c71bf099SYan, Zheng 39067a7eaa40SJosef Bacik trans = btrfs_join_transaction(root); 39073612b495STsutomu Itoh if (IS_ERR(trans)) 39083612b495STsutomu Itoh return PTR_ERR(trans); 3909d52c1bccSLiu Bo return btrfs_commit_transaction(trans, root); 3910c146afadSYan Zheng } 3911c146afadSYan Zheng 39126bccf3abSJeff Mahoney void close_ctree(struct btrfs_fs_info *fs_info) 3913eb60ceacSChris Mason { 39146bccf3abSJeff Mahoney struct btrfs_root *root = fs_info->tree_root; 3915c146afadSYan Zheng int ret; 3916e089f05cSChris Mason 3917afcdd129SJosef Bacik set_bit(BTRFS_FS_CLOSING_START, &fs_info->flags); 3918a2135011SChris Mason 39197343dd61SJustin Maggard /* wait for the qgroup rescan worker to stop */ 3920d06f23d6SJeff Mahoney btrfs_qgroup_wait_for_completion(fs_info, false); 39217343dd61SJustin Maggard 3922803b2f54SStefan Behrens /* wait for the uuid_scan task to finish */ 3923803b2f54SStefan Behrens down(&fs_info->uuid_tree_rescan_sem); 3924803b2f54SStefan Behrens /* avoid complains from lockdep et al., set sem back to initial state */ 3925803b2f54SStefan Behrens up(&fs_info->uuid_tree_rescan_sem); 3926803b2f54SStefan Behrens 3927837d5b6eSIlya Dryomov /* pause restriper - we want to resume on mount */ 3928aa1b8cd4SStefan Behrens btrfs_pause_balance(fs_info); 3929837d5b6eSIlya Dryomov 39308dabb742SStefan Behrens btrfs_dev_replace_suspend_for_unmount(fs_info); 39318dabb742SStefan Behrens 3932aa1b8cd4SStefan Behrens btrfs_scrub_cancel(fs_info); 39334cb5300bSChris Mason 39344cb5300bSChris Mason /* wait for any defraggers to finish */ 39354cb5300bSChris Mason wait_event(fs_info->transaction_wait, 39364cb5300bSChris Mason (atomic_read(&fs_info->defrag_running) == 0)); 39374cb5300bSChris Mason 39384cb5300bSChris Mason /* clear out the rbtree of defraggable inodes */ 393926176e7cSMiao Xie btrfs_cleanup_defrag_inodes(fs_info); 39404cb5300bSChris Mason 394121c7e756SMiao Xie cancel_work_sync(&fs_info->async_reclaim_work); 394221c7e756SMiao Xie 3943c146afadSYan Zheng if (!(fs_info->sb->s_flags & MS_RDONLY)) { 3944e44163e1SJeff Mahoney /* 3945e44163e1SJeff Mahoney * If the cleaner thread is stopped and there are 3946e44163e1SJeff Mahoney * block groups queued for removal, the deletion will be 3947e44163e1SJeff Mahoney * skipped when we quit the cleaner thread. 3948e44163e1SJeff Mahoney */ 39490b246afaSJeff Mahoney btrfs_delete_unused_bgs(fs_info); 3950e44163e1SJeff Mahoney 39516bccf3abSJeff Mahoney ret = btrfs_commit_super(fs_info); 3952d397712bSChris Mason if (ret) 395304892340SEric Sandeen btrfs_err(fs_info, "commit super ret %d", ret); 3954c146afadSYan Zheng } 3955ed2ff2cbSChris Mason 395687533c47SMiao Xie if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) 395768ce9682SStefan Behrens btrfs_error_commit_super(root); 3958acce952bSliubo 3959e3029d9fSAl Viro kthread_stop(fs_info->transaction_kthread); 3960e3029d9fSAl Viro kthread_stop(fs_info->cleaner_kthread); 39618929ecfaSYan, Zheng 3962afcdd129SJosef Bacik set_bit(BTRFS_FS_CLOSING_DONE, &fs_info->flags); 3963f25784b3SYan Zheng 396404892340SEric Sandeen btrfs_free_qgroup_config(fs_info); 3965bcef60f2SArne Jansen 3966963d678bSMiao Xie if (percpu_counter_sum(&fs_info->delalloc_bytes)) { 396704892340SEric Sandeen btrfs_info(fs_info, "at unmount delalloc count %lld", 3968963d678bSMiao Xie percpu_counter_sum(&fs_info->delalloc_bytes)); 3969b0c68f8bSChris Mason } 397031153d81SYan Zheng 39716618a59bSAnand Jain btrfs_sysfs_remove_mounted(fs_info); 3972b7c35e81SAnand Jain btrfs_sysfs_remove_fsid(fs_info->fs_devices); 39735ac1d209SJeff Mahoney 3974faa2dbf0SJosef Bacik btrfs_free_fs_roots(fs_info); 3975d10c5f31SChris Mason 39761a4319ccSLiu Bo btrfs_put_block_group_cache(fs_info); 39771a4319ccSLiu Bo 39782b1360daSJosef Bacik btrfs_free_block_groups(fs_info); 39792b1360daSJosef Bacik 3980de348ee0SWang Shilong /* 3981de348ee0SWang Shilong * we must make sure there is not any read request to 3982de348ee0SWang Shilong * submit after we stopping all workers. 3983de348ee0SWang Shilong */ 3984de348ee0SWang Shilong invalidate_inode_pages2(fs_info->btree_inode->i_mapping); 398596192499SJosef Bacik btrfs_stop_all_workers(fs_info); 398696192499SJosef Bacik 3987afcdd129SJosef Bacik clear_bit(BTRFS_FS_OPEN, &fs_info->flags); 398813e6c37bSJosef Bacik free_root_pointers(fs_info, 1); 39899ad6b7bcSChris Mason 399013e6c37bSJosef Bacik iput(fs_info->btree_inode); 3991d6bfde87SChris Mason 399221adbd5cSStefan Behrens #ifdef CONFIG_BTRFS_FS_CHECK_INTEGRITY 39930b246afaSJeff Mahoney if (btrfs_test_opt(fs_info, CHECK_INTEGRITY)) 399421adbd5cSStefan Behrens btrfsic_unmount(root, fs_info->fs_devices); 399521adbd5cSStefan Behrens #endif 399621adbd5cSStefan Behrens 3997dfe25020SChris Mason btrfs_close_devices(fs_info->fs_devices); 39980b86a832SChris Mason btrfs_mapping_tree_free(&fs_info->mapping_tree); 3999b248a415SChris Mason 4000e2d84521SMiao Xie percpu_counter_destroy(&fs_info->dirty_metadata_bytes); 4001963d678bSMiao Xie percpu_counter_destroy(&fs_info->delalloc_bytes); 4002c404e0dcSMiao Xie percpu_counter_destroy(&fs_info->bio_counter); 400304160088SChris Mason bdi_destroy(&fs_info->bdi); 400476dda93cSYan, Zheng cleanup_srcu_struct(&fs_info->subvol_srcu); 40050b86a832SChris Mason 400653b381b3SDavid Woodhouse btrfs_free_stripe_hash_table(fs_info); 400753b381b3SDavid Woodhouse 4008cdfb080eSChris Mason __btrfs_free_block_rsv(root->orphan_block_rsv); 40091cb048f5SFilipe David Borba Manana root->orphan_block_rsv = NULL; 401004216820SFilipe Manana 40110b246afaSJeff Mahoney lock_chunks(fs_info); 401204216820SFilipe Manana while (!list_empty(&fs_info->pinned_chunks)) { 401304216820SFilipe Manana struct extent_map *em; 401404216820SFilipe Manana 401504216820SFilipe Manana em = list_first_entry(&fs_info->pinned_chunks, 401604216820SFilipe Manana struct extent_map, list); 401704216820SFilipe Manana list_del_init(&em->list); 401804216820SFilipe Manana free_extent_map(em); 401904216820SFilipe Manana } 40200b246afaSJeff Mahoney unlock_chunks(fs_info); 4021eb60ceacSChris Mason } 4022eb60ceacSChris Mason 4023b9fab919SChris Mason int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid, 4024b9fab919SChris Mason int atomic) 4025ccd467d6SChris Mason { 40261259ab75SChris Mason int ret; 4027727011e0SChris Mason struct inode *btree_inode = buf->pages[0]->mapping->host; 40281259ab75SChris Mason 40290b32f4bbSJosef Bacik ret = extent_buffer_uptodate(buf); 40301259ab75SChris Mason if (!ret) 40311259ab75SChris Mason return ret; 40321259ab75SChris Mason 40331259ab75SChris Mason ret = verify_parent_transid(&BTRFS_I(btree_inode)->io_tree, buf, 4034b9fab919SChris Mason parent_transid, atomic); 4035b9fab919SChris Mason if (ret == -EAGAIN) 4036b9fab919SChris Mason return ret; 40371259ab75SChris Mason return !ret; 40385f39d397SChris Mason } 40396702ed49SChris Mason 40405f39d397SChris Mason void btrfs_mark_buffer_dirty(struct extent_buffer *buf) 40415f39d397SChris Mason { 40420b246afaSJeff Mahoney struct btrfs_fs_info *fs_info; 404306ea65a3SJosef Bacik struct btrfs_root *root; 40445f39d397SChris Mason u64 transid = btrfs_header_generation(buf); 4045b9473439SChris Mason int was_dirty; 4046b4ce94deSChris Mason 404706ea65a3SJosef Bacik #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS 404806ea65a3SJosef Bacik /* 404906ea65a3SJosef Bacik * This is a fast path so only do this check if we have sanity tests 405006ea65a3SJosef Bacik * enabled. Normal people shouldn't be marking dummy buffers as dirty 405106ea65a3SJosef Bacik * outside of the sanity tests. 405206ea65a3SJosef Bacik */ 405306ea65a3SJosef Bacik if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &buf->bflags))) 405406ea65a3SJosef Bacik return; 405506ea65a3SJosef Bacik #endif 405606ea65a3SJosef Bacik root = BTRFS_I(buf->pages[0]->mapping->host)->root; 40570b246afaSJeff Mahoney fs_info = root->fs_info; 4058b9447ef8SChris Mason btrfs_assert_tree_locked(buf); 40590b246afaSJeff Mahoney if (transid != fs_info->generation) 40605d163e0eSJeff Mahoney WARN(1, KERN_CRIT "btrfs transid mismatch buffer %llu, found %llu running %llu\n", 40610b246afaSJeff Mahoney buf->start, transid, fs_info->generation); 40620b32f4bbSJosef Bacik was_dirty = set_extent_buffer_dirty(buf); 4063e2d84521SMiao Xie if (!was_dirty) 40640b246afaSJeff Mahoney __percpu_counter_add(&fs_info->dirty_metadata_bytes, 4065e2d84521SMiao Xie buf->len, 40660b246afaSJeff Mahoney fs_info->dirty_metadata_batch); 40671f21ef0aSFilipe Manana #ifdef CONFIG_BTRFS_FS_CHECK_INTEGRITY 40681f21ef0aSFilipe Manana if (btrfs_header_level(buf) == 0 && check_leaf(root, buf)) { 40691f21ef0aSFilipe Manana btrfs_print_leaf(root, buf); 40701f21ef0aSFilipe Manana ASSERT(0); 40711f21ef0aSFilipe Manana } 40721f21ef0aSFilipe Manana #endif 4073eb60ceacSChris Mason } 4074eb60ceacSChris Mason 4075b53d3f5dSLiu Bo static void __btrfs_btree_balance_dirty(struct btrfs_root *root, 4076b53d3f5dSLiu Bo int flush_delayed) 407735b7e476SChris Mason { 40780b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4079188de649SChris Mason /* 4080188de649SChris Mason * looks as though older kernels can get into trouble with 4081188de649SChris Mason * this code, they end up stuck in balance_dirty_pages forever 4082188de649SChris Mason */ 4083e2d84521SMiao Xie int ret; 4084d6bfde87SChris Mason 40856933c02eSJens Axboe if (current->flags & PF_MEMALLOC) 4086d6bfde87SChris Mason return; 4087d6bfde87SChris Mason 4088b53d3f5dSLiu Bo if (flush_delayed) 408916cdcec7SMiao Xie btrfs_balance_delayed_items(root); 409016cdcec7SMiao Xie 40910b246afaSJeff Mahoney ret = percpu_counter_compare(&fs_info->dirty_metadata_bytes, 4092e2d84521SMiao Xie BTRFS_DIRTY_METADATA_THRESH); 4093e2d84521SMiao Xie if (ret > 0) { 40940b246afaSJeff Mahoney balance_dirty_pages_ratelimited(fs_info->btree_inode->i_mapping); 409516cdcec7SMiao Xie } 409616cdcec7SMiao Xie } 409716cdcec7SMiao Xie 4098b53d3f5dSLiu Bo void btrfs_btree_balance_dirty(struct btrfs_root *root) 409916cdcec7SMiao Xie { 4100b53d3f5dSLiu Bo __btrfs_btree_balance_dirty(root, 1); 410135b7e476SChris Mason } 4102b53d3f5dSLiu Bo 4103b53d3f5dSLiu Bo void btrfs_btree_balance_dirty_nodelay(struct btrfs_root *root) 4104b53d3f5dSLiu Bo { 4105b53d3f5dSLiu Bo __btrfs_btree_balance_dirty(root, 0); 4106d6bfde87SChris Mason } 41076b80053dSChris Mason 4108ca7a79adSChris Mason int btrfs_read_buffer(struct extent_buffer *buf, u64 parent_transid) 41096b80053dSChris Mason { 4110727011e0SChris Mason struct btrfs_root *root = BTRFS_I(buf->pages[0]->mapping->host)->root; 41118436ea91SJosef Bacik return btree_read_extent_buffer_pages(root, buf, parent_transid); 41126b80053dSChris Mason } 41130da5468fSChris Mason 4114fcd1f065SDavid Sterba static int btrfs_check_super_valid(struct btrfs_fs_info *fs_info, 4115acce952bSliubo int read_only) 4116acce952bSliubo { 4117c926093eSDavid Sterba struct btrfs_super_block *sb = fs_info->super_copy; 4118319e4d06SQu Wenruo u64 nodesize = btrfs_super_nodesize(sb); 4119319e4d06SQu Wenruo u64 sectorsize = btrfs_super_sectorsize(sb); 4120c926093eSDavid Sterba int ret = 0; 4121c926093eSDavid Sterba 4122319e4d06SQu Wenruo if (btrfs_super_magic(sb) != BTRFS_MAGIC) { 4123ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "no valid FS found"); 4124319e4d06SQu Wenruo ret = -EINVAL; 4125319e4d06SQu Wenruo } 4126319e4d06SQu Wenruo if (btrfs_super_flags(sb) & ~BTRFS_SUPER_FLAG_SUPP) 4127ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, "unrecognized super flag: %llu", 4128319e4d06SQu Wenruo btrfs_super_flags(sb) & ~BTRFS_SUPER_FLAG_SUPP); 412921e7626bSDavid Sterba if (btrfs_super_root_level(sb) >= BTRFS_MAX_LEVEL) { 4130ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "tree_root level too big: %d >= %d", 413121e7626bSDavid Sterba btrfs_super_root_level(sb), BTRFS_MAX_LEVEL); 4132c926093eSDavid Sterba ret = -EINVAL; 4133c926093eSDavid Sterba } 413421e7626bSDavid Sterba if (btrfs_super_chunk_root_level(sb) >= BTRFS_MAX_LEVEL) { 4135ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "chunk_root level too big: %d >= %d", 413621e7626bSDavid Sterba btrfs_super_chunk_root_level(sb), BTRFS_MAX_LEVEL); 4137c926093eSDavid Sterba ret = -EINVAL; 4138c926093eSDavid Sterba } 413921e7626bSDavid Sterba if (btrfs_super_log_root_level(sb) >= BTRFS_MAX_LEVEL) { 4140ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "log_root level too big: %d >= %d", 414121e7626bSDavid Sterba btrfs_super_log_root_level(sb), BTRFS_MAX_LEVEL); 4142c926093eSDavid Sterba ret = -EINVAL; 4143c926093eSDavid Sterba } 4144c926093eSDavid Sterba 41451104a885SDavid Sterba /* 4146319e4d06SQu Wenruo * Check sectorsize and nodesize first, other check will need it. 4147319e4d06SQu Wenruo * Check all possible sectorsize(4K, 8K, 16K, 32K, 64K) here. 41481104a885SDavid Sterba */ 4149319e4d06SQu Wenruo if (!is_power_of_2(sectorsize) || sectorsize < 4096 || 4150319e4d06SQu Wenruo sectorsize > BTRFS_MAX_METADATA_BLOCKSIZE) { 4151ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "invalid sectorsize %llu", sectorsize); 415275d6ad38SDavid Sterba ret = -EINVAL; 415375d6ad38SDavid Sterba } 4154319e4d06SQu Wenruo /* Only PAGE SIZE is supported yet */ 415509cbfeafSKirill A. Shutemov if (sectorsize != PAGE_SIZE) { 4156ab8d0fc4SJeff Mahoney btrfs_err(fs_info, 4157ab8d0fc4SJeff Mahoney "sectorsize %llu not supported yet, only support %lu", 415809cbfeafSKirill A. Shutemov sectorsize, PAGE_SIZE); 4159319e4d06SQu Wenruo ret = -EINVAL; 4160319e4d06SQu Wenruo } 4161319e4d06SQu Wenruo if (!is_power_of_2(nodesize) || nodesize < sectorsize || 4162319e4d06SQu Wenruo nodesize > BTRFS_MAX_METADATA_BLOCKSIZE) { 4163ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "invalid nodesize %llu", nodesize); 4164319e4d06SQu Wenruo ret = -EINVAL; 4165319e4d06SQu Wenruo } 4166319e4d06SQu Wenruo if (nodesize != le32_to_cpu(sb->__unused_leafsize)) { 4167ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "invalid leafsize %u, should be %llu", 4168ab8d0fc4SJeff Mahoney le32_to_cpu(sb->__unused_leafsize), nodesize); 4169319e4d06SQu Wenruo ret = -EINVAL; 4170319e4d06SQu Wenruo } 4171319e4d06SQu Wenruo 4172319e4d06SQu Wenruo /* Root alignment check */ 4173319e4d06SQu Wenruo if (!IS_ALIGNED(btrfs_super_root(sb), sectorsize)) { 4174ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, "tree_root block unaligned: %llu", 4175319e4d06SQu Wenruo btrfs_super_root(sb)); 4176319e4d06SQu Wenruo ret = -EINVAL; 4177319e4d06SQu Wenruo } 4178319e4d06SQu Wenruo if (!IS_ALIGNED(btrfs_super_chunk_root(sb), sectorsize)) { 4179ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, "chunk_root block unaligned: %llu", 4180319e4d06SQu Wenruo btrfs_super_chunk_root(sb)); 4181319e4d06SQu Wenruo ret = -EINVAL; 4182319e4d06SQu Wenruo } 4183319e4d06SQu Wenruo if (!IS_ALIGNED(btrfs_super_log_root(sb), sectorsize)) { 4184ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, "log_root block unaligned: %llu", 4185319e4d06SQu Wenruo btrfs_super_log_root(sb)); 418675d6ad38SDavid Sterba ret = -EINVAL; 418775d6ad38SDavid Sterba } 418875d6ad38SDavid Sterba 4189c926093eSDavid Sterba if (memcmp(fs_info->fsid, sb->dev_item.fsid, BTRFS_UUID_SIZE) != 0) { 4190ab8d0fc4SJeff Mahoney btrfs_err(fs_info, 4191ab8d0fc4SJeff Mahoney "dev_item UUID does not match fsid: %pU != %pU", 4192c926093eSDavid Sterba fs_info->fsid, sb->dev_item.fsid); 4193c926093eSDavid Sterba ret = -EINVAL; 4194c926093eSDavid Sterba } 4195c926093eSDavid Sterba 4196c926093eSDavid Sterba /* 4197c926093eSDavid Sterba * Hint to catch really bogus numbers, bitflips or so, more exact checks are 4198c926093eSDavid Sterba * done later 4199c926093eSDavid Sterba */ 420099e3ecfcSLiu Bo if (btrfs_super_bytes_used(sb) < 6 * btrfs_super_nodesize(sb)) { 420199e3ecfcSLiu Bo btrfs_err(fs_info, "bytes_used is too small %llu", 420299e3ecfcSLiu Bo btrfs_super_bytes_used(sb)); 420399e3ecfcSLiu Bo ret = -EINVAL; 420499e3ecfcSLiu Bo } 4205b7f67055SChandan Rajendra if (!is_power_of_2(btrfs_super_stripesize(sb))) { 420699e3ecfcSLiu Bo btrfs_err(fs_info, "invalid stripesize %u", 420799e3ecfcSLiu Bo btrfs_super_stripesize(sb)); 420899e3ecfcSLiu Bo ret = -EINVAL; 420999e3ecfcSLiu Bo } 421021e7626bSDavid Sterba if (btrfs_super_num_devices(sb) > (1UL << 31)) 4211ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, "suspicious number of devices: %llu", 421221e7626bSDavid Sterba btrfs_super_num_devices(sb)); 421375d6ad38SDavid Sterba if (btrfs_super_num_devices(sb) == 0) { 4214ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "number of devices is 0"); 421575d6ad38SDavid Sterba ret = -EINVAL; 421675d6ad38SDavid Sterba } 4217c926093eSDavid Sterba 421821e7626bSDavid Sterba if (btrfs_super_bytenr(sb) != BTRFS_SUPER_INFO_OFFSET) { 4219ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "super offset mismatch %llu != %u", 422021e7626bSDavid Sterba btrfs_super_bytenr(sb), BTRFS_SUPER_INFO_OFFSET); 4221c926093eSDavid Sterba ret = -EINVAL; 4222c926093eSDavid Sterba } 4223c926093eSDavid Sterba 4224c926093eSDavid Sterba /* 4225ce7fca5fSDavid Sterba * Obvious sys_chunk_array corruptions, it must hold at least one key 4226ce7fca5fSDavid Sterba * and one chunk 4227ce7fca5fSDavid Sterba */ 4228ce7fca5fSDavid Sterba if (btrfs_super_sys_array_size(sb) > BTRFS_SYSTEM_CHUNK_ARRAY_SIZE) { 4229ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "system chunk array too big %u > %u", 4230ce7fca5fSDavid Sterba btrfs_super_sys_array_size(sb), 4231ce7fca5fSDavid Sterba BTRFS_SYSTEM_CHUNK_ARRAY_SIZE); 4232ce7fca5fSDavid Sterba ret = -EINVAL; 4233ce7fca5fSDavid Sterba } 4234ce7fca5fSDavid Sterba if (btrfs_super_sys_array_size(sb) < sizeof(struct btrfs_disk_key) 4235ce7fca5fSDavid Sterba + sizeof(struct btrfs_chunk)) { 4236ab8d0fc4SJeff Mahoney btrfs_err(fs_info, "system chunk array too small %u < %zu", 4237ce7fca5fSDavid Sterba btrfs_super_sys_array_size(sb), 4238ce7fca5fSDavid Sterba sizeof(struct btrfs_disk_key) 4239ce7fca5fSDavid Sterba + sizeof(struct btrfs_chunk)); 4240ce7fca5fSDavid Sterba ret = -EINVAL; 4241ce7fca5fSDavid Sterba } 4242ce7fca5fSDavid Sterba 4243ce7fca5fSDavid Sterba /* 4244c926093eSDavid Sterba * The generation is a global counter, we'll trust it more than the others 4245c926093eSDavid Sterba * but it's still possible that it's the one that's wrong. 4246c926093eSDavid Sterba */ 424721e7626bSDavid Sterba if (btrfs_super_generation(sb) < btrfs_super_chunk_root_generation(sb)) 4248ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, 4249ab8d0fc4SJeff Mahoney "suspicious: generation < chunk_root_generation: %llu < %llu", 4250ab8d0fc4SJeff Mahoney btrfs_super_generation(sb), 4251ab8d0fc4SJeff Mahoney btrfs_super_chunk_root_generation(sb)); 425221e7626bSDavid Sterba if (btrfs_super_generation(sb) < btrfs_super_cache_generation(sb) 425321e7626bSDavid Sterba && btrfs_super_cache_generation(sb) != (u64)-1) 4254ab8d0fc4SJeff Mahoney btrfs_warn(fs_info, 4255ab8d0fc4SJeff Mahoney "suspicious: generation < cache_generation: %llu < %llu", 4256ab8d0fc4SJeff Mahoney btrfs_super_generation(sb), 4257ab8d0fc4SJeff Mahoney btrfs_super_cache_generation(sb)); 4258c926093eSDavid Sterba 4259c926093eSDavid Sterba return ret; 4260fcd1f065SDavid Sterba } 4261fcd1f065SDavid Sterba 426248a3b636SEric Sandeen static void btrfs_error_commit_super(struct btrfs_root *root) 4263acce952bSliubo { 42640b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4265acce952bSliubo 42660b246afaSJeff Mahoney mutex_lock(&fs_info->cleaner_mutex); 42670b246afaSJeff Mahoney btrfs_run_delayed_iputs(root); 42680b246afaSJeff Mahoney mutex_unlock(&fs_info->cleaner_mutex); 42690b246afaSJeff Mahoney 42700b246afaSJeff Mahoney down_write(&fs_info->cleanup_work_sem); 42710b246afaSJeff Mahoney up_write(&fs_info->cleanup_work_sem); 4272acce952bSliubo 4273acce952bSliubo /* cleanup FS via transaction */ 4274acce952bSliubo btrfs_cleanup_transaction(root); 4275acce952bSliubo } 4276acce952bSliubo 4277143bede5SJeff Mahoney static void btrfs_destroy_ordered_extents(struct btrfs_root *root) 4278acce952bSliubo { 4279acce952bSliubo struct btrfs_ordered_extent *ordered; 4280acce952bSliubo 4281199c2a9cSMiao Xie spin_lock(&root->ordered_extent_lock); 4282779880efSJosef Bacik /* 4283779880efSJosef Bacik * This will just short circuit the ordered completion stuff which will 4284779880efSJosef Bacik * make sure the ordered extent gets properly cleaned up. 4285779880efSJosef Bacik */ 4286199c2a9cSMiao Xie list_for_each_entry(ordered, &root->ordered_extents, 4287779880efSJosef Bacik root_extent_list) 4288779880efSJosef Bacik set_bit(BTRFS_ORDERED_IOERR, &ordered->flags); 4289199c2a9cSMiao Xie spin_unlock(&root->ordered_extent_lock); 4290199c2a9cSMiao Xie } 4291199c2a9cSMiao Xie 4292199c2a9cSMiao Xie static void btrfs_destroy_all_ordered_extents(struct btrfs_fs_info *fs_info) 4293199c2a9cSMiao Xie { 4294199c2a9cSMiao Xie struct btrfs_root *root; 4295199c2a9cSMiao Xie struct list_head splice; 4296199c2a9cSMiao Xie 4297199c2a9cSMiao Xie INIT_LIST_HEAD(&splice); 4298199c2a9cSMiao Xie 4299199c2a9cSMiao Xie spin_lock(&fs_info->ordered_root_lock); 4300199c2a9cSMiao Xie list_splice_init(&fs_info->ordered_roots, &splice); 4301199c2a9cSMiao Xie while (!list_empty(&splice)) { 4302199c2a9cSMiao Xie root = list_first_entry(&splice, struct btrfs_root, 4303199c2a9cSMiao Xie ordered_root); 43041de2cfdeSJosef Bacik list_move_tail(&root->ordered_root, 43051de2cfdeSJosef Bacik &fs_info->ordered_roots); 4306199c2a9cSMiao Xie 43072a85d9caSLiu Bo spin_unlock(&fs_info->ordered_root_lock); 4308199c2a9cSMiao Xie btrfs_destroy_ordered_extents(root); 4309199c2a9cSMiao Xie 43102a85d9caSLiu Bo cond_resched(); 43112a85d9caSLiu Bo spin_lock(&fs_info->ordered_root_lock); 4312199c2a9cSMiao Xie } 4313199c2a9cSMiao Xie spin_unlock(&fs_info->ordered_root_lock); 4314acce952bSliubo } 4315acce952bSliubo 431635a3621bSStefan Behrens static int btrfs_destroy_delayed_refs(struct btrfs_transaction *trans, 4317acce952bSliubo struct btrfs_root *root) 4318acce952bSliubo { 43190b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4320acce952bSliubo struct rb_node *node; 4321acce952bSliubo struct btrfs_delayed_ref_root *delayed_refs; 4322acce952bSliubo struct btrfs_delayed_ref_node *ref; 4323acce952bSliubo int ret = 0; 4324acce952bSliubo 4325acce952bSliubo delayed_refs = &trans->delayed_refs; 4326acce952bSliubo 4327acce952bSliubo spin_lock(&delayed_refs->lock); 4328d7df2c79SJosef Bacik if (atomic_read(&delayed_refs->num_entries) == 0) { 4329cfece4dbSDavid Sterba spin_unlock(&delayed_refs->lock); 43300b246afaSJeff Mahoney btrfs_info(fs_info, "delayed_refs has NO entry"); 4331acce952bSliubo return ret; 4332acce952bSliubo } 4333acce952bSliubo 4334d7df2c79SJosef Bacik while ((node = rb_first(&delayed_refs->href_root)) != NULL) { 4335d7df2c79SJosef Bacik struct btrfs_delayed_ref_head *head; 4336c6fc2454SQu Wenruo struct btrfs_delayed_ref_node *tmp; 4337e78417d1SJosef Bacik bool pin_bytes = false; 4338acce952bSliubo 4339d7df2c79SJosef Bacik head = rb_entry(node, struct btrfs_delayed_ref_head, 4340d7df2c79SJosef Bacik href_node); 4341b939d1abSJosef Bacik if (!mutex_trylock(&head->mutex)) { 4342d7df2c79SJosef Bacik atomic_inc(&head->node.refs); 434379787eaaSJeff Mahoney spin_unlock(&delayed_refs->lock); 4344b939d1abSJosef Bacik 4345acce952bSliubo mutex_lock(&head->mutex); 4346b939d1abSJosef Bacik mutex_unlock(&head->mutex); 4347d7df2c79SJosef Bacik btrfs_put_delayed_ref(&head->node); 4348e18fca73SJosef Bacik spin_lock(&delayed_refs->lock); 4349b939d1abSJosef Bacik continue; 4350b939d1abSJosef Bacik } 4351d7df2c79SJosef Bacik spin_lock(&head->lock); 4352c6fc2454SQu Wenruo list_for_each_entry_safe_reverse(ref, tmp, &head->ref_list, 4353c6fc2454SQu Wenruo list) { 4354d7df2c79SJosef Bacik ref->in_tree = 0; 4355c6fc2454SQu Wenruo list_del(&ref->list); 43561d57ee94SWang Xiaoguang if (!list_empty(&ref->add_list)) 43571d57ee94SWang Xiaoguang list_del(&ref->add_list); 4358d7df2c79SJosef Bacik atomic_dec(&delayed_refs->num_entries); 4359d7df2c79SJosef Bacik btrfs_put_delayed_ref(ref); 4360d7df2c79SJosef Bacik } 436154067ae9SJosef Bacik if (head->must_insert_reserved) 4362e78417d1SJosef Bacik pin_bytes = true; 436378a6184aSMiao Xie btrfs_free_delayed_extent_op(head->extent_op); 4364acce952bSliubo delayed_refs->num_heads--; 4365d7df2c79SJosef Bacik if (head->processing == 0) 4366acce952bSliubo delayed_refs->num_heads_ready--; 4367d7df2c79SJosef Bacik atomic_dec(&delayed_refs->num_entries); 4368d7df2c79SJosef Bacik head->node.in_tree = 0; 4369c46effa6SLiu Bo rb_erase(&head->href_node, &delayed_refs->href_root); 4370d7df2c79SJosef Bacik spin_unlock(&head->lock); 4371acce952bSliubo spin_unlock(&delayed_refs->lock); 4372e78417d1SJosef Bacik mutex_unlock(&head->mutex); 4373acce952bSliubo 4374d7df2c79SJosef Bacik if (pin_bytes) 4375d7df2c79SJosef Bacik btrfs_pin_extent(root, head->node.bytenr, 4376d7df2c79SJosef Bacik head->node.num_bytes, 1); 4377d7df2c79SJosef Bacik btrfs_put_delayed_ref(&head->node); 4378acce952bSliubo cond_resched(); 4379acce952bSliubo spin_lock(&delayed_refs->lock); 4380acce952bSliubo } 4381acce952bSliubo 4382acce952bSliubo spin_unlock(&delayed_refs->lock); 4383acce952bSliubo 4384acce952bSliubo return ret; 4385acce952bSliubo } 4386acce952bSliubo 4387143bede5SJeff Mahoney static void btrfs_destroy_delalloc_inodes(struct btrfs_root *root) 4388acce952bSliubo { 4389acce952bSliubo struct btrfs_inode *btrfs_inode; 4390acce952bSliubo struct list_head splice; 4391acce952bSliubo 4392acce952bSliubo INIT_LIST_HEAD(&splice); 4393acce952bSliubo 4394eb73c1b7SMiao Xie spin_lock(&root->delalloc_lock); 4395eb73c1b7SMiao Xie list_splice_init(&root->delalloc_inodes, &splice); 4396acce952bSliubo 4397acce952bSliubo while (!list_empty(&splice)) { 4398eb73c1b7SMiao Xie btrfs_inode = list_first_entry(&splice, struct btrfs_inode, 4399acce952bSliubo delalloc_inodes); 4400acce952bSliubo 4401acce952bSliubo list_del_init(&btrfs_inode->delalloc_inodes); 4402df0af1a5SMiao Xie clear_bit(BTRFS_INODE_IN_DELALLOC_LIST, 4403df0af1a5SMiao Xie &btrfs_inode->runtime_flags); 4404eb73c1b7SMiao Xie spin_unlock(&root->delalloc_lock); 4405acce952bSliubo 4406acce952bSliubo btrfs_invalidate_inodes(btrfs_inode->root); 4407b216cbfbSMiao Xie 4408eb73c1b7SMiao Xie spin_lock(&root->delalloc_lock); 4409acce952bSliubo } 4410acce952bSliubo 4411eb73c1b7SMiao Xie spin_unlock(&root->delalloc_lock); 4412eb73c1b7SMiao Xie } 4413eb73c1b7SMiao Xie 4414eb73c1b7SMiao Xie static void btrfs_destroy_all_delalloc_inodes(struct btrfs_fs_info *fs_info) 4415eb73c1b7SMiao Xie { 4416eb73c1b7SMiao Xie struct btrfs_root *root; 4417eb73c1b7SMiao Xie struct list_head splice; 4418eb73c1b7SMiao Xie 4419eb73c1b7SMiao Xie INIT_LIST_HEAD(&splice); 4420eb73c1b7SMiao Xie 4421eb73c1b7SMiao Xie spin_lock(&fs_info->delalloc_root_lock); 4422eb73c1b7SMiao Xie list_splice_init(&fs_info->delalloc_roots, &splice); 4423eb73c1b7SMiao Xie while (!list_empty(&splice)) { 4424eb73c1b7SMiao Xie root = list_first_entry(&splice, struct btrfs_root, 4425eb73c1b7SMiao Xie delalloc_root); 4426eb73c1b7SMiao Xie list_del_init(&root->delalloc_root); 4427eb73c1b7SMiao Xie root = btrfs_grab_fs_root(root); 4428eb73c1b7SMiao Xie BUG_ON(!root); 4429eb73c1b7SMiao Xie spin_unlock(&fs_info->delalloc_root_lock); 4430eb73c1b7SMiao Xie 4431eb73c1b7SMiao Xie btrfs_destroy_delalloc_inodes(root); 4432eb73c1b7SMiao Xie btrfs_put_fs_root(root); 4433eb73c1b7SMiao Xie 4434eb73c1b7SMiao Xie spin_lock(&fs_info->delalloc_root_lock); 4435eb73c1b7SMiao Xie } 4436eb73c1b7SMiao Xie spin_unlock(&fs_info->delalloc_root_lock); 4437acce952bSliubo } 4438acce952bSliubo 4439acce952bSliubo static int btrfs_destroy_marked_extents(struct btrfs_root *root, 4440acce952bSliubo struct extent_io_tree *dirty_pages, 4441acce952bSliubo int mark) 4442acce952bSliubo { 44430b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4444acce952bSliubo int ret; 4445acce952bSliubo struct extent_buffer *eb; 4446acce952bSliubo u64 start = 0; 4447acce952bSliubo u64 end; 4448acce952bSliubo 4449acce952bSliubo while (1) { 4450acce952bSliubo ret = find_first_extent_bit(dirty_pages, start, &start, &end, 4451e6138876SJosef Bacik mark, NULL); 4452acce952bSliubo if (ret) 4453acce952bSliubo break; 4454acce952bSliubo 445591166212SDavid Sterba clear_extent_bits(dirty_pages, start, end, mark); 4456acce952bSliubo while (start <= end) { 44570b246afaSJeff Mahoney eb = find_extent_buffer(fs_info, start); 44580b246afaSJeff Mahoney start += fs_info->nodesize; 4459fd8b2b61SJosef Bacik if (!eb) 4460acce952bSliubo continue; 4461fd8b2b61SJosef Bacik wait_on_extent_buffer_writeback(eb); 4462acce952bSliubo 4463fd8b2b61SJosef Bacik if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, 4464fd8b2b61SJosef Bacik &eb->bflags)) 4465fd8b2b61SJosef Bacik clear_extent_buffer_dirty(eb); 4466fd8b2b61SJosef Bacik free_extent_buffer_stale(eb); 4467acce952bSliubo } 4468acce952bSliubo } 4469acce952bSliubo 4470acce952bSliubo return ret; 4471acce952bSliubo } 4472acce952bSliubo 4473acce952bSliubo static int btrfs_destroy_pinned_extent(struct btrfs_root *root, 4474acce952bSliubo struct extent_io_tree *pinned_extents) 4475acce952bSliubo { 44760b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4477acce952bSliubo struct extent_io_tree *unpin; 4478acce952bSliubo u64 start; 4479acce952bSliubo u64 end; 4480acce952bSliubo int ret; 4481ed0eaa14SLiu Bo bool loop = true; 4482acce952bSliubo 4483acce952bSliubo unpin = pinned_extents; 4484ed0eaa14SLiu Bo again: 4485acce952bSliubo while (1) { 4486acce952bSliubo ret = find_first_extent_bit(unpin, 0, &start, &end, 4487e6138876SJosef Bacik EXTENT_DIRTY, NULL); 4488acce952bSliubo if (ret) 4489acce952bSliubo break; 4490acce952bSliubo 4491af6f8f60SDavid Sterba clear_extent_dirty(unpin, start, end); 4492acce952bSliubo btrfs_error_unpin_extent_range(root, start, end); 4493acce952bSliubo cond_resched(); 4494acce952bSliubo } 4495acce952bSliubo 4496ed0eaa14SLiu Bo if (loop) { 44970b246afaSJeff Mahoney if (unpin == &fs_info->freed_extents[0]) 44980b246afaSJeff Mahoney unpin = &fs_info->freed_extents[1]; 4499ed0eaa14SLiu Bo else 45000b246afaSJeff Mahoney unpin = &fs_info->freed_extents[0]; 4501ed0eaa14SLiu Bo loop = false; 4502ed0eaa14SLiu Bo goto again; 4503ed0eaa14SLiu Bo } 4504ed0eaa14SLiu Bo 4505acce952bSliubo return 0; 4506acce952bSliubo } 4507acce952bSliubo 4508c79a1751SLiu Bo static void btrfs_cleanup_bg_io(struct btrfs_block_group_cache *cache) 4509c79a1751SLiu Bo { 4510c79a1751SLiu Bo struct inode *inode; 4511c79a1751SLiu Bo 4512c79a1751SLiu Bo inode = cache->io_ctl.inode; 4513c79a1751SLiu Bo if (inode) { 4514c79a1751SLiu Bo invalidate_inode_pages2(inode->i_mapping); 4515c79a1751SLiu Bo BTRFS_I(inode)->generation = 0; 4516c79a1751SLiu Bo cache->io_ctl.inode = NULL; 4517c79a1751SLiu Bo iput(inode); 4518c79a1751SLiu Bo } 4519c79a1751SLiu Bo btrfs_put_block_group(cache); 4520c79a1751SLiu Bo } 4521c79a1751SLiu Bo 4522c79a1751SLiu Bo void btrfs_cleanup_dirty_bgs(struct btrfs_transaction *cur_trans, 4523c79a1751SLiu Bo struct btrfs_root *root) 4524c79a1751SLiu Bo { 45250b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4526c79a1751SLiu Bo struct btrfs_block_group_cache *cache; 4527c79a1751SLiu Bo 4528c79a1751SLiu Bo spin_lock(&cur_trans->dirty_bgs_lock); 4529c79a1751SLiu Bo while (!list_empty(&cur_trans->dirty_bgs)) { 4530c79a1751SLiu Bo cache = list_first_entry(&cur_trans->dirty_bgs, 4531c79a1751SLiu Bo struct btrfs_block_group_cache, 4532c79a1751SLiu Bo dirty_list); 4533c79a1751SLiu Bo if (!cache) { 45340b246afaSJeff Mahoney btrfs_err(fs_info, "orphan block group dirty_bgs list"); 4535c79a1751SLiu Bo spin_unlock(&cur_trans->dirty_bgs_lock); 4536c79a1751SLiu Bo return; 4537c79a1751SLiu Bo } 4538c79a1751SLiu Bo 4539c79a1751SLiu Bo if (!list_empty(&cache->io_list)) { 4540c79a1751SLiu Bo spin_unlock(&cur_trans->dirty_bgs_lock); 4541c79a1751SLiu Bo list_del_init(&cache->io_list); 4542c79a1751SLiu Bo btrfs_cleanup_bg_io(cache); 4543c79a1751SLiu Bo spin_lock(&cur_trans->dirty_bgs_lock); 4544c79a1751SLiu Bo } 4545c79a1751SLiu Bo 4546c79a1751SLiu Bo list_del_init(&cache->dirty_list); 4547c79a1751SLiu Bo spin_lock(&cache->lock); 4548c79a1751SLiu Bo cache->disk_cache_state = BTRFS_DC_ERROR; 4549c79a1751SLiu Bo spin_unlock(&cache->lock); 4550c79a1751SLiu Bo 4551c79a1751SLiu Bo spin_unlock(&cur_trans->dirty_bgs_lock); 4552c79a1751SLiu Bo btrfs_put_block_group(cache); 4553c79a1751SLiu Bo spin_lock(&cur_trans->dirty_bgs_lock); 4554c79a1751SLiu Bo } 4555c79a1751SLiu Bo spin_unlock(&cur_trans->dirty_bgs_lock); 4556c79a1751SLiu Bo 4557c79a1751SLiu Bo while (!list_empty(&cur_trans->io_bgs)) { 4558c79a1751SLiu Bo cache = list_first_entry(&cur_trans->io_bgs, 4559c79a1751SLiu Bo struct btrfs_block_group_cache, 4560c79a1751SLiu Bo io_list); 4561c79a1751SLiu Bo if (!cache) { 45620b246afaSJeff Mahoney btrfs_err(fs_info, "orphan block group on io_bgs list"); 4563c79a1751SLiu Bo return; 4564c79a1751SLiu Bo } 4565c79a1751SLiu Bo 4566c79a1751SLiu Bo list_del_init(&cache->io_list); 4567c79a1751SLiu Bo spin_lock(&cache->lock); 4568c79a1751SLiu Bo cache->disk_cache_state = BTRFS_DC_ERROR; 4569c79a1751SLiu Bo spin_unlock(&cache->lock); 4570c79a1751SLiu Bo btrfs_cleanup_bg_io(cache); 4571c79a1751SLiu Bo } 4572c79a1751SLiu Bo } 4573c79a1751SLiu Bo 457449b25e05SJeff Mahoney void btrfs_cleanup_one_transaction(struct btrfs_transaction *cur_trans, 457549b25e05SJeff Mahoney struct btrfs_root *root) 457649b25e05SJeff Mahoney { 45770b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4578c79a1751SLiu Bo btrfs_cleanup_dirty_bgs(cur_trans, root); 4579c79a1751SLiu Bo ASSERT(list_empty(&cur_trans->dirty_bgs)); 4580c79a1751SLiu Bo ASSERT(list_empty(&cur_trans->io_bgs)); 4581c79a1751SLiu Bo 458249b25e05SJeff Mahoney btrfs_destroy_delayed_refs(cur_trans, root); 458349b25e05SJeff Mahoney 45844a9d8bdeSMiao Xie cur_trans->state = TRANS_STATE_COMMIT_START; 45850b246afaSJeff Mahoney wake_up(&fs_info->transaction_blocked_wait); 458649b25e05SJeff Mahoney 45874a9d8bdeSMiao Xie cur_trans->state = TRANS_STATE_UNBLOCKED; 45880b246afaSJeff Mahoney wake_up(&fs_info->transaction_wait); 458949b25e05SJeff Mahoney 459067cde344SMiao Xie btrfs_destroy_delayed_inodes(root); 459167cde344SMiao Xie btrfs_assert_delayed_root_empty(root); 459267cde344SMiao Xie 459349b25e05SJeff Mahoney btrfs_destroy_marked_extents(root, &cur_trans->dirty_pages, 459449b25e05SJeff Mahoney EXTENT_DIRTY); 45956e841e32SLiu Bo btrfs_destroy_pinned_extent(root, 45960b246afaSJeff Mahoney fs_info->pinned_extents); 459749b25e05SJeff Mahoney 45984a9d8bdeSMiao Xie cur_trans->state =TRANS_STATE_COMPLETED; 45994a9d8bdeSMiao Xie wake_up(&cur_trans->commit_wait); 46004a9d8bdeSMiao Xie 460149b25e05SJeff Mahoney /* 460249b25e05SJeff Mahoney memset(cur_trans, 0, sizeof(*cur_trans)); 460349b25e05SJeff Mahoney kmem_cache_free(btrfs_transaction_cachep, cur_trans); 460449b25e05SJeff Mahoney */ 460549b25e05SJeff Mahoney } 460649b25e05SJeff Mahoney 460748a3b636SEric Sandeen static int btrfs_cleanup_transaction(struct btrfs_root *root) 4608acce952bSliubo { 46090b246afaSJeff Mahoney struct btrfs_fs_info *fs_info = root->fs_info; 4610acce952bSliubo struct btrfs_transaction *t; 4611acce952bSliubo 46120b246afaSJeff Mahoney mutex_lock(&fs_info->transaction_kthread_mutex); 4613acce952bSliubo 46140b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 46150b246afaSJeff Mahoney while (!list_empty(&fs_info->trans_list)) { 46160b246afaSJeff Mahoney t = list_first_entry(&fs_info->trans_list, 4617724e2315SJosef Bacik struct btrfs_transaction, list); 4618724e2315SJosef Bacik if (t->state >= TRANS_STATE_COMMIT_START) { 4619724e2315SJosef Bacik atomic_inc(&t->use_count); 46200b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 4621724e2315SJosef Bacik btrfs_wait_for_commit(root, t->transid); 4622724e2315SJosef Bacik btrfs_put_transaction(t); 46230b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 4624724e2315SJosef Bacik continue; 4625724e2315SJosef Bacik } 46260b246afaSJeff Mahoney if (t == fs_info->running_transaction) { 4627724e2315SJosef Bacik t->state = TRANS_STATE_COMMIT_DOING; 46280b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 4629724e2315SJosef Bacik /* 4630724e2315SJosef Bacik * We wait for 0 num_writers since we don't hold a trans 4631724e2315SJosef Bacik * handle open currently for this transaction. 4632724e2315SJosef Bacik */ 4633724e2315SJosef Bacik wait_event(t->writer_wait, 4634724e2315SJosef Bacik atomic_read(&t->num_writers) == 0); 4635724e2315SJosef Bacik } else { 46360b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 4637724e2315SJosef Bacik } 4638724e2315SJosef Bacik btrfs_cleanup_one_transaction(t, root); 4639724e2315SJosef Bacik 46400b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 46410b246afaSJeff Mahoney if (t == fs_info->running_transaction) 46420b246afaSJeff Mahoney fs_info->running_transaction = NULL; 4643724e2315SJosef Bacik list_del_init(&t->list); 46440b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 4645a4abeea4SJosef Bacik 4646724e2315SJosef Bacik btrfs_put_transaction(t); 4647724e2315SJosef Bacik trace_btrfs_transaction_commit(root); 46480b246afaSJeff Mahoney spin_lock(&fs_info->trans_lock); 4649724e2315SJosef Bacik } 46500b246afaSJeff Mahoney spin_unlock(&fs_info->trans_lock); 46510b246afaSJeff Mahoney btrfs_destroy_all_ordered_extents(fs_info); 465267cde344SMiao Xie btrfs_destroy_delayed_inodes(root); 465367cde344SMiao Xie btrfs_assert_delayed_root_empty(root); 46540b246afaSJeff Mahoney btrfs_destroy_pinned_extent(root, fs_info->pinned_extents); 46550b246afaSJeff Mahoney btrfs_destroy_all_delalloc_inodes(fs_info); 46560b246afaSJeff Mahoney mutex_unlock(&fs_info->transaction_kthread_mutex); 4657acce952bSliubo 4658acce952bSliubo return 0; 4659acce952bSliubo } 4660acce952bSliubo 4661e8c9f186SDavid Sterba static const struct extent_io_ops btree_extent_io_ops = { 4662ce9adaa5SChris Mason .readpage_end_io_hook = btree_readpage_end_io_hook, 46634bb31e92SArne Jansen .readpage_io_failed_hook = btree_io_failed_hook, 46640b86a832SChris Mason .submit_bio_hook = btree_submit_bio_hook, 4665239b14b3SChris Mason /* note we're sharing with inode.c for the merge bio hook */ 4666239b14b3SChris Mason .merge_bio_hook = btrfs_merge_bio_hook, 46670da5468fSChris Mason }; 4668