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> 28a74a4b97SChris Mason #include <linux/freezer.h> 29163e783eSDavid Woodhouse #include <linux/crc32c.h> 305a0e3ad6STejun Heo #include <linux/slab.h> 314b4e25f2SChris Mason #include "compat.h" 32eb60ceacSChris Mason #include "ctree.h" 33eb60ceacSChris Mason #include "disk-io.h" 34e089f05cSChris Mason #include "transaction.h" 350f7d52f4SChris Mason #include "btrfs_inode.h" 360b86a832SChris Mason #include "volumes.h" 37db94535dSChris Mason #include "print-tree.h" 388b712842SChris Mason #include "async-thread.h" 39925baeddSChris Mason #include "locking.h" 40e02119d5SChris Mason #include "tree-log.h" 41fa9c0d79SChris Mason #include "free-space-cache.h" 42eb60ceacSChris Mason 43d1310b2eSChris Mason static struct extent_io_ops btree_extent_io_ops; 448b712842SChris Mason static void end_workqueue_fn(struct btrfs_work *work); 454df27c4dSYan, Zheng static void free_fs_root(struct btrfs_root *root); 46ce9adaa5SChris Mason 47d352ac68SChris Mason /* 48d352ac68SChris Mason * end_io_wq structs are used to do processing in task context when an IO is 49d352ac68SChris Mason * complete. This is used during reads to verify checksums, and it is used 50d352ac68SChris Mason * by writes to insert metadata for new file extents after IO is complete. 51d352ac68SChris Mason */ 52ce9adaa5SChris Mason struct end_io_wq { 53ce9adaa5SChris Mason struct bio *bio; 54ce9adaa5SChris Mason bio_end_io_t *end_io; 55ce9adaa5SChris Mason void *private; 56ce9adaa5SChris Mason struct btrfs_fs_info *info; 57ce9adaa5SChris Mason int error; 5822c59948SChris Mason int metadata; 59ce9adaa5SChris Mason struct list_head list; 608b712842SChris Mason struct btrfs_work work; 61ce9adaa5SChris Mason }; 620da5468fSChris Mason 63d352ac68SChris Mason /* 64d352ac68SChris Mason * async submit bios are used to offload expensive checksumming 65d352ac68SChris Mason * onto the worker threads. They checksum file and metadata bios 66d352ac68SChris Mason * just before they are sent down the IO stack. 67d352ac68SChris Mason */ 6844b8bd7eSChris Mason struct async_submit_bio { 6944b8bd7eSChris Mason struct inode *inode; 7044b8bd7eSChris Mason struct bio *bio; 7144b8bd7eSChris Mason struct list_head list; 724a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_start; 734a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_done; 7444b8bd7eSChris Mason int rw; 7544b8bd7eSChris Mason int mirror_num; 76c8b97818SChris Mason unsigned long bio_flags; 778b712842SChris Mason struct btrfs_work work; 7844b8bd7eSChris Mason }; 7944b8bd7eSChris Mason 804008c04aSChris Mason /* These are used to set the lockdep class on the extent buffer locks. 814008c04aSChris Mason * The class is set by the readpage_end_io_hook after the buffer has 824008c04aSChris Mason * passed csum validation but before the pages are unlocked. 834008c04aSChris Mason * 844008c04aSChris Mason * The lockdep class is also set by btrfs_init_new_buffer on freshly 854008c04aSChris Mason * allocated blocks. 864008c04aSChris Mason * 874008c04aSChris Mason * The class is based on the level in the tree block, which allows lockdep 884008c04aSChris Mason * to know that lower nodes nest inside the locks of higher nodes. 894008c04aSChris Mason * 904008c04aSChris Mason * We also add a check to make sure the highest level of the tree is 914008c04aSChris Mason * the same as our lockdep setup here. If BTRFS_MAX_LEVEL changes, this 924008c04aSChris Mason * code needs update as well. 934008c04aSChris Mason */ 944008c04aSChris Mason #ifdef CONFIG_DEBUG_LOCK_ALLOC 954008c04aSChris Mason # if BTRFS_MAX_LEVEL != 8 964008c04aSChris Mason # error 974008c04aSChris Mason # endif 984008c04aSChris Mason static struct lock_class_key btrfs_eb_class[BTRFS_MAX_LEVEL + 1]; 994008c04aSChris Mason static const char *btrfs_eb_name[BTRFS_MAX_LEVEL + 1] = { 1004008c04aSChris Mason /* leaf */ 1014008c04aSChris Mason "btrfs-extent-00", 1024008c04aSChris Mason "btrfs-extent-01", 1034008c04aSChris Mason "btrfs-extent-02", 1044008c04aSChris Mason "btrfs-extent-03", 1054008c04aSChris Mason "btrfs-extent-04", 1064008c04aSChris Mason "btrfs-extent-05", 1074008c04aSChris Mason "btrfs-extent-06", 1084008c04aSChris Mason "btrfs-extent-07", 1094008c04aSChris Mason /* highest possible level */ 1104008c04aSChris Mason "btrfs-extent-08", 1114008c04aSChris Mason }; 1124008c04aSChris Mason #endif 1134008c04aSChris Mason 114d352ac68SChris Mason /* 115d352ac68SChris Mason * extents on the btree inode are pretty simple, there's one extent 116d352ac68SChris Mason * that covers the entire device 117d352ac68SChris Mason */ 118b2950863SChristoph Hellwig static struct extent_map *btree_get_extent(struct inode *inode, 119b2950863SChristoph Hellwig struct page *page, size_t page_offset, u64 start, u64 len, 1205f39d397SChris Mason int create) 1215f39d397SChris Mason { 1225f39d397SChris Mason struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 1235f39d397SChris Mason struct extent_map *em; 1245f39d397SChris Mason int ret; 1255f39d397SChris Mason 126890871beSChris Mason read_lock(&em_tree->lock); 127d1310b2eSChris Mason em = lookup_extent_mapping(em_tree, start, len); 128a061fc8dSChris Mason if (em) { 129a061fc8dSChris Mason em->bdev = 130a061fc8dSChris Mason BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev; 131890871beSChris Mason read_unlock(&em_tree->lock); 1325f39d397SChris Mason goto out; 133a061fc8dSChris Mason } 134890871beSChris Mason read_unlock(&em_tree->lock); 1357b13b7b1SChris Mason 1365f39d397SChris Mason em = alloc_extent_map(GFP_NOFS); 1375f39d397SChris Mason if (!em) { 1385f39d397SChris Mason em = ERR_PTR(-ENOMEM); 1395f39d397SChris Mason goto out; 1405f39d397SChris Mason } 1415f39d397SChris Mason em->start = 0; 1420afbaf8cSChris Mason em->len = (u64)-1; 143c8b97818SChris Mason em->block_len = (u64)-1; 1445f39d397SChris Mason em->block_start = 0; 145a061fc8dSChris Mason em->bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev; 146d1310b2eSChris Mason 147890871beSChris Mason write_lock(&em_tree->lock); 1485f39d397SChris Mason ret = add_extent_mapping(em_tree, em); 1495f39d397SChris Mason if (ret == -EEXIST) { 1500afbaf8cSChris Mason u64 failed_start = em->start; 1510afbaf8cSChris Mason u64 failed_len = em->len; 1520afbaf8cSChris Mason 1535f39d397SChris Mason free_extent_map(em); 1547b13b7b1SChris Mason em = lookup_extent_mapping(em_tree, start, len); 1550afbaf8cSChris Mason if (em) { 1567b13b7b1SChris Mason ret = 0; 1570afbaf8cSChris Mason } else { 1580afbaf8cSChris Mason em = lookup_extent_mapping(em_tree, failed_start, 1590afbaf8cSChris Mason failed_len); 1607b13b7b1SChris Mason ret = -EIO; 1610afbaf8cSChris Mason } 1625f39d397SChris Mason } else if (ret) { 1637b13b7b1SChris Mason free_extent_map(em); 1647b13b7b1SChris Mason em = NULL; 1655f39d397SChris Mason } 166890871beSChris Mason write_unlock(&em_tree->lock); 1677b13b7b1SChris Mason 1687b13b7b1SChris Mason if (ret) 1697b13b7b1SChris Mason em = ERR_PTR(ret); 1705f39d397SChris Mason out: 1715f39d397SChris Mason return em; 1725f39d397SChris Mason } 1735f39d397SChris Mason 17419c00ddcSChris Mason u32 btrfs_csum_data(struct btrfs_root *root, char *data, u32 seed, size_t len) 17519c00ddcSChris Mason { 176163e783eSDavid Woodhouse return crc32c(seed, data, len); 17719c00ddcSChris Mason } 17819c00ddcSChris Mason 17919c00ddcSChris Mason void btrfs_csum_final(u32 crc, char *result) 18019c00ddcSChris Mason { 18119c00ddcSChris Mason *(__le32 *)result = ~cpu_to_le32(crc); 18219c00ddcSChris Mason } 18319c00ddcSChris Mason 184d352ac68SChris Mason /* 185d352ac68SChris Mason * compute the csum for a btree block, and either verify it or write it 186d352ac68SChris Mason * into the csum field of the block. 187d352ac68SChris Mason */ 18819c00ddcSChris Mason static int csum_tree_block(struct btrfs_root *root, struct extent_buffer *buf, 18919c00ddcSChris Mason int verify) 19019c00ddcSChris Mason { 191607d432dSJosef Bacik u16 csum_size = 192607d432dSJosef Bacik btrfs_super_csum_size(&root->fs_info->super_copy); 193607d432dSJosef Bacik char *result = NULL; 19419c00ddcSChris Mason unsigned long len; 19519c00ddcSChris Mason unsigned long cur_len; 19619c00ddcSChris Mason unsigned long offset = BTRFS_CSUM_SIZE; 19719c00ddcSChris Mason char *map_token = NULL; 19819c00ddcSChris Mason char *kaddr; 19919c00ddcSChris Mason unsigned long map_start; 20019c00ddcSChris Mason unsigned long map_len; 20119c00ddcSChris Mason int err; 20219c00ddcSChris Mason u32 crc = ~(u32)0; 203607d432dSJosef Bacik unsigned long inline_result; 20419c00ddcSChris Mason 20519c00ddcSChris Mason len = buf->len - offset; 20619c00ddcSChris Mason while (len > 0) { 20719c00ddcSChris Mason err = map_private_extent_buffer(buf, offset, 32, 20819c00ddcSChris Mason &map_token, &kaddr, 20919c00ddcSChris Mason &map_start, &map_len, KM_USER0); 210d397712bSChris Mason if (err) 21119c00ddcSChris Mason return 1; 21219c00ddcSChris Mason cur_len = min(len, map_len - (offset - map_start)); 21319c00ddcSChris Mason crc = btrfs_csum_data(root, kaddr + offset - map_start, 21419c00ddcSChris Mason crc, cur_len); 21519c00ddcSChris Mason len -= cur_len; 21619c00ddcSChris Mason offset += cur_len; 21719c00ddcSChris Mason unmap_extent_buffer(buf, map_token, KM_USER0); 21819c00ddcSChris Mason } 219607d432dSJosef Bacik if (csum_size > sizeof(inline_result)) { 220607d432dSJosef Bacik result = kzalloc(csum_size * sizeof(char), GFP_NOFS); 221607d432dSJosef Bacik if (!result) 222607d432dSJosef Bacik return 1; 223607d432dSJosef Bacik } else { 224607d432dSJosef Bacik result = (char *)&inline_result; 225607d432dSJosef Bacik } 226607d432dSJosef Bacik 22719c00ddcSChris Mason btrfs_csum_final(crc, result); 22819c00ddcSChris Mason 22919c00ddcSChris Mason if (verify) { 230607d432dSJosef Bacik if (memcmp_extent_buffer(buf, result, 0, csum_size)) { 231e4204dedSChris Mason u32 val; 232e4204dedSChris Mason u32 found = 0; 233607d432dSJosef Bacik memcpy(&found, result, csum_size); 234e4204dedSChris Mason 235607d432dSJosef Bacik read_extent_buffer(buf, &val, 0, csum_size); 236193f284dSChris Mason if (printk_ratelimit()) { 237193f284dSChris Mason printk(KERN_INFO "btrfs: %s checksum verify " 238193f284dSChris Mason "failed on %llu wanted %X found %X " 239193f284dSChris Mason "level %d\n", 24019c00ddcSChris Mason root->fs_info->sb->s_id, 241193f284dSChris Mason (unsigned long long)buf->start, val, found, 242193f284dSChris Mason btrfs_header_level(buf)); 243193f284dSChris Mason } 244607d432dSJosef Bacik if (result != (char *)&inline_result) 245607d432dSJosef Bacik kfree(result); 24619c00ddcSChris Mason return 1; 24719c00ddcSChris Mason } 24819c00ddcSChris Mason } else { 249607d432dSJosef Bacik write_extent_buffer(buf, result, 0, csum_size); 25019c00ddcSChris Mason } 251607d432dSJosef Bacik if (result != (char *)&inline_result) 252607d432dSJosef Bacik kfree(result); 25319c00ddcSChris Mason return 0; 25419c00ddcSChris Mason } 25519c00ddcSChris Mason 256d352ac68SChris Mason /* 257d352ac68SChris Mason * we can't consider a given block up to date unless the transid of the 258d352ac68SChris Mason * block matches the transid in the parent node's pointer. This is how we 259d352ac68SChris Mason * detect blocks that either didn't get written at all or got written 260d352ac68SChris Mason * in the wrong place. 261d352ac68SChris Mason */ 2621259ab75SChris Mason static int verify_parent_transid(struct extent_io_tree *io_tree, 2631259ab75SChris Mason struct extent_buffer *eb, u64 parent_transid) 2641259ab75SChris Mason { 2652ac55d41SJosef Bacik struct extent_state *cached_state = NULL; 2661259ab75SChris Mason int ret; 2671259ab75SChris Mason 2681259ab75SChris Mason if (!parent_transid || btrfs_header_generation(eb) == parent_transid) 2691259ab75SChris Mason return 0; 2701259ab75SChris Mason 2712ac55d41SJosef Bacik lock_extent_bits(io_tree, eb->start, eb->start + eb->len - 1, 2722ac55d41SJosef Bacik 0, &cached_state, GFP_NOFS); 2732ac55d41SJosef Bacik if (extent_buffer_uptodate(io_tree, eb, cached_state) && 2741259ab75SChris Mason btrfs_header_generation(eb) == parent_transid) { 2751259ab75SChris Mason ret = 0; 2761259ab75SChris Mason goto out; 2771259ab75SChris Mason } 278193f284dSChris Mason if (printk_ratelimit()) { 279193f284dSChris Mason printk("parent transid verify failed on %llu wanted %llu " 280193f284dSChris Mason "found %llu\n", 2811259ab75SChris Mason (unsigned long long)eb->start, 2821259ab75SChris Mason (unsigned long long)parent_transid, 2831259ab75SChris Mason (unsigned long long)btrfs_header_generation(eb)); 284193f284dSChris Mason } 2851259ab75SChris Mason ret = 1; 2862ac55d41SJosef Bacik clear_extent_buffer_uptodate(io_tree, eb, &cached_state); 28733958dc6SChris Mason out: 2882ac55d41SJosef Bacik unlock_extent_cached(io_tree, eb->start, eb->start + eb->len - 1, 2892ac55d41SJosef Bacik &cached_state, GFP_NOFS); 2901259ab75SChris Mason return ret; 2911259ab75SChris Mason } 2921259ab75SChris Mason 293d352ac68SChris Mason /* 294d352ac68SChris Mason * helper to read a given tree block, doing retries as required when 295d352ac68SChris Mason * the checksums don't match and we have alternate mirrors to try. 296d352ac68SChris Mason */ 297f188591eSChris Mason static int btree_read_extent_buffer_pages(struct btrfs_root *root, 298f188591eSChris Mason struct extent_buffer *eb, 299ca7a79adSChris Mason u64 start, u64 parent_transid) 300f188591eSChris Mason { 301f188591eSChris Mason struct extent_io_tree *io_tree; 302f188591eSChris Mason int ret; 303f188591eSChris Mason int num_copies = 0; 304f188591eSChris Mason int mirror_num = 0; 305f188591eSChris Mason 306f188591eSChris Mason io_tree = &BTRFS_I(root->fs_info->btree_inode)->io_tree; 307f188591eSChris Mason while (1) { 308f188591eSChris Mason ret = read_extent_buffer_pages(io_tree, eb, start, 1, 309f188591eSChris Mason btree_get_extent, mirror_num); 3101259ab75SChris Mason if (!ret && 3111259ab75SChris Mason !verify_parent_transid(io_tree, eb, parent_transid)) 312f188591eSChris Mason return ret; 313d397712bSChris Mason 314f188591eSChris Mason num_copies = btrfs_num_copies(&root->fs_info->mapping_tree, 315f188591eSChris Mason eb->start, eb->len); 3164235298eSChris Mason if (num_copies == 1) 317f188591eSChris Mason return ret; 3184235298eSChris Mason 319f188591eSChris Mason mirror_num++; 3204235298eSChris Mason if (mirror_num > num_copies) 321f188591eSChris Mason return ret; 322f188591eSChris Mason } 323f188591eSChris Mason return -EIO; 324f188591eSChris Mason } 32519c00ddcSChris Mason 326d352ac68SChris Mason /* 327d397712bSChris Mason * checksum a dirty tree block before IO. This has extra checks to make sure 328d397712bSChris Mason * we only fill in the checksum field in the first page of a multi-page block 329d352ac68SChris Mason */ 330d397712bSChris Mason 331b2950863SChristoph Hellwig static int csum_dirty_buffer(struct btrfs_root *root, struct page *page) 33219c00ddcSChris Mason { 333d1310b2eSChris Mason struct extent_io_tree *tree; 33435ebb934SChris Mason u64 start = (u64)page->index << PAGE_CACHE_SHIFT; 33519c00ddcSChris Mason u64 found_start; 33619c00ddcSChris Mason int found_level; 33719c00ddcSChris Mason unsigned long len; 33819c00ddcSChris Mason struct extent_buffer *eb; 339f188591eSChris Mason int ret; 340f188591eSChris Mason 341d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 34219c00ddcSChris Mason 34319c00ddcSChris Mason if (page->private == EXTENT_PAGE_PRIVATE) 34419c00ddcSChris Mason goto out; 34519c00ddcSChris Mason if (!page->private) 34619c00ddcSChris Mason goto out; 34719c00ddcSChris Mason len = page->private >> 2; 348d397712bSChris Mason WARN_ON(len == 0); 349d397712bSChris Mason 35019c00ddcSChris Mason eb = alloc_extent_buffer(tree, start, len, page, GFP_NOFS); 351ca7a79adSChris Mason ret = btree_read_extent_buffer_pages(root, eb, start + PAGE_CACHE_SIZE, 352ca7a79adSChris Mason btrfs_header_generation(eb)); 353f188591eSChris Mason BUG_ON(ret); 35419c00ddcSChris Mason found_start = btrfs_header_bytenr(eb); 35519c00ddcSChris Mason if (found_start != start) { 35655c69072SChris Mason WARN_ON(1); 35755c69072SChris Mason goto err; 35855c69072SChris Mason } 35955c69072SChris Mason if (eb->first_page != page) { 36055c69072SChris Mason WARN_ON(1); 36155c69072SChris Mason goto err; 36255c69072SChris Mason } 36355c69072SChris Mason if (!PageUptodate(page)) { 36455c69072SChris Mason WARN_ON(1); 36555c69072SChris Mason goto err; 36619c00ddcSChris Mason } 36719c00ddcSChris Mason found_level = btrfs_header_level(eb); 3684bef0848SChris Mason 36919c00ddcSChris Mason csum_tree_block(root, eb, 0); 37055c69072SChris Mason err: 37119c00ddcSChris Mason free_extent_buffer(eb); 37219c00ddcSChris Mason out: 37319c00ddcSChris Mason return 0; 37419c00ddcSChris Mason } 37519c00ddcSChris Mason 3762b82032cSYan Zheng static int check_tree_block_fsid(struct btrfs_root *root, 3772b82032cSYan Zheng struct extent_buffer *eb) 3782b82032cSYan Zheng { 3792b82032cSYan Zheng struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices; 3802b82032cSYan Zheng u8 fsid[BTRFS_UUID_SIZE]; 3812b82032cSYan Zheng int ret = 1; 3822b82032cSYan Zheng 3832b82032cSYan Zheng read_extent_buffer(eb, fsid, (unsigned long)btrfs_header_fsid(eb), 3842b82032cSYan Zheng BTRFS_FSID_SIZE); 3852b82032cSYan Zheng while (fs_devices) { 3862b82032cSYan Zheng if (!memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE)) { 3872b82032cSYan Zheng ret = 0; 3882b82032cSYan Zheng break; 3892b82032cSYan Zheng } 3902b82032cSYan Zheng fs_devices = fs_devices->seed; 3912b82032cSYan Zheng } 3922b82032cSYan Zheng return ret; 3932b82032cSYan Zheng } 3942b82032cSYan Zheng 3954008c04aSChris Mason #ifdef CONFIG_DEBUG_LOCK_ALLOC 3964008c04aSChris Mason void btrfs_set_buffer_lockdep_class(struct extent_buffer *eb, int level) 3974008c04aSChris Mason { 3984008c04aSChris Mason lockdep_set_class_and_name(&eb->lock, 3994008c04aSChris Mason &btrfs_eb_class[level], 4004008c04aSChris Mason btrfs_eb_name[level]); 4014008c04aSChris Mason } 4024008c04aSChris Mason #endif 4034008c04aSChris Mason 404b2950863SChristoph Hellwig static int btree_readpage_end_io_hook(struct page *page, u64 start, u64 end, 405ce9adaa5SChris Mason struct extent_state *state) 406ce9adaa5SChris Mason { 407ce9adaa5SChris Mason struct extent_io_tree *tree; 408ce9adaa5SChris Mason u64 found_start; 409ce9adaa5SChris Mason int found_level; 410ce9adaa5SChris Mason unsigned long len; 411ce9adaa5SChris Mason struct extent_buffer *eb; 412ce9adaa5SChris Mason struct btrfs_root *root = BTRFS_I(page->mapping->host)->root; 413f188591eSChris Mason int ret = 0; 414ce9adaa5SChris Mason 415ce9adaa5SChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 416ce9adaa5SChris Mason if (page->private == EXTENT_PAGE_PRIVATE) 417ce9adaa5SChris Mason goto out; 418ce9adaa5SChris Mason if (!page->private) 419ce9adaa5SChris Mason goto out; 420d397712bSChris Mason 421ce9adaa5SChris Mason len = page->private >> 2; 422d397712bSChris Mason WARN_ON(len == 0); 423d397712bSChris Mason 424ce9adaa5SChris Mason eb = alloc_extent_buffer(tree, start, len, page, GFP_NOFS); 425f188591eSChris Mason 426ce9adaa5SChris Mason found_start = btrfs_header_bytenr(eb); 42723a07867SChris Mason if (found_start != start) { 428193f284dSChris Mason if (printk_ratelimit()) { 429193f284dSChris Mason printk(KERN_INFO "btrfs bad tree block start " 430193f284dSChris Mason "%llu %llu\n", 431a1b32a59SChris Mason (unsigned long long)found_start, 432a1b32a59SChris Mason (unsigned long long)eb->start); 433193f284dSChris Mason } 434f188591eSChris Mason ret = -EIO; 435ce9adaa5SChris Mason goto err; 436ce9adaa5SChris Mason } 437ce9adaa5SChris Mason if (eb->first_page != page) { 438d397712bSChris Mason printk(KERN_INFO "btrfs bad first page %lu %lu\n", 439d397712bSChris Mason eb->first_page->index, page->index); 440ce9adaa5SChris Mason WARN_ON(1); 441f188591eSChris Mason ret = -EIO; 442ce9adaa5SChris Mason goto err; 443ce9adaa5SChris Mason } 4442b82032cSYan Zheng if (check_tree_block_fsid(root, eb)) { 445193f284dSChris Mason if (printk_ratelimit()) { 446d397712bSChris Mason printk(KERN_INFO "btrfs bad fsid on block %llu\n", 447d397712bSChris Mason (unsigned long long)eb->start); 448193f284dSChris Mason } 4491259ab75SChris Mason ret = -EIO; 4501259ab75SChris Mason goto err; 4511259ab75SChris Mason } 452ce9adaa5SChris Mason found_level = btrfs_header_level(eb); 453ce9adaa5SChris Mason 4544008c04aSChris Mason btrfs_set_buffer_lockdep_class(eb, found_level); 4554008c04aSChris Mason 456ce9adaa5SChris Mason ret = csum_tree_block(root, eb, 1); 457f188591eSChris Mason if (ret) 458f188591eSChris Mason ret = -EIO; 459ce9adaa5SChris Mason 460ce9adaa5SChris Mason end = min_t(u64, eb->len, PAGE_CACHE_SIZE); 461ce9adaa5SChris Mason end = eb->start + end - 1; 462ce9adaa5SChris Mason err: 463ce9adaa5SChris Mason free_extent_buffer(eb); 464ce9adaa5SChris Mason out: 465f188591eSChris Mason return ret; 466ce9adaa5SChris Mason } 467ce9adaa5SChris Mason 468ce9adaa5SChris Mason static void end_workqueue_bio(struct bio *bio, int err) 469ce9adaa5SChris Mason { 470ce9adaa5SChris Mason struct end_io_wq *end_io_wq = bio->bi_private; 471ce9adaa5SChris Mason struct btrfs_fs_info *fs_info; 472ce9adaa5SChris Mason 473ce9adaa5SChris Mason fs_info = end_io_wq->info; 474ce9adaa5SChris Mason end_io_wq->error = err; 4758b712842SChris Mason end_io_wq->work.func = end_workqueue_fn; 4768b712842SChris Mason end_io_wq->work.flags = 0; 477d20f7043SChris Mason 478d20f7043SChris Mason if (bio->bi_rw & (1 << BIO_RW)) { 479cad321adSChris Mason if (end_io_wq->metadata) 480cad321adSChris Mason btrfs_queue_worker(&fs_info->endio_meta_write_workers, 481cad321adSChris Mason &end_io_wq->work); 482cad321adSChris Mason else 483e6dcd2dcSChris Mason btrfs_queue_worker(&fs_info->endio_write_workers, 484e6dcd2dcSChris Mason &end_io_wq->work); 485d20f7043SChris Mason } else { 486d20f7043SChris Mason if (end_io_wq->metadata) 487d20f7043SChris Mason btrfs_queue_worker(&fs_info->endio_meta_workers, 488d20f7043SChris Mason &end_io_wq->work); 489e6dcd2dcSChris Mason else 490d20f7043SChris Mason btrfs_queue_worker(&fs_info->endio_workers, 491d20f7043SChris Mason &end_io_wq->work); 492d20f7043SChris Mason } 493ce9adaa5SChris Mason } 494ce9adaa5SChris Mason 49522c59948SChris Mason int btrfs_bio_wq_end_io(struct btrfs_fs_info *info, struct bio *bio, 49622c59948SChris Mason int metadata) 4970b86a832SChris Mason { 498ce9adaa5SChris Mason struct end_io_wq *end_io_wq; 499ce9adaa5SChris Mason end_io_wq = kmalloc(sizeof(*end_io_wq), GFP_NOFS); 500ce9adaa5SChris Mason if (!end_io_wq) 501ce9adaa5SChris Mason return -ENOMEM; 502ce9adaa5SChris Mason 503ce9adaa5SChris Mason end_io_wq->private = bio->bi_private; 504ce9adaa5SChris Mason end_io_wq->end_io = bio->bi_end_io; 50522c59948SChris Mason end_io_wq->info = info; 506ce9adaa5SChris Mason end_io_wq->error = 0; 507ce9adaa5SChris Mason end_io_wq->bio = bio; 50822c59948SChris Mason end_io_wq->metadata = metadata; 509ce9adaa5SChris Mason 510ce9adaa5SChris Mason bio->bi_private = end_io_wq; 511ce9adaa5SChris Mason bio->bi_end_io = end_workqueue_bio; 51222c59948SChris Mason return 0; 51322c59948SChris Mason } 51422c59948SChris Mason 515b64a2851SChris Mason unsigned long btrfs_async_submit_limit(struct btrfs_fs_info *info) 5164854ddd0SChris Mason { 5174854ddd0SChris Mason unsigned long limit = min_t(unsigned long, 5184854ddd0SChris Mason info->workers.max_workers, 5194854ddd0SChris Mason info->fs_devices->open_devices); 5204854ddd0SChris Mason return 256 * limit; 5214854ddd0SChris Mason } 5224854ddd0SChris Mason 523777e6bd7SChris Mason int btrfs_congested_async(struct btrfs_fs_info *info, int iodone) 5240986fe9eSChris Mason { 525b64a2851SChris Mason return atomic_read(&info->nr_async_bios) > 526b64a2851SChris Mason btrfs_async_submit_limit(info); 5270986fe9eSChris Mason } 5280986fe9eSChris Mason 5294a69a410SChris Mason static void run_one_async_start(struct btrfs_work *work) 5304a69a410SChris Mason { 5314a69a410SChris Mason struct btrfs_fs_info *fs_info; 5324a69a410SChris Mason struct async_submit_bio *async; 5334a69a410SChris Mason 5344a69a410SChris Mason async = container_of(work, struct async_submit_bio, work); 5354a69a410SChris Mason fs_info = BTRFS_I(async->inode)->root->fs_info; 5364a69a410SChris Mason async->submit_bio_start(async->inode, async->rw, async->bio, 5374a69a410SChris Mason async->mirror_num, async->bio_flags); 5384a69a410SChris Mason } 5394a69a410SChris Mason 5404a69a410SChris Mason static void run_one_async_done(struct btrfs_work *work) 5418b712842SChris Mason { 5428b712842SChris Mason struct btrfs_fs_info *fs_info; 5438b712842SChris Mason struct async_submit_bio *async; 5444854ddd0SChris Mason int limit; 5458b712842SChris Mason 5468b712842SChris Mason async = container_of(work, struct async_submit_bio, work); 5478b712842SChris Mason fs_info = BTRFS_I(async->inode)->root->fs_info; 5484854ddd0SChris Mason 549b64a2851SChris Mason limit = btrfs_async_submit_limit(fs_info); 5504854ddd0SChris Mason limit = limit * 2 / 3; 5514854ddd0SChris Mason 5528b712842SChris Mason atomic_dec(&fs_info->nr_async_submits); 5530986fe9eSChris Mason 554b64a2851SChris Mason if (atomic_read(&fs_info->nr_async_submits) < limit && 555b64a2851SChris Mason waitqueue_active(&fs_info->async_submit_wait)) 5564854ddd0SChris Mason wake_up(&fs_info->async_submit_wait); 5574854ddd0SChris Mason 5584a69a410SChris Mason async->submit_bio_done(async->inode, async->rw, async->bio, 559c8b97818SChris Mason async->mirror_num, async->bio_flags); 5604a69a410SChris Mason } 5614a69a410SChris Mason 5624a69a410SChris Mason static void run_one_async_free(struct btrfs_work *work) 5634a69a410SChris Mason { 5644a69a410SChris Mason struct async_submit_bio *async; 5654a69a410SChris Mason 5664a69a410SChris Mason async = container_of(work, struct async_submit_bio, work); 5678b712842SChris Mason kfree(async); 5688b712842SChris Mason } 5698b712842SChris Mason 57044b8bd7eSChris Mason int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode, 57144b8bd7eSChris Mason int rw, struct bio *bio, int mirror_num, 572c8b97818SChris Mason unsigned long bio_flags, 5734a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_start, 5744a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_done) 57544b8bd7eSChris Mason { 57644b8bd7eSChris Mason struct async_submit_bio *async; 57744b8bd7eSChris Mason 57844b8bd7eSChris Mason async = kmalloc(sizeof(*async), GFP_NOFS); 57944b8bd7eSChris Mason if (!async) 58044b8bd7eSChris Mason return -ENOMEM; 58144b8bd7eSChris Mason 58244b8bd7eSChris Mason async->inode = inode; 58344b8bd7eSChris Mason async->rw = rw; 58444b8bd7eSChris Mason async->bio = bio; 58544b8bd7eSChris Mason async->mirror_num = mirror_num; 5864a69a410SChris Mason async->submit_bio_start = submit_bio_start; 5874a69a410SChris Mason async->submit_bio_done = submit_bio_done; 5884a69a410SChris Mason 5894a69a410SChris Mason async->work.func = run_one_async_start; 5904a69a410SChris Mason async->work.ordered_func = run_one_async_done; 5914a69a410SChris Mason async->work.ordered_free = run_one_async_free; 5924a69a410SChris Mason 5938b712842SChris Mason async->work.flags = 0; 594c8b97818SChris Mason async->bio_flags = bio_flags; 5958c8bee1dSChris Mason 596cb03c743SChris Mason atomic_inc(&fs_info->nr_async_submits); 597d313d7a3SChris Mason 598d313d7a3SChris Mason if (rw & (1 << BIO_RW_SYNCIO)) 599d313d7a3SChris Mason btrfs_set_work_high_prio(&async->work); 600d313d7a3SChris Mason 6018b712842SChris Mason btrfs_queue_worker(&fs_info->workers, &async->work); 6029473f16cSChris Mason 603771ed689SChris Mason while (atomic_read(&fs_info->async_submit_draining) && 604771ed689SChris Mason atomic_read(&fs_info->nr_async_submits)) { 605771ed689SChris Mason wait_event(fs_info->async_submit_wait, 606771ed689SChris Mason (atomic_read(&fs_info->nr_async_submits) == 0)); 607771ed689SChris Mason } 608771ed689SChris Mason 60944b8bd7eSChris Mason return 0; 61044b8bd7eSChris Mason } 61144b8bd7eSChris Mason 612ce3ed71aSChris Mason static int btree_csum_one_bio(struct bio *bio) 613ce3ed71aSChris Mason { 614ce3ed71aSChris Mason struct bio_vec *bvec = bio->bi_io_vec; 615ce3ed71aSChris Mason int bio_index = 0; 616ce3ed71aSChris Mason struct btrfs_root *root; 617ce3ed71aSChris Mason 618ce3ed71aSChris Mason WARN_ON(bio->bi_vcnt <= 0); 619ce3ed71aSChris Mason while (bio_index < bio->bi_vcnt) { 620ce3ed71aSChris Mason root = BTRFS_I(bvec->bv_page->mapping->host)->root; 621ce3ed71aSChris Mason csum_dirty_buffer(root, bvec->bv_page); 622ce3ed71aSChris Mason bio_index++; 623ce3ed71aSChris Mason bvec++; 624ce3ed71aSChris Mason } 625ce3ed71aSChris Mason return 0; 626ce3ed71aSChris Mason } 627ce3ed71aSChris Mason 6284a69a410SChris Mason static int __btree_submit_bio_start(struct inode *inode, int rw, 6294a69a410SChris Mason struct bio *bio, int mirror_num, 6304a69a410SChris Mason unsigned long bio_flags) 63122c59948SChris Mason { 6328b712842SChris Mason /* 6338b712842SChris Mason * when we're called for a write, we're already in the async 6345443be45SChris Mason * submission context. Just jump into btrfs_map_bio 6358b712842SChris Mason */ 636ce3ed71aSChris Mason btree_csum_one_bio(bio); 6374a69a410SChris Mason return 0; 63822c59948SChris Mason } 63922c59948SChris Mason 6404a69a410SChris Mason static int __btree_submit_bio_done(struct inode *inode, int rw, struct bio *bio, 6414a69a410SChris Mason int mirror_num, unsigned long bio_flags) 6424a69a410SChris Mason { 6438b712842SChris Mason /* 6444a69a410SChris Mason * when we're called for a write, we're already in the async 6454a69a410SChris Mason * submission context. Just jump into btrfs_map_bio 6468b712842SChris Mason */ 6478b712842SChris Mason return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, mirror_num, 1); 6480b86a832SChris Mason } 6490b86a832SChris Mason 65044b8bd7eSChris Mason static int btree_submit_bio_hook(struct inode *inode, int rw, struct bio *bio, 651c8b97818SChris Mason int mirror_num, unsigned long bio_flags) 65244b8bd7eSChris Mason { 6534a69a410SChris Mason int ret; 654cad321adSChris Mason 6554a69a410SChris Mason ret = btrfs_bio_wq_end_io(BTRFS_I(inode)->root->fs_info, 6564a69a410SChris Mason bio, 1); 6574a69a410SChris Mason BUG_ON(ret); 6584a69a410SChris Mason 659cad321adSChris Mason if (!(rw & (1 << BIO_RW))) { 660cad321adSChris Mason /* 661cad321adSChris Mason * called for a read, do the setup so that checksum validation 662cad321adSChris Mason * can happen in the async kernel threads 663cad321adSChris Mason */ 6644a69a410SChris Mason return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, 6656f3577bdSChris Mason mirror_num, 0); 66644b8bd7eSChris Mason } 667d313d7a3SChris Mason 668cad321adSChris Mason /* 669cad321adSChris Mason * kthread helpers are used to submit writes so that checksumming 670cad321adSChris Mason * can happen in parallel across all CPUs 671cad321adSChris Mason */ 67244b8bd7eSChris Mason return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info, 673c8b97818SChris Mason inode, rw, bio, mirror_num, 0, 6744a69a410SChris Mason __btree_submit_bio_start, 6754a69a410SChris Mason __btree_submit_bio_done); 67644b8bd7eSChris Mason } 67744b8bd7eSChris Mason 6785f39d397SChris Mason static int btree_writepage(struct page *page, struct writeback_control *wbc) 6795f39d397SChris Mason { 680d1310b2eSChris Mason struct extent_io_tree *tree; 681b9473439SChris Mason struct btrfs_root *root = BTRFS_I(page->mapping->host)->root; 682b9473439SChris Mason struct extent_buffer *eb; 683b9473439SChris Mason int was_dirty; 6845443be45SChris Mason 685b9473439SChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 686b9473439SChris Mason if (!(current->flags & PF_MEMALLOC)) { 687b9473439SChris Mason return extent_write_full_page(tree, page, 688b9473439SChris Mason btree_get_extent, wbc); 689b9473439SChris Mason } 690b9473439SChris Mason 6915443be45SChris Mason redirty_page_for_writepage(wbc, page); 692b9473439SChris Mason eb = btrfs_find_tree_block(root, page_offset(page), 693b9473439SChris Mason PAGE_CACHE_SIZE); 694b9473439SChris Mason WARN_ON(!eb); 695b9473439SChris Mason 696b9473439SChris Mason was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags); 697b9473439SChris Mason if (!was_dirty) { 698b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 699b9473439SChris Mason root->fs_info->dirty_metadata_bytes += PAGE_CACHE_SIZE; 700b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 701b9473439SChris Mason } 702b9473439SChris Mason free_extent_buffer(eb); 703b9473439SChris Mason 7045443be45SChris Mason unlock_page(page); 7055443be45SChris Mason return 0; 7065443be45SChris Mason } 7070da5468fSChris Mason 7080da5468fSChris Mason static int btree_writepages(struct address_space *mapping, 7090da5468fSChris Mason struct writeback_control *wbc) 7100da5468fSChris Mason { 711d1310b2eSChris Mason struct extent_io_tree *tree; 712d1310b2eSChris Mason tree = &BTRFS_I(mapping->host)->io_tree; 713d8d5f3e1SChris Mason if (wbc->sync_mode == WB_SYNC_NONE) { 714b9473439SChris Mason struct btrfs_root *root = BTRFS_I(mapping->host)->root; 715793955bcSChris Mason u64 num_dirty; 71624ab9cd8SChris Mason unsigned long thresh = 32 * 1024 * 1024; 717448d640bSChris Mason 718448d640bSChris Mason if (wbc->for_kupdate) 719448d640bSChris Mason return 0; 720448d640bSChris Mason 721b9473439SChris Mason /* this is a bit racy, but that's ok */ 722b9473439SChris Mason num_dirty = root->fs_info->dirty_metadata_bytes; 723d397712bSChris Mason if (num_dirty < thresh) 724793955bcSChris Mason return 0; 725793955bcSChris Mason } 7260da5468fSChris Mason return extent_writepages(tree, mapping, btree_get_extent, wbc); 7270da5468fSChris Mason } 7280da5468fSChris Mason 729b2950863SChristoph Hellwig static int btree_readpage(struct file *file, struct page *page) 7305f39d397SChris Mason { 731d1310b2eSChris Mason struct extent_io_tree *tree; 732d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 7335f39d397SChris Mason return extent_read_full_page(tree, page, btree_get_extent); 7345f39d397SChris Mason } 7355f39d397SChris Mason 73670dec807SChris Mason static int btree_releasepage(struct page *page, gfp_t gfp_flags) 7375f39d397SChris Mason { 738d1310b2eSChris Mason struct extent_io_tree *tree; 739d1310b2eSChris Mason struct extent_map_tree *map; 7405f39d397SChris Mason int ret; 7415f39d397SChris Mason 74298509cfcSChris Mason if (PageWriteback(page) || PageDirty(page)) 74398509cfcSChris Mason return 0; 74498509cfcSChris Mason 745d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 746d1310b2eSChris Mason map = &BTRFS_I(page->mapping->host)->extent_tree; 7476af118ceSChris Mason 7487b13b7b1SChris Mason ret = try_release_extent_state(map, tree, page, gfp_flags); 749d397712bSChris Mason if (!ret) 7506af118ceSChris Mason return 0; 7516af118ceSChris Mason 7526af118ceSChris Mason ret = try_release_extent_buffer(tree, page); 7535f39d397SChris Mason if (ret == 1) { 7545f39d397SChris Mason ClearPagePrivate(page); 7555f39d397SChris Mason set_page_private(page, 0); 756d98237b3SChris Mason page_cache_release(page); 7575f39d397SChris Mason } 7586af118ceSChris Mason 759d98237b3SChris Mason return ret; 760d98237b3SChris Mason } 761d98237b3SChris Mason 7625f39d397SChris Mason static void btree_invalidatepage(struct page *page, unsigned long offset) 763d98237b3SChris Mason { 764d1310b2eSChris Mason struct extent_io_tree *tree; 765d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 7665f39d397SChris Mason extent_invalidatepage(tree, page, offset); 7675f39d397SChris Mason btree_releasepage(page, GFP_NOFS); 7689ad6b7bcSChris Mason if (PagePrivate(page)) { 769d397712bSChris Mason printk(KERN_WARNING "btrfs warning page private not zero " 770d397712bSChris Mason "on page %llu\n", (unsigned long long)page_offset(page)); 7719ad6b7bcSChris Mason ClearPagePrivate(page); 7729ad6b7bcSChris Mason set_page_private(page, 0); 7739ad6b7bcSChris Mason page_cache_release(page); 7749ad6b7bcSChris Mason } 775d98237b3SChris Mason } 776d98237b3SChris Mason 7777f09410bSAlexey Dobriyan static const struct address_space_operations btree_aops = { 778d98237b3SChris Mason .readpage = btree_readpage, 779d98237b3SChris Mason .writepage = btree_writepage, 7800da5468fSChris Mason .writepages = btree_writepages, 7815f39d397SChris Mason .releasepage = btree_releasepage, 7825f39d397SChris Mason .invalidatepage = btree_invalidatepage, 783d98237b3SChris Mason .sync_page = block_sync_page, 784d98237b3SChris Mason }; 785123abc88SChris Mason 786ca7a79adSChris Mason int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize, 787ca7a79adSChris Mason u64 parent_transid) 788090d1875SChris Mason { 7895f39d397SChris Mason struct extent_buffer *buf = NULL; 7905f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 791de428b63SChris Mason int ret = 0; 792090d1875SChris Mason 793db94535dSChris Mason buf = btrfs_find_create_tree_block(root, bytenr, blocksize); 7945f39d397SChris Mason if (!buf) 795090d1875SChris Mason return 0; 796d1310b2eSChris Mason read_extent_buffer_pages(&BTRFS_I(btree_inode)->io_tree, 797f188591eSChris Mason buf, 0, 0, btree_get_extent, 0); 7985f39d397SChris Mason free_extent_buffer(buf); 799de428b63SChris Mason return ret; 800090d1875SChris Mason } 801090d1875SChris Mason 8020999df54SChris Mason struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root, 8030999df54SChris Mason u64 bytenr, u32 blocksize) 8040999df54SChris Mason { 8050999df54SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 8060999df54SChris Mason struct extent_buffer *eb; 8070999df54SChris Mason eb = find_extent_buffer(&BTRFS_I(btree_inode)->io_tree, 8080999df54SChris Mason bytenr, blocksize, GFP_NOFS); 8090999df54SChris Mason return eb; 8100999df54SChris Mason } 8110999df54SChris Mason 8120999df54SChris Mason struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root, 8130999df54SChris Mason u64 bytenr, u32 blocksize) 8140999df54SChris Mason { 8150999df54SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 8160999df54SChris Mason struct extent_buffer *eb; 8170999df54SChris Mason 8180999df54SChris Mason eb = alloc_extent_buffer(&BTRFS_I(btree_inode)->io_tree, 8190999df54SChris Mason bytenr, blocksize, NULL, GFP_NOFS); 8200999df54SChris Mason return eb; 8210999df54SChris Mason } 8220999df54SChris Mason 8230999df54SChris Mason 824e02119d5SChris Mason int btrfs_write_tree_block(struct extent_buffer *buf) 825e02119d5SChris Mason { 8268aa38c31SChristoph Hellwig return filemap_fdatawrite_range(buf->first_page->mapping, buf->start, 8278aa38c31SChristoph Hellwig buf->start + buf->len - 1); 828e02119d5SChris Mason } 829e02119d5SChris Mason 830e02119d5SChris Mason int btrfs_wait_tree_block_writeback(struct extent_buffer *buf) 831e02119d5SChris Mason { 8328aa38c31SChristoph Hellwig return filemap_fdatawait_range(buf->first_page->mapping, 8338aa38c31SChristoph Hellwig buf->start, buf->start + buf->len - 1); 834e02119d5SChris Mason } 835e02119d5SChris Mason 836db94535dSChris Mason struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr, 837ca7a79adSChris Mason u32 blocksize, u64 parent_transid) 838e20d96d6SChris Mason { 8395f39d397SChris Mason struct extent_buffer *buf = NULL; 8405f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 841d1310b2eSChris Mason struct extent_io_tree *io_tree; 84219c00ddcSChris Mason int ret; 84319c00ddcSChris Mason 844d1310b2eSChris Mason io_tree = &BTRFS_I(btree_inode)->io_tree; 845e20d96d6SChris Mason 846db94535dSChris Mason buf = btrfs_find_create_tree_block(root, bytenr, blocksize); 8475f39d397SChris Mason if (!buf) 848d98237b3SChris Mason return NULL; 849e4204dedSChris Mason 850ca7a79adSChris Mason ret = btree_read_extent_buffer_pages(root, buf, 0, parent_transid); 851ce9adaa5SChris Mason 852d397712bSChris Mason if (ret == 0) 853b4ce94deSChris Mason set_bit(EXTENT_BUFFER_UPTODATE, &buf->bflags); 8545f39d397SChris Mason return buf; 855ce9adaa5SChris Mason 856eb60ceacSChris Mason } 857eb60ceacSChris Mason 858e089f05cSChris Mason int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root, 8595f39d397SChris Mason struct extent_buffer *buf) 860ed2ff2cbSChris Mason { 8615f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 86255c69072SChris Mason if (btrfs_header_generation(buf) == 863925baeddSChris Mason root->fs_info->running_transaction->transid) { 864b9447ef8SChris Mason btrfs_assert_tree_locked(buf); 865b4ce94deSChris Mason 866b9473439SChris Mason if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &buf->bflags)) { 867b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 868b9473439SChris Mason if (root->fs_info->dirty_metadata_bytes >= buf->len) 869b9473439SChris Mason root->fs_info->dirty_metadata_bytes -= buf->len; 870b9473439SChris Mason else 871b9473439SChris Mason WARN_ON(1); 872b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 873b9473439SChris Mason } 874b4ce94deSChris Mason 875b9473439SChris Mason /* ugh, clear_extent_buffer_dirty needs to lock the page */ 876b9473439SChris Mason btrfs_set_lock_blocking(buf); 877d1310b2eSChris Mason clear_extent_buffer_dirty(&BTRFS_I(btree_inode)->io_tree, 87855c69072SChris Mason buf); 879925baeddSChris Mason } 8805f39d397SChris Mason return 0; 8815f39d397SChris Mason } 8825f39d397SChris Mason 883db94535dSChris Mason static int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize, 88487ee04ebSChris Mason u32 stripesize, struct btrfs_root *root, 8859f5fae2fSChris Mason struct btrfs_fs_info *fs_info, 886e20d96d6SChris Mason u64 objectid) 887d97e63b6SChris Mason { 888cfaa7295SChris Mason root->node = NULL; 889a28ec197SChris Mason root->commit_root = NULL; 890db94535dSChris Mason root->sectorsize = sectorsize; 891db94535dSChris Mason root->nodesize = nodesize; 892db94535dSChris Mason root->leafsize = leafsize; 89387ee04ebSChris Mason root->stripesize = stripesize; 894123abc88SChris Mason root->ref_cows = 0; 8950b86a832SChris Mason root->track_dirty = 0; 896c71bf099SYan, Zheng root->in_radix = 0; 897c71bf099SYan, Zheng root->clean_orphans = 0; 8980b86a832SChris Mason 8999f5fae2fSChris Mason root->fs_info = fs_info; 9000f7d52f4SChris Mason root->objectid = objectid; 9010f7d52f4SChris Mason root->last_trans = 0; 90213a8a7c8SYan, Zheng root->highest_objectid = 0; 90358176a96SJosef Bacik root->name = NULL; 9044313b399SChris Mason root->in_sysfs = 0; 9056bef4d31SEric Paris root->inode_tree = RB_ROOT; 906f0486c68SYan, Zheng root->block_rsv = NULL; 9070b86a832SChris Mason 9080b86a832SChris Mason INIT_LIST_HEAD(&root->dirty_list); 9097b128766SJosef Bacik INIT_LIST_HEAD(&root->orphan_list); 9105d4f98a2SYan Zheng INIT_LIST_HEAD(&root->root_list); 911925baeddSChris Mason spin_lock_init(&root->node_lock); 912bcc63abbSYan spin_lock_init(&root->list_lock); 9135d4f98a2SYan Zheng spin_lock_init(&root->inode_lock); 914f0486c68SYan, Zheng spin_lock_init(&root->accounting_lock); 915a2135011SChris Mason mutex_init(&root->objectid_mutex); 916e02119d5SChris Mason mutex_init(&root->log_mutex); 9177237f183SYan Zheng init_waitqueue_head(&root->log_writer_wait); 9187237f183SYan Zheng init_waitqueue_head(&root->log_commit_wait[0]); 9197237f183SYan Zheng init_waitqueue_head(&root->log_commit_wait[1]); 9207237f183SYan Zheng atomic_set(&root->log_commit[0], 0); 9217237f183SYan Zheng atomic_set(&root->log_commit[1], 0); 9227237f183SYan Zheng atomic_set(&root->log_writers, 0); 9237237f183SYan Zheng root->log_batch = 0; 9247237f183SYan Zheng root->log_transid = 0; 925257c62e1SChris Mason root->last_log_commit = 0; 926d0c803c4SChris Mason extent_io_tree_init(&root->dirty_log_pages, 927d0c803c4SChris Mason fs_info->btree_inode->i_mapping, GFP_NOFS); 928017e5369SChris Mason 9293768f368SChris Mason memset(&root->root_key, 0, sizeof(root->root_key)); 9303768f368SChris Mason memset(&root->root_item, 0, sizeof(root->root_item)); 9316702ed49SChris Mason memset(&root->defrag_progress, 0, sizeof(root->defrag_progress)); 93258176a96SJosef Bacik memset(&root->root_kobj, 0, sizeof(root->root_kobj)); 9333f157a2fSChris Mason root->defrag_trans_start = fs_info->generation; 93458176a96SJosef Bacik init_completion(&root->kobj_unregister); 9356702ed49SChris Mason root->defrag_running = 0; 9364d775673SChris Mason root->root_key.objectid = objectid; 9373394e160SChris Mason root->anon_super.s_root = NULL; 9383394e160SChris Mason root->anon_super.s_dev = 0; 9393394e160SChris Mason INIT_LIST_HEAD(&root->anon_super.s_list); 9403394e160SChris Mason INIT_LIST_HEAD(&root->anon_super.s_instances); 9413394e160SChris Mason init_rwsem(&root->anon_super.s_umount); 9423394e160SChris Mason 9433768f368SChris Mason return 0; 9443768f368SChris Mason } 9453768f368SChris Mason 946db94535dSChris Mason static int find_and_setup_root(struct btrfs_root *tree_root, 9479f5fae2fSChris Mason struct btrfs_fs_info *fs_info, 9489f5fae2fSChris Mason u64 objectid, 949e20d96d6SChris Mason struct btrfs_root *root) 9503768f368SChris Mason { 9513768f368SChris Mason int ret; 952db94535dSChris Mason u32 blocksize; 95384234f3aSYan Zheng u64 generation; 9543768f368SChris Mason 955db94535dSChris Mason __setup_root(tree_root->nodesize, tree_root->leafsize, 95687ee04ebSChris Mason tree_root->sectorsize, tree_root->stripesize, 95787ee04ebSChris Mason root, fs_info, objectid); 9583768f368SChris Mason ret = btrfs_find_last_root(tree_root, objectid, 9593768f368SChris Mason &root->root_item, &root->root_key); 9604df27c4dSYan, Zheng if (ret > 0) 9614df27c4dSYan, Zheng return -ENOENT; 9623768f368SChris Mason BUG_ON(ret); 9633768f368SChris Mason 96484234f3aSYan Zheng generation = btrfs_root_generation(&root->root_item); 965db94535dSChris Mason blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item)); 966db94535dSChris Mason root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item), 96784234f3aSYan Zheng blocksize, generation); 9683768f368SChris Mason BUG_ON(!root->node); 9694df27c4dSYan, Zheng root->commit_root = btrfs_root_node(root); 970d97e63b6SChris Mason return 0; 971d97e63b6SChris Mason } 972d97e63b6SChris Mason 973e02119d5SChris Mason int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans, 974e02119d5SChris Mason struct btrfs_fs_info *fs_info) 975e02119d5SChris Mason { 976e02119d5SChris Mason struct extent_buffer *eb; 977d0c803c4SChris Mason struct btrfs_root *log_root_tree = fs_info->log_root_tree; 978d0c803c4SChris Mason u64 start = 0; 979d0c803c4SChris Mason u64 end = 0; 980e02119d5SChris Mason int ret; 981e02119d5SChris Mason 982d0c803c4SChris Mason if (!log_root_tree) 983e02119d5SChris Mason return 0; 984e02119d5SChris Mason 985d0c803c4SChris Mason while (1) { 986d0c803c4SChris Mason ret = find_first_extent_bit(&log_root_tree->dirty_log_pages, 9878cef4e16SYan, Zheng 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW); 988d0c803c4SChris Mason if (ret) 989d0c803c4SChris Mason break; 990d0c803c4SChris Mason 9918cef4e16SYan, Zheng clear_extent_bits(&log_root_tree->dirty_log_pages, start, end, 9928cef4e16SYan, Zheng EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS); 993d0c803c4SChris Mason } 994e02119d5SChris Mason eb = fs_info->log_root_tree->node; 995e02119d5SChris Mason 996e02119d5SChris Mason WARN_ON(btrfs_header_level(eb) != 0); 997e02119d5SChris Mason WARN_ON(btrfs_header_nritems(eb) != 0); 998e02119d5SChris Mason 999d00aff00SChris Mason ret = btrfs_free_reserved_extent(fs_info->tree_root, 1000d00aff00SChris Mason eb->start, eb->len); 1001e02119d5SChris Mason BUG_ON(ret); 1002e02119d5SChris Mason 1003e02119d5SChris Mason free_extent_buffer(eb); 1004e02119d5SChris Mason kfree(fs_info->log_root_tree); 1005e02119d5SChris Mason fs_info->log_root_tree = NULL; 1006e02119d5SChris Mason return 0; 1007e02119d5SChris Mason } 1008e02119d5SChris Mason 10097237f183SYan Zheng static struct btrfs_root *alloc_log_tree(struct btrfs_trans_handle *trans, 1010e02119d5SChris Mason struct btrfs_fs_info *fs_info) 10110f7d52f4SChris Mason { 10120f7d52f4SChris Mason struct btrfs_root *root; 10130f7d52f4SChris Mason struct btrfs_root *tree_root = fs_info->tree_root; 10147237f183SYan Zheng struct extent_buffer *leaf; 1015e02119d5SChris Mason 1016e02119d5SChris Mason root = kzalloc(sizeof(*root), GFP_NOFS); 1017e02119d5SChris Mason if (!root) 10187237f183SYan Zheng return ERR_PTR(-ENOMEM); 1019e02119d5SChris Mason 1020e02119d5SChris Mason __setup_root(tree_root->nodesize, tree_root->leafsize, 1021e02119d5SChris Mason tree_root->sectorsize, tree_root->stripesize, 1022e02119d5SChris Mason root, fs_info, BTRFS_TREE_LOG_OBJECTID); 1023e02119d5SChris Mason 1024e02119d5SChris Mason root->root_key.objectid = BTRFS_TREE_LOG_OBJECTID; 1025e02119d5SChris Mason root->root_key.type = BTRFS_ROOT_ITEM_KEY; 1026e02119d5SChris Mason root->root_key.offset = BTRFS_TREE_LOG_OBJECTID; 10277237f183SYan Zheng /* 10287237f183SYan Zheng * log trees do not get reference counted because they go away 10297237f183SYan Zheng * before a real commit is actually done. They do store pointers 10307237f183SYan Zheng * to file data extents, and those reference counts still get 10317237f183SYan Zheng * updated (along with back refs to the log tree). 10327237f183SYan Zheng */ 1033e02119d5SChris Mason root->ref_cows = 0; 1034e02119d5SChris Mason 10355d4f98a2SYan Zheng leaf = btrfs_alloc_free_block(trans, root, root->leafsize, 0, 10365d4f98a2SYan Zheng BTRFS_TREE_LOG_OBJECTID, NULL, 0, 0, 0); 10377237f183SYan Zheng if (IS_ERR(leaf)) { 10387237f183SYan Zheng kfree(root); 10397237f183SYan Zheng return ERR_CAST(leaf); 10407237f183SYan Zheng } 1041e02119d5SChris Mason 10425d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, 0, sizeof(struct btrfs_header)); 10435d4f98a2SYan Zheng btrfs_set_header_bytenr(leaf, leaf->start); 10445d4f98a2SYan Zheng btrfs_set_header_generation(leaf, trans->transid); 10455d4f98a2SYan Zheng btrfs_set_header_backref_rev(leaf, BTRFS_MIXED_BACKREF_REV); 10465d4f98a2SYan Zheng btrfs_set_header_owner(leaf, BTRFS_TREE_LOG_OBJECTID); 10477237f183SYan Zheng root->node = leaf; 1048e02119d5SChris Mason 1049e02119d5SChris Mason write_extent_buffer(root->node, root->fs_info->fsid, 1050e02119d5SChris Mason (unsigned long)btrfs_header_fsid(root->node), 1051e02119d5SChris Mason BTRFS_FSID_SIZE); 1052e02119d5SChris Mason btrfs_mark_buffer_dirty(root->node); 1053e02119d5SChris Mason btrfs_tree_unlock(root->node); 10547237f183SYan Zheng return root; 10557237f183SYan Zheng } 10567237f183SYan Zheng 10577237f183SYan Zheng int btrfs_init_log_root_tree(struct btrfs_trans_handle *trans, 10587237f183SYan Zheng struct btrfs_fs_info *fs_info) 10597237f183SYan Zheng { 10607237f183SYan Zheng struct btrfs_root *log_root; 10617237f183SYan Zheng 10627237f183SYan Zheng log_root = alloc_log_tree(trans, fs_info); 10637237f183SYan Zheng if (IS_ERR(log_root)) 10647237f183SYan Zheng return PTR_ERR(log_root); 10657237f183SYan Zheng WARN_ON(fs_info->log_root_tree); 10667237f183SYan Zheng fs_info->log_root_tree = log_root; 10677237f183SYan Zheng return 0; 10687237f183SYan Zheng } 10697237f183SYan Zheng 10707237f183SYan Zheng int btrfs_add_log_tree(struct btrfs_trans_handle *trans, 10717237f183SYan Zheng struct btrfs_root *root) 10727237f183SYan Zheng { 10737237f183SYan Zheng struct btrfs_root *log_root; 10747237f183SYan Zheng struct btrfs_inode_item *inode_item; 10757237f183SYan Zheng 10767237f183SYan Zheng log_root = alloc_log_tree(trans, root->fs_info); 10777237f183SYan Zheng if (IS_ERR(log_root)) 10787237f183SYan Zheng return PTR_ERR(log_root); 10797237f183SYan Zheng 10807237f183SYan Zheng log_root->last_trans = trans->transid; 10817237f183SYan Zheng log_root->root_key.offset = root->root_key.objectid; 10827237f183SYan Zheng 10837237f183SYan Zheng inode_item = &log_root->root_item.inode; 10847237f183SYan Zheng inode_item->generation = cpu_to_le64(1); 10857237f183SYan Zheng inode_item->size = cpu_to_le64(3); 10867237f183SYan Zheng inode_item->nlink = cpu_to_le32(1); 10877237f183SYan Zheng inode_item->nbytes = cpu_to_le64(root->leafsize); 10887237f183SYan Zheng inode_item->mode = cpu_to_le32(S_IFDIR | 0755); 10897237f183SYan Zheng 10905d4f98a2SYan Zheng btrfs_set_root_node(&log_root->root_item, log_root->node); 10917237f183SYan Zheng 10927237f183SYan Zheng WARN_ON(root->log_root); 10937237f183SYan Zheng root->log_root = log_root; 10947237f183SYan Zheng root->log_transid = 0; 1095257c62e1SChris Mason root->last_log_commit = 0; 1096e02119d5SChris Mason return 0; 1097e02119d5SChris Mason } 1098e02119d5SChris Mason 1099e02119d5SChris Mason struct btrfs_root *btrfs_read_fs_root_no_radix(struct btrfs_root *tree_root, 1100e02119d5SChris Mason struct btrfs_key *location) 1101e02119d5SChris Mason { 1102e02119d5SChris Mason struct btrfs_root *root; 1103e02119d5SChris Mason struct btrfs_fs_info *fs_info = tree_root->fs_info; 11040f7d52f4SChris Mason struct btrfs_path *path; 11055f39d397SChris Mason struct extent_buffer *l; 110684234f3aSYan Zheng u64 generation; 1107db94535dSChris Mason u32 blocksize; 11080f7d52f4SChris Mason int ret = 0; 11090f7d52f4SChris Mason 11105eda7b5eSChris Mason root = kzalloc(sizeof(*root), GFP_NOFS); 11110cf6c620SChris Mason if (!root) 11120f7d52f4SChris Mason return ERR_PTR(-ENOMEM); 11130f7d52f4SChris Mason if (location->offset == (u64)-1) { 1114db94535dSChris Mason ret = find_and_setup_root(tree_root, fs_info, 11150f7d52f4SChris Mason location->objectid, root); 11160f7d52f4SChris Mason if (ret) { 11170f7d52f4SChris Mason kfree(root); 11180f7d52f4SChris Mason return ERR_PTR(ret); 11190f7d52f4SChris Mason } 112013a8a7c8SYan, Zheng goto out; 11210f7d52f4SChris Mason } 11220f7d52f4SChris Mason 1123db94535dSChris Mason __setup_root(tree_root->nodesize, tree_root->leafsize, 112487ee04ebSChris Mason tree_root->sectorsize, tree_root->stripesize, 112587ee04ebSChris Mason root, fs_info, location->objectid); 11260f7d52f4SChris Mason 11270f7d52f4SChris Mason path = btrfs_alloc_path(); 11280f7d52f4SChris Mason BUG_ON(!path); 11290f7d52f4SChris Mason ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0); 113013a8a7c8SYan, Zheng if (ret == 0) { 11315f39d397SChris Mason l = path->nodes[0]; 11325f39d397SChris Mason read_extent_buffer(l, &root->root_item, 11335f39d397SChris Mason btrfs_item_ptr_offset(l, path->slots[0]), 11340f7d52f4SChris Mason sizeof(root->root_item)); 113544b36eb2SYan Zheng memcpy(&root->root_key, location, sizeof(*location)); 113613a8a7c8SYan, Zheng } 11370f7d52f4SChris Mason btrfs_free_path(path); 11380f7d52f4SChris Mason if (ret) { 113913a8a7c8SYan, Zheng if (ret > 0) 114013a8a7c8SYan, Zheng ret = -ENOENT; 11410f7d52f4SChris Mason return ERR_PTR(ret); 11420f7d52f4SChris Mason } 114313a8a7c8SYan, Zheng 114484234f3aSYan Zheng generation = btrfs_root_generation(&root->root_item); 1145db94535dSChris Mason blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item)); 1146db94535dSChris Mason root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item), 114784234f3aSYan Zheng blocksize, generation); 11485d4f98a2SYan Zheng root->commit_root = btrfs_root_node(root); 11490f7d52f4SChris Mason BUG_ON(!root->node); 115013a8a7c8SYan, Zheng out: 115113a8a7c8SYan, Zheng if (location->objectid != BTRFS_TREE_LOG_OBJECTID) 11520f7d52f4SChris Mason root->ref_cows = 1; 115313a8a7c8SYan, Zheng 11545eda7b5eSChris Mason return root; 11555eda7b5eSChris Mason } 11565eda7b5eSChris Mason 1157dc17ff8fSChris Mason struct btrfs_root *btrfs_lookup_fs_root(struct btrfs_fs_info *fs_info, 1158dc17ff8fSChris Mason u64 root_objectid) 1159dc17ff8fSChris Mason { 1160dc17ff8fSChris Mason struct btrfs_root *root; 1161dc17ff8fSChris Mason 1162dc17ff8fSChris Mason if (root_objectid == BTRFS_ROOT_TREE_OBJECTID) 1163dc17ff8fSChris Mason return fs_info->tree_root; 1164dc17ff8fSChris Mason if (root_objectid == BTRFS_EXTENT_TREE_OBJECTID) 1165dc17ff8fSChris Mason return fs_info->extent_root; 1166dc17ff8fSChris Mason 1167dc17ff8fSChris Mason root = radix_tree_lookup(&fs_info->fs_roots_radix, 1168dc17ff8fSChris Mason (unsigned long)root_objectid); 1169dc17ff8fSChris Mason return root; 1170dc17ff8fSChris Mason } 1171dc17ff8fSChris Mason 1172edbd8d4eSChris Mason struct btrfs_root *btrfs_read_fs_root_no_name(struct btrfs_fs_info *fs_info, 1173edbd8d4eSChris Mason struct btrfs_key *location) 11745eda7b5eSChris Mason { 11755eda7b5eSChris Mason struct btrfs_root *root; 11765eda7b5eSChris Mason int ret; 11775eda7b5eSChris Mason 1178edbd8d4eSChris Mason if (location->objectid == BTRFS_ROOT_TREE_OBJECTID) 1179edbd8d4eSChris Mason return fs_info->tree_root; 1180edbd8d4eSChris Mason if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID) 1181edbd8d4eSChris Mason return fs_info->extent_root; 11828f18cf13SChris Mason if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID) 11838f18cf13SChris Mason return fs_info->chunk_root; 11848f18cf13SChris Mason if (location->objectid == BTRFS_DEV_TREE_OBJECTID) 11858f18cf13SChris Mason return fs_info->dev_root; 11860403e47eSYan Zheng if (location->objectid == BTRFS_CSUM_TREE_OBJECTID) 11870403e47eSYan Zheng return fs_info->csum_root; 11884df27c4dSYan, Zheng again: 11894df27c4dSYan, Zheng spin_lock(&fs_info->fs_roots_radix_lock); 11905eda7b5eSChris Mason root = radix_tree_lookup(&fs_info->fs_roots_radix, 11915eda7b5eSChris Mason (unsigned long)location->objectid); 11924df27c4dSYan, Zheng spin_unlock(&fs_info->fs_roots_radix_lock); 11935eda7b5eSChris Mason if (root) 11945eda7b5eSChris Mason return root; 11955eda7b5eSChris Mason 11964df27c4dSYan, Zheng ret = btrfs_find_orphan_item(fs_info->tree_root, location->objectid); 11974df27c4dSYan, Zheng if (ret == 0) 11984df27c4dSYan, Zheng ret = -ENOENT; 11994df27c4dSYan, Zheng if (ret < 0) 12004df27c4dSYan, Zheng return ERR_PTR(ret); 12014df27c4dSYan, Zheng 1202e02119d5SChris Mason root = btrfs_read_fs_root_no_radix(fs_info->tree_root, location); 12035eda7b5eSChris Mason if (IS_ERR(root)) 12045eda7b5eSChris Mason return root; 12053394e160SChris Mason 12064df27c4dSYan, Zheng WARN_ON(btrfs_root_refs(&root->root_item) == 0); 12073394e160SChris Mason set_anon_super(&root->anon_super, NULL); 12083394e160SChris Mason 12094df27c4dSYan, Zheng ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM); 12104df27c4dSYan, Zheng if (ret) 12114df27c4dSYan, Zheng goto fail; 12124df27c4dSYan, Zheng 12134df27c4dSYan, Zheng spin_lock(&fs_info->fs_roots_radix_lock); 12142619ba1fSChris Mason ret = radix_tree_insert(&fs_info->fs_roots_radix, 12152619ba1fSChris Mason (unsigned long)root->root_key.objectid, 12160f7d52f4SChris Mason root); 1217c71bf099SYan, Zheng if (ret == 0) { 12184df27c4dSYan, Zheng root->in_radix = 1; 1219c71bf099SYan, Zheng root->clean_orphans = 1; 1220c71bf099SYan, Zheng } 12214df27c4dSYan, Zheng spin_unlock(&fs_info->fs_roots_radix_lock); 12224df27c4dSYan, Zheng radix_tree_preload_end(); 12230f7d52f4SChris Mason if (ret) { 12244df27c4dSYan, Zheng if (ret == -EEXIST) { 12254df27c4dSYan, Zheng free_fs_root(root); 12264df27c4dSYan, Zheng goto again; 12270f7d52f4SChris Mason } 12284df27c4dSYan, Zheng goto fail; 12294df27c4dSYan, Zheng } 12304df27c4dSYan, Zheng 1231edbd8d4eSChris Mason ret = btrfs_find_dead_roots(fs_info->tree_root, 12325d4f98a2SYan Zheng root->root_key.objectid); 12334df27c4dSYan, Zheng WARN_ON(ret); 1234edbd8d4eSChris Mason return root; 12354df27c4dSYan, Zheng fail: 12364df27c4dSYan, Zheng free_fs_root(root); 12374df27c4dSYan, Zheng return ERR_PTR(ret); 1238edbd8d4eSChris Mason } 1239edbd8d4eSChris Mason 1240edbd8d4eSChris Mason struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info, 1241edbd8d4eSChris Mason struct btrfs_key *location, 1242edbd8d4eSChris Mason const char *name, int namelen) 1243edbd8d4eSChris Mason { 12444df27c4dSYan, Zheng return btrfs_read_fs_root_no_name(fs_info, location); 12454df27c4dSYan, Zheng #if 0 1246edbd8d4eSChris Mason struct btrfs_root *root; 1247edbd8d4eSChris Mason int ret; 1248edbd8d4eSChris Mason 1249edbd8d4eSChris Mason root = btrfs_read_fs_root_no_name(fs_info, location); 1250edbd8d4eSChris Mason if (!root) 1251edbd8d4eSChris Mason return NULL; 125258176a96SJosef Bacik 12534313b399SChris Mason if (root->in_sysfs) 12544313b399SChris Mason return root; 12554313b399SChris Mason 125658176a96SJosef Bacik ret = btrfs_set_root_name(root, name, namelen); 125758176a96SJosef Bacik if (ret) { 12585f39d397SChris Mason free_extent_buffer(root->node); 125958176a96SJosef Bacik kfree(root); 126058176a96SJosef Bacik return ERR_PTR(ret); 126158176a96SJosef Bacik } 12624df27c4dSYan, Zheng 126358176a96SJosef Bacik ret = btrfs_sysfs_add_root(root); 126458176a96SJosef Bacik if (ret) { 12655f39d397SChris Mason free_extent_buffer(root->node); 126658176a96SJosef Bacik kfree(root->name); 126758176a96SJosef Bacik kfree(root); 126858176a96SJosef Bacik return ERR_PTR(ret); 126958176a96SJosef Bacik } 12704313b399SChris Mason root->in_sysfs = 1; 12710f7d52f4SChris Mason return root; 12724df27c4dSYan, Zheng #endif 12730f7d52f4SChris Mason } 127404160088SChris Mason 127504160088SChris Mason static int btrfs_congested_fn(void *congested_data, int bdi_bits) 127604160088SChris Mason { 127704160088SChris Mason struct btrfs_fs_info *info = (struct btrfs_fs_info *)congested_data; 127804160088SChris Mason int ret = 0; 127904160088SChris Mason struct btrfs_device *device; 128004160088SChris Mason struct backing_dev_info *bdi; 1281b7967db7SChris Mason 1282c6e30871SQinghuang Feng list_for_each_entry(device, &info->fs_devices->devices, dev_list) { 1283dfe25020SChris Mason if (!device->bdev) 1284dfe25020SChris Mason continue; 128504160088SChris Mason bdi = blk_get_backing_dev_info(device->bdev); 128604160088SChris Mason if (bdi && bdi_congested(bdi, bdi_bits)) { 128704160088SChris Mason ret = 1; 128804160088SChris Mason break; 128904160088SChris Mason } 129004160088SChris Mason } 129104160088SChris Mason return ret; 129204160088SChris Mason } 129304160088SChris Mason 129438b66988SChris Mason /* 129538b66988SChris Mason * this unplugs every device on the box, and it is only used when page 129638b66988SChris Mason * is null 129738b66988SChris Mason */ 129838b66988SChris Mason static void __unplug_io_fn(struct backing_dev_info *bdi, struct page *page) 129938b66988SChris Mason { 130038b66988SChris Mason struct btrfs_device *device; 130138b66988SChris Mason struct btrfs_fs_info *info; 130238b66988SChris Mason 130338b66988SChris Mason info = (struct btrfs_fs_info *)bdi->unplug_io_data; 1304c6e30871SQinghuang Feng list_for_each_entry(device, &info->fs_devices->devices, dev_list) { 1305d20f7043SChris Mason if (!device->bdev) 1306d20f7043SChris Mason continue; 1307d20f7043SChris Mason 130838b66988SChris Mason bdi = blk_get_backing_dev_info(device->bdev); 1309d397712bSChris Mason if (bdi->unplug_io_fn) 131038b66988SChris Mason bdi->unplug_io_fn(bdi, page); 131138b66988SChris Mason } 131238b66988SChris Mason } 131338b66988SChris Mason 1314b2950863SChristoph Hellwig static void btrfs_unplug_io_fn(struct backing_dev_info *bdi, struct page *page) 131504160088SChris Mason { 131638b66988SChris Mason struct inode *inode; 1317f2d8d74dSChris Mason struct extent_map_tree *em_tree; 1318f2d8d74dSChris Mason struct extent_map *em; 1319bcbfce8aSChris Mason struct address_space *mapping; 132038b66988SChris Mason u64 offset; 132138b66988SChris Mason 1322bcbfce8aSChris Mason /* the generic O_DIRECT read code does this */ 13239f0ba5bdSChris Mason if (1 || !page) { 132438b66988SChris Mason __unplug_io_fn(bdi, page); 132538b66988SChris Mason return; 132638b66988SChris Mason } 132738b66988SChris Mason 1328bcbfce8aSChris Mason /* 1329bcbfce8aSChris Mason * page->mapping may change at any time. Get a consistent copy 1330bcbfce8aSChris Mason * and use that for everything below 1331bcbfce8aSChris Mason */ 1332bcbfce8aSChris Mason smp_mb(); 1333bcbfce8aSChris Mason mapping = page->mapping; 1334bcbfce8aSChris Mason if (!mapping) 1335bcbfce8aSChris Mason return; 1336bcbfce8aSChris Mason 1337bcbfce8aSChris Mason inode = mapping->host; 1338240d5d48SChris Mason 1339240d5d48SChris Mason /* 1340240d5d48SChris Mason * don't do the expensive searching for a small number of 1341240d5d48SChris Mason * devices 1342240d5d48SChris Mason */ 1343240d5d48SChris Mason if (BTRFS_I(inode)->root->fs_info->fs_devices->open_devices <= 2) { 1344240d5d48SChris Mason __unplug_io_fn(bdi, page); 1345240d5d48SChris Mason return; 1346240d5d48SChris Mason } 1347240d5d48SChris Mason 134838b66988SChris Mason offset = page_offset(page); 134904160088SChris Mason 1350f2d8d74dSChris Mason em_tree = &BTRFS_I(inode)->extent_tree; 1351890871beSChris Mason read_lock(&em_tree->lock); 1352f2d8d74dSChris Mason em = lookup_extent_mapping(em_tree, offset, PAGE_CACHE_SIZE); 1353890871beSChris Mason read_unlock(&em_tree->lock); 135489642229SChris Mason if (!em) { 135589642229SChris Mason __unplug_io_fn(bdi, page); 1356f2d8d74dSChris Mason return; 135789642229SChris Mason } 1358f2d8d74dSChris Mason 135989642229SChris Mason if (em->block_start >= EXTENT_MAP_LAST_BYTE) { 136089642229SChris Mason free_extent_map(em); 136189642229SChris Mason __unplug_io_fn(bdi, page); 136289642229SChris Mason return; 136389642229SChris Mason } 1364f2d8d74dSChris Mason offset = offset - em->start; 1365f2d8d74dSChris Mason btrfs_unplug_page(&BTRFS_I(inode)->root->fs_info->mapping_tree, 1366f2d8d74dSChris Mason em->block_start + offset, page); 1367f2d8d74dSChris Mason free_extent_map(em); 136804160088SChris Mason } 136904160088SChris Mason 1370ad081f14SJens Axboe /* 1371ad081f14SJens Axboe * If this fails, caller must call bdi_destroy() to get rid of the 1372ad081f14SJens Axboe * bdi again. 1373ad081f14SJens Axboe */ 137404160088SChris Mason static int setup_bdi(struct btrfs_fs_info *info, struct backing_dev_info *bdi) 137504160088SChris Mason { 1376ad081f14SJens Axboe int err; 1377ad081f14SJens Axboe 1378ad081f14SJens Axboe bdi->capabilities = BDI_CAP_MAP_COPY; 1379e6d086d8SJens Axboe err = bdi_setup_and_register(bdi, "btrfs", BDI_CAP_MAP_COPY); 1380ad081f14SJens Axboe if (err) 1381ad081f14SJens Axboe return err; 1382ad081f14SJens Axboe 13834575c9ccSChris Mason bdi->ra_pages = default_backing_dev_info.ra_pages; 138404160088SChris Mason bdi->unplug_io_fn = btrfs_unplug_io_fn; 138504160088SChris Mason bdi->unplug_io_data = info; 138604160088SChris Mason bdi->congested_fn = btrfs_congested_fn; 138704160088SChris Mason bdi->congested_data = info; 138804160088SChris Mason return 0; 138904160088SChris Mason } 139004160088SChris Mason 1391ce9adaa5SChris Mason static int bio_ready_for_csum(struct bio *bio) 1392ce9adaa5SChris Mason { 1393ce9adaa5SChris Mason u64 length = 0; 1394ce9adaa5SChris Mason u64 buf_len = 0; 1395ce9adaa5SChris Mason u64 start = 0; 1396ce9adaa5SChris Mason struct page *page; 1397ce9adaa5SChris Mason struct extent_io_tree *io_tree = NULL; 1398ce9adaa5SChris Mason struct btrfs_fs_info *info = NULL; 1399ce9adaa5SChris Mason struct bio_vec *bvec; 1400ce9adaa5SChris Mason int i; 1401ce9adaa5SChris Mason int ret; 1402ce9adaa5SChris Mason 1403ce9adaa5SChris Mason bio_for_each_segment(bvec, bio, i) { 1404ce9adaa5SChris Mason page = bvec->bv_page; 1405ce9adaa5SChris Mason if (page->private == EXTENT_PAGE_PRIVATE) { 1406ce9adaa5SChris Mason length += bvec->bv_len; 1407ce9adaa5SChris Mason continue; 1408ce9adaa5SChris Mason } 1409ce9adaa5SChris Mason if (!page->private) { 1410ce9adaa5SChris Mason length += bvec->bv_len; 1411ce9adaa5SChris Mason continue; 1412ce9adaa5SChris Mason } 1413ce9adaa5SChris Mason length = bvec->bv_len; 1414ce9adaa5SChris Mason buf_len = page->private >> 2; 1415ce9adaa5SChris Mason start = page_offset(page) + bvec->bv_offset; 1416ce9adaa5SChris Mason io_tree = &BTRFS_I(page->mapping->host)->io_tree; 1417ce9adaa5SChris Mason info = BTRFS_I(page->mapping->host)->root->fs_info; 1418ce9adaa5SChris Mason } 1419ce9adaa5SChris Mason /* are we fully contained in this bio? */ 1420ce9adaa5SChris Mason if (buf_len <= length) 1421ce9adaa5SChris Mason return 1; 1422ce9adaa5SChris Mason 1423ce9adaa5SChris Mason ret = extent_range_uptodate(io_tree, start + length, 1424ce9adaa5SChris Mason start + buf_len - 1); 1425ce9adaa5SChris Mason return ret; 1426ce9adaa5SChris Mason } 1427ce9adaa5SChris Mason 14288b712842SChris Mason /* 14298b712842SChris Mason * called by the kthread helper functions to finally call the bio end_io 14308b712842SChris Mason * functions. This is where read checksum verification actually happens 14318b712842SChris Mason */ 14328b712842SChris Mason static void end_workqueue_fn(struct btrfs_work *work) 1433ce9adaa5SChris Mason { 1434ce9adaa5SChris Mason struct bio *bio; 14358b712842SChris Mason struct end_io_wq *end_io_wq; 14368b712842SChris Mason struct btrfs_fs_info *fs_info; 1437ce9adaa5SChris Mason int error; 1438ce9adaa5SChris Mason 14398b712842SChris Mason end_io_wq = container_of(work, struct end_io_wq, work); 1440ce9adaa5SChris Mason bio = end_io_wq->bio; 14418b712842SChris Mason fs_info = end_io_wq->info; 14428b712842SChris Mason 1443cad321adSChris Mason /* metadata bio reads are special because the whole tree block must 14448b712842SChris Mason * be checksummed at once. This makes sure the entire block is in 14458b712842SChris Mason * ram and up to date before trying to verify things. For 14468b712842SChris Mason * blocksize <= pagesize, it is basically a noop 14478b712842SChris Mason */ 1448cad321adSChris Mason if (!(bio->bi_rw & (1 << BIO_RW)) && end_io_wq->metadata && 1449cad321adSChris Mason !bio_ready_for_csum(bio)) { 1450d20f7043SChris Mason btrfs_queue_worker(&fs_info->endio_meta_workers, 14518b712842SChris Mason &end_io_wq->work); 1452ce9adaa5SChris Mason return; 1453ce9adaa5SChris Mason } 1454ce9adaa5SChris Mason error = end_io_wq->error; 1455ce9adaa5SChris Mason bio->bi_private = end_io_wq->private; 1456ce9adaa5SChris Mason bio->bi_end_io = end_io_wq->end_io; 1457ce9adaa5SChris Mason kfree(end_io_wq); 1458ce9adaa5SChris Mason bio_endio(bio, error); 1459ce9adaa5SChris Mason } 146044b8bd7eSChris Mason 1461a74a4b97SChris Mason static int cleaner_kthread(void *arg) 1462a74a4b97SChris Mason { 1463a74a4b97SChris Mason struct btrfs_root *root = arg; 1464a74a4b97SChris Mason 1465a74a4b97SChris Mason do { 1466a74a4b97SChris Mason vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE); 146776dda93cSYan, Zheng 146876dda93cSYan, Zheng if (!(root->fs_info->sb->s_flags & MS_RDONLY) && 146976dda93cSYan, Zheng mutex_trylock(&root->fs_info->cleaner_mutex)) { 147024bbcf04SYan, Zheng btrfs_run_delayed_iputs(root); 1471a74a4b97SChris Mason btrfs_clean_old_snapshots(root); 1472a74a4b97SChris Mason mutex_unlock(&root->fs_info->cleaner_mutex); 147376dda93cSYan, Zheng } 1474a74a4b97SChris Mason 1475a74a4b97SChris Mason if (freezing(current)) { 1476a74a4b97SChris Mason refrigerator(); 1477a74a4b97SChris Mason } else { 1478a74a4b97SChris Mason set_current_state(TASK_INTERRUPTIBLE); 14798929ecfaSYan, Zheng if (!kthread_should_stop()) 1480a74a4b97SChris Mason schedule(); 1481a74a4b97SChris Mason __set_current_state(TASK_RUNNING); 1482a74a4b97SChris Mason } 1483a74a4b97SChris Mason } while (!kthread_should_stop()); 1484a74a4b97SChris Mason return 0; 1485a74a4b97SChris Mason } 1486a74a4b97SChris Mason 1487a74a4b97SChris Mason static int transaction_kthread(void *arg) 1488a74a4b97SChris Mason { 1489a74a4b97SChris Mason struct btrfs_root *root = arg; 1490a74a4b97SChris Mason struct btrfs_trans_handle *trans; 1491a74a4b97SChris Mason struct btrfs_transaction *cur; 14928929ecfaSYan, Zheng u64 transid; 1493a74a4b97SChris Mason unsigned long now; 1494a74a4b97SChris Mason unsigned long delay; 1495a74a4b97SChris Mason int ret; 1496a74a4b97SChris Mason 1497a74a4b97SChris Mason do { 1498a74a4b97SChris Mason delay = HZ * 30; 1499a74a4b97SChris Mason vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE); 1500a74a4b97SChris Mason mutex_lock(&root->fs_info->transaction_kthread_mutex); 1501a74a4b97SChris Mason 15028929ecfaSYan, Zheng spin_lock(&root->fs_info->new_trans_lock); 1503a74a4b97SChris Mason cur = root->fs_info->running_transaction; 1504a74a4b97SChris Mason if (!cur) { 15058929ecfaSYan, Zheng spin_unlock(&root->fs_info->new_trans_lock); 1506a74a4b97SChris Mason goto sleep; 1507a74a4b97SChris Mason } 150831153d81SYan Zheng 1509a74a4b97SChris Mason now = get_seconds(); 15108929ecfaSYan, Zheng if (!cur->blocked && 15118929ecfaSYan, Zheng (now < cur->start_time || now - cur->start_time < 30)) { 15128929ecfaSYan, Zheng spin_unlock(&root->fs_info->new_trans_lock); 1513a74a4b97SChris Mason delay = HZ * 5; 1514a74a4b97SChris Mason goto sleep; 1515a74a4b97SChris Mason } 15168929ecfaSYan, Zheng transid = cur->transid; 15178929ecfaSYan, Zheng spin_unlock(&root->fs_info->new_trans_lock); 151856bec294SChris Mason 15198929ecfaSYan, Zheng trans = btrfs_join_transaction(root, 1); 15208929ecfaSYan, Zheng if (transid == trans->transid) { 15218929ecfaSYan, Zheng ret = btrfs_commit_transaction(trans, root); 15228929ecfaSYan, Zheng BUG_ON(ret); 15238929ecfaSYan, Zheng } else { 15248929ecfaSYan, Zheng btrfs_end_transaction(trans, root); 15258929ecfaSYan, Zheng } 1526a74a4b97SChris Mason sleep: 1527a74a4b97SChris Mason wake_up_process(root->fs_info->cleaner_kthread); 1528a74a4b97SChris Mason mutex_unlock(&root->fs_info->transaction_kthread_mutex); 1529a74a4b97SChris Mason 1530a74a4b97SChris Mason if (freezing(current)) { 1531a74a4b97SChris Mason refrigerator(); 1532a74a4b97SChris Mason } else { 1533a74a4b97SChris Mason set_current_state(TASK_INTERRUPTIBLE); 15348929ecfaSYan, Zheng if (!kthread_should_stop() && 15358929ecfaSYan, Zheng !btrfs_transaction_blocked(root->fs_info)) 1536a74a4b97SChris Mason schedule_timeout(delay); 1537a74a4b97SChris Mason __set_current_state(TASK_RUNNING); 1538a74a4b97SChris Mason } 1539a74a4b97SChris Mason } while (!kthread_should_stop()); 1540a74a4b97SChris Mason return 0; 1541a74a4b97SChris Mason } 1542a74a4b97SChris Mason 15438a4b83ccSChris Mason struct btrfs_root *open_ctree(struct super_block *sb, 1544dfe25020SChris Mason struct btrfs_fs_devices *fs_devices, 1545dfe25020SChris Mason char *options) 1546eb60ceacSChris Mason { 1547db94535dSChris Mason u32 sectorsize; 1548db94535dSChris Mason u32 nodesize; 1549db94535dSChris Mason u32 leafsize; 1550db94535dSChris Mason u32 blocksize; 155187ee04ebSChris Mason u32 stripesize; 155284234f3aSYan Zheng u64 generation; 1553f2b636e8SJosef Bacik u64 features; 15543de4586cSChris Mason struct btrfs_key location; 1555a061fc8dSChris Mason struct buffer_head *bh; 1556e02119d5SChris Mason struct btrfs_root *extent_root = kzalloc(sizeof(struct btrfs_root), 1557e20d96d6SChris Mason GFP_NOFS); 1558d20f7043SChris Mason struct btrfs_root *csum_root = kzalloc(sizeof(struct btrfs_root), 1559d20f7043SChris Mason GFP_NOFS); 1560e02119d5SChris Mason struct btrfs_root *tree_root = kzalloc(sizeof(struct btrfs_root), 1561e20d96d6SChris Mason GFP_NOFS); 15628790d502SChris Mason struct btrfs_fs_info *fs_info = kzalloc(sizeof(*fs_info), 1563e20d96d6SChris Mason GFP_NOFS); 1564e02119d5SChris Mason struct btrfs_root *chunk_root = kzalloc(sizeof(struct btrfs_root), 15650b86a832SChris Mason GFP_NOFS); 1566e02119d5SChris Mason struct btrfs_root *dev_root = kzalloc(sizeof(struct btrfs_root), 15670b86a832SChris Mason GFP_NOFS); 1568e02119d5SChris Mason struct btrfs_root *log_tree_root; 1569e02119d5SChris Mason 1570eb60ceacSChris Mason int ret; 1571e58ca020SYan int err = -EINVAL; 15724543df7eSChris Mason 15732c90e5d6SChris Mason struct btrfs_super_block *disk_super; 15748790d502SChris Mason 15750463bb4eSJim Meyering if (!extent_root || !tree_root || !fs_info || 1576d20f7043SChris Mason !chunk_root || !dev_root || !csum_root) { 157739279cc3SChris Mason err = -ENOMEM; 157839279cc3SChris Mason goto fail; 157939279cc3SChris Mason } 158076dda93cSYan, Zheng 158176dda93cSYan, Zheng ret = init_srcu_struct(&fs_info->subvol_srcu); 158276dda93cSYan, Zheng if (ret) { 158376dda93cSYan, Zheng err = ret; 158476dda93cSYan, Zheng goto fail; 158576dda93cSYan, Zheng } 158676dda93cSYan, Zheng 158776dda93cSYan, Zheng ret = setup_bdi(fs_info, &fs_info->bdi); 158876dda93cSYan, Zheng if (ret) { 158976dda93cSYan, Zheng err = ret; 159076dda93cSYan, Zheng goto fail_srcu; 159176dda93cSYan, Zheng } 159276dda93cSYan, Zheng 159376dda93cSYan, Zheng fs_info->btree_inode = new_inode(sb); 159476dda93cSYan, Zheng if (!fs_info->btree_inode) { 159576dda93cSYan, Zheng err = -ENOMEM; 159676dda93cSYan, Zheng goto fail_bdi; 159776dda93cSYan, Zheng } 159876dda93cSYan, Zheng 159976dda93cSYan, Zheng INIT_RADIX_TREE(&fs_info->fs_roots_radix, GFP_ATOMIC); 16008fd17795SChris Mason INIT_LIST_HEAD(&fs_info->trans_list); 1601facda1e7SChris Mason INIT_LIST_HEAD(&fs_info->dead_roots); 160224bbcf04SYan, Zheng INIT_LIST_HEAD(&fs_info->delayed_iputs); 160319c00ddcSChris Mason INIT_LIST_HEAD(&fs_info->hashers); 1604ea8c2819SChris Mason INIT_LIST_HEAD(&fs_info->delalloc_inodes); 16055a3f23d5SChris Mason INIT_LIST_HEAD(&fs_info->ordered_operations); 160611833d66SYan Zheng INIT_LIST_HEAD(&fs_info->caching_block_groups); 16071832a6d5SChris Mason spin_lock_init(&fs_info->delalloc_lock); 1608cee36a03SChris Mason spin_lock_init(&fs_info->new_trans_lock); 160931153d81SYan Zheng spin_lock_init(&fs_info->ref_cache_lock); 161076dda93cSYan, Zheng spin_lock_init(&fs_info->fs_roots_radix_lock); 161124bbcf04SYan, Zheng spin_lock_init(&fs_info->delayed_iput_lock); 161219c00ddcSChris Mason 161358176a96SJosef Bacik init_completion(&fs_info->kobj_unregister); 16149f5fae2fSChris Mason fs_info->tree_root = tree_root; 16159f5fae2fSChris Mason fs_info->extent_root = extent_root; 1616d20f7043SChris Mason fs_info->csum_root = csum_root; 16170b86a832SChris Mason fs_info->chunk_root = chunk_root; 16180b86a832SChris Mason fs_info->dev_root = dev_root; 16198a4b83ccSChris Mason fs_info->fs_devices = fs_devices; 16200b86a832SChris Mason INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots); 16216324fbf3SChris Mason INIT_LIST_HEAD(&fs_info->space_info); 16220b86a832SChris Mason btrfs_mapping_init(&fs_info->mapping_tree); 1623f0486c68SYan, Zheng btrfs_init_block_rsv(&fs_info->global_block_rsv); 1624f0486c68SYan, Zheng btrfs_init_block_rsv(&fs_info->delalloc_block_rsv); 1625f0486c68SYan, Zheng btrfs_init_block_rsv(&fs_info->trans_block_rsv); 1626f0486c68SYan, Zheng btrfs_init_block_rsv(&fs_info->chunk_block_rsv); 1627f0486c68SYan, Zheng btrfs_init_block_rsv(&fs_info->empty_block_rsv); 1628f0486c68SYan, Zheng INIT_LIST_HEAD(&fs_info->durable_block_rsv_list); 1629f0486c68SYan, Zheng mutex_init(&fs_info->durable_block_rsv_mutex); 1630cb03c743SChris Mason atomic_set(&fs_info->nr_async_submits, 0); 1631771ed689SChris Mason atomic_set(&fs_info->async_delalloc_pages, 0); 16328c8bee1dSChris Mason atomic_set(&fs_info->async_submit_draining, 0); 16330986fe9eSChris Mason atomic_set(&fs_info->nr_async_bios, 0); 1634e20d96d6SChris Mason fs_info->sb = sb; 16356f568d35SChris Mason fs_info->max_inline = 8192 * 1024; 16369ed74f2dSJosef Bacik fs_info->metadata_ratio = 0; 1637c8b97818SChris Mason 1638b34b086cSChris Mason fs_info->thread_pool_size = min_t(unsigned long, 1639b34b086cSChris Mason num_online_cpus() + 2, 8); 16400afbaf8cSChris Mason 16413eaa2885SChris Mason INIT_LIST_HEAD(&fs_info->ordered_extents); 16423eaa2885SChris Mason spin_lock_init(&fs_info->ordered_extent_lock); 16433eaa2885SChris Mason 1644a061fc8dSChris Mason sb->s_blocksize = 4096; 1645a061fc8dSChris Mason sb->s_blocksize_bits = blksize_bits(4096); 164632a88aa1SJens Axboe sb->s_bdi = &fs_info->bdi; 1647a061fc8dSChris Mason 164876dda93cSYan, Zheng fs_info->btree_inode->i_ino = BTRFS_BTREE_INODE_OBJECTID; 164976dda93cSYan, Zheng fs_info->btree_inode->i_nlink = 1; 16500afbaf8cSChris Mason /* 16510afbaf8cSChris Mason * we set the i_size on the btree inode to the max possible int. 16520afbaf8cSChris Mason * the real end of the address space is determined by all of 16530afbaf8cSChris Mason * the devices in the system 16540afbaf8cSChris Mason */ 16550afbaf8cSChris Mason fs_info->btree_inode->i_size = OFFSET_MAX; 1656d98237b3SChris Mason fs_info->btree_inode->i_mapping->a_ops = &btree_aops; 165704160088SChris Mason fs_info->btree_inode->i_mapping->backing_dev_info = &fs_info->bdi; 165804160088SChris Mason 16595d4f98a2SYan Zheng RB_CLEAR_NODE(&BTRFS_I(fs_info->btree_inode)->rb_node); 1660d1310b2eSChris Mason extent_io_tree_init(&BTRFS_I(fs_info->btree_inode)->io_tree, 16615f39d397SChris Mason fs_info->btree_inode->i_mapping, 16625f39d397SChris Mason GFP_NOFS); 1663d1310b2eSChris Mason extent_map_tree_init(&BTRFS_I(fs_info->btree_inode)->extent_tree, 1664d1310b2eSChris Mason GFP_NOFS); 16650da5468fSChris Mason 1666d1310b2eSChris Mason BTRFS_I(fs_info->btree_inode)->io_tree.ops = &btree_extent_io_ops; 1667d1310b2eSChris Mason 16680f7d52f4SChris Mason BTRFS_I(fs_info->btree_inode)->root = tree_root; 16690f7d52f4SChris Mason memset(&BTRFS_I(fs_info->btree_inode)->location, 0, 16700f7d52f4SChris Mason sizeof(struct btrfs_key)); 167176dda93cSYan, Zheng BTRFS_I(fs_info->btree_inode)->dummy_inode = 1; 167222b0ebdaSChris Mason insert_inode_hash(fs_info->btree_inode); 167339279cc3SChris Mason 1674e02119d5SChris Mason spin_lock_init(&fs_info->block_group_cache_lock); 16756bef4d31SEric Paris fs_info->block_group_cache_tree = RB_ROOT; 1676925baeddSChris Mason 167711833d66SYan Zheng extent_io_tree_init(&fs_info->freed_extents[0], 16787d9eb12cSChris Mason fs_info->btree_inode->i_mapping, GFP_NOFS); 167911833d66SYan Zheng extent_io_tree_init(&fs_info->freed_extents[1], 168011833d66SYan Zheng fs_info->btree_inode->i_mapping, GFP_NOFS); 168111833d66SYan Zheng fs_info->pinned_extents = &fs_info->freed_extents[0]; 1682e6dcd2dcSChris Mason fs_info->do_barriers = 1; 1683f9295749SChris Mason 1684d98237b3SChris Mason 1685509659cdSChris Mason mutex_init(&fs_info->trans_mutex); 16865a3f23d5SChris Mason mutex_init(&fs_info->ordered_operations_mutex); 168730ae8467SChris Mason mutex_init(&fs_info->tree_log_mutex); 1688925baeddSChris Mason mutex_init(&fs_info->chunk_mutex); 1689925baeddSChris Mason mutex_init(&fs_info->transaction_kthread_mutex); 1690a74a4b97SChris Mason mutex_init(&fs_info->cleaner_mutex); 1691a74a4b97SChris Mason mutex_init(&fs_info->volume_mutex); 1692276e680dSYan Zheng init_rwsem(&fs_info->extent_commit_sem); 1693c71bf099SYan, Zheng init_rwsem(&fs_info->cleanup_work_sem); 169476dda93cSYan, Zheng init_rwsem(&fs_info->subvol_sem); 1695fa9c0d79SChris Mason 1696fa9c0d79SChris Mason btrfs_init_free_cluster(&fs_info->meta_alloc_cluster); 1697fa9c0d79SChris Mason btrfs_init_free_cluster(&fs_info->data_alloc_cluster); 1698fa9c0d79SChris Mason 16993768f368SChris Mason init_waitqueue_head(&fs_info->transaction_throttle); 170019c00ddcSChris Mason init_waitqueue_head(&fs_info->transaction_wait); 17014854ddd0SChris Mason init_waitqueue_head(&fs_info->async_submit_wait); 17029a8dd150SChris Mason 17030b86a832SChris Mason __setup_root(4096, 4096, 4096, 4096, tree_root, 17042c90e5d6SChris Mason fs_info, BTRFS_ROOT_TREE_OBJECTID); 17057eccb903SChris Mason 1706d98237b3SChris Mason 1707a512bbf8SYan Zheng bh = btrfs_read_dev_super(fs_devices->latest_bdev); 1708a061fc8dSChris Mason if (!bh) 170939279cc3SChris Mason goto fail_iput; 171039279cc3SChris Mason 1711a061fc8dSChris Mason memcpy(&fs_info->super_copy, bh->b_data, sizeof(fs_info->super_copy)); 17122d69a0f8SYan Zheng memcpy(&fs_info->super_for_commit, &fs_info->super_copy, 17132d69a0f8SYan Zheng sizeof(fs_info->super_for_commit)); 1714a061fc8dSChris Mason brelse(bh); 17155f39d397SChris Mason 1716a061fc8dSChris Mason memcpy(fs_info->fsid, fs_info->super_copy.fsid, BTRFS_FSID_SIZE); 17170b86a832SChris Mason 17185f39d397SChris Mason disk_super = &fs_info->super_copy; 17190f7d52f4SChris Mason if (!btrfs_super_root(disk_super)) 1720c6e2bac1SJosef Bacik goto fail_iput; 17210f7d52f4SChris Mason 17222b82032cSYan Zheng ret = btrfs_parse_options(tree_root, options); 17232b82032cSYan Zheng if (ret) { 17242b82032cSYan Zheng err = ret; 1725c6e2bac1SJosef Bacik goto fail_iput; 17262b82032cSYan Zheng } 1727dfe25020SChris Mason 1728f2b636e8SJosef Bacik features = btrfs_super_incompat_flags(disk_super) & 1729f2b636e8SJosef Bacik ~BTRFS_FEATURE_INCOMPAT_SUPP; 1730f2b636e8SJosef Bacik if (features) { 1731f2b636e8SJosef Bacik printk(KERN_ERR "BTRFS: couldn't mount because of " 1732f2b636e8SJosef Bacik "unsupported optional features (%Lx).\n", 173321380931SJoel Becker (unsigned long long)features); 1734f2b636e8SJosef Bacik err = -EINVAL; 1735c6e2bac1SJosef Bacik goto fail_iput; 1736f2b636e8SJosef Bacik } 1737f2b636e8SJosef Bacik 17385d4f98a2SYan Zheng features = btrfs_super_incompat_flags(disk_super); 17395d4f98a2SYan Zheng if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) { 17405d4f98a2SYan Zheng features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF; 17415d4f98a2SYan Zheng btrfs_set_super_incompat_flags(disk_super, features); 17425d4f98a2SYan Zheng } 17435d4f98a2SYan Zheng 1744f2b636e8SJosef Bacik features = btrfs_super_compat_ro_flags(disk_super) & 1745f2b636e8SJosef Bacik ~BTRFS_FEATURE_COMPAT_RO_SUPP; 1746f2b636e8SJosef Bacik if (!(sb->s_flags & MS_RDONLY) && features) { 1747f2b636e8SJosef Bacik printk(KERN_ERR "BTRFS: couldn't mount RDWR because of " 1748f2b636e8SJosef Bacik "unsupported option features (%Lx).\n", 174921380931SJoel Becker (unsigned long long)features); 1750f2b636e8SJosef Bacik err = -EINVAL; 1751c6e2bac1SJosef Bacik goto fail_iput; 1752f2b636e8SJosef Bacik } 175361d92c32SChris Mason 175461d92c32SChris Mason btrfs_init_workers(&fs_info->generic_worker, 175561d92c32SChris Mason "genwork", 1, NULL); 175661d92c32SChris Mason 17575443be45SChris Mason btrfs_init_workers(&fs_info->workers, "worker", 175861d92c32SChris Mason fs_info->thread_pool_size, 175961d92c32SChris Mason &fs_info->generic_worker); 1760c8b97818SChris Mason 1761771ed689SChris Mason btrfs_init_workers(&fs_info->delalloc_workers, "delalloc", 176261d92c32SChris Mason fs_info->thread_pool_size, 176361d92c32SChris Mason &fs_info->generic_worker); 1764771ed689SChris Mason 17655443be45SChris Mason btrfs_init_workers(&fs_info->submit_workers, "submit", 1766b720d209SChris Mason min_t(u64, fs_devices->num_devices, 176761d92c32SChris Mason fs_info->thread_pool_size), 176861d92c32SChris Mason &fs_info->generic_worker); 176961b49440SChris Mason 177061b49440SChris Mason /* a higher idle thresh on the submit workers makes it much more 177161b49440SChris Mason * likely that bios will be send down in a sane order to the 177261b49440SChris Mason * devices 177361b49440SChris Mason */ 177461b49440SChris Mason fs_info->submit_workers.idle_thresh = 64; 177553863232SChris Mason 1776771ed689SChris Mason fs_info->workers.idle_thresh = 16; 17774a69a410SChris Mason fs_info->workers.ordered = 1; 177861b49440SChris Mason 1779771ed689SChris Mason fs_info->delalloc_workers.idle_thresh = 2; 1780771ed689SChris Mason fs_info->delalloc_workers.ordered = 1; 1781771ed689SChris Mason 178261d92c32SChris Mason btrfs_init_workers(&fs_info->fixup_workers, "fixup", 1, 178361d92c32SChris Mason &fs_info->generic_worker); 17845443be45SChris Mason btrfs_init_workers(&fs_info->endio_workers, "endio", 178561d92c32SChris Mason fs_info->thread_pool_size, 178661d92c32SChris Mason &fs_info->generic_worker); 1787d20f7043SChris Mason btrfs_init_workers(&fs_info->endio_meta_workers, "endio-meta", 178861d92c32SChris Mason fs_info->thread_pool_size, 178961d92c32SChris Mason &fs_info->generic_worker); 1790cad321adSChris Mason btrfs_init_workers(&fs_info->endio_meta_write_workers, 179161d92c32SChris Mason "endio-meta-write", fs_info->thread_pool_size, 179261d92c32SChris Mason &fs_info->generic_worker); 17935443be45SChris Mason btrfs_init_workers(&fs_info->endio_write_workers, "endio-write", 179461d92c32SChris Mason fs_info->thread_pool_size, 179561d92c32SChris Mason &fs_info->generic_worker); 179661b49440SChris Mason 179761b49440SChris Mason /* 179861b49440SChris Mason * endios are largely parallel and should have a very 179961b49440SChris Mason * low idle thresh 180061b49440SChris Mason */ 180161b49440SChris Mason fs_info->endio_workers.idle_thresh = 4; 1802b51912c9SChris Mason fs_info->endio_meta_workers.idle_thresh = 4; 1803b51912c9SChris Mason 18049042846bSChris Mason fs_info->endio_write_workers.idle_thresh = 2; 18059042846bSChris Mason fs_info->endio_meta_write_workers.idle_thresh = 2; 18069042846bSChris Mason 18074543df7eSChris Mason btrfs_start_workers(&fs_info->workers, 1); 180861d92c32SChris Mason btrfs_start_workers(&fs_info->generic_worker, 1); 18091cc127b5SChris Mason btrfs_start_workers(&fs_info->submit_workers, 1); 1810771ed689SChris Mason btrfs_start_workers(&fs_info->delalloc_workers, 1); 1811247e743cSChris Mason btrfs_start_workers(&fs_info->fixup_workers, 1); 18129042846bSChris Mason btrfs_start_workers(&fs_info->endio_workers, 1); 18139042846bSChris Mason btrfs_start_workers(&fs_info->endio_meta_workers, 1); 18149042846bSChris Mason btrfs_start_workers(&fs_info->endio_meta_write_workers, 1); 18159042846bSChris Mason btrfs_start_workers(&fs_info->endio_write_workers, 1); 18164543df7eSChris Mason 18174575c9ccSChris Mason fs_info->bdi.ra_pages *= btrfs_super_num_devices(disk_super); 1818c8b97818SChris Mason fs_info->bdi.ra_pages = max(fs_info->bdi.ra_pages, 1819c8b97818SChris Mason 4 * 1024 * 1024 / PAGE_CACHE_SIZE); 18204575c9ccSChris Mason 1821db94535dSChris Mason nodesize = btrfs_super_nodesize(disk_super); 1822db94535dSChris Mason leafsize = btrfs_super_leafsize(disk_super); 1823db94535dSChris Mason sectorsize = btrfs_super_sectorsize(disk_super); 182487ee04ebSChris Mason stripesize = btrfs_super_stripesize(disk_super); 1825db94535dSChris Mason tree_root->nodesize = nodesize; 1826db94535dSChris Mason tree_root->leafsize = leafsize; 1827db94535dSChris Mason tree_root->sectorsize = sectorsize; 182887ee04ebSChris Mason tree_root->stripesize = stripesize; 1829a061fc8dSChris Mason 1830a061fc8dSChris Mason sb->s_blocksize = sectorsize; 1831a061fc8dSChris Mason sb->s_blocksize_bits = blksize_bits(sectorsize); 1832db94535dSChris Mason 183339279cc3SChris Mason if (strncmp((char *)(&disk_super->magic), BTRFS_MAGIC, 183439279cc3SChris Mason sizeof(disk_super->magic))) { 1835d397712bSChris Mason printk(KERN_INFO "btrfs: valid FS not found on %s\n", sb->s_id); 183639279cc3SChris Mason goto fail_sb_buffer; 183739279cc3SChris Mason } 183819c00ddcSChris Mason 1839925baeddSChris Mason mutex_lock(&fs_info->chunk_mutex); 1840e4404d6eSYan Zheng ret = btrfs_read_sys_array(tree_root); 1841925baeddSChris Mason mutex_unlock(&fs_info->chunk_mutex); 184284eed90fSChris Mason if (ret) { 1843d397712bSChris Mason printk(KERN_WARNING "btrfs: failed to read the system " 1844d397712bSChris Mason "array on %s\n", sb->s_id); 18455d4f98a2SYan Zheng goto fail_sb_buffer; 184684eed90fSChris Mason } 18470b86a832SChris Mason 18480b86a832SChris Mason blocksize = btrfs_level_size(tree_root, 18490b86a832SChris Mason btrfs_super_chunk_root_level(disk_super)); 185084234f3aSYan Zheng generation = btrfs_super_chunk_root_generation(disk_super); 18510b86a832SChris Mason 18520b86a832SChris Mason __setup_root(nodesize, leafsize, sectorsize, stripesize, 18530b86a832SChris Mason chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID); 18540b86a832SChris Mason 18550b86a832SChris Mason chunk_root->node = read_tree_block(chunk_root, 18560b86a832SChris Mason btrfs_super_chunk_root(disk_super), 185784234f3aSYan Zheng blocksize, generation); 18580b86a832SChris Mason BUG_ON(!chunk_root->node); 185983121942SDavid Woodhouse if (!test_bit(EXTENT_BUFFER_UPTODATE, &chunk_root->node->bflags)) { 186083121942SDavid Woodhouse printk(KERN_WARNING "btrfs: failed to read chunk root on %s\n", 186183121942SDavid Woodhouse sb->s_id); 186283121942SDavid Woodhouse goto fail_chunk_root; 186383121942SDavid Woodhouse } 18645d4f98a2SYan Zheng btrfs_set_root_node(&chunk_root->root_item, chunk_root->node); 18655d4f98a2SYan Zheng chunk_root->commit_root = btrfs_root_node(chunk_root); 18660b86a832SChris Mason 1867e17cade2SChris Mason read_extent_buffer(chunk_root->node, fs_info->chunk_tree_uuid, 1868e17cade2SChris Mason (unsigned long)btrfs_header_chunk_tree_uuid(chunk_root->node), 1869e17cade2SChris Mason BTRFS_UUID_SIZE); 1870e17cade2SChris Mason 1871925baeddSChris Mason mutex_lock(&fs_info->chunk_mutex); 18720b86a832SChris Mason ret = btrfs_read_chunk_tree(chunk_root); 1873925baeddSChris Mason mutex_unlock(&fs_info->chunk_mutex); 18742b82032cSYan Zheng if (ret) { 1875d397712bSChris Mason printk(KERN_WARNING "btrfs: failed to read chunk tree on %s\n", 1876d397712bSChris Mason sb->s_id); 18772b82032cSYan Zheng goto fail_chunk_root; 18782b82032cSYan Zheng } 18790b86a832SChris Mason 1880dfe25020SChris Mason btrfs_close_extra_devices(fs_devices); 1881dfe25020SChris Mason 1882db94535dSChris Mason blocksize = btrfs_level_size(tree_root, 1883db94535dSChris Mason btrfs_super_root_level(disk_super)); 188484234f3aSYan Zheng generation = btrfs_super_generation(disk_super); 18850b86a832SChris Mason 1886e20d96d6SChris Mason tree_root->node = read_tree_block(tree_root, 1887db94535dSChris Mason btrfs_super_root(disk_super), 188884234f3aSYan Zheng blocksize, generation); 188939279cc3SChris Mason if (!tree_root->node) 18902b82032cSYan Zheng goto fail_chunk_root; 189183121942SDavid Woodhouse if (!test_bit(EXTENT_BUFFER_UPTODATE, &tree_root->node->bflags)) { 189283121942SDavid Woodhouse printk(KERN_WARNING "btrfs: failed to read tree root on %s\n", 189383121942SDavid Woodhouse sb->s_id); 189483121942SDavid Woodhouse goto fail_tree_root; 189583121942SDavid Woodhouse } 18965d4f98a2SYan Zheng btrfs_set_root_node(&tree_root->root_item, tree_root->node); 18975d4f98a2SYan Zheng tree_root->commit_root = btrfs_root_node(tree_root); 1898db94535dSChris Mason 1899db94535dSChris Mason ret = find_and_setup_root(tree_root, fs_info, 1900e20d96d6SChris Mason BTRFS_EXTENT_TREE_OBJECTID, extent_root); 19010b86a832SChris Mason if (ret) 190239279cc3SChris Mason goto fail_tree_root; 19030b86a832SChris Mason extent_root->track_dirty = 1; 19040b86a832SChris Mason 19050b86a832SChris Mason ret = find_and_setup_root(tree_root, fs_info, 19060b86a832SChris Mason BTRFS_DEV_TREE_OBJECTID, dev_root); 19070b86a832SChris Mason if (ret) 19080b86a832SChris Mason goto fail_extent_root; 19095d4f98a2SYan Zheng dev_root->track_dirty = 1; 19103768f368SChris Mason 1911d20f7043SChris Mason ret = find_and_setup_root(tree_root, fs_info, 1912d20f7043SChris Mason BTRFS_CSUM_TREE_OBJECTID, csum_root); 1913d20f7043SChris Mason if (ret) 19145d4f98a2SYan Zheng goto fail_dev_root; 1915d20f7043SChris Mason 1916d20f7043SChris Mason csum_root->track_dirty = 1; 1917d20f7043SChris Mason 19188929ecfaSYan, Zheng fs_info->generation = generation; 19198929ecfaSYan, Zheng fs_info->last_trans_committed = generation; 19208929ecfaSYan, Zheng fs_info->data_alloc_profile = (u64)-1; 19218929ecfaSYan, Zheng fs_info->metadata_alloc_profile = (u64)-1; 19228929ecfaSYan, Zheng fs_info->system_alloc_profile = fs_info->metadata_alloc_profile; 19238929ecfaSYan, Zheng 19241b1d1f66SJosef Bacik ret = btrfs_read_block_groups(extent_root); 19251b1d1f66SJosef Bacik if (ret) { 19261b1d1f66SJosef Bacik printk(KERN_ERR "Failed to read block groups: %d\n", ret); 19271b1d1f66SJosef Bacik goto fail_block_groups; 19281b1d1f66SJosef Bacik } 19299078a3e1SChris Mason 1930a74a4b97SChris Mason fs_info->cleaner_kthread = kthread_run(cleaner_kthread, tree_root, 1931a74a4b97SChris Mason "btrfs-cleaner"); 193257506d50SQinghuang Feng if (IS_ERR(fs_info->cleaner_kthread)) 19331b1d1f66SJosef Bacik goto fail_block_groups; 1934a74a4b97SChris Mason 1935a74a4b97SChris Mason fs_info->transaction_kthread = kthread_run(transaction_kthread, 1936a74a4b97SChris Mason tree_root, 1937a74a4b97SChris Mason "btrfs-transaction"); 193857506d50SQinghuang Feng if (IS_ERR(fs_info->transaction_kthread)) 19393f157a2fSChris Mason goto fail_cleaner; 1940a74a4b97SChris Mason 1941c289811cSChris Mason if (!btrfs_test_opt(tree_root, SSD) && 1942c289811cSChris Mason !btrfs_test_opt(tree_root, NOSSD) && 1943c289811cSChris Mason !fs_info->fs_devices->rotating) { 1944c289811cSChris Mason printk(KERN_INFO "Btrfs detected SSD devices, enabling SSD " 1945c289811cSChris Mason "mode\n"); 1946c289811cSChris Mason btrfs_set_opt(fs_info->mount_opt, SSD); 1947c289811cSChris Mason } 1948c289811cSChris Mason 1949e02119d5SChris Mason if (btrfs_super_log_root(disk_super) != 0) { 1950e02119d5SChris Mason u64 bytenr = btrfs_super_log_root(disk_super); 1951d18a2c44SChris Mason 19527c2ca468SChris Mason if (fs_devices->rw_devices == 0) { 1953d397712bSChris Mason printk(KERN_WARNING "Btrfs log replay required " 1954d397712bSChris Mason "on RO media\n"); 19557c2ca468SChris Mason err = -EIO; 19567c2ca468SChris Mason goto fail_trans_kthread; 19577c2ca468SChris Mason } 1958e02119d5SChris Mason blocksize = 1959e02119d5SChris Mason btrfs_level_size(tree_root, 1960e02119d5SChris Mason btrfs_super_log_root_level(disk_super)); 1961e02119d5SChris Mason 1962e02119d5SChris Mason log_tree_root = kzalloc(sizeof(struct btrfs_root), 1963e02119d5SChris Mason GFP_NOFS); 1964e02119d5SChris Mason 1965e02119d5SChris Mason __setup_root(nodesize, leafsize, sectorsize, stripesize, 1966e02119d5SChris Mason log_tree_root, fs_info, BTRFS_TREE_LOG_OBJECTID); 1967e02119d5SChris Mason 1968e02119d5SChris Mason log_tree_root->node = read_tree_block(tree_root, bytenr, 196984234f3aSYan Zheng blocksize, 197084234f3aSYan Zheng generation + 1); 1971e02119d5SChris Mason ret = btrfs_recover_log_trees(log_tree_root); 1972e02119d5SChris Mason BUG_ON(ret); 1973e556ce2cSYan Zheng 1974e556ce2cSYan Zheng if (sb->s_flags & MS_RDONLY) { 1975e556ce2cSYan Zheng ret = btrfs_commit_super(tree_root); 1976e556ce2cSYan Zheng BUG_ON(ret); 1977e556ce2cSYan Zheng } 1978e02119d5SChris Mason } 19791a40e23bSZheng Yan 198076dda93cSYan, Zheng ret = btrfs_find_orphan_roots(tree_root); 198176dda93cSYan, Zheng BUG_ON(ret); 198276dda93cSYan, Zheng 19837c2ca468SChris Mason if (!(sb->s_flags & MS_RDONLY)) { 19845d4f98a2SYan Zheng ret = btrfs_recover_relocation(tree_root); 1985d7ce5843SMiao Xie if (ret < 0) { 1986d7ce5843SMiao Xie printk(KERN_WARNING 1987d7ce5843SMiao Xie "btrfs: failed to recover relocation\n"); 1988d7ce5843SMiao Xie err = -EINVAL; 1989d7ce5843SMiao Xie goto fail_trans_kthread; 1990d7ce5843SMiao Xie } 19917c2ca468SChris Mason } 19921a40e23bSZheng Yan 19933de4586cSChris Mason location.objectid = BTRFS_FS_TREE_OBJECTID; 19943de4586cSChris Mason location.type = BTRFS_ROOT_ITEM_KEY; 19953de4586cSChris Mason location.offset = (u64)-1; 19963de4586cSChris Mason 19973de4586cSChris Mason fs_info->fs_root = btrfs_read_fs_root_no_name(fs_info, &location); 19983de4586cSChris Mason if (!fs_info->fs_root) 19997c2ca468SChris Mason goto fail_trans_kthread; 2000c289811cSChris Mason 2001e3acc2a6SJosef Bacik if (!(sb->s_flags & MS_RDONLY)) { 2002e3acc2a6SJosef Bacik down_read(&fs_info->cleanup_work_sem); 2003e3acc2a6SJosef Bacik btrfs_orphan_cleanup(fs_info->fs_root); 2004e3acc2a6SJosef Bacik up_read(&fs_info->cleanup_work_sem); 2005e3acc2a6SJosef Bacik } 2006e3acc2a6SJosef Bacik 20070f7d52f4SChris Mason return tree_root; 200839279cc3SChris Mason 20097c2ca468SChris Mason fail_trans_kthread: 20107c2ca468SChris Mason kthread_stop(fs_info->transaction_kthread); 20113f157a2fSChris Mason fail_cleaner: 2012a74a4b97SChris Mason kthread_stop(fs_info->cleaner_kthread); 20137c2ca468SChris Mason 20147c2ca468SChris Mason /* 20157c2ca468SChris Mason * make sure we're done with the btree inode before we stop our 20167c2ca468SChris Mason * kthreads 20177c2ca468SChris Mason */ 20187c2ca468SChris Mason filemap_write_and_wait(fs_info->btree_inode->i_mapping); 20197c2ca468SChris Mason invalidate_inode_pages2(fs_info->btree_inode->i_mapping); 20207c2ca468SChris Mason 20211b1d1f66SJosef Bacik fail_block_groups: 20221b1d1f66SJosef Bacik btrfs_free_block_groups(fs_info); 2023d20f7043SChris Mason free_extent_buffer(csum_root->node); 20245d4f98a2SYan Zheng free_extent_buffer(csum_root->commit_root); 20255d4f98a2SYan Zheng fail_dev_root: 20265d4f98a2SYan Zheng free_extent_buffer(dev_root->node); 20275d4f98a2SYan Zheng free_extent_buffer(dev_root->commit_root); 20280b86a832SChris Mason fail_extent_root: 20290b86a832SChris Mason free_extent_buffer(extent_root->node); 20305d4f98a2SYan Zheng free_extent_buffer(extent_root->commit_root); 203139279cc3SChris Mason fail_tree_root: 20325f39d397SChris Mason free_extent_buffer(tree_root->node); 20335d4f98a2SYan Zheng free_extent_buffer(tree_root->commit_root); 20342b82032cSYan Zheng fail_chunk_root: 20352b82032cSYan Zheng free_extent_buffer(chunk_root->node); 20365d4f98a2SYan Zheng free_extent_buffer(chunk_root->commit_root); 203739279cc3SChris Mason fail_sb_buffer: 203861d92c32SChris Mason btrfs_stop_workers(&fs_info->generic_worker); 2039247e743cSChris Mason btrfs_stop_workers(&fs_info->fixup_workers); 2040771ed689SChris Mason btrfs_stop_workers(&fs_info->delalloc_workers); 20418b712842SChris Mason btrfs_stop_workers(&fs_info->workers); 20428b712842SChris Mason btrfs_stop_workers(&fs_info->endio_workers); 2043d20f7043SChris Mason btrfs_stop_workers(&fs_info->endio_meta_workers); 2044cad321adSChris Mason btrfs_stop_workers(&fs_info->endio_meta_write_workers); 2045e6dcd2dcSChris Mason btrfs_stop_workers(&fs_info->endio_write_workers); 20461cc127b5SChris Mason btrfs_stop_workers(&fs_info->submit_workers); 20474543df7eSChris Mason fail_iput: 20487c2ca468SChris Mason invalidate_inode_pages2(fs_info->btree_inode->i_mapping); 20494543df7eSChris Mason iput(fs_info->btree_inode); 20507e662854SQinghuang Feng 2051dfe25020SChris Mason btrfs_close_devices(fs_info->fs_devices); 205284eed90fSChris Mason btrfs_mapping_tree_free(&fs_info->mapping_tree); 2053ad081f14SJens Axboe fail_bdi: 20547e662854SQinghuang Feng bdi_destroy(&fs_info->bdi); 205576dda93cSYan, Zheng fail_srcu: 205676dda93cSYan, Zheng cleanup_srcu_struct(&fs_info->subvol_srcu); 20577e662854SQinghuang Feng fail: 205839279cc3SChris Mason kfree(extent_root); 205939279cc3SChris Mason kfree(tree_root); 206039279cc3SChris Mason kfree(fs_info); 206183afeac4SJim Meyering kfree(chunk_root); 206283afeac4SJim Meyering kfree(dev_root); 2063d20f7043SChris Mason kfree(csum_root); 206439279cc3SChris Mason return ERR_PTR(err); 2065eb60ceacSChris Mason } 2066eb60ceacSChris Mason 2067f2984462SChris Mason static void btrfs_end_buffer_write_sync(struct buffer_head *bh, int uptodate) 2068f2984462SChris Mason { 2069f2984462SChris Mason char b[BDEVNAME_SIZE]; 2070f2984462SChris Mason 2071f2984462SChris Mason if (uptodate) { 2072f2984462SChris Mason set_buffer_uptodate(bh); 2073f2984462SChris Mason } else { 2074f2984462SChris Mason if (!buffer_eopnotsupp(bh) && printk_ratelimit()) { 2075f2984462SChris Mason printk(KERN_WARNING "lost page write due to " 2076f2984462SChris Mason "I/O error on %s\n", 2077f2984462SChris Mason bdevname(bh->b_bdev, b)); 2078f2984462SChris Mason } 20791259ab75SChris Mason /* note, we dont' set_buffer_write_io_error because we have 20801259ab75SChris Mason * our own ways of dealing with the IO errors 20811259ab75SChris Mason */ 2082f2984462SChris Mason clear_buffer_uptodate(bh); 2083f2984462SChris Mason } 2084f2984462SChris Mason unlock_buffer(bh); 2085f2984462SChris Mason put_bh(bh); 2086f2984462SChris Mason } 2087f2984462SChris Mason 2088a512bbf8SYan Zheng struct buffer_head *btrfs_read_dev_super(struct block_device *bdev) 2089a512bbf8SYan Zheng { 2090a512bbf8SYan Zheng struct buffer_head *bh; 2091a512bbf8SYan Zheng struct buffer_head *latest = NULL; 2092a512bbf8SYan Zheng struct btrfs_super_block *super; 2093a512bbf8SYan Zheng int i; 2094a512bbf8SYan Zheng u64 transid = 0; 2095a512bbf8SYan Zheng u64 bytenr; 2096a512bbf8SYan Zheng 2097a512bbf8SYan Zheng /* we would like to check all the supers, but that would make 2098a512bbf8SYan Zheng * a btrfs mount succeed after a mkfs from a different FS. 2099a512bbf8SYan Zheng * So, we need to add a special mount option to scan for 2100a512bbf8SYan Zheng * later supers, using BTRFS_SUPER_MIRROR_MAX instead 2101a512bbf8SYan Zheng */ 2102a512bbf8SYan Zheng for (i = 0; i < 1; i++) { 2103a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 2104a512bbf8SYan Zheng if (bytenr + 4096 >= i_size_read(bdev->bd_inode)) 2105a512bbf8SYan Zheng break; 2106a512bbf8SYan Zheng bh = __bread(bdev, bytenr / 4096, 4096); 2107a512bbf8SYan Zheng if (!bh) 2108a512bbf8SYan Zheng continue; 2109a512bbf8SYan Zheng 2110a512bbf8SYan Zheng super = (struct btrfs_super_block *)bh->b_data; 2111a512bbf8SYan Zheng if (btrfs_super_bytenr(super) != bytenr || 2112a512bbf8SYan Zheng strncmp((char *)(&super->magic), BTRFS_MAGIC, 2113a512bbf8SYan Zheng sizeof(super->magic))) { 2114a512bbf8SYan Zheng brelse(bh); 2115a512bbf8SYan Zheng continue; 2116a512bbf8SYan Zheng } 2117a512bbf8SYan Zheng 2118a512bbf8SYan Zheng if (!latest || btrfs_super_generation(super) > transid) { 2119a512bbf8SYan Zheng brelse(latest); 2120a512bbf8SYan Zheng latest = bh; 2121a512bbf8SYan Zheng transid = btrfs_super_generation(super); 2122a512bbf8SYan Zheng } else { 2123a512bbf8SYan Zheng brelse(bh); 2124a512bbf8SYan Zheng } 2125a512bbf8SYan Zheng } 2126a512bbf8SYan Zheng return latest; 2127a512bbf8SYan Zheng } 2128a512bbf8SYan Zheng 21294eedeb75SHisashi Hifumi /* 21304eedeb75SHisashi Hifumi * this should be called twice, once with wait == 0 and 21314eedeb75SHisashi Hifumi * once with wait == 1. When wait == 0 is done, all the buffer heads 21324eedeb75SHisashi Hifumi * we write are pinned. 21334eedeb75SHisashi Hifumi * 21344eedeb75SHisashi Hifumi * They are released when wait == 1 is done. 21354eedeb75SHisashi Hifumi * max_mirrors must be the same for both runs, and it indicates how 21364eedeb75SHisashi Hifumi * many supers on this one device should be written. 21374eedeb75SHisashi Hifumi * 21384eedeb75SHisashi Hifumi * max_mirrors == 0 means to write them all. 21394eedeb75SHisashi Hifumi */ 2140a512bbf8SYan Zheng static int write_dev_supers(struct btrfs_device *device, 2141a512bbf8SYan Zheng struct btrfs_super_block *sb, 2142a512bbf8SYan Zheng int do_barriers, int wait, int max_mirrors) 2143a512bbf8SYan Zheng { 2144a512bbf8SYan Zheng struct buffer_head *bh; 2145a512bbf8SYan Zheng int i; 2146a512bbf8SYan Zheng int ret; 2147a512bbf8SYan Zheng int errors = 0; 2148a512bbf8SYan Zheng u32 crc; 2149a512bbf8SYan Zheng u64 bytenr; 2150a512bbf8SYan Zheng int last_barrier = 0; 2151a512bbf8SYan Zheng 2152a512bbf8SYan Zheng if (max_mirrors == 0) 2153a512bbf8SYan Zheng max_mirrors = BTRFS_SUPER_MIRROR_MAX; 2154a512bbf8SYan Zheng 2155a512bbf8SYan Zheng /* make sure only the last submit_bh does a barrier */ 2156a512bbf8SYan Zheng if (do_barriers) { 2157a512bbf8SYan Zheng for (i = 0; i < max_mirrors; i++) { 2158a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 2159a512bbf8SYan Zheng if (bytenr + BTRFS_SUPER_INFO_SIZE >= 2160a512bbf8SYan Zheng device->total_bytes) 2161a512bbf8SYan Zheng break; 2162a512bbf8SYan Zheng last_barrier = i; 2163a512bbf8SYan Zheng } 2164a512bbf8SYan Zheng } 2165a512bbf8SYan Zheng 2166a512bbf8SYan Zheng for (i = 0; i < max_mirrors; i++) { 2167a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 2168a512bbf8SYan Zheng if (bytenr + BTRFS_SUPER_INFO_SIZE >= device->total_bytes) 2169a512bbf8SYan Zheng break; 2170a512bbf8SYan Zheng 2171a512bbf8SYan Zheng if (wait) { 2172a512bbf8SYan Zheng bh = __find_get_block(device->bdev, bytenr / 4096, 2173a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE); 2174a512bbf8SYan Zheng BUG_ON(!bh); 2175a512bbf8SYan Zheng wait_on_buffer(bh); 21764eedeb75SHisashi Hifumi if (!buffer_uptodate(bh)) 21774eedeb75SHisashi Hifumi errors++; 21784eedeb75SHisashi Hifumi 21794eedeb75SHisashi Hifumi /* drop our reference */ 21804eedeb75SHisashi Hifumi brelse(bh); 21814eedeb75SHisashi Hifumi 21824eedeb75SHisashi Hifumi /* drop the reference from the wait == 0 run */ 2183a512bbf8SYan Zheng brelse(bh); 2184a512bbf8SYan Zheng continue; 2185a512bbf8SYan Zheng } else { 2186a512bbf8SYan Zheng btrfs_set_super_bytenr(sb, bytenr); 2187a512bbf8SYan Zheng 2188a512bbf8SYan Zheng crc = ~(u32)0; 2189a512bbf8SYan Zheng crc = btrfs_csum_data(NULL, (char *)sb + 2190a512bbf8SYan Zheng BTRFS_CSUM_SIZE, crc, 2191a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE - 2192a512bbf8SYan Zheng BTRFS_CSUM_SIZE); 2193a512bbf8SYan Zheng btrfs_csum_final(crc, sb->csum); 2194a512bbf8SYan Zheng 21954eedeb75SHisashi Hifumi /* 21964eedeb75SHisashi Hifumi * one reference for us, and we leave it for the 21974eedeb75SHisashi Hifumi * caller 21984eedeb75SHisashi Hifumi */ 2199a512bbf8SYan Zheng bh = __getblk(device->bdev, bytenr / 4096, 2200a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE); 2201a512bbf8SYan Zheng memcpy(bh->b_data, sb, BTRFS_SUPER_INFO_SIZE); 2202a512bbf8SYan Zheng 22034eedeb75SHisashi Hifumi /* one reference for submit_bh */ 2204a512bbf8SYan Zheng get_bh(bh); 22054eedeb75SHisashi Hifumi 22064eedeb75SHisashi Hifumi set_buffer_uptodate(bh); 2207a512bbf8SYan Zheng lock_buffer(bh); 2208a512bbf8SYan Zheng bh->b_end_io = btrfs_end_buffer_write_sync; 2209a512bbf8SYan Zheng } 2210a512bbf8SYan Zheng 2211a512bbf8SYan Zheng if (i == last_barrier && do_barriers && device->barriers) { 2212a512bbf8SYan Zheng ret = submit_bh(WRITE_BARRIER, bh); 2213a512bbf8SYan Zheng if (ret == -EOPNOTSUPP) { 2214a512bbf8SYan Zheng printk("btrfs: disabling barriers on dev %s\n", 2215a512bbf8SYan Zheng device->name); 2216a512bbf8SYan Zheng set_buffer_uptodate(bh); 2217a512bbf8SYan Zheng device->barriers = 0; 22184eedeb75SHisashi Hifumi /* one reference for submit_bh */ 2219a512bbf8SYan Zheng get_bh(bh); 2220a512bbf8SYan Zheng lock_buffer(bh); 2221ffbd517dSChris Mason ret = submit_bh(WRITE_SYNC, bh); 2222a512bbf8SYan Zheng } 2223a512bbf8SYan Zheng } else { 2224ffbd517dSChris Mason ret = submit_bh(WRITE_SYNC, bh); 2225a512bbf8SYan Zheng } 2226a512bbf8SYan Zheng 22274eedeb75SHisashi Hifumi if (ret) 2228a512bbf8SYan Zheng errors++; 2229a512bbf8SYan Zheng } 2230a512bbf8SYan Zheng return errors < i ? 0 : -1; 2231a512bbf8SYan Zheng } 2232a512bbf8SYan Zheng 2233a512bbf8SYan Zheng int write_all_supers(struct btrfs_root *root, int max_mirrors) 2234f2984462SChris Mason { 2235e5e9a520SChris Mason struct list_head *head; 2236f2984462SChris Mason struct btrfs_device *dev; 2237a061fc8dSChris Mason struct btrfs_super_block *sb; 2238f2984462SChris Mason struct btrfs_dev_item *dev_item; 2239f2984462SChris Mason int ret; 2240f2984462SChris Mason int do_barriers; 2241a236aed1SChris Mason int max_errors; 2242a236aed1SChris Mason int total_errors = 0; 2243a061fc8dSChris Mason u64 flags; 2244f2984462SChris Mason 2245a236aed1SChris Mason max_errors = btrfs_super_num_devices(&root->fs_info->super_copy) - 1; 2246f2984462SChris Mason do_barriers = !btrfs_test_opt(root, NOBARRIER); 2247f2984462SChris Mason 2248a061fc8dSChris Mason sb = &root->fs_info->super_for_commit; 2249a061fc8dSChris Mason dev_item = &sb->dev_item; 2250e5e9a520SChris Mason 2251e5e9a520SChris Mason mutex_lock(&root->fs_info->fs_devices->device_list_mutex); 2252e5e9a520SChris Mason head = &root->fs_info->fs_devices->devices; 2253c6e30871SQinghuang Feng list_for_each_entry(dev, head, dev_list) { 2254dfe25020SChris Mason if (!dev->bdev) { 2255dfe25020SChris Mason total_errors++; 2256dfe25020SChris Mason continue; 2257dfe25020SChris Mason } 22582b82032cSYan Zheng if (!dev->in_fs_metadata || !dev->writeable) 2259dfe25020SChris Mason continue; 2260dfe25020SChris Mason 22612b82032cSYan Zheng btrfs_set_stack_device_generation(dev_item, 0); 2262a061fc8dSChris Mason btrfs_set_stack_device_type(dev_item, dev->type); 2263a061fc8dSChris Mason btrfs_set_stack_device_id(dev_item, dev->devid); 2264a061fc8dSChris Mason btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes); 2265a061fc8dSChris Mason btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used); 2266a061fc8dSChris Mason btrfs_set_stack_device_io_align(dev_item, dev->io_align); 2267a061fc8dSChris Mason btrfs_set_stack_device_io_width(dev_item, dev->io_width); 2268a061fc8dSChris Mason btrfs_set_stack_device_sector_size(dev_item, dev->sector_size); 2269a061fc8dSChris Mason memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE); 22702b82032cSYan Zheng memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE); 2271a512bbf8SYan Zheng 2272a061fc8dSChris Mason flags = btrfs_super_flags(sb); 2273a061fc8dSChris Mason btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN); 2274f2984462SChris Mason 2275a512bbf8SYan Zheng ret = write_dev_supers(dev, sb, do_barriers, 0, max_mirrors); 2276a236aed1SChris Mason if (ret) 2277a236aed1SChris Mason total_errors++; 2278f2984462SChris Mason } 2279a236aed1SChris Mason if (total_errors > max_errors) { 2280d397712bSChris Mason printk(KERN_ERR "btrfs: %d errors while writing supers\n", 2281d397712bSChris Mason total_errors); 2282a236aed1SChris Mason BUG(); 2283a236aed1SChris Mason } 2284f2984462SChris Mason 2285a512bbf8SYan Zheng total_errors = 0; 2286c6e30871SQinghuang Feng list_for_each_entry(dev, head, dev_list) { 2287dfe25020SChris Mason if (!dev->bdev) 2288dfe25020SChris Mason continue; 22892b82032cSYan Zheng if (!dev->in_fs_metadata || !dev->writeable) 2290dfe25020SChris Mason continue; 2291dfe25020SChris Mason 2292a512bbf8SYan Zheng ret = write_dev_supers(dev, sb, do_barriers, 1, max_mirrors); 2293a512bbf8SYan Zheng if (ret) 22941259ab75SChris Mason total_errors++; 2295f2984462SChris Mason } 2296e5e9a520SChris Mason mutex_unlock(&root->fs_info->fs_devices->device_list_mutex); 2297a236aed1SChris Mason if (total_errors > max_errors) { 2298d397712bSChris Mason printk(KERN_ERR "btrfs: %d errors while writing supers\n", 2299d397712bSChris Mason total_errors); 2300a236aed1SChris Mason BUG(); 2301a236aed1SChris Mason } 2302f2984462SChris Mason return 0; 2303f2984462SChris Mason } 2304f2984462SChris Mason 2305a512bbf8SYan Zheng int write_ctree_super(struct btrfs_trans_handle *trans, 2306a512bbf8SYan Zheng struct btrfs_root *root, int max_mirrors) 2307cfaa7295SChris Mason { 2308e66f709bSChris Mason int ret; 23092c90e5d6SChris Mason 2310a512bbf8SYan Zheng ret = write_all_supers(root, max_mirrors); 23115f39d397SChris Mason return ret; 2312cfaa7295SChris Mason } 2313cfaa7295SChris Mason 23145eda7b5eSChris Mason int btrfs_free_fs_root(struct btrfs_fs_info *fs_info, struct btrfs_root *root) 23152619ba1fSChris Mason { 23164df27c4dSYan, Zheng spin_lock(&fs_info->fs_roots_radix_lock); 23172619ba1fSChris Mason radix_tree_delete(&fs_info->fs_roots_radix, 23182619ba1fSChris Mason (unsigned long)root->root_key.objectid); 23194df27c4dSYan, Zheng spin_unlock(&fs_info->fs_roots_radix_lock); 232076dda93cSYan, Zheng 232176dda93cSYan, Zheng if (btrfs_root_refs(&root->root_item) == 0) 232276dda93cSYan, Zheng synchronize_srcu(&fs_info->subvol_srcu); 232376dda93cSYan, Zheng 23244df27c4dSYan, Zheng free_fs_root(root); 23254df27c4dSYan, Zheng return 0; 23264df27c4dSYan, Zheng } 23274df27c4dSYan, Zheng 23284df27c4dSYan, Zheng static void free_fs_root(struct btrfs_root *root) 23294df27c4dSYan, Zheng { 23304df27c4dSYan, Zheng WARN_ON(!RB_EMPTY_ROOT(&root->inode_tree)); 23313394e160SChris Mason if (root->anon_super.s_dev) { 23323394e160SChris Mason down_write(&root->anon_super.s_umount); 23333394e160SChris Mason kill_anon_super(&root->anon_super); 23343394e160SChris Mason } 23355f39d397SChris Mason free_extent_buffer(root->node); 23365f39d397SChris Mason free_extent_buffer(root->commit_root); 233758176a96SJosef Bacik kfree(root->name); 23382619ba1fSChris Mason kfree(root); 23392619ba1fSChris Mason } 23402619ba1fSChris Mason 234135b7e476SChris Mason static int del_fs_roots(struct btrfs_fs_info *fs_info) 23420f7d52f4SChris Mason { 23430f7d52f4SChris Mason int ret; 23440f7d52f4SChris Mason struct btrfs_root *gang[8]; 23450f7d52f4SChris Mason int i; 23460f7d52f4SChris Mason 234776dda93cSYan, Zheng while (!list_empty(&fs_info->dead_roots)) { 234876dda93cSYan, Zheng gang[0] = list_entry(fs_info->dead_roots.next, 234976dda93cSYan, Zheng struct btrfs_root, root_list); 235076dda93cSYan, Zheng list_del(&gang[0]->root_list); 235176dda93cSYan, Zheng 235276dda93cSYan, Zheng if (gang[0]->in_radix) { 235376dda93cSYan, Zheng btrfs_free_fs_root(fs_info, gang[0]); 235476dda93cSYan, Zheng } else { 235576dda93cSYan, Zheng free_extent_buffer(gang[0]->node); 235676dda93cSYan, Zheng free_extent_buffer(gang[0]->commit_root); 235776dda93cSYan, Zheng kfree(gang[0]); 235876dda93cSYan, Zheng } 235976dda93cSYan, Zheng } 236076dda93cSYan, Zheng 23610f7d52f4SChris Mason while (1) { 23620f7d52f4SChris Mason ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix, 23630f7d52f4SChris Mason (void **)gang, 0, 23640f7d52f4SChris Mason ARRAY_SIZE(gang)); 23650f7d52f4SChris Mason if (!ret) 23660f7d52f4SChris Mason break; 23672619ba1fSChris Mason for (i = 0; i < ret; i++) 23685eda7b5eSChris Mason btrfs_free_fs_root(fs_info, gang[i]); 23690f7d52f4SChris Mason } 23700f7d52f4SChris Mason return 0; 23710f7d52f4SChris Mason } 2372b4100d64SChris Mason 2373c146afadSYan Zheng int btrfs_cleanup_fs_roots(struct btrfs_fs_info *fs_info) 2374c146afadSYan Zheng { 2375c146afadSYan Zheng u64 root_objectid = 0; 2376c146afadSYan Zheng struct btrfs_root *gang[8]; 2377c146afadSYan Zheng int i; 2378c146afadSYan Zheng int ret; 2379c146afadSYan Zheng 2380c146afadSYan Zheng while (1) { 2381c146afadSYan Zheng ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix, 2382c146afadSYan Zheng (void **)gang, root_objectid, 2383c146afadSYan Zheng ARRAY_SIZE(gang)); 2384c146afadSYan Zheng if (!ret) 2385c146afadSYan Zheng break; 23865d4f98a2SYan Zheng 23875d4f98a2SYan Zheng root_objectid = gang[ret - 1]->root_key.objectid + 1; 2388c146afadSYan Zheng for (i = 0; i < ret; i++) { 2389c146afadSYan Zheng root_objectid = gang[i]->root_key.objectid; 2390c146afadSYan Zheng btrfs_orphan_cleanup(gang[i]); 2391c146afadSYan Zheng } 2392c146afadSYan Zheng root_objectid++; 2393c146afadSYan Zheng } 2394c146afadSYan Zheng return 0; 2395c146afadSYan Zheng } 2396c146afadSYan Zheng 2397c146afadSYan Zheng int btrfs_commit_super(struct btrfs_root *root) 2398c146afadSYan Zheng { 2399c146afadSYan Zheng struct btrfs_trans_handle *trans; 2400c146afadSYan Zheng int ret; 2401c146afadSYan Zheng 2402c146afadSYan Zheng mutex_lock(&root->fs_info->cleaner_mutex); 240324bbcf04SYan, Zheng btrfs_run_delayed_iputs(root); 2404c146afadSYan Zheng btrfs_clean_old_snapshots(root); 2405c146afadSYan Zheng mutex_unlock(&root->fs_info->cleaner_mutex); 2406c71bf099SYan, Zheng 2407c71bf099SYan, Zheng /* wait until ongoing cleanup work done */ 2408c71bf099SYan, Zheng down_write(&root->fs_info->cleanup_work_sem); 2409c71bf099SYan, Zheng up_write(&root->fs_info->cleanup_work_sem); 2410c71bf099SYan, Zheng 2411a22285a6SYan, Zheng trans = btrfs_join_transaction(root, 1); 2412c146afadSYan Zheng ret = btrfs_commit_transaction(trans, root); 2413c146afadSYan Zheng BUG_ON(ret); 2414c146afadSYan Zheng /* run commit again to drop the original snapshot */ 2415a22285a6SYan, Zheng trans = btrfs_join_transaction(root, 1); 2416c146afadSYan Zheng btrfs_commit_transaction(trans, root); 2417c146afadSYan Zheng ret = btrfs_write_and_wait_transaction(NULL, root); 2418c146afadSYan Zheng BUG_ON(ret); 2419c146afadSYan Zheng 2420a512bbf8SYan Zheng ret = write_ctree_super(NULL, root, 0); 2421c146afadSYan Zheng return ret; 2422c146afadSYan Zheng } 2423c146afadSYan Zheng 2424e20d96d6SChris Mason int close_ctree(struct btrfs_root *root) 2425eb60ceacSChris Mason { 24260f7d52f4SChris Mason struct btrfs_fs_info *fs_info = root->fs_info; 2427c146afadSYan Zheng int ret; 2428e089f05cSChris Mason 2429facda1e7SChris Mason fs_info->closing = 1; 2430a2135011SChris Mason smp_mb(); 2431a2135011SChris Mason 2432c146afadSYan Zheng if (!(fs_info->sb->s_flags & MS_RDONLY)) { 2433c146afadSYan Zheng ret = btrfs_commit_super(root); 2434d397712bSChris Mason if (ret) 2435d397712bSChris Mason printk(KERN_ERR "btrfs: commit super ret %d\n", ret); 2436c146afadSYan Zheng } 2437ed2ff2cbSChris Mason 24388929ecfaSYan, Zheng kthread_stop(root->fs_info->transaction_kthread); 24398929ecfaSYan, Zheng kthread_stop(root->fs_info->cleaner_kthread); 24408929ecfaSYan, Zheng 2441f25784b3SYan Zheng fs_info->closing = 2; 2442f25784b3SYan Zheng smp_mb(); 2443f25784b3SYan Zheng 2444b0c68f8bSChris Mason if (fs_info->delalloc_bytes) { 2445d397712bSChris Mason printk(KERN_INFO "btrfs: at unmount delalloc count %llu\n", 244621380931SJoel Becker (unsigned long long)fs_info->delalloc_bytes); 2447b0c68f8bSChris Mason } 244831153d81SYan Zheng if (fs_info->total_ref_cache_size) { 2449d397712bSChris Mason printk(KERN_INFO "btrfs: at umount reference cache size %llu\n", 2450d397712bSChris Mason (unsigned long long)fs_info->total_ref_cache_size); 245131153d81SYan Zheng } 245231153d81SYan Zheng 24535f39d397SChris Mason free_extent_buffer(fs_info->extent_root->node); 24545d4f98a2SYan Zheng free_extent_buffer(fs_info->extent_root->commit_root); 24555f39d397SChris Mason free_extent_buffer(fs_info->tree_root->node); 24565d4f98a2SYan Zheng free_extent_buffer(fs_info->tree_root->commit_root); 24570b86a832SChris Mason free_extent_buffer(root->fs_info->chunk_root->node); 24585d4f98a2SYan Zheng free_extent_buffer(root->fs_info->chunk_root->commit_root); 24590b86a832SChris Mason free_extent_buffer(root->fs_info->dev_root->node); 24605d4f98a2SYan Zheng free_extent_buffer(root->fs_info->dev_root->commit_root); 2461d20f7043SChris Mason free_extent_buffer(root->fs_info->csum_root->node); 24625d4f98a2SYan Zheng free_extent_buffer(root->fs_info->csum_root->commit_root); 2463d20f7043SChris Mason 24649078a3e1SChris Mason btrfs_free_block_groups(root->fs_info); 2465c146afadSYan Zheng 24660f7d52f4SChris Mason del_fs_roots(fs_info); 2467d10c5f31SChris Mason 2468c146afadSYan Zheng iput(fs_info->btree_inode); 24699ad6b7bcSChris Mason 247061d92c32SChris Mason btrfs_stop_workers(&fs_info->generic_worker); 2471247e743cSChris Mason btrfs_stop_workers(&fs_info->fixup_workers); 2472771ed689SChris Mason btrfs_stop_workers(&fs_info->delalloc_workers); 24738b712842SChris Mason btrfs_stop_workers(&fs_info->workers); 24748b712842SChris Mason btrfs_stop_workers(&fs_info->endio_workers); 2475d20f7043SChris Mason btrfs_stop_workers(&fs_info->endio_meta_workers); 2476cad321adSChris Mason btrfs_stop_workers(&fs_info->endio_meta_write_workers); 2477e6dcd2dcSChris Mason btrfs_stop_workers(&fs_info->endio_write_workers); 24781cc127b5SChris Mason btrfs_stop_workers(&fs_info->submit_workers); 2479d6bfde87SChris Mason 2480dfe25020SChris Mason btrfs_close_devices(fs_info->fs_devices); 24810b86a832SChris Mason btrfs_mapping_tree_free(&fs_info->mapping_tree); 2482b248a415SChris Mason 248304160088SChris Mason bdi_destroy(&fs_info->bdi); 248476dda93cSYan, Zheng cleanup_srcu_struct(&fs_info->subvol_srcu); 24850b86a832SChris Mason 24860f7d52f4SChris Mason kfree(fs_info->extent_root); 24870f7d52f4SChris Mason kfree(fs_info->tree_root); 24880b86a832SChris Mason kfree(fs_info->chunk_root); 24890b86a832SChris Mason kfree(fs_info->dev_root); 2490d20f7043SChris Mason kfree(fs_info->csum_root); 2491eb60ceacSChris Mason return 0; 2492eb60ceacSChris Mason } 2493eb60ceacSChris Mason 24941259ab75SChris Mason int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid) 2495ccd467d6SChris Mason { 24961259ab75SChris Mason int ret; 2497810191ffSChris Mason struct inode *btree_inode = buf->first_page->mapping->host; 24981259ab75SChris Mason 24992ac55d41SJosef Bacik ret = extent_buffer_uptodate(&BTRFS_I(btree_inode)->io_tree, buf, 25002ac55d41SJosef Bacik NULL); 25011259ab75SChris Mason if (!ret) 25021259ab75SChris Mason return ret; 25031259ab75SChris Mason 25041259ab75SChris Mason ret = verify_parent_transid(&BTRFS_I(btree_inode)->io_tree, buf, 25051259ab75SChris Mason parent_transid); 25061259ab75SChris Mason return !ret; 25075f39d397SChris Mason } 25086702ed49SChris Mason 25095f39d397SChris Mason int btrfs_set_buffer_uptodate(struct extent_buffer *buf) 25105f39d397SChris Mason { 2511810191ffSChris Mason struct inode *btree_inode = buf->first_page->mapping->host; 2512d1310b2eSChris Mason return set_extent_buffer_uptodate(&BTRFS_I(btree_inode)->io_tree, 25135f39d397SChris Mason buf); 25145f39d397SChris Mason } 25155f39d397SChris Mason 25165f39d397SChris Mason void btrfs_mark_buffer_dirty(struct extent_buffer *buf) 25175f39d397SChris Mason { 2518810191ffSChris Mason struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root; 25195f39d397SChris Mason u64 transid = btrfs_header_generation(buf); 25205f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 2521b9473439SChris Mason int was_dirty; 2522b4ce94deSChris Mason 2523b9447ef8SChris Mason btrfs_assert_tree_locked(buf); 2524ccd467d6SChris Mason if (transid != root->fs_info->generation) { 2525d397712bSChris Mason printk(KERN_CRIT "btrfs transid mismatch buffer %llu, " 2526d397712bSChris Mason "found %llu running %llu\n", 2527db94535dSChris Mason (unsigned long long)buf->start, 2528d397712bSChris Mason (unsigned long long)transid, 2529d397712bSChris Mason (unsigned long long)root->fs_info->generation); 2530ccd467d6SChris Mason WARN_ON(1); 2531ccd467d6SChris Mason } 2532b9473439SChris Mason was_dirty = set_extent_buffer_dirty(&BTRFS_I(btree_inode)->io_tree, 2533b9473439SChris Mason buf); 2534b9473439SChris Mason if (!was_dirty) { 2535b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 2536b9473439SChris Mason root->fs_info->dirty_metadata_bytes += buf->len; 2537b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 2538b9473439SChris Mason } 2539eb60ceacSChris Mason } 2540eb60ceacSChris Mason 2541d3c2fdcfSChris Mason void btrfs_btree_balance_dirty(struct btrfs_root *root, unsigned long nr) 254235b7e476SChris Mason { 2543188de649SChris Mason /* 2544188de649SChris Mason * looks as though older kernels can get into trouble with 2545188de649SChris Mason * this code, they end up stuck in balance_dirty_pages forever 2546188de649SChris Mason */ 2547d6bfde87SChris Mason u64 num_dirty; 2548771ed689SChris Mason unsigned long thresh = 32 * 1024 * 1024; 2549d6bfde87SChris Mason 25506933c02eSJens Axboe if (current->flags & PF_MEMALLOC) 2551d6bfde87SChris Mason return; 2552d6bfde87SChris Mason 2553585ad2c3SChris Mason num_dirty = root->fs_info->dirty_metadata_bytes; 2554585ad2c3SChris Mason 2555d6bfde87SChris Mason if (num_dirty > thresh) { 2556d3c2fdcfSChris Mason balance_dirty_pages_ratelimited_nr( 2557304fced6SChris Mason root->fs_info->btree_inode->i_mapping, 1); 255835b7e476SChris Mason } 2559188de649SChris Mason return; 2560d6bfde87SChris Mason } 25616b80053dSChris Mason 2562ca7a79adSChris Mason int btrfs_read_buffer(struct extent_buffer *buf, u64 parent_transid) 25636b80053dSChris Mason { 2564810191ffSChris Mason struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root; 2565ce9adaa5SChris Mason int ret; 2566ca7a79adSChris Mason ret = btree_read_extent_buffer_pages(root, buf, 0, parent_transid); 2567d397712bSChris Mason if (ret == 0) 2568b4ce94deSChris Mason set_bit(EXTENT_BUFFER_UPTODATE, &buf->bflags); 2569ce9adaa5SChris Mason return ret; 25706b80053dSChris Mason } 25710da5468fSChris Mason 25724bef0848SChris Mason int btree_lock_page_hook(struct page *page) 25734bef0848SChris Mason { 25744bef0848SChris Mason struct inode *inode = page->mapping->host; 2575b9473439SChris Mason struct btrfs_root *root = BTRFS_I(inode)->root; 25764bef0848SChris Mason struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree; 25774bef0848SChris Mason struct extent_buffer *eb; 25784bef0848SChris Mason unsigned long len; 25794bef0848SChris Mason u64 bytenr = page_offset(page); 25804bef0848SChris Mason 25814bef0848SChris Mason if (page->private == EXTENT_PAGE_PRIVATE) 25824bef0848SChris Mason goto out; 25834bef0848SChris Mason 25844bef0848SChris Mason len = page->private >> 2; 25854bef0848SChris Mason eb = find_extent_buffer(io_tree, bytenr, len, GFP_NOFS); 25864bef0848SChris Mason if (!eb) 25874bef0848SChris Mason goto out; 25884bef0848SChris Mason 25894bef0848SChris Mason btrfs_tree_lock(eb); 25904bef0848SChris Mason btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN); 2591b9473439SChris Mason 2592b9473439SChris Mason if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) { 2593b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 2594b9473439SChris Mason if (root->fs_info->dirty_metadata_bytes >= eb->len) 2595b9473439SChris Mason root->fs_info->dirty_metadata_bytes -= eb->len; 2596b9473439SChris Mason else 2597b9473439SChris Mason WARN_ON(1); 2598b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 2599b9473439SChris Mason } 2600b9473439SChris Mason 26014bef0848SChris Mason btrfs_tree_unlock(eb); 26024bef0848SChris Mason free_extent_buffer(eb); 26034bef0848SChris Mason out: 26044bef0848SChris Mason lock_page(page); 26054bef0848SChris Mason return 0; 26064bef0848SChris Mason } 26074bef0848SChris Mason 2608d1310b2eSChris Mason static struct extent_io_ops btree_extent_io_ops = { 26094bef0848SChris Mason .write_cache_pages_lock_hook = btree_lock_page_hook, 2610ce9adaa5SChris Mason .readpage_end_io_hook = btree_readpage_end_io_hook, 26110b86a832SChris Mason .submit_bio_hook = btree_submit_bio_hook, 2612239b14b3SChris Mason /* note we're sharing with inode.c for the merge bio hook */ 2613239b14b3SChris Mason .merge_bio_hook = btrfs_merge_bio_hook, 26140da5468fSChris Mason }; 2615