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> 304b4e25f2SChris Mason #include "compat.h" 31eb60ceacSChris Mason #include "ctree.h" 32eb60ceacSChris Mason #include "disk-io.h" 33e089f05cSChris Mason #include "transaction.h" 340f7d52f4SChris Mason #include "btrfs_inode.h" 350b86a832SChris Mason #include "volumes.h" 36db94535dSChris Mason #include "print-tree.h" 378b712842SChris Mason #include "async-thread.h" 38925baeddSChris Mason #include "locking.h" 39e02119d5SChris Mason #include "tree-log.h" 40fa9c0d79SChris Mason #include "free-space-cache.h" 41eb60ceacSChris Mason 42d1310b2eSChris Mason static struct extent_io_ops btree_extent_io_ops; 438b712842SChris Mason static void end_workqueue_fn(struct btrfs_work *work); 44ce9adaa5SChris Mason 45d352ac68SChris Mason /* 46d352ac68SChris Mason * end_io_wq structs are used to do processing in task context when an IO is 47d352ac68SChris Mason * complete. This is used during reads to verify checksums, and it is used 48d352ac68SChris Mason * by writes to insert metadata for new file extents after IO is complete. 49d352ac68SChris Mason */ 50ce9adaa5SChris Mason struct end_io_wq { 51ce9adaa5SChris Mason struct bio *bio; 52ce9adaa5SChris Mason bio_end_io_t *end_io; 53ce9adaa5SChris Mason void *private; 54ce9adaa5SChris Mason struct btrfs_fs_info *info; 55ce9adaa5SChris Mason int error; 5622c59948SChris Mason int metadata; 57ce9adaa5SChris Mason struct list_head list; 588b712842SChris Mason struct btrfs_work work; 59ce9adaa5SChris Mason }; 600da5468fSChris Mason 61d352ac68SChris Mason /* 62d352ac68SChris Mason * async submit bios are used to offload expensive checksumming 63d352ac68SChris Mason * onto the worker threads. They checksum file and metadata bios 64d352ac68SChris Mason * just before they are sent down the IO stack. 65d352ac68SChris Mason */ 6644b8bd7eSChris Mason struct async_submit_bio { 6744b8bd7eSChris Mason struct inode *inode; 6844b8bd7eSChris Mason struct bio *bio; 6944b8bd7eSChris Mason struct list_head list; 704a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_start; 714a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_done; 7244b8bd7eSChris Mason int rw; 7344b8bd7eSChris Mason int mirror_num; 74c8b97818SChris Mason unsigned long bio_flags; 758b712842SChris Mason struct btrfs_work work; 7644b8bd7eSChris Mason }; 7744b8bd7eSChris Mason 784008c04aSChris Mason /* These are used to set the lockdep class on the extent buffer locks. 794008c04aSChris Mason * The class is set by the readpage_end_io_hook after the buffer has 804008c04aSChris Mason * passed csum validation but before the pages are unlocked. 814008c04aSChris Mason * 824008c04aSChris Mason * The lockdep class is also set by btrfs_init_new_buffer on freshly 834008c04aSChris Mason * allocated blocks. 844008c04aSChris Mason * 854008c04aSChris Mason * The class is based on the level in the tree block, which allows lockdep 864008c04aSChris Mason * to know that lower nodes nest inside the locks of higher nodes. 874008c04aSChris Mason * 884008c04aSChris Mason * We also add a check to make sure the highest level of the tree is 894008c04aSChris Mason * the same as our lockdep setup here. If BTRFS_MAX_LEVEL changes, this 904008c04aSChris Mason * code needs update as well. 914008c04aSChris Mason */ 924008c04aSChris Mason #ifdef CONFIG_DEBUG_LOCK_ALLOC 934008c04aSChris Mason # if BTRFS_MAX_LEVEL != 8 944008c04aSChris Mason # error 954008c04aSChris Mason # endif 964008c04aSChris Mason static struct lock_class_key btrfs_eb_class[BTRFS_MAX_LEVEL + 1]; 974008c04aSChris Mason static const char *btrfs_eb_name[BTRFS_MAX_LEVEL + 1] = { 984008c04aSChris Mason /* leaf */ 994008c04aSChris Mason "btrfs-extent-00", 1004008c04aSChris Mason "btrfs-extent-01", 1014008c04aSChris Mason "btrfs-extent-02", 1024008c04aSChris Mason "btrfs-extent-03", 1034008c04aSChris Mason "btrfs-extent-04", 1044008c04aSChris Mason "btrfs-extent-05", 1054008c04aSChris Mason "btrfs-extent-06", 1064008c04aSChris Mason "btrfs-extent-07", 1074008c04aSChris Mason /* highest possible level */ 1084008c04aSChris Mason "btrfs-extent-08", 1094008c04aSChris Mason }; 1104008c04aSChris Mason #endif 1114008c04aSChris Mason 112d352ac68SChris Mason /* 113d352ac68SChris Mason * extents on the btree inode are pretty simple, there's one extent 114d352ac68SChris Mason * that covers the entire device 115d352ac68SChris Mason */ 116b2950863SChristoph Hellwig static struct extent_map *btree_get_extent(struct inode *inode, 117b2950863SChristoph Hellwig struct page *page, size_t page_offset, u64 start, u64 len, 1185f39d397SChris Mason int create) 1195f39d397SChris Mason { 1205f39d397SChris Mason struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 1215f39d397SChris Mason struct extent_map *em; 1225f39d397SChris Mason int ret; 1235f39d397SChris Mason 124d1310b2eSChris Mason spin_lock(&em_tree->lock); 125d1310b2eSChris Mason em = lookup_extent_mapping(em_tree, start, len); 126a061fc8dSChris Mason if (em) { 127a061fc8dSChris Mason em->bdev = 128a061fc8dSChris Mason BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev; 129d1310b2eSChris Mason spin_unlock(&em_tree->lock); 1305f39d397SChris Mason goto out; 131a061fc8dSChris Mason } 132a061fc8dSChris Mason spin_unlock(&em_tree->lock); 1337b13b7b1SChris Mason 1345f39d397SChris Mason em = alloc_extent_map(GFP_NOFS); 1355f39d397SChris Mason if (!em) { 1365f39d397SChris Mason em = ERR_PTR(-ENOMEM); 1375f39d397SChris Mason goto out; 1385f39d397SChris Mason } 1395f39d397SChris Mason em->start = 0; 1400afbaf8cSChris Mason em->len = (u64)-1; 141c8b97818SChris Mason em->block_len = (u64)-1; 1425f39d397SChris Mason em->block_start = 0; 143a061fc8dSChris Mason em->bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev; 144d1310b2eSChris Mason 145d1310b2eSChris Mason spin_lock(&em_tree->lock); 1465f39d397SChris Mason ret = add_extent_mapping(em_tree, em); 1475f39d397SChris Mason if (ret == -EEXIST) { 1480afbaf8cSChris Mason u64 failed_start = em->start; 1490afbaf8cSChris Mason u64 failed_len = em->len; 1500afbaf8cSChris Mason 1515f39d397SChris Mason free_extent_map(em); 1527b13b7b1SChris Mason em = lookup_extent_mapping(em_tree, start, len); 1530afbaf8cSChris Mason if (em) { 1547b13b7b1SChris Mason ret = 0; 1550afbaf8cSChris Mason } else { 1560afbaf8cSChris Mason em = lookup_extent_mapping(em_tree, failed_start, 1570afbaf8cSChris Mason failed_len); 1587b13b7b1SChris Mason ret = -EIO; 1590afbaf8cSChris Mason } 1605f39d397SChris Mason } else if (ret) { 1617b13b7b1SChris Mason free_extent_map(em); 1627b13b7b1SChris Mason em = NULL; 1635f39d397SChris Mason } 1647b13b7b1SChris Mason spin_unlock(&em_tree->lock); 1657b13b7b1SChris Mason 1667b13b7b1SChris Mason if (ret) 1677b13b7b1SChris Mason em = ERR_PTR(ret); 1685f39d397SChris Mason out: 1695f39d397SChris Mason return em; 1705f39d397SChris Mason } 1715f39d397SChris Mason 17219c00ddcSChris Mason u32 btrfs_csum_data(struct btrfs_root *root, char *data, u32 seed, size_t len) 17319c00ddcSChris Mason { 174163e783eSDavid Woodhouse return crc32c(seed, data, len); 17519c00ddcSChris Mason } 17619c00ddcSChris Mason 17719c00ddcSChris Mason void btrfs_csum_final(u32 crc, char *result) 17819c00ddcSChris Mason { 17919c00ddcSChris Mason *(__le32 *)result = ~cpu_to_le32(crc); 18019c00ddcSChris Mason } 18119c00ddcSChris Mason 182d352ac68SChris Mason /* 183d352ac68SChris Mason * compute the csum for a btree block, and either verify it or write it 184d352ac68SChris Mason * into the csum field of the block. 185d352ac68SChris Mason */ 18619c00ddcSChris Mason static int csum_tree_block(struct btrfs_root *root, struct extent_buffer *buf, 18719c00ddcSChris Mason int verify) 18819c00ddcSChris Mason { 189607d432dSJosef Bacik u16 csum_size = 190607d432dSJosef Bacik btrfs_super_csum_size(&root->fs_info->super_copy); 191607d432dSJosef Bacik char *result = NULL; 19219c00ddcSChris Mason unsigned long len; 19319c00ddcSChris Mason unsigned long cur_len; 19419c00ddcSChris Mason unsigned long offset = BTRFS_CSUM_SIZE; 19519c00ddcSChris Mason char *map_token = NULL; 19619c00ddcSChris Mason char *kaddr; 19719c00ddcSChris Mason unsigned long map_start; 19819c00ddcSChris Mason unsigned long map_len; 19919c00ddcSChris Mason int err; 20019c00ddcSChris Mason u32 crc = ~(u32)0; 201607d432dSJosef Bacik unsigned long inline_result; 20219c00ddcSChris Mason 20319c00ddcSChris Mason len = buf->len - offset; 20419c00ddcSChris Mason while (len > 0) { 20519c00ddcSChris Mason err = map_private_extent_buffer(buf, offset, 32, 20619c00ddcSChris Mason &map_token, &kaddr, 20719c00ddcSChris Mason &map_start, &map_len, KM_USER0); 208d397712bSChris Mason if (err) 20919c00ddcSChris Mason return 1; 21019c00ddcSChris Mason cur_len = min(len, map_len - (offset - map_start)); 21119c00ddcSChris Mason crc = btrfs_csum_data(root, kaddr + offset - map_start, 21219c00ddcSChris Mason crc, cur_len); 21319c00ddcSChris Mason len -= cur_len; 21419c00ddcSChris Mason offset += cur_len; 21519c00ddcSChris Mason unmap_extent_buffer(buf, map_token, KM_USER0); 21619c00ddcSChris Mason } 217607d432dSJosef Bacik if (csum_size > sizeof(inline_result)) { 218607d432dSJosef Bacik result = kzalloc(csum_size * sizeof(char), GFP_NOFS); 219607d432dSJosef Bacik if (!result) 220607d432dSJosef Bacik return 1; 221607d432dSJosef Bacik } else { 222607d432dSJosef Bacik result = (char *)&inline_result; 223607d432dSJosef Bacik } 224607d432dSJosef Bacik 22519c00ddcSChris Mason btrfs_csum_final(crc, result); 22619c00ddcSChris Mason 22719c00ddcSChris Mason if (verify) { 228607d432dSJosef Bacik if (memcmp_extent_buffer(buf, result, 0, csum_size)) { 229e4204dedSChris Mason u32 val; 230e4204dedSChris Mason u32 found = 0; 231607d432dSJosef Bacik memcpy(&found, result, csum_size); 232e4204dedSChris Mason 233607d432dSJosef Bacik read_extent_buffer(buf, &val, 0, csum_size); 234193f284dSChris Mason if (printk_ratelimit()) { 235193f284dSChris Mason printk(KERN_INFO "btrfs: %s checksum verify " 236193f284dSChris Mason "failed on %llu wanted %X found %X " 237193f284dSChris Mason "level %d\n", 23819c00ddcSChris Mason root->fs_info->sb->s_id, 239193f284dSChris Mason (unsigned long long)buf->start, val, found, 240193f284dSChris Mason btrfs_header_level(buf)); 241193f284dSChris Mason } 242607d432dSJosef Bacik if (result != (char *)&inline_result) 243607d432dSJosef Bacik kfree(result); 24419c00ddcSChris Mason return 1; 24519c00ddcSChris Mason } 24619c00ddcSChris Mason } else { 247607d432dSJosef Bacik write_extent_buffer(buf, result, 0, csum_size); 24819c00ddcSChris Mason } 249607d432dSJosef Bacik if (result != (char *)&inline_result) 250607d432dSJosef Bacik kfree(result); 25119c00ddcSChris Mason return 0; 25219c00ddcSChris Mason } 25319c00ddcSChris Mason 254d352ac68SChris Mason /* 255d352ac68SChris Mason * we can't consider a given block up to date unless the transid of the 256d352ac68SChris Mason * block matches the transid in the parent node's pointer. This is how we 257d352ac68SChris Mason * detect blocks that either didn't get written at all or got written 258d352ac68SChris Mason * in the wrong place. 259d352ac68SChris Mason */ 2601259ab75SChris Mason static int verify_parent_transid(struct extent_io_tree *io_tree, 2611259ab75SChris Mason struct extent_buffer *eb, u64 parent_transid) 2621259ab75SChris Mason { 2631259ab75SChris Mason int ret; 2641259ab75SChris Mason 2651259ab75SChris Mason if (!parent_transid || btrfs_header_generation(eb) == parent_transid) 2661259ab75SChris Mason return 0; 2671259ab75SChris Mason 2681259ab75SChris Mason lock_extent(io_tree, eb->start, eb->start + eb->len - 1, GFP_NOFS); 2691259ab75SChris Mason if (extent_buffer_uptodate(io_tree, eb) && 2701259ab75SChris Mason btrfs_header_generation(eb) == parent_transid) { 2711259ab75SChris Mason ret = 0; 2721259ab75SChris Mason goto out; 2731259ab75SChris Mason } 274193f284dSChris Mason if (printk_ratelimit()) { 275193f284dSChris Mason printk("parent transid verify failed on %llu wanted %llu " 276193f284dSChris Mason "found %llu\n", 2771259ab75SChris Mason (unsigned long long)eb->start, 2781259ab75SChris Mason (unsigned long long)parent_transid, 2791259ab75SChris Mason (unsigned long long)btrfs_header_generation(eb)); 280193f284dSChris Mason } 2811259ab75SChris Mason ret = 1; 2821259ab75SChris Mason clear_extent_buffer_uptodate(io_tree, eb); 28333958dc6SChris Mason out: 2841259ab75SChris Mason unlock_extent(io_tree, eb->start, eb->start + eb->len - 1, 2851259ab75SChris Mason GFP_NOFS); 2861259ab75SChris Mason return ret; 2871259ab75SChris Mason } 2881259ab75SChris Mason 289d352ac68SChris Mason /* 290d352ac68SChris Mason * helper to read a given tree block, doing retries as required when 291d352ac68SChris Mason * the checksums don't match and we have alternate mirrors to try. 292d352ac68SChris Mason */ 293f188591eSChris Mason static int btree_read_extent_buffer_pages(struct btrfs_root *root, 294f188591eSChris Mason struct extent_buffer *eb, 295ca7a79adSChris Mason u64 start, u64 parent_transid) 296f188591eSChris Mason { 297f188591eSChris Mason struct extent_io_tree *io_tree; 298f188591eSChris Mason int ret; 299f188591eSChris Mason int num_copies = 0; 300f188591eSChris Mason int mirror_num = 0; 301f188591eSChris Mason 302f188591eSChris Mason io_tree = &BTRFS_I(root->fs_info->btree_inode)->io_tree; 303f188591eSChris Mason while (1) { 304f188591eSChris Mason ret = read_extent_buffer_pages(io_tree, eb, start, 1, 305f188591eSChris Mason btree_get_extent, mirror_num); 3061259ab75SChris Mason if (!ret && 3071259ab75SChris Mason !verify_parent_transid(io_tree, eb, parent_transid)) 308f188591eSChris Mason return ret; 309d397712bSChris Mason 310f188591eSChris Mason num_copies = btrfs_num_copies(&root->fs_info->mapping_tree, 311f188591eSChris Mason eb->start, eb->len); 3124235298eSChris Mason if (num_copies == 1) 313f188591eSChris Mason return ret; 3144235298eSChris Mason 315f188591eSChris Mason mirror_num++; 3164235298eSChris Mason if (mirror_num > num_copies) 317f188591eSChris Mason return ret; 318f188591eSChris Mason } 319f188591eSChris Mason return -EIO; 320f188591eSChris Mason } 32119c00ddcSChris Mason 322d352ac68SChris Mason /* 323d397712bSChris Mason * checksum a dirty tree block before IO. This has extra checks to make sure 324d397712bSChris Mason * we only fill in the checksum field in the first page of a multi-page block 325d352ac68SChris Mason */ 326d397712bSChris Mason 327b2950863SChristoph Hellwig static int csum_dirty_buffer(struct btrfs_root *root, struct page *page) 32819c00ddcSChris Mason { 329d1310b2eSChris Mason struct extent_io_tree *tree; 33035ebb934SChris Mason u64 start = (u64)page->index << PAGE_CACHE_SHIFT; 33119c00ddcSChris Mason u64 found_start; 33219c00ddcSChris Mason int found_level; 33319c00ddcSChris Mason unsigned long len; 33419c00ddcSChris Mason struct extent_buffer *eb; 335f188591eSChris Mason int ret; 336f188591eSChris Mason 337d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 33819c00ddcSChris Mason 33919c00ddcSChris Mason if (page->private == EXTENT_PAGE_PRIVATE) 34019c00ddcSChris Mason goto out; 34119c00ddcSChris Mason if (!page->private) 34219c00ddcSChris Mason goto out; 34319c00ddcSChris Mason len = page->private >> 2; 344d397712bSChris Mason WARN_ON(len == 0); 345d397712bSChris Mason 34619c00ddcSChris Mason eb = alloc_extent_buffer(tree, start, len, page, GFP_NOFS); 347ca7a79adSChris Mason ret = btree_read_extent_buffer_pages(root, eb, start + PAGE_CACHE_SIZE, 348ca7a79adSChris Mason btrfs_header_generation(eb)); 349f188591eSChris Mason BUG_ON(ret); 35019c00ddcSChris Mason found_start = btrfs_header_bytenr(eb); 35119c00ddcSChris Mason if (found_start != start) { 35255c69072SChris Mason WARN_ON(1); 35355c69072SChris Mason goto err; 35455c69072SChris Mason } 35555c69072SChris Mason if (eb->first_page != page) { 35655c69072SChris Mason WARN_ON(1); 35755c69072SChris Mason goto err; 35855c69072SChris Mason } 35955c69072SChris Mason if (!PageUptodate(page)) { 36055c69072SChris Mason WARN_ON(1); 36155c69072SChris Mason goto err; 36219c00ddcSChris Mason } 36319c00ddcSChris Mason found_level = btrfs_header_level(eb); 3644bef0848SChris Mason 36519c00ddcSChris Mason csum_tree_block(root, eb, 0); 36655c69072SChris Mason err: 36719c00ddcSChris Mason free_extent_buffer(eb); 36819c00ddcSChris Mason out: 36919c00ddcSChris Mason return 0; 37019c00ddcSChris Mason } 37119c00ddcSChris Mason 3722b82032cSYan Zheng static int check_tree_block_fsid(struct btrfs_root *root, 3732b82032cSYan Zheng struct extent_buffer *eb) 3742b82032cSYan Zheng { 3752b82032cSYan Zheng struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices; 3762b82032cSYan Zheng u8 fsid[BTRFS_UUID_SIZE]; 3772b82032cSYan Zheng int ret = 1; 3782b82032cSYan Zheng 3792b82032cSYan Zheng read_extent_buffer(eb, fsid, (unsigned long)btrfs_header_fsid(eb), 3802b82032cSYan Zheng BTRFS_FSID_SIZE); 3812b82032cSYan Zheng while (fs_devices) { 3822b82032cSYan Zheng if (!memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE)) { 3832b82032cSYan Zheng ret = 0; 3842b82032cSYan Zheng break; 3852b82032cSYan Zheng } 3862b82032cSYan Zheng fs_devices = fs_devices->seed; 3872b82032cSYan Zheng } 3882b82032cSYan Zheng return ret; 3892b82032cSYan Zheng } 3902b82032cSYan Zheng 3914008c04aSChris Mason #ifdef CONFIG_DEBUG_LOCK_ALLOC 3924008c04aSChris Mason void btrfs_set_buffer_lockdep_class(struct extent_buffer *eb, int level) 3934008c04aSChris Mason { 3944008c04aSChris Mason lockdep_set_class_and_name(&eb->lock, 3954008c04aSChris Mason &btrfs_eb_class[level], 3964008c04aSChris Mason btrfs_eb_name[level]); 3974008c04aSChris Mason } 3984008c04aSChris Mason #endif 3994008c04aSChris Mason 400b2950863SChristoph Hellwig static int btree_readpage_end_io_hook(struct page *page, u64 start, u64 end, 401ce9adaa5SChris Mason struct extent_state *state) 402ce9adaa5SChris Mason { 403ce9adaa5SChris Mason struct extent_io_tree *tree; 404ce9adaa5SChris Mason u64 found_start; 405ce9adaa5SChris Mason int found_level; 406ce9adaa5SChris Mason unsigned long len; 407ce9adaa5SChris Mason struct extent_buffer *eb; 408ce9adaa5SChris Mason struct btrfs_root *root = BTRFS_I(page->mapping->host)->root; 409f188591eSChris Mason int ret = 0; 410ce9adaa5SChris Mason 411ce9adaa5SChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 412ce9adaa5SChris Mason if (page->private == EXTENT_PAGE_PRIVATE) 413ce9adaa5SChris Mason goto out; 414ce9adaa5SChris Mason if (!page->private) 415ce9adaa5SChris Mason goto out; 416d397712bSChris Mason 417ce9adaa5SChris Mason len = page->private >> 2; 418d397712bSChris Mason WARN_ON(len == 0); 419d397712bSChris Mason 420ce9adaa5SChris Mason eb = alloc_extent_buffer(tree, start, len, page, GFP_NOFS); 421f188591eSChris Mason 422ce9adaa5SChris Mason found_start = btrfs_header_bytenr(eb); 42323a07867SChris Mason if (found_start != start) { 424193f284dSChris Mason if (printk_ratelimit()) { 425193f284dSChris Mason printk(KERN_INFO "btrfs bad tree block start " 426193f284dSChris Mason "%llu %llu\n", 427a1b32a59SChris Mason (unsigned long long)found_start, 428a1b32a59SChris Mason (unsigned long long)eb->start); 429193f284dSChris Mason } 430f188591eSChris Mason ret = -EIO; 431ce9adaa5SChris Mason goto err; 432ce9adaa5SChris Mason } 433ce9adaa5SChris Mason if (eb->first_page != page) { 434d397712bSChris Mason printk(KERN_INFO "btrfs bad first page %lu %lu\n", 435d397712bSChris Mason eb->first_page->index, page->index); 436ce9adaa5SChris Mason WARN_ON(1); 437f188591eSChris Mason ret = -EIO; 438ce9adaa5SChris Mason goto err; 439ce9adaa5SChris Mason } 4402b82032cSYan Zheng if (check_tree_block_fsid(root, eb)) { 441193f284dSChris Mason if (printk_ratelimit()) { 442d397712bSChris Mason printk(KERN_INFO "btrfs bad fsid on block %llu\n", 443d397712bSChris Mason (unsigned long long)eb->start); 444193f284dSChris Mason } 4451259ab75SChris Mason ret = -EIO; 4461259ab75SChris Mason goto err; 4471259ab75SChris Mason } 448ce9adaa5SChris Mason found_level = btrfs_header_level(eb); 449ce9adaa5SChris Mason 4504008c04aSChris Mason btrfs_set_buffer_lockdep_class(eb, found_level); 4514008c04aSChris Mason 452ce9adaa5SChris Mason ret = csum_tree_block(root, eb, 1); 453f188591eSChris Mason if (ret) 454f188591eSChris Mason ret = -EIO; 455ce9adaa5SChris Mason 456ce9adaa5SChris Mason end = min_t(u64, eb->len, PAGE_CACHE_SIZE); 457ce9adaa5SChris Mason end = eb->start + end - 1; 458ce9adaa5SChris Mason err: 459ce9adaa5SChris Mason free_extent_buffer(eb); 460ce9adaa5SChris Mason out: 461f188591eSChris Mason return ret; 462ce9adaa5SChris Mason } 463ce9adaa5SChris Mason 464ce9adaa5SChris Mason static void end_workqueue_bio(struct bio *bio, int err) 465ce9adaa5SChris Mason { 466ce9adaa5SChris Mason struct end_io_wq *end_io_wq = bio->bi_private; 467ce9adaa5SChris Mason struct btrfs_fs_info *fs_info; 468ce9adaa5SChris Mason 469ce9adaa5SChris Mason fs_info = end_io_wq->info; 470ce9adaa5SChris Mason end_io_wq->error = err; 4718b712842SChris Mason end_io_wq->work.func = end_workqueue_fn; 4728b712842SChris Mason end_io_wq->work.flags = 0; 473d20f7043SChris Mason 474d20f7043SChris Mason if (bio->bi_rw & (1 << BIO_RW)) { 475cad321adSChris Mason if (end_io_wq->metadata) 476cad321adSChris Mason btrfs_queue_worker(&fs_info->endio_meta_write_workers, 477cad321adSChris Mason &end_io_wq->work); 478cad321adSChris Mason else 479e6dcd2dcSChris Mason btrfs_queue_worker(&fs_info->endio_write_workers, 480e6dcd2dcSChris Mason &end_io_wq->work); 481d20f7043SChris Mason } else { 482d20f7043SChris Mason if (end_io_wq->metadata) 483d20f7043SChris Mason btrfs_queue_worker(&fs_info->endio_meta_workers, 484d20f7043SChris Mason &end_io_wq->work); 485e6dcd2dcSChris Mason else 486d20f7043SChris Mason btrfs_queue_worker(&fs_info->endio_workers, 487d20f7043SChris Mason &end_io_wq->work); 488d20f7043SChris Mason } 489ce9adaa5SChris Mason } 490ce9adaa5SChris Mason 49122c59948SChris Mason int btrfs_bio_wq_end_io(struct btrfs_fs_info *info, struct bio *bio, 49222c59948SChris Mason int metadata) 4930b86a832SChris Mason { 494ce9adaa5SChris Mason struct end_io_wq *end_io_wq; 495ce9adaa5SChris Mason end_io_wq = kmalloc(sizeof(*end_io_wq), GFP_NOFS); 496ce9adaa5SChris Mason if (!end_io_wq) 497ce9adaa5SChris Mason return -ENOMEM; 498ce9adaa5SChris Mason 499ce9adaa5SChris Mason end_io_wq->private = bio->bi_private; 500ce9adaa5SChris Mason end_io_wq->end_io = bio->bi_end_io; 50122c59948SChris Mason end_io_wq->info = info; 502ce9adaa5SChris Mason end_io_wq->error = 0; 503ce9adaa5SChris Mason end_io_wq->bio = bio; 50422c59948SChris Mason end_io_wq->metadata = metadata; 505ce9adaa5SChris Mason 506ce9adaa5SChris Mason bio->bi_private = end_io_wq; 507ce9adaa5SChris Mason bio->bi_end_io = end_workqueue_bio; 50822c59948SChris Mason return 0; 50922c59948SChris Mason } 51022c59948SChris Mason 511b64a2851SChris Mason unsigned long btrfs_async_submit_limit(struct btrfs_fs_info *info) 5124854ddd0SChris Mason { 5134854ddd0SChris Mason unsigned long limit = min_t(unsigned long, 5144854ddd0SChris Mason info->workers.max_workers, 5154854ddd0SChris Mason info->fs_devices->open_devices); 5164854ddd0SChris Mason return 256 * limit; 5174854ddd0SChris Mason } 5184854ddd0SChris Mason 519777e6bd7SChris Mason int btrfs_congested_async(struct btrfs_fs_info *info, int iodone) 5200986fe9eSChris Mason { 521b64a2851SChris Mason return atomic_read(&info->nr_async_bios) > 522b64a2851SChris Mason btrfs_async_submit_limit(info); 5230986fe9eSChris Mason } 5240986fe9eSChris Mason 5254a69a410SChris Mason static void run_one_async_start(struct btrfs_work *work) 5264a69a410SChris Mason { 5274a69a410SChris Mason struct btrfs_fs_info *fs_info; 5284a69a410SChris Mason struct async_submit_bio *async; 5294a69a410SChris Mason 5304a69a410SChris Mason async = container_of(work, struct async_submit_bio, work); 5314a69a410SChris Mason fs_info = BTRFS_I(async->inode)->root->fs_info; 5324a69a410SChris Mason async->submit_bio_start(async->inode, async->rw, async->bio, 5334a69a410SChris Mason async->mirror_num, async->bio_flags); 5344a69a410SChris Mason } 5354a69a410SChris Mason 5364a69a410SChris Mason static void run_one_async_done(struct btrfs_work *work) 5378b712842SChris Mason { 5388b712842SChris Mason struct btrfs_fs_info *fs_info; 5398b712842SChris Mason struct async_submit_bio *async; 5404854ddd0SChris Mason int limit; 5418b712842SChris Mason 5428b712842SChris Mason async = container_of(work, struct async_submit_bio, work); 5438b712842SChris Mason fs_info = BTRFS_I(async->inode)->root->fs_info; 5444854ddd0SChris Mason 545b64a2851SChris Mason limit = btrfs_async_submit_limit(fs_info); 5464854ddd0SChris Mason limit = limit * 2 / 3; 5474854ddd0SChris Mason 5488b712842SChris Mason atomic_dec(&fs_info->nr_async_submits); 5490986fe9eSChris Mason 550b64a2851SChris Mason if (atomic_read(&fs_info->nr_async_submits) < limit && 551b64a2851SChris Mason waitqueue_active(&fs_info->async_submit_wait)) 5524854ddd0SChris Mason wake_up(&fs_info->async_submit_wait); 5534854ddd0SChris Mason 5544a69a410SChris Mason async->submit_bio_done(async->inode, async->rw, async->bio, 555c8b97818SChris Mason async->mirror_num, async->bio_flags); 5564a69a410SChris Mason } 5574a69a410SChris Mason 5584a69a410SChris Mason static void run_one_async_free(struct btrfs_work *work) 5594a69a410SChris Mason { 5604a69a410SChris Mason struct async_submit_bio *async; 5614a69a410SChris Mason 5624a69a410SChris Mason async = container_of(work, struct async_submit_bio, work); 5638b712842SChris Mason kfree(async); 5648b712842SChris Mason } 5658b712842SChris Mason 56644b8bd7eSChris Mason int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode, 56744b8bd7eSChris Mason int rw, struct bio *bio, int mirror_num, 568c8b97818SChris Mason unsigned long bio_flags, 5694a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_start, 5704a69a410SChris Mason extent_submit_bio_hook_t *submit_bio_done) 57144b8bd7eSChris Mason { 57244b8bd7eSChris Mason struct async_submit_bio *async; 57344b8bd7eSChris Mason 57444b8bd7eSChris Mason async = kmalloc(sizeof(*async), GFP_NOFS); 57544b8bd7eSChris Mason if (!async) 57644b8bd7eSChris Mason return -ENOMEM; 57744b8bd7eSChris Mason 57844b8bd7eSChris Mason async->inode = inode; 57944b8bd7eSChris Mason async->rw = rw; 58044b8bd7eSChris Mason async->bio = bio; 58144b8bd7eSChris Mason async->mirror_num = mirror_num; 5824a69a410SChris Mason async->submit_bio_start = submit_bio_start; 5834a69a410SChris Mason async->submit_bio_done = submit_bio_done; 5844a69a410SChris Mason 5854a69a410SChris Mason async->work.func = run_one_async_start; 5864a69a410SChris Mason async->work.ordered_func = run_one_async_done; 5874a69a410SChris Mason async->work.ordered_free = run_one_async_free; 5884a69a410SChris Mason 5898b712842SChris Mason async->work.flags = 0; 590c8b97818SChris Mason async->bio_flags = bio_flags; 5918c8bee1dSChris Mason 592cb03c743SChris Mason atomic_inc(&fs_info->nr_async_submits); 593d313d7a3SChris Mason 594d313d7a3SChris Mason if (rw & (1 << BIO_RW_SYNCIO)) 595d313d7a3SChris Mason btrfs_set_work_high_prio(&async->work); 596d313d7a3SChris Mason 5978b712842SChris Mason btrfs_queue_worker(&fs_info->workers, &async->work); 5989473f16cSChris Mason 599771ed689SChris Mason while (atomic_read(&fs_info->async_submit_draining) && 600771ed689SChris Mason atomic_read(&fs_info->nr_async_submits)) { 601771ed689SChris Mason wait_event(fs_info->async_submit_wait, 602771ed689SChris Mason (atomic_read(&fs_info->nr_async_submits) == 0)); 603771ed689SChris Mason } 604771ed689SChris Mason 60544b8bd7eSChris Mason return 0; 60644b8bd7eSChris Mason } 60744b8bd7eSChris Mason 608ce3ed71aSChris Mason static int btree_csum_one_bio(struct bio *bio) 609ce3ed71aSChris Mason { 610ce3ed71aSChris Mason struct bio_vec *bvec = bio->bi_io_vec; 611ce3ed71aSChris Mason int bio_index = 0; 612ce3ed71aSChris Mason struct btrfs_root *root; 613ce3ed71aSChris Mason 614ce3ed71aSChris Mason WARN_ON(bio->bi_vcnt <= 0); 615ce3ed71aSChris Mason while (bio_index < bio->bi_vcnt) { 616ce3ed71aSChris Mason root = BTRFS_I(bvec->bv_page->mapping->host)->root; 617ce3ed71aSChris Mason csum_dirty_buffer(root, bvec->bv_page); 618ce3ed71aSChris Mason bio_index++; 619ce3ed71aSChris Mason bvec++; 620ce3ed71aSChris Mason } 621ce3ed71aSChris Mason return 0; 622ce3ed71aSChris Mason } 623ce3ed71aSChris Mason 6244a69a410SChris Mason static int __btree_submit_bio_start(struct inode *inode, int rw, 6254a69a410SChris Mason struct bio *bio, int mirror_num, 6264a69a410SChris Mason unsigned long bio_flags) 62722c59948SChris Mason { 6288b712842SChris Mason /* 6298b712842SChris Mason * when we're called for a write, we're already in the async 6305443be45SChris Mason * submission context. Just jump into btrfs_map_bio 6318b712842SChris Mason */ 632ce3ed71aSChris Mason btree_csum_one_bio(bio); 6334a69a410SChris Mason return 0; 63422c59948SChris Mason } 63522c59948SChris Mason 6364a69a410SChris Mason static int __btree_submit_bio_done(struct inode *inode, int rw, struct bio *bio, 6374a69a410SChris Mason int mirror_num, unsigned long bio_flags) 6384a69a410SChris Mason { 6398b712842SChris Mason /* 6404a69a410SChris Mason * when we're called for a write, we're already in the async 6414a69a410SChris Mason * submission context. Just jump into btrfs_map_bio 6428b712842SChris Mason */ 6438b712842SChris Mason return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, mirror_num, 1); 6440b86a832SChris Mason } 6450b86a832SChris Mason 64644b8bd7eSChris Mason static int btree_submit_bio_hook(struct inode *inode, int rw, struct bio *bio, 647c8b97818SChris Mason int mirror_num, unsigned long bio_flags) 64844b8bd7eSChris Mason { 6494a69a410SChris Mason int ret; 650cad321adSChris Mason 6514a69a410SChris Mason ret = btrfs_bio_wq_end_io(BTRFS_I(inode)->root->fs_info, 6524a69a410SChris Mason bio, 1); 6534a69a410SChris Mason BUG_ON(ret); 6544a69a410SChris Mason 655cad321adSChris Mason if (!(rw & (1 << BIO_RW))) { 656cad321adSChris Mason /* 657cad321adSChris Mason * called for a read, do the setup so that checksum validation 658cad321adSChris Mason * can happen in the async kernel threads 659cad321adSChris Mason */ 6604a69a410SChris Mason return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, 6616f3577bdSChris Mason mirror_num, 0); 66244b8bd7eSChris Mason } 663d313d7a3SChris Mason 664cad321adSChris Mason /* 665cad321adSChris Mason * kthread helpers are used to submit writes so that checksumming 666cad321adSChris Mason * can happen in parallel across all CPUs 667cad321adSChris Mason */ 66844b8bd7eSChris Mason return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info, 669c8b97818SChris Mason inode, rw, bio, mirror_num, 0, 6704a69a410SChris Mason __btree_submit_bio_start, 6714a69a410SChris Mason __btree_submit_bio_done); 67244b8bd7eSChris Mason } 67344b8bd7eSChris Mason 6745f39d397SChris Mason static int btree_writepage(struct page *page, struct writeback_control *wbc) 6755f39d397SChris Mason { 676d1310b2eSChris Mason struct extent_io_tree *tree; 677b9473439SChris Mason struct btrfs_root *root = BTRFS_I(page->mapping->host)->root; 678b9473439SChris Mason struct extent_buffer *eb; 679b9473439SChris Mason int was_dirty; 6805443be45SChris Mason 681b9473439SChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 682b9473439SChris Mason if (!(current->flags & PF_MEMALLOC)) { 683b9473439SChris Mason return extent_write_full_page(tree, page, 684b9473439SChris Mason btree_get_extent, wbc); 685b9473439SChris Mason } 686b9473439SChris Mason 6875443be45SChris Mason redirty_page_for_writepage(wbc, page); 688b9473439SChris Mason eb = btrfs_find_tree_block(root, page_offset(page), 689b9473439SChris Mason PAGE_CACHE_SIZE); 690b9473439SChris Mason WARN_ON(!eb); 691b9473439SChris Mason 692b9473439SChris Mason was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags); 693b9473439SChris Mason if (!was_dirty) { 694b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 695b9473439SChris Mason root->fs_info->dirty_metadata_bytes += PAGE_CACHE_SIZE; 696b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 697b9473439SChris Mason } 698b9473439SChris Mason free_extent_buffer(eb); 699b9473439SChris Mason 7005443be45SChris Mason unlock_page(page); 7015443be45SChris Mason return 0; 7025443be45SChris Mason } 7030da5468fSChris Mason 7040da5468fSChris Mason static int btree_writepages(struct address_space *mapping, 7050da5468fSChris Mason struct writeback_control *wbc) 7060da5468fSChris Mason { 707d1310b2eSChris Mason struct extent_io_tree *tree; 708d1310b2eSChris Mason tree = &BTRFS_I(mapping->host)->io_tree; 709d8d5f3e1SChris Mason if (wbc->sync_mode == WB_SYNC_NONE) { 710b9473439SChris Mason struct btrfs_root *root = BTRFS_I(mapping->host)->root; 711793955bcSChris Mason u64 num_dirty; 71224ab9cd8SChris Mason unsigned long thresh = 32 * 1024 * 1024; 713448d640bSChris Mason 714448d640bSChris Mason if (wbc->for_kupdate) 715448d640bSChris Mason return 0; 716448d640bSChris Mason 717b9473439SChris Mason /* this is a bit racy, but that's ok */ 718b9473439SChris Mason num_dirty = root->fs_info->dirty_metadata_bytes; 719d397712bSChris Mason if (num_dirty < thresh) 720793955bcSChris Mason return 0; 721793955bcSChris Mason } 7220da5468fSChris Mason return extent_writepages(tree, mapping, btree_get_extent, wbc); 7230da5468fSChris Mason } 7240da5468fSChris Mason 725b2950863SChristoph Hellwig static int btree_readpage(struct file *file, struct page *page) 7265f39d397SChris Mason { 727d1310b2eSChris Mason struct extent_io_tree *tree; 728d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 7295f39d397SChris Mason return extent_read_full_page(tree, page, btree_get_extent); 7305f39d397SChris Mason } 7315f39d397SChris Mason 73270dec807SChris Mason static int btree_releasepage(struct page *page, gfp_t gfp_flags) 7335f39d397SChris Mason { 734d1310b2eSChris Mason struct extent_io_tree *tree; 735d1310b2eSChris Mason struct extent_map_tree *map; 7365f39d397SChris Mason int ret; 7375f39d397SChris Mason 73898509cfcSChris Mason if (PageWriteback(page) || PageDirty(page)) 73998509cfcSChris Mason return 0; 74098509cfcSChris Mason 741d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 742d1310b2eSChris Mason map = &BTRFS_I(page->mapping->host)->extent_tree; 7436af118ceSChris Mason 7447b13b7b1SChris Mason ret = try_release_extent_state(map, tree, page, gfp_flags); 745d397712bSChris Mason if (!ret) 7466af118ceSChris Mason return 0; 7476af118ceSChris Mason 7486af118ceSChris Mason ret = try_release_extent_buffer(tree, page); 7495f39d397SChris Mason if (ret == 1) { 7505f39d397SChris Mason ClearPagePrivate(page); 7515f39d397SChris Mason set_page_private(page, 0); 752d98237b3SChris Mason page_cache_release(page); 7535f39d397SChris Mason } 7546af118ceSChris Mason 755d98237b3SChris Mason return ret; 756d98237b3SChris Mason } 757d98237b3SChris Mason 7585f39d397SChris Mason static void btree_invalidatepage(struct page *page, unsigned long offset) 759d98237b3SChris Mason { 760d1310b2eSChris Mason struct extent_io_tree *tree; 761d1310b2eSChris Mason tree = &BTRFS_I(page->mapping->host)->io_tree; 7625f39d397SChris Mason extent_invalidatepage(tree, page, offset); 7635f39d397SChris Mason btree_releasepage(page, GFP_NOFS); 7649ad6b7bcSChris Mason if (PagePrivate(page)) { 765d397712bSChris Mason printk(KERN_WARNING "btrfs warning page private not zero " 766d397712bSChris Mason "on page %llu\n", (unsigned long long)page_offset(page)); 7679ad6b7bcSChris Mason ClearPagePrivate(page); 7689ad6b7bcSChris Mason set_page_private(page, 0); 7699ad6b7bcSChris Mason page_cache_release(page); 7709ad6b7bcSChris Mason } 771d98237b3SChris Mason } 772d98237b3SChris Mason 773d98237b3SChris Mason static struct address_space_operations btree_aops = { 774d98237b3SChris Mason .readpage = btree_readpage, 775d98237b3SChris Mason .writepage = btree_writepage, 7760da5468fSChris Mason .writepages = btree_writepages, 7775f39d397SChris Mason .releasepage = btree_releasepage, 7785f39d397SChris Mason .invalidatepage = btree_invalidatepage, 779d98237b3SChris Mason .sync_page = block_sync_page, 780d98237b3SChris Mason }; 781123abc88SChris Mason 782ca7a79adSChris Mason int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize, 783ca7a79adSChris Mason u64 parent_transid) 784090d1875SChris Mason { 7855f39d397SChris Mason struct extent_buffer *buf = NULL; 7865f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 787de428b63SChris Mason int ret = 0; 788090d1875SChris Mason 789db94535dSChris Mason buf = btrfs_find_create_tree_block(root, bytenr, blocksize); 7905f39d397SChris Mason if (!buf) 791090d1875SChris Mason return 0; 792d1310b2eSChris Mason read_extent_buffer_pages(&BTRFS_I(btree_inode)->io_tree, 793f188591eSChris Mason buf, 0, 0, btree_get_extent, 0); 7945f39d397SChris Mason free_extent_buffer(buf); 795de428b63SChris Mason return ret; 796090d1875SChris Mason } 797090d1875SChris Mason 7980999df54SChris Mason struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root, 7990999df54SChris Mason u64 bytenr, u32 blocksize) 8000999df54SChris Mason { 8010999df54SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 8020999df54SChris Mason struct extent_buffer *eb; 8030999df54SChris Mason eb = find_extent_buffer(&BTRFS_I(btree_inode)->io_tree, 8040999df54SChris Mason bytenr, blocksize, GFP_NOFS); 8050999df54SChris Mason return eb; 8060999df54SChris Mason } 8070999df54SChris Mason 8080999df54SChris Mason struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root, 8090999df54SChris Mason u64 bytenr, u32 blocksize) 8100999df54SChris Mason { 8110999df54SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 8120999df54SChris Mason struct extent_buffer *eb; 8130999df54SChris Mason 8140999df54SChris Mason eb = alloc_extent_buffer(&BTRFS_I(btree_inode)->io_tree, 8150999df54SChris Mason bytenr, blocksize, NULL, GFP_NOFS); 8160999df54SChris Mason return eb; 8170999df54SChris Mason } 8180999df54SChris Mason 8190999df54SChris Mason 820e02119d5SChris Mason int btrfs_write_tree_block(struct extent_buffer *buf) 821e02119d5SChris Mason { 822e02119d5SChris Mason return btrfs_fdatawrite_range(buf->first_page->mapping, buf->start, 82324ab9cd8SChris Mason buf->start + buf->len - 1, WB_SYNC_ALL); 824e02119d5SChris Mason } 825e02119d5SChris Mason 826e02119d5SChris Mason int btrfs_wait_tree_block_writeback(struct extent_buffer *buf) 827e02119d5SChris Mason { 828e02119d5SChris Mason return btrfs_wait_on_page_writeback_range(buf->first_page->mapping, 829e02119d5SChris Mason buf->start, buf->start + buf->len - 1); 830e02119d5SChris Mason } 831e02119d5SChris Mason 832db94535dSChris Mason struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr, 833ca7a79adSChris Mason u32 blocksize, u64 parent_transid) 834e20d96d6SChris Mason { 8355f39d397SChris Mason struct extent_buffer *buf = NULL; 8365f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 837d1310b2eSChris Mason struct extent_io_tree *io_tree; 83819c00ddcSChris Mason int ret; 83919c00ddcSChris Mason 840d1310b2eSChris Mason io_tree = &BTRFS_I(btree_inode)->io_tree; 841e20d96d6SChris Mason 842db94535dSChris Mason buf = btrfs_find_create_tree_block(root, bytenr, blocksize); 8435f39d397SChris Mason if (!buf) 844d98237b3SChris Mason return NULL; 845e4204dedSChris Mason 846ca7a79adSChris Mason ret = btree_read_extent_buffer_pages(root, buf, 0, parent_transid); 847ce9adaa5SChris Mason 848d397712bSChris Mason if (ret == 0) 849b4ce94deSChris Mason set_bit(EXTENT_BUFFER_UPTODATE, &buf->bflags); 8505f39d397SChris Mason return buf; 851ce9adaa5SChris Mason 852eb60ceacSChris Mason } 853eb60ceacSChris Mason 854e089f05cSChris Mason int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root, 8555f39d397SChris Mason struct extent_buffer *buf) 856ed2ff2cbSChris Mason { 8575f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 85855c69072SChris Mason if (btrfs_header_generation(buf) == 859925baeddSChris Mason root->fs_info->running_transaction->transid) { 860b9447ef8SChris Mason btrfs_assert_tree_locked(buf); 861b4ce94deSChris Mason 862b9473439SChris Mason if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &buf->bflags)) { 863b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 864b9473439SChris Mason if (root->fs_info->dirty_metadata_bytes >= buf->len) 865b9473439SChris Mason root->fs_info->dirty_metadata_bytes -= buf->len; 866b9473439SChris Mason else 867b9473439SChris Mason WARN_ON(1); 868b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 869b9473439SChris Mason } 870b4ce94deSChris Mason 871b9473439SChris Mason /* ugh, clear_extent_buffer_dirty needs to lock the page */ 872b9473439SChris Mason btrfs_set_lock_blocking(buf); 873d1310b2eSChris Mason clear_extent_buffer_dirty(&BTRFS_I(btree_inode)->io_tree, 87455c69072SChris Mason buf); 875925baeddSChris Mason } 8765f39d397SChris Mason return 0; 8775f39d397SChris Mason } 8785f39d397SChris Mason 879db94535dSChris Mason static int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize, 88087ee04ebSChris Mason u32 stripesize, struct btrfs_root *root, 8819f5fae2fSChris Mason struct btrfs_fs_info *fs_info, 882e20d96d6SChris Mason u64 objectid) 883d97e63b6SChris Mason { 884cfaa7295SChris Mason root->node = NULL; 885a28ec197SChris Mason root->commit_root = NULL; 886db94535dSChris Mason root->sectorsize = sectorsize; 887db94535dSChris Mason root->nodesize = nodesize; 888db94535dSChris Mason root->leafsize = leafsize; 88987ee04ebSChris Mason root->stripesize = stripesize; 890123abc88SChris Mason root->ref_cows = 0; 8910b86a832SChris Mason root->track_dirty = 0; 8920b86a832SChris Mason 8939f5fae2fSChris Mason root->fs_info = fs_info; 8940f7d52f4SChris Mason root->objectid = objectid; 8950f7d52f4SChris Mason root->last_trans = 0; 8961b05da2eSChris Mason root->highest_inode = 0; 8971b05da2eSChris Mason root->last_inode_alloc = 0; 89858176a96SJosef Bacik root->name = NULL; 8994313b399SChris Mason root->in_sysfs = 0; 9005d4f98a2SYan Zheng root->inode_tree.rb_node = NULL; 9010b86a832SChris Mason 9020b86a832SChris Mason INIT_LIST_HEAD(&root->dirty_list); 9037b128766SJosef Bacik INIT_LIST_HEAD(&root->orphan_list); 9045d4f98a2SYan Zheng INIT_LIST_HEAD(&root->root_list); 905925baeddSChris Mason spin_lock_init(&root->node_lock); 906bcc63abbSYan spin_lock_init(&root->list_lock); 9075d4f98a2SYan Zheng spin_lock_init(&root->inode_lock); 908a2135011SChris Mason mutex_init(&root->objectid_mutex); 909e02119d5SChris Mason mutex_init(&root->log_mutex); 9107237f183SYan Zheng init_waitqueue_head(&root->log_writer_wait); 9117237f183SYan Zheng init_waitqueue_head(&root->log_commit_wait[0]); 9127237f183SYan Zheng init_waitqueue_head(&root->log_commit_wait[1]); 9137237f183SYan Zheng atomic_set(&root->log_commit[0], 0); 9147237f183SYan Zheng atomic_set(&root->log_commit[1], 0); 9157237f183SYan Zheng atomic_set(&root->log_writers, 0); 9167237f183SYan Zheng root->log_batch = 0; 9177237f183SYan Zheng root->log_transid = 0; 918d0c803c4SChris Mason extent_io_tree_init(&root->dirty_log_pages, 919d0c803c4SChris Mason fs_info->btree_inode->i_mapping, GFP_NOFS); 920017e5369SChris Mason 9213768f368SChris Mason memset(&root->root_key, 0, sizeof(root->root_key)); 9223768f368SChris Mason memset(&root->root_item, 0, sizeof(root->root_item)); 9236702ed49SChris Mason memset(&root->defrag_progress, 0, sizeof(root->defrag_progress)); 92458176a96SJosef Bacik memset(&root->root_kobj, 0, sizeof(root->root_kobj)); 9253f157a2fSChris Mason root->defrag_trans_start = fs_info->generation; 92658176a96SJosef Bacik init_completion(&root->kobj_unregister); 9276702ed49SChris Mason root->defrag_running = 0; 9286702ed49SChris Mason root->defrag_level = 0; 9294d775673SChris Mason root->root_key.objectid = objectid; 9303394e160SChris Mason root->anon_super.s_root = NULL; 9313394e160SChris Mason root->anon_super.s_dev = 0; 9323394e160SChris Mason INIT_LIST_HEAD(&root->anon_super.s_list); 9333394e160SChris Mason INIT_LIST_HEAD(&root->anon_super.s_instances); 9343394e160SChris Mason init_rwsem(&root->anon_super.s_umount); 9353394e160SChris Mason 9363768f368SChris Mason return 0; 9373768f368SChris Mason } 9383768f368SChris Mason 939db94535dSChris Mason static int find_and_setup_root(struct btrfs_root *tree_root, 9409f5fae2fSChris Mason struct btrfs_fs_info *fs_info, 9419f5fae2fSChris Mason u64 objectid, 942e20d96d6SChris Mason struct btrfs_root *root) 9433768f368SChris Mason { 9443768f368SChris Mason int ret; 945db94535dSChris Mason u32 blocksize; 94684234f3aSYan Zheng u64 generation; 9473768f368SChris Mason 948db94535dSChris Mason __setup_root(tree_root->nodesize, tree_root->leafsize, 94987ee04ebSChris Mason tree_root->sectorsize, tree_root->stripesize, 95087ee04ebSChris Mason root, fs_info, objectid); 9513768f368SChris Mason ret = btrfs_find_last_root(tree_root, objectid, 9523768f368SChris Mason &root->root_item, &root->root_key); 9533768f368SChris Mason BUG_ON(ret); 9543768f368SChris Mason 95584234f3aSYan Zheng generation = btrfs_root_generation(&root->root_item); 956db94535dSChris Mason blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item)); 957db94535dSChris Mason root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item), 95884234f3aSYan Zheng blocksize, generation); 9595d4f98a2SYan Zheng root->commit_root = btrfs_root_node(root); 9603768f368SChris Mason BUG_ON(!root->node); 961d97e63b6SChris Mason return 0; 962d97e63b6SChris Mason } 963d97e63b6SChris Mason 964e02119d5SChris Mason int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans, 965e02119d5SChris Mason struct btrfs_fs_info *fs_info) 966e02119d5SChris Mason { 967e02119d5SChris Mason struct extent_buffer *eb; 968d0c803c4SChris Mason struct btrfs_root *log_root_tree = fs_info->log_root_tree; 969d0c803c4SChris Mason u64 start = 0; 970d0c803c4SChris Mason u64 end = 0; 971e02119d5SChris Mason int ret; 972e02119d5SChris Mason 973d0c803c4SChris Mason if (!log_root_tree) 974e02119d5SChris Mason return 0; 975e02119d5SChris Mason 976d0c803c4SChris Mason while (1) { 977d0c803c4SChris Mason ret = find_first_extent_bit(&log_root_tree->dirty_log_pages, 978d0c803c4SChris Mason 0, &start, &end, EXTENT_DIRTY); 979d0c803c4SChris Mason if (ret) 980d0c803c4SChris Mason break; 981d0c803c4SChris Mason 982d0c803c4SChris Mason clear_extent_dirty(&log_root_tree->dirty_log_pages, 983d0c803c4SChris Mason start, end, GFP_NOFS); 984d0c803c4SChris Mason } 985e02119d5SChris Mason eb = fs_info->log_root_tree->node; 986e02119d5SChris Mason 987e02119d5SChris Mason WARN_ON(btrfs_header_level(eb) != 0); 988e02119d5SChris Mason WARN_ON(btrfs_header_nritems(eb) != 0); 989e02119d5SChris Mason 990d00aff00SChris Mason ret = btrfs_free_reserved_extent(fs_info->tree_root, 991d00aff00SChris Mason eb->start, eb->len); 992e02119d5SChris Mason BUG_ON(ret); 993e02119d5SChris Mason 994e02119d5SChris Mason free_extent_buffer(eb); 995e02119d5SChris Mason kfree(fs_info->log_root_tree); 996e02119d5SChris Mason fs_info->log_root_tree = NULL; 997e02119d5SChris Mason return 0; 998e02119d5SChris Mason } 999e02119d5SChris Mason 10007237f183SYan Zheng static struct btrfs_root *alloc_log_tree(struct btrfs_trans_handle *trans, 1001e02119d5SChris Mason struct btrfs_fs_info *fs_info) 10020f7d52f4SChris Mason { 10030f7d52f4SChris Mason struct btrfs_root *root; 10040f7d52f4SChris Mason struct btrfs_root *tree_root = fs_info->tree_root; 10057237f183SYan Zheng struct extent_buffer *leaf; 1006e02119d5SChris Mason 1007e02119d5SChris Mason root = kzalloc(sizeof(*root), GFP_NOFS); 1008e02119d5SChris Mason if (!root) 10097237f183SYan Zheng return ERR_PTR(-ENOMEM); 1010e02119d5SChris Mason 1011e02119d5SChris Mason __setup_root(tree_root->nodesize, tree_root->leafsize, 1012e02119d5SChris Mason tree_root->sectorsize, tree_root->stripesize, 1013e02119d5SChris Mason root, fs_info, BTRFS_TREE_LOG_OBJECTID); 1014e02119d5SChris Mason 1015e02119d5SChris Mason root->root_key.objectid = BTRFS_TREE_LOG_OBJECTID; 1016e02119d5SChris Mason root->root_key.type = BTRFS_ROOT_ITEM_KEY; 1017e02119d5SChris Mason root->root_key.offset = BTRFS_TREE_LOG_OBJECTID; 10187237f183SYan Zheng /* 10197237f183SYan Zheng * log trees do not get reference counted because they go away 10207237f183SYan Zheng * before a real commit is actually done. They do store pointers 10217237f183SYan Zheng * to file data extents, and those reference counts still get 10227237f183SYan Zheng * updated (along with back refs to the log tree). 10237237f183SYan Zheng */ 1024e02119d5SChris Mason root->ref_cows = 0; 1025e02119d5SChris Mason 10265d4f98a2SYan Zheng leaf = btrfs_alloc_free_block(trans, root, root->leafsize, 0, 10275d4f98a2SYan Zheng BTRFS_TREE_LOG_OBJECTID, NULL, 0, 0, 0); 10287237f183SYan Zheng if (IS_ERR(leaf)) { 10297237f183SYan Zheng kfree(root); 10307237f183SYan Zheng return ERR_CAST(leaf); 10317237f183SYan Zheng } 1032e02119d5SChris Mason 10335d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, 0, sizeof(struct btrfs_header)); 10345d4f98a2SYan Zheng btrfs_set_header_bytenr(leaf, leaf->start); 10355d4f98a2SYan Zheng btrfs_set_header_generation(leaf, trans->transid); 10365d4f98a2SYan Zheng btrfs_set_header_backref_rev(leaf, BTRFS_MIXED_BACKREF_REV); 10375d4f98a2SYan Zheng btrfs_set_header_owner(leaf, BTRFS_TREE_LOG_OBJECTID); 10387237f183SYan Zheng root->node = leaf; 1039e02119d5SChris Mason 1040e02119d5SChris Mason write_extent_buffer(root->node, root->fs_info->fsid, 1041e02119d5SChris Mason (unsigned long)btrfs_header_fsid(root->node), 1042e02119d5SChris Mason BTRFS_FSID_SIZE); 1043e02119d5SChris Mason btrfs_mark_buffer_dirty(root->node); 1044e02119d5SChris Mason btrfs_tree_unlock(root->node); 10457237f183SYan Zheng return root; 10467237f183SYan Zheng } 10477237f183SYan Zheng 10487237f183SYan Zheng int btrfs_init_log_root_tree(struct btrfs_trans_handle *trans, 10497237f183SYan Zheng struct btrfs_fs_info *fs_info) 10507237f183SYan Zheng { 10517237f183SYan Zheng struct btrfs_root *log_root; 10527237f183SYan Zheng 10537237f183SYan Zheng log_root = alloc_log_tree(trans, fs_info); 10547237f183SYan Zheng if (IS_ERR(log_root)) 10557237f183SYan Zheng return PTR_ERR(log_root); 10567237f183SYan Zheng WARN_ON(fs_info->log_root_tree); 10577237f183SYan Zheng fs_info->log_root_tree = log_root; 10587237f183SYan Zheng return 0; 10597237f183SYan Zheng } 10607237f183SYan Zheng 10617237f183SYan Zheng int btrfs_add_log_tree(struct btrfs_trans_handle *trans, 10627237f183SYan Zheng struct btrfs_root *root) 10637237f183SYan Zheng { 10647237f183SYan Zheng struct btrfs_root *log_root; 10657237f183SYan Zheng struct btrfs_inode_item *inode_item; 10667237f183SYan Zheng 10677237f183SYan Zheng log_root = alloc_log_tree(trans, root->fs_info); 10687237f183SYan Zheng if (IS_ERR(log_root)) 10697237f183SYan Zheng return PTR_ERR(log_root); 10707237f183SYan Zheng 10717237f183SYan Zheng log_root->last_trans = trans->transid; 10727237f183SYan Zheng log_root->root_key.offset = root->root_key.objectid; 10737237f183SYan Zheng 10747237f183SYan Zheng inode_item = &log_root->root_item.inode; 10757237f183SYan Zheng inode_item->generation = cpu_to_le64(1); 10767237f183SYan Zheng inode_item->size = cpu_to_le64(3); 10777237f183SYan Zheng inode_item->nlink = cpu_to_le32(1); 10787237f183SYan Zheng inode_item->nbytes = cpu_to_le64(root->leafsize); 10797237f183SYan Zheng inode_item->mode = cpu_to_le32(S_IFDIR | 0755); 10807237f183SYan Zheng 10815d4f98a2SYan Zheng btrfs_set_root_node(&log_root->root_item, log_root->node); 10827237f183SYan Zheng 10837237f183SYan Zheng WARN_ON(root->log_root); 10847237f183SYan Zheng root->log_root = log_root; 10857237f183SYan Zheng root->log_transid = 0; 1086e02119d5SChris Mason return 0; 1087e02119d5SChris Mason } 1088e02119d5SChris Mason 1089e02119d5SChris Mason struct btrfs_root *btrfs_read_fs_root_no_radix(struct btrfs_root *tree_root, 1090e02119d5SChris Mason struct btrfs_key *location) 1091e02119d5SChris Mason { 1092e02119d5SChris Mason struct btrfs_root *root; 1093e02119d5SChris Mason struct btrfs_fs_info *fs_info = tree_root->fs_info; 10940f7d52f4SChris Mason struct btrfs_path *path; 10955f39d397SChris Mason struct extent_buffer *l; 10961b05da2eSChris Mason u64 highest_inode; 109784234f3aSYan Zheng u64 generation; 1098db94535dSChris Mason u32 blocksize; 10990f7d52f4SChris Mason int ret = 0; 11000f7d52f4SChris Mason 11015eda7b5eSChris Mason root = kzalloc(sizeof(*root), GFP_NOFS); 11020cf6c620SChris Mason if (!root) 11030f7d52f4SChris Mason return ERR_PTR(-ENOMEM); 11040f7d52f4SChris Mason if (location->offset == (u64)-1) { 1105db94535dSChris Mason ret = find_and_setup_root(tree_root, fs_info, 11060f7d52f4SChris Mason location->objectid, root); 11070f7d52f4SChris Mason if (ret) { 11080f7d52f4SChris Mason kfree(root); 11090f7d52f4SChris Mason return ERR_PTR(ret); 11100f7d52f4SChris Mason } 11110f7d52f4SChris Mason goto insert; 11120f7d52f4SChris Mason } 11130f7d52f4SChris Mason 1114db94535dSChris Mason __setup_root(tree_root->nodesize, tree_root->leafsize, 111587ee04ebSChris Mason tree_root->sectorsize, tree_root->stripesize, 111687ee04ebSChris Mason root, fs_info, location->objectid); 11170f7d52f4SChris Mason 11180f7d52f4SChris Mason path = btrfs_alloc_path(); 11190f7d52f4SChris Mason BUG_ON(!path); 11200f7d52f4SChris Mason ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0); 11210f7d52f4SChris Mason if (ret != 0) { 11220f7d52f4SChris Mason if (ret > 0) 11230f7d52f4SChris Mason ret = -ENOENT; 11240f7d52f4SChris Mason goto out; 11250f7d52f4SChris Mason } 11265f39d397SChris Mason l = path->nodes[0]; 11275f39d397SChris Mason read_extent_buffer(l, &root->root_item, 11285f39d397SChris Mason btrfs_item_ptr_offset(l, path->slots[0]), 11290f7d52f4SChris Mason sizeof(root->root_item)); 113044b36eb2SYan Zheng memcpy(&root->root_key, location, sizeof(*location)); 11310f7d52f4SChris Mason ret = 0; 11320f7d52f4SChris Mason out: 11330f7d52f4SChris Mason btrfs_release_path(root, path); 11340f7d52f4SChris Mason btrfs_free_path(path); 11350f7d52f4SChris Mason if (ret) { 11360f7d52f4SChris Mason kfree(root); 11370f7d52f4SChris Mason return ERR_PTR(ret); 11380f7d52f4SChris Mason } 113984234f3aSYan Zheng generation = btrfs_root_generation(&root->root_item); 1140db94535dSChris Mason blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item)); 1141db94535dSChris Mason root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item), 114284234f3aSYan Zheng blocksize, generation); 11435d4f98a2SYan Zheng root->commit_root = btrfs_root_node(root); 11440f7d52f4SChris Mason BUG_ON(!root->node); 11450f7d52f4SChris Mason insert: 1146e02119d5SChris Mason if (location->objectid != BTRFS_TREE_LOG_OBJECTID) { 11470f7d52f4SChris Mason root->ref_cows = 1; 11485eda7b5eSChris Mason ret = btrfs_find_highest_inode(root, &highest_inode); 11495eda7b5eSChris Mason if (ret == 0) { 11505eda7b5eSChris Mason root->highest_inode = highest_inode; 11515eda7b5eSChris Mason root->last_inode_alloc = highest_inode; 11525eda7b5eSChris Mason } 1153e02119d5SChris Mason } 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; 1188edbd8d4eSChris Mason 11895eda7b5eSChris Mason root = radix_tree_lookup(&fs_info->fs_roots_radix, 11905eda7b5eSChris Mason (unsigned long)location->objectid); 11915eda7b5eSChris Mason if (root) 11925eda7b5eSChris Mason return root; 11935eda7b5eSChris Mason 1194e02119d5SChris Mason root = btrfs_read_fs_root_no_radix(fs_info->tree_root, location); 11955eda7b5eSChris Mason if (IS_ERR(root)) 11965eda7b5eSChris Mason return root; 11973394e160SChris Mason 11983394e160SChris Mason set_anon_super(&root->anon_super, NULL); 11993394e160SChris Mason 12002619ba1fSChris Mason ret = radix_tree_insert(&fs_info->fs_roots_radix, 12012619ba1fSChris Mason (unsigned long)root->root_key.objectid, 12020f7d52f4SChris Mason root); 12030f7d52f4SChris Mason if (ret) { 12045f39d397SChris Mason free_extent_buffer(root->node); 12050f7d52f4SChris Mason kfree(root); 12060f7d52f4SChris Mason return ERR_PTR(ret); 12070f7d52f4SChris Mason } 1208c146afadSYan Zheng if (!(fs_info->sb->s_flags & MS_RDONLY)) { 1209edbd8d4eSChris Mason ret = btrfs_find_dead_roots(fs_info->tree_root, 12105d4f98a2SYan Zheng root->root_key.objectid); 1211edbd8d4eSChris Mason BUG_ON(ret); 1212c146afadSYan Zheng btrfs_orphan_cleanup(root); 1213c146afadSYan Zheng } 1214edbd8d4eSChris Mason return root; 1215edbd8d4eSChris Mason } 1216edbd8d4eSChris Mason 1217edbd8d4eSChris Mason struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info, 1218edbd8d4eSChris Mason struct btrfs_key *location, 1219edbd8d4eSChris Mason const char *name, int namelen) 1220edbd8d4eSChris Mason { 1221edbd8d4eSChris Mason struct btrfs_root *root; 1222edbd8d4eSChris Mason int ret; 1223edbd8d4eSChris Mason 1224edbd8d4eSChris Mason root = btrfs_read_fs_root_no_name(fs_info, location); 1225edbd8d4eSChris Mason if (!root) 1226edbd8d4eSChris Mason return NULL; 122758176a96SJosef Bacik 12284313b399SChris Mason if (root->in_sysfs) 12294313b399SChris Mason return root; 12304313b399SChris Mason 123158176a96SJosef Bacik ret = btrfs_set_root_name(root, name, namelen); 123258176a96SJosef Bacik if (ret) { 12335f39d397SChris Mason free_extent_buffer(root->node); 123458176a96SJosef Bacik kfree(root); 123558176a96SJosef Bacik return ERR_PTR(ret); 123658176a96SJosef Bacik } 1237ea9e8b11SChris Mason #if 0 123858176a96SJosef Bacik ret = btrfs_sysfs_add_root(root); 123958176a96SJosef Bacik if (ret) { 12405f39d397SChris Mason free_extent_buffer(root->node); 124158176a96SJosef Bacik kfree(root->name); 124258176a96SJosef Bacik kfree(root); 124358176a96SJosef Bacik return ERR_PTR(ret); 124458176a96SJosef Bacik } 1245ea9e8b11SChris Mason #endif 12464313b399SChris Mason root->in_sysfs = 1; 12470f7d52f4SChris Mason return root; 12480f7d52f4SChris Mason } 124904160088SChris Mason 125004160088SChris Mason static int btrfs_congested_fn(void *congested_data, int bdi_bits) 125104160088SChris Mason { 125204160088SChris Mason struct btrfs_fs_info *info = (struct btrfs_fs_info *)congested_data; 125304160088SChris Mason int ret = 0; 125404160088SChris Mason struct btrfs_device *device; 125504160088SChris Mason struct backing_dev_info *bdi; 1256b7967db7SChris Mason 1257c6e30871SQinghuang Feng list_for_each_entry(device, &info->fs_devices->devices, dev_list) { 1258dfe25020SChris Mason if (!device->bdev) 1259dfe25020SChris Mason continue; 126004160088SChris Mason bdi = blk_get_backing_dev_info(device->bdev); 126104160088SChris Mason if (bdi && bdi_congested(bdi, bdi_bits)) { 126204160088SChris Mason ret = 1; 126304160088SChris Mason break; 126404160088SChris Mason } 126504160088SChris Mason } 126604160088SChris Mason return ret; 126704160088SChris Mason } 126804160088SChris Mason 126938b66988SChris Mason /* 127038b66988SChris Mason * this unplugs every device on the box, and it is only used when page 127138b66988SChris Mason * is null 127238b66988SChris Mason */ 127338b66988SChris Mason static void __unplug_io_fn(struct backing_dev_info *bdi, struct page *page) 127438b66988SChris Mason { 127538b66988SChris Mason struct btrfs_device *device; 127638b66988SChris Mason struct btrfs_fs_info *info; 127738b66988SChris Mason 127838b66988SChris Mason info = (struct btrfs_fs_info *)bdi->unplug_io_data; 1279c6e30871SQinghuang Feng list_for_each_entry(device, &info->fs_devices->devices, dev_list) { 1280d20f7043SChris Mason if (!device->bdev) 1281d20f7043SChris Mason continue; 1282d20f7043SChris Mason 128338b66988SChris Mason bdi = blk_get_backing_dev_info(device->bdev); 1284d397712bSChris Mason if (bdi->unplug_io_fn) 128538b66988SChris Mason bdi->unplug_io_fn(bdi, page); 128638b66988SChris Mason } 128738b66988SChris Mason } 128838b66988SChris Mason 1289b2950863SChristoph Hellwig static void btrfs_unplug_io_fn(struct backing_dev_info *bdi, struct page *page) 129004160088SChris Mason { 129138b66988SChris Mason struct inode *inode; 1292f2d8d74dSChris Mason struct extent_map_tree *em_tree; 1293f2d8d74dSChris Mason struct extent_map *em; 1294bcbfce8aSChris Mason struct address_space *mapping; 129538b66988SChris Mason u64 offset; 129638b66988SChris Mason 1297bcbfce8aSChris Mason /* the generic O_DIRECT read code does this */ 12989f0ba5bdSChris Mason if (1 || !page) { 129938b66988SChris Mason __unplug_io_fn(bdi, page); 130038b66988SChris Mason return; 130138b66988SChris Mason } 130238b66988SChris Mason 1303bcbfce8aSChris Mason /* 1304bcbfce8aSChris Mason * page->mapping may change at any time. Get a consistent copy 1305bcbfce8aSChris Mason * and use that for everything below 1306bcbfce8aSChris Mason */ 1307bcbfce8aSChris Mason smp_mb(); 1308bcbfce8aSChris Mason mapping = page->mapping; 1309bcbfce8aSChris Mason if (!mapping) 1310bcbfce8aSChris Mason return; 1311bcbfce8aSChris Mason 1312bcbfce8aSChris Mason inode = mapping->host; 1313240d5d48SChris Mason 1314240d5d48SChris Mason /* 1315240d5d48SChris Mason * don't do the expensive searching for a small number of 1316240d5d48SChris Mason * devices 1317240d5d48SChris Mason */ 1318240d5d48SChris Mason if (BTRFS_I(inode)->root->fs_info->fs_devices->open_devices <= 2) { 1319240d5d48SChris Mason __unplug_io_fn(bdi, page); 1320240d5d48SChris Mason return; 1321240d5d48SChris Mason } 1322240d5d48SChris Mason 132338b66988SChris Mason offset = page_offset(page); 132404160088SChris Mason 1325f2d8d74dSChris Mason em_tree = &BTRFS_I(inode)->extent_tree; 1326f2d8d74dSChris Mason spin_lock(&em_tree->lock); 1327f2d8d74dSChris Mason em = lookup_extent_mapping(em_tree, offset, PAGE_CACHE_SIZE); 1328f2d8d74dSChris Mason spin_unlock(&em_tree->lock); 132989642229SChris Mason if (!em) { 133089642229SChris Mason __unplug_io_fn(bdi, page); 1331f2d8d74dSChris Mason return; 133289642229SChris Mason } 1333f2d8d74dSChris Mason 133489642229SChris Mason if (em->block_start >= EXTENT_MAP_LAST_BYTE) { 133589642229SChris Mason free_extent_map(em); 133689642229SChris Mason __unplug_io_fn(bdi, page); 133789642229SChris Mason return; 133889642229SChris Mason } 1339f2d8d74dSChris Mason offset = offset - em->start; 1340f2d8d74dSChris Mason btrfs_unplug_page(&BTRFS_I(inode)->root->fs_info->mapping_tree, 1341f2d8d74dSChris Mason em->block_start + offset, page); 1342f2d8d74dSChris Mason free_extent_map(em); 134304160088SChris Mason } 134404160088SChris Mason 134504160088SChris Mason static int setup_bdi(struct btrfs_fs_info *info, struct backing_dev_info *bdi) 134604160088SChris Mason { 134704160088SChris Mason bdi_init(bdi); 13484575c9ccSChris Mason bdi->ra_pages = default_backing_dev_info.ra_pages; 134904160088SChris Mason bdi->state = 0; 135004160088SChris Mason bdi->capabilities = default_backing_dev_info.capabilities; 135104160088SChris Mason bdi->unplug_io_fn = btrfs_unplug_io_fn; 135204160088SChris Mason bdi->unplug_io_data = info; 135304160088SChris Mason bdi->congested_fn = btrfs_congested_fn; 135404160088SChris Mason bdi->congested_data = info; 135504160088SChris Mason return 0; 135604160088SChris Mason } 135704160088SChris Mason 1358ce9adaa5SChris Mason static int bio_ready_for_csum(struct bio *bio) 1359ce9adaa5SChris Mason { 1360ce9adaa5SChris Mason u64 length = 0; 1361ce9adaa5SChris Mason u64 buf_len = 0; 1362ce9adaa5SChris Mason u64 start = 0; 1363ce9adaa5SChris Mason struct page *page; 1364ce9adaa5SChris Mason struct extent_io_tree *io_tree = NULL; 1365ce9adaa5SChris Mason struct btrfs_fs_info *info = NULL; 1366ce9adaa5SChris Mason struct bio_vec *bvec; 1367ce9adaa5SChris Mason int i; 1368ce9adaa5SChris Mason int ret; 1369ce9adaa5SChris Mason 1370ce9adaa5SChris Mason bio_for_each_segment(bvec, bio, i) { 1371ce9adaa5SChris Mason page = bvec->bv_page; 1372ce9adaa5SChris Mason if (page->private == EXTENT_PAGE_PRIVATE) { 1373ce9adaa5SChris Mason length += bvec->bv_len; 1374ce9adaa5SChris Mason continue; 1375ce9adaa5SChris Mason } 1376ce9adaa5SChris Mason if (!page->private) { 1377ce9adaa5SChris Mason length += bvec->bv_len; 1378ce9adaa5SChris Mason continue; 1379ce9adaa5SChris Mason } 1380ce9adaa5SChris Mason length = bvec->bv_len; 1381ce9adaa5SChris Mason buf_len = page->private >> 2; 1382ce9adaa5SChris Mason start = page_offset(page) + bvec->bv_offset; 1383ce9adaa5SChris Mason io_tree = &BTRFS_I(page->mapping->host)->io_tree; 1384ce9adaa5SChris Mason info = BTRFS_I(page->mapping->host)->root->fs_info; 1385ce9adaa5SChris Mason } 1386ce9adaa5SChris Mason /* are we fully contained in this bio? */ 1387ce9adaa5SChris Mason if (buf_len <= length) 1388ce9adaa5SChris Mason return 1; 1389ce9adaa5SChris Mason 1390ce9adaa5SChris Mason ret = extent_range_uptodate(io_tree, start + length, 1391ce9adaa5SChris Mason start + buf_len - 1); 1392ce9adaa5SChris Mason return ret; 1393ce9adaa5SChris Mason } 1394ce9adaa5SChris Mason 13958b712842SChris Mason /* 13968b712842SChris Mason * called by the kthread helper functions to finally call the bio end_io 13978b712842SChris Mason * functions. This is where read checksum verification actually happens 13988b712842SChris Mason */ 13998b712842SChris Mason static void end_workqueue_fn(struct btrfs_work *work) 1400ce9adaa5SChris Mason { 1401ce9adaa5SChris Mason struct bio *bio; 14028b712842SChris Mason struct end_io_wq *end_io_wq; 14038b712842SChris Mason struct btrfs_fs_info *fs_info; 1404ce9adaa5SChris Mason int error; 1405ce9adaa5SChris Mason 14068b712842SChris Mason end_io_wq = container_of(work, struct end_io_wq, work); 1407ce9adaa5SChris Mason bio = end_io_wq->bio; 14088b712842SChris Mason fs_info = end_io_wq->info; 14098b712842SChris Mason 1410cad321adSChris Mason /* metadata bio reads are special because the whole tree block must 14118b712842SChris Mason * be checksummed at once. This makes sure the entire block is in 14128b712842SChris Mason * ram and up to date before trying to verify things. For 14138b712842SChris Mason * blocksize <= pagesize, it is basically a noop 14148b712842SChris Mason */ 1415cad321adSChris Mason if (!(bio->bi_rw & (1 << BIO_RW)) && end_io_wq->metadata && 1416cad321adSChris Mason !bio_ready_for_csum(bio)) { 1417d20f7043SChris Mason btrfs_queue_worker(&fs_info->endio_meta_workers, 14188b712842SChris Mason &end_io_wq->work); 1419ce9adaa5SChris Mason return; 1420ce9adaa5SChris Mason } 1421ce9adaa5SChris Mason error = end_io_wq->error; 1422ce9adaa5SChris Mason bio->bi_private = end_io_wq->private; 1423ce9adaa5SChris Mason bio->bi_end_io = end_io_wq->end_io; 1424ce9adaa5SChris Mason kfree(end_io_wq); 1425ce9adaa5SChris Mason bio_endio(bio, error); 1426ce9adaa5SChris Mason } 142744b8bd7eSChris Mason 1428a74a4b97SChris Mason static int cleaner_kthread(void *arg) 1429a74a4b97SChris Mason { 1430a74a4b97SChris Mason struct btrfs_root *root = arg; 1431a74a4b97SChris Mason 1432a74a4b97SChris Mason do { 1433a74a4b97SChris Mason smp_mb(); 1434a74a4b97SChris Mason if (root->fs_info->closing) 1435a74a4b97SChris Mason break; 1436a74a4b97SChris Mason 1437a74a4b97SChris Mason vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE); 1438a74a4b97SChris Mason mutex_lock(&root->fs_info->cleaner_mutex); 1439a74a4b97SChris Mason btrfs_clean_old_snapshots(root); 1440a74a4b97SChris Mason mutex_unlock(&root->fs_info->cleaner_mutex); 1441a74a4b97SChris Mason 1442a74a4b97SChris Mason if (freezing(current)) { 1443a74a4b97SChris Mason refrigerator(); 1444a74a4b97SChris Mason } else { 1445a74a4b97SChris Mason smp_mb(); 1446a74a4b97SChris Mason if (root->fs_info->closing) 1447a74a4b97SChris Mason break; 1448a74a4b97SChris Mason set_current_state(TASK_INTERRUPTIBLE); 1449a74a4b97SChris Mason schedule(); 1450a74a4b97SChris Mason __set_current_state(TASK_RUNNING); 1451a74a4b97SChris Mason } 1452a74a4b97SChris Mason } while (!kthread_should_stop()); 1453a74a4b97SChris Mason return 0; 1454a74a4b97SChris Mason } 1455a74a4b97SChris Mason 1456a74a4b97SChris Mason static int transaction_kthread(void *arg) 1457a74a4b97SChris Mason { 1458a74a4b97SChris Mason struct btrfs_root *root = arg; 1459a74a4b97SChris Mason struct btrfs_trans_handle *trans; 1460a74a4b97SChris Mason struct btrfs_transaction *cur; 1461a74a4b97SChris Mason unsigned long now; 1462a74a4b97SChris Mason unsigned long delay; 1463a74a4b97SChris Mason int ret; 1464a74a4b97SChris Mason 1465a74a4b97SChris Mason do { 1466a74a4b97SChris Mason smp_mb(); 1467a74a4b97SChris Mason if (root->fs_info->closing) 1468a74a4b97SChris Mason break; 1469a74a4b97SChris Mason 1470a74a4b97SChris Mason delay = HZ * 30; 1471a74a4b97SChris Mason vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE); 1472a74a4b97SChris Mason mutex_lock(&root->fs_info->transaction_kthread_mutex); 1473a74a4b97SChris Mason 1474a74a4b97SChris Mason mutex_lock(&root->fs_info->trans_mutex); 1475a74a4b97SChris Mason cur = root->fs_info->running_transaction; 1476a74a4b97SChris Mason if (!cur) { 1477a74a4b97SChris Mason mutex_unlock(&root->fs_info->trans_mutex); 1478a74a4b97SChris Mason goto sleep; 1479a74a4b97SChris Mason } 148031153d81SYan Zheng 1481a74a4b97SChris Mason now = get_seconds(); 1482a74a4b97SChris Mason if (now < cur->start_time || now - cur->start_time < 30) { 1483a74a4b97SChris Mason mutex_unlock(&root->fs_info->trans_mutex); 1484a74a4b97SChris Mason delay = HZ * 5; 1485a74a4b97SChris Mason goto sleep; 1486a74a4b97SChris Mason } 1487a74a4b97SChris Mason mutex_unlock(&root->fs_info->trans_mutex); 1488a74a4b97SChris Mason trans = btrfs_start_transaction(root, 1); 1489a74a4b97SChris Mason ret = btrfs_commit_transaction(trans, root); 149056bec294SChris Mason 1491a74a4b97SChris Mason sleep: 1492a74a4b97SChris Mason wake_up_process(root->fs_info->cleaner_kthread); 1493a74a4b97SChris Mason mutex_unlock(&root->fs_info->transaction_kthread_mutex); 1494a74a4b97SChris Mason 1495a74a4b97SChris Mason if (freezing(current)) { 1496a74a4b97SChris Mason refrigerator(); 1497a74a4b97SChris Mason } else { 1498a74a4b97SChris Mason if (root->fs_info->closing) 1499a74a4b97SChris Mason break; 1500a74a4b97SChris Mason set_current_state(TASK_INTERRUPTIBLE); 1501a74a4b97SChris Mason schedule_timeout(delay); 1502a74a4b97SChris Mason __set_current_state(TASK_RUNNING); 1503a74a4b97SChris Mason } 1504a74a4b97SChris Mason } while (!kthread_should_stop()); 1505a74a4b97SChris Mason return 0; 1506a74a4b97SChris Mason } 1507a74a4b97SChris Mason 15088a4b83ccSChris Mason struct btrfs_root *open_ctree(struct super_block *sb, 1509dfe25020SChris Mason struct btrfs_fs_devices *fs_devices, 1510dfe25020SChris Mason char *options) 1511eb60ceacSChris Mason { 1512db94535dSChris Mason u32 sectorsize; 1513db94535dSChris Mason u32 nodesize; 1514db94535dSChris Mason u32 leafsize; 1515db94535dSChris Mason u32 blocksize; 151687ee04ebSChris Mason u32 stripesize; 151784234f3aSYan Zheng u64 generation; 1518f2b636e8SJosef Bacik u64 features; 15193de4586cSChris Mason struct btrfs_key location; 1520a061fc8dSChris Mason struct buffer_head *bh; 1521e02119d5SChris Mason struct btrfs_root *extent_root = kzalloc(sizeof(struct btrfs_root), 1522e20d96d6SChris Mason GFP_NOFS); 1523d20f7043SChris Mason struct btrfs_root *csum_root = kzalloc(sizeof(struct btrfs_root), 1524d20f7043SChris Mason GFP_NOFS); 1525e02119d5SChris Mason struct btrfs_root *tree_root = kzalloc(sizeof(struct btrfs_root), 1526e20d96d6SChris Mason GFP_NOFS); 15278790d502SChris Mason struct btrfs_fs_info *fs_info = kzalloc(sizeof(*fs_info), 1528e20d96d6SChris Mason GFP_NOFS); 1529e02119d5SChris Mason struct btrfs_root *chunk_root = kzalloc(sizeof(struct btrfs_root), 15300b86a832SChris Mason GFP_NOFS); 1531e02119d5SChris Mason struct btrfs_root *dev_root = kzalloc(sizeof(struct btrfs_root), 15320b86a832SChris Mason GFP_NOFS); 1533e02119d5SChris Mason struct btrfs_root *log_tree_root; 1534e02119d5SChris Mason 1535eb60ceacSChris Mason int ret; 1536e58ca020SYan int err = -EINVAL; 15374543df7eSChris Mason 15382c90e5d6SChris Mason struct btrfs_super_block *disk_super; 15398790d502SChris Mason 15400463bb4eSJim Meyering if (!extent_root || !tree_root || !fs_info || 1541d20f7043SChris Mason !chunk_root || !dev_root || !csum_root) { 154239279cc3SChris Mason err = -ENOMEM; 154339279cc3SChris Mason goto fail; 154439279cc3SChris Mason } 15450f7d52f4SChris Mason INIT_RADIX_TREE(&fs_info->fs_roots_radix, GFP_NOFS); 15468fd17795SChris Mason INIT_LIST_HEAD(&fs_info->trans_list); 1547facda1e7SChris Mason INIT_LIST_HEAD(&fs_info->dead_roots); 154819c00ddcSChris Mason INIT_LIST_HEAD(&fs_info->hashers); 1549ea8c2819SChris Mason INIT_LIST_HEAD(&fs_info->delalloc_inodes); 15505a3f23d5SChris Mason INIT_LIST_HEAD(&fs_info->ordered_operations); 15511832a6d5SChris Mason spin_lock_init(&fs_info->delalloc_lock); 1552cee36a03SChris Mason spin_lock_init(&fs_info->new_trans_lock); 155331153d81SYan Zheng spin_lock_init(&fs_info->ref_cache_lock); 155419c00ddcSChris Mason 155558176a96SJosef Bacik init_completion(&fs_info->kobj_unregister); 15569f5fae2fSChris Mason fs_info->tree_root = tree_root; 15579f5fae2fSChris Mason fs_info->extent_root = extent_root; 1558d20f7043SChris Mason fs_info->csum_root = csum_root; 15590b86a832SChris Mason fs_info->chunk_root = chunk_root; 15600b86a832SChris Mason fs_info->dev_root = dev_root; 15618a4b83ccSChris Mason fs_info->fs_devices = fs_devices; 15620b86a832SChris Mason INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots); 15636324fbf3SChris Mason INIT_LIST_HEAD(&fs_info->space_info); 15640b86a832SChris Mason btrfs_mapping_init(&fs_info->mapping_tree); 1565cb03c743SChris Mason atomic_set(&fs_info->nr_async_submits, 0); 1566771ed689SChris Mason atomic_set(&fs_info->async_delalloc_pages, 0); 15678c8bee1dSChris Mason atomic_set(&fs_info->async_submit_draining, 0); 15680986fe9eSChris Mason atomic_set(&fs_info->nr_async_bios, 0); 1569e20d96d6SChris Mason fs_info->sb = sb; 1570c59f8951SChris Mason fs_info->max_extent = (u64)-1; 15716f568d35SChris Mason fs_info->max_inline = 8192 * 1024; 157204160088SChris Mason setup_bdi(fs_info, &fs_info->bdi); 1573d98237b3SChris Mason fs_info->btree_inode = new_inode(sb); 1574d98237b3SChris Mason fs_info->btree_inode->i_ino = 1; 15752c90e5d6SChris Mason fs_info->btree_inode->i_nlink = 1; 157697e728d4SJosef Bacik fs_info->metadata_ratio = 8; 1577c8b97818SChris Mason 1578b34b086cSChris Mason fs_info->thread_pool_size = min_t(unsigned long, 1579b34b086cSChris Mason num_online_cpus() + 2, 8); 15800afbaf8cSChris Mason 15813eaa2885SChris Mason INIT_LIST_HEAD(&fs_info->ordered_extents); 15823eaa2885SChris Mason spin_lock_init(&fs_info->ordered_extent_lock); 15833eaa2885SChris Mason 1584a061fc8dSChris Mason sb->s_blocksize = 4096; 1585a061fc8dSChris Mason sb->s_blocksize_bits = blksize_bits(4096); 1586a061fc8dSChris Mason 15870afbaf8cSChris Mason /* 15880afbaf8cSChris Mason * we set the i_size on the btree inode to the max possible int. 15890afbaf8cSChris Mason * the real end of the address space is determined by all of 15900afbaf8cSChris Mason * the devices in the system 15910afbaf8cSChris Mason */ 15920afbaf8cSChris Mason fs_info->btree_inode->i_size = OFFSET_MAX; 1593d98237b3SChris Mason fs_info->btree_inode->i_mapping->a_ops = &btree_aops; 159404160088SChris Mason fs_info->btree_inode->i_mapping->backing_dev_info = &fs_info->bdi; 159504160088SChris Mason 15965d4f98a2SYan Zheng RB_CLEAR_NODE(&BTRFS_I(fs_info->btree_inode)->rb_node); 1597d1310b2eSChris Mason extent_io_tree_init(&BTRFS_I(fs_info->btree_inode)->io_tree, 15985f39d397SChris Mason fs_info->btree_inode->i_mapping, 15995f39d397SChris Mason GFP_NOFS); 1600d1310b2eSChris Mason extent_map_tree_init(&BTRFS_I(fs_info->btree_inode)->extent_tree, 1601d1310b2eSChris Mason GFP_NOFS); 16020da5468fSChris Mason 1603d1310b2eSChris Mason BTRFS_I(fs_info->btree_inode)->io_tree.ops = &btree_extent_io_ops; 1604d1310b2eSChris Mason 16050f9dd46cSJosef Bacik spin_lock_init(&fs_info->block_group_cache_lock); 16060f9dd46cSJosef Bacik fs_info->block_group_cache_tree.rb_node = NULL; 16070f9dd46cSJosef Bacik 1608d1310b2eSChris Mason extent_io_tree_init(&fs_info->pinned_extents, 16091a5bc167SChris Mason fs_info->btree_inode->i_mapping, GFP_NOFS); 1610e66f709bSChris Mason fs_info->do_barriers = 1; 1611e18e4809SChris Mason 16120f7d52f4SChris Mason BTRFS_I(fs_info->btree_inode)->root = tree_root; 16130f7d52f4SChris Mason memset(&BTRFS_I(fs_info->btree_inode)->location, 0, 16140f7d52f4SChris Mason sizeof(struct btrfs_key)); 161522b0ebdaSChris Mason insert_inode_hash(fs_info->btree_inode); 161639279cc3SChris Mason 161779154b1bSChris Mason mutex_init(&fs_info->trans_mutex); 16185a3f23d5SChris Mason mutex_init(&fs_info->ordered_operations_mutex); 1619e02119d5SChris Mason mutex_init(&fs_info->tree_log_mutex); 1620a2135011SChris Mason mutex_init(&fs_info->drop_mutex); 1621925baeddSChris Mason mutex_init(&fs_info->chunk_mutex); 1622a74a4b97SChris Mason mutex_init(&fs_info->transaction_kthread_mutex); 1623a74a4b97SChris Mason mutex_init(&fs_info->cleaner_mutex); 16247d9eb12cSChris Mason mutex_init(&fs_info->volume_mutex); 16251a40e23bSZheng Yan mutex_init(&fs_info->tree_reloc_mutex); 1626fa9c0d79SChris Mason 1627fa9c0d79SChris Mason btrfs_init_free_cluster(&fs_info->meta_alloc_cluster); 1628fa9c0d79SChris Mason btrfs_init_free_cluster(&fs_info->data_alloc_cluster); 1629fa9c0d79SChris Mason 1630e6dcd2dcSChris Mason init_waitqueue_head(&fs_info->transaction_throttle); 1631f9295749SChris Mason init_waitqueue_head(&fs_info->transaction_wait); 16324854ddd0SChris Mason init_waitqueue_head(&fs_info->async_submit_wait); 16333768f368SChris Mason 16340b86a832SChris Mason __setup_root(4096, 4096, 4096, 4096, tree_root, 16352c90e5d6SChris Mason fs_info, BTRFS_ROOT_TREE_OBJECTID); 16367eccb903SChris Mason 1637d98237b3SChris Mason 1638a512bbf8SYan Zheng bh = btrfs_read_dev_super(fs_devices->latest_bdev); 1639a061fc8dSChris Mason if (!bh) 164039279cc3SChris Mason goto fail_iput; 164139279cc3SChris Mason 1642a061fc8dSChris Mason memcpy(&fs_info->super_copy, bh->b_data, sizeof(fs_info->super_copy)); 16432d69a0f8SYan Zheng memcpy(&fs_info->super_for_commit, &fs_info->super_copy, 16442d69a0f8SYan Zheng sizeof(fs_info->super_for_commit)); 1645a061fc8dSChris Mason brelse(bh); 16465f39d397SChris Mason 1647a061fc8dSChris Mason memcpy(fs_info->fsid, fs_info->super_copy.fsid, BTRFS_FSID_SIZE); 16480b86a832SChris Mason 16495f39d397SChris Mason disk_super = &fs_info->super_copy; 16500f7d52f4SChris Mason if (!btrfs_super_root(disk_super)) 1651c6e2bac1SJosef Bacik goto fail_iput; 16520f7d52f4SChris Mason 16532b82032cSYan Zheng ret = btrfs_parse_options(tree_root, options); 16542b82032cSYan Zheng if (ret) { 16552b82032cSYan Zheng err = ret; 1656c6e2bac1SJosef Bacik goto fail_iput; 16572b82032cSYan Zheng } 1658dfe25020SChris Mason 1659f2b636e8SJosef Bacik features = btrfs_super_incompat_flags(disk_super) & 1660f2b636e8SJosef Bacik ~BTRFS_FEATURE_INCOMPAT_SUPP; 1661f2b636e8SJosef Bacik if (features) { 1662f2b636e8SJosef Bacik printk(KERN_ERR "BTRFS: couldn't mount because of " 1663f2b636e8SJosef Bacik "unsupported optional features (%Lx).\n", 166421380931SJoel Becker (unsigned long long)features); 1665f2b636e8SJosef Bacik err = -EINVAL; 1666c6e2bac1SJosef Bacik goto fail_iput; 1667f2b636e8SJosef Bacik } 1668f2b636e8SJosef Bacik 16695d4f98a2SYan Zheng features = btrfs_super_incompat_flags(disk_super); 16705d4f98a2SYan Zheng if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) { 16715d4f98a2SYan Zheng features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF; 16725d4f98a2SYan Zheng btrfs_set_super_incompat_flags(disk_super, features); 16735d4f98a2SYan Zheng } 16745d4f98a2SYan Zheng 1675f2b636e8SJosef Bacik features = btrfs_super_compat_ro_flags(disk_super) & 1676f2b636e8SJosef Bacik ~BTRFS_FEATURE_COMPAT_RO_SUPP; 1677f2b636e8SJosef Bacik if (!(sb->s_flags & MS_RDONLY) && features) { 1678f2b636e8SJosef Bacik printk(KERN_ERR "BTRFS: couldn't mount RDWR because of " 1679f2b636e8SJosef Bacik "unsupported option features (%Lx).\n", 168021380931SJoel Becker (unsigned long long)features); 1681f2b636e8SJosef Bacik err = -EINVAL; 1682c6e2bac1SJosef Bacik goto fail_iput; 1683f2b636e8SJosef Bacik } 1684f2b636e8SJosef Bacik 16854543df7eSChris Mason /* 16864543df7eSChris Mason * we need to start all the end_io workers up front because the 16874543df7eSChris Mason * queue work function gets called at interrupt time, and so it 16884543df7eSChris Mason * cannot dynamically grow. 16894543df7eSChris Mason */ 16905443be45SChris Mason btrfs_init_workers(&fs_info->workers, "worker", 16915443be45SChris Mason fs_info->thread_pool_size); 1692c8b97818SChris Mason 1693771ed689SChris Mason btrfs_init_workers(&fs_info->delalloc_workers, "delalloc", 1694771ed689SChris Mason fs_info->thread_pool_size); 1695771ed689SChris Mason 16965443be45SChris Mason btrfs_init_workers(&fs_info->submit_workers, "submit", 1697b720d209SChris Mason min_t(u64, fs_devices->num_devices, 1698b720d209SChris Mason fs_info->thread_pool_size)); 169961b49440SChris Mason 170061b49440SChris Mason /* a higher idle thresh on the submit workers makes it much more 170161b49440SChris Mason * likely that bios will be send down in a sane order to the 170261b49440SChris Mason * devices 170361b49440SChris Mason */ 170461b49440SChris Mason fs_info->submit_workers.idle_thresh = 64; 170553863232SChris Mason 1706771ed689SChris Mason fs_info->workers.idle_thresh = 16; 17074a69a410SChris Mason fs_info->workers.ordered = 1; 170861b49440SChris Mason 1709771ed689SChris Mason fs_info->delalloc_workers.idle_thresh = 2; 1710771ed689SChris Mason fs_info->delalloc_workers.ordered = 1; 1711771ed689SChris Mason 17125443be45SChris Mason btrfs_init_workers(&fs_info->fixup_workers, "fixup", 1); 17135443be45SChris Mason btrfs_init_workers(&fs_info->endio_workers, "endio", 17145443be45SChris Mason fs_info->thread_pool_size); 1715d20f7043SChris Mason btrfs_init_workers(&fs_info->endio_meta_workers, "endio-meta", 1716d20f7043SChris Mason fs_info->thread_pool_size); 1717cad321adSChris Mason btrfs_init_workers(&fs_info->endio_meta_write_workers, 1718cad321adSChris Mason "endio-meta-write", fs_info->thread_pool_size); 17195443be45SChris Mason btrfs_init_workers(&fs_info->endio_write_workers, "endio-write", 1720e6dcd2dcSChris Mason fs_info->thread_pool_size); 172161b49440SChris Mason 172261b49440SChris Mason /* 172361b49440SChris Mason * endios are largely parallel and should have a very 172461b49440SChris Mason * low idle thresh 172561b49440SChris Mason */ 172661b49440SChris Mason fs_info->endio_workers.idle_thresh = 4; 1727b51912c9SChris Mason fs_info->endio_meta_workers.idle_thresh = 4; 1728b51912c9SChris Mason 17293a5f1d45SChris Mason fs_info->endio_write_workers.idle_thresh = 64; 1730cad321adSChris Mason fs_info->endio_meta_write_workers.idle_thresh = 64; 173161b49440SChris Mason 17324543df7eSChris Mason btrfs_start_workers(&fs_info->workers, 1); 17331cc127b5SChris Mason btrfs_start_workers(&fs_info->submit_workers, 1); 1734771ed689SChris Mason btrfs_start_workers(&fs_info->delalloc_workers, 1); 1735247e743cSChris Mason btrfs_start_workers(&fs_info->fixup_workers, 1); 17364543df7eSChris Mason btrfs_start_workers(&fs_info->endio_workers, fs_info->thread_pool_size); 1737d20f7043SChris Mason btrfs_start_workers(&fs_info->endio_meta_workers, 1738d20f7043SChris Mason fs_info->thread_pool_size); 1739cad321adSChris Mason btrfs_start_workers(&fs_info->endio_meta_write_workers, 1740cad321adSChris Mason fs_info->thread_pool_size); 1741e6dcd2dcSChris Mason btrfs_start_workers(&fs_info->endio_write_workers, 1742e6dcd2dcSChris Mason fs_info->thread_pool_size); 17434543df7eSChris Mason 17444575c9ccSChris Mason fs_info->bdi.ra_pages *= btrfs_super_num_devices(disk_super); 1745c8b97818SChris Mason fs_info->bdi.ra_pages = max(fs_info->bdi.ra_pages, 1746c8b97818SChris Mason 4 * 1024 * 1024 / PAGE_CACHE_SIZE); 17474575c9ccSChris Mason 1748db94535dSChris Mason nodesize = btrfs_super_nodesize(disk_super); 1749db94535dSChris Mason leafsize = btrfs_super_leafsize(disk_super); 1750db94535dSChris Mason sectorsize = btrfs_super_sectorsize(disk_super); 175187ee04ebSChris Mason stripesize = btrfs_super_stripesize(disk_super); 1752db94535dSChris Mason tree_root->nodesize = nodesize; 1753db94535dSChris Mason tree_root->leafsize = leafsize; 1754db94535dSChris Mason tree_root->sectorsize = sectorsize; 175587ee04ebSChris Mason tree_root->stripesize = stripesize; 1756a061fc8dSChris Mason 1757a061fc8dSChris Mason sb->s_blocksize = sectorsize; 1758a061fc8dSChris Mason sb->s_blocksize_bits = blksize_bits(sectorsize); 1759db94535dSChris Mason 176039279cc3SChris Mason if (strncmp((char *)(&disk_super->magic), BTRFS_MAGIC, 176139279cc3SChris Mason sizeof(disk_super->magic))) { 1762d397712bSChris Mason printk(KERN_INFO "btrfs: valid FS not found on %s\n", sb->s_id); 176339279cc3SChris Mason goto fail_sb_buffer; 176439279cc3SChris Mason } 176519c00ddcSChris Mason 1766925baeddSChris Mason mutex_lock(&fs_info->chunk_mutex); 1767e4404d6eSYan Zheng ret = btrfs_read_sys_array(tree_root); 1768925baeddSChris Mason mutex_unlock(&fs_info->chunk_mutex); 176984eed90fSChris Mason if (ret) { 1770d397712bSChris Mason printk(KERN_WARNING "btrfs: failed to read the system " 1771d397712bSChris Mason "array on %s\n", sb->s_id); 17725d4f98a2SYan Zheng goto fail_sb_buffer; 177384eed90fSChris Mason } 17740b86a832SChris Mason 17750b86a832SChris Mason blocksize = btrfs_level_size(tree_root, 17760b86a832SChris Mason btrfs_super_chunk_root_level(disk_super)); 177784234f3aSYan Zheng generation = btrfs_super_chunk_root_generation(disk_super); 17780b86a832SChris Mason 17790b86a832SChris Mason __setup_root(nodesize, leafsize, sectorsize, stripesize, 17800b86a832SChris Mason chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID); 17810b86a832SChris Mason 17820b86a832SChris Mason chunk_root->node = read_tree_block(chunk_root, 17830b86a832SChris Mason btrfs_super_chunk_root(disk_super), 178484234f3aSYan Zheng blocksize, generation); 17850b86a832SChris Mason BUG_ON(!chunk_root->node); 17865d4f98a2SYan Zheng btrfs_set_root_node(&chunk_root->root_item, chunk_root->node); 17875d4f98a2SYan Zheng chunk_root->commit_root = btrfs_root_node(chunk_root); 17880b86a832SChris Mason 1789e17cade2SChris Mason read_extent_buffer(chunk_root->node, fs_info->chunk_tree_uuid, 1790e17cade2SChris Mason (unsigned long)btrfs_header_chunk_tree_uuid(chunk_root->node), 1791e17cade2SChris Mason BTRFS_UUID_SIZE); 1792e17cade2SChris Mason 1793925baeddSChris Mason mutex_lock(&fs_info->chunk_mutex); 17940b86a832SChris Mason ret = btrfs_read_chunk_tree(chunk_root); 1795925baeddSChris Mason mutex_unlock(&fs_info->chunk_mutex); 17962b82032cSYan Zheng if (ret) { 1797d397712bSChris Mason printk(KERN_WARNING "btrfs: failed to read chunk tree on %s\n", 1798d397712bSChris Mason sb->s_id); 17992b82032cSYan Zheng goto fail_chunk_root; 18002b82032cSYan Zheng } 18010b86a832SChris Mason 1802dfe25020SChris Mason btrfs_close_extra_devices(fs_devices); 1803dfe25020SChris Mason 1804db94535dSChris Mason blocksize = btrfs_level_size(tree_root, 1805db94535dSChris Mason btrfs_super_root_level(disk_super)); 180684234f3aSYan Zheng generation = btrfs_super_generation(disk_super); 18070b86a832SChris Mason 1808e20d96d6SChris Mason tree_root->node = read_tree_block(tree_root, 1809db94535dSChris Mason btrfs_super_root(disk_super), 181084234f3aSYan Zheng blocksize, generation); 181139279cc3SChris Mason if (!tree_root->node) 18122b82032cSYan Zheng goto fail_chunk_root; 18135d4f98a2SYan Zheng btrfs_set_root_node(&tree_root->root_item, tree_root->node); 18145d4f98a2SYan Zheng tree_root->commit_root = btrfs_root_node(tree_root); 1815db94535dSChris Mason 1816db94535dSChris Mason ret = find_and_setup_root(tree_root, fs_info, 1817e20d96d6SChris Mason BTRFS_EXTENT_TREE_OBJECTID, extent_root); 18180b86a832SChris Mason if (ret) 181939279cc3SChris Mason goto fail_tree_root; 18200b86a832SChris Mason extent_root->track_dirty = 1; 18210b86a832SChris Mason 18220b86a832SChris Mason ret = find_and_setup_root(tree_root, fs_info, 18230b86a832SChris Mason BTRFS_DEV_TREE_OBJECTID, dev_root); 18240b86a832SChris Mason if (ret) 18250b86a832SChris Mason goto fail_extent_root; 18265d4f98a2SYan Zheng dev_root->track_dirty = 1; 18273768f368SChris Mason 1828d20f7043SChris Mason ret = find_and_setup_root(tree_root, fs_info, 1829d20f7043SChris Mason BTRFS_CSUM_TREE_OBJECTID, csum_root); 1830d20f7043SChris Mason if (ret) 18315d4f98a2SYan Zheng goto fail_dev_root; 1832d20f7043SChris Mason 1833d20f7043SChris Mason csum_root->track_dirty = 1; 1834d20f7043SChris Mason 18359078a3e1SChris Mason btrfs_read_block_groups(extent_root); 18369078a3e1SChris Mason 18372d69a0f8SYan Zheng fs_info->generation = generation; 1838c146afadSYan Zheng fs_info->last_trans_committed = generation; 1839d18a2c44SChris Mason fs_info->data_alloc_profile = (u64)-1; 1840d18a2c44SChris Mason fs_info->metadata_alloc_profile = (u64)-1; 1841611f0e00SChris Mason fs_info->system_alloc_profile = fs_info->metadata_alloc_profile; 1842a74a4b97SChris Mason fs_info->cleaner_kthread = kthread_run(cleaner_kthread, tree_root, 1843a74a4b97SChris Mason "btrfs-cleaner"); 184457506d50SQinghuang Feng if (IS_ERR(fs_info->cleaner_kthread)) 1845d20f7043SChris Mason goto fail_csum_root; 1846a74a4b97SChris Mason 1847a74a4b97SChris Mason fs_info->transaction_kthread = kthread_run(transaction_kthread, 1848a74a4b97SChris Mason tree_root, 1849a74a4b97SChris Mason "btrfs-transaction"); 185057506d50SQinghuang Feng if (IS_ERR(fs_info->transaction_kthread)) 18513f157a2fSChris Mason goto fail_cleaner; 1852a74a4b97SChris Mason 1853c289811cSChris Mason if (!btrfs_test_opt(tree_root, SSD) && 1854c289811cSChris Mason !btrfs_test_opt(tree_root, NOSSD) && 1855c289811cSChris Mason !fs_info->fs_devices->rotating) { 1856c289811cSChris Mason printk(KERN_INFO "Btrfs detected SSD devices, enabling SSD " 1857c289811cSChris Mason "mode\n"); 1858c289811cSChris Mason btrfs_set_opt(fs_info->mount_opt, SSD); 1859c289811cSChris Mason } 1860c289811cSChris Mason 1861e02119d5SChris Mason if (btrfs_super_log_root(disk_super) != 0) { 1862e02119d5SChris Mason u64 bytenr = btrfs_super_log_root(disk_super); 1863d18a2c44SChris Mason 18647c2ca468SChris Mason if (fs_devices->rw_devices == 0) { 1865d397712bSChris Mason printk(KERN_WARNING "Btrfs log replay required " 1866d397712bSChris Mason "on RO media\n"); 18677c2ca468SChris Mason err = -EIO; 18687c2ca468SChris Mason goto fail_trans_kthread; 18697c2ca468SChris Mason } 1870e02119d5SChris Mason blocksize = 1871e02119d5SChris Mason btrfs_level_size(tree_root, 1872e02119d5SChris Mason btrfs_super_log_root_level(disk_super)); 1873e02119d5SChris Mason 1874e02119d5SChris Mason log_tree_root = kzalloc(sizeof(struct btrfs_root), 1875e02119d5SChris Mason GFP_NOFS); 1876e02119d5SChris Mason 1877e02119d5SChris Mason __setup_root(nodesize, leafsize, sectorsize, stripesize, 1878e02119d5SChris Mason log_tree_root, fs_info, BTRFS_TREE_LOG_OBJECTID); 1879e02119d5SChris Mason 1880e02119d5SChris Mason log_tree_root->node = read_tree_block(tree_root, bytenr, 188184234f3aSYan Zheng blocksize, 188284234f3aSYan Zheng generation + 1); 1883e02119d5SChris Mason ret = btrfs_recover_log_trees(log_tree_root); 1884e02119d5SChris Mason BUG_ON(ret); 1885e556ce2cSYan Zheng 1886e556ce2cSYan Zheng if (sb->s_flags & MS_RDONLY) { 1887e556ce2cSYan Zheng ret = btrfs_commit_super(tree_root); 1888e556ce2cSYan Zheng BUG_ON(ret); 1889e556ce2cSYan Zheng } 1890e02119d5SChris Mason } 18911a40e23bSZheng Yan 18927c2ca468SChris Mason if (!(sb->s_flags & MS_RDONLY)) { 18935d4f98a2SYan Zheng ret = btrfs_recover_relocation(tree_root); 18941a40e23bSZheng Yan BUG_ON(ret); 18957c2ca468SChris Mason } 18961a40e23bSZheng Yan 18973de4586cSChris Mason location.objectid = BTRFS_FS_TREE_OBJECTID; 18983de4586cSChris Mason location.type = BTRFS_ROOT_ITEM_KEY; 18993de4586cSChris Mason location.offset = (u64)-1; 19003de4586cSChris Mason 19013de4586cSChris Mason fs_info->fs_root = btrfs_read_fs_root_no_name(fs_info, &location); 19023de4586cSChris Mason if (!fs_info->fs_root) 19037c2ca468SChris Mason goto fail_trans_kthread; 1904c289811cSChris Mason 19050f7d52f4SChris Mason return tree_root; 190639279cc3SChris Mason 19077c2ca468SChris Mason fail_trans_kthread: 19087c2ca468SChris Mason kthread_stop(fs_info->transaction_kthread); 19093f157a2fSChris Mason fail_cleaner: 1910a74a4b97SChris Mason kthread_stop(fs_info->cleaner_kthread); 19117c2ca468SChris Mason 19127c2ca468SChris Mason /* 19137c2ca468SChris Mason * make sure we're done with the btree inode before we stop our 19147c2ca468SChris Mason * kthreads 19157c2ca468SChris Mason */ 19167c2ca468SChris Mason filemap_write_and_wait(fs_info->btree_inode->i_mapping); 19177c2ca468SChris Mason invalidate_inode_pages2(fs_info->btree_inode->i_mapping); 19187c2ca468SChris Mason 1919d20f7043SChris Mason fail_csum_root: 1920d20f7043SChris Mason free_extent_buffer(csum_root->node); 19215d4f98a2SYan Zheng free_extent_buffer(csum_root->commit_root); 19225d4f98a2SYan Zheng fail_dev_root: 19235d4f98a2SYan Zheng free_extent_buffer(dev_root->node); 19245d4f98a2SYan Zheng free_extent_buffer(dev_root->commit_root); 19250b86a832SChris Mason fail_extent_root: 19260b86a832SChris Mason free_extent_buffer(extent_root->node); 19275d4f98a2SYan Zheng free_extent_buffer(extent_root->commit_root); 192839279cc3SChris Mason fail_tree_root: 19295f39d397SChris Mason free_extent_buffer(tree_root->node); 19305d4f98a2SYan Zheng free_extent_buffer(tree_root->commit_root); 19312b82032cSYan Zheng fail_chunk_root: 19322b82032cSYan Zheng free_extent_buffer(chunk_root->node); 19335d4f98a2SYan Zheng free_extent_buffer(chunk_root->commit_root); 193439279cc3SChris Mason fail_sb_buffer: 1935247e743cSChris Mason btrfs_stop_workers(&fs_info->fixup_workers); 1936771ed689SChris Mason btrfs_stop_workers(&fs_info->delalloc_workers); 19378b712842SChris Mason btrfs_stop_workers(&fs_info->workers); 19388b712842SChris Mason btrfs_stop_workers(&fs_info->endio_workers); 1939d20f7043SChris Mason btrfs_stop_workers(&fs_info->endio_meta_workers); 1940cad321adSChris Mason btrfs_stop_workers(&fs_info->endio_meta_write_workers); 1941e6dcd2dcSChris Mason btrfs_stop_workers(&fs_info->endio_write_workers); 19421cc127b5SChris Mason btrfs_stop_workers(&fs_info->submit_workers); 19434543df7eSChris Mason fail_iput: 19447c2ca468SChris Mason invalidate_inode_pages2(fs_info->btree_inode->i_mapping); 19454543df7eSChris Mason iput(fs_info->btree_inode); 19467e662854SQinghuang Feng 1947dfe25020SChris Mason btrfs_close_devices(fs_info->fs_devices); 194884eed90fSChris Mason btrfs_mapping_tree_free(&fs_info->mapping_tree); 19497e662854SQinghuang Feng bdi_destroy(&fs_info->bdi); 195084eed90fSChris Mason 19517e662854SQinghuang Feng fail: 195239279cc3SChris Mason kfree(extent_root); 195339279cc3SChris Mason kfree(tree_root); 195439279cc3SChris Mason kfree(fs_info); 195583afeac4SJim Meyering kfree(chunk_root); 195683afeac4SJim Meyering kfree(dev_root); 1957d20f7043SChris Mason kfree(csum_root); 195839279cc3SChris Mason return ERR_PTR(err); 1959eb60ceacSChris Mason } 1960eb60ceacSChris Mason 1961f2984462SChris Mason static void btrfs_end_buffer_write_sync(struct buffer_head *bh, int uptodate) 1962f2984462SChris Mason { 1963f2984462SChris Mason char b[BDEVNAME_SIZE]; 1964f2984462SChris Mason 1965f2984462SChris Mason if (uptodate) { 1966f2984462SChris Mason set_buffer_uptodate(bh); 1967f2984462SChris Mason } else { 1968f2984462SChris Mason if (!buffer_eopnotsupp(bh) && printk_ratelimit()) { 1969f2984462SChris Mason printk(KERN_WARNING "lost page write due to " 1970f2984462SChris Mason "I/O error on %s\n", 1971f2984462SChris Mason bdevname(bh->b_bdev, b)); 1972f2984462SChris Mason } 19731259ab75SChris Mason /* note, we dont' set_buffer_write_io_error because we have 19741259ab75SChris Mason * our own ways of dealing with the IO errors 19751259ab75SChris Mason */ 1976f2984462SChris Mason clear_buffer_uptodate(bh); 1977f2984462SChris Mason } 1978f2984462SChris Mason unlock_buffer(bh); 1979f2984462SChris Mason put_bh(bh); 1980f2984462SChris Mason } 1981f2984462SChris Mason 1982a512bbf8SYan Zheng struct buffer_head *btrfs_read_dev_super(struct block_device *bdev) 1983a512bbf8SYan Zheng { 1984a512bbf8SYan Zheng struct buffer_head *bh; 1985a512bbf8SYan Zheng struct buffer_head *latest = NULL; 1986a512bbf8SYan Zheng struct btrfs_super_block *super; 1987a512bbf8SYan Zheng int i; 1988a512bbf8SYan Zheng u64 transid = 0; 1989a512bbf8SYan Zheng u64 bytenr; 1990a512bbf8SYan Zheng 1991a512bbf8SYan Zheng /* we would like to check all the supers, but that would make 1992a512bbf8SYan Zheng * a btrfs mount succeed after a mkfs from a different FS. 1993a512bbf8SYan Zheng * So, we need to add a special mount option to scan for 1994a512bbf8SYan Zheng * later supers, using BTRFS_SUPER_MIRROR_MAX instead 1995a512bbf8SYan Zheng */ 1996a512bbf8SYan Zheng for (i = 0; i < 1; i++) { 1997a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 1998a512bbf8SYan Zheng if (bytenr + 4096 >= i_size_read(bdev->bd_inode)) 1999a512bbf8SYan Zheng break; 2000a512bbf8SYan Zheng bh = __bread(bdev, bytenr / 4096, 4096); 2001a512bbf8SYan Zheng if (!bh) 2002a512bbf8SYan Zheng continue; 2003a512bbf8SYan Zheng 2004a512bbf8SYan Zheng super = (struct btrfs_super_block *)bh->b_data; 2005a512bbf8SYan Zheng if (btrfs_super_bytenr(super) != bytenr || 2006a512bbf8SYan Zheng strncmp((char *)(&super->magic), BTRFS_MAGIC, 2007a512bbf8SYan Zheng sizeof(super->magic))) { 2008a512bbf8SYan Zheng brelse(bh); 2009a512bbf8SYan Zheng continue; 2010a512bbf8SYan Zheng } 2011a512bbf8SYan Zheng 2012a512bbf8SYan Zheng if (!latest || btrfs_super_generation(super) > transid) { 2013a512bbf8SYan Zheng brelse(latest); 2014a512bbf8SYan Zheng latest = bh; 2015a512bbf8SYan Zheng transid = btrfs_super_generation(super); 2016a512bbf8SYan Zheng } else { 2017a512bbf8SYan Zheng brelse(bh); 2018a512bbf8SYan Zheng } 2019a512bbf8SYan Zheng } 2020a512bbf8SYan Zheng return latest; 2021a512bbf8SYan Zheng } 2022a512bbf8SYan Zheng 2023a512bbf8SYan Zheng static int write_dev_supers(struct btrfs_device *device, 2024a512bbf8SYan Zheng struct btrfs_super_block *sb, 2025a512bbf8SYan Zheng int do_barriers, int wait, int max_mirrors) 2026a512bbf8SYan Zheng { 2027a512bbf8SYan Zheng struct buffer_head *bh; 2028a512bbf8SYan Zheng int i; 2029a512bbf8SYan Zheng int ret; 2030a512bbf8SYan Zheng int errors = 0; 2031a512bbf8SYan Zheng u32 crc; 2032a512bbf8SYan Zheng u64 bytenr; 2033a512bbf8SYan Zheng int last_barrier = 0; 2034a512bbf8SYan Zheng 2035a512bbf8SYan Zheng if (max_mirrors == 0) 2036a512bbf8SYan Zheng max_mirrors = BTRFS_SUPER_MIRROR_MAX; 2037a512bbf8SYan Zheng 2038a512bbf8SYan Zheng /* make sure only the last submit_bh does a barrier */ 2039a512bbf8SYan Zheng if (do_barriers) { 2040a512bbf8SYan Zheng for (i = 0; i < max_mirrors; i++) { 2041a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 2042a512bbf8SYan Zheng if (bytenr + BTRFS_SUPER_INFO_SIZE >= 2043a512bbf8SYan Zheng device->total_bytes) 2044a512bbf8SYan Zheng break; 2045a512bbf8SYan Zheng last_barrier = i; 2046a512bbf8SYan Zheng } 2047a512bbf8SYan Zheng } 2048a512bbf8SYan Zheng 2049a512bbf8SYan Zheng for (i = 0; i < max_mirrors; i++) { 2050a512bbf8SYan Zheng bytenr = btrfs_sb_offset(i); 2051a512bbf8SYan Zheng if (bytenr + BTRFS_SUPER_INFO_SIZE >= device->total_bytes) 2052a512bbf8SYan Zheng break; 2053a512bbf8SYan Zheng 2054a512bbf8SYan Zheng if (wait) { 2055a512bbf8SYan Zheng bh = __find_get_block(device->bdev, bytenr / 4096, 2056a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE); 2057a512bbf8SYan Zheng BUG_ON(!bh); 2058a512bbf8SYan Zheng brelse(bh); 2059a512bbf8SYan Zheng wait_on_buffer(bh); 2060a512bbf8SYan Zheng if (buffer_uptodate(bh)) { 2061a512bbf8SYan Zheng brelse(bh); 2062a512bbf8SYan Zheng continue; 2063a512bbf8SYan Zheng } 2064a512bbf8SYan Zheng } else { 2065a512bbf8SYan Zheng btrfs_set_super_bytenr(sb, bytenr); 2066a512bbf8SYan Zheng 2067a512bbf8SYan Zheng crc = ~(u32)0; 2068a512bbf8SYan Zheng crc = btrfs_csum_data(NULL, (char *)sb + 2069a512bbf8SYan Zheng BTRFS_CSUM_SIZE, crc, 2070a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE - 2071a512bbf8SYan Zheng BTRFS_CSUM_SIZE); 2072a512bbf8SYan Zheng btrfs_csum_final(crc, sb->csum); 2073a512bbf8SYan Zheng 2074a512bbf8SYan Zheng bh = __getblk(device->bdev, bytenr / 4096, 2075a512bbf8SYan Zheng BTRFS_SUPER_INFO_SIZE); 2076a512bbf8SYan Zheng memcpy(bh->b_data, sb, BTRFS_SUPER_INFO_SIZE); 2077a512bbf8SYan Zheng 2078a512bbf8SYan Zheng set_buffer_uptodate(bh); 2079a512bbf8SYan Zheng get_bh(bh); 2080a512bbf8SYan Zheng lock_buffer(bh); 2081a512bbf8SYan Zheng bh->b_end_io = btrfs_end_buffer_write_sync; 2082a512bbf8SYan Zheng } 2083a512bbf8SYan Zheng 2084a512bbf8SYan Zheng if (i == last_barrier && do_barriers && device->barriers) { 2085a512bbf8SYan Zheng ret = submit_bh(WRITE_BARRIER, bh); 2086a512bbf8SYan Zheng if (ret == -EOPNOTSUPP) { 2087a512bbf8SYan Zheng printk("btrfs: disabling barriers on dev %s\n", 2088a512bbf8SYan Zheng device->name); 2089a512bbf8SYan Zheng set_buffer_uptodate(bh); 2090a512bbf8SYan Zheng device->barriers = 0; 2091a512bbf8SYan Zheng get_bh(bh); 2092a512bbf8SYan Zheng lock_buffer(bh); 2093ffbd517dSChris Mason ret = submit_bh(WRITE_SYNC, bh); 2094a512bbf8SYan Zheng } 2095a512bbf8SYan Zheng } else { 2096ffbd517dSChris Mason ret = submit_bh(WRITE_SYNC, bh); 2097a512bbf8SYan Zheng } 2098a512bbf8SYan Zheng 2099a512bbf8SYan Zheng if (!ret && wait) { 2100a512bbf8SYan Zheng wait_on_buffer(bh); 2101a512bbf8SYan Zheng if (!buffer_uptodate(bh)) 2102a512bbf8SYan Zheng errors++; 2103a512bbf8SYan Zheng } else if (ret) { 2104a512bbf8SYan Zheng errors++; 2105a512bbf8SYan Zheng } 2106a512bbf8SYan Zheng if (wait) 2107a512bbf8SYan Zheng brelse(bh); 2108a512bbf8SYan Zheng } 2109a512bbf8SYan Zheng return errors < i ? 0 : -1; 2110a512bbf8SYan Zheng } 2111a512bbf8SYan Zheng 2112a512bbf8SYan Zheng int write_all_supers(struct btrfs_root *root, int max_mirrors) 2113f2984462SChris Mason { 2114e5e9a520SChris Mason struct list_head *head; 2115f2984462SChris Mason struct btrfs_device *dev; 2116a061fc8dSChris Mason struct btrfs_super_block *sb; 2117f2984462SChris Mason struct btrfs_dev_item *dev_item; 2118f2984462SChris Mason int ret; 2119f2984462SChris Mason int do_barriers; 2120a236aed1SChris Mason int max_errors; 2121a236aed1SChris Mason int total_errors = 0; 2122a061fc8dSChris Mason u64 flags; 2123f2984462SChris Mason 2124a236aed1SChris Mason max_errors = btrfs_super_num_devices(&root->fs_info->super_copy) - 1; 2125f2984462SChris Mason do_barriers = !btrfs_test_opt(root, NOBARRIER); 2126f2984462SChris Mason 2127a061fc8dSChris Mason sb = &root->fs_info->super_for_commit; 2128a061fc8dSChris Mason dev_item = &sb->dev_item; 2129e5e9a520SChris Mason 2130e5e9a520SChris Mason mutex_lock(&root->fs_info->fs_devices->device_list_mutex); 2131e5e9a520SChris Mason head = &root->fs_info->fs_devices->devices; 2132c6e30871SQinghuang Feng list_for_each_entry(dev, head, dev_list) { 2133dfe25020SChris Mason if (!dev->bdev) { 2134dfe25020SChris Mason total_errors++; 2135dfe25020SChris Mason continue; 2136dfe25020SChris Mason } 21372b82032cSYan Zheng if (!dev->in_fs_metadata || !dev->writeable) 2138dfe25020SChris Mason continue; 2139dfe25020SChris Mason 21402b82032cSYan Zheng btrfs_set_stack_device_generation(dev_item, 0); 2141a061fc8dSChris Mason btrfs_set_stack_device_type(dev_item, dev->type); 2142a061fc8dSChris Mason btrfs_set_stack_device_id(dev_item, dev->devid); 2143a061fc8dSChris Mason btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes); 2144a061fc8dSChris Mason btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used); 2145a061fc8dSChris Mason btrfs_set_stack_device_io_align(dev_item, dev->io_align); 2146a061fc8dSChris Mason btrfs_set_stack_device_io_width(dev_item, dev->io_width); 2147a061fc8dSChris Mason btrfs_set_stack_device_sector_size(dev_item, dev->sector_size); 2148a061fc8dSChris Mason memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE); 21492b82032cSYan Zheng memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE); 2150a512bbf8SYan Zheng 2151a061fc8dSChris Mason flags = btrfs_super_flags(sb); 2152a061fc8dSChris Mason btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN); 2153f2984462SChris Mason 2154a512bbf8SYan Zheng ret = write_dev_supers(dev, sb, do_barriers, 0, max_mirrors); 2155a236aed1SChris Mason if (ret) 2156a236aed1SChris Mason total_errors++; 2157f2984462SChris Mason } 2158a236aed1SChris Mason if (total_errors > max_errors) { 2159d397712bSChris Mason printk(KERN_ERR "btrfs: %d errors while writing supers\n", 2160d397712bSChris Mason total_errors); 2161a236aed1SChris Mason BUG(); 2162a236aed1SChris Mason } 2163f2984462SChris Mason 2164a512bbf8SYan Zheng total_errors = 0; 2165c6e30871SQinghuang Feng list_for_each_entry(dev, head, dev_list) { 2166dfe25020SChris Mason if (!dev->bdev) 2167dfe25020SChris Mason continue; 21682b82032cSYan Zheng if (!dev->in_fs_metadata || !dev->writeable) 2169dfe25020SChris Mason continue; 2170dfe25020SChris Mason 2171a512bbf8SYan Zheng ret = write_dev_supers(dev, sb, do_barriers, 1, max_mirrors); 2172a512bbf8SYan Zheng if (ret) 21731259ab75SChris Mason total_errors++; 2174f2984462SChris Mason } 2175e5e9a520SChris Mason mutex_unlock(&root->fs_info->fs_devices->device_list_mutex); 2176a236aed1SChris Mason if (total_errors > max_errors) { 2177d397712bSChris Mason printk(KERN_ERR "btrfs: %d errors while writing supers\n", 2178d397712bSChris Mason total_errors); 2179a236aed1SChris Mason BUG(); 2180a236aed1SChris Mason } 2181f2984462SChris Mason return 0; 2182f2984462SChris Mason } 2183f2984462SChris Mason 2184a512bbf8SYan Zheng int write_ctree_super(struct btrfs_trans_handle *trans, 2185a512bbf8SYan Zheng struct btrfs_root *root, int max_mirrors) 2186cfaa7295SChris Mason { 2187e66f709bSChris Mason int ret; 21882c90e5d6SChris Mason 2189a512bbf8SYan Zheng ret = write_all_supers(root, max_mirrors); 21905f39d397SChris Mason return ret; 2191cfaa7295SChris Mason } 2192cfaa7295SChris Mason 21935eda7b5eSChris Mason int btrfs_free_fs_root(struct btrfs_fs_info *fs_info, struct btrfs_root *root) 21942619ba1fSChris Mason { 21955d4f98a2SYan Zheng WARN_ON(!RB_EMPTY_ROOT(&root->inode_tree)); 21962619ba1fSChris Mason radix_tree_delete(&fs_info->fs_roots_radix, 21972619ba1fSChris Mason (unsigned long)root->root_key.objectid); 21983394e160SChris Mason if (root->anon_super.s_dev) { 21993394e160SChris Mason down_write(&root->anon_super.s_umount); 22003394e160SChris Mason kill_anon_super(&root->anon_super); 22013394e160SChris Mason } 22022619ba1fSChris Mason if (root->node) 22035f39d397SChris Mason free_extent_buffer(root->node); 22042619ba1fSChris Mason if (root->commit_root) 22055f39d397SChris Mason free_extent_buffer(root->commit_root); 220658176a96SJosef Bacik kfree(root->name); 22072619ba1fSChris Mason kfree(root); 22082619ba1fSChris Mason return 0; 22092619ba1fSChris Mason } 22102619ba1fSChris Mason 221135b7e476SChris Mason static int del_fs_roots(struct btrfs_fs_info *fs_info) 22120f7d52f4SChris Mason { 22130f7d52f4SChris Mason int ret; 22140f7d52f4SChris Mason struct btrfs_root *gang[8]; 22150f7d52f4SChris Mason int i; 22160f7d52f4SChris Mason 22170f7d52f4SChris Mason while (1) { 22180f7d52f4SChris Mason ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix, 22190f7d52f4SChris Mason (void **)gang, 0, 22200f7d52f4SChris Mason ARRAY_SIZE(gang)); 22210f7d52f4SChris Mason if (!ret) 22220f7d52f4SChris Mason break; 22232619ba1fSChris Mason for (i = 0; i < ret; i++) 22245eda7b5eSChris Mason btrfs_free_fs_root(fs_info, gang[i]); 22250f7d52f4SChris Mason } 22260f7d52f4SChris Mason return 0; 22270f7d52f4SChris Mason } 2228b4100d64SChris Mason 2229c146afadSYan Zheng int btrfs_cleanup_fs_roots(struct btrfs_fs_info *fs_info) 2230c146afadSYan Zheng { 2231c146afadSYan Zheng u64 root_objectid = 0; 2232c146afadSYan Zheng struct btrfs_root *gang[8]; 2233c146afadSYan Zheng int i; 2234c146afadSYan Zheng int ret; 2235c146afadSYan Zheng 2236c146afadSYan Zheng while (1) { 2237c146afadSYan Zheng ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix, 2238c146afadSYan Zheng (void **)gang, root_objectid, 2239c146afadSYan Zheng ARRAY_SIZE(gang)); 2240c146afadSYan Zheng if (!ret) 2241c146afadSYan Zheng break; 22425d4f98a2SYan Zheng 22435d4f98a2SYan Zheng root_objectid = gang[ret - 1]->root_key.objectid + 1; 2244c146afadSYan Zheng for (i = 0; i < ret; i++) { 2245c146afadSYan Zheng root_objectid = gang[i]->root_key.objectid; 2246c146afadSYan Zheng ret = btrfs_find_dead_roots(fs_info->tree_root, 22475d4f98a2SYan Zheng root_objectid); 2248c146afadSYan Zheng BUG_ON(ret); 2249c146afadSYan Zheng btrfs_orphan_cleanup(gang[i]); 2250c146afadSYan Zheng } 2251c146afadSYan Zheng root_objectid++; 2252c146afadSYan Zheng } 2253c146afadSYan Zheng return 0; 2254c146afadSYan Zheng } 2255c146afadSYan Zheng 2256c146afadSYan Zheng int btrfs_commit_super(struct btrfs_root *root) 2257c146afadSYan Zheng { 2258c146afadSYan Zheng struct btrfs_trans_handle *trans; 2259c146afadSYan Zheng int ret; 2260c146afadSYan Zheng 2261c146afadSYan Zheng mutex_lock(&root->fs_info->cleaner_mutex); 2262c146afadSYan Zheng btrfs_clean_old_snapshots(root); 2263c146afadSYan Zheng mutex_unlock(&root->fs_info->cleaner_mutex); 2264c146afadSYan Zheng trans = btrfs_start_transaction(root, 1); 2265c146afadSYan Zheng ret = btrfs_commit_transaction(trans, root); 2266c146afadSYan Zheng BUG_ON(ret); 2267c146afadSYan Zheng /* run commit again to drop the original snapshot */ 2268c146afadSYan Zheng trans = btrfs_start_transaction(root, 1); 2269c146afadSYan Zheng btrfs_commit_transaction(trans, root); 2270c146afadSYan Zheng ret = btrfs_write_and_wait_transaction(NULL, root); 2271c146afadSYan Zheng BUG_ON(ret); 2272c146afadSYan Zheng 2273a512bbf8SYan Zheng ret = write_ctree_super(NULL, root, 0); 2274c146afadSYan Zheng return ret; 2275c146afadSYan Zheng } 2276c146afadSYan Zheng 2277e20d96d6SChris Mason int close_ctree(struct btrfs_root *root) 2278eb60ceacSChris Mason { 22790f7d52f4SChris Mason struct btrfs_fs_info *fs_info = root->fs_info; 2280c146afadSYan Zheng int ret; 2281e089f05cSChris Mason 2282facda1e7SChris Mason fs_info->closing = 1; 2283a2135011SChris Mason smp_mb(); 2284a2135011SChris Mason 2285a74a4b97SChris Mason kthread_stop(root->fs_info->transaction_kthread); 2286a74a4b97SChris Mason kthread_stop(root->fs_info->cleaner_kthread); 2287a74a4b97SChris Mason 2288c146afadSYan Zheng if (!(fs_info->sb->s_flags & MS_RDONLY)) { 2289c146afadSYan Zheng ret = btrfs_commit_super(root); 2290d397712bSChris Mason if (ret) 2291d397712bSChris Mason printk(KERN_ERR "btrfs: commit super ret %d\n", ret); 2292c146afadSYan Zheng } 2293ed2ff2cbSChris Mason 2294b0c68f8bSChris Mason if (fs_info->delalloc_bytes) { 2295d397712bSChris Mason printk(KERN_INFO "btrfs: at unmount delalloc count %llu\n", 229621380931SJoel Becker (unsigned long long)fs_info->delalloc_bytes); 2297b0c68f8bSChris Mason } 229831153d81SYan Zheng if (fs_info->total_ref_cache_size) { 2299d397712bSChris Mason printk(KERN_INFO "btrfs: at umount reference cache size %llu\n", 2300d397712bSChris Mason (unsigned long long)fs_info->total_ref_cache_size); 230131153d81SYan Zheng } 230231153d81SYan Zheng 23035f39d397SChris Mason free_extent_buffer(fs_info->extent_root->node); 23045d4f98a2SYan Zheng free_extent_buffer(fs_info->extent_root->commit_root); 23055f39d397SChris Mason free_extent_buffer(fs_info->tree_root->node); 23065d4f98a2SYan Zheng free_extent_buffer(fs_info->tree_root->commit_root); 23070b86a832SChris Mason free_extent_buffer(root->fs_info->chunk_root->node); 23085d4f98a2SYan Zheng free_extent_buffer(root->fs_info->chunk_root->commit_root); 23090b86a832SChris Mason free_extent_buffer(root->fs_info->dev_root->node); 23105d4f98a2SYan Zheng free_extent_buffer(root->fs_info->dev_root->commit_root); 2311d20f7043SChris Mason free_extent_buffer(root->fs_info->csum_root->node); 23125d4f98a2SYan Zheng free_extent_buffer(root->fs_info->csum_root->commit_root); 2313d20f7043SChris Mason 23149078a3e1SChris Mason btrfs_free_block_groups(root->fs_info); 2315c146afadSYan Zheng 23160f7d52f4SChris Mason del_fs_roots(fs_info); 2317d10c5f31SChris Mason 2318c146afadSYan Zheng iput(fs_info->btree_inode); 23199ad6b7bcSChris Mason 2320247e743cSChris Mason btrfs_stop_workers(&fs_info->fixup_workers); 2321771ed689SChris Mason btrfs_stop_workers(&fs_info->delalloc_workers); 23228b712842SChris Mason btrfs_stop_workers(&fs_info->workers); 23238b712842SChris Mason btrfs_stop_workers(&fs_info->endio_workers); 2324d20f7043SChris Mason btrfs_stop_workers(&fs_info->endio_meta_workers); 2325cad321adSChris Mason btrfs_stop_workers(&fs_info->endio_meta_write_workers); 2326e6dcd2dcSChris Mason btrfs_stop_workers(&fs_info->endio_write_workers); 23271cc127b5SChris Mason btrfs_stop_workers(&fs_info->submit_workers); 2328d6bfde87SChris Mason 2329dfe25020SChris Mason btrfs_close_devices(fs_info->fs_devices); 23300b86a832SChris Mason btrfs_mapping_tree_free(&fs_info->mapping_tree); 2331b248a415SChris Mason 233204160088SChris Mason bdi_destroy(&fs_info->bdi); 23330b86a832SChris Mason 23340f7d52f4SChris Mason kfree(fs_info->extent_root); 23350f7d52f4SChris Mason kfree(fs_info->tree_root); 23360b86a832SChris Mason kfree(fs_info->chunk_root); 23370b86a832SChris Mason kfree(fs_info->dev_root); 2338d20f7043SChris Mason kfree(fs_info->csum_root); 2339eb60ceacSChris Mason return 0; 2340eb60ceacSChris Mason } 2341eb60ceacSChris Mason 23421259ab75SChris Mason int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid) 2343ccd467d6SChris Mason { 23441259ab75SChris Mason int ret; 2345810191ffSChris Mason struct inode *btree_inode = buf->first_page->mapping->host; 23461259ab75SChris Mason 23471259ab75SChris Mason ret = extent_buffer_uptodate(&BTRFS_I(btree_inode)->io_tree, buf); 23481259ab75SChris Mason if (!ret) 23491259ab75SChris Mason return ret; 23501259ab75SChris Mason 23511259ab75SChris Mason ret = verify_parent_transid(&BTRFS_I(btree_inode)->io_tree, buf, 23521259ab75SChris Mason parent_transid); 23531259ab75SChris Mason return !ret; 23545f39d397SChris Mason } 23556702ed49SChris Mason 23565f39d397SChris Mason int btrfs_set_buffer_uptodate(struct extent_buffer *buf) 23575f39d397SChris Mason { 2358810191ffSChris Mason struct inode *btree_inode = buf->first_page->mapping->host; 2359d1310b2eSChris Mason return set_extent_buffer_uptodate(&BTRFS_I(btree_inode)->io_tree, 23605f39d397SChris Mason buf); 23615f39d397SChris Mason } 23625f39d397SChris Mason 23635f39d397SChris Mason void btrfs_mark_buffer_dirty(struct extent_buffer *buf) 23645f39d397SChris Mason { 2365810191ffSChris Mason struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root; 23665f39d397SChris Mason u64 transid = btrfs_header_generation(buf); 23675f39d397SChris Mason struct inode *btree_inode = root->fs_info->btree_inode; 2368b9473439SChris Mason int was_dirty; 2369b4ce94deSChris Mason 2370b9447ef8SChris Mason btrfs_assert_tree_locked(buf); 2371ccd467d6SChris Mason if (transid != root->fs_info->generation) { 2372d397712bSChris Mason printk(KERN_CRIT "btrfs transid mismatch buffer %llu, " 2373d397712bSChris Mason "found %llu running %llu\n", 2374db94535dSChris Mason (unsigned long long)buf->start, 2375d397712bSChris Mason (unsigned long long)transid, 2376d397712bSChris Mason (unsigned long long)root->fs_info->generation); 2377ccd467d6SChris Mason WARN_ON(1); 2378ccd467d6SChris Mason } 2379b9473439SChris Mason was_dirty = set_extent_buffer_dirty(&BTRFS_I(btree_inode)->io_tree, 2380b9473439SChris Mason buf); 2381b9473439SChris Mason if (!was_dirty) { 2382b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 2383b9473439SChris Mason root->fs_info->dirty_metadata_bytes += buf->len; 2384b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 2385b9473439SChris Mason } 2386eb60ceacSChris Mason } 2387eb60ceacSChris Mason 2388d3c2fdcfSChris Mason void btrfs_btree_balance_dirty(struct btrfs_root *root, unsigned long nr) 238935b7e476SChris Mason { 2390188de649SChris Mason /* 2391188de649SChris Mason * looks as though older kernels can get into trouble with 2392188de649SChris Mason * this code, they end up stuck in balance_dirty_pages forever 2393188de649SChris Mason */ 2394d6bfde87SChris Mason u64 num_dirty; 2395771ed689SChris Mason unsigned long thresh = 32 * 1024 * 1024; 2396d6bfde87SChris Mason 23976933c02eSJens Axboe if (current->flags & PF_MEMALLOC) 2398d6bfde87SChris Mason return; 2399d6bfde87SChris Mason 2400585ad2c3SChris Mason num_dirty = root->fs_info->dirty_metadata_bytes; 2401585ad2c3SChris Mason 2402d6bfde87SChris Mason if (num_dirty > thresh) { 2403d3c2fdcfSChris Mason balance_dirty_pages_ratelimited_nr( 2404304fced6SChris Mason root->fs_info->btree_inode->i_mapping, 1); 240535b7e476SChris Mason } 2406188de649SChris Mason return; 2407d6bfde87SChris Mason } 24086b80053dSChris Mason 2409ca7a79adSChris Mason int btrfs_read_buffer(struct extent_buffer *buf, u64 parent_transid) 24106b80053dSChris Mason { 2411810191ffSChris Mason struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root; 2412ce9adaa5SChris Mason int ret; 2413ca7a79adSChris Mason ret = btree_read_extent_buffer_pages(root, buf, 0, parent_transid); 2414d397712bSChris Mason if (ret == 0) 2415b4ce94deSChris Mason set_bit(EXTENT_BUFFER_UPTODATE, &buf->bflags); 2416ce9adaa5SChris Mason return ret; 24176b80053dSChris Mason } 24180da5468fSChris Mason 24194bef0848SChris Mason int btree_lock_page_hook(struct page *page) 24204bef0848SChris Mason { 24214bef0848SChris Mason struct inode *inode = page->mapping->host; 2422b9473439SChris Mason struct btrfs_root *root = BTRFS_I(inode)->root; 24234bef0848SChris Mason struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree; 24244bef0848SChris Mason struct extent_buffer *eb; 24254bef0848SChris Mason unsigned long len; 24264bef0848SChris Mason u64 bytenr = page_offset(page); 24274bef0848SChris Mason 24284bef0848SChris Mason if (page->private == EXTENT_PAGE_PRIVATE) 24294bef0848SChris Mason goto out; 24304bef0848SChris Mason 24314bef0848SChris Mason len = page->private >> 2; 24324bef0848SChris Mason eb = find_extent_buffer(io_tree, bytenr, len, GFP_NOFS); 24334bef0848SChris Mason if (!eb) 24344bef0848SChris Mason goto out; 24354bef0848SChris Mason 24364bef0848SChris Mason btrfs_tree_lock(eb); 24374bef0848SChris Mason btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN); 2438b9473439SChris Mason 2439b9473439SChris Mason if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) { 2440b9473439SChris Mason spin_lock(&root->fs_info->delalloc_lock); 2441b9473439SChris Mason if (root->fs_info->dirty_metadata_bytes >= eb->len) 2442b9473439SChris Mason root->fs_info->dirty_metadata_bytes -= eb->len; 2443b9473439SChris Mason else 2444b9473439SChris Mason WARN_ON(1); 2445b9473439SChris Mason spin_unlock(&root->fs_info->delalloc_lock); 2446b9473439SChris Mason } 2447b9473439SChris Mason 24484bef0848SChris Mason btrfs_tree_unlock(eb); 24494bef0848SChris Mason free_extent_buffer(eb); 24504bef0848SChris Mason out: 24514bef0848SChris Mason lock_page(page); 24524bef0848SChris Mason return 0; 24534bef0848SChris Mason } 24544bef0848SChris Mason 2455d1310b2eSChris Mason static struct extent_io_ops btree_extent_io_ops = { 24564bef0848SChris Mason .write_cache_pages_lock_hook = btree_lock_page_hook, 2457ce9adaa5SChris Mason .readpage_end_io_hook = btree_readpage_end_io_hook, 24580b86a832SChris Mason .submit_bio_hook = btree_submit_bio_hook, 2459239b14b3SChris Mason /* note we're sharing with inode.c for the merge bio hook */ 2460239b14b3SChris Mason .merge_bio_hook = btrfs_merge_bio_hook, 24610da5468fSChris Mason }; 2462