1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2ac27a0ecSDave Kleikamp /* 3617ba13bSMingming Cao * linux/fs/ext4/ialloc.c 4ac27a0ecSDave Kleikamp * 5ac27a0ecSDave Kleikamp * Copyright (C) 1992, 1993, 1994, 1995 6ac27a0ecSDave Kleikamp * Remy Card (card@masi.ibp.fr) 7ac27a0ecSDave Kleikamp * Laboratoire MASI - Institut Blaise Pascal 8ac27a0ecSDave Kleikamp * Universite Pierre et Marie Curie (Paris VI) 9ac27a0ecSDave Kleikamp * 10ac27a0ecSDave Kleikamp * BSD ufs-inspired inode and directory allocation by 11ac27a0ecSDave Kleikamp * Stephen Tweedie (sct@redhat.com), 1993 12ac27a0ecSDave Kleikamp * Big-endian to little-endian byte-swapping/bitmaps by 13ac27a0ecSDave Kleikamp * David S. Miller (davem@caip.rutgers.edu), 1995 14ac27a0ecSDave Kleikamp */ 15ac27a0ecSDave Kleikamp 16ac27a0ecSDave Kleikamp #include <linux/time.h> 17ac27a0ecSDave Kleikamp #include <linux/fs.h> 18ac27a0ecSDave Kleikamp #include <linux/stat.h> 19ac27a0ecSDave Kleikamp #include <linux/string.h> 20ac27a0ecSDave Kleikamp #include <linux/quotaops.h> 21ac27a0ecSDave Kleikamp #include <linux/buffer_head.h> 22ac27a0ecSDave Kleikamp #include <linux/random.h> 23ac27a0ecSDave Kleikamp #include <linux/bitops.h> 243a5b2ecdSMingming Cao #include <linux/blkdev.h> 255b825c3aSIngo Molnar #include <linux/cred.h> 265b825c3aSIngo Molnar 27ac27a0ecSDave Kleikamp #include <asm/byteorder.h> 289bffad1eSTheodore Ts'o 293dcf5451SChristoph Hellwig #include "ext4.h" 303dcf5451SChristoph Hellwig #include "ext4_jbd2.h" 31ac27a0ecSDave Kleikamp #include "xattr.h" 32ac27a0ecSDave Kleikamp #include "acl.h" 33ac27a0ecSDave Kleikamp 349bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 359bffad1eSTheodore Ts'o 36ac27a0ecSDave Kleikamp /* 37ac27a0ecSDave Kleikamp * ialloc.c contains the inodes allocation and deallocation routines 38ac27a0ecSDave Kleikamp */ 39ac27a0ecSDave Kleikamp 40ac27a0ecSDave Kleikamp /* 41ac27a0ecSDave Kleikamp * The free inodes are managed by bitmaps. A file system contains several 42ac27a0ecSDave Kleikamp * blocks groups. Each group contains 1 bitmap block for blocks, 1 bitmap 43ac27a0ecSDave Kleikamp * block for inodes, N blocks for the inode table and data blocks. 44ac27a0ecSDave Kleikamp * 45ac27a0ecSDave Kleikamp * The file system contains group descriptors which are located after the 46ac27a0ecSDave Kleikamp * super block. Each descriptor contains the number of the bitmap block and 47ac27a0ecSDave Kleikamp * the free blocks count in the block. 48ac27a0ecSDave Kleikamp */ 49ac27a0ecSDave Kleikamp 50717d50e4SAndreas Dilger /* 51717d50e4SAndreas Dilger * To avoid calling the atomic setbit hundreds or thousands of times, we only 52717d50e4SAndreas Dilger * need to use it within a single byte (to ensure we get endianness right). 53717d50e4SAndreas Dilger * We can use memset for the rest of the bitmap as there are no other users. 54717d50e4SAndreas Dilger */ 5561d08673STheodore Ts'o void ext4_mark_bitmap_end(int start_bit, int end_bit, char *bitmap) 56717d50e4SAndreas Dilger { 57717d50e4SAndreas Dilger int i; 58717d50e4SAndreas Dilger 59717d50e4SAndreas Dilger if (start_bit >= end_bit) 60717d50e4SAndreas Dilger return; 61717d50e4SAndreas Dilger 62717d50e4SAndreas Dilger ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit); 63717d50e4SAndreas Dilger for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++) 64717d50e4SAndreas Dilger ext4_set_bit(i, bitmap); 65717d50e4SAndreas Dilger if (i < end_bit) 66717d50e4SAndreas Dilger memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3); 67717d50e4SAndreas Dilger } 68717d50e4SAndreas Dilger 69813e5727STheodore Ts'o void ext4_end_bitmap_read(struct buffer_head *bh, int uptodate) 70813e5727STheodore Ts'o { 71813e5727STheodore Ts'o if (uptodate) { 72813e5727STheodore Ts'o set_buffer_uptodate(bh); 73813e5727STheodore Ts'o set_bitmap_uptodate(bh); 74813e5727STheodore Ts'o } 75813e5727STheodore Ts'o unlock_buffer(bh); 76813e5727STheodore Ts'o put_bh(bh); 77813e5727STheodore Ts'o } 78813e5727STheodore Ts'o 799008a58eSDarrick J. Wong static int ext4_validate_inode_bitmap(struct super_block *sb, 809008a58eSDarrick J. Wong struct ext4_group_desc *desc, 819008a58eSDarrick J. Wong ext4_group_t block_group, 829008a58eSDarrick J. Wong struct buffer_head *bh) 839008a58eSDarrick J. Wong { 849008a58eSDarrick J. Wong ext4_fsblk_t blk; 858016e29fSHarshad Shirwadkar struct ext4_group_info *grp; 868016e29fSHarshad Shirwadkar 878016e29fSHarshad Shirwadkar if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) 888016e29fSHarshad Shirwadkar return 0; 898016e29fSHarshad Shirwadkar 908016e29fSHarshad Shirwadkar grp = ext4_get_group_info(sb, block_group); 919008a58eSDarrick J. Wong 929008a58eSDarrick J. Wong if (buffer_verified(bh)) 939008a58eSDarrick J. Wong return 0; 949008a58eSDarrick J. Wong if (EXT4_MB_GRP_IBITMAP_CORRUPT(grp)) 959008a58eSDarrick J. Wong return -EFSCORRUPTED; 969008a58eSDarrick J. Wong 979008a58eSDarrick J. Wong ext4_lock_group(sb, block_group); 988d5a803cSTheodore Ts'o if (buffer_verified(bh)) 998d5a803cSTheodore Ts'o goto verified; 1009008a58eSDarrick J. Wong blk = ext4_inode_bitmap(sb, desc); 1019008a58eSDarrick J. Wong if (!ext4_inode_bitmap_csum_verify(sb, block_group, desc, bh, 10246f870d6STheodore Ts'o EXT4_INODES_PER_GROUP(sb) / 8) || 10346f870d6STheodore Ts'o ext4_simulate_fail(sb, EXT4_SIM_IBITMAP_CRC)) { 1049008a58eSDarrick J. Wong ext4_unlock_group(sb, block_group); 1059008a58eSDarrick J. Wong ext4_error(sb, "Corrupt inode bitmap - block_group = %u, " 1069008a58eSDarrick J. Wong "inode_bitmap = %llu", block_group, blk); 107db79e6d1SWang Shilong ext4_mark_group_bitmap_corrupted(sb, block_group, 108db79e6d1SWang Shilong EXT4_GROUP_INFO_IBITMAP_CORRUPT); 1099008a58eSDarrick J. Wong return -EFSBADCRC; 1109008a58eSDarrick J. Wong } 1119008a58eSDarrick J. Wong set_buffer_verified(bh); 1128d5a803cSTheodore Ts'o verified: 1139008a58eSDarrick J. Wong ext4_unlock_group(sb, block_group); 1149008a58eSDarrick J. Wong return 0; 1159008a58eSDarrick J. Wong } 1169008a58eSDarrick J. Wong 117ac27a0ecSDave Kleikamp /* 118ac27a0ecSDave Kleikamp * Read the inode allocation bitmap for a given block_group, reading 119ac27a0ecSDave Kleikamp * into the specified slot in the superblock's bitmap cache. 120ac27a0ecSDave Kleikamp * 1219033783cSJosh Triplett * Return buffer_head of bitmap on success, or an ERR_PTR on error. 122ac27a0ecSDave Kleikamp */ 123ac27a0ecSDave Kleikamp static struct buffer_head * 124e29d1cdeSEric Sandeen ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group) 125ac27a0ecSDave Kleikamp { 126617ba13bSMingming Cao struct ext4_group_desc *desc; 1277dac4a17STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 128ac27a0ecSDave Kleikamp struct buffer_head *bh = NULL; 129e29d1cdeSEric Sandeen ext4_fsblk_t bitmap_blk; 1309008a58eSDarrick J. Wong int err; 131ac27a0ecSDave Kleikamp 132617ba13bSMingming Cao desc = ext4_get_group_desc(sb, block_group, NULL); 133ac27a0ecSDave Kleikamp if (!desc) 1349008a58eSDarrick J. Wong return ERR_PTR(-EFSCORRUPTED); 135bfff6873SLukas Czerner 136e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 1377dac4a17STheodore Ts'o if ((bitmap_blk <= le32_to_cpu(sbi->s_es->s_first_data_block)) || 1387dac4a17STheodore Ts'o (bitmap_blk >= ext4_blocks_count(sbi->s_es))) { 1397dac4a17STheodore Ts'o ext4_error(sb, "Invalid inode bitmap blk %llu in " 1407dac4a17STheodore Ts'o "block_group %u", bitmap_blk, block_group); 141206f6d55SWang Shilong ext4_mark_group_bitmap_corrupted(sb, block_group, 142206f6d55SWang Shilong EXT4_GROUP_INFO_IBITMAP_CORRUPT); 1437dac4a17STheodore Ts'o return ERR_PTR(-EFSCORRUPTED); 1447dac4a17STheodore Ts'o } 145e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 146e29d1cdeSEric Sandeen if (unlikely(!bh)) { 1475ef2a699SWang Shilong ext4_warning(sb, "Cannot read inode bitmap - " 148a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 149e29d1cdeSEric Sandeen block_group, bitmap_blk); 1500db9fdebSWang Shilong return ERR_PTR(-ENOMEM); 151e29d1cdeSEric Sandeen } 1522ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) 15341a246d1SDarrick J. Wong goto verify; 154e29d1cdeSEric Sandeen 155c806e68fSFrederic Bohe lock_buffer(bh); 1562ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) { 1572ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 15841a246d1SDarrick J. Wong goto verify; 1592ccb5fb9SAneesh Kumar K.V } 160bfff6873SLukas Czerner 161955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 1628844618dSTheodore Ts'o if (ext4_has_group_desc_csum(sb) && 1638844618dSTheodore Ts'o (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT))) { 1648844618dSTheodore Ts'o if (block_group == 0) { 1658844618dSTheodore Ts'o ext4_unlock_group(sb, block_group); 1668844618dSTheodore Ts'o unlock_buffer(bh); 1678844618dSTheodore Ts'o ext4_error(sb, "Inode bitmap for bg 0 marked " 1688844618dSTheodore Ts'o "uninitialized"); 1698844618dSTheodore Ts'o err = -EFSCORRUPTED; 1708844618dSTheodore Ts'o goto out; 1718844618dSTheodore Ts'o } 172044e6e3dSTheodore Ts'o memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8); 173044e6e3dSTheodore Ts'o ext4_mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), 174044e6e3dSTheodore Ts'o sb->s_blocksize * 8, bh->b_data); 1752ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 176e29d1cdeSEric Sandeen set_buffer_uptodate(bh); 17741a246d1SDarrick J. Wong set_buffer_verified(bh); 178955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1793300bedaSAneesh Kumar K.V unlock_buffer(bh); 180e29d1cdeSEric Sandeen return bh; 181e29d1cdeSEric Sandeen } 182955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 183bfff6873SLukas Czerner 1842ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh)) { 1852ccb5fb9SAneesh Kumar K.V /* 1862ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 1872ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 1882ccb5fb9SAneesh Kumar K.V */ 1892ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 1902ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 19141a246d1SDarrick J. Wong goto verify; 1922ccb5fb9SAneesh Kumar K.V } 1932ccb5fb9SAneesh Kumar K.V /* 194813e5727STheodore Ts'o * submit the buffer_head for reading 1952ccb5fb9SAneesh Kumar K.V */ 1960562e0baSJiaying Zhang trace_ext4_load_inode_bitmap(sb, block_group); 1972d069c08Szhangyi (F) ext4_read_bh(bh, REQ_META | REQ_PRIO, ext4_end_bitmap_read); 19846f870d6STheodore Ts'o ext4_simulate_fail_bh(sb, bh, EXT4_SIM_IBITMAP_EIO); 199813e5727STheodore Ts'o if (!buffer_uptodate(bh)) { 200e29d1cdeSEric Sandeen put_bh(bh); 20154d3adbcSTheodore Ts'o ext4_error_err(sb, EIO, "Cannot read inode bitmap - " 202a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 203e29d1cdeSEric Sandeen block_group, bitmap_blk); 204206f6d55SWang Shilong ext4_mark_group_bitmap_corrupted(sb, block_group, 205206f6d55SWang Shilong EXT4_GROUP_INFO_IBITMAP_CORRUPT); 2069008a58eSDarrick J. Wong return ERR_PTR(-EIO); 207e29d1cdeSEric Sandeen } 20841a246d1SDarrick J. Wong 20941a246d1SDarrick J. Wong verify: 2109008a58eSDarrick J. Wong err = ext4_validate_inode_bitmap(sb, desc, block_group, bh); 2119008a58eSDarrick J. Wong if (err) 2129008a58eSDarrick J. Wong goto out; 213ac27a0ecSDave Kleikamp return bh; 2149008a58eSDarrick J. Wong out: 2159008a58eSDarrick J. Wong put_bh(bh); 2169008a58eSDarrick J. Wong return ERR_PTR(err); 217ac27a0ecSDave Kleikamp } 218ac27a0ecSDave Kleikamp 219ac27a0ecSDave Kleikamp /* 220ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 221ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 222ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 223ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 224ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 225ac27a0ecSDave Kleikamp * has been deleted earlier. 226ac27a0ecSDave Kleikamp * 227ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 228ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 229ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 230ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 231ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 232ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 233ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 234ac27a0ecSDave Kleikamp */ 235617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 236ac27a0ecSDave Kleikamp { 237ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 238ac27a0ecSDave Kleikamp int is_directory; 239ac27a0ecSDave Kleikamp unsigned long ino; 240ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 241ac27a0ecSDave Kleikamp struct buffer_head *bh2; 242fd2d4291SAvantika Mathur ext4_group_t block_group; 243ac27a0ecSDave Kleikamp unsigned long bit; 244617ba13bSMingming Cao struct ext4_group_desc *gdp; 245617ba13bSMingming Cao struct ext4_super_block *es; 246617ba13bSMingming Cao struct ext4_sb_info *sbi; 2477ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 24887a39389SDarrick J. Wong struct ext4_group_info *grp; 249ac27a0ecSDave Kleikamp 25092b97816STheodore Ts'o if (!sb) { 25192b97816STheodore Ts'o printk(KERN_ERR "EXT4-fs: %s:%d: inode on " 25292b97816STheodore Ts'o "nonexistent device\n", __func__, __LINE__); 25392b97816STheodore Ts'o return; 25492b97816STheodore Ts'o } 255ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 25692b97816STheodore Ts'o ext4_msg(sb, KERN_ERR, "%s:%d: inode #%lu: count=%d", 25792b97816STheodore Ts'o __func__, __LINE__, inode->i_ino, 258ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 259ac27a0ecSDave Kleikamp return; 260ac27a0ecSDave Kleikamp } 261ac27a0ecSDave Kleikamp if (inode->i_nlink) { 26292b97816STheodore Ts'o ext4_msg(sb, KERN_ERR, "%s:%d: inode #%lu: nlink=%d\n", 26392b97816STheodore Ts'o __func__, __LINE__, inode->i_ino, inode->i_nlink); 264ac27a0ecSDave Kleikamp return; 265ac27a0ecSDave Kleikamp } 266617ba13bSMingming Cao sbi = EXT4_SB(sb); 267ac27a0ecSDave Kleikamp 268ac27a0ecSDave Kleikamp ino = inode->i_ino; 269617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 2709bffad1eSTheodore Ts'o trace_ext4_free_inode(inode); 271ac27a0ecSDave Kleikamp 272871a2931SChristoph Hellwig dquot_initialize(inode); 27363936ddaSChristoph Hellwig dquot_free_inode(inode); 274ac27a0ecSDave Kleikamp 275ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 276ac27a0ecSDave Kleikamp 277ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 2780930fcc1SAl Viro ext4_clear_inode(inode); 279ac27a0ecSDave Kleikamp 28049598e04SJun Piao es = sbi->s_es; 281617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 28212062dddSEric Sandeen ext4_error(sb, "reserved or nonexistent inode %lu", ino); 283ac27a0ecSDave Kleikamp goto error_return; 284ac27a0ecSDave Kleikamp } 285617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 286617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 287e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 28887a39389SDarrick J. Wong /* Don't bother if the inode bitmap is corrupt. */ 2899008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 2909008a58eSDarrick J. Wong fatal = PTR_ERR(bitmap_bh); 2919008a58eSDarrick J. Wong bitmap_bh = NULL; 292ac27a0ecSDave Kleikamp goto error_return; 2939008a58eSDarrick J. Wong } 2948016e29fSHarshad Shirwadkar if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { 2958016e29fSHarshad Shirwadkar grp = ext4_get_group_info(sb, block_group); 2969008a58eSDarrick J. Wong if (unlikely(EXT4_MB_GRP_IBITMAP_CORRUPT(grp))) { 2979008a58eSDarrick J. Wong fatal = -EFSCORRUPTED; 2989008a58eSDarrick J. Wong goto error_return; 2999008a58eSDarrick J. Wong } 3008016e29fSHarshad Shirwadkar } 301ac27a0ecSDave Kleikamp 302ac27a0ecSDave Kleikamp BUFFER_TRACE(bitmap_bh, "get_write_access"); 303617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bitmap_bh); 304ac27a0ecSDave Kleikamp if (fatal) 305ac27a0ecSDave Kleikamp goto error_return; 306ac27a0ecSDave Kleikamp 307d17413c0SDmitry Monakhov fatal = -ESRCH; 308617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 309d17413c0SDmitry Monakhov if (gdp) { 310ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 311617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 312d17413c0SDmitry Monakhov } 313955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 314597d508cSAkinobu Mita cleared = ext4_test_and_clear_bit(bit, bitmap_bh->b_data); 315d17413c0SDmitry Monakhov if (fatal || !cleared) { 316d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 317d17413c0SDmitry Monakhov goto out; 318d17413c0SDmitry Monakhov } 319d17413c0SDmitry Monakhov 320560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 321560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 322560671a0SAneesh Kumar K.V if (is_directory) { 323560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 324560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 325d17413c0SDmitry Monakhov percpu_counter_dec(&sbi->s_dirs_counter); 326d17413c0SDmitry Monakhov } 32741a246d1SDarrick J. Wong ext4_inode_bitmap_csum_set(sb, block_group, gdp, bitmap_bh, 32841a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8); 329feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, gdp); 330d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 3317d39db14STheodore Ts'o 332d17413c0SDmitry Monakhov percpu_counter_inc(&sbi->s_freeinodes_counter); 333d17413c0SDmitry Monakhov if (sbi->s_log_groups_per_flex) { 3347c990728SSuraj Jitindar Singh struct flex_groups *fg; 335d17413c0SDmitry Monakhov 3367c990728SSuraj Jitindar Singh fg = sbi_array_rcu_deref(sbi, s_flex_groups, 3377c990728SSuraj Jitindar Singh ext4_flex_group(sbi, block_group)); 3387c990728SSuraj Jitindar Singh atomic_inc(&fg->free_inodes); 339d17413c0SDmitry Monakhov if (is_directory) 3407c990728SSuraj Jitindar Singh atomic_dec(&fg->used_dirs); 3417d39db14STheodore Ts'o } 3420390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 343d17413c0SDmitry Monakhov fatal = ext4_handle_dirty_metadata(handle, NULL, bh2); 344d17413c0SDmitry Monakhov out: 345d17413c0SDmitry Monakhov if (cleared) { 3460390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 3470390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 348ac27a0ecSDave Kleikamp if (!fatal) 349ac27a0ecSDave Kleikamp fatal = err; 35087a39389SDarrick J. Wong } else { 351d17413c0SDmitry Monakhov ext4_error(sb, "bit already cleared for inode %lu", ino); 352db79e6d1SWang Shilong ext4_mark_group_bitmap_corrupted(sb, block_group, 353db79e6d1SWang Shilong EXT4_GROUP_INFO_IBITMAP_CORRUPT); 35487a39389SDarrick J. Wong } 355d17413c0SDmitry Monakhov 356ac27a0ecSDave Kleikamp error_return: 357ac27a0ecSDave Kleikamp brelse(bitmap_bh); 358617ba13bSMingming Cao ext4_std_error(sb, fatal); 359ac27a0ecSDave Kleikamp } 360ac27a0ecSDave Kleikamp 361a4912123STheodore Ts'o struct orlov_stats { 36290ba983fSTheodore Ts'o __u64 free_clusters; 363a4912123STheodore Ts'o __u32 free_inodes; 364a4912123STheodore Ts'o __u32 used_dirs; 365a4912123STheodore Ts'o }; 366a4912123STheodore Ts'o 367a4912123STheodore Ts'o /* 368a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 369a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 370a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 371a4912123STheodore Ts'o */ 3721f109d5aSTheodore Ts'o static void get_orlov_stats(struct super_block *sb, ext4_group_t g, 373a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 374a4912123STheodore Ts'o { 375a4912123STheodore Ts'o struct ext4_group_desc *desc; 376a4912123STheodore Ts'o 3777d39db14STheodore Ts'o if (flex_size > 1) { 3787c990728SSuraj Jitindar Singh struct flex_groups *fg = sbi_array_rcu_deref(EXT4_SB(sb), 3797c990728SSuraj Jitindar Singh s_flex_groups, g); 3807c990728SSuraj Jitindar Singh stats->free_inodes = atomic_read(&fg->free_inodes); 3817c990728SSuraj Jitindar Singh stats->free_clusters = atomic64_read(&fg->free_clusters); 3827c990728SSuraj Jitindar Singh stats->used_dirs = atomic_read(&fg->used_dirs); 3837d39db14STheodore Ts'o return; 3847d39db14STheodore Ts'o } 3857d39db14STheodore Ts'o 3867d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 3877d39db14STheodore Ts'o if (desc) { 3887d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 389021b65bbSTheodore Ts'o stats->free_clusters = ext4_free_group_clusters(sb, desc); 3907d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 3917d39db14STheodore Ts'o } else { 392a4912123STheodore Ts'o stats->free_inodes = 0; 39324aaa8efSTheodore Ts'o stats->free_clusters = 0; 394a4912123STheodore Ts'o stats->used_dirs = 0; 395a4912123STheodore Ts'o } 396a4912123STheodore Ts'o } 397a4912123STheodore Ts'o 398ac27a0ecSDave Kleikamp /* 399ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 400ac27a0ecSDave Kleikamp * 401ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 402ac27a0ecSDave Kleikamp * 403ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 404ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 405ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 406ac27a0ecSDave Kleikamp * 407ac27a0ecSDave Kleikamp * For the rest rules look so: 408ac27a0ecSDave Kleikamp * 409ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 410ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 411ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 412ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4131cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 414ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 415ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 416ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 417ac27a0ecSDave Kleikamp */ 418ac27a0ecSDave Kleikamp 4192aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 420dcca3fecSAl Viro ext4_group_t *group, umode_t mode, 421f157a4aaSTheodore Ts'o const struct qstr *qstr) 422ac27a0ecSDave Kleikamp { 423fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 424617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 4258df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 426617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 42714c83c9fSTheodore Ts'o unsigned int freei, avefreei, grp_free; 42824aaa8efSTheodore Ts'o ext4_fsblk_t freeb, avefreec; 429ac27a0ecSDave Kleikamp unsigned int ndirs; 430a4912123STheodore Ts'o int max_dirs, min_inodes; 43124aaa8efSTheodore Ts'o ext4_grpblk_t min_clusters; 4328df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 433617ba13bSMingming Cao struct ext4_group_desc *desc; 434a4912123STheodore Ts'o struct orlov_stats stats; 435a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 436f157a4aaSTheodore Ts'o struct dx_hash_info hinfo; 437a4912123STheodore Ts'o 4388df9675fSTheodore Ts'o ngroups = real_ngroups; 439a4912123STheodore Ts'o if (flex_size > 1) { 4408df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 441a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 442a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 443a4912123STheodore Ts'o } 444ac27a0ecSDave Kleikamp 445ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 446ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 44757042651STheodore Ts'o freeb = EXT4_C2B(sbi, 44857042651STheodore Ts'o percpu_counter_read_positive(&sbi->s_freeclusters_counter)); 44924aaa8efSTheodore Ts'o avefreec = freeb; 45024aaa8efSTheodore Ts'o do_div(avefreec, ngroups); 451ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 452ac27a0ecSDave Kleikamp 453a4912123STheodore Ts'o if (S_ISDIR(mode) && 4542b0143b5SDavid Howells ((parent == d_inode(sb->s_root)) || 45512e9b892SDmitry Monakhov (ext4_test_inode_flag(parent, EXT4_INODE_TOPDIR)))) { 456ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 4572aa9fc4cSAvantika Mathur int ret = -1; 458ac27a0ecSDave Kleikamp 459f157a4aaSTheodore Ts'o if (qstr) { 460f157a4aaSTheodore Ts'o hinfo.hash_version = DX_HASH_HALF_MD4; 461f157a4aaSTheodore Ts'o hinfo.seed = sbi->s_hash_seed; 462b886ee3eSGabriel Krisman Bertazi ext4fs_dirhash(parent, qstr->name, qstr->len, &hinfo); 463f157a4aaSTheodore Ts'o grp = hinfo.hash; 464f157a4aaSTheodore Ts'o } else 465dd1f723bSTheodore Ts'o grp = prandom_u32(); 4662aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 467ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 468a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 469a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 470a4912123STheodore Ts'o if (!stats.free_inodes) 471ac27a0ecSDave Kleikamp continue; 472a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 473ac27a0ecSDave Kleikamp continue; 474a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 475ac27a0ecSDave Kleikamp continue; 47624aaa8efSTheodore Ts'o if (stats.free_clusters < avefreec) 477ac27a0ecSDave Kleikamp continue; 478a4912123STheodore Ts'o grp = g; 4792aa9fc4cSAvantika Mathur ret = 0; 480a4912123STheodore Ts'o best_ndir = stats.used_dirs; 481ac27a0ecSDave Kleikamp } 482a4912123STheodore Ts'o if (ret) 483a4912123STheodore Ts'o goto fallback; 484a4912123STheodore Ts'o found_flex_bg: 485a4912123STheodore Ts'o if (flex_size == 1) { 486a4912123STheodore Ts'o *group = grp; 487a4912123STheodore Ts'o return 0; 488a4912123STheodore Ts'o } 489a4912123STheodore Ts'o 490a4912123STheodore Ts'o /* 491a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 492a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 493a4912123STheodore Ts'o * something similar, although regular files will 494a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 495a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 496a4912123STheodore Ts'o */ 497a4912123STheodore Ts'o grp *= flex_size; 498a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 4998df9675fSTheodore Ts'o if (grp+i >= real_ngroups) 500a4912123STheodore Ts'o break; 501a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 502a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 503a4912123STheodore Ts'o *group = grp+i; 504a4912123STheodore Ts'o return 0; 505a4912123STheodore Ts'o } 506a4912123STheodore Ts'o } 507ac27a0ecSDave Kleikamp goto fallback; 508ac27a0ecSDave Kleikamp } 509ac27a0ecSDave Kleikamp 510ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 511a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 512a4912123STheodore Ts'o if (min_inodes < 1) 513a4912123STheodore Ts'o min_inodes = 1; 51424aaa8efSTheodore Ts'o min_clusters = avefreec - EXT4_CLUSTERS_PER_GROUP(sb)*flex_size / 4; 515ac27a0ecSDave Kleikamp 516a4912123STheodore Ts'o /* 517a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 518a4912123STheodore Ts'o * inode for this parent directory 519a4912123STheodore Ts'o */ 520a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 521a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 522a4912123STheodore Ts'o if (flex_size > 1) 523a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 524a4912123STheodore Ts'o } 525ac27a0ecSDave Kleikamp 526ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 527a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 528a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 529a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 530ac27a0ecSDave Kleikamp continue; 531a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 532ac27a0ecSDave Kleikamp continue; 53324aaa8efSTheodore Ts'o if (stats.free_clusters < min_clusters) 534ac27a0ecSDave Kleikamp continue; 535a4912123STheodore Ts'o goto found_flex_bg; 536ac27a0ecSDave Kleikamp } 537ac27a0ecSDave Kleikamp 538ac27a0ecSDave Kleikamp fallback: 5398df9675fSTheodore Ts'o ngroups = real_ngroups; 540a4912123STheodore Ts'o avefreei = freei / ngroups; 541b5451f7bSTheodore Ts'o fallback_retry: 542a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 543ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 544a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 545a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 546bb3d132aSDan Carpenter if (desc) { 54714c83c9fSTheodore Ts'o grp_free = ext4_free_inodes_count(sb, desc); 548bb3d132aSDan Carpenter if (grp_free && grp_free >= avefreei) { 549a4912123STheodore Ts'o *group = grp; 5502aa9fc4cSAvantika Mathur return 0; 551ac27a0ecSDave Kleikamp } 552a4912123STheodore Ts'o } 553bb3d132aSDan Carpenter } 554ac27a0ecSDave Kleikamp 555ac27a0ecSDave Kleikamp if (avefreei) { 556ac27a0ecSDave Kleikamp /* 557ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 558ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 559ac27a0ecSDave Kleikamp */ 560ac27a0ecSDave Kleikamp avefreei = 0; 561b5451f7bSTheodore Ts'o goto fallback_retry; 562ac27a0ecSDave Kleikamp } 563ac27a0ecSDave Kleikamp 564ac27a0ecSDave Kleikamp return -1; 565ac27a0ecSDave Kleikamp } 566ac27a0ecSDave Kleikamp 5672aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 568dcca3fecSAl Viro ext4_group_t *group, umode_t mode) 569ac27a0ecSDave Kleikamp { 570fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 5718df9675fSTheodore Ts'o ext4_group_t i, last, ngroups = ext4_get_groups_count(sb); 572617ba13bSMingming Cao struct ext4_group_desc *desc; 573a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 574a4912123STheodore Ts'o 575a4912123STheodore Ts'o /* 576a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 577a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 578a4912123STheodore Ts'o * find another flex group, and store that information in the 579a4912123STheodore Ts'o * parent directory's inode information so that use that flex 580a4912123STheodore Ts'o * group for future allocations. 581a4912123STheodore Ts'o */ 582a4912123STheodore Ts'o if (flex_size > 1) { 583a4912123STheodore Ts'o int retry = 0; 584a4912123STheodore Ts'o 585a4912123STheodore Ts'o try_again: 586a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 587a4912123STheodore Ts'o last = parent_group + flex_size; 588a4912123STheodore Ts'o if (last > ngroups) 589a4912123STheodore Ts'o last = ngroups; 590a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 591a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 592a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 593a4912123STheodore Ts'o *group = i; 594a4912123STheodore Ts'o return 0; 595a4912123STheodore Ts'o } 596a4912123STheodore Ts'o } 597a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 598a4912123STheodore Ts'o retry = 1; 599a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 600a4912123STheodore Ts'o goto try_again; 601a4912123STheodore Ts'o } 602a4912123STheodore Ts'o /* 603a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 604a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 605a4912123STheodore Ts'o * avoid the topdir algorithms. 606a4912123STheodore Ts'o */ 607a4912123STheodore Ts'o *group = parent_group + flex_size; 608a4912123STheodore Ts'o if (*group > ngroups) 609a4912123STheodore Ts'o *group = 0; 6107dc57615SPeter Huewe return find_group_orlov(sb, parent, group, mode, NULL); 611a4912123STheodore Ts'o } 612ac27a0ecSDave Kleikamp 613ac27a0ecSDave Kleikamp /* 614ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 615ac27a0ecSDave Kleikamp */ 6162aa9fc4cSAvantika Mathur *group = parent_group; 6172aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 618560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 619021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc)) 6202aa9fc4cSAvantika Mathur return 0; 621ac27a0ecSDave Kleikamp 622ac27a0ecSDave Kleikamp /* 623ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 624ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 625ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 626ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 627ac27a0ecSDave Kleikamp * different blockgroup. 628ac27a0ecSDave Kleikamp * 629ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 630ac27a0ecSDave Kleikamp */ 6312aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 632ac27a0ecSDave Kleikamp 633ac27a0ecSDave Kleikamp /* 634ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 635ac27a0ecSDave Kleikamp * blocks. 636ac27a0ecSDave Kleikamp */ 637ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6382aa9fc4cSAvantika Mathur *group += i; 6392aa9fc4cSAvantika Mathur if (*group >= ngroups) 6402aa9fc4cSAvantika Mathur *group -= ngroups; 6412aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 642560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 643021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc)) 6442aa9fc4cSAvantika Mathur return 0; 645ac27a0ecSDave Kleikamp } 646ac27a0ecSDave Kleikamp 647ac27a0ecSDave Kleikamp /* 648ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 649ac27a0ecSDave Kleikamp * has no free blocks. 650ac27a0ecSDave Kleikamp */ 6512aa9fc4cSAvantika Mathur *group = parent_group; 652ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6532aa9fc4cSAvantika Mathur if (++*group >= ngroups) 6542aa9fc4cSAvantika Mathur *group = 0; 6552aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 656560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 6572aa9fc4cSAvantika Mathur return 0; 658ac27a0ecSDave Kleikamp } 659ac27a0ecSDave Kleikamp 660ac27a0ecSDave Kleikamp return -1; 661ac27a0ecSDave Kleikamp } 662ac27a0ecSDave Kleikamp 663ac27a0ecSDave Kleikamp /* 66419883bd9STheodore Ts'o * In no journal mode, if an inode has recently been deleted, we want 66519883bd9STheodore Ts'o * to avoid reusing it until we're reasonably sure the inode table 66619883bd9STheodore Ts'o * block has been written back to disk. (Yes, these values are 66719883bd9STheodore Ts'o * somewhat arbitrary...) 66819883bd9STheodore Ts'o */ 669a17a9d93STheodore Ts'o #define RECENTCY_MIN 60 670b5f51573SAndreas Dilger #define RECENTCY_DIRTY 300 67119883bd9STheodore Ts'o 67219883bd9STheodore Ts'o static int recently_deleted(struct super_block *sb, ext4_group_t group, int ino) 67319883bd9STheodore Ts'o { 67419883bd9STheodore Ts'o struct ext4_group_desc *gdp; 67519883bd9STheodore Ts'o struct ext4_inode *raw_inode; 67619883bd9STheodore Ts'o struct buffer_head *bh; 67719883bd9STheodore Ts'o int inodes_per_block = EXT4_SB(sb)->s_inodes_per_block; 678b5f51573SAndreas Dilger int offset, ret = 0; 679b5f51573SAndreas Dilger int recentcy = RECENTCY_MIN; 680b5f51573SAndreas Dilger u32 dtime, now; 68119883bd9STheodore Ts'o 68219883bd9STheodore Ts'o gdp = ext4_get_group_desc(sb, group, NULL); 68319883bd9STheodore Ts'o if (unlikely(!gdp)) 68419883bd9STheodore Ts'o return 0; 68519883bd9STheodore Ts'o 6864f9d956dSJan Kara bh = sb_find_get_block(sb, ext4_inode_table(sb, gdp) + 68719883bd9STheodore Ts'o (ino / inodes_per_block)); 6884f9d956dSJan Kara if (!bh || !buffer_uptodate(bh)) 68919883bd9STheodore Ts'o /* 69019883bd9STheodore Ts'o * If the block is not in the buffer cache, then it 69119883bd9STheodore Ts'o * must have been written out. 69219883bd9STheodore Ts'o */ 69319883bd9STheodore Ts'o goto out; 69419883bd9STheodore Ts'o 69519883bd9STheodore Ts'o offset = (ino % inodes_per_block) * EXT4_INODE_SIZE(sb); 69619883bd9STheodore Ts'o raw_inode = (struct ext4_inode *) (bh->b_data + offset); 697b5f51573SAndreas Dilger 698b5f51573SAndreas Dilger /* i_dtime is only 32 bits on disk, but we only care about relative 699b5f51573SAndreas Dilger * times in the range of a few minutes (i.e. long enough to sync a 700b5f51573SAndreas Dilger * recently-deleted inode to disk), so using the low 32 bits of the 701b5f51573SAndreas Dilger * clock (a 68 year range) is enough, see time_before32() */ 70219883bd9STheodore Ts'o dtime = le32_to_cpu(raw_inode->i_dtime); 703b5f51573SAndreas Dilger now = ktime_get_real_seconds(); 70419883bd9STheodore Ts'o if (buffer_dirty(bh)) 70519883bd9STheodore Ts'o recentcy += RECENTCY_DIRTY; 70619883bd9STheodore Ts'o 707b5f51573SAndreas Dilger if (dtime && time_before32(dtime, now) && 708b5f51573SAndreas Dilger time_before32(now, dtime + recentcy)) 70919883bd9STheodore Ts'o ret = 1; 71019883bd9STheodore Ts'o out: 71119883bd9STheodore Ts'o brelse(bh); 71219883bd9STheodore Ts'o return ret; 71319883bd9STheodore Ts'o } 71419883bd9STheodore Ts'o 715901ed070SWang Shilong static int find_inode_bit(struct super_block *sb, ext4_group_t group, 716901ed070SWang Shilong struct buffer_head *bitmap, unsigned long *ino) 717901ed070SWang Shilong { 718d05466b2SJan Kara bool check_recently_deleted = EXT4_SB(sb)->s_journal == NULL; 719d05466b2SJan Kara unsigned long recently_deleted_ino = EXT4_INODES_PER_GROUP(sb); 720d05466b2SJan Kara 721901ed070SWang Shilong next: 722901ed070SWang Shilong *ino = ext4_find_next_zero_bit((unsigned long *) 723901ed070SWang Shilong bitmap->b_data, 724901ed070SWang Shilong EXT4_INODES_PER_GROUP(sb), *ino); 725901ed070SWang Shilong if (*ino >= EXT4_INODES_PER_GROUP(sb)) 726d05466b2SJan Kara goto not_found; 727901ed070SWang Shilong 728d05466b2SJan Kara if (check_recently_deleted && recently_deleted(sb, group, *ino)) { 729d05466b2SJan Kara recently_deleted_ino = *ino; 730901ed070SWang Shilong *ino = *ino + 1; 731901ed070SWang Shilong if (*ino < EXT4_INODES_PER_GROUP(sb)) 732901ed070SWang Shilong goto next; 733d05466b2SJan Kara goto not_found; 734901ed070SWang Shilong } 735d05466b2SJan Kara return 1; 736d05466b2SJan Kara not_found: 737d05466b2SJan Kara if (recently_deleted_ino >= EXT4_INODES_PER_GROUP(sb)) 738d05466b2SJan Kara return 0; 739d05466b2SJan Kara /* 740d05466b2SJan Kara * Not reusing recently deleted inodes is mostly a preference. We don't 741d05466b2SJan Kara * want to report ENOSPC or skew allocation patterns because of that. 742d05466b2SJan Kara * So return even recently deleted inode if we could find better in the 743d05466b2SJan Kara * given range. 744d05466b2SJan Kara */ 745d05466b2SJan Kara *ino = recently_deleted_ino; 746901ed070SWang Shilong return 1; 747901ed070SWang Shilong } 748901ed070SWang Shilong 7498016e29fSHarshad Shirwadkar int ext4_mark_inode_used(struct super_block *sb, int ino) 7508016e29fSHarshad Shirwadkar { 7518016e29fSHarshad Shirwadkar unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 7528016e29fSHarshad Shirwadkar struct buffer_head *inode_bitmap_bh = NULL, *group_desc_bh = NULL; 7538016e29fSHarshad Shirwadkar struct ext4_group_desc *gdp; 7548016e29fSHarshad Shirwadkar ext4_group_t group; 7558016e29fSHarshad Shirwadkar int bit; 7568016e29fSHarshad Shirwadkar int err = -EFSCORRUPTED; 7578016e29fSHarshad Shirwadkar 7588016e29fSHarshad Shirwadkar if (ino < EXT4_FIRST_INO(sb) || ino > max_ino) 7598016e29fSHarshad Shirwadkar goto out; 7608016e29fSHarshad Shirwadkar 7618016e29fSHarshad Shirwadkar group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 7628016e29fSHarshad Shirwadkar bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 7638016e29fSHarshad Shirwadkar inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 7648016e29fSHarshad Shirwadkar if (IS_ERR(inode_bitmap_bh)) 7658016e29fSHarshad Shirwadkar return PTR_ERR(inode_bitmap_bh); 7668016e29fSHarshad Shirwadkar 7678016e29fSHarshad Shirwadkar if (ext4_test_bit(bit, inode_bitmap_bh->b_data)) { 7688016e29fSHarshad Shirwadkar err = 0; 7698016e29fSHarshad Shirwadkar goto out; 7708016e29fSHarshad Shirwadkar } 7718016e29fSHarshad Shirwadkar 7728016e29fSHarshad Shirwadkar gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 7738016e29fSHarshad Shirwadkar if (!gdp || !group_desc_bh) { 7748016e29fSHarshad Shirwadkar err = -EINVAL; 7758016e29fSHarshad Shirwadkar goto out; 7768016e29fSHarshad Shirwadkar } 7778016e29fSHarshad Shirwadkar 7788016e29fSHarshad Shirwadkar ext4_set_bit(bit, inode_bitmap_bh->b_data); 7798016e29fSHarshad Shirwadkar 7808016e29fSHarshad Shirwadkar BUFFER_TRACE(inode_bitmap_bh, "call ext4_handle_dirty_metadata"); 7818016e29fSHarshad Shirwadkar err = ext4_handle_dirty_metadata(NULL, NULL, inode_bitmap_bh); 7828016e29fSHarshad Shirwadkar if (err) { 7838016e29fSHarshad Shirwadkar ext4_std_error(sb, err); 7848016e29fSHarshad Shirwadkar goto out; 7858016e29fSHarshad Shirwadkar } 7868016e29fSHarshad Shirwadkar err = sync_dirty_buffer(inode_bitmap_bh); 7878016e29fSHarshad Shirwadkar if (err) { 7888016e29fSHarshad Shirwadkar ext4_std_error(sb, err); 7898016e29fSHarshad Shirwadkar goto out; 7908016e29fSHarshad Shirwadkar } 7918016e29fSHarshad Shirwadkar 7928016e29fSHarshad Shirwadkar /* We may have to initialize the block bitmap if it isn't already */ 7938016e29fSHarshad Shirwadkar if (ext4_has_group_desc_csum(sb) && 7948016e29fSHarshad Shirwadkar gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 7958016e29fSHarshad Shirwadkar struct buffer_head *block_bitmap_bh; 7968016e29fSHarshad Shirwadkar 7978016e29fSHarshad Shirwadkar block_bitmap_bh = ext4_read_block_bitmap(sb, group); 7988016e29fSHarshad Shirwadkar if (IS_ERR(block_bitmap_bh)) { 7998016e29fSHarshad Shirwadkar err = PTR_ERR(block_bitmap_bh); 8008016e29fSHarshad Shirwadkar goto out; 8018016e29fSHarshad Shirwadkar } 8028016e29fSHarshad Shirwadkar 8038016e29fSHarshad Shirwadkar BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 8048016e29fSHarshad Shirwadkar err = ext4_handle_dirty_metadata(NULL, NULL, block_bitmap_bh); 8058016e29fSHarshad Shirwadkar sync_dirty_buffer(block_bitmap_bh); 8068016e29fSHarshad Shirwadkar 8078016e29fSHarshad Shirwadkar /* recheck and clear flag under lock if we still need to */ 8088016e29fSHarshad Shirwadkar ext4_lock_group(sb, group); 8098016e29fSHarshad Shirwadkar if (ext4_has_group_desc_csum(sb) && 8108016e29fSHarshad Shirwadkar (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 8118016e29fSHarshad Shirwadkar gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 8128016e29fSHarshad Shirwadkar ext4_free_group_clusters_set(sb, gdp, 8138016e29fSHarshad Shirwadkar ext4_free_clusters_after_init(sb, group, gdp)); 8148016e29fSHarshad Shirwadkar ext4_block_bitmap_csum_set(sb, group, gdp, 8158016e29fSHarshad Shirwadkar block_bitmap_bh); 8168016e29fSHarshad Shirwadkar ext4_group_desc_csum_set(sb, group, gdp); 8178016e29fSHarshad Shirwadkar } 8188016e29fSHarshad Shirwadkar ext4_unlock_group(sb, group); 8198016e29fSHarshad Shirwadkar brelse(block_bitmap_bh); 8208016e29fSHarshad Shirwadkar 8218016e29fSHarshad Shirwadkar if (err) { 8228016e29fSHarshad Shirwadkar ext4_std_error(sb, err); 8238016e29fSHarshad Shirwadkar goto out; 8248016e29fSHarshad Shirwadkar } 8258016e29fSHarshad Shirwadkar } 8268016e29fSHarshad Shirwadkar 8278016e29fSHarshad Shirwadkar /* Update the relevant bg descriptor fields */ 8288016e29fSHarshad Shirwadkar if (ext4_has_group_desc_csum(sb)) { 8298016e29fSHarshad Shirwadkar int free; 8308016e29fSHarshad Shirwadkar 8318016e29fSHarshad Shirwadkar ext4_lock_group(sb, group); /* while we modify the bg desc */ 8328016e29fSHarshad Shirwadkar free = EXT4_INODES_PER_GROUP(sb) - 8338016e29fSHarshad Shirwadkar ext4_itable_unused_count(sb, gdp); 8348016e29fSHarshad Shirwadkar if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 8358016e29fSHarshad Shirwadkar gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 8368016e29fSHarshad Shirwadkar free = 0; 8378016e29fSHarshad Shirwadkar } 8388016e29fSHarshad Shirwadkar 8398016e29fSHarshad Shirwadkar /* 8408016e29fSHarshad Shirwadkar * Check the relative inode number against the last used 8418016e29fSHarshad Shirwadkar * relative inode number in this group. if it is greater 8428016e29fSHarshad Shirwadkar * we need to update the bg_itable_unused count 8438016e29fSHarshad Shirwadkar */ 8448016e29fSHarshad Shirwadkar if (bit >= free) 8458016e29fSHarshad Shirwadkar ext4_itable_unused_set(sb, gdp, 8468016e29fSHarshad Shirwadkar (EXT4_INODES_PER_GROUP(sb) - bit - 1)); 8478016e29fSHarshad Shirwadkar } else { 8488016e29fSHarshad Shirwadkar ext4_lock_group(sb, group); 8498016e29fSHarshad Shirwadkar } 8508016e29fSHarshad Shirwadkar 8518016e29fSHarshad Shirwadkar ext4_free_inodes_set(sb, gdp, ext4_free_inodes_count(sb, gdp) - 1); 8528016e29fSHarshad Shirwadkar if (ext4_has_group_desc_csum(sb)) { 8538016e29fSHarshad Shirwadkar ext4_inode_bitmap_csum_set(sb, group, gdp, inode_bitmap_bh, 8548016e29fSHarshad Shirwadkar EXT4_INODES_PER_GROUP(sb) / 8); 8558016e29fSHarshad Shirwadkar ext4_group_desc_csum_set(sb, group, gdp); 8568016e29fSHarshad Shirwadkar } 8578016e29fSHarshad Shirwadkar 8588016e29fSHarshad Shirwadkar ext4_unlock_group(sb, group); 8598016e29fSHarshad Shirwadkar err = ext4_handle_dirty_metadata(NULL, NULL, group_desc_bh); 8608016e29fSHarshad Shirwadkar sync_dirty_buffer(group_desc_bh); 8618016e29fSHarshad Shirwadkar out: 8628016e29fSHarshad Shirwadkar return err; 8638016e29fSHarshad Shirwadkar } 8648016e29fSHarshad Shirwadkar 865177cc0e7SEric Biggers static int ext4_xattr_credits_for_new_inode(struct inode *dir, mode_t mode, 866177cc0e7SEric Biggers bool encrypt) 867177cc0e7SEric Biggers { 868177cc0e7SEric Biggers struct super_block *sb = dir->i_sb; 869177cc0e7SEric Biggers int nblocks = 0; 870177cc0e7SEric Biggers #ifdef CONFIG_EXT4_FS_POSIX_ACL 871177cc0e7SEric Biggers struct posix_acl *p = get_acl(dir, ACL_TYPE_DEFAULT); 872177cc0e7SEric Biggers 873177cc0e7SEric Biggers if (IS_ERR(p)) 874177cc0e7SEric Biggers return PTR_ERR(p); 875177cc0e7SEric Biggers if (p) { 876177cc0e7SEric Biggers int acl_size = p->a_count * sizeof(ext4_acl_entry); 877177cc0e7SEric Biggers 878177cc0e7SEric Biggers nblocks += (S_ISDIR(mode) ? 2 : 1) * 879177cc0e7SEric Biggers __ext4_xattr_set_credits(sb, NULL /* inode */, 880177cc0e7SEric Biggers NULL /* block_bh */, acl_size, 881177cc0e7SEric Biggers true /* is_create */); 882177cc0e7SEric Biggers posix_acl_release(p); 883177cc0e7SEric Biggers } 884177cc0e7SEric Biggers #endif 885177cc0e7SEric Biggers 886177cc0e7SEric Biggers #ifdef CONFIG_SECURITY 887177cc0e7SEric Biggers { 888177cc0e7SEric Biggers int num_security_xattrs = 1; 889177cc0e7SEric Biggers 890177cc0e7SEric Biggers #ifdef CONFIG_INTEGRITY 891177cc0e7SEric Biggers num_security_xattrs++; 892177cc0e7SEric Biggers #endif 893177cc0e7SEric Biggers /* 894177cc0e7SEric Biggers * We assume that security xattrs are never more than 1k. 895177cc0e7SEric Biggers * In practice they are under 128 bytes. 896177cc0e7SEric Biggers */ 897177cc0e7SEric Biggers nblocks += num_security_xattrs * 898177cc0e7SEric Biggers __ext4_xattr_set_credits(sb, NULL /* inode */, 899177cc0e7SEric Biggers NULL /* block_bh */, 1024, 900177cc0e7SEric Biggers true /* is_create */); 901177cc0e7SEric Biggers } 902177cc0e7SEric Biggers #endif 903177cc0e7SEric Biggers if (encrypt) 904177cc0e7SEric Biggers nblocks += __ext4_xattr_set_credits(sb, 905177cc0e7SEric Biggers NULL /* inode */, 906177cc0e7SEric Biggers NULL /* block_bh */, 907177cc0e7SEric Biggers FSCRYPT_SET_CONTEXT_MAX_SIZE, 908177cc0e7SEric Biggers true /* is_create */); 909177cc0e7SEric Biggers return nblocks; 910177cc0e7SEric Biggers } 911177cc0e7SEric Biggers 91219883bd9STheodore Ts'o /* 913ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 914ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 915ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 916ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 917ac27a0ecSDave Kleikamp * directories already is chosen. 918ac27a0ecSDave Kleikamp * 919ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 920ac27a0ecSDave Kleikamp * group to find a free inode. 921ac27a0ecSDave Kleikamp */ 92214f3db55SChristian Brauner struct inode *__ext4_new_inode(struct user_namespace *mnt_userns, 92314f3db55SChristian Brauner handle_t *handle, struct inode *dir, 9241139575aSTheodore Ts'o umode_t mode, const struct qstr *qstr, 9251b917ed8STahsin Erdogan __u32 goal, uid_t *owner, __u32 i_flags, 9261b917ed8STahsin Erdogan int handle_type, unsigned int line_no, 9271b917ed8STahsin Erdogan int nblocks) 928ac27a0ecSDave Kleikamp { 929ac27a0ecSDave Kleikamp struct super_block *sb; 9303300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 9313300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 9328df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 933ac27a0ecSDave Kleikamp unsigned long ino = 0; 934ac27a0ecSDave Kleikamp struct inode *inode; 935617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 936617ba13bSMingming Cao struct ext4_inode_info *ei; 937617ba13bSMingming Cao struct ext4_sb_info *sbi; 938a7cdadeeSJan Kara int ret2, err; 939ac27a0ecSDave Kleikamp struct inode *ret; 9402aa9fc4cSAvantika Mathur ext4_group_t i; 941772cb7c8SJose R. Santos ext4_group_t flex_group; 9428016e29fSHarshad Shirwadkar struct ext4_group_info *grp = NULL; 94302ce5316SEric Biggers bool encrypt = false; 944ac27a0ecSDave Kleikamp 945ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 946ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 947ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 948ac27a0ecSDave Kleikamp 949af65207cSTahsin Erdogan sb = dir->i_sb; 950af65207cSTahsin Erdogan sbi = EXT4_SB(sb); 951af65207cSTahsin Erdogan 952af65207cSTahsin Erdogan if (unlikely(ext4_forced_shutdown(sbi))) 9530db1ff22STheodore Ts'o return ERR_PTR(-EIO); 9540db1ff22STheodore Ts'o 9558df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 9569bffad1eSTheodore Ts'o trace_ext4_request_inode(dir, mode); 957ac27a0ecSDave Kleikamp inode = new_inode(sb); 958ac27a0ecSDave Kleikamp if (!inode) 959ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 960617ba13bSMingming Cao ei = EXT4_I(inode); 961772cb7c8SJose R. Santos 962eb9cc7e1SJan Kara /* 963b8a07463SAdam Buchbinder * Initialize owners and quota early so that we don't have to account 964eb9cc7e1SJan Kara * for quota initialization worst case in standard inode creating 965eb9cc7e1SJan Kara * transaction 966eb9cc7e1SJan Kara */ 967eb9cc7e1SJan Kara if (owner) { 968eb9cc7e1SJan Kara inode->i_mode = mode; 969eb9cc7e1SJan Kara i_uid_write(inode, owner[0]); 970eb9cc7e1SJan Kara i_gid_write(inode, owner[1]); 971eb9cc7e1SJan Kara } else if (test_opt(sb, GRPID)) { 972eb9cc7e1SJan Kara inode->i_mode = mode; 973*db998553SChristian Brauner inode_fsuid_set(inode, mnt_userns); 974eb9cc7e1SJan Kara inode->i_gid = dir->i_gid; 975eb9cc7e1SJan Kara } else 97614f3db55SChristian Brauner inode_init_owner(mnt_userns, inode, dir, mode); 977040cb378SLi Xi 9780b7b7779SKaho Ng if (ext4_has_feature_project(sb) && 979040cb378SLi Xi ext4_test_inode_flag(dir, EXT4_INODE_PROJINHERIT)) 980040cb378SLi Xi ei->i_projid = EXT4_I(dir)->i_projid; 981040cb378SLi Xi else 982040cb378SLi Xi ei->i_projid = make_kprojid(&init_user_ns, EXT4_DEF_PROJID); 983040cb378SLi Xi 98402ce5316SEric Biggers if (!(i_flags & EXT4_EA_INODE_FL)) { 98502ce5316SEric Biggers err = fscrypt_prepare_new_inode(dir, inode, &encrypt); 98602ce5316SEric Biggers if (err) 98702ce5316SEric Biggers goto out; 98802ce5316SEric Biggers } 98902ce5316SEric Biggers 990a7cdadeeSJan Kara err = dquot_initialize(inode); 991a7cdadeeSJan Kara if (err) 992a7cdadeeSJan Kara goto out; 993eb9cc7e1SJan Kara 99402ce5316SEric Biggers if (!handle && sbi->s_journal && !(i_flags & EXT4_EA_INODE_FL)) { 99502ce5316SEric Biggers ret2 = ext4_xattr_credits_for_new_inode(dir, mode, encrypt); 99602ce5316SEric Biggers if (ret2 < 0) { 99702ce5316SEric Biggers err = ret2; 99802ce5316SEric Biggers goto out; 99902ce5316SEric Biggers } 100002ce5316SEric Biggers nblocks += ret2; 100102ce5316SEric Biggers } 100202ce5316SEric Biggers 100311013911SAndreas Dilger if (!goal) 100411013911SAndreas Dilger goal = sbi->s_inode_goal; 100511013911SAndreas Dilger 1006e6462869SJohann Lombardi if (goal && goal <= le32_to_cpu(sbi->s_es->s_inodes_count)) { 100711013911SAndreas Dilger group = (goal - 1) / EXT4_INODES_PER_GROUP(sb); 100811013911SAndreas Dilger ino = (goal - 1) % EXT4_INODES_PER_GROUP(sb); 100911013911SAndreas Dilger ret2 = 0; 101011013911SAndreas Dilger goto got_group; 101111013911SAndreas Dilger } 101211013911SAndreas Dilger 10134113c4caSLukas Czerner if (S_ISDIR(mode)) 1014f157a4aaSTheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode, qstr); 10154113c4caSLukas Czerner else 1016a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 1017ac27a0ecSDave Kleikamp 1018772cb7c8SJose R. Santos got_group: 1019a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 1020ac27a0ecSDave Kleikamp err = -ENOSPC; 10212aa9fc4cSAvantika Mathur if (ret2 == -1) 1022ac27a0ecSDave Kleikamp goto out; 1023ac27a0ecSDave Kleikamp 1024119c0d44STheodore Ts'o /* 1025119c0d44STheodore Ts'o * Normally we will only go through one pass of this loop, 1026119c0d44STheodore Ts'o * unless we get unlucky and it turns out the group we selected 1027119c0d44STheodore Ts'o * had its last inode grabbed by someone else. 1028119c0d44STheodore Ts'o */ 102911013911SAndreas Dilger for (i = 0; i < ngroups; i++, ino = 0) { 1030ac27a0ecSDave Kleikamp err = -EIO; 1031ac27a0ecSDave Kleikamp 10323300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 1033ac27a0ecSDave Kleikamp if (!gdp) 1034eb9cc7e1SJan Kara goto out; 1035ac27a0ecSDave Kleikamp 1036f2a09af6SYongqiang Yang /* 1037f2a09af6SYongqiang Yang * Check free inodes count before loading bitmap. 1038f2a09af6SYongqiang Yang */ 10392fe435d8SWang Shilong if (ext4_free_inodes_count(sb, gdp) == 0) 10402fe435d8SWang Shilong goto next_group; 1041f2a09af6SYongqiang Yang 10428016e29fSHarshad Shirwadkar if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { 104387a39389SDarrick J. Wong grp = ext4_get_group_info(sb, group); 10448016e29fSHarshad Shirwadkar /* 10458016e29fSHarshad Shirwadkar * Skip groups with already-known suspicious inode 10468016e29fSHarshad Shirwadkar * tables 10478016e29fSHarshad Shirwadkar */ 10482fe435d8SWang Shilong if (EXT4_MB_GRP_IBITMAP_CORRUPT(grp)) 10492fe435d8SWang Shilong goto next_group; 10508016e29fSHarshad Shirwadkar } 105187a39389SDarrick J. Wong 10523300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 10533300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 105487a39389SDarrick J. Wong /* Skip groups with suspicious inode tables */ 10558016e29fSHarshad Shirwadkar if (((!(sbi->s_mount_state & EXT4_FC_REPLAY)) 10568016e29fSHarshad Shirwadkar && EXT4_MB_GRP_IBITMAP_CORRUPT(grp)) || 10579008a58eSDarrick J. Wong IS_ERR(inode_bitmap_bh)) { 10589008a58eSDarrick J. Wong inode_bitmap_bh = NULL; 10592fe435d8SWang Shilong goto next_group; 106087a39389SDarrick J. Wong } 1061ac27a0ecSDave Kleikamp 1062ac27a0ecSDave Kleikamp repeat_in_this_group: 1063901ed070SWang Shilong ret2 = find_inode_bit(sb, group, inode_bitmap_bh, &ino); 1064901ed070SWang Shilong if (!ret2) 1065a34eb503STheodore Ts'o goto next_group; 1066901ed070SWang Shilong 1067119c0d44STheodore Ts'o if (group == 0 && (ino + 1) < EXT4_FIRST_INO(sb)) { 1068119c0d44STheodore Ts'o ext4_error(sb, "reserved inode found cleared - " 1069119c0d44STheodore Ts'o "inode=%lu", ino + 1); 1070206f6d55SWang Shilong ext4_mark_group_bitmap_corrupted(sb, group, 1071206f6d55SWang Shilong EXT4_GROUP_INFO_IBITMAP_CORRUPT); 10722fe435d8SWang Shilong goto next_group; 1073119c0d44STheodore Ts'o } 1074901ed070SWang Shilong 10758016e29fSHarshad Shirwadkar if ((!(sbi->s_mount_state & EXT4_FC_REPLAY)) && !handle) { 10761139575aSTheodore Ts'o BUG_ON(nblocks <= 0); 10771139575aSTheodore Ts'o handle = __ext4_journal_start_sb(dir->i_sb, line_no, 1078a70fd5acSyangerkun handle_type, nblocks, 0, 1079a70fd5acSyangerkun ext4_trans_default_revoke_credits(sb)); 10801139575aSTheodore Ts'o if (IS_ERR(handle)) { 10811139575aSTheodore Ts'o err = PTR_ERR(handle); 1082eb9cc7e1SJan Kara ext4_std_error(sb, err); 1083eb9cc7e1SJan Kara goto out; 10841139575aSTheodore Ts'o } 10851139575aSTheodore Ts'o } 1086ffb5387eSEric Sandeen BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 1087ffb5387eSEric Sandeen err = ext4_journal_get_write_access(handle, inode_bitmap_bh); 1088eb9cc7e1SJan Kara if (err) { 1089eb9cc7e1SJan Kara ext4_std_error(sb, err); 1090eb9cc7e1SJan Kara goto out; 1091eb9cc7e1SJan Kara } 1092119c0d44STheodore Ts'o ext4_lock_group(sb, group); 1093119c0d44STheodore Ts'o ret2 = ext4_test_and_set_bit(ino, inode_bitmap_bh->b_data); 1094901ed070SWang Shilong if (ret2) { 1095901ed070SWang Shilong /* Someone already took the bit. Repeat the search 1096901ed070SWang Shilong * with lock held. 1097901ed070SWang Shilong */ 1098901ed070SWang Shilong ret2 = find_inode_bit(sb, group, inode_bitmap_bh, &ino); 1099901ed070SWang Shilong if (ret2) { 1100901ed070SWang Shilong ext4_set_bit(ino, inode_bitmap_bh->b_data); 1101901ed070SWang Shilong ret2 = 0; 1102901ed070SWang Shilong } else { 1103901ed070SWang Shilong ret2 = 1; /* we didn't grab the inode */ 1104901ed070SWang Shilong } 1105901ed070SWang Shilong } 1106119c0d44STheodore Ts'o ext4_unlock_group(sb, group); 1107119c0d44STheodore Ts'o ino++; /* the inode bitmap is zero-based */ 1108119c0d44STheodore Ts'o if (!ret2) 1109119c0d44STheodore Ts'o goto got; /* we grabbed the inode! */ 1110901ed070SWang Shilong 1111119c0d44STheodore Ts'o if (ino < EXT4_INODES_PER_GROUP(sb)) 1112119c0d44STheodore Ts'o goto repeat_in_this_group; 1113a34eb503STheodore Ts'o next_group: 1114a34eb503STheodore Ts'o if (++group == ngroups) 1115a34eb503STheodore Ts'o group = 0; 1116ac27a0ecSDave Kleikamp } 1117ac27a0ecSDave Kleikamp err = -ENOSPC; 1118ac27a0ecSDave Kleikamp goto out; 1119ac27a0ecSDave Kleikamp 1120ac27a0ecSDave Kleikamp got: 1121ffb5387eSEric Sandeen BUFFER_TRACE(inode_bitmap_bh, "call ext4_handle_dirty_metadata"); 1122ffb5387eSEric Sandeen err = ext4_handle_dirty_metadata(handle, NULL, inode_bitmap_bh); 1123eb9cc7e1SJan Kara if (err) { 1124eb9cc7e1SJan Kara ext4_std_error(sb, err); 1125eb9cc7e1SJan Kara goto out; 1126eb9cc7e1SJan Kara } 1127ffb5387eSEric Sandeen 112861c219f5STheodore Ts'o BUFFER_TRACE(group_desc_bh, "get_write_access"); 112961c219f5STheodore Ts'o err = ext4_journal_get_write_access(handle, group_desc_bh); 113061c219f5STheodore Ts'o if (err) { 113161c219f5STheodore Ts'o ext4_std_error(sb, err); 113261c219f5STheodore Ts'o goto out; 113361c219f5STheodore Ts'o } 113461c219f5STheodore Ts'o 1135717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 1136feb0ab32SDarrick J. Wong if (ext4_has_group_desc_csum(sb) && 1137717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 11383300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 1139717d50e4SAndreas Dilger 11403300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 11419008a58eSDarrick J. Wong if (IS_ERR(block_bitmap_bh)) { 11429008a58eSDarrick J. Wong err = PTR_ERR(block_bitmap_bh); 1143599a9b77SJan Kara goto out; 1144599a9b77SJan Kara } 11453300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 11463300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 1147717d50e4SAndreas Dilger if (err) { 11483300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 1149eb9cc7e1SJan Kara ext4_std_error(sb, err); 1150eb9cc7e1SJan Kara goto out; 1151717d50e4SAndreas Dilger } 1152717d50e4SAndreas Dilger 1153fd034a84STheodore Ts'o BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 1154fd034a84STheodore Ts'o err = ext4_handle_dirty_metadata(handle, NULL, block_bitmap_bh); 1155fd034a84STheodore Ts'o 1156717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 1157fd034a84STheodore Ts'o ext4_lock_group(sb, group); 11588844618dSTheodore Ts'o if (ext4_has_group_desc_csum(sb) && 11598844618dSTheodore Ts'o (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 11603300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 1161021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 1162cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, group, gdp)); 1163fa77dcfaSDarrick J. Wong ext4_block_bitmap_csum_set(sb, group, gdp, 116479f1ba49STao Ma block_bitmap_bh); 1165feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, group, gdp); 1166717d50e4SAndreas Dilger } 1167955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 1168aeb1e5d6STheodore Ts'o brelse(block_bitmap_bh); 1169717d50e4SAndreas Dilger 1170eb9cc7e1SJan Kara if (err) { 1171eb9cc7e1SJan Kara ext4_std_error(sb, err); 1172eb9cc7e1SJan Kara goto out; 1173eb9cc7e1SJan Kara } 1174717d50e4SAndreas Dilger } 1175119c0d44STheodore Ts'o 1176119c0d44STheodore Ts'o /* Update the relevant bg descriptor fields */ 117741a246d1SDarrick J. Wong if (ext4_has_group_desc_csum(sb)) { 1178119c0d44STheodore Ts'o int free; 11798016e29fSHarshad Shirwadkar struct ext4_group_info *grp = NULL; 1180119c0d44STheodore Ts'o 11818016e29fSHarshad Shirwadkar if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { 11828016e29fSHarshad Shirwadkar grp = ext4_get_group_info(sb, group); 11838016e29fSHarshad Shirwadkar down_read(&grp->alloc_sem); /* 11848016e29fSHarshad Shirwadkar * protect vs itable 11858016e29fSHarshad Shirwadkar * lazyinit 11868016e29fSHarshad Shirwadkar */ 11878016e29fSHarshad Shirwadkar } 1188119c0d44STheodore Ts'o ext4_lock_group(sb, group); /* while we modify the bg desc */ 1189119c0d44STheodore Ts'o free = EXT4_INODES_PER_GROUP(sb) - 1190119c0d44STheodore Ts'o ext4_itable_unused_count(sb, gdp); 1191119c0d44STheodore Ts'o if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 1192119c0d44STheodore Ts'o gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 1193119c0d44STheodore Ts'o free = 0; 1194119c0d44STheodore Ts'o } 1195119c0d44STheodore Ts'o /* 1196119c0d44STheodore Ts'o * Check the relative inode number against the last used 1197119c0d44STheodore Ts'o * relative inode number in this group. if it is greater 1198119c0d44STheodore Ts'o * we need to update the bg_itable_unused count 1199119c0d44STheodore Ts'o */ 1200119c0d44STheodore Ts'o if (ino > free) 1201119c0d44STheodore Ts'o ext4_itable_unused_set(sb, gdp, 1202119c0d44STheodore Ts'o (EXT4_INODES_PER_GROUP(sb) - ino)); 12038016e29fSHarshad Shirwadkar if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) 1204119c0d44STheodore Ts'o up_read(&grp->alloc_sem); 12056f2e9f0eSTao Ma } else { 12066f2e9f0eSTao Ma ext4_lock_group(sb, group); 1207119c0d44STheodore Ts'o } 12086f2e9f0eSTao Ma 1209119c0d44STheodore Ts'o ext4_free_inodes_set(sb, gdp, ext4_free_inodes_count(sb, gdp) - 1); 1210119c0d44STheodore Ts'o if (S_ISDIR(mode)) { 1211119c0d44STheodore Ts'o ext4_used_dirs_set(sb, gdp, ext4_used_dirs_count(sb, gdp) + 1); 1212119c0d44STheodore Ts'o if (sbi->s_log_groups_per_flex) { 1213119c0d44STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 1214119c0d44STheodore Ts'o 12157c990728SSuraj Jitindar Singh atomic_inc(&sbi_array_rcu_deref(sbi, s_flex_groups, 12167c990728SSuraj Jitindar Singh f)->used_dirs); 1217119c0d44STheodore Ts'o } 1218119c0d44STheodore Ts'o } 121941a246d1SDarrick J. Wong if (ext4_has_group_desc_csum(sb)) { 122041a246d1SDarrick J. Wong ext4_inode_bitmap_csum_set(sb, group, gdp, inode_bitmap_bh, 122141a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8); 1222feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, group, gdp); 1223119c0d44STheodore Ts'o } 12246f2e9f0eSTao Ma ext4_unlock_group(sb, group); 1225119c0d44STheodore Ts'o 12263300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 12273300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 1228eb9cc7e1SJan Kara if (err) { 1229eb9cc7e1SJan Kara ext4_std_error(sb, err); 1230eb9cc7e1SJan Kara goto out; 1231eb9cc7e1SJan Kara } 1232ac27a0ecSDave Kleikamp 1233ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 1234ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 1235ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 1236ac27a0ecSDave Kleikamp 1237772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 1238772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 12397c990728SSuraj Jitindar Singh atomic_dec(&sbi_array_rcu_deref(sbi, s_flex_groups, 12407c990728SSuraj Jitindar Singh flex_group)->free_inodes); 1241772cb7c8SJose R. Santos } 1242ac27a0ecSDave Kleikamp 1243717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 1244ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 1245ac27a0ecSDave Kleikamp inode->i_blocks = 0; 124695582b00SDeepa Dinamani inode->i_mtime = inode->i_atime = inode->i_ctime = current_time(inode); 12477b62b293SArnd Bergmann ei->i_crtime = inode->i_mtime; 1248ac27a0ecSDave Kleikamp 1249ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 1250ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 1251ac27a0ecSDave Kleikamp ei->i_disksize = 0; 1252ac27a0ecSDave Kleikamp 12534af83508SEryu Guan /* Don't inherit extent flag from directory, amongst others. */ 12542dc6b0d4SDuane Griffin ei->i_flags = 12552dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 12561b917ed8STahsin Erdogan ei->i_flags |= i_flags; 1257ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 1258ac27a0ecSDave Kleikamp ei->i_dtime = 0; 1259ac27a0ecSDave Kleikamp ei->i_block_group = group; 1260a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 1261ac27a0ecSDave Kleikamp 1262043546e4SIra Weiny ext4_set_inode_flags(inode, true); 1263ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 12640390131bSFrank Mayhar ext4_handle_sync(handle); 12656b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 1266acd6ad83SJan Kara /* 1267acd6ad83SJan Kara * Likely a bitmap corruption causing inode to be allocated 1268acd6ad83SJan Kara * twice. 1269acd6ad83SJan Kara */ 1270acd6ad83SJan Kara err = -EIO; 1271eb9cc7e1SJan Kara ext4_error(sb, "failed to insert inode %lu: doubly allocated?", 1272eb9cc7e1SJan Kara inode->i_ino); 1273206f6d55SWang Shilong ext4_mark_group_bitmap_corrupted(sb, group, 1274206f6d55SWang Shilong EXT4_GROUP_INFO_IBITMAP_CORRUPT); 1275eb9cc7e1SJan Kara goto out; 12766b38e842SAl Viro } 127723253068STheodore Ts'o inode->i_generation = prandom_u32(); 1278ac27a0ecSDave Kleikamp 1279814525f4SDarrick J. Wong /* Precompute checksum seed for inode metadata */ 12809aa5d32bSDmitry Monakhov if (ext4_has_metadata_csum(sb)) { 1281814525f4SDarrick J. Wong __u32 csum; 1282814525f4SDarrick J. Wong __le32 inum = cpu_to_le32(inode->i_ino); 1283814525f4SDarrick J. Wong __le32 gen = cpu_to_le32(inode->i_generation); 1284814525f4SDarrick J. Wong csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum, 1285814525f4SDarrick J. Wong sizeof(inum)); 1286814525f4SDarrick J. Wong ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen, 1287814525f4SDarrick J. Wong sizeof(gen)); 1288814525f4SDarrick J. Wong } 1289814525f4SDarrick J. Wong 1290353eb83cSTheodore Ts'o ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */ 129119f5fb7aSTheodore Ts'o ext4_set_inode_state(inode, EXT4_STATE_NEW); 1292ef7f3835SKalpak Shah 129349598e04SJun Piao ei->i_extra_isize = sbi->s_want_extra_isize; 1294f08225d1STao Ma ei->i_inline_off = 0; 1295e2b911c5SDarrick J. Wong if (ext4_has_feature_inline_data(sb)) 1296f08225d1STao Ma ext4_set_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA); 1297ac27a0ecSDave Kleikamp ret = inode; 129863936ddaSChristoph Hellwig err = dquot_alloc_inode(inode); 129963936ddaSChristoph Hellwig if (err) 1300ac27a0ecSDave Kleikamp goto fail_drop; 1301ac27a0ecSDave Kleikamp 1302aa1dca3bSEric Biggers /* 1303aa1dca3bSEric Biggers * Since the encryption xattr will always be unique, create it first so 1304aa1dca3bSEric Biggers * that it's less likely to end up in an external xattr block and 1305aa1dca3bSEric Biggers * prevent its deduplication. 1306aa1dca3bSEric Biggers */ 1307aa1dca3bSEric Biggers if (encrypt) { 130802ce5316SEric Biggers err = fscrypt_set_context(inode, handle); 1309aa1dca3bSEric Biggers if (err) 1310aa1dca3bSEric Biggers goto fail_free_drop; 1311aa1dca3bSEric Biggers } 1312aa1dca3bSEric Biggers 13131b917ed8STahsin Erdogan if (!(ei->i_flags & EXT4_EA_INODE_FL)) { 1314617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1315ac27a0ecSDave Kleikamp if (err) 1316ac27a0ecSDave Kleikamp goto fail_free_drop; 1317ac27a0ecSDave Kleikamp 13182a7dba39SEric Paris err = ext4_init_security(handle, inode, dir, qstr); 1319ac27a0ecSDave Kleikamp if (err) 1320ac27a0ecSDave Kleikamp goto fail_free_drop; 1321ad47f953STahsin Erdogan } 1322ac27a0ecSDave Kleikamp 1323e2b911c5SDarrick J. Wong if (ext4_has_feature_extents(sb)) { 1324e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1325e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 132612e9b892SDmitry Monakhov ext4_set_inode_flag(inode, EXT4_INODE_EXTENTS); 1327a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1328a86c6181SAlex Tomas } 132942bf0383SAneesh Kumar K.V } 1330ac27a0ecSDave Kleikamp 1331688f869cSTheodore Ts'o if (ext4_handle_valid(handle)) { 1332688f869cSTheodore Ts'o ei->i_sync_tid = handle->h_transaction->t_tid; 1333688f869cSTheodore Ts'o ei->i_datasync_tid = handle->h_transaction->t_tid; 1334688f869cSTheodore Ts'o } 1335688f869cSTheodore Ts'o 13368753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 13378753e88fSAneesh Kumar K.V if (err) { 13388753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 13398753e88fSAneesh Kumar K.V goto fail_free_drop; 13408753e88fSAneesh Kumar K.V } 13418753e88fSAneesh Kumar K.V 1342617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 13439bffad1eSTheodore Ts'o trace_ext4_allocate_inode(inode, dir, mode); 13443300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1345ac27a0ecSDave Kleikamp return ret; 1346ac27a0ecSDave Kleikamp 1347ac27a0ecSDave Kleikamp fail_free_drop: 134863936ddaSChristoph Hellwig dquot_free_inode(inode); 1349ac27a0ecSDave Kleikamp fail_drop: 13506d6b77f1SMiklos Szeredi clear_nlink(inode); 13516b38e842SAl Viro unlock_new_inode(inode); 1352eb9cc7e1SJan Kara out: 1353eb9cc7e1SJan Kara dquot_drop(inode); 1354eb9cc7e1SJan Kara inode->i_flags |= S_NOQUOTA; 1355ac27a0ecSDave Kleikamp iput(inode); 13563300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1357ac27a0ecSDave Kleikamp return ERR_PTR(err); 1358ac27a0ecSDave Kleikamp } 1359ac27a0ecSDave Kleikamp 1360ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1361617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1362ac27a0ecSDave Kleikamp { 1363617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1364fd2d4291SAvantika Mathur ext4_group_t block_group; 1365ac27a0ecSDave Kleikamp int bit; 13667827a7f6STheodore Ts'o struct buffer_head *bitmap_bh = NULL; 1367ac27a0ecSDave Kleikamp struct inode *inode = NULL; 13687827a7f6STheodore Ts'o int err = -EFSCORRUPTED; 1369ac27a0ecSDave Kleikamp 13707827a7f6STheodore Ts'o if (ino < EXT4_FIRST_INO(sb) || ino > max_ino) 13717827a7f6STheodore Ts'o goto bad_orphan; 1372ac27a0ecSDave Kleikamp 1373617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1374617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1375e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1376206f6d55SWang Shilong if (IS_ERR(bitmap_bh)) 1377e884bce1SAl Viro return ERR_CAST(bitmap_bh); 1378ac27a0ecSDave Kleikamp 1379ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1380ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1381ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1382ac27a0ecSDave Kleikamp */ 13831d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 13841d1fe1eeSDavid Howells goto bad_orphan; 13851d1fe1eeSDavid Howells 13868a363970STheodore Ts'o inode = ext4_iget(sb, ino, EXT4_IGET_NORMAL); 13877827a7f6STheodore Ts'o if (IS_ERR(inode)) { 13887827a7f6STheodore Ts'o err = PTR_ERR(inode); 138954d3adbcSTheodore Ts'o ext4_error_err(sb, -err, 139054d3adbcSTheodore Ts'o "couldn't read orphan inode %lu (err %d)", 13917827a7f6STheodore Ts'o ino, err); 13923bbd0ef2SXiyu Yang brelse(bitmap_bh); 13937827a7f6STheodore Ts'o return inode; 13947827a7f6STheodore Ts'o } 13951d1fe1eeSDavid Howells 139691ef4cafSDuane Griffin /* 1397c9eb13a9STheodore Ts'o * If the orphans has i_nlinks > 0 then it should be able to 1398c9eb13a9STheodore Ts'o * be truncated, otherwise it won't be removed from the orphan 1399c9eb13a9STheodore Ts'o * list during processing and an infinite loop will result. 1400c9eb13a9STheodore Ts'o * Similarly, it must not be a bad inode. 140191ef4cafSDuane Griffin */ 1402c9eb13a9STheodore Ts'o if ((inode->i_nlink && !ext4_can_truncate(inode)) || 1403c9eb13a9STheodore Ts'o is_bad_inode(inode)) 140491ef4cafSDuane Griffin goto bad_orphan; 140591ef4cafSDuane Griffin 14061d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 14071d1fe1eeSDavid Howells goto bad_orphan; 14081d1fe1eeSDavid Howells brelse(bitmap_bh); 14091d1fe1eeSDavid Howells return inode; 14101d1fe1eeSDavid Howells 14111d1fe1eeSDavid Howells bad_orphan: 14127827a7f6STheodore Ts'o ext4_error(sb, "bad orphan inode %lu", ino); 14137827a7f6STheodore Ts'o if (bitmap_bh) 14147827a7f6STheodore Ts'o printk(KERN_ERR "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1415ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1416617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1417ac27a0ecSDave Kleikamp if (inode) { 14187827a7f6STheodore Ts'o printk(KERN_ERR "is_bad_inode(inode)=%d\n", 1419ac27a0ecSDave Kleikamp is_bad_inode(inode)); 14207827a7f6STheodore Ts'o printk(KERN_ERR "NEXT_ORPHAN(inode)=%u\n", 1421ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 14227827a7f6STheodore Ts'o printk(KERN_ERR "max_ino=%lu\n", max_ino); 14237827a7f6STheodore Ts'o printk(KERN_ERR "i_nlink=%u\n", inode->i_nlink); 1424ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 14251d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1426ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1427ac27a0ecSDave Kleikamp iput(inode); 1428ac27a0ecSDave Kleikamp } 1429ac27a0ecSDave Kleikamp brelse(bitmap_bh); 14301d1fe1eeSDavid Howells return ERR_PTR(err); 1431ac27a0ecSDave Kleikamp } 1432ac27a0ecSDave Kleikamp 1433617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1434ac27a0ecSDave Kleikamp { 1435ac27a0ecSDave Kleikamp unsigned long desc_count; 1436617ba13bSMingming Cao struct ext4_group_desc *gdp; 14378df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1438617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1439617ba13bSMingming Cao struct ext4_super_block *es; 1440ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1441ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1442ac27a0ecSDave Kleikamp 1443617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1444ac27a0ecSDave Kleikamp desc_count = 0; 1445ac27a0ecSDave Kleikamp bitmap_count = 0; 1446ac27a0ecSDave Kleikamp gdp = NULL; 14478df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1448617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1449ac27a0ecSDave Kleikamp if (!gdp) 1450ac27a0ecSDave Kleikamp continue; 1451560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1452ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1453e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 14549008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 14559008a58eSDarrick J. Wong bitmap_bh = NULL; 1456ac27a0ecSDave Kleikamp continue; 14579008a58eSDarrick J. Wong } 1458ac27a0ecSDave Kleikamp 1459f6fb99caSTheodore Ts'o x = ext4_count_free(bitmap_bh->b_data, 1460f6fb99caSTheodore Ts'o EXT4_INODES_PER_GROUP(sb) / 8); 1461c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1462785b4b3aSPeng Tao (unsigned long) i, ext4_free_inodes_count(sb, gdp), x); 1463ac27a0ecSDave Kleikamp bitmap_count += x; 1464ac27a0ecSDave Kleikamp } 1465ac27a0ecSDave Kleikamp brelse(bitmap_bh); 14664776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 14674776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1468ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1469ac27a0ecSDave Kleikamp return desc_count; 1470ac27a0ecSDave Kleikamp #else 1471ac27a0ecSDave Kleikamp desc_count = 0; 14728df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1473617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1474ac27a0ecSDave Kleikamp if (!gdp) 1475ac27a0ecSDave Kleikamp continue; 1476560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1477ac27a0ecSDave Kleikamp cond_resched(); 1478ac27a0ecSDave Kleikamp } 1479ac27a0ecSDave Kleikamp return desc_count; 1480ac27a0ecSDave Kleikamp #endif 1481ac27a0ecSDave Kleikamp } 1482ac27a0ecSDave Kleikamp 1483ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1484617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1485ac27a0ecSDave Kleikamp { 1486ac27a0ecSDave Kleikamp unsigned long count = 0; 14878df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1488ac27a0ecSDave Kleikamp 14898df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1490617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1491ac27a0ecSDave Kleikamp if (!gdp) 1492ac27a0ecSDave Kleikamp continue; 1493560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1494ac27a0ecSDave Kleikamp } 1495ac27a0ecSDave Kleikamp return count; 1496ac27a0ecSDave Kleikamp } 1497bfff6873SLukas Czerner 1498bfff6873SLukas Czerner /* 1499bfff6873SLukas Czerner * Zeroes not yet zeroed inode table - just write zeroes through the whole 1500bfff6873SLukas Czerner * inode table. Must be called without any spinlock held. The only place 1501bfff6873SLukas Czerner * where it is called from on active part of filesystem is ext4lazyinit 1502bfff6873SLukas Czerner * thread, so we do not need any special locks, however we have to prevent 1503bfff6873SLukas Czerner * inode allocation from the current group, so we take alloc_sem lock, to 1504119c0d44STheodore Ts'o * block ext4_new_inode() until we are finished. 1505bfff6873SLukas Czerner */ 1506e0cbee3eSH Hartley Sweeten int ext4_init_inode_table(struct super_block *sb, ext4_group_t group, 1507bfff6873SLukas Czerner int barrier) 1508bfff6873SLukas Czerner { 1509bfff6873SLukas Czerner struct ext4_group_info *grp = ext4_get_group_info(sb, group); 1510bfff6873SLukas Czerner struct ext4_sb_info *sbi = EXT4_SB(sb); 1511bfff6873SLukas Czerner struct ext4_group_desc *gdp = NULL; 1512bfff6873SLukas Czerner struct buffer_head *group_desc_bh; 1513bfff6873SLukas Czerner handle_t *handle; 1514bfff6873SLukas Czerner ext4_fsblk_t blk; 1515bfff6873SLukas Czerner int num, ret = 0, used_blks = 0; 1516bfff6873SLukas Czerner 1517bfff6873SLukas Czerner /* This should not happen, but just to be sure check this */ 1518bc98a42cSDavid Howells if (sb_rdonly(sb)) { 1519bfff6873SLukas Czerner ret = 1; 1520bfff6873SLukas Czerner goto out; 1521bfff6873SLukas Czerner } 1522bfff6873SLukas Czerner 1523bfff6873SLukas Czerner gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 1524bfff6873SLukas Czerner if (!gdp) 1525bfff6873SLukas Czerner goto out; 1526bfff6873SLukas Czerner 1527bfff6873SLukas Czerner /* 1528bfff6873SLukas Czerner * We do not need to lock this, because we are the only one 1529bfff6873SLukas Czerner * handling this flag. 1530bfff6873SLukas Czerner */ 1531bfff6873SLukas Czerner if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_ZEROED)) 1532bfff6873SLukas Czerner goto out; 1533bfff6873SLukas Czerner 15349924a92aSTheodore Ts'o handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1); 1535bfff6873SLukas Czerner if (IS_ERR(handle)) { 1536bfff6873SLukas Czerner ret = PTR_ERR(handle); 1537bfff6873SLukas Czerner goto out; 1538bfff6873SLukas Czerner } 1539bfff6873SLukas Czerner 1540bfff6873SLukas Czerner down_write(&grp->alloc_sem); 1541bfff6873SLukas Czerner /* 1542bfff6873SLukas Czerner * If inode bitmap was already initialized there may be some 1543bfff6873SLukas Czerner * used inodes so we need to skip blocks with used inodes in 1544bfff6873SLukas Czerner * inode table. 1545bfff6873SLukas Czerner */ 1546bfff6873SLukas Czerner if (!(gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT))) 1547bfff6873SLukas Czerner used_blks = DIV_ROUND_UP((EXT4_INODES_PER_GROUP(sb) - 1548bfff6873SLukas Czerner ext4_itable_unused_count(sb, gdp)), 1549bfff6873SLukas Czerner sbi->s_inodes_per_block); 1550bfff6873SLukas Czerner 155150122847STheodore Ts'o if ((used_blks < 0) || (used_blks > sbi->s_itb_per_group) || 155250122847STheodore Ts'o ((group == 0) && ((EXT4_INODES_PER_GROUP(sb) - 155350122847STheodore Ts'o ext4_itable_unused_count(sb, gdp)) < 155450122847STheodore Ts'o EXT4_FIRST_INO(sb)))) { 15551084f252STheodore Ts'o ext4_error(sb, "Something is wrong with group %u: " 15561084f252STheodore Ts'o "used itable blocks: %d; " 15571084f252STheodore Ts'o "itable unused count: %u", 1558857ac889SLukas Czerner group, used_blks, 1559857ac889SLukas Czerner ext4_itable_unused_count(sb, gdp)); 1560857ac889SLukas Czerner ret = 1; 156133853a0dSYongqiang Yang goto err_out; 1562857ac889SLukas Czerner } 1563857ac889SLukas Czerner 1564bfff6873SLukas Czerner blk = ext4_inode_table(sb, gdp) + used_blks; 1565bfff6873SLukas Czerner num = sbi->s_itb_per_group - used_blks; 1566bfff6873SLukas Czerner 1567bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, "get_write_access"); 1568bfff6873SLukas Czerner ret = ext4_journal_get_write_access(handle, 1569bfff6873SLukas Czerner group_desc_bh); 1570bfff6873SLukas Czerner if (ret) 1571bfff6873SLukas Czerner goto err_out; 1572bfff6873SLukas Czerner 1573bfff6873SLukas Czerner /* 1574bfff6873SLukas Czerner * Skip zeroout if the inode table is full. But we set the ZEROED 1575bfff6873SLukas Czerner * flag anyway, because obviously, when it is full it does not need 1576bfff6873SLukas Czerner * further zeroing. 1577bfff6873SLukas Czerner */ 1578bfff6873SLukas Czerner if (unlikely(num == 0)) 1579bfff6873SLukas Czerner goto skip_zeroout; 1580bfff6873SLukas Czerner 1581bfff6873SLukas Czerner ext4_debug("going to zero out inode table in group %d\n", 1582bfff6873SLukas Czerner group); 1583a107e5a3STheodore Ts'o ret = sb_issue_zeroout(sb, blk, num, GFP_NOFS); 1584bfff6873SLukas Czerner if (ret < 0) 1585bfff6873SLukas Czerner goto err_out; 1586a107e5a3STheodore Ts'o if (barrier) 1587c6bf3f0eSChristoph Hellwig blkdev_issue_flush(sb->s_bdev); 1588bfff6873SLukas Czerner 1589bfff6873SLukas Czerner skip_zeroout: 1590bfff6873SLukas Czerner ext4_lock_group(sb, group); 1591bfff6873SLukas Czerner gdp->bg_flags |= cpu_to_le16(EXT4_BG_INODE_ZEROED); 1592feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, group, gdp); 1593bfff6873SLukas Czerner ext4_unlock_group(sb, group); 1594bfff6873SLukas Czerner 1595bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, 1596bfff6873SLukas Czerner "call ext4_handle_dirty_metadata"); 1597bfff6873SLukas Czerner ret = ext4_handle_dirty_metadata(handle, NULL, 1598bfff6873SLukas Czerner group_desc_bh); 1599bfff6873SLukas Czerner 1600bfff6873SLukas Czerner err_out: 1601bfff6873SLukas Czerner up_write(&grp->alloc_sem); 1602bfff6873SLukas Czerner ext4_journal_stop(handle); 1603bfff6873SLukas Czerner out: 1604bfff6873SLukas Czerner return ret; 1605bfff6873SLukas Czerner } 1606