1ac27a0ecSDave Kleikamp /* 2617ba13bSMingming Cao * linux/fs/ext4/ialloc.c 3ac27a0ecSDave Kleikamp * 4ac27a0ecSDave Kleikamp * Copyright (C) 1992, 1993, 1994, 1995 5ac27a0ecSDave Kleikamp * Remy Card (card@masi.ibp.fr) 6ac27a0ecSDave Kleikamp * Laboratoire MASI - Institut Blaise Pascal 7ac27a0ecSDave Kleikamp * Universite Pierre et Marie Curie (Paris VI) 8ac27a0ecSDave Kleikamp * 9ac27a0ecSDave Kleikamp * BSD ufs-inspired inode and directory allocation by 10ac27a0ecSDave Kleikamp * Stephen Tweedie (sct@redhat.com), 1993 11ac27a0ecSDave Kleikamp * Big-endian to little-endian byte-swapping/bitmaps by 12ac27a0ecSDave Kleikamp * David S. Miller (davem@caip.rutgers.edu), 1995 13ac27a0ecSDave Kleikamp */ 14ac27a0ecSDave Kleikamp 15ac27a0ecSDave Kleikamp #include <linux/time.h> 16ac27a0ecSDave Kleikamp #include <linux/fs.h> 17dab291afSMingming Cao #include <linux/jbd2.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> 25ac27a0ecSDave Kleikamp #include <asm/byteorder.h> 269bffad1eSTheodore Ts'o 273dcf5451SChristoph Hellwig #include "ext4.h" 283dcf5451SChristoph Hellwig #include "ext4_jbd2.h" 29ac27a0ecSDave Kleikamp #include "xattr.h" 30ac27a0ecSDave Kleikamp #include "acl.h" 31ac27a0ecSDave Kleikamp 329bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 339bffad1eSTheodore Ts'o 34ac27a0ecSDave Kleikamp /* 35ac27a0ecSDave Kleikamp * ialloc.c contains the inodes allocation and deallocation routines 36ac27a0ecSDave Kleikamp */ 37ac27a0ecSDave Kleikamp 38ac27a0ecSDave Kleikamp /* 39ac27a0ecSDave Kleikamp * The free inodes are managed by bitmaps. A file system contains several 40ac27a0ecSDave Kleikamp * blocks groups. Each group contains 1 bitmap block for blocks, 1 bitmap 41ac27a0ecSDave Kleikamp * block for inodes, N blocks for the inode table and data blocks. 42ac27a0ecSDave Kleikamp * 43ac27a0ecSDave Kleikamp * The file system contains group descriptors which are located after the 44ac27a0ecSDave Kleikamp * super block. Each descriptor contains the number of the bitmap block and 45ac27a0ecSDave Kleikamp * the free blocks count in the block. 46ac27a0ecSDave Kleikamp */ 47ac27a0ecSDave Kleikamp 48717d50e4SAndreas Dilger /* 49717d50e4SAndreas Dilger * To avoid calling the atomic setbit hundreds or thousands of times, we only 50717d50e4SAndreas Dilger * need to use it within a single byte (to ensure we get endianness right). 51717d50e4SAndreas Dilger * We can use memset for the rest of the bitmap as there are no other users. 52717d50e4SAndreas Dilger */ 5361d08673STheodore Ts'o void ext4_mark_bitmap_end(int start_bit, int end_bit, char *bitmap) 54717d50e4SAndreas Dilger { 55717d50e4SAndreas Dilger int i; 56717d50e4SAndreas Dilger 57717d50e4SAndreas Dilger if (start_bit >= end_bit) 58717d50e4SAndreas Dilger return; 59717d50e4SAndreas Dilger 60717d50e4SAndreas Dilger ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit); 61717d50e4SAndreas Dilger for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++) 62717d50e4SAndreas Dilger ext4_set_bit(i, bitmap); 63717d50e4SAndreas Dilger if (i < end_bit) 64717d50e4SAndreas Dilger memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3); 65717d50e4SAndreas Dilger } 66717d50e4SAndreas Dilger 67717d50e4SAndreas Dilger /* Initializes an uninitialized inode bitmap */ 681f109d5aSTheodore Ts'o static unsigned ext4_init_inode_bitmap(struct super_block *sb, 691f109d5aSTheodore Ts'o struct buffer_head *bh, 70fd2d4291SAvantika Mathur ext4_group_t block_group, 71717d50e4SAndreas Dilger struct ext4_group_desc *gdp) 72717d50e4SAndreas Dilger { 73717d50e4SAndreas Dilger struct ext4_sb_info *sbi = EXT4_SB(sb); 74717d50e4SAndreas Dilger 75717d50e4SAndreas Dilger J_ASSERT_BH(bh, buffer_locked(bh)); 76717d50e4SAndreas Dilger 77717d50e4SAndreas Dilger /* If checksum is bad mark all blocks and inodes use to prevent 78717d50e4SAndreas Dilger * allocation, essentially implementing a per-group read-only flag. */ 79717d50e4SAndreas Dilger if (!ext4_group_desc_csum_verify(sbi, block_group, gdp)) { 8012062dddSEric Sandeen ext4_error(sb, "Checksum bad for group %u", block_group); 81560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, 0); 82560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, 0); 83560671a0SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 0); 84717d50e4SAndreas Dilger memset(bh->b_data, 0xff, sb->s_blocksize); 85717d50e4SAndreas Dilger return 0; 86717d50e4SAndreas Dilger } 87717d50e4SAndreas Dilger 88717d50e4SAndreas Dilger memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8); 8961d08673STheodore Ts'o ext4_mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), sb->s_blocksize * 8, 90717d50e4SAndreas Dilger bh->b_data); 91717d50e4SAndreas Dilger 92717d50e4SAndreas Dilger return EXT4_INODES_PER_GROUP(sb); 93717d50e4SAndreas Dilger } 94ac27a0ecSDave Kleikamp 95ac27a0ecSDave Kleikamp /* 96ac27a0ecSDave Kleikamp * Read the inode allocation bitmap for a given block_group, reading 97ac27a0ecSDave Kleikamp * into the specified slot in the superblock's bitmap cache. 98ac27a0ecSDave Kleikamp * 99ac27a0ecSDave Kleikamp * Return buffer_head of bitmap on success or NULL. 100ac27a0ecSDave Kleikamp */ 101ac27a0ecSDave Kleikamp static struct buffer_head * 102e29d1cdeSEric Sandeen ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group) 103ac27a0ecSDave Kleikamp { 104617ba13bSMingming Cao struct ext4_group_desc *desc; 105ac27a0ecSDave Kleikamp struct buffer_head *bh = NULL; 106e29d1cdeSEric Sandeen ext4_fsblk_t bitmap_blk; 107ac27a0ecSDave Kleikamp 108617ba13bSMingming Cao desc = ext4_get_group_desc(sb, block_group, NULL); 109ac27a0ecSDave Kleikamp if (!desc) 110e29d1cdeSEric Sandeen return NULL; 111bfff6873SLukas Czerner 112e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 113e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 114e29d1cdeSEric Sandeen if (unlikely(!bh)) { 11512062dddSEric Sandeen ext4_error(sb, "Cannot read inode bitmap - " 116a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 117e29d1cdeSEric Sandeen block_group, bitmap_blk); 118e29d1cdeSEric Sandeen return NULL; 119e29d1cdeSEric Sandeen } 1202ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) 121e29d1cdeSEric Sandeen return bh; 122e29d1cdeSEric Sandeen 123c806e68fSFrederic Bohe lock_buffer(bh); 1242ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) { 1252ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1262ccb5fb9SAneesh Kumar K.V return bh; 1272ccb5fb9SAneesh Kumar K.V } 128bfff6873SLukas Czerner 129955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 130e29d1cdeSEric Sandeen if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 131e29d1cdeSEric Sandeen ext4_init_inode_bitmap(sb, bh, block_group, desc); 1322ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 133e29d1cdeSEric Sandeen set_buffer_uptodate(bh); 134955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1353300bedaSAneesh Kumar K.V unlock_buffer(bh); 136e29d1cdeSEric Sandeen return bh; 137e29d1cdeSEric Sandeen } 138955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 139bfff6873SLukas Czerner 1402ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh)) { 1412ccb5fb9SAneesh Kumar K.V /* 1422ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 1432ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 1442ccb5fb9SAneesh Kumar K.V */ 1452ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 1462ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1472ccb5fb9SAneesh Kumar K.V return bh; 1482ccb5fb9SAneesh Kumar K.V } 1492ccb5fb9SAneesh Kumar K.V /* 1502ccb5fb9SAneesh Kumar K.V * submit the buffer_head for read. We can 1512ccb5fb9SAneesh Kumar K.V * safely mark the bitmap as uptodate now. 1522ccb5fb9SAneesh Kumar K.V * We do it here so the bitmap uptodate bit 1532ccb5fb9SAneesh Kumar K.V * get set with buffer lock held. 1542ccb5fb9SAneesh Kumar K.V */ 1552ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 156e29d1cdeSEric Sandeen if (bh_submit_read(bh) < 0) { 157e29d1cdeSEric Sandeen put_bh(bh); 15812062dddSEric Sandeen ext4_error(sb, "Cannot read inode bitmap - " 159a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 160e29d1cdeSEric Sandeen block_group, bitmap_blk); 161e29d1cdeSEric Sandeen return NULL; 162e29d1cdeSEric Sandeen } 163ac27a0ecSDave Kleikamp return bh; 164ac27a0ecSDave Kleikamp } 165ac27a0ecSDave Kleikamp 166ac27a0ecSDave Kleikamp /* 167ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 168ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 169ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 170ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 171ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 172ac27a0ecSDave Kleikamp * has been deleted earlier. 173ac27a0ecSDave Kleikamp * 174ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 175ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 176ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 177ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 178ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 179ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 180ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 181ac27a0ecSDave Kleikamp */ 182617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 183ac27a0ecSDave Kleikamp { 184ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 185ac27a0ecSDave Kleikamp int is_directory; 186ac27a0ecSDave Kleikamp unsigned long ino; 187ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 188ac27a0ecSDave Kleikamp struct buffer_head *bh2; 189fd2d4291SAvantika Mathur ext4_group_t block_group; 190ac27a0ecSDave Kleikamp unsigned long bit; 191617ba13bSMingming Cao struct ext4_group_desc *gdp; 192617ba13bSMingming Cao struct ext4_super_block *es; 193617ba13bSMingming Cao struct ext4_sb_info *sbi; 1947ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 195ac27a0ecSDave Kleikamp 196ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 1974776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has count=%d\n", 198ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 199ac27a0ecSDave Kleikamp return; 200ac27a0ecSDave Kleikamp } 201ac27a0ecSDave Kleikamp if (inode->i_nlink) { 2024776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n", 203ac27a0ecSDave Kleikamp inode->i_nlink); 204ac27a0ecSDave Kleikamp return; 205ac27a0ecSDave Kleikamp } 206ac27a0ecSDave Kleikamp if (!sb) { 2074776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode on " 2084776004fSTheodore Ts'o "nonexistent device\n"); 209ac27a0ecSDave Kleikamp return; 210ac27a0ecSDave Kleikamp } 211617ba13bSMingming Cao sbi = EXT4_SB(sb); 212ac27a0ecSDave Kleikamp 213ac27a0ecSDave Kleikamp ino = inode->i_ino; 214617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 2159bffad1eSTheodore Ts'o trace_ext4_free_inode(inode); 216ac27a0ecSDave Kleikamp 217ac27a0ecSDave Kleikamp /* 218ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 219ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 220ac27a0ecSDave Kleikamp */ 221871a2931SChristoph Hellwig dquot_initialize(inode); 222617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 22363936ddaSChristoph Hellwig dquot_free_inode(inode); 2249f754758SChristoph Hellwig dquot_drop(inode); 225ac27a0ecSDave Kleikamp 226ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 227ac27a0ecSDave Kleikamp 228ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 2290930fcc1SAl Viro ext4_clear_inode(inode); 230ac27a0ecSDave Kleikamp 231617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 232617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 23312062dddSEric Sandeen ext4_error(sb, "reserved or nonexistent inode %lu", ino); 234ac27a0ecSDave Kleikamp goto error_return; 235ac27a0ecSDave Kleikamp } 236617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 237617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 238e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 239ac27a0ecSDave Kleikamp if (!bitmap_bh) 240ac27a0ecSDave Kleikamp goto error_return; 241ac27a0ecSDave Kleikamp 242ac27a0ecSDave Kleikamp BUFFER_TRACE(bitmap_bh, "get_write_access"); 243617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bitmap_bh); 244ac27a0ecSDave Kleikamp if (fatal) 245ac27a0ecSDave Kleikamp goto error_return; 246ac27a0ecSDave Kleikamp 247d17413c0SDmitry Monakhov fatal = -ESRCH; 248617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 249d17413c0SDmitry Monakhov if (gdp) { 250ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 251617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 252d17413c0SDmitry Monakhov } 253955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 254d17413c0SDmitry Monakhov cleared = ext4_clear_bit(bit, bitmap_bh->b_data); 255d17413c0SDmitry Monakhov if (fatal || !cleared) { 256d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 257d17413c0SDmitry Monakhov goto out; 258d17413c0SDmitry Monakhov } 259d17413c0SDmitry Monakhov 260560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 261560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 262560671a0SAneesh Kumar K.V if (is_directory) { 263560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 264560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 265d17413c0SDmitry Monakhov percpu_counter_dec(&sbi->s_dirs_counter); 266d17413c0SDmitry Monakhov } 267d17413c0SDmitry Monakhov gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 268d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 2697d39db14STheodore Ts'o 270d17413c0SDmitry Monakhov percpu_counter_inc(&sbi->s_freeinodes_counter); 271d17413c0SDmitry Monakhov if (sbi->s_log_groups_per_flex) { 272d17413c0SDmitry Monakhov ext4_group_t f = ext4_flex_group(sbi, block_group); 273d17413c0SDmitry Monakhov 274d17413c0SDmitry Monakhov atomic_inc(&sbi->s_flex_groups[f].free_inodes); 275d17413c0SDmitry Monakhov if (is_directory) 276c4caae25SEric Sandeen atomic_dec(&sbi->s_flex_groups[f].used_dirs); 2777d39db14STheodore Ts'o } 2780390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 279d17413c0SDmitry Monakhov fatal = ext4_handle_dirty_metadata(handle, NULL, bh2); 280d17413c0SDmitry Monakhov out: 281d17413c0SDmitry Monakhov if (cleared) { 2820390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 2830390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 284ac27a0ecSDave Kleikamp if (!fatal) 285ac27a0ecSDave Kleikamp fatal = err; 286a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 287d17413c0SDmitry Monakhov } else 288d17413c0SDmitry Monakhov ext4_error(sb, "bit already cleared for inode %lu", ino); 289d17413c0SDmitry Monakhov 290ac27a0ecSDave Kleikamp error_return: 291ac27a0ecSDave Kleikamp brelse(bitmap_bh); 292617ba13bSMingming Cao ext4_std_error(sb, fatal); 293ac27a0ecSDave Kleikamp } 294ac27a0ecSDave Kleikamp 295ac27a0ecSDave Kleikamp /* 296ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 297ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 298ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 299ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 300ac27a0ecSDave Kleikamp * directories already is chosen. 301ac27a0ecSDave Kleikamp * 302ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory\'s block 303ac27a0ecSDave Kleikamp * group to find a free inode. 304ac27a0ecSDave Kleikamp */ 3052aa9fc4cSAvantika Mathur static int find_group_dir(struct super_block *sb, struct inode *parent, 3062aa9fc4cSAvantika Mathur ext4_group_t *best_group) 307ac27a0ecSDave Kleikamp { 3088df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 309ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 310617ba13bSMingming Cao struct ext4_group_desc *desc, *best_desc = NULL; 3112aa9fc4cSAvantika Mathur ext4_group_t group; 3122aa9fc4cSAvantika Mathur int ret = -1; 313ac27a0ecSDave Kleikamp 314617ba13bSMingming Cao freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter); 315ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 316ac27a0ecSDave Kleikamp 317ac27a0ecSDave Kleikamp for (group = 0; group < ngroups; group++) { 318ef2fb679SEric Sandeen desc = ext4_get_group_desc(sb, group, NULL); 319560671a0SAneesh Kumar K.V if (!desc || !ext4_free_inodes_count(sb, desc)) 320ac27a0ecSDave Kleikamp continue; 321560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc) < avefreei) 322ac27a0ecSDave Kleikamp continue; 323ac27a0ecSDave Kleikamp if (!best_desc || 324560671a0SAneesh Kumar K.V (ext4_free_blks_count(sb, desc) > 325560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, best_desc))) { 3262aa9fc4cSAvantika Mathur *best_group = group; 327ac27a0ecSDave Kleikamp best_desc = desc; 3282aa9fc4cSAvantika Mathur ret = 0; 329ac27a0ecSDave Kleikamp } 330ac27a0ecSDave Kleikamp } 3312aa9fc4cSAvantika Mathur return ret; 332ac27a0ecSDave Kleikamp } 333ac27a0ecSDave Kleikamp 334772cb7c8SJose R. Santos #define free_block_ratio 10 335772cb7c8SJose R. Santos 336772cb7c8SJose R. Santos static int find_group_flex(struct super_block *sb, struct inode *parent, 337772cb7c8SJose R. Santos ext4_group_t *best_group) 338772cb7c8SJose R. Santos { 339772cb7c8SJose R. Santos struct ext4_sb_info *sbi = EXT4_SB(sb); 340772cb7c8SJose R. Santos struct ext4_group_desc *desc; 341772cb7c8SJose R. Santos struct flex_groups *flex_group = sbi->s_flex_groups; 342772cb7c8SJose R. Santos ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 343772cb7c8SJose R. Santos ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group); 3448df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 345772cb7c8SJose R. Santos int flex_size = ext4_flex_bg_size(sbi); 346772cb7c8SJose R. Santos ext4_group_t best_flex = parent_fbg_group; 347772cb7c8SJose R. Santos int blocks_per_flex = sbi->s_blocks_per_group * flex_size; 348772cb7c8SJose R. Santos int flexbg_free_blocks; 349772cb7c8SJose R. Santos int flex_freeb_ratio; 350772cb7c8SJose R. Santos ext4_group_t n_fbg_groups; 351772cb7c8SJose R. Santos ext4_group_t i; 352772cb7c8SJose R. Santos 3538df9675fSTheodore Ts'o n_fbg_groups = (ngroups + flex_size - 1) >> 354772cb7c8SJose R. Santos sbi->s_log_groups_per_flex; 355772cb7c8SJose R. Santos 356772cb7c8SJose R. Santos find_close_to_parent: 3579f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[best_flex].free_blocks); 358772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 3599f24e420STheodore Ts'o if (atomic_read(&flex_group[best_flex].free_inodes) && 360772cb7c8SJose R. Santos flex_freeb_ratio > free_block_ratio) 361772cb7c8SJose R. Santos goto found_flexbg; 362772cb7c8SJose R. Santos 363772cb7c8SJose R. Santos if (best_flex && best_flex == parent_fbg_group) { 364772cb7c8SJose R. Santos best_flex--; 365772cb7c8SJose R. Santos goto find_close_to_parent; 366772cb7c8SJose R. Santos } 367772cb7c8SJose R. Santos 368772cb7c8SJose R. Santos for (i = 0; i < n_fbg_groups; i++) { 369772cb7c8SJose R. Santos if (i == parent_fbg_group || i == parent_fbg_group - 1) 370772cb7c8SJose R. Santos continue; 371772cb7c8SJose R. Santos 3729f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[i].free_blocks); 373772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 374772cb7c8SJose R. Santos 375772cb7c8SJose R. Santos if (flex_freeb_ratio > free_block_ratio && 3769f24e420STheodore Ts'o (atomic_read(&flex_group[i].free_inodes))) { 377772cb7c8SJose R. Santos best_flex = i; 378772cb7c8SJose R. Santos goto found_flexbg; 379772cb7c8SJose R. Santos } 380772cb7c8SJose R. Santos 3819f24e420STheodore Ts'o if ((atomic_read(&flex_group[best_flex].free_inodes) == 0) || 3829f24e420STheodore Ts'o ((atomic_read(&flex_group[i].free_blocks) > 3839f24e420STheodore Ts'o atomic_read(&flex_group[best_flex].free_blocks)) && 3849f24e420STheodore Ts'o atomic_read(&flex_group[i].free_inodes))) 385772cb7c8SJose R. Santos best_flex = i; 386772cb7c8SJose R. Santos } 387772cb7c8SJose R. Santos 3889f24e420STheodore Ts'o if (!atomic_read(&flex_group[best_flex].free_inodes) || 3899f24e420STheodore Ts'o !atomic_read(&flex_group[best_flex].free_blocks)) 390772cb7c8SJose R. Santos return -1; 391772cb7c8SJose R. Santos 392772cb7c8SJose R. Santos found_flexbg: 393772cb7c8SJose R. Santos for (i = best_flex * flex_size; i < ngroups && 394772cb7c8SJose R. Santos i < (best_flex + 1) * flex_size; i++) { 39588b6edd1STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 396560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc)) { 397772cb7c8SJose R. Santos *best_group = i; 398772cb7c8SJose R. Santos goto out; 399772cb7c8SJose R. Santos } 400772cb7c8SJose R. Santos } 401772cb7c8SJose R. Santos 402772cb7c8SJose R. Santos return -1; 403772cb7c8SJose R. Santos out: 404772cb7c8SJose R. Santos return 0; 405772cb7c8SJose R. Santos } 406772cb7c8SJose R. Santos 407a4912123STheodore Ts'o struct orlov_stats { 408a4912123STheodore Ts'o __u32 free_inodes; 409a4912123STheodore Ts'o __u32 free_blocks; 410a4912123STheodore Ts'o __u32 used_dirs; 411a4912123STheodore Ts'o }; 412a4912123STheodore Ts'o 413a4912123STheodore Ts'o /* 414a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 415a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 416a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 417a4912123STheodore Ts'o */ 4181f109d5aSTheodore Ts'o static void get_orlov_stats(struct super_block *sb, ext4_group_t g, 419a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 420a4912123STheodore Ts'o { 421a4912123STheodore Ts'o struct ext4_group_desc *desc; 4227d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 423a4912123STheodore Ts'o 4247d39db14STheodore Ts'o if (flex_size > 1) { 4257d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 4267d39db14STheodore Ts'o stats->free_blocks = atomic_read(&flex_group[g].free_blocks); 4277d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 4287d39db14STheodore Ts'o return; 4297d39db14STheodore Ts'o } 4307d39db14STheodore Ts'o 4317d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 4327d39db14STheodore Ts'o if (desc) { 4337d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 4347d39db14STheodore Ts'o stats->free_blocks = ext4_free_blks_count(sb, desc); 4357d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 4367d39db14STheodore Ts'o } else { 437a4912123STheodore Ts'o stats->free_inodes = 0; 438a4912123STheodore Ts'o stats->free_blocks = 0; 439a4912123STheodore Ts'o stats->used_dirs = 0; 440a4912123STheodore Ts'o } 441a4912123STheodore Ts'o } 442a4912123STheodore Ts'o 443ac27a0ecSDave Kleikamp /* 444ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 445ac27a0ecSDave Kleikamp * 446ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 447ac27a0ecSDave Kleikamp * 448ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 449ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 450ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 451ac27a0ecSDave Kleikamp * 452ac27a0ecSDave Kleikamp * For the rest rules look so: 453ac27a0ecSDave Kleikamp * 454ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 455ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 456ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 457ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4581cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 459ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 460ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 461ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 462ac27a0ecSDave Kleikamp */ 463ac27a0ecSDave Kleikamp 4642aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 465f157a4aaSTheodore Ts'o ext4_group_t *group, int mode, 466f157a4aaSTheodore Ts'o const struct qstr *qstr) 467ac27a0ecSDave Kleikamp { 468fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 469617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 4708df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 471617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 472ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 473617ba13bSMingming Cao ext4_fsblk_t freeb, avefreeb; 474ac27a0ecSDave Kleikamp unsigned int ndirs; 475a4912123STheodore Ts'o int max_dirs, min_inodes; 476617ba13bSMingming Cao ext4_grpblk_t min_blocks; 4778df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 478617ba13bSMingming Cao struct ext4_group_desc *desc; 479a4912123STheodore Ts'o struct orlov_stats stats; 480a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 481f157a4aaSTheodore Ts'o struct dx_hash_info hinfo; 482a4912123STheodore Ts'o 4838df9675fSTheodore Ts'o ngroups = real_ngroups; 484a4912123STheodore Ts'o if (flex_size > 1) { 4858df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 486a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 487a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 488a4912123STheodore Ts'o } 489ac27a0ecSDave Kleikamp 490ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 491ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 492ac27a0ecSDave Kleikamp freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter); 4933a5b2ecdSMingming Cao avefreeb = freeb; 494f4e5bc24SAndrew Morton do_div(avefreeb, ngroups); 495ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 496ac27a0ecSDave Kleikamp 497a4912123STheodore Ts'o if (S_ISDIR(mode) && 498a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 49912e9b892SDmitry Monakhov (ext4_test_inode_flag(parent, EXT4_INODE_TOPDIR)))) { 500ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 5012aa9fc4cSAvantika Mathur int ret = -1; 502ac27a0ecSDave Kleikamp 503f157a4aaSTheodore Ts'o if (qstr) { 504f157a4aaSTheodore Ts'o hinfo.hash_version = DX_HASH_HALF_MD4; 505f157a4aaSTheodore Ts'o hinfo.seed = sbi->s_hash_seed; 506f157a4aaSTheodore Ts'o ext4fs_dirhash(qstr->name, qstr->len, &hinfo); 507f157a4aaSTheodore Ts'o grp = hinfo.hash; 508f157a4aaSTheodore Ts'o } else 5092aa9fc4cSAvantika Mathur get_random_bytes(&grp, sizeof(grp)); 5102aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 511ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 512a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 513a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 514a4912123STheodore Ts'o if (!stats.free_inodes) 515ac27a0ecSDave Kleikamp continue; 516a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 517ac27a0ecSDave Kleikamp continue; 518a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 519ac27a0ecSDave Kleikamp continue; 520a4912123STheodore Ts'o if (stats.free_blocks < avefreeb) 521ac27a0ecSDave Kleikamp continue; 522a4912123STheodore Ts'o grp = g; 5232aa9fc4cSAvantika Mathur ret = 0; 524a4912123STheodore Ts'o best_ndir = stats.used_dirs; 525ac27a0ecSDave Kleikamp } 526a4912123STheodore Ts'o if (ret) 527a4912123STheodore Ts'o goto fallback; 528a4912123STheodore Ts'o found_flex_bg: 529a4912123STheodore Ts'o if (flex_size == 1) { 530a4912123STheodore Ts'o *group = grp; 531a4912123STheodore Ts'o return 0; 532a4912123STheodore Ts'o } 533a4912123STheodore Ts'o 534a4912123STheodore Ts'o /* 535a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 536a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 537a4912123STheodore Ts'o * something similar, although regular files will 538a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 539a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 540a4912123STheodore Ts'o */ 541a4912123STheodore Ts'o grp *= flex_size; 542a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 5438df9675fSTheodore Ts'o if (grp+i >= real_ngroups) 544a4912123STheodore Ts'o break; 545a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 546a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 547a4912123STheodore Ts'o *group = grp+i; 548a4912123STheodore Ts'o return 0; 549a4912123STheodore Ts'o } 550a4912123STheodore Ts'o } 551ac27a0ecSDave Kleikamp goto fallback; 552ac27a0ecSDave Kleikamp } 553ac27a0ecSDave Kleikamp 554ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 555a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 556a4912123STheodore Ts'o if (min_inodes < 1) 557a4912123STheodore Ts'o min_inodes = 1; 558a4912123STheodore Ts'o min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb)*flex_size / 4; 559ac27a0ecSDave Kleikamp 560a4912123STheodore Ts'o /* 561a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 562a4912123STheodore Ts'o * inode for this parent directory 563a4912123STheodore Ts'o */ 564a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 565a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 566a4912123STheodore Ts'o if (flex_size > 1) 567a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 568a4912123STheodore Ts'o } 569ac27a0ecSDave Kleikamp 570ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 571a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 572a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 573a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 574ac27a0ecSDave Kleikamp continue; 575a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 576ac27a0ecSDave Kleikamp continue; 577a4912123STheodore Ts'o if (stats.free_blocks < min_blocks) 578ac27a0ecSDave Kleikamp continue; 579a4912123STheodore Ts'o goto found_flex_bg; 580ac27a0ecSDave Kleikamp } 581ac27a0ecSDave Kleikamp 582ac27a0ecSDave Kleikamp fallback: 5838df9675fSTheodore Ts'o ngroups = real_ngroups; 584a4912123STheodore Ts'o avefreei = freei / ngroups; 585b5451f7bSTheodore Ts'o fallback_retry: 586a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 587ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 588a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 589a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 590560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 591a4912123STheodore Ts'o ext4_free_inodes_count(sb, desc) >= avefreei) { 592a4912123STheodore Ts'o *group = grp; 5932aa9fc4cSAvantika Mathur return 0; 594ac27a0ecSDave Kleikamp } 595a4912123STheodore Ts'o } 596ac27a0ecSDave Kleikamp 597ac27a0ecSDave Kleikamp if (avefreei) { 598ac27a0ecSDave Kleikamp /* 599ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 600ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 601ac27a0ecSDave Kleikamp */ 602ac27a0ecSDave Kleikamp avefreei = 0; 603b5451f7bSTheodore Ts'o goto fallback_retry; 604ac27a0ecSDave Kleikamp } 605ac27a0ecSDave Kleikamp 606ac27a0ecSDave Kleikamp return -1; 607ac27a0ecSDave Kleikamp } 608ac27a0ecSDave Kleikamp 6092aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 610a4912123STheodore Ts'o ext4_group_t *group, int mode) 611ac27a0ecSDave Kleikamp { 612fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 6138df9675fSTheodore Ts'o ext4_group_t i, last, ngroups = ext4_get_groups_count(sb); 614617ba13bSMingming Cao struct ext4_group_desc *desc; 615a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 616a4912123STheodore Ts'o 617a4912123STheodore Ts'o /* 618a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 619a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 620a4912123STheodore Ts'o * find another flex group, and store that information in the 621a4912123STheodore Ts'o * parent directory's inode information so that use that flex 622a4912123STheodore Ts'o * group for future allocations. 623a4912123STheodore Ts'o */ 624a4912123STheodore Ts'o if (flex_size > 1) { 625a4912123STheodore Ts'o int retry = 0; 626a4912123STheodore Ts'o 627a4912123STheodore Ts'o try_again: 628a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 629a4912123STheodore Ts'o last = parent_group + flex_size; 630a4912123STheodore Ts'o if (last > ngroups) 631a4912123STheodore Ts'o last = ngroups; 632a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 633a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 634a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 635a4912123STheodore Ts'o *group = i; 636a4912123STheodore Ts'o return 0; 637a4912123STheodore Ts'o } 638a4912123STheodore Ts'o } 639a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 640a4912123STheodore Ts'o retry = 1; 641a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 642a4912123STheodore Ts'o goto try_again; 643a4912123STheodore Ts'o } 644a4912123STheodore Ts'o /* 645a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 646a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 647a4912123STheodore Ts'o * avoid the topdir algorithms. 648a4912123STheodore Ts'o */ 649a4912123STheodore Ts'o *group = parent_group + flex_size; 650a4912123STheodore Ts'o if (*group > ngroups) 651a4912123STheodore Ts'o *group = 0; 652f157a4aaSTheodore Ts'o return find_group_orlov(sb, parent, group, mode, 0); 653a4912123STheodore Ts'o } 654ac27a0ecSDave Kleikamp 655ac27a0ecSDave Kleikamp /* 656ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 657ac27a0ecSDave Kleikamp */ 6582aa9fc4cSAvantika Mathur *group = parent_group; 6592aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 660560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 661560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6622aa9fc4cSAvantika Mathur return 0; 663ac27a0ecSDave Kleikamp 664ac27a0ecSDave Kleikamp /* 665ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 666ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 667ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 668ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 669ac27a0ecSDave Kleikamp * different blockgroup. 670ac27a0ecSDave Kleikamp * 671ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 672ac27a0ecSDave Kleikamp */ 6732aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 674ac27a0ecSDave Kleikamp 675ac27a0ecSDave Kleikamp /* 676ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 677ac27a0ecSDave Kleikamp * blocks. 678ac27a0ecSDave Kleikamp */ 679ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6802aa9fc4cSAvantika Mathur *group += i; 6812aa9fc4cSAvantika Mathur if (*group >= ngroups) 6822aa9fc4cSAvantika Mathur *group -= ngroups; 6832aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 684560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 685560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6862aa9fc4cSAvantika Mathur return 0; 687ac27a0ecSDave Kleikamp } 688ac27a0ecSDave Kleikamp 689ac27a0ecSDave Kleikamp /* 690ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 691ac27a0ecSDave Kleikamp * has no free blocks. 692ac27a0ecSDave Kleikamp */ 6932aa9fc4cSAvantika Mathur *group = parent_group; 694ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6952aa9fc4cSAvantika Mathur if (++*group >= ngroups) 6962aa9fc4cSAvantika Mathur *group = 0; 6972aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 698560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 6992aa9fc4cSAvantika Mathur return 0; 700ac27a0ecSDave Kleikamp } 701ac27a0ecSDave Kleikamp 702ac27a0ecSDave Kleikamp return -1; 703ac27a0ecSDave Kleikamp } 704ac27a0ecSDave Kleikamp 705ac27a0ecSDave Kleikamp /* 70639341867SAneesh Kumar K.V * claim the inode from the inode bitmap. If the group 707955ce5f5SAneesh Kumar K.V * is uninit we need to take the groups's ext4_group_lock 70839341867SAneesh Kumar K.V * and clear the uninit flag. The inode bitmap update 70939341867SAneesh Kumar K.V * and group desc uninit flag clear should be done 710955ce5f5SAneesh Kumar K.V * after holding ext4_group_lock so that ext4_read_inode_bitmap 71139341867SAneesh Kumar K.V * doesn't race with the ext4_claim_inode 71239341867SAneesh Kumar K.V */ 71339341867SAneesh Kumar K.V static int ext4_claim_inode(struct super_block *sb, 71439341867SAneesh Kumar K.V struct buffer_head *inode_bitmap_bh, 71539341867SAneesh Kumar K.V unsigned long ino, ext4_group_t group, int mode) 71639341867SAneesh Kumar K.V { 71739341867SAneesh Kumar K.V int free = 0, retval = 0, count; 71839341867SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 719bfff6873SLukas Czerner struct ext4_group_info *grp = ext4_get_group_info(sb, group); 72039341867SAneesh Kumar K.V struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL); 72139341867SAneesh Kumar K.V 722bfff6873SLukas Czerner /* 723bfff6873SLukas Czerner * We have to be sure that new inode allocation does not race with 724bfff6873SLukas Czerner * inode table initialization, because otherwise we may end up 725bfff6873SLukas Czerner * allocating and writing new inode right before sb_issue_zeroout 726bfff6873SLukas Czerner * takes place and overwriting our new inode with zeroes. So we 727bfff6873SLukas Czerner * take alloc_sem to prevent it. 728bfff6873SLukas Czerner */ 729bfff6873SLukas Czerner down_read(&grp->alloc_sem); 730955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 73139341867SAneesh Kumar K.V if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) { 73239341867SAneesh Kumar K.V /* not a free inode */ 73339341867SAneesh Kumar K.V retval = 1; 73439341867SAneesh Kumar K.V goto err_ret; 73539341867SAneesh Kumar K.V } 73639341867SAneesh Kumar K.V ino++; 73739341867SAneesh Kumar K.V if ((group == 0 && ino < EXT4_FIRST_INO(sb)) || 73839341867SAneesh Kumar K.V ino > EXT4_INODES_PER_GROUP(sb)) { 739955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 740bfff6873SLukas Czerner up_read(&grp->alloc_sem); 74112062dddSEric Sandeen ext4_error(sb, "reserved inode or inode > inodes count - " 74239341867SAneesh Kumar K.V "block_group = %u, inode=%lu", group, 74339341867SAneesh Kumar K.V ino + group * EXT4_INODES_PER_GROUP(sb)); 74439341867SAneesh Kumar K.V return 1; 74539341867SAneesh Kumar K.V } 74639341867SAneesh Kumar K.V /* If we didn't allocate from within the initialized part of the inode 74739341867SAneesh Kumar K.V * table then we need to initialize up to this inode. */ 74839341867SAneesh Kumar K.V if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) { 74939341867SAneesh Kumar K.V 75039341867SAneesh Kumar K.V if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 75139341867SAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 75239341867SAneesh Kumar K.V /* When marking the block group with 75339341867SAneesh Kumar K.V * ~EXT4_BG_INODE_UNINIT we don't want to depend 75439341867SAneesh Kumar K.V * on the value of bg_itable_unused even though 75539341867SAneesh Kumar K.V * mke2fs could have initialized the same for us. 75639341867SAneesh Kumar K.V * Instead we calculated the value below 75739341867SAneesh Kumar K.V */ 75839341867SAneesh Kumar K.V 75939341867SAneesh Kumar K.V free = 0; 76039341867SAneesh Kumar K.V } else { 76139341867SAneesh Kumar K.V free = EXT4_INODES_PER_GROUP(sb) - 76239341867SAneesh Kumar K.V ext4_itable_unused_count(sb, gdp); 76339341867SAneesh Kumar K.V } 76439341867SAneesh Kumar K.V 76539341867SAneesh Kumar K.V /* 76639341867SAneesh Kumar K.V * Check the relative inode number against the last used 76739341867SAneesh Kumar K.V * relative inode number in this group. if it is greater 76839341867SAneesh Kumar K.V * we need to update the bg_itable_unused count 76939341867SAneesh Kumar K.V * 77039341867SAneesh Kumar K.V */ 77139341867SAneesh Kumar K.V if (ino > free) 77239341867SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 77339341867SAneesh Kumar K.V (EXT4_INODES_PER_GROUP(sb) - ino)); 77439341867SAneesh Kumar K.V } 77539341867SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) - 1; 77639341867SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 77739341867SAneesh Kumar K.V if (S_ISDIR(mode)) { 77839341867SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) + 1; 77939341867SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 7807d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 7817d39db14STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 7827d39db14STheodore Ts'o 783c4caae25SEric Sandeen atomic_inc(&sbi->s_flex_groups[f].used_dirs); 7847d39db14STheodore Ts'o } 78539341867SAneesh Kumar K.V } 78639341867SAneesh Kumar K.V gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 78739341867SAneesh Kumar K.V err_ret: 788955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 789bfff6873SLukas Czerner up_read(&grp->alloc_sem); 79039341867SAneesh Kumar K.V return retval; 79139341867SAneesh Kumar K.V } 79239341867SAneesh Kumar K.V 79339341867SAneesh Kumar K.V /* 794ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 795ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 796ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 797ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 798ac27a0ecSDave Kleikamp * directories already is chosen. 799ac27a0ecSDave Kleikamp * 800ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 801ac27a0ecSDave Kleikamp * group to find a free inode. 802ac27a0ecSDave Kleikamp */ 803f157a4aaSTheodore Ts'o struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode, 80411013911SAndreas Dilger const struct qstr *qstr, __u32 goal) 805ac27a0ecSDave Kleikamp { 806ac27a0ecSDave Kleikamp struct super_block *sb; 8073300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 8083300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 8098df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 810ac27a0ecSDave Kleikamp unsigned long ino = 0; 811ac27a0ecSDave Kleikamp struct inode *inode; 812617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 813617ba13bSMingming Cao struct ext4_inode_info *ei; 814617ba13bSMingming Cao struct ext4_sb_info *sbi; 81539341867SAneesh Kumar K.V int ret2, err = 0; 816ac27a0ecSDave Kleikamp struct inode *ret; 8172aa9fc4cSAvantika Mathur ext4_group_t i; 8182aa9fc4cSAvantika Mathur int free = 0; 8192842c3b5STheodore Ts'o static int once = 1; 820772cb7c8SJose R. Santos ext4_group_t flex_group; 821ac27a0ecSDave Kleikamp 822ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 823ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 824ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 825ac27a0ecSDave Kleikamp 826ac27a0ecSDave Kleikamp sb = dir->i_sb; 8278df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 8289bffad1eSTheodore Ts'o trace_ext4_request_inode(dir, mode); 829ac27a0ecSDave Kleikamp inode = new_inode(sb); 830ac27a0ecSDave Kleikamp if (!inode) 831ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 832617ba13bSMingming Cao ei = EXT4_I(inode); 833617ba13bSMingming Cao sbi = EXT4_SB(sb); 834772cb7c8SJose R. Santos 83511013911SAndreas Dilger if (!goal) 83611013911SAndreas Dilger goal = sbi->s_inode_goal; 83711013911SAndreas Dilger 838e6462869SJohann Lombardi if (goal && goal <= le32_to_cpu(sbi->s_es->s_inodes_count)) { 83911013911SAndreas Dilger group = (goal - 1) / EXT4_INODES_PER_GROUP(sb); 84011013911SAndreas Dilger ino = (goal - 1) % EXT4_INODES_PER_GROUP(sb); 84111013911SAndreas Dilger ret2 = 0; 84211013911SAndreas Dilger goto got_group; 84311013911SAndreas Dilger } 84411013911SAndreas Dilger 845a4912123STheodore Ts'o if (sbi->s_log_groups_per_flex && test_opt(sb, OLDALLOC)) { 846772cb7c8SJose R. Santos ret2 = find_group_flex(sb, dir, &group); 84705bf9e83STheodore Ts'o if (ret2 == -1) { 848a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 8496b82f3cbSChuck Ebbert if (ret2 == 0 && once) { 8502842c3b5STheodore Ts'o once = 0; 85105bf9e83STheodore Ts'o printk(KERN_NOTICE "ext4: find_group_flex " 85205bf9e83STheodore Ts'o "failed, fallback succeeded dir %lu\n", 85305bf9e83STheodore Ts'o dir->i_ino); 85405bf9e83STheodore Ts'o } 8556b82f3cbSChuck Ebbert } 856772cb7c8SJose R. Santos goto got_group; 857772cb7c8SJose R. Santos } 858772cb7c8SJose R. Santos 859ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) { 860ac27a0ecSDave Kleikamp if (test_opt(sb, OLDALLOC)) 8612aa9fc4cSAvantika Mathur ret2 = find_group_dir(sb, dir, &group); 862ac27a0ecSDave Kleikamp else 863f157a4aaSTheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode, qstr); 864ac27a0ecSDave Kleikamp } else 865a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 866ac27a0ecSDave Kleikamp 867772cb7c8SJose R. Santos got_group: 868a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 869ac27a0ecSDave Kleikamp err = -ENOSPC; 8702aa9fc4cSAvantika Mathur if (ret2 == -1) 871ac27a0ecSDave Kleikamp goto out; 872ac27a0ecSDave Kleikamp 87311013911SAndreas Dilger for (i = 0; i < ngroups; i++, ino = 0) { 874ac27a0ecSDave Kleikamp err = -EIO; 875ac27a0ecSDave Kleikamp 8763300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 877ac27a0ecSDave Kleikamp if (!gdp) 878ac27a0ecSDave Kleikamp goto fail; 879ac27a0ecSDave Kleikamp 8803300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 8813300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 8823300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 883ac27a0ecSDave Kleikamp goto fail; 884ac27a0ecSDave Kleikamp 885ac27a0ecSDave Kleikamp repeat_in_this_group: 886617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 8873300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 8883300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 8893300bedaSAneesh Kumar K.V 890617ba13bSMingming Cao if (ino < EXT4_INODES_PER_GROUP(sb)) { 891ac27a0ecSDave Kleikamp 8923300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 8933300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 8943300bedaSAneesh Kumar K.V inode_bitmap_bh); 895ac27a0ecSDave Kleikamp if (err) 896ac27a0ecSDave Kleikamp goto fail; 897ac27a0ecSDave Kleikamp 89839341867SAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "get_write_access"); 89939341867SAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 90039341867SAneesh Kumar K.V group_desc_bh); 90139341867SAneesh Kumar K.V if (err) 90239341867SAneesh Kumar K.V goto fail; 90339341867SAneesh Kumar K.V if (!ext4_claim_inode(sb, inode_bitmap_bh, 90439341867SAneesh Kumar K.V ino, group, mode)) { 905ac27a0ecSDave Kleikamp /* we won it */ 9063300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, 9070390131bSFrank Mayhar "call ext4_handle_dirty_metadata"); 9080390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 90973b50c1cSCurt Wohlgemuth NULL, 9103300bedaSAneesh Kumar K.V inode_bitmap_bh); 911ac27a0ecSDave Kleikamp if (err) 912ac27a0ecSDave Kleikamp goto fail; 91339341867SAneesh Kumar K.V /* zero bit is inode number 1*/ 91439341867SAneesh Kumar K.V ino++; 915ac27a0ecSDave Kleikamp goto got; 916ac27a0ecSDave Kleikamp } 917ac27a0ecSDave Kleikamp /* we lost it */ 9183300bedaSAneesh Kumar K.V ext4_handle_release_buffer(handle, inode_bitmap_bh); 91939341867SAneesh Kumar K.V ext4_handle_release_buffer(handle, group_desc_bh); 920ac27a0ecSDave Kleikamp 921617ba13bSMingming Cao if (++ino < EXT4_INODES_PER_GROUP(sb)) 922ac27a0ecSDave Kleikamp goto repeat_in_this_group; 923ac27a0ecSDave Kleikamp } 924ac27a0ecSDave Kleikamp 925ac27a0ecSDave Kleikamp /* 926ac27a0ecSDave Kleikamp * This case is possible in concurrent environment. It is very 927ac27a0ecSDave Kleikamp * rare. We cannot repeat the find_group_xxx() call because 928ac27a0ecSDave Kleikamp * that will simply return the same blockgroup, because the 929ac27a0ecSDave Kleikamp * group descriptor metadata has not yet been updated. 930ac27a0ecSDave Kleikamp * So we just go onto the next blockgroup. 931ac27a0ecSDave Kleikamp */ 9328df9675fSTheodore Ts'o if (++group == ngroups) 933ac27a0ecSDave Kleikamp group = 0; 934ac27a0ecSDave Kleikamp } 935ac27a0ecSDave Kleikamp err = -ENOSPC; 936ac27a0ecSDave Kleikamp goto out; 937ac27a0ecSDave Kleikamp 938ac27a0ecSDave Kleikamp got: 939717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 940717d50e4SAndreas Dilger if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) && 941717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9423300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 943717d50e4SAndreas Dilger 9443300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 9453300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 9463300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 947717d50e4SAndreas Dilger if (err) { 9483300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 949717d50e4SAndreas Dilger goto fail; 950717d50e4SAndreas Dilger } 951717d50e4SAndreas Dilger 952717d50e4SAndreas Dilger free = 0; 953955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 954717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 955717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 956717d50e4SAndreas Dilger free = ext4_free_blocks_after_init(sb, group, gdp); 9573300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 958560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, free); 95923712a9cSFrederic Bohe gdp->bg_checksum = ext4_group_desc_csum(sbi, group, 96023712a9cSFrederic Bohe gdp); 961717d50e4SAndreas Dilger } 962955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 963717d50e4SAndreas Dilger 964717d50e4SAndreas Dilger /* Don't need to dirty bitmap block if we didn't change it */ 965717d50e4SAndreas Dilger if (free) { 9663300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 9670390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 9683300bedaSAneesh Kumar K.V NULL, block_bitmap_bh); 969717d50e4SAndreas Dilger } 970717d50e4SAndreas Dilger 9713300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 972717d50e4SAndreas Dilger if (err) 973717d50e4SAndreas Dilger goto fail; 974717d50e4SAndreas Dilger } 9753300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 9763300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 97739341867SAneesh Kumar K.V if (err) 97839341867SAneesh Kumar K.V goto fail; 979ac27a0ecSDave Kleikamp 980ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 981ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 982ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 983a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 984ac27a0ecSDave Kleikamp 985772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 986772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 9879f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 988772cb7c8SJose R. Santos } 989772cb7c8SJose R. Santos 990b10b8520SDmitry Monakhov if (test_opt(sb, GRPID)) { 991ac27a0ecSDave Kleikamp inode->i_mode = mode; 992b10b8520SDmitry Monakhov inode->i_uid = current_fsuid(); 993b10b8520SDmitry Monakhov inode->i_gid = dir->i_gid; 994b10b8520SDmitry Monakhov } else 995b10b8520SDmitry Monakhov inode_init_owner(inode, dir, mode); 996ac27a0ecSDave Kleikamp 997717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 998ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 999ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1000ef7f3835SKalpak Shah inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime = 1001ef7f3835SKalpak Shah ext4_current_time(inode); 1002ac27a0ecSDave Kleikamp 1003ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 1004ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 1005ac27a0ecSDave Kleikamp ei->i_disksize = 0; 1006ac27a0ecSDave Kleikamp 100742bf0383SAneesh Kumar K.V /* 10082dc6b0d4SDuane Griffin * Don't inherit extent flag from directory, amongst others. We set 10092dc6b0d4SDuane Griffin * extent flag on newly created directory and file only if -o extent 10102dc6b0d4SDuane Griffin * mount option is specified 101142bf0383SAneesh Kumar K.V */ 10122dc6b0d4SDuane Griffin ei->i_flags = 10132dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 1014ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 1015ac27a0ecSDave Kleikamp ei->i_dtime = 0; 1016ac27a0ecSDave Kleikamp ei->i_block_group = group; 1017a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 1018ac27a0ecSDave Kleikamp 1019617ba13bSMingming Cao ext4_set_inode_flags(inode); 1020ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 10210390131bSFrank Mayhar ext4_handle_sync(handle); 10226b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 10236b38e842SAl Viro err = -EINVAL; 10246b38e842SAl Viro goto fail_drop; 10256b38e842SAl Viro } 1026ac27a0ecSDave Kleikamp spin_lock(&sbi->s_next_gen_lock); 1027ac27a0ecSDave Kleikamp inode->i_generation = sbi->s_next_generation++; 1028ac27a0ecSDave Kleikamp spin_unlock(&sbi->s_next_gen_lock); 1029ac27a0ecSDave Kleikamp 1030353eb83cSTheodore Ts'o ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */ 103119f5fb7aSTheodore Ts'o ext4_set_inode_state(inode, EXT4_STATE_NEW); 1032ef7f3835SKalpak Shah 1033ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 1034ac27a0ecSDave Kleikamp 1035ac27a0ecSDave Kleikamp ret = inode; 1036871a2931SChristoph Hellwig dquot_initialize(inode); 103763936ddaSChristoph Hellwig err = dquot_alloc_inode(inode); 103863936ddaSChristoph Hellwig if (err) 1039ac27a0ecSDave Kleikamp goto fail_drop; 1040ac27a0ecSDave Kleikamp 1041617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1042ac27a0ecSDave Kleikamp if (err) 1043ac27a0ecSDave Kleikamp goto fail_free_drop; 1044ac27a0ecSDave Kleikamp 1045617ba13bSMingming Cao err = ext4_init_security(handle, inode, dir); 1046ac27a0ecSDave Kleikamp if (err) 1047ac27a0ecSDave Kleikamp goto fail_free_drop; 1048ac27a0ecSDave Kleikamp 104983982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 1050e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1051e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 105212e9b892SDmitry Monakhov ext4_set_inode_flag(inode, EXT4_INODE_EXTENTS); 1053a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1054a86c6181SAlex Tomas } 105542bf0383SAneesh Kumar K.V } 1056ac27a0ecSDave Kleikamp 10578753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 10588753e88fSAneesh Kumar K.V if (err) { 10598753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 10608753e88fSAneesh Kumar K.V goto fail_free_drop; 10618753e88fSAneesh Kumar K.V } 10628753e88fSAneesh Kumar K.V 1063617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 10649bffad1eSTheodore Ts'o trace_ext4_allocate_inode(inode, dir, mode); 1065ac27a0ecSDave Kleikamp goto really_out; 1066ac27a0ecSDave Kleikamp fail: 1067617ba13bSMingming Cao ext4_std_error(sb, err); 1068ac27a0ecSDave Kleikamp out: 1069ac27a0ecSDave Kleikamp iput(inode); 1070ac27a0ecSDave Kleikamp ret = ERR_PTR(err); 1071ac27a0ecSDave Kleikamp really_out: 10723300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1073ac27a0ecSDave Kleikamp return ret; 1074ac27a0ecSDave Kleikamp 1075ac27a0ecSDave Kleikamp fail_free_drop: 107663936ddaSChristoph Hellwig dquot_free_inode(inode); 1077ac27a0ecSDave Kleikamp 1078ac27a0ecSDave Kleikamp fail_drop: 10799f754758SChristoph Hellwig dquot_drop(inode); 1080ac27a0ecSDave Kleikamp inode->i_flags |= S_NOQUOTA; 1081ac27a0ecSDave Kleikamp inode->i_nlink = 0; 10826b38e842SAl Viro unlock_new_inode(inode); 1083ac27a0ecSDave Kleikamp iput(inode); 10843300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1085ac27a0ecSDave Kleikamp return ERR_PTR(err); 1086ac27a0ecSDave Kleikamp } 1087ac27a0ecSDave Kleikamp 1088ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1089617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1090ac27a0ecSDave Kleikamp { 1091617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1092fd2d4291SAvantika Mathur ext4_group_t block_group; 1093ac27a0ecSDave Kleikamp int bit; 10941d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 1095ac27a0ecSDave Kleikamp struct inode *inode = NULL; 10961d1fe1eeSDavid Howells long err = -EIO; 1097ac27a0ecSDave Kleikamp 1098ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1099ac27a0ecSDave Kleikamp if (ino > max_ino) { 110012062dddSEric Sandeen ext4_warning(sb, "bad orphan ino %lu! e2fsck was run?", ino); 11011d1fe1eeSDavid Howells goto error; 1102ac27a0ecSDave Kleikamp } 1103ac27a0ecSDave Kleikamp 1104617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1105617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1106e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1107ac27a0ecSDave Kleikamp if (!bitmap_bh) { 110812062dddSEric Sandeen ext4_warning(sb, "inode bitmap error for orphan %lu", ino); 11091d1fe1eeSDavid Howells goto error; 1110ac27a0ecSDave Kleikamp } 1111ac27a0ecSDave Kleikamp 1112ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1113ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1114ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1115ac27a0ecSDave Kleikamp */ 11161d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 11171d1fe1eeSDavid Howells goto bad_orphan; 11181d1fe1eeSDavid Howells 11191d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 11201d1fe1eeSDavid Howells if (IS_ERR(inode)) 11211d1fe1eeSDavid Howells goto iget_failed; 11221d1fe1eeSDavid Howells 112391ef4cafSDuane Griffin /* 112491ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 112591ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 112691ef4cafSDuane Griffin * during processing and an infinite loop will result. 112791ef4cafSDuane Griffin */ 112891ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 112991ef4cafSDuane Griffin goto bad_orphan; 113091ef4cafSDuane Griffin 11311d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 11321d1fe1eeSDavid Howells goto bad_orphan; 11331d1fe1eeSDavid Howells brelse(bitmap_bh); 11341d1fe1eeSDavid Howells return inode; 11351d1fe1eeSDavid Howells 11361d1fe1eeSDavid Howells iget_failed: 11371d1fe1eeSDavid Howells err = PTR_ERR(inode); 11381d1fe1eeSDavid Howells inode = NULL; 11391d1fe1eeSDavid Howells bad_orphan: 114012062dddSEric Sandeen ext4_warning(sb, "bad orphan inode %lu! e2fsck was run?", ino); 1141617ba13bSMingming Cao printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1142ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1143617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1144ac27a0ecSDave Kleikamp printk(KERN_NOTICE "inode=%p\n", inode); 1145ac27a0ecSDave Kleikamp if (inode) { 1146ac27a0ecSDave Kleikamp printk(KERN_NOTICE "is_bad_inode(inode)=%d\n", 1147ac27a0ecSDave Kleikamp is_bad_inode(inode)); 1148ac27a0ecSDave Kleikamp printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n", 1149ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 1150ac27a0ecSDave Kleikamp printk(KERN_NOTICE "max_ino=%lu\n", max_ino); 115191ef4cafSDuane Griffin printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink); 1152ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 11531d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1154ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1155ac27a0ecSDave Kleikamp iput(inode); 1156ac27a0ecSDave Kleikamp } 1157ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11581d1fe1eeSDavid Howells error: 11591d1fe1eeSDavid Howells return ERR_PTR(err); 1160ac27a0ecSDave Kleikamp } 1161ac27a0ecSDave Kleikamp 1162617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1163ac27a0ecSDave Kleikamp { 1164ac27a0ecSDave Kleikamp unsigned long desc_count; 1165617ba13bSMingming Cao struct ext4_group_desc *gdp; 11668df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1167617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1168617ba13bSMingming Cao struct ext4_super_block *es; 1169ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1170ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1171ac27a0ecSDave Kleikamp 1172617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1173ac27a0ecSDave Kleikamp desc_count = 0; 1174ac27a0ecSDave Kleikamp bitmap_count = 0; 1175ac27a0ecSDave Kleikamp gdp = NULL; 11768df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1177617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1178ac27a0ecSDave Kleikamp if (!gdp) 1179ac27a0ecSDave Kleikamp continue; 1180560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1181ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1182e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1183ac27a0ecSDave Kleikamp if (!bitmap_bh) 1184ac27a0ecSDave Kleikamp continue; 1185ac27a0ecSDave Kleikamp 1186617ba13bSMingming Cao x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8); 1187c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1188785b4b3aSPeng Tao (unsigned long) i, ext4_free_inodes_count(sb, gdp), x); 1189ac27a0ecSDave Kleikamp bitmap_count += x; 1190ac27a0ecSDave Kleikamp } 1191ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11924776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 11934776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1194ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1195ac27a0ecSDave Kleikamp return desc_count; 1196ac27a0ecSDave Kleikamp #else 1197ac27a0ecSDave Kleikamp desc_count = 0; 11988df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1199617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1200ac27a0ecSDave Kleikamp if (!gdp) 1201ac27a0ecSDave Kleikamp continue; 1202560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1203ac27a0ecSDave Kleikamp cond_resched(); 1204ac27a0ecSDave Kleikamp } 1205ac27a0ecSDave Kleikamp return desc_count; 1206ac27a0ecSDave Kleikamp #endif 1207ac27a0ecSDave Kleikamp } 1208ac27a0ecSDave Kleikamp 1209ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1210617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1211ac27a0ecSDave Kleikamp { 1212ac27a0ecSDave Kleikamp unsigned long count = 0; 12138df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1214ac27a0ecSDave Kleikamp 12158df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1216617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1217ac27a0ecSDave Kleikamp if (!gdp) 1218ac27a0ecSDave Kleikamp continue; 1219560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1220ac27a0ecSDave Kleikamp } 1221ac27a0ecSDave Kleikamp return count; 1222ac27a0ecSDave Kleikamp } 1223bfff6873SLukas Czerner 1224bfff6873SLukas Czerner /* 1225bfff6873SLukas Czerner * Zeroes not yet zeroed inode table - just write zeroes through the whole 1226bfff6873SLukas Czerner * inode table. Must be called without any spinlock held. The only place 1227bfff6873SLukas Czerner * where it is called from on active part of filesystem is ext4lazyinit 1228bfff6873SLukas Czerner * thread, so we do not need any special locks, however we have to prevent 1229bfff6873SLukas Czerner * inode allocation from the current group, so we take alloc_sem lock, to 1230bfff6873SLukas Czerner * block ext4_claim_inode until we are finished. 1231bfff6873SLukas Czerner */ 1232bfff6873SLukas Czerner extern int ext4_init_inode_table(struct super_block *sb, ext4_group_t group, 1233bfff6873SLukas Czerner int barrier) 1234bfff6873SLukas Czerner { 1235bfff6873SLukas Czerner struct ext4_group_info *grp = ext4_get_group_info(sb, group); 1236bfff6873SLukas Czerner struct ext4_sb_info *sbi = EXT4_SB(sb); 1237bfff6873SLukas Czerner struct ext4_group_desc *gdp = NULL; 1238bfff6873SLukas Czerner struct buffer_head *group_desc_bh; 1239bfff6873SLukas Czerner handle_t *handle; 1240bfff6873SLukas Czerner ext4_fsblk_t blk; 1241bfff6873SLukas Czerner int num, ret = 0, used_blks = 0; 1242bfff6873SLukas Czerner 1243bfff6873SLukas Czerner /* This should not happen, but just to be sure check this */ 1244bfff6873SLukas Czerner if (sb->s_flags & MS_RDONLY) { 1245bfff6873SLukas Czerner ret = 1; 1246bfff6873SLukas Czerner goto out; 1247bfff6873SLukas Czerner } 1248bfff6873SLukas Czerner 1249bfff6873SLukas Czerner gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 1250bfff6873SLukas Czerner if (!gdp) 1251bfff6873SLukas Czerner goto out; 1252bfff6873SLukas Czerner 1253bfff6873SLukas Czerner /* 1254bfff6873SLukas Czerner * We do not need to lock this, because we are the only one 1255bfff6873SLukas Czerner * handling this flag. 1256bfff6873SLukas Czerner */ 1257bfff6873SLukas Czerner if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_ZEROED)) 1258bfff6873SLukas Czerner goto out; 1259bfff6873SLukas Czerner 1260bfff6873SLukas Czerner handle = ext4_journal_start_sb(sb, 1); 1261bfff6873SLukas Czerner if (IS_ERR(handle)) { 1262bfff6873SLukas Czerner ret = PTR_ERR(handle); 1263bfff6873SLukas Czerner goto out; 1264bfff6873SLukas Czerner } 1265bfff6873SLukas Czerner 1266bfff6873SLukas Czerner down_write(&grp->alloc_sem); 1267bfff6873SLukas Czerner /* 1268bfff6873SLukas Czerner * If inode bitmap was already initialized there may be some 1269bfff6873SLukas Czerner * used inodes so we need to skip blocks with used inodes in 1270bfff6873SLukas Czerner * inode table. 1271bfff6873SLukas Czerner */ 1272bfff6873SLukas Czerner if (!(gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT))) 1273bfff6873SLukas Czerner used_blks = DIV_ROUND_UP((EXT4_INODES_PER_GROUP(sb) - 1274bfff6873SLukas Czerner ext4_itable_unused_count(sb, gdp)), 1275bfff6873SLukas Czerner sbi->s_inodes_per_block); 1276bfff6873SLukas Czerner 1277857ac889SLukas Czerner if ((used_blks < 0) || (used_blks > sbi->s_itb_per_group)) { 1278857ac889SLukas Czerner ext4_error(sb, "Something is wrong with group %u\n" 1279857ac889SLukas Czerner "Used itable blocks: %d" 1280857ac889SLukas Czerner "itable unused count: %u\n", 1281857ac889SLukas Czerner group, used_blks, 1282857ac889SLukas Czerner ext4_itable_unused_count(sb, gdp)); 1283857ac889SLukas Czerner ret = 1; 1284857ac889SLukas Czerner goto out; 1285857ac889SLukas Czerner } 1286857ac889SLukas Czerner 1287bfff6873SLukas Czerner blk = ext4_inode_table(sb, gdp) + used_blks; 1288bfff6873SLukas Czerner num = sbi->s_itb_per_group - used_blks; 1289bfff6873SLukas Czerner 1290bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, "get_write_access"); 1291bfff6873SLukas Czerner ret = ext4_journal_get_write_access(handle, 1292bfff6873SLukas Czerner group_desc_bh); 1293bfff6873SLukas Czerner if (ret) 1294bfff6873SLukas Czerner goto err_out; 1295bfff6873SLukas Czerner 1296bfff6873SLukas Czerner /* 1297bfff6873SLukas Czerner * Skip zeroout if the inode table is full. But we set the ZEROED 1298bfff6873SLukas Czerner * flag anyway, because obviously, when it is full it does not need 1299bfff6873SLukas Czerner * further zeroing. 1300bfff6873SLukas Czerner */ 1301bfff6873SLukas Czerner if (unlikely(num == 0)) 1302bfff6873SLukas Czerner goto skip_zeroout; 1303bfff6873SLukas Czerner 1304bfff6873SLukas Czerner ext4_debug("going to zero out inode table in group %d\n", 1305bfff6873SLukas Czerner group); 1306a107e5a3STheodore Ts'o ret = sb_issue_zeroout(sb, blk, num, GFP_NOFS); 1307bfff6873SLukas Czerner if (ret < 0) 1308bfff6873SLukas Czerner goto err_out; 1309a107e5a3STheodore Ts'o if (barrier) 1310a107e5a3STheodore Ts'o blkdev_issue_flush(sb->s_bdev, GFP_NOFS, NULL); 1311bfff6873SLukas Czerner 1312bfff6873SLukas Czerner skip_zeroout: 1313bfff6873SLukas Czerner ext4_lock_group(sb, group); 1314bfff6873SLukas Czerner gdp->bg_flags |= cpu_to_le16(EXT4_BG_INODE_ZEROED); 1315bfff6873SLukas Czerner gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 1316bfff6873SLukas Czerner ext4_unlock_group(sb, group); 1317bfff6873SLukas Czerner 1318bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, 1319bfff6873SLukas Czerner "call ext4_handle_dirty_metadata"); 1320bfff6873SLukas Czerner ret = ext4_handle_dirty_metadata(handle, NULL, 1321bfff6873SLukas Czerner group_desc_bh); 1322bfff6873SLukas Czerner 1323bfff6873SLukas Czerner err_out: 1324bfff6873SLukas Czerner up_write(&grp->alloc_sem); 1325bfff6873SLukas Czerner ext4_journal_stop(handle); 1326bfff6873SLukas Czerner out: 1327bfff6873SLukas Czerner return ret; 1328bfff6873SLukas Czerner } 1329