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); 81021b65bbSTheodore Ts'o ext4_free_group_clusters_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 */ 1550562e0baSJiaying Zhang trace_ext4_load_inode_bitmap(sb, block_group); 1562ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 157e29d1cdeSEric Sandeen if (bh_submit_read(bh) < 0) { 158e29d1cdeSEric Sandeen put_bh(bh); 15912062dddSEric Sandeen ext4_error(sb, "Cannot read inode bitmap - " 160a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 161e29d1cdeSEric Sandeen block_group, bitmap_blk); 162e29d1cdeSEric Sandeen return NULL; 163e29d1cdeSEric Sandeen } 164ac27a0ecSDave Kleikamp return bh; 165ac27a0ecSDave Kleikamp } 166ac27a0ecSDave Kleikamp 167ac27a0ecSDave Kleikamp /* 168ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 169ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 170ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 171ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 172ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 173ac27a0ecSDave Kleikamp * has been deleted earlier. 174ac27a0ecSDave Kleikamp * 175ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 176ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 177ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 178ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 179ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 180ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 181ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 182ac27a0ecSDave Kleikamp */ 183617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 184ac27a0ecSDave Kleikamp { 185ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 186ac27a0ecSDave Kleikamp int is_directory; 187ac27a0ecSDave Kleikamp unsigned long ino; 188ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 189ac27a0ecSDave Kleikamp struct buffer_head *bh2; 190fd2d4291SAvantika Mathur ext4_group_t block_group; 191ac27a0ecSDave Kleikamp unsigned long bit; 192617ba13bSMingming Cao struct ext4_group_desc *gdp; 193617ba13bSMingming Cao struct ext4_super_block *es; 194617ba13bSMingming Cao struct ext4_sb_info *sbi; 1957ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 196ac27a0ecSDave Kleikamp 197ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 1984776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has count=%d\n", 199ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 200ac27a0ecSDave Kleikamp return; 201ac27a0ecSDave Kleikamp } 202ac27a0ecSDave Kleikamp if (inode->i_nlink) { 2034776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n", 204ac27a0ecSDave Kleikamp inode->i_nlink); 205ac27a0ecSDave Kleikamp return; 206ac27a0ecSDave Kleikamp } 207ac27a0ecSDave Kleikamp if (!sb) { 2084776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode on " 2094776004fSTheodore Ts'o "nonexistent device\n"); 210ac27a0ecSDave Kleikamp return; 211ac27a0ecSDave Kleikamp } 212617ba13bSMingming Cao sbi = EXT4_SB(sb); 213ac27a0ecSDave Kleikamp 214ac27a0ecSDave Kleikamp ino = inode->i_ino; 215617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 2169bffad1eSTheodore Ts'o trace_ext4_free_inode(inode); 217ac27a0ecSDave Kleikamp 218ac27a0ecSDave Kleikamp /* 219ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 220ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 221ac27a0ecSDave Kleikamp */ 222871a2931SChristoph Hellwig dquot_initialize(inode); 223617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 22463936ddaSChristoph Hellwig dquot_free_inode(inode); 2259f754758SChristoph Hellwig dquot_drop(inode); 226ac27a0ecSDave Kleikamp 227ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 228ac27a0ecSDave Kleikamp 229ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 2300930fcc1SAl Viro ext4_clear_inode(inode); 231ac27a0ecSDave Kleikamp 232617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 233617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 23412062dddSEric Sandeen ext4_error(sb, "reserved or nonexistent inode %lu", ino); 235ac27a0ecSDave Kleikamp goto error_return; 236ac27a0ecSDave Kleikamp } 237617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 238617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 239e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 240ac27a0ecSDave Kleikamp if (!bitmap_bh) 241ac27a0ecSDave Kleikamp goto error_return; 242ac27a0ecSDave Kleikamp 243ac27a0ecSDave Kleikamp BUFFER_TRACE(bitmap_bh, "get_write_access"); 244617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bitmap_bh); 245ac27a0ecSDave Kleikamp if (fatal) 246ac27a0ecSDave Kleikamp goto error_return; 247ac27a0ecSDave Kleikamp 248d17413c0SDmitry Monakhov fatal = -ESRCH; 249617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 250d17413c0SDmitry Monakhov if (gdp) { 251ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 252617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 253d17413c0SDmitry Monakhov } 254955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 255d17413c0SDmitry Monakhov cleared = ext4_clear_bit(bit, bitmap_bh->b_data); 256d17413c0SDmitry Monakhov if (fatal || !cleared) { 257d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 258d17413c0SDmitry Monakhov goto out; 259d17413c0SDmitry Monakhov } 260d17413c0SDmitry Monakhov 261560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 262560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 263560671a0SAneesh Kumar K.V if (is_directory) { 264560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 265560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 266d17413c0SDmitry Monakhov percpu_counter_dec(&sbi->s_dirs_counter); 267d17413c0SDmitry Monakhov } 268d17413c0SDmitry Monakhov gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 269d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 2707d39db14STheodore Ts'o 271d17413c0SDmitry Monakhov percpu_counter_inc(&sbi->s_freeinodes_counter); 272d17413c0SDmitry Monakhov if (sbi->s_log_groups_per_flex) { 273d17413c0SDmitry Monakhov ext4_group_t f = ext4_flex_group(sbi, block_group); 274d17413c0SDmitry Monakhov 275d17413c0SDmitry Monakhov atomic_inc(&sbi->s_flex_groups[f].free_inodes); 276d17413c0SDmitry Monakhov if (is_directory) 277c4caae25SEric Sandeen atomic_dec(&sbi->s_flex_groups[f].used_dirs); 2787d39db14STheodore Ts'o } 2790390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 280d17413c0SDmitry Monakhov fatal = ext4_handle_dirty_metadata(handle, NULL, bh2); 281d17413c0SDmitry Monakhov out: 282d17413c0SDmitry Monakhov if (cleared) { 2830390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 2840390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 285ac27a0ecSDave Kleikamp if (!fatal) 286ac27a0ecSDave Kleikamp fatal = err; 287a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 288d17413c0SDmitry Monakhov } else 289d17413c0SDmitry Monakhov ext4_error(sb, "bit already cleared for inode %lu", ino); 290d17413c0SDmitry Monakhov 291ac27a0ecSDave Kleikamp error_return: 292ac27a0ecSDave Kleikamp brelse(bitmap_bh); 293617ba13bSMingming Cao ext4_std_error(sb, fatal); 294ac27a0ecSDave Kleikamp } 295ac27a0ecSDave Kleikamp 296ac27a0ecSDave Kleikamp /* 297ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 298ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 299ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 300ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 301ac27a0ecSDave Kleikamp * directories already is chosen. 302ac27a0ecSDave Kleikamp * 303ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory\'s block 304ac27a0ecSDave Kleikamp * group to find a free inode. 305ac27a0ecSDave Kleikamp */ 3062aa9fc4cSAvantika Mathur static int find_group_dir(struct super_block *sb, struct inode *parent, 3072aa9fc4cSAvantika Mathur ext4_group_t *best_group) 308ac27a0ecSDave Kleikamp { 3098df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 310ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 311617ba13bSMingming Cao struct ext4_group_desc *desc, *best_desc = NULL; 3122aa9fc4cSAvantika Mathur ext4_group_t group; 3132aa9fc4cSAvantika Mathur int ret = -1; 314ac27a0ecSDave Kleikamp 315617ba13bSMingming Cao freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter); 316ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 317ac27a0ecSDave Kleikamp 318ac27a0ecSDave Kleikamp for (group = 0; group < ngroups; group++) { 319ef2fb679SEric Sandeen desc = ext4_get_group_desc(sb, group, NULL); 320560671a0SAneesh Kumar K.V if (!desc || !ext4_free_inodes_count(sb, desc)) 321ac27a0ecSDave Kleikamp continue; 322560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc) < avefreei) 323ac27a0ecSDave Kleikamp continue; 324ac27a0ecSDave Kleikamp if (!best_desc || 325021b65bbSTheodore Ts'o (ext4_free_group_clusters(sb, desc) > 326021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, best_desc))) { 3272aa9fc4cSAvantika Mathur *best_group = group; 328ac27a0ecSDave Kleikamp best_desc = desc; 3292aa9fc4cSAvantika Mathur ret = 0; 330ac27a0ecSDave Kleikamp } 331ac27a0ecSDave Kleikamp } 3322aa9fc4cSAvantika Mathur return ret; 333ac27a0ecSDave Kleikamp } 334ac27a0ecSDave Kleikamp 335772cb7c8SJose R. Santos #define free_block_ratio 10 336772cb7c8SJose R. Santos 337772cb7c8SJose R. Santos static int find_group_flex(struct super_block *sb, struct inode *parent, 338772cb7c8SJose R. Santos ext4_group_t *best_group) 339772cb7c8SJose R. Santos { 340772cb7c8SJose R. Santos struct ext4_sb_info *sbi = EXT4_SB(sb); 341772cb7c8SJose R. Santos struct ext4_group_desc *desc; 342772cb7c8SJose R. Santos struct flex_groups *flex_group = sbi->s_flex_groups; 343772cb7c8SJose R. Santos ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 344772cb7c8SJose R. Santos ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group); 3458df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 346772cb7c8SJose R. Santos int flex_size = ext4_flex_bg_size(sbi); 347772cb7c8SJose R. Santos ext4_group_t best_flex = parent_fbg_group; 348772cb7c8SJose R. Santos int blocks_per_flex = sbi->s_blocks_per_group * flex_size; 34924aaa8efSTheodore Ts'o int flexbg_free_clusters; 350772cb7c8SJose R. Santos int flex_freeb_ratio; 351772cb7c8SJose R. Santos ext4_group_t n_fbg_groups; 352772cb7c8SJose R. Santos ext4_group_t i; 353772cb7c8SJose R. Santos 3548df9675fSTheodore Ts'o n_fbg_groups = (ngroups + flex_size - 1) >> 355772cb7c8SJose R. Santos sbi->s_log_groups_per_flex; 356772cb7c8SJose R. Santos 357772cb7c8SJose R. Santos find_close_to_parent: 35824aaa8efSTheodore Ts'o flexbg_free_clusters = atomic_read(&flex_group[best_flex].free_clusters); 35924aaa8efSTheodore Ts'o flex_freeb_ratio = EXT4_C2B(sbi, flexbg_free_clusters) * 100 / 36024aaa8efSTheodore Ts'o blocks_per_flex; 3619f24e420STheodore Ts'o if (atomic_read(&flex_group[best_flex].free_inodes) && 362772cb7c8SJose R. Santos flex_freeb_ratio > free_block_ratio) 363772cb7c8SJose R. Santos goto found_flexbg; 364772cb7c8SJose R. Santos 365772cb7c8SJose R. Santos if (best_flex && best_flex == parent_fbg_group) { 366772cb7c8SJose R. Santos best_flex--; 367772cb7c8SJose R. Santos goto find_close_to_parent; 368772cb7c8SJose R. Santos } 369772cb7c8SJose R. Santos 370772cb7c8SJose R. Santos for (i = 0; i < n_fbg_groups; i++) { 371772cb7c8SJose R. Santos if (i == parent_fbg_group || i == parent_fbg_group - 1) 372772cb7c8SJose R. Santos continue; 373772cb7c8SJose R. Santos 37424aaa8efSTheodore Ts'o flexbg_free_clusters = atomic_read(&flex_group[i].free_clusters); 37524aaa8efSTheodore Ts'o flex_freeb_ratio = EXT4_C2B(sbi, flexbg_free_clusters) * 100 / 37624aaa8efSTheodore Ts'o blocks_per_flex; 377772cb7c8SJose R. Santos 378772cb7c8SJose R. Santos if (flex_freeb_ratio > free_block_ratio && 3799f24e420STheodore Ts'o (atomic_read(&flex_group[i].free_inodes))) { 380772cb7c8SJose R. Santos best_flex = i; 381772cb7c8SJose R. Santos goto found_flexbg; 382772cb7c8SJose R. Santos } 383772cb7c8SJose R. Santos 3849f24e420STheodore Ts'o if ((atomic_read(&flex_group[best_flex].free_inodes) == 0) || 38524aaa8efSTheodore Ts'o ((atomic_read(&flex_group[i].free_clusters) > 38624aaa8efSTheodore Ts'o atomic_read(&flex_group[best_flex].free_clusters)) && 3879f24e420STheodore Ts'o atomic_read(&flex_group[i].free_inodes))) 388772cb7c8SJose R. Santos best_flex = i; 389772cb7c8SJose R. Santos } 390772cb7c8SJose R. Santos 3919f24e420STheodore Ts'o if (!atomic_read(&flex_group[best_flex].free_inodes) || 39224aaa8efSTheodore Ts'o !atomic_read(&flex_group[best_flex].free_clusters)) 393772cb7c8SJose R. Santos return -1; 394772cb7c8SJose R. Santos 395772cb7c8SJose R. Santos found_flexbg: 396772cb7c8SJose R. Santos for (i = best_flex * flex_size; i < ngroups && 397772cb7c8SJose R. Santos i < (best_flex + 1) * flex_size; i++) { 39888b6edd1STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 399560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc)) { 400772cb7c8SJose R. Santos *best_group = i; 401772cb7c8SJose R. Santos goto out; 402772cb7c8SJose R. Santos } 403772cb7c8SJose R. Santos } 404772cb7c8SJose R. Santos 405772cb7c8SJose R. Santos return -1; 406772cb7c8SJose R. Santos out: 407772cb7c8SJose R. Santos return 0; 408772cb7c8SJose R. Santos } 409772cb7c8SJose R. Santos 410a4912123STheodore Ts'o struct orlov_stats { 411a4912123STheodore Ts'o __u32 free_inodes; 41224aaa8efSTheodore Ts'o __u32 free_clusters; 413a4912123STheodore Ts'o __u32 used_dirs; 414a4912123STheodore Ts'o }; 415a4912123STheodore Ts'o 416a4912123STheodore Ts'o /* 417a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 418a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 419a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 420a4912123STheodore Ts'o */ 4211f109d5aSTheodore Ts'o static void get_orlov_stats(struct super_block *sb, ext4_group_t g, 422a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 423a4912123STheodore Ts'o { 424a4912123STheodore Ts'o struct ext4_group_desc *desc; 4257d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 426a4912123STheodore Ts'o 4277d39db14STheodore Ts'o if (flex_size > 1) { 4287d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 42924aaa8efSTheodore Ts'o stats->free_clusters = atomic_read(&flex_group[g].free_clusters); 4307d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 4317d39db14STheodore Ts'o return; 4327d39db14STheodore Ts'o } 4337d39db14STheodore Ts'o 4347d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 4357d39db14STheodore Ts'o if (desc) { 4367d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 437021b65bbSTheodore Ts'o stats->free_clusters = ext4_free_group_clusters(sb, desc); 4387d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 4397d39db14STheodore Ts'o } else { 440a4912123STheodore Ts'o stats->free_inodes = 0; 44124aaa8efSTheodore Ts'o stats->free_clusters = 0; 442a4912123STheodore Ts'o stats->used_dirs = 0; 443a4912123STheodore Ts'o } 444a4912123STheodore Ts'o } 445a4912123STheodore Ts'o 446ac27a0ecSDave Kleikamp /* 447ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 448ac27a0ecSDave Kleikamp * 449ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 450ac27a0ecSDave Kleikamp * 451ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 452ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 453ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 454ac27a0ecSDave Kleikamp * 455ac27a0ecSDave Kleikamp * For the rest rules look so: 456ac27a0ecSDave Kleikamp * 457ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 458ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 459ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 460ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4611cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 462ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 463ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 464ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 465ac27a0ecSDave Kleikamp */ 466ac27a0ecSDave Kleikamp 4672aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 468f157a4aaSTheodore Ts'o ext4_group_t *group, int mode, 469f157a4aaSTheodore Ts'o const struct qstr *qstr) 470ac27a0ecSDave Kleikamp { 471fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 472617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 4738df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 474617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 475ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 47624aaa8efSTheodore Ts'o ext4_fsblk_t freeb, avefreec; 477ac27a0ecSDave Kleikamp unsigned int ndirs; 478a4912123STheodore Ts'o int max_dirs, min_inodes; 47924aaa8efSTheodore Ts'o ext4_grpblk_t min_clusters; 4808df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 481617ba13bSMingming Cao struct ext4_group_desc *desc; 482a4912123STheodore Ts'o struct orlov_stats stats; 483a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 484f157a4aaSTheodore Ts'o struct dx_hash_info hinfo; 485a4912123STheodore Ts'o 4868df9675fSTheodore Ts'o ngroups = real_ngroups; 487a4912123STheodore Ts'o if (flex_size > 1) { 4888df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 489a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 490a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 491a4912123STheodore Ts'o } 492ac27a0ecSDave Kleikamp 493ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 494ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 49557042651STheodore Ts'o freeb = EXT4_C2B(sbi, 49657042651STheodore Ts'o percpu_counter_read_positive(&sbi->s_freeclusters_counter)); 49724aaa8efSTheodore Ts'o avefreec = freeb; 49824aaa8efSTheodore Ts'o do_div(avefreec, ngroups); 499ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 500ac27a0ecSDave Kleikamp 501a4912123STheodore Ts'o if (S_ISDIR(mode) && 502a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 50312e9b892SDmitry Monakhov (ext4_test_inode_flag(parent, EXT4_INODE_TOPDIR)))) { 504ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 5052aa9fc4cSAvantika Mathur int ret = -1; 506ac27a0ecSDave Kleikamp 507f157a4aaSTheodore Ts'o if (qstr) { 508f157a4aaSTheodore Ts'o hinfo.hash_version = DX_HASH_HALF_MD4; 509f157a4aaSTheodore Ts'o hinfo.seed = sbi->s_hash_seed; 510f157a4aaSTheodore Ts'o ext4fs_dirhash(qstr->name, qstr->len, &hinfo); 511f157a4aaSTheodore Ts'o grp = hinfo.hash; 512f157a4aaSTheodore Ts'o } else 5132aa9fc4cSAvantika Mathur get_random_bytes(&grp, sizeof(grp)); 5142aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 515ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 516a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 517a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 518a4912123STheodore Ts'o if (!stats.free_inodes) 519ac27a0ecSDave Kleikamp continue; 520a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 521ac27a0ecSDave Kleikamp continue; 522a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 523ac27a0ecSDave Kleikamp continue; 52424aaa8efSTheodore Ts'o if (stats.free_clusters < avefreec) 525ac27a0ecSDave Kleikamp continue; 526a4912123STheodore Ts'o grp = g; 5272aa9fc4cSAvantika Mathur ret = 0; 528a4912123STheodore Ts'o best_ndir = stats.used_dirs; 529ac27a0ecSDave Kleikamp } 530a4912123STheodore Ts'o if (ret) 531a4912123STheodore Ts'o goto fallback; 532a4912123STheodore Ts'o found_flex_bg: 533a4912123STheodore Ts'o if (flex_size == 1) { 534a4912123STheodore Ts'o *group = grp; 535a4912123STheodore Ts'o return 0; 536a4912123STheodore Ts'o } 537a4912123STheodore Ts'o 538a4912123STheodore Ts'o /* 539a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 540a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 541a4912123STheodore Ts'o * something similar, although regular files will 542a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 543a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 544a4912123STheodore Ts'o */ 545a4912123STheodore Ts'o grp *= flex_size; 546a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 5478df9675fSTheodore Ts'o if (grp+i >= real_ngroups) 548a4912123STheodore Ts'o break; 549a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 550a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 551a4912123STheodore Ts'o *group = grp+i; 552a4912123STheodore Ts'o return 0; 553a4912123STheodore Ts'o } 554a4912123STheodore Ts'o } 555ac27a0ecSDave Kleikamp goto fallback; 556ac27a0ecSDave Kleikamp } 557ac27a0ecSDave Kleikamp 558ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 559a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 560a4912123STheodore Ts'o if (min_inodes < 1) 561a4912123STheodore Ts'o min_inodes = 1; 56224aaa8efSTheodore Ts'o min_clusters = avefreec - EXT4_CLUSTERS_PER_GROUP(sb)*flex_size / 4; 563ac27a0ecSDave Kleikamp 564a4912123STheodore Ts'o /* 565a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 566a4912123STheodore Ts'o * inode for this parent directory 567a4912123STheodore Ts'o */ 568a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 569a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 570a4912123STheodore Ts'o if (flex_size > 1) 571a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 572a4912123STheodore Ts'o } 573ac27a0ecSDave Kleikamp 574ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 575a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 576a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 577a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 578ac27a0ecSDave Kleikamp continue; 579a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 580ac27a0ecSDave Kleikamp continue; 58124aaa8efSTheodore Ts'o if (stats.free_clusters < min_clusters) 582ac27a0ecSDave Kleikamp continue; 583a4912123STheodore Ts'o goto found_flex_bg; 584ac27a0ecSDave Kleikamp } 585ac27a0ecSDave Kleikamp 586ac27a0ecSDave Kleikamp fallback: 5878df9675fSTheodore Ts'o ngroups = real_ngroups; 588a4912123STheodore Ts'o avefreei = freei / ngroups; 589b5451f7bSTheodore Ts'o fallback_retry: 590a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 591ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 592a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 593a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 594560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 595a4912123STheodore Ts'o ext4_free_inodes_count(sb, desc) >= avefreei) { 596a4912123STheodore Ts'o *group = grp; 5972aa9fc4cSAvantika Mathur return 0; 598ac27a0ecSDave Kleikamp } 599a4912123STheodore Ts'o } 600ac27a0ecSDave Kleikamp 601ac27a0ecSDave Kleikamp if (avefreei) { 602ac27a0ecSDave Kleikamp /* 603ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 604ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 605ac27a0ecSDave Kleikamp */ 606ac27a0ecSDave Kleikamp avefreei = 0; 607b5451f7bSTheodore Ts'o goto fallback_retry; 608ac27a0ecSDave Kleikamp } 609ac27a0ecSDave Kleikamp 610ac27a0ecSDave Kleikamp return -1; 611ac27a0ecSDave Kleikamp } 612ac27a0ecSDave Kleikamp 6132aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 614a4912123STheodore Ts'o ext4_group_t *group, int mode) 615ac27a0ecSDave Kleikamp { 616fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 6178df9675fSTheodore Ts'o ext4_group_t i, last, ngroups = ext4_get_groups_count(sb); 618617ba13bSMingming Cao struct ext4_group_desc *desc; 619a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 620a4912123STheodore Ts'o 621a4912123STheodore Ts'o /* 622a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 623a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 624a4912123STheodore Ts'o * find another flex group, and store that information in the 625a4912123STheodore Ts'o * parent directory's inode information so that use that flex 626a4912123STheodore Ts'o * group for future allocations. 627a4912123STheodore Ts'o */ 628a4912123STheodore Ts'o if (flex_size > 1) { 629a4912123STheodore Ts'o int retry = 0; 630a4912123STheodore Ts'o 631a4912123STheodore Ts'o try_again: 632a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 633a4912123STheodore Ts'o last = parent_group + flex_size; 634a4912123STheodore Ts'o if (last > ngroups) 635a4912123STheodore Ts'o last = ngroups; 636a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 637a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 638a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 639a4912123STheodore Ts'o *group = i; 640a4912123STheodore Ts'o return 0; 641a4912123STheodore Ts'o } 642a4912123STheodore Ts'o } 643a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 644a4912123STheodore Ts'o retry = 1; 645a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 646a4912123STheodore Ts'o goto try_again; 647a4912123STheodore Ts'o } 648a4912123STheodore Ts'o /* 649a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 650a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 651a4912123STheodore Ts'o * avoid the topdir algorithms. 652a4912123STheodore Ts'o */ 653a4912123STheodore Ts'o *group = parent_group + flex_size; 654a4912123STheodore Ts'o if (*group > ngroups) 655a4912123STheodore Ts'o *group = 0; 6567dc57615SPeter Huewe return find_group_orlov(sb, parent, group, mode, NULL); 657a4912123STheodore Ts'o } 658ac27a0ecSDave Kleikamp 659ac27a0ecSDave Kleikamp /* 660ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 661ac27a0ecSDave Kleikamp */ 6622aa9fc4cSAvantika Mathur *group = parent_group; 6632aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 664560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 665021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc)) 6662aa9fc4cSAvantika Mathur return 0; 667ac27a0ecSDave Kleikamp 668ac27a0ecSDave Kleikamp /* 669ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 670ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 671ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 672ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 673ac27a0ecSDave Kleikamp * different blockgroup. 674ac27a0ecSDave Kleikamp * 675ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 676ac27a0ecSDave Kleikamp */ 6772aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 678ac27a0ecSDave Kleikamp 679ac27a0ecSDave Kleikamp /* 680ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 681ac27a0ecSDave Kleikamp * blocks. 682ac27a0ecSDave Kleikamp */ 683ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6842aa9fc4cSAvantika Mathur *group += i; 6852aa9fc4cSAvantika Mathur if (*group >= ngroups) 6862aa9fc4cSAvantika Mathur *group -= ngroups; 6872aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 688560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 689021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc)) 6902aa9fc4cSAvantika Mathur return 0; 691ac27a0ecSDave Kleikamp } 692ac27a0ecSDave Kleikamp 693ac27a0ecSDave Kleikamp /* 694ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 695ac27a0ecSDave Kleikamp * has no free blocks. 696ac27a0ecSDave Kleikamp */ 6972aa9fc4cSAvantika Mathur *group = parent_group; 698ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6992aa9fc4cSAvantika Mathur if (++*group >= ngroups) 7002aa9fc4cSAvantika Mathur *group = 0; 7012aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 702560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 7032aa9fc4cSAvantika Mathur return 0; 704ac27a0ecSDave Kleikamp } 705ac27a0ecSDave Kleikamp 706ac27a0ecSDave Kleikamp return -1; 707ac27a0ecSDave Kleikamp } 708ac27a0ecSDave Kleikamp 709ac27a0ecSDave Kleikamp /* 71039341867SAneesh Kumar K.V * claim the inode from the inode bitmap. If the group 711955ce5f5SAneesh Kumar K.V * is uninit we need to take the groups's ext4_group_lock 71239341867SAneesh Kumar K.V * and clear the uninit flag. The inode bitmap update 71339341867SAneesh Kumar K.V * and group desc uninit flag clear should be done 714955ce5f5SAneesh Kumar K.V * after holding ext4_group_lock so that ext4_read_inode_bitmap 71539341867SAneesh Kumar K.V * doesn't race with the ext4_claim_inode 71639341867SAneesh Kumar K.V */ 71739341867SAneesh Kumar K.V static int ext4_claim_inode(struct super_block *sb, 71839341867SAneesh Kumar K.V struct buffer_head *inode_bitmap_bh, 71939341867SAneesh Kumar K.V unsigned long ino, ext4_group_t group, int mode) 72039341867SAneesh Kumar K.V { 72139341867SAneesh Kumar K.V int free = 0, retval = 0, count; 72239341867SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 723bfff6873SLukas Czerner struct ext4_group_info *grp = ext4_get_group_info(sb, group); 72439341867SAneesh Kumar K.V struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL); 72539341867SAneesh Kumar K.V 726bfff6873SLukas Czerner /* 727bfff6873SLukas Czerner * We have to be sure that new inode allocation does not race with 728bfff6873SLukas Czerner * inode table initialization, because otherwise we may end up 729bfff6873SLukas Czerner * allocating and writing new inode right before sb_issue_zeroout 730bfff6873SLukas Czerner * takes place and overwriting our new inode with zeroes. So we 731bfff6873SLukas Czerner * take alloc_sem to prevent it. 732bfff6873SLukas Czerner */ 733bfff6873SLukas Czerner down_read(&grp->alloc_sem); 734955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 73539341867SAneesh Kumar K.V if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) { 73639341867SAneesh Kumar K.V /* not a free inode */ 73739341867SAneesh Kumar K.V retval = 1; 73839341867SAneesh Kumar K.V goto err_ret; 73939341867SAneesh Kumar K.V } 74039341867SAneesh Kumar K.V ino++; 74139341867SAneesh Kumar K.V if ((group == 0 && ino < EXT4_FIRST_INO(sb)) || 74239341867SAneesh Kumar K.V ino > EXT4_INODES_PER_GROUP(sb)) { 743955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 744bfff6873SLukas Czerner up_read(&grp->alloc_sem); 74512062dddSEric Sandeen ext4_error(sb, "reserved inode or inode > inodes count - " 74639341867SAneesh Kumar K.V "block_group = %u, inode=%lu", group, 74739341867SAneesh Kumar K.V ino + group * EXT4_INODES_PER_GROUP(sb)); 74839341867SAneesh Kumar K.V return 1; 74939341867SAneesh Kumar K.V } 75039341867SAneesh Kumar K.V /* If we didn't allocate from within the initialized part of the inode 75139341867SAneesh Kumar K.V * table then we need to initialize up to this inode. */ 75239341867SAneesh Kumar K.V if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) { 75339341867SAneesh Kumar K.V 75439341867SAneesh Kumar K.V if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 75539341867SAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 75639341867SAneesh Kumar K.V /* When marking the block group with 75739341867SAneesh Kumar K.V * ~EXT4_BG_INODE_UNINIT we don't want to depend 75839341867SAneesh Kumar K.V * on the value of bg_itable_unused even though 75939341867SAneesh Kumar K.V * mke2fs could have initialized the same for us. 76039341867SAneesh Kumar K.V * Instead we calculated the value below 76139341867SAneesh Kumar K.V */ 76239341867SAneesh Kumar K.V 76339341867SAneesh Kumar K.V free = 0; 76439341867SAneesh Kumar K.V } else { 76539341867SAneesh Kumar K.V free = EXT4_INODES_PER_GROUP(sb) - 76639341867SAneesh Kumar K.V ext4_itable_unused_count(sb, gdp); 76739341867SAneesh Kumar K.V } 76839341867SAneesh Kumar K.V 76939341867SAneesh Kumar K.V /* 77039341867SAneesh Kumar K.V * Check the relative inode number against the last used 77139341867SAneesh Kumar K.V * relative inode number in this group. if it is greater 77239341867SAneesh Kumar K.V * we need to update the bg_itable_unused count 77339341867SAneesh Kumar K.V * 77439341867SAneesh Kumar K.V */ 77539341867SAneesh Kumar K.V if (ino > free) 77639341867SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 77739341867SAneesh Kumar K.V (EXT4_INODES_PER_GROUP(sb) - ino)); 77839341867SAneesh Kumar K.V } 77939341867SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) - 1; 78039341867SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 78139341867SAneesh Kumar K.V if (S_ISDIR(mode)) { 78239341867SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) + 1; 78339341867SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 7847d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 7857d39db14STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 7867d39db14STheodore Ts'o 787c4caae25SEric Sandeen atomic_inc(&sbi->s_flex_groups[f].used_dirs); 7887d39db14STheodore Ts'o } 78939341867SAneesh Kumar K.V } 79039341867SAneesh Kumar K.V gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 79139341867SAneesh Kumar K.V err_ret: 792955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 793bfff6873SLukas Czerner up_read(&grp->alloc_sem); 79439341867SAneesh Kumar K.V return retval; 79539341867SAneesh Kumar K.V } 79639341867SAneesh Kumar K.V 79739341867SAneesh Kumar K.V /* 798ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 799ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 800ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 801ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 802ac27a0ecSDave Kleikamp * directories already is chosen. 803ac27a0ecSDave Kleikamp * 804ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 805ac27a0ecSDave Kleikamp * group to find a free inode. 806ac27a0ecSDave Kleikamp */ 807f157a4aaSTheodore Ts'o struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode, 80811013911SAndreas Dilger const struct qstr *qstr, __u32 goal) 809ac27a0ecSDave Kleikamp { 810ac27a0ecSDave Kleikamp struct super_block *sb; 8113300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 8123300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 8138df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 814ac27a0ecSDave Kleikamp unsigned long ino = 0; 815ac27a0ecSDave Kleikamp struct inode *inode; 816617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 817617ba13bSMingming Cao struct ext4_inode_info *ei; 818617ba13bSMingming Cao struct ext4_sb_info *sbi; 81939341867SAneesh Kumar K.V int ret2, err = 0; 820ac27a0ecSDave Kleikamp struct inode *ret; 8212aa9fc4cSAvantika Mathur ext4_group_t i; 8222842c3b5STheodore Ts'o static int once = 1; 823772cb7c8SJose R. Santos ext4_group_t flex_group; 824ac27a0ecSDave Kleikamp 825ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 826ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 827ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 828ac27a0ecSDave Kleikamp 829ac27a0ecSDave Kleikamp sb = dir->i_sb; 8308df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 8319bffad1eSTheodore Ts'o trace_ext4_request_inode(dir, mode); 832ac27a0ecSDave Kleikamp inode = new_inode(sb); 833ac27a0ecSDave Kleikamp if (!inode) 834ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 835617ba13bSMingming Cao ei = EXT4_I(inode); 836617ba13bSMingming Cao sbi = EXT4_SB(sb); 837772cb7c8SJose R. Santos 83811013911SAndreas Dilger if (!goal) 83911013911SAndreas Dilger goal = sbi->s_inode_goal; 84011013911SAndreas Dilger 841e6462869SJohann Lombardi if (goal && goal <= le32_to_cpu(sbi->s_es->s_inodes_count)) { 84211013911SAndreas Dilger group = (goal - 1) / EXT4_INODES_PER_GROUP(sb); 84311013911SAndreas Dilger ino = (goal - 1) % EXT4_INODES_PER_GROUP(sb); 84411013911SAndreas Dilger ret2 = 0; 84511013911SAndreas Dilger goto got_group; 84611013911SAndreas Dilger } 84711013911SAndreas Dilger 848a4912123STheodore Ts'o if (sbi->s_log_groups_per_flex && test_opt(sb, OLDALLOC)) { 849772cb7c8SJose R. Santos ret2 = find_group_flex(sb, dir, &group); 85005bf9e83STheodore Ts'o if (ret2 == -1) { 851a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 8526b82f3cbSChuck Ebbert if (ret2 == 0 && once) { 8532842c3b5STheodore Ts'o once = 0; 85405bf9e83STheodore Ts'o printk(KERN_NOTICE "ext4: find_group_flex " 85505bf9e83STheodore Ts'o "failed, fallback succeeded dir %lu\n", 85605bf9e83STheodore Ts'o dir->i_ino); 85705bf9e83STheodore Ts'o } 8586b82f3cbSChuck Ebbert } 859772cb7c8SJose R. Santos goto got_group; 860772cb7c8SJose R. Santos } 861772cb7c8SJose R. Santos 862ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) { 863ac27a0ecSDave Kleikamp if (test_opt(sb, OLDALLOC)) 8642aa9fc4cSAvantika Mathur ret2 = find_group_dir(sb, dir, &group); 865ac27a0ecSDave Kleikamp else 866f157a4aaSTheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode, qstr); 867ac27a0ecSDave Kleikamp } else 868a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 869ac27a0ecSDave Kleikamp 870772cb7c8SJose R. Santos got_group: 871a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 872ac27a0ecSDave Kleikamp err = -ENOSPC; 8732aa9fc4cSAvantika Mathur if (ret2 == -1) 874ac27a0ecSDave Kleikamp goto out; 875ac27a0ecSDave Kleikamp 87611013911SAndreas Dilger for (i = 0; i < ngroups; i++, ino = 0) { 877ac27a0ecSDave Kleikamp err = -EIO; 878ac27a0ecSDave Kleikamp 8793300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 880ac27a0ecSDave Kleikamp if (!gdp) 881ac27a0ecSDave Kleikamp goto fail; 882ac27a0ecSDave Kleikamp 8833300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 8843300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 8853300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 886ac27a0ecSDave Kleikamp goto fail; 887ac27a0ecSDave Kleikamp 888ac27a0ecSDave Kleikamp repeat_in_this_group: 889617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 8903300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 8913300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 8923300bedaSAneesh Kumar K.V 893617ba13bSMingming Cao if (ino < EXT4_INODES_PER_GROUP(sb)) { 894ac27a0ecSDave Kleikamp 8953300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 8963300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 8973300bedaSAneesh Kumar K.V inode_bitmap_bh); 898ac27a0ecSDave Kleikamp if (err) 899ac27a0ecSDave Kleikamp goto fail; 900ac27a0ecSDave Kleikamp 90139341867SAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "get_write_access"); 90239341867SAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 90339341867SAneesh Kumar K.V group_desc_bh); 90439341867SAneesh Kumar K.V if (err) 90539341867SAneesh Kumar K.V goto fail; 90639341867SAneesh Kumar K.V if (!ext4_claim_inode(sb, inode_bitmap_bh, 90739341867SAneesh Kumar K.V ino, group, mode)) { 908ac27a0ecSDave Kleikamp /* we won it */ 9093300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, 9100390131bSFrank Mayhar "call ext4_handle_dirty_metadata"); 9110390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 91273b50c1cSCurt Wohlgemuth NULL, 9133300bedaSAneesh Kumar K.V inode_bitmap_bh); 914ac27a0ecSDave Kleikamp if (err) 915ac27a0ecSDave Kleikamp goto fail; 91639341867SAneesh Kumar K.V /* zero bit is inode number 1*/ 91739341867SAneesh Kumar K.V ino++; 918ac27a0ecSDave Kleikamp goto got; 919ac27a0ecSDave Kleikamp } 920ac27a0ecSDave Kleikamp /* we lost it */ 9213300bedaSAneesh Kumar K.V ext4_handle_release_buffer(handle, inode_bitmap_bh); 92239341867SAneesh Kumar K.V ext4_handle_release_buffer(handle, group_desc_bh); 923ac27a0ecSDave Kleikamp 924617ba13bSMingming Cao if (++ino < EXT4_INODES_PER_GROUP(sb)) 925ac27a0ecSDave Kleikamp goto repeat_in_this_group; 926ac27a0ecSDave Kleikamp } 927ac27a0ecSDave Kleikamp 928ac27a0ecSDave Kleikamp /* 929ac27a0ecSDave Kleikamp * This case is possible in concurrent environment. It is very 930ac27a0ecSDave Kleikamp * rare. We cannot repeat the find_group_xxx() call because 931ac27a0ecSDave Kleikamp * that will simply return the same blockgroup, because the 932ac27a0ecSDave Kleikamp * group descriptor metadata has not yet been updated. 933ac27a0ecSDave Kleikamp * So we just go onto the next blockgroup. 934ac27a0ecSDave Kleikamp */ 9358df9675fSTheodore Ts'o if (++group == ngroups) 936ac27a0ecSDave Kleikamp group = 0; 937ac27a0ecSDave Kleikamp } 938ac27a0ecSDave Kleikamp err = -ENOSPC; 939ac27a0ecSDave Kleikamp goto out; 940ac27a0ecSDave Kleikamp 941ac27a0ecSDave Kleikamp got: 942717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 943717d50e4SAndreas Dilger if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) && 944717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9453300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 946717d50e4SAndreas Dilger 9473300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 9483300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 9493300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 950717d50e4SAndreas Dilger if (err) { 9513300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 952717d50e4SAndreas Dilger goto fail; 953717d50e4SAndreas Dilger } 954717d50e4SAndreas Dilger 955fd034a84STheodore Ts'o BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 956fd034a84STheodore Ts'o err = ext4_handle_dirty_metadata(handle, NULL, block_bitmap_bh); 957fd034a84STheodore Ts'o brelse(block_bitmap_bh); 958fd034a84STheodore Ts'o 959717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 960fd034a84STheodore Ts'o ext4_lock_group(sb, group); 961717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9623300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 963021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 964fd034a84STheodore Ts'o ext4_free_blocks_after_init(sb, group, gdp)); 96523712a9cSFrederic Bohe gdp->bg_checksum = ext4_group_desc_csum(sbi, group, 96623712a9cSFrederic Bohe gdp); 967717d50e4SAndreas Dilger } 968955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 969717d50e4SAndreas Dilger 970717d50e4SAndreas Dilger if (err) 971717d50e4SAndreas Dilger goto fail; 972717d50e4SAndreas Dilger } 9733300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 9743300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 97539341867SAneesh Kumar K.V if (err) 97639341867SAneesh Kumar K.V goto fail; 977ac27a0ecSDave Kleikamp 978ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 979ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 980ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 981a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 982ac27a0ecSDave Kleikamp 983772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 984772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 9859f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 986772cb7c8SJose R. Santos } 987772cb7c8SJose R. Santos 988b10b8520SDmitry Monakhov if (test_opt(sb, GRPID)) { 989ac27a0ecSDave Kleikamp inode->i_mode = mode; 990b10b8520SDmitry Monakhov inode->i_uid = current_fsuid(); 991b10b8520SDmitry Monakhov inode->i_gid = dir->i_gid; 992b10b8520SDmitry Monakhov } else 993b10b8520SDmitry Monakhov inode_init_owner(inode, dir, mode); 994ac27a0ecSDave Kleikamp 995717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 996ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 997ac27a0ecSDave Kleikamp inode->i_blocks = 0; 998ef7f3835SKalpak Shah inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime = 999ef7f3835SKalpak Shah ext4_current_time(inode); 1000ac27a0ecSDave Kleikamp 1001ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 1002ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 1003ac27a0ecSDave Kleikamp ei->i_disksize = 0; 1004ac27a0ecSDave Kleikamp 100542bf0383SAneesh Kumar K.V /* 10062dc6b0d4SDuane Griffin * Don't inherit extent flag from directory, amongst others. We set 10072dc6b0d4SDuane Griffin * extent flag on newly created directory and file only if -o extent 10082dc6b0d4SDuane Griffin * mount option is specified 100942bf0383SAneesh Kumar K.V */ 10102dc6b0d4SDuane Griffin ei->i_flags = 10112dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 1012ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 1013ac27a0ecSDave Kleikamp ei->i_dtime = 0; 1014ac27a0ecSDave Kleikamp ei->i_block_group = group; 1015a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 1016ac27a0ecSDave Kleikamp 1017617ba13bSMingming Cao ext4_set_inode_flags(inode); 1018ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 10190390131bSFrank Mayhar ext4_handle_sync(handle); 10206b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 10216b38e842SAl Viro err = -EINVAL; 10226b38e842SAl Viro goto fail_drop; 10236b38e842SAl Viro } 1024ac27a0ecSDave Kleikamp spin_lock(&sbi->s_next_gen_lock); 1025ac27a0ecSDave Kleikamp inode->i_generation = sbi->s_next_generation++; 1026ac27a0ecSDave Kleikamp spin_unlock(&sbi->s_next_gen_lock); 1027ac27a0ecSDave Kleikamp 1028353eb83cSTheodore Ts'o ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */ 102919f5fb7aSTheodore Ts'o ext4_set_inode_state(inode, EXT4_STATE_NEW); 1030ef7f3835SKalpak Shah 1031ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 1032ac27a0ecSDave Kleikamp 1033ac27a0ecSDave Kleikamp ret = inode; 1034871a2931SChristoph Hellwig dquot_initialize(inode); 103563936ddaSChristoph Hellwig err = dquot_alloc_inode(inode); 103663936ddaSChristoph Hellwig if (err) 1037ac27a0ecSDave Kleikamp goto fail_drop; 1038ac27a0ecSDave Kleikamp 1039617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1040ac27a0ecSDave Kleikamp if (err) 1041ac27a0ecSDave Kleikamp goto fail_free_drop; 1042ac27a0ecSDave Kleikamp 10432a7dba39SEric Paris err = ext4_init_security(handle, inode, dir, qstr); 1044ac27a0ecSDave Kleikamp if (err) 1045ac27a0ecSDave Kleikamp goto fail_free_drop; 1046ac27a0ecSDave Kleikamp 104783982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 1048e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1049e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 105012e9b892SDmitry Monakhov ext4_set_inode_flag(inode, EXT4_INODE_EXTENTS); 1051a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1052a86c6181SAlex Tomas } 105342bf0383SAneesh Kumar K.V } 1054ac27a0ecSDave Kleikamp 1055688f869cSTheodore Ts'o if (ext4_handle_valid(handle)) { 1056688f869cSTheodore Ts'o ei->i_sync_tid = handle->h_transaction->t_tid; 1057688f869cSTheodore Ts'o ei->i_datasync_tid = handle->h_transaction->t_tid; 1058688f869cSTheodore Ts'o } 1059688f869cSTheodore Ts'o 10608753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 10618753e88fSAneesh Kumar K.V if (err) { 10628753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 10638753e88fSAneesh Kumar K.V goto fail_free_drop; 10648753e88fSAneesh Kumar K.V } 10658753e88fSAneesh Kumar K.V 1066617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 10679bffad1eSTheodore Ts'o trace_ext4_allocate_inode(inode, dir, mode); 1068ac27a0ecSDave Kleikamp goto really_out; 1069ac27a0ecSDave Kleikamp fail: 1070617ba13bSMingming Cao ext4_std_error(sb, err); 1071ac27a0ecSDave Kleikamp out: 1072ac27a0ecSDave Kleikamp iput(inode); 1073ac27a0ecSDave Kleikamp ret = ERR_PTR(err); 1074ac27a0ecSDave Kleikamp really_out: 10753300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1076ac27a0ecSDave Kleikamp return ret; 1077ac27a0ecSDave Kleikamp 1078ac27a0ecSDave Kleikamp fail_free_drop: 107963936ddaSChristoph Hellwig dquot_free_inode(inode); 1080ac27a0ecSDave Kleikamp 1081ac27a0ecSDave Kleikamp fail_drop: 10829f754758SChristoph Hellwig dquot_drop(inode); 1083ac27a0ecSDave Kleikamp inode->i_flags |= S_NOQUOTA; 1084ac27a0ecSDave Kleikamp inode->i_nlink = 0; 10856b38e842SAl Viro unlock_new_inode(inode); 1086ac27a0ecSDave Kleikamp iput(inode); 10873300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1088ac27a0ecSDave Kleikamp return ERR_PTR(err); 1089ac27a0ecSDave Kleikamp } 1090ac27a0ecSDave Kleikamp 1091ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1092617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1093ac27a0ecSDave Kleikamp { 1094617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1095fd2d4291SAvantika Mathur ext4_group_t block_group; 1096ac27a0ecSDave Kleikamp int bit; 10971d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 1098ac27a0ecSDave Kleikamp struct inode *inode = NULL; 10991d1fe1eeSDavid Howells long err = -EIO; 1100ac27a0ecSDave Kleikamp 1101ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1102ac27a0ecSDave Kleikamp if (ino > max_ino) { 110312062dddSEric Sandeen ext4_warning(sb, "bad orphan ino %lu! e2fsck was run?", ino); 11041d1fe1eeSDavid Howells goto error; 1105ac27a0ecSDave Kleikamp } 1106ac27a0ecSDave Kleikamp 1107617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1108617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1109e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1110ac27a0ecSDave Kleikamp if (!bitmap_bh) { 111112062dddSEric Sandeen ext4_warning(sb, "inode bitmap error for orphan %lu", ino); 11121d1fe1eeSDavid Howells goto error; 1113ac27a0ecSDave Kleikamp } 1114ac27a0ecSDave Kleikamp 1115ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1116ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1117ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1118ac27a0ecSDave Kleikamp */ 11191d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 11201d1fe1eeSDavid Howells goto bad_orphan; 11211d1fe1eeSDavid Howells 11221d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 11231d1fe1eeSDavid Howells if (IS_ERR(inode)) 11241d1fe1eeSDavid Howells goto iget_failed; 11251d1fe1eeSDavid Howells 112691ef4cafSDuane Griffin /* 112791ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 112891ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 112991ef4cafSDuane Griffin * during processing and an infinite loop will result. 113091ef4cafSDuane Griffin */ 113191ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 113291ef4cafSDuane Griffin goto bad_orphan; 113391ef4cafSDuane Griffin 11341d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 11351d1fe1eeSDavid Howells goto bad_orphan; 11361d1fe1eeSDavid Howells brelse(bitmap_bh); 11371d1fe1eeSDavid Howells return inode; 11381d1fe1eeSDavid Howells 11391d1fe1eeSDavid Howells iget_failed: 11401d1fe1eeSDavid Howells err = PTR_ERR(inode); 11411d1fe1eeSDavid Howells inode = NULL; 11421d1fe1eeSDavid Howells bad_orphan: 114312062dddSEric Sandeen ext4_warning(sb, "bad orphan inode %lu! e2fsck was run?", ino); 1144617ba13bSMingming Cao printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1145ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1146617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1147ac27a0ecSDave Kleikamp printk(KERN_NOTICE "inode=%p\n", inode); 1148ac27a0ecSDave Kleikamp if (inode) { 1149ac27a0ecSDave Kleikamp printk(KERN_NOTICE "is_bad_inode(inode)=%d\n", 1150ac27a0ecSDave Kleikamp is_bad_inode(inode)); 1151ac27a0ecSDave Kleikamp printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n", 1152ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 1153ac27a0ecSDave Kleikamp printk(KERN_NOTICE "max_ino=%lu\n", max_ino); 115491ef4cafSDuane Griffin printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink); 1155ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 11561d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1157ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1158ac27a0ecSDave Kleikamp iput(inode); 1159ac27a0ecSDave Kleikamp } 1160ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11611d1fe1eeSDavid Howells error: 11621d1fe1eeSDavid Howells return ERR_PTR(err); 1163ac27a0ecSDave Kleikamp } 1164ac27a0ecSDave Kleikamp 1165617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1166ac27a0ecSDave Kleikamp { 1167ac27a0ecSDave Kleikamp unsigned long desc_count; 1168617ba13bSMingming Cao struct ext4_group_desc *gdp; 11698df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1170617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1171617ba13bSMingming Cao struct ext4_super_block *es; 1172ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1173ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1174ac27a0ecSDave Kleikamp 1175617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1176ac27a0ecSDave Kleikamp desc_count = 0; 1177ac27a0ecSDave Kleikamp bitmap_count = 0; 1178ac27a0ecSDave Kleikamp gdp = NULL; 11798df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1180617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1181ac27a0ecSDave Kleikamp if (!gdp) 1182ac27a0ecSDave Kleikamp continue; 1183560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1184ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1185e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1186ac27a0ecSDave Kleikamp if (!bitmap_bh) 1187ac27a0ecSDave Kleikamp continue; 1188ac27a0ecSDave Kleikamp 1189617ba13bSMingming Cao x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8); 1190c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1191785b4b3aSPeng Tao (unsigned long) i, ext4_free_inodes_count(sb, gdp), x); 1192ac27a0ecSDave Kleikamp bitmap_count += x; 1193ac27a0ecSDave Kleikamp } 1194ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11954776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 11964776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1197ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1198ac27a0ecSDave Kleikamp return desc_count; 1199ac27a0ecSDave Kleikamp #else 1200ac27a0ecSDave Kleikamp desc_count = 0; 12018df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1202617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1203ac27a0ecSDave Kleikamp if (!gdp) 1204ac27a0ecSDave Kleikamp continue; 1205560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1206ac27a0ecSDave Kleikamp cond_resched(); 1207ac27a0ecSDave Kleikamp } 1208ac27a0ecSDave Kleikamp return desc_count; 1209ac27a0ecSDave Kleikamp #endif 1210ac27a0ecSDave Kleikamp } 1211ac27a0ecSDave Kleikamp 1212ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1213617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1214ac27a0ecSDave Kleikamp { 1215ac27a0ecSDave Kleikamp unsigned long count = 0; 12168df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1217ac27a0ecSDave Kleikamp 12188df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1219617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1220ac27a0ecSDave Kleikamp if (!gdp) 1221ac27a0ecSDave Kleikamp continue; 1222560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1223ac27a0ecSDave Kleikamp } 1224ac27a0ecSDave Kleikamp return count; 1225ac27a0ecSDave Kleikamp } 1226bfff6873SLukas Czerner 1227bfff6873SLukas Czerner /* 1228bfff6873SLukas Czerner * Zeroes not yet zeroed inode table - just write zeroes through the whole 1229bfff6873SLukas Czerner * inode table. Must be called without any spinlock held. The only place 1230bfff6873SLukas Czerner * where it is called from on active part of filesystem is ext4lazyinit 1231bfff6873SLukas Czerner * thread, so we do not need any special locks, however we have to prevent 1232bfff6873SLukas Czerner * inode allocation from the current group, so we take alloc_sem lock, to 1233bfff6873SLukas Czerner * block ext4_claim_inode until we are finished. 1234bfff6873SLukas Czerner */ 1235bfff6873SLukas Czerner extern int ext4_init_inode_table(struct super_block *sb, ext4_group_t group, 1236bfff6873SLukas Czerner int barrier) 1237bfff6873SLukas Czerner { 1238bfff6873SLukas Czerner struct ext4_group_info *grp = ext4_get_group_info(sb, group); 1239bfff6873SLukas Czerner struct ext4_sb_info *sbi = EXT4_SB(sb); 1240bfff6873SLukas Czerner struct ext4_group_desc *gdp = NULL; 1241bfff6873SLukas Czerner struct buffer_head *group_desc_bh; 1242bfff6873SLukas Czerner handle_t *handle; 1243bfff6873SLukas Czerner ext4_fsblk_t blk; 1244bfff6873SLukas Czerner int num, ret = 0, used_blks = 0; 1245bfff6873SLukas Czerner 1246bfff6873SLukas Czerner /* This should not happen, but just to be sure check this */ 1247bfff6873SLukas Czerner if (sb->s_flags & MS_RDONLY) { 1248bfff6873SLukas Czerner ret = 1; 1249bfff6873SLukas Czerner goto out; 1250bfff6873SLukas Czerner } 1251bfff6873SLukas Czerner 1252bfff6873SLukas Czerner gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 1253bfff6873SLukas Czerner if (!gdp) 1254bfff6873SLukas Czerner goto out; 1255bfff6873SLukas Czerner 1256bfff6873SLukas Czerner /* 1257bfff6873SLukas Czerner * We do not need to lock this, because we are the only one 1258bfff6873SLukas Czerner * handling this flag. 1259bfff6873SLukas Czerner */ 1260bfff6873SLukas Czerner if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_ZEROED)) 1261bfff6873SLukas Czerner goto out; 1262bfff6873SLukas Czerner 1263bfff6873SLukas Czerner handle = ext4_journal_start_sb(sb, 1); 1264bfff6873SLukas Czerner if (IS_ERR(handle)) { 1265bfff6873SLukas Czerner ret = PTR_ERR(handle); 1266bfff6873SLukas Czerner goto out; 1267bfff6873SLukas Czerner } 1268bfff6873SLukas Czerner 1269bfff6873SLukas Czerner down_write(&grp->alloc_sem); 1270bfff6873SLukas Czerner /* 1271bfff6873SLukas Czerner * If inode bitmap was already initialized there may be some 1272bfff6873SLukas Czerner * used inodes so we need to skip blocks with used inodes in 1273bfff6873SLukas Czerner * inode table. 1274bfff6873SLukas Czerner */ 1275bfff6873SLukas Czerner if (!(gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT))) 1276bfff6873SLukas Czerner used_blks = DIV_ROUND_UP((EXT4_INODES_PER_GROUP(sb) - 1277bfff6873SLukas Czerner ext4_itable_unused_count(sb, gdp)), 1278bfff6873SLukas Czerner sbi->s_inodes_per_block); 1279bfff6873SLukas Czerner 1280857ac889SLukas Czerner if ((used_blks < 0) || (used_blks > sbi->s_itb_per_group)) { 1281857ac889SLukas Czerner ext4_error(sb, "Something is wrong with group %u\n" 1282857ac889SLukas Czerner "Used itable blocks: %d" 1283857ac889SLukas Czerner "itable unused count: %u\n", 1284857ac889SLukas Czerner group, used_blks, 1285857ac889SLukas Czerner ext4_itable_unused_count(sb, gdp)); 1286857ac889SLukas Czerner ret = 1; 128733853a0dSYongqiang Yang goto err_out; 1288857ac889SLukas Czerner } 1289857ac889SLukas Czerner 1290bfff6873SLukas Czerner blk = ext4_inode_table(sb, gdp) + used_blks; 1291bfff6873SLukas Czerner num = sbi->s_itb_per_group - used_blks; 1292bfff6873SLukas Czerner 1293bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, "get_write_access"); 1294bfff6873SLukas Czerner ret = ext4_journal_get_write_access(handle, 1295bfff6873SLukas Czerner group_desc_bh); 1296bfff6873SLukas Czerner if (ret) 1297bfff6873SLukas Czerner goto err_out; 1298bfff6873SLukas Czerner 1299bfff6873SLukas Czerner /* 1300bfff6873SLukas Czerner * Skip zeroout if the inode table is full. But we set the ZEROED 1301bfff6873SLukas Czerner * flag anyway, because obviously, when it is full it does not need 1302bfff6873SLukas Czerner * further zeroing. 1303bfff6873SLukas Czerner */ 1304bfff6873SLukas Czerner if (unlikely(num == 0)) 1305bfff6873SLukas Czerner goto skip_zeroout; 1306bfff6873SLukas Czerner 1307bfff6873SLukas Czerner ext4_debug("going to zero out inode table in group %d\n", 1308bfff6873SLukas Czerner group); 1309a107e5a3STheodore Ts'o ret = sb_issue_zeroout(sb, blk, num, GFP_NOFS); 1310bfff6873SLukas Czerner if (ret < 0) 1311bfff6873SLukas Czerner goto err_out; 1312a107e5a3STheodore Ts'o if (barrier) 1313a107e5a3STheodore Ts'o blkdev_issue_flush(sb->s_bdev, GFP_NOFS, NULL); 1314bfff6873SLukas Czerner 1315bfff6873SLukas Czerner skip_zeroout: 1316bfff6873SLukas Czerner ext4_lock_group(sb, group); 1317bfff6873SLukas Czerner gdp->bg_flags |= cpu_to_le16(EXT4_BG_INODE_ZEROED); 1318bfff6873SLukas Czerner gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 1319bfff6873SLukas Czerner ext4_unlock_group(sb, group); 1320bfff6873SLukas Czerner 1321bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, 1322bfff6873SLukas Czerner "call ext4_handle_dirty_metadata"); 1323bfff6873SLukas Czerner ret = ext4_handle_dirty_metadata(handle, NULL, 1324bfff6873SLukas Czerner group_desc_bh); 1325bfff6873SLukas Czerner 1326bfff6873SLukas Czerner err_out: 1327bfff6873SLukas Czerner up_write(&grp->alloc_sem); 1328bfff6873SLukas Czerner ext4_journal_stop(handle); 1329bfff6873SLukas Czerner out: 1330bfff6873SLukas Czerner return ret; 1331bfff6873SLukas Czerner } 1332