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 */ 53717d50e4SAndreas Dilger void 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 */ 68fd2d4291SAvantika Mathur unsigned ext4_init_inode_bitmap(struct super_block *sb, struct buffer_head *bh, 69fd2d4291SAvantika Mathur ext4_group_t block_group, 70717d50e4SAndreas Dilger struct ext4_group_desc *gdp) 71717d50e4SAndreas Dilger { 72717d50e4SAndreas Dilger struct ext4_sb_info *sbi = EXT4_SB(sb); 73717d50e4SAndreas Dilger 74717d50e4SAndreas Dilger J_ASSERT_BH(bh, buffer_locked(bh)); 75717d50e4SAndreas Dilger 76717d50e4SAndreas Dilger /* If checksum is bad mark all blocks and inodes use to prevent 77717d50e4SAndreas Dilger * allocation, essentially implementing a per-group read-only flag. */ 78717d50e4SAndreas Dilger if (!ext4_group_desc_csum_verify(sbi, block_group, gdp)) { 79a9df9a49STheodore Ts'o ext4_error(sb, __func__, "Checksum bad for group %u", 80717d50e4SAndreas Dilger 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); 89648f5879SAneesh Kumar K.V 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; 111e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 112e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 113e29d1cdeSEric Sandeen if (unlikely(!bh)) { 114e29d1cdeSEric Sandeen ext4_error(sb, __func__, 115ac27a0ecSDave Kleikamp "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 } 128955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 129e29d1cdeSEric Sandeen if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 130e29d1cdeSEric Sandeen ext4_init_inode_bitmap(sb, bh, block_group, desc); 1312ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 132e29d1cdeSEric Sandeen set_buffer_uptodate(bh); 133955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1343300bedaSAneesh Kumar K.V unlock_buffer(bh); 135e29d1cdeSEric Sandeen return bh; 136e29d1cdeSEric Sandeen } 137955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1382ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh)) { 1392ccb5fb9SAneesh Kumar K.V /* 1402ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 1412ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 1422ccb5fb9SAneesh Kumar K.V */ 1432ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 1442ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1452ccb5fb9SAneesh Kumar K.V return bh; 1462ccb5fb9SAneesh Kumar K.V } 1472ccb5fb9SAneesh Kumar K.V /* 1482ccb5fb9SAneesh Kumar K.V * submit the buffer_head for read. We can 1492ccb5fb9SAneesh Kumar K.V * safely mark the bitmap as uptodate now. 1502ccb5fb9SAneesh Kumar K.V * We do it here so the bitmap uptodate bit 1512ccb5fb9SAneesh Kumar K.V * get set with buffer lock held. 1522ccb5fb9SAneesh Kumar K.V */ 1532ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 154e29d1cdeSEric Sandeen if (bh_submit_read(bh) < 0) { 155e29d1cdeSEric Sandeen put_bh(bh); 156e29d1cdeSEric Sandeen ext4_error(sb, __func__, 157e29d1cdeSEric Sandeen "Cannot read inode bitmap - " 158a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 159e29d1cdeSEric Sandeen block_group, bitmap_blk); 160e29d1cdeSEric Sandeen return NULL; 161e29d1cdeSEric Sandeen } 162ac27a0ecSDave Kleikamp return bh; 163ac27a0ecSDave Kleikamp } 164ac27a0ecSDave Kleikamp 165ac27a0ecSDave Kleikamp /* 166ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 167ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 168ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 169ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 170ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 171ac27a0ecSDave Kleikamp * has been deleted earlier. 172ac27a0ecSDave Kleikamp * 173ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 174ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 175ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 176ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 177ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 178ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 179ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 180ac27a0ecSDave Kleikamp */ 181617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 182ac27a0ecSDave Kleikamp { 183ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 184ac27a0ecSDave Kleikamp int is_directory; 185ac27a0ecSDave Kleikamp unsigned long ino; 186ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 187ac27a0ecSDave Kleikamp struct buffer_head *bh2; 188fd2d4291SAvantika Mathur ext4_group_t block_group; 189ac27a0ecSDave Kleikamp unsigned long bit; 190617ba13bSMingming Cao struct ext4_group_desc *gdp; 191617ba13bSMingming Cao struct ext4_super_block *es; 192617ba13bSMingming Cao struct ext4_sb_info *sbi; 1937ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 194ac27a0ecSDave Kleikamp 195ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 1964776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has count=%d\n", 197ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 198ac27a0ecSDave Kleikamp return; 199ac27a0ecSDave Kleikamp } 200ac27a0ecSDave Kleikamp if (inode->i_nlink) { 2014776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n", 202ac27a0ecSDave Kleikamp inode->i_nlink); 203ac27a0ecSDave Kleikamp return; 204ac27a0ecSDave Kleikamp } 205ac27a0ecSDave Kleikamp if (!sb) { 2064776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode on " 2074776004fSTheodore Ts'o "nonexistent device\n"); 208ac27a0ecSDave Kleikamp return; 209ac27a0ecSDave Kleikamp } 210617ba13bSMingming Cao sbi = EXT4_SB(sb); 211ac27a0ecSDave Kleikamp 212ac27a0ecSDave Kleikamp ino = inode->i_ino; 213617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 2149bffad1eSTheodore Ts'o trace_ext4_free_inode(inode); 215ac27a0ecSDave Kleikamp 216ac27a0ecSDave Kleikamp /* 217ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 218ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 219ac27a0ecSDave Kleikamp */ 220a269eb18SJan Kara vfs_dq_init(inode); 221617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 222a269eb18SJan Kara vfs_dq_free_inode(inode); 223a269eb18SJan Kara vfs_dq_drop(inode); 224ac27a0ecSDave Kleikamp 225ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 226ac27a0ecSDave Kleikamp 227ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 228ac27a0ecSDave Kleikamp clear_inode(inode); 229ac27a0ecSDave Kleikamp 230617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 231617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 232617ba13bSMingming Cao ext4_error(sb, "ext4_free_inode", 233ac27a0ecSDave Kleikamp "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 247ac27a0ecSDave Kleikamp /* Ok, now we can actually update the inode bitmaps.. */ 248955ce5f5SAneesh Kumar K.V cleared = ext4_clear_bit_atomic(ext4_group_lock_ptr(sb, block_group), 249955ce5f5SAneesh Kumar K.V bit, bitmap_bh->b_data); 2507ce9d5d1SEric Sandeen if (!cleared) 251617ba13bSMingming Cao ext4_error(sb, "ext4_free_inode", 252ac27a0ecSDave Kleikamp "bit already cleared for inode %lu", ino); 253ac27a0ecSDave Kleikamp else { 254617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 255ac27a0ecSDave Kleikamp 256ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 257617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 258ac27a0ecSDave Kleikamp if (fatal) goto error_return; 259ac27a0ecSDave Kleikamp 260ac27a0ecSDave Kleikamp if (gdp) { 261955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 262560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 263560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 264560671a0SAneesh Kumar K.V if (is_directory) { 265560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 266560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 2677d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 2687d39db14STheodore Ts'o ext4_group_t f; 2697d39db14STheodore Ts'o 2707d39db14STheodore Ts'o f = ext4_flex_group(sbi, block_group); 2717d39db14STheodore Ts'o atomic_dec(&sbi->s_flex_groups[f].free_inodes); 2727d39db14STheodore Ts'o } 2737d39db14STheodore Ts'o 274560671a0SAneesh Kumar K.V } 275717d50e4SAndreas Dilger gdp->bg_checksum = ext4_group_desc_csum(sbi, 276717d50e4SAndreas Dilger block_group, gdp); 277955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 278ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_freeinodes_counter); 279ac27a0ecSDave Kleikamp if (is_directory) 280ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_dirs_counter); 281ac27a0ecSDave Kleikamp 282772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2839f24e420STheodore Ts'o ext4_group_t f; 2849f24e420STheodore Ts'o 2859f24e420STheodore Ts'o f = ext4_flex_group(sbi, block_group); 2869f24e420STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].free_inodes); 287772cb7c8SJose R. Santos } 288ac27a0ecSDave Kleikamp } 2890390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 2900390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bh2); 291ac27a0ecSDave Kleikamp if (!fatal) fatal = err; 292ac27a0ecSDave Kleikamp } 2930390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 2940390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 295ac27a0ecSDave Kleikamp if (!fatal) 296ac27a0ecSDave Kleikamp fatal = err; 297ac27a0ecSDave Kleikamp sb->s_dirt = 1; 298ac27a0ecSDave Kleikamp error_return: 299ac27a0ecSDave Kleikamp brelse(bitmap_bh); 300617ba13bSMingming Cao ext4_std_error(sb, fatal); 301ac27a0ecSDave Kleikamp } 302ac27a0ecSDave Kleikamp 303ac27a0ecSDave Kleikamp /* 304ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 305ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 306ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 307ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 308ac27a0ecSDave Kleikamp * directories already is chosen. 309ac27a0ecSDave Kleikamp * 310ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory\'s block 311ac27a0ecSDave Kleikamp * group to find a free inode. 312ac27a0ecSDave Kleikamp */ 3132aa9fc4cSAvantika Mathur static int find_group_dir(struct super_block *sb, struct inode *parent, 3142aa9fc4cSAvantika Mathur ext4_group_t *best_group) 315ac27a0ecSDave Kleikamp { 3168df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 317ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 318617ba13bSMingming Cao struct ext4_group_desc *desc, *best_desc = NULL; 3192aa9fc4cSAvantika Mathur ext4_group_t group; 3202aa9fc4cSAvantika Mathur int ret = -1; 321ac27a0ecSDave Kleikamp 322617ba13bSMingming Cao freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter); 323ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 324ac27a0ecSDave Kleikamp 325ac27a0ecSDave Kleikamp for (group = 0; group < ngroups; group++) { 326ef2fb679SEric Sandeen desc = ext4_get_group_desc(sb, group, NULL); 327560671a0SAneesh Kumar K.V if (!desc || !ext4_free_inodes_count(sb, desc)) 328ac27a0ecSDave Kleikamp continue; 329560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc) < avefreei) 330ac27a0ecSDave Kleikamp continue; 331ac27a0ecSDave Kleikamp if (!best_desc || 332560671a0SAneesh Kumar K.V (ext4_free_blks_count(sb, desc) > 333560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, best_desc))) { 3342aa9fc4cSAvantika Mathur *best_group = group; 335ac27a0ecSDave Kleikamp best_desc = desc; 3362aa9fc4cSAvantika Mathur ret = 0; 337ac27a0ecSDave Kleikamp } 338ac27a0ecSDave Kleikamp } 3392aa9fc4cSAvantika Mathur return ret; 340ac27a0ecSDave Kleikamp } 341ac27a0ecSDave Kleikamp 342772cb7c8SJose R. Santos #define free_block_ratio 10 343772cb7c8SJose R. Santos 344772cb7c8SJose R. Santos static int find_group_flex(struct super_block *sb, struct inode *parent, 345772cb7c8SJose R. Santos ext4_group_t *best_group) 346772cb7c8SJose R. Santos { 347772cb7c8SJose R. Santos struct ext4_sb_info *sbi = EXT4_SB(sb); 348772cb7c8SJose R. Santos struct ext4_group_desc *desc; 349772cb7c8SJose R. Santos struct flex_groups *flex_group = sbi->s_flex_groups; 350772cb7c8SJose R. Santos ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 351772cb7c8SJose R. Santos ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group); 3528df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 353772cb7c8SJose R. Santos int flex_size = ext4_flex_bg_size(sbi); 354772cb7c8SJose R. Santos ext4_group_t best_flex = parent_fbg_group; 355772cb7c8SJose R. Santos int blocks_per_flex = sbi->s_blocks_per_group * flex_size; 356772cb7c8SJose R. Santos int flexbg_free_blocks; 357772cb7c8SJose R. Santos int flex_freeb_ratio; 358772cb7c8SJose R. Santos ext4_group_t n_fbg_groups; 359772cb7c8SJose R. Santos ext4_group_t i; 360772cb7c8SJose R. Santos 3618df9675fSTheodore Ts'o n_fbg_groups = (ngroups + flex_size - 1) >> 362772cb7c8SJose R. Santos sbi->s_log_groups_per_flex; 363772cb7c8SJose R. Santos 364772cb7c8SJose R. Santos find_close_to_parent: 3659f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[best_flex].free_blocks); 366772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 3679f24e420STheodore Ts'o if (atomic_read(&flex_group[best_flex].free_inodes) && 368772cb7c8SJose R. Santos flex_freeb_ratio > free_block_ratio) 369772cb7c8SJose R. Santos goto found_flexbg; 370772cb7c8SJose R. Santos 371772cb7c8SJose R. Santos if (best_flex && best_flex == parent_fbg_group) { 372772cb7c8SJose R. Santos best_flex--; 373772cb7c8SJose R. Santos goto find_close_to_parent; 374772cb7c8SJose R. Santos } 375772cb7c8SJose R. Santos 376772cb7c8SJose R. Santos for (i = 0; i < n_fbg_groups; i++) { 377772cb7c8SJose R. Santos if (i == parent_fbg_group || i == parent_fbg_group - 1) 378772cb7c8SJose R. Santos continue; 379772cb7c8SJose R. Santos 3809f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[i].free_blocks); 381772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 382772cb7c8SJose R. Santos 383772cb7c8SJose R. Santos if (flex_freeb_ratio > free_block_ratio && 3849f24e420STheodore Ts'o (atomic_read(&flex_group[i].free_inodes))) { 385772cb7c8SJose R. Santos best_flex = i; 386772cb7c8SJose R. Santos goto found_flexbg; 387772cb7c8SJose R. Santos } 388772cb7c8SJose R. Santos 3899f24e420STheodore Ts'o if ((atomic_read(&flex_group[best_flex].free_inodes) == 0) || 3909f24e420STheodore Ts'o ((atomic_read(&flex_group[i].free_blocks) > 3919f24e420STheodore Ts'o atomic_read(&flex_group[best_flex].free_blocks)) && 3929f24e420STheodore Ts'o atomic_read(&flex_group[i].free_inodes))) 393772cb7c8SJose R. Santos best_flex = i; 394772cb7c8SJose R. Santos } 395772cb7c8SJose R. Santos 3969f24e420STheodore Ts'o if (!atomic_read(&flex_group[best_flex].free_inodes) || 3979f24e420STheodore Ts'o !atomic_read(&flex_group[best_flex].free_blocks)) 398772cb7c8SJose R. Santos return -1; 399772cb7c8SJose R. Santos 400772cb7c8SJose R. Santos found_flexbg: 401772cb7c8SJose R. Santos for (i = best_flex * flex_size; i < ngroups && 402772cb7c8SJose R. Santos i < (best_flex + 1) * flex_size; i++) { 40388b6edd1STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 404560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc)) { 405772cb7c8SJose R. Santos *best_group = i; 406772cb7c8SJose R. Santos goto out; 407772cb7c8SJose R. Santos } 408772cb7c8SJose R. Santos } 409772cb7c8SJose R. Santos 410772cb7c8SJose R. Santos return -1; 411772cb7c8SJose R. Santos out: 412772cb7c8SJose R. Santos return 0; 413772cb7c8SJose R. Santos } 414772cb7c8SJose R. Santos 415a4912123STheodore Ts'o struct orlov_stats { 416a4912123STheodore Ts'o __u32 free_inodes; 417a4912123STheodore Ts'o __u32 free_blocks; 418a4912123STheodore Ts'o __u32 used_dirs; 419a4912123STheodore Ts'o }; 420a4912123STheodore Ts'o 421a4912123STheodore Ts'o /* 422a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 423a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 424a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 425a4912123STheodore Ts'o */ 426a4912123STheodore Ts'o void get_orlov_stats(struct super_block *sb, ext4_group_t g, 427a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 428a4912123STheodore Ts'o { 429a4912123STheodore Ts'o struct ext4_group_desc *desc; 4307d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 431a4912123STheodore Ts'o 4327d39db14STheodore Ts'o if (flex_size > 1) { 4337d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 4347d39db14STheodore Ts'o stats->free_blocks = atomic_read(&flex_group[g].free_blocks); 4357d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 4367d39db14STheodore Ts'o return; 4377d39db14STheodore Ts'o } 4387d39db14STheodore Ts'o 4397d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 4407d39db14STheodore Ts'o if (desc) { 4417d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 4427d39db14STheodore Ts'o stats->free_blocks = ext4_free_blks_count(sb, desc); 4437d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 4447d39db14STheodore Ts'o } else { 445a4912123STheodore Ts'o stats->free_inodes = 0; 446a4912123STheodore Ts'o stats->free_blocks = 0; 447a4912123STheodore Ts'o stats->used_dirs = 0; 448a4912123STheodore Ts'o } 449a4912123STheodore Ts'o } 450a4912123STheodore Ts'o 451ac27a0ecSDave Kleikamp /* 452ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 453ac27a0ecSDave Kleikamp * 454ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 455ac27a0ecSDave Kleikamp * 456ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 457ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 458ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 459ac27a0ecSDave Kleikamp * 460ac27a0ecSDave Kleikamp * For the rest rules look so: 461ac27a0ecSDave Kleikamp * 462ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 463ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 464ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 465ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4661cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 467ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 468ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 469ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 470ac27a0ecSDave Kleikamp */ 471ac27a0ecSDave Kleikamp 4722aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 473a4912123STheodore Ts'o ext4_group_t *group, int mode) 474ac27a0ecSDave Kleikamp { 475fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 476617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 4778df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 478617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 479ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 480617ba13bSMingming Cao ext4_fsblk_t freeb, avefreeb; 481ac27a0ecSDave Kleikamp unsigned int ndirs; 482a4912123STheodore Ts'o int max_dirs, min_inodes; 483617ba13bSMingming Cao ext4_grpblk_t min_blocks; 4848df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 485617ba13bSMingming Cao struct ext4_group_desc *desc; 486a4912123STheodore Ts'o struct orlov_stats stats; 487a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 488a4912123STheodore Ts'o 4898df9675fSTheodore Ts'o ngroups = real_ngroups; 490a4912123STheodore Ts'o if (flex_size > 1) { 4918df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 492a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 493a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 494a4912123STheodore Ts'o } 495ac27a0ecSDave Kleikamp 496ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 497ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 498ac27a0ecSDave Kleikamp freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter); 4993a5b2ecdSMingming Cao avefreeb = freeb; 500f4e5bc24SAndrew Morton do_div(avefreeb, ngroups); 501ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 502ac27a0ecSDave Kleikamp 503a4912123STheodore Ts'o if (S_ISDIR(mode) && 504a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 505a4912123STheodore Ts'o (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL))) { 506ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 5072aa9fc4cSAvantika Mathur int ret = -1; 508ac27a0ecSDave Kleikamp 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; 652a4912123STheodore Ts'o return find_group_orlov(sb, parent, group, mode); 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); 71939341867SAneesh Kumar K.V struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL); 72039341867SAneesh Kumar K.V 721955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 72239341867SAneesh Kumar K.V if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) { 72339341867SAneesh Kumar K.V /* not a free inode */ 72439341867SAneesh Kumar K.V retval = 1; 72539341867SAneesh Kumar K.V goto err_ret; 72639341867SAneesh Kumar K.V } 72739341867SAneesh Kumar K.V ino++; 72839341867SAneesh Kumar K.V if ((group == 0 && ino < EXT4_FIRST_INO(sb)) || 72939341867SAneesh Kumar K.V ino > EXT4_INODES_PER_GROUP(sb)) { 730955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 73139341867SAneesh Kumar K.V ext4_error(sb, __func__, 73239341867SAneesh Kumar K.V "reserved inode or inode > inodes count - " 73339341867SAneesh Kumar K.V "block_group = %u, inode=%lu", group, 73439341867SAneesh Kumar K.V ino + group * EXT4_INODES_PER_GROUP(sb)); 73539341867SAneesh Kumar K.V return 1; 73639341867SAneesh Kumar K.V } 73739341867SAneesh Kumar K.V /* If we didn't allocate from within the initialized part of the inode 73839341867SAneesh Kumar K.V * table then we need to initialize up to this inode. */ 73939341867SAneesh Kumar K.V if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) { 74039341867SAneesh Kumar K.V 74139341867SAneesh Kumar K.V if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 74239341867SAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 74339341867SAneesh Kumar K.V /* When marking the block group with 74439341867SAneesh Kumar K.V * ~EXT4_BG_INODE_UNINIT we don't want to depend 74539341867SAneesh Kumar K.V * on the value of bg_itable_unused even though 74639341867SAneesh Kumar K.V * mke2fs could have initialized the same for us. 74739341867SAneesh Kumar K.V * Instead we calculated the value below 74839341867SAneesh Kumar K.V */ 74939341867SAneesh Kumar K.V 75039341867SAneesh Kumar K.V free = 0; 75139341867SAneesh Kumar K.V } else { 75239341867SAneesh Kumar K.V free = EXT4_INODES_PER_GROUP(sb) - 75339341867SAneesh Kumar K.V ext4_itable_unused_count(sb, gdp); 75439341867SAneesh Kumar K.V } 75539341867SAneesh Kumar K.V 75639341867SAneesh Kumar K.V /* 75739341867SAneesh Kumar K.V * Check the relative inode number against the last used 75839341867SAneesh Kumar K.V * relative inode number in this group. if it is greater 75939341867SAneesh Kumar K.V * we need to update the bg_itable_unused count 76039341867SAneesh Kumar K.V * 76139341867SAneesh Kumar K.V */ 76239341867SAneesh Kumar K.V if (ino > free) 76339341867SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 76439341867SAneesh Kumar K.V (EXT4_INODES_PER_GROUP(sb) - ino)); 76539341867SAneesh Kumar K.V } 76639341867SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) - 1; 76739341867SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 76839341867SAneesh Kumar K.V if (S_ISDIR(mode)) { 76939341867SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) + 1; 77039341867SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 7717d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 7727d39db14STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 7737d39db14STheodore Ts'o 7747d39db14STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].free_inodes); 7757d39db14STheodore Ts'o } 77639341867SAneesh Kumar K.V } 77739341867SAneesh Kumar K.V gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 77839341867SAneesh Kumar K.V err_ret: 779955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 78039341867SAneesh Kumar K.V return retval; 78139341867SAneesh Kumar K.V } 78239341867SAneesh Kumar K.V 78339341867SAneesh Kumar K.V /* 784ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 785ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 786ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 787ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 788ac27a0ecSDave Kleikamp * directories already is chosen. 789ac27a0ecSDave Kleikamp * 790ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 791ac27a0ecSDave Kleikamp * group to find a free inode. 792ac27a0ecSDave Kleikamp */ 793617ba13bSMingming Cao struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode) 794ac27a0ecSDave Kleikamp { 795ac27a0ecSDave Kleikamp struct super_block *sb; 7963300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 7973300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 7988df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 799ac27a0ecSDave Kleikamp unsigned long ino = 0; 800ac27a0ecSDave Kleikamp struct inode *inode; 801617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 802617ba13bSMingming Cao struct ext4_inode_info *ei; 803617ba13bSMingming Cao struct ext4_sb_info *sbi; 80439341867SAneesh Kumar K.V int ret2, err = 0; 805ac27a0ecSDave Kleikamp struct inode *ret; 8062aa9fc4cSAvantika Mathur ext4_group_t i; 8072aa9fc4cSAvantika Mathur int free = 0; 8082842c3b5STheodore Ts'o static int once = 1; 809772cb7c8SJose R. Santos ext4_group_t flex_group; 810ac27a0ecSDave Kleikamp 811ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 812ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 813ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 814ac27a0ecSDave Kleikamp 815ac27a0ecSDave Kleikamp sb = dir->i_sb; 8168df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 8179bffad1eSTheodore Ts'o trace_ext4_request_inode(dir, mode); 818ac27a0ecSDave Kleikamp inode = new_inode(sb); 819ac27a0ecSDave Kleikamp if (!inode) 820ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 821617ba13bSMingming Cao ei = EXT4_I(inode); 822617ba13bSMingming Cao sbi = EXT4_SB(sb); 823772cb7c8SJose R. Santos 824a4912123STheodore Ts'o if (sbi->s_log_groups_per_flex && test_opt(sb, OLDALLOC)) { 825772cb7c8SJose R. Santos ret2 = find_group_flex(sb, dir, &group); 82605bf9e83STheodore Ts'o if (ret2 == -1) { 827a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 8286b82f3cbSChuck Ebbert if (ret2 == 0 && once) { 8292842c3b5STheodore Ts'o once = 0; 83005bf9e83STheodore Ts'o printk(KERN_NOTICE "ext4: find_group_flex " 83105bf9e83STheodore Ts'o "failed, fallback succeeded dir %lu\n", 83205bf9e83STheodore Ts'o dir->i_ino); 83305bf9e83STheodore Ts'o } 8346b82f3cbSChuck Ebbert } 835772cb7c8SJose R. Santos goto got_group; 836772cb7c8SJose R. Santos } 837772cb7c8SJose R. Santos 838ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) { 839ac27a0ecSDave Kleikamp if (test_opt(sb, OLDALLOC)) 8402aa9fc4cSAvantika Mathur ret2 = find_group_dir(sb, dir, &group); 841ac27a0ecSDave Kleikamp else 842a4912123STheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode); 843ac27a0ecSDave Kleikamp } else 844a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 845ac27a0ecSDave Kleikamp 846772cb7c8SJose R. Santos got_group: 847a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 848ac27a0ecSDave Kleikamp err = -ENOSPC; 8492aa9fc4cSAvantika Mathur if (ret2 == -1) 850ac27a0ecSDave Kleikamp goto out; 851ac27a0ecSDave Kleikamp 8528df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 853ac27a0ecSDave Kleikamp err = -EIO; 854ac27a0ecSDave Kleikamp 8553300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 856ac27a0ecSDave Kleikamp if (!gdp) 857ac27a0ecSDave Kleikamp goto fail; 858ac27a0ecSDave Kleikamp 8593300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 8603300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 8613300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 862ac27a0ecSDave Kleikamp goto fail; 863ac27a0ecSDave Kleikamp 864ac27a0ecSDave Kleikamp ino = 0; 865ac27a0ecSDave Kleikamp 866ac27a0ecSDave Kleikamp repeat_in_this_group: 867617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 8683300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 8693300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 8703300bedaSAneesh Kumar K.V 871617ba13bSMingming Cao if (ino < EXT4_INODES_PER_GROUP(sb)) { 872ac27a0ecSDave Kleikamp 8733300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 8743300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 8753300bedaSAneesh Kumar K.V inode_bitmap_bh); 876ac27a0ecSDave Kleikamp if (err) 877ac27a0ecSDave Kleikamp goto fail; 878ac27a0ecSDave Kleikamp 87939341867SAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "get_write_access"); 88039341867SAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 88139341867SAneesh Kumar K.V group_desc_bh); 88239341867SAneesh Kumar K.V if (err) 88339341867SAneesh Kumar K.V goto fail; 88439341867SAneesh Kumar K.V if (!ext4_claim_inode(sb, inode_bitmap_bh, 88539341867SAneesh Kumar K.V ino, group, mode)) { 886ac27a0ecSDave Kleikamp /* we won it */ 8873300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, 8880390131bSFrank Mayhar "call ext4_handle_dirty_metadata"); 8890390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 8900390131bSFrank Mayhar inode, 8913300bedaSAneesh Kumar K.V inode_bitmap_bh); 892ac27a0ecSDave Kleikamp if (err) 893ac27a0ecSDave Kleikamp goto fail; 89439341867SAneesh Kumar K.V /* zero bit is inode number 1*/ 89539341867SAneesh Kumar K.V ino++; 896ac27a0ecSDave Kleikamp goto got; 897ac27a0ecSDave Kleikamp } 898ac27a0ecSDave Kleikamp /* we lost it */ 8993300bedaSAneesh Kumar K.V ext4_handle_release_buffer(handle, inode_bitmap_bh); 90039341867SAneesh Kumar K.V ext4_handle_release_buffer(handle, group_desc_bh); 901ac27a0ecSDave Kleikamp 902617ba13bSMingming Cao if (++ino < EXT4_INODES_PER_GROUP(sb)) 903ac27a0ecSDave Kleikamp goto repeat_in_this_group; 904ac27a0ecSDave Kleikamp } 905ac27a0ecSDave Kleikamp 906ac27a0ecSDave Kleikamp /* 907ac27a0ecSDave Kleikamp * This case is possible in concurrent environment. It is very 908ac27a0ecSDave Kleikamp * rare. We cannot repeat the find_group_xxx() call because 909ac27a0ecSDave Kleikamp * that will simply return the same blockgroup, because the 910ac27a0ecSDave Kleikamp * group descriptor metadata has not yet been updated. 911ac27a0ecSDave Kleikamp * So we just go onto the next blockgroup. 912ac27a0ecSDave Kleikamp */ 9138df9675fSTheodore Ts'o if (++group == ngroups) 914ac27a0ecSDave Kleikamp group = 0; 915ac27a0ecSDave Kleikamp } 916ac27a0ecSDave Kleikamp err = -ENOSPC; 917ac27a0ecSDave Kleikamp goto out; 918ac27a0ecSDave Kleikamp 919ac27a0ecSDave Kleikamp got: 920717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 921717d50e4SAndreas Dilger if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) && 922717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9233300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 924717d50e4SAndreas Dilger 9253300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 9263300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 9273300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 928717d50e4SAndreas Dilger if (err) { 9293300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 930717d50e4SAndreas Dilger goto fail; 931717d50e4SAndreas Dilger } 932717d50e4SAndreas Dilger 933717d50e4SAndreas Dilger free = 0; 934955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 935717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 936717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 937717d50e4SAndreas Dilger free = ext4_free_blocks_after_init(sb, group, gdp); 9383300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 939560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, free); 94023712a9cSFrederic Bohe gdp->bg_checksum = ext4_group_desc_csum(sbi, group, 94123712a9cSFrederic Bohe gdp); 942717d50e4SAndreas Dilger } 943955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 944717d50e4SAndreas Dilger 945717d50e4SAndreas Dilger /* Don't need to dirty bitmap block if we didn't change it */ 946717d50e4SAndreas Dilger if (free) { 9473300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 9480390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 9493300bedaSAneesh Kumar K.V NULL, block_bitmap_bh); 950717d50e4SAndreas Dilger } 951717d50e4SAndreas Dilger 9523300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 953717d50e4SAndreas Dilger if (err) 954717d50e4SAndreas Dilger goto fail; 955717d50e4SAndreas Dilger } 9563300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 9573300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 95839341867SAneesh Kumar K.V if (err) 95939341867SAneesh Kumar K.V goto fail; 960ac27a0ecSDave Kleikamp 961ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 962ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 963ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 964ac27a0ecSDave Kleikamp sb->s_dirt = 1; 965ac27a0ecSDave Kleikamp 966772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 967772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 9689f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 969772cb7c8SJose R. Santos } 970772cb7c8SJose R. Santos 9714c9c544eSDavid Howells inode->i_uid = current_fsuid(); 972ac27a0ecSDave Kleikamp if (test_opt(sb, GRPID)) 973ac27a0ecSDave Kleikamp inode->i_gid = dir->i_gid; 974ac27a0ecSDave Kleikamp else if (dir->i_mode & S_ISGID) { 975ac27a0ecSDave Kleikamp inode->i_gid = dir->i_gid; 976ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 977ac27a0ecSDave Kleikamp mode |= S_ISGID; 978ac27a0ecSDave Kleikamp } else 9794c9c544eSDavid Howells inode->i_gid = current_fsgid(); 980ac27a0ecSDave Kleikamp inode->i_mode = mode; 981ac27a0ecSDave Kleikamp 982717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 983ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 984ac27a0ecSDave Kleikamp inode->i_blocks = 0; 985ef7f3835SKalpak Shah inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime = 986ef7f3835SKalpak Shah ext4_current_time(inode); 987ac27a0ecSDave Kleikamp 988ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 989ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 990ac27a0ecSDave Kleikamp ei->i_disksize = 0; 991ac27a0ecSDave Kleikamp 99242bf0383SAneesh Kumar K.V /* 9932dc6b0d4SDuane Griffin * Don't inherit extent flag from directory, amongst others. We set 9942dc6b0d4SDuane Griffin * extent flag on newly created directory and file only if -o extent 9952dc6b0d4SDuane Griffin * mount option is specified 99642bf0383SAneesh Kumar K.V */ 9972dc6b0d4SDuane Griffin ei->i_flags = 9982dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 999ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 1000ac27a0ecSDave Kleikamp ei->i_dtime = 0; 1001ac27a0ecSDave Kleikamp ei->i_block_group = group; 1002a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 1003ac27a0ecSDave Kleikamp 1004617ba13bSMingming Cao ext4_set_inode_flags(inode); 1005ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 10060390131bSFrank Mayhar ext4_handle_sync(handle); 10076b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 10086b38e842SAl Viro err = -EINVAL; 10096b38e842SAl Viro goto fail_drop; 10106b38e842SAl Viro } 1011ac27a0ecSDave Kleikamp spin_lock(&sbi->s_next_gen_lock); 1012ac27a0ecSDave Kleikamp inode->i_generation = sbi->s_next_generation++; 1013ac27a0ecSDave Kleikamp spin_unlock(&sbi->s_next_gen_lock); 1014ac27a0ecSDave Kleikamp 1015617ba13bSMingming Cao ei->i_state = EXT4_STATE_NEW; 1016ef7f3835SKalpak Shah 1017ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 1018ac27a0ecSDave Kleikamp 1019ac27a0ecSDave Kleikamp ret = inode; 1020a269eb18SJan Kara if (vfs_dq_alloc_inode(inode)) { 1021ac27a0ecSDave Kleikamp err = -EDQUOT; 1022ac27a0ecSDave Kleikamp goto fail_drop; 1023ac27a0ecSDave Kleikamp } 1024ac27a0ecSDave Kleikamp 1025617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1026ac27a0ecSDave Kleikamp if (err) 1027ac27a0ecSDave Kleikamp goto fail_free_drop; 1028ac27a0ecSDave Kleikamp 1029617ba13bSMingming Cao err = ext4_init_security(handle, inode, dir); 1030ac27a0ecSDave Kleikamp if (err) 1031ac27a0ecSDave Kleikamp goto fail_free_drop; 1032ac27a0ecSDave Kleikamp 103383982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 1034e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1035e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 1036a86c6181SAlex Tomas EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL; 1037a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1038a86c6181SAlex Tomas } 103942bf0383SAneesh Kumar K.V } 1040ac27a0ecSDave Kleikamp 10418753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 10428753e88fSAneesh Kumar K.V if (err) { 10438753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 10448753e88fSAneesh Kumar K.V goto fail_free_drop; 10458753e88fSAneesh Kumar K.V } 10468753e88fSAneesh Kumar K.V 1047617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 10489bffad1eSTheodore Ts'o trace_ext4_allocate_inode(inode, dir, mode); 1049ac27a0ecSDave Kleikamp goto really_out; 1050ac27a0ecSDave Kleikamp fail: 1051617ba13bSMingming Cao ext4_std_error(sb, err); 1052ac27a0ecSDave Kleikamp out: 1053ac27a0ecSDave Kleikamp iput(inode); 1054ac27a0ecSDave Kleikamp ret = ERR_PTR(err); 1055ac27a0ecSDave Kleikamp really_out: 10563300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1057ac27a0ecSDave Kleikamp return ret; 1058ac27a0ecSDave Kleikamp 1059ac27a0ecSDave Kleikamp fail_free_drop: 1060a269eb18SJan Kara vfs_dq_free_inode(inode); 1061ac27a0ecSDave Kleikamp 1062ac27a0ecSDave Kleikamp fail_drop: 1063a269eb18SJan Kara vfs_dq_drop(inode); 1064ac27a0ecSDave Kleikamp inode->i_flags |= S_NOQUOTA; 1065ac27a0ecSDave Kleikamp inode->i_nlink = 0; 10666b38e842SAl Viro unlock_new_inode(inode); 1067ac27a0ecSDave Kleikamp iput(inode); 10683300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1069ac27a0ecSDave Kleikamp return ERR_PTR(err); 1070ac27a0ecSDave Kleikamp } 1071ac27a0ecSDave Kleikamp 1072ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1073617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1074ac27a0ecSDave Kleikamp { 1075617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1076fd2d4291SAvantika Mathur ext4_group_t block_group; 1077ac27a0ecSDave Kleikamp int bit; 10781d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 1079ac27a0ecSDave Kleikamp struct inode *inode = NULL; 10801d1fe1eeSDavid Howells long err = -EIO; 1081ac27a0ecSDave Kleikamp 1082ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1083ac27a0ecSDave Kleikamp if (ino > max_ino) { 108446e665e9SHarvey Harrison ext4_warning(sb, __func__, 1085ac27a0ecSDave Kleikamp "bad orphan ino %lu! e2fsck was run?", ino); 10861d1fe1eeSDavid Howells goto error; 1087ac27a0ecSDave Kleikamp } 1088ac27a0ecSDave Kleikamp 1089617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1090617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1091e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1092ac27a0ecSDave Kleikamp if (!bitmap_bh) { 109346e665e9SHarvey Harrison ext4_warning(sb, __func__, 1094ac27a0ecSDave Kleikamp "inode bitmap error for orphan %lu", ino); 10951d1fe1eeSDavid Howells goto error; 1096ac27a0ecSDave Kleikamp } 1097ac27a0ecSDave Kleikamp 1098ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1099ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1100ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1101ac27a0ecSDave Kleikamp */ 11021d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 11031d1fe1eeSDavid Howells goto bad_orphan; 11041d1fe1eeSDavid Howells 11051d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 11061d1fe1eeSDavid Howells if (IS_ERR(inode)) 11071d1fe1eeSDavid Howells goto iget_failed; 11081d1fe1eeSDavid Howells 110991ef4cafSDuane Griffin /* 111091ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 111191ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 111291ef4cafSDuane Griffin * during processing and an infinite loop will result. 111391ef4cafSDuane Griffin */ 111491ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 111591ef4cafSDuane Griffin goto bad_orphan; 111691ef4cafSDuane Griffin 11171d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 11181d1fe1eeSDavid Howells goto bad_orphan; 11191d1fe1eeSDavid Howells brelse(bitmap_bh); 11201d1fe1eeSDavid Howells return inode; 11211d1fe1eeSDavid Howells 11221d1fe1eeSDavid Howells iget_failed: 11231d1fe1eeSDavid Howells err = PTR_ERR(inode); 11241d1fe1eeSDavid Howells inode = NULL; 11251d1fe1eeSDavid Howells bad_orphan: 112646e665e9SHarvey Harrison ext4_warning(sb, __func__, 1127ac27a0ecSDave Kleikamp "bad orphan inode %lu! e2fsck was run?", ino); 1128617ba13bSMingming Cao printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1129ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1130617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1131ac27a0ecSDave Kleikamp printk(KERN_NOTICE "inode=%p\n", inode); 1132ac27a0ecSDave Kleikamp if (inode) { 1133ac27a0ecSDave Kleikamp printk(KERN_NOTICE "is_bad_inode(inode)=%d\n", 1134ac27a0ecSDave Kleikamp is_bad_inode(inode)); 1135ac27a0ecSDave Kleikamp printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n", 1136ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 1137ac27a0ecSDave Kleikamp printk(KERN_NOTICE "max_ino=%lu\n", max_ino); 113891ef4cafSDuane Griffin printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink); 1139ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 11401d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1141ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1142ac27a0ecSDave Kleikamp iput(inode); 1143ac27a0ecSDave Kleikamp } 1144ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11451d1fe1eeSDavid Howells error: 11461d1fe1eeSDavid Howells return ERR_PTR(err); 1147ac27a0ecSDave Kleikamp } 1148ac27a0ecSDave Kleikamp 1149617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1150ac27a0ecSDave Kleikamp { 1151ac27a0ecSDave Kleikamp unsigned long desc_count; 1152617ba13bSMingming Cao struct ext4_group_desc *gdp; 11538df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1154617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1155617ba13bSMingming Cao struct ext4_super_block *es; 1156ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1157ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1158ac27a0ecSDave Kleikamp 1159617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1160ac27a0ecSDave Kleikamp desc_count = 0; 1161ac27a0ecSDave Kleikamp bitmap_count = 0; 1162ac27a0ecSDave Kleikamp gdp = NULL; 11638df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1164617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1165ac27a0ecSDave Kleikamp if (!gdp) 1166ac27a0ecSDave Kleikamp continue; 1167560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1168ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1169e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1170ac27a0ecSDave Kleikamp if (!bitmap_bh) 1171ac27a0ecSDave Kleikamp continue; 1172ac27a0ecSDave Kleikamp 1173617ba13bSMingming Cao x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8); 1174c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1175560671a0SAneesh Kumar K.V i, ext4_free_inodes_count(sb, gdp), x); 1176ac27a0ecSDave Kleikamp bitmap_count += x; 1177ac27a0ecSDave Kleikamp } 1178ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11794776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 11804776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1181ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1182ac27a0ecSDave Kleikamp return desc_count; 1183ac27a0ecSDave Kleikamp #else 1184ac27a0ecSDave Kleikamp desc_count = 0; 11858df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1186617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1187ac27a0ecSDave Kleikamp if (!gdp) 1188ac27a0ecSDave Kleikamp continue; 1189560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1190ac27a0ecSDave Kleikamp cond_resched(); 1191ac27a0ecSDave Kleikamp } 1192ac27a0ecSDave Kleikamp return desc_count; 1193ac27a0ecSDave Kleikamp #endif 1194ac27a0ecSDave Kleikamp } 1195ac27a0ecSDave Kleikamp 1196ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1197617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1198ac27a0ecSDave Kleikamp { 1199ac27a0ecSDave Kleikamp unsigned long count = 0; 12008df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1201ac27a0ecSDave Kleikamp 12028df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1203617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1204ac27a0ecSDave Kleikamp if (!gdp) 1205ac27a0ecSDave Kleikamp continue; 1206560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1207ac27a0ecSDave Kleikamp } 1208ac27a0ecSDave Kleikamp return count; 1209ac27a0ecSDave Kleikamp } 1210