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> 263dcf5451SChristoph Hellwig #include "ext4.h" 273dcf5451SChristoph Hellwig #include "ext4_jbd2.h" 28ac27a0ecSDave Kleikamp #include "xattr.h" 29ac27a0ecSDave Kleikamp #include "acl.h" 30ac27a0ecSDave Kleikamp 31ac27a0ecSDave Kleikamp /* 32ac27a0ecSDave Kleikamp * ialloc.c contains the inodes allocation and deallocation routines 33ac27a0ecSDave Kleikamp */ 34ac27a0ecSDave Kleikamp 35ac27a0ecSDave Kleikamp /* 36ac27a0ecSDave Kleikamp * The free inodes are managed by bitmaps. A file system contains several 37ac27a0ecSDave Kleikamp * blocks groups. Each group contains 1 bitmap block for blocks, 1 bitmap 38ac27a0ecSDave Kleikamp * block for inodes, N blocks for the inode table and data blocks. 39ac27a0ecSDave Kleikamp * 40ac27a0ecSDave Kleikamp * The file system contains group descriptors which are located after the 41ac27a0ecSDave Kleikamp * super block. Each descriptor contains the number of the bitmap block and 42ac27a0ecSDave Kleikamp * the free blocks count in the block. 43ac27a0ecSDave Kleikamp */ 44ac27a0ecSDave Kleikamp 45717d50e4SAndreas Dilger /* 46717d50e4SAndreas Dilger * To avoid calling the atomic setbit hundreds or thousands of times, we only 47717d50e4SAndreas Dilger * need to use it within a single byte (to ensure we get endianness right). 48717d50e4SAndreas Dilger * We can use memset for the rest of the bitmap as there are no other users. 49717d50e4SAndreas Dilger */ 50717d50e4SAndreas Dilger void mark_bitmap_end(int start_bit, int end_bit, char *bitmap) 51717d50e4SAndreas Dilger { 52717d50e4SAndreas Dilger int i; 53717d50e4SAndreas Dilger 54717d50e4SAndreas Dilger if (start_bit >= end_bit) 55717d50e4SAndreas Dilger return; 56717d50e4SAndreas Dilger 57717d50e4SAndreas Dilger ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit); 58717d50e4SAndreas Dilger for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++) 59717d50e4SAndreas Dilger ext4_set_bit(i, bitmap); 60717d50e4SAndreas Dilger if (i < end_bit) 61717d50e4SAndreas Dilger memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3); 62717d50e4SAndreas Dilger } 63717d50e4SAndreas Dilger 64717d50e4SAndreas Dilger /* Initializes an uninitialized inode bitmap */ 65fd2d4291SAvantika Mathur unsigned ext4_init_inode_bitmap(struct super_block *sb, struct buffer_head *bh, 66fd2d4291SAvantika Mathur ext4_group_t block_group, 67717d50e4SAndreas Dilger struct ext4_group_desc *gdp) 68717d50e4SAndreas Dilger { 69717d50e4SAndreas Dilger struct ext4_sb_info *sbi = EXT4_SB(sb); 70717d50e4SAndreas Dilger 71717d50e4SAndreas Dilger J_ASSERT_BH(bh, buffer_locked(bh)); 72717d50e4SAndreas Dilger 73717d50e4SAndreas Dilger /* If checksum is bad mark all blocks and inodes use to prevent 74717d50e4SAndreas Dilger * allocation, essentially implementing a per-group read-only flag. */ 75717d50e4SAndreas Dilger if (!ext4_group_desc_csum_verify(sbi, block_group, gdp)) { 76a9df9a49STheodore Ts'o ext4_error(sb, __func__, "Checksum bad for group %u", 77717d50e4SAndreas Dilger block_group); 78560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, 0); 79560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, 0); 80560671a0SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 0); 81717d50e4SAndreas Dilger memset(bh->b_data, 0xff, sb->s_blocksize); 82717d50e4SAndreas Dilger return 0; 83717d50e4SAndreas Dilger } 84717d50e4SAndreas Dilger 85717d50e4SAndreas Dilger memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8); 86648f5879SAneesh Kumar K.V mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), sb->s_blocksize * 8, 87717d50e4SAndreas Dilger bh->b_data); 88717d50e4SAndreas Dilger 89717d50e4SAndreas Dilger return EXT4_INODES_PER_GROUP(sb); 90717d50e4SAndreas Dilger } 91ac27a0ecSDave Kleikamp 92ac27a0ecSDave Kleikamp /* 93ac27a0ecSDave Kleikamp * Read the inode allocation bitmap for a given block_group, reading 94ac27a0ecSDave Kleikamp * into the specified slot in the superblock's bitmap cache. 95ac27a0ecSDave Kleikamp * 96ac27a0ecSDave Kleikamp * Return buffer_head of bitmap on success or NULL. 97ac27a0ecSDave Kleikamp */ 98ac27a0ecSDave Kleikamp static struct buffer_head * 99e29d1cdeSEric Sandeen ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group) 100ac27a0ecSDave Kleikamp { 101617ba13bSMingming Cao struct ext4_group_desc *desc; 102ac27a0ecSDave Kleikamp struct buffer_head *bh = NULL; 103e29d1cdeSEric Sandeen ext4_fsblk_t bitmap_blk; 104ac27a0ecSDave Kleikamp 105617ba13bSMingming Cao desc = ext4_get_group_desc(sb, block_group, NULL); 106ac27a0ecSDave Kleikamp if (!desc) 107e29d1cdeSEric Sandeen return NULL; 108e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 109e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 110e29d1cdeSEric Sandeen if (unlikely(!bh)) { 111e29d1cdeSEric Sandeen ext4_error(sb, __func__, 112ac27a0ecSDave Kleikamp "Cannot read inode bitmap - " 113a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 114e29d1cdeSEric Sandeen block_group, bitmap_blk); 115e29d1cdeSEric Sandeen return NULL; 116e29d1cdeSEric Sandeen } 1172ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) 118e29d1cdeSEric Sandeen return bh; 119e29d1cdeSEric Sandeen 120c806e68fSFrederic Bohe lock_buffer(bh); 1212ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) { 1222ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1232ccb5fb9SAneesh Kumar K.V return bh; 1242ccb5fb9SAneesh Kumar K.V } 125955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 126e29d1cdeSEric Sandeen if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 127e29d1cdeSEric Sandeen ext4_init_inode_bitmap(sb, bh, block_group, desc); 1282ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 129e29d1cdeSEric Sandeen set_buffer_uptodate(bh); 130955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1313300bedaSAneesh Kumar K.V unlock_buffer(bh); 132e29d1cdeSEric Sandeen return bh; 133e29d1cdeSEric Sandeen } 134955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1352ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh)) { 1362ccb5fb9SAneesh Kumar K.V /* 1372ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 1382ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 1392ccb5fb9SAneesh Kumar K.V */ 1402ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 1412ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1422ccb5fb9SAneesh Kumar K.V return bh; 1432ccb5fb9SAneesh Kumar K.V } 1442ccb5fb9SAneesh Kumar K.V /* 1452ccb5fb9SAneesh Kumar K.V * submit the buffer_head for read. We can 1462ccb5fb9SAneesh Kumar K.V * safely mark the bitmap as uptodate now. 1472ccb5fb9SAneesh Kumar K.V * We do it here so the bitmap uptodate bit 1482ccb5fb9SAneesh Kumar K.V * get set with buffer lock held. 1492ccb5fb9SAneesh Kumar K.V */ 1502ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 151e29d1cdeSEric Sandeen if (bh_submit_read(bh) < 0) { 152e29d1cdeSEric Sandeen put_bh(bh); 153e29d1cdeSEric Sandeen ext4_error(sb, __func__, 154e29d1cdeSEric Sandeen "Cannot read inode bitmap - " 155a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 156e29d1cdeSEric Sandeen block_group, bitmap_blk); 157e29d1cdeSEric Sandeen return NULL; 158e29d1cdeSEric Sandeen } 159ac27a0ecSDave Kleikamp return bh; 160ac27a0ecSDave Kleikamp } 161ac27a0ecSDave Kleikamp 162ac27a0ecSDave Kleikamp /* 163ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 164ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 165ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 166ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 167ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 168ac27a0ecSDave Kleikamp * has been deleted earlier. 169ac27a0ecSDave Kleikamp * 170ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 171ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 172ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 173ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 174ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 175ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 176ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 177ac27a0ecSDave Kleikamp */ 178617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 179ac27a0ecSDave Kleikamp { 180ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 181ac27a0ecSDave Kleikamp int is_directory; 182ac27a0ecSDave Kleikamp unsigned long ino; 183ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 184ac27a0ecSDave Kleikamp struct buffer_head *bh2; 185fd2d4291SAvantika Mathur ext4_group_t block_group; 186ac27a0ecSDave Kleikamp unsigned long bit; 187617ba13bSMingming Cao struct ext4_group_desc *gdp; 188617ba13bSMingming Cao struct ext4_super_block *es; 189617ba13bSMingming Cao struct ext4_sb_info *sbi; 1907ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 191ac27a0ecSDave Kleikamp 192ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 1934776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has count=%d\n", 194ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 195ac27a0ecSDave Kleikamp return; 196ac27a0ecSDave Kleikamp } 197ac27a0ecSDave Kleikamp if (inode->i_nlink) { 1984776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n", 199ac27a0ecSDave Kleikamp inode->i_nlink); 200ac27a0ecSDave Kleikamp return; 201ac27a0ecSDave Kleikamp } 202ac27a0ecSDave Kleikamp if (!sb) { 2034776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode on " 2044776004fSTheodore Ts'o "nonexistent device\n"); 205ac27a0ecSDave Kleikamp return; 206ac27a0ecSDave Kleikamp } 207617ba13bSMingming Cao sbi = EXT4_SB(sb); 208ac27a0ecSDave Kleikamp 209ac27a0ecSDave Kleikamp ino = inode->i_ino; 210617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 211ba80b101STheodore Ts'o trace_mark(ext4_free_inode, 212ba80b101STheodore Ts'o "dev %s ino %lu mode %d uid %lu gid %lu bocks %llu", 213ba80b101STheodore Ts'o sb->s_id, inode->i_ino, inode->i_mode, 214ba80b101STheodore Ts'o (unsigned long) inode->i_uid, (unsigned long) inode->i_gid, 215ba80b101STheodore Ts'o (unsigned long long) inode->i_blocks); 216ac27a0ecSDave Kleikamp 217ac27a0ecSDave Kleikamp /* 218ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 219ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 220ac27a0ecSDave Kleikamp */ 221a269eb18SJan Kara vfs_dq_init(inode); 222617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 223a269eb18SJan Kara vfs_dq_free_inode(inode); 224a269eb18SJan Kara vfs_dq_drop(inode); 225ac27a0ecSDave Kleikamp 226ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 227ac27a0ecSDave Kleikamp 228ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 229ac27a0ecSDave Kleikamp clear_inode(inode); 230ac27a0ecSDave Kleikamp 231617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 232617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 233617ba13bSMingming Cao ext4_error(sb, "ext4_free_inode", 234ac27a0ecSDave Kleikamp "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 248ac27a0ecSDave Kleikamp /* Ok, now we can actually update the inode bitmaps.. */ 249955ce5f5SAneesh Kumar K.V cleared = ext4_clear_bit_atomic(ext4_group_lock_ptr(sb, block_group), 250955ce5f5SAneesh Kumar K.V bit, bitmap_bh->b_data); 2517ce9d5d1SEric Sandeen if (!cleared) 252617ba13bSMingming Cao ext4_error(sb, "ext4_free_inode", 253ac27a0ecSDave Kleikamp "bit already cleared for inode %lu", ino); 254ac27a0ecSDave Kleikamp else { 255617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 256ac27a0ecSDave Kleikamp 257ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 258617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 259ac27a0ecSDave Kleikamp if (fatal) goto error_return; 260ac27a0ecSDave Kleikamp 261ac27a0ecSDave Kleikamp if (gdp) { 262955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 263560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 264560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 265560671a0SAneesh Kumar K.V if (is_directory) { 266560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 267560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 2687d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 2697d39db14STheodore Ts'o ext4_group_t f; 2707d39db14STheodore Ts'o 2717d39db14STheodore Ts'o f = ext4_flex_group(sbi, block_group); 2727d39db14STheodore Ts'o atomic_dec(&sbi->s_flex_groups[f].free_inodes); 2737d39db14STheodore Ts'o } 2747d39db14STheodore Ts'o 275560671a0SAneesh Kumar K.V } 276717d50e4SAndreas Dilger gdp->bg_checksum = ext4_group_desc_csum(sbi, 277717d50e4SAndreas Dilger block_group, gdp); 278955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 279ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_freeinodes_counter); 280ac27a0ecSDave Kleikamp if (is_directory) 281ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_dirs_counter); 282ac27a0ecSDave Kleikamp 283772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2849f24e420STheodore Ts'o ext4_group_t f; 2859f24e420STheodore Ts'o 2869f24e420STheodore Ts'o f = ext4_flex_group(sbi, block_group); 2879f24e420STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].free_inodes); 288772cb7c8SJose R. Santos } 289ac27a0ecSDave Kleikamp } 2900390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 2910390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bh2); 292ac27a0ecSDave Kleikamp if (!fatal) fatal = err; 293ac27a0ecSDave Kleikamp } 2940390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 2950390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 296ac27a0ecSDave Kleikamp if (!fatal) 297ac27a0ecSDave Kleikamp fatal = err; 298ac27a0ecSDave Kleikamp sb->s_dirt = 1; 299ac27a0ecSDave Kleikamp error_return: 300ac27a0ecSDave Kleikamp brelse(bitmap_bh); 301617ba13bSMingming Cao ext4_std_error(sb, fatal); 302ac27a0ecSDave Kleikamp } 303ac27a0ecSDave Kleikamp 304ac27a0ecSDave Kleikamp /* 305ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 306ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 307ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 308ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 309ac27a0ecSDave Kleikamp * directories already is chosen. 310ac27a0ecSDave Kleikamp * 311ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory\'s block 312ac27a0ecSDave Kleikamp * group to find a free inode. 313ac27a0ecSDave Kleikamp */ 3142aa9fc4cSAvantika Mathur static int find_group_dir(struct super_block *sb, struct inode *parent, 3152aa9fc4cSAvantika Mathur ext4_group_t *best_group) 316ac27a0ecSDave Kleikamp { 3178df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 318ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 319617ba13bSMingming Cao struct ext4_group_desc *desc, *best_desc = NULL; 3202aa9fc4cSAvantika Mathur ext4_group_t group; 3212aa9fc4cSAvantika Mathur int ret = -1; 322ac27a0ecSDave Kleikamp 323617ba13bSMingming Cao freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter); 324ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 325ac27a0ecSDave Kleikamp 326ac27a0ecSDave Kleikamp for (group = 0; group < ngroups; group++) { 327ef2fb679SEric Sandeen desc = ext4_get_group_desc(sb, group, NULL); 328560671a0SAneesh Kumar K.V if (!desc || !ext4_free_inodes_count(sb, desc)) 329ac27a0ecSDave Kleikamp continue; 330560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc) < avefreei) 331ac27a0ecSDave Kleikamp continue; 332ac27a0ecSDave Kleikamp if (!best_desc || 333560671a0SAneesh Kumar K.V (ext4_free_blks_count(sb, desc) > 334560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, best_desc))) { 3352aa9fc4cSAvantika Mathur *best_group = group; 336ac27a0ecSDave Kleikamp best_desc = desc; 3372aa9fc4cSAvantika Mathur ret = 0; 338ac27a0ecSDave Kleikamp } 339ac27a0ecSDave Kleikamp } 3402aa9fc4cSAvantika Mathur return ret; 341ac27a0ecSDave Kleikamp } 342ac27a0ecSDave Kleikamp 343772cb7c8SJose R. Santos #define free_block_ratio 10 344772cb7c8SJose R. Santos 345772cb7c8SJose R. Santos static int find_group_flex(struct super_block *sb, struct inode *parent, 346772cb7c8SJose R. Santos ext4_group_t *best_group) 347772cb7c8SJose R. Santos { 348772cb7c8SJose R. Santos struct ext4_sb_info *sbi = EXT4_SB(sb); 349772cb7c8SJose R. Santos struct ext4_group_desc *desc; 350772cb7c8SJose R. Santos struct buffer_head *bh; 351772cb7c8SJose R. Santos struct flex_groups *flex_group = sbi->s_flex_groups; 352772cb7c8SJose R. Santos ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 353772cb7c8SJose R. Santos ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group); 3548df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 355772cb7c8SJose R. Santos int flex_size = ext4_flex_bg_size(sbi); 356772cb7c8SJose R. Santos ext4_group_t best_flex = parent_fbg_group; 357772cb7c8SJose R. Santos int blocks_per_flex = sbi->s_blocks_per_group * flex_size; 358772cb7c8SJose R. Santos int flexbg_free_blocks; 359772cb7c8SJose R. Santos int flex_freeb_ratio; 360772cb7c8SJose R. Santos ext4_group_t n_fbg_groups; 361772cb7c8SJose R. Santos ext4_group_t i; 362772cb7c8SJose R. Santos 3638df9675fSTheodore Ts'o n_fbg_groups = (ngroups + flex_size - 1) >> 364772cb7c8SJose R. Santos sbi->s_log_groups_per_flex; 365772cb7c8SJose R. Santos 366772cb7c8SJose R. Santos find_close_to_parent: 3679f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[best_flex].free_blocks); 368772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 3699f24e420STheodore Ts'o if (atomic_read(&flex_group[best_flex].free_inodes) && 370772cb7c8SJose R. Santos flex_freeb_ratio > free_block_ratio) 371772cb7c8SJose R. Santos goto found_flexbg; 372772cb7c8SJose R. Santos 373772cb7c8SJose R. Santos if (best_flex && best_flex == parent_fbg_group) { 374772cb7c8SJose R. Santos best_flex--; 375772cb7c8SJose R. Santos goto find_close_to_parent; 376772cb7c8SJose R. Santos } 377772cb7c8SJose R. Santos 378772cb7c8SJose R. Santos for (i = 0; i < n_fbg_groups; i++) { 379772cb7c8SJose R. Santos if (i == parent_fbg_group || i == parent_fbg_group - 1) 380772cb7c8SJose R. Santos continue; 381772cb7c8SJose R. Santos 3829f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[i].free_blocks); 383772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 384772cb7c8SJose R. Santos 385772cb7c8SJose R. Santos if (flex_freeb_ratio > free_block_ratio && 3869f24e420STheodore Ts'o (atomic_read(&flex_group[i].free_inodes))) { 387772cb7c8SJose R. Santos best_flex = i; 388772cb7c8SJose R. Santos goto found_flexbg; 389772cb7c8SJose R. Santos } 390772cb7c8SJose R. Santos 3919f24e420STheodore Ts'o if ((atomic_read(&flex_group[best_flex].free_inodes) == 0) || 3929f24e420STheodore Ts'o ((atomic_read(&flex_group[i].free_blocks) > 3939f24e420STheodore Ts'o atomic_read(&flex_group[best_flex].free_blocks)) && 3949f24e420STheodore Ts'o atomic_read(&flex_group[i].free_inodes))) 395772cb7c8SJose R. Santos best_flex = i; 396772cb7c8SJose R. Santos } 397772cb7c8SJose R. Santos 3989f24e420STheodore Ts'o if (!atomic_read(&flex_group[best_flex].free_inodes) || 3999f24e420STheodore Ts'o !atomic_read(&flex_group[best_flex].free_blocks)) 400772cb7c8SJose R. Santos return -1; 401772cb7c8SJose R. Santos 402772cb7c8SJose R. Santos found_flexbg: 403772cb7c8SJose R. Santos for (i = best_flex * flex_size; i < ngroups && 404772cb7c8SJose R. Santos i < (best_flex + 1) * flex_size; i++) { 405772cb7c8SJose R. Santos desc = ext4_get_group_desc(sb, i, &bh); 406560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc)) { 407772cb7c8SJose R. Santos *best_group = i; 408772cb7c8SJose R. Santos goto out; 409772cb7c8SJose R. Santos } 410772cb7c8SJose R. Santos } 411772cb7c8SJose R. Santos 412772cb7c8SJose R. Santos return -1; 413772cb7c8SJose R. Santos out: 414772cb7c8SJose R. Santos return 0; 415772cb7c8SJose R. Santos } 416772cb7c8SJose R. Santos 417a4912123STheodore Ts'o struct orlov_stats { 418a4912123STheodore Ts'o __u32 free_inodes; 419a4912123STheodore Ts'o __u32 free_blocks; 420a4912123STheodore Ts'o __u32 used_dirs; 421a4912123STheodore Ts'o }; 422a4912123STheodore Ts'o 423a4912123STheodore Ts'o /* 424a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 425a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 426a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 427a4912123STheodore Ts'o */ 428a4912123STheodore Ts'o void get_orlov_stats(struct super_block *sb, ext4_group_t g, 429a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 430a4912123STheodore Ts'o { 431a4912123STheodore Ts'o struct ext4_group_desc *desc; 4327d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 433a4912123STheodore Ts'o 4347d39db14STheodore Ts'o if (flex_size > 1) { 4357d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 4367d39db14STheodore Ts'o stats->free_blocks = atomic_read(&flex_group[g].free_blocks); 4377d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 4387d39db14STheodore Ts'o return; 4397d39db14STheodore Ts'o } 4407d39db14STheodore Ts'o 4417d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 4427d39db14STheodore Ts'o if (desc) { 4437d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 4447d39db14STheodore Ts'o stats->free_blocks = ext4_free_blks_count(sb, desc); 4457d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 4467d39db14STheodore Ts'o } else { 447a4912123STheodore Ts'o stats->free_inodes = 0; 448a4912123STheodore Ts'o stats->free_blocks = 0; 449a4912123STheodore Ts'o stats->used_dirs = 0; 450a4912123STheodore Ts'o } 451a4912123STheodore Ts'o } 452a4912123STheodore Ts'o 453ac27a0ecSDave Kleikamp /* 454ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 455ac27a0ecSDave Kleikamp * 456ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 457ac27a0ecSDave Kleikamp * 458ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 459ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 460ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 461ac27a0ecSDave Kleikamp * 462ac27a0ecSDave Kleikamp * For the rest rules look so: 463ac27a0ecSDave Kleikamp * 464ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 465ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 466ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 467ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4681cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 469ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 470ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 471ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 472ac27a0ecSDave Kleikamp */ 473ac27a0ecSDave Kleikamp 4742aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 475a4912123STheodore Ts'o ext4_group_t *group, int mode) 476ac27a0ecSDave Kleikamp { 477fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 478617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 4798df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 480617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 481ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 482617ba13bSMingming Cao ext4_fsblk_t freeb, avefreeb; 483ac27a0ecSDave Kleikamp unsigned int ndirs; 484a4912123STheodore Ts'o int max_dirs, min_inodes; 485617ba13bSMingming Cao ext4_grpblk_t min_blocks; 4868df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 487617ba13bSMingming Cao struct ext4_group_desc *desc; 488a4912123STheodore Ts'o struct orlov_stats stats; 489a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 490a4912123STheodore Ts'o 4918df9675fSTheodore Ts'o ngroups = real_ngroups; 492a4912123STheodore Ts'o if (flex_size > 1) { 4938df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 494a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 495a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 496a4912123STheodore Ts'o } 497ac27a0ecSDave Kleikamp 498ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 499ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 500ac27a0ecSDave Kleikamp freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter); 5013a5b2ecdSMingming Cao avefreeb = freeb; 502f4e5bc24SAndrew Morton do_div(avefreeb, ngroups); 503ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 504ac27a0ecSDave Kleikamp 505a4912123STheodore Ts'o if (S_ISDIR(mode) && 506a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 507a4912123STheodore Ts'o (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL))) { 508ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 5092aa9fc4cSAvantika Mathur int ret = -1; 510ac27a0ecSDave Kleikamp 5112aa9fc4cSAvantika Mathur get_random_bytes(&grp, sizeof(grp)); 5122aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 513ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 514a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 515a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 516a4912123STheodore Ts'o if (!stats.free_inodes) 517ac27a0ecSDave Kleikamp continue; 518a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 519ac27a0ecSDave Kleikamp continue; 520a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 521ac27a0ecSDave Kleikamp continue; 522a4912123STheodore Ts'o if (stats.free_blocks < avefreeb) 523ac27a0ecSDave Kleikamp continue; 524a4912123STheodore Ts'o grp = g; 5252aa9fc4cSAvantika Mathur ret = 0; 526a4912123STheodore Ts'o best_ndir = stats.used_dirs; 527ac27a0ecSDave Kleikamp } 528a4912123STheodore Ts'o if (ret) 529a4912123STheodore Ts'o goto fallback; 530a4912123STheodore Ts'o found_flex_bg: 531a4912123STheodore Ts'o if (flex_size == 1) { 532a4912123STheodore Ts'o *group = grp; 533a4912123STheodore Ts'o return 0; 534a4912123STheodore Ts'o } 535a4912123STheodore Ts'o 536a4912123STheodore Ts'o /* 537a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 538a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 539a4912123STheodore Ts'o * something similar, although regular files will 540a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 541a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 542a4912123STheodore Ts'o */ 543a4912123STheodore Ts'o grp *= flex_size; 544a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 5458df9675fSTheodore Ts'o if (grp+i >= real_ngroups) 546a4912123STheodore Ts'o break; 547a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 548a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 549a4912123STheodore Ts'o *group = grp+i; 550a4912123STheodore Ts'o return 0; 551a4912123STheodore Ts'o } 552a4912123STheodore Ts'o } 553ac27a0ecSDave Kleikamp goto fallback; 554ac27a0ecSDave Kleikamp } 555ac27a0ecSDave Kleikamp 556ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 557a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 558a4912123STheodore Ts'o if (min_inodes < 1) 559a4912123STheodore Ts'o min_inodes = 1; 560a4912123STheodore Ts'o min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb)*flex_size / 4; 561ac27a0ecSDave Kleikamp 562a4912123STheodore Ts'o /* 563a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 564a4912123STheodore Ts'o * inode for this parent directory 565a4912123STheodore Ts'o */ 566a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 567a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 568a4912123STheodore Ts'o if (flex_size > 1) 569a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 570a4912123STheodore Ts'o } 571ac27a0ecSDave Kleikamp 572ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 573a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 574a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 575a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 576ac27a0ecSDave Kleikamp continue; 577a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 578ac27a0ecSDave Kleikamp continue; 579a4912123STheodore Ts'o if (stats.free_blocks < min_blocks) 580ac27a0ecSDave Kleikamp continue; 581a4912123STheodore Ts'o goto found_flex_bg; 582ac27a0ecSDave Kleikamp } 583ac27a0ecSDave Kleikamp 584ac27a0ecSDave Kleikamp fallback: 5858df9675fSTheodore Ts'o ngroups = real_ngroups; 586a4912123STheodore Ts'o avefreei = freei / ngroups; 587b5451f7bSTheodore Ts'o fallback_retry: 588a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 589ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 590a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 591a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 592560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 593a4912123STheodore Ts'o ext4_free_inodes_count(sb, desc) >= avefreei) { 594a4912123STheodore Ts'o *group = grp; 5952aa9fc4cSAvantika Mathur return 0; 596ac27a0ecSDave Kleikamp } 597a4912123STheodore Ts'o } 598ac27a0ecSDave Kleikamp 599ac27a0ecSDave Kleikamp if (avefreei) { 600ac27a0ecSDave Kleikamp /* 601ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 602ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 603ac27a0ecSDave Kleikamp */ 604ac27a0ecSDave Kleikamp avefreei = 0; 605b5451f7bSTheodore Ts'o goto fallback_retry; 606ac27a0ecSDave Kleikamp } 607ac27a0ecSDave Kleikamp 608ac27a0ecSDave Kleikamp return -1; 609ac27a0ecSDave Kleikamp } 610ac27a0ecSDave Kleikamp 6112aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 612a4912123STheodore Ts'o ext4_group_t *group, int mode) 613ac27a0ecSDave Kleikamp { 614fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 6158df9675fSTheodore Ts'o ext4_group_t i, last, ngroups = ext4_get_groups_count(sb); 616617ba13bSMingming Cao struct ext4_group_desc *desc; 617a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 618a4912123STheodore Ts'o 619a4912123STheodore Ts'o /* 620a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 621a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 622a4912123STheodore Ts'o * find another flex group, and store that information in the 623a4912123STheodore Ts'o * parent directory's inode information so that use that flex 624a4912123STheodore Ts'o * group for future allocations. 625a4912123STheodore Ts'o */ 626a4912123STheodore Ts'o if (flex_size > 1) { 627a4912123STheodore Ts'o int retry = 0; 628a4912123STheodore Ts'o 629a4912123STheodore Ts'o try_again: 630a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 631a4912123STheodore Ts'o last = parent_group + flex_size; 632a4912123STheodore Ts'o if (last > ngroups) 633a4912123STheodore Ts'o last = ngroups; 634a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 635a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 636a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 637a4912123STheodore Ts'o *group = i; 638a4912123STheodore Ts'o return 0; 639a4912123STheodore Ts'o } 640a4912123STheodore Ts'o } 641a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 642a4912123STheodore Ts'o retry = 1; 643a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 644a4912123STheodore Ts'o goto try_again; 645a4912123STheodore Ts'o } 646a4912123STheodore Ts'o /* 647a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 648a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 649a4912123STheodore Ts'o * avoid the topdir algorithms. 650a4912123STheodore Ts'o */ 651a4912123STheodore Ts'o *group = parent_group + flex_size; 652a4912123STheodore Ts'o if (*group > ngroups) 653a4912123STheodore Ts'o *group = 0; 654a4912123STheodore Ts'o return find_group_orlov(sb, parent, group, mode); 655a4912123STheodore Ts'o } 656ac27a0ecSDave Kleikamp 657ac27a0ecSDave Kleikamp /* 658ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 659ac27a0ecSDave Kleikamp */ 6602aa9fc4cSAvantika Mathur *group = parent_group; 6612aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 662560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 663560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6642aa9fc4cSAvantika Mathur return 0; 665ac27a0ecSDave Kleikamp 666ac27a0ecSDave Kleikamp /* 667ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 668ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 669ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 670ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 671ac27a0ecSDave Kleikamp * different blockgroup. 672ac27a0ecSDave Kleikamp * 673ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 674ac27a0ecSDave Kleikamp */ 6752aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 676ac27a0ecSDave Kleikamp 677ac27a0ecSDave Kleikamp /* 678ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 679ac27a0ecSDave Kleikamp * blocks. 680ac27a0ecSDave Kleikamp */ 681ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6822aa9fc4cSAvantika Mathur *group += i; 6832aa9fc4cSAvantika Mathur if (*group >= ngroups) 6842aa9fc4cSAvantika Mathur *group -= ngroups; 6852aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 686560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 687560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6882aa9fc4cSAvantika Mathur return 0; 689ac27a0ecSDave Kleikamp } 690ac27a0ecSDave Kleikamp 691ac27a0ecSDave Kleikamp /* 692ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 693ac27a0ecSDave Kleikamp * has no free blocks. 694ac27a0ecSDave Kleikamp */ 6952aa9fc4cSAvantika Mathur *group = parent_group; 696ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6972aa9fc4cSAvantika Mathur if (++*group >= ngroups) 6982aa9fc4cSAvantika Mathur *group = 0; 6992aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 700560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 7012aa9fc4cSAvantika Mathur return 0; 702ac27a0ecSDave Kleikamp } 703ac27a0ecSDave Kleikamp 704ac27a0ecSDave Kleikamp return -1; 705ac27a0ecSDave Kleikamp } 706ac27a0ecSDave Kleikamp 707ac27a0ecSDave Kleikamp /* 70839341867SAneesh Kumar K.V * claim the inode from the inode bitmap. If the group 709955ce5f5SAneesh Kumar K.V * is uninit we need to take the groups's ext4_group_lock 71039341867SAneesh Kumar K.V * and clear the uninit flag. The inode bitmap update 71139341867SAneesh Kumar K.V * and group desc uninit flag clear should be done 712955ce5f5SAneesh Kumar K.V * after holding ext4_group_lock so that ext4_read_inode_bitmap 71339341867SAneesh Kumar K.V * doesn't race with the ext4_claim_inode 71439341867SAneesh Kumar K.V */ 71539341867SAneesh Kumar K.V static int ext4_claim_inode(struct super_block *sb, 71639341867SAneesh Kumar K.V struct buffer_head *inode_bitmap_bh, 71739341867SAneesh Kumar K.V unsigned long ino, ext4_group_t group, int mode) 71839341867SAneesh Kumar K.V { 71939341867SAneesh Kumar K.V int free = 0, retval = 0, count; 72039341867SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 72139341867SAneesh Kumar K.V struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL); 72239341867SAneesh Kumar K.V 723955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 72439341867SAneesh Kumar K.V if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) { 72539341867SAneesh Kumar K.V /* not a free inode */ 72639341867SAneesh Kumar K.V retval = 1; 72739341867SAneesh Kumar K.V goto err_ret; 72839341867SAneesh Kumar K.V } 72939341867SAneesh Kumar K.V ino++; 73039341867SAneesh Kumar K.V if ((group == 0 && ino < EXT4_FIRST_INO(sb)) || 73139341867SAneesh Kumar K.V ino > EXT4_INODES_PER_GROUP(sb)) { 732955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 73339341867SAneesh Kumar K.V ext4_error(sb, __func__, 73439341867SAneesh Kumar K.V "reserved inode or inode > inodes count - " 73539341867SAneesh Kumar K.V "block_group = %u, inode=%lu", group, 73639341867SAneesh Kumar K.V ino + group * EXT4_INODES_PER_GROUP(sb)); 73739341867SAneesh Kumar K.V return 1; 73839341867SAneesh Kumar K.V } 73939341867SAneesh Kumar K.V /* If we didn't allocate from within the initialized part of the inode 74039341867SAneesh Kumar K.V * table then we need to initialize up to this inode. */ 74139341867SAneesh Kumar K.V if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) { 74239341867SAneesh Kumar K.V 74339341867SAneesh Kumar K.V if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 74439341867SAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 74539341867SAneesh Kumar K.V /* When marking the block group with 74639341867SAneesh Kumar K.V * ~EXT4_BG_INODE_UNINIT we don't want to depend 74739341867SAneesh Kumar K.V * on the value of bg_itable_unused even though 74839341867SAneesh Kumar K.V * mke2fs could have initialized the same for us. 74939341867SAneesh Kumar K.V * Instead we calculated the value below 75039341867SAneesh Kumar K.V */ 75139341867SAneesh Kumar K.V 75239341867SAneesh Kumar K.V free = 0; 75339341867SAneesh Kumar K.V } else { 75439341867SAneesh Kumar K.V free = EXT4_INODES_PER_GROUP(sb) - 75539341867SAneesh Kumar K.V ext4_itable_unused_count(sb, gdp); 75639341867SAneesh Kumar K.V } 75739341867SAneesh Kumar K.V 75839341867SAneesh Kumar K.V /* 75939341867SAneesh Kumar K.V * Check the relative inode number against the last used 76039341867SAneesh Kumar K.V * relative inode number in this group. if it is greater 76139341867SAneesh Kumar K.V * we need to update the bg_itable_unused count 76239341867SAneesh Kumar K.V * 76339341867SAneesh Kumar K.V */ 76439341867SAneesh Kumar K.V if (ino > free) 76539341867SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 76639341867SAneesh Kumar K.V (EXT4_INODES_PER_GROUP(sb) - ino)); 76739341867SAneesh Kumar K.V } 76839341867SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) - 1; 76939341867SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 77039341867SAneesh Kumar K.V if (S_ISDIR(mode)) { 77139341867SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) + 1; 77239341867SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 7737d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 7747d39db14STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 7757d39db14STheodore Ts'o 7767d39db14STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].free_inodes); 7777d39db14STheodore Ts'o } 77839341867SAneesh Kumar K.V } 77939341867SAneesh Kumar K.V gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 78039341867SAneesh Kumar K.V err_ret: 781955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 78239341867SAneesh Kumar K.V return retval; 78339341867SAneesh Kumar K.V } 78439341867SAneesh Kumar K.V 78539341867SAneesh Kumar K.V /* 786ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 787ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 788ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 789ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 790ac27a0ecSDave Kleikamp * directories already is chosen. 791ac27a0ecSDave Kleikamp * 792ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 793ac27a0ecSDave Kleikamp * group to find a free inode. 794ac27a0ecSDave Kleikamp */ 795617ba13bSMingming Cao struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode) 796ac27a0ecSDave Kleikamp { 797ac27a0ecSDave Kleikamp struct super_block *sb; 7983300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 7993300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 8008df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 801ac27a0ecSDave Kleikamp unsigned long ino = 0; 802ac27a0ecSDave Kleikamp struct inode *inode; 803617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 804617ba13bSMingming Cao struct ext4_inode_info *ei; 805617ba13bSMingming Cao struct ext4_sb_info *sbi; 80639341867SAneesh Kumar K.V int ret2, err = 0; 807ac27a0ecSDave Kleikamp struct inode *ret; 8082aa9fc4cSAvantika Mathur ext4_group_t i; 8092aa9fc4cSAvantika Mathur int free = 0; 8102842c3b5STheodore Ts'o static int once = 1; 811772cb7c8SJose R. Santos ext4_group_t flex_group; 812ac27a0ecSDave Kleikamp 813ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 814ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 815ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 816ac27a0ecSDave Kleikamp 817ac27a0ecSDave Kleikamp sb = dir->i_sb; 8188df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 819ba80b101STheodore Ts'o trace_mark(ext4_request_inode, "dev %s dir %lu mode %d", sb->s_id, 820ba80b101STheodore Ts'o dir->i_ino, mode); 821ac27a0ecSDave Kleikamp inode = new_inode(sb); 822ac27a0ecSDave Kleikamp if (!inode) 823ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 824617ba13bSMingming Cao ei = EXT4_I(inode); 825617ba13bSMingming Cao sbi = EXT4_SB(sb); 826772cb7c8SJose R. Santos 827a4912123STheodore Ts'o if (sbi->s_log_groups_per_flex && test_opt(sb, OLDALLOC)) { 828772cb7c8SJose R. Santos ret2 = find_group_flex(sb, dir, &group); 82905bf9e83STheodore Ts'o if (ret2 == -1) { 830a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 8316b82f3cbSChuck Ebbert if (ret2 == 0 && once) { 8322842c3b5STheodore Ts'o once = 0; 83305bf9e83STheodore Ts'o printk(KERN_NOTICE "ext4: find_group_flex " 83405bf9e83STheodore Ts'o "failed, fallback succeeded dir %lu\n", 83505bf9e83STheodore Ts'o dir->i_ino); 83605bf9e83STheodore Ts'o } 8376b82f3cbSChuck Ebbert } 838772cb7c8SJose R. Santos goto got_group; 839772cb7c8SJose R. Santos } 840772cb7c8SJose R. Santos 841ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) { 842ac27a0ecSDave Kleikamp if (test_opt(sb, OLDALLOC)) 8432aa9fc4cSAvantika Mathur ret2 = find_group_dir(sb, dir, &group); 844ac27a0ecSDave Kleikamp else 845a4912123STheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode); 846ac27a0ecSDave Kleikamp } else 847a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 848ac27a0ecSDave Kleikamp 849772cb7c8SJose R. Santos got_group: 850a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 851ac27a0ecSDave Kleikamp err = -ENOSPC; 8522aa9fc4cSAvantika Mathur if (ret2 == -1) 853ac27a0ecSDave Kleikamp goto out; 854ac27a0ecSDave Kleikamp 8558df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 856ac27a0ecSDave Kleikamp err = -EIO; 857ac27a0ecSDave Kleikamp 8583300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 859ac27a0ecSDave Kleikamp if (!gdp) 860ac27a0ecSDave Kleikamp goto fail; 861ac27a0ecSDave Kleikamp 8623300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 8633300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 8643300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 865ac27a0ecSDave Kleikamp goto fail; 866ac27a0ecSDave Kleikamp 867ac27a0ecSDave Kleikamp ino = 0; 868ac27a0ecSDave Kleikamp 869ac27a0ecSDave Kleikamp repeat_in_this_group: 870617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 8713300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 8723300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 8733300bedaSAneesh Kumar K.V 874617ba13bSMingming Cao if (ino < EXT4_INODES_PER_GROUP(sb)) { 875ac27a0ecSDave Kleikamp 8763300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 8773300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 8783300bedaSAneesh Kumar K.V inode_bitmap_bh); 879ac27a0ecSDave Kleikamp if (err) 880ac27a0ecSDave Kleikamp goto fail; 881ac27a0ecSDave Kleikamp 88239341867SAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "get_write_access"); 88339341867SAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 88439341867SAneesh Kumar K.V group_desc_bh); 88539341867SAneesh Kumar K.V if (err) 88639341867SAneesh Kumar K.V goto fail; 88739341867SAneesh Kumar K.V if (!ext4_claim_inode(sb, inode_bitmap_bh, 88839341867SAneesh Kumar K.V ino, group, mode)) { 889ac27a0ecSDave Kleikamp /* we won it */ 8903300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, 8910390131bSFrank Mayhar "call ext4_handle_dirty_metadata"); 8920390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 8930390131bSFrank Mayhar inode, 8943300bedaSAneesh Kumar K.V inode_bitmap_bh); 895ac27a0ecSDave Kleikamp if (err) 896ac27a0ecSDave Kleikamp goto fail; 89739341867SAneesh Kumar K.V /* zero bit is inode number 1*/ 89839341867SAneesh Kumar K.V ino++; 899ac27a0ecSDave Kleikamp goto got; 900ac27a0ecSDave Kleikamp } 901ac27a0ecSDave Kleikamp /* we lost it */ 9023300bedaSAneesh Kumar K.V ext4_handle_release_buffer(handle, inode_bitmap_bh); 90339341867SAneesh Kumar K.V ext4_handle_release_buffer(handle, group_desc_bh); 904ac27a0ecSDave Kleikamp 905617ba13bSMingming Cao if (++ino < EXT4_INODES_PER_GROUP(sb)) 906ac27a0ecSDave Kleikamp goto repeat_in_this_group; 907ac27a0ecSDave Kleikamp } 908ac27a0ecSDave Kleikamp 909ac27a0ecSDave Kleikamp /* 910ac27a0ecSDave Kleikamp * This case is possible in concurrent environment. It is very 911ac27a0ecSDave Kleikamp * rare. We cannot repeat the find_group_xxx() call because 912ac27a0ecSDave Kleikamp * that will simply return the same blockgroup, because the 913ac27a0ecSDave Kleikamp * group descriptor metadata has not yet been updated. 914ac27a0ecSDave Kleikamp * So we just go onto the next blockgroup. 915ac27a0ecSDave Kleikamp */ 9168df9675fSTheodore Ts'o if (++group == ngroups) 917ac27a0ecSDave Kleikamp group = 0; 918ac27a0ecSDave Kleikamp } 919ac27a0ecSDave Kleikamp err = -ENOSPC; 920ac27a0ecSDave Kleikamp goto out; 921ac27a0ecSDave Kleikamp 922ac27a0ecSDave Kleikamp got: 923717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 924717d50e4SAndreas Dilger if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) && 925717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9263300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 927717d50e4SAndreas Dilger 9283300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 9293300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 9303300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 931717d50e4SAndreas Dilger if (err) { 9323300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 933717d50e4SAndreas Dilger goto fail; 934717d50e4SAndreas Dilger } 935717d50e4SAndreas Dilger 936717d50e4SAndreas Dilger free = 0; 937955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 938717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 939717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 940717d50e4SAndreas Dilger free = ext4_free_blocks_after_init(sb, group, gdp); 9413300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 942560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, free); 94323712a9cSFrederic Bohe gdp->bg_checksum = ext4_group_desc_csum(sbi, group, 94423712a9cSFrederic Bohe gdp); 945717d50e4SAndreas Dilger } 946955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 947717d50e4SAndreas Dilger 948717d50e4SAndreas Dilger /* Don't need to dirty bitmap block if we didn't change it */ 949717d50e4SAndreas Dilger if (free) { 9503300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 9510390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 9523300bedaSAneesh Kumar K.V NULL, block_bitmap_bh); 953717d50e4SAndreas Dilger } 954717d50e4SAndreas Dilger 9553300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 956717d50e4SAndreas Dilger if (err) 957717d50e4SAndreas Dilger goto fail; 958717d50e4SAndreas Dilger } 9593300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 9603300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 96139341867SAneesh Kumar K.V if (err) 96239341867SAneesh Kumar K.V goto fail; 963ac27a0ecSDave Kleikamp 964ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 965ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 966ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 967ac27a0ecSDave Kleikamp sb->s_dirt = 1; 968ac27a0ecSDave Kleikamp 969772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 970772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 9719f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 972772cb7c8SJose R. Santos } 973772cb7c8SJose R. Santos 9744c9c544eSDavid Howells inode->i_uid = current_fsuid(); 975ac27a0ecSDave Kleikamp if (test_opt(sb, GRPID)) 976ac27a0ecSDave Kleikamp inode->i_gid = dir->i_gid; 977ac27a0ecSDave Kleikamp else if (dir->i_mode & S_ISGID) { 978ac27a0ecSDave Kleikamp inode->i_gid = dir->i_gid; 979ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 980ac27a0ecSDave Kleikamp mode |= S_ISGID; 981ac27a0ecSDave Kleikamp } else 9824c9c544eSDavid Howells inode->i_gid = current_fsgid(); 983ac27a0ecSDave Kleikamp inode->i_mode = mode; 984ac27a0ecSDave Kleikamp 985717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 986ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 987ac27a0ecSDave Kleikamp inode->i_blocks = 0; 988ef7f3835SKalpak Shah inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime = 989ef7f3835SKalpak Shah ext4_current_time(inode); 990ac27a0ecSDave Kleikamp 991ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 992ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 993ac27a0ecSDave Kleikamp ei->i_disksize = 0; 994ac27a0ecSDave Kleikamp 99542bf0383SAneesh Kumar K.V /* 9962dc6b0d4SDuane Griffin * Don't inherit extent flag from directory, amongst others. We set 9972dc6b0d4SDuane Griffin * extent flag on newly created directory and file only if -o extent 9982dc6b0d4SDuane Griffin * mount option is specified 99942bf0383SAneesh Kumar K.V */ 10002dc6b0d4SDuane Griffin ei->i_flags = 10012dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 1002ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 1003ac27a0ecSDave Kleikamp ei->i_dtime = 0; 1004ac27a0ecSDave Kleikamp ei->i_block_group = group; 1005a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 1006ac27a0ecSDave Kleikamp 1007617ba13bSMingming Cao ext4_set_inode_flags(inode); 1008ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 10090390131bSFrank Mayhar ext4_handle_sync(handle); 10106b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 10116b38e842SAl Viro err = -EINVAL; 10126b38e842SAl Viro goto fail_drop; 10136b38e842SAl Viro } 1014ac27a0ecSDave Kleikamp spin_lock(&sbi->s_next_gen_lock); 1015ac27a0ecSDave Kleikamp inode->i_generation = sbi->s_next_generation++; 1016ac27a0ecSDave Kleikamp spin_unlock(&sbi->s_next_gen_lock); 1017ac27a0ecSDave Kleikamp 1018617ba13bSMingming Cao ei->i_state = EXT4_STATE_NEW; 1019ef7f3835SKalpak Shah 1020ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 1021ac27a0ecSDave Kleikamp 1022ac27a0ecSDave Kleikamp ret = inode; 1023a269eb18SJan Kara if (vfs_dq_alloc_inode(inode)) { 1024ac27a0ecSDave Kleikamp err = -EDQUOT; 1025ac27a0ecSDave Kleikamp goto fail_drop; 1026ac27a0ecSDave Kleikamp } 1027ac27a0ecSDave Kleikamp 1028617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1029ac27a0ecSDave Kleikamp if (err) 1030ac27a0ecSDave Kleikamp goto fail_free_drop; 1031ac27a0ecSDave Kleikamp 1032617ba13bSMingming Cao err = ext4_init_security(handle, inode, dir); 1033ac27a0ecSDave Kleikamp if (err) 1034ac27a0ecSDave Kleikamp goto fail_free_drop; 1035ac27a0ecSDave Kleikamp 103683982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 1037e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1038e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 1039a86c6181SAlex Tomas EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL; 1040a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1041a86c6181SAlex Tomas } 104242bf0383SAneesh Kumar K.V } 1043ac27a0ecSDave Kleikamp 10448753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 10458753e88fSAneesh Kumar K.V if (err) { 10468753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 10478753e88fSAneesh Kumar K.V goto fail_free_drop; 10488753e88fSAneesh Kumar K.V } 10498753e88fSAneesh Kumar K.V 1050617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 1051ba80b101STheodore Ts'o trace_mark(ext4_allocate_inode, "dev %s ino %lu dir %lu mode %d", 1052ba80b101STheodore Ts'o sb->s_id, inode->i_ino, dir->i_ino, mode); 1053ac27a0ecSDave Kleikamp goto really_out; 1054ac27a0ecSDave Kleikamp fail: 1055617ba13bSMingming Cao ext4_std_error(sb, err); 1056ac27a0ecSDave Kleikamp out: 1057ac27a0ecSDave Kleikamp iput(inode); 1058ac27a0ecSDave Kleikamp ret = ERR_PTR(err); 1059ac27a0ecSDave Kleikamp really_out: 10603300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1061ac27a0ecSDave Kleikamp return ret; 1062ac27a0ecSDave Kleikamp 1063ac27a0ecSDave Kleikamp fail_free_drop: 1064a269eb18SJan Kara vfs_dq_free_inode(inode); 1065ac27a0ecSDave Kleikamp 1066ac27a0ecSDave Kleikamp fail_drop: 1067a269eb18SJan Kara vfs_dq_drop(inode); 1068ac27a0ecSDave Kleikamp inode->i_flags |= S_NOQUOTA; 1069ac27a0ecSDave Kleikamp inode->i_nlink = 0; 10706b38e842SAl Viro unlock_new_inode(inode); 1071ac27a0ecSDave Kleikamp iput(inode); 10723300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1073ac27a0ecSDave Kleikamp return ERR_PTR(err); 1074ac27a0ecSDave Kleikamp } 1075ac27a0ecSDave Kleikamp 1076ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1077617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1078ac27a0ecSDave Kleikamp { 1079617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1080fd2d4291SAvantika Mathur ext4_group_t block_group; 1081ac27a0ecSDave Kleikamp int bit; 10821d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 1083ac27a0ecSDave Kleikamp struct inode *inode = NULL; 10841d1fe1eeSDavid Howells long err = -EIO; 1085ac27a0ecSDave Kleikamp 1086ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1087ac27a0ecSDave Kleikamp if (ino > max_ino) { 108846e665e9SHarvey Harrison ext4_warning(sb, __func__, 1089ac27a0ecSDave Kleikamp "bad orphan ino %lu! e2fsck was run?", ino); 10901d1fe1eeSDavid Howells goto error; 1091ac27a0ecSDave Kleikamp } 1092ac27a0ecSDave Kleikamp 1093617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1094617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1095e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1096ac27a0ecSDave Kleikamp if (!bitmap_bh) { 109746e665e9SHarvey Harrison ext4_warning(sb, __func__, 1098ac27a0ecSDave Kleikamp "inode bitmap error for orphan %lu", ino); 10991d1fe1eeSDavid Howells goto error; 1100ac27a0ecSDave Kleikamp } 1101ac27a0ecSDave Kleikamp 1102ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1103ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1104ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1105ac27a0ecSDave Kleikamp */ 11061d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 11071d1fe1eeSDavid Howells goto bad_orphan; 11081d1fe1eeSDavid Howells 11091d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 11101d1fe1eeSDavid Howells if (IS_ERR(inode)) 11111d1fe1eeSDavid Howells goto iget_failed; 11121d1fe1eeSDavid Howells 111391ef4cafSDuane Griffin /* 111491ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 111591ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 111691ef4cafSDuane Griffin * during processing and an infinite loop will result. 111791ef4cafSDuane Griffin */ 111891ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 111991ef4cafSDuane Griffin goto bad_orphan; 112091ef4cafSDuane Griffin 11211d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 11221d1fe1eeSDavid Howells goto bad_orphan; 11231d1fe1eeSDavid Howells brelse(bitmap_bh); 11241d1fe1eeSDavid Howells return inode; 11251d1fe1eeSDavid Howells 11261d1fe1eeSDavid Howells iget_failed: 11271d1fe1eeSDavid Howells err = PTR_ERR(inode); 11281d1fe1eeSDavid Howells inode = NULL; 11291d1fe1eeSDavid Howells bad_orphan: 113046e665e9SHarvey Harrison ext4_warning(sb, __func__, 1131ac27a0ecSDave Kleikamp "bad orphan inode %lu! e2fsck was run?", ino); 1132617ba13bSMingming Cao printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1133ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1134617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1135ac27a0ecSDave Kleikamp printk(KERN_NOTICE "inode=%p\n", inode); 1136ac27a0ecSDave Kleikamp if (inode) { 1137ac27a0ecSDave Kleikamp printk(KERN_NOTICE "is_bad_inode(inode)=%d\n", 1138ac27a0ecSDave Kleikamp is_bad_inode(inode)); 1139ac27a0ecSDave Kleikamp printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n", 1140ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 1141ac27a0ecSDave Kleikamp printk(KERN_NOTICE "max_ino=%lu\n", max_ino); 114291ef4cafSDuane Griffin printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink); 1143ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 11441d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1145ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1146ac27a0ecSDave Kleikamp iput(inode); 1147ac27a0ecSDave Kleikamp } 1148ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11491d1fe1eeSDavid Howells error: 11501d1fe1eeSDavid Howells return ERR_PTR(err); 1151ac27a0ecSDave Kleikamp } 1152ac27a0ecSDave Kleikamp 1153617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1154ac27a0ecSDave Kleikamp { 1155ac27a0ecSDave Kleikamp unsigned long desc_count; 1156617ba13bSMingming Cao struct ext4_group_desc *gdp; 11578df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1158617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1159617ba13bSMingming Cao struct ext4_super_block *es; 1160ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1161ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1162ac27a0ecSDave Kleikamp 1163617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1164ac27a0ecSDave Kleikamp desc_count = 0; 1165ac27a0ecSDave Kleikamp bitmap_count = 0; 1166ac27a0ecSDave Kleikamp gdp = NULL; 11678df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1168617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1169ac27a0ecSDave Kleikamp if (!gdp) 1170ac27a0ecSDave Kleikamp continue; 1171560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1172ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1173e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1174ac27a0ecSDave Kleikamp if (!bitmap_bh) 1175ac27a0ecSDave Kleikamp continue; 1176ac27a0ecSDave Kleikamp 1177617ba13bSMingming Cao x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8); 1178c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1179560671a0SAneesh Kumar K.V i, ext4_free_inodes_count(sb, gdp), x); 1180ac27a0ecSDave Kleikamp bitmap_count += x; 1181ac27a0ecSDave Kleikamp } 1182ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11834776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 11844776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1185ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1186ac27a0ecSDave Kleikamp return desc_count; 1187ac27a0ecSDave Kleikamp #else 1188ac27a0ecSDave Kleikamp desc_count = 0; 11898df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1190617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1191ac27a0ecSDave Kleikamp if (!gdp) 1192ac27a0ecSDave Kleikamp continue; 1193560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1194ac27a0ecSDave Kleikamp cond_resched(); 1195ac27a0ecSDave Kleikamp } 1196ac27a0ecSDave Kleikamp return desc_count; 1197ac27a0ecSDave Kleikamp #endif 1198ac27a0ecSDave Kleikamp } 1199ac27a0ecSDave Kleikamp 1200ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1201617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1202ac27a0ecSDave Kleikamp { 1203ac27a0ecSDave Kleikamp unsigned long count = 0; 12048df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1205ac27a0ecSDave Kleikamp 12068df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1207617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1208ac27a0ecSDave Kleikamp if (!gdp) 1209ac27a0ecSDave Kleikamp continue; 1210560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1211ac27a0ecSDave Kleikamp } 1212ac27a0ecSDave Kleikamp return count; 1213ac27a0ecSDave Kleikamp } 1214