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)) { 7912062dddSEric Sandeen ext4_error(sb, "Checksum bad for group %u", block_group); 80560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, 0); 81560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, 0); 82560671a0SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 0); 83717d50e4SAndreas Dilger memset(bh->b_data, 0xff, sb->s_blocksize); 84717d50e4SAndreas Dilger return 0; 85717d50e4SAndreas Dilger } 86717d50e4SAndreas Dilger 87717d50e4SAndreas Dilger memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8); 88648f5879SAneesh Kumar K.V mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), sb->s_blocksize * 8, 89717d50e4SAndreas Dilger bh->b_data); 90717d50e4SAndreas Dilger 91717d50e4SAndreas Dilger return EXT4_INODES_PER_GROUP(sb); 92717d50e4SAndreas Dilger } 93ac27a0ecSDave Kleikamp 94ac27a0ecSDave Kleikamp /* 95ac27a0ecSDave Kleikamp * Read the inode allocation bitmap for a given block_group, reading 96ac27a0ecSDave Kleikamp * into the specified slot in the superblock's bitmap cache. 97ac27a0ecSDave Kleikamp * 98ac27a0ecSDave Kleikamp * Return buffer_head of bitmap on success or NULL. 99ac27a0ecSDave Kleikamp */ 100ac27a0ecSDave Kleikamp static struct buffer_head * 101e29d1cdeSEric Sandeen ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group) 102ac27a0ecSDave Kleikamp { 103617ba13bSMingming Cao struct ext4_group_desc *desc; 104ac27a0ecSDave Kleikamp struct buffer_head *bh = NULL; 105e29d1cdeSEric Sandeen ext4_fsblk_t bitmap_blk; 106ac27a0ecSDave Kleikamp 107617ba13bSMingming Cao desc = ext4_get_group_desc(sb, block_group, NULL); 108ac27a0ecSDave Kleikamp if (!desc) 109e29d1cdeSEric Sandeen return NULL; 110e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 111e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 112e29d1cdeSEric Sandeen if (unlikely(!bh)) { 11312062dddSEric Sandeen ext4_error(sb, "Cannot read inode bitmap - " 114a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 115e29d1cdeSEric Sandeen block_group, bitmap_blk); 116e29d1cdeSEric Sandeen return NULL; 117e29d1cdeSEric Sandeen } 1182ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) 119e29d1cdeSEric Sandeen return bh; 120e29d1cdeSEric Sandeen 121c806e68fSFrederic Bohe lock_buffer(bh); 1222ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) { 1232ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1242ccb5fb9SAneesh Kumar K.V return bh; 1252ccb5fb9SAneesh Kumar K.V } 126955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 127e29d1cdeSEric Sandeen if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 128e29d1cdeSEric Sandeen ext4_init_inode_bitmap(sb, bh, block_group, desc); 1292ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 130e29d1cdeSEric Sandeen set_buffer_uptodate(bh); 131955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1323300bedaSAneesh Kumar K.V unlock_buffer(bh); 133e29d1cdeSEric Sandeen return bh; 134e29d1cdeSEric Sandeen } 135955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1362ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh)) { 1372ccb5fb9SAneesh Kumar K.V /* 1382ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 1392ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 1402ccb5fb9SAneesh Kumar K.V */ 1412ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 1422ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 1432ccb5fb9SAneesh Kumar K.V return bh; 1442ccb5fb9SAneesh Kumar K.V } 1452ccb5fb9SAneesh Kumar K.V /* 1462ccb5fb9SAneesh Kumar K.V * submit the buffer_head for read. We can 1472ccb5fb9SAneesh Kumar K.V * safely mark the bitmap as uptodate now. 1482ccb5fb9SAneesh Kumar K.V * We do it here so the bitmap uptodate bit 1492ccb5fb9SAneesh Kumar K.V * get set with buffer lock held. 1502ccb5fb9SAneesh Kumar K.V */ 1512ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 152e29d1cdeSEric Sandeen if (bh_submit_read(bh) < 0) { 153e29d1cdeSEric Sandeen put_bh(bh); 15412062dddSEric Sandeen ext4_error(sb, "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); 2119bffad1eSTheodore Ts'o trace_ext4_free_inode(inode); 212ac27a0ecSDave Kleikamp 213ac27a0ecSDave Kleikamp /* 214ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 215ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 216ac27a0ecSDave Kleikamp */ 217871a2931SChristoph Hellwig dquot_initialize(inode); 218617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 21963936ddaSChristoph Hellwig dquot_free_inode(inode); 2209f754758SChristoph Hellwig dquot_drop(inode); 221ac27a0ecSDave Kleikamp 222ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 223ac27a0ecSDave Kleikamp 224ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 2250930fcc1SAl Viro ext4_clear_inode(inode); 226ac27a0ecSDave Kleikamp 227617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 228617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 22912062dddSEric Sandeen ext4_error(sb, "reserved or nonexistent inode %lu", ino); 230ac27a0ecSDave Kleikamp goto error_return; 231ac27a0ecSDave Kleikamp } 232617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 233617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 234e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 235ac27a0ecSDave Kleikamp if (!bitmap_bh) 236ac27a0ecSDave Kleikamp goto error_return; 237ac27a0ecSDave Kleikamp 238ac27a0ecSDave Kleikamp BUFFER_TRACE(bitmap_bh, "get_write_access"); 239617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bitmap_bh); 240ac27a0ecSDave Kleikamp if (fatal) 241ac27a0ecSDave Kleikamp goto error_return; 242ac27a0ecSDave Kleikamp 243d17413c0SDmitry Monakhov fatal = -ESRCH; 244617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 245d17413c0SDmitry Monakhov if (gdp) { 246ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 247617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 248d17413c0SDmitry Monakhov } 249955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 250d17413c0SDmitry Monakhov cleared = ext4_clear_bit(bit, bitmap_bh->b_data); 251d17413c0SDmitry Monakhov if (fatal || !cleared) { 252d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 253d17413c0SDmitry Monakhov goto out; 254d17413c0SDmitry Monakhov } 255d17413c0SDmitry Monakhov 256560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 257560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 258560671a0SAneesh Kumar K.V if (is_directory) { 259560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 260560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 261d17413c0SDmitry Monakhov percpu_counter_dec(&sbi->s_dirs_counter); 262d17413c0SDmitry Monakhov } 263d17413c0SDmitry Monakhov gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 264d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 2657d39db14STheodore Ts'o 266d17413c0SDmitry Monakhov percpu_counter_inc(&sbi->s_freeinodes_counter); 267d17413c0SDmitry Monakhov if (sbi->s_log_groups_per_flex) { 268d17413c0SDmitry Monakhov ext4_group_t f = ext4_flex_group(sbi, block_group); 269d17413c0SDmitry Monakhov 270d17413c0SDmitry Monakhov atomic_inc(&sbi->s_flex_groups[f].free_inodes); 271d17413c0SDmitry Monakhov if (is_directory) 272c4caae25SEric Sandeen atomic_dec(&sbi->s_flex_groups[f].used_dirs); 2737d39db14STheodore Ts'o } 2740390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 275d17413c0SDmitry Monakhov fatal = ext4_handle_dirty_metadata(handle, NULL, bh2); 276d17413c0SDmitry Monakhov out: 277d17413c0SDmitry Monakhov if (cleared) { 2780390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 2790390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 280ac27a0ecSDave Kleikamp if (!fatal) 281ac27a0ecSDave Kleikamp fatal = err; 282ac27a0ecSDave Kleikamp sb->s_dirt = 1; 283d17413c0SDmitry Monakhov } else 284d17413c0SDmitry Monakhov ext4_error(sb, "bit already cleared for inode %lu", ino); 285d17413c0SDmitry Monakhov 286ac27a0ecSDave Kleikamp error_return: 287ac27a0ecSDave Kleikamp brelse(bitmap_bh); 288617ba13bSMingming Cao ext4_std_error(sb, fatal); 289ac27a0ecSDave Kleikamp } 290ac27a0ecSDave Kleikamp 291ac27a0ecSDave Kleikamp /* 292ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 293ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 294ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 295ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 296ac27a0ecSDave Kleikamp * directories already is chosen. 297ac27a0ecSDave Kleikamp * 298ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory\'s block 299ac27a0ecSDave Kleikamp * group to find a free inode. 300ac27a0ecSDave Kleikamp */ 3012aa9fc4cSAvantika Mathur static int find_group_dir(struct super_block *sb, struct inode *parent, 3022aa9fc4cSAvantika Mathur ext4_group_t *best_group) 303ac27a0ecSDave Kleikamp { 3048df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 305ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 306617ba13bSMingming Cao struct ext4_group_desc *desc, *best_desc = NULL; 3072aa9fc4cSAvantika Mathur ext4_group_t group; 3082aa9fc4cSAvantika Mathur int ret = -1; 309ac27a0ecSDave Kleikamp 310617ba13bSMingming Cao freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter); 311ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 312ac27a0ecSDave Kleikamp 313ac27a0ecSDave Kleikamp for (group = 0; group < ngroups; group++) { 314ef2fb679SEric Sandeen desc = ext4_get_group_desc(sb, group, NULL); 315560671a0SAneesh Kumar K.V if (!desc || !ext4_free_inodes_count(sb, desc)) 316ac27a0ecSDave Kleikamp continue; 317560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc) < avefreei) 318ac27a0ecSDave Kleikamp continue; 319ac27a0ecSDave Kleikamp if (!best_desc || 320560671a0SAneesh Kumar K.V (ext4_free_blks_count(sb, desc) > 321560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, best_desc))) { 3222aa9fc4cSAvantika Mathur *best_group = group; 323ac27a0ecSDave Kleikamp best_desc = desc; 3242aa9fc4cSAvantika Mathur ret = 0; 325ac27a0ecSDave Kleikamp } 326ac27a0ecSDave Kleikamp } 3272aa9fc4cSAvantika Mathur return ret; 328ac27a0ecSDave Kleikamp } 329ac27a0ecSDave Kleikamp 330772cb7c8SJose R. Santos #define free_block_ratio 10 331772cb7c8SJose R. Santos 332772cb7c8SJose R. Santos static int find_group_flex(struct super_block *sb, struct inode *parent, 333772cb7c8SJose R. Santos ext4_group_t *best_group) 334772cb7c8SJose R. Santos { 335772cb7c8SJose R. Santos struct ext4_sb_info *sbi = EXT4_SB(sb); 336772cb7c8SJose R. Santos struct ext4_group_desc *desc; 337772cb7c8SJose R. Santos struct flex_groups *flex_group = sbi->s_flex_groups; 338772cb7c8SJose R. Santos ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 339772cb7c8SJose R. Santos ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group); 3408df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 341772cb7c8SJose R. Santos int flex_size = ext4_flex_bg_size(sbi); 342772cb7c8SJose R. Santos ext4_group_t best_flex = parent_fbg_group; 343772cb7c8SJose R. Santos int blocks_per_flex = sbi->s_blocks_per_group * flex_size; 344772cb7c8SJose R. Santos int flexbg_free_blocks; 345772cb7c8SJose R. Santos int flex_freeb_ratio; 346772cb7c8SJose R. Santos ext4_group_t n_fbg_groups; 347772cb7c8SJose R. Santos ext4_group_t i; 348772cb7c8SJose R. Santos 3498df9675fSTheodore Ts'o n_fbg_groups = (ngroups + flex_size - 1) >> 350772cb7c8SJose R. Santos sbi->s_log_groups_per_flex; 351772cb7c8SJose R. Santos 352772cb7c8SJose R. Santos find_close_to_parent: 3539f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[best_flex].free_blocks); 354772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 3559f24e420STheodore Ts'o if (atomic_read(&flex_group[best_flex].free_inodes) && 356772cb7c8SJose R. Santos flex_freeb_ratio > free_block_ratio) 357772cb7c8SJose R. Santos goto found_flexbg; 358772cb7c8SJose R. Santos 359772cb7c8SJose R. Santos if (best_flex && best_flex == parent_fbg_group) { 360772cb7c8SJose R. Santos best_flex--; 361772cb7c8SJose R. Santos goto find_close_to_parent; 362772cb7c8SJose R. Santos } 363772cb7c8SJose R. Santos 364772cb7c8SJose R. Santos for (i = 0; i < n_fbg_groups; i++) { 365772cb7c8SJose R. Santos if (i == parent_fbg_group || i == parent_fbg_group - 1) 366772cb7c8SJose R. Santos continue; 367772cb7c8SJose R. Santos 3689f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[i].free_blocks); 369772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 370772cb7c8SJose R. Santos 371772cb7c8SJose R. Santos if (flex_freeb_ratio > free_block_ratio && 3729f24e420STheodore Ts'o (atomic_read(&flex_group[i].free_inodes))) { 373772cb7c8SJose R. Santos best_flex = i; 374772cb7c8SJose R. Santos goto found_flexbg; 375772cb7c8SJose R. Santos } 376772cb7c8SJose R. Santos 3779f24e420STheodore Ts'o if ((atomic_read(&flex_group[best_flex].free_inodes) == 0) || 3789f24e420STheodore Ts'o ((atomic_read(&flex_group[i].free_blocks) > 3799f24e420STheodore Ts'o atomic_read(&flex_group[best_flex].free_blocks)) && 3809f24e420STheodore Ts'o atomic_read(&flex_group[i].free_inodes))) 381772cb7c8SJose R. Santos best_flex = i; 382772cb7c8SJose R. Santos } 383772cb7c8SJose R. Santos 3849f24e420STheodore Ts'o if (!atomic_read(&flex_group[best_flex].free_inodes) || 3859f24e420STheodore Ts'o !atomic_read(&flex_group[best_flex].free_blocks)) 386772cb7c8SJose R. Santos return -1; 387772cb7c8SJose R. Santos 388772cb7c8SJose R. Santos found_flexbg: 389772cb7c8SJose R. Santos for (i = best_flex * flex_size; i < ngroups && 390772cb7c8SJose R. Santos i < (best_flex + 1) * flex_size; i++) { 39188b6edd1STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 392560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc)) { 393772cb7c8SJose R. Santos *best_group = i; 394772cb7c8SJose R. Santos goto out; 395772cb7c8SJose R. Santos } 396772cb7c8SJose R. Santos } 397772cb7c8SJose R. Santos 398772cb7c8SJose R. Santos return -1; 399772cb7c8SJose R. Santos out: 400772cb7c8SJose R. Santos return 0; 401772cb7c8SJose R. Santos } 402772cb7c8SJose R. Santos 403a4912123STheodore Ts'o struct orlov_stats { 404a4912123STheodore Ts'o __u32 free_inodes; 405a4912123STheodore Ts'o __u32 free_blocks; 406a4912123STheodore Ts'o __u32 used_dirs; 407a4912123STheodore Ts'o }; 408a4912123STheodore Ts'o 409a4912123STheodore Ts'o /* 410a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 411a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 412a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 413a4912123STheodore Ts'o */ 414a4912123STheodore Ts'o void get_orlov_stats(struct super_block *sb, ext4_group_t g, 415a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 416a4912123STheodore Ts'o { 417a4912123STheodore Ts'o struct ext4_group_desc *desc; 4187d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 419a4912123STheodore Ts'o 4207d39db14STheodore Ts'o if (flex_size > 1) { 4217d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 4227d39db14STheodore Ts'o stats->free_blocks = atomic_read(&flex_group[g].free_blocks); 4237d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 4247d39db14STheodore Ts'o return; 4257d39db14STheodore Ts'o } 4267d39db14STheodore Ts'o 4277d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 4287d39db14STheodore Ts'o if (desc) { 4297d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 4307d39db14STheodore Ts'o stats->free_blocks = ext4_free_blks_count(sb, desc); 4317d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 4327d39db14STheodore Ts'o } else { 433a4912123STheodore Ts'o stats->free_inodes = 0; 434a4912123STheodore Ts'o stats->free_blocks = 0; 435a4912123STheodore Ts'o stats->used_dirs = 0; 436a4912123STheodore Ts'o } 437a4912123STheodore Ts'o } 438a4912123STheodore Ts'o 439ac27a0ecSDave Kleikamp /* 440ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 441ac27a0ecSDave Kleikamp * 442ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 443ac27a0ecSDave Kleikamp * 444ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 445ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 446ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 447ac27a0ecSDave Kleikamp * 448ac27a0ecSDave Kleikamp * For the rest rules look so: 449ac27a0ecSDave Kleikamp * 450ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 451ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 452ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 453ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4541cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 455ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 456ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 457ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 458ac27a0ecSDave Kleikamp */ 459ac27a0ecSDave Kleikamp 4602aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 461f157a4aaSTheodore Ts'o ext4_group_t *group, int mode, 462f157a4aaSTheodore Ts'o const struct qstr *qstr) 463ac27a0ecSDave Kleikamp { 464fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 465617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 4668df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 467617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 468ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 469617ba13bSMingming Cao ext4_fsblk_t freeb, avefreeb; 470ac27a0ecSDave Kleikamp unsigned int ndirs; 471a4912123STheodore Ts'o int max_dirs, min_inodes; 472617ba13bSMingming Cao ext4_grpblk_t min_blocks; 4738df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 474617ba13bSMingming Cao struct ext4_group_desc *desc; 475a4912123STheodore Ts'o struct orlov_stats stats; 476a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 477f157a4aaSTheodore Ts'o struct dx_hash_info hinfo; 478a4912123STheodore Ts'o 4798df9675fSTheodore Ts'o ngroups = real_ngroups; 480a4912123STheodore Ts'o if (flex_size > 1) { 4818df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 482a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 483a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 484a4912123STheodore Ts'o } 485ac27a0ecSDave Kleikamp 486ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 487ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 488ac27a0ecSDave Kleikamp freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter); 4893a5b2ecdSMingming Cao avefreeb = freeb; 490f4e5bc24SAndrew Morton do_div(avefreeb, ngroups); 491ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 492ac27a0ecSDave Kleikamp 493a4912123STheodore Ts'o if (S_ISDIR(mode) && 494a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 49512e9b892SDmitry Monakhov (ext4_test_inode_flag(parent, EXT4_INODE_TOPDIR)))) { 496ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 4972aa9fc4cSAvantika Mathur int ret = -1; 498ac27a0ecSDave Kleikamp 499f157a4aaSTheodore Ts'o if (qstr) { 500f157a4aaSTheodore Ts'o hinfo.hash_version = DX_HASH_HALF_MD4; 501f157a4aaSTheodore Ts'o hinfo.seed = sbi->s_hash_seed; 502f157a4aaSTheodore Ts'o ext4fs_dirhash(qstr->name, qstr->len, &hinfo); 503f157a4aaSTheodore Ts'o grp = hinfo.hash; 504f157a4aaSTheodore Ts'o } else 5052aa9fc4cSAvantika Mathur get_random_bytes(&grp, sizeof(grp)); 5062aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 507ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 508a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 509a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 510a4912123STheodore Ts'o if (!stats.free_inodes) 511ac27a0ecSDave Kleikamp continue; 512a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 513ac27a0ecSDave Kleikamp continue; 514a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 515ac27a0ecSDave Kleikamp continue; 516a4912123STheodore Ts'o if (stats.free_blocks < avefreeb) 517ac27a0ecSDave Kleikamp continue; 518a4912123STheodore Ts'o grp = g; 5192aa9fc4cSAvantika Mathur ret = 0; 520a4912123STheodore Ts'o best_ndir = stats.used_dirs; 521ac27a0ecSDave Kleikamp } 522a4912123STheodore Ts'o if (ret) 523a4912123STheodore Ts'o goto fallback; 524a4912123STheodore Ts'o found_flex_bg: 525a4912123STheodore Ts'o if (flex_size == 1) { 526a4912123STheodore Ts'o *group = grp; 527a4912123STheodore Ts'o return 0; 528a4912123STheodore Ts'o } 529a4912123STheodore Ts'o 530a4912123STheodore Ts'o /* 531a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 532a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 533a4912123STheodore Ts'o * something similar, although regular files will 534a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 535a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 536a4912123STheodore Ts'o */ 537a4912123STheodore Ts'o grp *= flex_size; 538a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 5398df9675fSTheodore Ts'o if (grp+i >= real_ngroups) 540a4912123STheodore Ts'o break; 541a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 542a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 543a4912123STheodore Ts'o *group = grp+i; 544a4912123STheodore Ts'o return 0; 545a4912123STheodore Ts'o } 546a4912123STheodore Ts'o } 547ac27a0ecSDave Kleikamp goto fallback; 548ac27a0ecSDave Kleikamp } 549ac27a0ecSDave Kleikamp 550ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 551a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 552a4912123STheodore Ts'o if (min_inodes < 1) 553a4912123STheodore Ts'o min_inodes = 1; 554a4912123STheodore Ts'o min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb)*flex_size / 4; 555ac27a0ecSDave Kleikamp 556a4912123STheodore Ts'o /* 557a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 558a4912123STheodore Ts'o * inode for this parent directory 559a4912123STheodore Ts'o */ 560a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 561a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 562a4912123STheodore Ts'o if (flex_size > 1) 563a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 564a4912123STheodore Ts'o } 565ac27a0ecSDave Kleikamp 566ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 567a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 568a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 569a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 570ac27a0ecSDave Kleikamp continue; 571a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 572ac27a0ecSDave Kleikamp continue; 573a4912123STheodore Ts'o if (stats.free_blocks < min_blocks) 574ac27a0ecSDave Kleikamp continue; 575a4912123STheodore Ts'o goto found_flex_bg; 576ac27a0ecSDave Kleikamp } 577ac27a0ecSDave Kleikamp 578ac27a0ecSDave Kleikamp fallback: 5798df9675fSTheodore Ts'o ngroups = real_ngroups; 580a4912123STheodore Ts'o avefreei = freei / ngroups; 581b5451f7bSTheodore Ts'o fallback_retry: 582a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 583ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 584a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 585a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 586560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 587a4912123STheodore Ts'o ext4_free_inodes_count(sb, desc) >= avefreei) { 588a4912123STheodore Ts'o *group = grp; 5892aa9fc4cSAvantika Mathur return 0; 590ac27a0ecSDave Kleikamp } 591a4912123STheodore Ts'o } 592ac27a0ecSDave Kleikamp 593ac27a0ecSDave Kleikamp if (avefreei) { 594ac27a0ecSDave Kleikamp /* 595ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 596ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 597ac27a0ecSDave Kleikamp */ 598ac27a0ecSDave Kleikamp avefreei = 0; 599b5451f7bSTheodore Ts'o goto fallback_retry; 600ac27a0ecSDave Kleikamp } 601ac27a0ecSDave Kleikamp 602ac27a0ecSDave Kleikamp return -1; 603ac27a0ecSDave Kleikamp } 604ac27a0ecSDave Kleikamp 6052aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 606a4912123STheodore Ts'o ext4_group_t *group, int mode) 607ac27a0ecSDave Kleikamp { 608fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 6098df9675fSTheodore Ts'o ext4_group_t i, last, ngroups = ext4_get_groups_count(sb); 610617ba13bSMingming Cao struct ext4_group_desc *desc; 611a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 612a4912123STheodore Ts'o 613a4912123STheodore Ts'o /* 614a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 615a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 616a4912123STheodore Ts'o * find another flex group, and store that information in the 617a4912123STheodore Ts'o * parent directory's inode information so that use that flex 618a4912123STheodore Ts'o * group for future allocations. 619a4912123STheodore Ts'o */ 620a4912123STheodore Ts'o if (flex_size > 1) { 621a4912123STheodore Ts'o int retry = 0; 622a4912123STheodore Ts'o 623a4912123STheodore Ts'o try_again: 624a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 625a4912123STheodore Ts'o last = parent_group + flex_size; 626a4912123STheodore Ts'o if (last > ngroups) 627a4912123STheodore Ts'o last = ngroups; 628a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 629a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 630a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 631a4912123STheodore Ts'o *group = i; 632a4912123STheodore Ts'o return 0; 633a4912123STheodore Ts'o } 634a4912123STheodore Ts'o } 635a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 636a4912123STheodore Ts'o retry = 1; 637a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 638a4912123STheodore Ts'o goto try_again; 639a4912123STheodore Ts'o } 640a4912123STheodore Ts'o /* 641a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 642a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 643a4912123STheodore Ts'o * avoid the topdir algorithms. 644a4912123STheodore Ts'o */ 645a4912123STheodore Ts'o *group = parent_group + flex_size; 646a4912123STheodore Ts'o if (*group > ngroups) 647a4912123STheodore Ts'o *group = 0; 648f157a4aaSTheodore Ts'o return find_group_orlov(sb, parent, group, mode, 0); 649a4912123STheodore Ts'o } 650ac27a0ecSDave Kleikamp 651ac27a0ecSDave Kleikamp /* 652ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 653ac27a0ecSDave Kleikamp */ 6542aa9fc4cSAvantika Mathur *group = parent_group; 6552aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 656560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 657560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6582aa9fc4cSAvantika Mathur return 0; 659ac27a0ecSDave Kleikamp 660ac27a0ecSDave Kleikamp /* 661ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 662ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 663ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 664ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 665ac27a0ecSDave Kleikamp * different blockgroup. 666ac27a0ecSDave Kleikamp * 667ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 668ac27a0ecSDave Kleikamp */ 6692aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 670ac27a0ecSDave Kleikamp 671ac27a0ecSDave Kleikamp /* 672ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 673ac27a0ecSDave Kleikamp * blocks. 674ac27a0ecSDave Kleikamp */ 675ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6762aa9fc4cSAvantika Mathur *group += i; 6772aa9fc4cSAvantika Mathur if (*group >= ngroups) 6782aa9fc4cSAvantika Mathur *group -= ngroups; 6792aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 680560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 681560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6822aa9fc4cSAvantika Mathur return 0; 683ac27a0ecSDave Kleikamp } 684ac27a0ecSDave Kleikamp 685ac27a0ecSDave Kleikamp /* 686ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 687ac27a0ecSDave Kleikamp * has no free blocks. 688ac27a0ecSDave Kleikamp */ 6892aa9fc4cSAvantika Mathur *group = parent_group; 690ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6912aa9fc4cSAvantika Mathur if (++*group >= ngroups) 6922aa9fc4cSAvantika Mathur *group = 0; 6932aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 694560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 6952aa9fc4cSAvantika Mathur return 0; 696ac27a0ecSDave Kleikamp } 697ac27a0ecSDave Kleikamp 698ac27a0ecSDave Kleikamp return -1; 699ac27a0ecSDave Kleikamp } 700ac27a0ecSDave Kleikamp 701ac27a0ecSDave Kleikamp /* 70239341867SAneesh Kumar K.V * claim the inode from the inode bitmap. If the group 703955ce5f5SAneesh Kumar K.V * is uninit we need to take the groups's ext4_group_lock 70439341867SAneesh Kumar K.V * and clear the uninit flag. The inode bitmap update 70539341867SAneesh Kumar K.V * and group desc uninit flag clear should be done 706955ce5f5SAneesh Kumar K.V * after holding ext4_group_lock so that ext4_read_inode_bitmap 70739341867SAneesh Kumar K.V * doesn't race with the ext4_claim_inode 70839341867SAneesh Kumar K.V */ 70939341867SAneesh Kumar K.V static int ext4_claim_inode(struct super_block *sb, 71039341867SAneesh Kumar K.V struct buffer_head *inode_bitmap_bh, 71139341867SAneesh Kumar K.V unsigned long ino, ext4_group_t group, int mode) 71239341867SAneesh Kumar K.V { 71339341867SAneesh Kumar K.V int free = 0, retval = 0, count; 71439341867SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 71539341867SAneesh Kumar K.V struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL); 71639341867SAneesh Kumar K.V 717955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 71839341867SAneesh Kumar K.V if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) { 71939341867SAneesh Kumar K.V /* not a free inode */ 72039341867SAneesh Kumar K.V retval = 1; 72139341867SAneesh Kumar K.V goto err_ret; 72239341867SAneesh Kumar K.V } 72339341867SAneesh Kumar K.V ino++; 72439341867SAneesh Kumar K.V if ((group == 0 && ino < EXT4_FIRST_INO(sb)) || 72539341867SAneesh Kumar K.V ino > EXT4_INODES_PER_GROUP(sb)) { 726955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 72712062dddSEric Sandeen ext4_error(sb, "reserved inode or inode > inodes count - " 72839341867SAneesh Kumar K.V "block_group = %u, inode=%lu", group, 72939341867SAneesh Kumar K.V ino + group * EXT4_INODES_PER_GROUP(sb)); 73039341867SAneesh Kumar K.V return 1; 73139341867SAneesh Kumar K.V } 73239341867SAneesh Kumar K.V /* If we didn't allocate from within the initialized part of the inode 73339341867SAneesh Kumar K.V * table then we need to initialize up to this inode. */ 73439341867SAneesh Kumar K.V if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) { 73539341867SAneesh Kumar K.V 73639341867SAneesh Kumar K.V if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 73739341867SAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 73839341867SAneesh Kumar K.V /* When marking the block group with 73939341867SAneesh Kumar K.V * ~EXT4_BG_INODE_UNINIT we don't want to depend 74039341867SAneesh Kumar K.V * on the value of bg_itable_unused even though 74139341867SAneesh Kumar K.V * mke2fs could have initialized the same for us. 74239341867SAneesh Kumar K.V * Instead we calculated the value below 74339341867SAneesh Kumar K.V */ 74439341867SAneesh Kumar K.V 74539341867SAneesh Kumar K.V free = 0; 74639341867SAneesh Kumar K.V } else { 74739341867SAneesh Kumar K.V free = EXT4_INODES_PER_GROUP(sb) - 74839341867SAneesh Kumar K.V ext4_itable_unused_count(sb, gdp); 74939341867SAneesh Kumar K.V } 75039341867SAneesh Kumar K.V 75139341867SAneesh Kumar K.V /* 75239341867SAneesh Kumar K.V * Check the relative inode number against the last used 75339341867SAneesh Kumar K.V * relative inode number in this group. if it is greater 75439341867SAneesh Kumar K.V * we need to update the bg_itable_unused count 75539341867SAneesh Kumar K.V * 75639341867SAneesh Kumar K.V */ 75739341867SAneesh Kumar K.V if (ino > free) 75839341867SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 75939341867SAneesh Kumar K.V (EXT4_INODES_PER_GROUP(sb) - ino)); 76039341867SAneesh Kumar K.V } 76139341867SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) - 1; 76239341867SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 76339341867SAneesh Kumar K.V if (S_ISDIR(mode)) { 76439341867SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) + 1; 76539341867SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 7667d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 7677d39db14STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 7687d39db14STheodore Ts'o 769c4caae25SEric Sandeen atomic_inc(&sbi->s_flex_groups[f].used_dirs); 7707d39db14STheodore Ts'o } 77139341867SAneesh Kumar K.V } 77239341867SAneesh Kumar K.V gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 77339341867SAneesh Kumar K.V err_ret: 774955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 77539341867SAneesh Kumar K.V return retval; 77639341867SAneesh Kumar K.V } 77739341867SAneesh Kumar K.V 77839341867SAneesh Kumar K.V /* 779ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 780ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 781ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 782ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 783ac27a0ecSDave Kleikamp * directories already is chosen. 784ac27a0ecSDave Kleikamp * 785ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 786ac27a0ecSDave Kleikamp * group to find a free inode. 787ac27a0ecSDave Kleikamp */ 788f157a4aaSTheodore Ts'o struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode, 78911013911SAndreas Dilger const struct qstr *qstr, __u32 goal) 790ac27a0ecSDave Kleikamp { 791ac27a0ecSDave Kleikamp struct super_block *sb; 7923300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 7933300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 7948df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 795ac27a0ecSDave Kleikamp unsigned long ino = 0; 796ac27a0ecSDave Kleikamp struct inode *inode; 797617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 798617ba13bSMingming Cao struct ext4_inode_info *ei; 799617ba13bSMingming Cao struct ext4_sb_info *sbi; 80039341867SAneesh Kumar K.V int ret2, err = 0; 801ac27a0ecSDave Kleikamp struct inode *ret; 8022aa9fc4cSAvantika Mathur ext4_group_t i; 8032aa9fc4cSAvantika Mathur int free = 0; 8042842c3b5STheodore Ts'o static int once = 1; 805772cb7c8SJose R. Santos ext4_group_t flex_group; 806ac27a0ecSDave Kleikamp 807ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 808ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 809ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 810ac27a0ecSDave Kleikamp 811ac27a0ecSDave Kleikamp sb = dir->i_sb; 8128df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 8139bffad1eSTheodore Ts'o trace_ext4_request_inode(dir, mode); 814ac27a0ecSDave Kleikamp inode = new_inode(sb); 815ac27a0ecSDave Kleikamp if (!inode) 816ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 817617ba13bSMingming Cao ei = EXT4_I(inode); 818617ba13bSMingming Cao sbi = EXT4_SB(sb); 819772cb7c8SJose R. Santos 82011013911SAndreas Dilger if (!goal) 82111013911SAndreas Dilger goal = sbi->s_inode_goal; 82211013911SAndreas Dilger 823e6462869SJohann Lombardi if (goal && goal <= le32_to_cpu(sbi->s_es->s_inodes_count)) { 82411013911SAndreas Dilger group = (goal - 1) / EXT4_INODES_PER_GROUP(sb); 82511013911SAndreas Dilger ino = (goal - 1) % EXT4_INODES_PER_GROUP(sb); 82611013911SAndreas Dilger ret2 = 0; 82711013911SAndreas Dilger goto got_group; 82811013911SAndreas Dilger } 82911013911SAndreas Dilger 830a4912123STheodore Ts'o if (sbi->s_log_groups_per_flex && test_opt(sb, OLDALLOC)) { 831772cb7c8SJose R. Santos ret2 = find_group_flex(sb, dir, &group); 83205bf9e83STheodore Ts'o if (ret2 == -1) { 833a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 8346b82f3cbSChuck Ebbert if (ret2 == 0 && once) { 8352842c3b5STheodore Ts'o once = 0; 83605bf9e83STheodore Ts'o printk(KERN_NOTICE "ext4: find_group_flex " 83705bf9e83STheodore Ts'o "failed, fallback succeeded dir %lu\n", 83805bf9e83STheodore Ts'o dir->i_ino); 83905bf9e83STheodore Ts'o } 8406b82f3cbSChuck Ebbert } 841772cb7c8SJose R. Santos goto got_group; 842772cb7c8SJose R. Santos } 843772cb7c8SJose R. Santos 844ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) { 845ac27a0ecSDave Kleikamp if (test_opt(sb, OLDALLOC)) 8462aa9fc4cSAvantika Mathur ret2 = find_group_dir(sb, dir, &group); 847ac27a0ecSDave Kleikamp else 848f157a4aaSTheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode, qstr); 849ac27a0ecSDave Kleikamp } else 850a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 851ac27a0ecSDave Kleikamp 852772cb7c8SJose R. Santos got_group: 853a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 854ac27a0ecSDave Kleikamp err = -ENOSPC; 8552aa9fc4cSAvantika Mathur if (ret2 == -1) 856ac27a0ecSDave Kleikamp goto out; 857ac27a0ecSDave Kleikamp 85811013911SAndreas Dilger for (i = 0; i < ngroups; i++, ino = 0) { 859ac27a0ecSDave Kleikamp err = -EIO; 860ac27a0ecSDave Kleikamp 8613300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 862ac27a0ecSDave Kleikamp if (!gdp) 863ac27a0ecSDave Kleikamp goto fail; 864ac27a0ecSDave Kleikamp 8653300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 8663300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 8673300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 868ac27a0ecSDave Kleikamp goto fail; 869ac27a0ecSDave Kleikamp 870ac27a0ecSDave Kleikamp repeat_in_this_group: 871617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 8723300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 8733300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 8743300bedaSAneesh Kumar K.V 875617ba13bSMingming Cao if (ino < EXT4_INODES_PER_GROUP(sb)) { 876ac27a0ecSDave Kleikamp 8773300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 8783300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 8793300bedaSAneesh Kumar K.V inode_bitmap_bh); 880ac27a0ecSDave Kleikamp if (err) 881ac27a0ecSDave Kleikamp goto fail; 882ac27a0ecSDave Kleikamp 88339341867SAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "get_write_access"); 88439341867SAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 88539341867SAneesh Kumar K.V group_desc_bh); 88639341867SAneesh Kumar K.V if (err) 88739341867SAneesh Kumar K.V goto fail; 88839341867SAneesh Kumar K.V if (!ext4_claim_inode(sb, inode_bitmap_bh, 88939341867SAneesh Kumar K.V ino, group, mode)) { 890ac27a0ecSDave Kleikamp /* we won it */ 8913300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, 8920390131bSFrank Mayhar "call ext4_handle_dirty_metadata"); 8930390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 89473b50c1cSCurt Wohlgemuth NULL, 8953300bedaSAneesh Kumar K.V inode_bitmap_bh); 896ac27a0ecSDave Kleikamp if (err) 897ac27a0ecSDave Kleikamp goto fail; 89839341867SAneesh Kumar K.V /* zero bit is inode number 1*/ 89939341867SAneesh Kumar K.V ino++; 900ac27a0ecSDave Kleikamp goto got; 901ac27a0ecSDave Kleikamp } 902ac27a0ecSDave Kleikamp /* we lost it */ 9033300bedaSAneesh Kumar K.V ext4_handle_release_buffer(handle, inode_bitmap_bh); 90439341867SAneesh Kumar K.V ext4_handle_release_buffer(handle, group_desc_bh); 905ac27a0ecSDave Kleikamp 906617ba13bSMingming Cao if (++ino < EXT4_INODES_PER_GROUP(sb)) 907ac27a0ecSDave Kleikamp goto repeat_in_this_group; 908ac27a0ecSDave Kleikamp } 909ac27a0ecSDave Kleikamp 910ac27a0ecSDave Kleikamp /* 911ac27a0ecSDave Kleikamp * This case is possible in concurrent environment. It is very 912ac27a0ecSDave Kleikamp * rare. We cannot repeat the find_group_xxx() call because 913ac27a0ecSDave Kleikamp * that will simply return the same blockgroup, because the 914ac27a0ecSDave Kleikamp * group descriptor metadata has not yet been updated. 915ac27a0ecSDave Kleikamp * So we just go onto the next blockgroup. 916ac27a0ecSDave Kleikamp */ 9178df9675fSTheodore Ts'o if (++group == ngroups) 918ac27a0ecSDave Kleikamp group = 0; 919ac27a0ecSDave Kleikamp } 920ac27a0ecSDave Kleikamp err = -ENOSPC; 921ac27a0ecSDave Kleikamp goto out; 922ac27a0ecSDave Kleikamp 923ac27a0ecSDave Kleikamp got: 924717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 925717d50e4SAndreas Dilger if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) && 926717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9273300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 928717d50e4SAndreas Dilger 9293300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 9303300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 9313300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 932717d50e4SAndreas Dilger if (err) { 9333300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 934717d50e4SAndreas Dilger goto fail; 935717d50e4SAndreas Dilger } 936717d50e4SAndreas Dilger 937717d50e4SAndreas Dilger free = 0; 938955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, group); 939717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 940717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 941717d50e4SAndreas Dilger free = ext4_free_blocks_after_init(sb, group, gdp); 9423300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 943560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, free); 94423712a9cSFrederic Bohe gdp->bg_checksum = ext4_group_desc_csum(sbi, group, 94523712a9cSFrederic Bohe gdp); 946717d50e4SAndreas Dilger } 947955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 948717d50e4SAndreas Dilger 949717d50e4SAndreas Dilger /* Don't need to dirty bitmap block if we didn't change it */ 950717d50e4SAndreas Dilger if (free) { 9513300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 9520390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 9533300bedaSAneesh Kumar K.V NULL, block_bitmap_bh); 954717d50e4SAndreas Dilger } 955717d50e4SAndreas Dilger 9563300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 957717d50e4SAndreas Dilger if (err) 958717d50e4SAndreas Dilger goto fail; 959717d50e4SAndreas Dilger } 9603300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 9613300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 96239341867SAneesh Kumar K.V if (err) 96339341867SAneesh Kumar K.V goto fail; 964ac27a0ecSDave Kleikamp 965ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 966ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 967ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 968ac27a0ecSDave Kleikamp sb->s_dirt = 1; 969ac27a0ecSDave Kleikamp 970772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 971772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 9729f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 973772cb7c8SJose R. Santos } 974772cb7c8SJose R. Santos 975b10b8520SDmitry Monakhov if (test_opt(sb, GRPID)) { 976ac27a0ecSDave Kleikamp inode->i_mode = mode; 977b10b8520SDmitry Monakhov inode->i_uid = current_fsuid(); 978b10b8520SDmitry Monakhov inode->i_gid = dir->i_gid; 979b10b8520SDmitry Monakhov } else 980b10b8520SDmitry Monakhov inode_init_owner(inode, dir, 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 101519f5fb7aSTheodore Ts'o ei->i_state_flags = 0; 101619f5fb7aSTheodore Ts'o ext4_set_inode_state(inode, EXT4_STATE_NEW); 1017ef7f3835SKalpak Shah 1018ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 1019ac27a0ecSDave Kleikamp 1020ac27a0ecSDave Kleikamp ret = inode; 1021871a2931SChristoph Hellwig dquot_initialize(inode); 102263936ddaSChristoph Hellwig err = dquot_alloc_inode(inode); 102363936ddaSChristoph Hellwig if (err) 1024ac27a0ecSDave Kleikamp goto fail_drop; 1025ac27a0ecSDave Kleikamp 1026617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1027ac27a0ecSDave Kleikamp if (err) 1028ac27a0ecSDave Kleikamp goto fail_free_drop; 1029ac27a0ecSDave Kleikamp 1030617ba13bSMingming Cao err = ext4_init_security(handle, inode, dir); 1031ac27a0ecSDave Kleikamp if (err) 1032ac27a0ecSDave Kleikamp goto fail_free_drop; 1033ac27a0ecSDave Kleikamp 103483982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 1035e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1036e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 103712e9b892SDmitry Monakhov ext4_set_inode_flag(inode, EXT4_INODE_EXTENTS); 1038a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1039a86c6181SAlex Tomas } 104042bf0383SAneesh Kumar K.V } 1041ac27a0ecSDave Kleikamp 10428753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 10438753e88fSAneesh Kumar K.V if (err) { 10448753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 10458753e88fSAneesh Kumar K.V goto fail_free_drop; 10468753e88fSAneesh Kumar K.V } 10478753e88fSAneesh Kumar K.V 1048617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 10499bffad1eSTheodore Ts'o trace_ext4_allocate_inode(inode, dir, mode); 1050ac27a0ecSDave Kleikamp goto really_out; 1051ac27a0ecSDave Kleikamp fail: 1052617ba13bSMingming Cao ext4_std_error(sb, err); 1053ac27a0ecSDave Kleikamp out: 1054ac27a0ecSDave Kleikamp iput(inode); 1055ac27a0ecSDave Kleikamp ret = ERR_PTR(err); 1056ac27a0ecSDave Kleikamp really_out: 10573300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1058ac27a0ecSDave Kleikamp return ret; 1059ac27a0ecSDave Kleikamp 1060ac27a0ecSDave Kleikamp fail_free_drop: 106163936ddaSChristoph Hellwig dquot_free_inode(inode); 1062ac27a0ecSDave Kleikamp 1063ac27a0ecSDave Kleikamp fail_drop: 10649f754758SChristoph Hellwig dquot_drop(inode); 1065ac27a0ecSDave Kleikamp inode->i_flags |= S_NOQUOTA; 1066ac27a0ecSDave Kleikamp inode->i_nlink = 0; 10676b38e842SAl Viro unlock_new_inode(inode); 1068ac27a0ecSDave Kleikamp iput(inode); 10693300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1070ac27a0ecSDave Kleikamp return ERR_PTR(err); 1071ac27a0ecSDave Kleikamp } 1072ac27a0ecSDave Kleikamp 1073ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1074617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1075ac27a0ecSDave Kleikamp { 1076617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1077fd2d4291SAvantika Mathur ext4_group_t block_group; 1078ac27a0ecSDave Kleikamp int bit; 10791d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 1080ac27a0ecSDave Kleikamp struct inode *inode = NULL; 10811d1fe1eeSDavid Howells long err = -EIO; 1082ac27a0ecSDave Kleikamp 1083ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1084ac27a0ecSDave Kleikamp if (ino > max_ino) { 108512062dddSEric Sandeen ext4_warning(sb, "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) { 109312062dddSEric Sandeen ext4_warning(sb, "inode bitmap error for orphan %lu", ino); 10941d1fe1eeSDavid Howells goto error; 1095ac27a0ecSDave Kleikamp } 1096ac27a0ecSDave Kleikamp 1097ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1098ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1099ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1100ac27a0ecSDave Kleikamp */ 11011d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 11021d1fe1eeSDavid Howells goto bad_orphan; 11031d1fe1eeSDavid Howells 11041d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 11051d1fe1eeSDavid Howells if (IS_ERR(inode)) 11061d1fe1eeSDavid Howells goto iget_failed; 11071d1fe1eeSDavid Howells 110891ef4cafSDuane Griffin /* 110991ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 111091ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 111191ef4cafSDuane Griffin * during processing and an infinite loop will result. 111291ef4cafSDuane Griffin */ 111391ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 111491ef4cafSDuane Griffin goto bad_orphan; 111591ef4cafSDuane Griffin 11161d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 11171d1fe1eeSDavid Howells goto bad_orphan; 11181d1fe1eeSDavid Howells brelse(bitmap_bh); 11191d1fe1eeSDavid Howells return inode; 11201d1fe1eeSDavid Howells 11211d1fe1eeSDavid Howells iget_failed: 11221d1fe1eeSDavid Howells err = PTR_ERR(inode); 11231d1fe1eeSDavid Howells inode = NULL; 11241d1fe1eeSDavid Howells bad_orphan: 112512062dddSEric Sandeen ext4_warning(sb, "bad orphan inode %lu! e2fsck was run?", ino); 1126617ba13bSMingming Cao printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1127ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1128617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1129ac27a0ecSDave Kleikamp printk(KERN_NOTICE "inode=%p\n", inode); 1130ac27a0ecSDave Kleikamp if (inode) { 1131ac27a0ecSDave Kleikamp printk(KERN_NOTICE "is_bad_inode(inode)=%d\n", 1132ac27a0ecSDave Kleikamp is_bad_inode(inode)); 1133ac27a0ecSDave Kleikamp printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n", 1134ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 1135ac27a0ecSDave Kleikamp printk(KERN_NOTICE "max_ino=%lu\n", max_ino); 113691ef4cafSDuane Griffin printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink); 1137ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 11381d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1139ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1140ac27a0ecSDave Kleikamp iput(inode); 1141ac27a0ecSDave Kleikamp } 1142ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11431d1fe1eeSDavid Howells error: 11441d1fe1eeSDavid Howells return ERR_PTR(err); 1145ac27a0ecSDave Kleikamp } 1146ac27a0ecSDave Kleikamp 1147617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1148ac27a0ecSDave Kleikamp { 1149ac27a0ecSDave Kleikamp unsigned long desc_count; 1150617ba13bSMingming Cao struct ext4_group_desc *gdp; 11518df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1152617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1153617ba13bSMingming Cao struct ext4_super_block *es; 1154ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1155ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1156ac27a0ecSDave Kleikamp 1157617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1158ac27a0ecSDave Kleikamp desc_count = 0; 1159ac27a0ecSDave Kleikamp bitmap_count = 0; 1160ac27a0ecSDave Kleikamp gdp = NULL; 11618df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1162617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1163ac27a0ecSDave Kleikamp if (!gdp) 1164ac27a0ecSDave Kleikamp continue; 1165560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1166ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1167e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1168ac27a0ecSDave Kleikamp if (!bitmap_bh) 1169ac27a0ecSDave Kleikamp continue; 1170ac27a0ecSDave Kleikamp 1171617ba13bSMingming Cao x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8); 1172c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1173785b4b3aSPeng Tao (unsigned long) i, ext4_free_inodes_count(sb, gdp), x); 1174ac27a0ecSDave Kleikamp bitmap_count += x; 1175ac27a0ecSDave Kleikamp } 1176ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11774776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 11784776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1179ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1180ac27a0ecSDave Kleikamp return desc_count; 1181ac27a0ecSDave Kleikamp #else 1182ac27a0ecSDave Kleikamp desc_count = 0; 11838df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1184617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1185ac27a0ecSDave Kleikamp if (!gdp) 1186ac27a0ecSDave Kleikamp continue; 1187560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1188ac27a0ecSDave Kleikamp cond_resched(); 1189ac27a0ecSDave Kleikamp } 1190ac27a0ecSDave Kleikamp return desc_count; 1191ac27a0ecSDave Kleikamp #endif 1192ac27a0ecSDave Kleikamp } 1193ac27a0ecSDave Kleikamp 1194ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1195617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1196ac27a0ecSDave Kleikamp { 1197ac27a0ecSDave Kleikamp unsigned long count = 0; 11988df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1199ac27a0ecSDave Kleikamp 12008df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1201617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1202ac27a0ecSDave Kleikamp if (!gdp) 1203ac27a0ecSDave Kleikamp continue; 1204560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1205ac27a0ecSDave Kleikamp } 1206ac27a0ecSDave Kleikamp return count; 1207ac27a0ecSDave Kleikamp } 1208