1ac27a0ecSDave Kleikamp /* 2617ba13bSMingming Cao * linux/fs/ext4/ialloc.c 3ac27a0ecSDave Kleikamp * 4ac27a0ecSDave Kleikamp * Copyright (C) 1992, 1993, 1994, 1995 5ac27a0ecSDave Kleikamp * Remy Card (card@masi.ibp.fr) 6ac27a0ecSDave Kleikamp * Laboratoire MASI - Institut Blaise Pascal 7ac27a0ecSDave Kleikamp * Universite Pierre et Marie Curie (Paris VI) 8ac27a0ecSDave Kleikamp * 9ac27a0ecSDave Kleikamp * BSD ufs-inspired inode and directory allocation by 10ac27a0ecSDave Kleikamp * Stephen Tweedie (sct@redhat.com), 1993 11ac27a0ecSDave Kleikamp * Big-endian to little-endian byte-swapping/bitmaps by 12ac27a0ecSDave Kleikamp * David S. Miller (davem@caip.rutgers.edu), 1995 13ac27a0ecSDave Kleikamp */ 14ac27a0ecSDave Kleikamp 15ac27a0ecSDave Kleikamp #include <linux/time.h> 16ac27a0ecSDave Kleikamp #include <linux/fs.h> 17dab291afSMingming Cao #include <linux/jbd2.h> 18ac27a0ecSDave Kleikamp #include <linux/stat.h> 19ac27a0ecSDave Kleikamp #include <linux/string.h> 20ac27a0ecSDave Kleikamp #include <linux/quotaops.h> 21ac27a0ecSDave Kleikamp #include <linux/buffer_head.h> 22ac27a0ecSDave Kleikamp #include <linux/random.h> 23ac27a0ecSDave Kleikamp #include <linux/bitops.h> 243a5b2ecdSMingming Cao #include <linux/blkdev.h> 25ac27a0ecSDave Kleikamp #include <asm/byteorder.h> 269bffad1eSTheodore Ts'o 273dcf5451SChristoph Hellwig #include "ext4.h" 283dcf5451SChristoph Hellwig #include "ext4_jbd2.h" 29ac27a0ecSDave Kleikamp #include "xattr.h" 30ac27a0ecSDave Kleikamp #include "acl.h" 31ac27a0ecSDave Kleikamp 329bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 339bffad1eSTheodore Ts'o 34ac27a0ecSDave Kleikamp /* 35ac27a0ecSDave Kleikamp * ialloc.c contains the inodes allocation and deallocation routines 36ac27a0ecSDave Kleikamp */ 37ac27a0ecSDave Kleikamp 38ac27a0ecSDave Kleikamp /* 39ac27a0ecSDave Kleikamp * The free inodes are managed by bitmaps. A file system contains several 40ac27a0ecSDave Kleikamp * blocks groups. Each group contains 1 bitmap block for blocks, 1 bitmap 41ac27a0ecSDave Kleikamp * block for inodes, N blocks for the inode table and data blocks. 42ac27a0ecSDave Kleikamp * 43ac27a0ecSDave Kleikamp * The file system contains group descriptors which are located after the 44ac27a0ecSDave Kleikamp * super block. Each descriptor contains the number of the bitmap block and 45ac27a0ecSDave Kleikamp * the free blocks count in the block. 46ac27a0ecSDave Kleikamp */ 47ac27a0ecSDave Kleikamp 48717d50e4SAndreas Dilger /* 49717d50e4SAndreas Dilger * To avoid calling the atomic setbit hundreds or thousands of times, we only 50717d50e4SAndreas Dilger * need to use it within a single byte (to ensure we get endianness right). 51717d50e4SAndreas Dilger * We can use memset for the rest of the bitmap as there are no other users. 52717d50e4SAndreas Dilger */ 5361d08673STheodore Ts'o void ext4_mark_bitmap_end(int start_bit, int end_bit, char *bitmap) 54717d50e4SAndreas Dilger { 55717d50e4SAndreas Dilger int i; 56717d50e4SAndreas Dilger 57717d50e4SAndreas Dilger if (start_bit >= end_bit) 58717d50e4SAndreas Dilger return; 59717d50e4SAndreas Dilger 60717d50e4SAndreas Dilger ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit); 61717d50e4SAndreas Dilger for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++) 62717d50e4SAndreas Dilger ext4_set_bit(i, bitmap); 63717d50e4SAndreas Dilger if (i < end_bit) 64717d50e4SAndreas Dilger memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3); 65717d50e4SAndreas Dilger } 66717d50e4SAndreas Dilger 67717d50e4SAndreas Dilger /* Initializes an uninitialized inode bitmap */ 681f109d5aSTheodore Ts'o static unsigned ext4_init_inode_bitmap(struct super_block *sb, 691f109d5aSTheodore Ts'o struct buffer_head *bh, 70fd2d4291SAvantika Mathur ext4_group_t block_group, 71717d50e4SAndreas Dilger struct ext4_group_desc *gdp) 72717d50e4SAndreas Dilger { 73717d50e4SAndreas Dilger J_ASSERT_BH(bh, buffer_locked(bh)); 74717d50e4SAndreas Dilger 75717d50e4SAndreas Dilger /* If checksum is bad mark all blocks and inodes use to prevent 76717d50e4SAndreas Dilger * allocation, essentially implementing a per-group read-only flag. */ 77feb0ab32SDarrick J. Wong if (!ext4_group_desc_csum_verify(sb, block_group, gdp)) { 7812062dddSEric Sandeen ext4_error(sb, "Checksum bad for group %u", block_group); 79021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 0); 80560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, 0); 81560671a0SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 0); 82717d50e4SAndreas Dilger memset(bh->b_data, 0xff, sb->s_blocksize); 8341a246d1SDarrick J. Wong ext4_inode_bitmap_csum_set(sb, block_group, gdp, bh, 8441a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8); 85717d50e4SAndreas Dilger return 0; 86717d50e4SAndreas Dilger } 87717d50e4SAndreas Dilger 88717d50e4SAndreas Dilger memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8); 8961d08673STheodore Ts'o ext4_mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), sb->s_blocksize * 8, 90717d50e4SAndreas Dilger bh->b_data); 9141a246d1SDarrick J. Wong ext4_inode_bitmap_csum_set(sb, block_group, gdp, bh, 9241a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8); 93feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, gdp); 94717d50e4SAndreas Dilger 95717d50e4SAndreas Dilger return EXT4_INODES_PER_GROUP(sb); 96717d50e4SAndreas Dilger } 97ac27a0ecSDave Kleikamp 98813e5727STheodore Ts'o void ext4_end_bitmap_read(struct buffer_head *bh, int uptodate) 99813e5727STheodore Ts'o { 100813e5727STheodore Ts'o if (uptodate) { 101813e5727STheodore Ts'o set_buffer_uptodate(bh); 102813e5727STheodore Ts'o set_bitmap_uptodate(bh); 103813e5727STheodore Ts'o } 104813e5727STheodore Ts'o unlock_buffer(bh); 105813e5727STheodore Ts'o put_bh(bh); 106813e5727STheodore Ts'o } 107813e5727STheodore Ts'o 108ac27a0ecSDave Kleikamp /* 109ac27a0ecSDave Kleikamp * Read the inode allocation bitmap for a given block_group, reading 110ac27a0ecSDave Kleikamp * into the specified slot in the superblock's bitmap cache. 111ac27a0ecSDave Kleikamp * 112ac27a0ecSDave Kleikamp * Return buffer_head of bitmap on success or NULL. 113ac27a0ecSDave Kleikamp */ 114ac27a0ecSDave Kleikamp static struct buffer_head * 115e29d1cdeSEric Sandeen ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group) 116ac27a0ecSDave Kleikamp { 117617ba13bSMingming Cao struct ext4_group_desc *desc; 118ac27a0ecSDave Kleikamp struct buffer_head *bh = NULL; 119e29d1cdeSEric Sandeen ext4_fsblk_t bitmap_blk; 120ac27a0ecSDave Kleikamp 121617ba13bSMingming Cao desc = ext4_get_group_desc(sb, block_group, NULL); 122ac27a0ecSDave Kleikamp if (!desc) 123e29d1cdeSEric Sandeen return NULL; 124bfff6873SLukas Czerner 125e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 126e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 127e29d1cdeSEric Sandeen if (unlikely(!bh)) { 12812062dddSEric Sandeen ext4_error(sb, "Cannot read inode bitmap - " 129a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 130e29d1cdeSEric Sandeen block_group, bitmap_blk); 131e29d1cdeSEric Sandeen return NULL; 132e29d1cdeSEric Sandeen } 1332ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) 13441a246d1SDarrick J. Wong goto verify; 135e29d1cdeSEric Sandeen 136c806e68fSFrederic Bohe lock_buffer(bh); 1372ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh)) { 1382ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 13941a246d1SDarrick J. Wong goto verify; 1402ccb5fb9SAneesh Kumar K.V } 141bfff6873SLukas Czerner 142955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 143e29d1cdeSEric Sandeen if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 144e29d1cdeSEric Sandeen ext4_init_inode_bitmap(sb, bh, block_group, desc); 1452ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 146e29d1cdeSEric Sandeen set_buffer_uptodate(bh); 14741a246d1SDarrick J. Wong set_buffer_verified(bh); 148955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 1493300bedaSAneesh Kumar K.V unlock_buffer(bh); 150e29d1cdeSEric Sandeen return bh; 151e29d1cdeSEric Sandeen } 152955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 153bfff6873SLukas Czerner 1542ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh)) { 1552ccb5fb9SAneesh Kumar K.V /* 1562ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 1572ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 1582ccb5fb9SAneesh Kumar K.V */ 1592ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh); 1602ccb5fb9SAneesh Kumar K.V unlock_buffer(bh); 16141a246d1SDarrick J. Wong goto verify; 1622ccb5fb9SAneesh Kumar K.V } 1632ccb5fb9SAneesh Kumar K.V /* 164813e5727STheodore Ts'o * submit the buffer_head for reading 1652ccb5fb9SAneesh Kumar K.V */ 1660562e0baSJiaying Zhang trace_ext4_load_inode_bitmap(sb, block_group); 167813e5727STheodore Ts'o bh->b_end_io = ext4_end_bitmap_read; 168813e5727STheodore Ts'o get_bh(bh); 169813e5727STheodore Ts'o submit_bh(READ, bh); 170813e5727STheodore Ts'o wait_on_buffer(bh); 171813e5727STheodore Ts'o if (!buffer_uptodate(bh)) { 172e29d1cdeSEric Sandeen put_bh(bh); 17312062dddSEric Sandeen ext4_error(sb, "Cannot read inode bitmap - " 174a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 175e29d1cdeSEric Sandeen block_group, bitmap_blk); 176e29d1cdeSEric Sandeen return NULL; 177e29d1cdeSEric Sandeen } 17841a246d1SDarrick J. Wong 17941a246d1SDarrick J. Wong verify: 18041a246d1SDarrick J. Wong ext4_lock_group(sb, block_group); 18141a246d1SDarrick J. Wong if (!buffer_verified(bh) && 18241a246d1SDarrick J. Wong !ext4_inode_bitmap_csum_verify(sb, block_group, desc, bh, 18341a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8)) { 18441a246d1SDarrick J. Wong ext4_unlock_group(sb, block_group); 18541a246d1SDarrick J. Wong put_bh(bh); 18641a246d1SDarrick J. Wong ext4_error(sb, "Corrupt inode bitmap - block_group = %u, " 18741a246d1SDarrick J. Wong "inode_bitmap = %llu", block_group, bitmap_blk); 18841a246d1SDarrick J. Wong return NULL; 18941a246d1SDarrick J. Wong } 19041a246d1SDarrick J. Wong ext4_unlock_group(sb, block_group); 19141a246d1SDarrick J. Wong set_buffer_verified(bh); 192ac27a0ecSDave Kleikamp return bh; 193ac27a0ecSDave Kleikamp } 194ac27a0ecSDave Kleikamp 195ac27a0ecSDave Kleikamp /* 196ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 197ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 198ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 199ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 200ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 201ac27a0ecSDave Kleikamp * has been deleted earlier. 202ac27a0ecSDave Kleikamp * 203ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 204ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 205ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 206ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 207ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 208ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 209ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 210ac27a0ecSDave Kleikamp */ 211617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 212ac27a0ecSDave Kleikamp { 213ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 214ac27a0ecSDave Kleikamp int is_directory; 215ac27a0ecSDave Kleikamp unsigned long ino; 216ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 217ac27a0ecSDave Kleikamp struct buffer_head *bh2; 218fd2d4291SAvantika Mathur ext4_group_t block_group; 219ac27a0ecSDave Kleikamp unsigned long bit; 220617ba13bSMingming Cao struct ext4_group_desc *gdp; 221617ba13bSMingming Cao struct ext4_super_block *es; 222617ba13bSMingming Cao struct ext4_sb_info *sbi; 2237ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 224ac27a0ecSDave Kleikamp 22592b97816STheodore Ts'o if (!sb) { 22692b97816STheodore Ts'o printk(KERN_ERR "EXT4-fs: %s:%d: inode on " 22792b97816STheodore Ts'o "nonexistent device\n", __func__, __LINE__); 22892b97816STheodore Ts'o return; 22992b97816STheodore Ts'o } 230ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 23192b97816STheodore Ts'o ext4_msg(sb, KERN_ERR, "%s:%d: inode #%lu: count=%d", 23292b97816STheodore Ts'o __func__, __LINE__, inode->i_ino, 233ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 234ac27a0ecSDave Kleikamp return; 235ac27a0ecSDave Kleikamp } 236ac27a0ecSDave Kleikamp if (inode->i_nlink) { 23792b97816STheodore Ts'o ext4_msg(sb, KERN_ERR, "%s:%d: inode #%lu: nlink=%d\n", 23892b97816STheodore Ts'o __func__, __LINE__, inode->i_ino, inode->i_nlink); 239ac27a0ecSDave Kleikamp return; 240ac27a0ecSDave Kleikamp } 241617ba13bSMingming Cao sbi = EXT4_SB(sb); 242ac27a0ecSDave Kleikamp 243ac27a0ecSDave Kleikamp ino = inode->i_ino; 244617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 2459bffad1eSTheodore Ts'o trace_ext4_free_inode(inode); 246ac27a0ecSDave Kleikamp 247ac27a0ecSDave Kleikamp /* 248ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 249ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 250ac27a0ecSDave Kleikamp */ 251871a2931SChristoph Hellwig dquot_initialize(inode); 252617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 25363936ddaSChristoph Hellwig dquot_free_inode(inode); 2549f754758SChristoph Hellwig dquot_drop(inode); 255ac27a0ecSDave Kleikamp 256ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 257ac27a0ecSDave Kleikamp 258ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 2590930fcc1SAl Viro ext4_clear_inode(inode); 260ac27a0ecSDave Kleikamp 261617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 262617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 26312062dddSEric Sandeen ext4_error(sb, "reserved or nonexistent inode %lu", ino); 264ac27a0ecSDave Kleikamp goto error_return; 265ac27a0ecSDave Kleikamp } 266617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 267617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 268e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 269ac27a0ecSDave Kleikamp if (!bitmap_bh) 270ac27a0ecSDave Kleikamp goto error_return; 271ac27a0ecSDave Kleikamp 272ac27a0ecSDave Kleikamp BUFFER_TRACE(bitmap_bh, "get_write_access"); 273617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bitmap_bh); 274ac27a0ecSDave Kleikamp if (fatal) 275ac27a0ecSDave Kleikamp goto error_return; 276ac27a0ecSDave Kleikamp 277d17413c0SDmitry Monakhov fatal = -ESRCH; 278617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 279d17413c0SDmitry Monakhov if (gdp) { 280ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 281617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 282d17413c0SDmitry Monakhov } 283955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 284597d508cSAkinobu Mita cleared = ext4_test_and_clear_bit(bit, bitmap_bh->b_data); 285d17413c0SDmitry Monakhov if (fatal || !cleared) { 286d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 287d17413c0SDmitry Monakhov goto out; 288d17413c0SDmitry Monakhov } 289d17413c0SDmitry Monakhov 290560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 291560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 292560671a0SAneesh Kumar K.V if (is_directory) { 293560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 294560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 295d17413c0SDmitry Monakhov percpu_counter_dec(&sbi->s_dirs_counter); 296d17413c0SDmitry Monakhov } 29741a246d1SDarrick J. Wong ext4_inode_bitmap_csum_set(sb, block_group, gdp, bitmap_bh, 29841a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8); 299feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, gdp); 300d17413c0SDmitry Monakhov ext4_unlock_group(sb, block_group); 3017d39db14STheodore Ts'o 302d17413c0SDmitry Monakhov percpu_counter_inc(&sbi->s_freeinodes_counter); 303d17413c0SDmitry Monakhov if (sbi->s_log_groups_per_flex) { 304d17413c0SDmitry Monakhov ext4_group_t f = ext4_flex_group(sbi, block_group); 305d17413c0SDmitry Monakhov 306d17413c0SDmitry Monakhov atomic_inc(&sbi->s_flex_groups[f].free_inodes); 307d17413c0SDmitry Monakhov if (is_directory) 308c4caae25SEric Sandeen atomic_dec(&sbi->s_flex_groups[f].used_dirs); 3097d39db14STheodore Ts'o } 3100390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 311d17413c0SDmitry Monakhov fatal = ext4_handle_dirty_metadata(handle, NULL, bh2); 312d17413c0SDmitry Monakhov out: 313d17413c0SDmitry Monakhov if (cleared) { 3140390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 3150390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 316ac27a0ecSDave Kleikamp if (!fatal) 317ac27a0ecSDave Kleikamp fatal = err; 318d17413c0SDmitry Monakhov } else 319d17413c0SDmitry Monakhov ext4_error(sb, "bit already cleared for inode %lu", ino); 320d17413c0SDmitry Monakhov 321ac27a0ecSDave Kleikamp error_return: 322ac27a0ecSDave Kleikamp brelse(bitmap_bh); 323617ba13bSMingming Cao ext4_std_error(sb, fatal); 324ac27a0ecSDave Kleikamp } 325ac27a0ecSDave Kleikamp 326a4912123STheodore Ts'o struct orlov_stats { 32790ba983fSTheodore Ts'o __u64 free_clusters; 328a4912123STheodore Ts'o __u32 free_inodes; 329a4912123STheodore Ts'o __u32 used_dirs; 330a4912123STheodore Ts'o }; 331a4912123STheodore Ts'o 332a4912123STheodore Ts'o /* 333a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 334a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 335a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 336a4912123STheodore Ts'o */ 3371f109d5aSTheodore Ts'o static void get_orlov_stats(struct super_block *sb, ext4_group_t g, 338a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 339a4912123STheodore Ts'o { 340a4912123STheodore Ts'o struct ext4_group_desc *desc; 3417d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 342a4912123STheodore Ts'o 3437d39db14STheodore Ts'o if (flex_size > 1) { 3447d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 34590ba983fSTheodore Ts'o stats->free_clusters = atomic64_read(&flex_group[g].free_clusters); 3467d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 3477d39db14STheodore Ts'o return; 3487d39db14STheodore Ts'o } 3497d39db14STheodore Ts'o 3507d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 3517d39db14STheodore Ts'o if (desc) { 3527d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 353021b65bbSTheodore Ts'o stats->free_clusters = ext4_free_group_clusters(sb, desc); 3547d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 3557d39db14STheodore Ts'o } else { 356a4912123STheodore Ts'o stats->free_inodes = 0; 35724aaa8efSTheodore Ts'o stats->free_clusters = 0; 358a4912123STheodore Ts'o stats->used_dirs = 0; 359a4912123STheodore Ts'o } 360a4912123STheodore Ts'o } 361a4912123STheodore Ts'o 362ac27a0ecSDave Kleikamp /* 363ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 364ac27a0ecSDave Kleikamp * 365ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 366ac27a0ecSDave Kleikamp * 367ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 368ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 369ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 370ac27a0ecSDave Kleikamp * 371ac27a0ecSDave Kleikamp * For the rest rules look so: 372ac27a0ecSDave Kleikamp * 373ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 374ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 375ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 376ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 3771cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 378ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 379ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 380ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 381ac27a0ecSDave Kleikamp */ 382ac27a0ecSDave Kleikamp 3832aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 384dcca3fecSAl Viro ext4_group_t *group, umode_t mode, 385f157a4aaSTheodore Ts'o const struct qstr *qstr) 386ac27a0ecSDave Kleikamp { 387fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 388617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 3898df9675fSTheodore Ts'o ext4_group_t real_ngroups = ext4_get_groups_count(sb); 390617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 39114c83c9fSTheodore Ts'o unsigned int freei, avefreei, grp_free; 39224aaa8efSTheodore Ts'o ext4_fsblk_t freeb, avefreec; 393ac27a0ecSDave Kleikamp unsigned int ndirs; 394a4912123STheodore Ts'o int max_dirs, min_inodes; 39524aaa8efSTheodore Ts'o ext4_grpblk_t min_clusters; 3968df9675fSTheodore Ts'o ext4_group_t i, grp, g, ngroups; 397617ba13bSMingming Cao struct ext4_group_desc *desc; 398a4912123STheodore Ts'o struct orlov_stats stats; 399a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 400f157a4aaSTheodore Ts'o struct dx_hash_info hinfo; 401a4912123STheodore Ts'o 4028df9675fSTheodore Ts'o ngroups = real_ngroups; 403a4912123STheodore Ts'o if (flex_size > 1) { 4048df9675fSTheodore Ts'o ngroups = (real_ngroups + flex_size - 1) >> 405a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 406a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 407a4912123STheodore Ts'o } 408ac27a0ecSDave Kleikamp 409ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 410ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 41157042651STheodore Ts'o freeb = EXT4_C2B(sbi, 41257042651STheodore Ts'o percpu_counter_read_positive(&sbi->s_freeclusters_counter)); 41324aaa8efSTheodore Ts'o avefreec = freeb; 41424aaa8efSTheodore Ts'o do_div(avefreec, ngroups); 415ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 416ac27a0ecSDave Kleikamp 417a4912123STheodore Ts'o if (S_ISDIR(mode) && 418a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 41912e9b892SDmitry Monakhov (ext4_test_inode_flag(parent, EXT4_INODE_TOPDIR)))) { 420ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 4212aa9fc4cSAvantika Mathur int ret = -1; 422ac27a0ecSDave Kleikamp 423f157a4aaSTheodore Ts'o if (qstr) { 424f157a4aaSTheodore Ts'o hinfo.hash_version = DX_HASH_HALF_MD4; 425f157a4aaSTheodore Ts'o hinfo.seed = sbi->s_hash_seed; 426f157a4aaSTheodore Ts'o ext4fs_dirhash(qstr->name, qstr->len, &hinfo); 427f157a4aaSTheodore Ts'o grp = hinfo.hash; 428f157a4aaSTheodore Ts'o } else 4292aa9fc4cSAvantika Mathur get_random_bytes(&grp, sizeof(grp)); 4302aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 431ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 432a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 433a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 434a4912123STheodore Ts'o if (!stats.free_inodes) 435ac27a0ecSDave Kleikamp continue; 436a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 437ac27a0ecSDave Kleikamp continue; 438a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 439ac27a0ecSDave Kleikamp continue; 44024aaa8efSTheodore Ts'o if (stats.free_clusters < avefreec) 441ac27a0ecSDave Kleikamp continue; 442a4912123STheodore Ts'o grp = g; 4432aa9fc4cSAvantika Mathur ret = 0; 444a4912123STheodore Ts'o best_ndir = stats.used_dirs; 445ac27a0ecSDave Kleikamp } 446a4912123STheodore Ts'o if (ret) 447a4912123STheodore Ts'o goto fallback; 448a4912123STheodore Ts'o found_flex_bg: 449a4912123STheodore Ts'o if (flex_size == 1) { 450a4912123STheodore Ts'o *group = grp; 451a4912123STheodore Ts'o return 0; 452a4912123STheodore Ts'o } 453a4912123STheodore Ts'o 454a4912123STheodore Ts'o /* 455a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 456a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 457a4912123STheodore Ts'o * something similar, although regular files will 458a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 459a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 460a4912123STheodore Ts'o */ 461a4912123STheodore Ts'o grp *= flex_size; 462a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 4638df9675fSTheodore Ts'o if (grp+i >= real_ngroups) 464a4912123STheodore Ts'o break; 465a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 466a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 467a4912123STheodore Ts'o *group = grp+i; 468a4912123STheodore Ts'o return 0; 469a4912123STheodore Ts'o } 470a4912123STheodore Ts'o } 471ac27a0ecSDave Kleikamp goto fallback; 472ac27a0ecSDave Kleikamp } 473ac27a0ecSDave Kleikamp 474ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 475a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 476a4912123STheodore Ts'o if (min_inodes < 1) 477a4912123STheodore Ts'o min_inodes = 1; 47824aaa8efSTheodore Ts'o min_clusters = avefreec - EXT4_CLUSTERS_PER_GROUP(sb)*flex_size / 4; 479ac27a0ecSDave Kleikamp 480a4912123STheodore Ts'o /* 481a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 482a4912123STheodore Ts'o * inode for this parent directory 483a4912123STheodore Ts'o */ 484a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 485a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 486a4912123STheodore Ts'o if (flex_size > 1) 487a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 488a4912123STheodore Ts'o } 489ac27a0ecSDave Kleikamp 490ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 491a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 492a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 493a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 494ac27a0ecSDave Kleikamp continue; 495a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 496ac27a0ecSDave Kleikamp continue; 49724aaa8efSTheodore Ts'o if (stats.free_clusters < min_clusters) 498ac27a0ecSDave Kleikamp continue; 499a4912123STheodore Ts'o goto found_flex_bg; 500ac27a0ecSDave Kleikamp } 501ac27a0ecSDave Kleikamp 502ac27a0ecSDave Kleikamp fallback: 5038df9675fSTheodore Ts'o ngroups = real_ngroups; 504a4912123STheodore Ts'o avefreei = freei / ngroups; 505b5451f7bSTheodore Ts'o fallback_retry: 506a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 507ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 508a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 509a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 510bb3d132aSDan Carpenter if (desc) { 51114c83c9fSTheodore Ts'o grp_free = ext4_free_inodes_count(sb, desc); 512bb3d132aSDan Carpenter if (grp_free && grp_free >= avefreei) { 513a4912123STheodore Ts'o *group = grp; 5142aa9fc4cSAvantika Mathur return 0; 515ac27a0ecSDave Kleikamp } 516a4912123STheodore Ts'o } 517bb3d132aSDan Carpenter } 518ac27a0ecSDave Kleikamp 519ac27a0ecSDave Kleikamp if (avefreei) { 520ac27a0ecSDave Kleikamp /* 521ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 522ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 523ac27a0ecSDave Kleikamp */ 524ac27a0ecSDave Kleikamp avefreei = 0; 525b5451f7bSTheodore Ts'o goto fallback_retry; 526ac27a0ecSDave Kleikamp } 527ac27a0ecSDave Kleikamp 528ac27a0ecSDave Kleikamp return -1; 529ac27a0ecSDave Kleikamp } 530ac27a0ecSDave Kleikamp 5312aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 532dcca3fecSAl Viro ext4_group_t *group, umode_t mode) 533ac27a0ecSDave Kleikamp { 534fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 5358df9675fSTheodore Ts'o ext4_group_t i, last, ngroups = ext4_get_groups_count(sb); 536617ba13bSMingming Cao struct ext4_group_desc *desc; 537a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 538a4912123STheodore Ts'o 539a4912123STheodore Ts'o /* 540a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 541a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 542a4912123STheodore Ts'o * find another flex group, and store that information in the 543a4912123STheodore Ts'o * parent directory's inode information so that use that flex 544a4912123STheodore Ts'o * group for future allocations. 545a4912123STheodore Ts'o */ 546a4912123STheodore Ts'o if (flex_size > 1) { 547a4912123STheodore Ts'o int retry = 0; 548a4912123STheodore Ts'o 549a4912123STheodore Ts'o try_again: 550a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 551a4912123STheodore Ts'o last = parent_group + flex_size; 552a4912123STheodore Ts'o if (last > ngroups) 553a4912123STheodore Ts'o last = ngroups; 554a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 555a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 556a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 557a4912123STheodore Ts'o *group = i; 558a4912123STheodore Ts'o return 0; 559a4912123STheodore Ts'o } 560a4912123STheodore Ts'o } 561a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 562a4912123STheodore Ts'o retry = 1; 563a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 564a4912123STheodore Ts'o goto try_again; 565a4912123STheodore Ts'o } 566a4912123STheodore Ts'o /* 567a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 568a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 569a4912123STheodore Ts'o * avoid the topdir algorithms. 570a4912123STheodore Ts'o */ 571a4912123STheodore Ts'o *group = parent_group + flex_size; 572a4912123STheodore Ts'o if (*group > ngroups) 573a4912123STheodore Ts'o *group = 0; 5747dc57615SPeter Huewe return find_group_orlov(sb, parent, group, mode, NULL); 575a4912123STheodore Ts'o } 576ac27a0ecSDave Kleikamp 577ac27a0ecSDave Kleikamp /* 578ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 579ac27a0ecSDave Kleikamp */ 5802aa9fc4cSAvantika Mathur *group = parent_group; 5812aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 582560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 583021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc)) 5842aa9fc4cSAvantika Mathur return 0; 585ac27a0ecSDave Kleikamp 586ac27a0ecSDave Kleikamp /* 587ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 588ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 589ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 590ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 591ac27a0ecSDave Kleikamp * different blockgroup. 592ac27a0ecSDave Kleikamp * 593ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 594ac27a0ecSDave Kleikamp */ 5952aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 596ac27a0ecSDave Kleikamp 597ac27a0ecSDave Kleikamp /* 598ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 599ac27a0ecSDave Kleikamp * blocks. 600ac27a0ecSDave Kleikamp */ 601ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6022aa9fc4cSAvantika Mathur *group += i; 6032aa9fc4cSAvantika Mathur if (*group >= ngroups) 6042aa9fc4cSAvantika Mathur *group -= ngroups; 6052aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 606560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 607021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc)) 6082aa9fc4cSAvantika Mathur return 0; 609ac27a0ecSDave Kleikamp } 610ac27a0ecSDave Kleikamp 611ac27a0ecSDave Kleikamp /* 612ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 613ac27a0ecSDave Kleikamp * has no free blocks. 614ac27a0ecSDave Kleikamp */ 6152aa9fc4cSAvantika Mathur *group = parent_group; 616ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6172aa9fc4cSAvantika Mathur if (++*group >= ngroups) 6182aa9fc4cSAvantika Mathur *group = 0; 6192aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 620560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 6212aa9fc4cSAvantika Mathur return 0; 622ac27a0ecSDave Kleikamp } 623ac27a0ecSDave Kleikamp 624ac27a0ecSDave Kleikamp return -1; 625ac27a0ecSDave Kleikamp } 626ac27a0ecSDave Kleikamp 627ac27a0ecSDave Kleikamp /* 628ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 629ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 630ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 631ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 632ac27a0ecSDave Kleikamp * directories already is chosen. 633ac27a0ecSDave Kleikamp * 634ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 635ac27a0ecSDave Kleikamp * group to find a free inode. 636ac27a0ecSDave Kleikamp */ 6371139575aSTheodore Ts'o struct inode *__ext4_new_inode(handle_t *handle, struct inode *dir, 6381139575aSTheodore Ts'o umode_t mode, const struct qstr *qstr, 6391139575aSTheodore Ts'o __u32 goal, uid_t *owner, int handle_type, 6401139575aSTheodore Ts'o unsigned int line_no, int nblocks) 641ac27a0ecSDave Kleikamp { 642ac27a0ecSDave Kleikamp struct super_block *sb; 6433300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 6443300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 6458df9675fSTheodore Ts'o ext4_group_t ngroups, group = 0; 646ac27a0ecSDave Kleikamp unsigned long ino = 0; 647ac27a0ecSDave Kleikamp struct inode *inode; 648617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 649617ba13bSMingming Cao struct ext4_inode_info *ei; 650617ba13bSMingming Cao struct ext4_sb_info *sbi; 65139341867SAneesh Kumar K.V int ret2, err = 0; 652ac27a0ecSDave Kleikamp struct inode *ret; 6532aa9fc4cSAvantika Mathur ext4_group_t i; 654772cb7c8SJose R. Santos ext4_group_t flex_group; 655ac27a0ecSDave Kleikamp 656ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 657ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 658ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 659ac27a0ecSDave Kleikamp 660ac27a0ecSDave Kleikamp sb = dir->i_sb; 6618df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 6629bffad1eSTheodore Ts'o trace_ext4_request_inode(dir, mode); 663ac27a0ecSDave Kleikamp inode = new_inode(sb); 664ac27a0ecSDave Kleikamp if (!inode) 665ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 666617ba13bSMingming Cao ei = EXT4_I(inode); 667617ba13bSMingming Cao sbi = EXT4_SB(sb); 668772cb7c8SJose R. Santos 669eb9cc7e1SJan Kara /* 670eb9cc7e1SJan Kara * Initalize owners and quota early so that we don't have to account 671eb9cc7e1SJan Kara * for quota initialization worst case in standard inode creating 672eb9cc7e1SJan Kara * transaction 673eb9cc7e1SJan Kara */ 674eb9cc7e1SJan Kara if (owner) { 675eb9cc7e1SJan Kara inode->i_mode = mode; 676eb9cc7e1SJan Kara i_uid_write(inode, owner[0]); 677eb9cc7e1SJan Kara i_gid_write(inode, owner[1]); 678eb9cc7e1SJan Kara } else if (test_opt(sb, GRPID)) { 679eb9cc7e1SJan Kara inode->i_mode = mode; 680eb9cc7e1SJan Kara inode->i_uid = current_fsuid(); 681eb9cc7e1SJan Kara inode->i_gid = dir->i_gid; 682eb9cc7e1SJan Kara } else 683eb9cc7e1SJan Kara inode_init_owner(inode, dir, mode); 684eb9cc7e1SJan Kara dquot_initialize(inode); 685eb9cc7e1SJan Kara 68611013911SAndreas Dilger if (!goal) 68711013911SAndreas Dilger goal = sbi->s_inode_goal; 68811013911SAndreas Dilger 689e6462869SJohann Lombardi if (goal && goal <= le32_to_cpu(sbi->s_es->s_inodes_count)) { 69011013911SAndreas Dilger group = (goal - 1) / EXT4_INODES_PER_GROUP(sb); 69111013911SAndreas Dilger ino = (goal - 1) % EXT4_INODES_PER_GROUP(sb); 69211013911SAndreas Dilger ret2 = 0; 69311013911SAndreas Dilger goto got_group; 69411013911SAndreas Dilger } 69511013911SAndreas Dilger 6964113c4caSLukas Czerner if (S_ISDIR(mode)) 697f157a4aaSTheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode, qstr); 6984113c4caSLukas Czerner else 699a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 700ac27a0ecSDave Kleikamp 701772cb7c8SJose R. Santos got_group: 702a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 703ac27a0ecSDave Kleikamp err = -ENOSPC; 7042aa9fc4cSAvantika Mathur if (ret2 == -1) 705ac27a0ecSDave Kleikamp goto out; 706ac27a0ecSDave Kleikamp 707119c0d44STheodore Ts'o /* 708119c0d44STheodore Ts'o * Normally we will only go through one pass of this loop, 709119c0d44STheodore Ts'o * unless we get unlucky and it turns out the group we selected 710119c0d44STheodore Ts'o * had its last inode grabbed by someone else. 711119c0d44STheodore Ts'o */ 71211013911SAndreas Dilger for (i = 0; i < ngroups; i++, ino = 0) { 713ac27a0ecSDave Kleikamp err = -EIO; 714ac27a0ecSDave Kleikamp 7153300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 716ac27a0ecSDave Kleikamp if (!gdp) 717eb9cc7e1SJan Kara goto out; 718ac27a0ecSDave Kleikamp 719f2a09af6SYongqiang Yang /* 720f2a09af6SYongqiang Yang * Check free inodes count before loading bitmap. 721f2a09af6SYongqiang Yang */ 722f2a09af6SYongqiang Yang if (ext4_free_inodes_count(sb, gdp) == 0) { 723f2a09af6SYongqiang Yang if (++group == ngroups) 724f2a09af6SYongqiang Yang group = 0; 725f2a09af6SYongqiang Yang continue; 726f2a09af6SYongqiang Yang } 727f2a09af6SYongqiang Yang 7283300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 7293300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 7303300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 731eb9cc7e1SJan Kara goto out; 732ac27a0ecSDave Kleikamp 733ac27a0ecSDave Kleikamp repeat_in_this_group: 734617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 7353300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 7363300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 737119c0d44STheodore Ts'o if (ino >= EXT4_INODES_PER_GROUP(sb)) { 7388df9675fSTheodore Ts'o if (++group == ngroups) 739ac27a0ecSDave Kleikamp group = 0; 740119c0d44STheodore Ts'o continue; 741119c0d44STheodore Ts'o } 742119c0d44STheodore Ts'o if (group == 0 && (ino+1) < EXT4_FIRST_INO(sb)) { 743119c0d44STheodore Ts'o ext4_error(sb, "reserved inode found cleared - " 744119c0d44STheodore Ts'o "inode=%lu", ino + 1); 745119c0d44STheodore Ts'o continue; 746119c0d44STheodore Ts'o } 7471139575aSTheodore Ts'o if (!handle) { 7481139575aSTheodore Ts'o BUG_ON(nblocks <= 0); 7491139575aSTheodore Ts'o handle = __ext4_journal_start_sb(dir->i_sb, line_no, 7501139575aSTheodore Ts'o handle_type, nblocks); 7511139575aSTheodore Ts'o if (IS_ERR(handle)) { 7521139575aSTheodore Ts'o err = PTR_ERR(handle); 753eb9cc7e1SJan Kara ext4_std_error(sb, err); 754eb9cc7e1SJan Kara goto out; 7551139575aSTheodore Ts'o } 7561139575aSTheodore Ts'o } 757ffb5387eSEric Sandeen BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 758ffb5387eSEric Sandeen err = ext4_journal_get_write_access(handle, inode_bitmap_bh); 759eb9cc7e1SJan Kara if (err) { 760eb9cc7e1SJan Kara ext4_std_error(sb, err); 761eb9cc7e1SJan Kara goto out; 762eb9cc7e1SJan Kara } 763119c0d44STheodore Ts'o ext4_lock_group(sb, group); 764119c0d44STheodore Ts'o ret2 = ext4_test_and_set_bit(ino, inode_bitmap_bh->b_data); 765119c0d44STheodore Ts'o ext4_unlock_group(sb, group); 766119c0d44STheodore Ts'o ino++; /* the inode bitmap is zero-based */ 767119c0d44STheodore Ts'o if (!ret2) 768119c0d44STheodore Ts'o goto got; /* we grabbed the inode! */ 769119c0d44STheodore Ts'o if (ino < EXT4_INODES_PER_GROUP(sb)) 770119c0d44STheodore Ts'o goto repeat_in_this_group; 771ac27a0ecSDave Kleikamp } 772ac27a0ecSDave Kleikamp err = -ENOSPC; 773ac27a0ecSDave Kleikamp goto out; 774ac27a0ecSDave Kleikamp 775ac27a0ecSDave Kleikamp got: 776ffb5387eSEric Sandeen BUFFER_TRACE(inode_bitmap_bh, "call ext4_handle_dirty_metadata"); 777ffb5387eSEric Sandeen err = ext4_handle_dirty_metadata(handle, NULL, inode_bitmap_bh); 778eb9cc7e1SJan Kara if (err) { 779eb9cc7e1SJan Kara ext4_std_error(sb, err); 780eb9cc7e1SJan Kara goto out; 781eb9cc7e1SJan Kara } 782ffb5387eSEric Sandeen 783717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 784feb0ab32SDarrick J. Wong if (ext4_has_group_desc_csum(sb) && 785717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 7863300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 787717d50e4SAndreas Dilger 7883300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 7893300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 7903300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 791717d50e4SAndreas Dilger if (err) { 7923300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 793eb9cc7e1SJan Kara ext4_std_error(sb, err); 794eb9cc7e1SJan Kara goto out; 795717d50e4SAndreas Dilger } 796717d50e4SAndreas Dilger 797fd034a84STheodore Ts'o BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 798fd034a84STheodore Ts'o err = ext4_handle_dirty_metadata(handle, NULL, block_bitmap_bh); 799fd034a84STheodore Ts'o 800717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 801fd034a84STheodore Ts'o ext4_lock_group(sb, group); 802717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 8033300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 804021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 805cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, group, gdp)); 806fa77dcfaSDarrick J. Wong ext4_block_bitmap_csum_set(sb, group, gdp, 80779f1ba49STao Ma block_bitmap_bh); 808feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, group, gdp); 809717d50e4SAndreas Dilger } 810955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, group); 811aeb1e5d6STheodore Ts'o brelse(block_bitmap_bh); 812717d50e4SAndreas Dilger 813eb9cc7e1SJan Kara if (err) { 814eb9cc7e1SJan Kara ext4_std_error(sb, err); 815eb9cc7e1SJan Kara goto out; 816eb9cc7e1SJan Kara } 817717d50e4SAndreas Dilger } 818119c0d44STheodore Ts'o 819119c0d44STheodore Ts'o BUFFER_TRACE(group_desc_bh, "get_write_access"); 820119c0d44STheodore Ts'o err = ext4_journal_get_write_access(handle, group_desc_bh); 821eb9cc7e1SJan Kara if (err) { 822eb9cc7e1SJan Kara ext4_std_error(sb, err); 823eb9cc7e1SJan Kara goto out; 824eb9cc7e1SJan Kara } 825119c0d44STheodore Ts'o 826119c0d44STheodore Ts'o /* Update the relevant bg descriptor fields */ 82741a246d1SDarrick J. Wong if (ext4_has_group_desc_csum(sb)) { 828119c0d44STheodore Ts'o int free; 829119c0d44STheodore Ts'o struct ext4_group_info *grp = ext4_get_group_info(sb, group); 830119c0d44STheodore Ts'o 831119c0d44STheodore Ts'o down_read(&grp->alloc_sem); /* protect vs itable lazyinit */ 832119c0d44STheodore Ts'o ext4_lock_group(sb, group); /* while we modify the bg desc */ 833119c0d44STheodore Ts'o free = EXT4_INODES_PER_GROUP(sb) - 834119c0d44STheodore Ts'o ext4_itable_unused_count(sb, gdp); 835119c0d44STheodore Ts'o if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 836119c0d44STheodore Ts'o gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 837119c0d44STheodore Ts'o free = 0; 838119c0d44STheodore Ts'o } 839119c0d44STheodore Ts'o /* 840119c0d44STheodore Ts'o * Check the relative inode number against the last used 841119c0d44STheodore Ts'o * relative inode number in this group. if it is greater 842119c0d44STheodore Ts'o * we need to update the bg_itable_unused count 843119c0d44STheodore Ts'o */ 844119c0d44STheodore Ts'o if (ino > free) 845119c0d44STheodore Ts'o ext4_itable_unused_set(sb, gdp, 846119c0d44STheodore Ts'o (EXT4_INODES_PER_GROUP(sb) - ino)); 847119c0d44STheodore Ts'o up_read(&grp->alloc_sem); 8486f2e9f0eSTao Ma } else { 8496f2e9f0eSTao Ma ext4_lock_group(sb, group); 850119c0d44STheodore Ts'o } 8516f2e9f0eSTao Ma 852119c0d44STheodore Ts'o ext4_free_inodes_set(sb, gdp, ext4_free_inodes_count(sb, gdp) - 1); 853119c0d44STheodore Ts'o if (S_ISDIR(mode)) { 854119c0d44STheodore Ts'o ext4_used_dirs_set(sb, gdp, ext4_used_dirs_count(sb, gdp) + 1); 855119c0d44STheodore Ts'o if (sbi->s_log_groups_per_flex) { 856119c0d44STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 857119c0d44STheodore Ts'o 858119c0d44STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].used_dirs); 859119c0d44STheodore Ts'o } 860119c0d44STheodore Ts'o } 86141a246d1SDarrick J. Wong if (ext4_has_group_desc_csum(sb)) { 86241a246d1SDarrick J. Wong ext4_inode_bitmap_csum_set(sb, group, gdp, inode_bitmap_bh, 86341a246d1SDarrick J. Wong EXT4_INODES_PER_GROUP(sb) / 8); 864feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, group, gdp); 865119c0d44STheodore Ts'o } 8666f2e9f0eSTao Ma ext4_unlock_group(sb, group); 867119c0d44STheodore Ts'o 8683300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 8693300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 870eb9cc7e1SJan Kara if (err) { 871eb9cc7e1SJan Kara ext4_std_error(sb, err); 872eb9cc7e1SJan Kara goto out; 873eb9cc7e1SJan Kara } 874ac27a0ecSDave Kleikamp 875ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 876ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 877ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 878ac27a0ecSDave Kleikamp 879772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 880772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 8819f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 882772cb7c8SJose R. Santos } 883ac27a0ecSDave Kleikamp 884717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 885ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 886ac27a0ecSDave Kleikamp inode->i_blocks = 0; 887ef7f3835SKalpak Shah inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime = 888ef7f3835SKalpak Shah ext4_current_time(inode); 889ac27a0ecSDave Kleikamp 890ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 891ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 892ac27a0ecSDave Kleikamp ei->i_disksize = 0; 893ac27a0ecSDave Kleikamp 8944af83508SEryu Guan /* Don't inherit extent flag from directory, amongst others. */ 8952dc6b0d4SDuane Griffin ei->i_flags = 8962dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 897ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 898ac27a0ecSDave Kleikamp ei->i_dtime = 0; 899ac27a0ecSDave Kleikamp ei->i_block_group = group; 900a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 901ac27a0ecSDave Kleikamp 902617ba13bSMingming Cao ext4_set_inode_flags(inode); 903ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 9040390131bSFrank Mayhar ext4_handle_sync(handle); 9056b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 906acd6ad83SJan Kara /* 907acd6ad83SJan Kara * Likely a bitmap corruption causing inode to be allocated 908acd6ad83SJan Kara * twice. 909acd6ad83SJan Kara */ 910acd6ad83SJan Kara err = -EIO; 911eb9cc7e1SJan Kara ext4_error(sb, "failed to insert inode %lu: doubly allocated?", 912eb9cc7e1SJan Kara inode->i_ino); 913eb9cc7e1SJan Kara goto out; 9146b38e842SAl Viro } 915ac27a0ecSDave Kleikamp spin_lock(&sbi->s_next_gen_lock); 916ac27a0ecSDave Kleikamp inode->i_generation = sbi->s_next_generation++; 917ac27a0ecSDave Kleikamp spin_unlock(&sbi->s_next_gen_lock); 918ac27a0ecSDave Kleikamp 919814525f4SDarrick J. Wong /* Precompute checksum seed for inode metadata */ 920814525f4SDarrick J. Wong if (EXT4_HAS_RO_COMPAT_FEATURE(sb, 921814525f4SDarrick J. Wong EXT4_FEATURE_RO_COMPAT_METADATA_CSUM)) { 922814525f4SDarrick J. Wong __u32 csum; 923814525f4SDarrick J. Wong __le32 inum = cpu_to_le32(inode->i_ino); 924814525f4SDarrick J. Wong __le32 gen = cpu_to_le32(inode->i_generation); 925814525f4SDarrick J. Wong csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum, 926814525f4SDarrick J. Wong sizeof(inum)); 927814525f4SDarrick J. Wong ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen, 928814525f4SDarrick J. Wong sizeof(gen)); 929814525f4SDarrick J. Wong } 930814525f4SDarrick J. Wong 931353eb83cSTheodore Ts'o ext4_clear_state_flags(ei); /* Only relevant on 32-bit archs */ 93219f5fb7aSTheodore Ts'o ext4_set_inode_state(inode, EXT4_STATE_NEW); 933ef7f3835SKalpak Shah 934ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 935ac27a0ecSDave Kleikamp 936f08225d1STao Ma ei->i_inline_off = 0; 937f08225d1STao Ma if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_INLINE_DATA)) 938f08225d1STao Ma ext4_set_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA); 939f08225d1STao Ma 940ac27a0ecSDave Kleikamp ret = inode; 94163936ddaSChristoph Hellwig err = dquot_alloc_inode(inode); 94263936ddaSChristoph Hellwig if (err) 943ac27a0ecSDave Kleikamp goto fail_drop; 944ac27a0ecSDave Kleikamp 945617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 946ac27a0ecSDave Kleikamp if (err) 947ac27a0ecSDave Kleikamp goto fail_free_drop; 948ac27a0ecSDave Kleikamp 9492a7dba39SEric Paris err = ext4_init_security(handle, inode, dir, qstr); 950ac27a0ecSDave Kleikamp if (err) 951ac27a0ecSDave Kleikamp goto fail_free_drop; 952ac27a0ecSDave Kleikamp 95383982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 954e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 955e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 95612e9b892SDmitry Monakhov ext4_set_inode_flag(inode, EXT4_INODE_EXTENTS); 957a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 958a86c6181SAlex Tomas } 95942bf0383SAneesh Kumar K.V } 960ac27a0ecSDave Kleikamp 961688f869cSTheodore Ts'o if (ext4_handle_valid(handle)) { 962688f869cSTheodore Ts'o ei->i_sync_tid = handle->h_transaction->t_tid; 963688f869cSTheodore Ts'o ei->i_datasync_tid = handle->h_transaction->t_tid; 964688f869cSTheodore Ts'o } 965688f869cSTheodore Ts'o 9668753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 9678753e88fSAneesh Kumar K.V if (err) { 9688753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 9698753e88fSAneesh Kumar K.V goto fail_free_drop; 9708753e88fSAneesh Kumar K.V } 9718753e88fSAneesh Kumar K.V 972617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 9739bffad1eSTheodore Ts'o trace_ext4_allocate_inode(inode, dir, mode); 9743300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 975ac27a0ecSDave Kleikamp return ret; 976ac27a0ecSDave Kleikamp 977ac27a0ecSDave Kleikamp fail_free_drop: 97863936ddaSChristoph Hellwig dquot_free_inode(inode); 979ac27a0ecSDave Kleikamp fail_drop: 9806d6b77f1SMiklos Szeredi clear_nlink(inode); 9816b38e842SAl Viro unlock_new_inode(inode); 982eb9cc7e1SJan Kara out: 983eb9cc7e1SJan Kara dquot_drop(inode); 984eb9cc7e1SJan Kara inode->i_flags |= S_NOQUOTA; 985ac27a0ecSDave Kleikamp iput(inode); 9863300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 987ac27a0ecSDave Kleikamp return ERR_PTR(err); 988ac27a0ecSDave Kleikamp } 989ac27a0ecSDave Kleikamp 990ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 991617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 992ac27a0ecSDave Kleikamp { 993617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 994fd2d4291SAvantika Mathur ext4_group_t block_group; 995ac27a0ecSDave Kleikamp int bit; 9961d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 997ac27a0ecSDave Kleikamp struct inode *inode = NULL; 9981d1fe1eeSDavid Howells long err = -EIO; 999ac27a0ecSDave Kleikamp 1000ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1001ac27a0ecSDave Kleikamp if (ino > max_ino) { 100212062dddSEric Sandeen ext4_warning(sb, "bad orphan ino %lu! e2fsck was run?", ino); 10031d1fe1eeSDavid Howells goto error; 1004ac27a0ecSDave Kleikamp } 1005ac27a0ecSDave Kleikamp 1006617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1007617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1008e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1009ac27a0ecSDave Kleikamp if (!bitmap_bh) { 101012062dddSEric Sandeen ext4_warning(sb, "inode bitmap error for orphan %lu", ino); 10111d1fe1eeSDavid Howells goto error; 1012ac27a0ecSDave Kleikamp } 1013ac27a0ecSDave Kleikamp 1014ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1015ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1016ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1017ac27a0ecSDave Kleikamp */ 10181d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 10191d1fe1eeSDavid Howells goto bad_orphan; 10201d1fe1eeSDavid Howells 10211d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 10221d1fe1eeSDavid Howells if (IS_ERR(inode)) 10231d1fe1eeSDavid Howells goto iget_failed; 10241d1fe1eeSDavid Howells 102591ef4cafSDuane Griffin /* 102691ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 102791ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 102891ef4cafSDuane Griffin * during processing and an infinite loop will result. 102991ef4cafSDuane Griffin */ 103091ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 103191ef4cafSDuane Griffin goto bad_orphan; 103291ef4cafSDuane Griffin 10331d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 10341d1fe1eeSDavid Howells goto bad_orphan; 10351d1fe1eeSDavid Howells brelse(bitmap_bh); 10361d1fe1eeSDavid Howells return inode; 10371d1fe1eeSDavid Howells 10381d1fe1eeSDavid Howells iget_failed: 10391d1fe1eeSDavid Howells err = PTR_ERR(inode); 10401d1fe1eeSDavid Howells inode = NULL; 10411d1fe1eeSDavid Howells bad_orphan: 104212062dddSEric Sandeen ext4_warning(sb, "bad orphan inode %lu! e2fsck was run?", ino); 10438de5c325STheodore Ts'o printk(KERN_WARNING "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1044ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1045617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 10468de5c325STheodore Ts'o printk(KERN_WARNING "inode=%p\n", inode); 1047ac27a0ecSDave Kleikamp if (inode) { 10488de5c325STheodore Ts'o printk(KERN_WARNING "is_bad_inode(inode)=%d\n", 1049ac27a0ecSDave Kleikamp is_bad_inode(inode)); 10508de5c325STheodore Ts'o printk(KERN_WARNING "NEXT_ORPHAN(inode)=%u\n", 1051ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 10528de5c325STheodore Ts'o printk(KERN_WARNING "max_ino=%lu\n", max_ino); 10538de5c325STheodore Ts'o printk(KERN_WARNING "i_nlink=%u\n", inode->i_nlink); 1054ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 10551d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1056ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1057ac27a0ecSDave Kleikamp iput(inode); 1058ac27a0ecSDave Kleikamp } 1059ac27a0ecSDave Kleikamp brelse(bitmap_bh); 10601d1fe1eeSDavid Howells error: 10611d1fe1eeSDavid Howells return ERR_PTR(err); 1062ac27a0ecSDave Kleikamp } 1063ac27a0ecSDave Kleikamp 1064617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1065ac27a0ecSDave Kleikamp { 1066ac27a0ecSDave Kleikamp unsigned long desc_count; 1067617ba13bSMingming Cao struct ext4_group_desc *gdp; 10688df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1069617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1070617ba13bSMingming Cao struct ext4_super_block *es; 1071ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1072ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1073ac27a0ecSDave Kleikamp 1074617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1075ac27a0ecSDave Kleikamp desc_count = 0; 1076ac27a0ecSDave Kleikamp bitmap_count = 0; 1077ac27a0ecSDave Kleikamp gdp = NULL; 10788df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1079617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1080ac27a0ecSDave Kleikamp if (!gdp) 1081ac27a0ecSDave Kleikamp continue; 1082560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1083ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1084e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1085ac27a0ecSDave Kleikamp if (!bitmap_bh) 1086ac27a0ecSDave Kleikamp continue; 1087ac27a0ecSDave Kleikamp 1088f6fb99caSTheodore Ts'o x = ext4_count_free(bitmap_bh->b_data, 1089f6fb99caSTheodore Ts'o EXT4_INODES_PER_GROUP(sb) / 8); 1090c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1091785b4b3aSPeng Tao (unsigned long) i, ext4_free_inodes_count(sb, gdp), x); 1092ac27a0ecSDave Kleikamp bitmap_count += x; 1093ac27a0ecSDave Kleikamp } 1094ac27a0ecSDave Kleikamp brelse(bitmap_bh); 10954776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 10964776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1097ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1098ac27a0ecSDave Kleikamp return desc_count; 1099ac27a0ecSDave Kleikamp #else 1100ac27a0ecSDave Kleikamp desc_count = 0; 11018df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1102617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1103ac27a0ecSDave Kleikamp if (!gdp) 1104ac27a0ecSDave Kleikamp continue; 1105560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1106ac27a0ecSDave Kleikamp cond_resched(); 1107ac27a0ecSDave Kleikamp } 1108ac27a0ecSDave Kleikamp return desc_count; 1109ac27a0ecSDave Kleikamp #endif 1110ac27a0ecSDave Kleikamp } 1111ac27a0ecSDave Kleikamp 1112ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1113617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1114ac27a0ecSDave Kleikamp { 1115ac27a0ecSDave Kleikamp unsigned long count = 0; 11168df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 1117ac27a0ecSDave Kleikamp 11188df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 1119617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1120ac27a0ecSDave Kleikamp if (!gdp) 1121ac27a0ecSDave Kleikamp continue; 1122560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1123ac27a0ecSDave Kleikamp } 1124ac27a0ecSDave Kleikamp return count; 1125ac27a0ecSDave Kleikamp } 1126bfff6873SLukas Czerner 1127bfff6873SLukas Czerner /* 1128bfff6873SLukas Czerner * Zeroes not yet zeroed inode table - just write zeroes through the whole 1129bfff6873SLukas Czerner * inode table. Must be called without any spinlock held. The only place 1130bfff6873SLukas Czerner * where it is called from on active part of filesystem is ext4lazyinit 1131bfff6873SLukas Czerner * thread, so we do not need any special locks, however we have to prevent 1132bfff6873SLukas Czerner * inode allocation from the current group, so we take alloc_sem lock, to 1133119c0d44STheodore Ts'o * block ext4_new_inode() until we are finished. 1134bfff6873SLukas Czerner */ 1135e0cbee3eSH Hartley Sweeten int ext4_init_inode_table(struct super_block *sb, ext4_group_t group, 1136bfff6873SLukas Czerner int barrier) 1137bfff6873SLukas Czerner { 1138bfff6873SLukas Czerner struct ext4_group_info *grp = ext4_get_group_info(sb, group); 1139bfff6873SLukas Czerner struct ext4_sb_info *sbi = EXT4_SB(sb); 1140bfff6873SLukas Czerner struct ext4_group_desc *gdp = NULL; 1141bfff6873SLukas Czerner struct buffer_head *group_desc_bh; 1142bfff6873SLukas Czerner handle_t *handle; 1143bfff6873SLukas Czerner ext4_fsblk_t blk; 1144bfff6873SLukas Czerner int num, ret = 0, used_blks = 0; 1145bfff6873SLukas Czerner 1146bfff6873SLukas Czerner /* This should not happen, but just to be sure check this */ 1147bfff6873SLukas Czerner if (sb->s_flags & MS_RDONLY) { 1148bfff6873SLukas Czerner ret = 1; 1149bfff6873SLukas Czerner goto out; 1150bfff6873SLukas Czerner } 1151bfff6873SLukas Czerner 1152bfff6873SLukas Czerner gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 1153bfff6873SLukas Czerner if (!gdp) 1154bfff6873SLukas Czerner goto out; 1155bfff6873SLukas Czerner 1156bfff6873SLukas Czerner /* 1157bfff6873SLukas Czerner * We do not need to lock this, because we are the only one 1158bfff6873SLukas Czerner * handling this flag. 1159bfff6873SLukas Czerner */ 1160bfff6873SLukas Czerner if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_ZEROED)) 1161bfff6873SLukas Czerner goto out; 1162bfff6873SLukas Czerner 11639924a92aSTheodore Ts'o handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1); 1164bfff6873SLukas Czerner if (IS_ERR(handle)) { 1165bfff6873SLukas Czerner ret = PTR_ERR(handle); 1166bfff6873SLukas Czerner goto out; 1167bfff6873SLukas Czerner } 1168bfff6873SLukas Czerner 1169bfff6873SLukas Czerner down_write(&grp->alloc_sem); 1170bfff6873SLukas Czerner /* 1171bfff6873SLukas Czerner * If inode bitmap was already initialized there may be some 1172bfff6873SLukas Czerner * used inodes so we need to skip blocks with used inodes in 1173bfff6873SLukas Czerner * inode table. 1174bfff6873SLukas Czerner */ 1175bfff6873SLukas Czerner if (!(gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT))) 1176bfff6873SLukas Czerner used_blks = DIV_ROUND_UP((EXT4_INODES_PER_GROUP(sb) - 1177bfff6873SLukas Czerner ext4_itable_unused_count(sb, gdp)), 1178bfff6873SLukas Czerner sbi->s_inodes_per_block); 1179bfff6873SLukas Czerner 1180857ac889SLukas Czerner if ((used_blks < 0) || (used_blks > sbi->s_itb_per_group)) { 11811084f252STheodore Ts'o ext4_error(sb, "Something is wrong with group %u: " 11821084f252STheodore Ts'o "used itable blocks: %d; " 11831084f252STheodore Ts'o "itable unused count: %u", 1184857ac889SLukas Czerner group, used_blks, 1185857ac889SLukas Czerner ext4_itable_unused_count(sb, gdp)); 1186857ac889SLukas Czerner ret = 1; 118733853a0dSYongqiang Yang goto err_out; 1188857ac889SLukas Czerner } 1189857ac889SLukas Czerner 1190bfff6873SLukas Czerner blk = ext4_inode_table(sb, gdp) + used_blks; 1191bfff6873SLukas Czerner num = sbi->s_itb_per_group - used_blks; 1192bfff6873SLukas Czerner 1193bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, "get_write_access"); 1194bfff6873SLukas Czerner ret = ext4_journal_get_write_access(handle, 1195bfff6873SLukas Czerner group_desc_bh); 1196bfff6873SLukas Czerner if (ret) 1197bfff6873SLukas Czerner goto err_out; 1198bfff6873SLukas Czerner 1199bfff6873SLukas Czerner /* 1200bfff6873SLukas Czerner * Skip zeroout if the inode table is full. But we set the ZEROED 1201bfff6873SLukas Czerner * flag anyway, because obviously, when it is full it does not need 1202bfff6873SLukas Czerner * further zeroing. 1203bfff6873SLukas Czerner */ 1204bfff6873SLukas Czerner if (unlikely(num == 0)) 1205bfff6873SLukas Czerner goto skip_zeroout; 1206bfff6873SLukas Czerner 1207bfff6873SLukas Czerner ext4_debug("going to zero out inode table in group %d\n", 1208bfff6873SLukas Czerner group); 1209a107e5a3STheodore Ts'o ret = sb_issue_zeroout(sb, blk, num, GFP_NOFS); 1210bfff6873SLukas Czerner if (ret < 0) 1211bfff6873SLukas Czerner goto err_out; 1212a107e5a3STheodore Ts'o if (barrier) 1213a107e5a3STheodore Ts'o blkdev_issue_flush(sb->s_bdev, GFP_NOFS, NULL); 1214bfff6873SLukas Czerner 1215bfff6873SLukas Czerner skip_zeroout: 1216bfff6873SLukas Czerner ext4_lock_group(sb, group); 1217bfff6873SLukas Czerner gdp->bg_flags |= cpu_to_le16(EXT4_BG_INODE_ZEROED); 1218feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, group, gdp); 1219bfff6873SLukas Czerner ext4_unlock_group(sb, group); 1220bfff6873SLukas Czerner 1221bfff6873SLukas Czerner BUFFER_TRACE(group_desc_bh, 1222bfff6873SLukas Czerner "call ext4_handle_dirty_metadata"); 1223bfff6873SLukas Czerner ret = ext4_handle_dirty_metadata(handle, NULL, 1224bfff6873SLukas Czerner group_desc_bh); 1225bfff6873SLukas Czerner 1226bfff6873SLukas Czerner err_out: 1227bfff6873SLukas Czerner up_write(&grp->alloc_sem); 1228bfff6873SLukas Czerner ext4_journal_stop(handle); 1229bfff6873SLukas Czerner out: 1230bfff6873SLukas Czerner return ret; 1231bfff6873SLukas Czerner } 1232