1ac27a0ecSDave Kleikamp /* 2617ba13bSMingming Cao * linux/fs/ext4/ialloc.c 3ac27a0ecSDave Kleikamp * 4ac27a0ecSDave Kleikamp * Copyright (C) 1992, 1993, 1994, 1995 5ac27a0ecSDave Kleikamp * Remy Card (card@masi.ibp.fr) 6ac27a0ecSDave Kleikamp * Laboratoire MASI - Institut Blaise Pascal 7ac27a0ecSDave Kleikamp * Universite Pierre et Marie Curie (Paris VI) 8ac27a0ecSDave Kleikamp * 9ac27a0ecSDave Kleikamp * BSD ufs-inspired inode and directory allocation by 10ac27a0ecSDave Kleikamp * Stephen Tweedie (sct@redhat.com), 1993 11ac27a0ecSDave Kleikamp * Big-endian to little-endian byte-swapping/bitmaps by 12ac27a0ecSDave Kleikamp * David S. Miller (davem@caip.rutgers.edu), 1995 13ac27a0ecSDave Kleikamp */ 14ac27a0ecSDave Kleikamp 15ac27a0ecSDave Kleikamp #include <linux/time.h> 16ac27a0ecSDave Kleikamp #include <linux/fs.h> 17dab291afSMingming Cao #include <linux/jbd2.h> 18ac27a0ecSDave Kleikamp #include <linux/stat.h> 19ac27a0ecSDave Kleikamp #include <linux/string.h> 20ac27a0ecSDave Kleikamp #include <linux/quotaops.h> 21ac27a0ecSDave Kleikamp #include <linux/buffer_head.h> 22ac27a0ecSDave Kleikamp #include <linux/random.h> 23ac27a0ecSDave Kleikamp #include <linux/bitops.h> 243a5b2ecdSMingming Cao #include <linux/blkdev.h> 25ac27a0ecSDave Kleikamp #include <asm/byteorder.h> 263dcf5451SChristoph Hellwig #include "ext4.h" 273dcf5451SChristoph Hellwig #include "ext4_jbd2.h" 28ac27a0ecSDave Kleikamp #include "xattr.h" 29ac27a0ecSDave Kleikamp #include "acl.h" 30717d50e4SAndreas Dilger #include "group.h" 31ac27a0ecSDave Kleikamp 32ac27a0ecSDave Kleikamp /* 33ac27a0ecSDave Kleikamp * ialloc.c contains the inodes allocation and deallocation routines 34ac27a0ecSDave Kleikamp */ 35ac27a0ecSDave Kleikamp 36ac27a0ecSDave Kleikamp /* 37ac27a0ecSDave Kleikamp * The free inodes are managed by bitmaps. A file system contains several 38ac27a0ecSDave Kleikamp * blocks groups. Each group contains 1 bitmap block for blocks, 1 bitmap 39ac27a0ecSDave Kleikamp * block for inodes, N blocks for the inode table and data blocks. 40ac27a0ecSDave Kleikamp * 41ac27a0ecSDave Kleikamp * The file system contains group descriptors which are located after the 42ac27a0ecSDave Kleikamp * super block. Each descriptor contains the number of the bitmap block and 43ac27a0ecSDave Kleikamp * the free blocks count in the block. 44ac27a0ecSDave Kleikamp */ 45ac27a0ecSDave Kleikamp 46717d50e4SAndreas Dilger /* 47717d50e4SAndreas Dilger * To avoid calling the atomic setbit hundreds or thousands of times, we only 48717d50e4SAndreas Dilger * need to use it within a single byte (to ensure we get endianness right). 49717d50e4SAndreas Dilger * We can use memset for the rest of the bitmap as there are no other users. 50717d50e4SAndreas Dilger */ 51717d50e4SAndreas Dilger void mark_bitmap_end(int start_bit, int end_bit, char *bitmap) 52717d50e4SAndreas Dilger { 53717d50e4SAndreas Dilger int i; 54717d50e4SAndreas Dilger 55717d50e4SAndreas Dilger if (start_bit >= end_bit) 56717d50e4SAndreas Dilger return; 57717d50e4SAndreas Dilger 58717d50e4SAndreas Dilger ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit); 59717d50e4SAndreas Dilger for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++) 60717d50e4SAndreas Dilger ext4_set_bit(i, bitmap); 61717d50e4SAndreas Dilger if (i < end_bit) 62717d50e4SAndreas Dilger memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3); 63717d50e4SAndreas Dilger } 64717d50e4SAndreas Dilger 65717d50e4SAndreas Dilger /* Initializes an uninitialized inode bitmap */ 66fd2d4291SAvantika Mathur unsigned ext4_init_inode_bitmap(struct super_block *sb, struct buffer_head *bh, 67fd2d4291SAvantika Mathur ext4_group_t block_group, 68717d50e4SAndreas Dilger struct ext4_group_desc *gdp) 69717d50e4SAndreas Dilger { 70717d50e4SAndreas Dilger struct ext4_sb_info *sbi = EXT4_SB(sb); 71717d50e4SAndreas Dilger 72717d50e4SAndreas Dilger J_ASSERT_BH(bh, buffer_locked(bh)); 73717d50e4SAndreas Dilger 74717d50e4SAndreas Dilger /* If checksum is bad mark all blocks and inodes use to prevent 75717d50e4SAndreas Dilger * allocation, essentially implementing a per-group read-only flag. */ 76717d50e4SAndreas Dilger if (!ext4_group_desc_csum_verify(sbi, block_group, gdp)) { 77a9df9a49STheodore Ts'o ext4_error(sb, __func__, "Checksum bad for group %u", 78717d50e4SAndreas Dilger block_group); 79560671a0SAneesh Kumar K.V ext4_free_blks_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); 83717d50e4SAndreas Dilger return 0; 84717d50e4SAndreas Dilger } 85717d50e4SAndreas Dilger 86717d50e4SAndreas Dilger memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8); 87648f5879SAneesh Kumar K.V mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), sb->s_blocksize * 8, 88717d50e4SAndreas Dilger bh->b_data); 89717d50e4SAndreas Dilger 90717d50e4SAndreas Dilger return EXT4_INODES_PER_GROUP(sb); 91717d50e4SAndreas Dilger } 92ac27a0ecSDave Kleikamp 93ac27a0ecSDave Kleikamp /* 94ac27a0ecSDave Kleikamp * Read the inode allocation bitmap for a given block_group, reading 95ac27a0ecSDave Kleikamp * into the specified slot in the superblock's bitmap cache. 96ac27a0ecSDave Kleikamp * 97ac27a0ecSDave Kleikamp * Return buffer_head of bitmap on success or NULL. 98ac27a0ecSDave Kleikamp */ 99ac27a0ecSDave Kleikamp static struct buffer_head * 100e29d1cdeSEric Sandeen ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group) 101ac27a0ecSDave Kleikamp { 102617ba13bSMingming Cao struct ext4_group_desc *desc; 103ac27a0ecSDave Kleikamp struct buffer_head *bh = NULL; 104e29d1cdeSEric Sandeen ext4_fsblk_t bitmap_blk; 105ac27a0ecSDave Kleikamp 106617ba13bSMingming Cao desc = ext4_get_group_desc(sb, block_group, NULL); 107ac27a0ecSDave Kleikamp if (!desc) 108e29d1cdeSEric Sandeen return NULL; 109e29d1cdeSEric Sandeen bitmap_blk = ext4_inode_bitmap(sb, desc); 110e29d1cdeSEric Sandeen bh = sb_getblk(sb, bitmap_blk); 111e29d1cdeSEric Sandeen if (unlikely(!bh)) { 112e29d1cdeSEric Sandeen ext4_error(sb, __func__, 113ac27a0ecSDave Kleikamp "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 } 126b5f10eedSEric Sandeen spin_lock(sb_bgl_lock(EXT4_SB(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); 131b5f10eedSEric Sandeen spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group)); 1323300bedaSAneesh Kumar K.V unlock_buffer(bh); 133e29d1cdeSEric Sandeen return bh; 134e29d1cdeSEric Sandeen } 135b5f10eedSEric Sandeen spin_unlock(sb_bgl_lock(EXT4_SB(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); 154e29d1cdeSEric Sandeen ext4_error(sb, __func__, 155e29d1cdeSEric Sandeen "Cannot read inode bitmap - " 156a9df9a49STheodore Ts'o "block_group = %u, inode_bitmap = %llu", 157e29d1cdeSEric Sandeen block_group, bitmap_blk); 158e29d1cdeSEric Sandeen return NULL; 159e29d1cdeSEric Sandeen } 160ac27a0ecSDave Kleikamp return bh; 161ac27a0ecSDave Kleikamp } 162ac27a0ecSDave Kleikamp 163ac27a0ecSDave Kleikamp /* 164ac27a0ecSDave Kleikamp * NOTE! When we get the inode, we're the only people 165ac27a0ecSDave Kleikamp * that have access to it, and as such there are no 166ac27a0ecSDave Kleikamp * race conditions we have to worry about. The inode 167ac27a0ecSDave Kleikamp * is not on the hash-lists, and it cannot be reached 168ac27a0ecSDave Kleikamp * through the filesystem because the directory entry 169ac27a0ecSDave Kleikamp * has been deleted earlier. 170ac27a0ecSDave Kleikamp * 171ac27a0ecSDave Kleikamp * HOWEVER: we must make sure that we get no aliases, 172ac27a0ecSDave Kleikamp * which means that we have to call "clear_inode()" 173ac27a0ecSDave Kleikamp * _before_ we mark the inode not in use in the inode 174ac27a0ecSDave Kleikamp * bitmaps. Otherwise a newly created file might use 175ac27a0ecSDave Kleikamp * the same inode number (not actually the same pointer 176ac27a0ecSDave Kleikamp * though), and then we'd have two inodes sharing the 177ac27a0ecSDave Kleikamp * same inode number and space on the harddisk. 178ac27a0ecSDave Kleikamp */ 179617ba13bSMingming Cao void ext4_free_inode(handle_t *handle, struct inode *inode) 180ac27a0ecSDave Kleikamp { 181ac27a0ecSDave Kleikamp struct super_block *sb = inode->i_sb; 182ac27a0ecSDave Kleikamp int is_directory; 183ac27a0ecSDave Kleikamp unsigned long ino; 184ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 185ac27a0ecSDave Kleikamp struct buffer_head *bh2; 186fd2d4291SAvantika Mathur ext4_group_t block_group; 187ac27a0ecSDave Kleikamp unsigned long bit; 188617ba13bSMingming Cao struct ext4_group_desc *gdp; 189617ba13bSMingming Cao struct ext4_super_block *es; 190617ba13bSMingming Cao struct ext4_sb_info *sbi; 1917ce9d5d1SEric Sandeen int fatal = 0, err, count, cleared; 192ac27a0ecSDave Kleikamp 193ac27a0ecSDave Kleikamp if (atomic_read(&inode->i_count) > 1) { 1944776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has count=%d\n", 195ac27a0ecSDave Kleikamp atomic_read(&inode->i_count)); 196ac27a0ecSDave Kleikamp return; 197ac27a0ecSDave Kleikamp } 198ac27a0ecSDave Kleikamp if (inode->i_nlink) { 1994776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n", 200ac27a0ecSDave Kleikamp inode->i_nlink); 201ac27a0ecSDave Kleikamp return; 202ac27a0ecSDave Kleikamp } 203ac27a0ecSDave Kleikamp if (!sb) { 2044776004fSTheodore Ts'o printk(KERN_ERR "ext4_free_inode: inode on " 2054776004fSTheodore Ts'o "nonexistent device\n"); 206ac27a0ecSDave Kleikamp return; 207ac27a0ecSDave Kleikamp } 208617ba13bSMingming Cao sbi = EXT4_SB(sb); 209ac27a0ecSDave Kleikamp 210ac27a0ecSDave Kleikamp ino = inode->i_ino; 211617ba13bSMingming Cao ext4_debug("freeing inode %lu\n", ino); 212ba80b101STheodore Ts'o trace_mark(ext4_free_inode, 213ba80b101STheodore Ts'o "dev %s ino %lu mode %d uid %lu gid %lu bocks %llu", 214ba80b101STheodore Ts'o sb->s_id, inode->i_ino, inode->i_mode, 215ba80b101STheodore Ts'o (unsigned long) inode->i_uid, (unsigned long) inode->i_gid, 216ba80b101STheodore Ts'o (unsigned long long) inode->i_blocks); 217ac27a0ecSDave Kleikamp 218ac27a0ecSDave Kleikamp /* 219ac27a0ecSDave Kleikamp * Note: we must free any quota before locking the superblock, 220ac27a0ecSDave Kleikamp * as writing the quota to disk may need the lock as well. 221ac27a0ecSDave Kleikamp */ 222a269eb18SJan Kara vfs_dq_init(inode); 223617ba13bSMingming Cao ext4_xattr_delete_inode(handle, inode); 224a269eb18SJan Kara vfs_dq_free_inode(inode); 225a269eb18SJan Kara vfs_dq_drop(inode); 226ac27a0ecSDave Kleikamp 227ac27a0ecSDave Kleikamp is_directory = S_ISDIR(inode->i_mode); 228ac27a0ecSDave Kleikamp 229ac27a0ecSDave Kleikamp /* Do this BEFORE marking the inode not in use or returning an error */ 230ac27a0ecSDave Kleikamp clear_inode(inode); 231ac27a0ecSDave Kleikamp 232617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 233617ba13bSMingming Cao if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) { 234617ba13bSMingming Cao ext4_error(sb, "ext4_free_inode", 235ac27a0ecSDave Kleikamp "reserved or nonexistent inode %lu", ino); 236ac27a0ecSDave Kleikamp goto error_return; 237ac27a0ecSDave Kleikamp } 238617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 239617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 240e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 241ac27a0ecSDave Kleikamp if (!bitmap_bh) 242ac27a0ecSDave Kleikamp goto error_return; 243ac27a0ecSDave Kleikamp 244ac27a0ecSDave Kleikamp BUFFER_TRACE(bitmap_bh, "get_write_access"); 245617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bitmap_bh); 246ac27a0ecSDave Kleikamp if (fatal) 247ac27a0ecSDave Kleikamp goto error_return; 248ac27a0ecSDave Kleikamp 249ac27a0ecSDave Kleikamp /* Ok, now we can actually update the inode bitmaps.. */ 2507ce9d5d1SEric Sandeen spin_lock(sb_bgl_lock(sbi, block_group)); 2517ce9d5d1SEric Sandeen cleared = ext4_clear_bit(bit, bitmap_bh->b_data); 2527ce9d5d1SEric Sandeen spin_unlock(sb_bgl_lock(sbi, block_group)); 2537ce9d5d1SEric Sandeen if (!cleared) 254617ba13bSMingming Cao ext4_error(sb, "ext4_free_inode", 255ac27a0ecSDave Kleikamp "bit already cleared for inode %lu", ino); 256ac27a0ecSDave Kleikamp else { 257617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, block_group, &bh2); 258ac27a0ecSDave Kleikamp 259ac27a0ecSDave Kleikamp BUFFER_TRACE(bh2, "get_write_access"); 260617ba13bSMingming Cao fatal = ext4_journal_get_write_access(handle, bh2); 261ac27a0ecSDave Kleikamp if (fatal) goto error_return; 262ac27a0ecSDave Kleikamp 263ac27a0ecSDave Kleikamp if (gdp) { 264ac27a0ecSDave Kleikamp spin_lock(sb_bgl_lock(sbi, block_group)); 265560671a0SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) + 1; 266560671a0SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 267560671a0SAneesh Kumar K.V if (is_directory) { 268560671a0SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) - 1; 269560671a0SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 2707d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 2717d39db14STheodore Ts'o ext4_group_t f; 2727d39db14STheodore Ts'o 2737d39db14STheodore Ts'o f = ext4_flex_group(sbi, block_group); 2747d39db14STheodore Ts'o atomic_dec(&sbi->s_flex_groups[f].free_inodes); 2757d39db14STheodore Ts'o } 2767d39db14STheodore Ts'o 277560671a0SAneesh Kumar K.V } 278717d50e4SAndreas Dilger gdp->bg_checksum = ext4_group_desc_csum(sbi, 279717d50e4SAndreas Dilger block_group, gdp); 280ac27a0ecSDave Kleikamp spin_unlock(sb_bgl_lock(sbi, block_group)); 281ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_freeinodes_counter); 282ac27a0ecSDave Kleikamp if (is_directory) 283ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_dirs_counter); 284ac27a0ecSDave Kleikamp 285772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2869f24e420STheodore Ts'o ext4_group_t f; 2879f24e420STheodore Ts'o 2889f24e420STheodore Ts'o f = ext4_flex_group(sbi, block_group); 2899f24e420STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].free_inodes); 290772cb7c8SJose R. Santos } 291ac27a0ecSDave Kleikamp } 2920390131bSFrank Mayhar BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata"); 2930390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bh2); 294ac27a0ecSDave Kleikamp if (!fatal) fatal = err; 295ac27a0ecSDave Kleikamp } 2960390131bSFrank Mayhar BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata"); 2970390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 298ac27a0ecSDave Kleikamp if (!fatal) 299ac27a0ecSDave Kleikamp fatal = err; 300ac27a0ecSDave Kleikamp sb->s_dirt = 1; 301ac27a0ecSDave Kleikamp error_return: 302ac27a0ecSDave Kleikamp brelse(bitmap_bh); 303617ba13bSMingming Cao ext4_std_error(sb, fatal); 304ac27a0ecSDave Kleikamp } 305ac27a0ecSDave Kleikamp 306ac27a0ecSDave Kleikamp /* 307ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 308ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 309ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 310ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 311ac27a0ecSDave Kleikamp * directories already is chosen. 312ac27a0ecSDave Kleikamp * 313ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory\'s block 314ac27a0ecSDave Kleikamp * group to find a free inode. 315ac27a0ecSDave Kleikamp */ 3162aa9fc4cSAvantika Mathur static int find_group_dir(struct super_block *sb, struct inode *parent, 3172aa9fc4cSAvantika Mathur ext4_group_t *best_group) 318ac27a0ecSDave Kleikamp { 319fd2d4291SAvantika Mathur ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count; 320ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 321617ba13bSMingming Cao struct ext4_group_desc *desc, *best_desc = NULL; 3222aa9fc4cSAvantika Mathur ext4_group_t group; 3232aa9fc4cSAvantika Mathur int ret = -1; 324ac27a0ecSDave Kleikamp 325617ba13bSMingming Cao freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter); 326ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 327ac27a0ecSDave Kleikamp 328ac27a0ecSDave Kleikamp for (group = 0; group < ngroups; group++) { 329ef2fb679SEric Sandeen desc = ext4_get_group_desc(sb, group, NULL); 330560671a0SAneesh Kumar K.V if (!desc || !ext4_free_inodes_count(sb, desc)) 331ac27a0ecSDave Kleikamp continue; 332560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc) < avefreei) 333ac27a0ecSDave Kleikamp continue; 334ac27a0ecSDave Kleikamp if (!best_desc || 335560671a0SAneesh Kumar K.V (ext4_free_blks_count(sb, desc) > 336560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, best_desc))) { 3372aa9fc4cSAvantika Mathur *best_group = group; 338ac27a0ecSDave Kleikamp best_desc = desc; 3392aa9fc4cSAvantika Mathur ret = 0; 340ac27a0ecSDave Kleikamp } 341ac27a0ecSDave Kleikamp } 3422aa9fc4cSAvantika Mathur return ret; 343ac27a0ecSDave Kleikamp } 344ac27a0ecSDave Kleikamp 345772cb7c8SJose R. Santos #define free_block_ratio 10 346772cb7c8SJose R. Santos 347772cb7c8SJose R. Santos static int find_group_flex(struct super_block *sb, struct inode *parent, 348772cb7c8SJose R. Santos ext4_group_t *best_group) 349772cb7c8SJose R. Santos { 350772cb7c8SJose R. Santos struct ext4_sb_info *sbi = EXT4_SB(sb); 351772cb7c8SJose R. Santos struct ext4_group_desc *desc; 352772cb7c8SJose R. Santos struct buffer_head *bh; 353772cb7c8SJose R. Santos struct flex_groups *flex_group = sbi->s_flex_groups; 354772cb7c8SJose R. Santos ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 355772cb7c8SJose R. Santos ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group); 356772cb7c8SJose R. Santos ext4_group_t ngroups = sbi->s_groups_count; 357772cb7c8SJose R. Santos int flex_size = ext4_flex_bg_size(sbi); 358772cb7c8SJose R. Santos ext4_group_t best_flex = parent_fbg_group; 359772cb7c8SJose R. Santos int blocks_per_flex = sbi->s_blocks_per_group * flex_size; 360772cb7c8SJose R. Santos int flexbg_free_blocks; 361772cb7c8SJose R. Santos int flex_freeb_ratio; 362772cb7c8SJose R. Santos ext4_group_t n_fbg_groups; 363772cb7c8SJose R. Santos ext4_group_t i; 364772cb7c8SJose R. Santos 365772cb7c8SJose R. Santos n_fbg_groups = (sbi->s_groups_count + flex_size - 1) >> 366772cb7c8SJose R. Santos sbi->s_log_groups_per_flex; 367772cb7c8SJose R. Santos 368772cb7c8SJose R. Santos find_close_to_parent: 3699f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[best_flex].free_blocks); 370772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 3719f24e420STheodore Ts'o if (atomic_read(&flex_group[best_flex].free_inodes) && 372772cb7c8SJose R. Santos flex_freeb_ratio > free_block_ratio) 373772cb7c8SJose R. Santos goto found_flexbg; 374772cb7c8SJose R. Santos 375772cb7c8SJose R. Santos if (best_flex && best_flex == parent_fbg_group) { 376772cb7c8SJose R. Santos best_flex--; 377772cb7c8SJose R. Santos goto find_close_to_parent; 378772cb7c8SJose R. Santos } 379772cb7c8SJose R. Santos 380772cb7c8SJose R. Santos for (i = 0; i < n_fbg_groups; i++) { 381772cb7c8SJose R. Santos if (i == parent_fbg_group || i == parent_fbg_group - 1) 382772cb7c8SJose R. Santos continue; 383772cb7c8SJose R. Santos 3849f24e420STheodore Ts'o flexbg_free_blocks = atomic_read(&flex_group[i].free_blocks); 385772cb7c8SJose R. Santos flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex; 386772cb7c8SJose R. Santos 387772cb7c8SJose R. Santos if (flex_freeb_ratio > free_block_ratio && 3889f24e420STheodore Ts'o (atomic_read(&flex_group[i].free_inodes))) { 389772cb7c8SJose R. Santos best_flex = i; 390772cb7c8SJose R. Santos goto found_flexbg; 391772cb7c8SJose R. Santos } 392772cb7c8SJose R. Santos 3939f24e420STheodore Ts'o if ((atomic_read(&flex_group[best_flex].free_inodes) == 0) || 3949f24e420STheodore Ts'o ((atomic_read(&flex_group[i].free_blocks) > 3959f24e420STheodore Ts'o atomic_read(&flex_group[best_flex].free_blocks)) && 3969f24e420STheodore Ts'o atomic_read(&flex_group[i].free_inodes))) 397772cb7c8SJose R. Santos best_flex = i; 398772cb7c8SJose R. Santos } 399772cb7c8SJose R. Santos 4009f24e420STheodore Ts'o if (!atomic_read(&flex_group[best_flex].free_inodes) || 4019f24e420STheodore Ts'o !atomic_read(&flex_group[best_flex].free_blocks)) 402772cb7c8SJose R. Santos return -1; 403772cb7c8SJose R. Santos 404772cb7c8SJose R. Santos found_flexbg: 405772cb7c8SJose R. Santos for (i = best_flex * flex_size; i < ngroups && 406772cb7c8SJose R. Santos i < (best_flex + 1) * flex_size; i++) { 407772cb7c8SJose R. Santos desc = ext4_get_group_desc(sb, i, &bh); 408560671a0SAneesh Kumar K.V if (ext4_free_inodes_count(sb, desc)) { 409772cb7c8SJose R. Santos *best_group = i; 410772cb7c8SJose R. Santos goto out; 411772cb7c8SJose R. Santos } 412772cb7c8SJose R. Santos } 413772cb7c8SJose R. Santos 414772cb7c8SJose R. Santos return -1; 415772cb7c8SJose R. Santos out: 416772cb7c8SJose R. Santos return 0; 417772cb7c8SJose R. Santos } 418772cb7c8SJose R. Santos 419a4912123STheodore Ts'o struct orlov_stats { 420a4912123STheodore Ts'o __u32 free_inodes; 421a4912123STheodore Ts'o __u32 free_blocks; 422a4912123STheodore Ts'o __u32 used_dirs; 423a4912123STheodore Ts'o }; 424a4912123STheodore Ts'o 425a4912123STheodore Ts'o /* 426a4912123STheodore Ts'o * Helper function for Orlov's allocator; returns critical information 427a4912123STheodore Ts'o * for a particular block group or flex_bg. If flex_size is 1, then g 428a4912123STheodore Ts'o * is a block group number; otherwise it is flex_bg number. 429a4912123STheodore Ts'o */ 430a4912123STheodore Ts'o void get_orlov_stats(struct super_block *sb, ext4_group_t g, 431a4912123STheodore Ts'o int flex_size, struct orlov_stats *stats) 432a4912123STheodore Ts'o { 433a4912123STheodore Ts'o struct ext4_group_desc *desc; 4347d39db14STheodore Ts'o struct flex_groups *flex_group = EXT4_SB(sb)->s_flex_groups; 435a4912123STheodore Ts'o 4367d39db14STheodore Ts'o if (flex_size > 1) { 4377d39db14STheodore Ts'o stats->free_inodes = atomic_read(&flex_group[g].free_inodes); 4387d39db14STheodore Ts'o stats->free_blocks = atomic_read(&flex_group[g].free_blocks); 4397d39db14STheodore Ts'o stats->used_dirs = atomic_read(&flex_group[g].used_dirs); 4407d39db14STheodore Ts'o return; 4417d39db14STheodore Ts'o } 4427d39db14STheodore Ts'o 4437d39db14STheodore Ts'o desc = ext4_get_group_desc(sb, g, NULL); 4447d39db14STheodore Ts'o if (desc) { 4457d39db14STheodore Ts'o stats->free_inodes = ext4_free_inodes_count(sb, desc); 4467d39db14STheodore Ts'o stats->free_blocks = ext4_free_blks_count(sb, desc); 4477d39db14STheodore Ts'o stats->used_dirs = ext4_used_dirs_count(sb, desc); 4487d39db14STheodore Ts'o } else { 449a4912123STheodore Ts'o stats->free_inodes = 0; 450a4912123STheodore Ts'o stats->free_blocks = 0; 451a4912123STheodore Ts'o stats->used_dirs = 0; 452a4912123STheodore Ts'o } 453a4912123STheodore Ts'o } 454a4912123STheodore Ts'o 455ac27a0ecSDave Kleikamp /* 456ac27a0ecSDave Kleikamp * Orlov's allocator for directories. 457ac27a0ecSDave Kleikamp * 458ac27a0ecSDave Kleikamp * We always try to spread first-level directories. 459ac27a0ecSDave Kleikamp * 460ac27a0ecSDave Kleikamp * If there are blockgroups with both free inodes and free blocks counts 461ac27a0ecSDave Kleikamp * not worse than average we return one with smallest directory count. 462ac27a0ecSDave Kleikamp * Otherwise we simply return a random group. 463ac27a0ecSDave Kleikamp * 464ac27a0ecSDave Kleikamp * For the rest rules look so: 465ac27a0ecSDave Kleikamp * 466ac27a0ecSDave Kleikamp * It's OK to put directory into a group unless 467ac27a0ecSDave Kleikamp * it has too many directories already (max_dirs) or 468ac27a0ecSDave Kleikamp * it has too few free inodes left (min_inodes) or 469ac27a0ecSDave Kleikamp * it has too few free blocks left (min_blocks) or 4701cc8dcf5SBenoit Boissinot * Parent's group is preferred, if it doesn't satisfy these 471ac27a0ecSDave Kleikamp * conditions we search cyclically through the rest. If none 472ac27a0ecSDave Kleikamp * of the groups look good we just look for a group with more 473ac27a0ecSDave Kleikamp * free inodes than average (starting at parent's group). 474ac27a0ecSDave Kleikamp */ 475ac27a0ecSDave Kleikamp 4762aa9fc4cSAvantika Mathur static int find_group_orlov(struct super_block *sb, struct inode *parent, 477a4912123STheodore Ts'o ext4_group_t *group, int mode) 478ac27a0ecSDave Kleikamp { 479fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 480617ba13bSMingming Cao struct ext4_sb_info *sbi = EXT4_SB(sb); 481fd2d4291SAvantika Mathur ext4_group_t ngroups = sbi->s_groups_count; 482617ba13bSMingming Cao int inodes_per_group = EXT4_INODES_PER_GROUP(sb); 483ac27a0ecSDave Kleikamp unsigned int freei, avefreei; 484617ba13bSMingming Cao ext4_fsblk_t freeb, avefreeb; 485ac27a0ecSDave Kleikamp unsigned int ndirs; 486a4912123STheodore Ts'o int max_dirs, min_inodes; 487617ba13bSMingming Cao ext4_grpblk_t min_blocks; 488a4912123STheodore Ts'o ext4_group_t i, grp, g; 489617ba13bSMingming Cao struct ext4_group_desc *desc; 490a4912123STheodore Ts'o struct orlov_stats stats; 491a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(sbi); 492a4912123STheodore Ts'o 493a4912123STheodore Ts'o if (flex_size > 1) { 494a4912123STheodore Ts'o ngroups = (ngroups + flex_size - 1) >> 495a4912123STheodore Ts'o sbi->s_log_groups_per_flex; 496a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 497a4912123STheodore Ts'o } 498ac27a0ecSDave Kleikamp 499ac27a0ecSDave Kleikamp freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter); 500ac27a0ecSDave Kleikamp avefreei = freei / ngroups; 501ac27a0ecSDave Kleikamp freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter); 5023a5b2ecdSMingming Cao avefreeb = freeb; 503f4e5bc24SAndrew Morton do_div(avefreeb, ngroups); 504ac27a0ecSDave Kleikamp ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter); 505ac27a0ecSDave Kleikamp 506a4912123STheodore Ts'o if (S_ISDIR(mode) && 507a4912123STheodore Ts'o ((parent == sb->s_root->d_inode) || 508a4912123STheodore Ts'o (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL))) { 509ac27a0ecSDave Kleikamp int best_ndir = inodes_per_group; 5102aa9fc4cSAvantika Mathur int ret = -1; 511ac27a0ecSDave Kleikamp 5122aa9fc4cSAvantika Mathur get_random_bytes(&grp, sizeof(grp)); 5132aa9fc4cSAvantika Mathur parent_group = (unsigned)grp % ngroups; 514ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 515a4912123STheodore Ts'o g = (parent_group + i) % ngroups; 516a4912123STheodore Ts'o get_orlov_stats(sb, g, flex_size, &stats); 517a4912123STheodore Ts'o if (!stats.free_inodes) 518ac27a0ecSDave Kleikamp continue; 519a4912123STheodore Ts'o if (stats.used_dirs >= best_ndir) 520ac27a0ecSDave Kleikamp continue; 521a4912123STheodore Ts'o if (stats.free_inodes < avefreei) 522ac27a0ecSDave Kleikamp continue; 523a4912123STheodore Ts'o if (stats.free_blocks < avefreeb) 524ac27a0ecSDave Kleikamp continue; 525a4912123STheodore Ts'o grp = g; 5262aa9fc4cSAvantika Mathur ret = 0; 527a4912123STheodore Ts'o best_ndir = stats.used_dirs; 528ac27a0ecSDave Kleikamp } 529a4912123STheodore Ts'o if (ret) 530a4912123STheodore Ts'o goto fallback; 531a4912123STheodore Ts'o found_flex_bg: 532a4912123STheodore Ts'o if (flex_size == 1) { 533a4912123STheodore Ts'o *group = grp; 534a4912123STheodore Ts'o return 0; 535a4912123STheodore Ts'o } 536a4912123STheodore Ts'o 537a4912123STheodore Ts'o /* 538a4912123STheodore Ts'o * We pack inodes at the beginning of the flexgroup's 539a4912123STheodore Ts'o * inode tables. Block allocation decisions will do 540a4912123STheodore Ts'o * something similar, although regular files will 541a4912123STheodore Ts'o * start at 2nd block group of the flexgroup. See 542a4912123STheodore Ts'o * ext4_ext_find_goal() and ext4_find_near(). 543a4912123STheodore Ts'o */ 544a4912123STheodore Ts'o grp *= flex_size; 545a4912123STheodore Ts'o for (i = 0; i < flex_size; i++) { 546a4912123STheodore Ts'o if (grp+i >= sbi->s_groups_count) 547a4912123STheodore Ts'o break; 548a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp+i, NULL); 549a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 550a4912123STheodore Ts'o *group = grp+i; 551a4912123STheodore Ts'o return 0; 552a4912123STheodore Ts'o } 553a4912123STheodore Ts'o } 554ac27a0ecSDave Kleikamp goto fallback; 555ac27a0ecSDave Kleikamp } 556ac27a0ecSDave Kleikamp 557ac27a0ecSDave Kleikamp max_dirs = ndirs / ngroups + inodes_per_group / 16; 558a4912123STheodore Ts'o min_inodes = avefreei - inodes_per_group*flex_size / 4; 559a4912123STheodore Ts'o if (min_inodes < 1) 560a4912123STheodore Ts'o min_inodes = 1; 561a4912123STheodore Ts'o min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb)*flex_size / 4; 562ac27a0ecSDave Kleikamp 563a4912123STheodore Ts'o /* 564a4912123STheodore Ts'o * Start looking in the flex group where we last allocated an 565a4912123STheodore Ts'o * inode for this parent directory 566a4912123STheodore Ts'o */ 567a4912123STheodore Ts'o if (EXT4_I(parent)->i_last_alloc_group != ~0) { 568a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 569a4912123STheodore Ts'o if (flex_size > 1) 570a4912123STheodore Ts'o parent_group >>= sbi->s_log_groups_per_flex; 571a4912123STheodore Ts'o } 572ac27a0ecSDave Kleikamp 573ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 574a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 575a4912123STheodore Ts'o get_orlov_stats(sb, grp, flex_size, &stats); 576a4912123STheodore Ts'o if (stats.used_dirs >= max_dirs) 577ac27a0ecSDave Kleikamp continue; 578a4912123STheodore Ts'o if (stats.free_inodes < min_inodes) 579ac27a0ecSDave Kleikamp continue; 580a4912123STheodore Ts'o if (stats.free_blocks < min_blocks) 581ac27a0ecSDave Kleikamp continue; 582a4912123STheodore Ts'o goto found_flex_bg; 583ac27a0ecSDave Kleikamp } 584ac27a0ecSDave Kleikamp 585ac27a0ecSDave Kleikamp fallback: 586a4912123STheodore Ts'o ngroups = sbi->s_groups_count; 587a4912123STheodore Ts'o avefreei = freei / ngroups; 588a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_block_group; 589ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 590a4912123STheodore Ts'o grp = (parent_group + i) % ngroups; 591a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, grp, NULL); 592560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 593a4912123STheodore Ts'o ext4_free_inodes_count(sb, desc) >= avefreei) { 594a4912123STheodore Ts'o *group = grp; 5952aa9fc4cSAvantika Mathur return 0; 596ac27a0ecSDave Kleikamp } 597a4912123STheodore Ts'o } 598ac27a0ecSDave Kleikamp 599ac27a0ecSDave Kleikamp if (avefreei) { 600ac27a0ecSDave Kleikamp /* 601ac27a0ecSDave Kleikamp * The free-inodes counter is approximate, and for really small 602ac27a0ecSDave Kleikamp * filesystems the above test can fail to find any blockgroups 603ac27a0ecSDave Kleikamp */ 604ac27a0ecSDave Kleikamp avefreei = 0; 605ac27a0ecSDave Kleikamp goto fallback; 606ac27a0ecSDave Kleikamp } 607ac27a0ecSDave Kleikamp 608ac27a0ecSDave Kleikamp return -1; 609ac27a0ecSDave Kleikamp } 610ac27a0ecSDave Kleikamp 6112aa9fc4cSAvantika Mathur static int find_group_other(struct super_block *sb, struct inode *parent, 612a4912123STheodore Ts'o ext4_group_t *group, int mode) 613ac27a0ecSDave Kleikamp { 614fd2d4291SAvantika Mathur ext4_group_t parent_group = EXT4_I(parent)->i_block_group; 615fd2d4291SAvantika Mathur ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count; 616617ba13bSMingming Cao struct ext4_group_desc *desc; 617a4912123STheodore Ts'o ext4_group_t i, last; 618a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(sb)); 619a4912123STheodore Ts'o 620a4912123STheodore Ts'o /* 621a4912123STheodore Ts'o * Try to place the inode is the same flex group as its 622a4912123STheodore Ts'o * parent. If we can't find space, use the Orlov algorithm to 623a4912123STheodore Ts'o * find another flex group, and store that information in the 624a4912123STheodore Ts'o * parent directory's inode information so that use that flex 625a4912123STheodore Ts'o * group for future allocations. 626a4912123STheodore Ts'o */ 627a4912123STheodore Ts'o if (flex_size > 1) { 628a4912123STheodore Ts'o int retry = 0; 629a4912123STheodore Ts'o 630a4912123STheodore Ts'o try_again: 631a4912123STheodore Ts'o parent_group &= ~(flex_size-1); 632a4912123STheodore Ts'o last = parent_group + flex_size; 633a4912123STheodore Ts'o if (last > ngroups) 634a4912123STheodore Ts'o last = ngroups; 635a4912123STheodore Ts'o for (i = parent_group; i < last; i++) { 636a4912123STheodore Ts'o desc = ext4_get_group_desc(sb, i, NULL); 637a4912123STheodore Ts'o if (desc && ext4_free_inodes_count(sb, desc)) { 638a4912123STheodore Ts'o *group = i; 639a4912123STheodore Ts'o return 0; 640a4912123STheodore Ts'o } 641a4912123STheodore Ts'o } 642a4912123STheodore Ts'o if (!retry && EXT4_I(parent)->i_last_alloc_group != ~0) { 643a4912123STheodore Ts'o retry = 1; 644a4912123STheodore Ts'o parent_group = EXT4_I(parent)->i_last_alloc_group; 645a4912123STheodore Ts'o goto try_again; 646a4912123STheodore Ts'o } 647a4912123STheodore Ts'o /* 648a4912123STheodore Ts'o * If this didn't work, use the Orlov search algorithm 649a4912123STheodore Ts'o * to find a new flex group; we pass in the mode to 650a4912123STheodore Ts'o * avoid the topdir algorithms. 651a4912123STheodore Ts'o */ 652a4912123STheodore Ts'o *group = parent_group + flex_size; 653a4912123STheodore Ts'o if (*group > ngroups) 654a4912123STheodore Ts'o *group = 0; 655a4912123STheodore Ts'o return find_group_orlov(sb, parent, group, mode); 656a4912123STheodore Ts'o } 657ac27a0ecSDave Kleikamp 658ac27a0ecSDave Kleikamp /* 659ac27a0ecSDave Kleikamp * Try to place the inode in its parent directory 660ac27a0ecSDave Kleikamp */ 6612aa9fc4cSAvantika Mathur *group = parent_group; 6622aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 663560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 664560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6652aa9fc4cSAvantika Mathur return 0; 666ac27a0ecSDave Kleikamp 667ac27a0ecSDave Kleikamp /* 668ac27a0ecSDave Kleikamp * We're going to place this inode in a different blockgroup from its 669ac27a0ecSDave Kleikamp * parent. We want to cause files in a common directory to all land in 670ac27a0ecSDave Kleikamp * the same blockgroup. But we want files which are in a different 671ac27a0ecSDave Kleikamp * directory which shares a blockgroup with our parent to land in a 672ac27a0ecSDave Kleikamp * different blockgroup. 673ac27a0ecSDave Kleikamp * 674ac27a0ecSDave Kleikamp * So add our directory's i_ino into the starting point for the hash. 675ac27a0ecSDave Kleikamp */ 6762aa9fc4cSAvantika Mathur *group = (*group + parent->i_ino) % ngroups; 677ac27a0ecSDave Kleikamp 678ac27a0ecSDave Kleikamp /* 679ac27a0ecSDave Kleikamp * Use a quadratic hash to find a group with a free inode and some free 680ac27a0ecSDave Kleikamp * blocks. 681ac27a0ecSDave Kleikamp */ 682ac27a0ecSDave Kleikamp for (i = 1; i < ngroups; i <<= 1) { 6832aa9fc4cSAvantika Mathur *group += i; 6842aa9fc4cSAvantika Mathur if (*group >= ngroups) 6852aa9fc4cSAvantika Mathur *group -= ngroups; 6862aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 687560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc) && 688560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc)) 6892aa9fc4cSAvantika Mathur return 0; 690ac27a0ecSDave Kleikamp } 691ac27a0ecSDave Kleikamp 692ac27a0ecSDave Kleikamp /* 693ac27a0ecSDave Kleikamp * That failed: try linear search for a free inode, even if that group 694ac27a0ecSDave Kleikamp * has no free blocks. 695ac27a0ecSDave Kleikamp */ 6962aa9fc4cSAvantika Mathur *group = parent_group; 697ac27a0ecSDave Kleikamp for (i = 0; i < ngroups; i++) { 6982aa9fc4cSAvantika Mathur if (++*group >= ngroups) 6992aa9fc4cSAvantika Mathur *group = 0; 7002aa9fc4cSAvantika Mathur desc = ext4_get_group_desc(sb, *group, NULL); 701560671a0SAneesh Kumar K.V if (desc && ext4_free_inodes_count(sb, desc)) 7022aa9fc4cSAvantika Mathur return 0; 703ac27a0ecSDave Kleikamp } 704ac27a0ecSDave Kleikamp 705ac27a0ecSDave Kleikamp return -1; 706ac27a0ecSDave Kleikamp } 707ac27a0ecSDave Kleikamp 708ac27a0ecSDave Kleikamp /* 70939341867SAneesh Kumar K.V * claim the inode from the inode bitmap. If the group 71039341867SAneesh Kumar K.V * is uninit we need to take the groups's sb_bgl_lock 71139341867SAneesh Kumar K.V * and clear the uninit flag. The inode bitmap update 71239341867SAneesh Kumar K.V * and group desc uninit flag clear should be done 71339341867SAneesh Kumar K.V * after holding sb_bgl_lock so that ext4_read_inode_bitmap 71439341867SAneesh Kumar K.V * doesn't race with the ext4_claim_inode 71539341867SAneesh Kumar K.V */ 71639341867SAneesh Kumar K.V static int ext4_claim_inode(struct super_block *sb, 71739341867SAneesh Kumar K.V struct buffer_head *inode_bitmap_bh, 71839341867SAneesh Kumar K.V unsigned long ino, ext4_group_t group, int mode) 71939341867SAneesh Kumar K.V { 72039341867SAneesh Kumar K.V int free = 0, retval = 0, count; 72139341867SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 72239341867SAneesh Kumar K.V struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL); 72339341867SAneesh Kumar K.V 72439341867SAneesh Kumar K.V spin_lock(sb_bgl_lock(sbi, group)); 72539341867SAneesh Kumar K.V if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) { 72639341867SAneesh Kumar K.V /* not a free inode */ 72739341867SAneesh Kumar K.V retval = 1; 72839341867SAneesh Kumar K.V goto err_ret; 72939341867SAneesh Kumar K.V } 73039341867SAneesh Kumar K.V ino++; 73139341867SAneesh Kumar K.V if ((group == 0 && ino < EXT4_FIRST_INO(sb)) || 73239341867SAneesh Kumar K.V ino > EXT4_INODES_PER_GROUP(sb)) { 73339341867SAneesh Kumar K.V spin_unlock(sb_bgl_lock(sbi, group)); 73439341867SAneesh Kumar K.V ext4_error(sb, __func__, 73539341867SAneesh Kumar K.V "reserved inode or inode > inodes count - " 73639341867SAneesh Kumar K.V "block_group = %u, inode=%lu", group, 73739341867SAneesh Kumar K.V ino + group * EXT4_INODES_PER_GROUP(sb)); 73839341867SAneesh Kumar K.V return 1; 73939341867SAneesh Kumar K.V } 74039341867SAneesh Kumar K.V /* If we didn't allocate from within the initialized part of the inode 74139341867SAneesh Kumar K.V * table then we need to initialize up to this inode. */ 74239341867SAneesh Kumar K.V if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) { 74339341867SAneesh Kumar K.V 74439341867SAneesh Kumar K.V if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) { 74539341867SAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT); 74639341867SAneesh Kumar K.V /* When marking the block group with 74739341867SAneesh Kumar K.V * ~EXT4_BG_INODE_UNINIT we don't want to depend 74839341867SAneesh Kumar K.V * on the value of bg_itable_unused even though 74939341867SAneesh Kumar K.V * mke2fs could have initialized the same for us. 75039341867SAneesh Kumar K.V * Instead we calculated the value below 75139341867SAneesh Kumar K.V */ 75239341867SAneesh Kumar K.V 75339341867SAneesh Kumar K.V free = 0; 75439341867SAneesh Kumar K.V } else { 75539341867SAneesh Kumar K.V free = EXT4_INODES_PER_GROUP(sb) - 75639341867SAneesh Kumar K.V ext4_itable_unused_count(sb, gdp); 75739341867SAneesh Kumar K.V } 75839341867SAneesh Kumar K.V 75939341867SAneesh Kumar K.V /* 76039341867SAneesh Kumar K.V * Check the relative inode number against the last used 76139341867SAneesh Kumar K.V * relative inode number in this group. if it is greater 76239341867SAneesh Kumar K.V * we need to update the bg_itable_unused count 76339341867SAneesh Kumar K.V * 76439341867SAneesh Kumar K.V */ 76539341867SAneesh Kumar K.V if (ino > free) 76639341867SAneesh Kumar K.V ext4_itable_unused_set(sb, gdp, 76739341867SAneesh Kumar K.V (EXT4_INODES_PER_GROUP(sb) - ino)); 76839341867SAneesh Kumar K.V } 76939341867SAneesh Kumar K.V count = ext4_free_inodes_count(sb, gdp) - 1; 77039341867SAneesh Kumar K.V ext4_free_inodes_set(sb, gdp, count); 77139341867SAneesh Kumar K.V if (S_ISDIR(mode)) { 77239341867SAneesh Kumar K.V count = ext4_used_dirs_count(sb, gdp) + 1; 77339341867SAneesh Kumar K.V ext4_used_dirs_set(sb, gdp, count); 7747d39db14STheodore Ts'o if (sbi->s_log_groups_per_flex) { 7757d39db14STheodore Ts'o ext4_group_t f = ext4_flex_group(sbi, group); 7767d39db14STheodore Ts'o 7777d39db14STheodore Ts'o atomic_inc(&sbi->s_flex_groups[f].free_inodes); 7787d39db14STheodore Ts'o } 77939341867SAneesh Kumar K.V } 78039341867SAneesh Kumar K.V gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp); 78139341867SAneesh Kumar K.V err_ret: 78239341867SAneesh Kumar K.V spin_unlock(sb_bgl_lock(sbi, group)); 78339341867SAneesh Kumar K.V return retval; 78439341867SAneesh Kumar K.V } 78539341867SAneesh Kumar K.V 78639341867SAneesh Kumar K.V /* 787ac27a0ecSDave Kleikamp * There are two policies for allocating an inode. If the new inode is 788ac27a0ecSDave Kleikamp * a directory, then a forward search is made for a block group with both 789ac27a0ecSDave Kleikamp * free space and a low directory-to-inode ratio; if that fails, then of 790ac27a0ecSDave Kleikamp * the groups with above-average free space, that group with the fewest 791ac27a0ecSDave Kleikamp * directories already is chosen. 792ac27a0ecSDave Kleikamp * 793ac27a0ecSDave Kleikamp * For other inodes, search forward from the parent directory's block 794ac27a0ecSDave Kleikamp * group to find a free inode. 795ac27a0ecSDave Kleikamp */ 796617ba13bSMingming Cao struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode) 797ac27a0ecSDave Kleikamp { 798ac27a0ecSDave Kleikamp struct super_block *sb; 7993300bedaSAneesh Kumar K.V struct buffer_head *inode_bitmap_bh = NULL; 8003300bedaSAneesh Kumar K.V struct buffer_head *group_desc_bh; 8012aa9fc4cSAvantika Mathur ext4_group_t group = 0; 802ac27a0ecSDave Kleikamp unsigned long ino = 0; 803ac27a0ecSDave Kleikamp struct inode *inode; 804617ba13bSMingming Cao struct ext4_group_desc *gdp = NULL; 805617ba13bSMingming Cao struct ext4_super_block *es; 806617ba13bSMingming Cao struct ext4_inode_info *ei; 807617ba13bSMingming Cao struct ext4_sb_info *sbi; 80839341867SAneesh Kumar K.V int ret2, err = 0; 809ac27a0ecSDave Kleikamp struct inode *ret; 8102aa9fc4cSAvantika Mathur ext4_group_t i; 8112aa9fc4cSAvantika Mathur int free = 0; 8122842c3b5STheodore Ts'o static int once = 1; 813772cb7c8SJose R. Santos ext4_group_t flex_group; 814ac27a0ecSDave Kleikamp 815ac27a0ecSDave Kleikamp /* Cannot create files in a deleted directory */ 816ac27a0ecSDave Kleikamp if (!dir || !dir->i_nlink) 817ac27a0ecSDave Kleikamp return ERR_PTR(-EPERM); 818ac27a0ecSDave Kleikamp 819ac27a0ecSDave Kleikamp sb = dir->i_sb; 820ba80b101STheodore Ts'o trace_mark(ext4_request_inode, "dev %s dir %lu mode %d", sb->s_id, 821ba80b101STheodore Ts'o dir->i_ino, mode); 822ac27a0ecSDave Kleikamp inode = new_inode(sb); 823ac27a0ecSDave Kleikamp if (!inode) 824ac27a0ecSDave Kleikamp return ERR_PTR(-ENOMEM); 825617ba13bSMingming Cao ei = EXT4_I(inode); 826ac27a0ecSDave Kleikamp 827617ba13bSMingming Cao sbi = EXT4_SB(sb); 828ac27a0ecSDave Kleikamp es = sbi->s_es; 829772cb7c8SJose R. Santos 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); 8342842c3b5STheodore Ts'o 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 } 840772cb7c8SJose R. Santos goto got_group; 841772cb7c8SJose R. Santos } 842772cb7c8SJose R. Santos 843ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) { 844ac27a0ecSDave Kleikamp if (test_opt(sb, OLDALLOC)) 8452aa9fc4cSAvantika Mathur ret2 = find_group_dir(sb, dir, &group); 846ac27a0ecSDave Kleikamp else 847a4912123STheodore Ts'o ret2 = find_group_orlov(sb, dir, &group, mode); 848ac27a0ecSDave Kleikamp } else 849a4912123STheodore Ts'o ret2 = find_group_other(sb, dir, &group, mode); 850ac27a0ecSDave Kleikamp 851772cb7c8SJose R. Santos got_group: 852a4912123STheodore Ts'o EXT4_I(dir)->i_last_alloc_group = group; 853ac27a0ecSDave Kleikamp err = -ENOSPC; 8542aa9fc4cSAvantika Mathur if (ret2 == -1) 855ac27a0ecSDave Kleikamp goto out; 856ac27a0ecSDave Kleikamp 857ac27a0ecSDave Kleikamp for (i = 0; i < sbi->s_groups_count; i++) { 858ac27a0ecSDave Kleikamp err = -EIO; 859ac27a0ecSDave Kleikamp 8603300bedaSAneesh Kumar K.V gdp = ext4_get_group_desc(sb, group, &group_desc_bh); 861ac27a0ecSDave Kleikamp if (!gdp) 862ac27a0ecSDave Kleikamp goto fail; 863ac27a0ecSDave Kleikamp 8643300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 8653300bedaSAneesh Kumar K.V inode_bitmap_bh = ext4_read_inode_bitmap(sb, group); 8663300bedaSAneesh Kumar K.V if (!inode_bitmap_bh) 867ac27a0ecSDave Kleikamp goto fail; 868ac27a0ecSDave Kleikamp 869ac27a0ecSDave Kleikamp ino = 0; 870ac27a0ecSDave Kleikamp 871ac27a0ecSDave Kleikamp repeat_in_this_group: 872617ba13bSMingming Cao ino = ext4_find_next_zero_bit((unsigned long *) 8733300bedaSAneesh Kumar K.V inode_bitmap_bh->b_data, 8743300bedaSAneesh Kumar K.V EXT4_INODES_PER_GROUP(sb), ino); 8753300bedaSAneesh Kumar K.V 876617ba13bSMingming Cao if (ino < EXT4_INODES_PER_GROUP(sb)) { 877ac27a0ecSDave Kleikamp 8783300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, "get_write_access"); 8793300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 8803300bedaSAneesh Kumar K.V inode_bitmap_bh); 881ac27a0ecSDave Kleikamp if (err) 882ac27a0ecSDave Kleikamp goto fail; 883ac27a0ecSDave Kleikamp 88439341867SAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "get_write_access"); 88539341867SAneesh Kumar K.V err = ext4_journal_get_write_access(handle, 88639341867SAneesh Kumar K.V group_desc_bh); 88739341867SAneesh Kumar K.V if (err) 88839341867SAneesh Kumar K.V goto fail; 88939341867SAneesh Kumar K.V if (!ext4_claim_inode(sb, inode_bitmap_bh, 89039341867SAneesh Kumar K.V ino, group, mode)) { 891ac27a0ecSDave Kleikamp /* we won it */ 8923300bedaSAneesh Kumar K.V BUFFER_TRACE(inode_bitmap_bh, 8930390131bSFrank Mayhar "call ext4_handle_dirty_metadata"); 8940390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 8950390131bSFrank Mayhar inode, 8963300bedaSAneesh Kumar K.V inode_bitmap_bh); 897ac27a0ecSDave Kleikamp if (err) 898ac27a0ecSDave Kleikamp goto fail; 89939341867SAneesh Kumar K.V /* zero bit is inode number 1*/ 90039341867SAneesh Kumar K.V ino++; 901ac27a0ecSDave Kleikamp goto got; 902ac27a0ecSDave Kleikamp } 903ac27a0ecSDave Kleikamp /* we lost it */ 9043300bedaSAneesh Kumar K.V ext4_handle_release_buffer(handle, inode_bitmap_bh); 90539341867SAneesh Kumar K.V ext4_handle_release_buffer(handle, group_desc_bh); 906ac27a0ecSDave Kleikamp 907617ba13bSMingming Cao if (++ino < EXT4_INODES_PER_GROUP(sb)) 908ac27a0ecSDave Kleikamp goto repeat_in_this_group; 909ac27a0ecSDave Kleikamp } 910ac27a0ecSDave Kleikamp 911ac27a0ecSDave Kleikamp /* 912ac27a0ecSDave Kleikamp * This case is possible in concurrent environment. It is very 913ac27a0ecSDave Kleikamp * rare. We cannot repeat the find_group_xxx() call because 914ac27a0ecSDave Kleikamp * that will simply return the same blockgroup, because the 915ac27a0ecSDave Kleikamp * group descriptor metadata has not yet been updated. 916ac27a0ecSDave Kleikamp * So we just go onto the next blockgroup. 917ac27a0ecSDave Kleikamp */ 918ac27a0ecSDave Kleikamp if (++group == sbi->s_groups_count) 919ac27a0ecSDave Kleikamp group = 0; 920ac27a0ecSDave Kleikamp } 921ac27a0ecSDave Kleikamp err = -ENOSPC; 922ac27a0ecSDave Kleikamp goto out; 923ac27a0ecSDave Kleikamp 924ac27a0ecSDave Kleikamp got: 925717d50e4SAndreas Dilger /* We may have to initialize the block bitmap if it isn't already */ 926717d50e4SAndreas Dilger if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) && 927717d50e4SAndreas Dilger gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 9283300bedaSAneesh Kumar K.V struct buffer_head *block_bitmap_bh; 929717d50e4SAndreas Dilger 9303300bedaSAneesh Kumar K.V block_bitmap_bh = ext4_read_block_bitmap(sb, group); 9313300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "get block bitmap access"); 9323300bedaSAneesh Kumar K.V err = ext4_journal_get_write_access(handle, block_bitmap_bh); 933717d50e4SAndreas Dilger if (err) { 9343300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 935717d50e4SAndreas Dilger goto fail; 936717d50e4SAndreas Dilger } 937717d50e4SAndreas Dilger 938717d50e4SAndreas Dilger free = 0; 939ac27a0ecSDave Kleikamp spin_lock(sb_bgl_lock(sbi, group)); 940717d50e4SAndreas Dilger /* recheck and clear flag under lock if we still need to */ 941717d50e4SAndreas Dilger if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 942717d50e4SAndreas Dilger free = ext4_free_blocks_after_init(sb, group, gdp); 9433300bedaSAneesh Kumar K.V gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 944560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, free); 94523712a9cSFrederic Bohe gdp->bg_checksum = ext4_group_desc_csum(sbi, group, 94623712a9cSFrederic Bohe gdp); 947717d50e4SAndreas Dilger } 948717d50e4SAndreas Dilger spin_unlock(sb_bgl_lock(sbi, group)); 949717d50e4SAndreas Dilger 950717d50e4SAndreas Dilger /* Don't need to dirty bitmap block if we didn't change it */ 951717d50e4SAndreas Dilger if (free) { 9523300bedaSAneesh Kumar K.V BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap"); 9530390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, 9543300bedaSAneesh Kumar K.V NULL, block_bitmap_bh); 955717d50e4SAndreas Dilger } 956717d50e4SAndreas Dilger 9573300bedaSAneesh Kumar K.V brelse(block_bitmap_bh); 958717d50e4SAndreas Dilger if (err) 959717d50e4SAndreas Dilger goto fail; 960717d50e4SAndreas Dilger } 9613300bedaSAneesh Kumar K.V BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata"); 9623300bedaSAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh); 96339341867SAneesh Kumar K.V if (err) 96439341867SAneesh Kumar K.V goto fail; 965ac27a0ecSDave Kleikamp 966ac27a0ecSDave Kleikamp percpu_counter_dec(&sbi->s_freeinodes_counter); 967ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 968ac27a0ecSDave Kleikamp percpu_counter_inc(&sbi->s_dirs_counter); 969ac27a0ecSDave Kleikamp sb->s_dirt = 1; 970ac27a0ecSDave Kleikamp 971772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 972772cb7c8SJose R. Santos flex_group = ext4_flex_group(sbi, group); 9739f24e420STheodore Ts'o atomic_dec(&sbi->s_flex_groups[flex_group].free_inodes); 974772cb7c8SJose R. Santos } 975772cb7c8SJose R. Santos 9764c9c544eSDavid Howells inode->i_uid = current_fsuid(); 977ac27a0ecSDave Kleikamp if (test_opt(sb, GRPID)) 978ac27a0ecSDave Kleikamp inode->i_gid = dir->i_gid; 979ac27a0ecSDave Kleikamp else if (dir->i_mode & S_ISGID) { 980ac27a0ecSDave Kleikamp inode->i_gid = dir->i_gid; 981ac27a0ecSDave Kleikamp if (S_ISDIR(mode)) 982ac27a0ecSDave Kleikamp mode |= S_ISGID; 983ac27a0ecSDave Kleikamp } else 9844c9c544eSDavid Howells inode->i_gid = current_fsgid(); 985ac27a0ecSDave Kleikamp inode->i_mode = mode; 986ac27a0ecSDave Kleikamp 987717d50e4SAndreas Dilger inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb); 988ac27a0ecSDave Kleikamp /* This is the optimal IO size (for stat), not the fs block size */ 989ac27a0ecSDave Kleikamp inode->i_blocks = 0; 990ef7f3835SKalpak Shah inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime = 991ef7f3835SKalpak Shah ext4_current_time(inode); 992ac27a0ecSDave Kleikamp 993ac27a0ecSDave Kleikamp memset(ei->i_data, 0, sizeof(ei->i_data)); 994ac27a0ecSDave Kleikamp ei->i_dir_start_lookup = 0; 995ac27a0ecSDave Kleikamp ei->i_disksize = 0; 996ac27a0ecSDave Kleikamp 99742bf0383SAneesh Kumar K.V /* 9982dc6b0d4SDuane Griffin * Don't inherit extent flag from directory, amongst others. We set 9992dc6b0d4SDuane Griffin * extent flag on newly created directory and file only if -o extent 10002dc6b0d4SDuane Griffin * mount option is specified 100142bf0383SAneesh Kumar K.V */ 10022dc6b0d4SDuane Griffin ei->i_flags = 10032dc6b0d4SDuane Griffin ext4_mask_flags(mode, EXT4_I(dir)->i_flags & EXT4_FL_INHERITED); 1004ac27a0ecSDave Kleikamp ei->i_file_acl = 0; 1005ac27a0ecSDave Kleikamp ei->i_dtime = 0; 1006ac27a0ecSDave Kleikamp ei->i_block_group = group; 1007a4912123STheodore Ts'o ei->i_last_alloc_group = ~0; 1008ac27a0ecSDave Kleikamp 1009617ba13bSMingming Cao ext4_set_inode_flags(inode); 1010ac27a0ecSDave Kleikamp if (IS_DIRSYNC(inode)) 10110390131bSFrank Mayhar ext4_handle_sync(handle); 10126b38e842SAl Viro if (insert_inode_locked(inode) < 0) { 10136b38e842SAl Viro err = -EINVAL; 10146b38e842SAl Viro goto fail_drop; 10156b38e842SAl Viro } 1016ac27a0ecSDave Kleikamp spin_lock(&sbi->s_next_gen_lock); 1017ac27a0ecSDave Kleikamp inode->i_generation = sbi->s_next_generation++; 1018ac27a0ecSDave Kleikamp spin_unlock(&sbi->s_next_gen_lock); 1019ac27a0ecSDave Kleikamp 1020617ba13bSMingming Cao ei->i_state = EXT4_STATE_NEW; 1021ef7f3835SKalpak Shah 1022ef7f3835SKalpak Shah ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize; 1023ac27a0ecSDave Kleikamp 1024ac27a0ecSDave Kleikamp ret = inode; 1025a269eb18SJan Kara if (vfs_dq_alloc_inode(inode)) { 1026ac27a0ecSDave Kleikamp err = -EDQUOT; 1027ac27a0ecSDave Kleikamp goto fail_drop; 1028ac27a0ecSDave Kleikamp } 1029ac27a0ecSDave Kleikamp 1030617ba13bSMingming Cao err = ext4_init_acl(handle, inode, dir); 1031ac27a0ecSDave Kleikamp if (err) 1032ac27a0ecSDave Kleikamp goto fail_free_drop; 1033ac27a0ecSDave Kleikamp 1034617ba13bSMingming Cao err = ext4_init_security(handle, inode, dir); 1035ac27a0ecSDave Kleikamp if (err) 1036ac27a0ecSDave Kleikamp goto fail_free_drop; 1037ac27a0ecSDave Kleikamp 103883982b6fSTheodore Ts'o if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) { 1039e4079a11SEric Sandeen /* set extent flag only for directory, file and normal symlink*/ 1040e65187e6SAneesh Kumar K.V if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) { 1041a86c6181SAlex Tomas EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL; 1042a86c6181SAlex Tomas ext4_ext_tree_init(handle, inode); 1043a86c6181SAlex Tomas } 104442bf0383SAneesh Kumar K.V } 1045ac27a0ecSDave Kleikamp 10468753e88fSAneesh Kumar K.V err = ext4_mark_inode_dirty(handle, inode); 10478753e88fSAneesh Kumar K.V if (err) { 10488753e88fSAneesh Kumar K.V ext4_std_error(sb, err); 10498753e88fSAneesh Kumar K.V goto fail_free_drop; 10508753e88fSAneesh Kumar K.V } 10518753e88fSAneesh Kumar K.V 1052617ba13bSMingming Cao ext4_debug("allocating inode %lu\n", inode->i_ino); 1053ba80b101STheodore Ts'o trace_mark(ext4_allocate_inode, "dev %s ino %lu dir %lu mode %d", 1054ba80b101STheodore Ts'o sb->s_id, inode->i_ino, dir->i_ino, mode); 1055ac27a0ecSDave Kleikamp goto really_out; 1056ac27a0ecSDave Kleikamp fail: 1057617ba13bSMingming Cao ext4_std_error(sb, err); 1058ac27a0ecSDave Kleikamp out: 1059ac27a0ecSDave Kleikamp iput(inode); 1060ac27a0ecSDave Kleikamp ret = ERR_PTR(err); 1061ac27a0ecSDave Kleikamp really_out: 10623300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1063ac27a0ecSDave Kleikamp return ret; 1064ac27a0ecSDave Kleikamp 1065ac27a0ecSDave Kleikamp fail_free_drop: 1066a269eb18SJan Kara vfs_dq_free_inode(inode); 1067ac27a0ecSDave Kleikamp 1068ac27a0ecSDave Kleikamp fail_drop: 1069a269eb18SJan Kara vfs_dq_drop(inode); 1070ac27a0ecSDave Kleikamp inode->i_flags |= S_NOQUOTA; 1071ac27a0ecSDave Kleikamp inode->i_nlink = 0; 10726b38e842SAl Viro unlock_new_inode(inode); 1073ac27a0ecSDave Kleikamp iput(inode); 10743300bedaSAneesh Kumar K.V brelse(inode_bitmap_bh); 1075ac27a0ecSDave Kleikamp return ERR_PTR(err); 1076ac27a0ecSDave Kleikamp } 1077ac27a0ecSDave Kleikamp 1078ac27a0ecSDave Kleikamp /* Verify that we are loading a valid orphan from disk */ 1079617ba13bSMingming Cao struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino) 1080ac27a0ecSDave Kleikamp { 1081617ba13bSMingming Cao unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count); 1082fd2d4291SAvantika Mathur ext4_group_t block_group; 1083ac27a0ecSDave Kleikamp int bit; 10841d1fe1eeSDavid Howells struct buffer_head *bitmap_bh; 1085ac27a0ecSDave Kleikamp struct inode *inode = NULL; 10861d1fe1eeSDavid Howells long err = -EIO; 1087ac27a0ecSDave Kleikamp 1088ac27a0ecSDave Kleikamp /* Error cases - e2fsck has already cleaned up for us */ 1089ac27a0ecSDave Kleikamp if (ino > max_ino) { 109046e665e9SHarvey Harrison ext4_warning(sb, __func__, 1091ac27a0ecSDave Kleikamp "bad orphan ino %lu! e2fsck was run?", ino); 10921d1fe1eeSDavid Howells goto error; 1093ac27a0ecSDave Kleikamp } 1094ac27a0ecSDave Kleikamp 1095617ba13bSMingming Cao block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb); 1096617ba13bSMingming Cao bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb); 1097e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, block_group); 1098ac27a0ecSDave Kleikamp if (!bitmap_bh) { 109946e665e9SHarvey Harrison ext4_warning(sb, __func__, 1100ac27a0ecSDave Kleikamp "inode bitmap error for orphan %lu", ino); 11011d1fe1eeSDavid Howells goto error; 1102ac27a0ecSDave Kleikamp } 1103ac27a0ecSDave Kleikamp 1104ac27a0ecSDave Kleikamp /* Having the inode bit set should be a 100% indicator that this 1105ac27a0ecSDave Kleikamp * is a valid orphan (no e2fsck run on fs). Orphans also include 1106ac27a0ecSDave Kleikamp * inodes that were being truncated, so we can't check i_nlink==0. 1107ac27a0ecSDave Kleikamp */ 11081d1fe1eeSDavid Howells if (!ext4_test_bit(bit, bitmap_bh->b_data)) 11091d1fe1eeSDavid Howells goto bad_orphan; 11101d1fe1eeSDavid Howells 11111d1fe1eeSDavid Howells inode = ext4_iget(sb, ino); 11121d1fe1eeSDavid Howells if (IS_ERR(inode)) 11131d1fe1eeSDavid Howells goto iget_failed; 11141d1fe1eeSDavid Howells 111591ef4cafSDuane Griffin /* 111691ef4cafSDuane Griffin * If the orphans has i_nlinks > 0 then it should be able to be 111791ef4cafSDuane Griffin * truncated, otherwise it won't be removed from the orphan list 111891ef4cafSDuane Griffin * during processing and an infinite loop will result. 111991ef4cafSDuane Griffin */ 112091ef4cafSDuane Griffin if (inode->i_nlink && !ext4_can_truncate(inode)) 112191ef4cafSDuane Griffin goto bad_orphan; 112291ef4cafSDuane Griffin 11231d1fe1eeSDavid Howells if (NEXT_ORPHAN(inode) > max_ino) 11241d1fe1eeSDavid Howells goto bad_orphan; 11251d1fe1eeSDavid Howells brelse(bitmap_bh); 11261d1fe1eeSDavid Howells return inode; 11271d1fe1eeSDavid Howells 11281d1fe1eeSDavid Howells iget_failed: 11291d1fe1eeSDavid Howells err = PTR_ERR(inode); 11301d1fe1eeSDavid Howells inode = NULL; 11311d1fe1eeSDavid Howells bad_orphan: 113246e665e9SHarvey Harrison ext4_warning(sb, __func__, 1133ac27a0ecSDave Kleikamp "bad orphan inode %lu! e2fsck was run?", ino); 1134617ba13bSMingming Cao printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n", 1135ac27a0ecSDave Kleikamp bit, (unsigned long long)bitmap_bh->b_blocknr, 1136617ba13bSMingming Cao ext4_test_bit(bit, bitmap_bh->b_data)); 1137ac27a0ecSDave Kleikamp printk(KERN_NOTICE "inode=%p\n", inode); 1138ac27a0ecSDave Kleikamp if (inode) { 1139ac27a0ecSDave Kleikamp printk(KERN_NOTICE "is_bad_inode(inode)=%d\n", 1140ac27a0ecSDave Kleikamp is_bad_inode(inode)); 1141ac27a0ecSDave Kleikamp printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n", 1142ac27a0ecSDave Kleikamp NEXT_ORPHAN(inode)); 1143ac27a0ecSDave Kleikamp printk(KERN_NOTICE "max_ino=%lu\n", max_ino); 114491ef4cafSDuane Griffin printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink); 1145ac27a0ecSDave Kleikamp /* Avoid freeing blocks if we got a bad deleted inode */ 11461d1fe1eeSDavid Howells if (inode->i_nlink == 0) 1147ac27a0ecSDave Kleikamp inode->i_blocks = 0; 1148ac27a0ecSDave Kleikamp iput(inode); 1149ac27a0ecSDave Kleikamp } 1150ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11511d1fe1eeSDavid Howells error: 11521d1fe1eeSDavid Howells return ERR_PTR(err); 1153ac27a0ecSDave Kleikamp } 1154ac27a0ecSDave Kleikamp 1155617ba13bSMingming Cao unsigned long ext4_count_free_inodes(struct super_block *sb) 1156ac27a0ecSDave Kleikamp { 1157ac27a0ecSDave Kleikamp unsigned long desc_count; 1158617ba13bSMingming Cao struct ext4_group_desc *gdp; 1159fd2d4291SAvantika Mathur ext4_group_t i; 1160617ba13bSMingming Cao #ifdef EXT4FS_DEBUG 1161617ba13bSMingming Cao struct ext4_super_block *es; 1162ac27a0ecSDave Kleikamp unsigned long bitmap_count, x; 1163ac27a0ecSDave Kleikamp struct buffer_head *bitmap_bh = NULL; 1164ac27a0ecSDave Kleikamp 1165617ba13bSMingming Cao es = EXT4_SB(sb)->s_es; 1166ac27a0ecSDave Kleikamp desc_count = 0; 1167ac27a0ecSDave Kleikamp bitmap_count = 0; 1168ac27a0ecSDave Kleikamp gdp = NULL; 1169617ba13bSMingming Cao for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) { 1170617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1171ac27a0ecSDave Kleikamp if (!gdp) 1172ac27a0ecSDave Kleikamp continue; 1173560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1174ac27a0ecSDave Kleikamp brelse(bitmap_bh); 1175e29d1cdeSEric Sandeen bitmap_bh = ext4_read_inode_bitmap(sb, i); 1176ac27a0ecSDave Kleikamp if (!bitmap_bh) 1177ac27a0ecSDave Kleikamp continue; 1178ac27a0ecSDave Kleikamp 1179617ba13bSMingming Cao x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8); 1180c549a95dSEric Sandeen printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n", 1181560671a0SAneesh Kumar K.V i, ext4_free_inodes_count(sb, gdp), x); 1182ac27a0ecSDave Kleikamp bitmap_count += x; 1183ac27a0ecSDave Kleikamp } 1184ac27a0ecSDave Kleikamp brelse(bitmap_bh); 11854776004fSTheodore Ts'o printk(KERN_DEBUG "ext4_count_free_inodes: " 11864776004fSTheodore Ts'o "stored = %u, computed = %lu, %lu\n", 1187ac27a0ecSDave Kleikamp le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count); 1188ac27a0ecSDave Kleikamp return desc_count; 1189ac27a0ecSDave Kleikamp #else 1190ac27a0ecSDave Kleikamp desc_count = 0; 1191617ba13bSMingming Cao for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) { 1192617ba13bSMingming Cao gdp = ext4_get_group_desc(sb, i, NULL); 1193ac27a0ecSDave Kleikamp if (!gdp) 1194ac27a0ecSDave Kleikamp continue; 1195560671a0SAneesh Kumar K.V desc_count += ext4_free_inodes_count(sb, gdp); 1196ac27a0ecSDave Kleikamp cond_resched(); 1197ac27a0ecSDave Kleikamp } 1198ac27a0ecSDave Kleikamp return desc_count; 1199ac27a0ecSDave Kleikamp #endif 1200ac27a0ecSDave Kleikamp } 1201ac27a0ecSDave Kleikamp 1202ac27a0ecSDave Kleikamp /* Called at mount-time, super-block is locked */ 1203617ba13bSMingming Cao unsigned long ext4_count_dirs(struct super_block * sb) 1204ac27a0ecSDave Kleikamp { 1205ac27a0ecSDave Kleikamp unsigned long count = 0; 1206fd2d4291SAvantika Mathur ext4_group_t i; 1207ac27a0ecSDave Kleikamp 1208617ba13bSMingming Cao for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) { 1209617ba13bSMingming Cao struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL); 1210ac27a0ecSDave Kleikamp if (!gdp) 1211ac27a0ecSDave Kleikamp continue; 1212560671a0SAneesh Kumar K.V count += ext4_used_dirs_count(sb, gdp); 1213ac27a0ecSDave Kleikamp } 1214ac27a0ecSDave Kleikamp return count; 1215ac27a0ecSDave Kleikamp } 1216