1f5166768STheodore Ts'o // SPDX-License-Identifier: GPL-2.0 2a86c6181SAlex Tomas /* 3a86c6181SAlex Tomas * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 4a86c6181SAlex Tomas * Written by Alex Tomas <alex@clusterfs.com> 5a86c6181SAlex Tomas * 6a86c6181SAlex Tomas * Architecture independence: 7a86c6181SAlex Tomas * Copyright (c) 2005, Bull S.A. 8a86c6181SAlex Tomas * Written by Pierre Peiffer <pierre.peiffer@bull.net> 9a86c6181SAlex Tomas */ 10a86c6181SAlex Tomas 11a86c6181SAlex Tomas /* 12a86c6181SAlex Tomas * Extents support for EXT4 13a86c6181SAlex Tomas * 14a86c6181SAlex Tomas * TODO: 15a86c6181SAlex Tomas * - ext4*_error() should be used in some situations 16a86c6181SAlex Tomas * - analyze all BUG()/BUG_ON(), use -EIO where appropriate 17a86c6181SAlex Tomas * - smart tree reduction 18a86c6181SAlex Tomas */ 19a86c6181SAlex Tomas 20a86c6181SAlex Tomas #include <linux/fs.h> 21a86c6181SAlex Tomas #include <linux/time.h> 22cd02ff0bSMingming Cao #include <linux/jbd2.h> 23a86c6181SAlex Tomas #include <linux/highuid.h> 24a86c6181SAlex Tomas #include <linux/pagemap.h> 25a86c6181SAlex Tomas #include <linux/quotaops.h> 26a86c6181SAlex Tomas #include <linux/string.h> 27a86c6181SAlex Tomas #include <linux/slab.h> 287c0f6ba6SLinus Torvalds #include <linux/uaccess.h> 296873fa0dSEric Sandeen #include <linux/fiemap.h> 3066114cadSTejun Heo #include <linux/backing-dev.h> 31d3b6f23fSRitesh Harjani #include <linux/iomap.h> 323dcf5451SChristoph Hellwig #include "ext4_jbd2.h" 334a092d73STheodore Ts'o #include "ext4_extents.h" 34f19d5870STao Ma #include "xattr.h" 35a86c6181SAlex Tomas 360562e0baSJiaying Zhang #include <trace/events/ext4.h> 370562e0baSJiaying Zhang 385f95d21fSLukas Czerner /* 395f95d21fSLukas Czerner * used by extent splitting. 405f95d21fSLukas Czerner */ 415f95d21fSLukas Czerner #define EXT4_EXT_MAY_ZEROOUT 0x1 /* safe to zeroout if split fails \ 425f95d21fSLukas Czerner due to ENOSPC */ 43556615dcSLukas Czerner #define EXT4_EXT_MARK_UNWRIT1 0x2 /* mark first half unwritten */ 44556615dcSLukas Czerner #define EXT4_EXT_MARK_UNWRIT2 0x4 /* mark second half unwritten */ 455f95d21fSLukas Czerner 46dee1f973SDmitry Monakhov #define EXT4_EXT_DATA_VALID1 0x8 /* first half contains valid data */ 47dee1f973SDmitry Monakhov #define EXT4_EXT_DATA_VALID2 0x10 /* second half contains valid data */ 48dee1f973SDmitry Monakhov 497ac5990dSDarrick J. Wong static __le32 ext4_extent_block_csum(struct inode *inode, 507ac5990dSDarrick J. Wong struct ext4_extent_header *eh) 517ac5990dSDarrick J. Wong { 527ac5990dSDarrick J. Wong struct ext4_inode_info *ei = EXT4_I(inode); 537ac5990dSDarrick J. Wong struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 547ac5990dSDarrick J. Wong __u32 csum; 557ac5990dSDarrick J. Wong 567ac5990dSDarrick J. Wong csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh, 577ac5990dSDarrick J. Wong EXT4_EXTENT_TAIL_OFFSET(eh)); 587ac5990dSDarrick J. Wong return cpu_to_le32(csum); 597ac5990dSDarrick J. Wong } 607ac5990dSDarrick J. Wong 617ac5990dSDarrick J. Wong static int ext4_extent_block_csum_verify(struct inode *inode, 627ac5990dSDarrick J. Wong struct ext4_extent_header *eh) 637ac5990dSDarrick J. Wong { 647ac5990dSDarrick J. Wong struct ext4_extent_tail *et; 657ac5990dSDarrick J. Wong 669aa5d32bSDmitry Monakhov if (!ext4_has_metadata_csum(inode->i_sb)) 677ac5990dSDarrick J. Wong return 1; 687ac5990dSDarrick J. Wong 697ac5990dSDarrick J. Wong et = find_ext4_extent_tail(eh); 707ac5990dSDarrick J. Wong if (et->et_checksum != ext4_extent_block_csum(inode, eh)) 717ac5990dSDarrick J. Wong return 0; 727ac5990dSDarrick J. Wong return 1; 737ac5990dSDarrick J. Wong } 747ac5990dSDarrick J. Wong 757ac5990dSDarrick J. Wong static void ext4_extent_block_csum_set(struct inode *inode, 767ac5990dSDarrick J. Wong struct ext4_extent_header *eh) 777ac5990dSDarrick J. Wong { 787ac5990dSDarrick J. Wong struct ext4_extent_tail *et; 797ac5990dSDarrick J. Wong 809aa5d32bSDmitry Monakhov if (!ext4_has_metadata_csum(inode->i_sb)) 817ac5990dSDarrick J. Wong return; 827ac5990dSDarrick J. Wong 837ac5990dSDarrick J. Wong et = find_ext4_extent_tail(eh); 847ac5990dSDarrick J. Wong et->et_checksum = ext4_extent_block_csum(inode, eh); 857ac5990dSDarrick J. Wong } 867ac5990dSDarrick J. Wong 875f95d21fSLukas Czerner static int ext4_split_extent_at(handle_t *handle, 885f95d21fSLukas Czerner struct inode *inode, 89dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 905f95d21fSLukas Czerner ext4_lblk_t split, 915f95d21fSLukas Czerner int split_flag, 925f95d21fSLukas Czerner int flags); 935f95d21fSLukas Czerner 94a4130367SJan Kara static int ext4_ext_trunc_restart_fn(struct inode *inode, int *dropped) 95a86c6181SAlex Tomas { 967b808191STheodore Ts'o /* 97a4130367SJan Kara * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this 98a4130367SJan Kara * moment, get_block can be called only for blocks inside i_size since 99a4130367SJan Kara * page cache has been already dropped and writes are blocked by 100a4130367SJan Kara * i_mutex. So we can safely drop the i_data_sem here. 1017b808191STheodore Ts'o */ 102a4130367SJan Kara BUG_ON(EXT4_JOURNAL(inode) == NULL); 103a4130367SJan Kara ext4_discard_preallocations(inode); 104a4130367SJan Kara up_write(&EXT4_I(inode)->i_data_sem); 105a4130367SJan Kara *dropped = 1; 106a4130367SJan Kara return 0; 107a4130367SJan Kara } 108487caeefSJan Kara 109a4130367SJan Kara /* 110a4130367SJan Kara * Make sure 'handle' has at least 'check_cred' credits. If not, restart 111a4130367SJan Kara * transaction with 'restart_cred' credits. The function drops i_data_sem 112a4130367SJan Kara * when restarting transaction and gets it after transaction is restarted. 113a4130367SJan Kara * 114a4130367SJan Kara * The function returns 0 on success, 1 if transaction had to be restarted, 115a4130367SJan Kara * and < 0 in case of fatal error. 116a4130367SJan Kara */ 117a4130367SJan Kara int ext4_datasem_ensure_credits(handle_t *handle, struct inode *inode, 11883448bdfSJan Kara int check_cred, int restart_cred, 11983448bdfSJan Kara int revoke_cred) 120a4130367SJan Kara { 121a4130367SJan Kara int ret; 122a4130367SJan Kara int dropped = 0; 123a4130367SJan Kara 124a4130367SJan Kara ret = ext4_journal_ensure_credits_fn(handle, check_cred, restart_cred, 12583448bdfSJan Kara revoke_cred, ext4_ext_trunc_restart_fn(inode, &dropped)); 126a4130367SJan Kara if (dropped) 127a4130367SJan Kara down_write(&EXT4_I(inode)->i_data_sem); 128a4130367SJan Kara return ret; 129a86c6181SAlex Tomas } 130a86c6181SAlex Tomas 131a86c6181SAlex Tomas /* 132a86c6181SAlex Tomas * could return: 133a86c6181SAlex Tomas * - EROFS 134a86c6181SAlex Tomas * - ENOMEM 135a86c6181SAlex Tomas */ 136a86c6181SAlex Tomas static int ext4_ext_get_access(handle_t *handle, struct inode *inode, 137a86c6181SAlex Tomas struct ext4_ext_path *path) 138a86c6181SAlex Tomas { 139a86c6181SAlex Tomas if (path->p_bh) { 140a86c6181SAlex Tomas /* path points to block */ 1415d601255Sliang xie BUFFER_TRACE(path->p_bh, "get_write_access"); 142a86c6181SAlex Tomas return ext4_journal_get_write_access(handle, path->p_bh); 143a86c6181SAlex Tomas } 144a86c6181SAlex Tomas /* path points to leaf/index in inode body */ 145a86c6181SAlex Tomas /* we use in-core data, no need to protect them */ 146a86c6181SAlex Tomas return 0; 147a86c6181SAlex Tomas } 148a86c6181SAlex Tomas 149a86c6181SAlex Tomas /* 150a86c6181SAlex Tomas * could return: 151a86c6181SAlex Tomas * - EROFS 152a86c6181SAlex Tomas * - ENOMEM 153a86c6181SAlex Tomas * - EIO 154a86c6181SAlex Tomas */ 15543f81677SEric Biggers static int __ext4_ext_dirty(const char *where, unsigned int line, 15643f81677SEric Biggers handle_t *handle, struct inode *inode, 15743f81677SEric Biggers struct ext4_ext_path *path) 158a86c6181SAlex Tomas { 159a86c6181SAlex Tomas int err; 1604b1f1660SDmitry Monakhov 1614b1f1660SDmitry Monakhov WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem)); 162a86c6181SAlex Tomas if (path->p_bh) { 1637ac5990dSDarrick J. Wong ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh)); 164a86c6181SAlex Tomas /* path points to block */ 1659ea7a0dfSTheodore Ts'o err = __ext4_handle_dirty_metadata(where, line, handle, 1669ea7a0dfSTheodore Ts'o inode, path->p_bh); 167a86c6181SAlex Tomas } else { 168a86c6181SAlex Tomas /* path points to leaf/index in inode body */ 169a86c6181SAlex Tomas err = ext4_mark_inode_dirty(handle, inode); 170a86c6181SAlex Tomas } 171a86c6181SAlex Tomas return err; 172a86c6181SAlex Tomas } 173a86c6181SAlex Tomas 17443f81677SEric Biggers #define ext4_ext_dirty(handle, inode, path) \ 17543f81677SEric Biggers __ext4_ext_dirty(__func__, __LINE__, (handle), (inode), (path)) 17643f81677SEric Biggers 177f65e6fbaSAlex Tomas static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode, 178a86c6181SAlex Tomas struct ext4_ext_path *path, 179725d26d3SAneesh Kumar K.V ext4_lblk_t block) 180a86c6181SAlex Tomas { 181a86c6181SAlex Tomas if (path) { 18281fdbb4aSYongqiang Yang int depth = path->p_depth; 183a86c6181SAlex Tomas struct ext4_extent *ex; 184a86c6181SAlex Tomas 185ad4fb9caSKazuya Mio /* 186ad4fb9caSKazuya Mio * Try to predict block placement assuming that we are 187ad4fb9caSKazuya Mio * filling in a file which will eventually be 188ad4fb9caSKazuya Mio * non-sparse --- i.e., in the case of libbfd writing 189ad4fb9caSKazuya Mio * an ELF object sections out-of-order but in a way 190ad4fb9caSKazuya Mio * the eventually results in a contiguous object or 191ad4fb9caSKazuya Mio * executable file, or some database extending a table 192ad4fb9caSKazuya Mio * space file. However, this is actually somewhat 193ad4fb9caSKazuya Mio * non-ideal if we are writing a sparse file such as 194ad4fb9caSKazuya Mio * qemu or KVM writing a raw image file that is going 195ad4fb9caSKazuya Mio * to stay fairly sparse, since it will end up 196ad4fb9caSKazuya Mio * fragmenting the file system's free space. Maybe we 197ad4fb9caSKazuya Mio * should have some hueristics or some way to allow 198ad4fb9caSKazuya Mio * userspace to pass a hint to file system, 199b8d6568aSTao Ma * especially if the latter case turns out to be 200ad4fb9caSKazuya Mio * common. 201ad4fb9caSKazuya Mio */ 2027e028976SAvantika Mathur ex = path[depth].p_ext; 203ad4fb9caSKazuya Mio if (ex) { 204ad4fb9caSKazuya Mio ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex); 205ad4fb9caSKazuya Mio ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block); 206ad4fb9caSKazuya Mio 207ad4fb9caSKazuya Mio if (block > ext_block) 208ad4fb9caSKazuya Mio return ext_pblk + (block - ext_block); 209ad4fb9caSKazuya Mio else 210ad4fb9caSKazuya Mio return ext_pblk - (ext_block - block); 211ad4fb9caSKazuya Mio } 212a86c6181SAlex Tomas 213d0d856e8SRandy Dunlap /* it looks like index is empty; 214d0d856e8SRandy Dunlap * try to find starting block from index itself */ 215a86c6181SAlex Tomas if (path[depth].p_bh) 216a86c6181SAlex Tomas return path[depth].p_bh->b_blocknr; 217a86c6181SAlex Tomas } 218a86c6181SAlex Tomas 219a86c6181SAlex Tomas /* OK. use inode's group */ 220f86186b4SEric Sandeen return ext4_inode_to_goal_block(inode); 221a86c6181SAlex Tomas } 222a86c6181SAlex Tomas 223654b4908SAneesh Kumar K.V /* 224654b4908SAneesh Kumar K.V * Allocation for a meta data block 225654b4908SAneesh Kumar K.V */ 226f65e6fbaSAlex Tomas static ext4_fsblk_t 227654b4908SAneesh Kumar K.V ext4_ext_new_meta_block(handle_t *handle, struct inode *inode, 228a86c6181SAlex Tomas struct ext4_ext_path *path, 22955f020dbSAllison Henderson struct ext4_extent *ex, int *err, unsigned int flags) 230a86c6181SAlex Tomas { 231f65e6fbaSAlex Tomas ext4_fsblk_t goal, newblock; 232a86c6181SAlex Tomas 233a86c6181SAlex Tomas goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block)); 23455f020dbSAllison Henderson newblock = ext4_new_meta_blocks(handle, inode, goal, flags, 23555f020dbSAllison Henderson NULL, err); 236a86c6181SAlex Tomas return newblock; 237a86c6181SAlex Tomas } 238a86c6181SAlex Tomas 23955ad63bfSTheodore Ts'o static inline int ext4_ext_space_block(struct inode *inode, int check) 240a86c6181SAlex Tomas { 241a86c6181SAlex Tomas int size; 242a86c6181SAlex Tomas 243a86c6181SAlex Tomas size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header)) 244a86c6181SAlex Tomas / sizeof(struct ext4_extent); 245bbf2f9fbSRobert P. J. Day #ifdef AGGRESSIVE_TEST 24602dc62fbSYongqiang Yang if (!check && size > 6) 247a86c6181SAlex Tomas size = 6; 248a86c6181SAlex Tomas #endif 249a86c6181SAlex Tomas return size; 250a86c6181SAlex Tomas } 251a86c6181SAlex Tomas 25255ad63bfSTheodore Ts'o static inline int ext4_ext_space_block_idx(struct inode *inode, int check) 253a86c6181SAlex Tomas { 254a86c6181SAlex Tomas int size; 255a86c6181SAlex Tomas 256a86c6181SAlex Tomas size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header)) 257a86c6181SAlex Tomas / sizeof(struct ext4_extent_idx); 258bbf2f9fbSRobert P. J. Day #ifdef AGGRESSIVE_TEST 25902dc62fbSYongqiang Yang if (!check && size > 5) 260a86c6181SAlex Tomas size = 5; 261a86c6181SAlex Tomas #endif 262a86c6181SAlex Tomas return size; 263a86c6181SAlex Tomas } 264a86c6181SAlex Tomas 26555ad63bfSTheodore Ts'o static inline int ext4_ext_space_root(struct inode *inode, int check) 266a86c6181SAlex Tomas { 267a86c6181SAlex Tomas int size; 268a86c6181SAlex Tomas 269a86c6181SAlex Tomas size = sizeof(EXT4_I(inode)->i_data); 270a86c6181SAlex Tomas size -= sizeof(struct ext4_extent_header); 271a86c6181SAlex Tomas size /= sizeof(struct ext4_extent); 272bbf2f9fbSRobert P. J. Day #ifdef AGGRESSIVE_TEST 27302dc62fbSYongqiang Yang if (!check && size > 3) 274a86c6181SAlex Tomas size = 3; 275a86c6181SAlex Tomas #endif 276a86c6181SAlex Tomas return size; 277a86c6181SAlex Tomas } 278a86c6181SAlex Tomas 27955ad63bfSTheodore Ts'o static inline int ext4_ext_space_root_idx(struct inode *inode, int check) 280a86c6181SAlex Tomas { 281a86c6181SAlex Tomas int size; 282a86c6181SAlex Tomas 283a86c6181SAlex Tomas size = sizeof(EXT4_I(inode)->i_data); 284a86c6181SAlex Tomas size -= sizeof(struct ext4_extent_header); 285a86c6181SAlex Tomas size /= sizeof(struct ext4_extent_idx); 286bbf2f9fbSRobert P. J. Day #ifdef AGGRESSIVE_TEST 28702dc62fbSYongqiang Yang if (!check && size > 4) 288a86c6181SAlex Tomas size = 4; 289a86c6181SAlex Tomas #endif 290a86c6181SAlex Tomas return size; 291a86c6181SAlex Tomas } 292a86c6181SAlex Tomas 293fcf6b1b7SDmitry Monakhov static inline int 294fcf6b1b7SDmitry Monakhov ext4_force_split_extent_at(handle_t *handle, struct inode *inode, 295dfe50809STheodore Ts'o struct ext4_ext_path **ppath, ext4_lblk_t lblk, 296fcf6b1b7SDmitry Monakhov int nofail) 297fcf6b1b7SDmitry Monakhov { 298dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 299fcf6b1b7SDmitry Monakhov int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext); 30073c384c0STheodore Ts'o int flags = EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO; 30173c384c0STheodore Ts'o 30273c384c0STheodore Ts'o if (nofail) 30373c384c0STheodore Ts'o flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL | EXT4_EX_NOFAIL; 304fcf6b1b7SDmitry Monakhov 305dfe50809STheodore Ts'o return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ? 306fcf6b1b7SDmitry Monakhov EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0, 30773c384c0STheodore Ts'o flags); 308fcf6b1b7SDmitry Monakhov } 309fcf6b1b7SDmitry Monakhov 310c29c0ae7SAlex Tomas static int 311c29c0ae7SAlex Tomas ext4_ext_max_entries(struct inode *inode, int depth) 312c29c0ae7SAlex Tomas { 313c29c0ae7SAlex Tomas int max; 314c29c0ae7SAlex Tomas 315c29c0ae7SAlex Tomas if (depth == ext_depth(inode)) { 316c29c0ae7SAlex Tomas if (depth == 0) 31755ad63bfSTheodore Ts'o max = ext4_ext_space_root(inode, 1); 318c29c0ae7SAlex Tomas else 31955ad63bfSTheodore Ts'o max = ext4_ext_space_root_idx(inode, 1); 320c29c0ae7SAlex Tomas } else { 321c29c0ae7SAlex Tomas if (depth == 0) 32255ad63bfSTheodore Ts'o max = ext4_ext_space_block(inode, 1); 323c29c0ae7SAlex Tomas else 32455ad63bfSTheodore Ts'o max = ext4_ext_space_block_idx(inode, 1); 325c29c0ae7SAlex Tomas } 326c29c0ae7SAlex Tomas 327c29c0ae7SAlex Tomas return max; 328c29c0ae7SAlex Tomas } 329c29c0ae7SAlex Tomas 33056b19868SAneesh Kumar K.V static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext) 33156b19868SAneesh Kumar K.V { 332bf89d16fSTheodore Ts'o ext4_fsblk_t block = ext4_ext_pblock(ext); 33356b19868SAneesh Kumar K.V int len = ext4_ext_get_actual_len(ext); 3345946d089SEryu Guan ext4_lblk_t lblock = le32_to_cpu(ext->ee_block); 335e84a26ceSTheodore Ts'o 336f70749caSVegard Nossum /* 337f70749caSVegard Nossum * We allow neither: 338f70749caSVegard Nossum * - zero length 339f70749caSVegard Nossum * - overflow/wrap-around 340f70749caSVegard Nossum */ 341f70749caSVegard Nossum if (lblock + len <= lblock) 34231d4f3a2STheodore Ts'o return 0; 3436fd058f7STheodore Ts'o return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len); 34456b19868SAneesh Kumar K.V } 34556b19868SAneesh Kumar K.V 34656b19868SAneesh Kumar K.V static int ext4_valid_extent_idx(struct inode *inode, 34756b19868SAneesh Kumar K.V struct ext4_extent_idx *ext_idx) 34856b19868SAneesh Kumar K.V { 349bf89d16fSTheodore Ts'o ext4_fsblk_t block = ext4_idx_pblock(ext_idx); 350e84a26ceSTheodore Ts'o 3516fd058f7STheodore Ts'o return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1); 35256b19868SAneesh Kumar K.V } 35356b19868SAneesh Kumar K.V 35456b19868SAneesh Kumar K.V static int ext4_valid_extent_entries(struct inode *inode, 35556b19868SAneesh Kumar K.V struct ext4_extent_header *eh, 35654d3adbcSTheodore Ts'o ext4_fsblk_t *pblk, int depth) 35756b19868SAneesh Kumar K.V { 35856b19868SAneesh Kumar K.V unsigned short entries; 35956b19868SAneesh Kumar K.V if (eh->eh_entries == 0) 36056b19868SAneesh Kumar K.V return 1; 36156b19868SAneesh Kumar K.V 36256b19868SAneesh Kumar K.V entries = le16_to_cpu(eh->eh_entries); 36356b19868SAneesh Kumar K.V 36456b19868SAneesh Kumar K.V if (depth == 0) { 36556b19868SAneesh Kumar K.V /* leaf entries */ 36681fdbb4aSYongqiang Yang struct ext4_extent *ext = EXT_FIRST_EXTENT(eh); 3675946d089SEryu Guan ext4_lblk_t lblock = 0; 3685946d089SEryu Guan ext4_lblk_t prev = 0; 3695946d089SEryu Guan int len = 0; 37056b19868SAneesh Kumar K.V while (entries) { 37156b19868SAneesh Kumar K.V if (!ext4_valid_extent(inode, ext)) 37256b19868SAneesh Kumar K.V return 0; 3735946d089SEryu Guan 3745946d089SEryu Guan /* Check for overlapping extents */ 3755946d089SEryu Guan lblock = le32_to_cpu(ext->ee_block); 3765946d089SEryu Guan len = ext4_ext_get_actual_len(ext); 3775946d089SEryu Guan if ((lblock <= prev) && prev) { 37854d3adbcSTheodore Ts'o *pblk = ext4_ext_pblock(ext); 3795946d089SEryu Guan return 0; 3805946d089SEryu Guan } 38156b19868SAneesh Kumar K.V ext++; 38256b19868SAneesh Kumar K.V entries--; 3835946d089SEryu Guan prev = lblock + len - 1; 38456b19868SAneesh Kumar K.V } 38556b19868SAneesh Kumar K.V } else { 38681fdbb4aSYongqiang Yang struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh); 38756b19868SAneesh Kumar K.V while (entries) { 38856b19868SAneesh Kumar K.V if (!ext4_valid_extent_idx(inode, ext_idx)) 38956b19868SAneesh Kumar K.V return 0; 39056b19868SAneesh Kumar K.V ext_idx++; 39156b19868SAneesh Kumar K.V entries--; 39256b19868SAneesh Kumar K.V } 39356b19868SAneesh Kumar K.V } 39456b19868SAneesh Kumar K.V return 1; 39556b19868SAneesh Kumar K.V } 39656b19868SAneesh Kumar K.V 397c398eda0STheodore Ts'o static int __ext4_ext_check(const char *function, unsigned int line, 398c398eda0STheodore Ts'o struct inode *inode, struct ext4_extent_header *eh, 399c349179bSTheodore Ts'o int depth, ext4_fsblk_t pblk) 400c29c0ae7SAlex Tomas { 401c29c0ae7SAlex Tomas const char *error_msg; 4026a797d27SDarrick J. Wong int max = 0, err = -EFSCORRUPTED; 403c29c0ae7SAlex Tomas 404c29c0ae7SAlex Tomas if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) { 405c29c0ae7SAlex Tomas error_msg = "invalid magic"; 406c29c0ae7SAlex Tomas goto corrupted; 407c29c0ae7SAlex Tomas } 408c29c0ae7SAlex Tomas if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) { 409c29c0ae7SAlex Tomas error_msg = "unexpected eh_depth"; 410c29c0ae7SAlex Tomas goto corrupted; 411c29c0ae7SAlex Tomas } 412c29c0ae7SAlex Tomas if (unlikely(eh->eh_max == 0)) { 413c29c0ae7SAlex Tomas error_msg = "invalid eh_max"; 414c29c0ae7SAlex Tomas goto corrupted; 415c29c0ae7SAlex Tomas } 416c29c0ae7SAlex Tomas max = ext4_ext_max_entries(inode, depth); 417c29c0ae7SAlex Tomas if (unlikely(le16_to_cpu(eh->eh_max) > max)) { 418c29c0ae7SAlex Tomas error_msg = "too large eh_max"; 419c29c0ae7SAlex Tomas goto corrupted; 420c29c0ae7SAlex Tomas } 421c29c0ae7SAlex Tomas if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) { 422c29c0ae7SAlex Tomas error_msg = "invalid eh_entries"; 423c29c0ae7SAlex Tomas goto corrupted; 424c29c0ae7SAlex Tomas } 42554d3adbcSTheodore Ts'o if (!ext4_valid_extent_entries(inode, eh, &pblk, depth)) { 42656b19868SAneesh Kumar K.V error_msg = "invalid extent entries"; 42756b19868SAneesh Kumar K.V goto corrupted; 42856b19868SAneesh Kumar K.V } 4297bc94916SVegard Nossum if (unlikely(depth > 32)) { 4307bc94916SVegard Nossum error_msg = "too large eh_depth"; 4317bc94916SVegard Nossum goto corrupted; 4327bc94916SVegard Nossum } 4337ac5990dSDarrick J. Wong /* Verify checksum on non-root extent tree nodes */ 4347ac5990dSDarrick J. Wong if (ext_depth(inode) != depth && 4357ac5990dSDarrick J. Wong !ext4_extent_block_csum_verify(inode, eh)) { 4367ac5990dSDarrick J. Wong error_msg = "extent tree corrupted"; 4376a797d27SDarrick J. Wong err = -EFSBADCRC; 4387ac5990dSDarrick J. Wong goto corrupted; 4397ac5990dSDarrick J. Wong } 440c29c0ae7SAlex Tomas return 0; 441c29c0ae7SAlex Tomas 442c29c0ae7SAlex Tomas corrupted: 44354d3adbcSTheodore Ts'o ext4_error_inode_err(inode, function, line, 0, -err, 444c349179bSTheodore Ts'o "pblk %llu bad header/extent: %s - magic %x, " 445c29c0ae7SAlex Tomas "entries %u, max %u(%u), depth %u(%u)", 446c349179bSTheodore Ts'o (unsigned long long) pblk, error_msg, 447c349179bSTheodore Ts'o le16_to_cpu(eh->eh_magic), 44854d3adbcSTheodore Ts'o le16_to_cpu(eh->eh_entries), 44954d3adbcSTheodore Ts'o le16_to_cpu(eh->eh_max), 450c29c0ae7SAlex Tomas max, le16_to_cpu(eh->eh_depth), depth); 4516a797d27SDarrick J. Wong return err; 452c29c0ae7SAlex Tomas } 453c29c0ae7SAlex Tomas 454c349179bSTheodore Ts'o #define ext4_ext_check(inode, eh, depth, pblk) \ 455c349179bSTheodore Ts'o __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk)) 456c29c0ae7SAlex Tomas 4577a262f7cSAneesh Kumar K.V int ext4_ext_check_inode(struct inode *inode) 4587a262f7cSAneesh Kumar K.V { 459c349179bSTheodore Ts'o return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0); 4607a262f7cSAneesh Kumar K.V } 4617a262f7cSAneesh Kumar K.V 4624068664eSDmitry Monakhov static void ext4_cache_extents(struct inode *inode, 4634068664eSDmitry Monakhov struct ext4_extent_header *eh) 4644068664eSDmitry Monakhov { 4654068664eSDmitry Monakhov struct ext4_extent *ex = EXT_FIRST_EXTENT(eh); 4664068664eSDmitry Monakhov ext4_lblk_t prev = 0; 4674068664eSDmitry Monakhov int i; 4684068664eSDmitry Monakhov 4694068664eSDmitry Monakhov for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) { 4704068664eSDmitry Monakhov unsigned int status = EXTENT_STATUS_WRITTEN; 4714068664eSDmitry Monakhov ext4_lblk_t lblk = le32_to_cpu(ex->ee_block); 4724068664eSDmitry Monakhov int len = ext4_ext_get_actual_len(ex); 4734068664eSDmitry Monakhov 4744068664eSDmitry Monakhov if (prev && (prev != lblk)) 4754068664eSDmitry Monakhov ext4_es_cache_extent(inode, prev, lblk - prev, ~0, 4764068664eSDmitry Monakhov EXTENT_STATUS_HOLE); 4774068664eSDmitry Monakhov 4784068664eSDmitry Monakhov if (ext4_ext_is_unwritten(ex)) 4794068664eSDmitry Monakhov status = EXTENT_STATUS_UNWRITTEN; 4804068664eSDmitry Monakhov ext4_es_cache_extent(inode, lblk, len, 4814068664eSDmitry Monakhov ext4_ext_pblock(ex), status); 4824068664eSDmitry Monakhov prev = lblk + len; 4834068664eSDmitry Monakhov } 4844068664eSDmitry Monakhov } 4854068664eSDmitry Monakhov 4867d7ea89eSTheodore Ts'o static struct buffer_head * 4877d7ea89eSTheodore Ts'o __read_extent_tree_block(const char *function, unsigned int line, 488107a7bd3STheodore Ts'o struct inode *inode, ext4_fsblk_t pblk, int depth, 489107a7bd3STheodore Ts'o int flags) 490f8489128SDarrick J. Wong { 4917d7ea89eSTheodore Ts'o struct buffer_head *bh; 4927d7ea89eSTheodore Ts'o int err; 49373c384c0STheodore Ts'o gfp_t gfp_flags = __GFP_MOVABLE | GFP_NOFS; 494f8489128SDarrick J. Wong 49573c384c0STheodore Ts'o if (flags & EXT4_EX_NOFAIL) 49673c384c0STheodore Ts'o gfp_flags |= __GFP_NOFAIL; 49773c384c0STheodore Ts'o 49873c384c0STheodore Ts'o bh = sb_getblk_gfp(inode->i_sb, pblk, gfp_flags); 4997d7ea89eSTheodore Ts'o if (unlikely(!bh)) 5007d7ea89eSTheodore Ts'o return ERR_PTR(-ENOMEM); 5017d7ea89eSTheodore Ts'o 5027d7ea89eSTheodore Ts'o if (!bh_uptodate_or_lock(bh)) { 5037d7ea89eSTheodore Ts'o trace_ext4_ext_load_extent(inode, pblk, _RET_IP_); 5047d7ea89eSTheodore Ts'o err = bh_submit_read(bh); 5057d7ea89eSTheodore Ts'o if (err < 0) 5067d7ea89eSTheodore Ts'o goto errout; 5077d7ea89eSTheodore Ts'o } 5087869a4a6STheodore Ts'o if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE)) 5097d7ea89eSTheodore Ts'o return bh; 5100a944e8aSTheodore Ts'o if (!ext4_has_feature_journal(inode->i_sb) || 5110a944e8aSTheodore Ts'o (inode->i_ino != 5120a944e8aSTheodore Ts'o le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum))) { 5137d7ea89eSTheodore Ts'o err = __ext4_ext_check(function, line, inode, 514c349179bSTheodore Ts'o ext_block_hdr(bh), depth, pblk); 5157d7ea89eSTheodore Ts'o if (err) 5167d7ea89eSTheodore Ts'o goto errout; 5170a944e8aSTheodore Ts'o } 518f8489128SDarrick J. Wong set_buffer_verified(bh); 519107a7bd3STheodore Ts'o /* 520107a7bd3STheodore Ts'o * If this is a leaf block, cache all of its entries 521107a7bd3STheodore Ts'o */ 522107a7bd3STheodore Ts'o if (!(flags & EXT4_EX_NOCACHE) && depth == 0) { 523107a7bd3STheodore Ts'o struct ext4_extent_header *eh = ext_block_hdr(bh); 5244068664eSDmitry Monakhov ext4_cache_extents(inode, eh); 525107a7bd3STheodore Ts'o } 5267d7ea89eSTheodore Ts'o return bh; 5277d7ea89eSTheodore Ts'o errout: 5287d7ea89eSTheodore Ts'o put_bh(bh); 5297d7ea89eSTheodore Ts'o return ERR_PTR(err); 5307d7ea89eSTheodore Ts'o 531f8489128SDarrick J. Wong } 532f8489128SDarrick J. Wong 533107a7bd3STheodore Ts'o #define read_extent_tree_block(inode, pblk, depth, flags) \ 534107a7bd3STheodore Ts'o __read_extent_tree_block(__func__, __LINE__, (inode), (pblk), \ 535107a7bd3STheodore Ts'o (depth), (flags)) 536f8489128SDarrick J. Wong 5377869a4a6STheodore Ts'o /* 5387869a4a6STheodore Ts'o * This function is called to cache a file's extent information in the 5397869a4a6STheodore Ts'o * extent status tree 5407869a4a6STheodore Ts'o */ 5417869a4a6STheodore Ts'o int ext4_ext_precache(struct inode *inode) 5427869a4a6STheodore Ts'o { 5437869a4a6STheodore Ts'o struct ext4_inode_info *ei = EXT4_I(inode); 5447869a4a6STheodore Ts'o struct ext4_ext_path *path = NULL; 5457869a4a6STheodore Ts'o struct buffer_head *bh; 5467869a4a6STheodore Ts'o int i = 0, depth, ret = 0; 5477869a4a6STheodore Ts'o 5487869a4a6STheodore Ts'o if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) 5497869a4a6STheodore Ts'o return 0; /* not an extent-mapped inode */ 5507869a4a6STheodore Ts'o 5517869a4a6STheodore Ts'o down_read(&ei->i_data_sem); 5527869a4a6STheodore Ts'o depth = ext_depth(inode); 5537869a4a6STheodore Ts'o 5542f424a5aSRitesh Harjani /* Don't cache anything if there are no external extent blocks */ 5552f424a5aSRitesh Harjani if (!depth) { 5562f424a5aSRitesh Harjani up_read(&ei->i_data_sem); 5572f424a5aSRitesh Harjani return ret; 5582f424a5aSRitesh Harjani } 5592f424a5aSRitesh Harjani 5606396bb22SKees Cook path = kcalloc(depth + 1, sizeof(struct ext4_ext_path), 5617869a4a6STheodore Ts'o GFP_NOFS); 5627869a4a6STheodore Ts'o if (path == NULL) { 5637869a4a6STheodore Ts'o up_read(&ei->i_data_sem); 5647869a4a6STheodore Ts'o return -ENOMEM; 5657869a4a6STheodore Ts'o } 5667869a4a6STheodore Ts'o 5677869a4a6STheodore Ts'o path[0].p_hdr = ext_inode_hdr(inode); 5687869a4a6STheodore Ts'o ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0); 5697869a4a6STheodore Ts'o if (ret) 5707869a4a6STheodore Ts'o goto out; 5717869a4a6STheodore Ts'o path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr); 5727869a4a6STheodore Ts'o while (i >= 0) { 5737869a4a6STheodore Ts'o /* 5747869a4a6STheodore Ts'o * If this is a leaf block or we've reached the end of 5757869a4a6STheodore Ts'o * the index block, go up 5767869a4a6STheodore Ts'o */ 5777869a4a6STheodore Ts'o if ((i == depth) || 5787869a4a6STheodore Ts'o path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) { 5797869a4a6STheodore Ts'o brelse(path[i].p_bh); 5807869a4a6STheodore Ts'o path[i].p_bh = NULL; 5817869a4a6STheodore Ts'o i--; 5827869a4a6STheodore Ts'o continue; 5837869a4a6STheodore Ts'o } 5847869a4a6STheodore Ts'o bh = read_extent_tree_block(inode, 5857869a4a6STheodore Ts'o ext4_idx_pblock(path[i].p_idx++), 5867869a4a6STheodore Ts'o depth - i - 1, 5877869a4a6STheodore Ts'o EXT4_EX_FORCE_CACHE); 5887869a4a6STheodore Ts'o if (IS_ERR(bh)) { 5897869a4a6STheodore Ts'o ret = PTR_ERR(bh); 5907869a4a6STheodore Ts'o break; 5917869a4a6STheodore Ts'o } 5927869a4a6STheodore Ts'o i++; 5937869a4a6STheodore Ts'o path[i].p_bh = bh; 5947869a4a6STheodore Ts'o path[i].p_hdr = ext_block_hdr(bh); 5957869a4a6STheodore Ts'o path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr); 5967869a4a6STheodore Ts'o } 5977869a4a6STheodore Ts'o ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED); 5987869a4a6STheodore Ts'o out: 5997869a4a6STheodore Ts'o up_read(&ei->i_data_sem); 6007869a4a6STheodore Ts'o ext4_ext_drop_refs(path); 6017869a4a6STheodore Ts'o kfree(path); 6027869a4a6STheodore Ts'o return ret; 6037869a4a6STheodore Ts'o } 6047869a4a6STheodore Ts'o 605a86c6181SAlex Tomas #ifdef EXT_DEBUG 606a86c6181SAlex Tomas static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path) 607a86c6181SAlex Tomas { 608a86c6181SAlex Tomas int k, l = path->p_depth; 609a86c6181SAlex Tomas 610a86c6181SAlex Tomas ext_debug("path:"); 611a86c6181SAlex Tomas for (k = 0; k <= l; k++, path++) { 612a86c6181SAlex Tomas if (path->p_idx) { 6136e89bbb7SEric Biggers ext_debug(" %d->%llu", 6146e89bbb7SEric Biggers le32_to_cpu(path->p_idx->ei_block), 615bf89d16fSTheodore Ts'o ext4_idx_pblock(path->p_idx)); 616a86c6181SAlex Tomas } else if (path->p_ext) { 617553f9008SMingming ext_debug(" %d:[%d]%d:%llu ", 618a86c6181SAlex Tomas le32_to_cpu(path->p_ext->ee_block), 619556615dcSLukas Czerner ext4_ext_is_unwritten(path->p_ext), 620a2df2a63SAmit Arora ext4_ext_get_actual_len(path->p_ext), 621bf89d16fSTheodore Ts'o ext4_ext_pblock(path->p_ext)); 622a86c6181SAlex Tomas } else 623a86c6181SAlex Tomas ext_debug(" []"); 624a86c6181SAlex Tomas } 625a86c6181SAlex Tomas ext_debug("\n"); 626a86c6181SAlex Tomas } 627a86c6181SAlex Tomas 628a86c6181SAlex Tomas static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path) 629a86c6181SAlex Tomas { 630a86c6181SAlex Tomas int depth = ext_depth(inode); 631a86c6181SAlex Tomas struct ext4_extent_header *eh; 632a86c6181SAlex Tomas struct ext4_extent *ex; 633a86c6181SAlex Tomas int i; 634a86c6181SAlex Tomas 635a86c6181SAlex Tomas if (!path) 636a86c6181SAlex Tomas return; 637a86c6181SAlex Tomas 638a86c6181SAlex Tomas eh = path[depth].p_hdr; 639a86c6181SAlex Tomas ex = EXT_FIRST_EXTENT(eh); 640a86c6181SAlex Tomas 641553f9008SMingming ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino); 642553f9008SMingming 643a86c6181SAlex Tomas for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) { 644553f9008SMingming ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block), 645556615dcSLukas Czerner ext4_ext_is_unwritten(ex), 646bf89d16fSTheodore Ts'o ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex)); 647a86c6181SAlex Tomas } 648a86c6181SAlex Tomas ext_debug("\n"); 649a86c6181SAlex Tomas } 6501b16da77SYongqiang Yang 6511b16da77SYongqiang Yang static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path, 6521b16da77SYongqiang Yang ext4_fsblk_t newblock, int level) 6531b16da77SYongqiang Yang { 6541b16da77SYongqiang Yang int depth = ext_depth(inode); 6551b16da77SYongqiang Yang struct ext4_extent *ex; 6561b16da77SYongqiang Yang 6571b16da77SYongqiang Yang if (depth != level) { 6581b16da77SYongqiang Yang struct ext4_extent_idx *idx; 6591b16da77SYongqiang Yang idx = path[level].p_idx; 6601b16da77SYongqiang Yang while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) { 6611b16da77SYongqiang Yang ext_debug("%d: move %d:%llu in new index %llu\n", level, 6621b16da77SYongqiang Yang le32_to_cpu(idx->ei_block), 6631b16da77SYongqiang Yang ext4_idx_pblock(idx), 6641b16da77SYongqiang Yang newblock); 6651b16da77SYongqiang Yang idx++; 6661b16da77SYongqiang Yang } 6671b16da77SYongqiang Yang 6681b16da77SYongqiang Yang return; 6691b16da77SYongqiang Yang } 6701b16da77SYongqiang Yang 6711b16da77SYongqiang Yang ex = path[depth].p_ext; 6721b16da77SYongqiang Yang while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) { 6731b16da77SYongqiang Yang ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n", 6741b16da77SYongqiang Yang le32_to_cpu(ex->ee_block), 6751b16da77SYongqiang Yang ext4_ext_pblock(ex), 676556615dcSLukas Czerner ext4_ext_is_unwritten(ex), 6771b16da77SYongqiang Yang ext4_ext_get_actual_len(ex), 6781b16da77SYongqiang Yang newblock); 6791b16da77SYongqiang Yang ex++; 6801b16da77SYongqiang Yang } 6811b16da77SYongqiang Yang } 6821b16da77SYongqiang Yang 683a86c6181SAlex Tomas #else 684a86c6181SAlex Tomas #define ext4_ext_show_path(inode, path) 685a86c6181SAlex Tomas #define ext4_ext_show_leaf(inode, path) 6861b16da77SYongqiang Yang #define ext4_ext_show_move(inode, path, newblock, level) 687a86c6181SAlex Tomas #endif 688a86c6181SAlex Tomas 689b35905c1SAneesh Kumar K.V void ext4_ext_drop_refs(struct ext4_ext_path *path) 690a86c6181SAlex Tomas { 691b7ea89adSTheodore Ts'o int depth, i; 692a86c6181SAlex Tomas 693b7ea89adSTheodore Ts'o if (!path) 694b7ea89adSTheodore Ts'o return; 695b7ea89adSTheodore Ts'o depth = path->p_depth; 696de745485SEric Biggers for (i = 0; i <= depth; i++, path++) { 697a86c6181SAlex Tomas if (path->p_bh) { 698a86c6181SAlex Tomas brelse(path->p_bh); 699a86c6181SAlex Tomas path->p_bh = NULL; 700a86c6181SAlex Tomas } 701a86c6181SAlex Tomas } 702de745485SEric Biggers } 703a86c6181SAlex Tomas 704a86c6181SAlex Tomas /* 705d0d856e8SRandy Dunlap * ext4_ext_binsearch_idx: 706d0d856e8SRandy Dunlap * binary search for the closest index of the given block 707c29c0ae7SAlex Tomas * the header must be checked before calling this 708a86c6181SAlex Tomas */ 709a86c6181SAlex Tomas static void 710725d26d3SAneesh Kumar K.V ext4_ext_binsearch_idx(struct inode *inode, 711725d26d3SAneesh Kumar K.V struct ext4_ext_path *path, ext4_lblk_t block) 712a86c6181SAlex Tomas { 713a86c6181SAlex Tomas struct ext4_extent_header *eh = path->p_hdr; 714a86c6181SAlex Tomas struct ext4_extent_idx *r, *l, *m; 715a86c6181SAlex Tomas 716a86c6181SAlex Tomas 717bba90743SEric Sandeen ext_debug("binsearch for %u(idx): ", block); 718a86c6181SAlex Tomas 719a86c6181SAlex Tomas l = EXT_FIRST_INDEX(eh) + 1; 720e9f410b1SDmitry Monakhov r = EXT_LAST_INDEX(eh); 721a86c6181SAlex Tomas while (l <= r) { 722a86c6181SAlex Tomas m = l + (r - l) / 2; 723a86c6181SAlex Tomas if (block < le32_to_cpu(m->ei_block)) 724a86c6181SAlex Tomas r = m - 1; 725a86c6181SAlex Tomas else 726a86c6181SAlex Tomas l = m + 1; 72726d535edSDmitry Monakhov ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block), 72826d535edSDmitry Monakhov m, le32_to_cpu(m->ei_block), 72926d535edSDmitry Monakhov r, le32_to_cpu(r->ei_block)); 730a86c6181SAlex Tomas } 731a86c6181SAlex Tomas 732a86c6181SAlex Tomas path->p_idx = l - 1; 7334a3c3a51SZheng Liu ext_debug(" -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block), 734bf89d16fSTheodore Ts'o ext4_idx_pblock(path->p_idx)); 735a86c6181SAlex Tomas 736a86c6181SAlex Tomas #ifdef CHECK_BINSEARCH 737a86c6181SAlex Tomas { 738a86c6181SAlex Tomas struct ext4_extent_idx *chix, *ix; 739a86c6181SAlex Tomas int k; 740a86c6181SAlex Tomas 741a86c6181SAlex Tomas chix = ix = EXT_FIRST_INDEX(eh); 742a86c6181SAlex Tomas for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) { 7436e89bbb7SEric Biggers if (k != 0 && le32_to_cpu(ix->ei_block) <= 7446e89bbb7SEric Biggers le32_to_cpu(ix[-1].ei_block)) { 7454776004fSTheodore Ts'o printk(KERN_DEBUG "k=%d, ix=0x%p, " 7464776004fSTheodore Ts'o "first=0x%p\n", k, 747a86c6181SAlex Tomas ix, EXT_FIRST_INDEX(eh)); 7484776004fSTheodore Ts'o printk(KERN_DEBUG "%u <= %u\n", 749a86c6181SAlex Tomas le32_to_cpu(ix->ei_block), 750a86c6181SAlex Tomas le32_to_cpu(ix[-1].ei_block)); 751a86c6181SAlex Tomas } 752a86c6181SAlex Tomas BUG_ON(k && le32_to_cpu(ix->ei_block) 753a86c6181SAlex Tomas <= le32_to_cpu(ix[-1].ei_block)); 754a86c6181SAlex Tomas if (block < le32_to_cpu(ix->ei_block)) 755a86c6181SAlex Tomas break; 756a86c6181SAlex Tomas chix = ix; 757a86c6181SAlex Tomas } 758a86c6181SAlex Tomas BUG_ON(chix != path->p_idx); 759a86c6181SAlex Tomas } 760a86c6181SAlex Tomas #endif 761a86c6181SAlex Tomas 762a86c6181SAlex Tomas } 763a86c6181SAlex Tomas 764a86c6181SAlex Tomas /* 765d0d856e8SRandy Dunlap * ext4_ext_binsearch: 766d0d856e8SRandy Dunlap * binary search for closest extent of the given block 767c29c0ae7SAlex Tomas * the header must be checked before calling this 768a86c6181SAlex Tomas */ 769a86c6181SAlex Tomas static void 770725d26d3SAneesh Kumar K.V ext4_ext_binsearch(struct inode *inode, 771725d26d3SAneesh Kumar K.V struct ext4_ext_path *path, ext4_lblk_t block) 772a86c6181SAlex Tomas { 773a86c6181SAlex Tomas struct ext4_extent_header *eh = path->p_hdr; 774a86c6181SAlex Tomas struct ext4_extent *r, *l, *m; 775a86c6181SAlex Tomas 776a86c6181SAlex Tomas if (eh->eh_entries == 0) { 777a86c6181SAlex Tomas /* 778d0d856e8SRandy Dunlap * this leaf is empty: 779a86c6181SAlex Tomas * we get such a leaf in split/add case 780a86c6181SAlex Tomas */ 781a86c6181SAlex Tomas return; 782a86c6181SAlex Tomas } 783a86c6181SAlex Tomas 784bba90743SEric Sandeen ext_debug("binsearch for %u: ", block); 785a86c6181SAlex Tomas 786a86c6181SAlex Tomas l = EXT_FIRST_EXTENT(eh) + 1; 787e9f410b1SDmitry Monakhov r = EXT_LAST_EXTENT(eh); 788a86c6181SAlex Tomas 789a86c6181SAlex Tomas while (l <= r) { 790a86c6181SAlex Tomas m = l + (r - l) / 2; 791a86c6181SAlex Tomas if (block < le32_to_cpu(m->ee_block)) 792a86c6181SAlex Tomas r = m - 1; 793a86c6181SAlex Tomas else 794a86c6181SAlex Tomas l = m + 1; 79526d535edSDmitry Monakhov ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block), 79626d535edSDmitry Monakhov m, le32_to_cpu(m->ee_block), 79726d535edSDmitry Monakhov r, le32_to_cpu(r->ee_block)); 798a86c6181SAlex Tomas } 799a86c6181SAlex Tomas 800a86c6181SAlex Tomas path->p_ext = l - 1; 801553f9008SMingming ext_debug(" -> %d:%llu:[%d]%d ", 802a86c6181SAlex Tomas le32_to_cpu(path->p_ext->ee_block), 803bf89d16fSTheodore Ts'o ext4_ext_pblock(path->p_ext), 804556615dcSLukas Czerner ext4_ext_is_unwritten(path->p_ext), 805a2df2a63SAmit Arora ext4_ext_get_actual_len(path->p_ext)); 806a86c6181SAlex Tomas 807a86c6181SAlex Tomas #ifdef CHECK_BINSEARCH 808a86c6181SAlex Tomas { 809a86c6181SAlex Tomas struct ext4_extent *chex, *ex; 810a86c6181SAlex Tomas int k; 811a86c6181SAlex Tomas 812a86c6181SAlex Tomas chex = ex = EXT_FIRST_EXTENT(eh); 813a86c6181SAlex Tomas for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) { 814a86c6181SAlex Tomas BUG_ON(k && le32_to_cpu(ex->ee_block) 815a86c6181SAlex Tomas <= le32_to_cpu(ex[-1].ee_block)); 816a86c6181SAlex Tomas if (block < le32_to_cpu(ex->ee_block)) 817a86c6181SAlex Tomas break; 818a86c6181SAlex Tomas chex = ex; 819a86c6181SAlex Tomas } 820a86c6181SAlex Tomas BUG_ON(chex != path->p_ext); 821a86c6181SAlex Tomas } 822a86c6181SAlex Tomas #endif 823a86c6181SAlex Tomas 824a86c6181SAlex Tomas } 825a86c6181SAlex Tomas 8264209ae12SHarshad Shirwadkar void ext4_ext_tree_init(handle_t *handle, struct inode *inode) 827a86c6181SAlex Tomas { 828a86c6181SAlex Tomas struct ext4_extent_header *eh; 829a86c6181SAlex Tomas 830a86c6181SAlex Tomas eh = ext_inode_hdr(inode); 831a86c6181SAlex Tomas eh->eh_depth = 0; 832a86c6181SAlex Tomas eh->eh_entries = 0; 833a86c6181SAlex Tomas eh->eh_magic = EXT4_EXT_MAGIC; 83455ad63bfSTheodore Ts'o eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0)); 835a86c6181SAlex Tomas ext4_mark_inode_dirty(handle, inode); 836a86c6181SAlex Tomas } 837a86c6181SAlex Tomas 838a86c6181SAlex Tomas struct ext4_ext_path * 839ed8a1a76STheodore Ts'o ext4_find_extent(struct inode *inode, ext4_lblk_t block, 840705912caSTheodore Ts'o struct ext4_ext_path **orig_path, int flags) 841a86c6181SAlex Tomas { 842a86c6181SAlex Tomas struct ext4_extent_header *eh; 843a86c6181SAlex Tomas struct buffer_head *bh; 844705912caSTheodore Ts'o struct ext4_ext_path *path = orig_path ? *orig_path : NULL; 845705912caSTheodore Ts'o short int depth, i, ppos = 0; 846860d21e2STheodore Ts'o int ret; 84773c384c0STheodore Ts'o gfp_t gfp_flags = GFP_NOFS; 84873c384c0STheodore Ts'o 84973c384c0STheodore Ts'o if (flags & EXT4_EX_NOFAIL) 85073c384c0STheodore Ts'o gfp_flags |= __GFP_NOFAIL; 851a86c6181SAlex Tomas 852a86c6181SAlex Tomas eh = ext_inode_hdr(inode); 853c29c0ae7SAlex Tomas depth = ext_depth(inode); 854bc890a60STheodore Ts'o if (depth < 0 || depth > EXT4_MAX_EXTENT_DEPTH) { 855bc890a60STheodore Ts'o EXT4_ERROR_INODE(inode, "inode has invalid extent depth: %d", 856bc890a60STheodore Ts'o depth); 857bc890a60STheodore Ts'o ret = -EFSCORRUPTED; 858bc890a60STheodore Ts'o goto err; 859bc890a60STheodore Ts'o } 860a86c6181SAlex Tomas 86110809df8STheodore Ts'o if (path) { 862523f431cSTheodore Ts'o ext4_ext_drop_refs(path); 86310809df8STheodore Ts'o if (depth > path[0].p_maxdepth) { 86410809df8STheodore Ts'o kfree(path); 86510809df8STheodore Ts'o *orig_path = path = NULL; 86610809df8STheodore Ts'o } 86710809df8STheodore Ts'o } 86810809df8STheodore Ts'o if (!path) { 869a86c6181SAlex Tomas /* account possible depth increase */ 8706396bb22SKees Cook path = kcalloc(depth + 2, sizeof(struct ext4_ext_path), 87173c384c0STheodore Ts'o gfp_flags); 87219008f6dSTheodore Ts'o if (unlikely(!path)) 873a86c6181SAlex Tomas return ERR_PTR(-ENOMEM); 87410809df8STheodore Ts'o path[0].p_maxdepth = depth + 1; 875a86c6181SAlex Tomas } 876a86c6181SAlex Tomas path[0].p_hdr = eh; 8771973adcbSShen Feng path[0].p_bh = NULL; 878a86c6181SAlex Tomas 879c29c0ae7SAlex Tomas i = depth; 8804068664eSDmitry Monakhov if (!(flags & EXT4_EX_NOCACHE) && depth == 0) 8814068664eSDmitry Monakhov ext4_cache_extents(inode, eh); 882a86c6181SAlex Tomas /* walk through the tree */ 883a86c6181SAlex Tomas while (i) { 884a86c6181SAlex Tomas ext_debug("depth %d: num %d, max %d\n", 885a86c6181SAlex Tomas ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max)); 886c29c0ae7SAlex Tomas 887a86c6181SAlex Tomas ext4_ext_binsearch_idx(inode, path + ppos, block); 888bf89d16fSTheodore Ts'o path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx); 889a86c6181SAlex Tomas path[ppos].p_depth = i; 890a86c6181SAlex Tomas path[ppos].p_ext = NULL; 891a86c6181SAlex Tomas 892107a7bd3STheodore Ts'o bh = read_extent_tree_block(inode, path[ppos].p_block, --i, 893107a7bd3STheodore Ts'o flags); 894a1c83681SViresh Kumar if (IS_ERR(bh)) { 8957d7ea89eSTheodore Ts'o ret = PTR_ERR(bh); 896a86c6181SAlex Tomas goto err; 897860d21e2STheodore Ts'o } 8987d7ea89eSTheodore Ts'o 899a86c6181SAlex Tomas eh = ext_block_hdr(bh); 900a86c6181SAlex Tomas ppos++; 901a86c6181SAlex Tomas path[ppos].p_bh = bh; 902a86c6181SAlex Tomas path[ppos].p_hdr = eh; 903a86c6181SAlex Tomas } 904a86c6181SAlex Tomas 905a86c6181SAlex Tomas path[ppos].p_depth = i; 906a86c6181SAlex Tomas path[ppos].p_ext = NULL; 907a86c6181SAlex Tomas path[ppos].p_idx = NULL; 908a86c6181SAlex Tomas 909a86c6181SAlex Tomas /* find extent */ 910a86c6181SAlex Tomas ext4_ext_binsearch(inode, path + ppos, block); 9111973adcbSShen Feng /* if not an empty leaf */ 9121973adcbSShen Feng if (path[ppos].p_ext) 913bf89d16fSTheodore Ts'o path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext); 914a86c6181SAlex Tomas 915a86c6181SAlex Tomas ext4_ext_show_path(inode, path); 916a86c6181SAlex Tomas 917a86c6181SAlex Tomas return path; 918a86c6181SAlex Tomas 919a86c6181SAlex Tomas err: 920a86c6181SAlex Tomas ext4_ext_drop_refs(path); 921a86c6181SAlex Tomas kfree(path); 922705912caSTheodore Ts'o if (orig_path) 923705912caSTheodore Ts'o *orig_path = NULL; 924860d21e2STheodore Ts'o return ERR_PTR(ret); 925a86c6181SAlex Tomas } 926a86c6181SAlex Tomas 927a86c6181SAlex Tomas /* 928d0d856e8SRandy Dunlap * ext4_ext_insert_index: 929d0d856e8SRandy Dunlap * insert new index [@logical;@ptr] into the block at @curp; 930d0d856e8SRandy Dunlap * check where to insert: before @curp or after @curp 931a86c6181SAlex Tomas */ 9321f109d5aSTheodore Ts'o static int ext4_ext_insert_index(handle_t *handle, struct inode *inode, 933a86c6181SAlex Tomas struct ext4_ext_path *curp, 934f65e6fbaSAlex Tomas int logical, ext4_fsblk_t ptr) 935a86c6181SAlex Tomas { 936a86c6181SAlex Tomas struct ext4_extent_idx *ix; 937a86c6181SAlex Tomas int len, err; 938a86c6181SAlex Tomas 9397e028976SAvantika Mathur err = ext4_ext_get_access(handle, inode, curp); 9407e028976SAvantika Mathur if (err) 941a86c6181SAlex Tomas return err; 942a86c6181SAlex Tomas 943273df556SFrank Mayhar if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) { 944273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 945273df556SFrank Mayhar "logical %d == ei_block %d!", 946273df556SFrank Mayhar logical, le32_to_cpu(curp->p_idx->ei_block)); 9476a797d27SDarrick J. Wong return -EFSCORRUPTED; 948273df556SFrank Mayhar } 949d4620315SRobin Dong 950d4620315SRobin Dong if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries) 951d4620315SRobin Dong >= le16_to_cpu(curp->p_hdr->eh_max))) { 952d4620315SRobin Dong EXT4_ERROR_INODE(inode, 953d4620315SRobin Dong "eh_entries %d >= eh_max %d!", 954d4620315SRobin Dong le16_to_cpu(curp->p_hdr->eh_entries), 955d4620315SRobin Dong le16_to_cpu(curp->p_hdr->eh_max)); 9566a797d27SDarrick J. Wong return -EFSCORRUPTED; 957d4620315SRobin Dong } 958d4620315SRobin Dong 959a86c6181SAlex Tomas if (logical > le32_to_cpu(curp->p_idx->ei_block)) { 960a86c6181SAlex Tomas /* insert after */ 96180e675f9SEric Gouriou ext_debug("insert new index %d after: %llu\n", logical, ptr); 962a86c6181SAlex Tomas ix = curp->p_idx + 1; 963a86c6181SAlex Tomas } else { 964a86c6181SAlex Tomas /* insert before */ 96580e675f9SEric Gouriou ext_debug("insert new index %d before: %llu\n", logical, ptr); 966a86c6181SAlex Tomas ix = curp->p_idx; 967a86c6181SAlex Tomas } 968a86c6181SAlex Tomas 96980e675f9SEric Gouriou len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1; 97080e675f9SEric Gouriou BUG_ON(len < 0); 97180e675f9SEric Gouriou if (len > 0) { 97280e675f9SEric Gouriou ext_debug("insert new index %d: " 97380e675f9SEric Gouriou "move %d indices from 0x%p to 0x%p\n", 97480e675f9SEric Gouriou logical, len, ix, ix + 1); 97580e675f9SEric Gouriou memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx)); 97680e675f9SEric Gouriou } 97780e675f9SEric Gouriou 978f472e026STao Ma if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) { 979f472e026STao Ma EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!"); 9806a797d27SDarrick J. Wong return -EFSCORRUPTED; 981f472e026STao Ma } 982f472e026STao Ma 983a86c6181SAlex Tomas ix->ei_block = cpu_to_le32(logical); 984f65e6fbaSAlex Tomas ext4_idx_store_pblock(ix, ptr); 985e8546d06SMarcin Slusarz le16_add_cpu(&curp->p_hdr->eh_entries, 1); 986a86c6181SAlex Tomas 987273df556SFrank Mayhar if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) { 988273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!"); 9896a797d27SDarrick J. Wong return -EFSCORRUPTED; 990273df556SFrank Mayhar } 991a86c6181SAlex Tomas 992a86c6181SAlex Tomas err = ext4_ext_dirty(handle, inode, curp); 993a86c6181SAlex Tomas ext4_std_error(inode->i_sb, err); 994a86c6181SAlex Tomas 995a86c6181SAlex Tomas return err; 996a86c6181SAlex Tomas } 997a86c6181SAlex Tomas 998a86c6181SAlex Tomas /* 999d0d856e8SRandy Dunlap * ext4_ext_split: 1000d0d856e8SRandy Dunlap * inserts new subtree into the path, using free index entry 1001d0d856e8SRandy Dunlap * at depth @at: 1002a86c6181SAlex Tomas * - allocates all needed blocks (new leaf and all intermediate index blocks) 1003a86c6181SAlex Tomas * - makes decision where to split 1004d0d856e8SRandy Dunlap * - moves remaining extents and index entries (right to the split point) 1005a86c6181SAlex Tomas * into the newly allocated blocks 1006d0d856e8SRandy Dunlap * - initializes subtree 1007a86c6181SAlex Tomas */ 1008a86c6181SAlex Tomas static int ext4_ext_split(handle_t *handle, struct inode *inode, 100955f020dbSAllison Henderson unsigned int flags, 1010a86c6181SAlex Tomas struct ext4_ext_path *path, 1011a86c6181SAlex Tomas struct ext4_extent *newext, int at) 1012a86c6181SAlex Tomas { 1013a86c6181SAlex Tomas struct buffer_head *bh = NULL; 1014a86c6181SAlex Tomas int depth = ext_depth(inode); 1015a86c6181SAlex Tomas struct ext4_extent_header *neh; 1016a86c6181SAlex Tomas struct ext4_extent_idx *fidx; 1017a86c6181SAlex Tomas int i = at, k, m, a; 1018f65e6fbaSAlex Tomas ext4_fsblk_t newblock, oldblock; 1019a86c6181SAlex Tomas __le32 border; 1020f65e6fbaSAlex Tomas ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */ 102173c384c0STheodore Ts'o gfp_t gfp_flags = GFP_NOFS; 1022a86c6181SAlex Tomas int err = 0; 1023592acbf1SSriram Rajagopalan size_t ext_size = 0; 1024a86c6181SAlex Tomas 102573c384c0STheodore Ts'o if (flags & EXT4_EX_NOFAIL) 102673c384c0STheodore Ts'o gfp_flags |= __GFP_NOFAIL; 102773c384c0STheodore Ts'o 1028a86c6181SAlex Tomas /* make decision: where to split? */ 1029d0d856e8SRandy Dunlap /* FIXME: now decision is simplest: at current extent */ 1030a86c6181SAlex Tomas 1031d0d856e8SRandy Dunlap /* if current leaf will be split, then we should use 1032a86c6181SAlex Tomas * border from split point */ 1033273df556SFrank Mayhar if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) { 1034273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!"); 10356a797d27SDarrick J. Wong return -EFSCORRUPTED; 1036273df556SFrank Mayhar } 1037a86c6181SAlex Tomas if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) { 1038a86c6181SAlex Tomas border = path[depth].p_ext[1].ee_block; 1039d0d856e8SRandy Dunlap ext_debug("leaf will be split." 1040a86c6181SAlex Tomas " next leaf starts at %d\n", 1041a86c6181SAlex Tomas le32_to_cpu(border)); 1042a86c6181SAlex Tomas } else { 1043a86c6181SAlex Tomas border = newext->ee_block; 1044a86c6181SAlex Tomas ext_debug("leaf will be added." 1045a86c6181SAlex Tomas " next leaf starts at %d\n", 1046a86c6181SAlex Tomas le32_to_cpu(border)); 1047a86c6181SAlex Tomas } 1048a86c6181SAlex Tomas 1049a86c6181SAlex Tomas /* 1050d0d856e8SRandy Dunlap * If error occurs, then we break processing 1051d0d856e8SRandy Dunlap * and mark filesystem read-only. index won't 1052a86c6181SAlex Tomas * be inserted and tree will be in consistent 1053d0d856e8SRandy Dunlap * state. Next mount will repair buffers too. 1054a86c6181SAlex Tomas */ 1055a86c6181SAlex Tomas 1056a86c6181SAlex Tomas /* 1057d0d856e8SRandy Dunlap * Get array to track all allocated blocks. 1058d0d856e8SRandy Dunlap * We need this to handle errors and free blocks 1059d0d856e8SRandy Dunlap * upon them. 1060a86c6181SAlex Tomas */ 106173c384c0STheodore Ts'o ablocks = kcalloc(depth, sizeof(ext4_fsblk_t), gfp_flags); 1062a86c6181SAlex Tomas if (!ablocks) 1063a86c6181SAlex Tomas return -ENOMEM; 1064a86c6181SAlex Tomas 1065a86c6181SAlex Tomas /* allocate all needed blocks */ 1066a86c6181SAlex Tomas ext_debug("allocate %d blocks for indexes/leaf\n", depth - at); 1067a86c6181SAlex Tomas for (a = 0; a < depth - at; a++) { 1068654b4908SAneesh Kumar K.V newblock = ext4_ext_new_meta_block(handle, inode, path, 106955f020dbSAllison Henderson newext, &err, flags); 1070a86c6181SAlex Tomas if (newblock == 0) 1071a86c6181SAlex Tomas goto cleanup; 1072a86c6181SAlex Tomas ablocks[a] = newblock; 1073a86c6181SAlex Tomas } 1074a86c6181SAlex Tomas 1075a86c6181SAlex Tomas /* initialize new leaf */ 1076a86c6181SAlex Tomas newblock = ablocks[--a]; 1077273df556SFrank Mayhar if (unlikely(newblock == 0)) { 1078273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "newblock == 0!"); 10796a797d27SDarrick J. Wong err = -EFSCORRUPTED; 1080273df556SFrank Mayhar goto cleanup; 1081273df556SFrank Mayhar } 1082c45653c3SNikolay Borisov bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS); 1083aebf0243SWang Shilong if (unlikely(!bh)) { 1084860d21e2STheodore Ts'o err = -ENOMEM; 1085a86c6181SAlex Tomas goto cleanup; 1086a86c6181SAlex Tomas } 1087a86c6181SAlex Tomas lock_buffer(bh); 1088a86c6181SAlex Tomas 10897e028976SAvantika Mathur err = ext4_journal_get_create_access(handle, bh); 10907e028976SAvantika Mathur if (err) 1091a86c6181SAlex Tomas goto cleanup; 1092a86c6181SAlex Tomas 1093a86c6181SAlex Tomas neh = ext_block_hdr(bh); 1094a86c6181SAlex Tomas neh->eh_entries = 0; 109555ad63bfSTheodore Ts'o neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0)); 1096a86c6181SAlex Tomas neh->eh_magic = EXT4_EXT_MAGIC; 1097a86c6181SAlex Tomas neh->eh_depth = 0; 1098a86c6181SAlex Tomas 1099d0d856e8SRandy Dunlap /* move remainder of path[depth] to the new leaf */ 1100273df556SFrank Mayhar if (unlikely(path[depth].p_hdr->eh_entries != 1101273df556SFrank Mayhar path[depth].p_hdr->eh_max)) { 1102273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!", 1103273df556SFrank Mayhar path[depth].p_hdr->eh_entries, 1104273df556SFrank Mayhar path[depth].p_hdr->eh_max); 11056a797d27SDarrick J. Wong err = -EFSCORRUPTED; 1106273df556SFrank Mayhar goto cleanup; 1107273df556SFrank Mayhar } 1108a86c6181SAlex Tomas /* start copy from next extent */ 11091b16da77SYongqiang Yang m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++; 11101b16da77SYongqiang Yang ext4_ext_show_move(inode, path, newblock, depth); 1111a86c6181SAlex Tomas if (m) { 11121b16da77SYongqiang Yang struct ext4_extent *ex; 11131b16da77SYongqiang Yang ex = EXT_FIRST_EXTENT(neh); 11141b16da77SYongqiang Yang memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m); 1115e8546d06SMarcin Slusarz le16_add_cpu(&neh->eh_entries, m); 1116a86c6181SAlex Tomas } 1117a86c6181SAlex Tomas 1118592acbf1SSriram Rajagopalan /* zero out unused area in the extent block */ 1119592acbf1SSriram Rajagopalan ext_size = sizeof(struct ext4_extent_header) + 1120592acbf1SSriram Rajagopalan sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries); 1121592acbf1SSriram Rajagopalan memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size); 11227ac5990dSDarrick J. Wong ext4_extent_block_csum_set(inode, neh); 1123a86c6181SAlex Tomas set_buffer_uptodate(bh); 1124a86c6181SAlex Tomas unlock_buffer(bh); 1125a86c6181SAlex Tomas 11260390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, inode, bh); 11277e028976SAvantika Mathur if (err) 1128a86c6181SAlex Tomas goto cleanup; 1129a86c6181SAlex Tomas brelse(bh); 1130a86c6181SAlex Tomas bh = NULL; 1131a86c6181SAlex Tomas 1132a86c6181SAlex Tomas /* correct old leaf */ 1133a86c6181SAlex Tomas if (m) { 11347e028976SAvantika Mathur err = ext4_ext_get_access(handle, inode, path + depth); 11357e028976SAvantika Mathur if (err) 1136a86c6181SAlex Tomas goto cleanup; 1137e8546d06SMarcin Slusarz le16_add_cpu(&path[depth].p_hdr->eh_entries, -m); 11387e028976SAvantika Mathur err = ext4_ext_dirty(handle, inode, path + depth); 11397e028976SAvantika Mathur if (err) 1140a86c6181SAlex Tomas goto cleanup; 1141a86c6181SAlex Tomas 1142a86c6181SAlex Tomas } 1143a86c6181SAlex Tomas 1144a86c6181SAlex Tomas /* create intermediate indexes */ 1145a86c6181SAlex Tomas k = depth - at - 1; 1146273df556SFrank Mayhar if (unlikely(k < 0)) { 1147273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "k %d < 0!", k); 11486a797d27SDarrick J. Wong err = -EFSCORRUPTED; 1149273df556SFrank Mayhar goto cleanup; 1150273df556SFrank Mayhar } 1151a86c6181SAlex Tomas if (k) 1152a86c6181SAlex Tomas ext_debug("create %d intermediate indices\n", k); 1153a86c6181SAlex Tomas /* insert new index into current index block */ 1154a86c6181SAlex Tomas /* current depth stored in i var */ 1155a86c6181SAlex Tomas i = depth - 1; 1156a86c6181SAlex Tomas while (k--) { 1157a86c6181SAlex Tomas oldblock = newblock; 1158a86c6181SAlex Tomas newblock = ablocks[--a]; 1159bba90743SEric Sandeen bh = sb_getblk(inode->i_sb, newblock); 1160aebf0243SWang Shilong if (unlikely(!bh)) { 1161860d21e2STheodore Ts'o err = -ENOMEM; 1162a86c6181SAlex Tomas goto cleanup; 1163a86c6181SAlex Tomas } 1164a86c6181SAlex Tomas lock_buffer(bh); 1165a86c6181SAlex Tomas 11667e028976SAvantika Mathur err = ext4_journal_get_create_access(handle, bh); 11677e028976SAvantika Mathur if (err) 1168a86c6181SAlex Tomas goto cleanup; 1169a86c6181SAlex Tomas 1170a86c6181SAlex Tomas neh = ext_block_hdr(bh); 1171a86c6181SAlex Tomas neh->eh_entries = cpu_to_le16(1); 1172a86c6181SAlex Tomas neh->eh_magic = EXT4_EXT_MAGIC; 117355ad63bfSTheodore Ts'o neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0)); 1174a86c6181SAlex Tomas neh->eh_depth = cpu_to_le16(depth - i); 1175a86c6181SAlex Tomas fidx = EXT_FIRST_INDEX(neh); 1176a86c6181SAlex Tomas fidx->ei_block = border; 1177f65e6fbaSAlex Tomas ext4_idx_store_pblock(fidx, oldblock); 1178a86c6181SAlex Tomas 1179bba90743SEric Sandeen ext_debug("int.index at %d (block %llu): %u -> %llu\n", 1180bba90743SEric Sandeen i, newblock, le32_to_cpu(border), oldblock); 1181a86c6181SAlex Tomas 11821b16da77SYongqiang Yang /* move remainder of path[i] to the new index block */ 1183273df556SFrank Mayhar if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) != 1184273df556SFrank Mayhar EXT_LAST_INDEX(path[i].p_hdr))) { 1185273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1186273df556SFrank Mayhar "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!", 1187273df556SFrank Mayhar le32_to_cpu(path[i].p_ext->ee_block)); 11886a797d27SDarrick J. Wong err = -EFSCORRUPTED; 1189273df556SFrank Mayhar goto cleanup; 1190273df556SFrank Mayhar } 11911b16da77SYongqiang Yang /* start copy indexes */ 11921b16da77SYongqiang Yang m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++; 11931b16da77SYongqiang Yang ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx, 11941b16da77SYongqiang Yang EXT_MAX_INDEX(path[i].p_hdr)); 11951b16da77SYongqiang Yang ext4_ext_show_move(inode, path, newblock, i); 1196a86c6181SAlex Tomas if (m) { 11971b16da77SYongqiang Yang memmove(++fidx, path[i].p_idx, 1198a86c6181SAlex Tomas sizeof(struct ext4_extent_idx) * m); 1199e8546d06SMarcin Slusarz le16_add_cpu(&neh->eh_entries, m); 1200a86c6181SAlex Tomas } 1201592acbf1SSriram Rajagopalan /* zero out unused area in the extent block */ 1202592acbf1SSriram Rajagopalan ext_size = sizeof(struct ext4_extent_header) + 1203592acbf1SSriram Rajagopalan (sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries)); 1204592acbf1SSriram Rajagopalan memset(bh->b_data + ext_size, 0, 1205592acbf1SSriram Rajagopalan inode->i_sb->s_blocksize - ext_size); 12067ac5990dSDarrick J. Wong ext4_extent_block_csum_set(inode, neh); 1207a86c6181SAlex Tomas set_buffer_uptodate(bh); 1208a86c6181SAlex Tomas unlock_buffer(bh); 1209a86c6181SAlex Tomas 12100390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, inode, bh); 12117e028976SAvantika Mathur if (err) 1212a86c6181SAlex Tomas goto cleanup; 1213a86c6181SAlex Tomas brelse(bh); 1214a86c6181SAlex Tomas bh = NULL; 1215a86c6181SAlex Tomas 1216a86c6181SAlex Tomas /* correct old index */ 1217a86c6181SAlex Tomas if (m) { 1218a86c6181SAlex Tomas err = ext4_ext_get_access(handle, inode, path + i); 1219a86c6181SAlex Tomas if (err) 1220a86c6181SAlex Tomas goto cleanup; 1221e8546d06SMarcin Slusarz le16_add_cpu(&path[i].p_hdr->eh_entries, -m); 1222a86c6181SAlex Tomas err = ext4_ext_dirty(handle, inode, path + i); 1223a86c6181SAlex Tomas if (err) 1224a86c6181SAlex Tomas goto cleanup; 1225a86c6181SAlex Tomas } 1226a86c6181SAlex Tomas 1227a86c6181SAlex Tomas i--; 1228a86c6181SAlex Tomas } 1229a86c6181SAlex Tomas 1230a86c6181SAlex Tomas /* insert new index */ 1231a86c6181SAlex Tomas err = ext4_ext_insert_index(handle, inode, path + at, 1232a86c6181SAlex Tomas le32_to_cpu(border), newblock); 1233a86c6181SAlex Tomas 1234a86c6181SAlex Tomas cleanup: 1235a86c6181SAlex Tomas if (bh) { 1236a86c6181SAlex Tomas if (buffer_locked(bh)) 1237a86c6181SAlex Tomas unlock_buffer(bh); 1238a86c6181SAlex Tomas brelse(bh); 1239a86c6181SAlex Tomas } 1240a86c6181SAlex Tomas 1241a86c6181SAlex Tomas if (err) { 1242a86c6181SAlex Tomas /* free all allocated blocks in error case */ 1243a86c6181SAlex Tomas for (i = 0; i < depth; i++) { 1244a86c6181SAlex Tomas if (!ablocks[i]) 1245a86c6181SAlex Tomas continue; 12467dc57615SPeter Huewe ext4_free_blocks(handle, inode, NULL, ablocks[i], 1, 1247e6362609STheodore Ts'o EXT4_FREE_BLOCKS_METADATA); 1248a86c6181SAlex Tomas } 1249a86c6181SAlex Tomas } 1250a86c6181SAlex Tomas kfree(ablocks); 1251a86c6181SAlex Tomas 1252a86c6181SAlex Tomas return err; 1253a86c6181SAlex Tomas } 1254a86c6181SAlex Tomas 1255a86c6181SAlex Tomas /* 1256d0d856e8SRandy Dunlap * ext4_ext_grow_indepth: 1257d0d856e8SRandy Dunlap * implements tree growing procedure: 1258a86c6181SAlex Tomas * - allocates new block 1259a86c6181SAlex Tomas * - moves top-level data (index block or leaf) into the new block 1260d0d856e8SRandy Dunlap * - initializes new top-level, creating index that points to the 1261a86c6181SAlex Tomas * just created block 1262a86c6181SAlex Tomas */ 1263a86c6181SAlex Tomas static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode, 1264be5cd90dSDmitry Monakhov unsigned int flags) 1265a86c6181SAlex Tomas { 1266a86c6181SAlex Tomas struct ext4_extent_header *neh; 1267a86c6181SAlex Tomas struct buffer_head *bh; 1268be5cd90dSDmitry Monakhov ext4_fsblk_t newblock, goal = 0; 1269be5cd90dSDmitry Monakhov struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es; 1270a86c6181SAlex Tomas int err = 0; 1271592acbf1SSriram Rajagopalan size_t ext_size = 0; 1272a86c6181SAlex Tomas 1273be5cd90dSDmitry Monakhov /* Try to prepend new index to old one */ 1274be5cd90dSDmitry Monakhov if (ext_depth(inode)) 1275be5cd90dSDmitry Monakhov goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode))); 1276be5cd90dSDmitry Monakhov if (goal > le32_to_cpu(es->s_first_data_block)) { 1277be5cd90dSDmitry Monakhov flags |= EXT4_MB_HINT_TRY_GOAL; 1278be5cd90dSDmitry Monakhov goal--; 1279be5cd90dSDmitry Monakhov } else 1280be5cd90dSDmitry Monakhov goal = ext4_inode_to_goal_block(inode); 1281be5cd90dSDmitry Monakhov newblock = ext4_new_meta_blocks(handle, inode, goal, flags, 1282be5cd90dSDmitry Monakhov NULL, &err); 1283a86c6181SAlex Tomas if (newblock == 0) 1284a86c6181SAlex Tomas return err; 1285a86c6181SAlex Tomas 1286c45653c3SNikolay Borisov bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS); 1287aebf0243SWang Shilong if (unlikely(!bh)) 1288860d21e2STheodore Ts'o return -ENOMEM; 1289a86c6181SAlex Tomas lock_buffer(bh); 1290a86c6181SAlex Tomas 12917e028976SAvantika Mathur err = ext4_journal_get_create_access(handle, bh); 12927e028976SAvantika Mathur if (err) { 1293a86c6181SAlex Tomas unlock_buffer(bh); 1294a86c6181SAlex Tomas goto out; 1295a86c6181SAlex Tomas } 1296a86c6181SAlex Tomas 1297592acbf1SSriram Rajagopalan ext_size = sizeof(EXT4_I(inode)->i_data); 1298a86c6181SAlex Tomas /* move top-level index/leaf into new block */ 1299592acbf1SSriram Rajagopalan memmove(bh->b_data, EXT4_I(inode)->i_data, ext_size); 1300592acbf1SSriram Rajagopalan /* zero out unused area in the extent block */ 1301592acbf1SSriram Rajagopalan memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size); 1302a86c6181SAlex Tomas 1303a86c6181SAlex Tomas /* set size of new block */ 1304a86c6181SAlex Tomas neh = ext_block_hdr(bh); 1305a86c6181SAlex Tomas /* old root could have indexes or leaves 1306a86c6181SAlex Tomas * so calculate e_max right way */ 1307a86c6181SAlex Tomas if (ext_depth(inode)) 130855ad63bfSTheodore Ts'o neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0)); 1309a86c6181SAlex Tomas else 131055ad63bfSTheodore Ts'o neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0)); 1311a86c6181SAlex Tomas neh->eh_magic = EXT4_EXT_MAGIC; 13127ac5990dSDarrick J. Wong ext4_extent_block_csum_set(inode, neh); 1313a86c6181SAlex Tomas set_buffer_uptodate(bh); 1314a86c6181SAlex Tomas unlock_buffer(bh); 1315a86c6181SAlex Tomas 13160390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, inode, bh); 13177e028976SAvantika Mathur if (err) 1318a86c6181SAlex Tomas goto out; 1319a86c6181SAlex Tomas 13201939dd84SDmitry Monakhov /* Update top-level index: num,max,pointer */ 1321a86c6181SAlex Tomas neh = ext_inode_hdr(inode); 13221939dd84SDmitry Monakhov neh->eh_entries = cpu_to_le16(1); 13231939dd84SDmitry Monakhov ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock); 13241939dd84SDmitry Monakhov if (neh->eh_depth == 0) { 13251939dd84SDmitry Monakhov /* Root extent block becomes index block */ 13261939dd84SDmitry Monakhov neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0)); 13271939dd84SDmitry Monakhov EXT_FIRST_INDEX(neh)->ei_block = 13281939dd84SDmitry Monakhov EXT_FIRST_EXTENT(neh)->ee_block; 13291939dd84SDmitry Monakhov } 13302ae02107SMingming Cao ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n", 1331a86c6181SAlex Tomas le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max), 13325a0790c2SAndi Kleen le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block), 1333bf89d16fSTheodore Ts'o ext4_idx_pblock(EXT_FIRST_INDEX(neh))); 1334a86c6181SAlex Tomas 1335ba39ebb6SWei Yongjun le16_add_cpu(&neh->eh_depth, 1); 13364209ae12SHarshad Shirwadkar err = ext4_mark_inode_dirty(handle, inode); 1337a86c6181SAlex Tomas out: 1338a86c6181SAlex Tomas brelse(bh); 1339a86c6181SAlex Tomas 1340a86c6181SAlex Tomas return err; 1341a86c6181SAlex Tomas } 1342a86c6181SAlex Tomas 1343a86c6181SAlex Tomas /* 1344d0d856e8SRandy Dunlap * ext4_ext_create_new_leaf: 1345d0d856e8SRandy Dunlap * finds empty index and adds new leaf. 1346d0d856e8SRandy Dunlap * if no free index is found, then it requests in-depth growing. 1347a86c6181SAlex Tomas */ 1348a86c6181SAlex Tomas static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode, 1349107a7bd3STheodore Ts'o unsigned int mb_flags, 1350107a7bd3STheodore Ts'o unsigned int gb_flags, 1351dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 1352a86c6181SAlex Tomas struct ext4_extent *newext) 1353a86c6181SAlex Tomas { 1354dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 1355a86c6181SAlex Tomas struct ext4_ext_path *curp; 1356a86c6181SAlex Tomas int depth, i, err = 0; 1357a86c6181SAlex Tomas 1358a86c6181SAlex Tomas repeat: 1359a86c6181SAlex Tomas i = depth = ext_depth(inode); 1360a86c6181SAlex Tomas 1361a86c6181SAlex Tomas /* walk up to the tree and look for free index entry */ 1362a86c6181SAlex Tomas curp = path + depth; 1363a86c6181SAlex Tomas while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) { 1364a86c6181SAlex Tomas i--; 1365a86c6181SAlex Tomas curp--; 1366a86c6181SAlex Tomas } 1367a86c6181SAlex Tomas 1368d0d856e8SRandy Dunlap /* we use already allocated block for index block, 1369d0d856e8SRandy Dunlap * so subsequent data blocks should be contiguous */ 1370a86c6181SAlex Tomas if (EXT_HAS_FREE_INDEX(curp)) { 1371a86c6181SAlex Tomas /* if we found index with free entry, then use that 1372a86c6181SAlex Tomas * entry: create all needed subtree and add new leaf */ 1373107a7bd3STheodore Ts'o err = ext4_ext_split(handle, inode, mb_flags, path, newext, i); 1374787e0981SShen Feng if (err) 1375787e0981SShen Feng goto out; 1376a86c6181SAlex Tomas 1377a86c6181SAlex Tomas /* refill path */ 1378ed8a1a76STheodore Ts'o path = ext4_find_extent(inode, 1379725d26d3SAneesh Kumar K.V (ext4_lblk_t)le32_to_cpu(newext->ee_block), 1380dfe50809STheodore Ts'o ppath, gb_flags); 1381a86c6181SAlex Tomas if (IS_ERR(path)) 1382a86c6181SAlex Tomas err = PTR_ERR(path); 1383a86c6181SAlex Tomas } else { 1384a86c6181SAlex Tomas /* tree is full, time to grow in depth */ 1385be5cd90dSDmitry Monakhov err = ext4_ext_grow_indepth(handle, inode, mb_flags); 1386a86c6181SAlex Tomas if (err) 1387a86c6181SAlex Tomas goto out; 1388a86c6181SAlex Tomas 1389a86c6181SAlex Tomas /* refill path */ 1390ed8a1a76STheodore Ts'o path = ext4_find_extent(inode, 1391725d26d3SAneesh Kumar K.V (ext4_lblk_t)le32_to_cpu(newext->ee_block), 1392dfe50809STheodore Ts'o ppath, gb_flags); 1393a86c6181SAlex Tomas if (IS_ERR(path)) { 1394a86c6181SAlex Tomas err = PTR_ERR(path); 1395a86c6181SAlex Tomas goto out; 1396a86c6181SAlex Tomas } 1397a86c6181SAlex Tomas 1398a86c6181SAlex Tomas /* 1399d0d856e8SRandy Dunlap * only first (depth 0 -> 1) produces free space; 1400d0d856e8SRandy Dunlap * in all other cases we have to split the grown tree 1401a86c6181SAlex Tomas */ 1402a86c6181SAlex Tomas depth = ext_depth(inode); 1403a86c6181SAlex Tomas if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) { 1404d0d856e8SRandy Dunlap /* now we need to split */ 1405a86c6181SAlex Tomas goto repeat; 1406a86c6181SAlex Tomas } 1407a86c6181SAlex Tomas } 1408a86c6181SAlex Tomas 1409a86c6181SAlex Tomas out: 1410a86c6181SAlex Tomas return err; 1411a86c6181SAlex Tomas } 1412a86c6181SAlex Tomas 1413a86c6181SAlex Tomas /* 14141988b51eSAlex Tomas * search the closest allocated block to the left for *logical 14151988b51eSAlex Tomas * and returns it at @logical + it's physical address at @phys 14161988b51eSAlex Tomas * if *logical is the smallest allocated block, the function 14171988b51eSAlex Tomas * returns 0 at @phys 14181988b51eSAlex Tomas * return value contains 0 (success) or error code 14191988b51eSAlex Tomas */ 14201f109d5aSTheodore Ts'o static int ext4_ext_search_left(struct inode *inode, 14211f109d5aSTheodore Ts'o struct ext4_ext_path *path, 14221988b51eSAlex Tomas ext4_lblk_t *logical, ext4_fsblk_t *phys) 14231988b51eSAlex Tomas { 14241988b51eSAlex Tomas struct ext4_extent_idx *ix; 14251988b51eSAlex Tomas struct ext4_extent *ex; 1426b939e376SAneesh Kumar K.V int depth, ee_len; 14271988b51eSAlex Tomas 1428273df556SFrank Mayhar if (unlikely(path == NULL)) { 1429273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical); 14306a797d27SDarrick J. Wong return -EFSCORRUPTED; 1431273df556SFrank Mayhar } 14321988b51eSAlex Tomas depth = path->p_depth; 14331988b51eSAlex Tomas *phys = 0; 14341988b51eSAlex Tomas 14351988b51eSAlex Tomas if (depth == 0 && path->p_ext == NULL) 14361988b51eSAlex Tomas return 0; 14371988b51eSAlex Tomas 14381988b51eSAlex Tomas /* usually extent in the path covers blocks smaller 14391988b51eSAlex Tomas * then *logical, but it can be that extent is the 14401988b51eSAlex Tomas * first one in the file */ 14411988b51eSAlex Tomas 14421988b51eSAlex Tomas ex = path[depth].p_ext; 1443b939e376SAneesh Kumar K.V ee_len = ext4_ext_get_actual_len(ex); 14441988b51eSAlex Tomas if (*logical < le32_to_cpu(ex->ee_block)) { 1445273df556SFrank Mayhar if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) { 1446273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1447273df556SFrank Mayhar "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!", 1448273df556SFrank Mayhar *logical, le32_to_cpu(ex->ee_block)); 14496a797d27SDarrick J. Wong return -EFSCORRUPTED; 1450273df556SFrank Mayhar } 14511988b51eSAlex Tomas while (--depth >= 0) { 14521988b51eSAlex Tomas ix = path[depth].p_idx; 1453273df556SFrank Mayhar if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) { 1454273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1455273df556SFrank Mayhar "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!", 14566ee3b212STao Ma ix != NULL ? le32_to_cpu(ix->ei_block) : 0, 1457273df556SFrank Mayhar EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ? 14586ee3b212STao Ma le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0, 1459273df556SFrank Mayhar depth); 14606a797d27SDarrick J. Wong return -EFSCORRUPTED; 1461273df556SFrank Mayhar } 14621988b51eSAlex Tomas } 14631988b51eSAlex Tomas return 0; 14641988b51eSAlex Tomas } 14651988b51eSAlex Tomas 1466273df556SFrank Mayhar if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) { 1467273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1468273df556SFrank Mayhar "logical %d < ee_block %d + ee_len %d!", 1469273df556SFrank Mayhar *logical, le32_to_cpu(ex->ee_block), ee_len); 14706a797d27SDarrick J. Wong return -EFSCORRUPTED; 1471273df556SFrank Mayhar } 14721988b51eSAlex Tomas 1473b939e376SAneesh Kumar K.V *logical = le32_to_cpu(ex->ee_block) + ee_len - 1; 1474bf89d16fSTheodore Ts'o *phys = ext4_ext_pblock(ex) + ee_len - 1; 14751988b51eSAlex Tomas return 0; 14761988b51eSAlex Tomas } 14771988b51eSAlex Tomas 14781988b51eSAlex Tomas /* 14791988b51eSAlex Tomas * search the closest allocated block to the right for *logical 14801988b51eSAlex Tomas * and returns it at @logical + it's physical address at @phys 1481df3ab170STao Ma * if *logical is the largest allocated block, the function 14821988b51eSAlex Tomas * returns 0 at @phys 14831988b51eSAlex Tomas * return value contains 0 (success) or error code 14841988b51eSAlex Tomas */ 14851f109d5aSTheodore Ts'o static int ext4_ext_search_right(struct inode *inode, 14861f109d5aSTheodore Ts'o struct ext4_ext_path *path, 14874d33b1efSTheodore Ts'o ext4_lblk_t *logical, ext4_fsblk_t *phys, 14884d33b1efSTheodore Ts'o struct ext4_extent **ret_ex) 14891988b51eSAlex Tomas { 14901988b51eSAlex Tomas struct buffer_head *bh = NULL; 14911988b51eSAlex Tomas struct ext4_extent_header *eh; 14921988b51eSAlex Tomas struct ext4_extent_idx *ix; 14931988b51eSAlex Tomas struct ext4_extent *ex; 14941988b51eSAlex Tomas ext4_fsblk_t block; 1495395a87bfSEric Sandeen int depth; /* Note, NOT eh_depth; depth from top of tree */ 1496395a87bfSEric Sandeen int ee_len; 14971988b51eSAlex Tomas 1498273df556SFrank Mayhar if (unlikely(path == NULL)) { 1499273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical); 15006a797d27SDarrick J. Wong return -EFSCORRUPTED; 1501273df556SFrank Mayhar } 15021988b51eSAlex Tomas depth = path->p_depth; 15031988b51eSAlex Tomas *phys = 0; 15041988b51eSAlex Tomas 15051988b51eSAlex Tomas if (depth == 0 && path->p_ext == NULL) 15061988b51eSAlex Tomas return 0; 15071988b51eSAlex Tomas 15081988b51eSAlex Tomas /* usually extent in the path covers blocks smaller 15091988b51eSAlex Tomas * then *logical, but it can be that extent is the 15101988b51eSAlex Tomas * first one in the file */ 15111988b51eSAlex Tomas 15121988b51eSAlex Tomas ex = path[depth].p_ext; 1513b939e376SAneesh Kumar K.V ee_len = ext4_ext_get_actual_len(ex); 15141988b51eSAlex Tomas if (*logical < le32_to_cpu(ex->ee_block)) { 1515273df556SFrank Mayhar if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) { 1516273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1517273df556SFrank Mayhar "first_extent(path[%d].p_hdr) != ex", 1518273df556SFrank Mayhar depth); 15196a797d27SDarrick J. Wong return -EFSCORRUPTED; 1520273df556SFrank Mayhar } 15211988b51eSAlex Tomas while (--depth >= 0) { 15221988b51eSAlex Tomas ix = path[depth].p_idx; 1523273df556SFrank Mayhar if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) { 1524273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1525273df556SFrank Mayhar "ix != EXT_FIRST_INDEX *logical %d!", 1526273df556SFrank Mayhar *logical); 15276a797d27SDarrick J. Wong return -EFSCORRUPTED; 1528273df556SFrank Mayhar } 15291988b51eSAlex Tomas } 15304d33b1efSTheodore Ts'o goto found_extent; 15311988b51eSAlex Tomas } 15321988b51eSAlex Tomas 1533273df556SFrank Mayhar if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) { 1534273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1535273df556SFrank Mayhar "logical %d < ee_block %d + ee_len %d!", 1536273df556SFrank Mayhar *logical, le32_to_cpu(ex->ee_block), ee_len); 15376a797d27SDarrick J. Wong return -EFSCORRUPTED; 1538273df556SFrank Mayhar } 15391988b51eSAlex Tomas 15401988b51eSAlex Tomas if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) { 15411988b51eSAlex Tomas /* next allocated block in this leaf */ 15421988b51eSAlex Tomas ex++; 15434d33b1efSTheodore Ts'o goto found_extent; 15441988b51eSAlex Tomas } 15451988b51eSAlex Tomas 15461988b51eSAlex Tomas /* go up and search for index to the right */ 15471988b51eSAlex Tomas while (--depth >= 0) { 15481988b51eSAlex Tomas ix = path[depth].p_idx; 15491988b51eSAlex Tomas if (ix != EXT_LAST_INDEX(path[depth].p_hdr)) 155025f1ee3aSWu Fengguang goto got_index; 15511988b51eSAlex Tomas } 15521988b51eSAlex Tomas 155325f1ee3aSWu Fengguang /* we've gone up to the root and found no index to the right */ 15541988b51eSAlex Tomas return 0; 15551988b51eSAlex Tomas 155625f1ee3aSWu Fengguang got_index: 15571988b51eSAlex Tomas /* we've found index to the right, let's 15581988b51eSAlex Tomas * follow it and find the closest allocated 15591988b51eSAlex Tomas * block to the right */ 15601988b51eSAlex Tomas ix++; 1561bf89d16fSTheodore Ts'o block = ext4_idx_pblock(ix); 15621988b51eSAlex Tomas while (++depth < path->p_depth) { 1563395a87bfSEric Sandeen /* subtract from p_depth to get proper eh_depth */ 15647d7ea89eSTheodore Ts'o bh = read_extent_tree_block(inode, block, 1565107a7bd3STheodore Ts'o path->p_depth - depth, 0); 15667d7ea89eSTheodore Ts'o if (IS_ERR(bh)) 15677d7ea89eSTheodore Ts'o return PTR_ERR(bh); 15687d7ea89eSTheodore Ts'o eh = ext_block_hdr(bh); 15691988b51eSAlex Tomas ix = EXT_FIRST_INDEX(eh); 1570bf89d16fSTheodore Ts'o block = ext4_idx_pblock(ix); 15711988b51eSAlex Tomas put_bh(bh); 15721988b51eSAlex Tomas } 15731988b51eSAlex Tomas 1574107a7bd3STheodore Ts'o bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0); 15757d7ea89eSTheodore Ts'o if (IS_ERR(bh)) 15767d7ea89eSTheodore Ts'o return PTR_ERR(bh); 15771988b51eSAlex Tomas eh = ext_block_hdr(bh); 15781988b51eSAlex Tomas ex = EXT_FIRST_EXTENT(eh); 15794d33b1efSTheodore Ts'o found_extent: 15801988b51eSAlex Tomas *logical = le32_to_cpu(ex->ee_block); 1581bf89d16fSTheodore Ts'o *phys = ext4_ext_pblock(ex); 15824d33b1efSTheodore Ts'o *ret_ex = ex; 15834d33b1efSTheodore Ts'o if (bh) 15841988b51eSAlex Tomas put_bh(bh); 15851988b51eSAlex Tomas return 0; 15861988b51eSAlex Tomas } 15871988b51eSAlex Tomas 15881988b51eSAlex Tomas /* 1589d0d856e8SRandy Dunlap * ext4_ext_next_allocated_block: 1590f17722f9SLukas Czerner * returns allocated block in subsequent extent or EXT_MAX_BLOCKS. 1591d0d856e8SRandy Dunlap * NOTE: it considers block number from index entry as 1592d0d856e8SRandy Dunlap * allocated block. Thus, index entries have to be consistent 1593d0d856e8SRandy Dunlap * with leaves. 1594a86c6181SAlex Tomas */ 1595fcf6b1b7SDmitry Monakhov ext4_lblk_t 1596a86c6181SAlex Tomas ext4_ext_next_allocated_block(struct ext4_ext_path *path) 1597a86c6181SAlex Tomas { 1598a86c6181SAlex Tomas int depth; 1599a86c6181SAlex Tomas 1600a86c6181SAlex Tomas BUG_ON(path == NULL); 1601a86c6181SAlex Tomas depth = path->p_depth; 1602a86c6181SAlex Tomas 1603a86c6181SAlex Tomas if (depth == 0 && path->p_ext == NULL) 1604f17722f9SLukas Czerner return EXT_MAX_BLOCKS; 1605a86c6181SAlex Tomas 1606a86c6181SAlex Tomas while (depth >= 0) { 16076e89bbb7SEric Biggers struct ext4_ext_path *p = &path[depth]; 16086e89bbb7SEric Biggers 1609a86c6181SAlex Tomas if (depth == path->p_depth) { 1610a86c6181SAlex Tomas /* leaf */ 16116e89bbb7SEric Biggers if (p->p_ext && p->p_ext != EXT_LAST_EXTENT(p->p_hdr)) 16126e89bbb7SEric Biggers return le32_to_cpu(p->p_ext[1].ee_block); 1613a86c6181SAlex Tomas } else { 1614a86c6181SAlex Tomas /* index */ 16156e89bbb7SEric Biggers if (p->p_idx != EXT_LAST_INDEX(p->p_hdr)) 16166e89bbb7SEric Biggers return le32_to_cpu(p->p_idx[1].ei_block); 1617a86c6181SAlex Tomas } 1618a86c6181SAlex Tomas depth--; 1619a86c6181SAlex Tomas } 1620a86c6181SAlex Tomas 1621f17722f9SLukas Czerner return EXT_MAX_BLOCKS; 1622a86c6181SAlex Tomas } 1623a86c6181SAlex Tomas 1624a86c6181SAlex Tomas /* 1625d0d856e8SRandy Dunlap * ext4_ext_next_leaf_block: 1626f17722f9SLukas Czerner * returns first allocated block from next leaf or EXT_MAX_BLOCKS 1627a86c6181SAlex Tomas */ 16285718789dSRobin Dong static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path) 1629a86c6181SAlex Tomas { 1630a86c6181SAlex Tomas int depth; 1631a86c6181SAlex Tomas 1632a86c6181SAlex Tomas BUG_ON(path == NULL); 1633a86c6181SAlex Tomas depth = path->p_depth; 1634a86c6181SAlex Tomas 1635a86c6181SAlex Tomas /* zero-tree has no leaf blocks at all */ 1636a86c6181SAlex Tomas if (depth == 0) 1637f17722f9SLukas Czerner return EXT_MAX_BLOCKS; 1638a86c6181SAlex Tomas 1639a86c6181SAlex Tomas /* go to index block */ 1640a86c6181SAlex Tomas depth--; 1641a86c6181SAlex Tomas 1642a86c6181SAlex Tomas while (depth >= 0) { 1643a86c6181SAlex Tomas if (path[depth].p_idx != 1644a86c6181SAlex Tomas EXT_LAST_INDEX(path[depth].p_hdr)) 1645725d26d3SAneesh Kumar K.V return (ext4_lblk_t) 1646725d26d3SAneesh Kumar K.V le32_to_cpu(path[depth].p_idx[1].ei_block); 1647a86c6181SAlex Tomas depth--; 1648a86c6181SAlex Tomas } 1649a86c6181SAlex Tomas 1650f17722f9SLukas Czerner return EXT_MAX_BLOCKS; 1651a86c6181SAlex Tomas } 1652a86c6181SAlex Tomas 1653a86c6181SAlex Tomas /* 1654d0d856e8SRandy Dunlap * ext4_ext_correct_indexes: 1655d0d856e8SRandy Dunlap * if leaf gets modified and modified extent is first in the leaf, 1656d0d856e8SRandy Dunlap * then we have to correct all indexes above. 1657a86c6181SAlex Tomas * TODO: do we need to correct tree in all cases? 1658a86c6181SAlex Tomas */ 16591d03ec98SAneesh Kumar K.V static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode, 1660a86c6181SAlex Tomas struct ext4_ext_path *path) 1661a86c6181SAlex Tomas { 1662a86c6181SAlex Tomas struct ext4_extent_header *eh; 1663a86c6181SAlex Tomas int depth = ext_depth(inode); 1664a86c6181SAlex Tomas struct ext4_extent *ex; 1665a86c6181SAlex Tomas __le32 border; 1666a86c6181SAlex Tomas int k, err = 0; 1667a86c6181SAlex Tomas 1668a86c6181SAlex Tomas eh = path[depth].p_hdr; 1669a86c6181SAlex Tomas ex = path[depth].p_ext; 1670273df556SFrank Mayhar 1671273df556SFrank Mayhar if (unlikely(ex == NULL || eh == NULL)) { 1672273df556SFrank Mayhar EXT4_ERROR_INODE(inode, 1673273df556SFrank Mayhar "ex %p == NULL or eh %p == NULL", ex, eh); 16746a797d27SDarrick J. Wong return -EFSCORRUPTED; 1675273df556SFrank Mayhar } 1676a86c6181SAlex Tomas 1677a86c6181SAlex Tomas if (depth == 0) { 1678a86c6181SAlex Tomas /* there is no tree at all */ 1679a86c6181SAlex Tomas return 0; 1680a86c6181SAlex Tomas } 1681a86c6181SAlex Tomas 1682a86c6181SAlex Tomas if (ex != EXT_FIRST_EXTENT(eh)) { 1683a86c6181SAlex Tomas /* we correct tree if first leaf got modified only */ 1684a86c6181SAlex Tomas return 0; 1685a86c6181SAlex Tomas } 1686a86c6181SAlex Tomas 1687a86c6181SAlex Tomas /* 1688d0d856e8SRandy Dunlap * TODO: we need correction if border is smaller than current one 1689a86c6181SAlex Tomas */ 1690a86c6181SAlex Tomas k = depth - 1; 1691a86c6181SAlex Tomas border = path[depth].p_ext->ee_block; 16927e028976SAvantika Mathur err = ext4_ext_get_access(handle, inode, path + k); 16937e028976SAvantika Mathur if (err) 1694a86c6181SAlex Tomas return err; 1695a86c6181SAlex Tomas path[k].p_idx->ei_block = border; 16967e028976SAvantika Mathur err = ext4_ext_dirty(handle, inode, path + k); 16977e028976SAvantika Mathur if (err) 1698a86c6181SAlex Tomas return err; 1699a86c6181SAlex Tomas 1700a86c6181SAlex Tomas while (k--) { 1701a86c6181SAlex Tomas /* change all left-side indexes */ 1702a86c6181SAlex Tomas if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr)) 1703a86c6181SAlex Tomas break; 17047e028976SAvantika Mathur err = ext4_ext_get_access(handle, inode, path + k); 17057e028976SAvantika Mathur if (err) 1706a86c6181SAlex Tomas break; 1707a86c6181SAlex Tomas path[k].p_idx->ei_block = border; 17087e028976SAvantika Mathur err = ext4_ext_dirty(handle, inode, path + k); 17097e028976SAvantika Mathur if (err) 1710a86c6181SAlex Tomas break; 1711a86c6181SAlex Tomas } 1712a86c6181SAlex Tomas 1713a86c6181SAlex Tomas return err; 1714a86c6181SAlex Tomas } 1715a86c6181SAlex Tomas 171643f81677SEric Biggers static int ext4_can_extents_be_merged(struct inode *inode, 171743f81677SEric Biggers struct ext4_extent *ex1, 1718a86c6181SAlex Tomas struct ext4_extent *ex2) 1719a86c6181SAlex Tomas { 1720da0169b3SEric Sandeen unsigned short ext1_ee_len, ext2_ee_len; 1721a2df2a63SAmit Arora 1722556615dcSLukas Czerner if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2)) 1723a2df2a63SAmit Arora return 0; 1724a2df2a63SAmit Arora 1725a2df2a63SAmit Arora ext1_ee_len = ext4_ext_get_actual_len(ex1); 1726a2df2a63SAmit Arora ext2_ee_len = ext4_ext_get_actual_len(ex2); 1727a2df2a63SAmit Arora 1728a2df2a63SAmit Arora if (le32_to_cpu(ex1->ee_block) + ext1_ee_len != 172963f57933SAndrew Morton le32_to_cpu(ex2->ee_block)) 1730a86c6181SAlex Tomas return 0; 1731a86c6181SAlex Tomas 1732da0169b3SEric Sandeen if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN) 1733471d4011SSuparna Bhattacharya return 0; 1734378f32baSMatthew Bobrowski 1735556615dcSLukas Czerner if (ext4_ext_is_unwritten(ex1) && 1736378f32baSMatthew Bobrowski ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN) 1737a9b82415SDarrick J. Wong return 0; 1738bbf2f9fbSRobert P. J. Day #ifdef AGGRESSIVE_TEST 1739b939e376SAneesh Kumar K.V if (ext1_ee_len >= 4) 1740a86c6181SAlex Tomas return 0; 1741a86c6181SAlex Tomas #endif 1742a86c6181SAlex Tomas 1743bf89d16fSTheodore Ts'o if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2)) 1744a86c6181SAlex Tomas return 1; 1745a86c6181SAlex Tomas return 0; 1746a86c6181SAlex Tomas } 1747a86c6181SAlex Tomas 1748a86c6181SAlex Tomas /* 174956055d3aSAmit Arora * This function tries to merge the "ex" extent to the next extent in the tree. 175056055d3aSAmit Arora * It always tries to merge towards right. If you want to merge towards 175156055d3aSAmit Arora * left, pass "ex - 1" as argument instead of "ex". 175256055d3aSAmit Arora * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns 175356055d3aSAmit Arora * 1 if they got merged. 175456055d3aSAmit Arora */ 1755197217a5SYongqiang Yang static int ext4_ext_try_to_merge_right(struct inode *inode, 175656055d3aSAmit Arora struct ext4_ext_path *path, 175756055d3aSAmit Arora struct ext4_extent *ex) 175856055d3aSAmit Arora { 175956055d3aSAmit Arora struct ext4_extent_header *eh; 176056055d3aSAmit Arora unsigned int depth, len; 1761556615dcSLukas Czerner int merge_done = 0, unwritten; 176256055d3aSAmit Arora 176356055d3aSAmit Arora depth = ext_depth(inode); 176456055d3aSAmit Arora BUG_ON(path[depth].p_hdr == NULL); 176556055d3aSAmit Arora eh = path[depth].p_hdr; 176656055d3aSAmit Arora 176756055d3aSAmit Arora while (ex < EXT_LAST_EXTENT(eh)) { 176856055d3aSAmit Arora if (!ext4_can_extents_be_merged(inode, ex, ex + 1)) 176956055d3aSAmit Arora break; 177056055d3aSAmit Arora /* merge with next extent! */ 1771556615dcSLukas Czerner unwritten = ext4_ext_is_unwritten(ex); 177256055d3aSAmit Arora ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex) 177356055d3aSAmit Arora + ext4_ext_get_actual_len(ex + 1)); 1774556615dcSLukas Czerner if (unwritten) 1775556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); 177656055d3aSAmit Arora 177756055d3aSAmit Arora if (ex + 1 < EXT_LAST_EXTENT(eh)) { 177856055d3aSAmit Arora len = (EXT_LAST_EXTENT(eh) - ex - 1) 177956055d3aSAmit Arora * sizeof(struct ext4_extent); 178056055d3aSAmit Arora memmove(ex + 1, ex + 2, len); 178156055d3aSAmit Arora } 1782e8546d06SMarcin Slusarz le16_add_cpu(&eh->eh_entries, -1); 178356055d3aSAmit Arora merge_done = 1; 178456055d3aSAmit Arora WARN_ON(eh->eh_entries == 0); 178556055d3aSAmit Arora if (!eh->eh_entries) 178624676da4STheodore Ts'o EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!"); 178756055d3aSAmit Arora } 178856055d3aSAmit Arora 178956055d3aSAmit Arora return merge_done; 179056055d3aSAmit Arora } 179156055d3aSAmit Arora 179256055d3aSAmit Arora /* 1793ecb94f5fSTheodore Ts'o * This function does a very simple check to see if we can collapse 1794ecb94f5fSTheodore Ts'o * an extent tree with a single extent tree leaf block into the inode. 1795ecb94f5fSTheodore Ts'o */ 1796ecb94f5fSTheodore Ts'o static void ext4_ext_try_to_merge_up(handle_t *handle, 1797ecb94f5fSTheodore Ts'o struct inode *inode, 1798ecb94f5fSTheodore Ts'o struct ext4_ext_path *path) 1799ecb94f5fSTheodore Ts'o { 1800ecb94f5fSTheodore Ts'o size_t s; 1801ecb94f5fSTheodore Ts'o unsigned max_root = ext4_ext_space_root(inode, 0); 1802ecb94f5fSTheodore Ts'o ext4_fsblk_t blk; 1803ecb94f5fSTheodore Ts'o 1804ecb94f5fSTheodore Ts'o if ((path[0].p_depth != 1) || 1805ecb94f5fSTheodore Ts'o (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) || 1806ecb94f5fSTheodore Ts'o (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root)) 1807ecb94f5fSTheodore Ts'o return; 1808ecb94f5fSTheodore Ts'o 1809ecb94f5fSTheodore Ts'o /* 1810ecb94f5fSTheodore Ts'o * We need to modify the block allocation bitmap and the block 1811ecb94f5fSTheodore Ts'o * group descriptor to release the extent tree block. If we 1812ecb94f5fSTheodore Ts'o * can't get the journal credits, give up. 1813ecb94f5fSTheodore Ts'o */ 181483448bdfSJan Kara if (ext4_journal_extend(handle, 2, 181583448bdfSJan Kara ext4_free_metadata_revoke_credits(inode->i_sb, 1))) 1816ecb94f5fSTheodore Ts'o return; 1817ecb94f5fSTheodore Ts'o 1818ecb94f5fSTheodore Ts'o /* 1819ecb94f5fSTheodore Ts'o * Copy the extent data up to the inode 1820ecb94f5fSTheodore Ts'o */ 1821ecb94f5fSTheodore Ts'o blk = ext4_idx_pblock(path[0].p_idx); 1822ecb94f5fSTheodore Ts'o s = le16_to_cpu(path[1].p_hdr->eh_entries) * 1823ecb94f5fSTheodore Ts'o sizeof(struct ext4_extent_idx); 1824ecb94f5fSTheodore Ts'o s += sizeof(struct ext4_extent_header); 1825ecb94f5fSTheodore Ts'o 182610809df8STheodore Ts'o path[1].p_maxdepth = path[0].p_maxdepth; 1827ecb94f5fSTheodore Ts'o memcpy(path[0].p_hdr, path[1].p_hdr, s); 1828ecb94f5fSTheodore Ts'o path[0].p_depth = 0; 1829ecb94f5fSTheodore Ts'o path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) + 1830ecb94f5fSTheodore Ts'o (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr)); 1831ecb94f5fSTheodore Ts'o path[0].p_hdr->eh_max = cpu_to_le16(max_root); 1832ecb94f5fSTheodore Ts'o 1833ecb94f5fSTheodore Ts'o brelse(path[1].p_bh); 1834ecb94f5fSTheodore Ts'o ext4_free_blocks(handle, inode, NULL, blk, 1, 183571d4f7d0STheodore Ts'o EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET); 1836ecb94f5fSTheodore Ts'o } 1837ecb94f5fSTheodore Ts'o 1838ecb94f5fSTheodore Ts'o /* 1839adde81cfSEric Biggers * This function tries to merge the @ex extent to neighbours in the tree, then 1840adde81cfSEric Biggers * tries to collapse the extent tree into the inode. 1841197217a5SYongqiang Yang */ 1842ecb94f5fSTheodore Ts'o static void ext4_ext_try_to_merge(handle_t *handle, 1843ecb94f5fSTheodore Ts'o struct inode *inode, 1844197217a5SYongqiang Yang struct ext4_ext_path *path, 1845adde81cfSEric Biggers struct ext4_extent *ex) 1846adde81cfSEric Biggers { 1847197217a5SYongqiang Yang struct ext4_extent_header *eh; 1848197217a5SYongqiang Yang unsigned int depth; 1849197217a5SYongqiang Yang int merge_done = 0; 1850197217a5SYongqiang Yang 1851197217a5SYongqiang Yang depth = ext_depth(inode); 1852197217a5SYongqiang Yang BUG_ON(path[depth].p_hdr == NULL); 1853197217a5SYongqiang Yang eh = path[depth].p_hdr; 1854197217a5SYongqiang Yang 1855197217a5SYongqiang Yang if (ex > EXT_FIRST_EXTENT(eh)) 1856197217a5SYongqiang Yang merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1); 1857197217a5SYongqiang Yang 1858197217a5SYongqiang Yang if (!merge_done) 1859ecb94f5fSTheodore Ts'o (void) ext4_ext_try_to_merge_right(inode, path, ex); 1860197217a5SYongqiang Yang 1861ecb94f5fSTheodore Ts'o ext4_ext_try_to_merge_up(handle, inode, path); 1862197217a5SYongqiang Yang } 1863197217a5SYongqiang Yang 1864197217a5SYongqiang Yang /* 186525d14f98SAmit Arora * check if a portion of the "newext" extent overlaps with an 186625d14f98SAmit Arora * existing extent. 186725d14f98SAmit Arora * 186825d14f98SAmit Arora * If there is an overlap discovered, it updates the length of the newext 186925d14f98SAmit Arora * such that there will be no overlap, and then returns 1. 187025d14f98SAmit Arora * If there is no overlap found, it returns 0. 187125d14f98SAmit Arora */ 18724d33b1efSTheodore Ts'o static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi, 18734d33b1efSTheodore Ts'o struct inode *inode, 187425d14f98SAmit Arora struct ext4_extent *newext, 187525d14f98SAmit Arora struct ext4_ext_path *path) 187625d14f98SAmit Arora { 1877725d26d3SAneesh Kumar K.V ext4_lblk_t b1, b2; 187825d14f98SAmit Arora unsigned int depth, len1; 187925d14f98SAmit Arora unsigned int ret = 0; 188025d14f98SAmit Arora 188125d14f98SAmit Arora b1 = le32_to_cpu(newext->ee_block); 1882a2df2a63SAmit Arora len1 = ext4_ext_get_actual_len(newext); 188325d14f98SAmit Arora depth = ext_depth(inode); 188425d14f98SAmit Arora if (!path[depth].p_ext) 188525d14f98SAmit Arora goto out; 1886f5a44db5STheodore Ts'o b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block)); 188725d14f98SAmit Arora 188825d14f98SAmit Arora /* 188925d14f98SAmit Arora * get the next allocated block if the extent in the path 189025d14f98SAmit Arora * is before the requested block(s) 189125d14f98SAmit Arora */ 189225d14f98SAmit Arora if (b2 < b1) { 189325d14f98SAmit Arora b2 = ext4_ext_next_allocated_block(path); 1894f17722f9SLukas Czerner if (b2 == EXT_MAX_BLOCKS) 189525d14f98SAmit Arora goto out; 1896f5a44db5STheodore Ts'o b2 = EXT4_LBLK_CMASK(sbi, b2); 189725d14f98SAmit Arora } 189825d14f98SAmit Arora 1899725d26d3SAneesh Kumar K.V /* check for wrap through zero on extent logical start block*/ 190025d14f98SAmit Arora if (b1 + len1 < b1) { 1901f17722f9SLukas Czerner len1 = EXT_MAX_BLOCKS - b1; 190225d14f98SAmit Arora newext->ee_len = cpu_to_le16(len1); 190325d14f98SAmit Arora ret = 1; 190425d14f98SAmit Arora } 190525d14f98SAmit Arora 190625d14f98SAmit Arora /* check for overlap */ 190725d14f98SAmit Arora if (b1 + len1 > b2) { 190825d14f98SAmit Arora newext->ee_len = cpu_to_le16(b2 - b1); 190925d14f98SAmit Arora ret = 1; 191025d14f98SAmit Arora } 191125d14f98SAmit Arora out: 191225d14f98SAmit Arora return ret; 191325d14f98SAmit Arora } 191425d14f98SAmit Arora 191525d14f98SAmit Arora /* 1916d0d856e8SRandy Dunlap * ext4_ext_insert_extent: 1917d0d856e8SRandy Dunlap * tries to merge requsted extent into the existing extent or 1918d0d856e8SRandy Dunlap * inserts requested extent as new one into the tree, 1919d0d856e8SRandy Dunlap * creating new leaf in the no-space case. 1920a86c6181SAlex Tomas */ 1921a86c6181SAlex Tomas int ext4_ext_insert_extent(handle_t *handle, struct inode *inode, 1922dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 1923107a7bd3STheodore Ts'o struct ext4_extent *newext, int gb_flags) 1924a86c6181SAlex Tomas { 1925dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 1926a86c6181SAlex Tomas struct ext4_extent_header *eh; 1927a86c6181SAlex Tomas struct ext4_extent *ex, *fex; 1928a86c6181SAlex Tomas struct ext4_extent *nearex; /* nearest extent */ 1929a86c6181SAlex Tomas struct ext4_ext_path *npath = NULL; 1930725d26d3SAneesh Kumar K.V int depth, len, err; 1931725d26d3SAneesh Kumar K.V ext4_lblk_t next; 1932556615dcSLukas Czerner int mb_flags = 0, unwritten; 1933a86c6181SAlex Tomas 1934e3cf5d5dSTheodore Ts'o if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) 1935e3cf5d5dSTheodore Ts'o mb_flags |= EXT4_MB_DELALLOC_RESERVED; 1936273df556SFrank Mayhar if (unlikely(ext4_ext_get_actual_len(newext) == 0)) { 1937273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0"); 19386a797d27SDarrick J. Wong return -EFSCORRUPTED; 1939273df556SFrank Mayhar } 1940a86c6181SAlex Tomas depth = ext_depth(inode); 1941a86c6181SAlex Tomas ex = path[depth].p_ext; 1942be8981beSLukas Czerner eh = path[depth].p_hdr; 1943273df556SFrank Mayhar if (unlikely(path[depth].p_hdr == NULL)) { 1944273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth); 19456a797d27SDarrick J. Wong return -EFSCORRUPTED; 1946273df556SFrank Mayhar } 1947a86c6181SAlex Tomas 1948a86c6181SAlex Tomas /* try to insert block into found extent and return */ 1949107a7bd3STheodore Ts'o if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) { 1950be8981beSLukas Czerner 1951be8981beSLukas Czerner /* 1952be8981beSLukas Czerner * Try to see whether we should rather test the extent on 1953be8981beSLukas Czerner * right from ex, or from the left of ex. This is because 1954ed8a1a76STheodore Ts'o * ext4_find_extent() can return either extent on the 1955be8981beSLukas Czerner * left, or on the right from the searched position. This 1956be8981beSLukas Czerner * will make merging more effective. 1957be8981beSLukas Czerner */ 1958be8981beSLukas Czerner if (ex < EXT_LAST_EXTENT(eh) && 1959be8981beSLukas Czerner (le32_to_cpu(ex->ee_block) + 1960be8981beSLukas Czerner ext4_ext_get_actual_len(ex) < 1961be8981beSLukas Czerner le32_to_cpu(newext->ee_block))) { 1962be8981beSLukas Czerner ex += 1; 1963be8981beSLukas Czerner goto prepend; 1964be8981beSLukas Czerner } else if ((ex > EXT_FIRST_EXTENT(eh)) && 1965be8981beSLukas Czerner (le32_to_cpu(newext->ee_block) + 1966be8981beSLukas Czerner ext4_ext_get_actual_len(newext) < 1967be8981beSLukas Czerner le32_to_cpu(ex->ee_block))) 1968be8981beSLukas Czerner ex -= 1; 1969be8981beSLukas Czerner 1970be8981beSLukas Czerner /* Try to append newex to the ex */ 1971be8981beSLukas Czerner if (ext4_can_extents_be_merged(inode, ex, newext)) { 1972be8981beSLukas Czerner ext_debug("append [%d]%d block to %u:[%d]%d" 1973be8981beSLukas Czerner "(from %llu)\n", 1974556615dcSLukas Czerner ext4_ext_is_unwritten(newext), 1975a2df2a63SAmit Arora ext4_ext_get_actual_len(newext), 1976a86c6181SAlex Tomas le32_to_cpu(ex->ee_block), 1977556615dcSLukas Czerner ext4_ext_is_unwritten(ex), 1978bf89d16fSTheodore Ts'o ext4_ext_get_actual_len(ex), 1979bf89d16fSTheodore Ts'o ext4_ext_pblock(ex)); 1980be8981beSLukas Czerner err = ext4_ext_get_access(handle, inode, 1981be8981beSLukas Czerner path + depth); 19827e028976SAvantika Mathur if (err) 1983a86c6181SAlex Tomas return err; 1984556615dcSLukas Czerner unwritten = ext4_ext_is_unwritten(ex); 1985a2df2a63SAmit Arora ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex) 1986a2df2a63SAmit Arora + ext4_ext_get_actual_len(newext)); 1987556615dcSLukas Czerner if (unwritten) 1988556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); 1989a86c6181SAlex Tomas eh = path[depth].p_hdr; 1990a86c6181SAlex Tomas nearex = ex; 1991a86c6181SAlex Tomas goto merge; 1992a86c6181SAlex Tomas } 1993a86c6181SAlex Tomas 1994be8981beSLukas Czerner prepend: 1995be8981beSLukas Czerner /* Try to prepend newex to the ex */ 1996be8981beSLukas Czerner if (ext4_can_extents_be_merged(inode, newext, ex)) { 1997be8981beSLukas Czerner ext_debug("prepend %u[%d]%d block to %u:[%d]%d" 1998be8981beSLukas Czerner "(from %llu)\n", 1999be8981beSLukas Czerner le32_to_cpu(newext->ee_block), 2000556615dcSLukas Czerner ext4_ext_is_unwritten(newext), 2001be8981beSLukas Czerner ext4_ext_get_actual_len(newext), 2002be8981beSLukas Czerner le32_to_cpu(ex->ee_block), 2003556615dcSLukas Czerner ext4_ext_is_unwritten(ex), 2004be8981beSLukas Czerner ext4_ext_get_actual_len(ex), 2005be8981beSLukas Czerner ext4_ext_pblock(ex)); 2006be8981beSLukas Czerner err = ext4_ext_get_access(handle, inode, 2007be8981beSLukas Czerner path + depth); 2008be8981beSLukas Czerner if (err) 2009be8981beSLukas Czerner return err; 2010be8981beSLukas Czerner 2011556615dcSLukas Czerner unwritten = ext4_ext_is_unwritten(ex); 2012be8981beSLukas Czerner ex->ee_block = newext->ee_block; 2013be8981beSLukas Czerner ext4_ext_store_pblock(ex, ext4_ext_pblock(newext)); 2014be8981beSLukas Czerner ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex) 2015be8981beSLukas Czerner + ext4_ext_get_actual_len(newext)); 2016556615dcSLukas Czerner if (unwritten) 2017556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); 2018be8981beSLukas Czerner eh = path[depth].p_hdr; 2019be8981beSLukas Czerner nearex = ex; 2020be8981beSLukas Czerner goto merge; 2021be8981beSLukas Czerner } 2022be8981beSLukas Czerner } 2023be8981beSLukas Czerner 2024a86c6181SAlex Tomas depth = ext_depth(inode); 2025a86c6181SAlex Tomas eh = path[depth].p_hdr; 2026a86c6181SAlex Tomas if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) 2027a86c6181SAlex Tomas goto has_space; 2028a86c6181SAlex Tomas 2029a86c6181SAlex Tomas /* probably next leaf has space for us? */ 2030a86c6181SAlex Tomas fex = EXT_LAST_EXTENT(eh); 2031598dbdf2SRobin Dong next = EXT_MAX_BLOCKS; 2032598dbdf2SRobin Dong if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block)) 20335718789dSRobin Dong next = ext4_ext_next_leaf_block(path); 2034598dbdf2SRobin Dong if (next != EXT_MAX_BLOCKS) { 203532de6756SYongqiang Yang ext_debug("next leaf block - %u\n", next); 2036a86c6181SAlex Tomas BUG_ON(npath != NULL); 203773c384c0STheodore Ts'o npath = ext4_find_extent(inode, next, NULL, gb_flags); 2038a86c6181SAlex Tomas if (IS_ERR(npath)) 2039a86c6181SAlex Tomas return PTR_ERR(npath); 2040a86c6181SAlex Tomas BUG_ON(npath->p_depth != path->p_depth); 2041a86c6181SAlex Tomas eh = npath[depth].p_hdr; 2042a86c6181SAlex Tomas if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) { 204325985edcSLucas De Marchi ext_debug("next leaf isn't full(%d)\n", 2044a86c6181SAlex Tomas le16_to_cpu(eh->eh_entries)); 2045a86c6181SAlex Tomas path = npath; 2046ffb505ffSRobin Dong goto has_space; 2047a86c6181SAlex Tomas } 2048a86c6181SAlex Tomas ext_debug("next leaf has no free space(%d,%d)\n", 2049a86c6181SAlex Tomas le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max)); 2050a86c6181SAlex Tomas } 2051a86c6181SAlex Tomas 2052a86c6181SAlex Tomas /* 2053d0d856e8SRandy Dunlap * There is no free space in the found leaf. 2054d0d856e8SRandy Dunlap * We're gonna add a new leaf in the tree. 2055a86c6181SAlex Tomas */ 2056107a7bd3STheodore Ts'o if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL) 2057e3cf5d5dSTheodore Ts'o mb_flags |= EXT4_MB_USE_RESERVED; 2058107a7bd3STheodore Ts'o err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags, 2059dfe50809STheodore Ts'o ppath, newext); 2060a86c6181SAlex Tomas if (err) 2061a86c6181SAlex Tomas goto cleanup; 2062a86c6181SAlex Tomas depth = ext_depth(inode); 2063a86c6181SAlex Tomas eh = path[depth].p_hdr; 2064a86c6181SAlex Tomas 2065a86c6181SAlex Tomas has_space: 2066a86c6181SAlex Tomas nearex = path[depth].p_ext; 2067a86c6181SAlex Tomas 20687e028976SAvantika Mathur err = ext4_ext_get_access(handle, inode, path + depth); 20697e028976SAvantika Mathur if (err) 2070a86c6181SAlex Tomas goto cleanup; 2071a86c6181SAlex Tomas 2072a86c6181SAlex Tomas if (!nearex) { 2073a86c6181SAlex Tomas /* there is no extent in this leaf, create first one */ 207432de6756SYongqiang Yang ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n", 2075a86c6181SAlex Tomas le32_to_cpu(newext->ee_block), 2076bf89d16fSTheodore Ts'o ext4_ext_pblock(newext), 2077556615dcSLukas Czerner ext4_ext_is_unwritten(newext), 2078a2df2a63SAmit Arora ext4_ext_get_actual_len(newext)); 207980e675f9SEric Gouriou nearex = EXT_FIRST_EXTENT(eh); 2080a86c6181SAlex Tomas } else { 208180e675f9SEric Gouriou if (le32_to_cpu(newext->ee_block) 208280e675f9SEric Gouriou > le32_to_cpu(nearex->ee_block)) { 208380e675f9SEric Gouriou /* Insert after */ 208432de6756SYongqiang Yang ext_debug("insert %u:%llu:[%d]%d before: " 208532de6756SYongqiang Yang "nearest %p\n", 2086a86c6181SAlex Tomas le32_to_cpu(newext->ee_block), 2087bf89d16fSTheodore Ts'o ext4_ext_pblock(newext), 2088556615dcSLukas Czerner ext4_ext_is_unwritten(newext), 2089a2df2a63SAmit Arora ext4_ext_get_actual_len(newext), 209080e675f9SEric Gouriou nearex); 209180e675f9SEric Gouriou nearex++; 209280e675f9SEric Gouriou } else { 209380e675f9SEric Gouriou /* Insert before */ 209480e675f9SEric Gouriou BUG_ON(newext->ee_block == nearex->ee_block); 209532de6756SYongqiang Yang ext_debug("insert %u:%llu:[%d]%d after: " 209632de6756SYongqiang Yang "nearest %p\n", 209780e675f9SEric Gouriou le32_to_cpu(newext->ee_block), 209880e675f9SEric Gouriou ext4_ext_pblock(newext), 2099556615dcSLukas Czerner ext4_ext_is_unwritten(newext), 210080e675f9SEric Gouriou ext4_ext_get_actual_len(newext), 210180e675f9SEric Gouriou nearex); 210280e675f9SEric Gouriou } 210380e675f9SEric Gouriou len = EXT_LAST_EXTENT(eh) - nearex + 1; 210480e675f9SEric Gouriou if (len > 0) { 210532de6756SYongqiang Yang ext_debug("insert %u:%llu:[%d]%d: " 210680e675f9SEric Gouriou "move %d extents from 0x%p to 0x%p\n", 210780e675f9SEric Gouriou le32_to_cpu(newext->ee_block), 210880e675f9SEric Gouriou ext4_ext_pblock(newext), 2109556615dcSLukas Czerner ext4_ext_is_unwritten(newext), 211080e675f9SEric Gouriou ext4_ext_get_actual_len(newext), 211180e675f9SEric Gouriou len, nearex, nearex + 1); 211280e675f9SEric Gouriou memmove(nearex + 1, nearex, 211380e675f9SEric Gouriou len * sizeof(struct ext4_extent)); 211480e675f9SEric Gouriou } 2115a86c6181SAlex Tomas } 2116a86c6181SAlex Tomas 2117e8546d06SMarcin Slusarz le16_add_cpu(&eh->eh_entries, 1); 211880e675f9SEric Gouriou path[depth].p_ext = nearex; 2119a86c6181SAlex Tomas nearex->ee_block = newext->ee_block; 2120bf89d16fSTheodore Ts'o ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext)); 2121a86c6181SAlex Tomas nearex->ee_len = newext->ee_len; 2122a86c6181SAlex Tomas 2123a86c6181SAlex Tomas merge: 2124e7bcf823SHaiboLiu /* try to merge extents */ 2125107a7bd3STheodore Ts'o if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) 2126ecb94f5fSTheodore Ts'o ext4_ext_try_to_merge(handle, inode, path, nearex); 2127a86c6181SAlex Tomas 2128a86c6181SAlex Tomas 2129a86c6181SAlex Tomas /* time to correct all indexes above */ 2130a86c6181SAlex Tomas err = ext4_ext_correct_indexes(handle, inode, path); 2131a86c6181SAlex Tomas if (err) 2132a86c6181SAlex Tomas goto cleanup; 2133a86c6181SAlex Tomas 2134ecb94f5fSTheodore Ts'o err = ext4_ext_dirty(handle, inode, path + path->p_depth); 2135a86c6181SAlex Tomas 2136a86c6181SAlex Tomas cleanup: 2137a86c6181SAlex Tomas ext4_ext_drop_refs(npath); 2138a86c6181SAlex Tomas kfree(npath); 2139a86c6181SAlex Tomas return err; 2140a86c6181SAlex Tomas } 2141a86c6181SAlex Tomas 2142bb5835edSTheodore Ts'o static int ext4_fill_es_cache_info(struct inode *inode, 2143bb5835edSTheodore Ts'o ext4_lblk_t block, ext4_lblk_t num, 2144bb5835edSTheodore Ts'o struct fiemap_extent_info *fieinfo) 2145bb5835edSTheodore Ts'o { 2146bb5835edSTheodore Ts'o ext4_lblk_t next, end = block + num - 1; 2147bb5835edSTheodore Ts'o struct extent_status es; 2148bb5835edSTheodore Ts'o unsigned char blksize_bits = inode->i_sb->s_blocksize_bits; 2149bb5835edSTheodore Ts'o unsigned int flags; 2150bb5835edSTheodore Ts'o int err; 2151bb5835edSTheodore Ts'o 2152bb5835edSTheodore Ts'o while (block <= end) { 2153bb5835edSTheodore Ts'o next = 0; 2154bb5835edSTheodore Ts'o flags = 0; 2155bb5835edSTheodore Ts'o if (!ext4_es_lookup_extent(inode, block, &next, &es)) 2156bb5835edSTheodore Ts'o break; 2157bb5835edSTheodore Ts'o if (ext4_es_is_unwritten(&es)) 2158bb5835edSTheodore Ts'o flags |= FIEMAP_EXTENT_UNWRITTEN; 2159bb5835edSTheodore Ts'o if (ext4_es_is_delayed(&es)) 2160bb5835edSTheodore Ts'o flags |= (FIEMAP_EXTENT_DELALLOC | 2161bb5835edSTheodore Ts'o FIEMAP_EXTENT_UNKNOWN); 2162bb5835edSTheodore Ts'o if (ext4_es_is_hole(&es)) 2163bb5835edSTheodore Ts'o flags |= EXT4_FIEMAP_EXTENT_HOLE; 2164bb5835edSTheodore Ts'o if (next == 0) 2165bb5835edSTheodore Ts'o flags |= FIEMAP_EXTENT_LAST; 2166bb5835edSTheodore Ts'o if (flags & (FIEMAP_EXTENT_DELALLOC| 2167bb5835edSTheodore Ts'o EXT4_FIEMAP_EXTENT_HOLE)) 2168bb5835edSTheodore Ts'o es.es_pblk = 0; 2169bb5835edSTheodore Ts'o else 2170bb5835edSTheodore Ts'o es.es_pblk = ext4_es_pblock(&es); 2171bb5835edSTheodore Ts'o err = fiemap_fill_next_extent(fieinfo, 2172bb5835edSTheodore Ts'o (__u64)es.es_lblk << blksize_bits, 2173bb5835edSTheodore Ts'o (__u64)es.es_pblk << blksize_bits, 2174bb5835edSTheodore Ts'o (__u64)es.es_len << blksize_bits, 2175bb5835edSTheodore Ts'o flags); 2176bb5835edSTheodore Ts'o if (next == 0) 2177bb5835edSTheodore Ts'o break; 2178bb5835edSTheodore Ts'o block = next; 2179bb5835edSTheodore Ts'o if (err < 0) 2180bb5835edSTheodore Ts'o return err; 2181bb5835edSTheodore Ts'o if (err == 1) 2182bb5835edSTheodore Ts'o return 0; 2183bb5835edSTheodore Ts'o } 2184bb5835edSTheodore Ts'o return 0; 2185bb5835edSTheodore Ts'o } 2186bb5835edSTheodore Ts'o 2187bb5835edSTheodore Ts'o 2188a86c6181SAlex Tomas /* 2189140a5250SJan Kara * ext4_ext_determine_hole - determine hole around given block 2190140a5250SJan Kara * @inode: inode we lookup in 2191140a5250SJan Kara * @path: path in extent tree to @lblk 2192140a5250SJan Kara * @lblk: pointer to logical block around which we want to determine hole 2193140a5250SJan Kara * 2194140a5250SJan Kara * Determine hole length (and start if easily possible) around given logical 2195140a5250SJan Kara * block. We don't try too hard to find the beginning of the hole but @path 2196140a5250SJan Kara * actually points to extent before @lblk, we provide it. 2197140a5250SJan Kara * 2198140a5250SJan Kara * The function returns the length of a hole starting at @lblk. We update @lblk 2199140a5250SJan Kara * to the beginning of the hole if we managed to find it. 2200140a5250SJan Kara */ 2201140a5250SJan Kara static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode, 2202140a5250SJan Kara struct ext4_ext_path *path, 2203140a5250SJan Kara ext4_lblk_t *lblk) 2204140a5250SJan Kara { 2205140a5250SJan Kara int depth = ext_depth(inode); 2206140a5250SJan Kara struct ext4_extent *ex; 2207140a5250SJan Kara ext4_lblk_t len; 2208140a5250SJan Kara 2209140a5250SJan Kara ex = path[depth].p_ext; 2210140a5250SJan Kara if (ex == NULL) { 2211140a5250SJan Kara /* there is no extent yet, so gap is [0;-] */ 2212140a5250SJan Kara *lblk = 0; 2213140a5250SJan Kara len = EXT_MAX_BLOCKS; 2214140a5250SJan Kara } else if (*lblk < le32_to_cpu(ex->ee_block)) { 2215140a5250SJan Kara len = le32_to_cpu(ex->ee_block) - *lblk; 2216140a5250SJan Kara } else if (*lblk >= le32_to_cpu(ex->ee_block) 2217140a5250SJan Kara + ext4_ext_get_actual_len(ex)) { 2218140a5250SJan Kara ext4_lblk_t next; 2219140a5250SJan Kara 2220140a5250SJan Kara *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex); 2221140a5250SJan Kara next = ext4_ext_next_allocated_block(path); 2222140a5250SJan Kara BUG_ON(next == *lblk); 2223140a5250SJan Kara len = next - *lblk; 2224140a5250SJan Kara } else { 2225140a5250SJan Kara BUG(); 2226140a5250SJan Kara } 2227140a5250SJan Kara return len; 2228140a5250SJan Kara } 2229140a5250SJan Kara 2230140a5250SJan Kara /* 2231d0d856e8SRandy Dunlap * ext4_ext_put_gap_in_cache: 2232d0d856e8SRandy Dunlap * calculate boundaries of the gap that the requested block fits into 2233a86c6181SAlex Tomas * and cache this gap 2234a86c6181SAlex Tomas */ 223509b88252SAvantika Mathur static void 2236140a5250SJan Kara ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start, 2237140a5250SJan Kara ext4_lblk_t hole_len) 2238a86c6181SAlex Tomas { 22392f8e0a7cSZheng Liu struct extent_status es; 2240a86c6181SAlex Tomas 2241ad431025SEric Whitney ext4_es_find_extent_range(inode, &ext4_es_is_delayed, hole_start, 2242140a5250SJan Kara hole_start + hole_len - 1, &es); 22432f8e0a7cSZheng Liu if (es.es_len) { 22442f8e0a7cSZheng Liu /* There's delayed extent containing lblock? */ 2245140a5250SJan Kara if (es.es_lblk <= hole_start) 22462f8e0a7cSZheng Liu return; 2247140a5250SJan Kara hole_len = min(es.es_lblk - hole_start, hole_len); 22482f8e0a7cSZheng Liu } 2249140a5250SJan Kara ext_debug(" -> %u:%u\n", hole_start, hole_len); 2250140a5250SJan Kara ext4_es_insert_extent(inode, hole_start, hole_len, ~0, 2251140a5250SJan Kara EXTENT_STATUS_HOLE); 2252a86c6181SAlex Tomas } 2253a86c6181SAlex Tomas 2254a86c6181SAlex Tomas /* 2255d0d856e8SRandy Dunlap * ext4_ext_rm_idx: 2256d0d856e8SRandy Dunlap * removes index from the index block. 2257a86c6181SAlex Tomas */ 22581d03ec98SAneesh Kumar K.V static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode, 2259c36575e6SForrest Liu struct ext4_ext_path *path, int depth) 2260a86c6181SAlex Tomas { 2261a86c6181SAlex Tomas int err; 2262f65e6fbaSAlex Tomas ext4_fsblk_t leaf; 2263a86c6181SAlex Tomas 2264a86c6181SAlex Tomas /* free index block */ 2265c36575e6SForrest Liu depth--; 2266c36575e6SForrest Liu path = path + depth; 2267bf89d16fSTheodore Ts'o leaf = ext4_idx_pblock(path->p_idx); 2268273df556SFrank Mayhar if (unlikely(path->p_hdr->eh_entries == 0)) { 2269273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0"); 22706a797d27SDarrick J. Wong return -EFSCORRUPTED; 2271273df556SFrank Mayhar } 22727e028976SAvantika Mathur err = ext4_ext_get_access(handle, inode, path); 22737e028976SAvantika Mathur if (err) 2274a86c6181SAlex Tomas return err; 22750e1147b0SRobin Dong 22760e1147b0SRobin Dong if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) { 22770e1147b0SRobin Dong int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx; 22780e1147b0SRobin Dong len *= sizeof(struct ext4_extent_idx); 22790e1147b0SRobin Dong memmove(path->p_idx, path->p_idx + 1, len); 22800e1147b0SRobin Dong } 22810e1147b0SRobin Dong 2282e8546d06SMarcin Slusarz le16_add_cpu(&path->p_hdr->eh_entries, -1); 22837e028976SAvantika Mathur err = ext4_ext_dirty(handle, inode, path); 22847e028976SAvantika Mathur if (err) 2285a86c6181SAlex Tomas return err; 22862ae02107SMingming Cao ext_debug("index is empty, remove it, free block %llu\n", leaf); 2287d8990240SAditya Kali trace_ext4_ext_rm_idx(inode, leaf); 2288d8990240SAditya Kali 22897dc57615SPeter Huewe ext4_free_blocks(handle, inode, NULL, leaf, 1, 2290e6362609STheodore Ts'o EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET); 2291c36575e6SForrest Liu 2292c36575e6SForrest Liu while (--depth >= 0) { 2293c36575e6SForrest Liu if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr)) 2294c36575e6SForrest Liu break; 2295c36575e6SForrest Liu path--; 2296c36575e6SForrest Liu err = ext4_ext_get_access(handle, inode, path); 2297c36575e6SForrest Liu if (err) 2298c36575e6SForrest Liu break; 2299c36575e6SForrest Liu path->p_idx->ei_block = (path+1)->p_idx->ei_block; 2300c36575e6SForrest Liu err = ext4_ext_dirty(handle, inode, path); 2301c36575e6SForrest Liu if (err) 2302c36575e6SForrest Liu break; 2303c36575e6SForrest Liu } 2304a86c6181SAlex Tomas return err; 2305a86c6181SAlex Tomas } 2306a86c6181SAlex Tomas 2307a86c6181SAlex Tomas /* 2308ee12b630SMingming Cao * ext4_ext_calc_credits_for_single_extent: 2309ee12b630SMingming Cao * This routine returns max. credits that needed to insert an extent 2310ee12b630SMingming Cao * to the extent tree. 2311ee12b630SMingming Cao * When pass the actual path, the caller should calculate credits 2312ee12b630SMingming Cao * under i_data_sem. 2313a86c6181SAlex Tomas */ 2314525f4ed8SMingming Cao int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks, 2315a86c6181SAlex Tomas struct ext4_ext_path *path) 2316a86c6181SAlex Tomas { 2317a86c6181SAlex Tomas if (path) { 2318ee12b630SMingming Cao int depth = ext_depth(inode); 2319f3bd1f3fSMingming Cao int ret = 0; 2320ee12b630SMingming Cao 2321a86c6181SAlex Tomas /* probably there is space in leaf? */ 2322a86c6181SAlex Tomas if (le16_to_cpu(path[depth].p_hdr->eh_entries) 2323ee12b630SMingming Cao < le16_to_cpu(path[depth].p_hdr->eh_max)) { 2324ee12b630SMingming Cao 2325ee12b630SMingming Cao /* 2326ee12b630SMingming Cao * There are some space in the leaf tree, no 2327ee12b630SMingming Cao * need to account for leaf block credit 2328ee12b630SMingming Cao * 2329ee12b630SMingming Cao * bitmaps and block group descriptor blocks 2330df3ab170STao Ma * and other metadata blocks still need to be 2331ee12b630SMingming Cao * accounted. 2332ee12b630SMingming Cao */ 2333525f4ed8SMingming Cao /* 1 bitmap, 1 block group descriptor */ 2334ee12b630SMingming Cao ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb); 23355887e98bSAneesh Kumar K.V return ret; 2336ee12b630SMingming Cao } 2337ee12b630SMingming Cao } 2338ee12b630SMingming Cao 2339525f4ed8SMingming Cao return ext4_chunk_trans_blocks(inode, nrblocks); 2340a86c6181SAlex Tomas } 2341a86c6181SAlex Tomas 2342a86c6181SAlex Tomas /* 2343fffb2739SJan Kara * How many index/leaf blocks need to change/allocate to add @extents extents? 2344ee12b630SMingming Cao * 2345fffb2739SJan Kara * If we add a single extent, then in the worse case, each tree level 2346fffb2739SJan Kara * index/leaf need to be changed in case of the tree split. 2347ee12b630SMingming Cao * 2348fffb2739SJan Kara * If more extents are inserted, they could cause the whole tree split more 2349fffb2739SJan Kara * than once, but this is really rare. 2350a86c6181SAlex Tomas */ 2351fffb2739SJan Kara int ext4_ext_index_trans_blocks(struct inode *inode, int extents) 2352ee12b630SMingming Cao { 2353ee12b630SMingming Cao int index; 2354f19d5870STao Ma int depth; 2355f19d5870STao Ma 2356f19d5870STao Ma /* If we are converting the inline data, only one is needed here. */ 2357f19d5870STao Ma if (ext4_has_inline_data(inode)) 2358f19d5870STao Ma return 1; 2359f19d5870STao Ma 2360f19d5870STao Ma depth = ext_depth(inode); 2361a86c6181SAlex Tomas 2362fffb2739SJan Kara if (extents <= 1) 2363ee12b630SMingming Cao index = depth * 2; 2364ee12b630SMingming Cao else 2365ee12b630SMingming Cao index = depth * 3; 2366a86c6181SAlex Tomas 2367ee12b630SMingming Cao return index; 2368a86c6181SAlex Tomas } 2369a86c6181SAlex Tomas 2370981250caSTheodore Ts'o static inline int get_default_free_blocks_flags(struct inode *inode) 2371981250caSTheodore Ts'o { 2372ddfa17e4STahsin Erdogan if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) || 2373ddfa17e4STahsin Erdogan ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE)) 2374981250caSTheodore Ts'o return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET; 2375981250caSTheodore Ts'o else if (ext4_should_journal_data(inode)) 2376981250caSTheodore Ts'o return EXT4_FREE_BLOCKS_FORGET; 2377981250caSTheodore Ts'o return 0; 2378981250caSTheodore Ts'o } 2379981250caSTheodore Ts'o 23809fe67149SEric Whitney /* 23819fe67149SEric Whitney * ext4_rereserve_cluster - increment the reserved cluster count when 23829fe67149SEric Whitney * freeing a cluster with a pending reservation 23839fe67149SEric Whitney * 23849fe67149SEric Whitney * @inode - file containing the cluster 23859fe67149SEric Whitney * @lblk - logical block in cluster to be reserved 23869fe67149SEric Whitney * 23879fe67149SEric Whitney * Increments the reserved cluster count and adjusts quota in a bigalloc 23889fe67149SEric Whitney * file system when freeing a partial cluster containing at least one 23899fe67149SEric Whitney * delayed and unwritten block. A partial cluster meeting that 23909fe67149SEric Whitney * requirement will have a pending reservation. If so, the 23919fe67149SEric Whitney * RERESERVE_CLUSTER flag is used when calling ext4_free_blocks() to 23929fe67149SEric Whitney * defer reserved and allocated space accounting to a subsequent call 23939fe67149SEric Whitney * to this function. 23949fe67149SEric Whitney */ 23959fe67149SEric Whitney static void ext4_rereserve_cluster(struct inode *inode, ext4_lblk_t lblk) 23969fe67149SEric Whitney { 23979fe67149SEric Whitney struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 23989fe67149SEric Whitney struct ext4_inode_info *ei = EXT4_I(inode); 23999fe67149SEric Whitney 24009fe67149SEric Whitney dquot_reclaim_block(inode, EXT4_C2B(sbi, 1)); 24019fe67149SEric Whitney 24029fe67149SEric Whitney spin_lock(&ei->i_block_reservation_lock); 24039fe67149SEric Whitney ei->i_reserved_data_blocks++; 24049fe67149SEric Whitney percpu_counter_add(&sbi->s_dirtyclusters_counter, 1); 24059fe67149SEric Whitney spin_unlock(&ei->i_block_reservation_lock); 24069fe67149SEric Whitney 24079fe67149SEric Whitney percpu_counter_add(&sbi->s_freeclusters_counter, 1); 24089fe67149SEric Whitney ext4_remove_pending(inode, lblk); 24099fe67149SEric Whitney } 24109fe67149SEric Whitney 2411a86c6181SAlex Tomas static int ext4_remove_blocks(handle_t *handle, struct inode *inode, 2412a86c6181SAlex Tomas struct ext4_extent *ex, 24139fe67149SEric Whitney struct partial_cluster *partial, 2414725d26d3SAneesh Kumar K.V ext4_lblk_t from, ext4_lblk_t to) 2415a86c6181SAlex Tomas { 24160aa06000STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 2417a2df2a63SAmit Arora unsigned short ee_len = ext4_ext_get_actual_len(ex); 24189fe67149SEric Whitney ext4_fsblk_t last_pblk, pblk; 24199fe67149SEric Whitney ext4_lblk_t num; 24209fe67149SEric Whitney int flags; 242118888cf0SAndrey Sidorov 24229fe67149SEric Whitney /* only extent tail removal is allowed */ 24239fe67149SEric Whitney if (from < le32_to_cpu(ex->ee_block) || 24249fe67149SEric Whitney to != le32_to_cpu(ex->ee_block) + ee_len - 1) { 24259fe67149SEric Whitney ext4_error(sbi->s_sb, 24269fe67149SEric Whitney "strange request: removal(2) %u-%u from %u:%u", 24279fe67149SEric Whitney from, to, le32_to_cpu(ex->ee_block), ee_len); 24289fe67149SEric Whitney return 0; 24290aa06000STheodore Ts'o } 24300aa06000STheodore Ts'o 2431a86c6181SAlex Tomas #ifdef EXTENTS_STATS 2432a86c6181SAlex Tomas spin_lock(&sbi->s_ext_stats_lock); 2433a86c6181SAlex Tomas sbi->s_ext_blocks += ee_len; 2434a86c6181SAlex Tomas sbi->s_ext_extents++; 2435a86c6181SAlex Tomas if (ee_len < sbi->s_ext_min) 2436a86c6181SAlex Tomas sbi->s_ext_min = ee_len; 2437a86c6181SAlex Tomas if (ee_len > sbi->s_ext_max) 2438a86c6181SAlex Tomas sbi->s_ext_max = ee_len; 2439a86c6181SAlex Tomas if (ext_depth(inode) > sbi->s_depth_max) 2440a86c6181SAlex Tomas sbi->s_depth_max = ext_depth(inode); 2441a86c6181SAlex Tomas spin_unlock(&sbi->s_ext_stats_lock); 2442a86c6181SAlex Tomas #endif 24439fe67149SEric Whitney 24449fe67149SEric Whitney trace_ext4_remove_blocks(inode, ex, from, to, partial); 24459fe67149SEric Whitney 24469fe67149SEric Whitney /* 24479fe67149SEric Whitney * if we have a partial cluster, and it's different from the 24489fe67149SEric Whitney * cluster of the last block in the extent, we free it 24499fe67149SEric Whitney */ 24509fe67149SEric Whitney last_pblk = ext4_ext_pblock(ex) + ee_len - 1; 24519fe67149SEric Whitney 24529fe67149SEric Whitney if (partial->state != initial && 24539fe67149SEric Whitney partial->pclu != EXT4_B2C(sbi, last_pblk)) { 24549fe67149SEric Whitney if (partial->state == tofree) { 24559fe67149SEric Whitney flags = get_default_free_blocks_flags(inode); 24569fe67149SEric Whitney if (ext4_is_pending(inode, partial->lblk)) 24579fe67149SEric Whitney flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER; 24589fe67149SEric Whitney ext4_free_blocks(handle, inode, NULL, 24599fe67149SEric Whitney EXT4_C2B(sbi, partial->pclu), 24609fe67149SEric Whitney sbi->s_cluster_ratio, flags); 24619fe67149SEric Whitney if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER) 24629fe67149SEric Whitney ext4_rereserve_cluster(inode, partial->lblk); 24639fe67149SEric Whitney } 24649fe67149SEric Whitney partial->state = initial; 24659fe67149SEric Whitney } 2466725d26d3SAneesh Kumar K.V 2467a2df2a63SAmit Arora num = le32_to_cpu(ex->ee_block) + ee_len - from; 24680aa06000STheodore Ts'o pblk = ext4_ext_pblock(ex) + ee_len - num; 24699fe67149SEric Whitney 2470d23142c6SLukas Czerner /* 24719fe67149SEric Whitney * We free the partial cluster at the end of the extent (if any), 24729fe67149SEric Whitney * unless the cluster is used by another extent (partial_cluster 24739fe67149SEric Whitney * state is nofree). If a partial cluster exists here, it must be 24749fe67149SEric Whitney * shared with the last block in the extent. 2475d23142c6SLukas Czerner */ 24769fe67149SEric Whitney flags = get_default_free_blocks_flags(inode); 24779fe67149SEric Whitney 24789fe67149SEric Whitney /* partial, left end cluster aligned, right end unaligned */ 24799fe67149SEric Whitney if ((EXT4_LBLK_COFF(sbi, to) != sbi->s_cluster_ratio - 1) && 24809fe67149SEric Whitney (EXT4_LBLK_CMASK(sbi, to) >= from) && 24819fe67149SEric Whitney (partial->state != nofree)) { 24829fe67149SEric Whitney if (ext4_is_pending(inode, to)) 24839fe67149SEric Whitney flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER; 24849fe67149SEric Whitney ext4_free_blocks(handle, inode, NULL, 24859fe67149SEric Whitney EXT4_PBLK_CMASK(sbi, last_pblk), 24869fe67149SEric Whitney sbi->s_cluster_ratio, flags); 24879fe67149SEric Whitney if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER) 24889fe67149SEric Whitney ext4_rereserve_cluster(inode, to); 24899fe67149SEric Whitney partial->state = initial; 24909fe67149SEric Whitney flags = get_default_free_blocks_flags(inode); 24919fe67149SEric Whitney } 24929fe67149SEric Whitney 2493d23142c6SLukas Czerner flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER; 2494d23142c6SLukas Czerner 24950aa06000STheodore Ts'o /* 24969fe67149SEric Whitney * For bigalloc file systems, we never free a partial cluster 24979fe67149SEric Whitney * at the beginning of the extent. Instead, we check to see if we 24989fe67149SEric Whitney * need to free it on a subsequent call to ext4_remove_blocks, 24999fe67149SEric Whitney * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space. 25000aa06000STheodore Ts'o */ 25019fe67149SEric Whitney flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER; 25029fe67149SEric Whitney ext4_free_blocks(handle, inode, NULL, pblk, num, flags); 25039fe67149SEric Whitney 25049fe67149SEric Whitney /* reset the partial cluster if we've freed past it */ 25059fe67149SEric Whitney if (partial->state != initial && partial->pclu != EXT4_B2C(sbi, pblk)) 25069fe67149SEric Whitney partial->state = initial; 25079fe67149SEric Whitney 25089fe67149SEric Whitney /* 25099fe67149SEric Whitney * If we've freed the entire extent but the beginning is not left 25109fe67149SEric Whitney * cluster aligned and is not marked as ineligible for freeing we 25119fe67149SEric Whitney * record the partial cluster at the beginning of the extent. It 25129fe67149SEric Whitney * wasn't freed by the preceding ext4_free_blocks() call, and we 25139fe67149SEric Whitney * need to look farther to the left to determine if it's to be freed 25149fe67149SEric Whitney * (not shared with another extent). Else, reset the partial 25159fe67149SEric Whitney * cluster - we're either done freeing or the beginning of the 25169fe67149SEric Whitney * extent is left cluster aligned. 25179fe67149SEric Whitney */ 25189fe67149SEric Whitney if (EXT4_LBLK_COFF(sbi, from) && num == ee_len) { 25199fe67149SEric Whitney if (partial->state == initial) { 25209fe67149SEric Whitney partial->pclu = EXT4_B2C(sbi, pblk); 25219fe67149SEric Whitney partial->lblk = from; 25229fe67149SEric Whitney partial->state = tofree; 2523345ee947SEric Whitney } 25249fe67149SEric Whitney } else { 25259fe67149SEric Whitney partial->state = initial; 2526a86c6181SAlex Tomas } 2527a86c6181SAlex Tomas 25289fe67149SEric Whitney return 0; 25299fe67149SEric Whitney } 2530d583fb87SAllison Henderson 2531d583fb87SAllison Henderson /* 2532d583fb87SAllison Henderson * ext4_ext_rm_leaf() Removes the extents associated with the 25335bf43760SEric Whitney * blocks appearing between "start" and "end". Both "start" 25345bf43760SEric Whitney * and "end" must appear in the same extent or EIO is returned. 2535d583fb87SAllison Henderson * 2536d583fb87SAllison Henderson * @handle: The journal handle 2537d583fb87SAllison Henderson * @inode: The files inode 2538d583fb87SAllison Henderson * @path: The path to the leaf 2539d23142c6SLukas Czerner * @partial_cluster: The cluster which we'll have to free if all extents 25405bf43760SEric Whitney * has been released from it. However, if this value is 25415bf43760SEric Whitney * negative, it's a cluster just to the right of the 25425bf43760SEric Whitney * punched region and it must not be freed. 2543d583fb87SAllison Henderson * @start: The first block to remove 2544d583fb87SAllison Henderson * @end: The last block to remove 2545d583fb87SAllison Henderson */ 2546a86c6181SAlex Tomas static int 2547a86c6181SAlex Tomas ext4_ext_rm_leaf(handle_t *handle, struct inode *inode, 2548d23142c6SLukas Czerner struct ext4_ext_path *path, 25499fe67149SEric Whitney struct partial_cluster *partial, 25500aa06000STheodore Ts'o ext4_lblk_t start, ext4_lblk_t end) 2551a86c6181SAlex Tomas { 25520aa06000STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 2553a86c6181SAlex Tomas int err = 0, correct_index = 0; 255483448bdfSJan Kara int depth = ext_depth(inode), credits, revoke_credits; 2555a86c6181SAlex Tomas struct ext4_extent_header *eh; 2556750c9c47SDmitry Monakhov ext4_lblk_t a, b; 2557725d26d3SAneesh Kumar K.V unsigned num; 2558725d26d3SAneesh Kumar K.V ext4_lblk_t ex_ee_block; 2559a86c6181SAlex Tomas unsigned short ex_ee_len; 2560556615dcSLukas Czerner unsigned unwritten = 0; 2561a86c6181SAlex Tomas struct ext4_extent *ex; 2562d23142c6SLukas Czerner ext4_fsblk_t pblk; 2563a86c6181SAlex Tomas 2564c29c0ae7SAlex Tomas /* the header must be checked already in ext4_ext_remove_space() */ 25655f95d21fSLukas Czerner ext_debug("truncate since %u in leaf to %u\n", start, end); 2566a86c6181SAlex Tomas if (!path[depth].p_hdr) 2567a86c6181SAlex Tomas path[depth].p_hdr = ext_block_hdr(path[depth].p_bh); 2568a86c6181SAlex Tomas eh = path[depth].p_hdr; 2569273df556SFrank Mayhar if (unlikely(path[depth].p_hdr == NULL)) { 2570273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth); 25716a797d27SDarrick J. Wong return -EFSCORRUPTED; 2572273df556SFrank Mayhar } 2573a86c6181SAlex Tomas /* find where to start removing */ 25746ae06ff5SAshish Sangwan ex = path[depth].p_ext; 25756ae06ff5SAshish Sangwan if (!ex) 2576a86c6181SAlex Tomas ex = EXT_LAST_EXTENT(eh); 2577a86c6181SAlex Tomas 2578a86c6181SAlex Tomas ex_ee_block = le32_to_cpu(ex->ee_block); 2579a2df2a63SAmit Arora ex_ee_len = ext4_ext_get_actual_len(ex); 2580a86c6181SAlex Tomas 25819fe67149SEric Whitney trace_ext4_ext_rm_leaf(inode, start, ex, partial); 2582d8990240SAditya Kali 2583a86c6181SAlex Tomas while (ex >= EXT_FIRST_EXTENT(eh) && 2584a86c6181SAlex Tomas ex_ee_block + ex_ee_len > start) { 2585a41f2071SAneesh Kumar K.V 2586556615dcSLukas Czerner if (ext4_ext_is_unwritten(ex)) 2587556615dcSLukas Czerner unwritten = 1; 2588a41f2071SAneesh Kumar K.V else 2589556615dcSLukas Czerner unwritten = 0; 2590a41f2071SAneesh Kumar K.V 2591553f9008SMingming ext_debug("remove ext %u:[%d]%d\n", ex_ee_block, 2592556615dcSLukas Czerner unwritten, ex_ee_len); 2593a86c6181SAlex Tomas path[depth].p_ext = ex; 2594a86c6181SAlex Tomas 2595a86c6181SAlex Tomas a = ex_ee_block > start ? ex_ee_block : start; 2596d583fb87SAllison Henderson b = ex_ee_block+ex_ee_len - 1 < end ? 2597d583fb87SAllison Henderson ex_ee_block+ex_ee_len - 1 : end; 2598a86c6181SAlex Tomas 2599a86c6181SAlex Tomas ext_debug(" border %u:%u\n", a, b); 2600a86c6181SAlex Tomas 2601d583fb87SAllison Henderson /* If this extent is beyond the end of the hole, skip it */ 26025f95d21fSLukas Czerner if (end < ex_ee_block) { 2603d23142c6SLukas Czerner /* 2604d23142c6SLukas Czerner * We're going to skip this extent and move to another, 2605f4226d9eSEric Whitney * so note that its first cluster is in use to avoid 2606f4226d9eSEric Whitney * freeing it when removing blocks. Eventually, the 2607f4226d9eSEric Whitney * right edge of the truncated/punched region will 2608f4226d9eSEric Whitney * be just to the left. 2609d23142c6SLukas Czerner */ 2610f4226d9eSEric Whitney if (sbi->s_cluster_ratio > 1) { 2611d23142c6SLukas Czerner pblk = ext4_ext_pblock(ex); 26129fe67149SEric Whitney partial->pclu = EXT4_B2C(sbi, pblk); 26139fe67149SEric Whitney partial->state = nofree; 2614f4226d9eSEric Whitney } 2615d583fb87SAllison Henderson ex--; 2616d583fb87SAllison Henderson ex_ee_block = le32_to_cpu(ex->ee_block); 2617d583fb87SAllison Henderson ex_ee_len = ext4_ext_get_actual_len(ex); 2618d583fb87SAllison Henderson continue; 2619750c9c47SDmitry Monakhov } else if (b != ex_ee_block + ex_ee_len - 1) { 2620dc1841d6SLukas Czerner EXT4_ERROR_INODE(inode, 2621dc1841d6SLukas Czerner "can not handle truncate %u:%u " 2622dc1841d6SLukas Czerner "on extent %u:%u", 2623dc1841d6SLukas Czerner start, end, ex_ee_block, 2624dc1841d6SLukas Czerner ex_ee_block + ex_ee_len - 1); 26256a797d27SDarrick J. Wong err = -EFSCORRUPTED; 2626d583fb87SAllison Henderson goto out; 2627a86c6181SAlex Tomas } else if (a != ex_ee_block) { 2628a86c6181SAlex Tomas /* remove tail of the extent */ 2629750c9c47SDmitry Monakhov num = a - ex_ee_block; 2630a86c6181SAlex Tomas } else { 2631a86c6181SAlex Tomas /* remove whole extent: excellent! */ 2632a86c6181SAlex Tomas num = 0; 2633d583fb87SAllison Henderson } 263434071da7STheodore Ts'o /* 263534071da7STheodore Ts'o * 3 for leaf, sb, and inode plus 2 (bmap and group 263634071da7STheodore Ts'o * descriptor) for each block group; assume two block 263734071da7STheodore Ts'o * groups plus ex_ee_len/blocks_per_block_group for 263834071da7STheodore Ts'o * the worst case 263934071da7STheodore Ts'o */ 264034071da7STheodore Ts'o credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb)); 2641a86c6181SAlex Tomas if (ex == EXT_FIRST_EXTENT(eh)) { 2642a86c6181SAlex Tomas correct_index = 1; 2643a86c6181SAlex Tomas credits += (ext_depth(inode)) + 1; 2644a86c6181SAlex Tomas } 26455aca07ebSDmitry Monakhov credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb); 264683448bdfSJan Kara /* 264783448bdfSJan Kara * We may end up freeing some index blocks and data from the 264883448bdfSJan Kara * punched range. Note that partial clusters are accounted for 264983448bdfSJan Kara * by ext4_free_data_revoke_credits(). 265083448bdfSJan Kara */ 265183448bdfSJan Kara revoke_credits = 265283448bdfSJan Kara ext4_free_metadata_revoke_credits(inode->i_sb, 265383448bdfSJan Kara ext_depth(inode)) + 265483448bdfSJan Kara ext4_free_data_revoke_credits(inode, b - a + 1); 2655a86c6181SAlex Tomas 2656a4130367SJan Kara err = ext4_datasem_ensure_credits(handle, inode, credits, 265783448bdfSJan Kara credits, revoke_credits); 2658a4130367SJan Kara if (err) { 2659a4130367SJan Kara if (err > 0) 2660a4130367SJan Kara err = -EAGAIN; 2661a86c6181SAlex Tomas goto out; 2662a4130367SJan Kara } 2663a86c6181SAlex Tomas 2664a86c6181SAlex Tomas err = ext4_ext_get_access(handle, inode, path + depth); 2665a86c6181SAlex Tomas if (err) 2666a86c6181SAlex Tomas goto out; 2667a86c6181SAlex Tomas 26689fe67149SEric Whitney err = ext4_remove_blocks(handle, inode, ex, partial, a, b); 2669a86c6181SAlex Tomas if (err) 2670a86c6181SAlex Tomas goto out; 2671a86c6181SAlex Tomas 2672750c9c47SDmitry Monakhov if (num == 0) 2673d0d856e8SRandy Dunlap /* this extent is removed; mark slot entirely unused */ 2674f65e6fbaSAlex Tomas ext4_ext_store_pblock(ex, 0); 2675a86c6181SAlex Tomas 2676a86c6181SAlex Tomas ex->ee_len = cpu_to_le16(num); 2677749269faSAmit Arora /* 2678556615dcSLukas Czerner * Do not mark unwritten if all the blocks in the 2679749269faSAmit Arora * extent have been removed. 2680749269faSAmit Arora */ 2681556615dcSLukas Czerner if (unwritten && num) 2682556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); 2683d583fb87SAllison Henderson /* 2684d583fb87SAllison Henderson * If the extent was completely released, 2685d583fb87SAllison Henderson * we need to remove it from the leaf 2686d583fb87SAllison Henderson */ 2687d583fb87SAllison Henderson if (num == 0) { 2688f17722f9SLukas Czerner if (end != EXT_MAX_BLOCKS - 1) { 2689d583fb87SAllison Henderson /* 2690d583fb87SAllison Henderson * For hole punching, we need to scoot all the 2691d583fb87SAllison Henderson * extents up when an extent is removed so that 2692d583fb87SAllison Henderson * we dont have blank extents in the middle 2693d583fb87SAllison Henderson */ 2694d583fb87SAllison Henderson memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) * 2695d583fb87SAllison Henderson sizeof(struct ext4_extent)); 2696d583fb87SAllison Henderson 2697d583fb87SAllison Henderson /* Now get rid of the one at the end */ 2698d583fb87SAllison Henderson memset(EXT_LAST_EXTENT(eh), 0, 2699d583fb87SAllison Henderson sizeof(struct ext4_extent)); 2700d583fb87SAllison Henderson } 2701d583fb87SAllison Henderson le16_add_cpu(&eh->eh_entries, -1); 27025bf43760SEric Whitney } 2703d583fb87SAllison Henderson 2704750c9c47SDmitry Monakhov err = ext4_ext_dirty(handle, inode, path + depth); 2705750c9c47SDmitry Monakhov if (err) 2706750c9c47SDmitry Monakhov goto out; 2707750c9c47SDmitry Monakhov 2708bf52c6f7SYongqiang Yang ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num, 2709bf89d16fSTheodore Ts'o ext4_ext_pblock(ex)); 2710a86c6181SAlex Tomas ex--; 2711a86c6181SAlex Tomas ex_ee_block = le32_to_cpu(ex->ee_block); 2712a2df2a63SAmit Arora ex_ee_len = ext4_ext_get_actual_len(ex); 2713a86c6181SAlex Tomas } 2714a86c6181SAlex Tomas 2715a86c6181SAlex Tomas if (correct_index && eh->eh_entries) 2716a86c6181SAlex Tomas err = ext4_ext_correct_indexes(handle, inode, path); 2717a86c6181SAlex Tomas 27180aa06000STheodore Ts'o /* 2719ad6599abSEric Whitney * If there's a partial cluster and at least one extent remains in 2720ad6599abSEric Whitney * the leaf, free the partial cluster if it isn't shared with the 27215bf43760SEric Whitney * current extent. If it is shared with the current extent 27229fe67149SEric Whitney * we reset the partial cluster because we've reached the start of the 27235bf43760SEric Whitney * truncated/punched region and we're done removing blocks. 27240aa06000STheodore Ts'o */ 27259fe67149SEric Whitney if (partial->state == tofree && ex >= EXT_FIRST_EXTENT(eh)) { 27265bf43760SEric Whitney pblk = ext4_ext_pblock(ex) + ex_ee_len - 1; 27279fe67149SEric Whitney if (partial->pclu != EXT4_B2C(sbi, pblk)) { 27289fe67149SEric Whitney int flags = get_default_free_blocks_flags(inode); 27299fe67149SEric Whitney 27309fe67149SEric Whitney if (ext4_is_pending(inode, partial->lblk)) 27319fe67149SEric Whitney flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER; 27320aa06000STheodore Ts'o ext4_free_blocks(handle, inode, NULL, 27339fe67149SEric Whitney EXT4_C2B(sbi, partial->pclu), 27349fe67149SEric Whitney sbi->s_cluster_ratio, flags); 27359fe67149SEric Whitney if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER) 27369fe67149SEric Whitney ext4_rereserve_cluster(inode, partial->lblk); 27375bf43760SEric Whitney } 27389fe67149SEric Whitney partial->state = initial; 27390aa06000STheodore Ts'o } 27400aa06000STheodore Ts'o 2741a86c6181SAlex Tomas /* if this leaf is free, then we should 2742a86c6181SAlex Tomas * remove it from index block above */ 2743a86c6181SAlex Tomas if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL) 2744c36575e6SForrest Liu err = ext4_ext_rm_idx(handle, inode, path, depth); 2745a86c6181SAlex Tomas 2746a86c6181SAlex Tomas out: 2747a86c6181SAlex Tomas return err; 2748a86c6181SAlex Tomas } 2749a86c6181SAlex Tomas 2750a86c6181SAlex Tomas /* 2751d0d856e8SRandy Dunlap * ext4_ext_more_to_rm: 2752d0d856e8SRandy Dunlap * returns 1 if current index has to be freed (even partial) 2753a86c6181SAlex Tomas */ 275409b88252SAvantika Mathur static int 2755a86c6181SAlex Tomas ext4_ext_more_to_rm(struct ext4_ext_path *path) 2756a86c6181SAlex Tomas { 2757a86c6181SAlex Tomas BUG_ON(path->p_idx == NULL); 2758a86c6181SAlex Tomas 2759a86c6181SAlex Tomas if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr)) 2760a86c6181SAlex Tomas return 0; 2761a86c6181SAlex Tomas 2762a86c6181SAlex Tomas /* 2763d0d856e8SRandy Dunlap * if truncate on deeper level happened, it wasn't partial, 2764a86c6181SAlex Tomas * so we have to consider current index for truncation 2765a86c6181SAlex Tomas */ 2766a86c6181SAlex Tomas if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block) 2767a86c6181SAlex Tomas return 0; 2768a86c6181SAlex Tomas return 1; 2769a86c6181SAlex Tomas } 2770a86c6181SAlex Tomas 277126a4c0c6STheodore Ts'o int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start, 27725f95d21fSLukas Czerner ext4_lblk_t end) 2773a86c6181SAlex Tomas { 2774f4226d9eSEric Whitney struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 2775a86c6181SAlex Tomas int depth = ext_depth(inode); 2776968dee77SAshish Sangwan struct ext4_ext_path *path = NULL; 27779fe67149SEric Whitney struct partial_cluster partial; 2778a86c6181SAlex Tomas handle_t *handle; 27796f2080e6SDmitry Monakhov int i = 0, err = 0; 2780a86c6181SAlex Tomas 27819fe67149SEric Whitney partial.pclu = 0; 27829fe67149SEric Whitney partial.lblk = 0; 27839fe67149SEric Whitney partial.state = initial; 27849fe67149SEric Whitney 27855f95d21fSLukas Czerner ext_debug("truncate since %u to %u\n", start, end); 2786a86c6181SAlex Tomas 2787a86c6181SAlex Tomas /* probably first extent we're gonna free will be last in block */ 278883448bdfSJan Kara handle = ext4_journal_start_with_revoke(inode, EXT4_HT_TRUNCATE, 278983448bdfSJan Kara depth + 1, 279083448bdfSJan Kara ext4_free_metadata_revoke_credits(inode->i_sb, depth)); 2791a86c6181SAlex Tomas if (IS_ERR(handle)) 2792a86c6181SAlex Tomas return PTR_ERR(handle); 2793a86c6181SAlex Tomas 27940617b83fSDmitry Monakhov again: 279561801325SLukas Czerner trace_ext4_ext_remove_space(inode, start, end, depth); 2796d8990240SAditya Kali 2797a86c6181SAlex Tomas /* 27985f95d21fSLukas Czerner * Check if we are removing extents inside the extent tree. If that 27995f95d21fSLukas Czerner * is the case, we are going to punch a hole inside the extent tree 28005f95d21fSLukas Czerner * so we have to check whether we need to split the extent covering 28015f95d21fSLukas Czerner * the last block to remove so we can easily remove the part of it 28025f95d21fSLukas Czerner * in ext4_ext_rm_leaf(). 28035f95d21fSLukas Czerner */ 28045f95d21fSLukas Czerner if (end < EXT_MAX_BLOCKS - 1) { 28055f95d21fSLukas Czerner struct ext4_extent *ex; 2806f4226d9eSEric Whitney ext4_lblk_t ee_block, ex_end, lblk; 2807f4226d9eSEric Whitney ext4_fsblk_t pblk; 28085f95d21fSLukas Czerner 2809f4226d9eSEric Whitney /* find extent for or closest extent to this block */ 281073c384c0STheodore Ts'o path = ext4_find_extent(inode, end, NULL, 281173c384c0STheodore Ts'o EXT4_EX_NOCACHE | EXT4_EX_NOFAIL); 28125f95d21fSLukas Czerner if (IS_ERR(path)) { 28135f95d21fSLukas Czerner ext4_journal_stop(handle); 28145f95d21fSLukas Czerner return PTR_ERR(path); 28155f95d21fSLukas Czerner } 28165f95d21fSLukas Czerner depth = ext_depth(inode); 28176f2080e6SDmitry Monakhov /* Leaf not may not exist only if inode has no blocks at all */ 28185f95d21fSLukas Czerner ex = path[depth].p_ext; 2819968dee77SAshish Sangwan if (!ex) { 28206f2080e6SDmitry Monakhov if (depth) { 28216f2080e6SDmitry Monakhov EXT4_ERROR_INODE(inode, 28226f2080e6SDmitry Monakhov "path[%d].p_hdr == NULL", 28236f2080e6SDmitry Monakhov depth); 28246a797d27SDarrick J. Wong err = -EFSCORRUPTED; 28256f2080e6SDmitry Monakhov } 28266f2080e6SDmitry Monakhov goto out; 2827968dee77SAshish Sangwan } 28285f95d21fSLukas Czerner 28295f95d21fSLukas Czerner ee_block = le32_to_cpu(ex->ee_block); 2830f4226d9eSEric Whitney ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1; 28315f95d21fSLukas Czerner 28325f95d21fSLukas Czerner /* 28335f95d21fSLukas Czerner * See if the last block is inside the extent, if so split 28345f95d21fSLukas Czerner * the extent at 'end' block so we can easily remove the 28355f95d21fSLukas Czerner * tail of the first part of the split extent in 28365f95d21fSLukas Czerner * ext4_ext_rm_leaf(). 28375f95d21fSLukas Czerner */ 2838f4226d9eSEric Whitney if (end >= ee_block && end < ex_end) { 2839f4226d9eSEric Whitney 2840f4226d9eSEric Whitney /* 2841f4226d9eSEric Whitney * If we're going to split the extent, note that 2842f4226d9eSEric Whitney * the cluster containing the block after 'end' is 2843f4226d9eSEric Whitney * in use to avoid freeing it when removing blocks. 2844f4226d9eSEric Whitney */ 2845f4226d9eSEric Whitney if (sbi->s_cluster_ratio > 1) { 2846f4226d9eSEric Whitney pblk = ext4_ext_pblock(ex) + end - ee_block + 2; 28479fe67149SEric Whitney partial.pclu = EXT4_B2C(sbi, pblk); 28489fe67149SEric Whitney partial.state = nofree; 2849f4226d9eSEric Whitney } 2850f4226d9eSEric Whitney 28515f95d21fSLukas Czerner /* 28525f95d21fSLukas Czerner * Split the extent in two so that 'end' is the last 285327dd4385SLukas Czerner * block in the first new extent. Also we should not 285427dd4385SLukas Czerner * fail removing space due to ENOSPC so try to use 285527dd4385SLukas Czerner * reserved block if that happens. 28565f95d21fSLukas Czerner */ 2857dfe50809STheodore Ts'o err = ext4_force_split_extent_at(handle, inode, &path, 2858fcf6b1b7SDmitry Monakhov end + 1, 1); 28595f95d21fSLukas Czerner if (err < 0) 28605f95d21fSLukas Czerner goto out; 2861f4226d9eSEric Whitney 28627bd75230SEric Whitney } else if (sbi->s_cluster_ratio > 1 && end >= ex_end && 28637bd75230SEric Whitney partial.state == initial) { 2864f4226d9eSEric Whitney /* 28657bd75230SEric Whitney * If we're punching, there's an extent to the right. 28667bd75230SEric Whitney * If the partial cluster hasn't been set, set it to 28677bd75230SEric Whitney * that extent's first cluster and its state to nofree 28687bd75230SEric Whitney * so it won't be freed should it contain blocks to be 28697bd75230SEric Whitney * removed. If it's already set (tofree/nofree), we're 28707bd75230SEric Whitney * retrying and keep the original partial cluster info 28717bd75230SEric Whitney * so a cluster marked tofree as a result of earlier 28727bd75230SEric Whitney * extent removal is not lost. 2873f4226d9eSEric Whitney */ 2874f4226d9eSEric Whitney lblk = ex_end + 1; 2875f4226d9eSEric Whitney err = ext4_ext_search_right(inode, path, &lblk, &pblk, 2876f4226d9eSEric Whitney &ex); 2877f4226d9eSEric Whitney if (err) 2878f4226d9eSEric Whitney goto out; 28799fe67149SEric Whitney if (pblk) { 28809fe67149SEric Whitney partial.pclu = EXT4_B2C(sbi, pblk); 28819fe67149SEric Whitney partial.state = nofree; 28829fe67149SEric Whitney } 28835f95d21fSLukas Czerner } 28845f95d21fSLukas Czerner } 28855f95d21fSLukas Czerner /* 2886d0d856e8SRandy Dunlap * We start scanning from right side, freeing all the blocks 2887d0d856e8SRandy Dunlap * after i_size and walking into the tree depth-wise. 2888a86c6181SAlex Tomas */ 28890617b83fSDmitry Monakhov depth = ext_depth(inode); 2890968dee77SAshish Sangwan if (path) { 2891968dee77SAshish Sangwan int k = i = depth; 2892968dee77SAshish Sangwan while (--k > 0) 2893968dee77SAshish Sangwan path[k].p_block = 2894968dee77SAshish Sangwan le16_to_cpu(path[k].p_hdr->eh_entries)+1; 2895968dee77SAshish Sangwan } else { 28966396bb22SKees Cook path = kcalloc(depth + 1, sizeof(struct ext4_ext_path), 289773c384c0STheodore Ts'o GFP_NOFS | __GFP_NOFAIL); 2898a86c6181SAlex Tomas if (path == NULL) { 2899a86c6181SAlex Tomas ext4_journal_stop(handle); 2900a86c6181SAlex Tomas return -ENOMEM; 2901a86c6181SAlex Tomas } 290210809df8STheodore Ts'o path[0].p_maxdepth = path[0].p_depth = depth; 2903a86c6181SAlex Tomas path[0].p_hdr = ext_inode_hdr(inode); 290489a4e48fSTheodore Ts'o i = 0; 29055f95d21fSLukas Czerner 2906c349179bSTheodore Ts'o if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) { 29076a797d27SDarrick J. Wong err = -EFSCORRUPTED; 2908a86c6181SAlex Tomas goto out; 2909a86c6181SAlex Tomas } 2910968dee77SAshish Sangwan } 2911968dee77SAshish Sangwan err = 0; 2912a86c6181SAlex Tomas 2913a86c6181SAlex Tomas while (i >= 0 && err == 0) { 2914a86c6181SAlex Tomas if (i == depth) { 2915a86c6181SAlex Tomas /* this is leaf block */ 2916d583fb87SAllison Henderson err = ext4_ext_rm_leaf(handle, inode, path, 29179fe67149SEric Whitney &partial, start, end); 2918d0d856e8SRandy Dunlap /* root level has p_bh == NULL, brelse() eats this */ 2919a86c6181SAlex Tomas brelse(path[i].p_bh); 2920a86c6181SAlex Tomas path[i].p_bh = NULL; 2921a86c6181SAlex Tomas i--; 2922a86c6181SAlex Tomas continue; 2923a86c6181SAlex Tomas } 2924a86c6181SAlex Tomas 2925a86c6181SAlex Tomas /* this is index block */ 2926a86c6181SAlex Tomas if (!path[i].p_hdr) { 2927a86c6181SAlex Tomas ext_debug("initialize header\n"); 2928a86c6181SAlex Tomas path[i].p_hdr = ext_block_hdr(path[i].p_bh); 2929a86c6181SAlex Tomas } 2930a86c6181SAlex Tomas 2931a86c6181SAlex Tomas if (!path[i].p_idx) { 2932d0d856e8SRandy Dunlap /* this level hasn't been touched yet */ 2933a86c6181SAlex Tomas path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr); 2934a86c6181SAlex Tomas path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1; 2935a86c6181SAlex Tomas ext_debug("init index ptr: hdr 0x%p, num %d\n", 2936a86c6181SAlex Tomas path[i].p_hdr, 2937a86c6181SAlex Tomas le16_to_cpu(path[i].p_hdr->eh_entries)); 2938a86c6181SAlex Tomas } else { 2939d0d856e8SRandy Dunlap /* we were already here, see at next index */ 2940a86c6181SAlex Tomas path[i].p_idx--; 2941a86c6181SAlex Tomas } 2942a86c6181SAlex Tomas 2943a86c6181SAlex Tomas ext_debug("level %d - index, first 0x%p, cur 0x%p\n", 2944a86c6181SAlex Tomas i, EXT_FIRST_INDEX(path[i].p_hdr), 2945a86c6181SAlex Tomas path[i].p_idx); 2946a86c6181SAlex Tomas if (ext4_ext_more_to_rm(path + i)) { 2947c29c0ae7SAlex Tomas struct buffer_head *bh; 2948a86c6181SAlex Tomas /* go to the next level */ 29492ae02107SMingming Cao ext_debug("move to level %d (block %llu)\n", 2950bf89d16fSTheodore Ts'o i + 1, ext4_idx_pblock(path[i].p_idx)); 2951a86c6181SAlex Tomas memset(path + i + 1, 0, sizeof(*path)); 29527d7ea89eSTheodore Ts'o bh = read_extent_tree_block(inode, 2953107a7bd3STheodore Ts'o ext4_idx_pblock(path[i].p_idx), depth - i - 1, 2954107a7bd3STheodore Ts'o EXT4_EX_NOCACHE); 29557d7ea89eSTheodore Ts'o if (IS_ERR(bh)) { 2956a86c6181SAlex Tomas /* should we reset i_size? */ 29577d7ea89eSTheodore Ts'o err = PTR_ERR(bh); 2958a86c6181SAlex Tomas break; 2959a86c6181SAlex Tomas } 296076828c88STheodore Ts'o /* Yield here to deal with large extent trees. 296176828c88STheodore Ts'o * Should be a no-op if we did IO above. */ 296276828c88STheodore Ts'o cond_resched(); 2963c29c0ae7SAlex Tomas if (WARN_ON(i + 1 > depth)) { 29646a797d27SDarrick J. Wong err = -EFSCORRUPTED; 2965c29c0ae7SAlex Tomas break; 2966c29c0ae7SAlex Tomas } 2967c29c0ae7SAlex Tomas path[i + 1].p_bh = bh; 2968a86c6181SAlex Tomas 2969d0d856e8SRandy Dunlap /* save actual number of indexes since this 2970d0d856e8SRandy Dunlap * number is changed at the next iteration */ 2971a86c6181SAlex Tomas path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries); 2972a86c6181SAlex Tomas i++; 2973a86c6181SAlex Tomas } else { 2974d0d856e8SRandy Dunlap /* we finished processing this index, go up */ 2975a86c6181SAlex Tomas if (path[i].p_hdr->eh_entries == 0 && i > 0) { 2976d0d856e8SRandy Dunlap /* index is empty, remove it; 2977a86c6181SAlex Tomas * handle must be already prepared by the 2978a86c6181SAlex Tomas * truncatei_leaf() */ 2979c36575e6SForrest Liu err = ext4_ext_rm_idx(handle, inode, path, i); 2980a86c6181SAlex Tomas } 2981d0d856e8SRandy Dunlap /* root level has p_bh == NULL, brelse() eats this */ 2982a86c6181SAlex Tomas brelse(path[i].p_bh); 2983a86c6181SAlex Tomas path[i].p_bh = NULL; 2984a86c6181SAlex Tomas i--; 2985a86c6181SAlex Tomas ext_debug("return to level %d\n", i); 2986a86c6181SAlex Tomas } 2987a86c6181SAlex Tomas } 2988a86c6181SAlex Tomas 29899fe67149SEric Whitney trace_ext4_ext_remove_space_done(inode, start, end, depth, &partial, 29909fe67149SEric Whitney path->p_hdr->eh_entries); 2991d8990240SAditya Kali 29920756b908SEric Whitney /* 29939fe67149SEric Whitney * if there's a partial cluster and we have removed the first extent 29949fe67149SEric Whitney * in the file, then we also free the partial cluster, if any 29950756b908SEric Whitney */ 29969fe67149SEric Whitney if (partial.state == tofree && err == 0) { 29979fe67149SEric Whitney int flags = get_default_free_blocks_flags(inode); 29989fe67149SEric Whitney 29999fe67149SEric Whitney if (ext4_is_pending(inode, partial.lblk)) 30009fe67149SEric Whitney flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER; 30017b415bf6SAditya Kali ext4_free_blocks(handle, inode, NULL, 30029fe67149SEric Whitney EXT4_C2B(sbi, partial.pclu), 30039fe67149SEric Whitney sbi->s_cluster_ratio, flags); 30049fe67149SEric Whitney if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER) 30059fe67149SEric Whitney ext4_rereserve_cluster(inode, partial.lblk); 30069fe67149SEric Whitney partial.state = initial; 30077b415bf6SAditya Kali } 30087b415bf6SAditya Kali 3009a86c6181SAlex Tomas /* TODO: flexible tree reduction should be here */ 3010a86c6181SAlex Tomas if (path->p_hdr->eh_entries == 0) { 3011a86c6181SAlex Tomas /* 3012d0d856e8SRandy Dunlap * truncate to zero freed all the tree, 3013d0d856e8SRandy Dunlap * so we need to correct eh_depth 3014a86c6181SAlex Tomas */ 3015a86c6181SAlex Tomas err = ext4_ext_get_access(handle, inode, path); 3016a86c6181SAlex Tomas if (err == 0) { 3017a86c6181SAlex Tomas ext_inode_hdr(inode)->eh_depth = 0; 3018a86c6181SAlex Tomas ext_inode_hdr(inode)->eh_max = 301955ad63bfSTheodore Ts'o cpu_to_le16(ext4_ext_space_root(inode, 0)); 3020a86c6181SAlex Tomas err = ext4_ext_dirty(handle, inode, path); 3021a86c6181SAlex Tomas } 3022a86c6181SAlex Tomas } 3023a86c6181SAlex Tomas out: 3024a86c6181SAlex Tomas ext4_ext_drop_refs(path); 3025a86c6181SAlex Tomas kfree(path); 3026968dee77SAshish Sangwan path = NULL; 3027dfe50809STheodore Ts'o if (err == -EAGAIN) 3028dfe50809STheodore Ts'o goto again; 3029a86c6181SAlex Tomas ext4_journal_stop(handle); 3030a86c6181SAlex Tomas 3031a86c6181SAlex Tomas return err; 3032a86c6181SAlex Tomas } 3033a86c6181SAlex Tomas 3034a86c6181SAlex Tomas /* 3035a86c6181SAlex Tomas * called at mount time 3036a86c6181SAlex Tomas */ 3037a86c6181SAlex Tomas void ext4_ext_init(struct super_block *sb) 3038a86c6181SAlex Tomas { 3039a86c6181SAlex Tomas /* 3040a86c6181SAlex Tomas * possible initialization would be here 3041a86c6181SAlex Tomas */ 3042a86c6181SAlex Tomas 3043e2b911c5SDarrick J. Wong if (ext4_has_feature_extents(sb)) { 304490576c0bSTheodore Ts'o #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS) 304592b97816STheodore Ts'o printk(KERN_INFO "EXT4-fs: file extents enabled" 3046bbf2f9fbSRobert P. J. Day #ifdef AGGRESSIVE_TEST 304792b97816STheodore Ts'o ", aggressive tests" 3048a86c6181SAlex Tomas #endif 3049a86c6181SAlex Tomas #ifdef CHECK_BINSEARCH 305092b97816STheodore Ts'o ", check binsearch" 3051a86c6181SAlex Tomas #endif 3052a86c6181SAlex Tomas #ifdef EXTENTS_STATS 305392b97816STheodore Ts'o ", stats" 3054a86c6181SAlex Tomas #endif 305592b97816STheodore Ts'o "\n"); 305690576c0bSTheodore Ts'o #endif 3057a86c6181SAlex Tomas #ifdef EXTENTS_STATS 3058a86c6181SAlex Tomas spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock); 3059a86c6181SAlex Tomas EXT4_SB(sb)->s_ext_min = 1 << 30; 3060a86c6181SAlex Tomas EXT4_SB(sb)->s_ext_max = 0; 3061a86c6181SAlex Tomas #endif 3062a86c6181SAlex Tomas } 3063a86c6181SAlex Tomas } 3064a86c6181SAlex Tomas 3065a86c6181SAlex Tomas /* 3066a86c6181SAlex Tomas * called at umount time 3067a86c6181SAlex Tomas */ 3068a86c6181SAlex Tomas void ext4_ext_release(struct super_block *sb) 3069a86c6181SAlex Tomas { 3070e2b911c5SDarrick J. Wong if (!ext4_has_feature_extents(sb)) 3071a86c6181SAlex Tomas return; 3072a86c6181SAlex Tomas 3073a86c6181SAlex Tomas #ifdef EXTENTS_STATS 3074a86c6181SAlex Tomas if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) { 3075a86c6181SAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3076a86c6181SAlex Tomas printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n", 3077a86c6181SAlex Tomas sbi->s_ext_blocks, sbi->s_ext_extents, 3078a86c6181SAlex Tomas sbi->s_ext_blocks / sbi->s_ext_extents); 3079a86c6181SAlex Tomas printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n", 3080a86c6181SAlex Tomas sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max); 3081a86c6181SAlex Tomas } 3082a86c6181SAlex Tomas #endif 3083a86c6181SAlex Tomas } 3084a86c6181SAlex Tomas 3085d7b2a00cSZheng Liu static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex) 3086d7b2a00cSZheng Liu { 3087d7b2a00cSZheng Liu ext4_lblk_t ee_block; 3088d7b2a00cSZheng Liu ext4_fsblk_t ee_pblock; 3089d7b2a00cSZheng Liu unsigned int ee_len; 3090d7b2a00cSZheng Liu 3091d7b2a00cSZheng Liu ee_block = le32_to_cpu(ex->ee_block); 3092d7b2a00cSZheng Liu ee_len = ext4_ext_get_actual_len(ex); 3093d7b2a00cSZheng Liu ee_pblock = ext4_ext_pblock(ex); 3094d7b2a00cSZheng Liu 3095d7b2a00cSZheng Liu if (ee_len == 0) 3096d7b2a00cSZheng Liu return 0; 3097d7b2a00cSZheng Liu 3098d7b2a00cSZheng Liu return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock, 3099d7b2a00cSZheng Liu EXTENT_STATUS_WRITTEN); 3100d7b2a00cSZheng Liu } 3101d7b2a00cSZheng Liu 3102093a088bSAneesh Kumar K.V /* FIXME!! we need to try to merge to left or right after zero-out */ 3103093a088bSAneesh Kumar K.V static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex) 3104093a088bSAneesh Kumar K.V { 31052407518dSLukas Czerner ext4_fsblk_t ee_pblock; 31062407518dSLukas Czerner unsigned int ee_len; 3107093a088bSAneesh Kumar K.V 3108093a088bSAneesh Kumar K.V ee_len = ext4_ext_get_actual_len(ex); 3109bf89d16fSTheodore Ts'o ee_pblock = ext4_ext_pblock(ex); 311053085facSJan Kara return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock, 311153085facSJan Kara ee_len); 3112093a088bSAneesh Kumar K.V } 3113093a088bSAneesh Kumar K.V 311447ea3bb5SYongqiang Yang /* 311547ea3bb5SYongqiang Yang * ext4_split_extent_at() splits an extent at given block. 311647ea3bb5SYongqiang Yang * 311747ea3bb5SYongqiang Yang * @handle: the journal handle 311847ea3bb5SYongqiang Yang * @inode: the file inode 311947ea3bb5SYongqiang Yang * @path: the path to the extent 312047ea3bb5SYongqiang Yang * @split: the logical block where the extent is splitted. 312147ea3bb5SYongqiang Yang * @split_flags: indicates if the extent could be zeroout if split fails, and 3122556615dcSLukas Czerner * the states(init or unwritten) of new extents. 312347ea3bb5SYongqiang Yang * @flags: flags used to insert new extent to extent tree. 312447ea3bb5SYongqiang Yang * 312547ea3bb5SYongqiang Yang * 312647ea3bb5SYongqiang Yang * Splits extent [a, b] into two extents [a, @split) and [@split, b], states 312747ea3bb5SYongqiang Yang * of which are deterimined by split_flag. 312847ea3bb5SYongqiang Yang * 312947ea3bb5SYongqiang Yang * There are two cases: 313047ea3bb5SYongqiang Yang * a> the extent are splitted into two extent. 313147ea3bb5SYongqiang Yang * b> split is not needed, and just mark the extent. 313247ea3bb5SYongqiang Yang * 313347ea3bb5SYongqiang Yang * return 0 on success. 313447ea3bb5SYongqiang Yang */ 313547ea3bb5SYongqiang Yang static int ext4_split_extent_at(handle_t *handle, 313647ea3bb5SYongqiang Yang struct inode *inode, 3137dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 313847ea3bb5SYongqiang Yang ext4_lblk_t split, 313947ea3bb5SYongqiang Yang int split_flag, 314047ea3bb5SYongqiang Yang int flags) 314147ea3bb5SYongqiang Yang { 3142dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 314347ea3bb5SYongqiang Yang ext4_fsblk_t newblock; 314447ea3bb5SYongqiang Yang ext4_lblk_t ee_block; 3145adb23551SZheng Liu struct ext4_extent *ex, newex, orig_ex, zero_ex; 314647ea3bb5SYongqiang Yang struct ext4_extent *ex2 = NULL; 314747ea3bb5SYongqiang Yang unsigned int ee_len, depth; 314847ea3bb5SYongqiang Yang int err = 0; 314947ea3bb5SYongqiang Yang 3150dee1f973SDmitry Monakhov BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) == 3151dee1f973SDmitry Monakhov (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)); 3152dee1f973SDmitry Monakhov 315347ea3bb5SYongqiang Yang ext_debug("ext4_split_extents_at: inode %lu, logical" 315447ea3bb5SYongqiang Yang "block %llu\n", inode->i_ino, (unsigned long long)split); 315547ea3bb5SYongqiang Yang 315647ea3bb5SYongqiang Yang ext4_ext_show_leaf(inode, path); 315747ea3bb5SYongqiang Yang 315847ea3bb5SYongqiang Yang depth = ext_depth(inode); 315947ea3bb5SYongqiang Yang ex = path[depth].p_ext; 316047ea3bb5SYongqiang Yang ee_block = le32_to_cpu(ex->ee_block); 316147ea3bb5SYongqiang Yang ee_len = ext4_ext_get_actual_len(ex); 316247ea3bb5SYongqiang Yang newblock = split - ee_block + ext4_ext_pblock(ex); 316347ea3bb5SYongqiang Yang 316447ea3bb5SYongqiang Yang BUG_ON(split < ee_block || split >= (ee_block + ee_len)); 3165556615dcSLukas Czerner BUG_ON(!ext4_ext_is_unwritten(ex) && 3166357b66fdSDmitry Monakhov split_flag & (EXT4_EXT_MAY_ZEROOUT | 3167556615dcSLukas Czerner EXT4_EXT_MARK_UNWRIT1 | 3168556615dcSLukas Czerner EXT4_EXT_MARK_UNWRIT2)); 316947ea3bb5SYongqiang Yang 317047ea3bb5SYongqiang Yang err = ext4_ext_get_access(handle, inode, path + depth); 317147ea3bb5SYongqiang Yang if (err) 317247ea3bb5SYongqiang Yang goto out; 317347ea3bb5SYongqiang Yang 317447ea3bb5SYongqiang Yang if (split == ee_block) { 317547ea3bb5SYongqiang Yang /* 317647ea3bb5SYongqiang Yang * case b: block @split is the block that the extent begins with 317747ea3bb5SYongqiang Yang * then we just change the state of the extent, and splitting 317847ea3bb5SYongqiang Yang * is not needed. 317947ea3bb5SYongqiang Yang */ 3180556615dcSLukas Czerner if (split_flag & EXT4_EXT_MARK_UNWRIT2) 3181556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); 318247ea3bb5SYongqiang Yang else 318347ea3bb5SYongqiang Yang ext4_ext_mark_initialized(ex); 318447ea3bb5SYongqiang Yang 318547ea3bb5SYongqiang Yang if (!(flags & EXT4_GET_BLOCKS_PRE_IO)) 3186ecb94f5fSTheodore Ts'o ext4_ext_try_to_merge(handle, inode, path, ex); 318747ea3bb5SYongqiang Yang 3188ecb94f5fSTheodore Ts'o err = ext4_ext_dirty(handle, inode, path + path->p_depth); 318947ea3bb5SYongqiang Yang goto out; 319047ea3bb5SYongqiang Yang } 319147ea3bb5SYongqiang Yang 319247ea3bb5SYongqiang Yang /* case a */ 319347ea3bb5SYongqiang Yang memcpy(&orig_ex, ex, sizeof(orig_ex)); 319447ea3bb5SYongqiang Yang ex->ee_len = cpu_to_le16(split - ee_block); 3195556615dcSLukas Czerner if (split_flag & EXT4_EXT_MARK_UNWRIT1) 3196556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); 319747ea3bb5SYongqiang Yang 319847ea3bb5SYongqiang Yang /* 319947ea3bb5SYongqiang Yang * path may lead to new leaf, not to original leaf any more 320047ea3bb5SYongqiang Yang * after ext4_ext_insert_extent() returns, 320147ea3bb5SYongqiang Yang */ 320247ea3bb5SYongqiang Yang err = ext4_ext_dirty(handle, inode, path + depth); 320347ea3bb5SYongqiang Yang if (err) 320447ea3bb5SYongqiang Yang goto fix_extent_len; 320547ea3bb5SYongqiang Yang 320647ea3bb5SYongqiang Yang ex2 = &newex; 320747ea3bb5SYongqiang Yang ex2->ee_block = cpu_to_le32(split); 320847ea3bb5SYongqiang Yang ex2->ee_len = cpu_to_le16(ee_len - (split - ee_block)); 320947ea3bb5SYongqiang Yang ext4_ext_store_pblock(ex2, newblock); 3210556615dcSLukas Czerner if (split_flag & EXT4_EXT_MARK_UNWRIT2) 3211556615dcSLukas Czerner ext4_ext_mark_unwritten(ex2); 321247ea3bb5SYongqiang Yang 3213dfe50809STheodore Ts'o err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags); 321447ea3bb5SYongqiang Yang if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) { 3215dee1f973SDmitry Monakhov if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) { 3216adb23551SZheng Liu if (split_flag & EXT4_EXT_DATA_VALID1) { 3217dee1f973SDmitry Monakhov err = ext4_ext_zeroout(inode, ex2); 3218adb23551SZheng Liu zero_ex.ee_block = ex2->ee_block; 32198cde7ad1SZheng Liu zero_ex.ee_len = cpu_to_le16( 32208cde7ad1SZheng Liu ext4_ext_get_actual_len(ex2)); 3221adb23551SZheng Liu ext4_ext_store_pblock(&zero_ex, 3222adb23551SZheng Liu ext4_ext_pblock(ex2)); 3223adb23551SZheng Liu } else { 3224dee1f973SDmitry Monakhov err = ext4_ext_zeroout(inode, ex); 3225adb23551SZheng Liu zero_ex.ee_block = ex->ee_block; 32268cde7ad1SZheng Liu zero_ex.ee_len = cpu_to_le16( 32278cde7ad1SZheng Liu ext4_ext_get_actual_len(ex)); 3228adb23551SZheng Liu ext4_ext_store_pblock(&zero_ex, 3229adb23551SZheng Liu ext4_ext_pblock(ex)); 3230adb23551SZheng Liu } 3231adb23551SZheng Liu } else { 323247ea3bb5SYongqiang Yang err = ext4_ext_zeroout(inode, &orig_ex); 3233adb23551SZheng Liu zero_ex.ee_block = orig_ex.ee_block; 32348cde7ad1SZheng Liu zero_ex.ee_len = cpu_to_le16( 32358cde7ad1SZheng Liu ext4_ext_get_actual_len(&orig_ex)); 3236adb23551SZheng Liu ext4_ext_store_pblock(&zero_ex, 3237adb23551SZheng Liu ext4_ext_pblock(&orig_ex)); 3238adb23551SZheng Liu } 3239dee1f973SDmitry Monakhov 324047ea3bb5SYongqiang Yang if (err) 324147ea3bb5SYongqiang Yang goto fix_extent_len; 324247ea3bb5SYongqiang Yang /* update the extent length and mark as initialized */ 3243af1584f5SAl Viro ex->ee_len = cpu_to_le16(ee_len); 3244ecb94f5fSTheodore Ts'o ext4_ext_try_to_merge(handle, inode, path, ex); 3245ecb94f5fSTheodore Ts'o err = ext4_ext_dirty(handle, inode, path + path->p_depth); 3246adb23551SZheng Liu if (err) 3247adb23551SZheng Liu goto fix_extent_len; 3248adb23551SZheng Liu 3249adb23551SZheng Liu /* update extent status tree */ 3250d7b2a00cSZheng Liu err = ext4_zeroout_es(inode, &zero_ex); 3251adb23551SZheng Liu 325247ea3bb5SYongqiang Yang goto out; 325347ea3bb5SYongqiang Yang } else if (err) 325447ea3bb5SYongqiang Yang goto fix_extent_len; 325547ea3bb5SYongqiang Yang 325647ea3bb5SYongqiang Yang out: 325747ea3bb5SYongqiang Yang ext4_ext_show_leaf(inode, path); 325847ea3bb5SYongqiang Yang return err; 325947ea3bb5SYongqiang Yang 326047ea3bb5SYongqiang Yang fix_extent_len: 326147ea3bb5SYongqiang Yang ex->ee_len = orig_ex.ee_len; 3262b60ca334SHarshad Shirwadkar /* 3263b60ca334SHarshad Shirwadkar * Ignore ext4_ext_dirty return value since we are already in error path 3264b60ca334SHarshad Shirwadkar * and err is a non-zero error code. 3265b60ca334SHarshad Shirwadkar */ 326629faed16SDmitry Monakhov ext4_ext_dirty(handle, inode, path + path->p_depth); 326747ea3bb5SYongqiang Yang return err; 326847ea3bb5SYongqiang Yang } 326947ea3bb5SYongqiang Yang 327047ea3bb5SYongqiang Yang /* 327147ea3bb5SYongqiang Yang * ext4_split_extents() splits an extent and mark extent which is covered 327247ea3bb5SYongqiang Yang * by @map as split_flags indicates 327347ea3bb5SYongqiang Yang * 327447ea3bb5SYongqiang Yang * It may result in splitting the extent into multiple extents (up to three) 327547ea3bb5SYongqiang Yang * There are three possibilities: 327647ea3bb5SYongqiang Yang * a> There is no split required 327747ea3bb5SYongqiang Yang * b> Splits in two extents: Split is happening at either end of the extent 327847ea3bb5SYongqiang Yang * c> Splits in three extents: Somone is splitting in middle of the extent 327947ea3bb5SYongqiang Yang * 328047ea3bb5SYongqiang Yang */ 328147ea3bb5SYongqiang Yang static int ext4_split_extent(handle_t *handle, 328247ea3bb5SYongqiang Yang struct inode *inode, 3283dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 328447ea3bb5SYongqiang Yang struct ext4_map_blocks *map, 328547ea3bb5SYongqiang Yang int split_flag, 328647ea3bb5SYongqiang Yang int flags) 328747ea3bb5SYongqiang Yang { 3288dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 328947ea3bb5SYongqiang Yang ext4_lblk_t ee_block; 329047ea3bb5SYongqiang Yang struct ext4_extent *ex; 329147ea3bb5SYongqiang Yang unsigned int ee_len, depth; 329247ea3bb5SYongqiang Yang int err = 0; 3293556615dcSLukas Czerner int unwritten; 329447ea3bb5SYongqiang Yang int split_flag1, flags1; 32953a225670SZheng Liu int allocated = map->m_len; 329647ea3bb5SYongqiang Yang 329747ea3bb5SYongqiang Yang depth = ext_depth(inode); 329847ea3bb5SYongqiang Yang ex = path[depth].p_ext; 329947ea3bb5SYongqiang Yang ee_block = le32_to_cpu(ex->ee_block); 330047ea3bb5SYongqiang Yang ee_len = ext4_ext_get_actual_len(ex); 3301556615dcSLukas Czerner unwritten = ext4_ext_is_unwritten(ex); 330247ea3bb5SYongqiang Yang 330347ea3bb5SYongqiang Yang if (map->m_lblk + map->m_len < ee_block + ee_len) { 3304dee1f973SDmitry Monakhov split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT; 330547ea3bb5SYongqiang Yang flags1 = flags | EXT4_GET_BLOCKS_PRE_IO; 3306556615dcSLukas Czerner if (unwritten) 3307556615dcSLukas Czerner split_flag1 |= EXT4_EXT_MARK_UNWRIT1 | 3308556615dcSLukas Czerner EXT4_EXT_MARK_UNWRIT2; 3309dee1f973SDmitry Monakhov if (split_flag & EXT4_EXT_DATA_VALID2) 3310dee1f973SDmitry Monakhov split_flag1 |= EXT4_EXT_DATA_VALID1; 3311dfe50809STheodore Ts'o err = ext4_split_extent_at(handle, inode, ppath, 331247ea3bb5SYongqiang Yang map->m_lblk + map->m_len, split_flag1, flags1); 331393917411SYongqiang Yang if (err) 331493917411SYongqiang Yang goto out; 33153a225670SZheng Liu } else { 33163a225670SZheng Liu allocated = ee_len - (map->m_lblk - ee_block); 331747ea3bb5SYongqiang Yang } 3318357b66fdSDmitry Monakhov /* 3319357b66fdSDmitry Monakhov * Update path is required because previous ext4_split_extent_at() may 3320357b66fdSDmitry Monakhov * result in split of original leaf or extent zeroout. 3321357b66fdSDmitry Monakhov */ 332273c384c0STheodore Ts'o path = ext4_find_extent(inode, map->m_lblk, ppath, flags); 332347ea3bb5SYongqiang Yang if (IS_ERR(path)) 332447ea3bb5SYongqiang Yang return PTR_ERR(path); 3325357b66fdSDmitry Monakhov depth = ext_depth(inode); 3326357b66fdSDmitry Monakhov ex = path[depth].p_ext; 3327a18ed359SDmitry Monakhov if (!ex) { 3328a18ed359SDmitry Monakhov EXT4_ERROR_INODE(inode, "unexpected hole at %lu", 3329a18ed359SDmitry Monakhov (unsigned long) map->m_lblk); 33306a797d27SDarrick J. Wong return -EFSCORRUPTED; 3331a18ed359SDmitry Monakhov } 3332556615dcSLukas Czerner unwritten = ext4_ext_is_unwritten(ex); 3333357b66fdSDmitry Monakhov split_flag1 = 0; 333447ea3bb5SYongqiang Yang 333547ea3bb5SYongqiang Yang if (map->m_lblk >= ee_block) { 3336357b66fdSDmitry Monakhov split_flag1 = split_flag & EXT4_EXT_DATA_VALID2; 3337556615dcSLukas Czerner if (unwritten) { 3338556615dcSLukas Czerner split_flag1 |= EXT4_EXT_MARK_UNWRIT1; 3339357b66fdSDmitry Monakhov split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT | 3340556615dcSLukas Czerner EXT4_EXT_MARK_UNWRIT2); 3341357b66fdSDmitry Monakhov } 3342dfe50809STheodore Ts'o err = ext4_split_extent_at(handle, inode, ppath, 334347ea3bb5SYongqiang Yang map->m_lblk, split_flag1, flags); 334447ea3bb5SYongqiang Yang if (err) 334547ea3bb5SYongqiang Yang goto out; 334647ea3bb5SYongqiang Yang } 334747ea3bb5SYongqiang Yang 334847ea3bb5SYongqiang Yang ext4_ext_show_leaf(inode, path); 334947ea3bb5SYongqiang Yang out: 33503a225670SZheng Liu return err ? err : allocated; 335147ea3bb5SYongqiang Yang } 335247ea3bb5SYongqiang Yang 335356055d3aSAmit Arora /* 3354e35fd660STheodore Ts'o * This function is called by ext4_ext_map_blocks() if someone tries to write 3355556615dcSLukas Czerner * to an unwritten extent. It may result in splitting the unwritten 335656055d3aSAmit Arora * extent into multiple extents (up to three - one initialized and two 3357556615dcSLukas Czerner * unwritten). 335856055d3aSAmit Arora * There are three possibilities: 335956055d3aSAmit Arora * a> There is no split required: Entire extent should be initialized 336056055d3aSAmit Arora * b> Splits in two extents: Write is happening at either end of the extent 336156055d3aSAmit Arora * c> Splits in three extents: Somone is writing in middle of the extent 33626f91bc5fSEric Gouriou * 33636f91bc5fSEric Gouriou * Pre-conditions: 3364556615dcSLukas Czerner * - The extent pointed to by 'path' is unwritten. 33656f91bc5fSEric Gouriou * - The extent pointed to by 'path' contains a superset 33666f91bc5fSEric Gouriou * of the logical span [map->m_lblk, map->m_lblk + map->m_len). 33676f91bc5fSEric Gouriou * 33686f91bc5fSEric Gouriou * Post-conditions on success: 33696f91bc5fSEric Gouriou * - the returned value is the number of blocks beyond map->l_lblk 33706f91bc5fSEric Gouriou * that are allocated and initialized. 33716f91bc5fSEric Gouriou * It is guaranteed to be >= map->m_len. 337256055d3aSAmit Arora */ 3373725d26d3SAneesh Kumar K.V static int ext4_ext_convert_to_initialized(handle_t *handle, 3374725d26d3SAneesh Kumar K.V struct inode *inode, 3375e35fd660STheodore Ts'o struct ext4_map_blocks *map, 3376dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 337727dd4385SLukas Czerner int flags) 337856055d3aSAmit Arora { 3379dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 338067a5da56SZheng Liu struct ext4_sb_info *sbi; 33816f91bc5fSEric Gouriou struct ext4_extent_header *eh; 3382667eff35SYongqiang Yang struct ext4_map_blocks split_map; 33834f8caa60SJan Kara struct ext4_extent zero_ex1, zero_ex2; 3384bc2d9db4SLukas Czerner struct ext4_extent *ex, *abut_ex; 338521ca087aSDmitry Monakhov ext4_lblk_t ee_block, eof_block; 3386bc2d9db4SLukas Czerner unsigned int ee_len, depth, map_len = map->m_len; 3387bc2d9db4SLukas Czerner int allocated = 0, max_zeroout = 0; 338856055d3aSAmit Arora int err = 0; 33894f8caa60SJan Kara int split_flag = EXT4_EXT_DATA_VALID2; 339021ca087aSDmitry Monakhov 339121ca087aSDmitry Monakhov ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical" 339221ca087aSDmitry Monakhov "block %llu, max_blocks %u\n", inode->i_ino, 3393bc2d9db4SLukas Czerner (unsigned long long)map->m_lblk, map_len); 339421ca087aSDmitry Monakhov 339567a5da56SZheng Liu sbi = EXT4_SB(inode->i_sb); 3396801674f3SJan Kara eof_block = (EXT4_I(inode)->i_disksize + inode->i_sb->s_blocksize - 1) 3397801674f3SJan Kara >> inode->i_sb->s_blocksize_bits; 3398bc2d9db4SLukas Czerner if (eof_block < map->m_lblk + map_len) 3399bc2d9db4SLukas Czerner eof_block = map->m_lblk + map_len; 340056055d3aSAmit Arora 340156055d3aSAmit Arora depth = ext_depth(inode); 34026f91bc5fSEric Gouriou eh = path[depth].p_hdr; 340356055d3aSAmit Arora ex = path[depth].p_ext; 340456055d3aSAmit Arora ee_block = le32_to_cpu(ex->ee_block); 340556055d3aSAmit Arora ee_len = ext4_ext_get_actual_len(ex); 34064f8caa60SJan Kara zero_ex1.ee_len = 0; 34074f8caa60SJan Kara zero_ex2.ee_len = 0; 340821ca087aSDmitry Monakhov 34096f91bc5fSEric Gouriou trace_ext4_ext_convert_to_initialized_enter(inode, map, ex); 34106f91bc5fSEric Gouriou 34116f91bc5fSEric Gouriou /* Pre-conditions */ 3412556615dcSLukas Czerner BUG_ON(!ext4_ext_is_unwritten(ex)); 34136f91bc5fSEric Gouriou BUG_ON(!in_range(map->m_lblk, ee_block, ee_len)); 34146f91bc5fSEric Gouriou 34156f91bc5fSEric Gouriou /* 34166f91bc5fSEric Gouriou * Attempt to transfer newly initialized blocks from the currently 3417556615dcSLukas Czerner * unwritten extent to its neighbor. This is much cheaper 34186f91bc5fSEric Gouriou * than an insertion followed by a merge as those involve costly 3419bc2d9db4SLukas Czerner * memmove() calls. Transferring to the left is the common case in 3420bc2d9db4SLukas Czerner * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE) 3421bc2d9db4SLukas Czerner * followed by append writes. 34226f91bc5fSEric Gouriou * 34236f91bc5fSEric Gouriou * Limitations of the current logic: 3424bc2d9db4SLukas Czerner * - L1: we do not deal with writes covering the whole extent. 34256f91bc5fSEric Gouriou * This would require removing the extent if the transfer 34266f91bc5fSEric Gouriou * is possible. 3427bc2d9db4SLukas Czerner * - L2: we only attempt to merge with an extent stored in the 34286f91bc5fSEric Gouriou * same extent tree node. 34296f91bc5fSEric Gouriou */ 3430bc2d9db4SLukas Czerner if ((map->m_lblk == ee_block) && 3431bc2d9db4SLukas Czerner /* See if we can merge left */ 3432bc2d9db4SLukas Czerner (map_len < ee_len) && /*L1*/ 3433bc2d9db4SLukas Czerner (ex > EXT_FIRST_EXTENT(eh))) { /*L2*/ 34346f91bc5fSEric Gouriou ext4_lblk_t prev_lblk; 34356f91bc5fSEric Gouriou ext4_fsblk_t prev_pblk, ee_pblk; 3436bc2d9db4SLukas Czerner unsigned int prev_len; 34376f91bc5fSEric Gouriou 3438bc2d9db4SLukas Czerner abut_ex = ex - 1; 3439bc2d9db4SLukas Czerner prev_lblk = le32_to_cpu(abut_ex->ee_block); 3440bc2d9db4SLukas Czerner prev_len = ext4_ext_get_actual_len(abut_ex); 3441bc2d9db4SLukas Czerner prev_pblk = ext4_ext_pblock(abut_ex); 34426f91bc5fSEric Gouriou ee_pblk = ext4_ext_pblock(ex); 34436f91bc5fSEric Gouriou 34446f91bc5fSEric Gouriou /* 3445bc2d9db4SLukas Czerner * A transfer of blocks from 'ex' to 'abut_ex' is allowed 34466f91bc5fSEric Gouriou * upon those conditions: 3447bc2d9db4SLukas Czerner * - C1: abut_ex is initialized, 3448bc2d9db4SLukas Czerner * - C2: abut_ex is logically abutting ex, 3449bc2d9db4SLukas Czerner * - C3: abut_ex is physically abutting ex, 3450bc2d9db4SLukas Czerner * - C4: abut_ex can receive the additional blocks without 34516f91bc5fSEric Gouriou * overflowing the (initialized) length limit. 34526f91bc5fSEric Gouriou */ 3453556615dcSLukas Czerner if ((!ext4_ext_is_unwritten(abut_ex)) && /*C1*/ 34546f91bc5fSEric Gouriou ((prev_lblk + prev_len) == ee_block) && /*C2*/ 34556f91bc5fSEric Gouriou ((prev_pblk + prev_len) == ee_pblk) && /*C3*/ 3456bc2d9db4SLukas Czerner (prev_len < (EXT_INIT_MAX_LEN - map_len))) { /*C4*/ 34576f91bc5fSEric Gouriou err = ext4_ext_get_access(handle, inode, path + depth); 34586f91bc5fSEric Gouriou if (err) 34596f91bc5fSEric Gouriou goto out; 34606f91bc5fSEric Gouriou 34616f91bc5fSEric Gouriou trace_ext4_ext_convert_to_initialized_fastpath(inode, 3462bc2d9db4SLukas Czerner map, ex, abut_ex); 34636f91bc5fSEric Gouriou 3464bc2d9db4SLukas Czerner /* Shift the start of ex by 'map_len' blocks */ 3465bc2d9db4SLukas Czerner ex->ee_block = cpu_to_le32(ee_block + map_len); 3466bc2d9db4SLukas Czerner ext4_ext_store_pblock(ex, ee_pblk + map_len); 3467bc2d9db4SLukas Czerner ex->ee_len = cpu_to_le16(ee_len - map_len); 3468556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); /* Restore the flag */ 34696f91bc5fSEric Gouriou 3470bc2d9db4SLukas Czerner /* Extend abut_ex by 'map_len' blocks */ 3471bc2d9db4SLukas Czerner abut_ex->ee_len = cpu_to_le16(prev_len + map_len); 34726f91bc5fSEric Gouriou 3473bc2d9db4SLukas Czerner /* Result: number of initialized blocks past m_lblk */ 3474bc2d9db4SLukas Czerner allocated = map_len; 3475bc2d9db4SLukas Czerner } 3476bc2d9db4SLukas Czerner } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) && 3477bc2d9db4SLukas Czerner (map_len < ee_len) && /*L1*/ 3478bc2d9db4SLukas Czerner ex < EXT_LAST_EXTENT(eh)) { /*L2*/ 3479bc2d9db4SLukas Czerner /* See if we can merge right */ 3480bc2d9db4SLukas Czerner ext4_lblk_t next_lblk; 3481bc2d9db4SLukas Czerner ext4_fsblk_t next_pblk, ee_pblk; 3482bc2d9db4SLukas Czerner unsigned int next_len; 3483bc2d9db4SLukas Czerner 3484bc2d9db4SLukas Czerner abut_ex = ex + 1; 3485bc2d9db4SLukas Czerner next_lblk = le32_to_cpu(abut_ex->ee_block); 3486bc2d9db4SLukas Czerner next_len = ext4_ext_get_actual_len(abut_ex); 3487bc2d9db4SLukas Czerner next_pblk = ext4_ext_pblock(abut_ex); 3488bc2d9db4SLukas Czerner ee_pblk = ext4_ext_pblock(ex); 3489bc2d9db4SLukas Czerner 3490bc2d9db4SLukas Czerner /* 3491bc2d9db4SLukas Czerner * A transfer of blocks from 'ex' to 'abut_ex' is allowed 3492bc2d9db4SLukas Czerner * upon those conditions: 3493bc2d9db4SLukas Czerner * - C1: abut_ex is initialized, 3494bc2d9db4SLukas Czerner * - C2: abut_ex is logically abutting ex, 3495bc2d9db4SLukas Czerner * - C3: abut_ex is physically abutting ex, 3496bc2d9db4SLukas Czerner * - C4: abut_ex can receive the additional blocks without 3497bc2d9db4SLukas Czerner * overflowing the (initialized) length limit. 3498bc2d9db4SLukas Czerner */ 3499556615dcSLukas Czerner if ((!ext4_ext_is_unwritten(abut_ex)) && /*C1*/ 3500bc2d9db4SLukas Czerner ((map->m_lblk + map_len) == next_lblk) && /*C2*/ 3501bc2d9db4SLukas Czerner ((ee_pblk + ee_len) == next_pblk) && /*C3*/ 3502bc2d9db4SLukas Czerner (next_len < (EXT_INIT_MAX_LEN - map_len))) { /*C4*/ 3503bc2d9db4SLukas Czerner err = ext4_ext_get_access(handle, inode, path + depth); 3504bc2d9db4SLukas Czerner if (err) 3505bc2d9db4SLukas Czerner goto out; 3506bc2d9db4SLukas Czerner 3507bc2d9db4SLukas Czerner trace_ext4_ext_convert_to_initialized_fastpath(inode, 3508bc2d9db4SLukas Czerner map, ex, abut_ex); 3509bc2d9db4SLukas Czerner 3510bc2d9db4SLukas Czerner /* Shift the start of abut_ex by 'map_len' blocks */ 3511bc2d9db4SLukas Czerner abut_ex->ee_block = cpu_to_le32(next_lblk - map_len); 3512bc2d9db4SLukas Czerner ext4_ext_store_pblock(abut_ex, next_pblk - map_len); 3513bc2d9db4SLukas Czerner ex->ee_len = cpu_to_le16(ee_len - map_len); 3514556615dcSLukas Czerner ext4_ext_mark_unwritten(ex); /* Restore the flag */ 3515bc2d9db4SLukas Czerner 3516bc2d9db4SLukas Czerner /* Extend abut_ex by 'map_len' blocks */ 3517bc2d9db4SLukas Czerner abut_ex->ee_len = cpu_to_le16(next_len + map_len); 3518bc2d9db4SLukas Czerner 3519bc2d9db4SLukas Czerner /* Result: number of initialized blocks past m_lblk */ 3520bc2d9db4SLukas Czerner allocated = map_len; 3521bc2d9db4SLukas Czerner } 3522bc2d9db4SLukas Czerner } 3523bc2d9db4SLukas Czerner if (allocated) { 35246f91bc5fSEric Gouriou /* Mark the block containing both extents as dirty */ 3525b60ca334SHarshad Shirwadkar err = ext4_ext_dirty(handle, inode, path + depth); 35266f91bc5fSEric Gouriou 35276f91bc5fSEric Gouriou /* Update path to point to the right extent */ 3528bc2d9db4SLukas Czerner path[depth].p_ext = abut_ex; 35296f91bc5fSEric Gouriou goto out; 3530bc2d9db4SLukas Czerner } else 3531bc2d9db4SLukas Czerner allocated = ee_len - (map->m_lblk - ee_block); 35326f91bc5fSEric Gouriou 3533667eff35SYongqiang Yang WARN_ON(map->m_lblk < ee_block); 353421ca087aSDmitry Monakhov /* 353521ca087aSDmitry Monakhov * It is safe to convert extent to initialized via explicit 35369e740568SYongqiang Yang * zeroout only if extent is fully inside i_size or new_size. 353721ca087aSDmitry Monakhov */ 3538667eff35SYongqiang Yang split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0; 353921ca087aSDmitry Monakhov 354067a5da56SZheng Liu if (EXT4_EXT_MAY_ZEROOUT & split_flag) 354167a5da56SZheng Liu max_zeroout = sbi->s_extent_max_zeroout_kb >> 35424f42f80aSLukas Czerner (inode->i_sb->s_blocksize_bits - 10); 354367a5da56SZheng Liu 3544667eff35SYongqiang Yang /* 35454f8caa60SJan Kara * five cases: 3546667eff35SYongqiang Yang * 1. split the extent into three extents. 35474f8caa60SJan Kara * 2. split the extent into two extents, zeroout the head of the first 35484f8caa60SJan Kara * extent. 35494f8caa60SJan Kara * 3. split the extent into two extents, zeroout the tail of the second 35504f8caa60SJan Kara * extent. 3551667eff35SYongqiang Yang * 4. split the extent into two extents with out zeroout. 35524f8caa60SJan Kara * 5. no splitting needed, just possibly zeroout the head and / or the 35534f8caa60SJan Kara * tail of the extent. 3554667eff35SYongqiang Yang */ 3555667eff35SYongqiang Yang split_map.m_lblk = map->m_lblk; 3556667eff35SYongqiang Yang split_map.m_len = map->m_len; 3557667eff35SYongqiang Yang 35584f8caa60SJan Kara if (max_zeroout && (allocated > split_map.m_len)) { 355967a5da56SZheng Liu if (allocated <= max_zeroout) { 35604f8caa60SJan Kara /* case 3 or 5 */ 35614f8caa60SJan Kara zero_ex1.ee_block = 35624f8caa60SJan Kara cpu_to_le32(split_map.m_lblk + 35634f8caa60SJan Kara split_map.m_len); 35644f8caa60SJan Kara zero_ex1.ee_len = 35654f8caa60SJan Kara cpu_to_le16(allocated - split_map.m_len); 35664f8caa60SJan Kara ext4_ext_store_pblock(&zero_ex1, 35674f8caa60SJan Kara ext4_ext_pblock(ex) + split_map.m_lblk + 35684f8caa60SJan Kara split_map.m_len - ee_block); 35694f8caa60SJan Kara err = ext4_ext_zeroout(inode, &zero_ex1); 3570667eff35SYongqiang Yang if (err) 3571667eff35SYongqiang Yang goto out; 3572667eff35SYongqiang Yang split_map.m_len = allocated; 35734f8caa60SJan Kara } 35744f8caa60SJan Kara if (split_map.m_lblk - ee_block + split_map.m_len < 35754f8caa60SJan Kara max_zeroout) { 35764f8caa60SJan Kara /* case 2 or 5 */ 35774f8caa60SJan Kara if (split_map.m_lblk != ee_block) { 35784f8caa60SJan Kara zero_ex2.ee_block = ex->ee_block; 35794f8caa60SJan Kara zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk - 3580667eff35SYongqiang Yang ee_block); 35814f8caa60SJan Kara ext4_ext_store_pblock(&zero_ex2, 3582667eff35SYongqiang Yang ext4_ext_pblock(ex)); 35834f8caa60SJan Kara err = ext4_ext_zeroout(inode, &zero_ex2); 3584667eff35SYongqiang Yang if (err) 3585667eff35SYongqiang Yang goto out; 3586667eff35SYongqiang Yang } 3587667eff35SYongqiang Yang 35884f8caa60SJan Kara split_map.m_len += split_map.m_lblk - ee_block; 3589667eff35SYongqiang Yang split_map.m_lblk = ee_block; 35909b940f8eSAllison Henderson allocated = map->m_len; 3591667eff35SYongqiang Yang } 3592667eff35SYongqiang Yang } 3593667eff35SYongqiang Yang 3594ae9e9c6aSJan Kara err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag, 3595ae9e9c6aSJan Kara flags); 3596ae9e9c6aSJan Kara if (err > 0) 3597ae9e9c6aSJan Kara err = 0; 3598667eff35SYongqiang Yang out: 3599adb23551SZheng Liu /* If we have gotten a failure, don't zero out status tree */ 36004f8caa60SJan Kara if (!err) { 36014f8caa60SJan Kara err = ext4_zeroout_es(inode, &zero_ex1); 3602adb23551SZheng Liu if (!err) 36034f8caa60SJan Kara err = ext4_zeroout_es(inode, &zero_ex2); 36044f8caa60SJan Kara } 3605667eff35SYongqiang Yang return err ? err : allocated; 360656055d3aSAmit Arora } 360756055d3aSAmit Arora 3608c278bfecSAneesh Kumar K.V /* 3609e35fd660STheodore Ts'o * This function is called by ext4_ext_map_blocks() from 36100031462bSMingming Cao * ext4_get_blocks_dio_write() when DIO to write 3611556615dcSLukas Czerner * to an unwritten extent. 36120031462bSMingming Cao * 3613556615dcSLukas Czerner * Writing to an unwritten extent may result in splitting the unwritten 3614556615dcSLukas Czerner * extent into multiple initialized/unwritten extents (up to three) 36150031462bSMingming Cao * There are three possibilities: 3616556615dcSLukas Czerner * a> There is no split required: Entire extent should be unwritten 36170031462bSMingming Cao * b> Splits in two extents: Write is happening at either end of the extent 36180031462bSMingming Cao * c> Splits in three extents: Somone is writing in middle of the extent 36190031462bSMingming Cao * 3620b8a86845SLukas Czerner * This works the same way in the case of initialized -> unwritten conversion. 3621b8a86845SLukas Czerner * 36220031462bSMingming Cao * One of more index blocks maybe needed if the extent tree grow after 3623556615dcSLukas Czerner * the unwritten extent split. To prevent ENOSPC occur at the IO 3624556615dcSLukas Czerner * complete, we need to split the unwritten extent before DIO submit 3625556615dcSLukas Czerner * the IO. The unwritten extent called at this time will be split 3626556615dcSLukas Czerner * into three unwritten extent(at most). After IO complete, the part 36270031462bSMingming Cao * being filled will be convert to initialized by the end_io callback function 36280031462bSMingming Cao * via ext4_convert_unwritten_extents(). 3629ba230c3fSMingming * 3630556615dcSLukas Czerner * Returns the size of unwritten extent to be written on success. 36310031462bSMingming Cao */ 3632b8a86845SLukas Czerner static int ext4_split_convert_extents(handle_t *handle, 36330031462bSMingming Cao struct inode *inode, 3634e35fd660STheodore Ts'o struct ext4_map_blocks *map, 3635dfe50809STheodore Ts'o struct ext4_ext_path **ppath, 36360031462bSMingming Cao int flags) 36370031462bSMingming Cao { 3638dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 3639667eff35SYongqiang Yang ext4_lblk_t eof_block; 3640667eff35SYongqiang Yang ext4_lblk_t ee_block; 3641667eff35SYongqiang Yang struct ext4_extent *ex; 3642667eff35SYongqiang Yang unsigned int ee_len; 3643667eff35SYongqiang Yang int split_flag = 0, depth; 36440031462bSMingming Cao 3645b8a86845SLukas Czerner ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n", 3646b8a86845SLukas Czerner __func__, inode->i_ino, 3647e35fd660STheodore Ts'o (unsigned long long)map->m_lblk, map->m_len); 364821ca087aSDmitry Monakhov 3649801674f3SJan Kara eof_block = (EXT4_I(inode)->i_disksize + inode->i_sb->s_blocksize - 1) 3650801674f3SJan Kara >> inode->i_sb->s_blocksize_bits; 3651e35fd660STheodore Ts'o if (eof_block < map->m_lblk + map->m_len) 3652e35fd660STheodore Ts'o eof_block = map->m_lblk + map->m_len; 36530031462bSMingming Cao /* 365421ca087aSDmitry Monakhov * It is safe to convert extent to initialized via explicit 365521ca087aSDmitry Monakhov * zeroout only if extent is fully insde i_size or new_size. 365621ca087aSDmitry Monakhov */ 3657667eff35SYongqiang Yang depth = ext_depth(inode); 36580031462bSMingming Cao ex = path[depth].p_ext; 3659667eff35SYongqiang Yang ee_block = le32_to_cpu(ex->ee_block); 3660667eff35SYongqiang Yang ee_len = ext4_ext_get_actual_len(ex); 36610031462bSMingming Cao 3662b8a86845SLukas Czerner /* Convert to unwritten */ 3663b8a86845SLukas Czerner if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) { 3664b8a86845SLukas Czerner split_flag |= EXT4_EXT_DATA_VALID1; 3665b8a86845SLukas Czerner /* Convert to initialized */ 3666b8a86845SLukas Czerner } else if (flags & EXT4_GET_BLOCKS_CONVERT) { 3667b8a86845SLukas Czerner split_flag |= ee_block + ee_len <= eof_block ? 3668b8a86845SLukas Czerner EXT4_EXT_MAY_ZEROOUT : 0; 3669556615dcSLukas Czerner split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2); 3670b8a86845SLukas Czerner } 3671667eff35SYongqiang Yang flags |= EXT4_GET_BLOCKS_PRE_IO; 3672dfe50809STheodore Ts'o return ext4_split_extent(handle, inode, ppath, map, split_flag, flags); 36730031462bSMingming Cao } 3674197217a5SYongqiang Yang 3675c7064ef1SJiaying Zhang static int ext4_convert_unwritten_extents_endio(handle_t *handle, 36760031462bSMingming Cao struct inode *inode, 3677dee1f973SDmitry Monakhov struct ext4_map_blocks *map, 3678dfe50809STheodore Ts'o struct ext4_ext_path **ppath) 36790031462bSMingming Cao { 3680dfe50809STheodore Ts'o struct ext4_ext_path *path = *ppath; 36810031462bSMingming Cao struct ext4_extent *ex; 3682dee1f973SDmitry Monakhov ext4_lblk_t ee_block; 3683dee1f973SDmitry Monakhov unsigned int ee_len; 36840031462bSMingming Cao int depth; 36850031462bSMingming Cao int err = 0; 36860031462bSMingming Cao 36870031462bSMingming Cao depth = ext_depth(inode); 36880031462bSMingming Cao ex = path[depth].p_ext; 3689dee1f973SDmitry Monakhov ee_block = le32_to_cpu(ex->ee_block); 3690dee1f973SDmitry Monakhov ee_len = ext4_ext_get_actual_len(ex); 36910031462bSMingming Cao 3692197217a5SYongqiang Yang ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical" 3693197217a5SYongqiang Yang "block %llu, max_blocks %u\n", inode->i_ino, 3694dee1f973SDmitry Monakhov (unsigned long long)ee_block, ee_len); 3695dee1f973SDmitry Monakhov 3696ff95ec22SDmitry Monakhov /* If extent is larger than requested it is a clear sign that we still 3697ff95ec22SDmitry Monakhov * have some extent state machine issues left. So extent_split is still 3698ff95ec22SDmitry Monakhov * required. 3699ff95ec22SDmitry Monakhov * TODO: Once all related issues will be fixed this situation should be 3700ff95ec22SDmitry Monakhov * illegal. 3701ff95ec22SDmitry Monakhov */ 3702dee1f973SDmitry Monakhov if (ee_block != map->m_lblk || ee_len > map->m_len) { 3703e3d550c2SRakesh Pandit #ifdef CONFIG_EXT4_DEBUG 3704e3d550c2SRakesh Pandit ext4_warning(inode->i_sb, "Inode (%ld) finished: extent logical block %llu," 37058d2ae1cbSJakub Wilk " len %u; IO logical block %llu, len %u", 3706ff95ec22SDmitry Monakhov inode->i_ino, (unsigned long long)ee_block, ee_len, 3707ff95ec22SDmitry Monakhov (unsigned long long)map->m_lblk, map->m_len); 3708ff95ec22SDmitry Monakhov #endif 3709dfe50809STheodore Ts'o err = ext4_split_convert_extents(handle, inode, map, ppath, 3710dee1f973SDmitry Monakhov EXT4_GET_BLOCKS_CONVERT); 3711dee1f973SDmitry Monakhov if (err < 0) 3712dfe50809STheodore Ts'o return err; 3713ed8a1a76STheodore Ts'o path = ext4_find_extent(inode, map->m_lblk, ppath, 0); 3714dfe50809STheodore Ts'o if (IS_ERR(path)) 3715dfe50809STheodore Ts'o return PTR_ERR(path); 3716dee1f973SDmitry Monakhov depth = ext_depth(inode); 3717dee1f973SDmitry Monakhov ex = path[depth].p_ext; 3718dee1f973SDmitry Monakhov } 3719197217a5SYongqiang Yang 37200031462bSMingming Cao err = ext4_ext_get_access(handle, inode, path + depth); 37210031462bSMingming Cao if (err) 37220031462bSMingming Cao goto out; 37230031462bSMingming Cao /* first mark the extent as initialized */ 37240031462bSMingming Cao ext4_ext_mark_initialized(ex); 37250031462bSMingming Cao 3726197217a5SYongqiang Yang /* note: ext4_ext_correct_indexes() isn't needed here because 3727197217a5SYongqiang Yang * borders are not changed 37280031462bSMingming Cao */ 3729ecb94f5fSTheodore Ts'o ext4_ext_try_to_merge(handle, inode, path, ex); 3730197217a5SYongqiang Yang 37310031462bSMingming Cao /* Mark modified extent as dirty */ 3732ecb94f5fSTheodore Ts'o err = ext4_ext_dirty(handle, inode, path + path->p_depth); 37330031462bSMingming Cao out: 37340031462bSMingming Cao ext4_ext_show_leaf(inode, path); 37350031462bSMingming Cao return err; 37360031462bSMingming Cao } 37370031462bSMingming Cao 37380031462bSMingming Cao static int 3739e8b83d93STheodore Ts'o convert_initialized_extent(handle_t *handle, struct inode *inode, 3740b8a86845SLukas Czerner struct ext4_map_blocks *map, 374129c6eaffSEric Whitney struct ext4_ext_path **ppath, 3742f064a9d6SEric Whitney unsigned int *allocated) 3743b8a86845SLukas Czerner { 37444f224b8bSTheodore Ts'o struct ext4_ext_path *path = *ppath; 3745e8b83d93STheodore Ts'o struct ext4_extent *ex; 3746e8b83d93STheodore Ts'o ext4_lblk_t ee_block; 3747e8b83d93STheodore Ts'o unsigned int ee_len; 3748e8b83d93STheodore Ts'o int depth; 3749b8a86845SLukas Czerner int err = 0; 3750b8a86845SLukas Czerner 3751b8a86845SLukas Czerner /* 3752b8a86845SLukas Czerner * Make sure that the extent is no bigger than we support with 3753556615dcSLukas Czerner * unwritten extent 3754b8a86845SLukas Czerner */ 3755556615dcSLukas Czerner if (map->m_len > EXT_UNWRITTEN_MAX_LEN) 3756556615dcSLukas Czerner map->m_len = EXT_UNWRITTEN_MAX_LEN / 2; 3757b8a86845SLukas Czerner 3758e8b83d93STheodore Ts'o depth = ext_depth(inode); 3759e8b83d93STheodore Ts'o ex = path[depth].p_ext; 3760e8b83d93STheodore Ts'o ee_block = le32_to_cpu(ex->ee_block); 3761e8b83d93STheodore Ts'o ee_len = ext4_ext_get_actual_len(ex); 3762e8b83d93STheodore Ts'o 3763e8b83d93STheodore Ts'o ext_debug("%s: inode %lu, logical" 3764e8b83d93STheodore Ts'o "block %llu, max_blocks %u\n", __func__, inode->i_ino, 3765e8b83d93STheodore Ts'o (unsigned long long)ee_block, ee_len); 3766e8b83d93STheodore Ts'o 3767e8b83d93STheodore Ts'o if (ee_block != map->m_lblk || ee_len > map->m_len) { 3768dfe50809STheodore Ts'o err = ext4_split_convert_extents(handle, inode, map, ppath, 3769e8b83d93STheodore Ts'o EXT4_GET_BLOCKS_CONVERT_UNWRITTEN); 3770e8b83d93STheodore Ts'o if (err < 0) 3771e8b83d93STheodore Ts'o return err; 3772ed8a1a76STheodore Ts'o path = ext4_find_extent(inode, map->m_lblk, ppath, 0); 3773e8b83d93STheodore Ts'o if (IS_ERR(path)) 3774e8b83d93STheodore Ts'o return PTR_ERR(path); 3775e8b83d93STheodore Ts'o depth = ext_depth(inode); 3776e8b83d93STheodore Ts'o ex = path[depth].p_ext; 3777e8b83d93STheodore Ts'o if (!ex) { 3778e8b83d93STheodore Ts'o EXT4_ERROR_INODE(inode, "unexpected hole at %lu", 3779e8b83d93STheodore Ts'o (unsigned long) map->m_lblk); 37806a797d27SDarrick J. Wong return -EFSCORRUPTED; 3781e8b83d93STheodore Ts'o } 3782e8b83d93STheodore Ts'o } 3783e8b83d93STheodore Ts'o 3784e8b83d93STheodore Ts'o err = ext4_ext_get_access(handle, inode, path + depth); 3785e8b83d93STheodore Ts'o if (err) 3786e8b83d93STheodore Ts'o return err; 3787e8b83d93STheodore Ts'o /* first mark the extent as unwritten */ 3788e8b83d93STheodore Ts'o ext4_ext_mark_unwritten(ex); 3789e8b83d93STheodore Ts'o 3790e8b83d93STheodore Ts'o /* note: ext4_ext_correct_indexes() isn't needed here because 3791e8b83d93STheodore Ts'o * borders are not changed 3792e8b83d93STheodore Ts'o */ 3793e8b83d93STheodore Ts'o ext4_ext_try_to_merge(handle, inode, path, ex); 3794e8b83d93STheodore Ts'o 3795e8b83d93STheodore Ts'o /* Mark modified extent as dirty */ 3796e8b83d93STheodore Ts'o err = ext4_ext_dirty(handle, inode, path + path->p_depth); 3797e8b83d93STheodore Ts'o if (err) 3798e8b83d93STheodore Ts'o return err; 3799e8b83d93STheodore Ts'o ext4_ext_show_leaf(inode, path); 3800e8b83d93STheodore Ts'o 3801b8a86845SLukas Czerner ext4_update_inode_fsync_trans(handle, inode, 1); 38024337ecd1SEric Whitney 3803b8a86845SLukas Czerner map->m_flags |= EXT4_MAP_UNWRITTEN; 3804f064a9d6SEric Whitney if (*allocated > map->m_len) 3805f064a9d6SEric Whitney *allocated = map->m_len; 3806f064a9d6SEric Whitney map->m_len = *allocated; 3807f064a9d6SEric Whitney return 0; 3808b8a86845SLukas Czerner } 3809b8a86845SLukas Czerner 3810b8a86845SLukas Czerner static int 3811556615dcSLukas Czerner ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode, 3812e35fd660STheodore Ts'o struct ext4_map_blocks *map, 3813dfe50809STheodore Ts'o struct ext4_ext_path **ppath, int flags, 3814e35fd660STheodore Ts'o unsigned int allocated, ext4_fsblk_t newblock) 38150031462bSMingming Cao { 3816*8ec2d31bSRitesh Harjani struct ext4_ext_path __maybe_unused *path = *ppath; 38170031462bSMingming Cao int ret = 0; 38180031462bSMingming Cao int err = 0; 38190031462bSMingming Cao 3820556615dcSLukas Czerner ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical " 382188635ca2SZheng Liu "block %llu, max_blocks %u, flags %x, allocated %u\n", 3822e35fd660STheodore Ts'o inode->i_ino, (unsigned long long)map->m_lblk, map->m_len, 38230031462bSMingming Cao flags, allocated); 38240031462bSMingming Cao ext4_ext_show_leaf(inode, path); 38250031462bSMingming Cao 382627dd4385SLukas Czerner /* 3827556615dcSLukas Czerner * When writing into unwritten space, we should not fail to 382827dd4385SLukas Czerner * allocate metadata blocks for the new extent block if needed. 382927dd4385SLukas Czerner */ 383027dd4385SLukas Czerner flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL; 383127dd4385SLukas Czerner 3832556615dcSLukas Czerner trace_ext4_ext_handle_unwritten_extents(inode, map, flags, 3833b5645534SZheng Liu allocated, newblock); 3834d8990240SAditya Kali 3835779e2651SEric Whitney /* get_block() before submitting IO, split the extent */ 3836c8b459f4SLukas Czerner if (flags & EXT4_GET_BLOCKS_PRE_IO) { 3837dfe50809STheodore Ts'o ret = ext4_split_convert_extents(handle, inode, map, ppath, 3838dfe50809STheodore Ts'o flags | EXT4_GET_BLOCKS_CONVERT); 3839779e2651SEric Whitney if (ret < 0) { 3840779e2651SEric Whitney err = ret; 3841779e2651SEric Whitney goto out2; 3842779e2651SEric Whitney } 3843779e2651SEric Whitney /* 3844779e2651SEric Whitney * shouldn't get a 0 return when splitting an extent unless 3845779e2651SEric Whitney * m_len is 0 (bug) or extent has been corrupted 3846779e2651SEric Whitney */ 3847779e2651SEric Whitney if (unlikely(ret == 0)) { 3848779e2651SEric Whitney EXT4_ERROR_INODE(inode, 3849779e2651SEric Whitney "unexpected ret == 0, m_len = %u", 3850779e2651SEric Whitney map->m_len); 3851779e2651SEric Whitney err = -EFSCORRUPTED; 3852779e2651SEric Whitney goto out2; 3853779e2651SEric Whitney } 3854a25a4e1aSZheng Liu map->m_flags |= EXT4_MAP_UNWRITTEN; 38550031462bSMingming Cao goto out; 38560031462bSMingming Cao } 3857c7064ef1SJiaying Zhang /* IO end_io complete, convert the filled extent to written */ 3858c8b459f4SLukas Czerner if (flags & EXT4_GET_BLOCKS_CONVERT) { 3859bee6cf00SEric Whitney err = ext4_convert_unwritten_extents_endio(handle, inode, map, 3860dfe50809STheodore Ts'o ppath); 3861bee6cf00SEric Whitney if (err < 0) 38620031462bSMingming Cao goto out2; 3863bee6cf00SEric Whitney ext4_update_inode_fsync_trans(handle, inode, 1); 3864bee6cf00SEric Whitney goto map_out; 38650031462bSMingming Cao } 3866bee6cf00SEric Whitney /* buffered IO cases */ 38670031462bSMingming Cao /* 38680031462bSMingming Cao * repeat fallocate creation request 38690031462bSMingming Cao * we already have an unwritten extent 38700031462bSMingming Cao */ 3871556615dcSLukas Czerner if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) { 3872a25a4e1aSZheng Liu map->m_flags |= EXT4_MAP_UNWRITTEN; 38730031462bSMingming Cao goto map_out; 3874a25a4e1aSZheng Liu } 38750031462bSMingming Cao 38760031462bSMingming Cao /* buffered READ or buffered write_begin() lookup */ 38770031462bSMingming Cao if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) { 38780031462bSMingming Cao /* 38790031462bSMingming Cao * We have blocks reserved already. We 38800031462bSMingming Cao * return allocated blocks so that delalloc 38810031462bSMingming Cao * won't do block reservation for us. But 38820031462bSMingming Cao * the buffer head will be unmapped so that 38830031462bSMingming Cao * a read from the block returns 0s. 38840031462bSMingming Cao */ 3885e35fd660STheodore Ts'o map->m_flags |= EXT4_MAP_UNWRITTEN; 38860031462bSMingming Cao goto out1; 38870031462bSMingming Cao } 38880031462bSMingming Cao 3889be809e12SEric Whitney /* 3890be809e12SEric Whitney * Default case when (flags & EXT4_GET_BLOCKS_CREATE) == 1. 3891be809e12SEric Whitney * For buffered writes, at writepage time, etc. Convert a 3892be809e12SEric Whitney * discovered unwritten extent to written. 3893be809e12SEric Whitney */ 3894dfe50809STheodore Ts'o ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags); 3895be809e12SEric Whitney if (ret < 0) { 38960031462bSMingming Cao err = ret; 38970031462bSMingming Cao goto out2; 3898779e2651SEric Whitney } 3899be809e12SEric Whitney ext4_update_inode_fsync_trans(handle, inode, 1); 3900be809e12SEric Whitney /* 3901be809e12SEric Whitney * shouldn't get a 0 return when converting an unwritten extent 3902be809e12SEric Whitney * unless m_len is 0 (bug) or extent has been corrupted 3903be809e12SEric Whitney */ 3904be809e12SEric Whitney if (unlikely(ret == 0)) { 3905be809e12SEric Whitney EXT4_ERROR_INODE(inode, "unexpected ret == 0, m_len = %u", 3906be809e12SEric Whitney map->m_len); 3907be809e12SEric Whitney err = -EFSCORRUPTED; 3908be809e12SEric Whitney goto out2; 3909be809e12SEric Whitney } 3910be809e12SEric Whitney 3911779e2651SEric Whitney out: 39120031462bSMingming Cao allocated = ret; 3913e35fd660STheodore Ts'o map->m_flags |= EXT4_MAP_NEW; 39140031462bSMingming Cao map_out: 3915e35fd660STheodore Ts'o map->m_flags |= EXT4_MAP_MAPPED; 39160031462bSMingming Cao out1: 3917bee6cf00SEric Whitney map->m_pblk = newblock; 3918e35fd660STheodore Ts'o if (allocated > map->m_len) 3919e35fd660STheodore Ts'o allocated = map->m_len; 3920e35fd660STheodore Ts'o map->m_len = allocated; 3921bee6cf00SEric Whitney ext4_ext_show_leaf(inode, path); 39220031462bSMingming Cao out2: 39230031462bSMingming Cao return err ? err : allocated; 39240031462bSMingming Cao } 392558590b06STheodore Ts'o 39260031462bSMingming Cao /* 39274d33b1efSTheodore Ts'o * get_implied_cluster_alloc - check to see if the requested 39284d33b1efSTheodore Ts'o * allocation (in the map structure) overlaps with a cluster already 39294d33b1efSTheodore Ts'o * allocated in an extent. 3930d8990240SAditya Kali * @sb The filesystem superblock structure 39314d33b1efSTheodore Ts'o * @map The requested lblk->pblk mapping 39324d33b1efSTheodore Ts'o * @ex The extent structure which might contain an implied 39334d33b1efSTheodore Ts'o * cluster allocation 39344d33b1efSTheodore Ts'o * 39354d33b1efSTheodore Ts'o * This function is called by ext4_ext_map_blocks() after we failed to 39364d33b1efSTheodore Ts'o * find blocks that were already in the inode's extent tree. Hence, 39374d33b1efSTheodore Ts'o * we know that the beginning of the requested region cannot overlap 39384d33b1efSTheodore Ts'o * the extent from the inode's extent tree. There are three cases we 39394d33b1efSTheodore Ts'o * want to catch. The first is this case: 39404d33b1efSTheodore Ts'o * 39414d33b1efSTheodore Ts'o * |--- cluster # N--| 39424d33b1efSTheodore Ts'o * |--- extent ---| |---- requested region ---| 39434d33b1efSTheodore Ts'o * |==========| 39444d33b1efSTheodore Ts'o * 39454d33b1efSTheodore Ts'o * The second case that we need to test for is this one: 39464d33b1efSTheodore Ts'o * 39474d33b1efSTheodore Ts'o * |--------- cluster # N ----------------| 39484d33b1efSTheodore Ts'o * |--- requested region --| |------- extent ----| 39494d33b1efSTheodore Ts'o * |=======================| 39504d33b1efSTheodore Ts'o * 39514d33b1efSTheodore Ts'o * The third case is when the requested region lies between two extents 39524d33b1efSTheodore Ts'o * within the same cluster: 39534d33b1efSTheodore Ts'o * |------------- cluster # N-------------| 39544d33b1efSTheodore Ts'o * |----- ex -----| |---- ex_right ----| 39554d33b1efSTheodore Ts'o * |------ requested region ------| 39564d33b1efSTheodore Ts'o * |================| 39574d33b1efSTheodore Ts'o * 39584d33b1efSTheodore Ts'o * In each of the above cases, we need to set the map->m_pblk and 39594d33b1efSTheodore Ts'o * map->m_len so it corresponds to the return the extent labelled as 39604d33b1efSTheodore Ts'o * "|====|" from cluster #N, since it is already in use for data in 39614d33b1efSTheodore Ts'o * cluster EXT4_B2C(sbi, map->m_lblk). We will then return 1 to 39624d33b1efSTheodore Ts'o * signal to ext4_ext_map_blocks() that map->m_pblk should be treated 39634d33b1efSTheodore Ts'o * as a new "allocated" block region. Otherwise, we will return 0 and 39644d33b1efSTheodore Ts'o * ext4_ext_map_blocks() will then allocate one or more new clusters 39654d33b1efSTheodore Ts'o * by calling ext4_mb_new_blocks(). 39664d33b1efSTheodore Ts'o */ 3967d8990240SAditya Kali static int get_implied_cluster_alloc(struct super_block *sb, 39684d33b1efSTheodore Ts'o struct ext4_map_blocks *map, 39694d33b1efSTheodore Ts'o struct ext4_extent *ex, 39704d33b1efSTheodore Ts'o struct ext4_ext_path *path) 39714d33b1efSTheodore Ts'o { 3972d8990240SAditya Kali struct ext4_sb_info *sbi = EXT4_SB(sb); 3973f5a44db5STheodore Ts'o ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk); 39744d33b1efSTheodore Ts'o ext4_lblk_t ex_cluster_start, ex_cluster_end; 397514d7f3efSCurt Wohlgemuth ext4_lblk_t rr_cluster_start; 39764d33b1efSTheodore Ts'o ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block); 39774d33b1efSTheodore Ts'o ext4_fsblk_t ee_start = ext4_ext_pblock(ex); 39784d33b1efSTheodore Ts'o unsigned short ee_len = ext4_ext_get_actual_len(ex); 39794d33b1efSTheodore Ts'o 39804d33b1efSTheodore Ts'o /* The extent passed in that we are trying to match */ 39814d33b1efSTheodore Ts'o ex_cluster_start = EXT4_B2C(sbi, ee_block); 39824d33b1efSTheodore Ts'o ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1); 39834d33b1efSTheodore Ts'o 39844d33b1efSTheodore Ts'o /* The requested region passed into ext4_map_blocks() */ 39854d33b1efSTheodore Ts'o rr_cluster_start = EXT4_B2C(sbi, map->m_lblk); 39864d33b1efSTheodore Ts'o 39874d33b1efSTheodore Ts'o if ((rr_cluster_start == ex_cluster_end) || 39884d33b1efSTheodore Ts'o (rr_cluster_start == ex_cluster_start)) { 39894d33b1efSTheodore Ts'o if (rr_cluster_start == ex_cluster_end) 39904d33b1efSTheodore Ts'o ee_start += ee_len - 1; 3991f5a44db5STheodore Ts'o map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset; 39924d33b1efSTheodore Ts'o map->m_len = min(map->m_len, 39934d33b1efSTheodore Ts'o (unsigned) sbi->s_cluster_ratio - c_offset); 39944d33b1efSTheodore Ts'o /* 39954d33b1efSTheodore Ts'o * Check for and handle this case: 39964d33b1efSTheodore Ts'o * 39974d33b1efSTheodore Ts'o * |--------- cluster # N-------------| 39984d33b1efSTheodore Ts'o * |------- extent ----| 39994d33b1efSTheodore Ts'o * |--- requested region ---| 40004d33b1efSTheodore Ts'o * |===========| 40014d33b1efSTheodore Ts'o */ 40024d33b1efSTheodore Ts'o 40034d33b1efSTheodore Ts'o if (map->m_lblk < ee_block) 40044d33b1efSTheodore Ts'o map->m_len = min(map->m_len, ee_block - map->m_lblk); 40054d33b1efSTheodore Ts'o 40064d33b1efSTheodore Ts'o /* 40074d33b1efSTheodore Ts'o * Check for the case where there is already another allocated 40084d33b1efSTheodore Ts'o * block to the right of 'ex' but before the end of the cluster. 40094d33b1efSTheodore Ts'o * 40104d33b1efSTheodore Ts'o * |------------- cluster # N-------------| 40114d33b1efSTheodore Ts'o * |----- ex -----| |---- ex_right ----| 40124d33b1efSTheodore Ts'o * |------ requested region ------| 40134d33b1efSTheodore Ts'o * |================| 40144d33b1efSTheodore Ts'o */ 40154d33b1efSTheodore Ts'o if (map->m_lblk > ee_block) { 40164d33b1efSTheodore Ts'o ext4_lblk_t next = ext4_ext_next_allocated_block(path); 40174d33b1efSTheodore Ts'o map->m_len = min(map->m_len, next - map->m_lblk); 40184d33b1efSTheodore Ts'o } 4019d8990240SAditya Kali 4020d8990240SAditya Kali trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1); 40214d33b1efSTheodore Ts'o return 1; 40224d33b1efSTheodore Ts'o } 4023d8990240SAditya Kali 4024d8990240SAditya Kali trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0); 40254d33b1efSTheodore Ts'o return 0; 40264d33b1efSTheodore Ts'o } 40274d33b1efSTheodore Ts'o 40284d33b1efSTheodore Ts'o 40294d33b1efSTheodore Ts'o /* 4030f5ab0d1fSMingming Cao * Block allocation/map/preallocation routine for extents based files 4031f5ab0d1fSMingming Cao * 4032f5ab0d1fSMingming Cao * 4033c278bfecSAneesh Kumar K.V * Need to be called with 40340e855ac8SAneesh Kumar K.V * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block 40350e855ac8SAneesh Kumar K.V * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem) 4036f5ab0d1fSMingming Cao * 4037f5ab0d1fSMingming Cao * return > 0, number of of blocks already mapped/allocated 4038f5ab0d1fSMingming Cao * if create == 0 and these are pre-allocated blocks 4039f5ab0d1fSMingming Cao * buffer head is unmapped 4040f5ab0d1fSMingming Cao * otherwise blocks are mapped 4041f5ab0d1fSMingming Cao * 4042f5ab0d1fSMingming Cao * return = 0, if plain look up failed (blocks have not been allocated) 4043f5ab0d1fSMingming Cao * buffer head is unmapped 4044f5ab0d1fSMingming Cao * 4045f5ab0d1fSMingming Cao * return < 0, error case. 4046c278bfecSAneesh Kumar K.V */ 4047e35fd660STheodore Ts'o int ext4_ext_map_blocks(handle_t *handle, struct inode *inode, 4048e35fd660STheodore Ts'o struct ext4_map_blocks *map, int flags) 4049a86c6181SAlex Tomas { 4050a86c6181SAlex Tomas struct ext4_ext_path *path = NULL; 40514d33b1efSTheodore Ts'o struct ext4_extent newex, *ex, *ex2; 40524d33b1efSTheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 40530562e0baSJiaying Zhang ext4_fsblk_t newblock = 0; 405434990461SEric Whitney int err = 0, depth, ret; 40554d33b1efSTheodore Ts'o unsigned int allocated = 0, offset = 0; 405681fdbb4aSYongqiang Yang unsigned int allocated_clusters = 0; 4057c9de560dSAlex Tomas struct ext4_allocation_request ar; 40584d33b1efSTheodore Ts'o ext4_lblk_t cluster_offset; 4059a86c6181SAlex Tomas 406084fe3befSMingming ext_debug("blocks %u/%u requested for inode %lu\n", 4061e35fd660STheodore Ts'o map->m_lblk, map->m_len, inode->i_ino); 40620562e0baSJiaying Zhang trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags); 4063a86c6181SAlex Tomas 4064a86c6181SAlex Tomas /* find extent for this block */ 4065ed8a1a76STheodore Ts'o path = ext4_find_extent(inode, map->m_lblk, NULL, 0); 4066a86c6181SAlex Tomas if (IS_ERR(path)) { 4067a86c6181SAlex Tomas err = PTR_ERR(path); 4068a86c6181SAlex Tomas path = NULL; 4069a86c6181SAlex Tomas goto out2; 4070a86c6181SAlex Tomas } 4071a86c6181SAlex Tomas 4072a86c6181SAlex Tomas depth = ext_depth(inode); 4073a86c6181SAlex Tomas 4074a86c6181SAlex Tomas /* 4075d0d856e8SRandy Dunlap * consistent leaf must not be empty; 4076d0d856e8SRandy Dunlap * this situation is possible, though, _during_ tree modification; 4077ed8a1a76STheodore Ts'o * this is why assert can't be put in ext4_find_extent() 4078a86c6181SAlex Tomas */ 4079273df556SFrank Mayhar if (unlikely(path[depth].p_ext == NULL && depth != 0)) { 4080273df556SFrank Mayhar EXT4_ERROR_INODE(inode, "bad extent address " 4081f70f362bSTheodore Ts'o "lblock: %lu, depth: %d pblock %lld", 4082f70f362bSTheodore Ts'o (unsigned long) map->m_lblk, depth, 4083f70f362bSTheodore Ts'o path[depth].p_block); 40846a797d27SDarrick J. Wong err = -EFSCORRUPTED; 4085034fb4c9SSurbhi Palande goto out2; 4086034fb4c9SSurbhi Palande } 4087a86c6181SAlex Tomas 40887e028976SAvantika Mathur ex = path[depth].p_ext; 40897e028976SAvantika Mathur if (ex) { 4090725d26d3SAneesh Kumar K.V ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block); 4091bf89d16fSTheodore Ts'o ext4_fsblk_t ee_start = ext4_ext_pblock(ex); 4092a2df2a63SAmit Arora unsigned short ee_len; 4093471d4011SSuparna Bhattacharya 4094b8a86845SLukas Czerner 4095471d4011SSuparna Bhattacharya /* 4096556615dcSLukas Czerner * unwritten extents are treated as holes, except that 409756055d3aSAmit Arora * we split out initialized portions during a write. 4098471d4011SSuparna Bhattacharya */ 4099a2df2a63SAmit Arora ee_len = ext4_ext_get_actual_len(ex); 4100d8990240SAditya Kali 4101d8990240SAditya Kali trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len); 4102d8990240SAditya Kali 4103d0d856e8SRandy Dunlap /* if found extent covers block, simply return it */ 4104e35fd660STheodore Ts'o if (in_range(map->m_lblk, ee_block, ee_len)) { 4105e35fd660STheodore Ts'o newblock = map->m_lblk - ee_block + ee_start; 4106d0d856e8SRandy Dunlap /* number of remaining blocks in the extent */ 4107e35fd660STheodore Ts'o allocated = ee_len - (map->m_lblk - ee_block); 4108e35fd660STheodore Ts'o ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk, 4109a86c6181SAlex Tomas ee_block, ee_len, newblock); 411056055d3aSAmit Arora 4111b8a86845SLukas Czerner /* 4112b8a86845SLukas Czerner * If the extent is initialized check whether the 4113b8a86845SLukas Czerner * caller wants to convert it to unwritten. 4114b8a86845SLukas Czerner */ 4115556615dcSLukas Czerner if ((!ext4_ext_is_unwritten(ex)) && 4116b8a86845SLukas Czerner (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) { 4117f064a9d6SEric Whitney err = convert_initialized_extent(handle, 4118f064a9d6SEric Whitney inode, map, &path, &allocated); 4119b8a86845SLukas Czerner goto out2; 4120f064a9d6SEric Whitney } else if (!ext4_ext_is_unwritten(ex)) { 4121a86c6181SAlex Tomas goto out; 4122f064a9d6SEric Whitney } 412369eb33dcSZheng Liu 4124556615dcSLukas Czerner ret = ext4_ext_handle_unwritten_extents( 4125dfe50809STheodore Ts'o handle, inode, map, &path, flags, 4126e861304bSAllison Henderson allocated, newblock); 4127ce37c429SEric Whitney if (ret < 0) 4128ce37c429SEric Whitney err = ret; 4129ce37c429SEric Whitney else 4130ce37c429SEric Whitney allocated = ret; 413131cf0f2cSEric Whitney goto out2; 413256055d3aSAmit Arora } 4133a86c6181SAlex Tomas } 4134a86c6181SAlex Tomas 4135a86c6181SAlex Tomas /* 4136d0d856e8SRandy Dunlap * requested block isn't allocated yet; 4137a86c6181SAlex Tomas * we couldn't try to create block if create flag is zero 4138a86c6181SAlex Tomas */ 4139c2177057STheodore Ts'o if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) { 4140140a5250SJan Kara ext4_lblk_t hole_start, hole_len; 4141140a5250SJan Kara 4142facab4d9SJan Kara hole_start = map->m_lblk; 4143facab4d9SJan Kara hole_len = ext4_ext_determine_hole(inode, path, &hole_start); 414456055d3aSAmit Arora /* 414556055d3aSAmit Arora * put just found gap into cache to speed up 414656055d3aSAmit Arora * subsequent requests 414756055d3aSAmit Arora */ 4148140a5250SJan Kara ext4_ext_put_gap_in_cache(inode, hole_start, hole_len); 4149facab4d9SJan Kara 4150facab4d9SJan Kara /* Update hole_len to reflect hole size after map->m_lblk */ 4151facab4d9SJan Kara if (hole_start != map->m_lblk) 4152facab4d9SJan Kara hole_len -= map->m_lblk - hole_start; 4153facab4d9SJan Kara map->m_pblk = 0; 4154facab4d9SJan Kara map->m_len = min_t(unsigned int, map->m_len, hole_len); 4155facab4d9SJan Kara 4156a86c6181SAlex Tomas goto out2; 4157a86c6181SAlex Tomas } 41584d33b1efSTheodore Ts'o 4159a86c6181SAlex Tomas /* 4160c2ea3fdeSTheodore Ts'o * Okay, we need to do block allocation. 4161a86c6181SAlex Tomas */ 41624d33b1efSTheodore Ts'o newex.ee_block = cpu_to_le32(map->m_lblk); 4163d0abafacSEric Whitney cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk); 41644d33b1efSTheodore Ts'o 41654d33b1efSTheodore Ts'o /* 41664d33b1efSTheodore Ts'o * If we are doing bigalloc, check to see if the extent returned 4167ed8a1a76STheodore Ts'o * by ext4_find_extent() implies a cluster we can use. 41684d33b1efSTheodore Ts'o */ 41694d33b1efSTheodore Ts'o if (cluster_offset && ex && 4170d8990240SAditya Kali get_implied_cluster_alloc(inode->i_sb, map, ex, path)) { 41714d33b1efSTheodore Ts'o ar.len = allocated = map->m_len; 41724d33b1efSTheodore Ts'o newblock = map->m_pblk; 41734d33b1efSTheodore Ts'o goto got_allocated_blocks; 41744d33b1efSTheodore Ts'o } 4175a86c6181SAlex Tomas 4176c9de560dSAlex Tomas /* find neighbour allocated blocks */ 4177e35fd660STheodore Ts'o ar.lleft = map->m_lblk; 4178c9de560dSAlex Tomas err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft); 4179c9de560dSAlex Tomas if (err) 4180c9de560dSAlex Tomas goto out2; 4181e35fd660STheodore Ts'o ar.lright = map->m_lblk; 41824d33b1efSTheodore Ts'o ex2 = NULL; 41834d33b1efSTheodore Ts'o err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2); 4184c9de560dSAlex Tomas if (err) 4185c9de560dSAlex Tomas goto out2; 418625d14f98SAmit Arora 41874d33b1efSTheodore Ts'o /* Check if the extent after searching to the right implies a 41884d33b1efSTheodore Ts'o * cluster we can use. */ 41894d33b1efSTheodore Ts'o if ((sbi->s_cluster_ratio > 1) && ex2 && 4190d8990240SAditya Kali get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) { 41914d33b1efSTheodore Ts'o ar.len = allocated = map->m_len; 41924d33b1efSTheodore Ts'o newblock = map->m_pblk; 41934d33b1efSTheodore Ts'o goto got_allocated_blocks; 41944d33b1efSTheodore Ts'o } 41954d33b1efSTheodore Ts'o 4196749269faSAmit Arora /* 4197749269faSAmit Arora * See if request is beyond maximum number of blocks we can have in 4198749269faSAmit Arora * a single extent. For an initialized extent this limit is 4199556615dcSLukas Czerner * EXT_INIT_MAX_LEN and for an unwritten extent this limit is 4200556615dcSLukas Czerner * EXT_UNWRITTEN_MAX_LEN. 4201749269faSAmit Arora */ 4202e35fd660STheodore Ts'o if (map->m_len > EXT_INIT_MAX_LEN && 4203556615dcSLukas Czerner !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT)) 4204e35fd660STheodore Ts'o map->m_len = EXT_INIT_MAX_LEN; 4205556615dcSLukas Czerner else if (map->m_len > EXT_UNWRITTEN_MAX_LEN && 4206556615dcSLukas Czerner (flags & EXT4_GET_BLOCKS_UNWRIT_EXT)) 4207556615dcSLukas Czerner map->m_len = EXT_UNWRITTEN_MAX_LEN; 4208749269faSAmit Arora 4209e35fd660STheodore Ts'o /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */ 4210e35fd660STheodore Ts'o newex.ee_len = cpu_to_le16(map->m_len); 42114d33b1efSTheodore Ts'o err = ext4_ext_check_overlap(sbi, inode, &newex, path); 421225d14f98SAmit Arora if (err) 4213b939e376SAneesh Kumar K.V allocated = ext4_ext_get_actual_len(&newex); 421425d14f98SAmit Arora else 4215e35fd660STheodore Ts'o allocated = map->m_len; 4216c9de560dSAlex Tomas 4217c9de560dSAlex Tomas /* allocate new block */ 4218c9de560dSAlex Tomas ar.inode = inode; 4219e35fd660STheodore Ts'o ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk); 4220e35fd660STheodore Ts'o ar.logical = map->m_lblk; 42214d33b1efSTheodore Ts'o /* 42224d33b1efSTheodore Ts'o * We calculate the offset from the beginning of the cluster 42234d33b1efSTheodore Ts'o * for the logical block number, since when we allocate a 42244d33b1efSTheodore Ts'o * physical cluster, the physical block should start at the 42254d33b1efSTheodore Ts'o * same offset from the beginning of the cluster. This is 42264d33b1efSTheodore Ts'o * needed so that future calls to get_implied_cluster_alloc() 42274d33b1efSTheodore Ts'o * work correctly. 42284d33b1efSTheodore Ts'o */ 4229f5a44db5STheodore Ts'o offset = EXT4_LBLK_COFF(sbi, map->m_lblk); 42304d33b1efSTheodore Ts'o ar.len = EXT4_NUM_B2C(sbi, offset+allocated); 42314d33b1efSTheodore Ts'o ar.goal -= offset; 42324d33b1efSTheodore Ts'o ar.logical -= offset; 4233c9de560dSAlex Tomas if (S_ISREG(inode->i_mode)) 4234c9de560dSAlex Tomas ar.flags = EXT4_MB_HINT_DATA; 4235c9de560dSAlex Tomas else 4236c9de560dSAlex Tomas /* disable in-core preallocation for non-regular files */ 4237c9de560dSAlex Tomas ar.flags = 0; 4238556b27abSVivek Haldar if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE) 4239556b27abSVivek Haldar ar.flags |= EXT4_MB_HINT_NOPREALLOC; 4240e3cf5d5dSTheodore Ts'o if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) 4241e3cf5d5dSTheodore Ts'o ar.flags |= EXT4_MB_DELALLOC_RESERVED; 4242c5e298aeSTheodore Ts'o if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL) 4243c5e298aeSTheodore Ts'o ar.flags |= EXT4_MB_USE_RESERVED; 4244c9de560dSAlex Tomas newblock = ext4_mb_new_blocks(handle, &ar, &err); 4245a86c6181SAlex Tomas if (!newblock) 4246a86c6181SAlex Tomas goto out2; 42477b415bf6SAditya Kali allocated_clusters = ar.len; 42484d33b1efSTheodore Ts'o ar.len = EXT4_C2B(sbi, ar.len) - offset; 4249ec8c60beSRitesh Harjani ext_debug("allocate new block: goal %llu, found %llu/%u, requested %u\n", 4250ec8c60beSRitesh Harjani ar.goal, newblock, ar.len, allocated); 42514d33b1efSTheodore Ts'o if (ar.len > allocated) 42524d33b1efSTheodore Ts'o ar.len = allocated; 4253a86c6181SAlex Tomas 42544d33b1efSTheodore Ts'o got_allocated_blocks: 4255a86c6181SAlex Tomas /* try to insert new extent into found leaf and return */ 42564d33b1efSTheodore Ts'o ext4_ext_store_pblock(&newex, newblock + offset); 4257c9de560dSAlex Tomas newex.ee_len = cpu_to_le16(ar.len); 4258556615dcSLukas Czerner /* Mark unwritten */ 4259556615dcSLukas Czerner if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) { 4260556615dcSLukas Czerner ext4_ext_mark_unwritten(&newex); 4261a25a4e1aSZheng Liu map->m_flags |= EXT4_MAP_UNWRITTEN; 42628d5d02e6SMingming Cao } 4263c8d46e41SJiaying Zhang 42644337ecd1SEric Whitney err = ext4_ext_insert_extent(handle, inode, &path, &newex, flags); 426534990461SEric Whitney if (err) { 426634990461SEric Whitney if (allocated_clusters) { 426734990461SEric Whitney int fb_flags = 0; 426882e54229SDmitry Monakhov 426934990461SEric Whitney /* 427034990461SEric Whitney * free data blocks we just allocated. 427134990461SEric Whitney * not a good idea to call discard here directly, 427234990461SEric Whitney * but otherwise we'd need to call it every free(). 427334990461SEric Whitney */ 4274c2ea3fdeSTheodore Ts'o ext4_discard_preallocations(inode); 427534990461SEric Whitney if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) 427634990461SEric Whitney fb_flags = EXT4_FREE_BLOCKS_NO_QUOT_UPDATE; 4277c8e15130STheodore Ts'o ext4_free_blocks(handle, inode, NULL, newblock, 427834990461SEric Whitney EXT4_C2B(sbi, allocated_clusters), 427934990461SEric Whitney fb_flags); 428034990461SEric Whitney } 4281a86c6181SAlex Tomas goto out2; 4282315054f0SAlex Tomas } 4283a86c6181SAlex Tomas 4284a86c6181SAlex Tomas /* previous routine could use block we allocated */ 4285bf89d16fSTheodore Ts'o newblock = ext4_ext_pblock(&newex); 4286b939e376SAneesh Kumar K.V allocated = ext4_ext_get_actual_len(&newex); 4287e35fd660STheodore Ts'o if (allocated > map->m_len) 4288e35fd660STheodore Ts'o allocated = map->m_len; 4289e35fd660STheodore Ts'o map->m_flags |= EXT4_MAP_NEW; 4290a86c6181SAlex Tomas 4291b436b9beSJan Kara /* 4292b6bf9171SEric Whitney * Reduce the reserved cluster count to reflect successful deferred 4293b6bf9171SEric Whitney * allocation of delayed allocated clusters or direct allocation of 4294b6bf9171SEric Whitney * clusters discovered to be delayed allocated. Once allocated, a 4295b6bf9171SEric Whitney * cluster is not included in the reserved count. 42965f634d06SAneesh Kumar K.V */ 42972971148dSEric Whitney if (test_opt(inode->i_sb, DELALLOC) && allocated_clusters) { 42987b415bf6SAditya Kali if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) { 42997b415bf6SAditya Kali /* 4300b6bf9171SEric Whitney * When allocating delayed allocated clusters, simply 4301b6bf9171SEric Whitney * reduce the reserved cluster count and claim quota 4302232ec872SLukas Czerner */ 4303232ec872SLukas Czerner ext4_da_update_reserve_space(inode, allocated_clusters, 4304232ec872SLukas Czerner 1); 4305b6bf9171SEric Whitney } else { 4306b6bf9171SEric Whitney ext4_lblk_t lblk, len; 4307b6bf9171SEric Whitney unsigned int n; 4308b6bf9171SEric Whitney 4309b6bf9171SEric Whitney /* 4310b6bf9171SEric Whitney * When allocating non-delayed allocated clusters 4311b6bf9171SEric Whitney * (from fallocate, filemap, DIO, or clusters 4312b6bf9171SEric Whitney * allocated when delalloc has been disabled by 4313b6bf9171SEric Whitney * ext4_nonda_switch), reduce the reserved cluster 4314b6bf9171SEric Whitney * count by the number of allocated clusters that 4315b6bf9171SEric Whitney * have previously been delayed allocated. Quota 4316b6bf9171SEric Whitney * has been claimed by ext4_mb_new_blocks() above, 4317b6bf9171SEric Whitney * so release the quota reservations made for any 4318b6bf9171SEric Whitney * previously delayed allocated clusters. 4319b6bf9171SEric Whitney */ 4320b6bf9171SEric Whitney lblk = EXT4_LBLK_CMASK(sbi, map->m_lblk); 4321b6bf9171SEric Whitney len = allocated_clusters << sbi->s_cluster_bits; 4322b6bf9171SEric Whitney n = ext4_es_delayed_clu(inode, lblk, len); 4323b6bf9171SEric Whitney if (n > 0) 4324b6bf9171SEric Whitney ext4_da_update_reserve_space(inode, (int) n, 0); 43257b415bf6SAditya Kali } 43267b415bf6SAditya Kali } 43275f634d06SAneesh Kumar K.V 43285f634d06SAneesh Kumar K.V /* 4329b436b9beSJan Kara * Cache the extent and update transaction to commit on fdatasync only 4330556615dcSLukas Czerner * when it is _not_ an unwritten extent. 4331b436b9beSJan Kara */ 4332556615dcSLukas Czerner if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0) 4333b436b9beSJan Kara ext4_update_inode_fsync_trans(handle, inode, 1); 433469eb33dcSZheng Liu else 4335b436b9beSJan Kara ext4_update_inode_fsync_trans(handle, inode, 0); 4336a86c6181SAlex Tomas out: 4337e35fd660STheodore Ts'o if (allocated > map->m_len) 4338e35fd660STheodore Ts'o allocated = map->m_len; 4339a86c6181SAlex Tomas ext4_ext_show_leaf(inode, path); 4340e35fd660STheodore Ts'o map->m_flags |= EXT4_MAP_MAPPED; 4341e35fd660STheodore Ts'o map->m_pblk = newblock; 4342e35fd660STheodore Ts'o map->m_len = allocated; 4343a86c6181SAlex Tomas out2: 4344a86c6181SAlex Tomas ext4_ext_drop_refs(path); 4345a86c6181SAlex Tomas kfree(path); 4346e861304bSAllison Henderson 434763b99968STheodore Ts'o trace_ext4_ext_map_blocks_exit(inode, flags, map, 434863b99968STheodore Ts'o err ? err : allocated); 43497877191cSLukas Czerner return err ? err : allocated; 4350a86c6181SAlex Tomas } 4351a86c6181SAlex Tomas 4352d0abb36dSTheodore Ts'o int ext4_ext_truncate(handle_t *handle, struct inode *inode) 4353a86c6181SAlex Tomas { 4354a86c6181SAlex Tomas struct super_block *sb = inode->i_sb; 4355725d26d3SAneesh Kumar K.V ext4_lblk_t last_block; 4356a86c6181SAlex Tomas int err = 0; 4357a86c6181SAlex Tomas 4358a86c6181SAlex Tomas /* 4359d0d856e8SRandy Dunlap * TODO: optimization is possible here. 4360d0d856e8SRandy Dunlap * Probably we need not scan at all, 4361d0d856e8SRandy Dunlap * because page truncation is enough. 4362a86c6181SAlex Tomas */ 4363a86c6181SAlex Tomas 4364a86c6181SAlex Tomas /* we have to know where to truncate from in crash case */ 4365a86c6181SAlex Tomas EXT4_I(inode)->i_disksize = inode->i_size; 4366d0abb36dSTheodore Ts'o err = ext4_mark_inode_dirty(handle, inode); 4367d0abb36dSTheodore Ts'o if (err) 4368d0abb36dSTheodore Ts'o return err; 4369a86c6181SAlex Tomas 4370a86c6181SAlex Tomas last_block = (inode->i_size + sb->s_blocksize - 1) 4371a86c6181SAlex Tomas >> EXT4_BLOCK_SIZE_BITS(sb); 43728acd5e9bSTheodore Ts'o retry: 437351865fdaSZheng Liu err = ext4_es_remove_extent(inode, last_block, 437451865fdaSZheng Liu EXT_MAX_BLOCKS - last_block); 437594eec0fcSTheodore Ts'o if (err == -ENOMEM) { 43768acd5e9bSTheodore Ts'o cond_resched(); 43778acd5e9bSTheodore Ts'o congestion_wait(BLK_RW_ASYNC, HZ/50); 43788acd5e9bSTheodore Ts'o goto retry; 43798acd5e9bSTheodore Ts'o } 4380d0abb36dSTheodore Ts'o if (err) 4381d0abb36dSTheodore Ts'o return err; 438273c384c0STheodore Ts'o retry_remove_space: 438373c384c0STheodore Ts'o err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1); 438473c384c0STheodore Ts'o if (err == -ENOMEM) { 438573c384c0STheodore Ts'o cond_resched(); 438673c384c0STheodore Ts'o congestion_wait(BLK_RW_ASYNC, HZ/50); 438773c384c0STheodore Ts'o goto retry_remove_space; 438873c384c0STheodore Ts'o } 438973c384c0STheodore Ts'o return err; 4390a86c6181SAlex Tomas } 4391a86c6181SAlex Tomas 43920e8b6879SLukas Czerner static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset, 4393c174e6d6SDmitry Monakhov ext4_lblk_t len, loff_t new_size, 439477a2e84dSTahsin Erdogan int flags) 4395a2df2a63SAmit Arora { 4396496ad9aaSAl Viro struct inode *inode = file_inode(file); 4397a2df2a63SAmit Arora handle_t *handle; 4398a2df2a63SAmit Arora int ret = 0; 43994209ae12SHarshad Shirwadkar int ret2 = 0, ret3 = 0; 4400a2df2a63SAmit Arora int retries = 0; 44014134f5c8SLukas Czerner int depth = 0; 44022ed88685STheodore Ts'o struct ext4_map_blocks map; 44030e8b6879SLukas Czerner unsigned int credits; 4404c174e6d6SDmitry Monakhov loff_t epos; 4405a2df2a63SAmit Arora 4406c3fe493cSFabian Frederick BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)); 44070e8b6879SLukas Czerner map.m_lblk = offset; 4408c174e6d6SDmitry Monakhov map.m_len = len; 44093c6fe770SGreg Harm /* 44103c6fe770SGreg Harm * Don't normalize the request if it can fit in one extent so 44113c6fe770SGreg Harm * that it doesn't get unnecessarily split into multiple 44123c6fe770SGreg Harm * extents. 44133c6fe770SGreg Harm */ 4414556615dcSLukas Czerner if (len <= EXT_UNWRITTEN_MAX_LEN) 44153c6fe770SGreg Harm flags |= EXT4_GET_BLOCKS_NO_NORMALIZE; 441660d4616fSDmitry Monakhov 44170e8b6879SLukas Czerner /* 44180e8b6879SLukas Czerner * credits to insert 1 extent into extent tree 44190e8b6879SLukas Czerner */ 44200e8b6879SLukas Czerner credits = ext4_chunk_trans_blocks(inode, len); 44214134f5c8SLukas Czerner depth = ext_depth(inode); 44220e8b6879SLukas Czerner 4423a2df2a63SAmit Arora retry: 4424c174e6d6SDmitry Monakhov while (ret >= 0 && len) { 44254134f5c8SLukas Czerner /* 44264134f5c8SLukas Czerner * Recalculate credits when extent tree depth changes. 44274134f5c8SLukas Czerner */ 4428011c88e3SDan Carpenter if (depth != ext_depth(inode)) { 44294134f5c8SLukas Czerner credits = ext4_chunk_trans_blocks(inode, len); 44304134f5c8SLukas Czerner depth = ext_depth(inode); 44314134f5c8SLukas Czerner } 44324134f5c8SLukas Czerner 44339924a92aSTheodore Ts'o handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, 44349924a92aSTheodore Ts'o credits); 4435a2df2a63SAmit Arora if (IS_ERR(handle)) { 4436a2df2a63SAmit Arora ret = PTR_ERR(handle); 4437a2df2a63SAmit Arora break; 4438a2df2a63SAmit Arora } 4439a4e5d88bSDmitry Monakhov ret = ext4_map_blocks(handle, inode, &map, flags); 4440221879c9SAneesh Kumar K.V if (ret <= 0) { 4441f282ac19SLukas Czerner ext4_debug("inode #%lu: block %u: len %u: " 4442b06acd38SLukas Czerner "ext4_ext_map_blocks returned %d", 4443b06acd38SLukas Czerner inode->i_ino, map.m_lblk, 4444b06acd38SLukas Czerner map.m_len, ret); 4445a2df2a63SAmit Arora ext4_mark_inode_dirty(handle, inode); 4446a2df2a63SAmit Arora ret2 = ext4_journal_stop(handle); 4447a2df2a63SAmit Arora break; 4448a2df2a63SAmit Arora } 4449c174e6d6SDmitry Monakhov map.m_lblk += ret; 4450c174e6d6SDmitry Monakhov map.m_len = len = len - ret; 4451c174e6d6SDmitry Monakhov epos = (loff_t)map.m_lblk << inode->i_blkbits; 4452eeca7ea1SDeepa Dinamani inode->i_ctime = current_time(inode); 4453c174e6d6SDmitry Monakhov if (new_size) { 4454c174e6d6SDmitry Monakhov if (epos > new_size) 4455c174e6d6SDmitry Monakhov epos = new_size; 4456c174e6d6SDmitry Monakhov if (ext4_update_inode_size(inode, epos) & 0x1) 4457c174e6d6SDmitry Monakhov inode->i_mtime = inode->i_ctime; 4458c174e6d6SDmitry Monakhov } 44594209ae12SHarshad Shirwadkar ret2 = ext4_mark_inode_dirty(handle, inode); 4460c894aa97SEryu Guan ext4_update_inode_fsync_trans(handle, inode, 1); 44614209ae12SHarshad Shirwadkar ret3 = ext4_journal_stop(handle); 44624209ae12SHarshad Shirwadkar ret2 = ret3 ? ret3 : ret2; 44634209ae12SHarshad Shirwadkar if (unlikely(ret2)) 4464a2df2a63SAmit Arora break; 4465a2df2a63SAmit Arora } 4466fd28784aSAneesh Kumar K.V if (ret == -ENOSPC && 4467fd28784aSAneesh Kumar K.V ext4_should_retry_alloc(inode->i_sb, &retries)) { 4468fd28784aSAneesh Kumar K.V ret = 0; 4469a2df2a63SAmit Arora goto retry; 4470a2df2a63SAmit Arora } 4471f282ac19SLukas Czerner 44720e8b6879SLukas Czerner return ret > 0 ? ret2 : ret; 44730e8b6879SLukas Czerner } 44740e8b6879SLukas Czerner 447543f81677SEric Biggers static int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len); 447643f81677SEric Biggers 447743f81677SEric Biggers static int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len); 447843f81677SEric Biggers 4479b8a86845SLukas Czerner static long ext4_zero_range(struct file *file, loff_t offset, 4480b8a86845SLukas Czerner loff_t len, int mode) 4481b8a86845SLukas Czerner { 4482b8a86845SLukas Czerner struct inode *inode = file_inode(file); 4483b8a86845SLukas Czerner handle_t *handle = NULL; 4484b8a86845SLukas Czerner unsigned int max_blocks; 4485b8a86845SLukas Czerner loff_t new_size = 0; 4486b8a86845SLukas Czerner int ret = 0; 4487b8a86845SLukas Czerner int flags; 448869dc9536SDmitry Monakhov int credits; 4489c174e6d6SDmitry Monakhov int partial_begin, partial_end; 4490b8a86845SLukas Czerner loff_t start, end; 4491b8a86845SLukas Czerner ext4_lblk_t lblk; 4492b8a86845SLukas Czerner unsigned int blkbits = inode->i_blkbits; 4493b8a86845SLukas Czerner 4494b8a86845SLukas Czerner trace_ext4_zero_range(inode, offset, len, mode); 4495b8a86845SLukas Czerner 4496e1ee60fdSNamjae Jeon /* Call ext4_force_commit to flush all data in case of data=journal. */ 4497e1ee60fdSNamjae Jeon if (ext4_should_journal_data(inode)) { 4498e1ee60fdSNamjae Jeon ret = ext4_force_commit(inode->i_sb); 4499e1ee60fdSNamjae Jeon if (ret) 4500e1ee60fdSNamjae Jeon return ret; 4501e1ee60fdSNamjae Jeon } 4502e1ee60fdSNamjae Jeon 4503b8a86845SLukas Czerner /* 4504b8a86845SLukas Czerner * Round up offset. This is not fallocate, we neet to zero out 4505b8a86845SLukas Czerner * blocks, so convert interior block aligned part of the range to 4506b8a86845SLukas Czerner * unwritten and possibly manually zero out unaligned parts of the 4507b8a86845SLukas Czerner * range. 4508b8a86845SLukas Czerner */ 4509b8a86845SLukas Czerner start = round_up(offset, 1 << blkbits); 4510b8a86845SLukas Czerner end = round_down((offset + len), 1 << blkbits); 4511b8a86845SLukas Czerner 4512b8a86845SLukas Czerner if (start < offset || end > offset + len) 4513b8a86845SLukas Czerner return -EINVAL; 4514c174e6d6SDmitry Monakhov partial_begin = offset & ((1 << blkbits) - 1); 4515c174e6d6SDmitry Monakhov partial_end = (offset + len) & ((1 << blkbits) - 1); 4516b8a86845SLukas Czerner 4517b8a86845SLukas Czerner lblk = start >> blkbits; 4518b8a86845SLukas Czerner max_blocks = (end >> blkbits); 4519b8a86845SLukas Czerner if (max_blocks < lblk) 4520b8a86845SLukas Czerner max_blocks = 0; 4521b8a86845SLukas Czerner else 4522b8a86845SLukas Czerner max_blocks -= lblk; 4523b8a86845SLukas Czerner 45245955102cSAl Viro inode_lock(inode); 4525b8a86845SLukas Czerner 4526b8a86845SLukas Czerner /* 452780dd4978SChristophe JAILLET * Indirect files do not support unwritten extents 4528b8a86845SLukas Czerner */ 4529b8a86845SLukas Czerner if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) { 4530b8a86845SLukas Czerner ret = -EOPNOTSUPP; 4531b8a86845SLukas Czerner goto out_mutex; 4532b8a86845SLukas Czerner } 4533b8a86845SLukas Czerner 4534b8a86845SLukas Czerner if (!(mode & FALLOC_FL_KEEP_SIZE) && 45359b02e498SEric Biggers (offset + len > inode->i_size || 453651e3ae81STheodore Ts'o offset + len > EXT4_I(inode)->i_disksize)) { 4537b8a86845SLukas Czerner new_size = offset + len; 4538b8a86845SLukas Czerner ret = inode_newsize_ok(inode, new_size); 4539b8a86845SLukas Czerner if (ret) 4540b8a86845SLukas Czerner goto out_mutex; 4541b8a86845SLukas Czerner } 4542b8a86845SLukas Czerner 45430f2af21aSLukas Czerner flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT; 45440f2af21aSLukas Czerner 454517048e8aSJan Kara /* Wait all existing dio workers, newcomers will block on i_mutex */ 454617048e8aSJan Kara inode_dio_wait(inode); 454717048e8aSJan Kara 45480f2af21aSLukas Czerner /* Preallocate the range including the unaligned edges */ 45490f2af21aSLukas Czerner if (partial_begin || partial_end) { 45500f2af21aSLukas Czerner ret = ext4_alloc_file_blocks(file, 45510f2af21aSLukas Czerner round_down(offset, 1 << blkbits) >> blkbits, 45520f2af21aSLukas Czerner (round_up((offset + len), 1 << blkbits) - 45530f2af21aSLukas Czerner round_down(offset, 1 << blkbits)) >> blkbits, 455477a2e84dSTahsin Erdogan new_size, flags); 45550f2af21aSLukas Czerner if (ret) 45561d39834fSNikolay Borisov goto out_mutex; 45570f2af21aSLukas Czerner 45580f2af21aSLukas Czerner } 45590f2af21aSLukas Czerner 45600f2af21aSLukas Czerner /* Zero range excluding the unaligned edges */ 4561b8a86845SLukas Czerner if (max_blocks > 0) { 45620f2af21aSLukas Czerner flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN | 45630f2af21aSLukas Czerner EXT4_EX_NOCACHE); 4564b8a86845SLukas Czerner 4565ea3d7209SJan Kara /* 4566ea3d7209SJan Kara * Prevent page faults from reinstantiating pages we have 4567ea3d7209SJan Kara * released from page cache. 4568ea3d7209SJan Kara */ 4569ea3d7209SJan Kara down_write(&EXT4_I(inode)->i_mmap_sem); 4570430657b6SRoss Zwisler 4571430657b6SRoss Zwisler ret = ext4_break_layouts(inode); 4572430657b6SRoss Zwisler if (ret) { 4573430657b6SRoss Zwisler up_write(&EXT4_I(inode)->i_mmap_sem); 4574430657b6SRoss Zwisler goto out_mutex; 4575430657b6SRoss Zwisler } 4576430657b6SRoss Zwisler 457701127848SJan Kara ret = ext4_update_disksize_before_punch(inode, offset, len); 457801127848SJan Kara if (ret) { 457901127848SJan Kara up_write(&EXT4_I(inode)->i_mmap_sem); 45801d39834fSNikolay Borisov goto out_mutex; 458101127848SJan Kara } 4582ea3d7209SJan Kara /* Now release the pages and zero block aligned part of pages */ 4583ea3d7209SJan Kara truncate_pagecache_range(inode, start, end - 1); 4584eeca7ea1SDeepa Dinamani inode->i_mtime = inode->i_ctime = current_time(inode); 4585ea3d7209SJan Kara 4586c174e6d6SDmitry Monakhov ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, 458777a2e84dSTahsin Erdogan flags); 4588ea3d7209SJan Kara up_write(&EXT4_I(inode)->i_mmap_sem); 4589b8a86845SLukas Czerner if (ret) 45901d39834fSNikolay Borisov goto out_mutex; 4591b8a86845SLukas Czerner } 4592c174e6d6SDmitry Monakhov if (!partial_begin && !partial_end) 45931d39834fSNikolay Borisov goto out_mutex; 4594c174e6d6SDmitry Monakhov 459569dc9536SDmitry Monakhov /* 459669dc9536SDmitry Monakhov * In worst case we have to writeout two nonadjacent unwritten 459769dc9536SDmitry Monakhov * blocks and update the inode 459869dc9536SDmitry Monakhov */ 459969dc9536SDmitry Monakhov credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1; 460069dc9536SDmitry Monakhov if (ext4_should_journal_data(inode)) 460169dc9536SDmitry Monakhov credits += 2; 460269dc9536SDmitry Monakhov handle = ext4_journal_start(inode, EXT4_HT_MISC, credits); 4603b8a86845SLukas Czerner if (IS_ERR(handle)) { 4604b8a86845SLukas Czerner ret = PTR_ERR(handle); 4605b8a86845SLukas Czerner ext4_std_error(inode->i_sb, ret); 46061d39834fSNikolay Borisov goto out_mutex; 4607b8a86845SLukas Czerner } 4608b8a86845SLukas Czerner 4609eeca7ea1SDeepa Dinamani inode->i_mtime = inode->i_ctime = current_time(inode); 46104337ecd1SEric Whitney if (new_size) 46114631dbf6SDmitry Monakhov ext4_update_inode_size(inode, new_size); 46124209ae12SHarshad Shirwadkar ret = ext4_mark_inode_dirty(handle, inode); 46134209ae12SHarshad Shirwadkar if (unlikely(ret)) 46144209ae12SHarshad Shirwadkar goto out_handle; 4615b8a86845SLukas Czerner 4616b8a86845SLukas Czerner /* Zero out partial block at the edges of the range */ 4617b8a86845SLukas Czerner ret = ext4_zero_partial_blocks(handle, inode, offset, len); 461867a7d5f5SJan Kara if (ret >= 0) 461967a7d5f5SJan Kara ext4_update_inode_fsync_trans(handle, inode, 1); 4620b8a86845SLukas Czerner 4621b8a86845SLukas Czerner if (file->f_flags & O_SYNC) 4622b8a86845SLukas Czerner ext4_handle_sync(handle); 4623b8a86845SLukas Czerner 46244209ae12SHarshad Shirwadkar out_handle: 4625b8a86845SLukas Czerner ext4_journal_stop(handle); 4626b8a86845SLukas Czerner out_mutex: 46275955102cSAl Viro inode_unlock(inode); 4628b8a86845SLukas Czerner return ret; 4629b8a86845SLukas Czerner } 4630b8a86845SLukas Czerner 46310e8b6879SLukas Czerner /* 46320e8b6879SLukas Czerner * preallocate space for a file. This implements ext4's fallocate file 46330e8b6879SLukas Czerner * operation, which gets called from sys_fallocate system call. 46340e8b6879SLukas Czerner * For block-mapped files, posix_fallocate should fall back to the method 46350e8b6879SLukas Czerner * of writing zeroes to the required new blocks (the same behavior which is 46360e8b6879SLukas Czerner * expected for file systems which do not support fallocate() system call). 46370e8b6879SLukas Czerner */ 46380e8b6879SLukas Czerner long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len) 46390e8b6879SLukas Czerner { 46400e8b6879SLukas Czerner struct inode *inode = file_inode(file); 46410e8b6879SLukas Czerner loff_t new_size = 0; 46420e8b6879SLukas Czerner unsigned int max_blocks; 46430e8b6879SLukas Czerner int ret = 0; 46440e8b6879SLukas Czerner int flags; 46450e8b6879SLukas Czerner ext4_lblk_t lblk; 46460e8b6879SLukas Czerner unsigned int blkbits = inode->i_blkbits; 46470e8b6879SLukas Czerner 46482058f83aSMichael Halcrow /* 46492058f83aSMichael Halcrow * Encrypted inodes can't handle collapse range or insert 46502058f83aSMichael Halcrow * range since we would need to re-encrypt blocks with a 46512058f83aSMichael Halcrow * different IV or XTS tweak (which are based on the logical 46522058f83aSMichael Halcrow * block number). 46532058f83aSMichael Halcrow */ 4654592ddec7SChandan Rajendra if (IS_ENCRYPTED(inode) && 4655457b1e35SEric Biggers (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE))) 46562058f83aSMichael Halcrow return -EOPNOTSUPP; 46572058f83aSMichael Halcrow 46580e8b6879SLukas Czerner /* Return error if mode is not supported */ 46590e8b6879SLukas Czerner if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | 4660331573feSNamjae Jeon FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | 4661331573feSNamjae Jeon FALLOC_FL_INSERT_RANGE)) 46620e8b6879SLukas Czerner return -EOPNOTSUPP; 46630e8b6879SLukas Czerner 46640e8b6879SLukas Czerner if (mode & FALLOC_FL_PUNCH_HOLE) 46650e8b6879SLukas Czerner return ext4_punch_hole(inode, offset, len); 46660e8b6879SLukas Czerner 46670e8b6879SLukas Czerner ret = ext4_convert_inline_data(inode); 46680e8b6879SLukas Czerner if (ret) 46690e8b6879SLukas Czerner return ret; 46700e8b6879SLukas Czerner 467140c406c7STheodore Ts'o if (mode & FALLOC_FL_COLLAPSE_RANGE) 467240c406c7STheodore Ts'o return ext4_collapse_range(inode, offset, len); 467340c406c7STheodore Ts'o 4674331573feSNamjae Jeon if (mode & FALLOC_FL_INSERT_RANGE) 4675331573feSNamjae Jeon return ext4_insert_range(inode, offset, len); 4676331573feSNamjae Jeon 4677b8a86845SLukas Czerner if (mode & FALLOC_FL_ZERO_RANGE) 4678b8a86845SLukas Czerner return ext4_zero_range(file, offset, len, mode); 4679b8a86845SLukas Czerner 46800e8b6879SLukas Czerner trace_ext4_fallocate_enter(inode, offset, len, mode); 46810e8b6879SLukas Czerner lblk = offset >> blkbits; 46820e8b6879SLukas Czerner 4683518eaa63SFabian Frederick max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits); 4684556615dcSLukas Czerner flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT; 46850e8b6879SLukas Czerner 46865955102cSAl Viro inode_lock(inode); 46870e8b6879SLukas Czerner 4688280227a7SDavide Italiano /* 4689280227a7SDavide Italiano * We only support preallocation for extent-based files only 4690280227a7SDavide Italiano */ 4691280227a7SDavide Italiano if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) { 4692280227a7SDavide Italiano ret = -EOPNOTSUPP; 4693280227a7SDavide Italiano goto out; 4694280227a7SDavide Italiano } 4695280227a7SDavide Italiano 46960e8b6879SLukas Czerner if (!(mode & FALLOC_FL_KEEP_SIZE) && 46979b02e498SEric Biggers (offset + len > inode->i_size || 469851e3ae81STheodore Ts'o offset + len > EXT4_I(inode)->i_disksize)) { 46990e8b6879SLukas Czerner new_size = offset + len; 47000e8b6879SLukas Czerner ret = inode_newsize_ok(inode, new_size); 47010e8b6879SLukas Czerner if (ret) 47020e8b6879SLukas Czerner goto out; 47030e8b6879SLukas Czerner } 47040e8b6879SLukas Czerner 470517048e8aSJan Kara /* Wait all existing dio workers, newcomers will block on i_mutex */ 470617048e8aSJan Kara inode_dio_wait(inode); 470717048e8aSJan Kara 470877a2e84dSTahsin Erdogan ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags); 47090e8b6879SLukas Czerner if (ret) 47100e8b6879SLukas Czerner goto out; 47110e8b6879SLukas Czerner 4712c174e6d6SDmitry Monakhov if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) { 4713c174e6d6SDmitry Monakhov ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal, 4714c174e6d6SDmitry Monakhov EXT4_I(inode)->i_sync_tid); 4715f282ac19SLukas Czerner } 4716f282ac19SLukas Czerner out: 47175955102cSAl Viro inode_unlock(inode); 47180e8b6879SLukas Czerner trace_ext4_fallocate_exit(inode, offset, max_blocks, ret); 47190e8b6879SLukas Czerner return ret; 4720a2df2a63SAmit Arora } 47216873fa0dSEric Sandeen 47226873fa0dSEric Sandeen /* 47230031462bSMingming Cao * This function convert a range of blocks to written extents 47240031462bSMingming Cao * The caller of this function will pass the start offset and the size. 47250031462bSMingming Cao * all unwritten extents within this range will be converted to 47260031462bSMingming Cao * written extents. 47270031462bSMingming Cao * 47280031462bSMingming Cao * This function is called from the direct IO end io call back 47290031462bSMingming Cao * function, to convert the fallocated extents after IO is completed. 4730109f5565SMingming * Returns 0 on success. 47310031462bSMingming Cao */ 47326b523df4SJan Kara int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode, 47336b523df4SJan Kara loff_t offset, ssize_t len) 47340031462bSMingming Cao { 47350031462bSMingming Cao unsigned int max_blocks; 47364209ae12SHarshad Shirwadkar int ret = 0, ret2 = 0, ret3 = 0; 47372ed88685STheodore Ts'o struct ext4_map_blocks map; 4738a00713eaSRitesh Harjani unsigned int blkbits = inode->i_blkbits; 4739a00713eaSRitesh Harjani unsigned int credits = 0; 47400031462bSMingming Cao 47412ed88685STheodore Ts'o map.m_lblk = offset >> blkbits; 4742518eaa63SFabian Frederick max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits); 4743518eaa63SFabian Frederick 4744a00713eaSRitesh Harjani if (!handle) { 47456b523df4SJan Kara /* 47460031462bSMingming Cao * credits to insert 1 extent into extent tree 47470031462bSMingming Cao */ 47480031462bSMingming Cao credits = ext4_chunk_trans_blocks(inode, max_blocks); 47496b523df4SJan Kara } 47500031462bSMingming Cao while (ret >= 0 && ret < max_blocks) { 47512ed88685STheodore Ts'o map.m_lblk += ret; 47522ed88685STheodore Ts'o map.m_len = (max_blocks -= ret); 47536b523df4SJan Kara if (credits) { 47546b523df4SJan Kara handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, 47556b523df4SJan Kara credits); 47560031462bSMingming Cao if (IS_ERR(handle)) { 47570031462bSMingming Cao ret = PTR_ERR(handle); 47580031462bSMingming Cao break; 47590031462bSMingming Cao } 47606b523df4SJan Kara } 47612ed88685STheodore Ts'o ret = ext4_map_blocks(handle, inode, &map, 4762c7064ef1SJiaying Zhang EXT4_GET_BLOCKS_IO_CONVERT_EXT); 4763b06acd38SLukas Czerner if (ret <= 0) 4764b06acd38SLukas Czerner ext4_warning(inode->i_sb, 4765b06acd38SLukas Czerner "inode #%lu: block %u: len %u: " 476692b97816STheodore Ts'o "ext4_ext_map_blocks returned %d", 4767b06acd38SLukas Czerner inode->i_ino, map.m_lblk, 476892b97816STheodore Ts'o map.m_len, ret); 47694209ae12SHarshad Shirwadkar ret2 = ext4_mark_inode_dirty(handle, inode); 47704209ae12SHarshad Shirwadkar if (credits) { 47714209ae12SHarshad Shirwadkar ret3 = ext4_journal_stop(handle); 47724209ae12SHarshad Shirwadkar if (unlikely(ret3)) 47734209ae12SHarshad Shirwadkar ret2 = ret3; 47744209ae12SHarshad Shirwadkar } 47754209ae12SHarshad Shirwadkar 47760031462bSMingming Cao if (ret <= 0 || ret2) 47770031462bSMingming Cao break; 47780031462bSMingming Cao } 47790031462bSMingming Cao return ret > 0 ? ret2 : ret; 47800031462bSMingming Cao } 47816d9c85ebSYongqiang Yang 4782a00713eaSRitesh Harjani int ext4_convert_unwritten_io_end_vec(handle_t *handle, ext4_io_end_t *io_end) 4783a00713eaSRitesh Harjani { 4784a00713eaSRitesh Harjani int ret, err = 0; 4785c8cc8816SRitesh Harjani struct ext4_io_end_vec *io_end_vec; 4786a00713eaSRitesh Harjani 4787a00713eaSRitesh Harjani /* 4788a00713eaSRitesh Harjani * This is somewhat ugly but the idea is clear: When transaction is 4789a00713eaSRitesh Harjani * reserved, everything goes into it. Otherwise we rather start several 4790a00713eaSRitesh Harjani * smaller transactions for conversion of each extent separately. 4791a00713eaSRitesh Harjani */ 4792a00713eaSRitesh Harjani if (handle) { 4793a00713eaSRitesh Harjani handle = ext4_journal_start_reserved(handle, 4794a00713eaSRitesh Harjani EXT4_HT_EXT_CONVERT); 4795a00713eaSRitesh Harjani if (IS_ERR(handle)) 4796a00713eaSRitesh Harjani return PTR_ERR(handle); 4797a00713eaSRitesh Harjani } 4798a00713eaSRitesh Harjani 4799c8cc8816SRitesh Harjani list_for_each_entry(io_end_vec, &io_end->list_vec, list) { 4800a00713eaSRitesh Harjani ret = ext4_convert_unwritten_extents(handle, io_end->inode, 4801c8cc8816SRitesh Harjani io_end_vec->offset, 4802c8cc8816SRitesh Harjani io_end_vec->size); 4803c8cc8816SRitesh Harjani if (ret) 4804c8cc8816SRitesh Harjani break; 4805c8cc8816SRitesh Harjani } 4806c8cc8816SRitesh Harjani 4807a00713eaSRitesh Harjani if (handle) 4808a00713eaSRitesh Harjani err = ext4_journal_stop(handle); 4809a00713eaSRitesh Harjani 4810a00713eaSRitesh Harjani return ret < 0 ? ret : err; 4811a00713eaSRitesh Harjani } 4812a00713eaSRitesh Harjani 4813d3b6f23fSRitesh Harjani static int ext4_iomap_xattr_fiemap(struct inode *inode, struct iomap *iomap) 48146873fa0dSEric Sandeen { 48156873fa0dSEric Sandeen __u64 physical = 0; 4816d3b6f23fSRitesh Harjani __u64 length = 0; 48176873fa0dSEric Sandeen int blockbits = inode->i_sb->s_blocksize_bits; 48186873fa0dSEric Sandeen int error = 0; 4819d3b6f23fSRitesh Harjani u16 iomap_type; 48206873fa0dSEric Sandeen 48216873fa0dSEric Sandeen /* in-inode? */ 482219f5fb7aSTheodore Ts'o if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) { 48236873fa0dSEric Sandeen struct ext4_iloc iloc; 48246873fa0dSEric Sandeen int offset; /* offset of xattr in inode */ 48256873fa0dSEric Sandeen 48266873fa0dSEric Sandeen error = ext4_get_inode_loc(inode, &iloc); 48276873fa0dSEric Sandeen if (error) 48286873fa0dSEric Sandeen return error; 4829a60697f4SJan Kara physical = (__u64)iloc.bh->b_blocknr << blockbits; 48306873fa0dSEric Sandeen offset = EXT4_GOOD_OLD_INODE_SIZE + 48316873fa0dSEric Sandeen EXT4_I(inode)->i_extra_isize; 48326873fa0dSEric Sandeen physical += offset; 48336873fa0dSEric Sandeen length = EXT4_SB(inode->i_sb)->s_inode_size - offset; 4834fd2dd9fbSCurt Wohlgemuth brelse(iloc.bh); 4835d3b6f23fSRitesh Harjani iomap_type = IOMAP_INLINE; 4836d3b6f23fSRitesh Harjani } else if (EXT4_I(inode)->i_file_acl) { /* external block */ 4837a60697f4SJan Kara physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits; 48386873fa0dSEric Sandeen length = inode->i_sb->s_blocksize; 4839d3b6f23fSRitesh Harjani iomap_type = IOMAP_MAPPED; 4840d3b6f23fSRitesh Harjani } else { 4841d3b6f23fSRitesh Harjani /* no in-inode or external block for xattr, so return -ENOENT */ 4842d3b6f23fSRitesh Harjani error = -ENOENT; 4843d3b6f23fSRitesh Harjani goto out; 48446873fa0dSEric Sandeen } 48456873fa0dSEric Sandeen 4846d3b6f23fSRitesh Harjani iomap->addr = physical; 4847d3b6f23fSRitesh Harjani iomap->offset = 0; 4848d3b6f23fSRitesh Harjani iomap->length = length; 4849d3b6f23fSRitesh Harjani iomap->type = iomap_type; 4850d3b6f23fSRitesh Harjani iomap->flags = 0; 4851d3b6f23fSRitesh Harjani out: 4852d3b6f23fSRitesh Harjani return error; 48536873fa0dSEric Sandeen } 48546873fa0dSEric Sandeen 4855d3b6f23fSRitesh Harjani static int ext4_iomap_xattr_begin(struct inode *inode, loff_t offset, 4856d3b6f23fSRitesh Harjani loff_t length, unsigned flags, 4857d3b6f23fSRitesh Harjani struct iomap *iomap, struct iomap *srcmap) 4858d3b6f23fSRitesh Harjani { 4859d3b6f23fSRitesh Harjani int error; 4860d3b6f23fSRitesh Harjani 4861d3b6f23fSRitesh Harjani error = ext4_iomap_xattr_fiemap(inode, iomap); 4862d3b6f23fSRitesh Harjani if (error == 0 && (offset >= iomap->length)) 4863d3b6f23fSRitesh Harjani error = -ENOENT; 4864d3b6f23fSRitesh Harjani return error; 4865d3b6f23fSRitesh Harjani } 4866d3b6f23fSRitesh Harjani 4867d3b6f23fSRitesh Harjani static const struct iomap_ops ext4_iomap_xattr_ops = { 4868d3b6f23fSRitesh Harjani .iomap_begin = ext4_iomap_xattr_begin, 4869d3b6f23fSRitesh Harjani }; 4870d3b6f23fSRitesh Harjani 4871d3b6f23fSRitesh Harjani static int _ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo, 4872d3b6f23fSRitesh Harjani __u64 start, __u64 len, bool from_es_cache) 48736873fa0dSEric Sandeen { 48746873fa0dSEric Sandeen ext4_lblk_t start_blk; 4875bb5835edSTheodore Ts'o u32 ext4_fiemap_flags = FIEMAP_FLAG_SYNC | FIEMAP_FLAG_XATTR; 48766873fa0dSEric Sandeen int error = 0; 48776873fa0dSEric Sandeen 48787869a4a6STheodore Ts'o if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) { 48797869a4a6STheodore Ts'o error = ext4_ext_precache(inode); 48807869a4a6STheodore Ts'o if (error) 48817869a4a6STheodore Ts'o return error; 4882bb5835edSTheodore Ts'o fieinfo->fi_flags &= ~FIEMAP_FLAG_CACHE; 48837869a4a6STheodore Ts'o } 48847869a4a6STheodore Ts'o 4885d3b6f23fSRitesh Harjani if (from_es_cache) 4886bb5835edSTheodore Ts'o ext4_fiemap_flags &= FIEMAP_FLAG_XATTR; 4887d3b6f23fSRitesh Harjani 4888bb5835edSTheodore Ts'o if (fiemap_check_flags(fieinfo, ext4_fiemap_flags)) 48896873fa0dSEric Sandeen return -EBADR; 48906873fa0dSEric Sandeen 48916873fa0dSEric Sandeen if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) { 4892d3b6f23fSRitesh Harjani fieinfo->fi_flags &= ~FIEMAP_FLAG_XATTR; 4893d3b6f23fSRitesh Harjani error = iomap_fiemap(inode, fieinfo, start, len, 4894d3b6f23fSRitesh Harjani &ext4_iomap_xattr_ops); 4895d3b6f23fSRitesh Harjani } else if (!from_es_cache) { 4896d3b6f23fSRitesh Harjani error = iomap_fiemap(inode, fieinfo, start, len, 4897d3b6f23fSRitesh Harjani &ext4_iomap_report_ops); 48986873fa0dSEric Sandeen } else { 4899aca92ff6SLeonard Michlmayr ext4_lblk_t len_blks; 4900aca92ff6SLeonard Michlmayr __u64 last_blk; 4901aca92ff6SLeonard Michlmayr 49026873fa0dSEric Sandeen start_blk = start >> inode->i_sb->s_blocksize_bits; 4903aca92ff6SLeonard Michlmayr last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits; 4904f17722f9SLukas Czerner if (last_blk >= EXT_MAX_BLOCKS) 4905f17722f9SLukas Czerner last_blk = EXT_MAX_BLOCKS-1; 4906aca92ff6SLeonard Michlmayr len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1; 49076873fa0dSEric Sandeen 49086873fa0dSEric Sandeen /* 490991dd8c11SLukas Czerner * Walk the extent tree gathering extent information 491091dd8c11SLukas Czerner * and pushing extents back to the user. 49116873fa0dSEric Sandeen */ 4912d3b6f23fSRitesh Harjani error = ext4_fill_es_cache_info(inode, start_blk, len_blks, 4913d3b6f23fSRitesh Harjani fieinfo); 49146873fa0dSEric Sandeen } 49156873fa0dSEric Sandeen return error; 49166873fa0dSEric Sandeen } 49179eb79482SNamjae Jeon 4918bb5835edSTheodore Ts'o int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo, 4919bb5835edSTheodore Ts'o __u64 start, __u64 len) 4920bb5835edSTheodore Ts'o { 4921d3b6f23fSRitesh Harjani return _ext4_fiemap(inode, fieinfo, start, len, false); 4922bb5835edSTheodore Ts'o } 4923bb5835edSTheodore Ts'o 4924bb5835edSTheodore Ts'o int ext4_get_es_cache(struct inode *inode, struct fiemap_extent_info *fieinfo, 4925bb5835edSTheodore Ts'o __u64 start, __u64 len) 4926bb5835edSTheodore Ts'o { 4927bb5835edSTheodore Ts'o if (ext4_has_inline_data(inode)) { 4928bb5835edSTheodore Ts'o int has_inline; 4929bb5835edSTheodore Ts'o 4930bb5835edSTheodore Ts'o down_read(&EXT4_I(inode)->xattr_sem); 4931bb5835edSTheodore Ts'o has_inline = ext4_has_inline_data(inode); 4932bb5835edSTheodore Ts'o up_read(&EXT4_I(inode)->xattr_sem); 4933bb5835edSTheodore Ts'o if (has_inline) 4934bb5835edSTheodore Ts'o return 0; 4935bb5835edSTheodore Ts'o } 4936bb5835edSTheodore Ts'o 4937d3b6f23fSRitesh Harjani return _ext4_fiemap(inode, fieinfo, start, len, true); 4938bb5835edSTheodore Ts'o } 4939bb5835edSTheodore Ts'o 4940bb5835edSTheodore Ts'o 49419eb79482SNamjae Jeon /* 49429eb79482SNamjae Jeon * ext4_access_path: 49439eb79482SNamjae Jeon * Function to access the path buffer for marking it dirty. 49449eb79482SNamjae Jeon * It also checks if there are sufficient credits left in the journal handle 49459eb79482SNamjae Jeon * to update path. 49469eb79482SNamjae Jeon */ 49479eb79482SNamjae Jeon static int 49489eb79482SNamjae Jeon ext4_access_path(handle_t *handle, struct inode *inode, 49499eb79482SNamjae Jeon struct ext4_ext_path *path) 49509eb79482SNamjae Jeon { 49519eb79482SNamjae Jeon int credits, err; 49529eb79482SNamjae Jeon 49539eb79482SNamjae Jeon if (!ext4_handle_valid(handle)) 49549eb79482SNamjae Jeon return 0; 49559eb79482SNamjae Jeon 49569eb79482SNamjae Jeon /* 49579eb79482SNamjae Jeon * Check if need to extend journal credits 49589eb79482SNamjae Jeon * 3 for leaf, sb, and inode plus 2 (bmap and group 49599eb79482SNamjae Jeon * descriptor) for each block group; assume two block 49609eb79482SNamjae Jeon * groups 49619eb79482SNamjae Jeon */ 49629eb79482SNamjae Jeon credits = ext4_writepage_trans_blocks(inode); 496383448bdfSJan Kara err = ext4_datasem_ensure_credits(handle, inode, 7, credits, 0); 4964a4130367SJan Kara if (err < 0) 49659eb79482SNamjae Jeon return err; 49669eb79482SNamjae Jeon 49679eb79482SNamjae Jeon err = ext4_ext_get_access(handle, inode, path); 49689eb79482SNamjae Jeon return err; 49699eb79482SNamjae Jeon } 49709eb79482SNamjae Jeon 49719eb79482SNamjae Jeon /* 49729eb79482SNamjae Jeon * ext4_ext_shift_path_extents: 49739eb79482SNamjae Jeon * Shift the extents of a path structure lying between path[depth].p_ext 4974331573feSNamjae Jeon * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells 4975331573feSNamjae Jeon * if it is right shift or left shift operation. 49769eb79482SNamjae Jeon */ 49779eb79482SNamjae Jeon static int 49789eb79482SNamjae Jeon ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift, 49799eb79482SNamjae Jeon struct inode *inode, handle_t *handle, 4980331573feSNamjae Jeon enum SHIFT_DIRECTION SHIFT) 49819eb79482SNamjae Jeon { 49829eb79482SNamjae Jeon int depth, err = 0; 49839eb79482SNamjae Jeon struct ext4_extent *ex_start, *ex_last; 49844756ee18Szhengbin bool update = false; 49859eb79482SNamjae Jeon depth = path->p_depth; 49869eb79482SNamjae Jeon 49879eb79482SNamjae Jeon while (depth >= 0) { 49889eb79482SNamjae Jeon if (depth == path->p_depth) { 49899eb79482SNamjae Jeon ex_start = path[depth].p_ext; 49909eb79482SNamjae Jeon if (!ex_start) 49916a797d27SDarrick J. Wong return -EFSCORRUPTED; 49929eb79482SNamjae Jeon 49939eb79482SNamjae Jeon ex_last = EXT_LAST_EXTENT(path[depth].p_hdr); 49949eb79482SNamjae Jeon 49959eb79482SNamjae Jeon err = ext4_access_path(handle, inode, path + depth); 49969eb79482SNamjae Jeon if (err) 49979eb79482SNamjae Jeon goto out; 49989eb79482SNamjae Jeon 49999eb79482SNamjae Jeon if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr)) 50004756ee18Szhengbin update = true; 50019eb79482SNamjae Jeon 50029eb79482SNamjae Jeon while (ex_start <= ex_last) { 5003331573feSNamjae Jeon if (SHIFT == SHIFT_LEFT) { 5004331573feSNamjae Jeon le32_add_cpu(&ex_start->ee_block, 5005331573feSNamjae Jeon -shift); 50066dd834efSLukas Czerner /* Try to merge to the left. */ 50076dd834efSLukas Czerner if ((ex_start > 5008331573feSNamjae Jeon EXT_FIRST_EXTENT(path[depth].p_hdr)) 5009331573feSNamjae Jeon && 50106dd834efSLukas Czerner ext4_ext_try_to_merge_right(inode, 50119eb79482SNamjae Jeon path, ex_start - 1)) 50129eb79482SNamjae Jeon ex_last--; 50136dd834efSLukas Czerner else 50149eb79482SNamjae Jeon ex_start++; 5015331573feSNamjae Jeon } else { 5016331573feSNamjae Jeon le32_add_cpu(&ex_last->ee_block, shift); 5017331573feSNamjae Jeon ext4_ext_try_to_merge_right(inode, path, 5018331573feSNamjae Jeon ex_last); 5019331573feSNamjae Jeon ex_last--; 5020331573feSNamjae Jeon } 50219eb79482SNamjae Jeon } 50229eb79482SNamjae Jeon err = ext4_ext_dirty(handle, inode, path + depth); 50239eb79482SNamjae Jeon if (err) 50249eb79482SNamjae Jeon goto out; 50259eb79482SNamjae Jeon 50269eb79482SNamjae Jeon if (--depth < 0 || !update) 50279eb79482SNamjae Jeon break; 50289eb79482SNamjae Jeon } 50299eb79482SNamjae Jeon 50309eb79482SNamjae Jeon /* Update index too */ 50319eb79482SNamjae Jeon err = ext4_access_path(handle, inode, path + depth); 50329eb79482SNamjae Jeon if (err) 50339eb79482SNamjae Jeon goto out; 50349eb79482SNamjae Jeon 5035331573feSNamjae Jeon if (SHIFT == SHIFT_LEFT) 5036847c6c42SZheng Liu le32_add_cpu(&path[depth].p_idx->ei_block, -shift); 5037331573feSNamjae Jeon else 5038331573feSNamjae Jeon le32_add_cpu(&path[depth].p_idx->ei_block, shift); 50399eb79482SNamjae Jeon err = ext4_ext_dirty(handle, inode, path + depth); 50409eb79482SNamjae Jeon if (err) 50419eb79482SNamjae Jeon goto out; 50429eb79482SNamjae Jeon 50439eb79482SNamjae Jeon /* we are done if current index is not a starting index */ 50449eb79482SNamjae Jeon if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr)) 50459eb79482SNamjae Jeon break; 50469eb79482SNamjae Jeon 50479eb79482SNamjae Jeon depth--; 50489eb79482SNamjae Jeon } 50499eb79482SNamjae Jeon 50509eb79482SNamjae Jeon out: 50519eb79482SNamjae Jeon return err; 50529eb79482SNamjae Jeon } 50539eb79482SNamjae Jeon 50549eb79482SNamjae Jeon /* 50559eb79482SNamjae Jeon * ext4_ext_shift_extents: 5056331573feSNamjae Jeon * All the extents which lies in the range from @start to the last allocated 5057331573feSNamjae Jeon * block for the @inode are shifted either towards left or right (depending 5058331573feSNamjae Jeon * upon @SHIFT) by @shift blocks. 50599eb79482SNamjae Jeon * On success, 0 is returned, error otherwise. 50609eb79482SNamjae Jeon */ 50619eb79482SNamjae Jeon static int 50629eb79482SNamjae Jeon ext4_ext_shift_extents(struct inode *inode, handle_t *handle, 5063331573feSNamjae Jeon ext4_lblk_t start, ext4_lblk_t shift, 5064331573feSNamjae Jeon enum SHIFT_DIRECTION SHIFT) 50659eb79482SNamjae Jeon { 50669eb79482SNamjae Jeon struct ext4_ext_path *path; 50679eb79482SNamjae Jeon int ret = 0, depth; 50689eb79482SNamjae Jeon struct ext4_extent *extent; 5069331573feSNamjae Jeon ext4_lblk_t stop, *iterator, ex_start, ex_end; 50709eb79482SNamjae Jeon 50719eb79482SNamjae Jeon /* Let path point to the last extent */ 507203e916faSRoman Pen path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 507303e916faSRoman Pen EXT4_EX_NOCACHE); 50749eb79482SNamjae Jeon if (IS_ERR(path)) 50759eb79482SNamjae Jeon return PTR_ERR(path); 50769eb79482SNamjae Jeon 50779eb79482SNamjae Jeon depth = path->p_depth; 50789eb79482SNamjae Jeon extent = path[depth].p_ext; 5079ee4bd0d9STheodore Ts'o if (!extent) 5080ee4bd0d9STheodore Ts'o goto out; 50819eb79482SNamjae Jeon 50822a9b8cbaSRoman Pen stop = le32_to_cpu(extent->ee_block); 50839eb79482SNamjae Jeon 50849eb79482SNamjae Jeon /* 5085349fa7d6SEric Biggers * For left shifts, make sure the hole on the left is big enough to 5086349fa7d6SEric Biggers * accommodate the shift. For right shifts, make sure the last extent 5087349fa7d6SEric Biggers * won't be shifted beyond EXT_MAX_BLOCKS. 50889eb79482SNamjae Jeon */ 5089331573feSNamjae Jeon if (SHIFT == SHIFT_LEFT) { 509003e916faSRoman Pen path = ext4_find_extent(inode, start - 1, &path, 509103e916faSRoman Pen EXT4_EX_NOCACHE); 50928dc79ec4SDmitry Monakhov if (IS_ERR(path)) 50938dc79ec4SDmitry Monakhov return PTR_ERR(path); 50949eb79482SNamjae Jeon depth = path->p_depth; 50959eb79482SNamjae Jeon extent = path[depth].p_ext; 50968dc79ec4SDmitry Monakhov if (extent) { 5097847c6c42SZheng Liu ex_start = le32_to_cpu(extent->ee_block); 5098847c6c42SZheng Liu ex_end = le32_to_cpu(extent->ee_block) + 5099847c6c42SZheng Liu ext4_ext_get_actual_len(extent); 51008dc79ec4SDmitry Monakhov } else { 51018dc79ec4SDmitry Monakhov ex_start = 0; 51028dc79ec4SDmitry Monakhov ex_end = 0; 51038dc79ec4SDmitry Monakhov } 51049eb79482SNamjae Jeon 51059eb79482SNamjae Jeon if ((start == ex_start && shift > ex_start) || 5106331573feSNamjae Jeon (shift > start - ex_end)) { 5107349fa7d6SEric Biggers ret = -EINVAL; 5108349fa7d6SEric Biggers goto out; 5109349fa7d6SEric Biggers } 5110349fa7d6SEric Biggers } else { 5111349fa7d6SEric Biggers if (shift > EXT_MAX_BLOCKS - 5112349fa7d6SEric Biggers (stop + ext4_ext_get_actual_len(extent))) { 5113349fa7d6SEric Biggers ret = -EINVAL; 5114349fa7d6SEric Biggers goto out; 5115331573feSNamjae Jeon } 5116331573feSNamjae Jeon } 5117331573feSNamjae Jeon 5118331573feSNamjae Jeon /* 5119331573feSNamjae Jeon * In case of left shift, iterator points to start and it is increased 5120331573feSNamjae Jeon * till we reach stop. In case of right shift, iterator points to stop 5121331573feSNamjae Jeon * and it is decreased till we reach start. 5122331573feSNamjae Jeon */ 5123331573feSNamjae Jeon if (SHIFT == SHIFT_LEFT) 5124331573feSNamjae Jeon iterator = &start; 5125331573feSNamjae Jeon else 5126331573feSNamjae Jeon iterator = &stop; 51279eb79482SNamjae Jeon 51282a9b8cbaSRoman Pen /* 51292a9b8cbaSRoman Pen * Its safe to start updating extents. Start and stop are unsigned, so 51302a9b8cbaSRoman Pen * in case of right shift if extent with 0 block is reached, iterator 51312a9b8cbaSRoman Pen * becomes NULL to indicate the end of the loop. 51322a9b8cbaSRoman Pen */ 51332a9b8cbaSRoman Pen while (iterator && start <= stop) { 513403e916faSRoman Pen path = ext4_find_extent(inode, *iterator, &path, 513503e916faSRoman Pen EXT4_EX_NOCACHE); 51369eb79482SNamjae Jeon if (IS_ERR(path)) 51379eb79482SNamjae Jeon return PTR_ERR(path); 51389eb79482SNamjae Jeon depth = path->p_depth; 51399eb79482SNamjae Jeon extent = path[depth].p_ext; 5140a18ed359SDmitry Monakhov if (!extent) { 5141a18ed359SDmitry Monakhov EXT4_ERROR_INODE(inode, "unexpected hole at %lu", 5142331573feSNamjae Jeon (unsigned long) *iterator); 51436a797d27SDarrick J. Wong return -EFSCORRUPTED; 5144a18ed359SDmitry Monakhov } 5145331573feSNamjae Jeon if (SHIFT == SHIFT_LEFT && *iterator > 5146331573feSNamjae Jeon le32_to_cpu(extent->ee_block)) { 51479eb79482SNamjae Jeon /* Hole, move to the next extent */ 5148f8fb4f41SDmitry Monakhov if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) { 5149f8fb4f41SDmitry Monakhov path[depth].p_ext++; 5150f8fb4f41SDmitry Monakhov } else { 5151331573feSNamjae Jeon *iterator = ext4_ext_next_allocated_block(path); 5152f8fb4f41SDmitry Monakhov continue; 51539eb79482SNamjae Jeon } 51549eb79482SNamjae Jeon } 5155331573feSNamjae Jeon 5156331573feSNamjae Jeon if (SHIFT == SHIFT_LEFT) { 5157331573feSNamjae Jeon extent = EXT_LAST_EXTENT(path[depth].p_hdr); 5158331573feSNamjae Jeon *iterator = le32_to_cpu(extent->ee_block) + 5159331573feSNamjae Jeon ext4_ext_get_actual_len(extent); 5160331573feSNamjae Jeon } else { 5161331573feSNamjae Jeon extent = EXT_FIRST_EXTENT(path[depth].p_hdr); 51622a9b8cbaSRoman Pen if (le32_to_cpu(extent->ee_block) > 0) 51632a9b8cbaSRoman Pen *iterator = le32_to_cpu(extent->ee_block) - 1; 51642a9b8cbaSRoman Pen else 51652a9b8cbaSRoman Pen /* Beginning is reached, end of the loop */ 51662a9b8cbaSRoman Pen iterator = NULL; 5167331573feSNamjae Jeon /* Update path extent in case we need to stop */ 5168331573feSNamjae Jeon while (le32_to_cpu(extent->ee_block) < start) 5169331573feSNamjae Jeon extent++; 5170331573feSNamjae Jeon path[depth].p_ext = extent; 5171331573feSNamjae Jeon } 51729eb79482SNamjae Jeon ret = ext4_ext_shift_path_extents(path, shift, inode, 5173331573feSNamjae Jeon handle, SHIFT); 51749eb79482SNamjae Jeon if (ret) 51759eb79482SNamjae Jeon break; 51769eb79482SNamjae Jeon } 5177ee4bd0d9STheodore Ts'o out: 5178ee4bd0d9STheodore Ts'o ext4_ext_drop_refs(path); 5179ee4bd0d9STheodore Ts'o kfree(path); 51809eb79482SNamjae Jeon return ret; 51819eb79482SNamjae Jeon } 51829eb79482SNamjae Jeon 51839eb79482SNamjae Jeon /* 51849eb79482SNamjae Jeon * ext4_collapse_range: 51859eb79482SNamjae Jeon * This implements the fallocate's collapse range functionality for ext4 51869eb79482SNamjae Jeon * Returns: 0 and non-zero on error. 51879eb79482SNamjae Jeon */ 518843f81677SEric Biggers static int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len) 51899eb79482SNamjae Jeon { 51909eb79482SNamjae Jeon struct super_block *sb = inode->i_sb; 51919eb79482SNamjae Jeon ext4_lblk_t punch_start, punch_stop; 51929eb79482SNamjae Jeon handle_t *handle; 51939eb79482SNamjae Jeon unsigned int credits; 5194a8680e0dSNamjae Jeon loff_t new_size, ioffset; 51959eb79482SNamjae Jeon int ret; 51969eb79482SNamjae Jeon 5197b9576fc3STheodore Ts'o /* 5198b9576fc3STheodore Ts'o * We need to test this early because xfstests assumes that a 5199b9576fc3STheodore Ts'o * collapse range of (0, 1) will return EOPNOTSUPP if the file 5200b9576fc3STheodore Ts'o * system does not support collapse range. 5201b9576fc3STheodore Ts'o */ 5202b9576fc3STheodore Ts'o if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) 5203b9576fc3STheodore Ts'o return -EOPNOTSUPP; 5204b9576fc3STheodore Ts'o 52059b02e498SEric Biggers /* Collapse range works only on fs cluster size aligned regions. */ 52069b02e498SEric Biggers if (!IS_ALIGNED(offset | len, EXT4_CLUSTER_SIZE(sb))) 52079eb79482SNamjae Jeon return -EINVAL; 52089eb79482SNamjae Jeon 52099eb79482SNamjae Jeon trace_ext4_collapse_range(inode, offset, len); 52109eb79482SNamjae Jeon 52119eb79482SNamjae Jeon punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb); 52129eb79482SNamjae Jeon punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb); 52139eb79482SNamjae Jeon 52141ce01c4aSNamjae Jeon /* Call ext4_force_commit to flush all data in case of data=journal. */ 52151ce01c4aSNamjae Jeon if (ext4_should_journal_data(inode)) { 52161ce01c4aSNamjae Jeon ret = ext4_force_commit(inode->i_sb); 52171ce01c4aSNamjae Jeon if (ret) 52181ce01c4aSNamjae Jeon return ret; 52191ce01c4aSNamjae Jeon } 52201ce01c4aSNamjae Jeon 52215955102cSAl Viro inode_lock(inode); 522223fffa92SLukas Czerner /* 522323fffa92SLukas Czerner * There is no need to overlap collapse range with EOF, in which case 522423fffa92SLukas Czerner * it is effectively a truncate operation 522523fffa92SLukas Czerner */ 52269b02e498SEric Biggers if (offset + len >= inode->i_size) { 522723fffa92SLukas Czerner ret = -EINVAL; 522823fffa92SLukas Czerner goto out_mutex; 522923fffa92SLukas Czerner } 523023fffa92SLukas Czerner 52319eb79482SNamjae Jeon /* Currently just for extent based files */ 52329eb79482SNamjae Jeon if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) { 52339eb79482SNamjae Jeon ret = -EOPNOTSUPP; 52349eb79482SNamjae Jeon goto out_mutex; 52359eb79482SNamjae Jeon } 52369eb79482SNamjae Jeon 52379eb79482SNamjae Jeon /* Wait for existing dio to complete */ 52389eb79482SNamjae Jeon inode_dio_wait(inode); 52399eb79482SNamjae Jeon 5240ea3d7209SJan Kara /* 5241ea3d7209SJan Kara * Prevent page faults from reinstantiating pages we have released from 5242ea3d7209SJan Kara * page cache. 5243ea3d7209SJan Kara */ 5244ea3d7209SJan Kara down_write(&EXT4_I(inode)->i_mmap_sem); 5245430657b6SRoss Zwisler 5246430657b6SRoss Zwisler ret = ext4_break_layouts(inode); 5247430657b6SRoss Zwisler if (ret) 5248430657b6SRoss Zwisler goto out_mmap; 5249430657b6SRoss Zwisler 525032ebffd3SJan Kara /* 525132ebffd3SJan Kara * Need to round down offset to be aligned with page size boundary 525232ebffd3SJan Kara * for page size > block size. 525332ebffd3SJan Kara */ 525432ebffd3SJan Kara ioffset = round_down(offset, PAGE_SIZE); 525532ebffd3SJan Kara /* 525632ebffd3SJan Kara * Write tail of the last page before removed range since it will get 525732ebffd3SJan Kara * removed from the page cache below. 525832ebffd3SJan Kara */ 525932ebffd3SJan Kara ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset); 526032ebffd3SJan Kara if (ret) 526132ebffd3SJan Kara goto out_mmap; 526232ebffd3SJan Kara /* 526332ebffd3SJan Kara * Write data that will be shifted to preserve them when discarding 526432ebffd3SJan Kara * page cache below. We are also protected from pages becoming dirty 526532ebffd3SJan Kara * by i_mmap_sem. 526632ebffd3SJan Kara */ 526732ebffd3SJan Kara ret = filemap_write_and_wait_range(inode->i_mapping, offset + len, 526832ebffd3SJan Kara LLONG_MAX); 526932ebffd3SJan Kara if (ret) 527032ebffd3SJan Kara goto out_mmap; 5271ea3d7209SJan Kara truncate_pagecache(inode, ioffset); 5272ea3d7209SJan Kara 52739eb79482SNamjae Jeon credits = ext4_writepage_trans_blocks(inode); 52749eb79482SNamjae Jeon handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits); 52759eb79482SNamjae Jeon if (IS_ERR(handle)) { 52769eb79482SNamjae Jeon ret = PTR_ERR(handle); 5277ea3d7209SJan Kara goto out_mmap; 52789eb79482SNamjae Jeon } 52799eb79482SNamjae Jeon 52809eb79482SNamjae Jeon down_write(&EXT4_I(inode)->i_data_sem); 52819eb79482SNamjae Jeon ext4_discard_preallocations(inode); 52829eb79482SNamjae Jeon 52839eb79482SNamjae Jeon ret = ext4_es_remove_extent(inode, punch_start, 52842c1d2328SLukas Czerner EXT_MAX_BLOCKS - punch_start); 52859eb79482SNamjae Jeon if (ret) { 52869eb79482SNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 52879eb79482SNamjae Jeon goto out_stop; 52889eb79482SNamjae Jeon } 52899eb79482SNamjae Jeon 52909eb79482SNamjae Jeon ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1); 52919eb79482SNamjae Jeon if (ret) { 52929eb79482SNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 52939eb79482SNamjae Jeon goto out_stop; 52949eb79482SNamjae Jeon } 5295ef24f6c2SLukas Czerner ext4_discard_preallocations(inode); 52969eb79482SNamjae Jeon 52979eb79482SNamjae Jeon ret = ext4_ext_shift_extents(inode, handle, punch_stop, 5298331573feSNamjae Jeon punch_stop - punch_start, SHIFT_LEFT); 52999eb79482SNamjae Jeon if (ret) { 53009eb79482SNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 53019eb79482SNamjae Jeon goto out_stop; 53029eb79482SNamjae Jeon } 53039eb79482SNamjae Jeon 53049b02e498SEric Biggers new_size = inode->i_size - len; 53059337d5d3SLukas Czerner i_size_write(inode, new_size); 53069eb79482SNamjae Jeon EXT4_I(inode)->i_disksize = new_size; 53079eb79482SNamjae Jeon 53089eb79482SNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 53099eb79482SNamjae Jeon if (IS_SYNC(inode)) 53109eb79482SNamjae Jeon ext4_handle_sync(handle); 5311eeca7ea1SDeepa Dinamani inode->i_mtime = inode->i_ctime = current_time(inode); 53124209ae12SHarshad Shirwadkar ret = ext4_mark_inode_dirty(handle, inode); 531367a7d5f5SJan Kara ext4_update_inode_fsync_trans(handle, inode, 1); 53149eb79482SNamjae Jeon 53159eb79482SNamjae Jeon out_stop: 53169eb79482SNamjae Jeon ext4_journal_stop(handle); 5317ea3d7209SJan Kara out_mmap: 5318ea3d7209SJan Kara up_write(&EXT4_I(inode)->i_mmap_sem); 53199eb79482SNamjae Jeon out_mutex: 53205955102cSAl Viro inode_unlock(inode); 53219eb79482SNamjae Jeon return ret; 53229eb79482SNamjae Jeon } 5323fcf6b1b7SDmitry Monakhov 5324331573feSNamjae Jeon /* 5325331573feSNamjae Jeon * ext4_insert_range: 5326331573feSNamjae Jeon * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate. 5327331573feSNamjae Jeon * The data blocks starting from @offset to the EOF are shifted by @len 5328331573feSNamjae Jeon * towards right to create a hole in the @inode. Inode size is increased 5329331573feSNamjae Jeon * by len bytes. 5330331573feSNamjae Jeon * Returns 0 on success, error otherwise. 5331331573feSNamjae Jeon */ 533243f81677SEric Biggers static int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len) 5333331573feSNamjae Jeon { 5334331573feSNamjae Jeon struct super_block *sb = inode->i_sb; 5335331573feSNamjae Jeon handle_t *handle; 5336331573feSNamjae Jeon struct ext4_ext_path *path; 5337331573feSNamjae Jeon struct ext4_extent *extent; 5338331573feSNamjae Jeon ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0; 5339331573feSNamjae Jeon unsigned int credits, ee_len; 5340331573feSNamjae Jeon int ret = 0, depth, split_flag = 0; 5341331573feSNamjae Jeon loff_t ioffset; 5342331573feSNamjae Jeon 5343331573feSNamjae Jeon /* 5344331573feSNamjae Jeon * We need to test this early because xfstests assumes that an 5345331573feSNamjae Jeon * insert range of (0, 1) will return EOPNOTSUPP if the file 5346331573feSNamjae Jeon * system does not support insert range. 5347331573feSNamjae Jeon */ 5348331573feSNamjae Jeon if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) 5349331573feSNamjae Jeon return -EOPNOTSUPP; 5350331573feSNamjae Jeon 53519b02e498SEric Biggers /* Insert range works only on fs cluster size aligned regions. */ 53529b02e498SEric Biggers if (!IS_ALIGNED(offset | len, EXT4_CLUSTER_SIZE(sb))) 5353331573feSNamjae Jeon return -EINVAL; 5354331573feSNamjae Jeon 5355331573feSNamjae Jeon trace_ext4_insert_range(inode, offset, len); 5356331573feSNamjae Jeon 5357331573feSNamjae Jeon offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb); 5358331573feSNamjae Jeon len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb); 5359331573feSNamjae Jeon 5360331573feSNamjae Jeon /* Call ext4_force_commit to flush all data in case of data=journal */ 5361331573feSNamjae Jeon if (ext4_should_journal_data(inode)) { 5362331573feSNamjae Jeon ret = ext4_force_commit(inode->i_sb); 5363331573feSNamjae Jeon if (ret) 5364331573feSNamjae Jeon return ret; 5365331573feSNamjae Jeon } 5366331573feSNamjae Jeon 53675955102cSAl Viro inode_lock(inode); 5368331573feSNamjae Jeon /* Currently just for extent based files */ 5369331573feSNamjae Jeon if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) { 5370331573feSNamjae Jeon ret = -EOPNOTSUPP; 5371331573feSNamjae Jeon goto out_mutex; 5372331573feSNamjae Jeon } 5373331573feSNamjae Jeon 53749b02e498SEric Biggers /* Check whether the maximum file size would be exceeded */ 53759b02e498SEric Biggers if (len > inode->i_sb->s_maxbytes - inode->i_size) { 5376331573feSNamjae Jeon ret = -EFBIG; 5377331573feSNamjae Jeon goto out_mutex; 5378331573feSNamjae Jeon } 5379331573feSNamjae Jeon 53809b02e498SEric Biggers /* Offset must be less than i_size */ 53819b02e498SEric Biggers if (offset >= inode->i_size) { 5382331573feSNamjae Jeon ret = -EINVAL; 5383331573feSNamjae Jeon goto out_mutex; 5384331573feSNamjae Jeon } 5385331573feSNamjae Jeon 5386331573feSNamjae Jeon /* Wait for existing dio to complete */ 5387331573feSNamjae Jeon inode_dio_wait(inode); 5388331573feSNamjae Jeon 5389ea3d7209SJan Kara /* 5390ea3d7209SJan Kara * Prevent page faults from reinstantiating pages we have released from 5391ea3d7209SJan Kara * page cache. 5392ea3d7209SJan Kara */ 5393ea3d7209SJan Kara down_write(&EXT4_I(inode)->i_mmap_sem); 5394430657b6SRoss Zwisler 5395430657b6SRoss Zwisler ret = ext4_break_layouts(inode); 5396430657b6SRoss Zwisler if (ret) 5397430657b6SRoss Zwisler goto out_mmap; 5398430657b6SRoss Zwisler 539932ebffd3SJan Kara /* 540032ebffd3SJan Kara * Need to round down to align start offset to page size boundary 540132ebffd3SJan Kara * for page size > block size. 540232ebffd3SJan Kara */ 540332ebffd3SJan Kara ioffset = round_down(offset, PAGE_SIZE); 540432ebffd3SJan Kara /* Write out all dirty pages */ 540532ebffd3SJan Kara ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, 540632ebffd3SJan Kara LLONG_MAX); 540732ebffd3SJan Kara if (ret) 540832ebffd3SJan Kara goto out_mmap; 5409ea3d7209SJan Kara truncate_pagecache(inode, ioffset); 5410ea3d7209SJan Kara 5411331573feSNamjae Jeon credits = ext4_writepage_trans_blocks(inode); 5412331573feSNamjae Jeon handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits); 5413331573feSNamjae Jeon if (IS_ERR(handle)) { 5414331573feSNamjae Jeon ret = PTR_ERR(handle); 5415ea3d7209SJan Kara goto out_mmap; 5416331573feSNamjae Jeon } 5417331573feSNamjae Jeon 5418331573feSNamjae Jeon /* Expand file to avoid data loss if there is error while shifting */ 5419331573feSNamjae Jeon inode->i_size += len; 5420331573feSNamjae Jeon EXT4_I(inode)->i_disksize += len; 5421eeca7ea1SDeepa Dinamani inode->i_mtime = inode->i_ctime = current_time(inode); 5422331573feSNamjae Jeon ret = ext4_mark_inode_dirty(handle, inode); 5423331573feSNamjae Jeon if (ret) 5424331573feSNamjae Jeon goto out_stop; 5425331573feSNamjae Jeon 5426331573feSNamjae Jeon down_write(&EXT4_I(inode)->i_data_sem); 5427331573feSNamjae Jeon ext4_discard_preallocations(inode); 5428331573feSNamjae Jeon 5429331573feSNamjae Jeon path = ext4_find_extent(inode, offset_lblk, NULL, 0); 5430331573feSNamjae Jeon if (IS_ERR(path)) { 5431331573feSNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 5432331573feSNamjae Jeon goto out_stop; 5433331573feSNamjae Jeon } 5434331573feSNamjae Jeon 5435331573feSNamjae Jeon depth = ext_depth(inode); 5436331573feSNamjae Jeon extent = path[depth].p_ext; 5437331573feSNamjae Jeon if (extent) { 5438331573feSNamjae Jeon ee_start_lblk = le32_to_cpu(extent->ee_block); 5439331573feSNamjae Jeon ee_len = ext4_ext_get_actual_len(extent); 5440331573feSNamjae Jeon 5441331573feSNamjae Jeon /* 5442331573feSNamjae Jeon * If offset_lblk is not the starting block of extent, split 5443331573feSNamjae Jeon * the extent @offset_lblk 5444331573feSNamjae Jeon */ 5445331573feSNamjae Jeon if ((offset_lblk > ee_start_lblk) && 5446331573feSNamjae Jeon (offset_lblk < (ee_start_lblk + ee_len))) { 5447331573feSNamjae Jeon if (ext4_ext_is_unwritten(extent)) 5448331573feSNamjae Jeon split_flag = EXT4_EXT_MARK_UNWRIT1 | 5449331573feSNamjae Jeon EXT4_EXT_MARK_UNWRIT2; 5450331573feSNamjae Jeon ret = ext4_split_extent_at(handle, inode, &path, 5451331573feSNamjae Jeon offset_lblk, split_flag, 5452331573feSNamjae Jeon EXT4_EX_NOCACHE | 5453331573feSNamjae Jeon EXT4_GET_BLOCKS_PRE_IO | 5454331573feSNamjae Jeon EXT4_GET_BLOCKS_METADATA_NOFAIL); 5455331573feSNamjae Jeon } 5456331573feSNamjae Jeon 5457331573feSNamjae Jeon ext4_ext_drop_refs(path); 5458331573feSNamjae Jeon kfree(path); 5459331573feSNamjae Jeon if (ret < 0) { 5460331573feSNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 5461331573feSNamjae Jeon goto out_stop; 5462331573feSNamjae Jeon } 5463edf15aa1SFabian Frederick } else { 5464edf15aa1SFabian Frederick ext4_ext_drop_refs(path); 5465edf15aa1SFabian Frederick kfree(path); 5466331573feSNamjae Jeon } 5467331573feSNamjae Jeon 5468331573feSNamjae Jeon ret = ext4_es_remove_extent(inode, offset_lblk, 5469331573feSNamjae Jeon EXT_MAX_BLOCKS - offset_lblk); 5470331573feSNamjae Jeon if (ret) { 5471331573feSNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 5472331573feSNamjae Jeon goto out_stop; 5473331573feSNamjae Jeon } 5474331573feSNamjae Jeon 5475331573feSNamjae Jeon /* 5476331573feSNamjae Jeon * if offset_lblk lies in a hole which is at start of file, use 5477331573feSNamjae Jeon * ee_start_lblk to shift extents 5478331573feSNamjae Jeon */ 5479331573feSNamjae Jeon ret = ext4_ext_shift_extents(inode, handle, 5480331573feSNamjae Jeon ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk, 5481331573feSNamjae Jeon len_lblk, SHIFT_RIGHT); 5482331573feSNamjae Jeon 5483331573feSNamjae Jeon up_write(&EXT4_I(inode)->i_data_sem); 5484331573feSNamjae Jeon if (IS_SYNC(inode)) 5485331573feSNamjae Jeon ext4_handle_sync(handle); 548667a7d5f5SJan Kara if (ret >= 0) 548767a7d5f5SJan Kara ext4_update_inode_fsync_trans(handle, inode, 1); 5488331573feSNamjae Jeon 5489331573feSNamjae Jeon out_stop: 5490331573feSNamjae Jeon ext4_journal_stop(handle); 5491ea3d7209SJan Kara out_mmap: 5492ea3d7209SJan Kara up_write(&EXT4_I(inode)->i_mmap_sem); 5493331573feSNamjae Jeon out_mutex: 54945955102cSAl Viro inode_unlock(inode); 5495331573feSNamjae Jeon return ret; 5496331573feSNamjae Jeon } 5497331573feSNamjae Jeon 5498fcf6b1b7SDmitry Monakhov /** 5499c60990b3STheodore Ts'o * ext4_swap_extents() - Swap extents between two inodes 5500c60990b3STheodore Ts'o * @handle: handle for this transaction 5501fcf6b1b7SDmitry Monakhov * @inode1: First inode 5502fcf6b1b7SDmitry Monakhov * @inode2: Second inode 5503fcf6b1b7SDmitry Monakhov * @lblk1: Start block for first inode 5504fcf6b1b7SDmitry Monakhov * @lblk2: Start block for second inode 5505fcf6b1b7SDmitry Monakhov * @count: Number of blocks to swap 5506dcae058aSzhenwei.pi * @unwritten: Mark second inode's extents as unwritten after swap 5507fcf6b1b7SDmitry Monakhov * @erp: Pointer to save error value 5508fcf6b1b7SDmitry Monakhov * 5509fcf6b1b7SDmitry Monakhov * This helper routine does exactly what is promise "swap extents". All other 5510fcf6b1b7SDmitry Monakhov * stuff such as page-cache locking consistency, bh mapping consistency or 5511fcf6b1b7SDmitry Monakhov * extent's data copying must be performed by caller. 5512fcf6b1b7SDmitry Monakhov * Locking: 5513fcf6b1b7SDmitry Monakhov * i_mutex is held for both inodes 5514fcf6b1b7SDmitry Monakhov * i_data_sem is locked for write for both inodes 5515fcf6b1b7SDmitry Monakhov * Assumptions: 5516fcf6b1b7SDmitry Monakhov * All pages from requested range are locked for both inodes 5517fcf6b1b7SDmitry Monakhov */ 5518fcf6b1b7SDmitry Monakhov int 5519fcf6b1b7SDmitry Monakhov ext4_swap_extents(handle_t *handle, struct inode *inode1, 5520fcf6b1b7SDmitry Monakhov struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2, 5521fcf6b1b7SDmitry Monakhov ext4_lblk_t count, int unwritten, int *erp) 5522fcf6b1b7SDmitry Monakhov { 5523fcf6b1b7SDmitry Monakhov struct ext4_ext_path *path1 = NULL; 5524fcf6b1b7SDmitry Monakhov struct ext4_ext_path *path2 = NULL; 5525fcf6b1b7SDmitry Monakhov int replaced_count = 0; 5526fcf6b1b7SDmitry Monakhov 5527fcf6b1b7SDmitry Monakhov BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem)); 5528fcf6b1b7SDmitry Monakhov BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem)); 55295955102cSAl Viro BUG_ON(!inode_is_locked(inode1)); 55305955102cSAl Viro BUG_ON(!inode_is_locked(inode2)); 5531fcf6b1b7SDmitry Monakhov 5532fcf6b1b7SDmitry Monakhov *erp = ext4_es_remove_extent(inode1, lblk1, count); 553319008f6dSTheodore Ts'o if (unlikely(*erp)) 5534fcf6b1b7SDmitry Monakhov return 0; 5535fcf6b1b7SDmitry Monakhov *erp = ext4_es_remove_extent(inode2, lblk2, count); 553619008f6dSTheodore Ts'o if (unlikely(*erp)) 5537fcf6b1b7SDmitry Monakhov return 0; 5538fcf6b1b7SDmitry Monakhov 5539fcf6b1b7SDmitry Monakhov while (count) { 5540fcf6b1b7SDmitry Monakhov struct ext4_extent *ex1, *ex2, tmp_ex; 5541fcf6b1b7SDmitry Monakhov ext4_lblk_t e1_blk, e2_blk; 5542fcf6b1b7SDmitry Monakhov int e1_len, e2_len, len; 5543fcf6b1b7SDmitry Monakhov int split = 0; 5544fcf6b1b7SDmitry Monakhov 5545ed8a1a76STheodore Ts'o path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE); 5546a1c83681SViresh Kumar if (IS_ERR(path1)) { 5547fcf6b1b7SDmitry Monakhov *erp = PTR_ERR(path1); 554819008f6dSTheodore Ts'o path1 = NULL; 554919008f6dSTheodore Ts'o finish: 555019008f6dSTheodore Ts'o count = 0; 555119008f6dSTheodore Ts'o goto repeat; 5552fcf6b1b7SDmitry Monakhov } 5553ed8a1a76STheodore Ts'o path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE); 5554a1c83681SViresh Kumar if (IS_ERR(path2)) { 5555fcf6b1b7SDmitry Monakhov *erp = PTR_ERR(path2); 555619008f6dSTheodore Ts'o path2 = NULL; 555719008f6dSTheodore Ts'o goto finish; 5558fcf6b1b7SDmitry Monakhov } 5559fcf6b1b7SDmitry Monakhov ex1 = path1[path1->p_depth].p_ext; 5560fcf6b1b7SDmitry Monakhov ex2 = path2[path2->p_depth].p_ext; 5561fcf6b1b7SDmitry Monakhov /* Do we have somthing to swap ? */ 5562fcf6b1b7SDmitry Monakhov if (unlikely(!ex2 || !ex1)) 556319008f6dSTheodore Ts'o goto finish; 5564fcf6b1b7SDmitry Monakhov 5565fcf6b1b7SDmitry Monakhov e1_blk = le32_to_cpu(ex1->ee_block); 5566fcf6b1b7SDmitry Monakhov e2_blk = le32_to_cpu(ex2->ee_block); 5567fcf6b1b7SDmitry Monakhov e1_len = ext4_ext_get_actual_len(ex1); 5568fcf6b1b7SDmitry Monakhov e2_len = ext4_ext_get_actual_len(ex2); 5569fcf6b1b7SDmitry Monakhov 5570fcf6b1b7SDmitry Monakhov /* Hole handling */ 5571fcf6b1b7SDmitry Monakhov if (!in_range(lblk1, e1_blk, e1_len) || 5572fcf6b1b7SDmitry Monakhov !in_range(lblk2, e2_blk, e2_len)) { 5573fcf6b1b7SDmitry Monakhov ext4_lblk_t next1, next2; 5574fcf6b1b7SDmitry Monakhov 5575fcf6b1b7SDmitry Monakhov /* if hole after extent, then go to next extent */ 5576fcf6b1b7SDmitry Monakhov next1 = ext4_ext_next_allocated_block(path1); 5577fcf6b1b7SDmitry Monakhov next2 = ext4_ext_next_allocated_block(path2); 5578fcf6b1b7SDmitry Monakhov /* If hole before extent, then shift to that extent */ 5579fcf6b1b7SDmitry Monakhov if (e1_blk > lblk1) 5580fcf6b1b7SDmitry Monakhov next1 = e1_blk; 5581fcf6b1b7SDmitry Monakhov if (e2_blk > lblk2) 55824e562013SManinder Singh next2 = e2_blk; 5583fcf6b1b7SDmitry Monakhov /* Do we have something to swap */ 5584fcf6b1b7SDmitry Monakhov if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS) 558519008f6dSTheodore Ts'o goto finish; 5586fcf6b1b7SDmitry Monakhov /* Move to the rightest boundary */ 5587fcf6b1b7SDmitry Monakhov len = next1 - lblk1; 5588fcf6b1b7SDmitry Monakhov if (len < next2 - lblk2) 5589fcf6b1b7SDmitry Monakhov len = next2 - lblk2; 5590fcf6b1b7SDmitry Monakhov if (len > count) 5591fcf6b1b7SDmitry Monakhov len = count; 5592fcf6b1b7SDmitry Monakhov lblk1 += len; 5593fcf6b1b7SDmitry Monakhov lblk2 += len; 5594fcf6b1b7SDmitry Monakhov count -= len; 5595fcf6b1b7SDmitry Monakhov goto repeat; 5596fcf6b1b7SDmitry Monakhov } 5597fcf6b1b7SDmitry Monakhov 5598fcf6b1b7SDmitry Monakhov /* Prepare left boundary */ 5599fcf6b1b7SDmitry Monakhov if (e1_blk < lblk1) { 5600fcf6b1b7SDmitry Monakhov split = 1; 5601fcf6b1b7SDmitry Monakhov *erp = ext4_force_split_extent_at(handle, inode1, 5602dfe50809STheodore Ts'o &path1, lblk1, 0); 560319008f6dSTheodore Ts'o if (unlikely(*erp)) 560419008f6dSTheodore Ts'o goto finish; 5605fcf6b1b7SDmitry Monakhov } 5606fcf6b1b7SDmitry Monakhov if (e2_blk < lblk2) { 5607fcf6b1b7SDmitry Monakhov split = 1; 5608fcf6b1b7SDmitry Monakhov *erp = ext4_force_split_extent_at(handle, inode2, 5609dfe50809STheodore Ts'o &path2, lblk2, 0); 561019008f6dSTheodore Ts'o if (unlikely(*erp)) 561119008f6dSTheodore Ts'o goto finish; 5612fcf6b1b7SDmitry Monakhov } 5613dfe50809STheodore Ts'o /* ext4_split_extent_at() may result in leaf extent split, 5614fcf6b1b7SDmitry Monakhov * path must to be revalidated. */ 5615fcf6b1b7SDmitry Monakhov if (split) 5616fcf6b1b7SDmitry Monakhov goto repeat; 5617fcf6b1b7SDmitry Monakhov 5618fcf6b1b7SDmitry Monakhov /* Prepare right boundary */ 5619fcf6b1b7SDmitry Monakhov len = count; 5620fcf6b1b7SDmitry Monakhov if (len > e1_blk + e1_len - lblk1) 5621fcf6b1b7SDmitry Monakhov len = e1_blk + e1_len - lblk1; 5622fcf6b1b7SDmitry Monakhov if (len > e2_blk + e2_len - lblk2) 5623fcf6b1b7SDmitry Monakhov len = e2_blk + e2_len - lblk2; 5624fcf6b1b7SDmitry Monakhov 5625fcf6b1b7SDmitry Monakhov if (len != e1_len) { 5626fcf6b1b7SDmitry Monakhov split = 1; 5627fcf6b1b7SDmitry Monakhov *erp = ext4_force_split_extent_at(handle, inode1, 5628dfe50809STheodore Ts'o &path1, lblk1 + len, 0); 562919008f6dSTheodore Ts'o if (unlikely(*erp)) 563019008f6dSTheodore Ts'o goto finish; 5631fcf6b1b7SDmitry Monakhov } 5632fcf6b1b7SDmitry Monakhov if (len != e2_len) { 5633fcf6b1b7SDmitry Monakhov split = 1; 5634fcf6b1b7SDmitry Monakhov *erp = ext4_force_split_extent_at(handle, inode2, 5635dfe50809STheodore Ts'o &path2, lblk2 + len, 0); 5636fcf6b1b7SDmitry Monakhov if (*erp) 563719008f6dSTheodore Ts'o goto finish; 5638fcf6b1b7SDmitry Monakhov } 5639dfe50809STheodore Ts'o /* ext4_split_extent_at() may result in leaf extent split, 5640fcf6b1b7SDmitry Monakhov * path must to be revalidated. */ 5641fcf6b1b7SDmitry Monakhov if (split) 5642fcf6b1b7SDmitry Monakhov goto repeat; 5643fcf6b1b7SDmitry Monakhov 5644fcf6b1b7SDmitry Monakhov BUG_ON(e2_len != e1_len); 5645fcf6b1b7SDmitry Monakhov *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth); 564619008f6dSTheodore Ts'o if (unlikely(*erp)) 564719008f6dSTheodore Ts'o goto finish; 5648fcf6b1b7SDmitry Monakhov *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth); 564919008f6dSTheodore Ts'o if (unlikely(*erp)) 565019008f6dSTheodore Ts'o goto finish; 5651fcf6b1b7SDmitry Monakhov 5652fcf6b1b7SDmitry Monakhov /* Both extents are fully inside boundaries. Swap it now */ 5653fcf6b1b7SDmitry Monakhov tmp_ex = *ex1; 5654fcf6b1b7SDmitry Monakhov ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2)); 5655fcf6b1b7SDmitry Monakhov ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex)); 5656fcf6b1b7SDmitry Monakhov ex1->ee_len = cpu_to_le16(e2_len); 5657fcf6b1b7SDmitry Monakhov ex2->ee_len = cpu_to_le16(e1_len); 5658fcf6b1b7SDmitry Monakhov if (unwritten) 5659fcf6b1b7SDmitry Monakhov ext4_ext_mark_unwritten(ex2); 5660fcf6b1b7SDmitry Monakhov if (ext4_ext_is_unwritten(&tmp_ex)) 5661fcf6b1b7SDmitry Monakhov ext4_ext_mark_unwritten(ex1); 5662fcf6b1b7SDmitry Monakhov 5663fcf6b1b7SDmitry Monakhov ext4_ext_try_to_merge(handle, inode2, path2, ex2); 5664fcf6b1b7SDmitry Monakhov ext4_ext_try_to_merge(handle, inode1, path1, ex1); 5665fcf6b1b7SDmitry Monakhov *erp = ext4_ext_dirty(handle, inode2, path2 + 5666fcf6b1b7SDmitry Monakhov path2->p_depth); 566719008f6dSTheodore Ts'o if (unlikely(*erp)) 566819008f6dSTheodore Ts'o goto finish; 5669fcf6b1b7SDmitry Monakhov *erp = ext4_ext_dirty(handle, inode1, path1 + 5670fcf6b1b7SDmitry Monakhov path1->p_depth); 5671fcf6b1b7SDmitry Monakhov /* 5672fcf6b1b7SDmitry Monakhov * Looks scarry ah..? second inode already points to new blocks, 5673fcf6b1b7SDmitry Monakhov * and it was successfully dirtied. But luckily error may happen 5674fcf6b1b7SDmitry Monakhov * only due to journal error, so full transaction will be 5675fcf6b1b7SDmitry Monakhov * aborted anyway. 5676fcf6b1b7SDmitry Monakhov */ 567719008f6dSTheodore Ts'o if (unlikely(*erp)) 567819008f6dSTheodore Ts'o goto finish; 5679fcf6b1b7SDmitry Monakhov lblk1 += len; 5680fcf6b1b7SDmitry Monakhov lblk2 += len; 5681fcf6b1b7SDmitry Monakhov replaced_count += len; 5682fcf6b1b7SDmitry Monakhov count -= len; 5683fcf6b1b7SDmitry Monakhov 5684fcf6b1b7SDmitry Monakhov repeat: 5685fcf6b1b7SDmitry Monakhov ext4_ext_drop_refs(path1); 5686fcf6b1b7SDmitry Monakhov kfree(path1); 5687fcf6b1b7SDmitry Monakhov ext4_ext_drop_refs(path2); 5688fcf6b1b7SDmitry Monakhov kfree(path2); 5689b7ea89adSTheodore Ts'o path1 = path2 = NULL; 5690fcf6b1b7SDmitry Monakhov } 5691fcf6b1b7SDmitry Monakhov return replaced_count; 5692fcf6b1b7SDmitry Monakhov } 56930b02f4c0SEric Whitney 56940b02f4c0SEric Whitney /* 56950b02f4c0SEric Whitney * ext4_clu_mapped - determine whether any block in a logical cluster has 56960b02f4c0SEric Whitney * been mapped to a physical cluster 56970b02f4c0SEric Whitney * 56980b02f4c0SEric Whitney * @inode - file containing the logical cluster 56990b02f4c0SEric Whitney * @lclu - logical cluster of interest 57000b02f4c0SEric Whitney * 57010b02f4c0SEric Whitney * Returns 1 if any block in the logical cluster is mapped, signifying 57020b02f4c0SEric Whitney * that a physical cluster has been allocated for it. Otherwise, 57030b02f4c0SEric Whitney * returns 0. Can also return negative error codes. Derived from 57040b02f4c0SEric Whitney * ext4_ext_map_blocks(). 57050b02f4c0SEric Whitney */ 57060b02f4c0SEric Whitney int ext4_clu_mapped(struct inode *inode, ext4_lblk_t lclu) 57070b02f4c0SEric Whitney { 57080b02f4c0SEric Whitney struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb); 57090b02f4c0SEric Whitney struct ext4_ext_path *path; 57100b02f4c0SEric Whitney int depth, mapped = 0, err = 0; 57110b02f4c0SEric Whitney struct ext4_extent *extent; 57120b02f4c0SEric Whitney ext4_lblk_t first_lblk, first_lclu, last_lclu; 57130b02f4c0SEric Whitney 57140b02f4c0SEric Whitney /* search for the extent closest to the first block in the cluster */ 57150b02f4c0SEric Whitney path = ext4_find_extent(inode, EXT4_C2B(sbi, lclu), NULL, 0); 57160b02f4c0SEric Whitney if (IS_ERR(path)) { 57170b02f4c0SEric Whitney err = PTR_ERR(path); 57180b02f4c0SEric Whitney path = NULL; 57190b02f4c0SEric Whitney goto out; 57200b02f4c0SEric Whitney } 57210b02f4c0SEric Whitney 57220b02f4c0SEric Whitney depth = ext_depth(inode); 57230b02f4c0SEric Whitney 57240b02f4c0SEric Whitney /* 57250b02f4c0SEric Whitney * A consistent leaf must not be empty. This situation is possible, 57260b02f4c0SEric Whitney * though, _during_ tree modification, and it's why an assert can't 57270b02f4c0SEric Whitney * be put in ext4_find_extent(). 57280b02f4c0SEric Whitney */ 57290b02f4c0SEric Whitney if (unlikely(path[depth].p_ext == NULL && depth != 0)) { 57300b02f4c0SEric Whitney EXT4_ERROR_INODE(inode, 57310b02f4c0SEric Whitney "bad extent address - lblock: %lu, depth: %d, pblock: %lld", 57320b02f4c0SEric Whitney (unsigned long) EXT4_C2B(sbi, lclu), 57330b02f4c0SEric Whitney depth, path[depth].p_block); 57340b02f4c0SEric Whitney err = -EFSCORRUPTED; 57350b02f4c0SEric Whitney goto out; 57360b02f4c0SEric Whitney } 57370b02f4c0SEric Whitney 57380b02f4c0SEric Whitney extent = path[depth].p_ext; 57390b02f4c0SEric Whitney 57400b02f4c0SEric Whitney /* can't be mapped if the extent tree is empty */ 57410b02f4c0SEric Whitney if (extent == NULL) 57420b02f4c0SEric Whitney goto out; 57430b02f4c0SEric Whitney 57440b02f4c0SEric Whitney first_lblk = le32_to_cpu(extent->ee_block); 57450b02f4c0SEric Whitney first_lclu = EXT4_B2C(sbi, first_lblk); 57460b02f4c0SEric Whitney 57470b02f4c0SEric Whitney /* 57480b02f4c0SEric Whitney * Three possible outcomes at this point - found extent spanning 57490b02f4c0SEric Whitney * the target cluster, to the left of the target cluster, or to the 57500b02f4c0SEric Whitney * right of the target cluster. The first two cases are handled here. 57510b02f4c0SEric Whitney * The last case indicates the target cluster is not mapped. 57520b02f4c0SEric Whitney */ 57530b02f4c0SEric Whitney if (lclu >= first_lclu) { 57540b02f4c0SEric Whitney last_lclu = EXT4_B2C(sbi, first_lblk + 57550b02f4c0SEric Whitney ext4_ext_get_actual_len(extent) - 1); 57560b02f4c0SEric Whitney if (lclu <= last_lclu) { 57570b02f4c0SEric Whitney mapped = 1; 57580b02f4c0SEric Whitney } else { 57590b02f4c0SEric Whitney first_lblk = ext4_ext_next_allocated_block(path); 57600b02f4c0SEric Whitney first_lclu = EXT4_B2C(sbi, first_lblk); 57610b02f4c0SEric Whitney if (lclu == first_lclu) 57620b02f4c0SEric Whitney mapped = 1; 57630b02f4c0SEric Whitney } 57640b02f4c0SEric Whitney } 57650b02f4c0SEric Whitney 57660b02f4c0SEric Whitney out: 57670b02f4c0SEric Whitney ext4_ext_drop_refs(path); 57680b02f4c0SEric Whitney kfree(path); 57690b02f4c0SEric Whitney 57700b02f4c0SEric Whitney return err ? err : mapped; 57710b02f4c0SEric Whitney } 5772