fs/ext4/extents.c

   1 /*
   2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3  * Written by Alex Tomas <alex@clusterfs.com>
   4  *
   5  * Architecture independence:
   6  *   Copyright (c) 2005, Bull S.A.
   7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <asm/uaccess.h>
  41 #include <linux/fiemap.h>
  42 #include <linux/backing-dev.h>
  43 #include "ext4_jbd2.h"
  44 #include "ext4_extents.h"
  45 #include "xattr.h"
  46
  47 #include <trace/events/ext4.h>
  48
  49 /*
  50  * used by extent splitting.
  51  */
  52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  53                                         due to ENOSPC */
  54 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  55 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  56
  57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  59
  60 static __le32 ext4_extent_block_csum(struct inode *inode,
  61                                      struct ext4_extent_header *eh)
  62 {
  63         struct ext4_inode_info *ei = EXT4_I(inode);
  64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  65         __u32 csum;
  66
  67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  68                            EXT4_EXTENT_TAIL_OFFSET(eh));
  69         return cpu_to_le32(csum);
  70 }
  71
  72 static int ext4_extent_block_csum_verify(struct inode *inode,
  73                                          struct ext4_extent_header *eh)
  74 {
  75         struct ext4_extent_tail *et;
  76
  77         if (!ext4_has_metadata_csum(inode->i_sb))
  78                 return 1;
  79
  80         et = find_ext4_extent_tail(eh);
  81         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  82                 return 0;
  83         return 1;
  84 }
  85
  86 static void ext4_extent_block_csum_set(struct inode *inode,
  87                                        struct ext4_extent_header *eh)
  88 {
  89         struct ext4_extent_tail *et;
  90
  91         if (!ext4_has_metadata_csum(inode->i_sb))
  92                 return;
  93
  94         et = find_ext4_extent_tail(eh);
  95         et->et_checksum = ext4_extent_block_csum(inode, eh);
  96 }
  97
  98 static int ext4_split_extent(handle_t *handle,
  99                                 struct inode *inode,
 100                                 struct ext4_ext_path **ppath,
 101                                 struct ext4_map_blocks *map,
 102                                 int split_flag,
 103                                 int flags);
 104
 105 static int ext4_split_extent_at(handle_t *handle,
 106                              struct inode *inode,
 107                              struct ext4_ext_path **ppath,
 108                              ext4_lblk_t split,
 109                              int split_flag,
 110                              int flags);
 111
 112 static int ext4_find_delayed_extent(struct inode *inode,
 113                                     struct extent_status *newes);
 114
 115 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 116                                             struct inode *inode,
 117                                             int needed)
 118 {
 119         int err;
 120
 121         if (!ext4_handle_valid(handle))
 122                 return 0;
 123         if (handle->h_buffer_credits >= needed)
 124                 return 0;
 125         /*
 126          * If we need to extend the journal get a few extra blocks
 127          * while we're at it for efficiency's sake.
 128          */
 129         needed += 3;
 130         err = ext4_journal_extend(handle, needed - handle->h_buffer_credits);
 131         if (err <= 0)
 132                 return err;
 133         err = ext4_truncate_restart_trans(handle, inode, needed);
 134         if (err == 0)
 135                 err = -EAGAIN;
 136
 137         return err;
 138 }
 139
 140 /*
 141  * could return:
 142  *  - EROFS
 143  *  - ENOMEM
 144  */
 145 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 146                                 struct ext4_ext_path *path)
 147 {
 148         if (path->p_bh) {
 149                 /* path points to block */
 150                 BUFFER_TRACE(path->p_bh, "get_write_access");
 151                 return ext4_journal_get_write_access(handle, path->p_bh);
 152         }
 153         /* path points to leaf/index in inode body */
 154         /* we use in-core data, no need to protect them */
 155         return 0;
 156 }
 157
 158 /*
 159  * could return:
 160  *  - EROFS
 161  *  - ENOMEM
 162  *  - EIO
 163  */
 164 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 165                      struct inode *inode, struct ext4_ext_path *path)
 166 {
 167         int err;
 168
 169         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 170         if (path->p_bh) {
 171                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 172                 /* path points to block */
 173                 err = __ext4_handle_dirty_metadata(where, line, handle,
 174                                                    inode, path->p_bh);
 175         } else {
 176                 /* path points to leaf/index in inode body */
 177                 err = ext4_mark_inode_dirty(handle, inode);
 178         }
 179         return err;
 180 }
 181
 182 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 183                               struct ext4_ext_path *path,
 184                               ext4_lblk_t block)
 185 {
 186         if (path) {
 187                 int depth = path->p_depth;
 188                 struct ext4_extent *ex;
 189
 190                 /*
 191                  * Try to predict block placement assuming that we are
 192                  * filling in a file which will eventually be
 193                  * non-sparse --- i.e., in the case of libbfd writing
 194                  * an ELF object sections out-of-order but in a way
 195                  * the eventually results in a contiguous object or
 196                  * executable file, or some database extending a table
 197                  * space file.  However, this is actually somewhat
 198                  * non-ideal if we are writing a sparse file such as
 199                  * qemu or KVM writing a raw image file that is going
 200                  * to stay fairly sparse, since it will end up
 201                  * fragmenting the file system's free space.  Maybe we
 202                  * should have some hueristics or some way to allow
 203                  * userspace to pass a hint to file system,
 204                  * especially if the latter case turns out to be
 205                  * common.
 206                  */
 207                 ex = path[depth].p_ext;
 208                 if (ex) {
 209                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 210                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 211
 212                         if (block > ext_block)
 213                                 return ext_pblk + (block - ext_block);
 214                         else
 215                                 return ext_pblk - (ext_block - block);
 216                 }
 217
 218                 /* it looks like index is empty;
 219                  * try to find starting block from index itself */
 220                 if (path[depth].p_bh)
 221                         return path[depth].p_bh->b_blocknr;
 222         }
 223
 224         /* OK. use inode's group */
 225         return ext4_inode_to_goal_block(inode);
 226 }
 227
 228 /*
 229  * Allocation for a meta data block
 230  */
 231 static ext4_fsblk_t
 232 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 233                         struct ext4_ext_path *path,
 234                         struct ext4_extent *ex, int *err, unsigned int flags)
 235 {
 236         ext4_fsblk_t goal, newblock;
 237
 238         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 239         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 240                                         NULL, err);
 241         return newblock;
 242 }
 243
 244 static inline int ext4_ext_space_block(struct inode *inode, int check)
 245 {
 246         int size;
 247
 248         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 249                         / sizeof(struct ext4_extent);
 250 #ifdef AGGRESSIVE_TEST
 251         if (!check && size > 6)
 252                 size = 6;
 253 #endif
 254         return size;
 255 }
 256
 257 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 258 {
 259         int size;
 260
 261         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 262                         / sizeof(struct ext4_extent_idx);
 263 #ifdef AGGRESSIVE_TEST
 264         if (!check && size > 5)
 265                 size = 5;
 266 #endif
 267         return size;
 268 }
 269
 270 static inline int ext4_ext_space_root(struct inode *inode, int check)
 271 {
 272         int size;
 273
 274         size = sizeof(EXT4_I(inode)->i_data);
 275         size -= sizeof(struct ext4_extent_header);
 276         size /= sizeof(struct ext4_extent);
 277 #ifdef AGGRESSIVE_TEST
 278         if (!check && size > 3)
 279                 size = 3;
 280 #endif
 281         return size;
 282 }
 283
 284 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 285 {
 286         int size;
 287
 288         size = sizeof(EXT4_I(inode)->i_data);
 289         size -= sizeof(struct ext4_extent_header);
 290         size /= sizeof(struct ext4_extent_idx);
 291 #ifdef AGGRESSIVE_TEST
 292         if (!check && size > 4)
 293                 size = 4;
 294 #endif
 295         return size;
 296 }
 297
 298 static inline int
 299 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 300                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 301                            int nofail)
 302 {
 303         struct ext4_ext_path *path = *ppath;
 304         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 305
 306         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 307                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 308                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 309                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 310 }
 311
 312 /*
 313  * Calculate the number of metadata blocks needed
 314  * to allocate @blocks
 315  * Worse case is one block per extent
 316  */
 317 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 318 {
 319         struct ext4_inode_info *ei = EXT4_I(inode);
 320         int idxs;
 321
 322         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 323                 / sizeof(struct ext4_extent_idx));
 324
 325         /*
 326          * If the new delayed allocation block is contiguous with the
 327          * previous da block, it can share index blocks with the
 328          * previous block, so we only need to allocate a new index
 329          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 330          * an additional index block, and at ldxs**3 blocks, yet
 331          * another index blocks.
 332          */
 333         if (ei->i_da_metadata_calc_len &&
 334             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 335                 int num = 0;
 336
 337                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 338                         num++;
 339                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 340                         num++;
 341                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 342                         num++;
 343                         ei->i_da_metadata_calc_len = 0;
 344                 } else
 345                         ei->i_da_metadata_calc_len++;
 346                 ei->i_da_metadata_calc_last_lblock++;
 347                 return num;
 348         }
 349
 350         /*
 351          * In the worst case we need a new set of index blocks at
 352          * every level of the inode's extent tree.
 353          */
 354         ei->i_da_metadata_calc_len = 1;
 355         ei->i_da_metadata_calc_last_lblock = lblock;
 356         return ext_depth(inode) + 1;
 357 }
 358
 359 static int
 360 ext4_ext_max_entries(struct inode *inode, int depth)
 361 {
 362         int max;
 363
 364         if (depth == ext_depth(inode)) {
 365                 if (depth == 0)
 366                         max = ext4_ext_space_root(inode, 1);
 367                 else
 368                         max = ext4_ext_space_root_idx(inode, 1);
 369         } else {
 370                 if (depth == 0)
 371                         max = ext4_ext_space_block(inode, 1);
 372                 else
 373                         max = ext4_ext_space_block_idx(inode, 1);
 374         }
 375
 376         return max;
 377 }
 378
 379 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 380 {
 381         ext4_fsblk_t block = ext4_ext_pblock(ext);
 382         int len = ext4_ext_get_actual_len(ext);
 383         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 384         ext4_lblk_t last = lblock + len - 1;
 385
 386         if (len == 0 || lblock > last)
 387                 return 0;
 388         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 389 }
 390
 391 static int ext4_valid_extent_idx(struct inode *inode,
 392                                 struct ext4_extent_idx *ext_idx)
 393 {
 394         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 395
 396         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 397 }
 398
 399 static int ext4_valid_extent_entries(struct inode *inode,
 400                                 struct ext4_extent_header *eh,
 401                                 int depth)
 402 {
 403         unsigned short entries;
 404         if (eh->eh_entries == 0)
 405                 return 1;
 406
 407         entries = le16_to_cpu(eh->eh_entries);
 408
 409         if (depth == 0) {
 410                 /* leaf entries */
 411                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 412                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 413                 ext4_fsblk_t pblock = 0;
 414                 ext4_lblk_t lblock = 0;
 415                 ext4_lblk_t prev = 0;
 416                 int len = 0;
 417                 while (entries) {
 418                         if (!ext4_valid_extent(inode, ext))
 419                                 return 0;
 420
 421                         /* Check for overlapping extents */
 422                         lblock = le32_to_cpu(ext->ee_block);
 423                         len = ext4_ext_get_actual_len(ext);
 424                         if ((lblock <= prev) && prev) {
 425                                 pblock = ext4_ext_pblock(ext);
 426                                 es->s_last_error_block = cpu_to_le64(pblock);
 427                                 return 0;
 428                         }
 429                         ext++;
 430                         entries--;
 431                         prev = lblock + len - 1;
 432                 }
 433         } else {
 434                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 435                 while (entries) {
 436                         if (!ext4_valid_extent_idx(inode, ext_idx))
 437                                 return 0;
 438                         ext_idx++;
 439                         entries--;
 440                 }
 441         }
 442         return 1;
 443 }
 444
 445 static int __ext4_ext_check(const char *function, unsigned int line,
 446                             struct inode *inode, struct ext4_extent_header *eh,
 447                             int depth, ext4_fsblk_t pblk)
 448 {
 449         const char *error_msg;
 450         int max = 0, err = -EFSCORRUPTED;
 451
 452         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 453                 error_msg = "invalid magic";
 454                 goto corrupted;
 455         }
 456         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 457                 error_msg = "unexpected eh_depth";
 458                 goto corrupted;
 459         }
 460         if (unlikely(eh->eh_max == 0)) {
 461                 error_msg = "invalid eh_max";
 462                 goto corrupted;
 463         }
 464         max = ext4_ext_max_entries(inode, depth);
 465         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 466                 error_msg = "too large eh_max";
 467                 goto corrupted;
 468         }
 469         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 470                 error_msg = "invalid eh_entries";
 471                 goto corrupted;
 472         }
 473         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 474                 error_msg = "invalid extent entries";
 475                 goto corrupted;
 476         }
 477         /* Verify checksum on non-root extent tree nodes */
 478         if (ext_depth(inode) != depth &&
 479             !ext4_extent_block_csum_verify(inode, eh)) {
 480                 error_msg = "extent tree corrupted";
 481                 err = -EFSBADCRC;
 482                 goto corrupted;
 483         }
 484         return 0;
 485
 486 corrupted:
 487         ext4_error_inode(inode, function, line, 0,
 488                          "pblk %llu bad header/extent: %s - magic %x, "
 489                          "entries %u, max %u(%u), depth %u(%u)",
 490                          (unsigned long long) pblk, error_msg,
 491                          le16_to_cpu(eh->eh_magic),
 492                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 493                          max, le16_to_cpu(eh->eh_depth), depth);
 494         return err;
 495 }
 496
 497 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 498         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 499
 500 int ext4_ext_check_inode(struct inode *inode)
 501 {
 502         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 503 }
 504
 505 static struct buffer_head *
 506 __read_extent_tree_block(const char *function, unsigned int line,
 507                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 508                          int flags)
 509 {
 510         struct buffer_head              *bh;
 511         int                             err;
 512
 513         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
 514         if (unlikely(!bh))
 515                 return ERR_PTR(-ENOMEM);
 516
 517         if (!bh_uptodate_or_lock(bh)) {
 518                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 519                 err = bh_submit_read(bh);
 520                 if (err < 0)
 521                         goto errout;
 522         }
 523         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 524                 return bh;
 525         err = __ext4_ext_check(function, line, inode,
 526                                ext_block_hdr(bh), depth, pblk);
 527         if (err)
 528                 goto errout;
 529         set_buffer_verified(bh);
 530         /*
 531          * If this is a leaf block, cache all of its entries
 532          */
 533         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 534                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 535                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 536                 ext4_lblk_t prev = 0;
 537                 int i;
 538
 539                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 540                         unsigned int status = EXTENT_STATUS_WRITTEN;
 541                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 542                         int len = ext4_ext_get_actual_len(ex);
 543
 544                         if (prev && (prev != lblk))
 545                                 ext4_es_cache_extent(inode, prev,
 546                                                      lblk - prev, ~0,
 547                                                      EXTENT_STATUS_HOLE);
 548
 549                         if (ext4_ext_is_unwritten(ex))
 550                                 status = EXTENT_STATUS_UNWRITTEN;
 551                         ext4_es_cache_extent(inode, lblk, len,
 552                                              ext4_ext_pblock(ex), status);
 553                         prev = lblk + len;
 554                 }
 555         }
 556         return bh;
 557 errout:
 558         put_bh(bh);
 559         return ERR_PTR(err);
 560
 561 }
 562
 563 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 564         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 565                                  (depth), (flags))
 566
 567 /*
 568  * This function is called to cache a file's extent information in the
 569  * extent status tree
 570  */
 571 int ext4_ext_precache(struct inode *inode)
 572 {
 573         struct ext4_inode_info *ei = EXT4_I(inode);
 574         struct ext4_ext_path *path = NULL;
 575         struct buffer_head *bh;
 576         int i = 0, depth, ret = 0;
 577
 578         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 579                 return 0;       /* not an extent-mapped inode */
 580
 581         down_read(&ei->i_data_sem);
 582         depth = ext_depth(inode);
 583
 584         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 585                        GFP_NOFS);
 586         if (path == NULL) {
 587                 up_read(&ei->i_data_sem);
 588                 return -ENOMEM;
 589         }
 590
 591         /* Don't cache anything if there are no external extent blocks */
 592         if (depth == 0)
 593                 goto out;
 594         path[0].p_hdr = ext_inode_hdr(inode);
 595         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 596         if (ret)
 597                 goto out;
 598         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 599         while (i >= 0) {
 600                 /*
 601                  * If this is a leaf block or we've reached the end of
 602                  * the index block, go up
 603                  */
 604                 if ((i == depth) ||
 605                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 606                         brelse(path[i].p_bh);
 607                         path[i].p_bh = NULL;
 608                         i--;
 609                         continue;
 610                 }
 611                 bh = read_extent_tree_block(inode,
 612                                             ext4_idx_pblock(path[i].p_idx++),
 613                                             depth - i - 1,
 614                                             EXT4_EX_FORCE_CACHE);
 615                 if (IS_ERR(bh)) {
 616                         ret = PTR_ERR(bh);
 617                         break;
 618                 }
 619                 i++;
 620                 path[i].p_bh = bh;
 621                 path[i].p_hdr = ext_block_hdr(bh);
 622                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 623         }
 624         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 625 out:
 626         up_read(&ei->i_data_sem);
 627         ext4_ext_drop_refs(path);
 628         kfree(path);
 629         return ret;
 630 }
 631
 632 #ifdef EXT_DEBUG
 633 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 634 {
 635         int k, l = path->p_depth;
 636
 637         ext_debug("path:");
 638         for (k = 0; k <= l; k++, path++) {
 639                 if (path->p_idx) {
 640                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 641                             ext4_idx_pblock(path->p_idx));
 642                 } else if (path->p_ext) {
 643                         ext_debug("  %d:[%d]%d:%llu ",
 644                                   le32_to_cpu(path->p_ext->ee_block),
 645                                   ext4_ext_is_unwritten(path->p_ext),
 646                                   ext4_ext_get_actual_len(path->p_ext),
 647                                   ext4_ext_pblock(path->p_ext));
 648                 } else
 649                         ext_debug("  []");
 650         }
 651         ext_debug("\n");
 652 }
 653
 654 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 655 {
 656         int depth = ext_depth(inode);
 657         struct ext4_extent_header *eh;
 658         struct ext4_extent *ex;
 659         int i;
 660
 661         if (!path)
 662                 return;
 663
 664         eh = path[depth].p_hdr;
 665         ex = EXT_FIRST_EXTENT(eh);
 666
 667         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 668
 669         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 670                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 671                           ext4_ext_is_unwritten(ex),
 672                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 673         }
 674         ext_debug("\n");
 675 }
 676
 677 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 678                         ext4_fsblk_t newblock, int level)
 679 {
 680         int depth = ext_depth(inode);
 681         struct ext4_extent *ex;
 682
 683         if (depth != level) {
 684                 struct ext4_extent_idx *idx;
 685                 idx = path[level].p_idx;
 686                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 687                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 688                                         le32_to_cpu(idx->ei_block),
 689                                         ext4_idx_pblock(idx),
 690                                         newblock);
 691                         idx++;
 692                 }
 693
 694                 return;
 695         }
 696
 697         ex = path[depth].p_ext;
 698         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 699                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 700                                 le32_to_cpu(ex->ee_block),
 701                                 ext4_ext_pblock(ex),
 702                                 ext4_ext_is_unwritten(ex),
 703                                 ext4_ext_get_actual_len(ex),
 704                                 newblock);
 705                 ex++;
 706         }
 707 }
 708
 709 #else
 710 #define ext4_ext_show_path(inode, path)
 711 #define ext4_ext_show_leaf(inode, path)
 712 #define ext4_ext_show_move(inode, path, newblock, level)
 713 #endif
 714
 715 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 716 {
 717         int depth, i;
 718
 719         if (!path)
 720                 return;
 721         depth = path->p_depth;
 722         for (i = 0; i <= depth; i++, path++)
 723                 if (path->p_bh) {
 724                         brelse(path->p_bh);
 725                         path->p_bh = NULL;
 726                 }
 727 }
 728
 729 /*
 730  * ext4_ext_binsearch_idx:
 731  * binary search for the closest index of the given block
 732  * the header must be checked before calling this
 733  */
 734 static void
 735 ext4_ext_binsearch_idx(struct inode *inode,
 736                         struct ext4_ext_path *path, ext4_lblk_t block)
 737 {
 738         struct ext4_extent_header *eh = path->p_hdr;
 739         struct ext4_extent_idx *r, *l, *m;
 740
 741
 742         ext_debug("binsearch for %u(idx):  ", block);
 743
 744         l = EXT_FIRST_INDEX(eh) + 1;
 745         r = EXT_LAST_INDEX(eh);
 746         while (l <= r) {
 747                 m = l + (r - l) / 2;
 748                 if (block < le32_to_cpu(m->ei_block))
 749                         r = m - 1;
 750                 else
 751                         l = m + 1;
 752                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 753                                 m, le32_to_cpu(m->ei_block),
 754                                 r, le32_to_cpu(r->ei_block));
 755         }
 756
 757         path->p_idx = l - 1;
 758         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 759                   ext4_idx_pblock(path->p_idx));
 760
 761 #ifdef CHECK_BINSEARCH
 762         {
 763                 struct ext4_extent_idx *chix, *ix;
 764                 int k;
 765
 766                 chix = ix = EXT_FIRST_INDEX(eh);
 767                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 768                   if (k != 0 &&
 769                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 770                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 771                                        "first=0x%p\n", k,
 772                                        ix, EXT_FIRST_INDEX(eh));
 773                                 printk(KERN_DEBUG "%u <= %u\n",
 774                                        le32_to_cpu(ix->ei_block),
 775                                        le32_to_cpu(ix[-1].ei_block));
 776                         }
 777                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 778                                            <= le32_to_cpu(ix[-1].ei_block));
 779                         if (block < le32_to_cpu(ix->ei_block))
 780                                 break;
 781                         chix = ix;
 782                 }
 783                 BUG_ON(chix != path->p_idx);
 784         }
 785 #endif
 786
 787 }
 788
 789 /*
 790  * ext4_ext_binsearch:
 791  * binary search for closest extent of the given block
 792  * the header must be checked before calling this
 793  */
 794 static void
 795 ext4_ext_binsearch(struct inode *inode,
 796                 struct ext4_ext_path *path, ext4_lblk_t block)
 797 {
 798         struct ext4_extent_header *eh = path->p_hdr;
 799         struct ext4_extent *r, *l, *m;
 800
 801         if (eh->eh_entries == 0) {
 802                 /*
 803                  * this leaf is empty:
 804                  * we get such a leaf in split/add case
 805                  */
 806                 return;
 807         }
 808
 809         ext_debug("binsearch for %u:  ", block);
 810
 811         l = EXT_FIRST_EXTENT(eh) + 1;
 812         r = EXT_LAST_EXTENT(eh);
 813
 814         while (l <= r) {
 815                 m = l + (r - l) / 2;
 816                 if (block < le32_to_cpu(m->ee_block))
 817                         r = m - 1;
 818                 else
 819                         l = m + 1;
 820                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 821                                 m, le32_to_cpu(m->ee_block),
 822                                 r, le32_to_cpu(r->ee_block));
 823         }
 824
 825         path->p_ext = l - 1;
 826         ext_debug("  -> %d:%llu:[%d]%d ",
 827                         le32_to_cpu(path->p_ext->ee_block),
 828                         ext4_ext_pblock(path->p_ext),
 829                         ext4_ext_is_unwritten(path->p_ext),
 830                         ext4_ext_get_actual_len(path->p_ext));
 831
 832 #ifdef CHECK_BINSEARCH
 833         {
 834                 struct ext4_extent *chex, *ex;
 835                 int k;
 836
 837                 chex = ex = EXT_FIRST_EXTENT(eh);
 838                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 839                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 840                                           <= le32_to_cpu(ex[-1].ee_block));
 841                         if (block < le32_to_cpu(ex->ee_block))
 842                                 break;
 843                         chex = ex;
 844                 }
 845                 BUG_ON(chex != path->p_ext);
 846         }
 847 #endif
 848
 849 }
 850
 851 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 852 {
 853         struct ext4_extent_header *eh;
 854
 855         eh = ext_inode_hdr(inode);
 856         eh->eh_depth = 0;
 857         eh->eh_entries = 0;
 858         eh->eh_magic = EXT4_EXT_MAGIC;
 859         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 860         ext4_mark_inode_dirty(handle, inode);
 861         return 0;
 862 }
 863
 864 struct ext4_ext_path *
 865 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 866                  struct ext4_ext_path **orig_path, int flags)
 867 {
 868         struct ext4_extent_header *eh;
 869         struct buffer_head *bh;
 870         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 871         short int depth, i, ppos = 0;
 872         int ret;
 873
 874         eh = ext_inode_hdr(inode);
 875         depth = ext_depth(inode);
 876
 877         if (path) {
 878                 ext4_ext_drop_refs(path);
 879                 if (depth > path[0].p_maxdepth) {
 880                         kfree(path);
 881                         *orig_path = path = NULL;
 882                 }
 883         }
 884         if (!path) {
 885                 /* account possible depth increase */
 886                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 887                                 GFP_NOFS);
 888                 if (unlikely(!path))
 889                         return ERR_PTR(-ENOMEM);
 890                 path[0].p_maxdepth = depth + 1;
 891         }
 892         path[0].p_hdr = eh;
 893         path[0].p_bh = NULL;
 894
 895         i = depth;
 896         /* walk through the tree */
 897         while (i) {
 898                 ext_debug("depth %d: num %d, max %d\n",
 899                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 900
 901                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 902                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 903                 path[ppos].p_depth = i;
 904                 path[ppos].p_ext = NULL;
 905
 906                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 907                                             flags);
 908                 if (IS_ERR(bh)) {
 909                         ret = PTR_ERR(bh);
 910                         goto err;
 911                 }
 912
 913                 eh = ext_block_hdr(bh);
 914                 ppos++;
 915                 path[ppos].p_bh = bh;
 916                 path[ppos].p_hdr = eh;
 917         }
 918
 919         path[ppos].p_depth = i;
 920         path[ppos].p_ext = NULL;
 921         path[ppos].p_idx = NULL;
 922
 923         /* find extent */
 924         ext4_ext_binsearch(inode, path + ppos, block);
 925         /* if not an empty leaf */
 926         if (path[ppos].p_ext)
 927                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 928
 929         ext4_ext_show_path(inode, path);
 930
 931         return path;
 932
 933 err:
 934         ext4_ext_drop_refs(path);
 935         kfree(path);
 936         if (orig_path)
 937                 *orig_path = NULL;
 938         return ERR_PTR(ret);
 939 }
 940
 941 /*
 942  * ext4_ext_insert_index:
 943  * insert new index [@logical;@ptr] into the block at @curp;
 944  * check where to insert: before @curp or after @curp
 945  */
 946 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 947                                  struct ext4_ext_path *curp,
 948                                  int logical, ext4_fsblk_t ptr)
 949 {
 950         struct ext4_extent_idx *ix;
 951         int len, err;
 952
 953         err = ext4_ext_get_access(handle, inode, curp);
 954         if (err)
 955                 return err;
 956
 957         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 958                 EXT4_ERROR_INODE(inode,
 959                                  "logical %d == ei_block %d!",
 960                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 961                 return -EFSCORRUPTED;
 962         }
 963
 964         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 965                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 966                 EXT4_ERROR_INODE(inode,
 967                                  "eh_entries %d >= eh_max %d!",
 968                                  le16_to_cpu(curp->p_hdr->eh_entries),
 969                                  le16_to_cpu(curp->p_hdr->eh_max));
 970                 return -EFSCORRUPTED;
 971         }
 972
 973         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 974                 /* insert after */
 975                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 976                 ix = curp->p_idx + 1;
 977         } else {
 978                 /* insert before */
 979                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 980                 ix = curp->p_idx;
 981         }
 982
 983         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 984         BUG_ON(len < 0);
 985         if (len > 0) {
 986                 ext_debug("insert new index %d: "
 987                                 "move %d indices from 0x%p to 0x%p\n",
 988                                 logical, len, ix, ix + 1);
 989                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 990         }
 991
 992         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 993                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 994                 return -EFSCORRUPTED;
 995         }
 996
 997         ix->ei_block = cpu_to_le32(logical);
 998         ext4_idx_store_pblock(ix, ptr);
 999         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1000
1001         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1002                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1003                 return -EFSCORRUPTED;
1004         }
1005
1006         err = ext4_ext_dirty(handle, inode, curp);
1007         ext4_std_error(inode->i_sb, err);
1008
1009         return err;
1010 }
1011
1012 /*
1013  * ext4_ext_split:
1014  * inserts new subtree into the path, using free index entry
1015  * at depth @at:
1016  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1017  * - makes decision where to split
1018  * - moves remaining extents and index entries (right to the split point)
1019  *   into the newly allocated blocks
1020  * - initializes subtree
1021  */
1022 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1023                           unsigned int flags,
1024                           struct ext4_ext_path *path,
1025                           struct ext4_extent *newext, int at)
1026 {
1027         struct buffer_head *bh = NULL;
1028         int depth = ext_depth(inode);
1029         struct ext4_extent_header *neh;
1030         struct ext4_extent_idx *fidx;
1031         int i = at, k, m, a;
1032         ext4_fsblk_t newblock, oldblock;
1033         __le32 border;
1034         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1035         int err = 0;
1036
1037         /* make decision: where to split? */
1038         /* FIXME: now decision is simplest: at current extent */
1039
1040         /* if current leaf will be split, then we should use
1041          * border from split point */
1042         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1043                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1044                 return -EFSCORRUPTED;
1045         }
1046         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1047                 border = path[depth].p_ext[1].ee_block;
1048                 ext_debug("leaf will be split."
1049                                 " next leaf starts at %d\n",
1050                                   le32_to_cpu(border));
1051         } else {
1052                 border = newext->ee_block;
1053                 ext_debug("leaf will be added."
1054                                 " next leaf starts at %d\n",
1055                                 le32_to_cpu(border));
1056         }
1057
1058         /*
1059          * If error occurs, then we break processing
1060          * and mark filesystem read-only. index won't
1061          * be inserted and tree will be in consistent
1062          * state. Next mount will repair buffers too.
1063          */
1064
1065         /*
1066          * Get array to track all allocated blocks.
1067          * We need this to handle errors and free blocks
1068          * upon them.
1069          */
1070         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1071         if (!ablocks)
1072                 return -ENOMEM;
1073
1074         /* allocate all needed blocks */
1075         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1076         for (a = 0; a < depth - at; a++) {
1077                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1078                                                    newext, &err, flags);
1079                 if (newblock == 0)
1080                         goto cleanup;
1081                 ablocks[a] = newblock;
1082         }
1083
1084         /* initialize new leaf */
1085         newblock = ablocks[--a];
1086         if (unlikely(newblock == 0)) {
1087                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1088                 err = -EFSCORRUPTED;
1089                 goto cleanup;
1090         }
1091         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1092         if (unlikely(!bh)) {
1093                 err = -ENOMEM;
1094                 goto cleanup;
1095         }
1096         lock_buffer(bh);
1097
1098         err = ext4_journal_get_create_access(handle, bh);
1099         if (err)
1100                 goto cleanup;
1101
1102         neh = ext_block_hdr(bh);
1103         neh->eh_entries = 0;
1104         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1105         neh->eh_magic = EXT4_EXT_MAGIC;
1106         neh->eh_depth = 0;
1107
1108         /* move remainder of path[depth] to the new leaf */
1109         if (unlikely(path[depth].p_hdr->eh_entries !=
1110                      path[depth].p_hdr->eh_max)) {
1111                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1112                                  path[depth].p_hdr->eh_entries,
1113                                  path[depth].p_hdr->eh_max);
1114                 err = -EFSCORRUPTED;
1115                 goto cleanup;
1116         }
1117         /* start copy from next extent */
1118         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1119         ext4_ext_show_move(inode, path, newblock, depth);
1120         if (m) {
1121                 struct ext4_extent *ex;
1122                 ex = EXT_FIRST_EXTENT(neh);
1123                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1124                 le16_add_cpu(&neh->eh_entries, m);
1125         }
1126
1127         ext4_extent_block_csum_set(inode, neh);
1128         set_buffer_uptodate(bh);
1129         unlock_buffer(bh);
1130
1131         err = ext4_handle_dirty_metadata(handle, inode, bh);
1132         if (err)
1133                 goto cleanup;
1134         brelse(bh);
1135         bh = NULL;
1136
1137         /* correct old leaf */
1138         if (m) {
1139                 err = ext4_ext_get_access(handle, inode, path + depth);
1140                 if (err)
1141                         goto cleanup;
1142                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1143                 err = ext4_ext_dirty(handle, inode, path + depth);
1144                 if (err)
1145                         goto cleanup;
1146
1147         }
1148
1149         /* create intermediate indexes */
1150         k = depth - at - 1;
1151         if (unlikely(k < 0)) {
1152                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1153                 err = -EFSCORRUPTED;
1154                 goto cleanup;
1155         }
1156         if (k)
1157                 ext_debug("create %d intermediate indices\n", k);
1158         /* insert new index into current index block */
1159         /* current depth stored in i var */
1160         i = depth - 1;
1161         while (k--) {
1162                 oldblock = newblock;
1163                 newblock = ablocks[--a];
1164                 bh = sb_getblk(inode->i_sb, newblock);
1165                 if (unlikely(!bh)) {
1166                         err = -ENOMEM;
1167                         goto cleanup;
1168                 }
1169                 lock_buffer(bh);
1170
1171                 err = ext4_journal_get_create_access(handle, bh);
1172                 if (err)
1173                         goto cleanup;
1174
1175                 neh = ext_block_hdr(bh);
1176                 neh->eh_entries = cpu_to_le16(1);
1177                 neh->eh_magic = EXT4_EXT_MAGIC;
1178                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1179                 neh->eh_depth = cpu_to_le16(depth - i);
1180                 fidx = EXT_FIRST_INDEX(neh);
1181                 fidx->ei_block = border;
1182                 ext4_idx_store_pblock(fidx, oldblock);
1183
1184                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1185                                 i, newblock, le32_to_cpu(border), oldblock);
1186
1187                 /* move remainder of path[i] to the new index block */
1188                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1189                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1190                         EXT4_ERROR_INODE(inode,
1191                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1192                                          le32_to_cpu(path[i].p_ext->ee_block));
1193                         err = -EFSCORRUPTED;
1194                         goto cleanup;
1195                 }
1196                 /* start copy indexes */
1197                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1198                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1199                                 EXT_MAX_INDEX(path[i].p_hdr));
1200                 ext4_ext_show_move(inode, path, newblock, i);
1201                 if (m) {
1202                         memmove(++fidx, path[i].p_idx,
1203                                 sizeof(struct ext4_extent_idx) * m);
1204                         le16_add_cpu(&neh->eh_entries, m);
1205                 }
1206                 ext4_extent_block_csum_set(inode, neh);
1207                 set_buffer_uptodate(bh);
1208                 unlock_buffer(bh);
1209
1210                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1211                 if (err)
1212                         goto cleanup;
1213                 brelse(bh);
1214                 bh = NULL;
1215
1216                 /* correct old index */
1217                 if (m) {
1218                         err = ext4_ext_get_access(handle, inode, path + i);
1219                         if (err)
1220                                 goto cleanup;
1221                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1222                         err = ext4_ext_dirty(handle, inode, path + i);
1223                         if (err)
1224                                 goto cleanup;
1225                 }
1226
1227                 i--;
1228         }
1229
1230         /* insert new index */
1231         err = ext4_ext_insert_index(handle, inode, path + at,
1232                                     le32_to_cpu(border), newblock);
1233
1234 cleanup:
1235         if (bh) {
1236                 if (buffer_locked(bh))
1237                         unlock_buffer(bh);
1238                 brelse(bh);
1239         }
1240
1241         if (err) {
1242                 /* free all allocated blocks in error case */
1243                 for (i = 0; i < depth; i++) {
1244                         if (!ablocks[i])
1245                                 continue;
1246                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1247                                          EXT4_FREE_BLOCKS_METADATA);
1248                 }
1249         }
1250         kfree(ablocks);
1251
1252         return err;
1253 }
1254
1255 /*
1256  * ext4_ext_grow_indepth:
1257  * implements tree growing procedure:
1258  * - allocates new block
1259  * - moves top-level data (index block or leaf) into the new block
1260  * - initializes new top-level, creating index that points to the
1261  *   just created block
1262  */
1263 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1264                                  unsigned int flags)
1265 {
1266         struct ext4_extent_header *neh;
1267         struct buffer_head *bh;
1268         ext4_fsblk_t newblock, goal = 0;
1269         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1270         int err = 0;
1271
1272         /* Try to prepend new index to old one */
1273         if (ext_depth(inode))
1274                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1275         if (goal > le32_to_cpu(es->s_first_data_block)) {
1276                 flags |= EXT4_MB_HINT_TRY_GOAL;
1277                 goal--;
1278         } else
1279                 goal = ext4_inode_to_goal_block(inode);
1280         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1281                                         NULL, &err);
1282         if (newblock == 0)
1283                 return err;
1284
1285         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1286         if (unlikely(!bh))
1287                 return -ENOMEM;
1288         lock_buffer(bh);
1289
1290         err = ext4_journal_get_create_access(handle, bh);
1291         if (err) {
1292                 unlock_buffer(bh);
1293                 goto out;
1294         }
1295
1296         /* move top-level index/leaf into new block */
1297         memmove(bh->b_data, EXT4_I(inode)->i_data,
1298                 sizeof(EXT4_I(inode)->i_data));
1299
1300         /* set size of new block */
1301         neh = ext_block_hdr(bh);
1302         /* old root could have indexes or leaves
1303          * so calculate e_max right way */
1304         if (ext_depth(inode))
1305                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1306         else
1307                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1308         neh->eh_magic = EXT4_EXT_MAGIC;
1309         ext4_extent_block_csum_set(inode, neh);
1310         set_buffer_uptodate(bh);
1311         unlock_buffer(bh);
1312
1313         err = ext4_handle_dirty_metadata(handle, inode, bh);
1314         if (err)
1315                 goto out;
1316
1317         /* Update top-level index: num,max,pointer */
1318         neh = ext_inode_hdr(inode);
1319         neh->eh_entries = cpu_to_le16(1);
1320         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1321         if (neh->eh_depth == 0) {
1322                 /* Root extent block becomes index block */
1323                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1324                 EXT_FIRST_INDEX(neh)->ei_block =
1325                         EXT_FIRST_EXTENT(neh)->ee_block;
1326         }
1327         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1328                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1329                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1330                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1331
1332         le16_add_cpu(&neh->eh_depth, 1);
1333         ext4_mark_inode_dirty(handle, inode);
1334 out:
1335         brelse(bh);
1336
1337         return err;
1338 }
1339
1340 /*
1341  * ext4_ext_create_new_leaf:
1342  * finds empty index and adds new leaf.
1343  * if no free index is found, then it requests in-depth growing.
1344  */
1345 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1346                                     unsigned int mb_flags,
1347                                     unsigned int gb_flags,
1348                                     struct ext4_ext_path **ppath,
1349                                     struct ext4_extent *newext)
1350 {
1351         struct ext4_ext_path *path = *ppath;
1352         struct ext4_ext_path *curp;
1353         int depth, i, err = 0;
1354
1355 repeat:
1356         i = depth = ext_depth(inode);
1357
1358         /* walk up to the tree and look for free index entry */
1359         curp = path + depth;
1360         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1361                 i--;
1362                 curp--;
1363         }
1364
1365         /* we use already allocated block for index block,
1366          * so subsequent data blocks should be contiguous */
1367         if (EXT_HAS_FREE_INDEX(curp)) {
1368                 /* if we found index with free entry, then use that
1369                  * entry: create all needed subtree and add new leaf */
1370                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1371                 if (err)
1372                         goto out;
1373
1374                 /* refill path */
1375                 path = ext4_find_extent(inode,
1376                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1377                                     ppath, gb_flags);
1378                 if (IS_ERR(path))
1379                         err = PTR_ERR(path);
1380         } else {
1381                 /* tree is full, time to grow in depth */
1382                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1383                 if (err)
1384                         goto out;
1385
1386                 /* refill path */
1387                 path = ext4_find_extent(inode,
1388                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1389                                     ppath, gb_flags);
1390                 if (IS_ERR(path)) {
1391                         err = PTR_ERR(path);
1392                         goto out;
1393                 }
1394
1395                 /*
1396                  * only first (depth 0 -> 1) produces free space;
1397                  * in all other cases we have to split the grown tree
1398                  */
1399                 depth = ext_depth(inode);
1400                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1401                         /* now we need to split */
1402                         goto repeat;
1403                 }
1404         }
1405
1406 out:
1407         return err;
1408 }
1409
1410 /*
1411  * search the closest allocated block to the left for *logical
1412  * and returns it at @logical + it's physical address at @phys
1413  * if *logical is the smallest allocated block, the function
1414  * returns 0 at @phys
1415  * return value contains 0 (success) or error code
1416  */
1417 static int ext4_ext_search_left(struct inode *inode,
1418                                 struct ext4_ext_path *path,
1419                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1420 {
1421         struct ext4_extent_idx *ix;
1422         struct ext4_extent *ex;
1423         int depth, ee_len;
1424
1425         if (unlikely(path == NULL)) {
1426                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1427                 return -EFSCORRUPTED;
1428         }
1429         depth = path->p_depth;
1430         *phys = 0;
1431
1432         if (depth == 0 && path->p_ext == NULL)
1433                 return 0;
1434
1435         /* usually extent in the path covers blocks smaller
1436          * then *logical, but it can be that extent is the
1437          * first one in the file */
1438
1439         ex = path[depth].p_ext;
1440         ee_len = ext4_ext_get_actual_len(ex);
1441         if (*logical < le32_to_cpu(ex->ee_block)) {
1442                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1443                         EXT4_ERROR_INODE(inode,
1444                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1445                                          *logical, le32_to_cpu(ex->ee_block));
1446                         return -EFSCORRUPTED;
1447                 }
1448                 while (--depth >= 0) {
1449                         ix = path[depth].p_idx;
1450                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1451                                 EXT4_ERROR_INODE(inode,
1452                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1453                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1454                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1455                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1456                                   depth);
1457                                 return -EFSCORRUPTED;
1458                         }
1459                 }
1460                 return 0;
1461         }
1462
1463         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1464                 EXT4_ERROR_INODE(inode,
1465                                  "logical %d < ee_block %d + ee_len %d!",
1466                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1467                 return -EFSCORRUPTED;
1468         }
1469
1470         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1471         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1472         return 0;
1473 }
1474
1475 /*
1476  * search the closest allocated block to the right for *logical
1477  * and returns it at @logical + it's physical address at @phys
1478  * if *logical is the largest allocated block, the function
1479  * returns 0 at @phys
1480  * return value contains 0 (success) or error code
1481  */
1482 static int ext4_ext_search_right(struct inode *inode,
1483                                  struct ext4_ext_path *path,
1484                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1485                                  struct ext4_extent **ret_ex)
1486 {
1487         struct buffer_head *bh = NULL;
1488         struct ext4_extent_header *eh;
1489         struct ext4_extent_idx *ix;
1490         struct ext4_extent *ex;
1491         ext4_fsblk_t block;
1492         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1493         int ee_len;
1494
1495         if (unlikely(path == NULL)) {
1496                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1497                 return -EFSCORRUPTED;
1498         }
1499         depth = path->p_depth;
1500         *phys = 0;
1501
1502         if (depth == 0 && path->p_ext == NULL)
1503                 return 0;
1504
1505         /* usually extent in the path covers blocks smaller
1506          * then *logical, but it can be that extent is the
1507          * first one in the file */
1508
1509         ex = path[depth].p_ext;
1510         ee_len = ext4_ext_get_actual_len(ex);
1511         if (*logical < le32_to_cpu(ex->ee_block)) {
1512                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1513                         EXT4_ERROR_INODE(inode,
1514                                          "first_extent(path[%d].p_hdr) != ex",
1515                                          depth);
1516                         return -EFSCORRUPTED;
1517                 }
1518                 while (--depth >= 0) {
1519                         ix = path[depth].p_idx;
1520                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1521                                 EXT4_ERROR_INODE(inode,
1522                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1523                                                  *logical);
1524                                 return -EFSCORRUPTED;
1525                         }
1526                 }
1527                 goto found_extent;
1528         }
1529
1530         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1531                 EXT4_ERROR_INODE(inode,
1532                                  "logical %d < ee_block %d + ee_len %d!",
1533                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1534                 return -EFSCORRUPTED;
1535         }
1536
1537         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1538                 /* next allocated block in this leaf */
1539                 ex++;
1540                 goto found_extent;
1541         }
1542
1543         /* go up and search for index to the right */
1544         while (--depth >= 0) {
1545                 ix = path[depth].p_idx;
1546                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1547                         goto got_index;
1548         }
1549
1550         /* we've gone up to the root and found no index to the right */
1551         return 0;
1552
1553 got_index:
1554         /* we've found index to the right, let's
1555          * follow it and find the closest allocated
1556          * block to the right */
1557         ix++;
1558         block = ext4_idx_pblock(ix);
1559         while (++depth < path->p_depth) {
1560                 /* subtract from p_depth to get proper eh_depth */
1561                 bh = read_extent_tree_block(inode, block,
1562                                             path->p_depth - depth, 0);
1563                 if (IS_ERR(bh))
1564                         return PTR_ERR(bh);
1565                 eh = ext_block_hdr(bh);
1566                 ix = EXT_FIRST_INDEX(eh);
1567                 block = ext4_idx_pblock(ix);
1568                 put_bh(bh);
1569         }
1570
1571         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1572         if (IS_ERR(bh))
1573                 return PTR_ERR(bh);
1574         eh = ext_block_hdr(bh);
1575         ex = EXT_FIRST_EXTENT(eh);
1576 found_extent:
1577         *logical = le32_to_cpu(ex->ee_block);
1578         *phys = ext4_ext_pblock(ex);
1579         *ret_ex = ex;
1580         if (bh)
1581                 put_bh(bh);
1582         return 0;
1583 }
1584
1585 /*
1586  * ext4_ext_next_allocated_block:
1587  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1588  * NOTE: it considers block number from index entry as
1589  * allocated block. Thus, index entries have to be consistent
1590  * with leaves.
1591  */
1592 ext4_lblk_t
1593 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1594 {
1595         int depth;
1596
1597         BUG_ON(path == NULL);
1598         depth = path->p_depth;
1599
1600         if (depth == 0 && path->p_ext == NULL)
1601                 return EXT_MAX_BLOCKS;
1602
1603         while (depth >= 0) {
1604                 if (depth == path->p_depth) {
1605                         /* leaf */
1606                         if (path[depth].p_ext &&
1607                                 path[depth].p_ext !=
1608                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1609                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1610                 } else {
1611                         /* index */
1612                         if (path[depth].p_idx !=
1613                                         EXT_LAST_INDEX(path[depth].p_hdr))
1614                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1615                 }
1616                 depth--;
1617         }
1618
1619         return EXT_MAX_BLOCKS;
1620 }
1621
1622 /*
1623  * ext4_ext_next_leaf_block:
1624  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1625  */
1626 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1627 {
1628         int depth;
1629
1630         BUG_ON(path == NULL);
1631         depth = path->p_depth;
1632
1633         /* zero-tree has no leaf blocks at all */
1634         if (depth == 0)
1635                 return EXT_MAX_BLOCKS;
1636
1637         /* go to index block */
1638         depth--;
1639
1640         while (depth >= 0) {
1641                 if (path[depth].p_idx !=
1642                                 EXT_LAST_INDEX(path[depth].p_hdr))
1643                         return (ext4_lblk_t)
1644                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1645                 depth--;
1646         }
1647
1648         return EXT_MAX_BLOCKS;
1649 }
1650
1651 /*
1652  * ext4_ext_correct_indexes:
1653  * if leaf gets modified and modified extent is first in the leaf,
1654  * then we have to correct all indexes above.
1655  * TODO: do we need to correct tree in all cases?
1656  */
1657 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1658                                 struct ext4_ext_path *path)
1659 {
1660         struct ext4_extent_header *eh;
1661         int depth = ext_depth(inode);
1662         struct ext4_extent *ex;
1663         __le32 border;
1664         int k, err = 0;
1665
1666         eh = path[depth].p_hdr;
1667         ex = path[depth].p_ext;
1668
1669         if (unlikely(ex == NULL || eh == NULL)) {
1670                 EXT4_ERROR_INODE(inode,
1671                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1672                 return -EFSCORRUPTED;
1673         }
1674
1675         if (depth == 0) {
1676                 /* there is no tree at all */
1677                 return 0;
1678         }
1679
1680         if (ex != EXT_FIRST_EXTENT(eh)) {
1681                 /* we correct tree if first leaf got modified only */
1682                 return 0;
1683         }
1684
1685         /*
1686          * TODO: we need correction if border is smaller than current one
1687          */
1688         k = depth - 1;
1689         border = path[depth].p_ext->ee_block;
1690         err = ext4_ext_get_access(handle, inode, path + k);
1691         if (err)
1692                 return err;
1693         path[k].p_idx->ei_block = border;
1694         err = ext4_ext_dirty(handle, inode, path + k);
1695         if (err)
1696                 return err;
1697
1698         while (k--) {
1699                 /* change all left-side indexes */
1700                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1701                         break;
1702                 err = ext4_ext_get_access(handle, inode, path + k);
1703                 if (err)
1704                         break;
1705                 path[k].p_idx->ei_block = border;
1706                 err = ext4_ext_dirty(handle, inode, path + k);
1707                 if (err)
1708                         break;
1709         }
1710
1711         return err;
1712 }
1713
1714 int
1715 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1716                                 struct ext4_extent *ex2)
1717 {
1718         unsigned short ext1_ee_len, ext2_ee_len;
1719
1720         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1721                 return 0;
1722
1723         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1724         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1725
1726         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1727                         le32_to_cpu(ex2->ee_block))
1728                 return 0;
1729
1730         /*
1731          * To allow future support for preallocated extents to be added
1732          * as an RO_COMPAT feature, refuse to merge to extents if
1733          * this can result in the top bit of ee_len being set.
1734          */
1735         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1736                 return 0;
1737         /*
1738          * The check for IO to unwritten extent is somewhat racy as we
1739          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1740          * dropping i_data_sem. But reserved blocks should save us in that
1741          * case.
1742          */
1743         if (ext4_ext_is_unwritten(ex1) &&
1744             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1745              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1746              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1747                 return 0;
1748 #ifdef AGGRESSIVE_TEST
1749         if (ext1_ee_len >= 4)
1750                 return 0;
1751 #endif
1752
1753         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1754                 return 1;
1755         return 0;
1756 }
1757
1758 /*
1759  * This function tries to merge the "ex" extent to the next extent in the tree.
1760  * It always tries to merge towards right. If you want to merge towards
1761  * left, pass "ex - 1" as argument instead of "ex".
1762  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1763  * 1 if they got merged.
1764  */
1765 static int ext4_ext_try_to_merge_right(struct inode *inode,
1766                                  struct ext4_ext_path *path,
1767                                  struct ext4_extent *ex)
1768 {
1769         struct ext4_extent_header *eh;
1770         unsigned int depth, len;
1771         int merge_done = 0, unwritten;
1772
1773         depth = ext_depth(inode);
1774         BUG_ON(path[depth].p_hdr == NULL);
1775         eh = path[depth].p_hdr;
1776
1777         while (ex < EXT_LAST_EXTENT(eh)) {
1778                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1779                         break;
1780                 /* merge with next extent! */
1781                 unwritten = ext4_ext_is_unwritten(ex);
1782                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1783                                 + ext4_ext_get_actual_len(ex + 1));
1784                 if (unwritten)
1785                         ext4_ext_mark_unwritten(ex);
1786
1787                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1788                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1789                                 * sizeof(struct ext4_extent);
1790                         memmove(ex + 1, ex + 2, len);
1791                 }
1792                 le16_add_cpu(&eh->eh_entries, -1);
1793                 merge_done = 1;
1794                 WARN_ON(eh->eh_entries == 0);
1795                 if (!eh->eh_entries)
1796                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1797         }
1798
1799         return merge_done;
1800 }
1801
1802 /*
1803  * This function does a very simple check to see if we can collapse
1804  * an extent tree with a single extent tree leaf block into the inode.
1805  */
1806 static void ext4_ext_try_to_merge_up(handle_t *handle,
1807                                      struct inode *inode,
1808                                      struct ext4_ext_path *path)
1809 {
1810         size_t s;
1811         unsigned max_root = ext4_ext_space_root(inode, 0);
1812         ext4_fsblk_t blk;
1813
1814         if ((path[0].p_depth != 1) ||
1815             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1816             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1817                 return;
1818
1819         /*
1820          * We need to modify the block allocation bitmap and the block
1821          * group descriptor to release the extent tree block.  If we
1822          * can't get the journal credits, give up.
1823          */
1824         if (ext4_journal_extend(handle, 2))
1825                 return;
1826
1827         /*
1828          * Copy the extent data up to the inode
1829          */
1830         blk = ext4_idx_pblock(path[0].p_idx);
1831         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1832                 sizeof(struct ext4_extent_idx);
1833         s += sizeof(struct ext4_extent_header);
1834
1835         path[1].p_maxdepth = path[0].p_maxdepth;
1836         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1837         path[0].p_depth = 0;
1838         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1839                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1840         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1841
1842         brelse(path[1].p_bh);
1843         ext4_free_blocks(handle, inode, NULL, blk, 1,
1844                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1845 }
1846
1847 /*
1848  * This function tries to merge the @ex extent to neighbours in the tree.
1849  * return 1 if merge left else 0.
1850  */
1851 static void ext4_ext_try_to_merge(handle_t *handle,
1852                                   struct inode *inode,
1853                                   struct ext4_ext_path *path,
1854                                   struct ext4_extent *ex) {
1855         struct ext4_extent_header *eh;
1856         unsigned int depth;
1857         int merge_done = 0;
1858
1859         depth = ext_depth(inode);
1860         BUG_ON(path[depth].p_hdr == NULL);
1861         eh = path[depth].p_hdr;
1862
1863         if (ex > EXT_FIRST_EXTENT(eh))
1864                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1865
1866         if (!merge_done)
1867                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1868
1869         ext4_ext_try_to_merge_up(handle, inode, path);
1870 }
1871
1872 /*
1873  * check if a portion of the "newext" extent overlaps with an
1874  * existing extent.
1875  *
1876  * If there is an overlap discovered, it updates the length of the newext
1877  * such that there will be no overlap, and then returns 1.
1878  * If there is no overlap found, it returns 0.
1879  */
1880 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1881                                            struct inode *inode,
1882                                            struct ext4_extent *newext,
1883                                            struct ext4_ext_path *path)
1884 {
1885         ext4_lblk_t b1, b2;
1886         unsigned int depth, len1;
1887         unsigned int ret = 0;
1888
1889         b1 = le32_to_cpu(newext->ee_block);
1890         len1 = ext4_ext_get_actual_len(newext);
1891         depth = ext_depth(inode);
1892         if (!path[depth].p_ext)
1893                 goto out;
1894         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1895
1896         /*
1897          * get the next allocated block if the extent in the path
1898          * is before the requested block(s)
1899          */
1900         if (b2 < b1) {
1901                 b2 = ext4_ext_next_allocated_block(path);
1902                 if (b2 == EXT_MAX_BLOCKS)
1903                         goto out;
1904                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1905         }
1906
1907         /* check for wrap through zero on extent logical start block*/
1908         if (b1 + len1 < b1) {
1909                 len1 = EXT_MAX_BLOCKS - b1;
1910                 newext->ee_len = cpu_to_le16(len1);
1911                 ret = 1;
1912         }
1913
1914         /* check for overlap */
1915         if (b1 + len1 > b2) {
1916                 newext->ee_len = cpu_to_le16(b2 - b1);
1917                 ret = 1;
1918         }
1919 out:
1920         return ret;
1921 }
1922
1923 /*
1924  * ext4_ext_insert_extent:
1925  * tries to merge requsted extent into the existing extent or
1926  * inserts requested extent as new one into the tree,
1927  * creating new leaf in the no-space case.
1928  */
1929 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1930                                 struct ext4_ext_path **ppath,
1931                                 struct ext4_extent *newext, int gb_flags)
1932 {
1933         struct ext4_ext_path *path = *ppath;
1934         struct ext4_extent_header *eh;
1935         struct ext4_extent *ex, *fex;
1936         struct ext4_extent *nearex; /* nearest extent */
1937         struct ext4_ext_path *npath = NULL;
1938         int depth, len, err;
1939         ext4_lblk_t next;
1940         int mb_flags = 0, unwritten;
1941
1942         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1943                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1944         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1945                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1946                 return -EFSCORRUPTED;
1947         }
1948         depth = ext_depth(inode);
1949         ex = path[depth].p_ext;
1950         eh = path[depth].p_hdr;
1951         if (unlikely(path[depth].p_hdr == NULL)) {
1952                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1953                 return -EFSCORRUPTED;
1954         }
1955
1956         /* try to insert block into found extent and return */
1957         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1958
1959                 /*
1960                  * Try to see whether we should rather test the extent on
1961                  * right from ex, or from the left of ex. This is because
1962                  * ext4_find_extent() can return either extent on the
1963                  * left, or on the right from the searched position. This
1964                  * will make merging more effective.
1965                  */
1966                 if (ex < EXT_LAST_EXTENT(eh) &&
1967                     (le32_to_cpu(ex->ee_block) +
1968                     ext4_ext_get_actual_len(ex) <
1969                     le32_to_cpu(newext->ee_block))) {
1970                         ex += 1;
1971                         goto prepend;
1972                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1973                            (le32_to_cpu(newext->ee_block) +
1974                            ext4_ext_get_actual_len(newext) <
1975                            le32_to_cpu(ex->ee_block)))
1976                         ex -= 1;
1977
1978                 /* Try to append newex to the ex */
1979                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1980                         ext_debug("append [%d]%d block to %u:[%d]%d"
1981                                   "(from %llu)\n",
1982                                   ext4_ext_is_unwritten(newext),
1983                                   ext4_ext_get_actual_len(newext),
1984                                   le32_to_cpu(ex->ee_block),
1985                                   ext4_ext_is_unwritten(ex),
1986                                   ext4_ext_get_actual_len(ex),
1987                                   ext4_ext_pblock(ex));
1988                         err = ext4_ext_get_access(handle, inode,
1989                                                   path + depth);
1990                         if (err)
1991                                 return err;
1992                         unwritten = ext4_ext_is_unwritten(ex);
1993                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1994                                         + ext4_ext_get_actual_len(newext));
1995                         if (unwritten)
1996                                 ext4_ext_mark_unwritten(ex);
1997                         eh = path[depth].p_hdr;
1998                         nearex = ex;
1999                         goto merge;
2000                 }
2001
2002 prepend:
2003                 /* Try to prepend newex to the ex */
2004                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2005                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2006                                   "(from %llu)\n",
2007                                   le32_to_cpu(newext->ee_block),
2008                                   ext4_ext_is_unwritten(newext),
2009                                   ext4_ext_get_actual_len(newext),
2010                                   le32_to_cpu(ex->ee_block),
2011                                   ext4_ext_is_unwritten(ex),
2012                                   ext4_ext_get_actual_len(ex),
2013                                   ext4_ext_pblock(ex));
2014                         err = ext4_ext_get_access(handle, inode,
2015                                                   path + depth);
2016                         if (err)
2017                                 return err;
2018
2019                         unwritten = ext4_ext_is_unwritten(ex);
2020                         ex->ee_block = newext->ee_block;
2021                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2022                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2023                                         + ext4_ext_get_actual_len(newext));
2024                         if (unwritten)
2025                                 ext4_ext_mark_unwritten(ex);
2026                         eh = path[depth].p_hdr;
2027                         nearex = ex;
2028                         goto merge;
2029                 }
2030         }
2031
2032         depth = ext_depth(inode);
2033         eh = path[depth].p_hdr;
2034         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2035                 goto has_space;
2036
2037         /* probably next leaf has space for us? */
2038         fex = EXT_LAST_EXTENT(eh);
2039         next = EXT_MAX_BLOCKS;
2040         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2041                 next = ext4_ext_next_leaf_block(path);
2042         if (next != EXT_MAX_BLOCKS) {
2043                 ext_debug("next leaf block - %u\n", next);
2044                 BUG_ON(npath != NULL);
2045                 npath = ext4_find_extent(inode, next, NULL, 0);
2046                 if (IS_ERR(npath))
2047                         return PTR_ERR(npath);
2048                 BUG_ON(npath->p_depth != path->p_depth);
2049                 eh = npath[depth].p_hdr;
2050                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2051                         ext_debug("next leaf isn't full(%d)\n",
2052                                   le16_to_cpu(eh->eh_entries));
2053                         path = npath;
2054                         goto has_space;
2055                 }
2056                 ext_debug("next leaf has no free space(%d,%d)\n",
2057                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2058         }
2059
2060         /*
2061          * There is no free space in the found leaf.
2062          * We're gonna add a new leaf in the tree.
2063          */
2064         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2065                 mb_flags |= EXT4_MB_USE_RESERVED;
2066         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2067                                        ppath, newext);
2068         if (err)
2069                 goto cleanup;
2070         depth = ext_depth(inode);
2071         eh = path[depth].p_hdr;
2072
2073 has_space:
2074         nearex = path[depth].p_ext;
2075
2076         err = ext4_ext_get_access(handle, inode, path + depth);
2077         if (err)
2078                 goto cleanup;
2079
2080         if (!nearex) {
2081                 /* there is no extent in this leaf, create first one */
2082                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2083                                 le32_to_cpu(newext->ee_block),
2084                                 ext4_ext_pblock(newext),
2085                                 ext4_ext_is_unwritten(newext),
2086                                 ext4_ext_get_actual_len(newext));
2087                 nearex = EXT_FIRST_EXTENT(eh);
2088         } else {
2089                 if (le32_to_cpu(newext->ee_block)
2090                            > le32_to_cpu(nearex->ee_block)) {
2091                         /* Insert after */
2092                         ext_debug("insert %u:%llu:[%d]%d before: "
2093                                         "nearest %p\n",
2094                                         le32_to_cpu(newext->ee_block),
2095                                         ext4_ext_pblock(newext),
2096                                         ext4_ext_is_unwritten(newext),
2097                                         ext4_ext_get_actual_len(newext),
2098                                         nearex);
2099                         nearex++;
2100                 } else {
2101                         /* Insert before */
2102                         BUG_ON(newext->ee_block == nearex->ee_block);
2103                         ext_debug("insert %u:%llu:[%d]%d after: "
2104                                         "nearest %p\n",
2105                                         le32_to_cpu(newext->ee_block),
2106                                         ext4_ext_pblock(newext),
2107                                         ext4_ext_is_unwritten(newext),
2108                                         ext4_ext_get_actual_len(newext),
2109                                         nearex);
2110                 }
2111                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2112                 if (len > 0) {
2113                         ext_debug("insert %u:%llu:[%d]%d: "
2114                                         "move %d extents from 0x%p to 0x%p\n",
2115                                         le32_to_cpu(newext->ee_block),
2116                                         ext4_ext_pblock(newext),
2117                                         ext4_ext_is_unwritten(newext),
2118                                         ext4_ext_get_actual_len(newext),
2119                                         len, nearex, nearex + 1);
2120                         memmove(nearex + 1, nearex,
2121                                 len * sizeof(struct ext4_extent));
2122                 }
2123         }
2124
2125         le16_add_cpu(&eh->eh_entries, 1);
2126         path[depth].p_ext = nearex;
2127         nearex->ee_block = newext->ee_block;
2128         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2129         nearex->ee_len = newext->ee_len;
2130
2131 merge:
2132         /* try to merge extents */
2133         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2134                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2135
2136
2137         /* time to correct all indexes above */
2138         err = ext4_ext_correct_indexes(handle, inode, path);
2139         if (err)
2140                 goto cleanup;
2141
2142         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2143
2144 cleanup:
2145         ext4_ext_drop_refs(npath);
2146         kfree(npath);
2147         return err;
2148 }
2149
2150 static int ext4_fill_fiemap_extents(struct inode *inode,
2151                                     ext4_lblk_t block, ext4_lblk_t num,
2152                                     struct fiemap_extent_info *fieinfo)
2153 {
2154         struct ext4_ext_path *path = NULL;
2155         struct ext4_extent *ex;
2156         struct extent_status es;
2157         ext4_lblk_t next, next_del, start = 0, end = 0;
2158         ext4_lblk_t last = block + num;
2159         int exists, depth = 0, err = 0;
2160         unsigned int flags = 0;
2161         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2162
2163         while (block < last && block != EXT_MAX_BLOCKS) {
2164                 num = last - block;
2165                 /* find extent for this block */
2166                 down_read(&EXT4_I(inode)->i_data_sem);
2167
2168                 path = ext4_find_extent(inode, block, &path, 0);
2169                 if (IS_ERR(path)) {
2170                         up_read(&EXT4_I(inode)->i_data_sem);
2171                         err = PTR_ERR(path);
2172                         path = NULL;
2173                         break;
2174                 }
2175
2176                 depth = ext_depth(inode);
2177                 if (unlikely(path[depth].p_hdr == NULL)) {
2178                         up_read(&EXT4_I(inode)->i_data_sem);
2179                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2180                         err = -EFSCORRUPTED;
2181                         break;
2182                 }
2183                 ex = path[depth].p_ext;
2184                 next = ext4_ext_next_allocated_block(path);
2185
2186                 flags = 0;
2187                 exists = 0;
2188                 if (!ex) {
2189                         /* there is no extent yet, so try to allocate
2190                          * all requested space */
2191                         start = block;
2192                         end = block + num;
2193                 } else if (le32_to_cpu(ex->ee_block) > block) {
2194                         /* need to allocate space before found extent */
2195                         start = block;
2196                         end = le32_to_cpu(ex->ee_block);
2197                         if (block + num < end)
2198                                 end = block + num;
2199                 } else if (block >= le32_to_cpu(ex->ee_block)
2200                                         + ext4_ext_get_actual_len(ex)) {
2201                         /* need to allocate space after found extent */
2202                         start = block;
2203                         end = block + num;
2204                         if (end >= next)
2205                                 end = next;
2206                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2207                         /*
2208                          * some part of requested space is covered
2209                          * by found extent
2210                          */
2211                         start = block;
2212                         end = le32_to_cpu(ex->ee_block)
2213                                 + ext4_ext_get_actual_len(ex);
2214                         if (block + num < end)
2215                                 end = block + num;
2216                         exists = 1;
2217                 } else {
2218                         BUG();
2219                 }
2220                 BUG_ON(end <= start);
2221
2222                 if (!exists) {
2223                         es.es_lblk = start;
2224                         es.es_len = end - start;
2225                         es.es_pblk = 0;
2226                 } else {
2227                         es.es_lblk = le32_to_cpu(ex->ee_block);
2228                         es.es_len = ext4_ext_get_actual_len(ex);
2229                         es.es_pblk = ext4_ext_pblock(ex);
2230                         if (ext4_ext_is_unwritten(ex))
2231                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2232                 }
2233
2234                 /*
2235                  * Find delayed extent and update es accordingly. We call
2236                  * it even in !exists case to find out whether es is the
2237                  * last existing extent or not.
2238                  */
2239                 next_del = ext4_find_delayed_extent(inode, &es);
2240                 if (!exists && next_del) {
2241                         exists = 1;
2242                         flags |= (FIEMAP_EXTENT_DELALLOC |
2243                                   FIEMAP_EXTENT_UNKNOWN);
2244                 }
2245                 up_read(&EXT4_I(inode)->i_data_sem);
2246
2247                 if (unlikely(es.es_len == 0)) {
2248                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2249                         err = -EFSCORRUPTED;
2250                         break;
2251                 }
2252
2253                 /*
2254                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2255                  * we need to check next == EXT_MAX_BLOCKS because it is
2256                  * possible that an extent is with unwritten and delayed
2257                  * status due to when an extent is delayed allocated and
2258                  * is allocated by fallocate status tree will track both of
2259                  * them in a extent.
2260                  *
2261                  * So we could return a unwritten and delayed extent, and
2262                  * its block is equal to 'next'.
2263                  */
2264                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2265                         flags |= FIEMAP_EXTENT_LAST;
2266                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2267                                      next != EXT_MAX_BLOCKS)) {
2268                                 EXT4_ERROR_INODE(inode,
2269                                                  "next extent == %u, next "
2270                                                  "delalloc extent = %u",
2271                                                  next, next_del);
2272                                 err = -EFSCORRUPTED;
2273                                 break;
2274                         }
2275                 }
2276
2277                 if (exists) {
2278                         err = fiemap_fill_next_extent(fieinfo,
2279                                 (__u64)es.es_lblk << blksize_bits,
2280                                 (__u64)es.es_pblk << blksize_bits,
2281                                 (__u64)es.es_len << blksize_bits,
2282                                 flags);
2283                         if (err < 0)
2284                                 break;
2285                         if (err == 1) {
2286                                 err = 0;
2287                                 break;
2288                         }
2289                 }
2290
2291                 block = es.es_lblk + es.es_len;
2292         }
2293
2294         ext4_ext_drop_refs(path);
2295         kfree(path);
2296         return err;
2297 }
2298
2299 /*
2300  * ext4_ext_determine_hole - determine hole around given block
2301  * @inode:      inode we lookup in
2302  * @path:       path in extent tree to @lblk
2303  * @lblk:       pointer to logical block around which we want to determine hole
2304  *
2305  * Determine hole length (and start if easily possible) around given logical
2306  * block. We don't try too hard to find the beginning of the hole but @path
2307  * actually points to extent before @lblk, we provide it.
2308  *
2309  * The function returns the length of a hole starting at @lblk. We update @lblk
2310  * to the beginning of the hole if we managed to find it.
2311  */
2312 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2313                                            struct ext4_ext_path *path,
2314                                            ext4_lblk_t *lblk)
2315 {
2316         int depth = ext_depth(inode);
2317         struct ext4_extent *ex;
2318         ext4_lblk_t len;
2319
2320         ex = path[depth].p_ext;
2321         if (ex == NULL) {
2322                 /* there is no extent yet, so gap is [0;-] */
2323                 *lblk = 0;
2324                 len = EXT_MAX_BLOCKS;
2325         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2326                 len = le32_to_cpu(ex->ee_block) - *lblk;
2327         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2328                         + ext4_ext_get_actual_len(ex)) {
2329                 ext4_lblk_t next;
2330
2331                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2332                 next = ext4_ext_next_allocated_block(path);
2333                 BUG_ON(next == *lblk);
2334                 len = next - *lblk;
2335         } else {
2336                 BUG();
2337         }
2338         return len;
2339 }
2340
2341 /*
2342  * ext4_ext_put_gap_in_cache:
2343  * calculate boundaries of the gap that the requested block fits into
2344  * and cache this gap
2345  */
2346 static void
2347 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2348                           ext4_lblk_t hole_len)
2349 {
2350         struct extent_status es;
2351
2352         ext4_es_find_delayed_extent_range(inode, hole_start,
2353                                           hole_start + hole_len - 1, &es);
2354         if (es.es_len) {
2355                 /* There's delayed extent containing lblock? */
2356                 if (es.es_lblk <= hole_start)
2357                         return;
2358                 hole_len = min(es.es_lblk - hole_start, hole_len);
2359         }
2360         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2361         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2362                               EXTENT_STATUS_HOLE);
2363 }
2364
2365 /*
2366  * ext4_ext_rm_idx:
2367  * removes index from the index block.
2368  */
2369 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2370                         struct ext4_ext_path *path, int depth)
2371 {
2372         int err;
2373         ext4_fsblk_t leaf;
2374
2375         /* free index block */
2376         depth--;
2377         path = path + depth;
2378         leaf = ext4_idx_pblock(path->p_idx);
2379         if (unlikely(path->p_hdr->eh_entries == 0)) {
2380                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2381                 return -EFSCORRUPTED;
2382         }
2383         err = ext4_ext_get_access(handle, inode, path);
2384         if (err)
2385                 return err;
2386
2387         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2388                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2389                 len *= sizeof(struct ext4_extent_idx);
2390                 memmove(path->p_idx, path->p_idx + 1, len);
2391         }
2392
2393         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2394         err = ext4_ext_dirty(handle, inode, path);
2395         if (err)
2396                 return err;
2397         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2398         trace_ext4_ext_rm_idx(inode, leaf);
2399
2400         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2401                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2402
2403         while (--depth >= 0) {
2404                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2405                         break;
2406                 path--;
2407                 err = ext4_ext_get_access(handle, inode, path);
2408                 if (err)
2409                         break;
2410                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2411                 err = ext4_ext_dirty(handle, inode, path);
2412                 if (err)
2413                         break;
2414         }
2415         return err;
2416 }
2417
2418 /*
2419  * ext4_ext_calc_credits_for_single_extent:
2420  * This routine returns max. credits that needed to insert an extent
2421  * to the extent tree.
2422  * When pass the actual path, the caller should calculate credits
2423  * under i_data_sem.
2424  */
2425 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2426                                                 struct ext4_ext_path *path)
2427 {
2428         if (path) {
2429                 int depth = ext_depth(inode);
2430                 int ret = 0;
2431
2432                 /* probably there is space in leaf? */
2433                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2434                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2435
2436                         /*
2437                          *  There are some space in the leaf tree, no
2438                          *  need to account for leaf block credit
2439                          *
2440                          *  bitmaps and block group descriptor blocks
2441                          *  and other metadata blocks still need to be
2442                          *  accounted.
2443                          */
2444                         /* 1 bitmap, 1 block group descriptor */
2445                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2446                         return ret;
2447                 }
2448         }
2449
2450         return ext4_chunk_trans_blocks(inode, nrblocks);
2451 }
2452
2453 /*
2454  * How many index/leaf blocks need to change/allocate to add @extents extents?
2455  *
2456  * If we add a single extent, then in the worse case, each tree level
2457  * index/leaf need to be changed in case of the tree split.
2458  *
2459  * If more extents are inserted, they could cause the whole tree split more
2460  * than once, but this is really rare.
2461  */
2462 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2463 {
2464         int index;
2465         int depth;
2466
2467         /* If we are converting the inline data, only one is needed here. */
2468         if (ext4_has_inline_data(inode))
2469                 return 1;
2470
2471         depth = ext_depth(inode);
2472
2473         if (extents <= 1)
2474                 index = depth * 2;
2475         else
2476                 index = depth * 3;
2477
2478         return index;
2479 }
2480
2481 static inline int get_default_free_blocks_flags(struct inode *inode)
2482 {
2483         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2484                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2485         else if (ext4_should_journal_data(inode))
2486                 return EXT4_FREE_BLOCKS_FORGET;
2487         return 0;
2488 }
2489
2490 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2491                               struct ext4_extent *ex,
2492                               long long *partial_cluster,
2493                               ext4_lblk_t from, ext4_lblk_t to)
2494 {
2495         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2496         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2497         ext4_fsblk_t pblk;
2498         int flags = get_default_free_blocks_flags(inode);
2499
2500         /*
2501          * For bigalloc file systems, we never free a partial cluster
2502          * at the beginning of the extent.  Instead, we make a note
2503          * that we tried freeing the cluster, and check to see if we
2504          * need to free it on a subsequent call to ext4_remove_blocks,
2505          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2506          */
2507         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2508
2509         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2510         /*
2511          * If we have a partial cluster, and it's different from the
2512          * cluster of the last block, we need to explicitly free the
2513          * partial cluster here.
2514          */
2515         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2516         if (*partial_cluster > 0 &&
2517             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2518                 ext4_free_blocks(handle, inode, NULL,
2519                                  EXT4_C2B(sbi, *partial_cluster),
2520                                  sbi->s_cluster_ratio, flags);
2521                 *partial_cluster = 0;
2522         }
2523
2524 #ifdef EXTENTS_STATS
2525         {
2526                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2527                 spin_lock(&sbi->s_ext_stats_lock);
2528                 sbi->s_ext_blocks += ee_len;
2529                 sbi->s_ext_extents++;
2530                 if (ee_len < sbi->s_ext_min)
2531                         sbi->s_ext_min = ee_len;
2532                 if (ee_len > sbi->s_ext_max)
2533                         sbi->s_ext_max = ee_len;
2534                 if (ext_depth(inode) > sbi->s_depth_max)
2535                         sbi->s_depth_max = ext_depth(inode);
2536                 spin_unlock(&sbi->s_ext_stats_lock);
2537         }
2538 #endif
2539         if (from >= le32_to_cpu(ex->ee_block)
2540             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2541                 /* tail removal */
2542                 ext4_lblk_t num;
2543                 long long first_cluster;
2544
2545                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2546                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2547                 /*
2548                  * Usually we want to free partial cluster at the end of the
2549                  * extent, except for the situation when the cluster is still
2550                  * used by any other extent (partial_cluster is negative).
2551                  */
2552                 if (*partial_cluster < 0 &&
2553                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2554                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2555
2556                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2557                           num, pblk, *partial_cluster);
2558                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2559                 /*
2560                  * If the block range to be freed didn't start at the
2561                  * beginning of a cluster, and we removed the entire
2562                  * extent and the cluster is not used by any other extent,
2563                  * save the partial cluster here, since we might need to
2564                  * delete if we determine that the truncate or punch hole
2565                  * operation has removed all of the blocks in the cluster.
2566                  * If that cluster is used by another extent, preserve its
2567                  * negative value so it isn't freed later on.
2568                  *
2569                  * If the whole extent wasn't freed, we've reached the
2570                  * start of the truncated/punched region and have finished
2571                  * removing blocks.  If there's a partial cluster here it's
2572                  * shared with the remainder of the extent and is no longer
2573                  * a candidate for removal.
2574                  */
2575                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2576                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2577                         if (first_cluster != -*partial_cluster)
2578                                 *partial_cluster = first_cluster;
2579                 } else {
2580                         *partial_cluster = 0;
2581                 }
2582         } else
2583                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2584                            "%u-%u from %u:%u",
2585                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2586         return 0;
2587 }
2588
2589
2590 /*
2591  * ext4_ext_rm_leaf() Removes the extents associated with the
2592  * blocks appearing between "start" and "end".  Both "start"
2593  * and "end" must appear in the same extent or EIO is returned.
2594  *
2595  * @handle: The journal handle
2596  * @inode:  The files inode
2597  * @path:   The path to the leaf
2598  * @partial_cluster: The cluster which we'll have to free if all extents
2599  *                   has been released from it.  However, if this value is
2600  *                   negative, it's a cluster just to the right of the
2601  *                   punched region and it must not be freed.
2602  * @start:  The first block to remove
2603  * @end:   The last block to remove
2604  */
2605 static int
2606 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2607                  struct ext4_ext_path *path,
2608                  long long *partial_cluster,
2609                  ext4_lblk_t start, ext4_lblk_t end)
2610 {
2611         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2612         int err = 0, correct_index = 0;
2613         int depth = ext_depth(inode), credits;
2614         struct ext4_extent_header *eh;
2615         ext4_lblk_t a, b;
2616         unsigned num;
2617         ext4_lblk_t ex_ee_block;
2618         unsigned short ex_ee_len;
2619         unsigned unwritten = 0;
2620         struct ext4_extent *ex;
2621         ext4_fsblk_t pblk;
2622
2623         /* the header must be checked already in ext4_ext_remove_space() */
2624         ext_debug("truncate since %u in leaf to %u\n", start, end);
2625         if (!path[depth].p_hdr)
2626                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2627         eh = path[depth].p_hdr;
2628         if (unlikely(path[depth].p_hdr == NULL)) {
2629                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2630                 return -EFSCORRUPTED;
2631         }
2632         /* find where to start removing */
2633         ex = path[depth].p_ext;
2634         if (!ex)
2635                 ex = EXT_LAST_EXTENT(eh);
2636
2637         ex_ee_block = le32_to_cpu(ex->ee_block);
2638         ex_ee_len = ext4_ext_get_actual_len(ex);
2639
2640         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2641
2642         while (ex >= EXT_FIRST_EXTENT(eh) &&
2643                         ex_ee_block + ex_ee_len > start) {
2644
2645                 if (ext4_ext_is_unwritten(ex))
2646                         unwritten = 1;
2647                 else
2648                         unwritten = 0;
2649
2650                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2651                           unwritten, ex_ee_len);
2652                 path[depth].p_ext = ex;
2653
2654                 a = ex_ee_block > start ? ex_ee_block : start;
2655                 b = ex_ee_block+ex_ee_len - 1 < end ?
2656                         ex_ee_block+ex_ee_len - 1 : end;
2657
2658                 ext_debug("  border %u:%u\n", a, b);
2659
2660                 /* If this extent is beyond the end of the hole, skip it */
2661                 if (end < ex_ee_block) {
2662                         /*
2663                          * We're going to skip this extent and move to another,
2664                          * so note that its first cluster is in use to avoid
2665                          * freeing it when removing blocks.  Eventually, the
2666                          * right edge of the truncated/punched region will
2667                          * be just to the left.
2668                          */
2669                         if (sbi->s_cluster_ratio > 1) {
2670                                 pblk = ext4_ext_pblock(ex);
2671                                 *partial_cluster =
2672                                         -(long long) EXT4_B2C(sbi, pblk);
2673                         }
2674                         ex--;
2675                         ex_ee_block = le32_to_cpu(ex->ee_block);
2676                         ex_ee_len = ext4_ext_get_actual_len(ex);
2677                         continue;
2678                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2679                         EXT4_ERROR_INODE(inode,
2680                                          "can not handle truncate %u:%u "
2681                                          "on extent %u:%u",
2682                                          start, end, ex_ee_block,
2683                                          ex_ee_block + ex_ee_len - 1);
2684                         err = -EFSCORRUPTED;
2685                         goto out;
2686                 } else if (a != ex_ee_block) {
2687                         /* remove tail of the extent */
2688                         num = a - ex_ee_block;
2689                 } else {
2690                         /* remove whole extent: excellent! */
2691                         num = 0;
2692                 }
2693                 /*
2694                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2695                  * descriptor) for each block group; assume two block
2696                  * groups plus ex_ee_len/blocks_per_block_group for
2697                  * the worst case
2698                  */
2699                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2700                 if (ex == EXT_FIRST_EXTENT(eh)) {
2701                         correct_index = 1;
2702                         credits += (ext_depth(inode)) + 1;
2703                 }
2704                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2705
2706                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2707                 if (err)
2708                         goto out;
2709
2710                 err = ext4_ext_get_access(handle, inode, path + depth);
2711                 if (err)
2712                         goto out;
2713
2714                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2715                                          a, b);
2716                 if (err)
2717                         goto out;
2718
2719                 if (num == 0)
2720                         /* this extent is removed; mark slot entirely unused */
2721                         ext4_ext_store_pblock(ex, 0);
2722
2723                 ex->ee_len = cpu_to_le16(num);
2724                 /*
2725                  * Do not mark unwritten if all the blocks in the
2726                  * extent have been removed.
2727                  */
2728                 if (unwritten && num)
2729                         ext4_ext_mark_unwritten(ex);
2730                 /*
2731                  * If the extent was completely released,
2732                  * we need to remove it from the leaf
2733                  */
2734                 if (num == 0) {
2735                         if (end != EXT_MAX_BLOCKS - 1) {
2736                                 /*
2737                                  * For hole punching, we need to scoot all the
2738                                  * extents up when an extent is removed so that
2739                                  * we dont have blank extents in the middle
2740                                  */
2741                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2742                                         sizeof(struct ext4_extent));
2743
2744                                 /* Now get rid of the one at the end */
2745                                 memset(EXT_LAST_EXTENT(eh), 0,
2746                                         sizeof(struct ext4_extent));
2747                         }
2748                         le16_add_cpu(&eh->eh_entries, -1);
2749                 }
2750
2751                 err = ext4_ext_dirty(handle, inode, path + depth);
2752                 if (err)
2753                         goto out;
2754
2755                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2756                                 ext4_ext_pblock(ex));
2757                 ex--;
2758                 ex_ee_block = le32_to_cpu(ex->ee_block);
2759                 ex_ee_len = ext4_ext_get_actual_len(ex);
2760         }
2761
2762         if (correct_index && eh->eh_entries)
2763                 err = ext4_ext_correct_indexes(handle, inode, path);
2764
2765         /*
2766          * If there's a partial cluster and at least one extent remains in
2767          * the leaf, free the partial cluster if it isn't shared with the
2768          * current extent.  If it is shared with the current extent
2769          * we zero partial_cluster because we've reached the start of the
2770          * truncated/punched region and we're done removing blocks.
2771          */
2772         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2773                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2774                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2775                         ext4_free_blocks(handle, inode, NULL,
2776                                          EXT4_C2B(sbi, *partial_cluster),
2777                                          sbi->s_cluster_ratio,
2778                                          get_default_free_blocks_flags(inode));
2779                 }
2780                 *partial_cluster = 0;
2781         }
2782
2783         /* if this leaf is free, then we should
2784          * remove it from index block above */
2785         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2786                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2787
2788 out:
2789         return err;
2790 }
2791
2792 /*
2793  * ext4_ext_more_to_rm:
2794  * returns 1 if current index has to be freed (even partial)
2795  */
2796 static int
2797 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2798 {
2799         BUG_ON(path->p_idx == NULL);
2800
2801         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2802                 return 0;
2803
2804         /*
2805          * if truncate on deeper level happened, it wasn't partial,
2806          * so we have to consider current index for truncation
2807          */
2808         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2809                 return 0;
2810         return 1;
2811 }
2812
2813 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2814                           ext4_lblk_t end)
2815 {
2816         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2817         int depth = ext_depth(inode);
2818         struct ext4_ext_path *path = NULL;
2819         long long partial_cluster = 0;
2820         handle_t *handle;
2821         int i = 0, err = 0;
2822
2823         ext_debug("truncate since %u to %u\n", start, end);
2824
2825         /* probably first extent we're gonna free will be last in block */
2826         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2827         if (IS_ERR(handle))
2828                 return PTR_ERR(handle);
2829
2830 again:
2831         trace_ext4_ext_remove_space(inode, start, end, depth);
2832
2833         /*
2834          * Check if we are removing extents inside the extent tree. If that
2835          * is the case, we are going to punch a hole inside the extent tree
2836          * so we have to check whether we need to split the extent covering
2837          * the last block to remove so we can easily remove the part of it
2838          * in ext4_ext_rm_leaf().
2839          */
2840         if (end < EXT_MAX_BLOCKS - 1) {
2841                 struct ext4_extent *ex;
2842                 ext4_lblk_t ee_block, ex_end, lblk;
2843                 ext4_fsblk_t pblk;
2844
2845                 /* find extent for or closest extent to this block */
2846                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2847                 if (IS_ERR(path)) {
2848                         ext4_journal_stop(handle);
2849                         return PTR_ERR(path);
2850                 }
2851                 depth = ext_depth(inode);
2852                 /* Leaf not may not exist only if inode has no blocks at all */
2853                 ex = path[depth].p_ext;
2854                 if (!ex) {
2855                         if (depth) {
2856                                 EXT4_ERROR_INODE(inode,
2857                                                  "path[%d].p_hdr == NULL",
2858                                                  depth);
2859                                 err = -EFSCORRUPTED;
2860                         }
2861                         goto out;
2862                 }
2863
2864                 ee_block = le32_to_cpu(ex->ee_block);
2865                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2866
2867                 /*
2868                  * See if the last block is inside the extent, if so split
2869                  * the extent at 'end' block so we can easily remove the
2870                  * tail of the first part of the split extent in
2871                  * ext4_ext_rm_leaf().
2872                  */
2873                 if (end >= ee_block && end < ex_end) {
2874
2875                         /*
2876                          * If we're going to split the extent, note that
2877                          * the cluster containing the block after 'end' is
2878                          * in use to avoid freeing it when removing blocks.
2879                          */
2880                         if (sbi->s_cluster_ratio > 1) {
2881                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2882                                 partial_cluster =
2883                                         -(long long) EXT4_B2C(sbi, pblk);
2884                         }
2885
2886                         /*
2887                          * Split the extent in two so that 'end' is the last
2888                          * block in the first new extent. Also we should not
2889                          * fail removing space due to ENOSPC so try to use
2890                          * reserved block if that happens.
2891                          */
2892                         err = ext4_force_split_extent_at(handle, inode, &path,
2893                                                          end + 1, 1);
2894                         if (err < 0)
2895                                 goto out;
2896
2897                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2898                         /*
2899                          * If there's an extent to the right its first cluster
2900                          * contains the immediate right boundary of the
2901                          * truncated/punched region.  Set partial_cluster to
2902                          * its negative value so it won't be freed if shared
2903                          * with the current extent.  The end < ee_block case
2904                          * is handled in ext4_ext_rm_leaf().
2905                          */
2906                         lblk = ex_end + 1;
2907                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2908                                                     &ex);
2909                         if (err)
2910                                 goto out;
2911                         if (pblk)
2912                                 partial_cluster =
2913                                         -(long long) EXT4_B2C(sbi, pblk);
2914                 }
2915         }
2916         /*
2917          * We start scanning from right side, freeing all the blocks
2918          * after i_size and walking into the tree depth-wise.
2919          */
2920         depth = ext_depth(inode);
2921         if (path) {
2922                 int k = i = depth;
2923                 while (--k > 0)
2924                         path[k].p_block =
2925                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2926         } else {
2927                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2928                                GFP_NOFS);
2929                 if (path == NULL) {
2930                         ext4_journal_stop(handle);
2931                         return -ENOMEM;
2932                 }
2933                 path[0].p_maxdepth = path[0].p_depth = depth;
2934                 path[0].p_hdr = ext_inode_hdr(inode);
2935                 i = 0;
2936
2937                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2938                         err = -EFSCORRUPTED;
2939                         goto out;
2940                 }
2941         }
2942         err = 0;
2943
2944         while (i >= 0 && err == 0) {
2945                 if (i == depth) {
2946                         /* this is leaf block */
2947                         err = ext4_ext_rm_leaf(handle, inode, path,
2948                                                &partial_cluster, start,
2949                                                end);
2950                         /* root level has p_bh == NULL, brelse() eats this */
2951                         brelse(path[i].p_bh);
2952                         path[i].p_bh = NULL;
2953                         i--;
2954                         continue;
2955                 }
2956
2957                 /* this is index block */
2958                 if (!path[i].p_hdr) {
2959                         ext_debug("initialize header\n");
2960                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2961                 }
2962
2963                 if (!path[i].p_idx) {
2964                         /* this level hasn't been touched yet */
2965                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2966                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2967                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2968                                   path[i].p_hdr,
2969                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2970                 } else {
2971                         /* we were already here, see at next index */
2972                         path[i].p_idx--;
2973                 }
2974
2975                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2976                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2977                                 path[i].p_idx);
2978                 if (ext4_ext_more_to_rm(path + i)) {
2979                         struct buffer_head *bh;
2980                         /* go to the next level */
2981                         ext_debug("move to level %d (block %llu)\n",
2982                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2983                         memset(path + i + 1, 0, sizeof(*path));
2984                         bh = read_extent_tree_block(inode,
2985                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2986                                 EXT4_EX_NOCACHE);
2987                         if (IS_ERR(bh)) {
2988                                 /* should we reset i_size? */
2989                                 err = PTR_ERR(bh);
2990                                 break;
2991                         }
2992                         /* Yield here to deal with large extent trees.
2993                          * Should be a no-op if we did IO above. */
2994                         cond_resched();
2995                         if (WARN_ON(i + 1 > depth)) {
2996                                 err = -EFSCORRUPTED;
2997                                 break;
2998                         }
2999                         path[i + 1].p_bh = bh;
3000
3001                         /* save actual number of indexes since this
3002                          * number is changed at the next iteration */
3003                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3004                         i++;
3005                 } else {
3006                         /* we finished processing this index, go up */
3007                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3008                                 /* index is empty, remove it;
3009                                  * handle must be already prepared by the
3010                                  * truncatei_leaf() */
3011                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3012                         }
3013                         /* root level has p_bh == NULL, brelse() eats this */
3014                         brelse(path[i].p_bh);
3015                         path[i].p_bh = NULL;
3016                         i--;
3017                         ext_debug("return to level %d\n", i);
3018                 }
3019         }
3020
3021         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3022                         partial_cluster, path->p_hdr->eh_entries);
3023
3024         /*
3025          * If we still have something in the partial cluster and we have removed
3026          * even the first extent, then we should free the blocks in the partial
3027          * cluster as well.  (This code will only run when there are no leaves
3028          * to the immediate left of the truncated/punched region.)
3029          */
3030         if (partial_cluster > 0 && err == 0) {
3031                 /* don't zero partial_cluster since it's not used afterwards */
3032                 ext4_free_blocks(handle, inode, NULL,
3033                                  EXT4_C2B(sbi, partial_cluster),
3034                                  sbi->s_cluster_ratio,
3035                                  get_default_free_blocks_flags(inode));
3036         }
3037
3038         /* TODO: flexible tree reduction should be here */
3039         if (path->p_hdr->eh_entries == 0) {
3040                 /*
3041                  * truncate to zero freed all the tree,
3042                  * so we need to correct eh_depth
3043                  */
3044                 err = ext4_ext_get_access(handle, inode, path);
3045                 if (err == 0) {
3046                         ext_inode_hdr(inode)->eh_depth = 0;
3047                         ext_inode_hdr(inode)->eh_max =
3048                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3049                         err = ext4_ext_dirty(handle, inode, path);
3050                 }
3051         }
3052 out:
3053         ext4_ext_drop_refs(path);
3054         kfree(path);
3055         path = NULL;
3056         if (err == -EAGAIN)
3057                 goto again;
3058         ext4_journal_stop(handle);
3059
3060         return err;
3061 }
3062
3063 /*
3064  * called at mount time
3065  */
3066 void ext4_ext_init(struct super_block *sb)
3067 {
3068         /*
3069          * possible initialization would be here
3070          */
3071
3072         if (ext4_has_feature_extents(sb)) {
3073 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3074                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3075 #ifdef AGGRESSIVE_TEST
3076                        ", aggressive tests"
3077 #endif
3078 #ifdef CHECK_BINSEARCH
3079                        ", check binsearch"
3080 #endif
3081 #ifdef EXTENTS_STATS
3082                        ", stats"
3083 #endif
3084                        "\n");
3085 #endif
3086 #ifdef EXTENTS_STATS
3087                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3088                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3089                 EXT4_SB(sb)->s_ext_max = 0;
3090 #endif
3091         }
3092 }
3093
3094 /*
3095  * called at umount time
3096  */
3097 void ext4_ext_release(struct super_block *sb)
3098 {
3099         if (!ext4_has_feature_extents(sb))
3100                 return;
3101
3102 #ifdef EXTENTS_STATS
3103         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3104                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3105                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3106                         sbi->s_ext_blocks, sbi->s_ext_extents,
3107                         sbi->s_ext_blocks / sbi->s_ext_extents);
3108                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3109                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3110         }
3111 #endif
3112 }
3113
3114 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3115 {
3116         ext4_lblk_t  ee_block;
3117         ext4_fsblk_t ee_pblock;
3118         unsigned int ee_len;
3119
3120         ee_block  = le32_to_cpu(ex->ee_block);
3121         ee_len    = ext4_ext_get_actual_len(ex);
3122         ee_pblock = ext4_ext_pblock(ex);
3123
3124         if (ee_len == 0)
3125                 return 0;
3126
3127         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3128                                      EXTENT_STATUS_WRITTEN);
3129 }
3130
3131 /* FIXME!! we need to try to merge to left or right after zero-out  */
3132 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3133 {
3134         ext4_fsblk_t ee_pblock;
3135         unsigned int ee_len;
3136
3137         ee_len    = ext4_ext_get_actual_len(ex);
3138         ee_pblock = ext4_ext_pblock(ex);
3139         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3140                                   ee_len);
3141 }
3142
3143 /*
3144  * ext4_split_extent_at() splits an extent at given block.
3145  *
3146  * @handle: the journal handle
3147  * @inode: the file inode
3148  * @path: the path to the extent
3149  * @split: the logical block where the extent is splitted.
3150  * @split_flags: indicates if the extent could be zeroout if split fails, and
3151  *               the states(init or unwritten) of new extents.
3152  * @flags: flags used to insert new extent to extent tree.
3153  *
3154  *
3155  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3156  * of which are deterimined by split_flag.
3157  *
3158  * There are two cases:
3159  *  a> the extent are splitted into two extent.
3160  *  b> split is not needed, and just mark the extent.
3161  *
3162  * return 0 on success.
3163  */
3164 static int ext4_split_extent_at(handle_t *handle,
3165                              struct inode *inode,
3166                              struct ext4_ext_path **ppath,
3167                              ext4_lblk_t split,
3168                              int split_flag,
3169                              int flags)
3170 {
3171         struct ext4_ext_path *path = *ppath;
3172         ext4_fsblk_t newblock;
3173         ext4_lblk_t ee_block;
3174         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3175         struct ext4_extent *ex2 = NULL;
3176         unsigned int ee_len, depth;
3177         int err = 0;
3178
3179         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3180                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3181
3182         ext_debug("ext4_split_extents_at: inode %lu, logical"
3183                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3184
3185         ext4_ext_show_leaf(inode, path);
3186
3187         depth = ext_depth(inode);
3188         ex = path[depth].p_ext;
3189         ee_block = le32_to_cpu(ex->ee_block);
3190         ee_len = ext4_ext_get_actual_len(ex);
3191         newblock = split - ee_block + ext4_ext_pblock(ex);
3192
3193         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3194         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3195                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3196                              EXT4_EXT_MARK_UNWRIT1 |
3197                              EXT4_EXT_MARK_UNWRIT2));
3198
3199         err = ext4_ext_get_access(handle, inode, path + depth);
3200         if (err)
3201                 goto out;
3202
3203         if (split == ee_block) {
3204                 /*
3205                  * case b: block @split is the block that the extent begins with
3206                  * then we just change the state of the extent, and splitting
3207                  * is not needed.
3208                  */
3209                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3210                         ext4_ext_mark_unwritten(ex);
3211                 else
3212                         ext4_ext_mark_initialized(ex);
3213
3214                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3215                         ext4_ext_try_to_merge(handle, inode, path, ex);
3216
3217                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3218                 goto out;
3219         }
3220
3221         /* case a */
3222         memcpy(&orig_ex, ex, sizeof(orig_ex));
3223         ex->ee_len = cpu_to_le16(split - ee_block);
3224         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3225                 ext4_ext_mark_unwritten(ex);
3226
3227         /*
3228          * path may lead to new leaf, not to original leaf any more
3229          * after ext4_ext_insert_extent() returns,
3230          */
3231         err = ext4_ext_dirty(handle, inode, path + depth);
3232         if (err)
3233                 goto fix_extent_len;
3234
3235         ex2 = &newex;
3236         ex2->ee_block = cpu_to_le32(split);
3237         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3238         ext4_ext_store_pblock(ex2, newblock);
3239         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3240                 ext4_ext_mark_unwritten(ex2);
3241
3242         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3243         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3244                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3245                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3246                                 err = ext4_ext_zeroout(inode, ex2);
3247                                 zero_ex.ee_block = ex2->ee_block;
3248                                 zero_ex.ee_len = cpu_to_le16(
3249                                                 ext4_ext_get_actual_len(ex2));
3250                                 ext4_ext_store_pblock(&zero_ex,
3251                                                       ext4_ext_pblock(ex2));
3252                         } else {
3253                                 err = ext4_ext_zeroout(inode, ex);
3254                                 zero_ex.ee_block = ex->ee_block;
3255                                 zero_ex.ee_len = cpu_to_le16(
3256                                                 ext4_ext_get_actual_len(ex));
3257                                 ext4_ext_store_pblock(&zero_ex,
3258                                                       ext4_ext_pblock(ex));
3259                         }
3260                 } else {
3261                         err = ext4_ext_zeroout(inode, &orig_ex);
3262                         zero_ex.ee_block = orig_ex.ee_block;
3263                         zero_ex.ee_len = cpu_to_le16(
3264                                                 ext4_ext_get_actual_len(&orig_ex));
3265                         ext4_ext_store_pblock(&zero_ex,
3266                                               ext4_ext_pblock(&orig_ex));
3267                 }
3268
3269                 if (err)
3270                         goto fix_extent_len;
3271                 /* update the extent length and mark as initialized */
3272                 ex->ee_len = cpu_to_le16(ee_len);
3273                 ext4_ext_try_to_merge(handle, inode, path, ex);
3274                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3275                 if (err)
3276                         goto fix_extent_len;
3277
3278                 /* update extent status tree */
3279                 err = ext4_zeroout_es(inode, &zero_ex);
3280
3281                 goto out;
3282         } else if (err)
3283                 goto fix_extent_len;
3284
3285 out:
3286         ext4_ext_show_leaf(inode, path);
3287         return err;
3288
3289 fix_extent_len:
3290         ex->ee_len = orig_ex.ee_len;
3291         ext4_ext_dirty(handle, inode, path + path->p_depth);
3292         return err;
3293 }
3294
3295 /*
3296  * ext4_split_extents() splits an extent and mark extent which is covered
3297  * by @map as split_flags indicates
3298  *
3299  * It may result in splitting the extent into multiple extents (up to three)
3300  * There are three possibilities:
3301  *   a> There is no split required
3302  *   b> Splits in two extents: Split is happening at either end of the extent
3303  *   c> Splits in three extents: Somone is splitting in middle of the extent
3304  *
3305  */
3306 static int ext4_split_extent(handle_t *handle,
3307                               struct inode *inode,
3308                               struct ext4_ext_path **ppath,
3309                               struct ext4_map_blocks *map,
3310                               int split_flag,
3311                               int flags)
3312 {
3313         struct ext4_ext_path *path = *ppath;
3314         ext4_lblk_t ee_block;
3315         struct ext4_extent *ex;
3316         unsigned int ee_len, depth;
3317         int err = 0;
3318         int unwritten;
3319         int split_flag1, flags1;
3320         int allocated = map->m_len;
3321
3322         depth = ext_depth(inode);
3323         ex = path[depth].p_ext;
3324         ee_block = le32_to_cpu(ex->ee_block);
3325         ee_len = ext4_ext_get_actual_len(ex);
3326         unwritten = ext4_ext_is_unwritten(ex);
3327
3328         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3329                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3330                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3331                 if (unwritten)
3332                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3333                                        EXT4_EXT_MARK_UNWRIT2;
3334                 if (split_flag & EXT4_EXT_DATA_VALID2)
3335                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3336                 err = ext4_split_extent_at(handle, inode, ppath,
3337                                 map->m_lblk + map->m_len, split_flag1, flags1);
3338                 if (err)
3339                         goto out;
3340         } else {
3341                 allocated = ee_len - (map->m_lblk - ee_block);
3342         }
3343         /*
3344          * Update path is required because previous ext4_split_extent_at() may
3345          * result in split of original leaf or extent zeroout.
3346          */
3347         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3348         if (IS_ERR(path))
3349                 return PTR_ERR(path);
3350         depth = ext_depth(inode);
3351         ex = path[depth].p_ext;
3352         if (!ex) {
3353                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3354                                  (unsigned long) map->m_lblk);
3355                 return -EFSCORRUPTED;
3356         }
3357         unwritten = ext4_ext_is_unwritten(ex);
3358         split_flag1 = 0;
3359
3360         if (map->m_lblk >= ee_block) {
3361                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3362                 if (unwritten) {
3363                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3364                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3365                                                      EXT4_EXT_MARK_UNWRIT2);
3366                 }
3367                 err = ext4_split_extent_at(handle, inode, ppath,
3368                                 map->m_lblk, split_flag1, flags);
3369                 if (err)
3370                         goto out;
3371         }
3372
3373         ext4_ext_show_leaf(inode, path);
3374 out:
3375         return err ? err : allocated;
3376 }
3377
3378 /*
3379  * This function is called by ext4_ext_map_blocks() if someone tries to write
3380  * to an unwritten extent. It may result in splitting the unwritten
3381  * extent into multiple extents (up to three - one initialized and two
3382  * unwritten).
3383  * There are three possibilities:
3384  *   a> There is no split required: Entire extent should be initialized
3385  *   b> Splits in two extents: Write is happening at either end of the extent
3386  *   c> Splits in three extents: Somone is writing in middle of the extent
3387  *
3388  * Pre-conditions:
3389  *  - The extent pointed to by 'path' is unwritten.
3390  *  - The extent pointed to by 'path' contains a superset
3391  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3392  *
3393  * Post-conditions on success:
3394  *  - the returned value is the number of blocks beyond map->l_lblk
3395  *    that are allocated and initialized.
3396  *    It is guaranteed to be >= map->m_len.
3397  */
3398 static int ext4_ext_convert_to_initialized(handle_t *handle,
3399                                            struct inode *inode,
3400                                            struct ext4_map_blocks *map,
3401                                            struct ext4_ext_path **ppath,
3402                                            int flags)
3403 {
3404         struct ext4_ext_path *path = *ppath;
3405         struct ext4_sb_info *sbi;
3406         struct ext4_extent_header *eh;
3407         struct ext4_map_blocks split_map;
3408         struct ext4_extent zero_ex;
3409         struct ext4_extent *ex, *abut_ex;
3410         ext4_lblk_t ee_block, eof_block;
3411         unsigned int ee_len, depth, map_len = map->m_len;
3412         int allocated = 0, max_zeroout = 0;
3413         int err = 0;
3414         int split_flag = 0;
3415
3416         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3417                 "block %llu, max_blocks %u\n", inode->i_ino,
3418                 (unsigned long long)map->m_lblk, map_len);
3419
3420         sbi = EXT4_SB(inode->i_sb);
3421         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3422                 inode->i_sb->s_blocksize_bits;
3423         if (eof_block < map->m_lblk + map_len)
3424                 eof_block = map->m_lblk + map_len;
3425
3426         depth = ext_depth(inode);
3427         eh = path[depth].p_hdr;
3428         ex = path[depth].p_ext;
3429         ee_block = le32_to_cpu(ex->ee_block);
3430         ee_len = ext4_ext_get_actual_len(ex);
3431         zero_ex.ee_len = 0;
3432
3433         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3434
3435         /* Pre-conditions */
3436         BUG_ON(!ext4_ext_is_unwritten(ex));
3437         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3438
3439         /*
3440          * Attempt to transfer newly initialized blocks from the currently
3441          * unwritten extent to its neighbor. This is much cheaper
3442          * than an insertion followed by a merge as those involve costly
3443          * memmove() calls. Transferring to the left is the common case in
3444          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3445          * followed by append writes.
3446          *
3447          * Limitations of the current logic:
3448          *  - L1: we do not deal with writes covering the whole extent.
3449          *    This would require removing the extent if the transfer
3450          *    is possible.
3451          *  - L2: we only attempt to merge with an extent stored in the
3452          *    same extent tree node.
3453          */
3454         if ((map->m_lblk == ee_block) &&
3455                 /* See if we can merge left */
3456                 (map_len < ee_len) &&           /*L1*/
3457                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3458                 ext4_lblk_t prev_lblk;
3459                 ext4_fsblk_t prev_pblk, ee_pblk;
3460                 unsigned int prev_len;
3461
3462                 abut_ex = ex - 1;
3463                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3464                 prev_len = ext4_ext_get_actual_len(abut_ex);
3465                 prev_pblk = ext4_ext_pblock(abut_ex);
3466                 ee_pblk = ext4_ext_pblock(ex);
3467
3468                 /*
3469                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3470                  * upon those conditions:
3471                  * - C1: abut_ex is initialized,
3472                  * - C2: abut_ex is logically abutting ex,
3473                  * - C3: abut_ex is physically abutting ex,
3474                  * - C4: abut_ex can receive the additional blocks without
3475                  *   overflowing the (initialized) length limit.
3476                  */
3477                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3478                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3479                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3480                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3481                         err = ext4_ext_get_access(handle, inode, path + depth);
3482                         if (err)
3483                                 goto out;
3484
3485                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3486                                 map, ex, abut_ex);
3487
3488                         /* Shift the start of ex by 'map_len' blocks */
3489                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3490                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3491                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3492                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3493
3494                         /* Extend abut_ex by 'map_len' blocks */
3495                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3496
3497                         /* Result: number of initialized blocks past m_lblk */
3498                         allocated = map_len;
3499                 }
3500         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3501                    (map_len < ee_len) &&        /*L1*/
3502                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3503                 /* See if we can merge right */
3504                 ext4_lblk_t next_lblk;
3505                 ext4_fsblk_t next_pblk, ee_pblk;
3506                 unsigned int next_len;
3507
3508                 abut_ex = ex + 1;
3509                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3510                 next_len = ext4_ext_get_actual_len(abut_ex);
3511                 next_pblk = ext4_ext_pblock(abut_ex);
3512                 ee_pblk = ext4_ext_pblock(ex);
3513
3514                 /*
3515                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3516                  * upon those conditions:
3517                  * - C1: abut_ex is initialized,
3518                  * - C2: abut_ex is logically abutting ex,
3519                  * - C3: abut_ex is physically abutting ex,
3520                  * - C4: abut_ex can receive the additional blocks without
3521                  *   overflowing the (initialized) length limit.
3522                  */
3523                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3524                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3525                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3526                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3527                         err = ext4_ext_get_access(handle, inode, path + depth);
3528                         if (err)
3529                                 goto out;
3530
3531                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3532                                 map, ex, abut_ex);
3533
3534                         /* Shift the start of abut_ex by 'map_len' blocks */
3535                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3536                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3537                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3538                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3539
3540                         /* Extend abut_ex by 'map_len' blocks */
3541                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3542
3543                         /* Result: number of initialized blocks past m_lblk */
3544                         allocated = map_len;
3545                 }
3546         }
3547         if (allocated) {
3548                 /* Mark the block containing both extents as dirty */
3549                 ext4_ext_dirty(handle, inode, path + depth);
3550
3551                 /* Update path to point to the right extent */
3552                 path[depth].p_ext = abut_ex;
3553                 goto out;
3554         } else
3555                 allocated = ee_len - (map->m_lblk - ee_block);
3556
3557         WARN_ON(map->m_lblk < ee_block);
3558         /*
3559          * It is safe to convert extent to initialized via explicit
3560          * zeroout only if extent is fully inside i_size or new_size.
3561          */
3562         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3563
3564         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3565                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3566                         (inode->i_sb->s_blocksize_bits - 10);
3567
3568         if (ext4_encrypted_inode(inode))
3569                 max_zeroout = 0;
3570
3571         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3572         if (max_zeroout && (ee_len <= max_zeroout)) {
3573                 err = ext4_ext_zeroout(inode, ex);
3574                 if (err)
3575                         goto out;
3576                 zero_ex.ee_block = ex->ee_block;
3577                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3578                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3579
3580                 err = ext4_ext_get_access(handle, inode, path + depth);
3581                 if (err)
3582                         goto out;
3583                 ext4_ext_mark_initialized(ex);
3584                 ext4_ext_try_to_merge(handle, inode, path, ex);
3585                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3586                 goto out;
3587         }
3588
3589         /*
3590          * four cases:
3591          * 1. split the extent into three extents.
3592          * 2. split the extent into two extents, zeroout the first half.
3593          * 3. split the extent into two extents, zeroout the second half.
3594          * 4. split the extent into two extents with out zeroout.
3595          */
3596         split_map.m_lblk = map->m_lblk;
3597         split_map.m_len = map->m_len;
3598
3599         if (max_zeroout && (allocated > map->m_len)) {
3600                 if (allocated <= max_zeroout) {
3601                         /* case 3 */
3602                         zero_ex.ee_block =
3603                                          cpu_to_le32(map->m_lblk);
3604                         zero_ex.ee_len = cpu_to_le16(allocated);
3605                         ext4_ext_store_pblock(&zero_ex,
3606                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3607                         err = ext4_ext_zeroout(inode, &zero_ex);
3608                         if (err)
3609                                 goto out;
3610                         split_map.m_lblk = map->m_lblk;
3611                         split_map.m_len = allocated;
3612                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3613                         /* case 2 */
3614                         if (map->m_lblk != ee_block) {
3615                                 zero_ex.ee_block = ex->ee_block;
3616                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3617                                                         ee_block);
3618                                 ext4_ext_store_pblock(&zero_ex,
3619                                                       ext4_ext_pblock(ex));
3620                                 err = ext4_ext_zeroout(inode, &zero_ex);
3621                                 if (err)
3622                                         goto out;
3623                         }
3624
3625                         split_map.m_lblk = ee_block;
3626                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3627                         allocated = map->m_len;
3628                 }
3629         }
3630
3631         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3632                                 flags);
3633         if (err > 0)
3634                 err = 0;
3635 out:
3636         /* If we have gotten a failure, don't zero out status tree */
3637         if (!err)
3638                 err = ext4_zeroout_es(inode, &zero_ex);
3639         return err ? err : allocated;
3640 }
3641
3642 /*
3643  * This function is called by ext4_ext_map_blocks() from
3644  * ext4_get_blocks_dio_write() when DIO to write
3645  * to an unwritten extent.
3646  *
3647  * Writing to an unwritten extent may result in splitting the unwritten
3648  * extent into multiple initialized/unwritten extents (up to three)
3649  * There are three possibilities:
3650  *   a> There is no split required: Entire extent should be unwritten
3651  *   b> Splits in two extents: Write is happening at either end of the extent
3652  *   c> Splits in three extents: Somone is writing in middle of the extent
3653  *
3654  * This works the same way in the case of initialized -> unwritten conversion.
3655  *
3656  * One of more index blocks maybe needed if the extent tree grow after
3657  * the unwritten extent split. To prevent ENOSPC occur at the IO
3658  * complete, we need to split the unwritten extent before DIO submit
3659  * the IO. The unwritten extent called at this time will be split
3660  * into three unwritten extent(at most). After IO complete, the part
3661  * being filled will be convert to initialized by the end_io callback function
3662  * via ext4_convert_unwritten_extents().
3663  *
3664  * Returns the size of unwritten extent to be written on success.
3665  */
3666 static int ext4_split_convert_extents(handle_t *handle,
3667                                         struct inode *inode,
3668                                         struct ext4_map_blocks *map,
3669                                         struct ext4_ext_path **ppath,
3670                                         int flags)
3671 {
3672         struct ext4_ext_path *path = *ppath;
3673         ext4_lblk_t eof_block;
3674         ext4_lblk_t ee_block;
3675         struct ext4_extent *ex;
3676         unsigned int ee_len;
3677         int split_flag = 0, depth;
3678
3679         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3680                   __func__, inode->i_ino,
3681                   (unsigned long long)map->m_lblk, map->m_len);
3682
3683         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3684                 inode->i_sb->s_blocksize_bits;
3685         if (eof_block < map->m_lblk + map->m_len)
3686                 eof_block = map->m_lblk + map->m_len;
3687         /*
3688          * It is safe to convert extent to initialized via explicit
3689          * zeroout only if extent is fully insde i_size or new_size.
3690          */
3691         depth = ext_depth(inode);
3692         ex = path[depth].p_ext;
3693         ee_block = le32_to_cpu(ex->ee_block);
3694         ee_len = ext4_ext_get_actual_len(ex);
3695
3696         /* Convert to unwritten */
3697         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3698                 split_flag |= EXT4_EXT_DATA_VALID1;
3699         /* Convert to initialized */
3700         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3701                 split_flag |= ee_block + ee_len <= eof_block ?
3702                               EXT4_EXT_MAY_ZEROOUT : 0;
3703                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3704         }
3705         flags |= EXT4_GET_BLOCKS_PRE_IO;
3706         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3707 }
3708
3709 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3710                                                 struct inode *inode,
3711                                                 struct ext4_map_blocks *map,
3712                                                 struct ext4_ext_path **ppath)
3713 {
3714         struct ext4_ext_path *path = *ppath;
3715         struct ext4_extent *ex;
3716         ext4_lblk_t ee_block;
3717         unsigned int ee_len;
3718         int depth;
3719         int err = 0;
3720
3721         depth = ext_depth(inode);
3722         ex = path[depth].p_ext;
3723         ee_block = le32_to_cpu(ex->ee_block);
3724         ee_len = ext4_ext_get_actual_len(ex);
3725
3726         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3727                 "block %llu, max_blocks %u\n", inode->i_ino,
3728                   (unsigned long long)ee_block, ee_len);
3729
3730         /* If extent is larger than requested it is a clear sign that we still
3731          * have some extent state machine issues left. So extent_split is still
3732          * required.
3733          * TODO: Once all related issues will be fixed this situation should be
3734          * illegal.
3735          */
3736         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3737 #ifdef EXT4_DEBUG
3738                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3739                              " len %u; IO logical block %llu, len %u",
3740                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3741                              (unsigned long long)map->m_lblk, map->m_len);
3742 #endif
3743                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3744                                                  EXT4_GET_BLOCKS_CONVERT);
3745                 if (err < 0)
3746                         return err;
3747                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3748                 if (IS_ERR(path))
3749                         return PTR_ERR(path);
3750                 depth = ext_depth(inode);
3751                 ex = path[depth].p_ext;
3752         }
3753
3754         err = ext4_ext_get_access(handle, inode, path + depth);
3755         if (err)
3756                 goto out;
3757         /* first mark the extent as initialized */
3758         ext4_ext_mark_initialized(ex);
3759
3760         /* note: ext4_ext_correct_indexes() isn't needed here because
3761          * borders are not changed
3762          */
3763         ext4_ext_try_to_merge(handle, inode, path, ex);
3764
3765         /* Mark modified extent as dirty */
3766         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3767 out:
3768         ext4_ext_show_leaf(inode, path);
3769         return err;
3770 }
3771
3772 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3773                         sector_t block, int count)
3774 {
3775         int i;
3776         for (i = 0; i < count; i++)
3777                 unmap_underlying_metadata(bdev, block + i);
3778 }
3779
3780 /*
3781  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3782  */
3783 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3784                               ext4_lblk_t lblk,
3785                               struct ext4_ext_path *path,
3786                               unsigned int len)
3787 {
3788         int i, depth;
3789         struct ext4_extent_header *eh;
3790         struct ext4_extent *last_ex;
3791
3792         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3793                 return 0;
3794
3795         depth = ext_depth(inode);
3796         eh = path[depth].p_hdr;
3797
3798         /*
3799          * We're going to remove EOFBLOCKS_FL entirely in future so we
3800          * do not care for this case anymore. Simply remove the flag
3801          * if there are no extents.
3802          */
3803         if (unlikely(!eh->eh_entries))
3804                 goto out;
3805         last_ex = EXT_LAST_EXTENT(eh);
3806         /*
3807          * We should clear the EOFBLOCKS_FL flag if we are writing the
3808          * last block in the last extent in the file.  We test this by
3809          * first checking to see if the caller to
3810          * ext4_ext_get_blocks() was interested in the last block (or
3811          * a block beyond the last block) in the current extent.  If
3812          * this turns out to be false, we can bail out from this
3813          * function immediately.
3814          */
3815         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3816             ext4_ext_get_actual_len(last_ex))
3817                 return 0;
3818         /*
3819          * If the caller does appear to be planning to write at or
3820          * beyond the end of the current extent, we then test to see
3821          * if the current extent is the last extent in the file, by
3822          * checking to make sure it was reached via the rightmost node
3823          * at each level of the tree.
3824          */
3825         for (i = depth-1; i >= 0; i--)
3826                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3827                         return 0;
3828 out:
3829         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3830         return ext4_mark_inode_dirty(handle, inode);
3831 }
3832
3833 /**
3834  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3835  *
3836  * Return 1 if there is a delalloc block in the range, otherwise 0.
3837  */
3838 int ext4_find_delalloc_range(struct inode *inode,
3839                              ext4_lblk_t lblk_start,
3840                              ext4_lblk_t lblk_end)
3841 {
3842         struct extent_status es;
3843
3844         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3845         if (es.es_len == 0)
3846                 return 0; /* there is no delay extent in this tree */
3847         else if (es.es_lblk <= lblk_start &&
3848                  lblk_start < es.es_lblk + es.es_len)
3849                 return 1;
3850         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3851                 return 1;
3852         else
3853                 return 0;
3854 }
3855
3856 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3857 {
3858         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3859         ext4_lblk_t lblk_start, lblk_end;
3860         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3861         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3862
3863         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3864 }
3865
3866 /**
3867  * Determines how many complete clusters (out of those specified by the 'map')
3868  * are under delalloc and were reserved quota for.
3869  * This function is called when we are writing out the blocks that were
3870  * originally written with their allocation delayed, but then the space was
3871  * allocated using fallocate() before the delayed allocation could be resolved.
3872  * The cases to look for are:
3873  * ('=' indicated delayed allocated blocks
3874  *  '-' indicates non-delayed allocated blocks)
3875  * (a) partial clusters towards beginning and/or end outside of allocated range
3876  *     are not delalloc'ed.
3877  *      Ex:
3878  *      |----c---=|====c====|====c====|===-c----|
3879  *               |++++++ allocated ++++++|
3880  *      ==> 4 complete clusters in above example
3881  *
3882  * (b) partial cluster (outside of allocated range) towards either end is
3883  *     marked for delayed allocation. In this case, we will exclude that
3884  *     cluster.
3885  *      Ex:
3886  *      |----====c========|========c========|
3887  *           |++++++ allocated ++++++|
3888  *      ==> 1 complete clusters in above example
3889  *
3890  *      Ex:
3891  *      |================c================|
3892  *            |++++++ allocated ++++++|
3893  *      ==> 0 complete clusters in above example
3894  *
3895  * The ext4_da_update_reserve_space will be called only if we
3896  * determine here that there were some "entire" clusters that span
3897  * this 'allocated' range.
3898  * In the non-bigalloc case, this function will just end up returning num_blks
3899  * without ever calling ext4_find_delalloc_range.
3900  */
3901 static unsigned int
3902 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3903                            unsigned int num_blks)
3904 {
3905         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3906         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3907         ext4_lblk_t lblk_from, lblk_to, c_offset;
3908         unsigned int allocated_clusters = 0;
3909
3910         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3911         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3912
3913         /* max possible clusters for this allocation */
3914         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3915
3916         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3917
3918         /* Check towards left side */
3919         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3920         if (c_offset) {
3921                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3922                 lblk_to = lblk_from + c_offset - 1;
3923
3924                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3925                         allocated_clusters--;
3926         }
3927
3928         /* Now check towards right. */
3929         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3930         if (allocated_clusters && c_offset) {
3931                 lblk_from = lblk_start + num_blks;
3932                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3933
3934                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3935                         allocated_clusters--;
3936         }
3937
3938         return allocated_clusters;
3939 }
3940
3941 static int
3942 convert_initialized_extent(handle_t *handle, struct inode *inode,
3943                            struct ext4_map_blocks *map,
3944                            struct ext4_ext_path **ppath,
3945                            unsigned int allocated)
3946 {
3947         struct ext4_ext_path *path = *ppath;
3948         struct ext4_extent *ex;
3949         ext4_lblk_t ee_block;
3950         unsigned int ee_len;
3951         int depth;
3952         int err = 0;
3953
3954         /*
3955          * Make sure that the extent is no bigger than we support with
3956          * unwritten extent
3957          */
3958         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3959                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3960
3961         depth = ext_depth(inode);
3962         ex = path[depth].p_ext;
3963         ee_block = le32_to_cpu(ex->ee_block);
3964         ee_len = ext4_ext_get_actual_len(ex);
3965
3966         ext_debug("%s: inode %lu, logical"
3967                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3968                   (unsigned long long)ee_block, ee_len);
3969
3970         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3971                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3972                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3973                 if (err < 0)
3974                         return err;
3975                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3976                 if (IS_ERR(path))
3977                         return PTR_ERR(path);
3978                 depth = ext_depth(inode);
3979                 ex = path[depth].p_ext;
3980                 if (!ex) {
3981                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3982                                          (unsigned long) map->m_lblk);
3983                         return -EFSCORRUPTED;
3984                 }
3985         }
3986
3987         err = ext4_ext_get_access(handle, inode, path + depth);
3988         if (err)
3989                 return err;
3990         /* first mark the extent as unwritten */
3991         ext4_ext_mark_unwritten(ex);
3992
3993         /* note: ext4_ext_correct_indexes() isn't needed here because
3994          * borders are not changed
3995          */
3996         ext4_ext_try_to_merge(handle, inode, path, ex);
3997
3998         /* Mark modified extent as dirty */
3999         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
4000         if (err)
4001                 return err;
4002         ext4_ext_show_leaf(inode, path);
4003
4004         ext4_update_inode_fsync_trans(handle, inode, 1);
4005         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
4006         if (err)
4007                 return err;
4008         map->m_flags |= EXT4_MAP_UNWRITTEN;
4009         if (allocated > map->m_len)
4010                 allocated = map->m_len;
4011         map->m_len = allocated;
4012         return allocated;
4013 }
4014
4015 static int
4016 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4017                         struct ext4_map_blocks *map,
4018                         struct ext4_ext_path **ppath, int flags,
4019                         unsigned int allocated, ext4_fsblk_t newblock)
4020 {
4021         struct ext4_ext_path *path = *ppath;
4022         int ret = 0;
4023         int err = 0;
4024
4025         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4026                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4027                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4028                   flags, allocated);
4029         ext4_ext_show_leaf(inode, path);
4030
4031         /*
4032          * When writing into unwritten space, we should not fail to
4033          * allocate metadata blocks for the new extent block if needed.
4034          */
4035         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4036
4037         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4038                                                     allocated, newblock);
4039
4040         /* get_block() before submit the IO, split the extent */
4041         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4042                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4043                                          flags | EXT4_GET_BLOCKS_CONVERT);
4044                 if (ret <= 0)
4045                         goto out;
4046                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4047                 goto out;
4048         }
4049         /* IO end_io complete, convert the filled extent to written */
4050         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4051                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4052                         if (allocated > map->m_len)
4053                                 allocated = map->m_len;
4054                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4055                                                  allocated);
4056                         if (err < 0)
4057                                 goto out2;
4058                 }
4059                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4060                                                            ppath);
4061                 if (ret >= 0) {
4062                         ext4_update_inode_fsync_trans(handle, inode, 1);
4063                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4064                                                  path, map->m_len);
4065                 } else
4066                         err = ret;
4067                 map->m_flags |= EXT4_MAP_MAPPED;
4068                 map->m_pblk = newblock;
4069                 if (allocated > map->m_len)
4070                         allocated = map->m_len;
4071                 map->m_len = allocated;
4072                 goto out2;
4073         }
4074         /* buffered IO case */
4075         /*
4076          * repeat fallocate creation request
4077          * we already have an unwritten extent
4078          */
4079         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4080                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4081                 goto map_out;
4082         }
4083
4084         /* buffered READ or buffered write_begin() lookup */
4085         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4086                 /*
4087                  * We have blocks reserved already.  We
4088                  * return allocated blocks so that delalloc
4089                  * won't do block reservation for us.  But
4090                  * the buffer head will be unmapped so that
4091                  * a read from the block returns 0s.
4092                  */
4093                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4094                 goto out1;
4095         }
4096
4097         /* buffered write, writepage time, convert*/
4098         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4099         if (ret >= 0)
4100                 ext4_update_inode_fsync_trans(handle, inode, 1);
4101 out:
4102         if (ret <= 0) {
4103                 err = ret;
4104                 goto out2;
4105         } else
4106                 allocated = ret;
4107         map->m_flags |= EXT4_MAP_NEW;
4108         /*
4109          * if we allocated more blocks than requested
4110          * we need to make sure we unmap the extra block
4111          * allocated. The actual needed block will get
4112          * unmapped later when we find the buffer_head marked
4113          * new.
4114          */
4115         if (allocated > map->m_len) {
4116                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
4117                                         newblock + map->m_len,
4118                                         allocated - map->m_len);
4119                 allocated = map->m_len;
4120         }
4121         map->m_len = allocated;
4122
4123         /*
4124          * If we have done fallocate with the offset that is already
4125          * delayed allocated, we would have block reservation
4126          * and quota reservation done in the delayed write path.
4127          * But fallocate would have already updated quota and block
4128          * count for this offset. So cancel these reservation
4129          */
4130         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4131                 unsigned int reserved_clusters;
4132                 reserved_clusters = get_reserved_cluster_alloc(inode,
4133                                 map->m_lblk, map->m_len);
4134                 if (reserved_clusters)
4135                         ext4_da_update_reserve_space(inode,
4136                                                      reserved_clusters,
4137                                                      0);
4138         }
4139
4140 map_out:
4141         map->m_flags |= EXT4_MAP_MAPPED;
4142         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4143                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4144                                          map->m_len);
4145                 if (err < 0)
4146                         goto out2;
4147         }
4148 out1:
4149         if (allocated > map->m_len)
4150                 allocated = map->m_len;
4151         ext4_ext_show_leaf(inode, path);
4152         map->m_pblk = newblock;
4153         map->m_len = allocated;
4154 out2:
4155         return err ? err : allocated;
4156 }
4157
4158 /*
4159  * get_implied_cluster_alloc - check to see if the requested
4160  * allocation (in the map structure) overlaps with a cluster already
4161  * allocated in an extent.
4162  *      @sb     The filesystem superblock structure
4163  *      @map    The requested lblk->pblk mapping
4164  *      @ex     The extent structure which might contain an implied
4165  *                      cluster allocation
4166  *
4167  * This function is called by ext4_ext_map_blocks() after we failed to
4168  * find blocks that were already in the inode's extent tree.  Hence,
4169  * we know that the beginning of the requested region cannot overlap
4170  * the extent from the inode's extent tree.  There are three cases we
4171  * want to catch.  The first is this case:
4172  *
4173  *               |--- cluster # N--|
4174  *    |--- extent ---|  |---- requested region ---|
4175  *                      |==========|
4176  *
4177  * The second case that we need to test for is this one:
4178  *
4179  *   |--------- cluster # N ----------------|
4180  *         |--- requested region --|   |------- extent ----|
4181  *         |=======================|
4182  *
4183  * The third case is when the requested region lies between two extents
4184  * within the same cluster:
4185  *          |------------- cluster # N-------------|
4186  * |----- ex -----|                  |---- ex_right ----|
4187  *                  |------ requested region ------|
4188  *                  |================|
4189  *
4190  * In each of the above cases, we need to set the map->m_pblk and
4191  * map->m_len so it corresponds to the return the extent labelled as
4192  * "|====|" from cluster #N, since it is already in use for data in
4193  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4194  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4195  * as a new "allocated" block region.  Otherwise, we will return 0 and
4196  * ext4_ext_map_blocks() will then allocate one or more new clusters
4197  * by calling ext4_mb_new_blocks().
4198  */
4199 static int get_implied_cluster_alloc(struct super_block *sb,
4200                                      struct ext4_map_blocks *map,
4201                                      struct ext4_extent *ex,
4202                                      struct ext4_ext_path *path)
4203 {
4204         struct ext4_sb_info *sbi = EXT4_SB(sb);
4205         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4206         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4207         ext4_lblk_t rr_cluster_start;
4208         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4209         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4210         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4211
4212         /* The extent passed in that we are trying to match */
4213         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4214         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4215
4216         /* The requested region passed into ext4_map_blocks() */
4217         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4218
4219         if ((rr_cluster_start == ex_cluster_end) ||
4220             (rr_cluster_start == ex_cluster_start)) {
4221                 if (rr_cluster_start == ex_cluster_end)
4222                         ee_start += ee_len - 1;
4223                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4224                 map->m_len = min(map->m_len,
4225                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4226                 /*
4227                  * Check for and handle this case:
4228                  *
4229                  *   |--------- cluster # N-------------|
4230                  *                     |------- extent ----|
4231                  *         |--- requested region ---|
4232                  *         |===========|
4233                  */
4234
4235                 if (map->m_lblk < ee_block)
4236                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4237
4238                 /*
4239                  * Check for the case where there is already another allocated
4240                  * block to the right of 'ex' but before the end of the cluster.
4241                  *
4242                  *          |------------- cluster # N-------------|
4243                  * |----- ex -----|                  |---- ex_right ----|
4244                  *                  |------ requested region ------|
4245                  *                  |================|
4246                  */
4247                 if (map->m_lblk > ee_block) {
4248                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4249                         map->m_len = min(map->m_len, next - map->m_lblk);
4250                 }
4251
4252                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4253                 return 1;
4254         }
4255
4256         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4257         return 0;
4258 }
4259
4260
4261 /*
4262  * Block allocation/map/preallocation routine for extents based files
4263  *
4264  *
4265  * Need to be called with
4266  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4267  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4268  *
4269  * return > 0, number of of blocks already mapped/allocated
4270  *          if create == 0 and these are pre-allocated blocks
4271  *              buffer head is unmapped
4272  *          otherwise blocks are mapped
4273  *
4274  * return = 0, if plain look up failed (blocks have not been allocated)
4275  *          buffer head is unmapped
4276  *
4277  * return < 0, error case.
4278  */
4279 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4280                         struct ext4_map_blocks *map, int flags)
4281 {
4282         struct ext4_ext_path *path = NULL;
4283         struct ext4_extent newex, *ex, *ex2;
4284         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4285         ext4_fsblk_t newblock = 0;
4286         int free_on_err = 0, err = 0, depth, ret;
4287         unsigned int allocated = 0, offset = 0;
4288         unsigned int allocated_clusters = 0;
4289         struct ext4_allocation_request ar;
4290         ext4_lblk_t cluster_offset;
4291         bool map_from_cluster = false;
4292
4293         ext_debug("blocks %u/%u requested for inode %lu\n",
4294                   map->m_lblk, map->m_len, inode->i_ino);
4295         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4296
4297         /* find extent for this block */
4298         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4299         if (IS_ERR(path)) {
4300                 err = PTR_ERR(path);
4301                 path = NULL;
4302                 goto out2;
4303         }
4304
4305         depth = ext_depth(inode);
4306
4307         /*
4308          * consistent leaf must not be empty;
4309          * this situation is possible, though, _during_ tree modification;
4310          * this is why assert can't be put in ext4_find_extent()
4311          */
4312         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4313                 EXT4_ERROR_INODE(inode, "bad extent address "
4314                                  "lblock: %lu, depth: %d pblock %lld",
4315                                  (unsigned long) map->m_lblk, depth,
4316                                  path[depth].p_block);
4317                 err = -EFSCORRUPTED;
4318                 goto out2;
4319         }
4320
4321         ex = path[depth].p_ext;
4322         if (ex) {
4323                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4324                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4325                 unsigned short ee_len;
4326
4327
4328                 /*
4329                  * unwritten extents are treated as holes, except that
4330                  * we split out initialized portions during a write.
4331                  */
4332                 ee_len = ext4_ext_get_actual_len(ex);
4333
4334                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4335
4336                 /* if found extent covers block, simply return it */
4337                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4338                         newblock = map->m_lblk - ee_block + ee_start;
4339                         /* number of remaining blocks in the extent */
4340                         allocated = ee_len - (map->m_lblk - ee_block);
4341                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4342                                   ee_block, ee_len, newblock);
4343
4344                         /*
4345                          * If the extent is initialized check whether the
4346                          * caller wants to convert it to unwritten.
4347                          */
4348                         if ((!ext4_ext_is_unwritten(ex)) &&
4349                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4350                                 allocated = convert_initialized_extent(
4351                                                 handle, inode, map, &path,
4352                                                 allocated);
4353                                 goto out2;
4354                         } else if (!ext4_ext_is_unwritten(ex))
4355                                 goto out;
4356
4357                         ret = ext4_ext_handle_unwritten_extents(
4358                                 handle, inode, map, &path, flags,
4359                                 allocated, newblock);
4360                         if (ret < 0)
4361                                 err = ret;
4362                         else
4363                                 allocated = ret;
4364                         goto out2;
4365                 }
4366         }
4367
4368         /*
4369          * requested block isn't allocated yet;
4370          * we couldn't try to create block if create flag is zero
4371          */
4372         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4373                 ext4_lblk_t hole_start, hole_len;
4374
4375                 hole_start = map->m_lblk;
4376                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4377                 /*
4378                  * put just found gap into cache to speed up
4379                  * subsequent requests
4380                  */
4381                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4382
4383                 /* Update hole_len to reflect hole size after map->m_lblk */
4384                 if (hole_start != map->m_lblk)
4385                         hole_len -= map->m_lblk - hole_start;
4386                 map->m_pblk = 0;
4387                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4388
4389                 goto out2;
4390         }
4391
4392         /*
4393          * Okay, we need to do block allocation.
4394          */
4395         newex.ee_block = cpu_to_le32(map->m_lblk);
4396         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4397
4398         /*
4399          * If we are doing bigalloc, check to see if the extent returned
4400          * by ext4_find_extent() implies a cluster we can use.
4401          */
4402         if (cluster_offset && ex &&
4403             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4404                 ar.len = allocated = map->m_len;
4405                 newblock = map->m_pblk;
4406                 map_from_cluster = true;
4407                 goto got_allocated_blocks;
4408         }
4409
4410         /* find neighbour allocated blocks */
4411         ar.lleft = map->m_lblk;
4412         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4413         if (err)
4414                 goto out2;
4415         ar.lright = map->m_lblk;
4416         ex2 = NULL;
4417         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4418         if (err)
4419                 goto out2;
4420
4421         /* Check if the extent after searching to the right implies a
4422          * cluster we can use. */
4423         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4424             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4425                 ar.len = allocated = map->m_len;
4426                 newblock = map->m_pblk;
4427                 map_from_cluster = true;
4428                 goto got_allocated_blocks;
4429         }
4430
4431         /*
4432          * See if request is beyond maximum number of blocks we can have in
4433          * a single extent. For an initialized extent this limit is
4434          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4435          * EXT_UNWRITTEN_MAX_LEN.
4436          */
4437         if (map->m_len > EXT_INIT_MAX_LEN &&
4438             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4439                 map->m_len = EXT_INIT_MAX_LEN;
4440         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4441                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4442                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4443
4444         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4445         newex.ee_len = cpu_to_le16(map->m_len);
4446         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4447         if (err)
4448                 allocated = ext4_ext_get_actual_len(&newex);
4449         else
4450                 allocated = map->m_len;
4451
4452         /* allocate new block */
4453         ar.inode = inode;
4454         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4455         ar.logical = map->m_lblk;
4456         /*
4457          * We calculate the offset from the beginning of the cluster
4458          * for the logical block number, since when we allocate a
4459          * physical cluster, the physical block should start at the
4460          * same offset from the beginning of the cluster.  This is
4461          * needed so that future calls to get_implied_cluster_alloc()
4462          * work correctly.
4463          */
4464         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4465         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4466         ar.goal -= offset;
4467         ar.logical -= offset;
4468         if (S_ISREG(inode->i_mode))
4469                 ar.flags = EXT4_MB_HINT_DATA;
4470         else
4471                 /* disable in-core preallocation for non-regular files */
4472                 ar.flags = 0;
4473         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4474                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4475         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4476                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4477         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4478                 ar.flags |= EXT4_MB_USE_RESERVED;
4479         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4480         if (!newblock)
4481                 goto out2;
4482         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4483                   ar.goal, newblock, allocated);
4484         free_on_err = 1;
4485         allocated_clusters = ar.len;
4486         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4487         if (ar.len > allocated)
4488                 ar.len = allocated;
4489
4490 got_allocated_blocks:
4491         /* try to insert new extent into found leaf and return */
4492         ext4_ext_store_pblock(&newex, newblock + offset);
4493         newex.ee_len = cpu_to_le16(ar.len);
4494         /* Mark unwritten */
4495         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4496                 ext4_ext_mark_unwritten(&newex);
4497                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4498         }
4499
4500         err = 0;
4501         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4502                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4503                                          path, ar.len);
4504         if (!err)
4505                 err = ext4_ext_insert_extent(handle, inode, &path,
4506                                              &newex, flags);
4507
4508         if (err && free_on_err) {
4509                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4510                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4511                 /* free data blocks we just allocated */
4512                 /* not a good idea to call discard here directly,
4513                  * but otherwise we'd need to call it every free() */
4514                 ext4_discard_preallocations(inode);
4515                 ext4_free_blocks(handle, inode, NULL, newblock,
4516                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4517                 goto out2;
4518         }
4519
4520         /* previous routine could use block we allocated */
4521         newblock = ext4_ext_pblock(&newex);
4522         allocated = ext4_ext_get_actual_len(&newex);
4523         if (allocated > map->m_len)
4524                 allocated = map->m_len;
4525         map->m_flags |= EXT4_MAP_NEW;
4526
4527         /*
4528          * Update reserved blocks/metadata blocks after successful
4529          * block allocation which had been deferred till now.
4530          */
4531         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4532                 unsigned int reserved_clusters;
4533                 /*
4534                  * Check how many clusters we had reserved this allocated range
4535                  */
4536                 reserved_clusters = get_reserved_cluster_alloc(inode,
4537                                                 map->m_lblk, allocated);
4538                 if (!map_from_cluster) {
4539                         BUG_ON(allocated_clusters < reserved_clusters);
4540                         if (reserved_clusters < allocated_clusters) {
4541                                 struct ext4_inode_info *ei = EXT4_I(inode);
4542                                 int reservation = allocated_clusters -
4543                                                   reserved_clusters;
4544                                 /*
4545                                  * It seems we claimed few clusters outside of
4546                                  * the range of this allocation. We should give
4547                                  * it back to the reservation pool. This can
4548                                  * happen in the following case:
4549                                  *
4550                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4551                                  *   cluster has 4 blocks. Thus, the clusters
4552                                  *   are [0-3],[4-7],[8-11]...
4553                                  * * First comes delayed allocation write for
4554                                  *   logical blocks 10 & 11. Since there were no
4555                                  *   previous delayed allocated blocks in the
4556                                  *   range [8-11], we would reserve 1 cluster
4557                                  *   for this write.
4558                                  * * Next comes write for logical blocks 3 to 8.
4559                                  *   In this case, we will reserve 2 clusters
4560                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4561                                  *   that range has a delayed allocated blocks.
4562                                  *   Thus total reserved clusters now becomes 3.
4563                                  * * Now, during the delayed allocation writeout
4564                                  *   time, we will first write blocks [3-8] and
4565                                  *   allocate 3 clusters for writing these
4566                                  *   blocks. Also, we would claim all these
4567                                  *   three clusters above.
4568                                  * * Now when we come here to writeout the
4569                                  *   blocks [10-11], we would expect to claim
4570                                  *   the reservation of 1 cluster we had made
4571                                  *   (and we would claim it since there are no
4572                                  *   more delayed allocated blocks in the range
4573                                  *   [8-11]. But our reserved cluster count had
4574                                  *   already gone to 0.
4575                                  *
4576                                  *   Thus, at the step 4 above when we determine
4577                                  *   that there are still some unwritten delayed
4578                                  *   allocated blocks outside of our current
4579                                  *   block range, we should increment the
4580                                  *   reserved clusters count so that when the
4581                                  *   remaining blocks finally gets written, we
4582                                  *   could claim them.
4583                                  */
4584                                 dquot_reserve_block(inode,
4585                                                 EXT4_C2B(sbi, reservation));
4586                                 spin_lock(&ei->i_block_reservation_lock);
4587                                 ei->i_reserved_data_blocks += reservation;
4588                                 spin_unlock(&ei->i_block_reservation_lock);
4589                         }
4590                         /*
4591                          * We will claim quota for all newly allocated blocks.
4592                          * We're updating the reserved space *after* the
4593                          * correction above so we do not accidentally free
4594                          * all the metadata reservation because we might
4595                          * actually need it later on.
4596                          */
4597                         ext4_da_update_reserve_space(inode, allocated_clusters,
4598                                                         1);
4599                 }
4600         }
4601
4602         /*
4603          * Cache the extent and update transaction to commit on fdatasync only
4604          * when it is _not_ an unwritten extent.
4605          */
4606         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4607                 ext4_update_inode_fsync_trans(handle, inode, 1);
4608         else
4609                 ext4_update_inode_fsync_trans(handle, inode, 0);
4610 out:
4611         if (allocated > map->m_len)
4612                 allocated = map->m_len;
4613         ext4_ext_show_leaf(inode, path);
4614         map->m_flags |= EXT4_MAP_MAPPED;
4615         map->m_pblk = newblock;
4616         map->m_len = allocated;
4617 out2:
4618         ext4_ext_drop_refs(path);
4619         kfree(path);
4620
4621         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4622                                        err ? err : allocated);
4623         return err ? err : allocated;
4624 }
4625
4626 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4627 {
4628         struct super_block *sb = inode->i_sb;
4629         ext4_lblk_t last_block;
4630         int err = 0;
4631
4632         /*
4633          * TODO: optimization is possible here.
4634          * Probably we need not scan at all,
4635          * because page truncation is enough.
4636          */
4637
4638         /* we have to know where to truncate from in crash case */
4639         EXT4_I(inode)->i_disksize = inode->i_size;
4640         ext4_mark_inode_dirty(handle, inode);
4641
4642         last_block = (inode->i_size + sb->s_blocksize - 1)
4643                         >> EXT4_BLOCK_SIZE_BITS(sb);
4644 retry:
4645         err = ext4_es_remove_extent(inode, last_block,
4646                                     EXT_MAX_BLOCKS - last_block);
4647         if (err == -ENOMEM) {
4648                 cond_resched();
4649                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4650                 goto retry;
4651         }
4652         if (err) {
4653                 ext4_std_error(inode->i_sb, err);
4654                 return;
4655         }
4656         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4657         ext4_std_error(inode->i_sb, err);
4658 }
4659
4660 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4661                                   ext4_lblk_t len, loff_t new_size,
4662                                   int flags, int mode)
4663 {
4664         struct inode *inode = file_inode(file);
4665         handle_t *handle;
4666         int ret = 0;
4667         int ret2 = 0;
4668         int retries = 0;
4669         int depth = 0;
4670         struct ext4_map_blocks map;
4671         unsigned int credits;
4672         loff_t epos;
4673
4674         map.m_lblk = offset;
4675         map.m_len = len;
4676         /*
4677          * Don't normalize the request if it can fit in one extent so
4678          * that it doesn't get unnecessarily split into multiple
4679          * extents.
4680          */
4681         if (len <= EXT_UNWRITTEN_MAX_LEN)
4682                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4683
4684         /*
4685          * credits to insert 1 extent into extent tree
4686          */
4687         credits = ext4_chunk_trans_blocks(inode, len);
4688         /*
4689          * We can only call ext_depth() on extent based inodes
4690          */
4691         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4692                 depth = ext_depth(inode);
4693         else
4694                 depth = -1;
4695
4696 retry:
4697         while (ret >= 0 && len) {
4698                 /*
4699                  * Recalculate credits when extent tree depth changes.
4700                  */
4701                 if (depth >= 0 && depth != ext_depth(inode)) {
4702                         credits = ext4_chunk_trans_blocks(inode, len);
4703                         depth = ext_depth(inode);
4704                 }
4705
4706                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4707                                             credits);
4708                 if (IS_ERR(handle)) {
4709                         ret = PTR_ERR(handle);
4710                         break;
4711                 }
4712                 ret = ext4_map_blocks(handle, inode, &map, flags);
4713                 if (ret <= 0) {
4714                         ext4_debug("inode #%lu: block %u: len %u: "
4715                                    "ext4_ext_map_blocks returned %d",
4716                                    inode->i_ino, map.m_lblk,
4717                                    map.m_len, ret);
4718                         ext4_mark_inode_dirty(handle, inode);
4719                         ret2 = ext4_journal_stop(handle);
4720                         break;
4721                 }
4722                 map.m_lblk += ret;
4723                 map.m_len = len = len - ret;
4724                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4725                 inode->i_ctime = ext4_current_time(inode);
4726                 if (new_size) {
4727                         if (epos > new_size)
4728                                 epos = new_size;
4729                         if (ext4_update_inode_size(inode, epos) & 0x1)
4730                                 inode->i_mtime = inode->i_ctime;
4731                 } else {
4732                         if (epos > inode->i_size)
4733                                 ext4_set_inode_flag(inode,
4734                                                     EXT4_INODE_EOFBLOCKS);
4735                 }
4736                 ext4_mark_inode_dirty(handle, inode);
4737                 ret2 = ext4_journal_stop(handle);
4738                 if (ret2)
4739                         break;
4740         }
4741         if (ret == -ENOSPC &&
4742                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4743                 ret = 0;
4744                 goto retry;
4745         }
4746
4747         return ret > 0 ? ret2 : ret;
4748 }
4749
4750 static long ext4_zero_range(struct file *file, loff_t offset,
4751                             loff_t len, int mode)
4752 {
4753         struct inode *inode = file_inode(file);
4754         handle_t *handle = NULL;
4755         unsigned int max_blocks;
4756         loff_t new_size = 0;
4757         int ret = 0;
4758         int flags;
4759         int credits;
4760         int partial_begin, partial_end;
4761         loff_t start, end;
4762         ext4_lblk_t lblk;
4763         unsigned int blkbits = inode->i_blkbits;
4764
4765         trace_ext4_zero_range(inode, offset, len, mode);
4766
4767         if (!S_ISREG(inode->i_mode))
4768                 return -EINVAL;
4769
4770         /* Call ext4_force_commit to flush all data in case of data=journal. */
4771         if (ext4_should_journal_data(inode)) {
4772                 ret = ext4_force_commit(inode->i_sb);
4773                 if (ret)
4774                         return ret;
4775         }
4776
4777         /*
4778          * Round up offset. This is not fallocate, we neet to zero out
4779          * blocks, so convert interior block aligned part of the range to
4780          * unwritten and possibly manually zero out unaligned parts of the
4781          * range.
4782          */
4783         start = round_up(offset, 1 << blkbits);
4784         end = round_down((offset + len), 1 << blkbits);
4785
4786         if (start < offset || end > offset + len)
4787                 return -EINVAL;
4788         partial_begin = offset & ((1 << blkbits) - 1);
4789         partial_end = (offset + len) & ((1 << blkbits) - 1);
4790
4791         lblk = start >> blkbits;
4792         max_blocks = (end >> blkbits);
4793         if (max_blocks < lblk)
4794                 max_blocks = 0;
4795         else
4796                 max_blocks -= lblk;
4797
4798         inode_lock(inode);
4799
4800         /*
4801          * Indirect files do not support unwritten extnets
4802          */
4803         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4804                 ret = -EOPNOTSUPP;
4805                 goto out_mutex;
4806         }
4807
4808         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4809              offset + len > i_size_read(inode)) {
4810                 new_size = offset + len;
4811                 ret = inode_newsize_ok(inode, new_size);
4812                 if (ret)
4813                         goto out_mutex;
4814         }
4815
4816         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4817         if (mode & FALLOC_FL_KEEP_SIZE)
4818                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4819
4820         /* Wait all existing dio workers, newcomers will block on i_mutex */
4821         ext4_inode_block_unlocked_dio(inode);
4822         inode_dio_wait(inode);
4823
4824         /* Preallocate the range including the unaligned edges */
4825         if (partial_begin || partial_end) {
4826                 ret = ext4_alloc_file_blocks(file,
4827                                 round_down(offset, 1 << blkbits) >> blkbits,
4828                                 (round_up((offset + len), 1 << blkbits) -
4829                                  round_down(offset, 1 << blkbits)) >> blkbits,
4830                                 new_size, flags, mode);
4831                 if (ret)
4832                         goto out_dio;
4833
4834         }
4835
4836         /* Zero range excluding the unaligned edges */
4837         if (max_blocks > 0) {
4838                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4839                           EXT4_EX_NOCACHE);
4840
4841                 /*
4842                  * Prevent page faults from reinstantiating pages we have
4843                  * released from page cache.
4844                  */
4845                 down_write(&EXT4_I(inode)->i_mmap_sem);
4846                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4847                 if (ret) {
4848                         up_write(&EXT4_I(inode)->i_mmap_sem);
4849                         goto out_dio;
4850                 }
4851                 /* Now release the pages and zero block aligned part of pages */
4852                 truncate_pagecache_range(inode, start, end - 1);
4853                 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4854
4855                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4856                                              flags, mode);
4857                 up_write(&EXT4_I(inode)->i_mmap_sem);
4858                 if (ret)
4859                         goto out_dio;
4860         }
4861         if (!partial_begin && !partial_end)
4862                 goto out_dio;
4863
4864         /*
4865          * In worst case we have to writeout two nonadjacent unwritten
4866          * blocks and update the inode
4867          */
4868         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4869         if (ext4_should_journal_data(inode))
4870                 credits += 2;
4871         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4872         if (IS_ERR(handle)) {
4873                 ret = PTR_ERR(handle);
4874                 ext4_std_error(inode->i_sb, ret);
4875                 goto out_dio;
4876         }
4877
4878         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4879         if (new_size) {
4880                 ext4_update_inode_size(inode, new_size);
4881         } else {
4882                 /*
4883                 * Mark that we allocate beyond EOF so the subsequent truncate
4884                 * can proceed even if the new size is the same as i_size.
4885                 */
4886                 if ((offset + len) > i_size_read(inode))
4887                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4888         }
4889         ext4_mark_inode_dirty(handle, inode);
4890
4891         /* Zero out partial block at the edges of the range */
4892         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4893
4894         if (file->f_flags & O_SYNC)
4895                 ext4_handle_sync(handle);
4896
4897         ext4_journal_stop(handle);
4898 out_dio:
4899         ext4_inode_resume_unlocked_dio(inode);
4900 out_mutex:
4901         inode_unlock(inode);
4902         return ret;
4903 }
4904
4905 /*
4906  * preallocate space for a file. This implements ext4's fallocate file
4907  * operation, which gets called from sys_fallocate system call.
4908  * For block-mapped files, posix_fallocate should fall back to the method
4909  * of writing zeroes to the required new blocks (the same behavior which is
4910  * expected for file systems which do not support fallocate() system call).
4911  */
4912 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4913 {
4914         struct inode *inode = file_inode(file);
4915         loff_t new_size = 0;
4916         unsigned int max_blocks;
4917         int ret = 0;
4918         int flags;
4919         ext4_lblk_t lblk;
4920         unsigned int blkbits = inode->i_blkbits;
4921
4922         /*
4923          * Encrypted inodes can't handle collapse range or insert
4924          * range since we would need to re-encrypt blocks with a
4925          * different IV or XTS tweak (which are based on the logical
4926          * block number).
4927          *
4928          * XXX It's not clear why zero range isn't working, but we'll
4929          * leave it disabled for encrypted inodes for now.  This is a
4930          * bug we should fix....
4931          */
4932         if (ext4_encrypted_inode(inode) &&
4933             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4934                      FALLOC_FL_ZERO_RANGE)))
4935                 return -EOPNOTSUPP;
4936
4937         /* Return error if mode is not supported */
4938         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4939                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4940                      FALLOC_FL_INSERT_RANGE))
4941                 return -EOPNOTSUPP;
4942
4943         if (mode & FALLOC_FL_PUNCH_HOLE)
4944                 return ext4_punch_hole(inode, offset, len);
4945
4946         ret = ext4_convert_inline_data(inode);
4947         if (ret)
4948                 return ret;
4949
4950         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4951                 return ext4_collapse_range(inode, offset, len);
4952
4953         if (mode & FALLOC_FL_INSERT_RANGE)
4954                 return ext4_insert_range(inode, offset, len);
4955
4956         if (mode & FALLOC_FL_ZERO_RANGE)
4957                 return ext4_zero_range(file, offset, len, mode);
4958
4959         trace_ext4_fallocate_enter(inode, offset, len, mode);
4960         lblk = offset >> blkbits;
4961         /*
4962          * We can't just convert len to max_blocks because
4963          * If blocksize = 4096 offset = 3072 and len = 2048
4964          */
4965         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4966                 - lblk;
4967
4968         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4969         if (mode & FALLOC_FL_KEEP_SIZE)
4970                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4971
4972         inode_lock(inode);
4973
4974         /*
4975          * We only support preallocation for extent-based files only
4976          */
4977         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4978                 ret = -EOPNOTSUPP;
4979                 goto out;
4980         }
4981
4982         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4983              offset + len > i_size_read(inode)) {
4984                 new_size = offset + len;
4985                 ret = inode_newsize_ok(inode, new_size);
4986                 if (ret)
4987                         goto out;
4988         }
4989
4990         /* Wait all existing dio workers, newcomers will block on i_mutex */
4991         ext4_inode_block_unlocked_dio(inode);
4992         inode_dio_wait(inode);
4993
4994         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4995                                      flags, mode);
4996         ext4_inode_resume_unlocked_dio(inode);
4997         if (ret)
4998                 goto out;
4999
5000         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
5001                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
5002                                                 EXT4_I(inode)->i_sync_tid);
5003         }
5004 out:
5005         inode_unlock(inode);
5006         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
5007         return ret;
5008 }
5009
5010 /*
5011  * This function convert a range of blocks to written extents
5012  * The caller of this function will pass the start offset and the size.
5013  * all unwritten extents within this range will be converted to
5014  * written extents.
5015  *
5016  * This function is called from the direct IO end io call back
5017  * function, to convert the fallocated extents after IO is completed.
5018  * Returns 0 on success.
5019  */
5020 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
5021                                    loff_t offset, ssize_t len)
5022 {
5023         unsigned int max_blocks;
5024         int ret = 0;
5025         int ret2 = 0;
5026         struct ext4_map_blocks map;
5027         unsigned int credits, blkbits = inode->i_blkbits;
5028
5029         map.m_lblk = offset >> blkbits;
5030         /*
5031          * We can't just convert len to max_blocks because
5032          * If blocksize = 4096 offset = 3072 and len = 2048
5033          */
5034         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
5035                       map.m_lblk);
5036         /*
5037          * This is somewhat ugly but the idea is clear: When transaction is
5038          * reserved, everything goes into it. Otherwise we rather start several
5039          * smaller transactions for conversion of each extent separately.
5040          */
5041         if (handle) {
5042                 handle = ext4_journal_start_reserved(handle,
5043                                                      EXT4_HT_EXT_CONVERT);
5044                 if (IS_ERR(handle))
5045                         return PTR_ERR(handle);
5046                 credits = 0;
5047         } else {
5048                 /*
5049                  * credits to insert 1 extent into extent tree
5050                  */
5051                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5052         }
5053         while (ret >= 0 && ret < max_blocks) {
5054                 map.m_lblk += ret;
5055                 map.m_len = (max_blocks -= ret);
5056                 if (credits) {
5057                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5058                                                     credits);
5059                         if (IS_ERR(handle)) {
5060                                 ret = PTR_ERR(handle);
5061                                 break;
5062                         }
5063                 }
5064                 ret = ext4_map_blocks(handle, inode, &map,
5065                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5066                 if (ret <= 0)
5067                         ext4_warning(inode->i_sb,
5068                                      "inode #%lu: block %u: len %u: "
5069                                      "ext4_ext_map_blocks returned %d",
5070                                      inode->i_ino, map.m_lblk,
5071                                      map.m_len, ret);
5072                 ext4_mark_inode_dirty(handle, inode);
5073                 if (credits)
5074                         ret2 = ext4_journal_stop(handle);
5075                 if (ret <= 0 || ret2)
5076                         break;
5077         }
5078         if (!credits)
5079                 ret2 = ext4_journal_stop(handle);
5080         return ret > 0 ? ret2 : ret;
5081 }
5082
5083 /*
5084  * If newes is not existing extent (newes->ec_pblk equals zero) find
5085  * delayed extent at start of newes and update newes accordingly and
5086  * return start of the next delayed extent.
5087  *
5088  * If newes is existing extent (newes->ec_pblk is not equal zero)
5089  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5090  * extent found. Leave newes unmodified.
5091  */
5092 static int ext4_find_delayed_extent(struct inode *inode,
5093                                     struct extent_status *newes)
5094 {
5095         struct extent_status es;
5096         ext4_lblk_t block, next_del;
5097
5098         if (newes->es_pblk == 0) {
5099                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5100                                 newes->es_lblk + newes->es_len - 1, &es);
5101
5102                 /*
5103                  * No extent in extent-tree contains block @newes->es_pblk,
5104                  * then the block may stay in 1)a hole or 2)delayed-extent.
5105                  */
5106                 if (es.es_len == 0)
5107                         /* A hole found. */
5108                         return 0;
5109
5110                 if (es.es_lblk > newes->es_lblk) {
5111                         /* A hole found. */
5112                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5113                                             newes->es_len);
5114                         return 0;
5115                 }
5116
5117                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5118         }
5119
5120         block = newes->es_lblk + newes->es_len;
5121         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5122         if (es.es_len == 0)
5123                 next_del = EXT_MAX_BLOCKS;
5124         else
5125                 next_del = es.es_lblk;
5126
5127         return next_del;
5128 }
5129 /* fiemap flags we can handle specified here */
5130 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5131
5132 static int ext4_xattr_fiemap(struct inode *inode,
5133                                 struct fiemap_extent_info *fieinfo)
5134 {
5135         __u64 physical = 0;
5136         __u64 length;
5137         __u32 flags = FIEMAP_EXTENT_LAST;
5138         int blockbits = inode->i_sb->s_blocksize_bits;
5139         int error = 0;
5140
5141         /* in-inode? */
5142         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5143                 struct ext4_iloc iloc;
5144                 int offset;     /* offset of xattr in inode */
5145
5146                 error = ext4_get_inode_loc(inode, &iloc);
5147                 if (error)
5148                         return error;
5149                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5150                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5151                                 EXT4_I(inode)->i_extra_isize;
5152                 physical += offset;
5153                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5154                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5155                 brelse(iloc.bh);
5156         } else { /* external block */
5157                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5158                 length = inode->i_sb->s_blocksize;
5159         }
5160
5161         if (physical)
5162                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5163                                                 length, flags);
5164         return (error < 0 ? error : 0);
5165 }
5166
5167 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5168                 __u64 start, __u64 len)
5169 {
5170         ext4_lblk_t start_blk;
5171         int error = 0;
5172
5173         if (ext4_has_inline_data(inode)) {
5174                 int has_inline = 1;
5175
5176                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5177                                                 start, len);
5178
5179                 if (has_inline)
5180                         return error;
5181         }
5182
5183         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5184                 error = ext4_ext_precache(inode);
5185                 if (error)
5186                         return error;
5187         }
5188
5189         /* fallback to generic here if not in extents fmt */
5190         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5191                 return generic_block_fiemap(inode, fieinfo, start, len,
5192                         ext4_get_block);
5193
5194         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5195                 return -EBADR;
5196
5197         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5198                 error = ext4_xattr_fiemap(inode, fieinfo);
5199         } else {
5200                 ext4_lblk_t len_blks;
5201                 __u64 last_blk;
5202
5203                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5204                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5205                 if (last_blk >= EXT_MAX_BLOCKS)
5206                         last_blk = EXT_MAX_BLOCKS-1;
5207                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5208
5209                 /*
5210                  * Walk the extent tree gathering extent information
5211                  * and pushing extents back to the user.
5212                  */
5213                 error = ext4_fill_fiemap_extents(inode, start_blk,
5214                                                  len_blks, fieinfo);
5215         }
5216         return error;
5217 }
5218
5219 /*
5220  * ext4_access_path:
5221  * Function to access the path buffer for marking it dirty.
5222  * It also checks if there are sufficient credits left in the journal handle
5223  * to update path.
5224  */
5225 static int
5226 ext4_access_path(handle_t *handle, struct inode *inode,
5227                 struct ext4_ext_path *path)
5228 {
5229         int credits, err;
5230
5231         if (!ext4_handle_valid(handle))
5232                 return 0;
5233
5234         /*
5235          * Check if need to extend journal credits
5236          * 3 for leaf, sb, and inode plus 2 (bmap and group
5237          * descriptor) for each block group; assume two block
5238          * groups
5239          */
5240         if (handle->h_buffer_credits < 7) {
5241                 credits = ext4_writepage_trans_blocks(inode);
5242                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5243                 /* EAGAIN is success */
5244                 if (err && err != -EAGAIN)
5245                         return err;
5246         }
5247
5248         err = ext4_ext_get_access(handle, inode, path);
5249         return err;
5250 }
5251
5252 /*
5253  * ext4_ext_shift_path_extents:
5254  * Shift the extents of a path structure lying between path[depth].p_ext
5255  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5256  * if it is right shift or left shift operation.
5257  */
5258 static int
5259 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5260                             struct inode *inode, handle_t *handle,
5261                             enum SHIFT_DIRECTION SHIFT)
5262 {
5263         int depth, err = 0;
5264         struct ext4_extent *ex_start, *ex_last;
5265         bool update = 0;
5266         depth = path->p_depth;
5267
5268         while (depth >= 0) {
5269                 if (depth == path->p_depth) {
5270                         ex_start = path[depth].p_ext;
5271                         if (!ex_start)
5272                                 return -EFSCORRUPTED;
5273
5274                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5275
5276                         err = ext4_access_path(handle, inode, path + depth);
5277                         if (err)
5278                                 goto out;
5279
5280                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5281                                 update = 1;
5282
5283                         while (ex_start <= ex_last) {
5284                                 if (SHIFT == SHIFT_LEFT) {
5285                                         le32_add_cpu(&ex_start->ee_block,
5286                                                 -shift);
5287                                         /* Try to merge to the left. */
5288                                         if ((ex_start >
5289                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5290                                             &&
5291                                             ext4_ext_try_to_merge_right(inode,
5292                                             path, ex_start - 1))
5293                                                 ex_last--;
5294                                         else
5295                                                 ex_start++;
5296                                 } else {
5297                                         le32_add_cpu(&ex_last->ee_block, shift);
5298                                         ext4_ext_try_to_merge_right(inode, path,
5299                                                 ex_last);
5300                                         ex_last--;
5301                                 }
5302                         }
5303                         err = ext4_ext_dirty(handle, inode, path + depth);
5304                         if (err)
5305                                 goto out;
5306
5307                         if (--depth < 0 || !update)
5308                                 break;
5309                 }
5310
5311                 /* Update index too */
5312                 err = ext4_access_path(handle, inode, path + depth);
5313                 if (err)
5314                         goto out;
5315
5316                 if (SHIFT == SHIFT_LEFT)
5317                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5318                 else
5319                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5320                 err = ext4_ext_dirty(handle, inode, path + depth);
5321                 if (err)
5322                         goto out;
5323
5324                 /* we are done if current index is not a starting index */
5325                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5326                         break;
5327
5328                 depth--;
5329         }
5330
5331 out:
5332         return err;
5333 }
5334
5335 /*
5336  * ext4_ext_shift_extents:
5337  * All the extents which lies in the range from @start to the last allocated
5338  * block for the @inode are shifted either towards left or right (depending
5339  * upon @SHIFT) by @shift blocks.
5340  * On success, 0 is returned, error otherwise.
5341  */
5342 static int
5343 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5344                        ext4_lblk_t start, ext4_lblk_t shift,
5345                        enum SHIFT_DIRECTION SHIFT)
5346 {
5347         struct ext4_ext_path *path;
5348         int ret = 0, depth;
5349         struct ext4_extent *extent;
5350         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5351
5352         /* Let path point to the last extent */
5353         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 0);
5354         if (IS_ERR(path))
5355                 return PTR_ERR(path);
5356
5357         depth = path->p_depth;
5358         extent = path[depth].p_ext;
5359         if (!extent)
5360                 goto out;
5361
5362         stop = le32_to_cpu(extent->ee_block) +
5363                         ext4_ext_get_actual_len(extent);
5364
5365        /*
5366          * In case of left shift, Don't start shifting extents until we make
5367          * sure the hole is big enough to accommodate the shift.
5368         */
5369         if (SHIFT == SHIFT_LEFT) {
5370                 path = ext4_find_extent(inode, start - 1, &path, 0);
5371                 if (IS_ERR(path))
5372                         return PTR_ERR(path);
5373                 depth = path->p_depth;
5374                 extent =  path[depth].p_ext;
5375                 if (extent) {
5376                         ex_start = le32_to_cpu(extent->ee_block);
5377                         ex_end = le32_to_cpu(extent->ee_block) +
5378                                 ext4_ext_get_actual_len(extent);
5379                 } else {
5380                         ex_start = 0;
5381                         ex_end = 0;
5382                 }
5383
5384                 if ((start == ex_start && shift > ex_start) ||
5385                     (shift > start - ex_end)) {
5386                         ext4_ext_drop_refs(path);
5387                         kfree(path);
5388                         return -EINVAL;
5389                 }
5390         }
5391
5392         /*
5393          * In case of left shift, iterator points to start and it is increased
5394          * till we reach stop. In case of right shift, iterator points to stop
5395          * and it is decreased till we reach start.
5396          */
5397         if (SHIFT == SHIFT_LEFT)
5398                 iterator = &start;
5399         else
5400                 iterator = &stop;
5401
5402         /* Its safe to start updating extents */
5403         while (start < stop) {
5404                 path = ext4_find_extent(inode, *iterator, &path, 0);
5405                 if (IS_ERR(path))
5406                         return PTR_ERR(path);
5407                 depth = path->p_depth;
5408                 extent = path[depth].p_ext;
5409                 if (!extent) {
5410                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5411                                          (unsigned long) *iterator);
5412                         return -EFSCORRUPTED;
5413                 }
5414                 if (SHIFT == SHIFT_LEFT && *iterator >
5415                     le32_to_cpu(extent->ee_block)) {
5416                         /* Hole, move to the next extent */
5417                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5418                                 path[depth].p_ext++;
5419                         } else {
5420                                 *iterator = ext4_ext_next_allocated_block(path);
5421                                 continue;
5422                         }
5423                 }
5424
5425                 if (SHIFT == SHIFT_LEFT) {
5426                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5427                         *iterator = le32_to_cpu(extent->ee_block) +
5428                                         ext4_ext_get_actual_len(extent);
5429                 } else {
5430                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5431                         *iterator =  le32_to_cpu(extent->ee_block) > 0 ?
5432                                 le32_to_cpu(extent->ee_block) - 1 : 0;
5433                         /* Update path extent in case we need to stop */
5434                         while (le32_to_cpu(extent->ee_block) < start)
5435                                 extent++;
5436                         path[depth].p_ext = extent;
5437                 }
5438                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5439                                 handle, SHIFT);
5440                 if (ret)
5441                         break;
5442         }
5443 out:
5444         ext4_ext_drop_refs(path);
5445         kfree(path);
5446         return ret;
5447 }
5448
5449 /*
5450  * ext4_collapse_range:
5451  * This implements the fallocate's collapse range functionality for ext4
5452  * Returns: 0 and non-zero on error.
5453  */
5454 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5455 {
5456         struct super_block *sb = inode->i_sb;
5457         ext4_lblk_t punch_start, punch_stop;
5458         handle_t *handle;
5459         unsigned int credits;
5460         loff_t new_size, ioffset;
5461         int ret;
5462
5463         /*
5464          * We need to test this early because xfstests assumes that a
5465          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5466          * system does not support collapse range.
5467          */
5468         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5469                 return -EOPNOTSUPP;
5470
5471         /* Collapse range works only on fs block size aligned offsets. */
5472         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5473             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5474                 return -EINVAL;
5475
5476         if (!S_ISREG(inode->i_mode))
5477                 return -EINVAL;
5478
5479         trace_ext4_collapse_range(inode, offset, len);
5480
5481         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5482         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5483
5484         /* Call ext4_force_commit to flush all data in case of data=journal. */
5485         if (ext4_should_journal_data(inode)) {
5486                 ret = ext4_force_commit(inode->i_sb);
5487                 if (ret)
5488                         return ret;
5489         }
5490
5491         inode_lock(inode);
5492         /*
5493          * There is no need to overlap collapse range with EOF, in which case
5494          * it is effectively a truncate operation
5495          */
5496         if (offset + len >= i_size_read(inode)) {
5497                 ret = -EINVAL;
5498                 goto out_mutex;
5499         }
5500
5501         /* Currently just for extent based files */
5502         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5503                 ret = -EOPNOTSUPP;
5504                 goto out_mutex;
5505         }
5506
5507         /* Wait for existing dio to complete */
5508         ext4_inode_block_unlocked_dio(inode);
5509         inode_dio_wait(inode);
5510
5511         /*
5512          * Prevent page faults from reinstantiating pages we have released from
5513          * page cache.
5514          */
5515         down_write(&EXT4_I(inode)->i_mmap_sem);
5516         /*
5517          * Need to round down offset to be aligned with page size boundary
5518          * for page size > block size.
5519          */
5520         ioffset = round_down(offset, PAGE_SIZE);
5521         /*
5522          * Write tail of the last page before removed range since it will get
5523          * removed from the page cache below.
5524          */
5525         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5526         if (ret)
5527                 goto out_mmap;
5528         /*
5529          * Write data that will be shifted to preserve them when discarding
5530          * page cache below. We are also protected from pages becoming dirty
5531          * by i_mmap_sem.
5532          */
5533         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5534                                            LLONG_MAX);
5535         if (ret)
5536                 goto out_mmap;
5537         truncate_pagecache(inode, ioffset);
5538
5539         credits = ext4_writepage_trans_blocks(inode);
5540         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5541         if (IS_ERR(handle)) {
5542                 ret = PTR_ERR(handle);
5543                 goto out_mmap;
5544         }
5545
5546         down_write(&EXT4_I(inode)->i_data_sem);
5547         ext4_discard_preallocations(inode);
5548
5549         ret = ext4_es_remove_extent(inode, punch_start,
5550                                     EXT_MAX_BLOCKS - punch_start);
5551         if (ret) {
5552                 up_write(&EXT4_I(inode)->i_data_sem);
5553                 goto out_stop;
5554         }
5555
5556         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5557         if (ret) {
5558                 up_write(&EXT4_I(inode)->i_data_sem);
5559                 goto out_stop;
5560         }
5561         ext4_discard_preallocations(inode);
5562
5563         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5564                                      punch_stop - punch_start, SHIFT_LEFT);
5565         if (ret) {
5566                 up_write(&EXT4_I(inode)->i_data_sem);
5567                 goto out_stop;
5568         }
5569
5570         new_size = i_size_read(inode) - len;
5571         i_size_write(inode, new_size);
5572         EXT4_I(inode)->i_disksize = new_size;
5573
5574         up_write(&EXT4_I(inode)->i_data_sem);
5575         if (IS_SYNC(inode))
5576                 ext4_handle_sync(handle);
5577         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5578         ext4_mark_inode_dirty(handle, inode);
5579
5580 out_stop:
5581         ext4_journal_stop(handle);
5582 out_mmap:
5583         up_write(&EXT4_I(inode)->i_mmap_sem);
5584         ext4_inode_resume_unlocked_dio(inode);
5585 out_mutex:
5586         inode_unlock(inode);
5587         return ret;
5588 }
5589
5590 /*
5591  * ext4_insert_range:
5592  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5593  * The data blocks starting from @offset to the EOF are shifted by @len
5594  * towards right to create a hole in the @inode. Inode size is increased
5595  * by len bytes.
5596  * Returns 0 on success, error otherwise.
5597  */
5598 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5599 {
5600         struct super_block *sb = inode->i_sb;
5601         handle_t *handle;
5602         struct ext4_ext_path *path;
5603         struct ext4_extent *extent;
5604         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5605         unsigned int credits, ee_len;
5606         int ret = 0, depth, split_flag = 0;
5607         loff_t ioffset;
5608
5609         /*
5610          * We need to test this early because xfstests assumes that an
5611          * insert range of (0, 1) will return EOPNOTSUPP if the file
5612          * system does not support insert range.
5613          */
5614         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5615                 return -EOPNOTSUPP;
5616
5617         /* Insert range works only on fs block size aligned offsets. */
5618         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5619                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5620                 return -EINVAL;
5621
5622         if (!S_ISREG(inode->i_mode))
5623                 return -EOPNOTSUPP;
5624
5625         trace_ext4_insert_range(inode, offset, len);
5626
5627         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5628         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5629
5630         /* Call ext4_force_commit to flush all data in case of data=journal */
5631         if (ext4_should_journal_data(inode)) {
5632                 ret = ext4_force_commit(inode->i_sb);
5633                 if (ret)
5634                         return ret;
5635         }
5636
5637         inode_lock(inode);
5638         /* Currently just for extent based files */
5639         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5640                 ret = -EOPNOTSUPP;
5641                 goto out_mutex;
5642         }
5643
5644         /* Check for wrap through zero */
5645         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5646                 ret = -EFBIG;
5647                 goto out_mutex;
5648         }
5649
5650         /* Offset should be less than i_size */
5651         if (offset >= i_size_read(inode)) {
5652                 ret = -EINVAL;
5653                 goto out_mutex;
5654         }
5655
5656         /* Wait for existing dio to complete */
5657         ext4_inode_block_unlocked_dio(inode);
5658         inode_dio_wait(inode);
5659
5660         /*
5661          * Prevent page faults from reinstantiating pages we have released from
5662          * page cache.
5663          */
5664         down_write(&EXT4_I(inode)->i_mmap_sem);
5665         /*
5666          * Need to round down to align start offset to page size boundary
5667          * for page size > block size.
5668          */
5669         ioffset = round_down(offset, PAGE_SIZE);
5670         /* Write out all dirty pages */
5671         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5672                         LLONG_MAX);
5673         if (ret)
5674                 goto out_mmap;
5675         truncate_pagecache(inode, ioffset);
5676
5677         credits = ext4_writepage_trans_blocks(inode);
5678         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5679         if (IS_ERR(handle)) {
5680                 ret = PTR_ERR(handle);
5681                 goto out_mmap;
5682         }
5683
5684         /* Expand file to avoid data loss if there is error while shifting */
5685         inode->i_size += len;
5686         EXT4_I(inode)->i_disksize += len;
5687         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5688         ret = ext4_mark_inode_dirty(handle, inode);
5689         if (ret)
5690                 goto out_stop;
5691
5692         down_write(&EXT4_I(inode)->i_data_sem);
5693         ext4_discard_preallocations(inode);
5694
5695         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5696         if (IS_ERR(path)) {
5697                 up_write(&EXT4_I(inode)->i_data_sem);
5698                 goto out_stop;
5699         }
5700
5701         depth = ext_depth(inode);
5702         extent = path[depth].p_ext;
5703         if (extent) {
5704                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5705                 ee_len = ext4_ext_get_actual_len(extent);
5706
5707                 /*
5708                  * If offset_lblk is not the starting block of extent, split
5709                  * the extent @offset_lblk
5710                  */
5711                 if ((offset_lblk > ee_start_lblk) &&
5712                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5713                         if (ext4_ext_is_unwritten(extent))
5714                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5715                                         EXT4_EXT_MARK_UNWRIT2;
5716                         ret = ext4_split_extent_at(handle, inode, &path,
5717                                         offset_lblk, split_flag,
5718                                         EXT4_EX_NOCACHE |
5719                                         EXT4_GET_BLOCKS_PRE_IO |
5720                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5721                 }
5722
5723                 ext4_ext_drop_refs(path);
5724                 kfree(path);
5725                 if (ret < 0) {
5726                         up_write(&EXT4_I(inode)->i_data_sem);
5727                         goto out_stop;
5728                 }
5729         }
5730
5731         ret = ext4_es_remove_extent(inode, offset_lblk,
5732                         EXT_MAX_BLOCKS - offset_lblk);
5733         if (ret) {
5734                 up_write(&EXT4_I(inode)->i_data_sem);
5735                 goto out_stop;
5736         }
5737
5738         /*
5739          * if offset_lblk lies in a hole which is at start of file, use
5740          * ee_start_lblk to shift extents
5741          */
5742         ret = ext4_ext_shift_extents(inode, handle,
5743                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5744                 len_lblk, SHIFT_RIGHT);
5745
5746         up_write(&EXT4_I(inode)->i_data_sem);
5747         if (IS_SYNC(inode))
5748                 ext4_handle_sync(handle);
5749
5750 out_stop:
5751         ext4_journal_stop(handle);
5752 out_mmap:
5753         up_write(&EXT4_I(inode)->i_mmap_sem);
5754         ext4_inode_resume_unlocked_dio(inode);
5755 out_mutex:
5756         inode_unlock(inode);
5757         return ret;
5758 }
5759
5760 /**
5761  * ext4_swap_extents - Swap extents between two inodes
5762  *
5763  * @inode1:     First inode
5764  * @inode2:     Second inode
5765  * @lblk1:      Start block for first inode
5766  * @lblk2:      Start block for second inode
5767  * @count:      Number of blocks to swap
5768  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5769  * @erp:        Pointer to save error value
5770  *
5771  * This helper routine does exactly what is promise "swap extents". All other
5772  * stuff such as page-cache locking consistency, bh mapping consistency or
5773  * extent's data copying must be performed by caller.
5774  * Locking:
5775  *              i_mutex is held for both inodes
5776  *              i_data_sem is locked for write for both inodes
5777  * Assumptions:
5778  *              All pages from requested range are locked for both inodes
5779  */
5780 int
5781 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5782                      struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5783                   ext4_lblk_t count, int unwritten, int *erp)
5784 {
5785         struct ext4_ext_path *path1 = NULL;
5786         struct ext4_ext_path *path2 = NULL;
5787         int replaced_count = 0;
5788
5789         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5790         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5791         BUG_ON(!inode_is_locked(inode1));
5792         BUG_ON(!inode_is_locked(inode2));
5793
5794         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5795         if (unlikely(*erp))
5796                 return 0;
5797         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5798         if (unlikely(*erp))
5799                 return 0;
5800
5801         while (count) {
5802                 struct ext4_extent *ex1, *ex2, tmp_ex;
5803                 ext4_lblk_t e1_blk, e2_blk;
5804                 int e1_len, e2_len, len;
5805                 int split = 0;
5806
5807                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5808                 if (IS_ERR(path1)) {
5809                         *erp = PTR_ERR(path1);
5810                         path1 = NULL;
5811                 finish:
5812                         count = 0;
5813                         goto repeat;
5814                 }
5815                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5816                 if (IS_ERR(path2)) {
5817                         *erp = PTR_ERR(path2);
5818                         path2 = NULL;
5819                         goto finish;
5820                 }
5821                 ex1 = path1[path1->p_depth].p_ext;
5822                 ex2 = path2[path2->p_depth].p_ext;
5823                 /* Do we have somthing to swap ? */
5824                 if (unlikely(!ex2 || !ex1))
5825                         goto finish;
5826
5827                 e1_blk = le32_to_cpu(ex1->ee_block);
5828                 e2_blk = le32_to_cpu(ex2->ee_block);
5829                 e1_len = ext4_ext_get_actual_len(ex1);
5830                 e2_len = ext4_ext_get_actual_len(ex2);
5831
5832                 /* Hole handling */
5833                 if (!in_range(lblk1, e1_blk, e1_len) ||
5834                     !in_range(lblk2, e2_blk, e2_len)) {
5835                         ext4_lblk_t next1, next2;
5836
5837                         /* if hole after extent, then go to next extent */
5838                         next1 = ext4_ext_next_allocated_block(path1);
5839                         next2 = ext4_ext_next_allocated_block(path2);
5840                         /* If hole before extent, then shift to that extent */
5841                         if (e1_blk > lblk1)
5842                                 next1 = e1_blk;
5843                         if (e2_blk > lblk2)
5844                                 next2 = e1_blk;
5845                         /* Do we have something to swap */
5846                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5847                                 goto finish;
5848                         /* Move to the rightest boundary */
5849                         len = next1 - lblk1;
5850                         if (len < next2 - lblk2)
5851                                 len = next2 - lblk2;
5852                         if (len > count)
5853                                 len = count;
5854                         lblk1 += len;
5855                         lblk2 += len;
5856                         count -= len;
5857                         goto repeat;
5858                 }
5859
5860                 /* Prepare left boundary */
5861                 if (e1_blk < lblk1) {
5862                         split = 1;
5863                         *erp = ext4_force_split_extent_at(handle, inode1,
5864                                                 &path1, lblk1, 0);
5865                         if (unlikely(*erp))
5866                                 goto finish;
5867                 }
5868                 if (e2_blk < lblk2) {
5869                         split = 1;
5870                         *erp = ext4_force_split_extent_at(handle, inode2,
5871                                                 &path2,  lblk2, 0);
5872                         if (unlikely(*erp))
5873                                 goto finish;
5874                 }
5875                 /* ext4_split_extent_at() may result in leaf extent split,
5876                  * path must to be revalidated. */
5877                 if (split)
5878                         goto repeat;
5879
5880                 /* Prepare right boundary */
5881                 len = count;
5882                 if (len > e1_blk + e1_len - lblk1)
5883                         len = e1_blk + e1_len - lblk1;
5884                 if (len > e2_blk + e2_len - lblk2)
5885                         len = e2_blk + e2_len - lblk2;
5886
5887                 if (len != e1_len) {
5888                         split = 1;
5889                         *erp = ext4_force_split_extent_at(handle, inode1,
5890                                                 &path1, lblk1 + len, 0);
5891                         if (unlikely(*erp))
5892                                 goto finish;
5893                 }
5894                 if (len != e2_len) {
5895                         split = 1;
5896                         *erp = ext4_force_split_extent_at(handle, inode2,
5897                                                 &path2, lblk2 + len, 0);
5898                         if (*erp)
5899                                 goto finish;
5900                 }
5901                 /* ext4_split_extent_at() may result in leaf extent split,
5902                  * path must to be revalidated. */
5903                 if (split)
5904                         goto repeat;
5905
5906                 BUG_ON(e2_len != e1_len);
5907                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5908                 if (unlikely(*erp))
5909                         goto finish;
5910                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5911                 if (unlikely(*erp))
5912                         goto finish;
5913
5914                 /* Both extents are fully inside boundaries. Swap it now */
5915                 tmp_ex = *ex1;
5916                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5917                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5918                 ex1->ee_len = cpu_to_le16(e2_len);
5919                 ex2->ee_len = cpu_to_le16(e1_len);
5920                 if (unwritten)
5921                         ext4_ext_mark_unwritten(ex2);
5922                 if (ext4_ext_is_unwritten(&tmp_ex))
5923                         ext4_ext_mark_unwritten(ex1);
5924
5925                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5926                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5927                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5928                                       path2->p_depth);
5929                 if (unlikely(*erp))
5930                         goto finish;
5931                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5932                                       path1->p_depth);
5933                 /*
5934                  * Looks scarry ah..? second inode already points to new blocks,
5935                  * and it was successfully dirtied. But luckily error may happen
5936                  * only due to journal error, so full transaction will be
5937                  * aborted anyway.
5938                  */
5939                 if (unlikely(*erp))
5940                         goto finish;
5941                 lblk1 += len;
5942                 lblk2 += len;
5943                 replaced_count += len;
5944                 count -= len;
5945
5946         repeat:
5947                 ext4_ext_drop_refs(path1);
5948                 kfree(path1);
5949                 ext4_ext_drop_refs(path2);
5950                 kfree(path2);
5951                 path1 = path2 = NULL;
5952         }
5953         return replaced_count;
5954 }