fs/ext4/extents.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   4  * Written by Alex Tomas <alex@clusterfs.com>
   5  *
   6  * Architecture independence:
   7  *   Copyright (c) 2005, Bull S.A.
   8  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   9  */
  10
  11 /*
  12  * Extents support for EXT4
  13  *
  14  * TODO:
  15  *   - ext4*_error() should be used in some situations
  16  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  17  *   - smart tree reduction
  18  */
  19
  20 #include <linux/fs.h>
  21 #include <linux/time.h>
  22 #include <linux/jbd2.h>
  23 #include <linux/highuid.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/quotaops.h>
  26 #include <linux/string.h>
  27 #include <linux/slab.h>
  28 #include <linux/uaccess.h>
  29 #include <linux/fiemap.h>
  30 #include <linux/backing-dev.h>
  31 #include "ext4_jbd2.h"
  32 #include "ext4_extents.h"
  33 #include "xattr.h"
  34
  35 #include <trace/events/ext4.h>
  36
  37 /*
  38  * used by extent splitting.
  39  */
  40 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  41                                         due to ENOSPC */
  42 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  43 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  44
  45 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  46 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  47
  48 static __le32 ext4_extent_block_csum(struct inode *inode,
  49                                      struct ext4_extent_header *eh)
  50 {
  51         struct ext4_inode_info *ei = EXT4_I(inode);
  52         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  53         __u32 csum;
  54
  55         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  56                            EXT4_EXTENT_TAIL_OFFSET(eh));
  57         return cpu_to_le32(csum);
  58 }
  59
  60 static int ext4_extent_block_csum_verify(struct inode *inode,
  61                                          struct ext4_extent_header *eh)
  62 {
  63         struct ext4_extent_tail *et;
  64
  65         if (!ext4_has_metadata_csum(inode->i_sb))
  66                 return 1;
  67
  68         et = find_ext4_extent_tail(eh);
  69         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  70                 return 0;
  71         return 1;
  72 }
  73
  74 static void ext4_extent_block_csum_set(struct inode *inode,
  75                                        struct ext4_extent_header *eh)
  76 {
  77         struct ext4_extent_tail *et;
  78
  79         if (!ext4_has_metadata_csum(inode->i_sb))
  80                 return;
  81
  82         et = find_ext4_extent_tail(eh);
  83         et->et_checksum = ext4_extent_block_csum(inode, eh);
  84 }
  85
  86 static int ext4_split_extent(handle_t *handle,
  87                                 struct inode *inode,
  88                                 struct ext4_ext_path **ppath,
  89                                 struct ext4_map_blocks *map,
  90                                 int split_flag,
  91                                 int flags);
  92
  93 static int ext4_split_extent_at(handle_t *handle,
  94                              struct inode *inode,
  95                              struct ext4_ext_path **ppath,
  96                              ext4_lblk_t split,
  97                              int split_flag,
  98                              int flags);
  99
 100 static int ext4_find_delayed_extent(struct inode *inode,
 101                                     struct extent_status *newes);
 102
 103 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 104                                             struct inode *inode,
 105                                             int needed)
 106 {
 107         int err;
 108
 109         if (!ext4_handle_valid(handle))
 110                 return 0;
 111         if (handle->h_buffer_credits >= needed)
 112                 return 0;
 113         /*
 114          * If we need to extend the journal get a few extra blocks
 115          * while we're at it for efficiency's sake.
 116          */
 117         needed += 3;
 118         err = ext4_journal_extend(handle, needed - handle->h_buffer_credits);
 119         if (err <= 0)
 120                 return err;
 121         err = ext4_truncate_restart_trans(handle, inode, needed);
 122         if (err == 0)
 123                 err = -EAGAIN;
 124
 125         return err;
 126 }
 127
 128 /*
 129  * could return:
 130  *  - EROFS
 131  *  - ENOMEM
 132  */
 133 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 134                                 struct ext4_ext_path *path)
 135 {
 136         if (path->p_bh) {
 137                 /* path points to block */
 138                 BUFFER_TRACE(path->p_bh, "get_write_access");
 139                 return ext4_journal_get_write_access(handle, path->p_bh);
 140         }
 141         /* path points to leaf/index in inode body */
 142         /* we use in-core data, no need to protect them */
 143         return 0;
 144 }
 145
 146 /*
 147  * could return:
 148  *  - EROFS
 149  *  - ENOMEM
 150  *  - EIO
 151  */
 152 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 153                      struct inode *inode, struct ext4_ext_path *path)
 154 {
 155         int err;
 156
 157         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 158         if (path->p_bh) {
 159                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 160                 /* path points to block */
 161                 err = __ext4_handle_dirty_metadata(where, line, handle,
 162                                                    inode, path->p_bh);
 163         } else {
 164                 /* path points to leaf/index in inode body */
 165                 err = ext4_mark_inode_dirty(handle, inode);
 166         }
 167         return err;
 168 }
 169
 170 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 171                               struct ext4_ext_path *path,
 172                               ext4_lblk_t block)
 173 {
 174         if (path) {
 175                 int depth = path->p_depth;
 176                 struct ext4_extent *ex;
 177
 178                 /*
 179                  * Try to predict block placement assuming that we are
 180                  * filling in a file which will eventually be
 181                  * non-sparse --- i.e., in the case of libbfd writing
 182                  * an ELF object sections out-of-order but in a way
 183                  * the eventually results in a contiguous object or
 184                  * executable file, or some database extending a table
 185                  * space file.  However, this is actually somewhat
 186                  * non-ideal if we are writing a sparse file such as
 187                  * qemu or KVM writing a raw image file that is going
 188                  * to stay fairly sparse, since it will end up
 189                  * fragmenting the file system's free space.  Maybe we
 190                  * should have some hueristics or some way to allow
 191                  * userspace to pass a hint to file system,
 192                  * especially if the latter case turns out to be
 193                  * common.
 194                  */
 195                 ex = path[depth].p_ext;
 196                 if (ex) {
 197                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 198                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 199
 200                         if (block > ext_block)
 201                                 return ext_pblk + (block - ext_block);
 202                         else
 203                                 return ext_pblk - (ext_block - block);
 204                 }
 205
 206                 /* it looks like index is empty;
 207                  * try to find starting block from index itself */
 208                 if (path[depth].p_bh)
 209                         return path[depth].p_bh->b_blocknr;
 210         }
 211
 212         /* OK. use inode's group */
 213         return ext4_inode_to_goal_block(inode);
 214 }
 215
 216 /*
 217  * Allocation for a meta data block
 218  */
 219 static ext4_fsblk_t
 220 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 221                         struct ext4_ext_path *path,
 222                         struct ext4_extent *ex, int *err, unsigned int flags)
 223 {
 224         ext4_fsblk_t goal, newblock;
 225
 226         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 227         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 228                                         NULL, err);
 229         return newblock;
 230 }
 231
 232 static inline int ext4_ext_space_block(struct inode *inode, int check)
 233 {
 234         int size;
 235
 236         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 237                         / sizeof(struct ext4_extent);
 238 #ifdef AGGRESSIVE_TEST
 239         if (!check && size > 6)
 240                 size = 6;
 241 #endif
 242         return size;
 243 }
 244
 245 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 246 {
 247         int size;
 248
 249         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 250                         / sizeof(struct ext4_extent_idx);
 251 #ifdef AGGRESSIVE_TEST
 252         if (!check && size > 5)
 253                 size = 5;
 254 #endif
 255         return size;
 256 }
 257
 258 static inline int ext4_ext_space_root(struct inode *inode, int check)
 259 {
 260         int size;
 261
 262         size = sizeof(EXT4_I(inode)->i_data);
 263         size -= sizeof(struct ext4_extent_header);
 264         size /= sizeof(struct ext4_extent);
 265 #ifdef AGGRESSIVE_TEST
 266         if (!check && size > 3)
 267                 size = 3;
 268 #endif
 269         return size;
 270 }
 271
 272 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 273 {
 274         int size;
 275
 276         size = sizeof(EXT4_I(inode)->i_data);
 277         size -= sizeof(struct ext4_extent_header);
 278         size /= sizeof(struct ext4_extent_idx);
 279 #ifdef AGGRESSIVE_TEST
 280         if (!check && size > 4)
 281                 size = 4;
 282 #endif
 283         return size;
 284 }
 285
 286 static inline int
 287 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 288                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 289                            int nofail)
 290 {
 291         struct ext4_ext_path *path = *ppath;
 292         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 293
 294         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 295                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 296                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 297                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 298 }
 299
 300 /*
 301  * Calculate the number of metadata blocks needed
 302  * to allocate @blocks
 303  * Worse case is one block per extent
 304  */
 305 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 306 {
 307         struct ext4_inode_info *ei = EXT4_I(inode);
 308         int idxs;
 309
 310         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 311                 / sizeof(struct ext4_extent_idx));
 312
 313         /*
 314          * If the new delayed allocation block is contiguous with the
 315          * previous da block, it can share index blocks with the
 316          * previous block, so we only need to allocate a new index
 317          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 318          * an additional index block, and at ldxs**3 blocks, yet
 319          * another index blocks.
 320          */
 321         if (ei->i_da_metadata_calc_len &&
 322             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 323                 int num = 0;
 324
 325                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 326                         num++;
 327                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 328                         num++;
 329                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 330                         num++;
 331                         ei->i_da_metadata_calc_len = 0;
 332                 } else
 333                         ei->i_da_metadata_calc_len++;
 334                 ei->i_da_metadata_calc_last_lblock++;
 335                 return num;
 336         }
 337
 338         /*
 339          * In the worst case we need a new set of index blocks at
 340          * every level of the inode's extent tree.
 341          */
 342         ei->i_da_metadata_calc_len = 1;
 343         ei->i_da_metadata_calc_last_lblock = lblock;
 344         return ext_depth(inode) + 1;
 345 }
 346
 347 static int
 348 ext4_ext_max_entries(struct inode *inode, int depth)
 349 {
 350         int max;
 351
 352         if (depth == ext_depth(inode)) {
 353                 if (depth == 0)
 354                         max = ext4_ext_space_root(inode, 1);
 355                 else
 356                         max = ext4_ext_space_root_idx(inode, 1);
 357         } else {
 358                 if (depth == 0)
 359                         max = ext4_ext_space_block(inode, 1);
 360                 else
 361                         max = ext4_ext_space_block_idx(inode, 1);
 362         }
 363
 364         return max;
 365 }
 366
 367 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 368 {
 369         ext4_fsblk_t block = ext4_ext_pblock(ext);
 370         int len = ext4_ext_get_actual_len(ext);
 371         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 372
 373         /*
 374          * We allow neither:
 375          *  - zero length
 376          *  - overflow/wrap-around
 377          */
 378         if (lblock + len <= lblock)
 379                 return 0;
 380         return ext4_inode_block_valid(inode, block, len);
 381 }
 382
 383 static int ext4_valid_extent_idx(struct inode *inode,
 384                                 struct ext4_extent_idx *ext_idx)
 385 {
 386         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 387
 388         return ext4_inode_block_valid(inode, block, 1);
 389 }
 390
 391 static int ext4_valid_extent_entries(struct inode *inode,
 392                                 struct ext4_extent_header *eh,
 393                                 ext4_fsblk_t *pblk, int depth)
 394 {
 395         unsigned short entries;
 396         ext4_lblk_t lblock = 0;
 397         ext4_lblk_t prev = 0;
 398
 399         if (eh->eh_entries == 0)
 400                 return 1;
 401
 402         entries = le16_to_cpu(eh->eh_entries);
 403
 404         if (depth == 0) {
 405                 /* leaf entries */
 406                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 407                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 408                 ext4_fsblk_t pblock = 0;
 409                 while (entries) {
 410                         if (!ext4_valid_extent(inode, ext))
 411                                 return 0;
 412
 413                         /* Check for overlapping extents */
 414                         lblock = le32_to_cpu(ext->ee_block);
 415                         if ((lblock <= prev) && prev) {
 416                                 pblock = ext4_ext_pblock(ext);
 417                                 es->s_last_error_block = cpu_to_le64(pblock);
 418                                 return 0;
 419                         }
 420                         prev = lblock + ext4_ext_get_actual_len(ext) - 1;
 421                         ext++;
 422                         entries--;
 423                 }
 424         } else {
 425                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 426                 while (entries) {
 427                         if (!ext4_valid_extent_idx(inode, ext_idx))
 428                                 return 0;
 429
 430                         /* Check for overlapping index extents */
 431                         lblock = le32_to_cpu(ext_idx->ei_block);
 432                         if ((lblock <= prev) && prev) {
 433                                 *pblk = ext4_idx_pblock(ext_idx);
 434                                 return 0;
 435                         }
 436                         ext_idx++;
 437                         entries--;
 438                         prev = lblock;
 439                 }
 440         }
 441         return 1;
 442 }
 443
 444 static int __ext4_ext_check(const char *function, unsigned int line,
 445                             struct inode *inode, struct ext4_extent_header *eh,
 446                             int depth, ext4_fsblk_t pblk)
 447 {
 448         const char *error_msg;
 449         int max = 0, err = -EFSCORRUPTED;
 450
 451         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 452                 error_msg = "invalid magic";
 453                 goto corrupted;
 454         }
 455         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 456                 error_msg = "unexpected eh_depth";
 457                 goto corrupted;
 458         }
 459         if (unlikely(eh->eh_max == 0)) {
 460                 error_msg = "invalid eh_max";
 461                 goto corrupted;
 462         }
 463         max = ext4_ext_max_entries(inode, depth);
 464         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 465                 error_msg = "too large eh_max";
 466                 goto corrupted;
 467         }
 468         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 469                 error_msg = "invalid eh_entries";
 470                 goto corrupted;
 471         }
 472         if (!ext4_valid_extent_entries(inode, eh, &pblk, depth)) {
 473                 error_msg = "invalid extent entries";
 474                 goto corrupted;
 475         }
 476         if (unlikely(depth > 32)) {
 477                 error_msg = "too large eh_depth";
 478                 goto corrupted;
 479         }
 480         /* Verify checksum on non-root extent tree nodes */
 481         if (ext_depth(inode) != depth &&
 482             !ext4_extent_block_csum_verify(inode, eh)) {
 483                 error_msg = "extent tree corrupted";
 484                 err = -EFSBADCRC;
 485                 goto corrupted;
 486         }
 487         return 0;
 488
 489 corrupted:
 490         ext4_error_inode(inode, function, line, 0,
 491                          "pblk %llu bad header/extent: %s - magic %x, "
 492                          "entries %u, max %u(%u), depth %u(%u)",
 493                          (unsigned long long) pblk, error_msg,
 494                          le16_to_cpu(eh->eh_magic),
 495                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 496                          max, le16_to_cpu(eh->eh_depth), depth);
 497         return err;
 498 }
 499
 500 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 501         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 502
 503 int ext4_ext_check_inode(struct inode *inode)
 504 {
 505         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 506 }
 507
 508 static void ext4_cache_extents(struct inode *inode,
 509                                struct ext4_extent_header *eh)
 510 {
 511         struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 512         ext4_lblk_t prev = 0;
 513         int i;
 514
 515         for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 516                 unsigned int status = EXTENT_STATUS_WRITTEN;
 517                 ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 518                 int len = ext4_ext_get_actual_len(ex);
 519
 520                 if (prev && (prev != lblk))
 521                         ext4_es_cache_extent(inode, prev, lblk - prev, ~0,
 522                                              EXTENT_STATUS_HOLE);
 523
 524                 if (ext4_ext_is_unwritten(ex))
 525                         status = EXTENT_STATUS_UNWRITTEN;
 526                 ext4_es_cache_extent(inode, lblk, len,
 527                                      ext4_ext_pblock(ex), status);
 528                 prev = lblk + len;
 529         }
 530 }
 531
 532 static struct buffer_head *
 533 __read_extent_tree_block(const char *function, unsigned int line,
 534                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 535                          int flags)
 536 {
 537         struct buffer_head              *bh;
 538         int                             err;
 539
 540         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
 541         if (unlikely(!bh))
 542                 return ERR_PTR(-ENOMEM);
 543
 544         if (!bh_uptodate_or_lock(bh)) {
 545                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 546                 err = bh_submit_read(bh);
 547                 if (err < 0)
 548                         goto errout;
 549         }
 550         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 551                 return bh;
 552         err = __ext4_ext_check(function, line, inode,
 553                                ext_block_hdr(bh), depth, pblk);
 554         if (err)
 555                 goto errout;
 556         set_buffer_verified(bh);
 557         /*
 558          * If this is a leaf block, cache all of its entries
 559          */
 560         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 561                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 562                 ext4_cache_extents(inode, eh);
 563         }
 564         return bh;
 565 errout:
 566         put_bh(bh);
 567         return ERR_PTR(err);
 568
 569 }
 570
 571 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 572         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 573                                  (depth), (flags))
 574
 575 /*
 576  * This function is called to cache a file's extent information in the
 577  * extent status tree
 578  */
 579 int ext4_ext_precache(struct inode *inode)
 580 {
 581         struct ext4_inode_info *ei = EXT4_I(inode);
 582         struct ext4_ext_path *path = NULL;
 583         struct buffer_head *bh;
 584         int i = 0, depth, ret = 0;
 585
 586         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 587                 return 0;       /* not an extent-mapped inode */
 588
 589         down_read(&ei->i_data_sem);
 590         depth = ext_depth(inode);
 591
 592         path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
 593                        GFP_NOFS);
 594         if (path == NULL) {
 595                 up_read(&ei->i_data_sem);
 596                 return -ENOMEM;
 597         }
 598
 599         /* Don't cache anything if there are no external extent blocks */
 600         if (depth == 0)
 601                 goto out;
 602         path[0].p_hdr = ext_inode_hdr(inode);
 603         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 604         if (ret)
 605                 goto out;
 606         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 607         while (i >= 0) {
 608                 /*
 609                  * If this is a leaf block or we've reached the end of
 610                  * the index block, go up
 611                  */
 612                 if ((i == depth) ||
 613                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 614                         brelse(path[i].p_bh);
 615                         path[i].p_bh = NULL;
 616                         i--;
 617                         continue;
 618                 }
 619                 bh = read_extent_tree_block(inode,
 620                                             ext4_idx_pblock(path[i].p_idx++),
 621                                             depth - i - 1,
 622                                             EXT4_EX_FORCE_CACHE);
 623                 if (IS_ERR(bh)) {
 624                         ret = PTR_ERR(bh);
 625                         break;
 626                 }
 627                 i++;
 628                 path[i].p_bh = bh;
 629                 path[i].p_hdr = ext_block_hdr(bh);
 630                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 631         }
 632         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 633 out:
 634         up_read(&ei->i_data_sem);
 635         ext4_ext_drop_refs(path);
 636         kfree(path);
 637         return ret;
 638 }
 639
 640 #ifdef EXT_DEBUG
 641 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 642 {
 643         int k, l = path->p_depth;
 644
 645         ext_debug("path:");
 646         for (k = 0; k <= l; k++, path++) {
 647                 if (path->p_idx) {
 648                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 649                             ext4_idx_pblock(path->p_idx));
 650                 } else if (path->p_ext) {
 651                         ext_debug("  %d:[%d]%d:%llu ",
 652                                   le32_to_cpu(path->p_ext->ee_block),
 653                                   ext4_ext_is_unwritten(path->p_ext),
 654                                   ext4_ext_get_actual_len(path->p_ext),
 655                                   ext4_ext_pblock(path->p_ext));
 656                 } else
 657                         ext_debug("  []");
 658         }
 659         ext_debug("\n");
 660 }
 661
 662 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 663 {
 664         int depth = ext_depth(inode);
 665         struct ext4_extent_header *eh;
 666         struct ext4_extent *ex;
 667         int i;
 668
 669         if (!path)
 670                 return;
 671
 672         eh = path[depth].p_hdr;
 673         ex = EXT_FIRST_EXTENT(eh);
 674
 675         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 676
 677         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 678                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 679                           ext4_ext_is_unwritten(ex),
 680                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 681         }
 682         ext_debug("\n");
 683 }
 684
 685 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 686                         ext4_fsblk_t newblock, int level)
 687 {
 688         int depth = ext_depth(inode);
 689         struct ext4_extent *ex;
 690
 691         if (depth != level) {
 692                 struct ext4_extent_idx *idx;
 693                 idx = path[level].p_idx;
 694                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 695                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 696                                         le32_to_cpu(idx->ei_block),
 697                                         ext4_idx_pblock(idx),
 698                                         newblock);
 699                         idx++;
 700                 }
 701
 702                 return;
 703         }
 704
 705         ex = path[depth].p_ext;
 706         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 707                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 708                                 le32_to_cpu(ex->ee_block),
 709                                 ext4_ext_pblock(ex),
 710                                 ext4_ext_is_unwritten(ex),
 711                                 ext4_ext_get_actual_len(ex),
 712                                 newblock);
 713                 ex++;
 714         }
 715 }
 716
 717 #else
 718 #define ext4_ext_show_path(inode, path)
 719 #define ext4_ext_show_leaf(inode, path)
 720 #define ext4_ext_show_move(inode, path, newblock, level)
 721 #endif
 722
 723 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 724 {
 725         int depth, i;
 726
 727         if (!path)
 728                 return;
 729         depth = path->p_depth;
 730         for (i = 0; i <= depth; i++, path++)
 731                 if (path->p_bh) {
 732                         brelse(path->p_bh);
 733                         path->p_bh = NULL;
 734                 }
 735 }
 736
 737 /*
 738  * ext4_ext_binsearch_idx:
 739  * binary search for the closest index of the given block
 740  * the header must be checked before calling this
 741  */
 742 static void
 743 ext4_ext_binsearch_idx(struct inode *inode,
 744                         struct ext4_ext_path *path, ext4_lblk_t block)
 745 {
 746         struct ext4_extent_header *eh = path->p_hdr;
 747         struct ext4_extent_idx *r, *l, *m;
 748
 749
 750         ext_debug("binsearch for %u(idx):  ", block);
 751
 752         l = EXT_FIRST_INDEX(eh) + 1;
 753         r = EXT_LAST_INDEX(eh);
 754         while (l <= r) {
 755                 m = l + (r - l) / 2;
 756                 if (block < le32_to_cpu(m->ei_block))
 757                         r = m - 1;
 758                 else
 759                         l = m + 1;
 760                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 761                                 m, le32_to_cpu(m->ei_block),
 762                                 r, le32_to_cpu(r->ei_block));
 763         }
 764
 765         path->p_idx = l - 1;
 766         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 767                   ext4_idx_pblock(path->p_idx));
 768
 769 #ifdef CHECK_BINSEARCH
 770         {
 771                 struct ext4_extent_idx *chix, *ix;
 772                 int k;
 773
 774                 chix = ix = EXT_FIRST_INDEX(eh);
 775                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 776                   if (k != 0 &&
 777                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 778                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 779                                        "first=0x%p\n", k,
 780                                        ix, EXT_FIRST_INDEX(eh));
 781                                 printk(KERN_DEBUG "%u <= %u\n",
 782                                        le32_to_cpu(ix->ei_block),
 783                                        le32_to_cpu(ix[-1].ei_block));
 784                         }
 785                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 786                                            <= le32_to_cpu(ix[-1].ei_block));
 787                         if (block < le32_to_cpu(ix->ei_block))
 788                                 break;
 789                         chix = ix;
 790                 }
 791                 BUG_ON(chix != path->p_idx);
 792         }
 793 #endif
 794
 795 }
 796
 797 /*
 798  * ext4_ext_binsearch:
 799  * binary search for closest extent of the given block
 800  * the header must be checked before calling this
 801  */
 802 static void
 803 ext4_ext_binsearch(struct inode *inode,
 804                 struct ext4_ext_path *path, ext4_lblk_t block)
 805 {
 806         struct ext4_extent_header *eh = path->p_hdr;
 807         struct ext4_extent *r, *l, *m;
 808
 809         if (eh->eh_entries == 0) {
 810                 /*
 811                  * this leaf is empty:
 812                  * we get such a leaf in split/add case
 813                  */
 814                 return;
 815         }
 816
 817         ext_debug("binsearch for %u:  ", block);
 818
 819         l = EXT_FIRST_EXTENT(eh) + 1;
 820         r = EXT_LAST_EXTENT(eh);
 821
 822         while (l <= r) {
 823                 m = l + (r - l) / 2;
 824                 if (block < le32_to_cpu(m->ee_block))
 825                         r = m - 1;
 826                 else
 827                         l = m + 1;
 828                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 829                                 m, le32_to_cpu(m->ee_block),
 830                                 r, le32_to_cpu(r->ee_block));
 831         }
 832
 833         path->p_ext = l - 1;
 834         ext_debug("  -> %d:%llu:[%d]%d ",
 835                         le32_to_cpu(path->p_ext->ee_block),
 836                         ext4_ext_pblock(path->p_ext),
 837                         ext4_ext_is_unwritten(path->p_ext),
 838                         ext4_ext_get_actual_len(path->p_ext));
 839
 840 #ifdef CHECK_BINSEARCH
 841         {
 842                 struct ext4_extent *chex, *ex;
 843                 int k;
 844
 845                 chex = ex = EXT_FIRST_EXTENT(eh);
 846                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 847                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 848                                           <= le32_to_cpu(ex[-1].ee_block));
 849                         if (block < le32_to_cpu(ex->ee_block))
 850                                 break;
 851                         chex = ex;
 852                 }
 853                 BUG_ON(chex != path->p_ext);
 854         }
 855 #endif
 856
 857 }
 858
 859 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 860 {
 861         struct ext4_extent_header *eh;
 862
 863         eh = ext_inode_hdr(inode);
 864         eh->eh_depth = 0;
 865         eh->eh_entries = 0;
 866         eh->eh_magic = EXT4_EXT_MAGIC;
 867         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 868         eh->eh_generation = 0;
 869         ext4_mark_inode_dirty(handle, inode);
 870         return 0;
 871 }
 872
 873 struct ext4_ext_path *
 874 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 875                  struct ext4_ext_path **orig_path, int flags)
 876 {
 877         struct ext4_extent_header *eh;
 878         struct buffer_head *bh;
 879         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 880         short int depth, i, ppos = 0;
 881         int ret;
 882
 883         eh = ext_inode_hdr(inode);
 884         depth = ext_depth(inode);
 885         if (depth < 0 || depth > EXT4_MAX_EXTENT_DEPTH) {
 886                 EXT4_ERROR_INODE(inode, "inode has invalid extent depth: %d",
 887                                  depth);
 888                 ret = -EFSCORRUPTED;
 889                 goto err;
 890         }
 891
 892         if (path) {
 893                 ext4_ext_drop_refs(path);
 894                 if (depth > path[0].p_maxdepth) {
 895                         kfree(path);
 896                         *orig_path = path = NULL;
 897                 }
 898         }
 899         if (!path) {
 900                 /* account possible depth increase */
 901                 path = kcalloc(depth + 2, sizeof(struct ext4_ext_path),
 902                                 GFP_NOFS);
 903                 if (unlikely(!path))
 904                         return ERR_PTR(-ENOMEM);
 905                 path[0].p_maxdepth = depth + 1;
 906         }
 907         path[0].p_hdr = eh;
 908         path[0].p_bh = NULL;
 909
 910         i = depth;
 911         if (!(flags & EXT4_EX_NOCACHE) && depth == 0)
 912                 ext4_cache_extents(inode, eh);
 913         /* walk through the tree */
 914         while (i) {
 915                 ext_debug("depth %d: num %d, max %d\n",
 916                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 917
 918                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 919                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 920                 path[ppos].p_depth = i;
 921                 path[ppos].p_ext = NULL;
 922
 923                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 924                                             flags);
 925                 if (IS_ERR(bh)) {
 926                         ret = PTR_ERR(bh);
 927                         goto err;
 928                 }
 929
 930                 eh = ext_block_hdr(bh);
 931                 ppos++;
 932                 path[ppos].p_bh = bh;
 933                 path[ppos].p_hdr = eh;
 934         }
 935
 936         path[ppos].p_depth = i;
 937         path[ppos].p_ext = NULL;
 938         path[ppos].p_idx = NULL;
 939
 940         /* find extent */
 941         ext4_ext_binsearch(inode, path + ppos, block);
 942         /* if not an empty leaf */
 943         if (path[ppos].p_ext)
 944                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 945
 946         ext4_ext_show_path(inode, path);
 947
 948         return path;
 949
 950 err:
 951         ext4_ext_drop_refs(path);
 952         kfree(path);
 953         if (orig_path)
 954                 *orig_path = NULL;
 955         return ERR_PTR(ret);
 956 }
 957
 958 /*
 959  * ext4_ext_insert_index:
 960  * insert new index [@logical;@ptr] into the block at @curp;
 961  * check where to insert: before @curp or after @curp
 962  */
 963 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 964                                  struct ext4_ext_path *curp,
 965                                  int logical, ext4_fsblk_t ptr)
 966 {
 967         struct ext4_extent_idx *ix;
 968         int len, err;
 969
 970         err = ext4_ext_get_access(handle, inode, curp);
 971         if (err)
 972                 return err;
 973
 974         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 975                 EXT4_ERROR_INODE(inode,
 976                                  "logical %d == ei_block %d!",
 977                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 978                 return -EFSCORRUPTED;
 979         }
 980
 981         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 982                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 983                 EXT4_ERROR_INODE(inode,
 984                                  "eh_entries %d >= eh_max %d!",
 985                                  le16_to_cpu(curp->p_hdr->eh_entries),
 986                                  le16_to_cpu(curp->p_hdr->eh_max));
 987                 return -EFSCORRUPTED;
 988         }
 989
 990         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 991                 /* insert after */
 992                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 993                 ix = curp->p_idx + 1;
 994         } else {
 995                 /* insert before */
 996                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 997                 ix = curp->p_idx;
 998         }
 999
1000         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
1001         BUG_ON(len < 0);
1002         if (len > 0) {
1003                 ext_debug("insert new index %d: "
1004                                 "move %d indices from 0x%p to 0x%p\n",
1005                                 logical, len, ix, ix + 1);
1006                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
1007         }
1008
1009         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
1010                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
1011                 return -EFSCORRUPTED;
1012         }
1013
1014         ix->ei_block = cpu_to_le32(logical);
1015         ext4_idx_store_pblock(ix, ptr);
1016         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1017
1018         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1019                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1020                 return -EFSCORRUPTED;
1021         }
1022
1023         err = ext4_ext_dirty(handle, inode, curp);
1024         ext4_std_error(inode->i_sb, err);
1025
1026         return err;
1027 }
1028
1029 /*
1030  * ext4_ext_split:
1031  * inserts new subtree into the path, using free index entry
1032  * at depth @at:
1033  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1034  * - makes decision where to split
1035  * - moves remaining extents and index entries (right to the split point)
1036  *   into the newly allocated blocks
1037  * - initializes subtree
1038  */
1039 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1040                           unsigned int flags,
1041                           struct ext4_ext_path *path,
1042                           struct ext4_extent *newext, int at)
1043 {
1044         struct buffer_head *bh = NULL;
1045         int depth = ext_depth(inode);
1046         struct ext4_extent_header *neh;
1047         struct ext4_extent_idx *fidx;
1048         int i = at, k, m, a;
1049         ext4_fsblk_t newblock, oldblock;
1050         __le32 border;
1051         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1052         int err = 0;
1053         size_t ext_size = 0;
1054
1055         /* make decision: where to split? */
1056         /* FIXME: now decision is simplest: at current extent */
1057
1058         /* if current leaf will be split, then we should use
1059          * border from split point */
1060         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1061                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1062                 return -EFSCORRUPTED;
1063         }
1064         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1065                 border = path[depth].p_ext[1].ee_block;
1066                 ext_debug("leaf will be split."
1067                                 " next leaf starts at %d\n",
1068                                   le32_to_cpu(border));
1069         } else {
1070                 border = newext->ee_block;
1071                 ext_debug("leaf will be added."
1072                                 " next leaf starts at %d\n",
1073                                 le32_to_cpu(border));
1074         }
1075
1076         /*
1077          * If error occurs, then we break processing
1078          * and mark filesystem read-only. index won't
1079          * be inserted and tree will be in consistent
1080          * state. Next mount will repair buffers too.
1081          */
1082
1083         /*
1084          * Get array to track all allocated blocks.
1085          * We need this to handle errors and free blocks
1086          * upon them.
1087          */
1088         ablocks = kcalloc(depth, sizeof(ext4_fsblk_t), GFP_NOFS);
1089         if (!ablocks)
1090                 return -ENOMEM;
1091
1092         /* allocate all needed blocks */
1093         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1094         for (a = 0; a < depth - at; a++) {
1095                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1096                                                    newext, &err, flags);
1097                 if (newblock == 0)
1098                         goto cleanup;
1099                 ablocks[a] = newblock;
1100         }
1101
1102         /* initialize new leaf */
1103         newblock = ablocks[--a];
1104         if (unlikely(newblock == 0)) {
1105                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1106                 err = -EFSCORRUPTED;
1107                 goto cleanup;
1108         }
1109         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1110         if (unlikely(!bh)) {
1111                 err = -ENOMEM;
1112                 goto cleanup;
1113         }
1114         lock_buffer(bh);
1115
1116         err = ext4_journal_get_create_access(handle, bh);
1117         if (err)
1118                 goto cleanup;
1119
1120         neh = ext_block_hdr(bh);
1121         neh->eh_entries = 0;
1122         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1123         neh->eh_magic = EXT4_EXT_MAGIC;
1124         neh->eh_depth = 0;
1125         neh->eh_generation = 0;
1126
1127         /* move remainder of path[depth] to the new leaf */
1128         if (unlikely(path[depth].p_hdr->eh_entries !=
1129                      path[depth].p_hdr->eh_max)) {
1130                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1131                                  path[depth].p_hdr->eh_entries,
1132                                  path[depth].p_hdr->eh_max);
1133                 err = -EFSCORRUPTED;
1134                 goto cleanup;
1135         }
1136         /* start copy from next extent */
1137         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1138         ext4_ext_show_move(inode, path, newblock, depth);
1139         if (m) {
1140                 struct ext4_extent *ex;
1141                 ex = EXT_FIRST_EXTENT(neh);
1142                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1143                 le16_add_cpu(&neh->eh_entries, m);
1144         }
1145
1146         /* zero out unused area in the extent block */
1147         ext_size = sizeof(struct ext4_extent_header) +
1148                 sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries);
1149         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1150         ext4_extent_block_csum_set(inode, neh);
1151         set_buffer_uptodate(bh);
1152         unlock_buffer(bh);
1153
1154         err = ext4_handle_dirty_metadata(handle, inode, bh);
1155         if (err)
1156                 goto cleanup;
1157         brelse(bh);
1158         bh = NULL;
1159
1160         /* correct old leaf */
1161         if (m) {
1162                 err = ext4_ext_get_access(handle, inode, path + depth);
1163                 if (err)
1164                         goto cleanup;
1165                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1166                 err = ext4_ext_dirty(handle, inode, path + depth);
1167                 if (err)
1168                         goto cleanup;
1169
1170         }
1171
1172         /* create intermediate indexes */
1173         k = depth - at - 1;
1174         if (unlikely(k < 0)) {
1175                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1176                 err = -EFSCORRUPTED;
1177                 goto cleanup;
1178         }
1179         if (k)
1180                 ext_debug("create %d intermediate indices\n", k);
1181         /* insert new index into current index block */
1182         /* current depth stored in i var */
1183         i = depth - 1;
1184         while (k--) {
1185                 oldblock = newblock;
1186                 newblock = ablocks[--a];
1187                 bh = sb_getblk(inode->i_sb, newblock);
1188                 if (unlikely(!bh)) {
1189                         err = -ENOMEM;
1190                         goto cleanup;
1191                 }
1192                 lock_buffer(bh);
1193
1194                 err = ext4_journal_get_create_access(handle, bh);
1195                 if (err)
1196                         goto cleanup;
1197
1198                 neh = ext_block_hdr(bh);
1199                 neh->eh_entries = cpu_to_le16(1);
1200                 neh->eh_magic = EXT4_EXT_MAGIC;
1201                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1202                 neh->eh_depth = cpu_to_le16(depth - i);
1203                 neh->eh_generation = 0;
1204                 fidx = EXT_FIRST_INDEX(neh);
1205                 fidx->ei_block = border;
1206                 ext4_idx_store_pblock(fidx, oldblock);
1207
1208                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1209                                 i, newblock, le32_to_cpu(border), oldblock);
1210
1211                 /* move remainder of path[i] to the new index block */
1212                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1213                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1214                         EXT4_ERROR_INODE(inode,
1215                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1216                                          le32_to_cpu(path[i].p_ext->ee_block));
1217                         err = -EFSCORRUPTED;
1218                         goto cleanup;
1219                 }
1220                 /* start copy indexes */
1221                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1222                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1223                                 EXT_MAX_INDEX(path[i].p_hdr));
1224                 ext4_ext_show_move(inode, path, newblock, i);
1225                 if (m) {
1226                         memmove(++fidx, path[i].p_idx,
1227                                 sizeof(struct ext4_extent_idx) * m);
1228                         le16_add_cpu(&neh->eh_entries, m);
1229                 }
1230                 /* zero out unused area in the extent block */
1231                 ext_size = sizeof(struct ext4_extent_header) +
1232                    (sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries));
1233                 memset(bh->b_data + ext_size, 0,
1234                         inode->i_sb->s_blocksize - ext_size);
1235                 ext4_extent_block_csum_set(inode, neh);
1236                 set_buffer_uptodate(bh);
1237                 unlock_buffer(bh);
1238
1239                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1240                 if (err)
1241                         goto cleanup;
1242                 brelse(bh);
1243                 bh = NULL;
1244
1245                 /* correct old index */
1246                 if (m) {
1247                         err = ext4_ext_get_access(handle, inode, path + i);
1248                         if (err)
1249                                 goto cleanup;
1250                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1251                         err = ext4_ext_dirty(handle, inode, path + i);
1252                         if (err)
1253                                 goto cleanup;
1254                 }
1255
1256                 i--;
1257         }
1258
1259         /* insert new index */
1260         err = ext4_ext_insert_index(handle, inode, path + at,
1261                                     le32_to_cpu(border), newblock);
1262
1263 cleanup:
1264         if (bh) {
1265                 if (buffer_locked(bh))
1266                         unlock_buffer(bh);
1267                 brelse(bh);
1268         }
1269
1270         if (err) {
1271                 /* free all allocated blocks in error case */
1272                 for (i = 0; i < depth; i++) {
1273                         if (!ablocks[i])
1274                                 continue;
1275                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1276                                          EXT4_FREE_BLOCKS_METADATA);
1277                 }
1278         }
1279         kfree(ablocks);
1280
1281         return err;
1282 }
1283
1284 /*
1285  * ext4_ext_grow_indepth:
1286  * implements tree growing procedure:
1287  * - allocates new block
1288  * - moves top-level data (index block or leaf) into the new block
1289  * - initializes new top-level, creating index that points to the
1290  *   just created block
1291  */
1292 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1293                                  unsigned int flags)
1294 {
1295         struct ext4_extent_header *neh;
1296         struct buffer_head *bh;
1297         ext4_fsblk_t newblock, goal = 0;
1298         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1299         int err = 0;
1300         size_t ext_size = 0;
1301
1302         /* Try to prepend new index to old one */
1303         if (ext_depth(inode))
1304                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1305         if (goal > le32_to_cpu(es->s_first_data_block)) {
1306                 flags |= EXT4_MB_HINT_TRY_GOAL;
1307                 goal--;
1308         } else
1309                 goal = ext4_inode_to_goal_block(inode);
1310         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1311                                         NULL, &err);
1312         if (newblock == 0)
1313                 return err;
1314
1315         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1316         if (unlikely(!bh))
1317                 return -ENOMEM;
1318         lock_buffer(bh);
1319
1320         err = ext4_journal_get_create_access(handle, bh);
1321         if (err) {
1322                 unlock_buffer(bh);
1323                 goto out;
1324         }
1325
1326         ext_size = sizeof(EXT4_I(inode)->i_data);
1327         /* move top-level index/leaf into new block */
1328         memmove(bh->b_data, EXT4_I(inode)->i_data, ext_size);
1329         /* zero out unused area in the extent block */
1330         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1331
1332         /* set size of new block */
1333         neh = ext_block_hdr(bh);
1334         /* old root could have indexes or leaves
1335          * so calculate e_max right way */
1336         if (ext_depth(inode))
1337                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1338         else
1339                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1340         neh->eh_magic = EXT4_EXT_MAGIC;
1341         ext4_extent_block_csum_set(inode, neh);
1342         set_buffer_uptodate(bh);
1343         unlock_buffer(bh);
1344
1345         err = ext4_handle_dirty_metadata(handle, inode, bh);
1346         if (err)
1347                 goto out;
1348
1349         /* Update top-level index: num,max,pointer */
1350         neh = ext_inode_hdr(inode);
1351         neh->eh_entries = cpu_to_le16(1);
1352         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1353         if (neh->eh_depth == 0) {
1354                 /* Root extent block becomes index block */
1355                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1356                 EXT_FIRST_INDEX(neh)->ei_block =
1357                         EXT_FIRST_EXTENT(neh)->ee_block;
1358         }
1359         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1360                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1361                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1362                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1363
1364         le16_add_cpu(&neh->eh_depth, 1);
1365         ext4_mark_inode_dirty(handle, inode);
1366 out:
1367         brelse(bh);
1368
1369         return err;
1370 }
1371
1372 /*
1373  * ext4_ext_create_new_leaf:
1374  * finds empty index and adds new leaf.
1375  * if no free index is found, then it requests in-depth growing.
1376  */
1377 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1378                                     unsigned int mb_flags,
1379                                     unsigned int gb_flags,
1380                                     struct ext4_ext_path **ppath,
1381                                     struct ext4_extent *newext)
1382 {
1383         struct ext4_ext_path *path = *ppath;
1384         struct ext4_ext_path *curp;
1385         int depth, i, err = 0;
1386
1387 repeat:
1388         i = depth = ext_depth(inode);
1389
1390         /* walk up to the tree and look for free index entry */
1391         curp = path + depth;
1392         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1393                 i--;
1394                 curp--;
1395         }
1396
1397         /* we use already allocated block for index block,
1398          * so subsequent data blocks should be contiguous */
1399         if (EXT_HAS_FREE_INDEX(curp)) {
1400                 /* if we found index with free entry, then use that
1401                  * entry: create all needed subtree and add new leaf */
1402                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1403                 if (err)
1404                         goto out;
1405
1406                 /* refill path */
1407                 path = ext4_find_extent(inode,
1408                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1409                                     ppath, gb_flags);
1410                 if (IS_ERR(path))
1411                         err = PTR_ERR(path);
1412         } else {
1413                 /* tree is full, time to grow in depth */
1414                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1415                 if (err)
1416                         goto out;
1417
1418                 /* refill path */
1419                 path = ext4_find_extent(inode,
1420                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1421                                     ppath, gb_flags);
1422                 if (IS_ERR(path)) {
1423                         err = PTR_ERR(path);
1424                         goto out;
1425                 }
1426
1427                 /*
1428                  * only first (depth 0 -> 1) produces free space;
1429                  * in all other cases we have to split the grown tree
1430                  */
1431                 depth = ext_depth(inode);
1432                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1433                         /* now we need to split */
1434                         goto repeat;
1435                 }
1436         }
1437
1438 out:
1439         return err;
1440 }
1441
1442 /*
1443  * search the closest allocated block to the left for *logical
1444  * and returns it at @logical + it's physical address at @phys
1445  * if *logical is the smallest allocated block, the function
1446  * returns 0 at @phys
1447  * return value contains 0 (success) or error code
1448  */
1449 static int ext4_ext_search_left(struct inode *inode,
1450                                 struct ext4_ext_path *path,
1451                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1452 {
1453         struct ext4_extent_idx *ix;
1454         struct ext4_extent *ex;
1455         int depth, ee_len;
1456
1457         if (unlikely(path == NULL)) {
1458                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1459                 return -EFSCORRUPTED;
1460         }
1461         depth = path->p_depth;
1462         *phys = 0;
1463
1464         if (depth == 0 && path->p_ext == NULL)
1465                 return 0;
1466
1467         /* usually extent in the path covers blocks smaller
1468          * then *logical, but it can be that extent is the
1469          * first one in the file */
1470
1471         ex = path[depth].p_ext;
1472         ee_len = ext4_ext_get_actual_len(ex);
1473         if (*logical < le32_to_cpu(ex->ee_block)) {
1474                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1475                         EXT4_ERROR_INODE(inode,
1476                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1477                                          *logical, le32_to_cpu(ex->ee_block));
1478                         return -EFSCORRUPTED;
1479                 }
1480                 while (--depth >= 0) {
1481                         ix = path[depth].p_idx;
1482                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1483                                 EXT4_ERROR_INODE(inode,
1484                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1485                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1486                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1487                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1488                                   depth);
1489                                 return -EFSCORRUPTED;
1490                         }
1491                 }
1492                 return 0;
1493         }
1494
1495         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1496                 EXT4_ERROR_INODE(inode,
1497                                  "logical %d < ee_block %d + ee_len %d!",
1498                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1499                 return -EFSCORRUPTED;
1500         }
1501
1502         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1503         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1504         return 0;
1505 }
1506
1507 /*
1508  * search the closest allocated block to the right for *logical
1509  * and returns it at @logical + it's physical address at @phys
1510  * if *logical is the largest allocated block, the function
1511  * returns 0 at @phys
1512  * return value contains 0 (success) or error code
1513  */
1514 static int ext4_ext_search_right(struct inode *inode,
1515                                  struct ext4_ext_path *path,
1516                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1517                                  struct ext4_extent **ret_ex)
1518 {
1519         struct buffer_head *bh = NULL;
1520         struct ext4_extent_header *eh;
1521         struct ext4_extent_idx *ix;
1522         struct ext4_extent *ex;
1523         ext4_fsblk_t block;
1524         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1525         int ee_len;
1526
1527         if (unlikely(path == NULL)) {
1528                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1529                 return -EFSCORRUPTED;
1530         }
1531         depth = path->p_depth;
1532         *phys = 0;
1533
1534         if (depth == 0 && path->p_ext == NULL)
1535                 return 0;
1536
1537         /* usually extent in the path covers blocks smaller
1538          * then *logical, but it can be that extent is the
1539          * first one in the file */
1540
1541         ex = path[depth].p_ext;
1542         ee_len = ext4_ext_get_actual_len(ex);
1543         if (*logical < le32_to_cpu(ex->ee_block)) {
1544                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1545                         EXT4_ERROR_INODE(inode,
1546                                          "first_extent(path[%d].p_hdr) != ex",
1547                                          depth);
1548                         return -EFSCORRUPTED;
1549                 }
1550                 while (--depth >= 0) {
1551                         ix = path[depth].p_idx;
1552                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1553                                 EXT4_ERROR_INODE(inode,
1554                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1555                                                  *logical);
1556                                 return -EFSCORRUPTED;
1557                         }
1558                 }
1559                 goto found_extent;
1560         }
1561
1562         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1563                 EXT4_ERROR_INODE(inode,
1564                                  "logical %d < ee_block %d + ee_len %d!",
1565                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1566                 return -EFSCORRUPTED;
1567         }
1568
1569         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1570                 /* next allocated block in this leaf */
1571                 ex++;
1572                 goto found_extent;
1573         }
1574
1575         /* go up and search for index to the right */
1576         while (--depth >= 0) {
1577                 ix = path[depth].p_idx;
1578                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1579                         goto got_index;
1580         }
1581
1582         /* we've gone up to the root and found no index to the right */
1583         return 0;
1584
1585 got_index:
1586         /* we've found index to the right, let's
1587          * follow it and find the closest allocated
1588          * block to the right */
1589         ix++;
1590         block = ext4_idx_pblock(ix);
1591         while (++depth < path->p_depth) {
1592                 /* subtract from p_depth to get proper eh_depth */
1593                 bh = read_extent_tree_block(inode, block,
1594                                             path->p_depth - depth, 0);
1595                 if (IS_ERR(bh))
1596                         return PTR_ERR(bh);
1597                 eh = ext_block_hdr(bh);
1598                 ix = EXT_FIRST_INDEX(eh);
1599                 block = ext4_idx_pblock(ix);
1600                 put_bh(bh);
1601         }
1602
1603         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1604         if (IS_ERR(bh))
1605                 return PTR_ERR(bh);
1606         eh = ext_block_hdr(bh);
1607         ex = EXT_FIRST_EXTENT(eh);
1608 found_extent:
1609         *logical = le32_to_cpu(ex->ee_block);
1610         *phys = ext4_ext_pblock(ex);
1611         *ret_ex = ex;
1612         if (bh)
1613                 put_bh(bh);
1614         return 0;
1615 }
1616
1617 /*
1618  * ext4_ext_next_allocated_block:
1619  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1620  * NOTE: it considers block number from index entry as
1621  * allocated block. Thus, index entries have to be consistent
1622  * with leaves.
1623  */
1624 ext4_lblk_t
1625 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1626 {
1627         int depth;
1628
1629         BUG_ON(path == NULL);
1630         depth = path->p_depth;
1631
1632         if (depth == 0 && path->p_ext == NULL)
1633                 return EXT_MAX_BLOCKS;
1634
1635         while (depth >= 0) {
1636                 if (depth == path->p_depth) {
1637                         /* leaf */
1638                         if (path[depth].p_ext &&
1639                                 path[depth].p_ext !=
1640                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1641                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1642                 } else {
1643                         /* index */
1644                         if (path[depth].p_idx !=
1645                                         EXT_LAST_INDEX(path[depth].p_hdr))
1646                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1647                 }
1648                 depth--;
1649         }
1650
1651         return EXT_MAX_BLOCKS;
1652 }
1653
1654 /*
1655  * ext4_ext_next_leaf_block:
1656  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1657  */
1658 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1659 {
1660         int depth;
1661
1662         BUG_ON(path == NULL);
1663         depth = path->p_depth;
1664
1665         /* zero-tree has no leaf blocks at all */
1666         if (depth == 0)
1667                 return EXT_MAX_BLOCKS;
1668
1669         /* go to index block */
1670         depth--;
1671
1672         while (depth >= 0) {
1673                 if (path[depth].p_idx !=
1674                                 EXT_LAST_INDEX(path[depth].p_hdr))
1675                         return (ext4_lblk_t)
1676                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1677                 depth--;
1678         }
1679
1680         return EXT_MAX_BLOCKS;
1681 }
1682
1683 /*
1684  * ext4_ext_correct_indexes:
1685  * if leaf gets modified and modified extent is first in the leaf,
1686  * then we have to correct all indexes above.
1687  * TODO: do we need to correct tree in all cases?
1688  */
1689 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1690                                 struct ext4_ext_path *path)
1691 {
1692         struct ext4_extent_header *eh;
1693         int depth = ext_depth(inode);
1694         struct ext4_extent *ex;
1695         __le32 border;
1696         int k, err = 0;
1697
1698         eh = path[depth].p_hdr;
1699         ex = path[depth].p_ext;
1700
1701         if (unlikely(ex == NULL || eh == NULL)) {
1702                 EXT4_ERROR_INODE(inode,
1703                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1704                 return -EFSCORRUPTED;
1705         }
1706
1707         if (depth == 0) {
1708                 /* there is no tree at all */
1709                 return 0;
1710         }
1711
1712         if (ex != EXT_FIRST_EXTENT(eh)) {
1713                 /* we correct tree if first leaf got modified only */
1714                 return 0;
1715         }
1716
1717         /*
1718          * TODO: we need correction if border is smaller than current one
1719          */
1720         k = depth - 1;
1721         border = path[depth].p_ext->ee_block;
1722         err = ext4_ext_get_access(handle, inode, path + k);
1723         if (err)
1724                 return err;
1725         path[k].p_idx->ei_block = border;
1726         err = ext4_ext_dirty(handle, inode, path + k);
1727         if (err)
1728                 return err;
1729
1730         while (k--) {
1731                 /* change all left-side indexes */
1732                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1733                         break;
1734                 err = ext4_ext_get_access(handle, inode, path + k);
1735                 if (err)
1736                         break;
1737                 path[k].p_idx->ei_block = border;
1738                 err = ext4_ext_dirty(handle, inode, path + k);
1739                 if (err)
1740                         break;
1741         }
1742
1743         return err;
1744 }
1745
1746 int
1747 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1748                                 struct ext4_extent *ex2)
1749 {
1750         unsigned short ext1_ee_len, ext2_ee_len;
1751
1752         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1753                 return 0;
1754
1755         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1756         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1757
1758         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1759                         le32_to_cpu(ex2->ee_block))
1760                 return 0;
1761
1762         /*
1763          * To allow future support for preallocated extents to be added
1764          * as an RO_COMPAT feature, refuse to merge to extents if
1765          * this can result in the top bit of ee_len being set.
1766          */
1767         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1768                 return 0;
1769         /*
1770          * The check for IO to unwritten extent is somewhat racy as we
1771          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1772          * dropping i_data_sem. But reserved blocks should save us in that
1773          * case.
1774          */
1775         if (ext4_ext_is_unwritten(ex1) &&
1776             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1777              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1778              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1779                 return 0;
1780 #ifdef AGGRESSIVE_TEST
1781         if (ext1_ee_len >= 4)
1782                 return 0;
1783 #endif
1784
1785         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1786                 return 1;
1787         return 0;
1788 }
1789
1790 /*
1791  * This function tries to merge the "ex" extent to the next extent in the tree.
1792  * It always tries to merge towards right. If you want to merge towards
1793  * left, pass "ex - 1" as argument instead of "ex".
1794  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1795  * 1 if they got merged.
1796  */
1797 static int ext4_ext_try_to_merge_right(struct inode *inode,
1798                                  struct ext4_ext_path *path,
1799                                  struct ext4_extent *ex)
1800 {
1801         struct ext4_extent_header *eh;
1802         unsigned int depth, len;
1803         int merge_done = 0, unwritten;
1804
1805         depth = ext_depth(inode);
1806         BUG_ON(path[depth].p_hdr == NULL);
1807         eh = path[depth].p_hdr;
1808
1809         while (ex < EXT_LAST_EXTENT(eh)) {
1810                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1811                         break;
1812                 /* merge with next extent! */
1813                 unwritten = ext4_ext_is_unwritten(ex);
1814                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1815                                 + ext4_ext_get_actual_len(ex + 1));
1816                 if (unwritten)
1817                         ext4_ext_mark_unwritten(ex);
1818
1819                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1820                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1821                                 * sizeof(struct ext4_extent);
1822                         memmove(ex + 1, ex + 2, len);
1823                 }
1824                 le16_add_cpu(&eh->eh_entries, -1);
1825                 merge_done = 1;
1826                 WARN_ON(eh->eh_entries == 0);
1827                 if (!eh->eh_entries)
1828                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1829         }
1830
1831         return merge_done;
1832 }
1833
1834 /*
1835  * This function does a very simple check to see if we can collapse
1836  * an extent tree with a single extent tree leaf block into the inode.
1837  */
1838 static void ext4_ext_try_to_merge_up(handle_t *handle,
1839                                      struct inode *inode,
1840                                      struct ext4_ext_path *path)
1841 {
1842         size_t s;
1843         unsigned max_root = ext4_ext_space_root(inode, 0);
1844         ext4_fsblk_t blk;
1845
1846         if ((path[0].p_depth != 1) ||
1847             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1848             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1849                 return;
1850
1851         /*
1852          * We need to modify the block allocation bitmap and the block
1853          * group descriptor to release the extent tree block.  If we
1854          * can't get the journal credits, give up.
1855          */
1856         if (ext4_journal_extend(handle, 2))
1857                 return;
1858
1859         /*
1860          * Copy the extent data up to the inode
1861          */
1862         blk = ext4_idx_pblock(path[0].p_idx);
1863         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1864                 sizeof(struct ext4_extent_idx);
1865         s += sizeof(struct ext4_extent_header);
1866
1867         path[1].p_maxdepth = path[0].p_maxdepth;
1868         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1869         path[0].p_depth = 0;
1870         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1871                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1872         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1873
1874         brelse(path[1].p_bh);
1875         ext4_free_blocks(handle, inode, NULL, blk, 1,
1876                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1877 }
1878
1879 /*
1880  * This function tries to merge the @ex extent to neighbours in the tree.
1881  * return 1 if merge left else 0.
1882  */
1883 static void ext4_ext_try_to_merge(handle_t *handle,
1884                                   struct inode *inode,
1885                                   struct ext4_ext_path *path,
1886                                   struct ext4_extent *ex) {
1887         struct ext4_extent_header *eh;
1888         unsigned int depth;
1889         int merge_done = 0;
1890
1891         depth = ext_depth(inode);
1892         BUG_ON(path[depth].p_hdr == NULL);
1893         eh = path[depth].p_hdr;
1894
1895         if (ex > EXT_FIRST_EXTENT(eh))
1896                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1897
1898         if (!merge_done)
1899                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1900
1901         ext4_ext_try_to_merge_up(handle, inode, path);
1902 }
1903
1904 /*
1905  * check if a portion of the "newext" extent overlaps with an
1906  * existing extent.
1907  *
1908  * If there is an overlap discovered, it updates the length of the newext
1909  * such that there will be no overlap, and then returns 1.
1910  * If there is no overlap found, it returns 0.
1911  */
1912 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1913                                            struct inode *inode,
1914                                            struct ext4_extent *newext,
1915                                            struct ext4_ext_path *path)
1916 {
1917         ext4_lblk_t b1, b2;
1918         unsigned int depth, len1;
1919         unsigned int ret = 0;
1920
1921         b1 = le32_to_cpu(newext->ee_block);
1922         len1 = ext4_ext_get_actual_len(newext);
1923         depth = ext_depth(inode);
1924         if (!path[depth].p_ext)
1925                 goto out;
1926         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1927
1928         /*
1929          * get the next allocated block if the extent in the path
1930          * is before the requested block(s)
1931          */
1932         if (b2 < b1) {
1933                 b2 = ext4_ext_next_allocated_block(path);
1934                 if (b2 == EXT_MAX_BLOCKS)
1935                         goto out;
1936                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1937         }
1938
1939         /* check for wrap through zero on extent logical start block*/
1940         if (b1 + len1 < b1) {
1941                 len1 = EXT_MAX_BLOCKS - b1;
1942                 newext->ee_len = cpu_to_le16(len1);
1943                 ret = 1;
1944         }
1945
1946         /* check for overlap */
1947         if (b1 + len1 > b2) {
1948                 newext->ee_len = cpu_to_le16(b2 - b1);
1949                 ret = 1;
1950         }
1951 out:
1952         return ret;
1953 }
1954
1955 /*
1956  * ext4_ext_insert_extent:
1957  * tries to merge requsted extent into the existing extent or
1958  * inserts requested extent as new one into the tree,
1959  * creating new leaf in the no-space case.
1960  */
1961 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1962                                 struct ext4_ext_path **ppath,
1963                                 struct ext4_extent *newext, int gb_flags)
1964 {
1965         struct ext4_ext_path *path = *ppath;
1966         struct ext4_extent_header *eh;
1967         struct ext4_extent *ex, *fex;
1968         struct ext4_extent *nearex; /* nearest extent */
1969         struct ext4_ext_path *npath = NULL;
1970         int depth, len, err;
1971         ext4_lblk_t next;
1972         int mb_flags = 0, unwritten;
1973
1974         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1975                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1976         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1977                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1978                 return -EFSCORRUPTED;
1979         }
1980         depth = ext_depth(inode);
1981         ex = path[depth].p_ext;
1982         eh = path[depth].p_hdr;
1983         if (unlikely(path[depth].p_hdr == NULL)) {
1984                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1985                 return -EFSCORRUPTED;
1986         }
1987
1988         /* try to insert block into found extent and return */
1989         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1990
1991                 /*
1992                  * Try to see whether we should rather test the extent on
1993                  * right from ex, or from the left of ex. This is because
1994                  * ext4_find_extent() can return either extent on the
1995                  * left, or on the right from the searched position. This
1996                  * will make merging more effective.
1997                  */
1998                 if (ex < EXT_LAST_EXTENT(eh) &&
1999                     (le32_to_cpu(ex->ee_block) +
2000                     ext4_ext_get_actual_len(ex) <
2001                     le32_to_cpu(newext->ee_block))) {
2002                         ex += 1;
2003                         goto prepend;
2004                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
2005                            (le32_to_cpu(newext->ee_block) +
2006                            ext4_ext_get_actual_len(newext) <
2007                            le32_to_cpu(ex->ee_block)))
2008                         ex -= 1;
2009
2010                 /* Try to append newex to the ex */
2011                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
2012                         ext_debug("append [%d]%d block to %u:[%d]%d"
2013                                   "(from %llu)\n",
2014                                   ext4_ext_is_unwritten(newext),
2015                                   ext4_ext_get_actual_len(newext),
2016                                   le32_to_cpu(ex->ee_block),
2017                                   ext4_ext_is_unwritten(ex),
2018                                   ext4_ext_get_actual_len(ex),
2019                                   ext4_ext_pblock(ex));
2020                         err = ext4_ext_get_access(handle, inode,
2021                                                   path + depth);
2022                         if (err)
2023                                 return err;
2024                         unwritten = ext4_ext_is_unwritten(ex);
2025                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2026                                         + ext4_ext_get_actual_len(newext));
2027                         if (unwritten)
2028                                 ext4_ext_mark_unwritten(ex);
2029                         eh = path[depth].p_hdr;
2030                         nearex = ex;
2031                         goto merge;
2032                 }
2033
2034 prepend:
2035                 /* Try to prepend newex to the ex */
2036                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2037                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2038                                   "(from %llu)\n",
2039                                   le32_to_cpu(newext->ee_block),
2040                                   ext4_ext_is_unwritten(newext),
2041                                   ext4_ext_get_actual_len(newext),
2042                                   le32_to_cpu(ex->ee_block),
2043                                   ext4_ext_is_unwritten(ex),
2044                                   ext4_ext_get_actual_len(ex),
2045                                   ext4_ext_pblock(ex));
2046                         err = ext4_ext_get_access(handle, inode,
2047                                                   path + depth);
2048                         if (err)
2049                                 return err;
2050
2051                         unwritten = ext4_ext_is_unwritten(ex);
2052                         ex->ee_block = newext->ee_block;
2053                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2054                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2055                                         + ext4_ext_get_actual_len(newext));
2056                         if (unwritten)
2057                                 ext4_ext_mark_unwritten(ex);
2058                         eh = path[depth].p_hdr;
2059                         nearex = ex;
2060                         goto merge;
2061                 }
2062         }
2063
2064         depth = ext_depth(inode);
2065         eh = path[depth].p_hdr;
2066         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2067                 goto has_space;
2068
2069         /* probably next leaf has space for us? */
2070         fex = EXT_LAST_EXTENT(eh);
2071         next = EXT_MAX_BLOCKS;
2072         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2073                 next = ext4_ext_next_leaf_block(path);
2074         if (next != EXT_MAX_BLOCKS) {
2075                 ext_debug("next leaf block - %u\n", next);
2076                 BUG_ON(npath != NULL);
2077                 npath = ext4_find_extent(inode, next, NULL, 0);
2078                 if (IS_ERR(npath))
2079                         return PTR_ERR(npath);
2080                 BUG_ON(npath->p_depth != path->p_depth);
2081                 eh = npath[depth].p_hdr;
2082                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2083                         ext_debug("next leaf isn't full(%d)\n",
2084                                   le16_to_cpu(eh->eh_entries));
2085                         path = npath;
2086                         goto has_space;
2087                 }
2088                 ext_debug("next leaf has no free space(%d,%d)\n",
2089                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2090         }
2091
2092         /*
2093          * There is no free space in the found leaf.
2094          * We're gonna add a new leaf in the tree.
2095          */
2096         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2097                 mb_flags |= EXT4_MB_USE_RESERVED;
2098         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2099                                        ppath, newext);
2100         if (err)
2101                 goto cleanup;
2102         depth = ext_depth(inode);
2103         eh = path[depth].p_hdr;
2104
2105 has_space:
2106         nearex = path[depth].p_ext;
2107
2108         err = ext4_ext_get_access(handle, inode, path + depth);
2109         if (err)
2110                 goto cleanup;
2111
2112         if (!nearex) {
2113                 /* there is no extent in this leaf, create first one */
2114                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\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                 nearex = EXT_FIRST_EXTENT(eh);
2120         } else {
2121                 if (le32_to_cpu(newext->ee_block)
2122                            > le32_to_cpu(nearex->ee_block)) {
2123                         /* Insert after */
2124                         ext_debug("insert %u:%llu:[%d]%d before: "
2125                                         "nearest %p\n",
2126                                         le32_to_cpu(newext->ee_block),
2127                                         ext4_ext_pblock(newext),
2128                                         ext4_ext_is_unwritten(newext),
2129                                         ext4_ext_get_actual_len(newext),
2130                                         nearex);
2131                         nearex++;
2132                 } else {
2133                         /* Insert before */
2134                         BUG_ON(newext->ee_block == nearex->ee_block);
2135                         ext_debug("insert %u:%llu:[%d]%d after: "
2136                                         "nearest %p\n",
2137                                         le32_to_cpu(newext->ee_block),
2138                                         ext4_ext_pblock(newext),
2139                                         ext4_ext_is_unwritten(newext),
2140                                         ext4_ext_get_actual_len(newext),
2141                                         nearex);
2142                 }
2143                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2144                 if (len > 0) {
2145                         ext_debug("insert %u:%llu:[%d]%d: "
2146                                         "move %d extents from 0x%p to 0x%p\n",
2147                                         le32_to_cpu(newext->ee_block),
2148                                         ext4_ext_pblock(newext),
2149                                         ext4_ext_is_unwritten(newext),
2150                                         ext4_ext_get_actual_len(newext),
2151                                         len, nearex, nearex + 1);
2152                         memmove(nearex + 1, nearex,
2153                                 len * sizeof(struct ext4_extent));
2154                 }
2155         }
2156
2157         le16_add_cpu(&eh->eh_entries, 1);
2158         path[depth].p_ext = nearex;
2159         nearex->ee_block = newext->ee_block;
2160         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2161         nearex->ee_len = newext->ee_len;
2162
2163 merge:
2164         /* try to merge extents */
2165         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2166                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2167
2168
2169         /* time to correct all indexes above */
2170         err = ext4_ext_correct_indexes(handle, inode, path);
2171         if (err)
2172                 goto cleanup;
2173
2174         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2175
2176 cleanup:
2177         ext4_ext_drop_refs(npath);
2178         kfree(npath);
2179         return err;
2180 }
2181
2182 static int ext4_fill_fiemap_extents(struct inode *inode,
2183                                     ext4_lblk_t block, ext4_lblk_t num,
2184                                     struct fiemap_extent_info *fieinfo)
2185 {
2186         struct ext4_ext_path *path = NULL;
2187         struct ext4_extent *ex;
2188         struct extent_status es;
2189         ext4_lblk_t next, next_del, start = 0, end = 0;
2190         ext4_lblk_t last = block + num;
2191         int exists, depth = 0, err = 0;
2192         unsigned int flags = 0;
2193         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2194
2195         while (block < last && block != EXT_MAX_BLOCKS) {
2196                 num = last - block;
2197                 /* find extent for this block */
2198                 down_read(&EXT4_I(inode)->i_data_sem);
2199
2200                 path = ext4_find_extent(inode, block, &path, 0);
2201                 if (IS_ERR(path)) {
2202                         up_read(&EXT4_I(inode)->i_data_sem);
2203                         err = PTR_ERR(path);
2204                         path = NULL;
2205                         break;
2206                 }
2207
2208                 depth = ext_depth(inode);
2209                 if (unlikely(path[depth].p_hdr == NULL)) {
2210                         up_read(&EXT4_I(inode)->i_data_sem);
2211                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2212                         err = -EFSCORRUPTED;
2213                         break;
2214                 }
2215                 ex = path[depth].p_ext;
2216                 next = ext4_ext_next_allocated_block(path);
2217
2218                 flags = 0;
2219                 exists = 0;
2220                 if (!ex) {
2221                         /* there is no extent yet, so try to allocate
2222                          * all requested space */
2223                         start = block;
2224                         end = block + num;
2225                 } else if (le32_to_cpu(ex->ee_block) > block) {
2226                         /* need to allocate space before found extent */
2227                         start = block;
2228                         end = le32_to_cpu(ex->ee_block);
2229                         if (block + num < end)
2230                                 end = block + num;
2231                 } else if (block >= le32_to_cpu(ex->ee_block)
2232                                         + ext4_ext_get_actual_len(ex)) {
2233                         /* need to allocate space after found extent */
2234                         start = block;
2235                         end = block + num;
2236                         if (end >= next)
2237                                 end = next;
2238                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2239                         /*
2240                          * some part of requested space is covered
2241                          * by found extent
2242                          */
2243                         start = block;
2244                         end = le32_to_cpu(ex->ee_block)
2245                                 + ext4_ext_get_actual_len(ex);
2246                         if (block + num < end)
2247                                 end = block + num;
2248                         exists = 1;
2249                 } else {
2250                         BUG();
2251                 }
2252                 BUG_ON(end <= start);
2253
2254                 if (!exists) {
2255                         es.es_lblk = start;
2256                         es.es_len = end - start;
2257                         es.es_pblk = 0;
2258                 } else {
2259                         es.es_lblk = le32_to_cpu(ex->ee_block);
2260                         es.es_len = ext4_ext_get_actual_len(ex);
2261                         es.es_pblk = ext4_ext_pblock(ex);
2262                         if (ext4_ext_is_unwritten(ex))
2263                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2264                 }
2265
2266                 /*
2267                  * Find delayed extent and update es accordingly. We call
2268                  * it even in !exists case to find out whether es is the
2269                  * last existing extent or not.
2270                  */
2271                 next_del = ext4_find_delayed_extent(inode, &es);
2272                 if (!exists && next_del) {
2273                         exists = 1;
2274                         flags |= (FIEMAP_EXTENT_DELALLOC |
2275                                   FIEMAP_EXTENT_UNKNOWN);
2276                 }
2277                 up_read(&EXT4_I(inode)->i_data_sem);
2278
2279                 if (unlikely(es.es_len == 0)) {
2280                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2281                         err = -EFSCORRUPTED;
2282                         break;
2283                 }
2284
2285                 /*
2286                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2287                  * we need to check next == EXT_MAX_BLOCKS because it is
2288                  * possible that an extent is with unwritten and delayed
2289                  * status due to when an extent is delayed allocated and
2290                  * is allocated by fallocate status tree will track both of
2291                  * them in a extent.
2292                  *
2293                  * So we could return a unwritten and delayed extent, and
2294                  * its block is equal to 'next'.
2295                  */
2296                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2297                         flags |= FIEMAP_EXTENT_LAST;
2298                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2299                                      next != EXT_MAX_BLOCKS)) {
2300                                 EXT4_ERROR_INODE(inode,
2301                                                  "next extent == %u, next "
2302                                                  "delalloc extent = %u",
2303                                                  next, next_del);
2304                                 err = -EFSCORRUPTED;
2305                                 break;
2306                         }
2307                 }
2308
2309                 if (exists) {
2310                         err = fiemap_fill_next_extent(fieinfo,
2311                                 (__u64)es.es_lblk << blksize_bits,
2312                                 (__u64)es.es_pblk << blksize_bits,
2313                                 (__u64)es.es_len << blksize_bits,
2314                                 flags);
2315                         if (err < 0)
2316                                 break;
2317                         if (err == 1) {
2318                                 err = 0;
2319                                 break;
2320                         }
2321                 }
2322
2323                 block = es.es_lblk + es.es_len;
2324         }
2325
2326         ext4_ext_drop_refs(path);
2327         kfree(path);
2328         return err;
2329 }
2330
2331 /*
2332  * ext4_ext_determine_hole - determine hole around given block
2333  * @inode:      inode we lookup in
2334  * @path:       path in extent tree to @lblk
2335  * @lblk:       pointer to logical block around which we want to determine hole
2336  *
2337  * Determine hole length (and start if easily possible) around given logical
2338  * block. We don't try too hard to find the beginning of the hole but @path
2339  * actually points to extent before @lblk, we provide it.
2340  *
2341  * The function returns the length of a hole starting at @lblk. We update @lblk
2342  * to the beginning of the hole if we managed to find it.
2343  */
2344 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2345                                            struct ext4_ext_path *path,
2346                                            ext4_lblk_t *lblk)
2347 {
2348         int depth = ext_depth(inode);
2349         struct ext4_extent *ex;
2350         ext4_lblk_t len;
2351
2352         ex = path[depth].p_ext;
2353         if (ex == NULL) {
2354                 /* there is no extent yet, so gap is [0;-] */
2355                 *lblk = 0;
2356                 len = EXT_MAX_BLOCKS;
2357         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2358                 len = le32_to_cpu(ex->ee_block) - *lblk;
2359         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2360                         + ext4_ext_get_actual_len(ex)) {
2361                 ext4_lblk_t next;
2362
2363                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2364                 next = ext4_ext_next_allocated_block(path);
2365                 BUG_ON(next == *lblk);
2366                 len = next - *lblk;
2367         } else {
2368                 BUG();
2369         }
2370         return len;
2371 }
2372
2373 /*
2374  * ext4_ext_put_gap_in_cache:
2375  * calculate boundaries of the gap that the requested block fits into
2376  * and cache this gap
2377  */
2378 static void
2379 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2380                           ext4_lblk_t hole_len)
2381 {
2382         struct extent_status es;
2383
2384         ext4_es_find_delayed_extent_range(inode, hole_start,
2385                                           hole_start + hole_len - 1, &es);
2386         if (es.es_len) {
2387                 /* There's delayed extent containing lblock? */
2388                 if (es.es_lblk <= hole_start)
2389                         return;
2390                 hole_len = min(es.es_lblk - hole_start, hole_len);
2391         }
2392         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2393         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2394                               EXTENT_STATUS_HOLE);
2395 }
2396
2397 /*
2398  * ext4_ext_rm_idx:
2399  * removes index from the index block.
2400  */
2401 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2402                         struct ext4_ext_path *path, int depth)
2403 {
2404         int err;
2405         ext4_fsblk_t leaf;
2406
2407         /* free index block */
2408         depth--;
2409         path = path + depth;
2410         leaf = ext4_idx_pblock(path->p_idx);
2411         if (unlikely(path->p_hdr->eh_entries == 0)) {
2412                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2413                 return -EFSCORRUPTED;
2414         }
2415         err = ext4_ext_get_access(handle, inode, path);
2416         if (err)
2417                 return err;
2418
2419         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2420                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2421                 len *= sizeof(struct ext4_extent_idx);
2422                 memmove(path->p_idx, path->p_idx + 1, len);
2423         }
2424
2425         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2426         err = ext4_ext_dirty(handle, inode, path);
2427         if (err)
2428                 return err;
2429         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2430         trace_ext4_ext_rm_idx(inode, leaf);
2431
2432         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2433                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2434
2435         while (--depth >= 0) {
2436                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2437                         break;
2438                 path--;
2439                 err = ext4_ext_get_access(handle, inode, path);
2440                 if (err)
2441                         break;
2442                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2443                 err = ext4_ext_dirty(handle, inode, path);
2444                 if (err)
2445                         break;
2446         }
2447         return err;
2448 }
2449
2450 /*
2451  * ext4_ext_calc_credits_for_single_extent:
2452  * This routine returns max. credits that needed to insert an extent
2453  * to the extent tree.
2454  * When pass the actual path, the caller should calculate credits
2455  * under i_data_sem.
2456  */
2457 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2458                                                 struct ext4_ext_path *path)
2459 {
2460         if (path) {
2461                 int depth = ext_depth(inode);
2462                 int ret = 0;
2463
2464                 /* probably there is space in leaf? */
2465                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2466                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2467
2468                         /*
2469                          *  There are some space in the leaf tree, no
2470                          *  need to account for leaf block credit
2471                          *
2472                          *  bitmaps and block group descriptor blocks
2473                          *  and other metadata blocks still need to be
2474                          *  accounted.
2475                          */
2476                         /* 1 bitmap, 1 block group descriptor */
2477                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2478                         return ret;
2479                 }
2480         }
2481
2482         return ext4_chunk_trans_blocks(inode, nrblocks);
2483 }
2484
2485 /*
2486  * How many index/leaf blocks need to change/allocate to add @extents extents?
2487  *
2488  * If we add a single extent, then in the worse case, each tree level
2489  * index/leaf need to be changed in case of the tree split.
2490  *
2491  * If more extents are inserted, they could cause the whole tree split more
2492  * than once, but this is really rare.
2493  */
2494 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2495 {
2496         int index;
2497         int depth;
2498
2499         /* If we are converting the inline data, only one is needed here. */
2500         if (ext4_has_inline_data(inode))
2501                 return 1;
2502
2503         depth = ext_depth(inode);
2504
2505         if (extents <= 1)
2506                 index = depth * 2;
2507         else
2508                 index = depth * 3;
2509
2510         return index;
2511 }
2512
2513 static inline int get_default_free_blocks_flags(struct inode *inode)
2514 {
2515         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2516             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2517                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2518         else if (ext4_should_journal_data(inode))
2519                 return EXT4_FREE_BLOCKS_FORGET;
2520         return 0;
2521 }
2522
2523 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2524                               struct ext4_extent *ex,
2525                               long long *partial_cluster,
2526                               ext4_lblk_t from, ext4_lblk_t to)
2527 {
2528         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2529         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2530         ext4_fsblk_t pblk;
2531         int flags = get_default_free_blocks_flags(inode);
2532
2533         /*
2534          * For bigalloc file systems, we never free a partial cluster
2535          * at the beginning of the extent.  Instead, we make a note
2536          * that we tried freeing the cluster, and check to see if we
2537          * need to free it on a subsequent call to ext4_remove_blocks,
2538          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2539          */
2540         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2541
2542         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2543         /*
2544          * If we have a partial cluster, and it's different from the
2545          * cluster of the last block, we need to explicitly free the
2546          * partial cluster here.
2547          */
2548         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2549         if (*partial_cluster > 0 &&
2550             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2551                 ext4_free_blocks(handle, inode, NULL,
2552                                  EXT4_C2B(sbi, *partial_cluster),
2553                                  sbi->s_cluster_ratio, flags);
2554                 *partial_cluster = 0;
2555         }
2556
2557 #ifdef EXTENTS_STATS
2558         {
2559                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2560                 spin_lock(&sbi->s_ext_stats_lock);
2561                 sbi->s_ext_blocks += ee_len;
2562                 sbi->s_ext_extents++;
2563                 if (ee_len < sbi->s_ext_min)
2564                         sbi->s_ext_min = ee_len;
2565                 if (ee_len > sbi->s_ext_max)
2566                         sbi->s_ext_max = ee_len;
2567                 if (ext_depth(inode) > sbi->s_depth_max)
2568                         sbi->s_depth_max = ext_depth(inode);
2569                 spin_unlock(&sbi->s_ext_stats_lock);
2570         }
2571 #endif
2572         if (from >= le32_to_cpu(ex->ee_block)
2573             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2574                 /* tail removal */
2575                 ext4_lblk_t num;
2576                 long long first_cluster;
2577
2578                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2579                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2580                 /*
2581                  * Usually we want to free partial cluster at the end of the
2582                  * extent, except for the situation when the cluster is still
2583                  * used by any other extent (partial_cluster is negative).
2584                  */
2585                 if (*partial_cluster < 0 &&
2586                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2587                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2588
2589                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2590                           num, pblk, *partial_cluster);
2591                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2592                 /*
2593                  * If the block range to be freed didn't start at the
2594                  * beginning of a cluster, and we removed the entire
2595                  * extent and the cluster is not used by any other extent,
2596                  * save the partial cluster here, since we might need to
2597                  * delete if we determine that the truncate or punch hole
2598                  * operation has removed all of the blocks in the cluster.
2599                  * If that cluster is used by another extent, preserve its
2600                  * negative value so it isn't freed later on.
2601                  *
2602                  * If the whole extent wasn't freed, we've reached the
2603                  * start of the truncated/punched region and have finished
2604                  * removing blocks.  If there's a partial cluster here it's
2605                  * shared with the remainder of the extent and is no longer
2606                  * a candidate for removal.
2607                  */
2608                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2609                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2610                         if (first_cluster != -*partial_cluster)
2611                                 *partial_cluster = first_cluster;
2612                 } else {
2613                         *partial_cluster = 0;
2614                 }
2615         } else
2616                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2617                            "%u-%u from %u:%u",
2618                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2619         return 0;
2620 }
2621
2622
2623 /*
2624  * ext4_ext_rm_leaf() Removes the extents associated with the
2625  * blocks appearing between "start" and "end".  Both "start"
2626  * and "end" must appear in the same extent or EIO is returned.
2627  *
2628  * @handle: The journal handle
2629  * @inode:  The files inode
2630  * @path:   The path to the leaf
2631  * @partial_cluster: The cluster which we'll have to free if all extents
2632  *                   has been released from it.  However, if this value is
2633  *                   negative, it's a cluster just to the right of the
2634  *                   punched region and it must not be freed.
2635  * @start:  The first block to remove
2636  * @end:   The last block to remove
2637  */
2638 static int
2639 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2640                  struct ext4_ext_path *path,
2641                  long long *partial_cluster,
2642                  ext4_lblk_t start, ext4_lblk_t end)
2643 {
2644         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2645         int err = 0, correct_index = 0;
2646         int depth = ext_depth(inode), credits;
2647         struct ext4_extent_header *eh;
2648         ext4_lblk_t a, b;
2649         unsigned num;
2650         ext4_lblk_t ex_ee_block;
2651         unsigned short ex_ee_len;
2652         unsigned unwritten = 0;
2653         struct ext4_extent *ex;
2654         ext4_fsblk_t pblk;
2655
2656         /* the header must be checked already in ext4_ext_remove_space() */
2657         ext_debug("truncate since %u in leaf to %u\n", start, end);
2658         if (!path[depth].p_hdr)
2659                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2660         eh = path[depth].p_hdr;
2661         if (unlikely(path[depth].p_hdr == NULL)) {
2662                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2663                 return -EFSCORRUPTED;
2664         }
2665         /* find where to start removing */
2666         ex = path[depth].p_ext;
2667         if (!ex)
2668                 ex = EXT_LAST_EXTENT(eh);
2669
2670         ex_ee_block = le32_to_cpu(ex->ee_block);
2671         ex_ee_len = ext4_ext_get_actual_len(ex);
2672
2673         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2674
2675         while (ex >= EXT_FIRST_EXTENT(eh) &&
2676                         ex_ee_block + ex_ee_len > start) {
2677
2678                 if (ext4_ext_is_unwritten(ex))
2679                         unwritten = 1;
2680                 else
2681                         unwritten = 0;
2682
2683                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2684                           unwritten, ex_ee_len);
2685                 path[depth].p_ext = ex;
2686
2687                 a = ex_ee_block > start ? ex_ee_block : start;
2688                 b = ex_ee_block+ex_ee_len - 1 < end ?
2689                         ex_ee_block+ex_ee_len - 1 : end;
2690
2691                 ext_debug("  border %u:%u\n", a, b);
2692
2693                 /* If this extent is beyond the end of the hole, skip it */
2694                 if (end < ex_ee_block) {
2695                         /*
2696                          * We're going to skip this extent and move to another,
2697                          * so note that its first cluster is in use to avoid
2698                          * freeing it when removing blocks.  Eventually, the
2699                          * right edge of the truncated/punched region will
2700                          * be just to the left.
2701                          */
2702                         if (sbi->s_cluster_ratio > 1) {
2703                                 pblk = ext4_ext_pblock(ex);
2704                                 *partial_cluster =
2705                                         -(long long) EXT4_B2C(sbi, pblk);
2706                         }
2707                         ex--;
2708                         ex_ee_block = le32_to_cpu(ex->ee_block);
2709                         ex_ee_len = ext4_ext_get_actual_len(ex);
2710                         continue;
2711                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2712                         EXT4_ERROR_INODE(inode,
2713                                          "can not handle truncate %u:%u "
2714                                          "on extent %u:%u",
2715                                          start, end, ex_ee_block,
2716                                          ex_ee_block + ex_ee_len - 1);
2717                         err = -EFSCORRUPTED;
2718                         goto out;
2719                 } else if (a != ex_ee_block) {
2720                         /* remove tail of the extent */
2721                         num = a - ex_ee_block;
2722                 } else {
2723                         /* remove whole extent: excellent! */
2724                         num = 0;
2725                 }
2726                 /*
2727                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2728                  * descriptor) for each block group; assume two block
2729                  * groups plus ex_ee_len/blocks_per_block_group for
2730                  * the worst case
2731                  */
2732                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2733                 if (ex == EXT_FIRST_EXTENT(eh)) {
2734                         correct_index = 1;
2735                         credits += (ext_depth(inode)) + 1;
2736                 }
2737                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2738
2739                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2740                 if (err)
2741                         goto out;
2742
2743                 err = ext4_ext_get_access(handle, inode, path + depth);
2744                 if (err)
2745                         goto out;
2746
2747                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2748                                          a, b);
2749                 if (err)
2750                         goto out;
2751
2752                 if (num == 0)
2753                         /* this extent is removed; mark slot entirely unused */
2754                         ext4_ext_store_pblock(ex, 0);
2755
2756                 ex->ee_len = cpu_to_le16(num);
2757                 /*
2758                  * Do not mark unwritten if all the blocks in the
2759                  * extent have been removed.
2760                  */
2761                 if (unwritten && num)
2762                         ext4_ext_mark_unwritten(ex);
2763                 /*
2764                  * If the extent was completely released,
2765                  * we need to remove it from the leaf
2766                  */
2767                 if (num == 0) {
2768                         if (end != EXT_MAX_BLOCKS - 1) {
2769                                 /*
2770                                  * For hole punching, we need to scoot all the
2771                                  * extents up when an extent is removed so that
2772                                  * we dont have blank extents in the middle
2773                                  */
2774                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2775                                         sizeof(struct ext4_extent));
2776
2777                                 /* Now get rid of the one at the end */
2778                                 memset(EXT_LAST_EXTENT(eh), 0,
2779                                         sizeof(struct ext4_extent));
2780                         }
2781                         le16_add_cpu(&eh->eh_entries, -1);
2782                 }
2783
2784                 err = ext4_ext_dirty(handle, inode, path + depth);
2785                 if (err)
2786                         goto out;
2787
2788                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2789                                 ext4_ext_pblock(ex));
2790                 ex--;
2791                 ex_ee_block = le32_to_cpu(ex->ee_block);
2792                 ex_ee_len = ext4_ext_get_actual_len(ex);
2793         }
2794
2795         if (correct_index && eh->eh_entries)
2796                 err = ext4_ext_correct_indexes(handle, inode, path);
2797
2798         /*
2799          * If there's a partial cluster and at least one extent remains in
2800          * the leaf, free the partial cluster if it isn't shared with the
2801          * current extent.  If it is shared with the current extent
2802          * we zero partial_cluster because we've reached the start of the
2803          * truncated/punched region and we're done removing blocks.
2804          */
2805         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2806                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2807                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2808                         ext4_free_blocks(handle, inode, NULL,
2809                                          EXT4_C2B(sbi, *partial_cluster),
2810                                          sbi->s_cluster_ratio,
2811                                          get_default_free_blocks_flags(inode));
2812                 }
2813                 *partial_cluster = 0;
2814         }
2815
2816         /* if this leaf is free, then we should
2817          * remove it from index block above */
2818         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2819                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2820
2821 out:
2822         return err;
2823 }
2824
2825 /*
2826  * ext4_ext_more_to_rm:
2827  * returns 1 if current index has to be freed (even partial)
2828  */
2829 static int
2830 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2831 {
2832         BUG_ON(path->p_idx == NULL);
2833
2834         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2835                 return 0;
2836
2837         /*
2838          * if truncate on deeper level happened, it wasn't partial,
2839          * so we have to consider current index for truncation
2840          */
2841         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2842                 return 0;
2843         return 1;
2844 }
2845
2846 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2847                           ext4_lblk_t end)
2848 {
2849         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2850         int depth = ext_depth(inode);
2851         struct ext4_ext_path *path = NULL;
2852         long long partial_cluster = 0;
2853         handle_t *handle;
2854         int i = 0, err = 0;
2855
2856         ext_debug("truncate since %u to %u\n", start, end);
2857
2858         /* probably first extent we're gonna free will be last in block */
2859         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2860         if (IS_ERR(handle))
2861                 return PTR_ERR(handle);
2862
2863 again:
2864         trace_ext4_ext_remove_space(inode, start, end, depth);
2865
2866         /*
2867          * Check if we are removing extents inside the extent tree. If that
2868          * is the case, we are going to punch a hole inside the extent tree
2869          * so we have to check whether we need to split the extent covering
2870          * the last block to remove so we can easily remove the part of it
2871          * in ext4_ext_rm_leaf().
2872          */
2873         if (end < EXT_MAX_BLOCKS - 1) {
2874                 struct ext4_extent *ex;
2875                 ext4_lblk_t ee_block, ex_end, lblk;
2876                 ext4_fsblk_t pblk;
2877
2878                 /* find extent for or closest extent to this block */
2879                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2880                 if (IS_ERR(path)) {
2881                         ext4_journal_stop(handle);
2882                         return PTR_ERR(path);
2883                 }
2884                 depth = ext_depth(inode);
2885                 /* Leaf not may not exist only if inode has no blocks at all */
2886                 ex = path[depth].p_ext;
2887                 if (!ex) {
2888                         if (depth) {
2889                                 EXT4_ERROR_INODE(inode,
2890                                                  "path[%d].p_hdr == NULL",
2891                                                  depth);
2892                                 err = -EFSCORRUPTED;
2893                         }
2894                         goto out;
2895                 }
2896
2897                 ee_block = le32_to_cpu(ex->ee_block);
2898                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2899
2900                 /*
2901                  * See if the last block is inside the extent, if so split
2902                  * the extent at 'end' block so we can easily remove the
2903                  * tail of the first part of the split extent in
2904                  * ext4_ext_rm_leaf().
2905                  */
2906                 if (end >= ee_block && end < ex_end) {
2907
2908                         /*
2909                          * If we're going to split the extent, note that
2910                          * the cluster containing the block after 'end' is
2911                          * in use to avoid freeing it when removing blocks.
2912                          */
2913                         if (sbi->s_cluster_ratio > 1) {
2914                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 1;
2915                                 partial_cluster =
2916                                         -(long long) EXT4_B2C(sbi, pblk);
2917                         }
2918
2919                         /*
2920                          * Split the extent in two so that 'end' is the last
2921                          * block in the first new extent. Also we should not
2922                          * fail removing space due to ENOSPC so try to use
2923                          * reserved block if that happens.
2924                          */
2925                         err = ext4_force_split_extent_at(handle, inode, &path,
2926                                                          end + 1, 1);
2927                         if (err < 0)
2928                                 goto out;
2929
2930                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2931                         /*
2932                          * If there's an extent to the right its first cluster
2933                          * contains the immediate right boundary of the
2934                          * truncated/punched region.  Set partial_cluster to
2935                          * its negative value so it won't be freed if shared
2936                          * with the current extent.  The end < ee_block case
2937                          * is handled in ext4_ext_rm_leaf().
2938                          */
2939                         lblk = ex_end + 1;
2940                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2941                                                     &ex);
2942                         if (err)
2943                                 goto out;
2944                         if (pblk)
2945                                 partial_cluster =
2946                                         -(long long) EXT4_B2C(sbi, pblk);
2947                 }
2948         }
2949         /*
2950          * We start scanning from right side, freeing all the blocks
2951          * after i_size and walking into the tree depth-wise.
2952          */
2953         depth = ext_depth(inode);
2954         if (path) {
2955                 int k = i = depth;
2956                 while (--k > 0)
2957                         path[k].p_block =
2958                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2959         } else {
2960                 path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
2961                                GFP_NOFS);
2962                 if (path == NULL) {
2963                         ext4_journal_stop(handle);
2964                         return -ENOMEM;
2965                 }
2966                 path[0].p_maxdepth = path[0].p_depth = depth;
2967                 path[0].p_hdr = ext_inode_hdr(inode);
2968                 i = 0;
2969
2970                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2971                         err = -EFSCORRUPTED;
2972                         goto out;
2973                 }
2974         }
2975         err = 0;
2976
2977         while (i >= 0 && err == 0) {
2978                 if (i == depth) {
2979                         /* this is leaf block */
2980                         err = ext4_ext_rm_leaf(handle, inode, path,
2981                                                &partial_cluster, start,
2982                                                end);
2983                         /* root level has p_bh == NULL, brelse() eats this */
2984                         brelse(path[i].p_bh);
2985                         path[i].p_bh = NULL;
2986                         i--;
2987                         continue;
2988                 }
2989
2990                 /* this is index block */
2991                 if (!path[i].p_hdr) {
2992                         ext_debug("initialize header\n");
2993                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2994                 }
2995
2996                 if (!path[i].p_idx) {
2997                         /* this level hasn't been touched yet */
2998                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2999                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
3000                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
3001                                   path[i].p_hdr,
3002                                   le16_to_cpu(path[i].p_hdr->eh_entries));
3003                 } else {
3004                         /* we were already here, see at next index */
3005                         path[i].p_idx--;
3006                 }
3007
3008                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
3009                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
3010                                 path[i].p_idx);
3011                 if (ext4_ext_more_to_rm(path + i)) {
3012                         struct buffer_head *bh;
3013                         /* go to the next level */
3014                         ext_debug("move to level %d (block %llu)\n",
3015                                   i + 1, ext4_idx_pblock(path[i].p_idx));
3016                         memset(path + i + 1, 0, sizeof(*path));
3017                         bh = read_extent_tree_block(inode,
3018                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
3019                                 EXT4_EX_NOCACHE);
3020                         if (IS_ERR(bh)) {
3021                                 /* should we reset i_size? */
3022                                 err = PTR_ERR(bh);
3023                                 break;
3024                         }
3025                         /* Yield here to deal with large extent trees.
3026                          * Should be a no-op if we did IO above. */
3027                         cond_resched();
3028                         if (WARN_ON(i + 1 > depth)) {
3029                                 err = -EFSCORRUPTED;
3030                                 break;
3031                         }
3032                         path[i + 1].p_bh = bh;
3033
3034                         /* save actual number of indexes since this
3035                          * number is changed at the next iteration */
3036                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3037                         i++;
3038                 } else {
3039                         /* we finished processing this index, go up */
3040                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3041                                 /* index is empty, remove it;
3042                                  * handle must be already prepared by the
3043                                  * truncatei_leaf() */
3044                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3045                         }
3046                         /* root level has p_bh == NULL, brelse() eats this */
3047                         brelse(path[i].p_bh);
3048                         path[i].p_bh = NULL;
3049                         i--;
3050                         ext_debug("return to level %d\n", i);
3051                 }
3052         }
3053
3054         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3055                         partial_cluster, path->p_hdr->eh_entries);
3056
3057         /*
3058          * If we still have something in the partial cluster and we have removed
3059          * even the first extent, then we should free the blocks in the partial
3060          * cluster as well.  (This code will only run when there are no leaves
3061          * to the immediate left of the truncated/punched region.)
3062          */
3063         if (partial_cluster > 0 && err == 0) {
3064                 /* don't zero partial_cluster since it's not used afterwards */
3065                 ext4_free_blocks(handle, inode, NULL,
3066                                  EXT4_C2B(sbi, partial_cluster),
3067                                  sbi->s_cluster_ratio,
3068                                  get_default_free_blocks_flags(inode));
3069         }
3070
3071         /* TODO: flexible tree reduction should be here */
3072         if (path->p_hdr->eh_entries == 0) {
3073                 /*
3074                  * truncate to zero freed all the tree,
3075                  * so we need to correct eh_depth
3076                  */
3077                 err = ext4_ext_get_access(handle, inode, path);
3078                 if (err == 0) {
3079                         ext_inode_hdr(inode)->eh_depth = 0;
3080                         ext_inode_hdr(inode)->eh_max =
3081                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3082                         err = ext4_ext_dirty(handle, inode, path);
3083                 }
3084         }
3085 out:
3086         ext4_ext_drop_refs(path);
3087         kfree(path);
3088         path = NULL;
3089         if (err == -EAGAIN)
3090                 goto again;
3091         ext4_journal_stop(handle);
3092
3093         return err;
3094 }
3095
3096 /*
3097  * called at mount time
3098  */
3099 void ext4_ext_init(struct super_block *sb)
3100 {
3101         /*
3102          * possible initialization would be here
3103          */
3104
3105         if (ext4_has_feature_extents(sb)) {
3106 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3107                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3108 #ifdef AGGRESSIVE_TEST
3109                        ", aggressive tests"
3110 #endif
3111 #ifdef CHECK_BINSEARCH
3112                        ", check binsearch"
3113 #endif
3114 #ifdef EXTENTS_STATS
3115                        ", stats"
3116 #endif
3117                        "\n");
3118 #endif
3119 #ifdef EXTENTS_STATS
3120                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3121                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3122                 EXT4_SB(sb)->s_ext_max = 0;
3123 #endif
3124         }
3125 }
3126
3127 /*
3128  * called at umount time
3129  */
3130 void ext4_ext_release(struct super_block *sb)
3131 {
3132         if (!ext4_has_feature_extents(sb))
3133                 return;
3134
3135 #ifdef EXTENTS_STATS
3136         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3137                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3138                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3139                         sbi->s_ext_blocks, sbi->s_ext_extents,
3140                         sbi->s_ext_blocks / sbi->s_ext_extents);
3141                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3142                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3143         }
3144 #endif
3145 }
3146
3147 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3148 {
3149         ext4_lblk_t  ee_block;
3150         ext4_fsblk_t ee_pblock;
3151         unsigned int ee_len;
3152
3153         ee_block  = le32_to_cpu(ex->ee_block);
3154         ee_len    = ext4_ext_get_actual_len(ex);
3155         ee_pblock = ext4_ext_pblock(ex);
3156
3157         if (ee_len == 0)
3158                 return 0;
3159
3160         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3161                                      EXTENT_STATUS_WRITTEN);
3162 }
3163
3164 /* FIXME!! we need to try to merge to left or right after zero-out  */
3165 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3166 {
3167         ext4_fsblk_t ee_pblock;
3168         unsigned int ee_len;
3169
3170         ee_len    = ext4_ext_get_actual_len(ex);
3171         ee_pblock = ext4_ext_pblock(ex);
3172         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3173                                   ee_len);
3174 }
3175
3176 /*
3177  * ext4_split_extent_at() splits an extent at given block.
3178  *
3179  * @handle: the journal handle
3180  * @inode: the file inode
3181  * @path: the path to the extent
3182  * @split: the logical block where the extent is splitted.
3183  * @split_flags: indicates if the extent could be zeroout if split fails, and
3184  *               the states(init or unwritten) of new extents.
3185  * @flags: flags used to insert new extent to extent tree.
3186  *
3187  *
3188  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3189  * of which are deterimined by split_flag.
3190  *
3191  * There are two cases:
3192  *  a> the extent are splitted into two extent.
3193  *  b> split is not needed, and just mark the extent.
3194  *
3195  * return 0 on success.
3196  */
3197 static int ext4_split_extent_at(handle_t *handle,
3198                              struct inode *inode,
3199                              struct ext4_ext_path **ppath,
3200                              ext4_lblk_t split,
3201                              int split_flag,
3202                              int flags)
3203 {
3204         struct ext4_ext_path *path = *ppath;
3205         ext4_fsblk_t newblock;
3206         ext4_lblk_t ee_block;
3207         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3208         struct ext4_extent *ex2 = NULL;
3209         unsigned int ee_len, depth;
3210         int err = 0;
3211
3212         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3213                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3214
3215         ext_debug("ext4_split_extents_at: inode %lu, logical"
3216                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3217
3218         ext4_ext_show_leaf(inode, path);
3219
3220         depth = ext_depth(inode);
3221         ex = path[depth].p_ext;
3222         ee_block = le32_to_cpu(ex->ee_block);
3223         ee_len = ext4_ext_get_actual_len(ex);
3224         newblock = split - ee_block + ext4_ext_pblock(ex);
3225
3226         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3227         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3228                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3229                              EXT4_EXT_MARK_UNWRIT1 |
3230                              EXT4_EXT_MARK_UNWRIT2));
3231
3232         err = ext4_ext_get_access(handle, inode, path + depth);
3233         if (err)
3234                 goto out;
3235
3236         if (split == ee_block) {
3237                 /*
3238                  * case b: block @split is the block that the extent begins with
3239                  * then we just change the state of the extent, and splitting
3240                  * is not needed.
3241                  */
3242                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3243                         ext4_ext_mark_unwritten(ex);
3244                 else
3245                         ext4_ext_mark_initialized(ex);
3246
3247                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3248                         ext4_ext_try_to_merge(handle, inode, path, ex);
3249
3250                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3251                 goto out;
3252         }
3253
3254         /* case a */
3255         memcpy(&orig_ex, ex, sizeof(orig_ex));
3256         ex->ee_len = cpu_to_le16(split - ee_block);
3257         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3258                 ext4_ext_mark_unwritten(ex);
3259
3260         /*
3261          * path may lead to new leaf, not to original leaf any more
3262          * after ext4_ext_insert_extent() returns,
3263          */
3264         err = ext4_ext_dirty(handle, inode, path + depth);
3265         if (err)
3266                 goto fix_extent_len;
3267
3268         ex2 = &newex;
3269         ex2->ee_block = cpu_to_le32(split);
3270         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3271         ext4_ext_store_pblock(ex2, newblock);
3272         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3273                 ext4_ext_mark_unwritten(ex2);
3274
3275         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3276         if (err != -ENOSPC && err != -EDQUOT)
3277                 goto out;
3278
3279         if (EXT4_EXT_MAY_ZEROOUT & split_flag) {
3280                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3281                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3282                                 err = ext4_ext_zeroout(inode, ex2);
3283                                 zero_ex.ee_block = ex2->ee_block;
3284                                 zero_ex.ee_len = cpu_to_le16(
3285                                                 ext4_ext_get_actual_len(ex2));
3286                                 ext4_ext_store_pblock(&zero_ex,
3287                                                       ext4_ext_pblock(ex2));
3288                         } else {
3289                                 err = ext4_ext_zeroout(inode, ex);
3290                                 zero_ex.ee_block = ex->ee_block;
3291                                 zero_ex.ee_len = cpu_to_le16(
3292                                                 ext4_ext_get_actual_len(ex));
3293                                 ext4_ext_store_pblock(&zero_ex,
3294                                                       ext4_ext_pblock(ex));
3295                         }
3296                 } else {
3297                         err = ext4_ext_zeroout(inode, &orig_ex);
3298                         zero_ex.ee_block = orig_ex.ee_block;
3299                         zero_ex.ee_len = cpu_to_le16(
3300                                                 ext4_ext_get_actual_len(&orig_ex));
3301                         ext4_ext_store_pblock(&zero_ex,
3302                                               ext4_ext_pblock(&orig_ex));
3303                 }
3304
3305                 if (!err) {
3306                         /* update the extent length and mark as initialized */
3307                         ex->ee_len = cpu_to_le16(ee_len);
3308                         ext4_ext_try_to_merge(handle, inode, path, ex);
3309                         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3310                         if (!err)
3311                                 /* update extent status tree */
3312                                 err = ext4_zeroout_es(inode, &zero_ex);
3313                         /* If we failed at this point, we don't know in which
3314                          * state the extent tree exactly is so don't try to fix
3315                          * length of the original extent as it may do even more
3316                          * damage.
3317                          */
3318                         goto out;
3319                 }
3320         }
3321
3322 fix_extent_len:
3323         ex->ee_len = orig_ex.ee_len;
3324         ext4_ext_dirty(handle, inode, path + path->p_depth);
3325         return err;
3326 out:
3327         ext4_ext_show_leaf(inode, path);
3328         return err;
3329 }
3330
3331 /*
3332  * ext4_split_extents() splits an extent and mark extent which is covered
3333  * by @map as split_flags indicates
3334  *
3335  * It may result in splitting the extent into multiple extents (up to three)
3336  * There are three possibilities:
3337  *   a> There is no split required
3338  *   b> Splits in two extents: Split is happening at either end of the extent
3339  *   c> Splits in three extents: Somone is splitting in middle of the extent
3340  *
3341  */
3342 static int ext4_split_extent(handle_t *handle,
3343                               struct inode *inode,
3344                               struct ext4_ext_path **ppath,
3345                               struct ext4_map_blocks *map,
3346                               int split_flag,
3347                               int flags)
3348 {
3349         struct ext4_ext_path *path = *ppath;
3350         ext4_lblk_t ee_block;
3351         struct ext4_extent *ex;
3352         unsigned int ee_len, depth;
3353         int err = 0;
3354         int unwritten;
3355         int split_flag1, flags1;
3356         int allocated = map->m_len;
3357
3358         depth = ext_depth(inode);
3359         ex = path[depth].p_ext;
3360         ee_block = le32_to_cpu(ex->ee_block);
3361         ee_len = ext4_ext_get_actual_len(ex);
3362         unwritten = ext4_ext_is_unwritten(ex);
3363
3364         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3365                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3366                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3367                 if (unwritten)
3368                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3369                                        EXT4_EXT_MARK_UNWRIT2;
3370                 if (split_flag & EXT4_EXT_DATA_VALID2)
3371                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3372                 err = ext4_split_extent_at(handle, inode, ppath,
3373                                 map->m_lblk + map->m_len, split_flag1, flags1);
3374                 if (err)
3375                         goto out;
3376         } else {
3377                 allocated = ee_len - (map->m_lblk - ee_block);
3378         }
3379         /*
3380          * Update path is required because previous ext4_split_extent_at() may
3381          * result in split of original leaf or extent zeroout.
3382          */
3383         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3384         if (IS_ERR(path))
3385                 return PTR_ERR(path);
3386         depth = ext_depth(inode);
3387         ex = path[depth].p_ext;
3388         if (!ex) {
3389                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3390                                  (unsigned long) map->m_lblk);
3391                 return -EFSCORRUPTED;
3392         }
3393         unwritten = ext4_ext_is_unwritten(ex);
3394         split_flag1 = 0;
3395
3396         if (map->m_lblk >= ee_block) {
3397                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3398                 if (unwritten) {
3399                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3400                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3401                                                      EXT4_EXT_MARK_UNWRIT2);
3402                 }
3403                 err = ext4_split_extent_at(handle, inode, ppath,
3404                                 map->m_lblk, split_flag1, flags);
3405                 if (err)
3406                         goto out;
3407         }
3408
3409         ext4_ext_show_leaf(inode, path);
3410 out:
3411         return err ? err : allocated;
3412 }
3413
3414 /*
3415  * This function is called by ext4_ext_map_blocks() if someone tries to write
3416  * to an unwritten extent. It may result in splitting the unwritten
3417  * extent into multiple extents (up to three - one initialized and two
3418  * unwritten).
3419  * There are three possibilities:
3420  *   a> There is no split required: Entire extent should be initialized
3421  *   b> Splits in two extents: Write is happening at either end of the extent
3422  *   c> Splits in three extents: Somone is writing in middle of the extent
3423  *
3424  * Pre-conditions:
3425  *  - The extent pointed to by 'path' is unwritten.
3426  *  - The extent pointed to by 'path' contains a superset
3427  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3428  *
3429  * Post-conditions on success:
3430  *  - the returned value is the number of blocks beyond map->l_lblk
3431  *    that are allocated and initialized.
3432  *    It is guaranteed to be >= map->m_len.
3433  */
3434 static int ext4_ext_convert_to_initialized(handle_t *handle,
3435                                            struct inode *inode,
3436                                            struct ext4_map_blocks *map,
3437                                            struct ext4_ext_path **ppath,
3438                                            int flags)
3439 {
3440         struct ext4_ext_path *path = *ppath;
3441         struct ext4_sb_info *sbi;
3442         struct ext4_extent_header *eh;
3443         struct ext4_map_blocks split_map;
3444         struct ext4_extent zero_ex1, zero_ex2;
3445         struct ext4_extent *ex, *abut_ex;
3446         ext4_lblk_t ee_block, eof_block;
3447         unsigned int ee_len, depth, map_len = map->m_len;
3448         int allocated = 0, max_zeroout = 0;
3449         int err = 0;
3450         int split_flag = EXT4_EXT_DATA_VALID2;
3451
3452         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3453                 "block %llu, max_blocks %u\n", inode->i_ino,
3454                 (unsigned long long)map->m_lblk, map_len);
3455
3456         sbi = EXT4_SB(inode->i_sb);
3457         eof_block = (EXT4_I(inode)->i_disksize + inode->i_sb->s_blocksize - 1)
3458                         >> inode->i_sb->s_blocksize_bits;
3459         if (eof_block < map->m_lblk + map_len)
3460                 eof_block = map->m_lblk + map_len;
3461
3462         depth = ext_depth(inode);
3463         eh = path[depth].p_hdr;
3464         ex = path[depth].p_ext;
3465         ee_block = le32_to_cpu(ex->ee_block);
3466         ee_len = ext4_ext_get_actual_len(ex);
3467         zero_ex1.ee_len = 0;
3468         zero_ex2.ee_len = 0;
3469
3470         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3471
3472         /* Pre-conditions */
3473         BUG_ON(!ext4_ext_is_unwritten(ex));
3474         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3475
3476         /*
3477          * Attempt to transfer newly initialized blocks from the currently
3478          * unwritten extent to its neighbor. This is much cheaper
3479          * than an insertion followed by a merge as those involve costly
3480          * memmove() calls. Transferring to the left is the common case in
3481          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3482          * followed by append writes.
3483          *
3484          * Limitations of the current logic:
3485          *  - L1: we do not deal with writes covering the whole extent.
3486          *    This would require removing the extent if the transfer
3487          *    is possible.
3488          *  - L2: we only attempt to merge with an extent stored in the
3489          *    same extent tree node.
3490          */
3491         if ((map->m_lblk == ee_block) &&
3492                 /* See if we can merge left */
3493                 (map_len < ee_len) &&           /*L1*/
3494                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3495                 ext4_lblk_t prev_lblk;
3496                 ext4_fsblk_t prev_pblk, ee_pblk;
3497                 unsigned int prev_len;
3498
3499                 abut_ex = ex - 1;
3500                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3501                 prev_len = ext4_ext_get_actual_len(abut_ex);
3502                 prev_pblk = ext4_ext_pblock(abut_ex);
3503                 ee_pblk = ext4_ext_pblock(ex);
3504
3505                 /*
3506                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3507                  * upon those conditions:
3508                  * - C1: abut_ex is initialized,
3509                  * - C2: abut_ex is logically abutting ex,
3510                  * - C3: abut_ex is physically abutting ex,
3511                  * - C4: abut_ex can receive the additional blocks without
3512                  *   overflowing the (initialized) length limit.
3513                  */
3514                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3515                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3516                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3517                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3518                         err = ext4_ext_get_access(handle, inode, path + depth);
3519                         if (err)
3520                                 goto out;
3521
3522                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3523                                 map, ex, abut_ex);
3524
3525                         /* Shift the start of ex by 'map_len' blocks */
3526                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3527                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3528                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3529                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3530
3531                         /* Extend abut_ex by 'map_len' blocks */
3532                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3533
3534                         /* Result: number of initialized blocks past m_lblk */
3535                         allocated = map_len;
3536                 }
3537         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3538                    (map_len < ee_len) &&        /*L1*/
3539                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3540                 /* See if we can merge right */
3541                 ext4_lblk_t next_lblk;
3542                 ext4_fsblk_t next_pblk, ee_pblk;
3543                 unsigned int next_len;
3544
3545                 abut_ex = ex + 1;
3546                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3547                 next_len = ext4_ext_get_actual_len(abut_ex);
3548                 next_pblk = ext4_ext_pblock(abut_ex);
3549                 ee_pblk = ext4_ext_pblock(ex);
3550
3551                 /*
3552                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3553                  * upon those conditions:
3554                  * - C1: abut_ex is initialized,
3555                  * - C2: abut_ex is logically abutting ex,
3556                  * - C3: abut_ex is physically abutting ex,
3557                  * - C4: abut_ex can receive the additional blocks without
3558                  *   overflowing the (initialized) length limit.
3559                  */
3560                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3561                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3562                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3563                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3564                         err = ext4_ext_get_access(handle, inode, path + depth);
3565                         if (err)
3566                                 goto out;
3567
3568                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3569                                 map, ex, abut_ex);
3570
3571                         /* Shift the start of abut_ex by 'map_len' blocks */
3572                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3573                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3574                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3575                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3576
3577                         /* Extend abut_ex by 'map_len' blocks */
3578                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3579
3580                         /* Result: number of initialized blocks past m_lblk */
3581                         allocated = map_len;
3582                 }
3583         }
3584         if (allocated) {
3585                 /* Mark the block containing both extents as dirty */
3586                 ext4_ext_dirty(handle, inode, path + depth);
3587
3588                 /* Update path to point to the right extent */
3589                 path[depth].p_ext = abut_ex;
3590                 goto out;
3591         } else
3592                 allocated = ee_len - (map->m_lblk - ee_block);
3593
3594         WARN_ON(map->m_lblk < ee_block);
3595         /*
3596          * It is safe to convert extent to initialized via explicit
3597          * zeroout only if extent is fully inside i_size or new_size.
3598          */
3599         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3600
3601         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3602                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3603                         (inode->i_sb->s_blocksize_bits - 10);
3604
3605         if (ext4_encrypted_inode(inode))
3606                 max_zeroout = 0;
3607
3608         /*
3609          * five cases:
3610          * 1. split the extent into three extents.
3611          * 2. split the extent into two extents, zeroout the head of the first
3612          *    extent.
3613          * 3. split the extent into two extents, zeroout the tail of the second
3614          *    extent.
3615          * 4. split the extent into two extents with out zeroout.
3616          * 5. no splitting needed, just possibly zeroout the head and / or the
3617          *    tail of the extent.
3618          */
3619         split_map.m_lblk = map->m_lblk;
3620         split_map.m_len = map->m_len;
3621
3622         if (max_zeroout && (allocated > split_map.m_len)) {
3623                 if (allocated <= max_zeroout) {
3624                         /* case 3 or 5 */
3625                         zero_ex1.ee_block =
3626                                  cpu_to_le32(split_map.m_lblk +
3627                                              split_map.m_len);
3628                         zero_ex1.ee_len =
3629                                 cpu_to_le16(allocated - split_map.m_len);
3630                         ext4_ext_store_pblock(&zero_ex1,
3631                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3632                                 split_map.m_len - ee_block);
3633                         err = ext4_ext_zeroout(inode, &zero_ex1);
3634                         if (err)
3635                                 goto out;
3636                         split_map.m_len = allocated;
3637                 }
3638                 if (split_map.m_lblk - ee_block + split_map.m_len <
3639                                                                 max_zeroout) {
3640                         /* case 2 or 5 */
3641                         if (split_map.m_lblk != ee_block) {
3642                                 zero_ex2.ee_block = ex->ee_block;
3643                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3644                                                         ee_block);
3645                                 ext4_ext_store_pblock(&zero_ex2,
3646                                                       ext4_ext_pblock(ex));
3647                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3648                                 if (err)
3649                                         goto out;
3650                         }
3651
3652                         split_map.m_len += split_map.m_lblk - ee_block;
3653                         split_map.m_lblk = ee_block;
3654                         allocated = map->m_len;
3655                 }
3656         }
3657
3658         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3659                                 flags);
3660         if (err > 0)
3661                 err = 0;
3662 out:
3663         /* If we have gotten a failure, don't zero out status tree */
3664         if (!err) {
3665                 err = ext4_zeroout_es(inode, &zero_ex1);
3666                 if (!err)
3667                         err = ext4_zeroout_es(inode, &zero_ex2);
3668         }
3669         return err ? err : allocated;
3670 }
3671
3672 /*
3673  * This function is called by ext4_ext_map_blocks() from
3674  * ext4_get_blocks_dio_write() when DIO to write
3675  * to an unwritten extent.
3676  *
3677  * Writing to an unwritten extent may result in splitting the unwritten
3678  * extent into multiple initialized/unwritten extents (up to three)
3679  * There are three possibilities:
3680  *   a> There is no split required: Entire extent should be unwritten
3681  *   b> Splits in two extents: Write is happening at either end of the extent
3682  *   c> Splits in three extents: Somone is writing in middle of the extent
3683  *
3684  * This works the same way in the case of initialized -> unwritten conversion.
3685  *
3686  * One of more index blocks maybe needed if the extent tree grow after
3687  * the unwritten extent split. To prevent ENOSPC occur at the IO
3688  * complete, we need to split the unwritten extent before DIO submit
3689  * the IO. The unwritten extent called at this time will be split
3690  * into three unwritten extent(at most). After IO complete, the part
3691  * being filled will be convert to initialized by the end_io callback function
3692  * via ext4_convert_unwritten_extents().
3693  *
3694  * Returns the size of unwritten extent to be written on success.
3695  */
3696 static int ext4_split_convert_extents(handle_t *handle,
3697                                         struct inode *inode,
3698                                         struct ext4_map_blocks *map,
3699                                         struct ext4_ext_path **ppath,
3700                                         int flags)
3701 {
3702         struct ext4_ext_path *path = *ppath;
3703         ext4_lblk_t eof_block;
3704         ext4_lblk_t ee_block;
3705         struct ext4_extent *ex;
3706         unsigned int ee_len;
3707         int split_flag = 0, depth;
3708
3709         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3710                   __func__, inode->i_ino,
3711                   (unsigned long long)map->m_lblk, map->m_len);
3712
3713         eof_block = (EXT4_I(inode)->i_disksize + inode->i_sb->s_blocksize - 1)
3714                         >> inode->i_sb->s_blocksize_bits;
3715         if (eof_block < map->m_lblk + map->m_len)
3716                 eof_block = map->m_lblk + map->m_len;
3717         /*
3718          * It is safe to convert extent to initialized via explicit
3719          * zeroout only if extent is fully insde i_size or new_size.
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         /* Convert to unwritten */
3727         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3728                 split_flag |= EXT4_EXT_DATA_VALID1;
3729         /* Convert to initialized */
3730         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3731                 split_flag |= ee_block + ee_len <= eof_block ?
3732                               EXT4_EXT_MAY_ZEROOUT : 0;
3733                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3734         }
3735         flags |= EXT4_GET_BLOCKS_PRE_IO;
3736         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3737 }
3738
3739 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3740                                                 struct inode *inode,
3741                                                 struct ext4_map_blocks *map,
3742                                                 struct ext4_ext_path **ppath)
3743 {
3744         struct ext4_ext_path *path = *ppath;
3745         struct ext4_extent *ex;
3746         ext4_lblk_t ee_block;
3747         unsigned int ee_len;
3748         int depth;
3749         int err = 0;
3750
3751         depth = ext_depth(inode);
3752         ex = path[depth].p_ext;
3753         ee_block = le32_to_cpu(ex->ee_block);
3754         ee_len = ext4_ext_get_actual_len(ex);
3755
3756         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3757                 "block %llu, max_blocks %u\n", inode->i_ino,
3758                   (unsigned long long)ee_block, ee_len);
3759
3760         /* If extent is larger than requested it is a clear sign that we still
3761          * have some extent state machine issues left. So extent_split is still
3762          * required.
3763          * TODO: Once all related issues will be fixed this situation should be
3764          * illegal.
3765          */
3766         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3767 #ifdef CONFIG_EXT4_DEBUG
3768                 ext4_warning(inode->i_sb, "Inode (%ld) finished: extent logical block %llu,"
3769                              " len %u; IO logical block %llu, len %u",
3770                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3771                              (unsigned long long)map->m_lblk, map->m_len);
3772 #endif
3773                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3774                                                  EXT4_GET_BLOCKS_CONVERT);
3775                 if (err < 0)
3776                         return err;
3777                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3778                 if (IS_ERR(path))
3779                         return PTR_ERR(path);
3780                 depth = ext_depth(inode);
3781                 ex = path[depth].p_ext;
3782         }
3783
3784         err = ext4_ext_get_access(handle, inode, path + depth);
3785         if (err)
3786                 goto out;
3787         /* first mark the extent as initialized */
3788         ext4_ext_mark_initialized(ex);
3789
3790         /* note: ext4_ext_correct_indexes() isn't needed here because
3791          * borders are not changed
3792          */
3793         ext4_ext_try_to_merge(handle, inode, path, ex);
3794
3795         /* Mark modified extent as dirty */
3796         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3797 out:
3798         ext4_ext_show_leaf(inode, path);
3799         return err;
3800 }
3801
3802 /*
3803  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3804  */
3805 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3806                               ext4_lblk_t lblk,
3807                               struct ext4_ext_path *path,
3808                               unsigned int len)
3809 {
3810         int i, depth;
3811         struct ext4_extent_header *eh;
3812         struct ext4_extent *last_ex;
3813
3814         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3815                 return 0;
3816
3817         depth = ext_depth(inode);
3818         eh = path[depth].p_hdr;
3819
3820         /*
3821          * We're going to remove EOFBLOCKS_FL entirely in future so we
3822          * do not care for this case anymore. Simply remove the flag
3823          * if there are no extents.
3824          */
3825         if (unlikely(!eh->eh_entries))
3826                 goto out;
3827         last_ex = EXT_LAST_EXTENT(eh);
3828         /*
3829          * We should clear the EOFBLOCKS_FL flag if we are writing the
3830          * last block in the last extent in the file.  We test this by
3831          * first checking to see if the caller to
3832          * ext4_ext_get_blocks() was interested in the last block (or
3833          * a block beyond the last block) in the current extent.  If
3834          * this turns out to be false, we can bail out from this
3835          * function immediately.
3836          */
3837         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3838             ext4_ext_get_actual_len(last_ex))
3839                 return 0;
3840         /*
3841          * If the caller does appear to be planning to write at or
3842          * beyond the end of the current extent, we then test to see
3843          * if the current extent is the last extent in the file, by
3844          * checking to make sure it was reached via the rightmost node
3845          * at each level of the tree.
3846          */
3847         for (i = depth-1; i >= 0; i--)
3848                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3849                         return 0;
3850 out:
3851         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3852         return ext4_mark_inode_dirty(handle, inode);
3853 }
3854
3855 /**
3856  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3857  *
3858  * Return 1 if there is a delalloc block in the range, otherwise 0.
3859  */
3860 int ext4_find_delalloc_range(struct inode *inode,
3861                              ext4_lblk_t lblk_start,
3862                              ext4_lblk_t lblk_end)
3863 {
3864         struct extent_status es;
3865
3866         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3867         if (es.es_len == 0)
3868                 return 0; /* there is no delay extent in this tree */
3869         else if (es.es_lblk <= lblk_start &&
3870                  lblk_start < es.es_lblk + es.es_len)
3871                 return 1;
3872         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3873                 return 1;
3874         else
3875                 return 0;
3876 }
3877
3878 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3879 {
3880         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3881         ext4_lblk_t lblk_start, lblk_end;
3882         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3883         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3884
3885         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3886 }
3887
3888 /**
3889  * Determines how many complete clusters (out of those specified by the 'map')
3890  * are under delalloc and were reserved quota for.
3891  * This function is called when we are writing out the blocks that were
3892  * originally written with their allocation delayed, but then the space was
3893  * allocated using fallocate() before the delayed allocation could be resolved.
3894  * The cases to look for are:
3895  * ('=' indicated delayed allocated blocks
3896  *  '-' indicates non-delayed allocated blocks)
3897  * (a) partial clusters towards beginning and/or end outside of allocated range
3898  *     are not delalloc'ed.
3899  *      Ex:
3900  *      |----c---=|====c====|====c====|===-c----|
3901  *               |++++++ allocated ++++++|
3902  *      ==> 4 complete clusters in above example
3903  *
3904  * (b) partial cluster (outside of allocated range) towards either end is
3905  *     marked for delayed allocation. In this case, we will exclude that
3906  *     cluster.
3907  *      Ex:
3908  *      |----====c========|========c========|
3909  *           |++++++ allocated ++++++|
3910  *      ==> 1 complete clusters in above example
3911  *
3912  *      Ex:
3913  *      |================c================|
3914  *            |++++++ allocated ++++++|
3915  *      ==> 0 complete clusters in above example
3916  *
3917  * The ext4_da_update_reserve_space will be called only if we
3918  * determine here that there were some "entire" clusters that span
3919  * this 'allocated' range.
3920  * In the non-bigalloc case, this function will just end up returning num_blks
3921  * without ever calling ext4_find_delalloc_range.
3922  */
3923 static unsigned int
3924 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3925                            unsigned int num_blks)
3926 {
3927         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3928         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3929         ext4_lblk_t lblk_from, lblk_to, c_offset;
3930         unsigned int allocated_clusters = 0;
3931
3932         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3933         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3934
3935         /* max possible clusters for this allocation */
3936         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3937
3938         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3939
3940         /* Check towards left side */
3941         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3942         if (c_offset) {
3943                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3944                 lblk_to = lblk_from + c_offset - 1;
3945
3946                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3947                         allocated_clusters--;
3948         }
3949
3950         /* Now check towards right. */
3951         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3952         if (allocated_clusters && c_offset) {
3953                 lblk_from = lblk_start + num_blks;
3954                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3955
3956                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3957                         allocated_clusters--;
3958         }
3959
3960         return allocated_clusters;
3961 }
3962
3963 static int
3964 convert_initialized_extent(handle_t *handle, struct inode *inode,
3965                            struct ext4_map_blocks *map,
3966                            struct ext4_ext_path **ppath,
3967                            unsigned int allocated)
3968 {
3969         struct ext4_ext_path *path = *ppath;
3970         struct ext4_extent *ex;
3971         ext4_lblk_t ee_block;
3972         unsigned int ee_len;
3973         int depth;
3974         int err = 0;
3975
3976         /*
3977          * Make sure that the extent is no bigger than we support with
3978          * unwritten extent
3979          */
3980         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3981                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3982
3983         depth = ext_depth(inode);
3984         ex = path[depth].p_ext;
3985         ee_block = le32_to_cpu(ex->ee_block);
3986         ee_len = ext4_ext_get_actual_len(ex);
3987
3988         ext_debug("%s: inode %lu, logical"
3989                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3990                   (unsigned long long)ee_block, ee_len);
3991
3992         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3993                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3994                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3995                 if (err < 0)
3996                         return err;
3997                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3998                 if (IS_ERR(path))
3999                         return PTR_ERR(path);
4000                 depth = ext_depth(inode);
4001                 ex = path[depth].p_ext;
4002                 if (!ex) {
4003                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
4004                                          (unsigned long) map->m_lblk);
4005                         return -EFSCORRUPTED;
4006                 }
4007         }
4008
4009         err = ext4_ext_get_access(handle, inode, path + depth);
4010         if (err)
4011                 return err;
4012         /* first mark the extent as unwritten */
4013         ext4_ext_mark_unwritten(ex);
4014
4015         /* note: ext4_ext_correct_indexes() isn't needed here because
4016          * borders are not changed
4017          */
4018         ext4_ext_try_to_merge(handle, inode, path, ex);
4019
4020         /* Mark modified extent as dirty */
4021         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
4022         if (err)
4023                 return err;
4024         ext4_ext_show_leaf(inode, path);
4025
4026         ext4_update_inode_fsync_trans(handle, inode, 1);
4027         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
4028         if (err)
4029                 return err;
4030         map->m_flags |= EXT4_MAP_UNWRITTEN;
4031         if (allocated > map->m_len)
4032                 allocated = map->m_len;
4033         map->m_len = allocated;
4034         return allocated;
4035 }
4036
4037 static int
4038 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4039                         struct ext4_map_blocks *map,
4040                         struct ext4_ext_path **ppath, int flags,
4041                         unsigned int allocated, ext4_fsblk_t newblock)
4042 {
4043         struct ext4_ext_path *path = *ppath;
4044         int ret = 0;
4045         int err = 0;
4046
4047         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4048                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4049                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4050                   flags, allocated);
4051         ext4_ext_show_leaf(inode, path);
4052
4053         /*
4054          * When writing into unwritten space, we should not fail to
4055          * allocate metadata blocks for the new extent block if needed.
4056          */
4057         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4058
4059         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4060                                                     allocated, newblock);
4061
4062         /* get_block() before submit the IO, split the extent */
4063         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4064                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4065                                          flags | EXT4_GET_BLOCKS_CONVERT);
4066                 if (ret <= 0)
4067                         goto out;
4068                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4069                 goto out;
4070         }
4071         /* IO end_io complete, convert the filled extent to written */
4072         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4073                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4074                         if (allocated > map->m_len)
4075                                 allocated = map->m_len;
4076                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4077                                                  allocated);
4078                         if (err < 0)
4079                                 goto out2;
4080                 }
4081                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4082                                                            ppath);
4083                 if (ret >= 0) {
4084                         ext4_update_inode_fsync_trans(handle, inode, 1);
4085                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4086                                                  path, map->m_len);
4087                 } else
4088                         err = ret;
4089                 map->m_flags |= EXT4_MAP_MAPPED;
4090                 map->m_pblk = newblock;
4091                 if (allocated > map->m_len)
4092                         allocated = map->m_len;
4093                 map->m_len = allocated;
4094                 goto out2;
4095         }
4096         /* buffered IO case */
4097         /*
4098          * repeat fallocate creation request
4099          * we already have an unwritten extent
4100          */
4101         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4102                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4103                 goto map_out;
4104         }
4105
4106         /* buffered READ or buffered write_begin() lookup */
4107         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4108                 /*
4109                  * We have blocks reserved already.  We
4110                  * return allocated blocks so that delalloc
4111                  * won't do block reservation for us.  But
4112                  * the buffer head will be unmapped so that
4113                  * a read from the block returns 0s.
4114                  */
4115                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4116                 goto out1;
4117         }
4118
4119         /* buffered write, writepage time, convert*/
4120         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4121         if (ret >= 0)
4122                 ext4_update_inode_fsync_trans(handle, inode, 1);
4123 out:
4124         if (ret <= 0) {
4125                 err = ret;
4126                 goto out2;
4127         } else
4128                 allocated = ret;
4129         map->m_flags |= EXT4_MAP_NEW;
4130         /*
4131          * if we allocated more blocks than requested
4132          * we need to make sure we unmap the extra block
4133          * allocated. The actual needed block will get
4134          * unmapped later when we find the buffer_head marked
4135          * new.
4136          */
4137         if (allocated > map->m_len) {
4138                 clean_bdev_aliases(inode->i_sb->s_bdev, newblock + map->m_len,
4139                                    allocated - map->m_len);
4140                 allocated = map->m_len;
4141         }
4142         map->m_len = allocated;
4143
4144         /*
4145          * If we have done fallocate with the offset that is already
4146          * delayed allocated, we would have block reservation
4147          * and quota reservation done in the delayed write path.
4148          * But fallocate would have already updated quota and block
4149          * count for this offset. So cancel these reservation
4150          */
4151         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4152                 unsigned int reserved_clusters;
4153                 reserved_clusters = get_reserved_cluster_alloc(inode,
4154                                 map->m_lblk, map->m_len);
4155                 if (reserved_clusters)
4156                         ext4_da_update_reserve_space(inode,
4157                                                      reserved_clusters,
4158                                                      0);
4159         }
4160
4161 map_out:
4162         map->m_flags |= EXT4_MAP_MAPPED;
4163         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4164                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4165                                          map->m_len);
4166                 if (err < 0)
4167                         goto out2;
4168         }
4169 out1:
4170         if (allocated > map->m_len)
4171                 allocated = map->m_len;
4172         ext4_ext_show_leaf(inode, path);
4173         map->m_pblk = newblock;
4174         map->m_len = allocated;
4175 out2:
4176         return err ? err : allocated;
4177 }
4178
4179 /*
4180  * get_implied_cluster_alloc - check to see if the requested
4181  * allocation (in the map structure) overlaps with a cluster already
4182  * allocated in an extent.
4183  *      @sb     The filesystem superblock structure
4184  *      @map    The requested lblk->pblk mapping
4185  *      @ex     The extent structure which might contain an implied
4186  *                      cluster allocation
4187  *
4188  * This function is called by ext4_ext_map_blocks() after we failed to
4189  * find blocks that were already in the inode's extent tree.  Hence,
4190  * we know that the beginning of the requested region cannot overlap
4191  * the extent from the inode's extent tree.  There are three cases we
4192  * want to catch.  The first is this case:
4193  *
4194  *               |--- cluster # N--|
4195  *    |--- extent ---|  |---- requested region ---|
4196  *                      |==========|
4197  *
4198  * The second case that we need to test for is this one:
4199  *
4200  *   |--------- cluster # N ----------------|
4201  *         |--- requested region --|   |------- extent ----|
4202  *         |=======================|
4203  *
4204  * The third case is when the requested region lies between two extents
4205  * within the same cluster:
4206  *          |------------- cluster # N-------------|
4207  * |----- ex -----|                  |---- ex_right ----|
4208  *                  |------ requested region ------|
4209  *                  |================|
4210  *
4211  * In each of the above cases, we need to set the map->m_pblk and
4212  * map->m_len so it corresponds to the return the extent labelled as
4213  * "|====|" from cluster #N, since it is already in use for data in
4214  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4215  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4216  * as a new "allocated" block region.  Otherwise, we will return 0 and
4217  * ext4_ext_map_blocks() will then allocate one or more new clusters
4218  * by calling ext4_mb_new_blocks().
4219  */
4220 static int get_implied_cluster_alloc(struct super_block *sb,
4221                                      struct ext4_map_blocks *map,
4222                                      struct ext4_extent *ex,
4223                                      struct ext4_ext_path *path)
4224 {
4225         struct ext4_sb_info *sbi = EXT4_SB(sb);
4226         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4227         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4228         ext4_lblk_t rr_cluster_start;
4229         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4230         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4231         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4232
4233         /* The extent passed in that we are trying to match */
4234         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4235         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4236
4237         /* The requested region passed into ext4_map_blocks() */
4238         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4239
4240         if ((rr_cluster_start == ex_cluster_end) ||
4241             (rr_cluster_start == ex_cluster_start)) {
4242                 if (rr_cluster_start == ex_cluster_end)
4243                         ee_start += ee_len - 1;
4244                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4245                 map->m_len = min(map->m_len,
4246                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4247                 /*
4248                  * Check for and handle this case:
4249                  *
4250                  *   |--------- cluster # N-------------|
4251                  *                     |------- extent ----|
4252                  *         |--- requested region ---|
4253                  *         |===========|
4254                  */
4255
4256                 if (map->m_lblk < ee_block)
4257                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4258
4259                 /*
4260                  * Check for the case where there is already another allocated
4261                  * block to the right of 'ex' but before the end of the cluster.
4262                  *
4263                  *          |------------- cluster # N-------------|
4264                  * |----- ex -----|                  |---- ex_right ----|
4265                  *                  |------ requested region ------|
4266                  *                  |================|
4267                  */
4268                 if (map->m_lblk > ee_block) {
4269                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4270                         map->m_len = min(map->m_len, next - map->m_lblk);
4271                 }
4272
4273                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4274                 return 1;
4275         }
4276
4277         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4278         return 0;
4279 }
4280
4281
4282 /*
4283  * Block allocation/map/preallocation routine for extents based files
4284  *
4285  *
4286  * Need to be called with
4287  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4288  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4289  *
4290  * return > 0, number of of blocks already mapped/allocated
4291  *          if create == 0 and these are pre-allocated blocks
4292  *              buffer head is unmapped
4293  *          otherwise blocks are mapped
4294  *
4295  * return = 0, if plain look up failed (blocks have not been allocated)
4296  *          buffer head is unmapped
4297  *
4298  * return < 0, error case.
4299  */
4300 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4301                         struct ext4_map_blocks *map, int flags)
4302 {
4303         struct ext4_ext_path *path = NULL;
4304         struct ext4_extent newex, *ex, *ex2;
4305         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4306         ext4_fsblk_t newblock = 0;
4307         int free_on_err = 0, err = 0, depth, ret;
4308         unsigned int allocated = 0, offset = 0;
4309         unsigned int allocated_clusters = 0;
4310         struct ext4_allocation_request ar;
4311         ext4_lblk_t cluster_offset;
4312         bool map_from_cluster = false;
4313
4314         ext_debug("blocks %u/%u requested for inode %lu\n",
4315                   map->m_lblk, map->m_len, inode->i_ino);
4316         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4317
4318         /* find extent for this block */
4319         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4320         if (IS_ERR(path)) {
4321                 err = PTR_ERR(path);
4322                 path = NULL;
4323                 goto out2;
4324         }
4325
4326         depth = ext_depth(inode);
4327
4328         /*
4329          * consistent leaf must not be empty;
4330          * this situation is possible, though, _during_ tree modification;
4331          * this is why assert can't be put in ext4_find_extent()
4332          */
4333         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4334                 EXT4_ERROR_INODE(inode, "bad extent address "
4335                                  "lblock: %lu, depth: %d pblock %lld",
4336                                  (unsigned long) map->m_lblk, depth,
4337                                  path[depth].p_block);
4338                 err = -EFSCORRUPTED;
4339                 goto out2;
4340         }
4341
4342         ex = path[depth].p_ext;
4343         if (ex) {
4344                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4345                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4346                 unsigned short ee_len;
4347
4348
4349                 /*
4350                  * unwritten extents are treated as holes, except that
4351                  * we split out initialized portions during a write.
4352                  */
4353                 ee_len = ext4_ext_get_actual_len(ex);
4354
4355                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4356
4357                 /* if found extent covers block, simply return it */
4358                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4359                         newblock = map->m_lblk - ee_block + ee_start;
4360                         /* number of remaining blocks in the extent */
4361                         allocated = ee_len - (map->m_lblk - ee_block);
4362                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4363                                   ee_block, ee_len, newblock);
4364
4365                         /*
4366                          * If the extent is initialized check whether the
4367                          * caller wants to convert it to unwritten.
4368                          */
4369                         if ((!ext4_ext_is_unwritten(ex)) &&
4370                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4371                                 allocated = convert_initialized_extent(
4372                                                 handle, inode, map, &path,
4373                                                 allocated);
4374                                 goto out2;
4375                         } else if (!ext4_ext_is_unwritten(ex))
4376                                 goto out;
4377
4378                         ret = ext4_ext_handle_unwritten_extents(
4379                                 handle, inode, map, &path, flags,
4380                                 allocated, newblock);
4381                         if (ret < 0)
4382                                 err = ret;
4383                         else
4384                                 allocated = ret;
4385                         goto out2;
4386                 }
4387         }
4388
4389         /*
4390          * requested block isn't allocated yet;
4391          * we couldn't try to create block if create flag is zero
4392          */
4393         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4394                 ext4_lblk_t hole_start, hole_len;
4395
4396                 hole_start = map->m_lblk;
4397                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4398                 /*
4399                  * put just found gap into cache to speed up
4400                  * subsequent requests
4401                  */
4402                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4403
4404                 /* Update hole_len to reflect hole size after map->m_lblk */
4405                 if (hole_start != map->m_lblk)
4406                         hole_len -= map->m_lblk - hole_start;
4407                 map->m_pblk = 0;
4408                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4409
4410                 goto out2;
4411         }
4412
4413         /*
4414          * Okay, we need to do block allocation.
4415          */
4416         newex.ee_block = cpu_to_le32(map->m_lblk);
4417         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4418
4419         /*
4420          * If we are doing bigalloc, check to see if the extent returned
4421          * by ext4_find_extent() implies a cluster we can use.
4422          */
4423         if (cluster_offset && ex &&
4424             get_implied_cluster_alloc(inode->i_sb, map, ex, 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         /* find neighbour allocated blocks */
4432         ar.lleft = map->m_lblk;
4433         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4434         if (err)
4435                 goto out2;
4436         ar.lright = map->m_lblk;
4437         ex2 = NULL;
4438         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4439         if (err)
4440                 goto out2;
4441
4442         /* Check if the extent after searching to the right implies a
4443          * cluster we can use. */
4444         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4445             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4446                 ar.len = allocated = map->m_len;
4447                 newblock = map->m_pblk;
4448                 map_from_cluster = true;
4449                 goto got_allocated_blocks;
4450         }
4451
4452         /*
4453          * See if request is beyond maximum number of blocks we can have in
4454          * a single extent. For an initialized extent this limit is
4455          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4456          * EXT_UNWRITTEN_MAX_LEN.
4457          */
4458         if (map->m_len > EXT_INIT_MAX_LEN &&
4459             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4460                 map->m_len = EXT_INIT_MAX_LEN;
4461         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4462                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4463                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4464
4465         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4466         newex.ee_len = cpu_to_le16(map->m_len);
4467         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4468         if (err)
4469                 allocated = ext4_ext_get_actual_len(&newex);
4470         else
4471                 allocated = map->m_len;
4472
4473         /* allocate new block */
4474         ar.inode = inode;
4475         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4476         ar.logical = map->m_lblk;
4477         /*
4478          * We calculate the offset from the beginning of the cluster
4479          * for the logical block number, since when we allocate a
4480          * physical cluster, the physical block should start at the
4481          * same offset from the beginning of the cluster.  This is
4482          * needed so that future calls to get_implied_cluster_alloc()
4483          * work correctly.
4484          */
4485         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4486         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4487         ar.goal -= offset;
4488         ar.logical -= offset;
4489         if (S_ISREG(inode->i_mode))
4490                 ar.flags = EXT4_MB_HINT_DATA;
4491         else
4492                 /* disable in-core preallocation for non-regular files */
4493                 ar.flags = 0;
4494         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4495                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4496         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4497                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4498         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4499                 ar.flags |= EXT4_MB_USE_RESERVED;
4500         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4501         if (!newblock)
4502                 goto out2;
4503         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4504                   ar.goal, newblock, allocated);
4505         free_on_err = 1;
4506         allocated_clusters = ar.len;
4507         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4508         if (ar.len > allocated)
4509                 ar.len = allocated;
4510
4511 got_allocated_blocks:
4512         /* try to insert new extent into found leaf and return */
4513         ext4_ext_store_pblock(&newex, newblock + offset);
4514         newex.ee_len = cpu_to_le16(ar.len);
4515         /* Mark unwritten */
4516         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4517                 ext4_ext_mark_unwritten(&newex);
4518                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4519         }
4520
4521         err = 0;
4522         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4523                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4524                                          path, ar.len);
4525         if (!err)
4526                 err = ext4_ext_insert_extent(handle, inode, &path,
4527                                              &newex, flags);
4528
4529         if (err && free_on_err) {
4530                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4531                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4532                 /* free data blocks we just allocated */
4533                 /* not a good idea to call discard here directly,
4534                  * but otherwise we'd need to call it every free() */
4535                 ext4_discard_preallocations(inode);
4536                 ext4_free_blocks(handle, inode, NULL, newblock,
4537                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4538                 goto out2;
4539         }
4540
4541         /* previous routine could use block we allocated */
4542         newblock = ext4_ext_pblock(&newex);
4543         allocated = ext4_ext_get_actual_len(&newex);
4544         if (allocated > map->m_len)
4545                 allocated = map->m_len;
4546         map->m_flags |= EXT4_MAP_NEW;
4547
4548         /*
4549          * Update reserved blocks/metadata blocks after successful
4550          * block allocation which had been deferred till now.
4551          */
4552         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4553                 unsigned int reserved_clusters;
4554                 /*
4555                  * Check how many clusters we had reserved this allocated range
4556                  */
4557                 reserved_clusters = get_reserved_cluster_alloc(inode,
4558                                                 map->m_lblk, allocated);
4559                 if (!map_from_cluster) {
4560                         BUG_ON(allocated_clusters < reserved_clusters);
4561                         if (reserved_clusters < allocated_clusters) {
4562                                 struct ext4_inode_info *ei = EXT4_I(inode);
4563                                 int reservation = allocated_clusters -
4564                                                   reserved_clusters;
4565                                 /*
4566                                  * It seems we claimed few clusters outside of
4567                                  * the range of this allocation. We should give
4568                                  * it back to the reservation pool. This can
4569                                  * happen in the following case:
4570                                  *
4571                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4572                                  *   cluster has 4 blocks. Thus, the clusters
4573                                  *   are [0-3],[4-7],[8-11]...
4574                                  * * First comes delayed allocation write for
4575                                  *   logical blocks 10 & 11. Since there were no
4576                                  *   previous delayed allocated blocks in the
4577                                  *   range [8-11], we would reserve 1 cluster
4578                                  *   for this write.
4579                                  * * Next comes write for logical blocks 3 to 8.
4580                                  *   In this case, we will reserve 2 clusters
4581                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4582                                  *   that range has a delayed allocated blocks.
4583                                  *   Thus total reserved clusters now becomes 3.
4584                                  * * Now, during the delayed allocation writeout
4585                                  *   time, we will first write blocks [3-8] and
4586                                  *   allocate 3 clusters for writing these
4587                                  *   blocks. Also, we would claim all these
4588                                  *   three clusters above.
4589                                  * * Now when we come here to writeout the
4590                                  *   blocks [10-11], we would expect to claim
4591                                  *   the reservation of 1 cluster we had made
4592                                  *   (and we would claim it since there are no
4593                                  *   more delayed allocated blocks in the range
4594                                  *   [8-11]. But our reserved cluster count had
4595                                  *   already gone to 0.
4596                                  *
4597                                  *   Thus, at the step 4 above when we determine
4598                                  *   that there are still some unwritten delayed
4599                                  *   allocated blocks outside of our current
4600                                  *   block range, we should increment the
4601                                  *   reserved clusters count so that when the
4602                                  *   remaining blocks finally gets written, we
4603                                  *   could claim them.
4604                                  */
4605                                 dquot_reserve_block(inode,
4606                                                 EXT4_C2B(sbi, reservation));
4607                                 spin_lock(&ei->i_block_reservation_lock);
4608                                 ei->i_reserved_data_blocks += reservation;
4609                                 spin_unlock(&ei->i_block_reservation_lock);
4610                         }
4611                         /*
4612                          * We will claim quota for all newly allocated blocks.
4613                          * We're updating the reserved space *after* the
4614                          * correction above so we do not accidentally free
4615                          * all the metadata reservation because we might
4616                          * actually need it later on.
4617                          */
4618                         ext4_da_update_reserve_space(inode, allocated_clusters,
4619                                                         1);
4620                 }
4621         }
4622
4623         /*
4624          * Cache the extent and update transaction to commit on fdatasync only
4625          * when it is _not_ an unwritten extent.
4626          */
4627         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4628                 ext4_update_inode_fsync_trans(handle, inode, 1);
4629         else
4630                 ext4_update_inode_fsync_trans(handle, inode, 0);
4631 out:
4632         if (allocated > map->m_len)
4633                 allocated = map->m_len;
4634         ext4_ext_show_leaf(inode, path);
4635         map->m_flags |= EXT4_MAP_MAPPED;
4636         map->m_pblk = newblock;
4637         map->m_len = allocated;
4638 out2:
4639         ext4_ext_drop_refs(path);
4640         kfree(path);
4641
4642         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4643                                        err ? err : allocated);
4644         return err ? err : allocated;
4645 }
4646
4647 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4648 {
4649         struct super_block *sb = inode->i_sb;
4650         ext4_lblk_t last_block;
4651         int err = 0;
4652
4653         /*
4654          * TODO: optimization is possible here.
4655          * Probably we need not scan at all,
4656          * because page truncation is enough.
4657          */
4658
4659         /* we have to know where to truncate from in crash case */
4660         EXT4_I(inode)->i_disksize = inode->i_size;
4661         err = ext4_mark_inode_dirty(handle, inode);
4662         if (err)
4663                 return err;
4664
4665         last_block = (inode->i_size + sb->s_blocksize - 1)
4666                         >> EXT4_BLOCK_SIZE_BITS(sb);
4667 retry:
4668         err = ext4_es_remove_extent(inode, last_block,
4669                                     EXT_MAX_BLOCKS - last_block);
4670         if (err == -ENOMEM) {
4671                 cond_resched();
4672                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4673                 goto retry;
4674         }
4675         if (err)
4676                 return err;
4677         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4678 }
4679
4680 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4681                                   ext4_lblk_t len, loff_t new_size,
4682                                   int flags)
4683 {
4684         struct inode *inode = file_inode(file);
4685         handle_t *handle;
4686         int ret = 0;
4687         int ret2 = 0;
4688         int retries = 0;
4689         int depth = 0;
4690         struct ext4_map_blocks map;
4691         unsigned int credits;
4692         loff_t epos;
4693
4694         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4695         map.m_lblk = offset;
4696         map.m_len = len;
4697         /*
4698          * Don't normalize the request if it can fit in one extent so
4699          * that it doesn't get unnecessarily split into multiple
4700          * extents.
4701          */
4702         if (len <= EXT_UNWRITTEN_MAX_LEN)
4703                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4704
4705         /*
4706          * credits to insert 1 extent into extent tree
4707          */
4708         credits = ext4_chunk_trans_blocks(inode, len);
4709         depth = ext_depth(inode);
4710
4711 retry:
4712         while (ret >= 0 && len) {
4713                 /*
4714                  * Recalculate credits when extent tree depth changes.
4715                  */
4716                 if (depth != ext_depth(inode)) {
4717                         credits = ext4_chunk_trans_blocks(inode, len);
4718                         depth = ext_depth(inode);
4719                 }
4720
4721                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4722                                             credits);
4723                 if (IS_ERR(handle)) {
4724                         ret = PTR_ERR(handle);
4725                         break;
4726                 }
4727                 ret = ext4_map_blocks(handle, inode, &map, flags);
4728                 if (ret <= 0) {
4729                         ext4_debug("inode #%lu: block %u: len %u: "
4730                                    "ext4_ext_map_blocks returned %d",
4731                                    inode->i_ino, map.m_lblk,
4732                                    map.m_len, ret);
4733                         ext4_mark_inode_dirty(handle, inode);
4734                         ret2 = ext4_journal_stop(handle);
4735                         break;
4736                 }
4737                 map.m_lblk += ret;
4738                 map.m_len = len = len - ret;
4739                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4740                 inode->i_ctime = current_time(inode);
4741                 if (new_size) {
4742                         if (epos > new_size)
4743                                 epos = new_size;
4744                         if (ext4_update_inode_size(inode, epos) & 0x1)
4745                                 inode->i_mtime = inode->i_ctime;
4746                 } else {
4747                         if (epos > inode->i_size)
4748                                 ext4_set_inode_flag(inode,
4749                                                     EXT4_INODE_EOFBLOCKS);
4750                 }
4751                 ext4_mark_inode_dirty(handle, inode);
4752                 ext4_update_inode_fsync_trans(handle, inode, 1);
4753                 ret2 = ext4_journal_stop(handle);
4754                 if (ret2)
4755                         break;
4756         }
4757         if (ret == -ENOSPC &&
4758                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4759                 ret = 0;
4760                 goto retry;
4761         }
4762
4763         return ret > 0 ? ret2 : ret;
4764 }
4765
4766 static long ext4_zero_range(struct file *file, loff_t offset,
4767                             loff_t len, int mode)
4768 {
4769         struct inode *inode = file_inode(file);
4770         handle_t *handle = NULL;
4771         unsigned int max_blocks;
4772         loff_t new_size = 0;
4773         int ret = 0;
4774         int flags;
4775         int credits;
4776         int partial_begin, partial_end;
4777         loff_t start, end;
4778         ext4_lblk_t lblk;
4779         unsigned int blkbits = inode->i_blkbits;
4780
4781         trace_ext4_zero_range(inode, offset, len, mode);
4782
4783         if (!S_ISREG(inode->i_mode))
4784                 return -EINVAL;
4785
4786         /* Call ext4_force_commit to flush all data in case of data=journal. */
4787         if (ext4_should_journal_data(inode)) {
4788                 ret = ext4_force_commit(inode->i_sb);
4789                 if (ret)
4790                         return ret;
4791         }
4792
4793         /*
4794          * Round up offset. This is not fallocate, we neet to zero out
4795          * blocks, so convert interior block aligned part of the range to
4796          * unwritten and possibly manually zero out unaligned parts of the
4797          * range.
4798          */
4799         start = round_up(offset, 1 << blkbits);
4800         end = round_down((offset + len), 1 << blkbits);
4801
4802         if (start < offset || end > offset + len)
4803                 return -EINVAL;
4804         partial_begin = offset & ((1 << blkbits) - 1);
4805         partial_end = (offset + len) & ((1 << blkbits) - 1);
4806
4807         lblk = start >> blkbits;
4808         max_blocks = (end >> blkbits);
4809         if (max_blocks < lblk)
4810                 max_blocks = 0;
4811         else
4812                 max_blocks -= lblk;
4813
4814         inode_lock(inode);
4815
4816         /*
4817          * Indirect files do not support unwritten extnets
4818          */
4819         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4820                 ret = -EOPNOTSUPP;
4821                 goto out_mutex;
4822         }
4823
4824         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4825             (offset + len > i_size_read(inode) ||
4826              offset + len > EXT4_I(inode)->i_disksize)) {
4827                 new_size = offset + len;
4828                 ret = inode_newsize_ok(inode, new_size);
4829                 if (ret)
4830                         goto out_mutex;
4831         }
4832
4833         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4834         if (mode & FALLOC_FL_KEEP_SIZE)
4835                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4836
4837         /* Wait all existing dio workers, newcomers will block on i_mutex */
4838         inode_dio_wait(inode);
4839
4840         /* Preallocate the range including the unaligned edges */
4841         if (partial_begin || partial_end) {
4842                 ret = ext4_alloc_file_blocks(file,
4843                                 round_down(offset, 1 << blkbits) >> blkbits,
4844                                 (round_up((offset + len), 1 << blkbits) -
4845                                  round_down(offset, 1 << blkbits)) >> blkbits,
4846                                 new_size, flags);
4847                 if (ret)
4848                         goto out_mutex;
4849
4850         }
4851
4852         /* Zero range excluding the unaligned edges */
4853         if (max_blocks > 0) {
4854                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4855                           EXT4_EX_NOCACHE);
4856
4857                 /*
4858                  * Prevent page faults from reinstantiating pages we have
4859                  * released from page cache.
4860                  */
4861                 down_write(&EXT4_I(inode)->i_mmap_sem);
4862
4863                 ret = ext4_break_layouts(inode);
4864                 if (ret) {
4865                         up_write(&EXT4_I(inode)->i_mmap_sem);
4866                         goto out_mutex;
4867                 }
4868
4869                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4870                 if (ret) {
4871                         up_write(&EXT4_I(inode)->i_mmap_sem);
4872                         goto out_mutex;
4873                 }
4874                 /* Now release the pages and zero block aligned part of pages */
4875                 truncate_pagecache_range(inode, start, end - 1);
4876                 inode->i_mtime = inode->i_ctime = current_time(inode);
4877
4878                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4879                                              flags);
4880                 up_write(&EXT4_I(inode)->i_mmap_sem);
4881                 if (ret)
4882                         goto out_mutex;
4883         }
4884         if (!partial_begin && !partial_end)
4885                 goto out_mutex;
4886
4887         /*
4888          * In worst case we have to writeout two nonadjacent unwritten
4889          * blocks and update the inode
4890          */
4891         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4892         if (ext4_should_journal_data(inode))
4893                 credits += 2;
4894         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4895         if (IS_ERR(handle)) {
4896                 ret = PTR_ERR(handle);
4897                 ext4_std_error(inode->i_sb, ret);
4898                 goto out_mutex;
4899         }
4900
4901         inode->i_mtime = inode->i_ctime = current_time(inode);
4902         if (new_size) {
4903                 ext4_update_inode_size(inode, new_size);
4904         } else {
4905                 /*
4906                 * Mark that we allocate beyond EOF so the subsequent truncate
4907                 * can proceed even if the new size is the same as i_size.
4908                 */
4909                 if ((offset + len) > i_size_read(inode))
4910                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4911         }
4912         ext4_mark_inode_dirty(handle, inode);
4913
4914         /* Zero out partial block at the edges of the range */
4915         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4916         if (ret >= 0)
4917                 ext4_update_inode_fsync_trans(handle, inode, 1);
4918
4919         if (file->f_flags & O_SYNC)
4920                 ext4_handle_sync(handle);
4921
4922         ext4_journal_stop(handle);
4923 out_mutex:
4924         inode_unlock(inode);
4925         return ret;
4926 }
4927
4928 /*
4929  * preallocate space for a file. This implements ext4's fallocate file
4930  * operation, which gets called from sys_fallocate system call.
4931  * For block-mapped files, posix_fallocate should fall back to the method
4932  * of writing zeroes to the required new blocks (the same behavior which is
4933  * expected for file systems which do not support fallocate() system call).
4934  */
4935 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4936 {
4937         struct inode *inode = file_inode(file);
4938         loff_t new_size = 0;
4939         unsigned int max_blocks;
4940         int ret = 0;
4941         int flags;
4942         ext4_lblk_t lblk;
4943         unsigned int blkbits = inode->i_blkbits;
4944
4945         /*
4946          * Encrypted inodes can't handle collapse range or insert
4947          * range since we would need to re-encrypt blocks with a
4948          * different IV or XTS tweak (which are based on the logical
4949          * block number).
4950          *
4951          * XXX It's not clear why zero range isn't working, but we'll
4952          * leave it disabled for encrypted inodes for now.  This is a
4953          * bug we should fix....
4954          */
4955         if (ext4_encrypted_inode(inode) &&
4956             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4957                      FALLOC_FL_ZERO_RANGE)))
4958                 return -EOPNOTSUPP;
4959
4960         /* Return error if mode is not supported */
4961         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4962                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4963                      FALLOC_FL_INSERT_RANGE))
4964                 return -EOPNOTSUPP;
4965
4966         if (mode & FALLOC_FL_PUNCH_HOLE)
4967                 return ext4_punch_hole(inode, offset, len);
4968
4969         ret = ext4_convert_inline_data(inode);
4970         if (ret)
4971                 return ret;
4972
4973         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4974                 return ext4_collapse_range(inode, offset, len);
4975
4976         if (mode & FALLOC_FL_INSERT_RANGE)
4977                 return ext4_insert_range(inode, offset, len);
4978
4979         if (mode & FALLOC_FL_ZERO_RANGE)
4980                 return ext4_zero_range(file, offset, len, mode);
4981
4982         trace_ext4_fallocate_enter(inode, offset, len, mode);
4983         lblk = offset >> blkbits;
4984
4985         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4986         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4987         if (mode & FALLOC_FL_KEEP_SIZE)
4988                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4989
4990         inode_lock(inode);
4991
4992         /*
4993          * We only support preallocation for extent-based files only
4994          */
4995         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4996                 ret = -EOPNOTSUPP;
4997                 goto out;
4998         }
4999
5000         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
5001             (offset + len > i_size_read(inode) ||
5002              offset + len > EXT4_I(inode)->i_disksize)) {
5003                 new_size = offset + len;
5004                 ret = inode_newsize_ok(inode, new_size);
5005                 if (ret)
5006                         goto out;
5007         }
5008
5009         /* Wait all existing dio workers, newcomers will block on i_mutex */
5010         inode_dio_wait(inode);
5011
5012         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
5013         if (ret)
5014                 goto out;
5015
5016         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
5017                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
5018                                                 EXT4_I(inode)->i_sync_tid);
5019         }
5020 out:
5021         inode_unlock(inode);
5022         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
5023         return ret;
5024 }
5025
5026 /*
5027  * This function convert a range of blocks to written extents
5028  * The caller of this function will pass the start offset and the size.
5029  * all unwritten extents within this range will be converted to
5030  * written extents.
5031  *
5032  * This function is called from the direct IO end io call back
5033  * function, to convert the fallocated extents after IO is completed.
5034  * Returns 0 on success.
5035  */
5036 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
5037                                    loff_t offset, ssize_t len)
5038 {
5039         unsigned int max_blocks;
5040         int ret = 0;
5041         int ret2 = 0;
5042         struct ext4_map_blocks map;
5043         unsigned int credits, blkbits = inode->i_blkbits;
5044
5045         map.m_lblk = offset >> blkbits;
5046         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
5047
5048         /*
5049          * This is somewhat ugly but the idea is clear: When transaction is
5050          * reserved, everything goes into it. Otherwise we rather start several
5051          * smaller transactions for conversion of each extent separately.
5052          */
5053         if (handle) {
5054                 handle = ext4_journal_start_reserved(handle,
5055                                                      EXT4_HT_EXT_CONVERT);
5056                 if (IS_ERR(handle))
5057                         return PTR_ERR(handle);
5058                 credits = 0;
5059         } else {
5060                 /*
5061                  * credits to insert 1 extent into extent tree
5062                  */
5063                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5064         }
5065         while (ret >= 0 && ret < max_blocks) {
5066                 map.m_lblk += ret;
5067                 map.m_len = (max_blocks -= ret);
5068                 if (credits) {
5069                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5070                                                     credits);
5071                         if (IS_ERR(handle)) {
5072                                 ret = PTR_ERR(handle);
5073                                 break;
5074                         }
5075                 }
5076                 ret = ext4_map_blocks(handle, inode, &map,
5077                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5078                 if (ret <= 0)
5079                         ext4_warning(inode->i_sb,
5080                                      "inode #%lu: block %u: len %u: "
5081                                      "ext4_ext_map_blocks returned %d",
5082                                      inode->i_ino, map.m_lblk,
5083                                      map.m_len, ret);
5084                 ext4_mark_inode_dirty(handle, inode);
5085                 if (credits)
5086                         ret2 = ext4_journal_stop(handle);
5087                 if (ret <= 0 || ret2)
5088                         break;
5089         }
5090         if (!credits)
5091                 ret2 = ext4_journal_stop(handle);
5092         return ret > 0 ? ret2 : ret;
5093 }
5094
5095 /*
5096  * If newes is not existing extent (newes->ec_pblk equals zero) find
5097  * delayed extent at start of newes and update newes accordingly and
5098  * return start of the next delayed extent.
5099  *
5100  * If newes is existing extent (newes->ec_pblk is not equal zero)
5101  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5102  * extent found. Leave newes unmodified.
5103  */
5104 static int ext4_find_delayed_extent(struct inode *inode,
5105                                     struct extent_status *newes)
5106 {
5107         struct extent_status es;
5108         ext4_lblk_t block, next_del;
5109
5110         if (newes->es_pblk == 0) {
5111                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5112                                 newes->es_lblk + newes->es_len - 1, &es);
5113
5114                 /*
5115                  * No extent in extent-tree contains block @newes->es_pblk,
5116                  * then the block may stay in 1)a hole or 2)delayed-extent.
5117                  */
5118                 if (es.es_len == 0)
5119                         /* A hole found. */
5120                         return 0;
5121
5122                 if (es.es_lblk > newes->es_lblk) {
5123                         /* A hole found. */
5124                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5125                                             newes->es_len);
5126                         return 0;
5127                 }
5128
5129                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5130         }
5131
5132         block = newes->es_lblk + newes->es_len;
5133         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5134         if (es.es_len == 0)
5135                 next_del = EXT_MAX_BLOCKS;
5136         else
5137                 next_del = es.es_lblk;
5138
5139         return next_del;
5140 }
5141 /* fiemap flags we can handle specified here */
5142 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5143
5144 static int ext4_xattr_fiemap(struct inode *inode,
5145                                 struct fiemap_extent_info *fieinfo)
5146 {
5147         __u64 physical = 0;
5148         __u64 length;
5149         __u32 flags = FIEMAP_EXTENT_LAST;
5150         int blockbits = inode->i_sb->s_blocksize_bits;
5151         int error = 0;
5152
5153         /* in-inode? */
5154         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5155                 struct ext4_iloc iloc;
5156                 int offset;     /* offset of xattr in inode */
5157
5158                 error = ext4_get_inode_loc(inode, &iloc);
5159                 if (error)
5160                         return error;
5161                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5162                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5163                                 EXT4_I(inode)->i_extra_isize;
5164                 physical += offset;
5165                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5166                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5167                 brelse(iloc.bh);
5168         } else { /* external block */
5169                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5170                 length = inode->i_sb->s_blocksize;
5171         }
5172
5173         if (physical)
5174                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5175                                                 length, flags);
5176         return (error < 0 ? error : 0);
5177 }
5178
5179 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5180                 __u64 start, __u64 len)
5181 {
5182         ext4_lblk_t start_blk;
5183         int error = 0;
5184
5185         if (ext4_has_inline_data(inode)) {
5186                 int has_inline = 1;
5187
5188                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5189                                                 start, len);
5190
5191                 if (has_inline)
5192                         return error;
5193         }
5194
5195         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5196                 error = ext4_ext_precache(inode);
5197                 if (error)
5198                         return error;
5199         }
5200
5201         /* fallback to generic here if not in extents fmt */
5202         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5203                 return generic_block_fiemap(inode, fieinfo, start, len,
5204                         ext4_get_block);
5205
5206         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5207                 return -EBADR;
5208
5209         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5210                 error = ext4_xattr_fiemap(inode, fieinfo);
5211         } else {
5212                 ext4_lblk_t len_blks;
5213                 __u64 last_blk;
5214
5215                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5216                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5217                 if (last_blk >= EXT_MAX_BLOCKS)
5218                         last_blk = EXT_MAX_BLOCKS-1;
5219                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5220
5221                 /*
5222                  * Walk the extent tree gathering extent information
5223                  * and pushing extents back to the user.
5224                  */
5225                 error = ext4_fill_fiemap_extents(inode, start_blk,
5226                                                  len_blks, fieinfo);
5227         }
5228         return error;
5229 }
5230
5231 /*
5232  * ext4_access_path:
5233  * Function to access the path buffer for marking it dirty.
5234  * It also checks if there are sufficient credits left in the journal handle
5235  * to update path.
5236  */
5237 static int
5238 ext4_access_path(handle_t *handle, struct inode *inode,
5239                 struct ext4_ext_path *path)
5240 {
5241         int credits, err;
5242
5243         if (!ext4_handle_valid(handle))
5244                 return 0;
5245
5246         /*
5247          * Check if need to extend journal credits
5248          * 3 for leaf, sb, and inode plus 2 (bmap and group
5249          * descriptor) for each block group; assume two block
5250          * groups
5251          */
5252         if (handle->h_buffer_credits < 7) {
5253                 credits = ext4_writepage_trans_blocks(inode);
5254                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5255                 /* EAGAIN is success */
5256                 if (err && err != -EAGAIN)
5257                         return err;
5258         }
5259
5260         err = ext4_ext_get_access(handle, inode, path);
5261         return err;
5262 }
5263
5264 /*
5265  * ext4_ext_shift_path_extents:
5266  * Shift the extents of a path structure lying between path[depth].p_ext
5267  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5268  * if it is right shift or left shift operation.
5269  */
5270 static int
5271 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5272                             struct inode *inode, handle_t *handle,
5273                             enum SHIFT_DIRECTION SHIFT)
5274 {
5275         int depth, err = 0;
5276         struct ext4_extent *ex_start, *ex_last;
5277         bool update = 0;
5278         depth = path->p_depth;
5279
5280         while (depth >= 0) {
5281                 if (depth == path->p_depth) {
5282                         ex_start = path[depth].p_ext;
5283                         if (!ex_start)
5284                                 return -EFSCORRUPTED;
5285
5286                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5287
5288                         err = ext4_access_path(handle, inode, path + depth);
5289                         if (err)
5290                                 goto out;
5291
5292                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5293                                 update = 1;
5294
5295                         while (ex_start <= ex_last) {
5296                                 if (SHIFT == SHIFT_LEFT) {
5297                                         le32_add_cpu(&ex_start->ee_block,
5298                                                 -shift);
5299                                         /* Try to merge to the left. */
5300                                         if ((ex_start >
5301                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5302                                             &&
5303                                             ext4_ext_try_to_merge_right(inode,
5304                                             path, ex_start - 1))
5305                                                 ex_last--;
5306                                         else
5307                                                 ex_start++;
5308                                 } else {
5309                                         le32_add_cpu(&ex_last->ee_block, shift);
5310                                         ext4_ext_try_to_merge_right(inode, path,
5311                                                 ex_last);
5312                                         ex_last--;
5313                                 }
5314                         }
5315                         err = ext4_ext_dirty(handle, inode, path + depth);
5316                         if (err)
5317                                 goto out;
5318
5319                         if (--depth < 0 || !update)
5320                                 break;
5321                 }
5322
5323                 /* Update index too */
5324                 err = ext4_access_path(handle, inode, path + depth);
5325                 if (err)
5326                         goto out;
5327
5328                 if (SHIFT == SHIFT_LEFT)
5329                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5330                 else
5331                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5332                 err = ext4_ext_dirty(handle, inode, path + depth);
5333                 if (err)
5334                         goto out;
5335
5336                 /* we are done if current index is not a starting index */
5337                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5338                         break;
5339
5340                 depth--;
5341         }
5342
5343 out:
5344         return err;
5345 }
5346
5347 /*
5348  * ext4_ext_shift_extents:
5349  * All the extents which lies in the range from @start to the last allocated
5350  * block for the @inode are shifted either towards left or right (depending
5351  * upon @SHIFT) by @shift blocks.
5352  * On success, 0 is returned, error otherwise.
5353  */
5354 static int
5355 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5356                        ext4_lblk_t start, ext4_lblk_t shift,
5357                        enum SHIFT_DIRECTION SHIFT)
5358 {
5359         struct ext4_ext_path *path;
5360         int ret = 0, depth;
5361         struct ext4_extent *extent;
5362         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5363
5364         /* Let path point to the last extent */
5365         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5366                                 EXT4_EX_NOCACHE);
5367         if (IS_ERR(path))
5368                 return PTR_ERR(path);
5369
5370         depth = path->p_depth;
5371         extent = path[depth].p_ext;
5372         if (!extent)
5373                 goto out;
5374
5375         stop = le32_to_cpu(extent->ee_block);
5376
5377        /*
5378         * For left shifts, make sure the hole on the left is big enough to
5379         * accommodate the shift.  For right shifts, make sure the last extent
5380         * won't be shifted beyond EXT_MAX_BLOCKS.
5381         */
5382         if (SHIFT == SHIFT_LEFT) {
5383                 path = ext4_find_extent(inode, start - 1, &path,
5384                                         EXT4_EX_NOCACHE);
5385                 if (IS_ERR(path))
5386                         return PTR_ERR(path);
5387                 depth = path->p_depth;
5388                 extent =  path[depth].p_ext;
5389                 if (extent) {
5390                         ex_start = le32_to_cpu(extent->ee_block);
5391                         ex_end = le32_to_cpu(extent->ee_block) +
5392                                 ext4_ext_get_actual_len(extent);
5393                 } else {
5394                         ex_start = 0;
5395                         ex_end = 0;
5396                 }
5397
5398                 if ((start == ex_start && shift > ex_start) ||
5399                     (shift > start - ex_end)) {
5400                         ret = -EINVAL;
5401                         goto out;
5402                 }
5403         } else {
5404                 if (shift > EXT_MAX_BLOCKS -
5405                     (stop + ext4_ext_get_actual_len(extent))) {
5406                         ret = -EINVAL;
5407                         goto out;
5408                 }
5409         }
5410
5411         /*
5412          * In case of left shift, iterator points to start and it is increased
5413          * till we reach stop. In case of right shift, iterator points to stop
5414          * and it is decreased till we reach start.
5415          */
5416         if (SHIFT == SHIFT_LEFT)
5417                 iterator = &start;
5418         else
5419                 iterator = &stop;
5420
5421         /*
5422          * Its safe to start updating extents.  Start and stop are unsigned, so
5423          * in case of right shift if extent with 0 block is reached, iterator
5424          * becomes NULL to indicate the end of the loop.
5425          */
5426         while (iterator && start <= stop) {
5427                 path = ext4_find_extent(inode, *iterator, &path,
5428                                         EXT4_EX_NOCACHE);
5429                 if (IS_ERR(path))
5430                         return PTR_ERR(path);
5431                 depth = path->p_depth;
5432                 extent = path[depth].p_ext;
5433                 if (!extent) {
5434                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5435                                          (unsigned long) *iterator);
5436                         return -EFSCORRUPTED;
5437                 }
5438                 if (SHIFT == SHIFT_LEFT && *iterator >
5439                     le32_to_cpu(extent->ee_block)) {
5440                         /* Hole, move to the next extent */
5441                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5442                                 path[depth].p_ext++;
5443                         } else {
5444                                 *iterator = ext4_ext_next_allocated_block(path);
5445                                 continue;
5446                         }
5447                 }
5448
5449                 if (SHIFT == SHIFT_LEFT) {
5450                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5451                         *iterator = le32_to_cpu(extent->ee_block) +
5452                                         ext4_ext_get_actual_len(extent);
5453                 } else {
5454                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5455                         if (le32_to_cpu(extent->ee_block) > 0)
5456                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5457                         else
5458                                 /* Beginning is reached, end of the loop */
5459                                 iterator = NULL;
5460                         /* Update path extent in case we need to stop */
5461                         while (le32_to_cpu(extent->ee_block) < start)
5462                                 extent++;
5463                         path[depth].p_ext = extent;
5464                 }
5465                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5466                                 handle, SHIFT);
5467                 if (ret)
5468                         break;
5469         }
5470 out:
5471         ext4_ext_drop_refs(path);
5472         kfree(path);
5473         return ret;
5474 }
5475
5476 /*
5477  * ext4_collapse_range:
5478  * This implements the fallocate's collapse range functionality for ext4
5479  * Returns: 0 and non-zero on error.
5480  */
5481 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5482 {
5483         struct super_block *sb = inode->i_sb;
5484         ext4_lblk_t punch_start, punch_stop;
5485         handle_t *handle;
5486         unsigned int credits;
5487         loff_t new_size, ioffset;
5488         int ret;
5489
5490         /*
5491          * We need to test this early because xfstests assumes that a
5492          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5493          * system does not support collapse range.
5494          */
5495         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5496                 return -EOPNOTSUPP;
5497
5498         /* Collapse range works only on fs block size aligned offsets. */
5499         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5500             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5501                 return -EINVAL;
5502
5503         if (!S_ISREG(inode->i_mode))
5504                 return -EINVAL;
5505
5506         trace_ext4_collapse_range(inode, offset, len);
5507
5508         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5509         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5510
5511         /* Call ext4_force_commit to flush all data in case of data=journal. */
5512         if (ext4_should_journal_data(inode)) {
5513                 ret = ext4_force_commit(inode->i_sb);
5514                 if (ret)
5515                         return ret;
5516         }
5517
5518         inode_lock(inode);
5519         /*
5520          * There is no need to overlap collapse range with EOF, in which case
5521          * it is effectively a truncate operation
5522          */
5523         if (offset + len >= i_size_read(inode)) {
5524                 ret = -EINVAL;
5525                 goto out_mutex;
5526         }
5527
5528         /* Currently just for extent based files */
5529         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5530                 ret = -EOPNOTSUPP;
5531                 goto out_mutex;
5532         }
5533
5534         /* Wait for existing dio to complete */
5535         inode_dio_wait(inode);
5536
5537         /*
5538          * Prevent page faults from reinstantiating pages we have released from
5539          * page cache.
5540          */
5541         down_write(&EXT4_I(inode)->i_mmap_sem);
5542
5543         ret = ext4_break_layouts(inode);
5544         if (ret)
5545                 goto out_mmap;
5546
5547         /*
5548          * Need to round down offset to be aligned with page size boundary
5549          * for page size > block size.
5550          */
5551         ioffset = round_down(offset, PAGE_SIZE);
5552         /*
5553          * Write tail of the last page before removed range since it will get
5554          * removed from the page cache below.
5555          */
5556         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5557         if (ret)
5558                 goto out_mmap;
5559         /*
5560          * Write data that will be shifted to preserve them when discarding
5561          * page cache below. We are also protected from pages becoming dirty
5562          * by i_mmap_sem.
5563          */
5564         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5565                                            LLONG_MAX);
5566         if (ret)
5567                 goto out_mmap;
5568         truncate_pagecache(inode, ioffset);
5569
5570         credits = ext4_writepage_trans_blocks(inode);
5571         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5572         if (IS_ERR(handle)) {
5573                 ret = PTR_ERR(handle);
5574                 goto out_mmap;
5575         }
5576
5577         down_write(&EXT4_I(inode)->i_data_sem);
5578         ext4_discard_preallocations(inode);
5579
5580         ret = ext4_es_remove_extent(inode, punch_start,
5581                                     EXT_MAX_BLOCKS - punch_start);
5582         if (ret) {
5583                 up_write(&EXT4_I(inode)->i_data_sem);
5584                 goto out_stop;
5585         }
5586
5587         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5588         if (ret) {
5589                 up_write(&EXT4_I(inode)->i_data_sem);
5590                 goto out_stop;
5591         }
5592         ext4_discard_preallocations(inode);
5593
5594         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5595                                      punch_stop - punch_start, SHIFT_LEFT);
5596         if (ret) {
5597                 up_write(&EXT4_I(inode)->i_data_sem);
5598                 goto out_stop;
5599         }
5600
5601         new_size = i_size_read(inode) - len;
5602         i_size_write(inode, new_size);
5603         EXT4_I(inode)->i_disksize = new_size;
5604
5605         up_write(&EXT4_I(inode)->i_data_sem);
5606         if (IS_SYNC(inode))
5607                 ext4_handle_sync(handle);
5608         inode->i_mtime = inode->i_ctime = current_time(inode);
5609         ext4_mark_inode_dirty(handle, inode);
5610         ext4_update_inode_fsync_trans(handle, inode, 1);
5611
5612 out_stop:
5613         ext4_journal_stop(handle);
5614 out_mmap:
5615         up_write(&EXT4_I(inode)->i_mmap_sem);
5616 out_mutex:
5617         inode_unlock(inode);
5618         return ret;
5619 }
5620
5621 /*
5622  * ext4_insert_range:
5623  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5624  * The data blocks starting from @offset to the EOF are shifted by @len
5625  * towards right to create a hole in the @inode. Inode size is increased
5626  * by len bytes.
5627  * Returns 0 on success, error otherwise.
5628  */
5629 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5630 {
5631         struct super_block *sb = inode->i_sb;
5632         handle_t *handle;
5633         struct ext4_ext_path *path;
5634         struct ext4_extent *extent;
5635         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5636         unsigned int credits, ee_len;
5637         int ret = 0, depth, split_flag = 0;
5638         loff_t ioffset;
5639
5640         /*
5641          * We need to test this early because xfstests assumes that an
5642          * insert range of (0, 1) will return EOPNOTSUPP if the file
5643          * system does not support insert range.
5644          */
5645         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5646                 return -EOPNOTSUPP;
5647
5648         /* Insert range works only on fs block size aligned offsets. */
5649         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5650                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5651                 return -EINVAL;
5652
5653         if (!S_ISREG(inode->i_mode))
5654                 return -EOPNOTSUPP;
5655
5656         trace_ext4_insert_range(inode, offset, len);
5657
5658         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5659         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5660
5661         /* Call ext4_force_commit to flush all data in case of data=journal */
5662         if (ext4_should_journal_data(inode)) {
5663                 ret = ext4_force_commit(inode->i_sb);
5664                 if (ret)
5665                         return ret;
5666         }
5667
5668         inode_lock(inode);
5669         /* Currently just for extent based files */
5670         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5671                 ret = -EOPNOTSUPP;
5672                 goto out_mutex;
5673         }
5674
5675         /* Check for wrap through zero */
5676         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5677                 ret = -EFBIG;
5678                 goto out_mutex;
5679         }
5680
5681         /* Offset should be less than i_size */
5682         if (offset >= i_size_read(inode)) {
5683                 ret = -EINVAL;
5684                 goto out_mutex;
5685         }
5686
5687         /* Wait for existing dio to complete */
5688         inode_dio_wait(inode);
5689
5690         /*
5691          * Prevent page faults from reinstantiating pages we have released from
5692          * page cache.
5693          */
5694         down_write(&EXT4_I(inode)->i_mmap_sem);
5695
5696         ret = ext4_break_layouts(inode);
5697         if (ret)
5698                 goto out_mmap;
5699
5700         /*
5701          * Need to round down to align start offset to page size boundary
5702          * for page size > block size.
5703          */
5704         ioffset = round_down(offset, PAGE_SIZE);
5705         /* Write out all dirty pages */
5706         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5707                         LLONG_MAX);
5708         if (ret)
5709                 goto out_mmap;
5710         truncate_pagecache(inode, ioffset);
5711
5712         credits = ext4_writepage_trans_blocks(inode);
5713         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5714         if (IS_ERR(handle)) {
5715                 ret = PTR_ERR(handle);
5716                 goto out_mmap;
5717         }
5718
5719         /* Expand file to avoid data loss if there is error while shifting */
5720         inode->i_size += len;
5721         EXT4_I(inode)->i_disksize += len;
5722         inode->i_mtime = inode->i_ctime = current_time(inode);
5723         ret = ext4_mark_inode_dirty(handle, inode);
5724         if (ret)
5725                 goto out_stop;
5726
5727         down_write(&EXT4_I(inode)->i_data_sem);
5728         ext4_discard_preallocations(inode);
5729
5730         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5731         if (IS_ERR(path)) {
5732                 up_write(&EXT4_I(inode)->i_data_sem);
5733                 goto out_stop;
5734         }
5735
5736         depth = ext_depth(inode);
5737         extent = path[depth].p_ext;
5738         if (extent) {
5739                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5740                 ee_len = ext4_ext_get_actual_len(extent);
5741
5742                 /*
5743                  * If offset_lblk is not the starting block of extent, split
5744                  * the extent @offset_lblk
5745                  */
5746                 if ((offset_lblk > ee_start_lblk) &&
5747                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5748                         if (ext4_ext_is_unwritten(extent))
5749                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5750                                         EXT4_EXT_MARK_UNWRIT2;
5751                         ret = ext4_split_extent_at(handle, inode, &path,
5752                                         offset_lblk, split_flag,
5753                                         EXT4_EX_NOCACHE |
5754                                         EXT4_GET_BLOCKS_PRE_IO |
5755                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5756                 }
5757
5758                 ext4_ext_drop_refs(path);
5759                 kfree(path);
5760                 if (ret < 0) {
5761                         up_write(&EXT4_I(inode)->i_data_sem);
5762                         goto out_stop;
5763                 }
5764         } else {
5765                 ext4_ext_drop_refs(path);
5766                 kfree(path);
5767         }
5768
5769         ret = ext4_es_remove_extent(inode, offset_lblk,
5770                         EXT_MAX_BLOCKS - offset_lblk);
5771         if (ret) {
5772                 up_write(&EXT4_I(inode)->i_data_sem);
5773                 goto out_stop;
5774         }
5775
5776         /*
5777          * if offset_lblk lies in a hole which is at start of file, use
5778          * ee_start_lblk to shift extents
5779          */
5780         ret = ext4_ext_shift_extents(inode, handle,
5781                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5782                 len_lblk, SHIFT_RIGHT);
5783
5784         up_write(&EXT4_I(inode)->i_data_sem);
5785         if (IS_SYNC(inode))
5786                 ext4_handle_sync(handle);
5787         if (ret >= 0)
5788                 ext4_update_inode_fsync_trans(handle, inode, 1);
5789
5790 out_stop:
5791         ext4_journal_stop(handle);
5792 out_mmap:
5793         up_write(&EXT4_I(inode)->i_mmap_sem);
5794 out_mutex:
5795         inode_unlock(inode);
5796         return ret;
5797 }
5798
5799 /**
5800  * ext4_swap_extents - Swap extents between two inodes
5801  *
5802  * @inode1:     First inode
5803  * @inode2:     Second inode
5804  * @lblk1:      Start block for first inode
5805  * @lblk2:      Start block for second inode
5806  * @count:      Number of blocks to swap
5807  * @unwritten: Mark second inode's extents as unwritten after swap
5808  * @erp:        Pointer to save error value
5809  *
5810  * This helper routine does exactly what is promise "swap extents". All other
5811  * stuff such as page-cache locking consistency, bh mapping consistency or
5812  * extent's data copying must be performed by caller.
5813  * Locking:
5814  *              i_mutex is held for both inodes
5815  *              i_data_sem is locked for write for both inodes
5816  * Assumptions:
5817  *              All pages from requested range are locked for both inodes
5818  */
5819 int
5820 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5821                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5822                   ext4_lblk_t count, int unwritten, int *erp)
5823 {
5824         struct ext4_ext_path *path1 = NULL;
5825         struct ext4_ext_path *path2 = NULL;
5826         int replaced_count = 0;
5827
5828         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5829         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5830         BUG_ON(!inode_is_locked(inode1));
5831         BUG_ON(!inode_is_locked(inode2));
5832
5833         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5834         if (unlikely(*erp))
5835                 return 0;
5836         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5837         if (unlikely(*erp))
5838                 return 0;
5839
5840         while (count) {
5841                 struct ext4_extent *ex1, *ex2, tmp_ex;
5842                 ext4_lblk_t e1_blk, e2_blk;
5843                 int e1_len, e2_len, len;
5844                 int split = 0;
5845
5846                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5847                 if (IS_ERR(path1)) {
5848                         *erp = PTR_ERR(path1);
5849                         path1 = NULL;
5850                 finish:
5851                         count = 0;
5852                         goto repeat;
5853                 }
5854                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5855                 if (IS_ERR(path2)) {
5856                         *erp = PTR_ERR(path2);
5857                         path2 = NULL;
5858                         goto finish;
5859                 }
5860                 ex1 = path1[path1->p_depth].p_ext;
5861                 ex2 = path2[path2->p_depth].p_ext;
5862                 /* Do we have somthing to swap ? */
5863                 if (unlikely(!ex2 || !ex1))
5864                         goto finish;
5865
5866                 e1_blk = le32_to_cpu(ex1->ee_block);
5867                 e2_blk = le32_to_cpu(ex2->ee_block);
5868                 e1_len = ext4_ext_get_actual_len(ex1);
5869                 e2_len = ext4_ext_get_actual_len(ex2);
5870
5871                 /* Hole handling */
5872                 if (!in_range(lblk1, e1_blk, e1_len) ||
5873                     !in_range(lblk2, e2_blk, e2_len)) {
5874                         ext4_lblk_t next1, next2;
5875
5876                         /* if hole after extent, then go to next extent */
5877                         next1 = ext4_ext_next_allocated_block(path1);
5878                         next2 = ext4_ext_next_allocated_block(path2);
5879                         /* If hole before extent, then shift to that extent */
5880                         if (e1_blk > lblk1)
5881                                 next1 = e1_blk;
5882                         if (e2_blk > lblk2)
5883                                 next2 = e2_blk;
5884                         /* Do we have something to swap */
5885                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5886                                 goto finish;
5887                         /* Move to the rightest boundary */
5888                         len = next1 - lblk1;
5889                         if (len < next2 - lblk2)
5890                                 len = next2 - lblk2;
5891                         if (len > count)
5892                                 len = count;
5893                         lblk1 += len;
5894                         lblk2 += len;
5895                         count -= len;
5896                         goto repeat;
5897                 }
5898
5899                 /* Prepare left boundary */
5900                 if (e1_blk < lblk1) {
5901                         split = 1;
5902                         *erp = ext4_force_split_extent_at(handle, inode1,
5903                                                 &path1, lblk1, 0);
5904                         if (unlikely(*erp))
5905                                 goto finish;
5906                 }
5907                 if (e2_blk < lblk2) {
5908                         split = 1;
5909                         *erp = ext4_force_split_extent_at(handle, inode2,
5910                                                 &path2,  lblk2, 0);
5911                         if (unlikely(*erp))
5912                                 goto finish;
5913                 }
5914                 /* ext4_split_extent_at() may result in leaf extent split,
5915                  * path must to be revalidated. */
5916                 if (split)
5917                         goto repeat;
5918
5919                 /* Prepare right boundary */
5920                 len = count;
5921                 if (len > e1_blk + e1_len - lblk1)
5922                         len = e1_blk + e1_len - lblk1;
5923                 if (len > e2_blk + e2_len - lblk2)
5924                         len = e2_blk + e2_len - lblk2;
5925
5926                 if (len != e1_len) {
5927                         split = 1;
5928                         *erp = ext4_force_split_extent_at(handle, inode1,
5929                                                 &path1, lblk1 + len, 0);
5930                         if (unlikely(*erp))
5931                                 goto finish;
5932                 }
5933                 if (len != e2_len) {
5934                         split = 1;
5935                         *erp = ext4_force_split_extent_at(handle, inode2,
5936                                                 &path2, lblk2 + len, 0);
5937                         if (*erp)
5938                                 goto finish;
5939                 }
5940                 /* ext4_split_extent_at() may result in leaf extent split,
5941                  * path must to be revalidated. */
5942                 if (split)
5943                         goto repeat;
5944
5945                 BUG_ON(e2_len != e1_len);
5946                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5947                 if (unlikely(*erp))
5948                         goto finish;
5949                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5950                 if (unlikely(*erp))
5951                         goto finish;
5952
5953                 /* Both extents are fully inside boundaries. Swap it now */
5954                 tmp_ex = *ex1;
5955                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5956                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5957                 ex1->ee_len = cpu_to_le16(e2_len);
5958                 ex2->ee_len = cpu_to_le16(e1_len);
5959                 if (unwritten)
5960                         ext4_ext_mark_unwritten(ex2);
5961                 if (ext4_ext_is_unwritten(&tmp_ex))
5962                         ext4_ext_mark_unwritten(ex1);
5963
5964                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5965                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5966                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5967                                       path2->p_depth);
5968                 if (unlikely(*erp))
5969                         goto finish;
5970                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5971                                       path1->p_depth);
5972                 /*
5973                  * Looks scarry ah..? second inode already points to new blocks,
5974                  * and it was successfully dirtied. But luckily error may happen
5975                  * only due to journal error, so full transaction will be
5976                  * aborted anyway.
5977                  */
5978                 if (unlikely(*erp))
5979                         goto finish;
5980                 lblk1 += len;
5981                 lblk2 += len;
5982                 replaced_count += len;
5983                 count -= len;
5984
5985         repeat:
5986                 ext4_ext_drop_refs(path1);
5987                 kfree(path1);
5988                 ext4_ext_drop_refs(path2);
5989                 kfree(path2);
5990                 path1 = path2 = NULL;
5991         }
5992         return replaced_count;
5993 }