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_data_block_valid(EXT4_SB(inode->i_sb), 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_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 389 }
 390
 391 static int ext4_valid_extent_entries(struct inode *inode,
 392                                 struct ext4_extent_header *eh,
 393                                 int depth)
 394 {
 395         unsigned short entries;
 396         if (eh->eh_entries == 0)
 397                 return 1;
 398
 399         entries = le16_to_cpu(eh->eh_entries);
 400
 401         if (depth == 0) {
 402                 /* leaf entries */
 403                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 404                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 405                 ext4_fsblk_t pblock = 0;
 406                 ext4_lblk_t lblock = 0;
 407                 ext4_lblk_t prev = 0;
 408                 int len = 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                         len = ext4_ext_get_actual_len(ext);
 416                         if ((lblock <= prev) && prev) {
 417                                 pblock = ext4_ext_pblock(ext);
 418                                 es->s_last_error_block = cpu_to_le64(pblock);
 419                                 return 0;
 420                         }
 421                         ext++;
 422                         entries--;
 423                         prev = lblock + len - 1;
 424                 }
 425         } else {
 426                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 427                 while (entries) {
 428                         if (!ext4_valid_extent_idx(inode, ext_idx))
 429                                 return 0;
 430                         ext_idx++;
 431                         entries--;
 432                 }
 433         }
 434         return 1;
 435 }
 436
 437 static int __ext4_ext_check(const char *function, unsigned int line,
 438                             struct inode *inode, struct ext4_extent_header *eh,
 439                             int depth, ext4_fsblk_t pblk)
 440 {
 441         const char *error_msg;
 442         int max = 0, err = -EFSCORRUPTED;
 443
 444         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 445                 error_msg = "invalid magic";
 446                 goto corrupted;
 447         }
 448         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 449                 error_msg = "unexpected eh_depth";
 450                 goto corrupted;
 451         }
 452         if (unlikely(eh->eh_max == 0)) {
 453                 error_msg = "invalid eh_max";
 454                 goto corrupted;
 455         }
 456         max = ext4_ext_max_entries(inode, depth);
 457         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 458                 error_msg = "too large eh_max";
 459                 goto corrupted;
 460         }
 461         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 462                 error_msg = "invalid eh_entries";
 463                 goto corrupted;
 464         }
 465         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 466                 error_msg = "invalid extent entries";
 467                 goto corrupted;
 468         }
 469         if (unlikely(depth > 32)) {
 470                 error_msg = "too large eh_depth";
 471                 goto corrupted;
 472         }
 473         /* Verify checksum on non-root extent tree nodes */
 474         if (ext_depth(inode) != depth &&
 475             !ext4_extent_block_csum_verify(inode, eh)) {
 476                 error_msg = "extent tree corrupted";
 477                 err = -EFSBADCRC;
 478                 goto corrupted;
 479         }
 480         return 0;
 481
 482 corrupted:
 483         ext4_error_inode(inode, function, line, 0,
 484                          "pblk %llu bad header/extent: %s - magic %x, "
 485                          "entries %u, max %u(%u), depth %u(%u)",
 486                          (unsigned long long) pblk, error_msg,
 487                          le16_to_cpu(eh->eh_magic),
 488                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 489                          max, le16_to_cpu(eh->eh_depth), depth);
 490         return err;
 491 }
 492
 493 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 494         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 495
 496 int ext4_ext_check_inode(struct inode *inode)
 497 {
 498         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 499 }
 500
 501 static struct buffer_head *
 502 __read_extent_tree_block(const char *function, unsigned int line,
 503                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 504                          int flags)
 505 {
 506         struct buffer_head              *bh;
 507         int                             err;
 508
 509         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
 510         if (unlikely(!bh))
 511                 return ERR_PTR(-ENOMEM);
 512
 513         if (!bh_uptodate_or_lock(bh)) {
 514                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 515                 err = bh_submit_read(bh);
 516                 if (err < 0)
 517                         goto errout;
 518         }
 519         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 520                 return bh;
 521         if (!ext4_has_feature_journal(inode->i_sb) ||
 522             (inode->i_ino !=
 523              le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum))) {
 524                 err = __ext4_ext_check(function, line, inode,
 525                                        ext_block_hdr(bh), depth, pblk);
 526                 if (err)
 527                         goto errout;
 528         }
 529         set_buffer_verified(bh);
 530         /*
 531          * If this is a leaf block, cache all of its entries
 532          */
 533         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 534                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 535                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 536                 ext4_lblk_t prev = 0;
 537                 int i;
 538
 539                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 540                         unsigned int status = EXTENT_STATUS_WRITTEN;
 541                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 542                         int len = ext4_ext_get_actual_len(ex);
 543
 544                         if (prev && (prev != lblk))
 545                                 ext4_es_cache_extent(inode, prev,
 546                                                      lblk - prev, ~0,
 547                                                      EXTENT_STATUS_HOLE);
 548
 549                         if (ext4_ext_is_unwritten(ex))
 550                                 status = EXTENT_STATUS_UNWRITTEN;
 551                         ext4_es_cache_extent(inode, lblk, len,
 552                                              ext4_ext_pblock(ex), status);
 553                         prev = lblk + len;
 554                 }
 555         }
 556         return bh;
 557 errout:
 558         put_bh(bh);
 559         return ERR_PTR(err);
 560
 561 }
 562
 563 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 564         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 565                                  (depth), (flags))
 566
 567 /*
 568  * This function is called to cache a file's extent information in the
 569  * extent status tree
 570  */
 571 int ext4_ext_precache(struct inode *inode)
 572 {
 573         struct ext4_inode_info *ei = EXT4_I(inode);
 574         struct ext4_ext_path *path = NULL;
 575         struct buffer_head *bh;
 576         int i = 0, depth, ret = 0;
 577
 578         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 579                 return 0;       /* not an extent-mapped inode */
 580
 581         down_read(&ei->i_data_sem);
 582         depth = ext_depth(inode);
 583
 584         path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
 585                        GFP_NOFS);
 586         if (path == NULL) {
 587                 up_read(&ei->i_data_sem);
 588                 return -ENOMEM;
 589         }
 590
 591         /* Don't cache anything if there are no external extent blocks */
 592         if (depth == 0)
 593                 goto out;
 594         path[0].p_hdr = ext_inode_hdr(inode);
 595         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 596         if (ret)
 597                 goto out;
 598         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 599         while (i >= 0) {
 600                 /*
 601                  * If this is a leaf block or we've reached the end of
 602                  * the index block, go up
 603                  */
 604                 if ((i == depth) ||
 605                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 606                         brelse(path[i].p_bh);
 607                         path[i].p_bh = NULL;
 608                         i--;
 609                         continue;
 610                 }
 611                 bh = read_extent_tree_block(inode,
 612                                             ext4_idx_pblock(path[i].p_idx++),
 613                                             depth - i - 1,
 614                                             EXT4_EX_FORCE_CACHE);
 615                 if (IS_ERR(bh)) {
 616                         ret = PTR_ERR(bh);
 617                         break;
 618                 }
 619                 i++;
 620                 path[i].p_bh = bh;
 621                 path[i].p_hdr = ext_block_hdr(bh);
 622                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 623         }
 624         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 625 out:
 626         up_read(&ei->i_data_sem);
 627         ext4_ext_drop_refs(path);
 628         kfree(path);
 629         return ret;
 630 }
 631
 632 #ifdef EXT_DEBUG
 633 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 634 {
 635         int k, l = path->p_depth;
 636
 637         ext_debug("path:");
 638         for (k = 0; k <= l; k++, path++) {
 639                 if (path->p_idx) {
 640                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 641                             ext4_idx_pblock(path->p_idx));
 642                 } else if (path->p_ext) {
 643                         ext_debug("  %d:[%d]%d:%llu ",
 644                                   le32_to_cpu(path->p_ext->ee_block),
 645                                   ext4_ext_is_unwritten(path->p_ext),
 646                                   ext4_ext_get_actual_len(path->p_ext),
 647                                   ext4_ext_pblock(path->p_ext));
 648                 } else
 649                         ext_debug("  []");
 650         }
 651         ext_debug("\n");
 652 }
 653
 654 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 655 {
 656         int depth = ext_depth(inode);
 657         struct ext4_extent_header *eh;
 658         struct ext4_extent *ex;
 659         int i;
 660
 661         if (!path)
 662                 return;
 663
 664         eh = path[depth].p_hdr;
 665         ex = EXT_FIRST_EXTENT(eh);
 666
 667         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 668
 669         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 670                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 671                           ext4_ext_is_unwritten(ex),
 672                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 673         }
 674         ext_debug("\n");
 675 }
 676
 677 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 678                         ext4_fsblk_t newblock, int level)
 679 {
 680         int depth = ext_depth(inode);
 681         struct ext4_extent *ex;
 682
 683         if (depth != level) {
 684                 struct ext4_extent_idx *idx;
 685                 idx = path[level].p_idx;
 686                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 687                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 688                                         le32_to_cpu(idx->ei_block),
 689                                         ext4_idx_pblock(idx),
 690                                         newblock);
 691                         idx++;
 692                 }
 693
 694                 return;
 695         }
 696
 697         ex = path[depth].p_ext;
 698         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 699                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 700                                 le32_to_cpu(ex->ee_block),
 701                                 ext4_ext_pblock(ex),
 702                                 ext4_ext_is_unwritten(ex),
 703                                 ext4_ext_get_actual_len(ex),
 704                                 newblock);
 705                 ex++;
 706         }
 707 }
 708
 709 #else
 710 #define ext4_ext_show_path(inode, path)
 711 #define ext4_ext_show_leaf(inode, path)
 712 #define ext4_ext_show_move(inode, path, newblock, level)
 713 #endif
 714
 715 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 716 {
 717         int depth, i;
 718
 719         if (!path)
 720                 return;
 721         depth = path->p_depth;
 722         for (i = 0; i <= depth; i++, path++)
 723                 if (path->p_bh) {
 724                         brelse(path->p_bh);
 725                         path->p_bh = NULL;
 726                 }
 727 }
 728
 729 /*
 730  * ext4_ext_binsearch_idx:
 731  * binary search for the closest index of the given block
 732  * the header must be checked before calling this
 733  */
 734 static void
 735 ext4_ext_binsearch_idx(struct inode *inode,
 736                         struct ext4_ext_path *path, ext4_lblk_t block)
 737 {
 738         struct ext4_extent_header *eh = path->p_hdr;
 739         struct ext4_extent_idx *r, *l, *m;
 740
 741
 742         ext_debug("binsearch for %u(idx):  ", block);
 743
 744         l = EXT_FIRST_INDEX(eh) + 1;
 745         r = EXT_LAST_INDEX(eh);
 746         while (l <= r) {
 747                 m = l + (r - l) / 2;
 748                 if (block < le32_to_cpu(m->ei_block))
 749                         r = m - 1;
 750                 else
 751                         l = m + 1;
 752                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 753                                 m, le32_to_cpu(m->ei_block),
 754                                 r, le32_to_cpu(r->ei_block));
 755         }
 756
 757         path->p_idx = l - 1;
 758         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 759                   ext4_idx_pblock(path->p_idx));
 760
 761 #ifdef CHECK_BINSEARCH
 762         {
 763                 struct ext4_extent_idx *chix, *ix;
 764                 int k;
 765
 766                 chix = ix = EXT_FIRST_INDEX(eh);
 767                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 768                   if (k != 0 &&
 769                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 770                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 771                                        "first=0x%p\n", k,
 772                                        ix, EXT_FIRST_INDEX(eh));
 773                                 printk(KERN_DEBUG "%u <= %u\n",
 774                                        le32_to_cpu(ix->ei_block),
 775                                        le32_to_cpu(ix[-1].ei_block));
 776                         }
 777                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 778                                            <= le32_to_cpu(ix[-1].ei_block));
 779                         if (block < le32_to_cpu(ix->ei_block))
 780                                 break;
 781                         chix = ix;
 782                 }
 783                 BUG_ON(chix != path->p_idx);
 784         }
 785 #endif
 786
 787 }
 788
 789 /*
 790  * ext4_ext_binsearch:
 791  * binary search for closest extent of the given block
 792  * the header must be checked before calling this
 793  */
 794 static void
 795 ext4_ext_binsearch(struct inode *inode,
 796                 struct ext4_ext_path *path, ext4_lblk_t block)
 797 {
 798         struct ext4_extent_header *eh = path->p_hdr;
 799         struct ext4_extent *r, *l, *m;
 800
 801         if (eh->eh_entries == 0) {
 802                 /*
 803                  * this leaf is empty:
 804                  * we get such a leaf in split/add case
 805                  */
 806                 return;
 807         }
 808
 809         ext_debug("binsearch for %u:  ", block);
 810
 811         l = EXT_FIRST_EXTENT(eh) + 1;
 812         r = EXT_LAST_EXTENT(eh);
 813
 814         while (l <= r) {
 815                 m = l + (r - l) / 2;
 816                 if (block < le32_to_cpu(m->ee_block))
 817                         r = m - 1;
 818                 else
 819                         l = m + 1;
 820                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 821                                 m, le32_to_cpu(m->ee_block),
 822                                 r, le32_to_cpu(r->ee_block));
 823         }
 824
 825         path->p_ext = l - 1;
 826         ext_debug("  -> %d:%llu:[%d]%d ",
 827                         le32_to_cpu(path->p_ext->ee_block),
 828                         ext4_ext_pblock(path->p_ext),
 829                         ext4_ext_is_unwritten(path->p_ext),
 830                         ext4_ext_get_actual_len(path->p_ext));
 831
 832 #ifdef CHECK_BINSEARCH
 833         {
 834                 struct ext4_extent *chex, *ex;
 835                 int k;
 836
 837                 chex = ex = EXT_FIRST_EXTENT(eh);
 838                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 839                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 840                                           <= le32_to_cpu(ex[-1].ee_block));
 841                         if (block < le32_to_cpu(ex->ee_block))
 842                                 break;
 843                         chex = ex;
 844                 }
 845                 BUG_ON(chex != path->p_ext);
 846         }
 847 #endif
 848
 849 }
 850
 851 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 852 {
 853         struct ext4_extent_header *eh;
 854
 855         eh = ext_inode_hdr(inode);
 856         eh->eh_depth = 0;
 857         eh->eh_entries = 0;
 858         eh->eh_magic = EXT4_EXT_MAGIC;
 859         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 860         ext4_mark_inode_dirty(handle, inode);
 861         return 0;
 862 }
 863
 864 struct ext4_ext_path *
 865 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 866                  struct ext4_ext_path **orig_path, int flags)
 867 {
 868         struct ext4_extent_header *eh;
 869         struct buffer_head *bh;
 870         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 871         short int depth, i, ppos = 0;
 872         int ret;
 873
 874         eh = ext_inode_hdr(inode);
 875         depth = ext_depth(inode);
 876         if (depth < 0 || depth > EXT4_MAX_EXTENT_DEPTH) {
 877                 EXT4_ERROR_INODE(inode, "inode has invalid extent depth: %d",
 878                                  depth);
 879                 ret = -EFSCORRUPTED;
 880                 goto err;
 881         }
 882
 883         if (path) {
 884                 ext4_ext_drop_refs(path);
 885                 if (depth > path[0].p_maxdepth) {
 886                         kfree(path);
 887                         *orig_path = path = NULL;
 888                 }
 889         }
 890         if (!path) {
 891                 /* account possible depth increase */
 892                 path = kcalloc(depth + 2, sizeof(struct ext4_ext_path),
 893                                 GFP_NOFS);
 894                 if (unlikely(!path))
 895                         return ERR_PTR(-ENOMEM);
 896                 path[0].p_maxdepth = depth + 1;
 897         }
 898         path[0].p_hdr = eh;
 899         path[0].p_bh = NULL;
 900
 901         i = depth;
 902         /* walk through the tree */
 903         while (i) {
 904                 ext_debug("depth %d: num %d, max %d\n",
 905                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 906
 907                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 908                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 909                 path[ppos].p_depth = i;
 910                 path[ppos].p_ext = NULL;
 911
 912                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 913                                             flags);
 914                 if (IS_ERR(bh)) {
 915                         ret = PTR_ERR(bh);
 916                         goto err;
 917                 }
 918
 919                 eh = ext_block_hdr(bh);
 920                 ppos++;
 921                 path[ppos].p_bh = bh;
 922                 path[ppos].p_hdr = eh;
 923         }
 924
 925         path[ppos].p_depth = i;
 926         path[ppos].p_ext = NULL;
 927         path[ppos].p_idx = NULL;
 928
 929         /* find extent */
 930         ext4_ext_binsearch(inode, path + ppos, block);
 931         /* if not an empty leaf */
 932         if (path[ppos].p_ext)
 933                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 934
 935         ext4_ext_show_path(inode, path);
 936
 937         return path;
 938
 939 err:
 940         ext4_ext_drop_refs(path);
 941         kfree(path);
 942         if (orig_path)
 943                 *orig_path = NULL;
 944         return ERR_PTR(ret);
 945 }
 946
 947 /*
 948  * ext4_ext_insert_index:
 949  * insert new index [@logical;@ptr] into the block at @curp;
 950  * check where to insert: before @curp or after @curp
 951  */
 952 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 953                                  struct ext4_ext_path *curp,
 954                                  int logical, ext4_fsblk_t ptr)
 955 {
 956         struct ext4_extent_idx *ix;
 957         int len, err;
 958
 959         err = ext4_ext_get_access(handle, inode, curp);
 960         if (err)
 961                 return err;
 962
 963         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 964                 EXT4_ERROR_INODE(inode,
 965                                  "logical %d == ei_block %d!",
 966                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 967                 return -EFSCORRUPTED;
 968         }
 969
 970         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 971                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 972                 EXT4_ERROR_INODE(inode,
 973                                  "eh_entries %d >= eh_max %d!",
 974                                  le16_to_cpu(curp->p_hdr->eh_entries),
 975                                  le16_to_cpu(curp->p_hdr->eh_max));
 976                 return -EFSCORRUPTED;
 977         }
 978
 979         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 980                 /* insert after */
 981                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 982                 ix = curp->p_idx + 1;
 983         } else {
 984                 /* insert before */
 985                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 986                 ix = curp->p_idx;
 987         }
 988
 989         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 990         BUG_ON(len < 0);
 991         if (len > 0) {
 992                 ext_debug("insert new index %d: "
 993                                 "move %d indices from 0x%p to 0x%p\n",
 994                                 logical, len, ix, ix + 1);
 995                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 996         }
 997
 998         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 999                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
1000                 return -EFSCORRUPTED;
1001         }
1002
1003         ix->ei_block = cpu_to_le32(logical);
1004         ext4_idx_store_pblock(ix, ptr);
1005         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1006
1007         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1008                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1009                 return -EFSCORRUPTED;
1010         }
1011
1012         err = ext4_ext_dirty(handle, inode, curp);
1013         ext4_std_error(inode->i_sb, err);
1014
1015         return err;
1016 }
1017
1018 /*
1019  * ext4_ext_split:
1020  * inserts new subtree into the path, using free index entry
1021  * at depth @at:
1022  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1023  * - makes decision where to split
1024  * - moves remaining extents and index entries (right to the split point)
1025  *   into the newly allocated blocks
1026  * - initializes subtree
1027  */
1028 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1029                           unsigned int flags,
1030                           struct ext4_ext_path *path,
1031                           struct ext4_extent *newext, int at)
1032 {
1033         struct buffer_head *bh = NULL;
1034         int depth = ext_depth(inode);
1035         struct ext4_extent_header *neh;
1036         struct ext4_extent_idx *fidx;
1037         int i = at, k, m, a;
1038         ext4_fsblk_t newblock, oldblock;
1039         __le32 border;
1040         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1041         int err = 0;
1042         size_t ext_size = 0;
1043
1044         /* make decision: where to split? */
1045         /* FIXME: now decision is simplest: at current extent */
1046
1047         /* if current leaf will be split, then we should use
1048          * border from split point */
1049         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1050                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1051                 return -EFSCORRUPTED;
1052         }
1053         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1054                 border = path[depth].p_ext[1].ee_block;
1055                 ext_debug("leaf will be split."
1056                                 " next leaf starts at %d\n",
1057                                   le32_to_cpu(border));
1058         } else {
1059                 border = newext->ee_block;
1060                 ext_debug("leaf will be added."
1061                                 " next leaf starts at %d\n",
1062                                 le32_to_cpu(border));
1063         }
1064
1065         /*
1066          * If error occurs, then we break processing
1067          * and mark filesystem read-only. index won't
1068          * be inserted and tree will be in consistent
1069          * state. Next mount will repair buffers too.
1070          */
1071
1072         /*
1073          * Get array to track all allocated blocks.
1074          * We need this to handle errors and free blocks
1075          * upon them.
1076          */
1077         ablocks = kcalloc(depth, sizeof(ext4_fsblk_t), GFP_NOFS);
1078         if (!ablocks)
1079                 return -ENOMEM;
1080
1081         /* allocate all needed blocks */
1082         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1083         for (a = 0; a < depth - at; a++) {
1084                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1085                                                    newext, &err, flags);
1086                 if (newblock == 0)
1087                         goto cleanup;
1088                 ablocks[a] = newblock;
1089         }
1090
1091         /* initialize new leaf */
1092         newblock = ablocks[--a];
1093         if (unlikely(newblock == 0)) {
1094                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1095                 err = -EFSCORRUPTED;
1096                 goto cleanup;
1097         }
1098         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1099         if (unlikely(!bh)) {
1100                 err = -ENOMEM;
1101                 goto cleanup;
1102         }
1103         lock_buffer(bh);
1104
1105         err = ext4_journal_get_create_access(handle, bh);
1106         if (err)
1107                 goto cleanup;
1108
1109         neh = ext_block_hdr(bh);
1110         neh->eh_entries = 0;
1111         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1112         neh->eh_magic = EXT4_EXT_MAGIC;
1113         neh->eh_depth = 0;
1114
1115         /* move remainder of path[depth] to the new leaf */
1116         if (unlikely(path[depth].p_hdr->eh_entries !=
1117                      path[depth].p_hdr->eh_max)) {
1118                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1119                                  path[depth].p_hdr->eh_entries,
1120                                  path[depth].p_hdr->eh_max);
1121                 err = -EFSCORRUPTED;
1122                 goto cleanup;
1123         }
1124         /* start copy from next extent */
1125         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1126         ext4_ext_show_move(inode, path, newblock, depth);
1127         if (m) {
1128                 struct ext4_extent *ex;
1129                 ex = EXT_FIRST_EXTENT(neh);
1130                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1131                 le16_add_cpu(&neh->eh_entries, m);
1132         }
1133
1134         /* zero out unused area in the extent block */
1135         ext_size = sizeof(struct ext4_extent_header) +
1136                 sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries);
1137         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1138         ext4_extent_block_csum_set(inode, neh);
1139         set_buffer_uptodate(bh);
1140         unlock_buffer(bh);
1141
1142         err = ext4_handle_dirty_metadata(handle, inode, bh);
1143         if (err)
1144                 goto cleanup;
1145         brelse(bh);
1146         bh = NULL;
1147
1148         /* correct old leaf */
1149         if (m) {
1150                 err = ext4_ext_get_access(handle, inode, path + depth);
1151                 if (err)
1152                         goto cleanup;
1153                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1154                 err = ext4_ext_dirty(handle, inode, path + depth);
1155                 if (err)
1156                         goto cleanup;
1157
1158         }
1159
1160         /* create intermediate indexes */
1161         k = depth - at - 1;
1162         if (unlikely(k < 0)) {
1163                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1164                 err = -EFSCORRUPTED;
1165                 goto cleanup;
1166         }
1167         if (k)
1168                 ext_debug("create %d intermediate indices\n", k);
1169         /* insert new index into current index block */
1170         /* current depth stored in i var */
1171         i = depth - 1;
1172         while (k--) {
1173                 oldblock = newblock;
1174                 newblock = ablocks[--a];
1175                 bh = sb_getblk(inode->i_sb, newblock);
1176                 if (unlikely(!bh)) {
1177                         err = -ENOMEM;
1178                         goto cleanup;
1179                 }
1180                 lock_buffer(bh);
1181
1182                 err = ext4_journal_get_create_access(handle, bh);
1183                 if (err)
1184                         goto cleanup;
1185
1186                 neh = ext_block_hdr(bh);
1187                 neh->eh_entries = cpu_to_le16(1);
1188                 neh->eh_magic = EXT4_EXT_MAGIC;
1189                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1190                 neh->eh_depth = cpu_to_le16(depth - i);
1191                 fidx = EXT_FIRST_INDEX(neh);
1192                 fidx->ei_block = border;
1193                 ext4_idx_store_pblock(fidx, oldblock);
1194
1195                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1196                                 i, newblock, le32_to_cpu(border), oldblock);
1197
1198                 /* move remainder of path[i] to the new index block */
1199                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1200                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1201                         EXT4_ERROR_INODE(inode,
1202                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1203                                          le32_to_cpu(path[i].p_ext->ee_block));
1204                         err = -EFSCORRUPTED;
1205                         goto cleanup;
1206                 }
1207                 /* start copy indexes */
1208                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1209                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1210                                 EXT_MAX_INDEX(path[i].p_hdr));
1211                 ext4_ext_show_move(inode, path, newblock, i);
1212                 if (m) {
1213                         memmove(++fidx, path[i].p_idx,
1214                                 sizeof(struct ext4_extent_idx) * m);
1215                         le16_add_cpu(&neh->eh_entries, m);
1216                 }
1217                 /* zero out unused area in the extent block */
1218                 ext_size = sizeof(struct ext4_extent_header) +
1219                    (sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries));
1220                 memset(bh->b_data + ext_size, 0,
1221                         inode->i_sb->s_blocksize - ext_size);
1222                 ext4_extent_block_csum_set(inode, neh);
1223                 set_buffer_uptodate(bh);
1224                 unlock_buffer(bh);
1225
1226                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1227                 if (err)
1228                         goto cleanup;
1229                 brelse(bh);
1230                 bh = NULL;
1231
1232                 /* correct old index */
1233                 if (m) {
1234                         err = ext4_ext_get_access(handle, inode, path + i);
1235                         if (err)
1236                                 goto cleanup;
1237                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1238                         err = ext4_ext_dirty(handle, inode, path + i);
1239                         if (err)
1240                                 goto cleanup;
1241                 }
1242
1243                 i--;
1244         }
1245
1246         /* insert new index */
1247         err = ext4_ext_insert_index(handle, inode, path + at,
1248                                     le32_to_cpu(border), newblock);
1249
1250 cleanup:
1251         if (bh) {
1252                 if (buffer_locked(bh))
1253                         unlock_buffer(bh);
1254                 brelse(bh);
1255         }
1256
1257         if (err) {
1258                 /* free all allocated blocks in error case */
1259                 for (i = 0; i < depth; i++) {
1260                         if (!ablocks[i])
1261                                 continue;
1262                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1263                                          EXT4_FREE_BLOCKS_METADATA);
1264                 }
1265         }
1266         kfree(ablocks);
1267
1268         return err;
1269 }
1270
1271 /*
1272  * ext4_ext_grow_indepth:
1273  * implements tree growing procedure:
1274  * - allocates new block
1275  * - moves top-level data (index block or leaf) into the new block
1276  * - initializes new top-level, creating index that points to the
1277  *   just created block
1278  */
1279 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1280                                  unsigned int flags)
1281 {
1282         struct ext4_extent_header *neh;
1283         struct buffer_head *bh;
1284         ext4_fsblk_t newblock, goal = 0;
1285         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1286         int err = 0;
1287         size_t ext_size = 0;
1288
1289         /* Try to prepend new index to old one */
1290         if (ext_depth(inode))
1291                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1292         if (goal > le32_to_cpu(es->s_first_data_block)) {
1293                 flags |= EXT4_MB_HINT_TRY_GOAL;
1294                 goal--;
1295         } else
1296                 goal = ext4_inode_to_goal_block(inode);
1297         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1298                                         NULL, &err);
1299         if (newblock == 0)
1300                 return err;
1301
1302         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1303         if (unlikely(!bh))
1304                 return -ENOMEM;
1305         lock_buffer(bh);
1306
1307         err = ext4_journal_get_create_access(handle, bh);
1308         if (err) {
1309                 unlock_buffer(bh);
1310                 goto out;
1311         }
1312
1313         ext_size = sizeof(EXT4_I(inode)->i_data);
1314         /* move top-level index/leaf into new block */
1315         memmove(bh->b_data, EXT4_I(inode)->i_data, ext_size);
1316         /* zero out unused area in the extent block */
1317         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1318
1319         /* set size of new block */
1320         neh = ext_block_hdr(bh);
1321         /* old root could have indexes or leaves
1322          * so calculate e_max right way */
1323         if (ext_depth(inode))
1324                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1325         else
1326                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1327         neh->eh_magic = EXT4_EXT_MAGIC;
1328         ext4_extent_block_csum_set(inode, neh);
1329         set_buffer_uptodate(bh);
1330         unlock_buffer(bh);
1331
1332         err = ext4_handle_dirty_metadata(handle, inode, bh);
1333         if (err)
1334                 goto out;
1335
1336         /* Update top-level index: num,max,pointer */
1337         neh = ext_inode_hdr(inode);
1338         neh->eh_entries = cpu_to_le16(1);
1339         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1340         if (neh->eh_depth == 0) {
1341                 /* Root extent block becomes index block */
1342                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1343                 EXT_FIRST_INDEX(neh)->ei_block =
1344                         EXT_FIRST_EXTENT(neh)->ee_block;
1345         }
1346         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1347                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1348                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1349                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1350
1351         le16_add_cpu(&neh->eh_depth, 1);
1352         ext4_mark_inode_dirty(handle, inode);
1353 out:
1354         brelse(bh);
1355
1356         return err;
1357 }
1358
1359 /*
1360  * ext4_ext_create_new_leaf:
1361  * finds empty index and adds new leaf.
1362  * if no free index is found, then it requests in-depth growing.
1363  */
1364 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1365                                     unsigned int mb_flags,
1366                                     unsigned int gb_flags,
1367                                     struct ext4_ext_path **ppath,
1368                                     struct ext4_extent *newext)
1369 {
1370         struct ext4_ext_path *path = *ppath;
1371         struct ext4_ext_path *curp;
1372         int depth, i, err = 0;
1373
1374 repeat:
1375         i = depth = ext_depth(inode);
1376
1377         /* walk up to the tree and look for free index entry */
1378         curp = path + depth;
1379         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1380                 i--;
1381                 curp--;
1382         }
1383
1384         /* we use already allocated block for index block,
1385          * so subsequent data blocks should be contiguous */
1386         if (EXT_HAS_FREE_INDEX(curp)) {
1387                 /* if we found index with free entry, then use that
1388                  * entry: create all needed subtree and add new leaf */
1389                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1390                 if (err)
1391                         goto out;
1392
1393                 /* refill path */
1394                 path = ext4_find_extent(inode,
1395                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1396                                     ppath, gb_flags);
1397                 if (IS_ERR(path))
1398                         err = PTR_ERR(path);
1399         } else {
1400                 /* tree is full, time to grow in depth */
1401                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1402                 if (err)
1403                         goto out;
1404
1405                 /* refill path */
1406                 path = ext4_find_extent(inode,
1407                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1408                                     ppath, gb_flags);
1409                 if (IS_ERR(path)) {
1410                         err = PTR_ERR(path);
1411                         goto out;
1412                 }
1413
1414                 /*
1415                  * only first (depth 0 -> 1) produces free space;
1416                  * in all other cases we have to split the grown tree
1417                  */
1418                 depth = ext_depth(inode);
1419                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1420                         /* now we need to split */
1421                         goto repeat;
1422                 }
1423         }
1424
1425 out:
1426         return err;
1427 }
1428
1429 /*
1430  * search the closest allocated block to the left for *logical
1431  * and returns it at @logical + it's physical address at @phys
1432  * if *logical is the smallest allocated block, the function
1433  * returns 0 at @phys
1434  * return value contains 0 (success) or error code
1435  */
1436 static int ext4_ext_search_left(struct inode *inode,
1437                                 struct ext4_ext_path *path,
1438                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1439 {
1440         struct ext4_extent_idx *ix;
1441         struct ext4_extent *ex;
1442         int depth, ee_len;
1443
1444         if (unlikely(path == NULL)) {
1445                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1446                 return -EFSCORRUPTED;
1447         }
1448         depth = path->p_depth;
1449         *phys = 0;
1450
1451         if (depth == 0 && path->p_ext == NULL)
1452                 return 0;
1453
1454         /* usually extent in the path covers blocks smaller
1455          * then *logical, but it can be that extent is the
1456          * first one in the file */
1457
1458         ex = path[depth].p_ext;
1459         ee_len = ext4_ext_get_actual_len(ex);
1460         if (*logical < le32_to_cpu(ex->ee_block)) {
1461                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1462                         EXT4_ERROR_INODE(inode,
1463                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1464                                          *logical, le32_to_cpu(ex->ee_block));
1465                         return -EFSCORRUPTED;
1466                 }
1467                 while (--depth >= 0) {
1468                         ix = path[depth].p_idx;
1469                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1470                                 EXT4_ERROR_INODE(inode,
1471                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1472                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1473                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1474                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1475                                   depth);
1476                                 return -EFSCORRUPTED;
1477                         }
1478                 }
1479                 return 0;
1480         }
1481
1482         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1483                 EXT4_ERROR_INODE(inode,
1484                                  "logical %d < ee_block %d + ee_len %d!",
1485                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1486                 return -EFSCORRUPTED;
1487         }
1488
1489         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1490         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1491         return 0;
1492 }
1493
1494 /*
1495  * search the closest allocated block to the right for *logical
1496  * and returns it at @logical + it's physical address at @phys
1497  * if *logical is the largest allocated block, the function
1498  * returns 0 at @phys
1499  * return value contains 0 (success) or error code
1500  */
1501 static int ext4_ext_search_right(struct inode *inode,
1502                                  struct ext4_ext_path *path,
1503                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1504                                  struct ext4_extent **ret_ex)
1505 {
1506         struct buffer_head *bh = NULL;
1507         struct ext4_extent_header *eh;
1508         struct ext4_extent_idx *ix;
1509         struct ext4_extent *ex;
1510         ext4_fsblk_t block;
1511         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1512         int ee_len;
1513
1514         if (unlikely(path == NULL)) {
1515                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1516                 return -EFSCORRUPTED;
1517         }
1518         depth = path->p_depth;
1519         *phys = 0;
1520
1521         if (depth == 0 && path->p_ext == NULL)
1522                 return 0;
1523
1524         /* usually extent in the path covers blocks smaller
1525          * then *logical, but it can be that extent is the
1526          * first one in the file */
1527
1528         ex = path[depth].p_ext;
1529         ee_len = ext4_ext_get_actual_len(ex);
1530         if (*logical < le32_to_cpu(ex->ee_block)) {
1531                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1532                         EXT4_ERROR_INODE(inode,
1533                                          "first_extent(path[%d].p_hdr) != ex",
1534                                          depth);
1535                         return -EFSCORRUPTED;
1536                 }
1537                 while (--depth >= 0) {
1538                         ix = path[depth].p_idx;
1539                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1540                                 EXT4_ERROR_INODE(inode,
1541                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1542                                                  *logical);
1543                                 return -EFSCORRUPTED;
1544                         }
1545                 }
1546                 goto found_extent;
1547         }
1548
1549         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1550                 EXT4_ERROR_INODE(inode,
1551                                  "logical %d < ee_block %d + ee_len %d!",
1552                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1553                 return -EFSCORRUPTED;
1554         }
1555
1556         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1557                 /* next allocated block in this leaf */
1558                 ex++;
1559                 goto found_extent;
1560         }
1561
1562         /* go up and search for index to the right */
1563         while (--depth >= 0) {
1564                 ix = path[depth].p_idx;
1565                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1566                         goto got_index;
1567         }
1568
1569         /* we've gone up to the root and found no index to the right */
1570         return 0;
1571
1572 got_index:
1573         /* we've found index to the right, let's
1574          * follow it and find the closest allocated
1575          * block to the right */
1576         ix++;
1577         block = ext4_idx_pblock(ix);
1578         while (++depth < path->p_depth) {
1579                 /* subtract from p_depth to get proper eh_depth */
1580                 bh = read_extent_tree_block(inode, block,
1581                                             path->p_depth - depth, 0);
1582                 if (IS_ERR(bh))
1583                         return PTR_ERR(bh);
1584                 eh = ext_block_hdr(bh);
1585                 ix = EXT_FIRST_INDEX(eh);
1586                 block = ext4_idx_pblock(ix);
1587                 put_bh(bh);
1588         }
1589
1590         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1591         if (IS_ERR(bh))
1592                 return PTR_ERR(bh);
1593         eh = ext_block_hdr(bh);
1594         ex = EXT_FIRST_EXTENT(eh);
1595 found_extent:
1596         *logical = le32_to_cpu(ex->ee_block);
1597         *phys = ext4_ext_pblock(ex);
1598         *ret_ex = ex;
1599         if (bh)
1600                 put_bh(bh);
1601         return 0;
1602 }
1603
1604 /*
1605  * ext4_ext_next_allocated_block:
1606  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1607  * NOTE: it considers block number from index entry as
1608  * allocated block. Thus, index entries have to be consistent
1609  * with leaves.
1610  */
1611 ext4_lblk_t
1612 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1613 {
1614         int depth;
1615
1616         BUG_ON(path == NULL);
1617         depth = path->p_depth;
1618
1619         if (depth == 0 && path->p_ext == NULL)
1620                 return EXT_MAX_BLOCKS;
1621
1622         while (depth >= 0) {
1623                 if (depth == path->p_depth) {
1624                         /* leaf */
1625                         if (path[depth].p_ext &&
1626                                 path[depth].p_ext !=
1627                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1628                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1629                 } else {
1630                         /* index */
1631                         if (path[depth].p_idx !=
1632                                         EXT_LAST_INDEX(path[depth].p_hdr))
1633                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1634                 }
1635                 depth--;
1636         }
1637
1638         return EXT_MAX_BLOCKS;
1639 }
1640
1641 /*
1642  * ext4_ext_next_leaf_block:
1643  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1644  */
1645 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1646 {
1647         int depth;
1648
1649         BUG_ON(path == NULL);
1650         depth = path->p_depth;
1651
1652         /* zero-tree has no leaf blocks at all */
1653         if (depth == 0)
1654                 return EXT_MAX_BLOCKS;
1655
1656         /* go to index block */
1657         depth--;
1658
1659         while (depth >= 0) {
1660                 if (path[depth].p_idx !=
1661                                 EXT_LAST_INDEX(path[depth].p_hdr))
1662                         return (ext4_lblk_t)
1663                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1664                 depth--;
1665         }
1666
1667         return EXT_MAX_BLOCKS;
1668 }
1669
1670 /*
1671  * ext4_ext_correct_indexes:
1672  * if leaf gets modified and modified extent is first in the leaf,
1673  * then we have to correct all indexes above.
1674  * TODO: do we need to correct tree in all cases?
1675  */
1676 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1677                                 struct ext4_ext_path *path)
1678 {
1679         struct ext4_extent_header *eh;
1680         int depth = ext_depth(inode);
1681         struct ext4_extent *ex;
1682         __le32 border;
1683         int k, err = 0;
1684
1685         eh = path[depth].p_hdr;
1686         ex = path[depth].p_ext;
1687
1688         if (unlikely(ex == NULL || eh == NULL)) {
1689                 EXT4_ERROR_INODE(inode,
1690                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1691                 return -EFSCORRUPTED;
1692         }
1693
1694         if (depth == 0) {
1695                 /* there is no tree at all */
1696                 return 0;
1697         }
1698
1699         if (ex != EXT_FIRST_EXTENT(eh)) {
1700                 /* we correct tree if first leaf got modified only */
1701                 return 0;
1702         }
1703
1704         /*
1705          * TODO: we need correction if border is smaller than current one
1706          */
1707         k = depth - 1;
1708         border = path[depth].p_ext->ee_block;
1709         err = ext4_ext_get_access(handle, inode, path + k);
1710         if (err)
1711                 return err;
1712         path[k].p_idx->ei_block = border;
1713         err = ext4_ext_dirty(handle, inode, path + k);
1714         if (err)
1715                 return err;
1716
1717         while (k--) {
1718                 /* change all left-side indexes */
1719                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1720                         break;
1721                 err = ext4_ext_get_access(handle, inode, path + k);
1722                 if (err)
1723                         break;
1724                 path[k].p_idx->ei_block = border;
1725                 err = ext4_ext_dirty(handle, inode, path + k);
1726                 if (err)
1727                         break;
1728         }
1729
1730         return err;
1731 }
1732
1733 int
1734 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1735                                 struct ext4_extent *ex2)
1736 {
1737         unsigned short ext1_ee_len, ext2_ee_len;
1738
1739         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1740                 return 0;
1741
1742         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1743         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1744
1745         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1746                         le32_to_cpu(ex2->ee_block))
1747                 return 0;
1748
1749         /*
1750          * To allow future support for preallocated extents to be added
1751          * as an RO_COMPAT feature, refuse to merge to extents if
1752          * this can result in the top bit of ee_len being set.
1753          */
1754         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1755                 return 0;
1756         /*
1757          * The check for IO to unwritten extent is somewhat racy as we
1758          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1759          * dropping i_data_sem. But reserved blocks should save us in that
1760          * case.
1761          */
1762         if (ext4_ext_is_unwritten(ex1) &&
1763             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1764              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1765              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1766                 return 0;
1767 #ifdef AGGRESSIVE_TEST
1768         if (ext1_ee_len >= 4)
1769                 return 0;
1770 #endif
1771
1772         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1773                 return 1;
1774         return 0;
1775 }
1776
1777 /*
1778  * This function tries to merge the "ex" extent to the next extent in the tree.
1779  * It always tries to merge towards right. If you want to merge towards
1780  * left, pass "ex - 1" as argument instead of "ex".
1781  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1782  * 1 if they got merged.
1783  */
1784 static int ext4_ext_try_to_merge_right(struct inode *inode,
1785                                  struct ext4_ext_path *path,
1786                                  struct ext4_extent *ex)
1787 {
1788         struct ext4_extent_header *eh;
1789         unsigned int depth, len;
1790         int merge_done = 0, unwritten;
1791
1792         depth = ext_depth(inode);
1793         BUG_ON(path[depth].p_hdr == NULL);
1794         eh = path[depth].p_hdr;
1795
1796         while (ex < EXT_LAST_EXTENT(eh)) {
1797                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1798                         break;
1799                 /* merge with next extent! */
1800                 unwritten = ext4_ext_is_unwritten(ex);
1801                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1802                                 + ext4_ext_get_actual_len(ex + 1));
1803                 if (unwritten)
1804                         ext4_ext_mark_unwritten(ex);
1805
1806                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1807                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1808                                 * sizeof(struct ext4_extent);
1809                         memmove(ex + 1, ex + 2, len);
1810                 }
1811                 le16_add_cpu(&eh->eh_entries, -1);
1812                 merge_done = 1;
1813                 WARN_ON(eh->eh_entries == 0);
1814                 if (!eh->eh_entries)
1815                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1816         }
1817
1818         return merge_done;
1819 }
1820
1821 /*
1822  * This function does a very simple check to see if we can collapse
1823  * an extent tree with a single extent tree leaf block into the inode.
1824  */
1825 static void ext4_ext_try_to_merge_up(handle_t *handle,
1826                                      struct inode *inode,
1827                                      struct ext4_ext_path *path)
1828 {
1829         size_t s;
1830         unsigned max_root = ext4_ext_space_root(inode, 0);
1831         ext4_fsblk_t blk;
1832
1833         if ((path[0].p_depth != 1) ||
1834             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1835             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1836                 return;
1837
1838         /*
1839          * We need to modify the block allocation bitmap and the block
1840          * group descriptor to release the extent tree block.  If we
1841          * can't get the journal credits, give up.
1842          */
1843         if (ext4_journal_extend(handle, 2))
1844                 return;
1845
1846         /*
1847          * Copy the extent data up to the inode
1848          */
1849         blk = ext4_idx_pblock(path[0].p_idx);
1850         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1851                 sizeof(struct ext4_extent_idx);
1852         s += sizeof(struct ext4_extent_header);
1853
1854         path[1].p_maxdepth = path[0].p_maxdepth;
1855         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1856         path[0].p_depth = 0;
1857         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1858                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1859         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1860
1861         brelse(path[1].p_bh);
1862         ext4_free_blocks(handle, inode, NULL, blk, 1,
1863                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1864 }
1865
1866 /*
1867  * This function tries to merge the @ex extent to neighbours in the tree.
1868  * return 1 if merge left else 0.
1869  */
1870 static void ext4_ext_try_to_merge(handle_t *handle,
1871                                   struct inode *inode,
1872                                   struct ext4_ext_path *path,
1873                                   struct ext4_extent *ex) {
1874         struct ext4_extent_header *eh;
1875         unsigned int depth;
1876         int merge_done = 0;
1877
1878         depth = ext_depth(inode);
1879         BUG_ON(path[depth].p_hdr == NULL);
1880         eh = path[depth].p_hdr;
1881
1882         if (ex > EXT_FIRST_EXTENT(eh))
1883                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1884
1885         if (!merge_done)
1886                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1887
1888         ext4_ext_try_to_merge_up(handle, inode, path);
1889 }
1890
1891 /*
1892  * check if a portion of the "newext" extent overlaps with an
1893  * existing extent.
1894  *
1895  * If there is an overlap discovered, it updates the length of the newext
1896  * such that there will be no overlap, and then returns 1.
1897  * If there is no overlap found, it returns 0.
1898  */
1899 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1900                                            struct inode *inode,
1901                                            struct ext4_extent *newext,
1902                                            struct ext4_ext_path *path)
1903 {
1904         ext4_lblk_t b1, b2;
1905         unsigned int depth, len1;
1906         unsigned int ret = 0;
1907
1908         b1 = le32_to_cpu(newext->ee_block);
1909         len1 = ext4_ext_get_actual_len(newext);
1910         depth = ext_depth(inode);
1911         if (!path[depth].p_ext)
1912                 goto out;
1913         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1914
1915         /*
1916          * get the next allocated block if the extent in the path
1917          * is before the requested block(s)
1918          */
1919         if (b2 < b1) {
1920                 b2 = ext4_ext_next_allocated_block(path);
1921                 if (b2 == EXT_MAX_BLOCKS)
1922                         goto out;
1923                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1924         }
1925
1926         /* check for wrap through zero on extent logical start block*/
1927         if (b1 + len1 < b1) {
1928                 len1 = EXT_MAX_BLOCKS - b1;
1929                 newext->ee_len = cpu_to_le16(len1);
1930                 ret = 1;
1931         }
1932
1933         /* check for overlap */
1934         if (b1 + len1 > b2) {
1935                 newext->ee_len = cpu_to_le16(b2 - b1);
1936                 ret = 1;
1937         }
1938 out:
1939         return ret;
1940 }
1941
1942 /*
1943  * ext4_ext_insert_extent:
1944  * tries to merge requsted extent into the existing extent or
1945  * inserts requested extent as new one into the tree,
1946  * creating new leaf in the no-space case.
1947  */
1948 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1949                                 struct ext4_ext_path **ppath,
1950                                 struct ext4_extent *newext, int gb_flags)
1951 {
1952         struct ext4_ext_path *path = *ppath;
1953         struct ext4_extent_header *eh;
1954         struct ext4_extent *ex, *fex;
1955         struct ext4_extent *nearex; /* nearest extent */
1956         struct ext4_ext_path *npath = NULL;
1957         int depth, len, err;
1958         ext4_lblk_t next;
1959         int mb_flags = 0, unwritten;
1960
1961         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1962                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1963         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1964                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1965                 return -EFSCORRUPTED;
1966         }
1967         depth = ext_depth(inode);
1968         ex = path[depth].p_ext;
1969         eh = path[depth].p_hdr;
1970         if (unlikely(path[depth].p_hdr == NULL)) {
1971                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1972                 return -EFSCORRUPTED;
1973         }
1974
1975         /* try to insert block into found extent and return */
1976         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1977
1978                 /*
1979                  * Try to see whether we should rather test the extent on
1980                  * right from ex, or from the left of ex. This is because
1981                  * ext4_find_extent() can return either extent on the
1982                  * left, or on the right from the searched position. This
1983                  * will make merging more effective.
1984                  */
1985                 if (ex < EXT_LAST_EXTENT(eh) &&
1986                     (le32_to_cpu(ex->ee_block) +
1987                     ext4_ext_get_actual_len(ex) <
1988                     le32_to_cpu(newext->ee_block))) {
1989                         ex += 1;
1990                         goto prepend;
1991                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1992                            (le32_to_cpu(newext->ee_block) +
1993                            ext4_ext_get_actual_len(newext) <
1994                            le32_to_cpu(ex->ee_block)))
1995                         ex -= 1;
1996
1997                 /* Try to append newex to the ex */
1998                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1999                         ext_debug("append [%d]%d block to %u:[%d]%d"
2000                                   "(from %llu)\n",
2001                                   ext4_ext_is_unwritten(newext),
2002                                   ext4_ext_get_actual_len(newext),
2003                                   le32_to_cpu(ex->ee_block),
2004                                   ext4_ext_is_unwritten(ex),
2005                                   ext4_ext_get_actual_len(ex),
2006                                   ext4_ext_pblock(ex));
2007                         err = ext4_ext_get_access(handle, inode,
2008                                                   path + depth);
2009                         if (err)
2010                                 return err;
2011                         unwritten = ext4_ext_is_unwritten(ex);
2012                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2013                                         + ext4_ext_get_actual_len(newext));
2014                         if (unwritten)
2015                                 ext4_ext_mark_unwritten(ex);
2016                         eh = path[depth].p_hdr;
2017                         nearex = ex;
2018                         goto merge;
2019                 }
2020
2021 prepend:
2022                 /* Try to prepend newex to the ex */
2023                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2024                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2025                                   "(from %llu)\n",
2026                                   le32_to_cpu(newext->ee_block),
2027                                   ext4_ext_is_unwritten(newext),
2028                                   ext4_ext_get_actual_len(newext),
2029                                   le32_to_cpu(ex->ee_block),
2030                                   ext4_ext_is_unwritten(ex),
2031                                   ext4_ext_get_actual_len(ex),
2032                                   ext4_ext_pblock(ex));
2033                         err = ext4_ext_get_access(handle, inode,
2034                                                   path + depth);
2035                         if (err)
2036                                 return err;
2037
2038                         unwritten = ext4_ext_is_unwritten(ex);
2039                         ex->ee_block = newext->ee_block;
2040                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2041                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2042                                         + ext4_ext_get_actual_len(newext));
2043                         if (unwritten)
2044                                 ext4_ext_mark_unwritten(ex);
2045                         eh = path[depth].p_hdr;
2046                         nearex = ex;
2047                         goto merge;
2048                 }
2049         }
2050
2051         depth = ext_depth(inode);
2052         eh = path[depth].p_hdr;
2053         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2054                 goto has_space;
2055
2056         /* probably next leaf has space for us? */
2057         fex = EXT_LAST_EXTENT(eh);
2058         next = EXT_MAX_BLOCKS;
2059         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2060                 next = ext4_ext_next_leaf_block(path);
2061         if (next != EXT_MAX_BLOCKS) {
2062                 ext_debug("next leaf block - %u\n", next);
2063                 BUG_ON(npath != NULL);
2064                 npath = ext4_find_extent(inode, next, NULL, 0);
2065                 if (IS_ERR(npath))
2066                         return PTR_ERR(npath);
2067                 BUG_ON(npath->p_depth != path->p_depth);
2068                 eh = npath[depth].p_hdr;
2069                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2070                         ext_debug("next leaf isn't full(%d)\n",
2071                                   le16_to_cpu(eh->eh_entries));
2072                         path = npath;
2073                         goto has_space;
2074                 }
2075                 ext_debug("next leaf has no free space(%d,%d)\n",
2076                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2077         }
2078
2079         /*
2080          * There is no free space in the found leaf.
2081          * We're gonna add a new leaf in the tree.
2082          */
2083         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2084                 mb_flags |= EXT4_MB_USE_RESERVED;
2085         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2086                                        ppath, newext);
2087         if (err)
2088                 goto cleanup;
2089         depth = ext_depth(inode);
2090         eh = path[depth].p_hdr;
2091
2092 has_space:
2093         nearex = path[depth].p_ext;
2094
2095         err = ext4_ext_get_access(handle, inode, path + depth);
2096         if (err)
2097                 goto cleanup;
2098
2099         if (!nearex) {
2100                 /* there is no extent in this leaf, create first one */
2101                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2102                                 le32_to_cpu(newext->ee_block),
2103                                 ext4_ext_pblock(newext),
2104                                 ext4_ext_is_unwritten(newext),
2105                                 ext4_ext_get_actual_len(newext));
2106                 nearex = EXT_FIRST_EXTENT(eh);
2107         } else {
2108                 if (le32_to_cpu(newext->ee_block)
2109                            > le32_to_cpu(nearex->ee_block)) {
2110                         /* Insert after */
2111                         ext_debug("insert %u:%llu:[%d]%d before: "
2112                                         "nearest %p\n",
2113                                         le32_to_cpu(newext->ee_block),
2114                                         ext4_ext_pblock(newext),
2115                                         ext4_ext_is_unwritten(newext),
2116                                         ext4_ext_get_actual_len(newext),
2117                                         nearex);
2118                         nearex++;
2119                 } else {
2120                         /* Insert before */
2121                         BUG_ON(newext->ee_block == nearex->ee_block);
2122                         ext_debug("insert %u:%llu:[%d]%d after: "
2123                                         "nearest %p\n",
2124                                         le32_to_cpu(newext->ee_block),
2125                                         ext4_ext_pblock(newext),
2126                                         ext4_ext_is_unwritten(newext),
2127                                         ext4_ext_get_actual_len(newext),
2128                                         nearex);
2129                 }
2130                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2131                 if (len > 0) {
2132                         ext_debug("insert %u:%llu:[%d]%d: "
2133                                         "move %d extents from 0x%p to 0x%p\n",
2134                                         le32_to_cpu(newext->ee_block),
2135                                         ext4_ext_pblock(newext),
2136                                         ext4_ext_is_unwritten(newext),
2137                                         ext4_ext_get_actual_len(newext),
2138                                         len, nearex, nearex + 1);
2139                         memmove(nearex + 1, nearex,
2140                                 len * sizeof(struct ext4_extent));
2141                 }
2142         }
2143
2144         le16_add_cpu(&eh->eh_entries, 1);
2145         path[depth].p_ext = nearex;
2146         nearex->ee_block = newext->ee_block;
2147         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2148         nearex->ee_len = newext->ee_len;
2149
2150 merge:
2151         /* try to merge extents */
2152         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2153                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2154
2155
2156         /* time to correct all indexes above */
2157         err = ext4_ext_correct_indexes(handle, inode, path);
2158         if (err)
2159                 goto cleanup;
2160
2161         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2162
2163 cleanup:
2164         ext4_ext_drop_refs(npath);
2165         kfree(npath);
2166         return err;
2167 }
2168
2169 static int ext4_fill_fiemap_extents(struct inode *inode,
2170                                     ext4_lblk_t block, ext4_lblk_t num,
2171                                     struct fiemap_extent_info *fieinfo)
2172 {
2173         struct ext4_ext_path *path = NULL;
2174         struct ext4_extent *ex;
2175         struct extent_status es;
2176         ext4_lblk_t next, next_del, start = 0, end = 0;
2177         ext4_lblk_t last = block + num;
2178         int exists, depth = 0, err = 0;
2179         unsigned int flags = 0;
2180         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2181
2182         while (block < last && block != EXT_MAX_BLOCKS) {
2183                 num = last - block;
2184                 /* find extent for this block */
2185                 down_read(&EXT4_I(inode)->i_data_sem);
2186
2187                 path = ext4_find_extent(inode, block, &path, 0);
2188                 if (IS_ERR(path)) {
2189                         up_read(&EXT4_I(inode)->i_data_sem);
2190                         err = PTR_ERR(path);
2191                         path = NULL;
2192                         break;
2193                 }
2194
2195                 depth = ext_depth(inode);
2196                 if (unlikely(path[depth].p_hdr == NULL)) {
2197                         up_read(&EXT4_I(inode)->i_data_sem);
2198                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2199                         err = -EFSCORRUPTED;
2200                         break;
2201                 }
2202                 ex = path[depth].p_ext;
2203                 next = ext4_ext_next_allocated_block(path);
2204
2205                 flags = 0;
2206                 exists = 0;
2207                 if (!ex) {
2208                         /* there is no extent yet, so try to allocate
2209                          * all requested space */
2210                         start = block;
2211                         end = block + num;
2212                 } else if (le32_to_cpu(ex->ee_block) > block) {
2213                         /* need to allocate space before found extent */
2214                         start = block;
2215                         end = le32_to_cpu(ex->ee_block);
2216                         if (block + num < end)
2217                                 end = block + num;
2218                 } else if (block >= le32_to_cpu(ex->ee_block)
2219                                         + ext4_ext_get_actual_len(ex)) {
2220                         /* need to allocate space after found extent */
2221                         start = block;
2222                         end = block + num;
2223                         if (end >= next)
2224                                 end = next;
2225                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2226                         /*
2227                          * some part of requested space is covered
2228                          * by found extent
2229                          */
2230                         start = block;
2231                         end = le32_to_cpu(ex->ee_block)
2232                                 + ext4_ext_get_actual_len(ex);
2233                         if (block + num < end)
2234                                 end = block + num;
2235                         exists = 1;
2236                 } else {
2237                         BUG();
2238                 }
2239                 BUG_ON(end <= start);
2240
2241                 if (!exists) {
2242                         es.es_lblk = start;
2243                         es.es_len = end - start;
2244                         es.es_pblk = 0;
2245                 } else {
2246                         es.es_lblk = le32_to_cpu(ex->ee_block);
2247                         es.es_len = ext4_ext_get_actual_len(ex);
2248                         es.es_pblk = ext4_ext_pblock(ex);
2249                         if (ext4_ext_is_unwritten(ex))
2250                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2251                 }
2252
2253                 /*
2254                  * Find delayed extent and update es accordingly. We call
2255                  * it even in !exists case to find out whether es is the
2256                  * last existing extent or not.
2257                  */
2258                 next_del = ext4_find_delayed_extent(inode, &es);
2259                 if (!exists && next_del) {
2260                         exists = 1;
2261                         flags |= (FIEMAP_EXTENT_DELALLOC |
2262                                   FIEMAP_EXTENT_UNKNOWN);
2263                 }
2264                 up_read(&EXT4_I(inode)->i_data_sem);
2265
2266                 if (unlikely(es.es_len == 0)) {
2267                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2268                         err = -EFSCORRUPTED;
2269                         break;
2270                 }
2271
2272                 /*
2273                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2274                  * we need to check next == EXT_MAX_BLOCKS because it is
2275                  * possible that an extent is with unwritten and delayed
2276                  * status due to when an extent is delayed allocated and
2277                  * is allocated by fallocate status tree will track both of
2278                  * them in a extent.
2279                  *
2280                  * So we could return a unwritten and delayed extent, and
2281                  * its block is equal to 'next'.
2282                  */
2283                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2284                         flags |= FIEMAP_EXTENT_LAST;
2285                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2286                                      next != EXT_MAX_BLOCKS)) {
2287                                 EXT4_ERROR_INODE(inode,
2288                                                  "next extent == %u, next "
2289                                                  "delalloc extent = %u",
2290                                                  next, next_del);
2291                                 err = -EFSCORRUPTED;
2292                                 break;
2293                         }
2294                 }
2295
2296                 if (exists) {
2297                         err = fiemap_fill_next_extent(fieinfo,
2298                                 (__u64)es.es_lblk << blksize_bits,
2299                                 (__u64)es.es_pblk << blksize_bits,
2300                                 (__u64)es.es_len << blksize_bits,
2301                                 flags);
2302                         if (err < 0)
2303                                 break;
2304                         if (err == 1) {
2305                                 err = 0;
2306                                 break;
2307                         }
2308                 }
2309
2310                 block = es.es_lblk + es.es_len;
2311         }
2312
2313         ext4_ext_drop_refs(path);
2314         kfree(path);
2315         return err;
2316 }
2317
2318 static int ext4_fill_es_cache_info(struct inode *inode,
2319                                    ext4_lblk_t block, ext4_lblk_t num,
2320                                    struct fiemap_extent_info *fieinfo)
2321 {
2322         ext4_lblk_t next, end = block + num - 1;
2323         struct extent_status es;
2324         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2325         unsigned int flags;
2326         int err;
2327
2328         while (block <= end) {
2329                 next = 0;
2330                 flags = 0;
2331                 if (!ext4_es_lookup_extent(inode, block, &next, &es))
2332                         break;
2333                 if (ext4_es_is_unwritten(&es))
2334                         flags |= FIEMAP_EXTENT_UNWRITTEN;
2335                 if (ext4_es_is_delayed(&es))
2336                         flags |= (FIEMAP_EXTENT_DELALLOC |
2337                                   FIEMAP_EXTENT_UNKNOWN);
2338                 if (ext4_es_is_hole(&es))
2339                         flags |= EXT4_FIEMAP_EXTENT_HOLE;
2340                 if (next == 0)
2341                         flags |= FIEMAP_EXTENT_LAST;
2342                 if (flags & (FIEMAP_EXTENT_DELALLOC|
2343                              EXT4_FIEMAP_EXTENT_HOLE))
2344                         es.es_pblk = 0;
2345                 else
2346                         es.es_pblk = ext4_es_pblock(&es);
2347                 err = fiemap_fill_next_extent(fieinfo,
2348                                 (__u64)es.es_lblk << blksize_bits,
2349                                 (__u64)es.es_pblk << blksize_bits,
2350                                 (__u64)es.es_len << blksize_bits,
2351                                 flags);
2352                 if (next == 0)
2353                         break;
2354                 block = next;
2355                 if (err < 0)
2356                         return err;
2357                 if (err == 1)
2358                         return 0;
2359         }
2360         return 0;
2361 }
2362
2363
2364 /*
2365  * ext4_ext_determine_hole - determine hole around given block
2366  * @inode:      inode we lookup in
2367  * @path:       path in extent tree to @lblk
2368  * @lblk:       pointer to logical block around which we want to determine hole
2369  *
2370  * Determine hole length (and start if easily possible) around given logical
2371  * block. We don't try too hard to find the beginning of the hole but @path
2372  * actually points to extent before @lblk, we provide it.
2373  *
2374  * The function returns the length of a hole starting at @lblk. We update @lblk
2375  * to the beginning of the hole if we managed to find it.
2376  */
2377 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2378                                            struct ext4_ext_path *path,
2379                                            ext4_lblk_t *lblk)
2380 {
2381         int depth = ext_depth(inode);
2382         struct ext4_extent *ex;
2383         ext4_lblk_t len;
2384
2385         ex = path[depth].p_ext;
2386         if (ex == NULL) {
2387                 /* there is no extent yet, so gap is [0;-] */
2388                 *lblk = 0;
2389                 len = EXT_MAX_BLOCKS;
2390         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2391                 len = le32_to_cpu(ex->ee_block) - *lblk;
2392         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2393                         + ext4_ext_get_actual_len(ex)) {
2394                 ext4_lblk_t next;
2395
2396                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2397                 next = ext4_ext_next_allocated_block(path);
2398                 BUG_ON(next == *lblk);
2399                 len = next - *lblk;
2400         } else {
2401                 BUG();
2402         }
2403         return len;
2404 }
2405
2406 /*
2407  * ext4_ext_put_gap_in_cache:
2408  * calculate boundaries of the gap that the requested block fits into
2409  * and cache this gap
2410  */
2411 static void
2412 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2413                           ext4_lblk_t hole_len)
2414 {
2415         struct extent_status es;
2416
2417         ext4_es_find_extent_range(inode, &ext4_es_is_delayed, hole_start,
2418                                   hole_start + hole_len - 1, &es);
2419         if (es.es_len) {
2420                 /* There's delayed extent containing lblock? */
2421                 if (es.es_lblk <= hole_start)
2422                         return;
2423                 hole_len = min(es.es_lblk - hole_start, hole_len);
2424         }
2425         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2426         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2427                               EXTENT_STATUS_HOLE);
2428 }
2429
2430 /*
2431  * ext4_ext_rm_idx:
2432  * removes index from the index block.
2433  */
2434 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2435                         struct ext4_ext_path *path, int depth)
2436 {
2437         int err;
2438         ext4_fsblk_t leaf;
2439
2440         /* free index block */
2441         depth--;
2442         path = path + depth;
2443         leaf = ext4_idx_pblock(path->p_idx);
2444         if (unlikely(path->p_hdr->eh_entries == 0)) {
2445                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2446                 return -EFSCORRUPTED;
2447         }
2448         err = ext4_ext_get_access(handle, inode, path);
2449         if (err)
2450                 return err;
2451
2452         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2453                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2454                 len *= sizeof(struct ext4_extent_idx);
2455                 memmove(path->p_idx, path->p_idx + 1, len);
2456         }
2457
2458         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2459         err = ext4_ext_dirty(handle, inode, path);
2460         if (err)
2461                 return err;
2462         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2463         trace_ext4_ext_rm_idx(inode, leaf);
2464
2465         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2466                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2467
2468         while (--depth >= 0) {
2469                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2470                         break;
2471                 path--;
2472                 err = ext4_ext_get_access(handle, inode, path);
2473                 if (err)
2474                         break;
2475                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2476                 err = ext4_ext_dirty(handle, inode, path);
2477                 if (err)
2478                         break;
2479         }
2480         return err;
2481 }
2482
2483 /*
2484  * ext4_ext_calc_credits_for_single_extent:
2485  * This routine returns max. credits that needed to insert an extent
2486  * to the extent tree.
2487  * When pass the actual path, the caller should calculate credits
2488  * under i_data_sem.
2489  */
2490 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2491                                                 struct ext4_ext_path *path)
2492 {
2493         if (path) {
2494                 int depth = ext_depth(inode);
2495                 int ret = 0;
2496
2497                 /* probably there is space in leaf? */
2498                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2499                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2500
2501                         /*
2502                          *  There are some space in the leaf tree, no
2503                          *  need to account for leaf block credit
2504                          *
2505                          *  bitmaps and block group descriptor blocks
2506                          *  and other metadata blocks still need to be
2507                          *  accounted.
2508                          */
2509                         /* 1 bitmap, 1 block group descriptor */
2510                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2511                         return ret;
2512                 }
2513         }
2514
2515         return ext4_chunk_trans_blocks(inode, nrblocks);
2516 }
2517
2518 /*
2519  * How many index/leaf blocks need to change/allocate to add @extents extents?
2520  *
2521  * If we add a single extent, then in the worse case, each tree level
2522  * index/leaf need to be changed in case of the tree split.
2523  *
2524  * If more extents are inserted, they could cause the whole tree split more
2525  * than once, but this is really rare.
2526  */
2527 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2528 {
2529         int index;
2530         int depth;
2531
2532         /* If we are converting the inline data, only one is needed here. */
2533         if (ext4_has_inline_data(inode))
2534                 return 1;
2535
2536         depth = ext_depth(inode);
2537
2538         if (extents <= 1)
2539                 index = depth * 2;
2540         else
2541                 index = depth * 3;
2542
2543         return index;
2544 }
2545
2546 static inline int get_default_free_blocks_flags(struct inode *inode)
2547 {
2548         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2549             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2550                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2551         else if (ext4_should_journal_data(inode))
2552                 return EXT4_FREE_BLOCKS_FORGET;
2553         return 0;
2554 }
2555
2556 /*
2557  * ext4_rereserve_cluster - increment the reserved cluster count when
2558  *                          freeing a cluster with a pending reservation
2559  *
2560  * @inode - file containing the cluster
2561  * @lblk - logical block in cluster to be reserved
2562  *
2563  * Increments the reserved cluster count and adjusts quota in a bigalloc
2564  * file system when freeing a partial cluster containing at least one
2565  * delayed and unwritten block.  A partial cluster meeting that
2566  * requirement will have a pending reservation.  If so, the
2567  * RERESERVE_CLUSTER flag is used when calling ext4_free_blocks() to
2568  * defer reserved and allocated space accounting to a subsequent call
2569  * to this function.
2570  */
2571 static void ext4_rereserve_cluster(struct inode *inode, ext4_lblk_t lblk)
2572 {
2573         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2574         struct ext4_inode_info *ei = EXT4_I(inode);
2575
2576         dquot_reclaim_block(inode, EXT4_C2B(sbi, 1));
2577
2578         spin_lock(&ei->i_block_reservation_lock);
2579         ei->i_reserved_data_blocks++;
2580         percpu_counter_add(&sbi->s_dirtyclusters_counter, 1);
2581         spin_unlock(&ei->i_block_reservation_lock);
2582
2583         percpu_counter_add(&sbi->s_freeclusters_counter, 1);
2584         ext4_remove_pending(inode, lblk);
2585 }
2586
2587 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2588                               struct ext4_extent *ex,
2589                               struct partial_cluster *partial,
2590                               ext4_lblk_t from, ext4_lblk_t to)
2591 {
2592         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2593         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2594         ext4_fsblk_t last_pblk, pblk;
2595         ext4_lblk_t num;
2596         int flags;
2597
2598         /* only extent tail removal is allowed */
2599         if (from < le32_to_cpu(ex->ee_block) ||
2600             to != le32_to_cpu(ex->ee_block) + ee_len - 1) {
2601                 ext4_error(sbi->s_sb,
2602                            "strange request: removal(2) %u-%u from %u:%u",
2603                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2604                 return 0;
2605         }
2606
2607 #ifdef EXTENTS_STATS
2608         spin_lock(&sbi->s_ext_stats_lock);
2609         sbi->s_ext_blocks += ee_len;
2610         sbi->s_ext_extents++;
2611         if (ee_len < sbi->s_ext_min)
2612                 sbi->s_ext_min = ee_len;
2613         if (ee_len > sbi->s_ext_max)
2614                 sbi->s_ext_max = ee_len;
2615         if (ext_depth(inode) > sbi->s_depth_max)
2616                 sbi->s_depth_max = ext_depth(inode);
2617         spin_unlock(&sbi->s_ext_stats_lock);
2618 #endif
2619
2620         trace_ext4_remove_blocks(inode, ex, from, to, partial);
2621
2622         /*
2623          * if we have a partial cluster, and it's different from the
2624          * cluster of the last block in the extent, we free it
2625          */
2626         last_pblk = ext4_ext_pblock(ex) + ee_len - 1;
2627
2628         if (partial->state != initial &&
2629             partial->pclu != EXT4_B2C(sbi, last_pblk)) {
2630                 if (partial->state == tofree) {
2631                         flags = get_default_free_blocks_flags(inode);
2632                         if (ext4_is_pending(inode, partial->lblk))
2633                                 flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2634                         ext4_free_blocks(handle, inode, NULL,
2635                                          EXT4_C2B(sbi, partial->pclu),
2636                                          sbi->s_cluster_ratio, flags);
2637                         if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2638                                 ext4_rereserve_cluster(inode, partial->lblk);
2639                 }
2640                 partial->state = initial;
2641         }
2642
2643         num = le32_to_cpu(ex->ee_block) + ee_len - from;
2644         pblk = ext4_ext_pblock(ex) + ee_len - num;
2645
2646         /*
2647          * We free the partial cluster at the end of the extent (if any),
2648          * unless the cluster is used by another extent (partial_cluster
2649          * state is nofree).  If a partial cluster exists here, it must be
2650          * shared with the last block in the extent.
2651          */
2652         flags = get_default_free_blocks_flags(inode);
2653
2654         /* partial, left end cluster aligned, right end unaligned */
2655         if ((EXT4_LBLK_COFF(sbi, to) != sbi->s_cluster_ratio - 1) &&
2656             (EXT4_LBLK_CMASK(sbi, to) >= from) &&
2657             (partial->state != nofree)) {
2658                 if (ext4_is_pending(inode, to))
2659                         flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2660                 ext4_free_blocks(handle, inode, NULL,
2661                                  EXT4_PBLK_CMASK(sbi, last_pblk),
2662                                  sbi->s_cluster_ratio, flags);
2663                 if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2664                         ext4_rereserve_cluster(inode, to);
2665                 partial->state = initial;
2666                 flags = get_default_free_blocks_flags(inode);
2667         }
2668
2669         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2670
2671         /*
2672          * For bigalloc file systems, we never free a partial cluster
2673          * at the beginning of the extent.  Instead, we check to see if we
2674          * need to free it on a subsequent call to ext4_remove_blocks,
2675          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2676          */
2677         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2678         ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2679
2680         /* reset the partial cluster if we've freed past it */
2681         if (partial->state != initial && partial->pclu != EXT4_B2C(sbi, pblk))
2682                 partial->state = initial;
2683
2684         /*
2685          * If we've freed the entire extent but the beginning is not left
2686          * cluster aligned and is not marked as ineligible for freeing we
2687          * record the partial cluster at the beginning of the extent.  It
2688          * wasn't freed by the preceding ext4_free_blocks() call, and we
2689          * need to look farther to the left to determine if it's to be freed
2690          * (not shared with another extent). Else, reset the partial
2691          * cluster - we're either  done freeing or the beginning of the
2692          * extent is left cluster aligned.
2693          */
2694         if (EXT4_LBLK_COFF(sbi, from) && num == ee_len) {
2695                 if (partial->state == initial) {
2696                         partial->pclu = EXT4_B2C(sbi, pblk);
2697                         partial->lblk = from;
2698                         partial->state = tofree;
2699                 }
2700         } else {
2701                 partial->state = initial;
2702         }
2703
2704         return 0;
2705 }
2706
2707 /*
2708  * ext4_ext_rm_leaf() Removes the extents associated with the
2709  * blocks appearing between "start" and "end".  Both "start"
2710  * and "end" must appear in the same extent or EIO is returned.
2711  *
2712  * @handle: The journal handle
2713  * @inode:  The files inode
2714  * @path:   The path to the leaf
2715  * @partial_cluster: The cluster which we'll have to free if all extents
2716  *                   has been released from it.  However, if this value is
2717  *                   negative, it's a cluster just to the right of the
2718  *                   punched region and it must not be freed.
2719  * @start:  The first block to remove
2720  * @end:   The last block to remove
2721  */
2722 static int
2723 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2724                  struct ext4_ext_path *path,
2725                  struct partial_cluster *partial,
2726                  ext4_lblk_t start, ext4_lblk_t end)
2727 {
2728         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2729         int err = 0, correct_index = 0;
2730         int depth = ext_depth(inode), credits;
2731         struct ext4_extent_header *eh;
2732         ext4_lblk_t a, b;
2733         unsigned num;
2734         ext4_lblk_t ex_ee_block;
2735         unsigned short ex_ee_len;
2736         unsigned unwritten = 0;
2737         struct ext4_extent *ex;
2738         ext4_fsblk_t pblk;
2739
2740         /* the header must be checked already in ext4_ext_remove_space() */
2741         ext_debug("truncate since %u in leaf to %u\n", start, end);
2742         if (!path[depth].p_hdr)
2743                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2744         eh = path[depth].p_hdr;
2745         if (unlikely(path[depth].p_hdr == NULL)) {
2746                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2747                 return -EFSCORRUPTED;
2748         }
2749         /* find where to start removing */
2750         ex = path[depth].p_ext;
2751         if (!ex)
2752                 ex = EXT_LAST_EXTENT(eh);
2753
2754         ex_ee_block = le32_to_cpu(ex->ee_block);
2755         ex_ee_len = ext4_ext_get_actual_len(ex);
2756
2757         trace_ext4_ext_rm_leaf(inode, start, ex, partial);
2758
2759         while (ex >= EXT_FIRST_EXTENT(eh) &&
2760                         ex_ee_block + ex_ee_len > start) {
2761
2762                 if (ext4_ext_is_unwritten(ex))
2763                         unwritten = 1;
2764                 else
2765                         unwritten = 0;
2766
2767                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2768                           unwritten, ex_ee_len);
2769                 path[depth].p_ext = ex;
2770
2771                 a = ex_ee_block > start ? ex_ee_block : start;
2772                 b = ex_ee_block+ex_ee_len - 1 < end ?
2773                         ex_ee_block+ex_ee_len - 1 : end;
2774
2775                 ext_debug("  border %u:%u\n", a, b);
2776
2777                 /* If this extent is beyond the end of the hole, skip it */
2778                 if (end < ex_ee_block) {
2779                         /*
2780                          * We're going to skip this extent and move to another,
2781                          * so note that its first cluster is in use to avoid
2782                          * freeing it when removing blocks.  Eventually, the
2783                          * right edge of the truncated/punched region will
2784                          * be just to the left.
2785                          */
2786                         if (sbi->s_cluster_ratio > 1) {
2787                                 pblk = ext4_ext_pblock(ex);
2788                                 partial->pclu = EXT4_B2C(sbi, pblk);
2789                                 partial->state = nofree;
2790                         }
2791                         ex--;
2792                         ex_ee_block = le32_to_cpu(ex->ee_block);
2793                         ex_ee_len = ext4_ext_get_actual_len(ex);
2794                         continue;
2795                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2796                         EXT4_ERROR_INODE(inode,
2797                                          "can not handle truncate %u:%u "
2798                                          "on extent %u:%u",
2799                                          start, end, ex_ee_block,
2800                                          ex_ee_block + ex_ee_len - 1);
2801                         err = -EFSCORRUPTED;
2802                         goto out;
2803                 } else if (a != ex_ee_block) {
2804                         /* remove tail of the extent */
2805                         num = a - ex_ee_block;
2806                 } else {
2807                         /* remove whole extent: excellent! */
2808                         num = 0;
2809                 }
2810                 /*
2811                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2812                  * descriptor) for each block group; assume two block
2813                  * groups plus ex_ee_len/blocks_per_block_group for
2814                  * the worst case
2815                  */
2816                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2817                 if (ex == EXT_FIRST_EXTENT(eh)) {
2818                         correct_index = 1;
2819                         credits += (ext_depth(inode)) + 1;
2820                 }
2821                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2822
2823                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2824                 if (err)
2825                         goto out;
2826
2827                 err = ext4_ext_get_access(handle, inode, path + depth);
2828                 if (err)
2829                         goto out;
2830
2831                 err = ext4_remove_blocks(handle, inode, ex, partial, a, b);
2832                 if (err)
2833                         goto out;
2834
2835                 if (num == 0)
2836                         /* this extent is removed; mark slot entirely unused */
2837                         ext4_ext_store_pblock(ex, 0);
2838
2839                 ex->ee_len = cpu_to_le16(num);
2840                 /*
2841                  * Do not mark unwritten if all the blocks in the
2842                  * extent have been removed.
2843                  */
2844                 if (unwritten && num)
2845                         ext4_ext_mark_unwritten(ex);
2846                 /*
2847                  * If the extent was completely released,
2848                  * we need to remove it from the leaf
2849                  */
2850                 if (num == 0) {
2851                         if (end != EXT_MAX_BLOCKS - 1) {
2852                                 /*
2853                                  * For hole punching, we need to scoot all the
2854                                  * extents up when an extent is removed so that
2855                                  * we dont have blank extents in the middle
2856                                  */
2857                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2858                                         sizeof(struct ext4_extent));
2859
2860                                 /* Now get rid of the one at the end */
2861                                 memset(EXT_LAST_EXTENT(eh), 0,
2862                                         sizeof(struct ext4_extent));
2863                         }
2864                         le16_add_cpu(&eh->eh_entries, -1);
2865                 }
2866
2867                 err = ext4_ext_dirty(handle, inode, path + depth);
2868                 if (err)
2869                         goto out;
2870
2871                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2872                                 ext4_ext_pblock(ex));
2873                 ex--;
2874                 ex_ee_block = le32_to_cpu(ex->ee_block);
2875                 ex_ee_len = ext4_ext_get_actual_len(ex);
2876         }
2877
2878         if (correct_index && eh->eh_entries)
2879                 err = ext4_ext_correct_indexes(handle, inode, path);
2880
2881         /*
2882          * If there's a partial cluster and at least one extent remains in
2883          * the leaf, free the partial cluster if it isn't shared with the
2884          * current extent.  If it is shared with the current extent
2885          * we reset the partial cluster because we've reached the start of the
2886          * truncated/punched region and we're done removing blocks.
2887          */
2888         if (partial->state == tofree && ex >= EXT_FIRST_EXTENT(eh)) {
2889                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2890                 if (partial->pclu != EXT4_B2C(sbi, pblk)) {
2891                         int flags = get_default_free_blocks_flags(inode);
2892
2893                         if (ext4_is_pending(inode, partial->lblk))
2894                                 flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2895                         ext4_free_blocks(handle, inode, NULL,
2896                                          EXT4_C2B(sbi, partial->pclu),
2897                                          sbi->s_cluster_ratio, flags);
2898                         if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2899                                 ext4_rereserve_cluster(inode, partial->lblk);
2900                 }
2901                 partial->state = initial;
2902         }
2903
2904         /* if this leaf is free, then we should
2905          * remove it from index block above */
2906         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2907                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2908
2909 out:
2910         return err;
2911 }
2912
2913 /*
2914  * ext4_ext_more_to_rm:
2915  * returns 1 if current index has to be freed (even partial)
2916  */
2917 static int
2918 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2919 {
2920         BUG_ON(path->p_idx == NULL);
2921
2922         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2923                 return 0;
2924
2925         /*
2926          * if truncate on deeper level happened, it wasn't partial,
2927          * so we have to consider current index for truncation
2928          */
2929         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2930                 return 0;
2931         return 1;
2932 }
2933
2934 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2935                           ext4_lblk_t end)
2936 {
2937         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2938         int depth = ext_depth(inode);
2939         struct ext4_ext_path *path = NULL;
2940         struct partial_cluster partial;
2941         handle_t *handle;
2942         int i = 0, err = 0;
2943
2944         partial.pclu = 0;
2945         partial.lblk = 0;
2946         partial.state = initial;
2947
2948         ext_debug("truncate since %u to %u\n", start, end);
2949
2950         /* probably first extent we're gonna free will be last in block */
2951         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2952         if (IS_ERR(handle))
2953                 return PTR_ERR(handle);
2954
2955 again:
2956         trace_ext4_ext_remove_space(inode, start, end, depth);
2957
2958         /*
2959          * Check if we are removing extents inside the extent tree. If that
2960          * is the case, we are going to punch a hole inside the extent tree
2961          * so we have to check whether we need to split the extent covering
2962          * the last block to remove so we can easily remove the part of it
2963          * in ext4_ext_rm_leaf().
2964          */
2965         if (end < EXT_MAX_BLOCKS - 1) {
2966                 struct ext4_extent *ex;
2967                 ext4_lblk_t ee_block, ex_end, lblk;
2968                 ext4_fsblk_t pblk;
2969
2970                 /* find extent for or closest extent to this block */
2971                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2972                 if (IS_ERR(path)) {
2973                         ext4_journal_stop(handle);
2974                         return PTR_ERR(path);
2975                 }
2976                 depth = ext_depth(inode);
2977                 /* Leaf not may not exist only if inode has no blocks at all */
2978                 ex = path[depth].p_ext;
2979                 if (!ex) {
2980                         if (depth) {
2981                                 EXT4_ERROR_INODE(inode,
2982                                                  "path[%d].p_hdr == NULL",
2983                                                  depth);
2984                                 err = -EFSCORRUPTED;
2985                         }
2986                         goto out;
2987                 }
2988
2989                 ee_block = le32_to_cpu(ex->ee_block);
2990                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2991
2992                 /*
2993                  * See if the last block is inside the extent, if so split
2994                  * the extent at 'end' block so we can easily remove the
2995                  * tail of the first part of the split extent in
2996                  * ext4_ext_rm_leaf().
2997                  */
2998                 if (end >= ee_block && end < ex_end) {
2999
3000                         /*
3001                          * If we're going to split the extent, note that
3002                          * the cluster containing the block after 'end' is
3003                          * in use to avoid freeing it when removing blocks.
3004                          */
3005                         if (sbi->s_cluster_ratio > 1) {
3006                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
3007                                 partial.pclu = EXT4_B2C(sbi, pblk);
3008                                 partial.state = nofree;
3009                         }
3010
3011                         /*
3012                          * Split the extent in two so that 'end' is the last
3013                          * block in the first new extent. Also we should not
3014                          * fail removing space due to ENOSPC so try to use
3015                          * reserved block if that happens.
3016                          */
3017                         err = ext4_force_split_extent_at(handle, inode, &path,
3018                                                          end + 1, 1);
3019                         if (err < 0)
3020                                 goto out;
3021
3022                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end &&
3023                            partial.state == initial) {
3024                         /*
3025                          * If we're punching, there's an extent to the right.
3026                          * If the partial cluster hasn't been set, set it to
3027                          * that extent's first cluster and its state to nofree
3028                          * so it won't be freed should it contain blocks to be
3029                          * removed. If it's already set (tofree/nofree), we're
3030                          * retrying and keep the original partial cluster info
3031                          * so a cluster marked tofree as a result of earlier
3032                          * extent removal is not lost.
3033                          */
3034                         lblk = ex_end + 1;
3035                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
3036                                                     &ex);
3037                         if (err)
3038                                 goto out;
3039                         if (pblk) {
3040                                 partial.pclu = EXT4_B2C(sbi, pblk);
3041                                 partial.state = nofree;
3042                         }
3043                 }
3044         }
3045         /*
3046          * We start scanning from right side, freeing all the blocks
3047          * after i_size and walking into the tree depth-wise.
3048          */
3049         depth = ext_depth(inode);
3050         if (path) {
3051                 int k = i = depth;
3052                 while (--k > 0)
3053                         path[k].p_block =
3054                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
3055         } else {
3056                 path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
3057                                GFP_NOFS);
3058                 if (path == NULL) {
3059                         ext4_journal_stop(handle);
3060                         return -ENOMEM;
3061                 }
3062                 path[0].p_maxdepth = path[0].p_depth = depth;
3063                 path[0].p_hdr = ext_inode_hdr(inode);
3064                 i = 0;
3065
3066                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
3067                         err = -EFSCORRUPTED;
3068                         goto out;
3069                 }
3070         }
3071         err = 0;
3072
3073         while (i >= 0 && err == 0) {
3074                 if (i == depth) {
3075                         /* this is leaf block */
3076                         err = ext4_ext_rm_leaf(handle, inode, path,
3077                                                &partial, start, end);
3078                         /* root level has p_bh == NULL, brelse() eats this */
3079                         brelse(path[i].p_bh);
3080                         path[i].p_bh = NULL;
3081                         i--;
3082                         continue;
3083                 }
3084
3085                 /* this is index block */
3086                 if (!path[i].p_hdr) {
3087                         ext_debug("initialize header\n");
3088                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
3089                 }
3090
3091                 if (!path[i].p_idx) {
3092                         /* this level hasn't been touched yet */
3093                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
3094                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
3095                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
3096                                   path[i].p_hdr,
3097                                   le16_to_cpu(path[i].p_hdr->eh_entries));
3098                 } else {
3099                         /* we were already here, see at next index */
3100                         path[i].p_idx--;
3101                 }
3102
3103                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
3104                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
3105                                 path[i].p_idx);
3106                 if (ext4_ext_more_to_rm(path + i)) {
3107                         struct buffer_head *bh;
3108                         /* go to the next level */
3109                         ext_debug("move to level %d (block %llu)\n",
3110                                   i + 1, ext4_idx_pblock(path[i].p_idx));
3111                         memset(path + i + 1, 0, sizeof(*path));
3112                         bh = read_extent_tree_block(inode,
3113                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
3114                                 EXT4_EX_NOCACHE);
3115                         if (IS_ERR(bh)) {
3116                                 /* should we reset i_size? */
3117                                 err = PTR_ERR(bh);
3118                                 break;
3119                         }
3120                         /* Yield here to deal with large extent trees.
3121                          * Should be a no-op if we did IO above. */
3122                         cond_resched();
3123                         if (WARN_ON(i + 1 > depth)) {
3124                                 err = -EFSCORRUPTED;
3125                                 break;
3126                         }
3127                         path[i + 1].p_bh = bh;
3128
3129                         /* save actual number of indexes since this
3130                          * number is changed at the next iteration */
3131                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3132                         i++;
3133                 } else {
3134                         /* we finished processing this index, go up */
3135                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3136                                 /* index is empty, remove it;
3137                                  * handle must be already prepared by the
3138                                  * truncatei_leaf() */
3139                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3140                         }
3141                         /* root level has p_bh == NULL, brelse() eats this */
3142                         brelse(path[i].p_bh);
3143                         path[i].p_bh = NULL;
3144                         i--;
3145                         ext_debug("return to level %d\n", i);
3146                 }
3147         }
3148
3149         trace_ext4_ext_remove_space_done(inode, start, end, depth, &partial,
3150                                          path->p_hdr->eh_entries);
3151
3152         /*
3153          * if there's a partial cluster and we have removed the first extent
3154          * in the file, then we also free the partial cluster, if any
3155          */
3156         if (partial.state == tofree && err == 0) {
3157                 int flags = get_default_free_blocks_flags(inode);
3158
3159                 if (ext4_is_pending(inode, partial.lblk))
3160                         flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
3161                 ext4_free_blocks(handle, inode, NULL,
3162                                  EXT4_C2B(sbi, partial.pclu),
3163                                  sbi->s_cluster_ratio, flags);
3164                 if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
3165                         ext4_rereserve_cluster(inode, partial.lblk);
3166                 partial.state = initial;
3167         }
3168
3169         /* TODO: flexible tree reduction should be here */
3170         if (path->p_hdr->eh_entries == 0) {
3171                 /*
3172                  * truncate to zero freed all the tree,
3173                  * so we need to correct eh_depth
3174                  */
3175                 err = ext4_ext_get_access(handle, inode, path);
3176                 if (err == 0) {
3177                         ext_inode_hdr(inode)->eh_depth = 0;
3178                         ext_inode_hdr(inode)->eh_max =
3179                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3180                         err = ext4_ext_dirty(handle, inode, path);
3181                 }
3182         }
3183 out:
3184         ext4_ext_drop_refs(path);
3185         kfree(path);
3186         path = NULL;
3187         if (err == -EAGAIN)
3188                 goto again;
3189         ext4_journal_stop(handle);
3190
3191         return err;
3192 }
3193
3194 /*
3195  * called at mount time
3196  */
3197 void ext4_ext_init(struct super_block *sb)
3198 {
3199         /*
3200          * possible initialization would be here
3201          */
3202
3203         if (ext4_has_feature_extents(sb)) {
3204 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3205                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3206 #ifdef AGGRESSIVE_TEST
3207                        ", aggressive tests"
3208 #endif
3209 #ifdef CHECK_BINSEARCH
3210                        ", check binsearch"
3211 #endif
3212 #ifdef EXTENTS_STATS
3213                        ", stats"
3214 #endif
3215                        "\n");
3216 #endif
3217 #ifdef EXTENTS_STATS
3218                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3219                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3220                 EXT4_SB(sb)->s_ext_max = 0;
3221 #endif
3222         }
3223 }
3224
3225 /*
3226  * called at umount time
3227  */
3228 void ext4_ext_release(struct super_block *sb)
3229 {
3230         if (!ext4_has_feature_extents(sb))
3231                 return;
3232
3233 #ifdef EXTENTS_STATS
3234         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3235                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3236                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3237                         sbi->s_ext_blocks, sbi->s_ext_extents,
3238                         sbi->s_ext_blocks / sbi->s_ext_extents);
3239                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3240                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3241         }
3242 #endif
3243 }
3244
3245 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3246 {
3247         ext4_lblk_t  ee_block;
3248         ext4_fsblk_t ee_pblock;
3249         unsigned int ee_len;
3250
3251         ee_block  = le32_to_cpu(ex->ee_block);
3252         ee_len    = ext4_ext_get_actual_len(ex);
3253         ee_pblock = ext4_ext_pblock(ex);
3254
3255         if (ee_len == 0)
3256                 return 0;
3257
3258         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3259                                      EXTENT_STATUS_WRITTEN);
3260 }
3261
3262 /* FIXME!! we need to try to merge to left or right after zero-out  */
3263 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3264 {
3265         ext4_fsblk_t ee_pblock;
3266         unsigned int ee_len;
3267
3268         ee_len    = ext4_ext_get_actual_len(ex);
3269         ee_pblock = ext4_ext_pblock(ex);
3270         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3271                                   ee_len);
3272 }
3273
3274 /*
3275  * ext4_split_extent_at() splits an extent at given block.
3276  *
3277  * @handle: the journal handle
3278  * @inode: the file inode
3279  * @path: the path to the extent
3280  * @split: the logical block where the extent is splitted.
3281  * @split_flags: indicates if the extent could be zeroout if split fails, and
3282  *               the states(init or unwritten) of new extents.
3283  * @flags: flags used to insert new extent to extent tree.
3284  *
3285  *
3286  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3287  * of which are deterimined by split_flag.
3288  *
3289  * There are two cases:
3290  *  a> the extent are splitted into two extent.
3291  *  b> split is not needed, and just mark the extent.
3292  *
3293  * return 0 on success.
3294  */
3295 static int ext4_split_extent_at(handle_t *handle,
3296                              struct inode *inode,
3297                              struct ext4_ext_path **ppath,
3298                              ext4_lblk_t split,
3299                              int split_flag,
3300                              int flags)
3301 {
3302         struct ext4_ext_path *path = *ppath;
3303         ext4_fsblk_t newblock;
3304         ext4_lblk_t ee_block;
3305         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3306         struct ext4_extent *ex2 = NULL;
3307         unsigned int ee_len, depth;
3308         int err = 0;
3309
3310         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3311                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3312
3313         ext_debug("ext4_split_extents_at: inode %lu, logical"
3314                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3315
3316         ext4_ext_show_leaf(inode, path);
3317
3318         depth = ext_depth(inode);
3319         ex = path[depth].p_ext;
3320         ee_block = le32_to_cpu(ex->ee_block);
3321         ee_len = ext4_ext_get_actual_len(ex);
3322         newblock = split - ee_block + ext4_ext_pblock(ex);
3323
3324         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3325         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3326                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3327                              EXT4_EXT_MARK_UNWRIT1 |
3328                              EXT4_EXT_MARK_UNWRIT2));
3329
3330         err = ext4_ext_get_access(handle, inode, path + depth);
3331         if (err)
3332                 goto out;
3333
3334         if (split == ee_block) {
3335                 /*
3336                  * case b: block @split is the block that the extent begins with
3337                  * then we just change the state of the extent, and splitting
3338                  * is not needed.
3339                  */
3340                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3341                         ext4_ext_mark_unwritten(ex);
3342                 else
3343                         ext4_ext_mark_initialized(ex);
3344
3345                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3346                         ext4_ext_try_to_merge(handle, inode, path, ex);
3347
3348                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3349                 goto out;
3350         }
3351
3352         /* case a */
3353         memcpy(&orig_ex, ex, sizeof(orig_ex));
3354         ex->ee_len = cpu_to_le16(split - ee_block);
3355         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3356                 ext4_ext_mark_unwritten(ex);
3357
3358         /*
3359          * path may lead to new leaf, not to original leaf any more
3360          * after ext4_ext_insert_extent() returns,
3361          */
3362         err = ext4_ext_dirty(handle, inode, path + depth);
3363         if (err)
3364                 goto fix_extent_len;
3365
3366         ex2 = &newex;
3367         ex2->ee_block = cpu_to_le32(split);
3368         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3369         ext4_ext_store_pblock(ex2, newblock);
3370         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3371                 ext4_ext_mark_unwritten(ex2);
3372
3373         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3374         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3375                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3376                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3377                                 err = ext4_ext_zeroout(inode, ex2);
3378                                 zero_ex.ee_block = ex2->ee_block;
3379                                 zero_ex.ee_len = cpu_to_le16(
3380                                                 ext4_ext_get_actual_len(ex2));
3381                                 ext4_ext_store_pblock(&zero_ex,
3382                                                       ext4_ext_pblock(ex2));
3383                         } else {
3384                                 err = ext4_ext_zeroout(inode, ex);
3385                                 zero_ex.ee_block = ex->ee_block;
3386                                 zero_ex.ee_len = cpu_to_le16(
3387                                                 ext4_ext_get_actual_len(ex));
3388                                 ext4_ext_store_pblock(&zero_ex,
3389                                                       ext4_ext_pblock(ex));
3390                         }
3391                 } else {
3392                         err = ext4_ext_zeroout(inode, &orig_ex);
3393                         zero_ex.ee_block = orig_ex.ee_block;
3394                         zero_ex.ee_len = cpu_to_le16(
3395                                                 ext4_ext_get_actual_len(&orig_ex));
3396                         ext4_ext_store_pblock(&zero_ex,
3397                                               ext4_ext_pblock(&orig_ex));
3398                 }
3399
3400                 if (err)
3401                         goto fix_extent_len;
3402                 /* update the extent length and mark as initialized */
3403                 ex->ee_len = cpu_to_le16(ee_len);
3404                 ext4_ext_try_to_merge(handle, inode, path, ex);
3405                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3406                 if (err)
3407                         goto fix_extent_len;
3408
3409                 /* update extent status tree */
3410                 err = ext4_zeroout_es(inode, &zero_ex);
3411
3412                 goto out;
3413         } else if (err)
3414                 goto fix_extent_len;
3415
3416 out:
3417         ext4_ext_show_leaf(inode, path);
3418         return err;
3419
3420 fix_extent_len:
3421         ex->ee_len = orig_ex.ee_len;
3422         ext4_ext_dirty(handle, inode, path + path->p_depth);
3423         return err;
3424 }
3425
3426 /*
3427  * ext4_split_extents() splits an extent and mark extent which is covered
3428  * by @map as split_flags indicates
3429  *
3430  * It may result in splitting the extent into multiple extents (up to three)
3431  * There are three possibilities:
3432  *   a> There is no split required
3433  *   b> Splits in two extents: Split is happening at either end of the extent
3434  *   c> Splits in three extents: Somone is splitting in middle of the extent
3435  *
3436  */
3437 static int ext4_split_extent(handle_t *handle,
3438                               struct inode *inode,
3439                               struct ext4_ext_path **ppath,
3440                               struct ext4_map_blocks *map,
3441                               int split_flag,
3442                               int flags)
3443 {
3444         struct ext4_ext_path *path = *ppath;
3445         ext4_lblk_t ee_block;
3446         struct ext4_extent *ex;
3447         unsigned int ee_len, depth;
3448         int err = 0;
3449         int unwritten;
3450         int split_flag1, flags1;
3451         int allocated = map->m_len;
3452
3453         depth = ext_depth(inode);
3454         ex = path[depth].p_ext;
3455         ee_block = le32_to_cpu(ex->ee_block);
3456         ee_len = ext4_ext_get_actual_len(ex);
3457         unwritten = ext4_ext_is_unwritten(ex);
3458
3459         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3460                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3461                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3462                 if (unwritten)
3463                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3464                                        EXT4_EXT_MARK_UNWRIT2;
3465                 if (split_flag & EXT4_EXT_DATA_VALID2)
3466                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3467                 err = ext4_split_extent_at(handle, inode, ppath,
3468                                 map->m_lblk + map->m_len, split_flag1, flags1);
3469                 if (err)
3470                         goto out;
3471         } else {
3472                 allocated = ee_len - (map->m_lblk - ee_block);
3473         }
3474         /*
3475          * Update path is required because previous ext4_split_extent_at() may
3476          * result in split of original leaf or extent zeroout.
3477          */
3478         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3479         if (IS_ERR(path))
3480                 return PTR_ERR(path);
3481         depth = ext_depth(inode);
3482         ex = path[depth].p_ext;
3483         if (!ex) {
3484                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3485                                  (unsigned long) map->m_lblk);
3486                 return -EFSCORRUPTED;
3487         }
3488         unwritten = ext4_ext_is_unwritten(ex);
3489         split_flag1 = 0;
3490
3491         if (map->m_lblk >= ee_block) {
3492                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3493                 if (unwritten) {
3494                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3495                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3496                                                      EXT4_EXT_MARK_UNWRIT2);
3497                 }
3498                 err = ext4_split_extent_at(handle, inode, ppath,
3499                                 map->m_lblk, split_flag1, flags);
3500                 if (err)
3501                         goto out;
3502         }
3503
3504         ext4_ext_show_leaf(inode, path);
3505 out:
3506         return err ? err : allocated;
3507 }
3508
3509 /*
3510  * This function is called by ext4_ext_map_blocks() if someone tries to write
3511  * to an unwritten extent. It may result in splitting the unwritten
3512  * extent into multiple extents (up to three - one initialized and two
3513  * unwritten).
3514  * There are three possibilities:
3515  *   a> There is no split required: Entire extent should be initialized
3516  *   b> Splits in two extents: Write is happening at either end of the extent
3517  *   c> Splits in three extents: Somone is writing in middle of the extent
3518  *
3519  * Pre-conditions:
3520  *  - The extent pointed to by 'path' is unwritten.
3521  *  - The extent pointed to by 'path' contains a superset
3522  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3523  *
3524  * Post-conditions on success:
3525  *  - the returned value is the number of blocks beyond map->l_lblk
3526  *    that are allocated and initialized.
3527  *    It is guaranteed to be >= map->m_len.
3528  */
3529 static int ext4_ext_convert_to_initialized(handle_t *handle,
3530                                            struct inode *inode,
3531                                            struct ext4_map_blocks *map,
3532                                            struct ext4_ext_path **ppath,
3533                                            int flags)
3534 {
3535         struct ext4_ext_path *path = *ppath;
3536         struct ext4_sb_info *sbi;
3537         struct ext4_extent_header *eh;
3538         struct ext4_map_blocks split_map;
3539         struct ext4_extent zero_ex1, zero_ex2;
3540         struct ext4_extent *ex, *abut_ex;
3541         ext4_lblk_t ee_block, eof_block;
3542         unsigned int ee_len, depth, map_len = map->m_len;
3543         int allocated = 0, max_zeroout = 0;
3544         int err = 0;
3545         int split_flag = EXT4_EXT_DATA_VALID2;
3546
3547         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3548                 "block %llu, max_blocks %u\n", inode->i_ino,
3549                 (unsigned long long)map->m_lblk, map_len);
3550
3551         sbi = EXT4_SB(inode->i_sb);
3552         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3553                 inode->i_sb->s_blocksize_bits;
3554         if (eof_block < map->m_lblk + map_len)
3555                 eof_block = map->m_lblk + map_len;
3556
3557         depth = ext_depth(inode);
3558         eh = path[depth].p_hdr;
3559         ex = path[depth].p_ext;
3560         ee_block = le32_to_cpu(ex->ee_block);
3561         ee_len = ext4_ext_get_actual_len(ex);
3562         zero_ex1.ee_len = 0;
3563         zero_ex2.ee_len = 0;
3564
3565         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3566
3567         /* Pre-conditions */
3568         BUG_ON(!ext4_ext_is_unwritten(ex));
3569         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3570
3571         /*
3572          * Attempt to transfer newly initialized blocks from the currently
3573          * unwritten extent to its neighbor. This is much cheaper
3574          * than an insertion followed by a merge as those involve costly
3575          * memmove() calls. Transferring to the left is the common case in
3576          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3577          * followed by append writes.
3578          *
3579          * Limitations of the current logic:
3580          *  - L1: we do not deal with writes covering the whole extent.
3581          *    This would require removing the extent if the transfer
3582          *    is possible.
3583          *  - L2: we only attempt to merge with an extent stored in the
3584          *    same extent tree node.
3585          */
3586         if ((map->m_lblk == ee_block) &&
3587                 /* See if we can merge left */
3588                 (map_len < ee_len) &&           /*L1*/
3589                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3590                 ext4_lblk_t prev_lblk;
3591                 ext4_fsblk_t prev_pblk, ee_pblk;
3592                 unsigned int prev_len;
3593
3594                 abut_ex = ex - 1;
3595                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3596                 prev_len = ext4_ext_get_actual_len(abut_ex);
3597                 prev_pblk = ext4_ext_pblock(abut_ex);
3598                 ee_pblk = ext4_ext_pblock(ex);
3599
3600                 /*
3601                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3602                  * upon those conditions:
3603                  * - C1: abut_ex is initialized,
3604                  * - C2: abut_ex is logically abutting ex,
3605                  * - C3: abut_ex is physically abutting ex,
3606                  * - C4: abut_ex can receive the additional blocks without
3607                  *   overflowing the (initialized) length limit.
3608                  */
3609                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3610                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3611                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3612                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3613                         err = ext4_ext_get_access(handle, inode, path + depth);
3614                         if (err)
3615                                 goto out;
3616
3617                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3618                                 map, ex, abut_ex);
3619
3620                         /* Shift the start of ex by 'map_len' blocks */
3621                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3622                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3623                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3624                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3625
3626                         /* Extend abut_ex by 'map_len' blocks */
3627                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3628
3629                         /* Result: number of initialized blocks past m_lblk */
3630                         allocated = map_len;
3631                 }
3632         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3633                    (map_len < ee_len) &&        /*L1*/
3634                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3635                 /* See if we can merge right */
3636                 ext4_lblk_t next_lblk;
3637                 ext4_fsblk_t next_pblk, ee_pblk;
3638                 unsigned int next_len;
3639
3640                 abut_ex = ex + 1;
3641                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3642                 next_len = ext4_ext_get_actual_len(abut_ex);
3643                 next_pblk = ext4_ext_pblock(abut_ex);
3644                 ee_pblk = ext4_ext_pblock(ex);
3645
3646                 /*
3647                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3648                  * upon those conditions:
3649                  * - C1: abut_ex is initialized,
3650                  * - C2: abut_ex is logically abutting ex,
3651                  * - C3: abut_ex is physically abutting ex,
3652                  * - C4: abut_ex can receive the additional blocks without
3653                  *   overflowing the (initialized) length limit.
3654                  */
3655                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3656                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3657                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3658                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3659                         err = ext4_ext_get_access(handle, inode, path + depth);
3660                         if (err)
3661                                 goto out;
3662
3663                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3664                                 map, ex, abut_ex);
3665
3666                         /* Shift the start of abut_ex by 'map_len' blocks */
3667                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3668                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3669                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3670                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3671
3672                         /* Extend abut_ex by 'map_len' blocks */
3673                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3674
3675                         /* Result: number of initialized blocks past m_lblk */
3676                         allocated = map_len;
3677                 }
3678         }
3679         if (allocated) {
3680                 /* Mark the block containing both extents as dirty */
3681                 ext4_ext_dirty(handle, inode, path + depth);
3682
3683                 /* Update path to point to the right extent */
3684                 path[depth].p_ext = abut_ex;
3685                 goto out;
3686         } else
3687                 allocated = ee_len - (map->m_lblk - ee_block);
3688
3689         WARN_ON(map->m_lblk < ee_block);
3690         /*
3691          * It is safe to convert extent to initialized via explicit
3692          * zeroout only if extent is fully inside i_size or new_size.
3693          */
3694         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3695
3696         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3697                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3698                         (inode->i_sb->s_blocksize_bits - 10);
3699
3700         if (IS_ENCRYPTED(inode))
3701                 max_zeroout = 0;
3702
3703         /*
3704          * five cases:
3705          * 1. split the extent into three extents.
3706          * 2. split the extent into two extents, zeroout the head of the first
3707          *    extent.
3708          * 3. split the extent into two extents, zeroout the tail of the second
3709          *    extent.
3710          * 4. split the extent into two extents with out zeroout.
3711          * 5. no splitting needed, just possibly zeroout the head and / or the
3712          *    tail of the extent.
3713          */
3714         split_map.m_lblk = map->m_lblk;
3715         split_map.m_len = map->m_len;
3716
3717         if (max_zeroout && (allocated > split_map.m_len)) {
3718                 if (allocated <= max_zeroout) {
3719                         /* case 3 or 5 */
3720                         zero_ex1.ee_block =
3721                                  cpu_to_le32(split_map.m_lblk +
3722                                              split_map.m_len);
3723                         zero_ex1.ee_len =
3724                                 cpu_to_le16(allocated - split_map.m_len);
3725                         ext4_ext_store_pblock(&zero_ex1,
3726                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3727                                 split_map.m_len - ee_block);
3728                         err = ext4_ext_zeroout(inode, &zero_ex1);
3729                         if (err)
3730                                 goto out;
3731                         split_map.m_len = allocated;
3732                 }
3733                 if (split_map.m_lblk - ee_block + split_map.m_len <
3734                                                                 max_zeroout) {
3735                         /* case 2 or 5 */
3736                         if (split_map.m_lblk != ee_block) {
3737                                 zero_ex2.ee_block = ex->ee_block;
3738                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3739                                                         ee_block);
3740                                 ext4_ext_store_pblock(&zero_ex2,
3741                                                       ext4_ext_pblock(ex));
3742                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3743                                 if (err)
3744                                         goto out;
3745                         }
3746
3747                         split_map.m_len += split_map.m_lblk - ee_block;
3748                         split_map.m_lblk = ee_block;
3749                         allocated = map->m_len;
3750                 }
3751         }
3752
3753         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3754                                 flags);
3755         if (err > 0)
3756                 err = 0;
3757 out:
3758         /* If we have gotten a failure, don't zero out status tree */
3759         if (!err) {
3760                 err = ext4_zeroout_es(inode, &zero_ex1);
3761                 if (!err)
3762                         err = ext4_zeroout_es(inode, &zero_ex2);
3763         }
3764         return err ? err : allocated;
3765 }
3766
3767 /*
3768  * This function is called by ext4_ext_map_blocks() from
3769  * ext4_get_blocks_dio_write() when DIO to write
3770  * to an unwritten extent.
3771  *
3772  * Writing to an unwritten extent may result in splitting the unwritten
3773  * extent into multiple initialized/unwritten extents (up to three)
3774  * There are three possibilities:
3775  *   a> There is no split required: Entire extent should be unwritten
3776  *   b> Splits in two extents: Write is happening at either end of the extent
3777  *   c> Splits in three extents: Somone is writing in middle of the extent
3778  *
3779  * This works the same way in the case of initialized -> unwritten conversion.
3780  *
3781  * One of more index blocks maybe needed if the extent tree grow after
3782  * the unwritten extent split. To prevent ENOSPC occur at the IO
3783  * complete, we need to split the unwritten extent before DIO submit
3784  * the IO. The unwritten extent called at this time will be split
3785  * into three unwritten extent(at most). After IO complete, the part
3786  * being filled will be convert to initialized by the end_io callback function
3787  * via ext4_convert_unwritten_extents().
3788  *
3789  * Returns the size of unwritten extent to be written on success.
3790  */
3791 static int ext4_split_convert_extents(handle_t *handle,
3792                                         struct inode *inode,
3793                                         struct ext4_map_blocks *map,
3794                                         struct ext4_ext_path **ppath,
3795                                         int flags)
3796 {
3797         struct ext4_ext_path *path = *ppath;
3798         ext4_lblk_t eof_block;
3799         ext4_lblk_t ee_block;
3800         struct ext4_extent *ex;
3801         unsigned int ee_len;
3802         int split_flag = 0, depth;
3803
3804         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3805                   __func__, inode->i_ino,
3806                   (unsigned long long)map->m_lblk, map->m_len);
3807
3808         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3809                 inode->i_sb->s_blocksize_bits;
3810         if (eof_block < map->m_lblk + map->m_len)
3811                 eof_block = map->m_lblk + map->m_len;
3812         /*
3813          * It is safe to convert extent to initialized via explicit
3814          * zeroout only if extent is fully insde i_size or new_size.
3815          */
3816         depth = ext_depth(inode);
3817         ex = path[depth].p_ext;
3818         ee_block = le32_to_cpu(ex->ee_block);
3819         ee_len = ext4_ext_get_actual_len(ex);
3820
3821         /* Convert to unwritten */
3822         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3823                 split_flag |= EXT4_EXT_DATA_VALID1;
3824         /* Convert to initialized */
3825         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3826                 split_flag |= ee_block + ee_len <= eof_block ?
3827                               EXT4_EXT_MAY_ZEROOUT : 0;
3828                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3829         }
3830         flags |= EXT4_GET_BLOCKS_PRE_IO;
3831         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3832 }
3833
3834 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3835                                                 struct inode *inode,
3836                                                 struct ext4_map_blocks *map,
3837                                                 struct ext4_ext_path **ppath)
3838 {
3839         struct ext4_ext_path *path = *ppath;
3840         struct ext4_extent *ex;
3841         ext4_lblk_t ee_block;
3842         unsigned int ee_len;
3843         int depth;
3844         int err = 0;
3845
3846         depth = ext_depth(inode);
3847         ex = path[depth].p_ext;
3848         ee_block = le32_to_cpu(ex->ee_block);
3849         ee_len = ext4_ext_get_actual_len(ex);
3850
3851         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3852                 "block %llu, max_blocks %u\n", inode->i_ino,
3853                   (unsigned long long)ee_block, ee_len);
3854
3855         /* If extent is larger than requested it is a clear sign that we still
3856          * have some extent state machine issues left. So extent_split is still
3857          * required.
3858          * TODO: Once all related issues will be fixed this situation should be
3859          * illegal.
3860          */
3861         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3862 #ifdef CONFIG_EXT4_DEBUG
3863                 ext4_warning(inode->i_sb, "Inode (%ld) finished: extent logical block %llu,"
3864                              " len %u; IO logical block %llu, len %u",
3865                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3866                              (unsigned long long)map->m_lblk, map->m_len);
3867 #endif
3868                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3869                                                  EXT4_GET_BLOCKS_CONVERT);
3870                 if (err < 0)
3871                         return err;
3872                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3873                 if (IS_ERR(path))
3874                         return PTR_ERR(path);
3875                 depth = ext_depth(inode);
3876                 ex = path[depth].p_ext;
3877         }
3878
3879         err = ext4_ext_get_access(handle, inode, path + depth);
3880         if (err)
3881                 goto out;
3882         /* first mark the extent as initialized */
3883         ext4_ext_mark_initialized(ex);
3884
3885         /* note: ext4_ext_correct_indexes() isn't needed here because
3886          * borders are not changed
3887          */
3888         ext4_ext_try_to_merge(handle, inode, path, ex);
3889
3890         /* Mark modified extent as dirty */
3891         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3892 out:
3893         ext4_ext_show_leaf(inode, path);
3894         return err;
3895 }
3896
3897 /*
3898  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3899  */
3900 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3901                               ext4_lblk_t lblk,
3902                               struct ext4_ext_path *path,
3903                               unsigned int len)
3904 {
3905         int i, depth;
3906         struct ext4_extent_header *eh;
3907         struct ext4_extent *last_ex;
3908
3909         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3910                 return 0;
3911
3912         depth = ext_depth(inode);
3913         eh = path[depth].p_hdr;
3914
3915         /*
3916          * We're going to remove EOFBLOCKS_FL entirely in future so we
3917          * do not care for this case anymore. Simply remove the flag
3918          * if there are no extents.
3919          */
3920         if (unlikely(!eh->eh_entries))
3921                 goto out;
3922         last_ex = EXT_LAST_EXTENT(eh);
3923         /*
3924          * We should clear the EOFBLOCKS_FL flag if we are writing the
3925          * last block in the last extent in the file.  We test this by
3926          * first checking to see if the caller to
3927          * ext4_ext_get_blocks() was interested in the last block (or
3928          * a block beyond the last block) in the current extent.  If
3929          * this turns out to be false, we can bail out from this
3930          * function immediately.
3931          */
3932         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3933             ext4_ext_get_actual_len(last_ex))
3934                 return 0;
3935         /*
3936          * If the caller does appear to be planning to write at or
3937          * beyond the end of the current extent, we then test to see
3938          * if the current extent is the last extent in the file, by
3939          * checking to make sure it was reached via the rightmost node
3940          * at each level of the tree.
3941          */
3942         for (i = depth-1; i >= 0; i--)
3943                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3944                         return 0;
3945 out:
3946         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3947         return ext4_mark_inode_dirty(handle, inode);
3948 }
3949
3950 static int
3951 convert_initialized_extent(handle_t *handle, struct inode *inode,
3952                            struct ext4_map_blocks *map,
3953                            struct ext4_ext_path **ppath,
3954                            unsigned int allocated)
3955 {
3956         struct ext4_ext_path *path = *ppath;
3957         struct ext4_extent *ex;
3958         ext4_lblk_t ee_block;
3959         unsigned int ee_len;
3960         int depth;
3961         int err = 0;
3962
3963         /*
3964          * Make sure that the extent is no bigger than we support with
3965          * unwritten extent
3966          */
3967         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3968                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3969
3970         depth = ext_depth(inode);
3971         ex = path[depth].p_ext;
3972         ee_block = le32_to_cpu(ex->ee_block);
3973         ee_len = ext4_ext_get_actual_len(ex);
3974
3975         ext_debug("%s: inode %lu, logical"
3976                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3977                   (unsigned long long)ee_block, ee_len);
3978
3979         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3980                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3981                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3982                 if (err < 0)
3983                         return err;
3984                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3985                 if (IS_ERR(path))
3986                         return PTR_ERR(path);
3987                 depth = ext_depth(inode);
3988                 ex = path[depth].p_ext;
3989                 if (!ex) {
3990                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3991                                          (unsigned long) map->m_lblk);
3992                         return -EFSCORRUPTED;
3993                 }
3994         }
3995
3996         err = ext4_ext_get_access(handle, inode, path + depth);
3997         if (err)
3998                 return err;
3999         /* first mark the extent as unwritten */
4000         ext4_ext_mark_unwritten(ex);
4001
4002         /* note: ext4_ext_correct_indexes() isn't needed here because
4003          * borders are not changed
4004          */
4005         ext4_ext_try_to_merge(handle, inode, path, ex);
4006
4007         /* Mark modified extent as dirty */
4008         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
4009         if (err)
4010                 return err;
4011         ext4_ext_show_leaf(inode, path);
4012
4013         ext4_update_inode_fsync_trans(handle, inode, 1);
4014         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
4015         if (err)
4016                 return err;
4017         map->m_flags |= EXT4_MAP_UNWRITTEN;
4018         if (allocated > map->m_len)
4019                 allocated = map->m_len;
4020         map->m_len = allocated;
4021         return allocated;
4022 }
4023
4024 static int
4025 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4026                         struct ext4_map_blocks *map,
4027                         struct ext4_ext_path **ppath, int flags,
4028                         unsigned int allocated, ext4_fsblk_t newblock)
4029 {
4030         struct ext4_ext_path *path = *ppath;
4031         int ret = 0;
4032         int err = 0;
4033
4034         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4035                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4036                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4037                   flags, allocated);
4038         ext4_ext_show_leaf(inode, path);
4039
4040         /*
4041          * When writing into unwritten space, we should not fail to
4042          * allocate metadata blocks for the new extent block if needed.
4043          */
4044         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4045
4046         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4047                                                     allocated, newblock);
4048
4049         /* get_block() before submit the IO, split the extent */
4050         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4051                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4052                                          flags | EXT4_GET_BLOCKS_CONVERT);
4053                 if (ret <= 0)
4054                         goto out;
4055                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4056                 goto out;
4057         }
4058         /* IO end_io complete, convert the filled extent to written */
4059         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4060                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4061                         if (allocated > map->m_len)
4062                                 allocated = map->m_len;
4063                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4064                                                  allocated);
4065                         if (err < 0)
4066                                 goto out2;
4067                 }
4068                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4069                                                            ppath);
4070                 if (ret >= 0) {
4071                         ext4_update_inode_fsync_trans(handle, inode, 1);
4072                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4073                                                  path, map->m_len);
4074                 } else
4075                         err = ret;
4076                 map->m_flags |= EXT4_MAP_MAPPED;
4077                 map->m_pblk = newblock;
4078                 if (allocated > map->m_len)
4079                         allocated = map->m_len;
4080                 map->m_len = allocated;
4081                 goto out2;
4082         }
4083         /* buffered IO case */
4084         /*
4085          * repeat fallocate creation request
4086          * we already have an unwritten extent
4087          */
4088         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4089                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4090                 goto map_out;
4091         }
4092
4093         /* buffered READ or buffered write_begin() lookup */
4094         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4095                 /*
4096                  * We have blocks reserved already.  We
4097                  * return allocated blocks so that delalloc
4098                  * won't do block reservation for us.  But
4099                  * the buffer head will be unmapped so that
4100                  * a read from the block returns 0s.
4101                  */
4102                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4103                 goto out1;
4104         }
4105
4106         /* buffered write, writepage time, convert*/
4107         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4108         if (ret >= 0)
4109                 ext4_update_inode_fsync_trans(handle, inode, 1);
4110 out:
4111         if (ret <= 0) {
4112                 err = ret;
4113                 goto out2;
4114         } else
4115                 allocated = ret;
4116         map->m_flags |= EXT4_MAP_NEW;
4117         if (allocated > map->m_len)
4118                 allocated = map->m_len;
4119         map->m_len = allocated;
4120
4121 map_out:
4122         map->m_flags |= EXT4_MAP_MAPPED;
4123         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4124                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4125                                          map->m_len);
4126                 if (err < 0)
4127                         goto out2;
4128         }
4129 out1:
4130         if (allocated > map->m_len)
4131                 allocated = map->m_len;
4132         ext4_ext_show_leaf(inode, path);
4133         map->m_pblk = newblock;
4134         map->m_len = allocated;
4135 out2:
4136         return err ? err : allocated;
4137 }
4138
4139 /*
4140  * get_implied_cluster_alloc - check to see if the requested
4141  * allocation (in the map structure) overlaps with a cluster already
4142  * allocated in an extent.
4143  *      @sb     The filesystem superblock structure
4144  *      @map    The requested lblk->pblk mapping
4145  *      @ex     The extent structure which might contain an implied
4146  *                      cluster allocation
4147  *
4148  * This function is called by ext4_ext_map_blocks() after we failed to
4149  * find blocks that were already in the inode's extent tree.  Hence,
4150  * we know that the beginning of the requested region cannot overlap
4151  * the extent from the inode's extent tree.  There are three cases we
4152  * want to catch.  The first is this case:
4153  *
4154  *               |--- cluster # N--|
4155  *    |--- extent ---|  |---- requested region ---|
4156  *                      |==========|
4157  *
4158  * The second case that we need to test for is this one:
4159  *
4160  *   |--------- cluster # N ----------------|
4161  *         |--- requested region --|   |------- extent ----|
4162  *         |=======================|
4163  *
4164  * The third case is when the requested region lies between two extents
4165  * within the same cluster:
4166  *          |------------- cluster # N-------------|
4167  * |----- ex -----|                  |---- ex_right ----|
4168  *                  |------ requested region ------|
4169  *                  |================|
4170  *
4171  * In each of the above cases, we need to set the map->m_pblk and
4172  * map->m_len so it corresponds to the return the extent labelled as
4173  * "|====|" from cluster #N, since it is already in use for data in
4174  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4175  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4176  * as a new "allocated" block region.  Otherwise, we will return 0 and
4177  * ext4_ext_map_blocks() will then allocate one or more new clusters
4178  * by calling ext4_mb_new_blocks().
4179  */
4180 static int get_implied_cluster_alloc(struct super_block *sb,
4181                                      struct ext4_map_blocks *map,
4182                                      struct ext4_extent *ex,
4183                                      struct ext4_ext_path *path)
4184 {
4185         struct ext4_sb_info *sbi = EXT4_SB(sb);
4186         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4187         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4188         ext4_lblk_t rr_cluster_start;
4189         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4190         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4191         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4192
4193         /* The extent passed in that we are trying to match */
4194         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4195         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4196
4197         /* The requested region passed into ext4_map_blocks() */
4198         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4199
4200         if ((rr_cluster_start == ex_cluster_end) ||
4201             (rr_cluster_start == ex_cluster_start)) {
4202                 if (rr_cluster_start == ex_cluster_end)
4203                         ee_start += ee_len - 1;
4204                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4205                 map->m_len = min(map->m_len,
4206                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4207                 /*
4208                  * Check for and handle this case:
4209                  *
4210                  *   |--------- cluster # N-------------|
4211                  *                     |------- extent ----|
4212                  *         |--- requested region ---|
4213                  *         |===========|
4214                  */
4215
4216                 if (map->m_lblk < ee_block)
4217                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4218
4219                 /*
4220                  * Check for the case where there is already another allocated
4221                  * block to the right of 'ex' but before the end of the cluster.
4222                  *
4223                  *          |------------- cluster # N-------------|
4224                  * |----- ex -----|                  |---- ex_right ----|
4225                  *                  |------ requested region ------|
4226                  *                  |================|
4227                  */
4228                 if (map->m_lblk > ee_block) {
4229                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4230                         map->m_len = min(map->m_len, next - map->m_lblk);
4231                 }
4232
4233                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4234                 return 1;
4235         }
4236
4237         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4238         return 0;
4239 }
4240
4241
4242 /*
4243  * Block allocation/map/preallocation routine for extents based files
4244  *
4245  *
4246  * Need to be called with
4247  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4248  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4249  *
4250  * return > 0, number of of blocks already mapped/allocated
4251  *          if create == 0 and these are pre-allocated blocks
4252  *              buffer head is unmapped
4253  *          otherwise blocks are mapped
4254  *
4255  * return = 0, if plain look up failed (blocks have not been allocated)
4256  *          buffer head is unmapped
4257  *
4258  * return < 0, error case.
4259  */
4260 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4261                         struct ext4_map_blocks *map, int flags)
4262 {
4263         struct ext4_ext_path *path = NULL;
4264         struct ext4_extent newex, *ex, *ex2;
4265         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4266         ext4_fsblk_t newblock = 0;
4267         int free_on_err = 0, err = 0, depth, ret;
4268         unsigned int allocated = 0, offset = 0;
4269         unsigned int allocated_clusters = 0;
4270         struct ext4_allocation_request ar;
4271         ext4_lblk_t cluster_offset;
4272         bool map_from_cluster = false;
4273
4274         ext_debug("blocks %u/%u requested for inode %lu\n",
4275                   map->m_lblk, map->m_len, inode->i_ino);
4276         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4277
4278         /* find extent for this block */
4279         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4280         if (IS_ERR(path)) {
4281                 err = PTR_ERR(path);
4282                 path = NULL;
4283                 goto out2;
4284         }
4285
4286         depth = ext_depth(inode);
4287
4288         /*
4289          * consistent leaf must not be empty;
4290          * this situation is possible, though, _during_ tree modification;
4291          * this is why assert can't be put in ext4_find_extent()
4292          */
4293         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4294                 EXT4_ERROR_INODE(inode, "bad extent address "
4295                                  "lblock: %lu, depth: %d pblock %lld",
4296                                  (unsigned long) map->m_lblk, depth,
4297                                  path[depth].p_block);
4298                 err = -EFSCORRUPTED;
4299                 goto out2;
4300         }
4301
4302         ex = path[depth].p_ext;
4303         if (ex) {
4304                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4305                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4306                 unsigned short ee_len;
4307
4308
4309                 /*
4310                  * unwritten extents are treated as holes, except that
4311                  * we split out initialized portions during a write.
4312                  */
4313                 ee_len = ext4_ext_get_actual_len(ex);
4314
4315                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4316
4317                 /* if found extent covers block, simply return it */
4318                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4319                         newblock = map->m_lblk - ee_block + ee_start;
4320                         /* number of remaining blocks in the extent */
4321                         allocated = ee_len - (map->m_lblk - ee_block);
4322                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4323                                   ee_block, ee_len, newblock);
4324
4325                         /*
4326                          * If the extent is initialized check whether the
4327                          * caller wants to convert it to unwritten.
4328                          */
4329                         if ((!ext4_ext_is_unwritten(ex)) &&
4330                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4331                                 allocated = convert_initialized_extent(
4332                                                 handle, inode, map, &path,
4333                                                 allocated);
4334                                 goto out2;
4335                         } else if (!ext4_ext_is_unwritten(ex))
4336                                 goto out;
4337
4338                         ret = ext4_ext_handle_unwritten_extents(
4339                                 handle, inode, map, &path, flags,
4340                                 allocated, newblock);
4341                         if (ret < 0)
4342                                 err = ret;
4343                         else
4344                                 allocated = ret;
4345                         goto out2;
4346                 }
4347         }
4348
4349         /*
4350          * requested block isn't allocated yet;
4351          * we couldn't try to create block if create flag is zero
4352          */
4353         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4354                 ext4_lblk_t hole_start, hole_len;
4355
4356                 hole_start = map->m_lblk;
4357                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4358                 /*
4359                  * put just found gap into cache to speed up
4360                  * subsequent requests
4361                  */
4362                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4363
4364                 /* Update hole_len to reflect hole size after map->m_lblk */
4365                 if (hole_start != map->m_lblk)
4366                         hole_len -= map->m_lblk - hole_start;
4367                 map->m_pblk = 0;
4368                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4369
4370                 goto out2;
4371         }
4372
4373         /*
4374          * Okay, we need to do block allocation.
4375          */
4376         newex.ee_block = cpu_to_le32(map->m_lblk);
4377         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4378
4379         /*
4380          * If we are doing bigalloc, check to see if the extent returned
4381          * by ext4_find_extent() implies a cluster we can use.
4382          */
4383         if (cluster_offset && ex &&
4384             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4385                 ar.len = allocated = map->m_len;
4386                 newblock = map->m_pblk;
4387                 map_from_cluster = true;
4388                 goto got_allocated_blocks;
4389         }
4390
4391         /* find neighbour allocated blocks */
4392         ar.lleft = map->m_lblk;
4393         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4394         if (err)
4395                 goto out2;
4396         ar.lright = map->m_lblk;
4397         ex2 = NULL;
4398         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4399         if (err)
4400                 goto out2;
4401
4402         /* Check if the extent after searching to the right implies a
4403          * cluster we can use. */
4404         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4405             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4406                 ar.len = allocated = map->m_len;
4407                 newblock = map->m_pblk;
4408                 map_from_cluster = true;
4409                 goto got_allocated_blocks;
4410         }
4411
4412         /*
4413          * See if request is beyond maximum number of blocks we can have in
4414          * a single extent. For an initialized extent this limit is
4415          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4416          * EXT_UNWRITTEN_MAX_LEN.
4417          */
4418         if (map->m_len > EXT_INIT_MAX_LEN &&
4419             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4420                 map->m_len = EXT_INIT_MAX_LEN;
4421         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4422                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4423                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4424
4425         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4426         newex.ee_len = cpu_to_le16(map->m_len);
4427         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4428         if (err)
4429                 allocated = ext4_ext_get_actual_len(&newex);
4430         else
4431                 allocated = map->m_len;
4432
4433         /* allocate new block */
4434         ar.inode = inode;
4435         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4436         ar.logical = map->m_lblk;
4437         /*
4438          * We calculate the offset from the beginning of the cluster
4439          * for the logical block number, since when we allocate a
4440          * physical cluster, the physical block should start at the
4441          * same offset from the beginning of the cluster.  This is
4442          * needed so that future calls to get_implied_cluster_alloc()
4443          * work correctly.
4444          */
4445         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4446         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4447         ar.goal -= offset;
4448         ar.logical -= offset;
4449         if (S_ISREG(inode->i_mode))
4450                 ar.flags = EXT4_MB_HINT_DATA;
4451         else
4452                 /* disable in-core preallocation for non-regular files */
4453                 ar.flags = 0;
4454         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4455                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4456         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4457                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4458         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4459                 ar.flags |= EXT4_MB_USE_RESERVED;
4460         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4461         if (!newblock)
4462                 goto out2;
4463         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4464                   ar.goal, newblock, allocated);
4465         free_on_err = 1;
4466         allocated_clusters = ar.len;
4467         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4468         if (ar.len > allocated)
4469                 ar.len = allocated;
4470
4471 got_allocated_blocks:
4472         /* try to insert new extent into found leaf and return */
4473         ext4_ext_store_pblock(&newex, newblock + offset);
4474         newex.ee_len = cpu_to_le16(ar.len);
4475         /* Mark unwritten */
4476         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4477                 ext4_ext_mark_unwritten(&newex);
4478                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4479         }
4480
4481         err = 0;
4482         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4483                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4484                                          path, ar.len);
4485         if (!err)
4486                 err = ext4_ext_insert_extent(handle, inode, &path,
4487                                              &newex, flags);
4488
4489         if (err && free_on_err) {
4490                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4491                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4492                 /* free data blocks we just allocated */
4493                 /* not a good idea to call discard here directly,
4494                  * but otherwise we'd need to call it every free() */
4495                 ext4_discard_preallocations(inode);
4496                 ext4_free_blocks(handle, inode, NULL, newblock,
4497                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4498                 goto out2;
4499         }
4500
4501         /* previous routine could use block we allocated */
4502         newblock = ext4_ext_pblock(&newex);
4503         allocated = ext4_ext_get_actual_len(&newex);
4504         if (allocated > map->m_len)
4505                 allocated = map->m_len;
4506         map->m_flags |= EXT4_MAP_NEW;
4507
4508         /*
4509          * Reduce the reserved cluster count to reflect successful deferred
4510          * allocation of delayed allocated clusters or direct allocation of
4511          * clusters discovered to be delayed allocated.  Once allocated, a
4512          * cluster is not included in the reserved count.
4513          */
4514         if (test_opt(inode->i_sb, DELALLOC) && !map_from_cluster) {
4515                 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4516                         /*
4517                          * When allocating delayed allocated clusters, simply
4518                          * reduce the reserved cluster count and claim quota
4519                          */
4520                         ext4_da_update_reserve_space(inode, allocated_clusters,
4521                                                         1);
4522                 } else {
4523                         ext4_lblk_t lblk, len;
4524                         unsigned int n;
4525
4526                         /*
4527                          * When allocating non-delayed allocated clusters
4528                          * (from fallocate, filemap, DIO, or clusters
4529                          * allocated when delalloc has been disabled by
4530                          * ext4_nonda_switch), reduce the reserved cluster
4531                          * count by the number of allocated clusters that
4532                          * have previously been delayed allocated.  Quota
4533                          * has been claimed by ext4_mb_new_blocks() above,
4534                          * so release the quota reservations made for any
4535                          * previously delayed allocated clusters.
4536                          */
4537                         lblk = EXT4_LBLK_CMASK(sbi, map->m_lblk);
4538                         len = allocated_clusters << sbi->s_cluster_bits;
4539                         n = ext4_es_delayed_clu(inode, lblk, len);
4540                         if (n > 0)
4541                                 ext4_da_update_reserve_space(inode, (int) n, 0);
4542                 }
4543         }
4544
4545         /*
4546          * Cache the extent and update transaction to commit on fdatasync only
4547          * when it is _not_ an unwritten extent.
4548          */
4549         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4550                 ext4_update_inode_fsync_trans(handle, inode, 1);
4551         else
4552                 ext4_update_inode_fsync_trans(handle, inode, 0);
4553 out:
4554         if (allocated > map->m_len)
4555                 allocated = map->m_len;
4556         ext4_ext_show_leaf(inode, path);
4557         map->m_flags |= EXT4_MAP_MAPPED;
4558         map->m_pblk = newblock;
4559         map->m_len = allocated;
4560 out2:
4561         ext4_ext_drop_refs(path);
4562         kfree(path);
4563
4564         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4565                                        err ? err : allocated);
4566         return err ? err : allocated;
4567 }
4568
4569 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4570 {
4571         struct super_block *sb = inode->i_sb;
4572         ext4_lblk_t last_block;
4573         int err = 0;
4574
4575         /*
4576          * TODO: optimization is possible here.
4577          * Probably we need not scan at all,
4578          * because page truncation is enough.
4579          */
4580
4581         /* we have to know where to truncate from in crash case */
4582         EXT4_I(inode)->i_disksize = inode->i_size;
4583         err = ext4_mark_inode_dirty(handle, inode);
4584         if (err)
4585                 return err;
4586
4587         last_block = (inode->i_size + sb->s_blocksize - 1)
4588                         >> EXT4_BLOCK_SIZE_BITS(sb);
4589 retry:
4590         err = ext4_es_remove_extent(inode, last_block,
4591                                     EXT_MAX_BLOCKS - last_block);
4592         if (err == -ENOMEM) {
4593                 cond_resched();
4594                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4595                 goto retry;
4596         }
4597         if (err)
4598                 return err;
4599         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4600 }
4601
4602 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4603                                   ext4_lblk_t len, loff_t new_size,
4604                                   int flags)
4605 {
4606         struct inode *inode = file_inode(file);
4607         handle_t *handle;
4608         int ret = 0;
4609         int ret2 = 0;
4610         int retries = 0;
4611         int depth = 0;
4612         struct ext4_map_blocks map;
4613         unsigned int credits;
4614         loff_t epos;
4615
4616         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4617         map.m_lblk = offset;
4618         map.m_len = len;
4619         /*
4620          * Don't normalize the request if it can fit in one extent so
4621          * that it doesn't get unnecessarily split into multiple
4622          * extents.
4623          */
4624         if (len <= EXT_UNWRITTEN_MAX_LEN)
4625                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4626
4627         /*
4628          * credits to insert 1 extent into extent tree
4629          */
4630         credits = ext4_chunk_trans_blocks(inode, len);
4631         depth = ext_depth(inode);
4632
4633 retry:
4634         while (ret >= 0 && len) {
4635                 /*
4636                  * Recalculate credits when extent tree depth changes.
4637                  */
4638                 if (depth != ext_depth(inode)) {
4639                         credits = ext4_chunk_trans_blocks(inode, len);
4640                         depth = ext_depth(inode);
4641                 }
4642
4643                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4644                                             credits);
4645                 if (IS_ERR(handle)) {
4646                         ret = PTR_ERR(handle);
4647                         break;
4648                 }
4649                 ret = ext4_map_blocks(handle, inode, &map, flags);
4650                 if (ret <= 0) {
4651                         ext4_debug("inode #%lu: block %u: len %u: "
4652                                    "ext4_ext_map_blocks returned %d",
4653                                    inode->i_ino, map.m_lblk,
4654                                    map.m_len, ret);
4655                         ext4_mark_inode_dirty(handle, inode);
4656                         ret2 = ext4_journal_stop(handle);
4657                         break;
4658                 }
4659                 map.m_lblk += ret;
4660                 map.m_len = len = len - ret;
4661                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4662                 inode->i_ctime = current_time(inode);
4663                 if (new_size) {
4664                         if (epos > new_size)
4665                                 epos = new_size;
4666                         if (ext4_update_inode_size(inode, epos) & 0x1)
4667                                 inode->i_mtime = inode->i_ctime;
4668                 } else {
4669                         if (epos > inode->i_size)
4670                                 ext4_set_inode_flag(inode,
4671                                                     EXT4_INODE_EOFBLOCKS);
4672                 }
4673                 ext4_mark_inode_dirty(handle, inode);
4674                 ext4_update_inode_fsync_trans(handle, inode, 1);
4675                 ret2 = ext4_journal_stop(handle);
4676                 if (ret2)
4677                         break;
4678         }
4679         if (ret == -ENOSPC &&
4680                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4681                 ret = 0;
4682                 goto retry;
4683         }
4684
4685         return ret > 0 ? ret2 : ret;
4686 }
4687
4688 static long ext4_zero_range(struct file *file, loff_t offset,
4689                             loff_t len, int mode)
4690 {
4691         struct inode *inode = file_inode(file);
4692         handle_t *handle = NULL;
4693         unsigned int max_blocks;
4694         loff_t new_size = 0;
4695         int ret = 0;
4696         int flags;
4697         int credits;
4698         int partial_begin, partial_end;
4699         loff_t start, end;
4700         ext4_lblk_t lblk;
4701         unsigned int blkbits = inode->i_blkbits;
4702
4703         trace_ext4_zero_range(inode, offset, len, mode);
4704
4705         if (!S_ISREG(inode->i_mode))
4706                 return -EINVAL;
4707
4708         /* Call ext4_force_commit to flush all data in case of data=journal. */
4709         if (ext4_should_journal_data(inode)) {
4710                 ret = ext4_force_commit(inode->i_sb);
4711                 if (ret)
4712                         return ret;
4713         }
4714
4715         /*
4716          * Round up offset. This is not fallocate, we neet to zero out
4717          * blocks, so convert interior block aligned part of the range to
4718          * unwritten and possibly manually zero out unaligned parts of the
4719          * range.
4720          */
4721         start = round_up(offset, 1 << blkbits);
4722         end = round_down((offset + len), 1 << blkbits);
4723
4724         if (start < offset || end > offset + len)
4725                 return -EINVAL;
4726         partial_begin = offset & ((1 << blkbits) - 1);
4727         partial_end = (offset + len) & ((1 << blkbits) - 1);
4728
4729         lblk = start >> blkbits;
4730         max_blocks = (end >> blkbits);
4731         if (max_blocks < lblk)
4732                 max_blocks = 0;
4733         else
4734                 max_blocks -= lblk;
4735
4736         inode_lock(inode);
4737
4738         /*
4739          * Indirect files do not support unwritten extnets
4740          */
4741         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4742                 ret = -EOPNOTSUPP;
4743                 goto out_mutex;
4744         }
4745
4746         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4747             (offset + len > i_size_read(inode) ||
4748              offset + len > EXT4_I(inode)->i_disksize)) {
4749                 new_size = offset + len;
4750                 ret = inode_newsize_ok(inode, new_size);
4751                 if (ret)
4752                         goto out_mutex;
4753         }
4754
4755         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4756         if (mode & FALLOC_FL_KEEP_SIZE)
4757                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4758
4759         /* Wait all existing dio workers, newcomers will block on i_mutex */
4760         inode_dio_wait(inode);
4761
4762         /* Preallocate the range including the unaligned edges */
4763         if (partial_begin || partial_end) {
4764                 ret = ext4_alloc_file_blocks(file,
4765                                 round_down(offset, 1 << blkbits) >> blkbits,
4766                                 (round_up((offset + len), 1 << blkbits) -
4767                                  round_down(offset, 1 << blkbits)) >> blkbits,
4768                                 new_size, flags);
4769                 if (ret)
4770                         goto out_mutex;
4771
4772         }
4773
4774         /* Zero range excluding the unaligned edges */
4775         if (max_blocks > 0) {
4776                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4777                           EXT4_EX_NOCACHE);
4778
4779                 /*
4780                  * Prevent page faults from reinstantiating pages we have
4781                  * released from page cache.
4782                  */
4783                 down_write(&EXT4_I(inode)->i_mmap_sem);
4784
4785                 ret = ext4_break_layouts(inode);
4786                 if (ret) {
4787                         up_write(&EXT4_I(inode)->i_mmap_sem);
4788                         goto out_mutex;
4789                 }
4790
4791                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4792                 if (ret) {
4793                         up_write(&EXT4_I(inode)->i_mmap_sem);
4794                         goto out_mutex;
4795                 }
4796                 /* Now release the pages and zero block aligned part of pages */
4797                 truncate_pagecache_range(inode, start, end - 1);
4798                 inode->i_mtime = inode->i_ctime = current_time(inode);
4799
4800                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4801                                              flags);
4802                 up_write(&EXT4_I(inode)->i_mmap_sem);
4803                 if (ret)
4804                         goto out_mutex;
4805         }
4806         if (!partial_begin && !partial_end)
4807                 goto out_mutex;
4808
4809         /*
4810          * In worst case we have to writeout two nonadjacent unwritten
4811          * blocks and update the inode
4812          */
4813         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4814         if (ext4_should_journal_data(inode))
4815                 credits += 2;
4816         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4817         if (IS_ERR(handle)) {
4818                 ret = PTR_ERR(handle);
4819                 ext4_std_error(inode->i_sb, ret);
4820                 goto out_mutex;
4821         }
4822
4823         inode->i_mtime = inode->i_ctime = current_time(inode);
4824         if (new_size) {
4825                 ext4_update_inode_size(inode, new_size);
4826         } else {
4827                 /*
4828                 * Mark that we allocate beyond EOF so the subsequent truncate
4829                 * can proceed even if the new size is the same as i_size.
4830                 */
4831                 if ((offset + len) > i_size_read(inode))
4832                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4833         }
4834         ext4_mark_inode_dirty(handle, inode);
4835
4836         /* Zero out partial block at the edges of the range */
4837         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4838         if (ret >= 0)
4839                 ext4_update_inode_fsync_trans(handle, inode, 1);
4840
4841         if (file->f_flags & O_SYNC)
4842                 ext4_handle_sync(handle);
4843
4844         ext4_journal_stop(handle);
4845 out_mutex:
4846         inode_unlock(inode);
4847         return ret;
4848 }
4849
4850 /*
4851  * preallocate space for a file. This implements ext4's fallocate file
4852  * operation, which gets called from sys_fallocate system call.
4853  * For block-mapped files, posix_fallocate should fall back to the method
4854  * of writing zeroes to the required new blocks (the same behavior which is
4855  * expected for file systems which do not support fallocate() system call).
4856  */
4857 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4858 {
4859         struct inode *inode = file_inode(file);
4860         loff_t new_size = 0;
4861         unsigned int max_blocks;
4862         int ret = 0;
4863         int flags;
4864         ext4_lblk_t lblk;
4865         unsigned int blkbits = inode->i_blkbits;
4866
4867         /*
4868          * Encrypted inodes can't handle collapse range or insert
4869          * range since we would need to re-encrypt blocks with a
4870          * different IV or XTS tweak (which are based on the logical
4871          * block number).
4872          *
4873          * XXX It's not clear why zero range isn't working, but we'll
4874          * leave it disabled for encrypted inodes for now.  This is a
4875          * bug we should fix....
4876          */
4877         if (IS_ENCRYPTED(inode) &&
4878             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4879                      FALLOC_FL_ZERO_RANGE)))
4880                 return -EOPNOTSUPP;
4881
4882         /* Return error if mode is not supported */
4883         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4884                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4885                      FALLOC_FL_INSERT_RANGE))
4886                 return -EOPNOTSUPP;
4887
4888         if (mode & FALLOC_FL_PUNCH_HOLE)
4889                 return ext4_punch_hole(inode, offset, len);
4890
4891         ret = ext4_convert_inline_data(inode);
4892         if (ret)
4893                 return ret;
4894
4895         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4896                 return ext4_collapse_range(inode, offset, len);
4897
4898         if (mode & FALLOC_FL_INSERT_RANGE)
4899                 return ext4_insert_range(inode, offset, len);
4900
4901         if (mode & FALLOC_FL_ZERO_RANGE)
4902                 return ext4_zero_range(file, offset, len, mode);
4903
4904         trace_ext4_fallocate_enter(inode, offset, len, mode);
4905         lblk = offset >> blkbits;
4906
4907         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4908         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4909         if (mode & FALLOC_FL_KEEP_SIZE)
4910                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4911
4912         inode_lock(inode);
4913
4914         /*
4915          * We only support preallocation for extent-based files only
4916          */
4917         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4918                 ret = -EOPNOTSUPP;
4919                 goto out;
4920         }
4921
4922         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4923             (offset + len > i_size_read(inode) ||
4924              offset + len > EXT4_I(inode)->i_disksize)) {
4925                 new_size = offset + len;
4926                 ret = inode_newsize_ok(inode, new_size);
4927                 if (ret)
4928                         goto out;
4929         }
4930
4931         /* Wait all existing dio workers, newcomers will block on i_mutex */
4932         inode_dio_wait(inode);
4933
4934         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
4935         if (ret)
4936                 goto out;
4937
4938         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4939                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4940                                                 EXT4_I(inode)->i_sync_tid);
4941         }
4942 out:
4943         inode_unlock(inode);
4944         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4945         return ret;
4946 }
4947
4948 /*
4949  * This function convert a range of blocks to written extents
4950  * The caller of this function will pass the start offset and the size.
4951  * all unwritten extents within this range will be converted to
4952  * written extents.
4953  *
4954  * This function is called from the direct IO end io call back
4955  * function, to convert the fallocated extents after IO is completed.
4956  * Returns 0 on success.
4957  */
4958 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4959                                    loff_t offset, ssize_t len)
4960 {
4961         unsigned int max_blocks;
4962         int ret = 0;
4963         int ret2 = 0;
4964         struct ext4_map_blocks map;
4965         unsigned int blkbits = inode->i_blkbits;
4966         unsigned int credits = 0;
4967
4968         map.m_lblk = offset >> blkbits;
4969         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4970
4971         if (!handle) {
4972                 /*
4973                  * credits to insert 1 extent into extent tree
4974                  */
4975                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
4976         }
4977         while (ret >= 0 && ret < max_blocks) {
4978                 map.m_lblk += ret;
4979                 map.m_len = (max_blocks -= ret);
4980                 if (credits) {
4981                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4982                                                     credits);
4983                         if (IS_ERR(handle)) {
4984                                 ret = PTR_ERR(handle);
4985                                 break;
4986                         }
4987                 }
4988                 ret = ext4_map_blocks(handle, inode, &map,
4989                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4990                 if (ret <= 0)
4991                         ext4_warning(inode->i_sb,
4992                                      "inode #%lu: block %u: len %u: "
4993                                      "ext4_ext_map_blocks returned %d",
4994                                      inode->i_ino, map.m_lblk,
4995                                      map.m_len, ret);
4996                 ext4_mark_inode_dirty(handle, inode);
4997                 if (credits)
4998                         ret2 = ext4_journal_stop(handle);
4999                 if (ret <= 0 || ret2)
5000                         break;
5001         }
5002         return ret > 0 ? ret2 : ret;
5003 }
5004
5005 int ext4_convert_unwritten_io_end_vec(handle_t *handle, ext4_io_end_t *io_end)
5006 {
5007         int ret, err = 0;
5008
5009         /*
5010          * This is somewhat ugly but the idea is clear: When transaction is
5011          * reserved, everything goes into it. Otherwise we rather start several
5012          * smaller transactions for conversion of each extent separately.
5013          */
5014         if (handle) {
5015                 handle = ext4_journal_start_reserved(handle,
5016                                                      EXT4_HT_EXT_CONVERT);
5017                 if (IS_ERR(handle))
5018                         return PTR_ERR(handle);
5019         }
5020
5021         ret = ext4_convert_unwritten_extents(handle, io_end->inode,
5022                                              io_end->offset, io_end->size);
5023         if (handle)
5024                 err = ext4_journal_stop(handle);
5025
5026         return ret < 0 ? ret : err;
5027 }
5028
5029 /*
5030  * If newes is not existing extent (newes->ec_pblk equals zero) find
5031  * delayed extent at start of newes and update newes accordingly and
5032  * return start of the next delayed extent.
5033  *
5034  * If newes is existing extent (newes->ec_pblk is not equal zero)
5035  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5036  * extent found. Leave newes unmodified.
5037  */
5038 static int ext4_find_delayed_extent(struct inode *inode,
5039                                     struct extent_status *newes)
5040 {
5041         struct extent_status es;
5042         ext4_lblk_t block, next_del;
5043
5044         if (newes->es_pblk == 0) {
5045                 ext4_es_find_extent_range(inode, &ext4_es_is_delayed,
5046                                           newes->es_lblk,
5047                                           newes->es_lblk + newes->es_len - 1,
5048                                           &es);
5049
5050                 /*
5051                  * No extent in extent-tree contains block @newes->es_pblk,
5052                  * then the block may stay in 1)a hole or 2)delayed-extent.
5053                  */
5054                 if (es.es_len == 0)
5055                         /* A hole found. */
5056                         return 0;
5057
5058                 if (es.es_lblk > newes->es_lblk) {
5059                         /* A hole found. */
5060                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5061                                             newes->es_len);
5062                         return 0;
5063                 }
5064
5065                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5066         }
5067
5068         block = newes->es_lblk + newes->es_len;
5069         ext4_es_find_extent_range(inode, &ext4_es_is_delayed, block,
5070                                   EXT_MAX_BLOCKS, &es);
5071         if (es.es_len == 0)
5072                 next_del = EXT_MAX_BLOCKS;
5073         else
5074                 next_del = es.es_lblk;
5075
5076         return next_del;
5077 }
5078
5079 static int ext4_xattr_fiemap(struct inode *inode,
5080                                 struct fiemap_extent_info *fieinfo)
5081 {
5082         __u64 physical = 0;
5083         __u64 length;
5084         __u32 flags = FIEMAP_EXTENT_LAST;
5085         int blockbits = inode->i_sb->s_blocksize_bits;
5086         int error = 0;
5087
5088         /* in-inode? */
5089         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5090                 struct ext4_iloc iloc;
5091                 int offset;     /* offset of xattr in inode */
5092
5093                 error = ext4_get_inode_loc(inode, &iloc);
5094                 if (error)
5095                         return error;
5096                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5097                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5098                                 EXT4_I(inode)->i_extra_isize;
5099                 physical += offset;
5100                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5101                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5102                 brelse(iloc.bh);
5103         } else { /* external block */
5104                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5105                 length = inode->i_sb->s_blocksize;
5106         }
5107
5108         if (physical)
5109                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5110                                                 length, flags);
5111         return (error < 0 ? error : 0);
5112 }
5113
5114 static int _ext4_fiemap(struct inode *inode,
5115                         struct fiemap_extent_info *fieinfo,
5116                         __u64 start, __u64 len,
5117                         int (*fill)(struct inode *, ext4_lblk_t,
5118                                     ext4_lblk_t,
5119                                     struct fiemap_extent_info *))
5120 {
5121         ext4_lblk_t start_blk;
5122         u32 ext4_fiemap_flags = FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR;
5123
5124         int error = 0;
5125
5126         if (ext4_has_inline_data(inode)) {
5127                 int has_inline = 1;
5128
5129                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5130                                                 start, len);
5131
5132                 if (has_inline)
5133                         return error;
5134         }
5135
5136         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5137                 error = ext4_ext_precache(inode);
5138                 if (error)
5139                         return error;
5140                 fieinfo->fi_flags &= ~FIEMAP_FLAG_CACHE;
5141         }
5142
5143         /* fallback to generic here if not in extents fmt */
5144         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
5145             fill == ext4_fill_fiemap_extents)
5146                 return generic_block_fiemap(inode, fieinfo, start, len,
5147                         ext4_get_block);
5148
5149         if (fill == ext4_fill_es_cache_info)
5150                 ext4_fiemap_flags &= FIEMAP_FLAG_XATTR;
5151         if (fiemap_check_flags(fieinfo, ext4_fiemap_flags))
5152                 return -EBADR;
5153
5154         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5155                 error = ext4_xattr_fiemap(inode, fieinfo);
5156         } else {
5157                 ext4_lblk_t len_blks;
5158                 __u64 last_blk;
5159
5160                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5161                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5162                 if (last_blk >= EXT_MAX_BLOCKS)
5163                         last_blk = EXT_MAX_BLOCKS-1;
5164                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5165
5166                 /*
5167                  * Walk the extent tree gathering extent information
5168                  * and pushing extents back to the user.
5169                  */
5170                 error = fill(inode, start_blk, len_blks, fieinfo);
5171         }
5172         return error;
5173 }
5174
5175 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5176                 __u64 start, __u64 len)
5177 {
5178         return _ext4_fiemap(inode, fieinfo, start, len,
5179                             ext4_fill_fiemap_extents);
5180 }
5181
5182 int ext4_get_es_cache(struct inode *inode, struct fiemap_extent_info *fieinfo,
5183                       __u64 start, __u64 len)
5184 {
5185         if (ext4_has_inline_data(inode)) {
5186                 int has_inline;
5187
5188                 down_read(&EXT4_I(inode)->xattr_sem);
5189                 has_inline = ext4_has_inline_data(inode);
5190                 up_read(&EXT4_I(inode)->xattr_sem);
5191                 if (has_inline)
5192                         return 0;
5193         }
5194
5195         return _ext4_fiemap(inode, fieinfo, start, len,
5196                             ext4_fill_es_cache_info);
5197 }
5198
5199
5200 /*
5201  * ext4_access_path:
5202  * Function to access the path buffer for marking it dirty.
5203  * It also checks if there are sufficient credits left in the journal handle
5204  * to update path.
5205  */
5206 static int
5207 ext4_access_path(handle_t *handle, struct inode *inode,
5208                 struct ext4_ext_path *path)
5209 {
5210         int credits, err;
5211
5212         if (!ext4_handle_valid(handle))
5213                 return 0;
5214
5215         /*
5216          * Check if need to extend journal credits
5217          * 3 for leaf, sb, and inode plus 2 (bmap and group
5218          * descriptor) for each block group; assume two block
5219          * groups
5220          */
5221         if (handle->h_buffer_credits < 7) {
5222                 credits = ext4_writepage_trans_blocks(inode);
5223                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5224                 /* EAGAIN is success */
5225                 if (err && err != -EAGAIN)
5226                         return err;
5227         }
5228
5229         err = ext4_ext_get_access(handle, inode, path);
5230         return err;
5231 }
5232
5233 /*
5234  * ext4_ext_shift_path_extents:
5235  * Shift the extents of a path structure lying between path[depth].p_ext
5236  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5237  * if it is right shift or left shift operation.
5238  */
5239 static int
5240 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5241                             struct inode *inode, handle_t *handle,
5242                             enum SHIFT_DIRECTION SHIFT)
5243 {
5244         int depth, err = 0;
5245         struct ext4_extent *ex_start, *ex_last;
5246         bool update = 0;
5247         depth = path->p_depth;
5248
5249         while (depth >= 0) {
5250                 if (depth == path->p_depth) {
5251                         ex_start = path[depth].p_ext;
5252                         if (!ex_start)
5253                                 return -EFSCORRUPTED;
5254
5255                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5256
5257                         err = ext4_access_path(handle, inode, path + depth);
5258                         if (err)
5259                                 goto out;
5260
5261                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5262                                 update = 1;
5263
5264                         while (ex_start <= ex_last) {
5265                                 if (SHIFT == SHIFT_LEFT) {
5266                                         le32_add_cpu(&ex_start->ee_block,
5267                                                 -shift);
5268                                         /* Try to merge to the left. */
5269                                         if ((ex_start >
5270                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5271                                             &&
5272                                             ext4_ext_try_to_merge_right(inode,
5273                                             path, ex_start - 1))
5274                                                 ex_last--;
5275                                         else
5276                                                 ex_start++;
5277                                 } else {
5278                                         le32_add_cpu(&ex_last->ee_block, shift);
5279                                         ext4_ext_try_to_merge_right(inode, path,
5280                                                 ex_last);
5281                                         ex_last--;
5282                                 }
5283                         }
5284                         err = ext4_ext_dirty(handle, inode, path + depth);
5285                         if (err)
5286                                 goto out;
5287
5288                         if (--depth < 0 || !update)
5289                                 break;
5290                 }
5291
5292                 /* Update index too */
5293                 err = ext4_access_path(handle, inode, path + depth);
5294                 if (err)
5295                         goto out;
5296
5297                 if (SHIFT == SHIFT_LEFT)
5298                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5299                 else
5300                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5301                 err = ext4_ext_dirty(handle, inode, path + depth);
5302                 if (err)
5303                         goto out;
5304
5305                 /* we are done if current index is not a starting index */
5306                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5307                         break;
5308
5309                 depth--;
5310         }
5311
5312 out:
5313         return err;
5314 }
5315
5316 /*
5317  * ext4_ext_shift_extents:
5318  * All the extents which lies in the range from @start to the last allocated
5319  * block for the @inode are shifted either towards left or right (depending
5320  * upon @SHIFT) by @shift blocks.
5321  * On success, 0 is returned, error otherwise.
5322  */
5323 static int
5324 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5325                        ext4_lblk_t start, ext4_lblk_t shift,
5326                        enum SHIFT_DIRECTION SHIFT)
5327 {
5328         struct ext4_ext_path *path;
5329         int ret = 0, depth;
5330         struct ext4_extent *extent;
5331         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5332
5333         /* Let path point to the last extent */
5334         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5335                                 EXT4_EX_NOCACHE);
5336         if (IS_ERR(path))
5337                 return PTR_ERR(path);
5338
5339         depth = path->p_depth;
5340         extent = path[depth].p_ext;
5341         if (!extent)
5342                 goto out;
5343
5344         stop = le32_to_cpu(extent->ee_block);
5345
5346        /*
5347         * For left shifts, make sure the hole on the left is big enough to
5348         * accommodate the shift.  For right shifts, make sure the last extent
5349         * won't be shifted beyond EXT_MAX_BLOCKS.
5350         */
5351         if (SHIFT == SHIFT_LEFT) {
5352                 path = ext4_find_extent(inode, start - 1, &path,
5353                                         EXT4_EX_NOCACHE);
5354                 if (IS_ERR(path))
5355                         return PTR_ERR(path);
5356                 depth = path->p_depth;
5357                 extent =  path[depth].p_ext;
5358                 if (extent) {
5359                         ex_start = le32_to_cpu(extent->ee_block);
5360                         ex_end = le32_to_cpu(extent->ee_block) +
5361                                 ext4_ext_get_actual_len(extent);
5362                 } else {
5363                         ex_start = 0;
5364                         ex_end = 0;
5365                 }
5366
5367                 if ((start == ex_start && shift > ex_start) ||
5368                     (shift > start - ex_end)) {
5369                         ret = -EINVAL;
5370                         goto out;
5371                 }
5372         } else {
5373                 if (shift > EXT_MAX_BLOCKS -
5374                     (stop + ext4_ext_get_actual_len(extent))) {
5375                         ret = -EINVAL;
5376                         goto out;
5377                 }
5378         }
5379
5380         /*
5381          * In case of left shift, iterator points to start and it is increased
5382          * till we reach stop. In case of right shift, iterator points to stop
5383          * and it is decreased till we reach start.
5384          */
5385         if (SHIFT == SHIFT_LEFT)
5386                 iterator = &start;
5387         else
5388                 iterator = &stop;
5389
5390         /*
5391          * Its safe to start updating extents.  Start and stop are unsigned, so
5392          * in case of right shift if extent with 0 block is reached, iterator
5393          * becomes NULL to indicate the end of the loop.
5394          */
5395         while (iterator && start <= stop) {
5396                 path = ext4_find_extent(inode, *iterator, &path,
5397                                         EXT4_EX_NOCACHE);
5398                 if (IS_ERR(path))
5399                         return PTR_ERR(path);
5400                 depth = path->p_depth;
5401                 extent = path[depth].p_ext;
5402                 if (!extent) {
5403                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5404                                          (unsigned long) *iterator);
5405                         return -EFSCORRUPTED;
5406                 }
5407                 if (SHIFT == SHIFT_LEFT && *iterator >
5408                     le32_to_cpu(extent->ee_block)) {
5409                         /* Hole, move to the next extent */
5410                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5411                                 path[depth].p_ext++;
5412                         } else {
5413                                 *iterator = ext4_ext_next_allocated_block(path);
5414                                 continue;
5415                         }
5416                 }
5417
5418                 if (SHIFT == SHIFT_LEFT) {
5419                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5420                         *iterator = le32_to_cpu(extent->ee_block) +
5421                                         ext4_ext_get_actual_len(extent);
5422                 } else {
5423                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5424                         if (le32_to_cpu(extent->ee_block) > 0)
5425                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5426                         else
5427                                 /* Beginning is reached, end of the loop */
5428                                 iterator = NULL;
5429                         /* Update path extent in case we need to stop */
5430                         while (le32_to_cpu(extent->ee_block) < start)
5431                                 extent++;
5432                         path[depth].p_ext = extent;
5433                 }
5434                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5435                                 handle, SHIFT);
5436                 if (ret)
5437                         break;
5438         }
5439 out:
5440         ext4_ext_drop_refs(path);
5441         kfree(path);
5442         return ret;
5443 }
5444
5445 /*
5446  * ext4_collapse_range:
5447  * This implements the fallocate's collapse range functionality for ext4
5448  * Returns: 0 and non-zero on error.
5449  */
5450 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5451 {
5452         struct super_block *sb = inode->i_sb;
5453         ext4_lblk_t punch_start, punch_stop;
5454         handle_t *handle;
5455         unsigned int credits;
5456         loff_t new_size, ioffset;
5457         int ret;
5458
5459         /*
5460          * We need to test this early because xfstests assumes that a
5461          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5462          * system does not support collapse range.
5463          */
5464         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5465                 return -EOPNOTSUPP;
5466
5467         /* Collapse range works only on fs block size aligned offsets. */
5468         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5469             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5470                 return -EINVAL;
5471
5472         if (!S_ISREG(inode->i_mode))
5473                 return -EINVAL;
5474
5475         trace_ext4_collapse_range(inode, offset, len);
5476
5477         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5478         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5479
5480         /* Call ext4_force_commit to flush all data in case of data=journal. */
5481         if (ext4_should_journal_data(inode)) {
5482                 ret = ext4_force_commit(inode->i_sb);
5483                 if (ret)
5484                         return ret;
5485         }
5486
5487         inode_lock(inode);
5488         /*
5489          * There is no need to overlap collapse range with EOF, in which case
5490          * it is effectively a truncate operation
5491          */
5492         if (offset + len >= i_size_read(inode)) {
5493                 ret = -EINVAL;
5494                 goto out_mutex;
5495         }
5496
5497         /* Currently just for extent based files */
5498         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5499                 ret = -EOPNOTSUPP;
5500                 goto out_mutex;
5501         }
5502
5503         /* Wait for existing dio to complete */
5504         inode_dio_wait(inode);
5505
5506         /*
5507          * Prevent page faults from reinstantiating pages we have released from
5508          * page cache.
5509          */
5510         down_write(&EXT4_I(inode)->i_mmap_sem);
5511
5512         ret = ext4_break_layouts(inode);
5513         if (ret)
5514                 goto out_mmap;
5515
5516         /*
5517          * Need to round down offset to be aligned with page size boundary
5518          * for page size > block size.
5519          */
5520         ioffset = round_down(offset, PAGE_SIZE);
5521         /*
5522          * Write tail of the last page before removed range since it will get
5523          * removed from the page cache below.
5524          */
5525         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5526         if (ret)
5527                 goto out_mmap;
5528         /*
5529          * Write data that will be shifted to preserve them when discarding
5530          * page cache below. We are also protected from pages becoming dirty
5531          * by i_mmap_sem.
5532          */
5533         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5534                                            LLONG_MAX);
5535         if (ret)
5536                 goto out_mmap;
5537         truncate_pagecache(inode, ioffset);
5538
5539         credits = ext4_writepage_trans_blocks(inode);
5540         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5541         if (IS_ERR(handle)) {
5542                 ret = PTR_ERR(handle);
5543                 goto out_mmap;
5544         }
5545
5546         down_write(&EXT4_I(inode)->i_data_sem);
5547         ext4_discard_preallocations(inode);
5548
5549         ret = ext4_es_remove_extent(inode, punch_start,
5550                                     EXT_MAX_BLOCKS - punch_start);
5551         if (ret) {
5552                 up_write(&EXT4_I(inode)->i_data_sem);
5553                 goto out_stop;
5554         }
5555
5556         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5557         if (ret) {
5558                 up_write(&EXT4_I(inode)->i_data_sem);
5559                 goto out_stop;
5560         }
5561         ext4_discard_preallocations(inode);
5562
5563         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5564                                      punch_stop - punch_start, SHIFT_LEFT);
5565         if (ret) {
5566                 up_write(&EXT4_I(inode)->i_data_sem);
5567                 goto out_stop;
5568         }
5569
5570         new_size = i_size_read(inode) - len;
5571         i_size_write(inode, new_size);
5572         EXT4_I(inode)->i_disksize = new_size;
5573
5574         up_write(&EXT4_I(inode)->i_data_sem);
5575         if (IS_SYNC(inode))
5576                 ext4_handle_sync(handle);
5577         inode->i_mtime = inode->i_ctime = current_time(inode);
5578         ext4_mark_inode_dirty(handle, inode);
5579         ext4_update_inode_fsync_trans(handle, inode, 1);
5580
5581 out_stop:
5582         ext4_journal_stop(handle);
5583 out_mmap:
5584         up_write(&EXT4_I(inode)->i_mmap_sem);
5585 out_mutex:
5586         inode_unlock(inode);
5587         return ret;
5588 }
5589
5590 /*
5591  * ext4_insert_range:
5592  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5593  * The data blocks starting from @offset to the EOF are shifted by @len
5594  * towards right to create a hole in the @inode. Inode size is increased
5595  * by len bytes.
5596  * Returns 0 on success, error otherwise.
5597  */
5598 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5599 {
5600         struct super_block *sb = inode->i_sb;
5601         handle_t *handle;
5602         struct ext4_ext_path *path;
5603         struct ext4_extent *extent;
5604         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5605         unsigned int credits, ee_len;
5606         int ret = 0, depth, split_flag = 0;
5607         loff_t ioffset;
5608
5609         /*
5610          * We need to test this early because xfstests assumes that an
5611          * insert range of (0, 1) will return EOPNOTSUPP if the file
5612          * system does not support insert range.
5613          */
5614         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5615                 return -EOPNOTSUPP;
5616
5617         /* Insert range works only on fs block size aligned offsets. */
5618         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5619                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5620                 return -EINVAL;
5621
5622         if (!S_ISREG(inode->i_mode))
5623                 return -EOPNOTSUPP;
5624
5625         trace_ext4_insert_range(inode, offset, len);
5626
5627         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5628         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5629
5630         /* Call ext4_force_commit to flush all data in case of data=journal */
5631         if (ext4_should_journal_data(inode)) {
5632                 ret = ext4_force_commit(inode->i_sb);
5633                 if (ret)
5634                         return ret;
5635         }
5636
5637         inode_lock(inode);
5638         /* Currently just for extent based files */
5639         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5640                 ret = -EOPNOTSUPP;
5641                 goto out_mutex;
5642         }
5643
5644         /* Check for wrap through zero */
5645         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5646                 ret = -EFBIG;
5647                 goto out_mutex;
5648         }
5649
5650         /* Offset should be less than i_size */
5651         if (offset >= i_size_read(inode)) {
5652                 ret = -EINVAL;
5653                 goto out_mutex;
5654         }
5655
5656         /* Wait for existing dio to complete */
5657         inode_dio_wait(inode);
5658
5659         /*
5660          * Prevent page faults from reinstantiating pages we have released from
5661          * page cache.
5662          */
5663         down_write(&EXT4_I(inode)->i_mmap_sem);
5664
5665         ret = ext4_break_layouts(inode);
5666         if (ret)
5667                 goto out_mmap;
5668
5669         /*
5670          * Need to round down to align start offset to page size boundary
5671          * for page size > block size.
5672          */
5673         ioffset = round_down(offset, PAGE_SIZE);
5674         /* Write out all dirty pages */
5675         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5676                         LLONG_MAX);
5677         if (ret)
5678                 goto out_mmap;
5679         truncate_pagecache(inode, ioffset);
5680
5681         credits = ext4_writepage_trans_blocks(inode);
5682         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5683         if (IS_ERR(handle)) {
5684                 ret = PTR_ERR(handle);
5685                 goto out_mmap;
5686         }
5687
5688         /* Expand file to avoid data loss if there is error while shifting */
5689         inode->i_size += len;
5690         EXT4_I(inode)->i_disksize += len;
5691         inode->i_mtime = inode->i_ctime = current_time(inode);
5692         ret = ext4_mark_inode_dirty(handle, inode);
5693         if (ret)
5694                 goto out_stop;
5695
5696         down_write(&EXT4_I(inode)->i_data_sem);
5697         ext4_discard_preallocations(inode);
5698
5699         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5700         if (IS_ERR(path)) {
5701                 up_write(&EXT4_I(inode)->i_data_sem);
5702                 goto out_stop;
5703         }
5704
5705         depth = ext_depth(inode);
5706         extent = path[depth].p_ext;
5707         if (extent) {
5708                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5709                 ee_len = ext4_ext_get_actual_len(extent);
5710
5711                 /*
5712                  * If offset_lblk is not the starting block of extent, split
5713                  * the extent @offset_lblk
5714                  */
5715                 if ((offset_lblk > ee_start_lblk) &&
5716                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5717                         if (ext4_ext_is_unwritten(extent))
5718                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5719                                         EXT4_EXT_MARK_UNWRIT2;
5720                         ret = ext4_split_extent_at(handle, inode, &path,
5721                                         offset_lblk, split_flag,
5722                                         EXT4_EX_NOCACHE |
5723                                         EXT4_GET_BLOCKS_PRE_IO |
5724                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5725                 }
5726
5727                 ext4_ext_drop_refs(path);
5728                 kfree(path);
5729                 if (ret < 0) {
5730                         up_write(&EXT4_I(inode)->i_data_sem);
5731                         goto out_stop;
5732                 }
5733         } else {
5734                 ext4_ext_drop_refs(path);
5735                 kfree(path);
5736         }
5737
5738         ret = ext4_es_remove_extent(inode, offset_lblk,
5739                         EXT_MAX_BLOCKS - offset_lblk);
5740         if (ret) {
5741                 up_write(&EXT4_I(inode)->i_data_sem);
5742                 goto out_stop;
5743         }
5744
5745         /*
5746          * if offset_lblk lies in a hole which is at start of file, use
5747          * ee_start_lblk to shift extents
5748          */
5749         ret = ext4_ext_shift_extents(inode, handle,
5750                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5751                 len_lblk, SHIFT_RIGHT);
5752
5753         up_write(&EXT4_I(inode)->i_data_sem);
5754         if (IS_SYNC(inode))
5755                 ext4_handle_sync(handle);
5756         if (ret >= 0)
5757                 ext4_update_inode_fsync_trans(handle, inode, 1);
5758
5759 out_stop:
5760         ext4_journal_stop(handle);
5761 out_mmap:
5762         up_write(&EXT4_I(inode)->i_mmap_sem);
5763 out_mutex:
5764         inode_unlock(inode);
5765         return ret;
5766 }
5767
5768 /**
5769  * ext4_swap_extents() - Swap extents between two inodes
5770  * @handle: handle for this transaction
5771  * @inode1:     First inode
5772  * @inode2:     Second inode
5773  * @lblk1:      Start block for first inode
5774  * @lblk2:      Start block for second inode
5775  * @count:      Number of blocks to swap
5776  * @unwritten: Mark second inode's extents as unwritten after swap
5777  * @erp:        Pointer to save error value
5778  *
5779  * This helper routine does exactly what is promise "swap extents". All other
5780  * stuff such as page-cache locking consistency, bh mapping consistency or
5781  * extent's data copying must be performed by caller.
5782  * Locking:
5783  *              i_mutex is held for both inodes
5784  *              i_data_sem is locked for write for both inodes
5785  * Assumptions:
5786  *              All pages from requested range are locked for both inodes
5787  */
5788 int
5789 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5790                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5791                   ext4_lblk_t count, int unwritten, int *erp)
5792 {
5793         struct ext4_ext_path *path1 = NULL;
5794         struct ext4_ext_path *path2 = NULL;
5795         int replaced_count = 0;
5796
5797         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5798         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5799         BUG_ON(!inode_is_locked(inode1));
5800         BUG_ON(!inode_is_locked(inode2));
5801
5802         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5803         if (unlikely(*erp))
5804                 return 0;
5805         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5806         if (unlikely(*erp))
5807                 return 0;
5808
5809         while (count) {
5810                 struct ext4_extent *ex1, *ex2, tmp_ex;
5811                 ext4_lblk_t e1_blk, e2_blk;
5812                 int e1_len, e2_len, len;
5813                 int split = 0;
5814
5815                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5816                 if (IS_ERR(path1)) {
5817                         *erp = PTR_ERR(path1);
5818                         path1 = NULL;
5819                 finish:
5820                         count = 0;
5821                         goto repeat;
5822                 }
5823                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5824                 if (IS_ERR(path2)) {
5825                         *erp = PTR_ERR(path2);
5826                         path2 = NULL;
5827                         goto finish;
5828                 }
5829                 ex1 = path1[path1->p_depth].p_ext;
5830                 ex2 = path2[path2->p_depth].p_ext;
5831                 /* Do we have somthing to swap ? */
5832                 if (unlikely(!ex2 || !ex1))
5833                         goto finish;
5834
5835                 e1_blk = le32_to_cpu(ex1->ee_block);
5836                 e2_blk = le32_to_cpu(ex2->ee_block);
5837                 e1_len = ext4_ext_get_actual_len(ex1);
5838                 e2_len = ext4_ext_get_actual_len(ex2);
5839
5840                 /* Hole handling */
5841                 if (!in_range(lblk1, e1_blk, e1_len) ||
5842                     !in_range(lblk2, e2_blk, e2_len)) {
5843                         ext4_lblk_t next1, next2;
5844
5845                         /* if hole after extent, then go to next extent */
5846                         next1 = ext4_ext_next_allocated_block(path1);
5847                         next2 = ext4_ext_next_allocated_block(path2);
5848                         /* If hole before extent, then shift to that extent */
5849                         if (e1_blk > lblk1)
5850                                 next1 = e1_blk;
5851                         if (e2_blk > lblk2)
5852                                 next2 = e2_blk;
5853                         /* Do we have something to swap */
5854                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5855                                 goto finish;
5856                         /* Move to the rightest boundary */
5857                         len = next1 - lblk1;
5858                         if (len < next2 - lblk2)
5859                                 len = next2 - lblk2;
5860                         if (len > count)
5861                                 len = count;
5862                         lblk1 += len;
5863                         lblk2 += len;
5864                         count -= len;
5865                         goto repeat;
5866                 }
5867
5868                 /* Prepare left boundary */
5869                 if (e1_blk < lblk1) {
5870                         split = 1;
5871                         *erp = ext4_force_split_extent_at(handle, inode1,
5872                                                 &path1, lblk1, 0);
5873                         if (unlikely(*erp))
5874                                 goto finish;
5875                 }
5876                 if (e2_blk < lblk2) {
5877                         split = 1;
5878                         *erp = ext4_force_split_extent_at(handle, inode2,
5879                                                 &path2,  lblk2, 0);
5880                         if (unlikely(*erp))
5881                                 goto finish;
5882                 }
5883                 /* ext4_split_extent_at() may result in leaf extent split,
5884                  * path must to be revalidated. */
5885                 if (split)
5886                         goto repeat;
5887
5888                 /* Prepare right boundary */
5889                 len = count;
5890                 if (len > e1_blk + e1_len - lblk1)
5891                         len = e1_blk + e1_len - lblk1;
5892                 if (len > e2_blk + e2_len - lblk2)
5893                         len = e2_blk + e2_len - lblk2;
5894
5895                 if (len != e1_len) {
5896                         split = 1;
5897                         *erp = ext4_force_split_extent_at(handle, inode1,
5898                                                 &path1, lblk1 + len, 0);
5899                         if (unlikely(*erp))
5900                                 goto finish;
5901                 }
5902                 if (len != e2_len) {
5903                         split = 1;
5904                         *erp = ext4_force_split_extent_at(handle, inode2,
5905                                                 &path2, lblk2 + len, 0);
5906                         if (*erp)
5907                                 goto finish;
5908                 }
5909                 /* ext4_split_extent_at() may result in leaf extent split,
5910                  * path must to be revalidated. */
5911                 if (split)
5912                         goto repeat;
5913
5914                 BUG_ON(e2_len != e1_len);
5915                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5916                 if (unlikely(*erp))
5917                         goto finish;
5918                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5919                 if (unlikely(*erp))
5920                         goto finish;
5921
5922                 /* Both extents are fully inside boundaries. Swap it now */
5923                 tmp_ex = *ex1;
5924                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5925                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5926                 ex1->ee_len = cpu_to_le16(e2_len);
5927                 ex2->ee_len = cpu_to_le16(e1_len);
5928                 if (unwritten)
5929                         ext4_ext_mark_unwritten(ex2);
5930                 if (ext4_ext_is_unwritten(&tmp_ex))
5931                         ext4_ext_mark_unwritten(ex1);
5932
5933                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5934                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5935                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5936                                       path2->p_depth);
5937                 if (unlikely(*erp))
5938                         goto finish;
5939                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5940                                       path1->p_depth);
5941                 /*
5942                  * Looks scarry ah..? second inode already points to new blocks,
5943                  * and it was successfully dirtied. But luckily error may happen
5944                  * only due to journal error, so full transaction will be
5945                  * aborted anyway.
5946                  */
5947                 if (unlikely(*erp))
5948                         goto finish;
5949                 lblk1 += len;
5950                 lblk2 += len;
5951                 replaced_count += len;
5952                 count -= len;
5953
5954         repeat:
5955                 ext4_ext_drop_refs(path1);
5956                 kfree(path1);
5957                 ext4_ext_drop_refs(path2);
5958                 kfree(path2);
5959                 path1 = path2 = NULL;
5960         }
5961         return replaced_count;
5962 }
5963
5964 /*
5965  * ext4_clu_mapped - determine whether any block in a logical cluster has
5966  *                   been mapped to a physical cluster
5967  *
5968  * @inode - file containing the logical cluster
5969  * @lclu - logical cluster of interest
5970  *
5971  * Returns 1 if any block in the logical cluster is mapped, signifying
5972  * that a physical cluster has been allocated for it.  Otherwise,
5973  * returns 0.  Can also return negative error codes.  Derived from
5974  * ext4_ext_map_blocks().
5975  */
5976 int ext4_clu_mapped(struct inode *inode, ext4_lblk_t lclu)
5977 {
5978         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5979         struct ext4_ext_path *path;
5980         int depth, mapped = 0, err = 0;
5981         struct ext4_extent *extent;
5982         ext4_lblk_t first_lblk, first_lclu, last_lclu;
5983
5984         /* search for the extent closest to the first block in the cluster */
5985         path = ext4_find_extent(inode, EXT4_C2B(sbi, lclu), NULL, 0);
5986         if (IS_ERR(path)) {
5987                 err = PTR_ERR(path);
5988                 path = NULL;
5989                 goto out;
5990         }
5991
5992         depth = ext_depth(inode);
5993
5994         /*
5995          * A consistent leaf must not be empty.  This situation is possible,
5996          * though, _during_ tree modification, and it's why an assert can't
5997          * be put in ext4_find_extent().
5998          */
5999         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
6000                 EXT4_ERROR_INODE(inode,
6001                     "bad extent address - lblock: %lu, depth: %d, pblock: %lld",
6002                                  (unsigned long) EXT4_C2B(sbi, lclu),
6003                                  depth, path[depth].p_block);
6004                 err = -EFSCORRUPTED;
6005                 goto out;
6006         }
6007
6008         extent = path[depth].p_ext;
6009
6010         /* can't be mapped if the extent tree is empty */
6011         if (extent == NULL)
6012                 goto out;
6013
6014         first_lblk = le32_to_cpu(extent->ee_block);
6015         first_lclu = EXT4_B2C(sbi, first_lblk);
6016
6017         /*
6018          * Three possible outcomes at this point - found extent spanning
6019          * the target cluster, to the left of the target cluster, or to the
6020          * right of the target cluster.  The first two cases are handled here.
6021          * The last case indicates the target cluster is not mapped.
6022          */
6023         if (lclu >= first_lclu) {
6024                 last_lclu = EXT4_B2C(sbi, first_lblk +
6025                                      ext4_ext_get_actual_len(extent) - 1);
6026                 if (lclu <= last_lclu) {
6027                         mapped = 1;
6028                 } else {
6029                         first_lblk = ext4_ext_next_allocated_block(path);
6030                         first_lclu = EXT4_B2C(sbi, first_lblk);
6031                         if (lclu == first_lclu)
6032                                 mapped = 1;
6033                 }
6034         }
6035
6036 out:
6037         ext4_ext_drop_refs(path);
6038         kfree(path);
6039
6040         return err ? err : mapped;
6041 }