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 /*
2319  * ext4_ext_determine_hole - determine hole around given block
2320  * @inode:      inode we lookup in
2321  * @path:       path in extent tree to @lblk
2322  * @lblk:       pointer to logical block around which we want to determine hole
2323  *
2324  * Determine hole length (and start if easily possible) around given logical
2325  * block. We don't try too hard to find the beginning of the hole but @path
2326  * actually points to extent before @lblk, we provide it.
2327  *
2328  * The function returns the length of a hole starting at @lblk. We update @lblk
2329  * to the beginning of the hole if we managed to find it.
2330  */
2331 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2332                                            struct ext4_ext_path *path,
2333                                            ext4_lblk_t *lblk)
2334 {
2335         int depth = ext_depth(inode);
2336         struct ext4_extent *ex;
2337         ext4_lblk_t len;
2338
2339         ex = path[depth].p_ext;
2340         if (ex == NULL) {
2341                 /* there is no extent yet, so gap is [0;-] */
2342                 *lblk = 0;
2343                 len = EXT_MAX_BLOCKS;
2344         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2345                 len = le32_to_cpu(ex->ee_block) - *lblk;
2346         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2347                         + ext4_ext_get_actual_len(ex)) {
2348                 ext4_lblk_t next;
2349
2350                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2351                 next = ext4_ext_next_allocated_block(path);
2352                 BUG_ON(next == *lblk);
2353                 len = next - *lblk;
2354         } else {
2355                 BUG();
2356         }
2357         return len;
2358 }
2359
2360 /*
2361  * ext4_ext_put_gap_in_cache:
2362  * calculate boundaries of the gap that the requested block fits into
2363  * and cache this gap
2364  */
2365 static void
2366 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2367                           ext4_lblk_t hole_len)
2368 {
2369         struct extent_status es;
2370
2371         ext4_es_find_extent_range(inode, &ext4_es_is_delayed, hole_start,
2372                                   hole_start + hole_len - 1, &es);
2373         if (es.es_len) {
2374                 /* There's delayed extent containing lblock? */
2375                 if (es.es_lblk <= hole_start)
2376                         return;
2377                 hole_len = min(es.es_lblk - hole_start, hole_len);
2378         }
2379         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2380         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2381                               EXTENT_STATUS_HOLE);
2382 }
2383
2384 /*
2385  * ext4_ext_rm_idx:
2386  * removes index from the index block.
2387  */
2388 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2389                         struct ext4_ext_path *path, int depth)
2390 {
2391         int err;
2392         ext4_fsblk_t leaf;
2393
2394         /* free index block */
2395         depth--;
2396         path = path + depth;
2397         leaf = ext4_idx_pblock(path->p_idx);
2398         if (unlikely(path->p_hdr->eh_entries == 0)) {
2399                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2400                 return -EFSCORRUPTED;
2401         }
2402         err = ext4_ext_get_access(handle, inode, path);
2403         if (err)
2404                 return err;
2405
2406         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2407                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2408                 len *= sizeof(struct ext4_extent_idx);
2409                 memmove(path->p_idx, path->p_idx + 1, len);
2410         }
2411
2412         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2413         err = ext4_ext_dirty(handle, inode, path);
2414         if (err)
2415                 return err;
2416         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2417         trace_ext4_ext_rm_idx(inode, leaf);
2418
2419         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2420                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2421
2422         while (--depth >= 0) {
2423                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2424                         break;
2425                 path--;
2426                 err = ext4_ext_get_access(handle, inode, path);
2427                 if (err)
2428                         break;
2429                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2430                 err = ext4_ext_dirty(handle, inode, path);
2431                 if (err)
2432                         break;
2433         }
2434         return err;
2435 }
2436
2437 /*
2438  * ext4_ext_calc_credits_for_single_extent:
2439  * This routine returns max. credits that needed to insert an extent
2440  * to the extent tree.
2441  * When pass the actual path, the caller should calculate credits
2442  * under i_data_sem.
2443  */
2444 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2445                                                 struct ext4_ext_path *path)
2446 {
2447         if (path) {
2448                 int depth = ext_depth(inode);
2449                 int ret = 0;
2450
2451                 /* probably there is space in leaf? */
2452                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2453                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2454
2455                         /*
2456                          *  There are some space in the leaf tree, no
2457                          *  need to account for leaf block credit
2458                          *
2459                          *  bitmaps and block group descriptor blocks
2460                          *  and other metadata blocks still need to be
2461                          *  accounted.
2462                          */
2463                         /* 1 bitmap, 1 block group descriptor */
2464                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2465                         return ret;
2466                 }
2467         }
2468
2469         return ext4_chunk_trans_blocks(inode, nrblocks);
2470 }
2471
2472 /*
2473  * How many index/leaf blocks need to change/allocate to add @extents extents?
2474  *
2475  * If we add a single extent, then in the worse case, each tree level
2476  * index/leaf need to be changed in case of the tree split.
2477  *
2478  * If more extents are inserted, they could cause the whole tree split more
2479  * than once, but this is really rare.
2480  */
2481 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2482 {
2483         int index;
2484         int depth;
2485
2486         /* If we are converting the inline data, only one is needed here. */
2487         if (ext4_has_inline_data(inode))
2488                 return 1;
2489
2490         depth = ext_depth(inode);
2491
2492         if (extents <= 1)
2493                 index = depth * 2;
2494         else
2495                 index = depth * 3;
2496
2497         return index;
2498 }
2499
2500 static inline int get_default_free_blocks_flags(struct inode *inode)
2501 {
2502         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2503             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2504                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2505         else if (ext4_should_journal_data(inode))
2506                 return EXT4_FREE_BLOCKS_FORGET;
2507         return 0;
2508 }
2509
2510 /*
2511  * ext4_rereserve_cluster - increment the reserved cluster count when
2512  *                          freeing a cluster with a pending reservation
2513  *
2514  * @inode - file containing the cluster
2515  * @lblk - logical block in cluster to be reserved
2516  *
2517  * Increments the reserved cluster count and adjusts quota in a bigalloc
2518  * file system when freeing a partial cluster containing at least one
2519  * delayed and unwritten block.  A partial cluster meeting that
2520  * requirement will have a pending reservation.  If so, the
2521  * RERESERVE_CLUSTER flag is used when calling ext4_free_blocks() to
2522  * defer reserved and allocated space accounting to a subsequent call
2523  * to this function.
2524  */
2525 static void ext4_rereserve_cluster(struct inode *inode, ext4_lblk_t lblk)
2526 {
2527         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2528         struct ext4_inode_info *ei = EXT4_I(inode);
2529
2530         dquot_reclaim_block(inode, EXT4_C2B(sbi, 1));
2531
2532         spin_lock(&ei->i_block_reservation_lock);
2533         ei->i_reserved_data_blocks++;
2534         percpu_counter_add(&sbi->s_dirtyclusters_counter, 1);
2535         spin_unlock(&ei->i_block_reservation_lock);
2536
2537         percpu_counter_add(&sbi->s_freeclusters_counter, 1);
2538         ext4_remove_pending(inode, lblk);
2539 }
2540
2541 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2542                               struct ext4_extent *ex,
2543                               struct partial_cluster *partial,
2544                               ext4_lblk_t from, ext4_lblk_t to)
2545 {
2546         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2547         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2548         ext4_fsblk_t last_pblk, pblk;
2549         ext4_lblk_t num;
2550         int flags;
2551
2552         /* only extent tail removal is allowed */
2553         if (from < le32_to_cpu(ex->ee_block) ||
2554             to != le32_to_cpu(ex->ee_block) + ee_len - 1) {
2555                 ext4_error(sbi->s_sb,
2556                            "strange request: removal(2) %u-%u from %u:%u",
2557                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2558                 return 0;
2559         }
2560
2561 #ifdef EXTENTS_STATS
2562         spin_lock(&sbi->s_ext_stats_lock);
2563         sbi->s_ext_blocks += ee_len;
2564         sbi->s_ext_extents++;
2565         if (ee_len < sbi->s_ext_min)
2566                 sbi->s_ext_min = ee_len;
2567         if (ee_len > sbi->s_ext_max)
2568                 sbi->s_ext_max = ee_len;
2569         if (ext_depth(inode) > sbi->s_depth_max)
2570                 sbi->s_depth_max = ext_depth(inode);
2571         spin_unlock(&sbi->s_ext_stats_lock);
2572 #endif
2573
2574         trace_ext4_remove_blocks(inode, ex, from, to, partial);
2575
2576         /*
2577          * if we have a partial cluster, and it's different from the
2578          * cluster of the last block in the extent, we free it
2579          */
2580         last_pblk = ext4_ext_pblock(ex) + ee_len - 1;
2581
2582         if (partial->state != initial &&
2583             partial->pclu != EXT4_B2C(sbi, last_pblk)) {
2584                 if (partial->state == tofree) {
2585                         flags = get_default_free_blocks_flags(inode);
2586                         if (ext4_is_pending(inode, partial->lblk))
2587                                 flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2588                         ext4_free_blocks(handle, inode, NULL,
2589                                          EXT4_C2B(sbi, partial->pclu),
2590                                          sbi->s_cluster_ratio, flags);
2591                         if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2592                                 ext4_rereserve_cluster(inode, partial->lblk);
2593                 }
2594                 partial->state = initial;
2595         }
2596
2597         num = le32_to_cpu(ex->ee_block) + ee_len - from;
2598         pblk = ext4_ext_pblock(ex) + ee_len - num;
2599
2600         /*
2601          * We free the partial cluster at the end of the extent (if any),
2602          * unless the cluster is used by another extent (partial_cluster
2603          * state is nofree).  If a partial cluster exists here, it must be
2604          * shared with the last block in the extent.
2605          */
2606         flags = get_default_free_blocks_flags(inode);
2607
2608         /* partial, left end cluster aligned, right end unaligned */
2609         if ((EXT4_LBLK_COFF(sbi, to) != sbi->s_cluster_ratio - 1) &&
2610             (EXT4_LBLK_CMASK(sbi, to) >= from) &&
2611             (partial->state != nofree)) {
2612                 if (ext4_is_pending(inode, to))
2613                         flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2614                 ext4_free_blocks(handle, inode, NULL,
2615                                  EXT4_PBLK_CMASK(sbi, last_pblk),
2616                                  sbi->s_cluster_ratio, flags);
2617                 if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2618                         ext4_rereserve_cluster(inode, to);
2619                 partial->state = initial;
2620                 flags = get_default_free_blocks_flags(inode);
2621         }
2622
2623         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2624
2625         /*
2626          * For bigalloc file systems, we never free a partial cluster
2627          * at the beginning of the extent.  Instead, we check to see if we
2628          * need to free it on a subsequent call to ext4_remove_blocks,
2629          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2630          */
2631         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2632         ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2633
2634         /* reset the partial cluster if we've freed past it */
2635         if (partial->state != initial && partial->pclu != EXT4_B2C(sbi, pblk))
2636                 partial->state = initial;
2637
2638         /*
2639          * If we've freed the entire extent but the beginning is not left
2640          * cluster aligned and is not marked as ineligible for freeing we
2641          * record the partial cluster at the beginning of the extent.  It
2642          * wasn't freed by the preceding ext4_free_blocks() call, and we
2643          * need to look farther to the left to determine if it's to be freed
2644          * (not shared with another extent). Else, reset the partial
2645          * cluster - we're either  done freeing or the beginning of the
2646          * extent is left cluster aligned.
2647          */
2648         if (EXT4_LBLK_COFF(sbi, from) && num == ee_len) {
2649                 if (partial->state == initial) {
2650                         partial->pclu = EXT4_B2C(sbi, pblk);
2651                         partial->lblk = from;
2652                         partial->state = tofree;
2653                 }
2654         } else {
2655                 partial->state = initial;
2656         }
2657
2658         return 0;
2659 }
2660
2661 /*
2662  * ext4_ext_rm_leaf() Removes the extents associated with the
2663  * blocks appearing between "start" and "end".  Both "start"
2664  * and "end" must appear in the same extent or EIO is returned.
2665  *
2666  * @handle: The journal handle
2667  * @inode:  The files inode
2668  * @path:   The path to the leaf
2669  * @partial_cluster: The cluster which we'll have to free if all extents
2670  *                   has been released from it.  However, if this value is
2671  *                   negative, it's a cluster just to the right of the
2672  *                   punched region and it must not be freed.
2673  * @start:  The first block to remove
2674  * @end:   The last block to remove
2675  */
2676 static int
2677 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2678                  struct ext4_ext_path *path,
2679                  struct partial_cluster *partial,
2680                  ext4_lblk_t start, ext4_lblk_t end)
2681 {
2682         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2683         int err = 0, correct_index = 0;
2684         int depth = ext_depth(inode), credits;
2685         struct ext4_extent_header *eh;
2686         ext4_lblk_t a, b;
2687         unsigned num;
2688         ext4_lblk_t ex_ee_block;
2689         unsigned short ex_ee_len;
2690         unsigned unwritten = 0;
2691         struct ext4_extent *ex;
2692         ext4_fsblk_t pblk;
2693
2694         /* the header must be checked already in ext4_ext_remove_space() */
2695         ext_debug("truncate since %u in leaf to %u\n", start, end);
2696         if (!path[depth].p_hdr)
2697                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2698         eh = path[depth].p_hdr;
2699         if (unlikely(path[depth].p_hdr == NULL)) {
2700                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2701                 return -EFSCORRUPTED;
2702         }
2703         /* find where to start removing */
2704         ex = path[depth].p_ext;
2705         if (!ex)
2706                 ex = EXT_LAST_EXTENT(eh);
2707
2708         ex_ee_block = le32_to_cpu(ex->ee_block);
2709         ex_ee_len = ext4_ext_get_actual_len(ex);
2710
2711         trace_ext4_ext_rm_leaf(inode, start, ex, partial);
2712
2713         while (ex >= EXT_FIRST_EXTENT(eh) &&
2714                         ex_ee_block + ex_ee_len > start) {
2715
2716                 if (ext4_ext_is_unwritten(ex))
2717                         unwritten = 1;
2718                 else
2719                         unwritten = 0;
2720
2721                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2722                           unwritten, ex_ee_len);
2723                 path[depth].p_ext = ex;
2724
2725                 a = ex_ee_block > start ? ex_ee_block : start;
2726                 b = ex_ee_block+ex_ee_len - 1 < end ?
2727                         ex_ee_block+ex_ee_len - 1 : end;
2728
2729                 ext_debug("  border %u:%u\n", a, b);
2730
2731                 /* If this extent is beyond the end of the hole, skip it */
2732                 if (end < ex_ee_block) {
2733                         /*
2734                          * We're going to skip this extent and move to another,
2735                          * so note that its first cluster is in use to avoid
2736                          * freeing it when removing blocks.  Eventually, the
2737                          * right edge of the truncated/punched region will
2738                          * be just to the left.
2739                          */
2740                         if (sbi->s_cluster_ratio > 1) {
2741                                 pblk = ext4_ext_pblock(ex);
2742                                 partial->pclu = EXT4_B2C(sbi, pblk);
2743                                 partial->state = nofree;
2744                         }
2745                         ex--;
2746                         ex_ee_block = le32_to_cpu(ex->ee_block);
2747                         ex_ee_len = ext4_ext_get_actual_len(ex);
2748                         continue;
2749                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2750                         EXT4_ERROR_INODE(inode,
2751                                          "can not handle truncate %u:%u "
2752                                          "on extent %u:%u",
2753                                          start, end, ex_ee_block,
2754                                          ex_ee_block + ex_ee_len - 1);
2755                         err = -EFSCORRUPTED;
2756                         goto out;
2757                 } else if (a != ex_ee_block) {
2758                         /* remove tail of the extent */
2759                         num = a - ex_ee_block;
2760                 } else {
2761                         /* remove whole extent: excellent! */
2762                         num = 0;
2763                 }
2764                 /*
2765                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2766                  * descriptor) for each block group; assume two block
2767                  * groups plus ex_ee_len/blocks_per_block_group for
2768                  * the worst case
2769                  */
2770                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2771                 if (ex == EXT_FIRST_EXTENT(eh)) {
2772                         correct_index = 1;
2773                         credits += (ext_depth(inode)) + 1;
2774                 }
2775                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2776
2777                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2778                 if (err)
2779                         goto out;
2780
2781                 err = ext4_ext_get_access(handle, inode, path + depth);
2782                 if (err)
2783                         goto out;
2784
2785                 err = ext4_remove_blocks(handle, inode, ex, partial, a, b);
2786                 if (err)
2787                         goto out;
2788
2789                 if (num == 0)
2790                         /* this extent is removed; mark slot entirely unused */
2791                         ext4_ext_store_pblock(ex, 0);
2792
2793                 ex->ee_len = cpu_to_le16(num);
2794                 /*
2795                  * Do not mark unwritten if all the blocks in the
2796                  * extent have been removed.
2797                  */
2798                 if (unwritten && num)
2799                         ext4_ext_mark_unwritten(ex);
2800                 /*
2801                  * If the extent was completely released,
2802                  * we need to remove it from the leaf
2803                  */
2804                 if (num == 0) {
2805                         if (end != EXT_MAX_BLOCKS - 1) {
2806                                 /*
2807                                  * For hole punching, we need to scoot all the
2808                                  * extents up when an extent is removed so that
2809                                  * we dont have blank extents in the middle
2810                                  */
2811                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2812                                         sizeof(struct ext4_extent));
2813
2814                                 /* Now get rid of the one at the end */
2815                                 memset(EXT_LAST_EXTENT(eh), 0,
2816                                         sizeof(struct ext4_extent));
2817                         }
2818                         le16_add_cpu(&eh->eh_entries, -1);
2819                 }
2820
2821                 err = ext4_ext_dirty(handle, inode, path + depth);
2822                 if (err)
2823                         goto out;
2824
2825                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2826                                 ext4_ext_pblock(ex));
2827                 ex--;
2828                 ex_ee_block = le32_to_cpu(ex->ee_block);
2829                 ex_ee_len = ext4_ext_get_actual_len(ex);
2830         }
2831
2832         if (correct_index && eh->eh_entries)
2833                 err = ext4_ext_correct_indexes(handle, inode, path);
2834
2835         /*
2836          * If there's a partial cluster and at least one extent remains in
2837          * the leaf, free the partial cluster if it isn't shared with the
2838          * current extent.  If it is shared with the current extent
2839          * we reset the partial cluster because we've reached the start of the
2840          * truncated/punched region and we're done removing blocks.
2841          */
2842         if (partial->state == tofree && ex >= EXT_FIRST_EXTENT(eh)) {
2843                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2844                 if (partial->pclu != EXT4_B2C(sbi, pblk)) {
2845                         int flags = get_default_free_blocks_flags(inode);
2846
2847                         if (ext4_is_pending(inode, partial->lblk))
2848                                 flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2849                         ext4_free_blocks(handle, inode, NULL,
2850                                          EXT4_C2B(sbi, partial->pclu),
2851                                          sbi->s_cluster_ratio, flags);
2852                         if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2853                                 ext4_rereserve_cluster(inode, partial->lblk);
2854                 }
2855                 partial->state = initial;
2856         }
2857
2858         /* if this leaf is free, then we should
2859          * remove it from index block above */
2860         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2861                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2862
2863 out:
2864         return err;
2865 }
2866
2867 /*
2868  * ext4_ext_more_to_rm:
2869  * returns 1 if current index has to be freed (even partial)
2870  */
2871 static int
2872 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2873 {
2874         BUG_ON(path->p_idx == NULL);
2875
2876         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2877                 return 0;
2878
2879         /*
2880          * if truncate on deeper level happened, it wasn't partial,
2881          * so we have to consider current index for truncation
2882          */
2883         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2884                 return 0;
2885         return 1;
2886 }
2887
2888 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2889                           ext4_lblk_t end)
2890 {
2891         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2892         int depth = ext_depth(inode);
2893         struct ext4_ext_path *path = NULL;
2894         struct partial_cluster partial;
2895         handle_t *handle;
2896         int i = 0, err = 0;
2897
2898         partial.pclu = 0;
2899         partial.lblk = 0;
2900         partial.state = initial;
2901
2902         ext_debug("truncate since %u to %u\n", start, end);
2903
2904         /* probably first extent we're gonna free will be last in block */
2905         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2906         if (IS_ERR(handle))
2907                 return PTR_ERR(handle);
2908
2909 again:
2910         trace_ext4_ext_remove_space(inode, start, end, depth);
2911
2912         /*
2913          * Check if we are removing extents inside the extent tree. If that
2914          * is the case, we are going to punch a hole inside the extent tree
2915          * so we have to check whether we need to split the extent covering
2916          * the last block to remove so we can easily remove the part of it
2917          * in ext4_ext_rm_leaf().
2918          */
2919         if (end < EXT_MAX_BLOCKS - 1) {
2920                 struct ext4_extent *ex;
2921                 ext4_lblk_t ee_block, ex_end, lblk;
2922                 ext4_fsblk_t pblk;
2923
2924                 /* find extent for or closest extent to this block */
2925                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2926                 if (IS_ERR(path)) {
2927                         ext4_journal_stop(handle);
2928                         return PTR_ERR(path);
2929                 }
2930                 depth = ext_depth(inode);
2931                 /* Leaf not may not exist only if inode has no blocks at all */
2932                 ex = path[depth].p_ext;
2933                 if (!ex) {
2934                         if (depth) {
2935                                 EXT4_ERROR_INODE(inode,
2936                                                  "path[%d].p_hdr == NULL",
2937                                                  depth);
2938                                 err = -EFSCORRUPTED;
2939                         }
2940                         goto out;
2941                 }
2942
2943                 ee_block = le32_to_cpu(ex->ee_block);
2944                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2945
2946                 /*
2947                  * See if the last block is inside the extent, if so split
2948                  * the extent at 'end' block so we can easily remove the
2949                  * tail of the first part of the split extent in
2950                  * ext4_ext_rm_leaf().
2951                  */
2952                 if (end >= ee_block && end < ex_end) {
2953
2954                         /*
2955                          * If we're going to split the extent, note that
2956                          * the cluster containing the block after 'end' is
2957                          * in use to avoid freeing it when removing blocks.
2958                          */
2959                         if (sbi->s_cluster_ratio > 1) {
2960                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2961                                 partial.pclu = EXT4_B2C(sbi, pblk);
2962                                 partial.state = nofree;
2963                         }
2964
2965                         /*
2966                          * Split the extent in two so that 'end' is the last
2967                          * block in the first new extent. Also we should not
2968                          * fail removing space due to ENOSPC so try to use
2969                          * reserved block if that happens.
2970                          */
2971                         err = ext4_force_split_extent_at(handle, inode, &path,
2972                                                          end + 1, 1);
2973                         if (err < 0)
2974                                 goto out;
2975
2976                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end &&
2977                            partial.state == initial) {
2978                         /*
2979                          * If we're punching, there's an extent to the right.
2980                          * If the partial cluster hasn't been set, set it to
2981                          * that extent's first cluster and its state to nofree
2982                          * so it won't be freed should it contain blocks to be
2983                          * removed. If it's already set (tofree/nofree), we're
2984                          * retrying and keep the original partial cluster info
2985                          * so a cluster marked tofree as a result of earlier
2986                          * extent removal is not lost.
2987                          */
2988                         lblk = ex_end + 1;
2989                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2990                                                     &ex);
2991                         if (err)
2992                                 goto out;
2993                         if (pblk) {
2994                                 partial.pclu = EXT4_B2C(sbi, pblk);
2995                                 partial.state = nofree;
2996                         }
2997                 }
2998         }
2999         /*
3000          * We start scanning from right side, freeing all the blocks
3001          * after i_size and walking into the tree depth-wise.
3002          */
3003         depth = ext_depth(inode);
3004         if (path) {
3005                 int k = i = depth;
3006                 while (--k > 0)
3007                         path[k].p_block =
3008                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
3009         } else {
3010                 path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
3011                                GFP_NOFS);
3012                 if (path == NULL) {
3013                         ext4_journal_stop(handle);
3014                         return -ENOMEM;
3015                 }
3016                 path[0].p_maxdepth = path[0].p_depth = depth;
3017                 path[0].p_hdr = ext_inode_hdr(inode);
3018                 i = 0;
3019
3020                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
3021                         err = -EFSCORRUPTED;
3022                         goto out;
3023                 }
3024         }
3025         err = 0;
3026
3027         while (i >= 0 && err == 0) {
3028                 if (i == depth) {
3029                         /* this is leaf block */
3030                         err = ext4_ext_rm_leaf(handle, inode, path,
3031                                                &partial, start, end);
3032                         /* root level has p_bh == NULL, brelse() eats this */
3033                         brelse(path[i].p_bh);
3034                         path[i].p_bh = NULL;
3035                         i--;
3036                         continue;
3037                 }
3038
3039                 /* this is index block */
3040                 if (!path[i].p_hdr) {
3041                         ext_debug("initialize header\n");
3042                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
3043                 }
3044
3045                 if (!path[i].p_idx) {
3046                         /* this level hasn't been touched yet */
3047                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
3048                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
3049                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
3050                                   path[i].p_hdr,
3051                                   le16_to_cpu(path[i].p_hdr->eh_entries));
3052                 } else {
3053                         /* we were already here, see at next index */
3054                         path[i].p_idx--;
3055                 }
3056
3057                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
3058                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
3059                                 path[i].p_idx);
3060                 if (ext4_ext_more_to_rm(path + i)) {
3061                         struct buffer_head *bh;
3062                         /* go to the next level */
3063                         ext_debug("move to level %d (block %llu)\n",
3064                                   i + 1, ext4_idx_pblock(path[i].p_idx));
3065                         memset(path + i + 1, 0, sizeof(*path));
3066                         bh = read_extent_tree_block(inode,
3067                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
3068                                 EXT4_EX_NOCACHE);
3069                         if (IS_ERR(bh)) {
3070                                 /* should we reset i_size? */
3071                                 err = PTR_ERR(bh);
3072                                 break;
3073                         }
3074                         /* Yield here to deal with large extent trees.
3075                          * Should be a no-op if we did IO above. */
3076                         cond_resched();
3077                         if (WARN_ON(i + 1 > depth)) {
3078                                 err = -EFSCORRUPTED;
3079                                 break;
3080                         }
3081                         path[i + 1].p_bh = bh;
3082
3083                         /* save actual number of indexes since this
3084                          * number is changed at the next iteration */
3085                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3086                         i++;
3087                 } else {
3088                         /* we finished processing this index, go up */
3089                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3090                                 /* index is empty, remove it;
3091                                  * handle must be already prepared by the
3092                                  * truncatei_leaf() */
3093                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3094                         }
3095                         /* root level has p_bh == NULL, brelse() eats this */
3096                         brelse(path[i].p_bh);
3097                         path[i].p_bh = NULL;
3098                         i--;
3099                         ext_debug("return to level %d\n", i);
3100                 }
3101         }
3102
3103         trace_ext4_ext_remove_space_done(inode, start, end, depth, &partial,
3104                                          path->p_hdr->eh_entries);
3105
3106         /*
3107          * if there's a partial cluster and we have removed the first extent
3108          * in the file, then we also free the partial cluster, if any
3109          */
3110         if (partial.state == tofree && err == 0) {
3111                 int flags = get_default_free_blocks_flags(inode);
3112
3113                 if (ext4_is_pending(inode, partial.lblk))
3114                         flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
3115                 ext4_free_blocks(handle, inode, NULL,
3116                                  EXT4_C2B(sbi, partial.pclu),
3117                                  sbi->s_cluster_ratio, flags);
3118                 if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
3119                         ext4_rereserve_cluster(inode, partial.lblk);
3120                 partial.state = initial;
3121         }
3122
3123         /* TODO: flexible tree reduction should be here */
3124         if (path->p_hdr->eh_entries == 0) {
3125                 /*
3126                  * truncate to zero freed all the tree,
3127                  * so we need to correct eh_depth
3128                  */
3129                 err = ext4_ext_get_access(handle, inode, path);
3130                 if (err == 0) {
3131                         ext_inode_hdr(inode)->eh_depth = 0;
3132                         ext_inode_hdr(inode)->eh_max =
3133                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3134                         err = ext4_ext_dirty(handle, inode, path);
3135                 }
3136         }
3137 out:
3138         ext4_ext_drop_refs(path);
3139         kfree(path);
3140         path = NULL;
3141         if (err == -EAGAIN)
3142                 goto again;
3143         ext4_journal_stop(handle);
3144
3145         return err;
3146 }
3147
3148 /*
3149  * called at mount time
3150  */
3151 void ext4_ext_init(struct super_block *sb)
3152 {
3153         /*
3154          * possible initialization would be here
3155          */
3156
3157         if (ext4_has_feature_extents(sb)) {
3158 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3159                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3160 #ifdef AGGRESSIVE_TEST
3161                        ", aggressive tests"
3162 #endif
3163 #ifdef CHECK_BINSEARCH
3164                        ", check binsearch"
3165 #endif
3166 #ifdef EXTENTS_STATS
3167                        ", stats"
3168 #endif
3169                        "\n");
3170 #endif
3171 #ifdef EXTENTS_STATS
3172                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3173                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3174                 EXT4_SB(sb)->s_ext_max = 0;
3175 #endif
3176         }
3177 }
3178
3179 /*
3180  * called at umount time
3181  */
3182 void ext4_ext_release(struct super_block *sb)
3183 {
3184         if (!ext4_has_feature_extents(sb))
3185                 return;
3186
3187 #ifdef EXTENTS_STATS
3188         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3189                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3190                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3191                         sbi->s_ext_blocks, sbi->s_ext_extents,
3192                         sbi->s_ext_blocks / sbi->s_ext_extents);
3193                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3194                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3195         }
3196 #endif
3197 }
3198
3199 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3200 {
3201         ext4_lblk_t  ee_block;
3202         ext4_fsblk_t ee_pblock;
3203         unsigned int ee_len;
3204
3205         ee_block  = le32_to_cpu(ex->ee_block);
3206         ee_len    = ext4_ext_get_actual_len(ex);
3207         ee_pblock = ext4_ext_pblock(ex);
3208
3209         if (ee_len == 0)
3210                 return 0;
3211
3212         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3213                                      EXTENT_STATUS_WRITTEN);
3214 }
3215
3216 /* FIXME!! we need to try to merge to left or right after zero-out  */
3217 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3218 {
3219         ext4_fsblk_t ee_pblock;
3220         unsigned int ee_len;
3221
3222         ee_len    = ext4_ext_get_actual_len(ex);
3223         ee_pblock = ext4_ext_pblock(ex);
3224         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3225                                   ee_len);
3226 }
3227
3228 /*
3229  * ext4_split_extent_at() splits an extent at given block.
3230  *
3231  * @handle: the journal handle
3232  * @inode: the file inode
3233  * @path: the path to the extent
3234  * @split: the logical block where the extent is splitted.
3235  * @split_flags: indicates if the extent could be zeroout if split fails, and
3236  *               the states(init or unwritten) of new extents.
3237  * @flags: flags used to insert new extent to extent tree.
3238  *
3239  *
3240  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3241  * of which are deterimined by split_flag.
3242  *
3243  * There are two cases:
3244  *  a> the extent are splitted into two extent.
3245  *  b> split is not needed, and just mark the extent.
3246  *
3247  * return 0 on success.
3248  */
3249 static int ext4_split_extent_at(handle_t *handle,
3250                              struct inode *inode,
3251                              struct ext4_ext_path **ppath,
3252                              ext4_lblk_t split,
3253                              int split_flag,
3254                              int flags)
3255 {
3256         struct ext4_ext_path *path = *ppath;
3257         ext4_fsblk_t newblock;
3258         ext4_lblk_t ee_block;
3259         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3260         struct ext4_extent *ex2 = NULL;
3261         unsigned int ee_len, depth;
3262         int err = 0;
3263
3264         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3265                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3266
3267         ext_debug("ext4_split_extents_at: inode %lu, logical"
3268                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3269
3270         ext4_ext_show_leaf(inode, path);
3271
3272         depth = ext_depth(inode);
3273         ex = path[depth].p_ext;
3274         ee_block = le32_to_cpu(ex->ee_block);
3275         ee_len = ext4_ext_get_actual_len(ex);
3276         newblock = split - ee_block + ext4_ext_pblock(ex);
3277
3278         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3279         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3280                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3281                              EXT4_EXT_MARK_UNWRIT1 |
3282                              EXT4_EXT_MARK_UNWRIT2));
3283
3284         err = ext4_ext_get_access(handle, inode, path + depth);
3285         if (err)
3286                 goto out;
3287
3288         if (split == ee_block) {
3289                 /*
3290                  * case b: block @split is the block that the extent begins with
3291                  * then we just change the state of the extent, and splitting
3292                  * is not needed.
3293                  */
3294                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3295                         ext4_ext_mark_unwritten(ex);
3296                 else
3297                         ext4_ext_mark_initialized(ex);
3298
3299                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3300                         ext4_ext_try_to_merge(handle, inode, path, ex);
3301
3302                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3303                 goto out;
3304         }
3305
3306         /* case a */
3307         memcpy(&orig_ex, ex, sizeof(orig_ex));
3308         ex->ee_len = cpu_to_le16(split - ee_block);
3309         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3310                 ext4_ext_mark_unwritten(ex);
3311
3312         /*
3313          * path may lead to new leaf, not to original leaf any more
3314          * after ext4_ext_insert_extent() returns,
3315          */
3316         err = ext4_ext_dirty(handle, inode, path + depth);
3317         if (err)
3318                 goto fix_extent_len;
3319
3320         ex2 = &newex;
3321         ex2->ee_block = cpu_to_le32(split);
3322         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3323         ext4_ext_store_pblock(ex2, newblock);
3324         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3325                 ext4_ext_mark_unwritten(ex2);
3326
3327         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3328         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3329                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3330                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3331                                 err = ext4_ext_zeroout(inode, ex2);
3332                                 zero_ex.ee_block = ex2->ee_block;
3333                                 zero_ex.ee_len = cpu_to_le16(
3334                                                 ext4_ext_get_actual_len(ex2));
3335                                 ext4_ext_store_pblock(&zero_ex,
3336                                                       ext4_ext_pblock(ex2));
3337                         } else {
3338                                 err = ext4_ext_zeroout(inode, ex);
3339                                 zero_ex.ee_block = ex->ee_block;
3340                                 zero_ex.ee_len = cpu_to_le16(
3341                                                 ext4_ext_get_actual_len(ex));
3342                                 ext4_ext_store_pblock(&zero_ex,
3343                                                       ext4_ext_pblock(ex));
3344                         }
3345                 } else {
3346                         err = ext4_ext_zeroout(inode, &orig_ex);
3347                         zero_ex.ee_block = orig_ex.ee_block;
3348                         zero_ex.ee_len = cpu_to_le16(
3349                                                 ext4_ext_get_actual_len(&orig_ex));
3350                         ext4_ext_store_pblock(&zero_ex,
3351                                               ext4_ext_pblock(&orig_ex));
3352                 }
3353
3354                 if (err)
3355                         goto fix_extent_len;
3356                 /* update the extent length and mark as initialized */
3357                 ex->ee_len = cpu_to_le16(ee_len);
3358                 ext4_ext_try_to_merge(handle, inode, path, ex);
3359                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3360                 if (err)
3361                         goto fix_extent_len;
3362
3363                 /* update extent status tree */
3364                 err = ext4_zeroout_es(inode, &zero_ex);
3365
3366                 goto out;
3367         } else if (err)
3368                 goto fix_extent_len;
3369
3370 out:
3371         ext4_ext_show_leaf(inode, path);
3372         return err;
3373
3374 fix_extent_len:
3375         ex->ee_len = orig_ex.ee_len;
3376         ext4_ext_dirty(handle, inode, path + path->p_depth);
3377         return err;
3378 }
3379
3380 /*
3381  * ext4_split_extents() splits an extent and mark extent which is covered
3382  * by @map as split_flags indicates
3383  *
3384  * It may result in splitting the extent into multiple extents (up to three)
3385  * There are three possibilities:
3386  *   a> There is no split required
3387  *   b> Splits in two extents: Split is happening at either end of the extent
3388  *   c> Splits in three extents: Somone is splitting in middle of the extent
3389  *
3390  */
3391 static int ext4_split_extent(handle_t *handle,
3392                               struct inode *inode,
3393                               struct ext4_ext_path **ppath,
3394                               struct ext4_map_blocks *map,
3395                               int split_flag,
3396                               int flags)
3397 {
3398         struct ext4_ext_path *path = *ppath;
3399         ext4_lblk_t ee_block;
3400         struct ext4_extent *ex;
3401         unsigned int ee_len, depth;
3402         int err = 0;
3403         int unwritten;
3404         int split_flag1, flags1;
3405         int allocated = map->m_len;
3406
3407         depth = ext_depth(inode);
3408         ex = path[depth].p_ext;
3409         ee_block = le32_to_cpu(ex->ee_block);
3410         ee_len = ext4_ext_get_actual_len(ex);
3411         unwritten = ext4_ext_is_unwritten(ex);
3412
3413         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3414                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3415                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3416                 if (unwritten)
3417                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3418                                        EXT4_EXT_MARK_UNWRIT2;
3419                 if (split_flag & EXT4_EXT_DATA_VALID2)
3420                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3421                 err = ext4_split_extent_at(handle, inode, ppath,
3422                                 map->m_lblk + map->m_len, split_flag1, flags1);
3423                 if (err)
3424                         goto out;
3425         } else {
3426                 allocated = ee_len - (map->m_lblk - ee_block);
3427         }
3428         /*
3429          * Update path is required because previous ext4_split_extent_at() may
3430          * result in split of original leaf or extent zeroout.
3431          */
3432         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3433         if (IS_ERR(path))
3434                 return PTR_ERR(path);
3435         depth = ext_depth(inode);
3436         ex = path[depth].p_ext;
3437         if (!ex) {
3438                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3439                                  (unsigned long) map->m_lblk);
3440                 return -EFSCORRUPTED;
3441         }
3442         unwritten = ext4_ext_is_unwritten(ex);
3443         split_flag1 = 0;
3444
3445         if (map->m_lblk >= ee_block) {
3446                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3447                 if (unwritten) {
3448                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3449                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3450                                                      EXT4_EXT_MARK_UNWRIT2);
3451                 }
3452                 err = ext4_split_extent_at(handle, inode, ppath,
3453                                 map->m_lblk, split_flag1, flags);
3454                 if (err)
3455                         goto out;
3456         }
3457
3458         ext4_ext_show_leaf(inode, path);
3459 out:
3460         return err ? err : allocated;
3461 }
3462
3463 /*
3464  * This function is called by ext4_ext_map_blocks() if someone tries to write
3465  * to an unwritten extent. It may result in splitting the unwritten
3466  * extent into multiple extents (up to three - one initialized and two
3467  * unwritten).
3468  * There are three possibilities:
3469  *   a> There is no split required: Entire extent should be initialized
3470  *   b> Splits in two extents: Write is happening at either end of the extent
3471  *   c> Splits in three extents: Somone is writing in middle of the extent
3472  *
3473  * Pre-conditions:
3474  *  - The extent pointed to by 'path' is unwritten.
3475  *  - The extent pointed to by 'path' contains a superset
3476  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3477  *
3478  * Post-conditions on success:
3479  *  - the returned value is the number of blocks beyond map->l_lblk
3480  *    that are allocated and initialized.
3481  *    It is guaranteed to be >= map->m_len.
3482  */
3483 static int ext4_ext_convert_to_initialized(handle_t *handle,
3484                                            struct inode *inode,
3485                                            struct ext4_map_blocks *map,
3486                                            struct ext4_ext_path **ppath,
3487                                            int flags)
3488 {
3489         struct ext4_ext_path *path = *ppath;
3490         struct ext4_sb_info *sbi;
3491         struct ext4_extent_header *eh;
3492         struct ext4_map_blocks split_map;
3493         struct ext4_extent zero_ex1, zero_ex2;
3494         struct ext4_extent *ex, *abut_ex;
3495         ext4_lblk_t ee_block, eof_block;
3496         unsigned int ee_len, depth, map_len = map->m_len;
3497         int allocated = 0, max_zeroout = 0;
3498         int err = 0;
3499         int split_flag = EXT4_EXT_DATA_VALID2;
3500
3501         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3502                 "block %llu, max_blocks %u\n", inode->i_ino,
3503                 (unsigned long long)map->m_lblk, map_len);
3504
3505         sbi = EXT4_SB(inode->i_sb);
3506         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3507                 inode->i_sb->s_blocksize_bits;
3508         if (eof_block < map->m_lblk + map_len)
3509                 eof_block = map->m_lblk + map_len;
3510
3511         depth = ext_depth(inode);
3512         eh = path[depth].p_hdr;
3513         ex = path[depth].p_ext;
3514         ee_block = le32_to_cpu(ex->ee_block);
3515         ee_len = ext4_ext_get_actual_len(ex);
3516         zero_ex1.ee_len = 0;
3517         zero_ex2.ee_len = 0;
3518
3519         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3520
3521         /* Pre-conditions */
3522         BUG_ON(!ext4_ext_is_unwritten(ex));
3523         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3524
3525         /*
3526          * Attempt to transfer newly initialized blocks from the currently
3527          * unwritten extent to its neighbor. This is much cheaper
3528          * than an insertion followed by a merge as those involve costly
3529          * memmove() calls. Transferring to the left is the common case in
3530          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3531          * followed by append writes.
3532          *
3533          * Limitations of the current logic:
3534          *  - L1: we do not deal with writes covering the whole extent.
3535          *    This would require removing the extent if the transfer
3536          *    is possible.
3537          *  - L2: we only attempt to merge with an extent stored in the
3538          *    same extent tree node.
3539          */
3540         if ((map->m_lblk == ee_block) &&
3541                 /* See if we can merge left */
3542                 (map_len < ee_len) &&           /*L1*/
3543                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3544                 ext4_lblk_t prev_lblk;
3545                 ext4_fsblk_t prev_pblk, ee_pblk;
3546                 unsigned int prev_len;
3547
3548                 abut_ex = ex - 1;
3549                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3550                 prev_len = ext4_ext_get_actual_len(abut_ex);
3551                 prev_pblk = ext4_ext_pblock(abut_ex);
3552                 ee_pblk = ext4_ext_pblock(ex);
3553
3554                 /*
3555                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3556                  * upon those conditions:
3557                  * - C1: abut_ex is initialized,
3558                  * - C2: abut_ex is logically abutting ex,
3559                  * - C3: abut_ex is physically abutting ex,
3560                  * - C4: abut_ex can receive the additional blocks without
3561                  *   overflowing the (initialized) length limit.
3562                  */
3563                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3564                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3565                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3566                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3567                         err = ext4_ext_get_access(handle, inode, path + depth);
3568                         if (err)
3569                                 goto out;
3570
3571                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3572                                 map, ex, abut_ex);
3573
3574                         /* Shift the start of ex by 'map_len' blocks */
3575                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3576                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3577                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3578                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3579
3580                         /* Extend abut_ex by 'map_len' blocks */
3581                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3582
3583                         /* Result: number of initialized blocks past m_lblk */
3584                         allocated = map_len;
3585                 }
3586         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3587                    (map_len < ee_len) &&        /*L1*/
3588                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3589                 /* See if we can merge right */
3590                 ext4_lblk_t next_lblk;
3591                 ext4_fsblk_t next_pblk, ee_pblk;
3592                 unsigned int next_len;
3593
3594                 abut_ex = ex + 1;
3595                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3596                 next_len = ext4_ext_get_actual_len(abut_ex);
3597                 next_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                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3611                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3612                     (next_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 abut_ex by 'map_len' blocks */
3621                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3622                         ext4_ext_store_pblock(abut_ex, next_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(next_len + map_len);
3628
3629                         /* Result: number of initialized blocks past m_lblk */
3630                         allocated = map_len;
3631                 }
3632         }
3633         if (allocated) {
3634                 /* Mark the block containing both extents as dirty */
3635                 ext4_ext_dirty(handle, inode, path + depth);
3636
3637                 /* Update path to point to the right extent */
3638                 path[depth].p_ext = abut_ex;
3639                 goto out;
3640         } else
3641                 allocated = ee_len - (map->m_lblk - ee_block);
3642
3643         WARN_ON(map->m_lblk < ee_block);
3644         /*
3645          * It is safe to convert extent to initialized via explicit
3646          * zeroout only if extent is fully inside i_size or new_size.
3647          */
3648         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3649
3650         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3651                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3652                         (inode->i_sb->s_blocksize_bits - 10);
3653
3654         if (IS_ENCRYPTED(inode))
3655                 max_zeroout = 0;
3656
3657         /*
3658          * five cases:
3659          * 1. split the extent into three extents.
3660          * 2. split the extent into two extents, zeroout the head of the first
3661          *    extent.
3662          * 3. split the extent into two extents, zeroout the tail of the second
3663          *    extent.
3664          * 4. split the extent into two extents with out zeroout.
3665          * 5. no splitting needed, just possibly zeroout the head and / or the
3666          *    tail of the extent.
3667          */
3668         split_map.m_lblk = map->m_lblk;
3669         split_map.m_len = map->m_len;
3670
3671         if (max_zeroout && (allocated > split_map.m_len)) {
3672                 if (allocated <= max_zeroout) {
3673                         /* case 3 or 5 */
3674                         zero_ex1.ee_block =
3675                                  cpu_to_le32(split_map.m_lblk +
3676                                              split_map.m_len);
3677                         zero_ex1.ee_len =
3678                                 cpu_to_le16(allocated - split_map.m_len);
3679                         ext4_ext_store_pblock(&zero_ex1,
3680                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3681                                 split_map.m_len - ee_block);
3682                         err = ext4_ext_zeroout(inode, &zero_ex1);
3683                         if (err)
3684                                 goto out;
3685                         split_map.m_len = allocated;
3686                 }
3687                 if (split_map.m_lblk - ee_block + split_map.m_len <
3688                                                                 max_zeroout) {
3689                         /* case 2 or 5 */
3690                         if (split_map.m_lblk != ee_block) {
3691                                 zero_ex2.ee_block = ex->ee_block;
3692                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3693                                                         ee_block);
3694                                 ext4_ext_store_pblock(&zero_ex2,
3695                                                       ext4_ext_pblock(ex));
3696                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3697                                 if (err)
3698                                         goto out;
3699                         }
3700
3701                         split_map.m_len += split_map.m_lblk - ee_block;
3702                         split_map.m_lblk = ee_block;
3703                         allocated = map->m_len;
3704                 }
3705         }
3706
3707         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3708                                 flags);
3709         if (err > 0)
3710                 err = 0;
3711 out:
3712         /* If we have gotten a failure, don't zero out status tree */
3713         if (!err) {
3714                 err = ext4_zeroout_es(inode, &zero_ex1);
3715                 if (!err)
3716                         err = ext4_zeroout_es(inode, &zero_ex2);
3717         }
3718         return err ? err : allocated;
3719 }
3720
3721 /*
3722  * This function is called by ext4_ext_map_blocks() from
3723  * ext4_get_blocks_dio_write() when DIO to write
3724  * to an unwritten extent.
3725  *
3726  * Writing to an unwritten extent may result in splitting the unwritten
3727  * extent into multiple initialized/unwritten extents (up to three)
3728  * There are three possibilities:
3729  *   a> There is no split required: Entire extent should be unwritten
3730  *   b> Splits in two extents: Write is happening at either end of the extent
3731  *   c> Splits in three extents: Somone is writing in middle of the extent
3732  *
3733  * This works the same way in the case of initialized -> unwritten conversion.
3734  *
3735  * One of more index blocks maybe needed if the extent tree grow after
3736  * the unwritten extent split. To prevent ENOSPC occur at the IO
3737  * complete, we need to split the unwritten extent before DIO submit
3738  * the IO. The unwritten extent called at this time will be split
3739  * into three unwritten extent(at most). After IO complete, the part
3740  * being filled will be convert to initialized by the end_io callback function
3741  * via ext4_convert_unwritten_extents().
3742  *
3743  * Returns the size of unwritten extent to be written on success.
3744  */
3745 static int ext4_split_convert_extents(handle_t *handle,
3746                                         struct inode *inode,
3747                                         struct ext4_map_blocks *map,
3748                                         struct ext4_ext_path **ppath,
3749                                         int flags)
3750 {
3751         struct ext4_ext_path *path = *ppath;
3752         ext4_lblk_t eof_block;
3753         ext4_lblk_t ee_block;
3754         struct ext4_extent *ex;
3755         unsigned int ee_len;
3756         int split_flag = 0, depth;
3757
3758         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3759                   __func__, inode->i_ino,
3760                   (unsigned long long)map->m_lblk, map->m_len);
3761
3762         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3763                 inode->i_sb->s_blocksize_bits;
3764         if (eof_block < map->m_lblk + map->m_len)
3765                 eof_block = map->m_lblk + map->m_len;
3766         /*
3767          * It is safe to convert extent to initialized via explicit
3768          * zeroout only if extent is fully insde i_size or new_size.
3769          */
3770         depth = ext_depth(inode);
3771         ex = path[depth].p_ext;
3772         ee_block = le32_to_cpu(ex->ee_block);
3773         ee_len = ext4_ext_get_actual_len(ex);
3774
3775         /* Convert to unwritten */
3776         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3777                 split_flag |= EXT4_EXT_DATA_VALID1;
3778         /* Convert to initialized */
3779         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3780                 split_flag |= ee_block + ee_len <= eof_block ?
3781                               EXT4_EXT_MAY_ZEROOUT : 0;
3782                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3783         }
3784         flags |= EXT4_GET_BLOCKS_PRE_IO;
3785         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3786 }
3787
3788 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3789                                                 struct inode *inode,
3790                                                 struct ext4_map_blocks *map,
3791                                                 struct ext4_ext_path **ppath)
3792 {
3793         struct ext4_ext_path *path = *ppath;
3794         struct ext4_extent *ex;
3795         ext4_lblk_t ee_block;
3796         unsigned int ee_len;
3797         int depth;
3798         int err = 0;
3799
3800         depth = ext_depth(inode);
3801         ex = path[depth].p_ext;
3802         ee_block = le32_to_cpu(ex->ee_block);
3803         ee_len = ext4_ext_get_actual_len(ex);
3804
3805         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3806                 "block %llu, max_blocks %u\n", inode->i_ino,
3807                   (unsigned long long)ee_block, ee_len);
3808
3809         /* If extent is larger than requested it is a clear sign that we still
3810          * have some extent state machine issues left. So extent_split is still
3811          * required.
3812          * TODO: Once all related issues will be fixed this situation should be
3813          * illegal.
3814          */
3815         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3816 #ifdef EXT4_DEBUG
3817                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3818                              " len %u; IO logical block %llu, len %u",
3819                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3820                              (unsigned long long)map->m_lblk, map->m_len);
3821 #endif
3822                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3823                                                  EXT4_GET_BLOCKS_CONVERT);
3824                 if (err < 0)
3825                         return err;
3826                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3827                 if (IS_ERR(path))
3828                         return PTR_ERR(path);
3829                 depth = ext_depth(inode);
3830                 ex = path[depth].p_ext;
3831         }
3832
3833         err = ext4_ext_get_access(handle, inode, path + depth);
3834         if (err)
3835                 goto out;
3836         /* first mark the extent as initialized */
3837         ext4_ext_mark_initialized(ex);
3838
3839         /* note: ext4_ext_correct_indexes() isn't needed here because
3840          * borders are not changed
3841          */
3842         ext4_ext_try_to_merge(handle, inode, path, ex);
3843
3844         /* Mark modified extent as dirty */
3845         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3846 out:
3847         ext4_ext_show_leaf(inode, path);
3848         return err;
3849 }
3850
3851 /*
3852  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3853  */
3854 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3855                               ext4_lblk_t lblk,
3856                               struct ext4_ext_path *path,
3857                               unsigned int len)
3858 {
3859         int i, depth;
3860         struct ext4_extent_header *eh;
3861         struct ext4_extent *last_ex;
3862
3863         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3864                 return 0;
3865
3866         depth = ext_depth(inode);
3867         eh = path[depth].p_hdr;
3868
3869         /*
3870          * We're going to remove EOFBLOCKS_FL entirely in future so we
3871          * do not care for this case anymore. Simply remove the flag
3872          * if there are no extents.
3873          */
3874         if (unlikely(!eh->eh_entries))
3875                 goto out;
3876         last_ex = EXT_LAST_EXTENT(eh);
3877         /*
3878          * We should clear the EOFBLOCKS_FL flag if we are writing the
3879          * last block in the last extent in the file.  We test this by
3880          * first checking to see if the caller to
3881          * ext4_ext_get_blocks() was interested in the last block (or
3882          * a block beyond the last block) in the current extent.  If
3883          * this turns out to be false, we can bail out from this
3884          * function immediately.
3885          */
3886         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3887             ext4_ext_get_actual_len(last_ex))
3888                 return 0;
3889         /*
3890          * If the caller does appear to be planning to write at or
3891          * beyond the end of the current extent, we then test to see
3892          * if the current extent is the last extent in the file, by
3893          * checking to make sure it was reached via the rightmost node
3894          * at each level of the tree.
3895          */
3896         for (i = depth-1; i >= 0; i--)
3897                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3898                         return 0;
3899 out:
3900         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3901         return ext4_mark_inode_dirty(handle, inode);
3902 }
3903
3904 static int
3905 convert_initialized_extent(handle_t *handle, struct inode *inode,
3906                            struct ext4_map_blocks *map,
3907                            struct ext4_ext_path **ppath,
3908                            unsigned int allocated)
3909 {
3910         struct ext4_ext_path *path = *ppath;
3911         struct ext4_extent *ex;
3912         ext4_lblk_t ee_block;
3913         unsigned int ee_len;
3914         int depth;
3915         int err = 0;
3916
3917         /*
3918          * Make sure that the extent is no bigger than we support with
3919          * unwritten extent
3920          */
3921         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3922                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3923
3924         depth = ext_depth(inode);
3925         ex = path[depth].p_ext;
3926         ee_block = le32_to_cpu(ex->ee_block);
3927         ee_len = ext4_ext_get_actual_len(ex);
3928
3929         ext_debug("%s: inode %lu, logical"
3930                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3931                   (unsigned long long)ee_block, ee_len);
3932
3933         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3934                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3935                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3936                 if (err < 0)
3937                         return err;
3938                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3939                 if (IS_ERR(path))
3940                         return PTR_ERR(path);
3941                 depth = ext_depth(inode);
3942                 ex = path[depth].p_ext;
3943                 if (!ex) {
3944                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3945                                          (unsigned long) map->m_lblk);
3946                         return -EFSCORRUPTED;
3947                 }
3948         }
3949
3950         err = ext4_ext_get_access(handle, inode, path + depth);
3951         if (err)
3952                 return err;
3953         /* first mark the extent as unwritten */
3954         ext4_ext_mark_unwritten(ex);
3955
3956         /* note: ext4_ext_correct_indexes() isn't needed here because
3957          * borders are not changed
3958          */
3959         ext4_ext_try_to_merge(handle, inode, path, ex);
3960
3961         /* Mark modified extent as dirty */
3962         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3963         if (err)
3964                 return err;
3965         ext4_ext_show_leaf(inode, path);
3966
3967         ext4_update_inode_fsync_trans(handle, inode, 1);
3968         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3969         if (err)
3970                 return err;
3971         map->m_flags |= EXT4_MAP_UNWRITTEN;
3972         if (allocated > map->m_len)
3973                 allocated = map->m_len;
3974         map->m_len = allocated;
3975         return allocated;
3976 }
3977
3978 static int
3979 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
3980                         struct ext4_map_blocks *map,
3981                         struct ext4_ext_path **ppath, int flags,
3982                         unsigned int allocated, ext4_fsblk_t newblock)
3983 {
3984         struct ext4_ext_path *path = *ppath;
3985         int ret = 0;
3986         int err = 0;
3987
3988         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
3989                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3990                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3991                   flags, allocated);
3992         ext4_ext_show_leaf(inode, path);
3993
3994         /*
3995          * When writing into unwritten space, we should not fail to
3996          * allocate metadata blocks for the new extent block if needed.
3997          */
3998         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
3999
4000         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4001                                                     allocated, newblock);
4002
4003         /* get_block() before submit the IO, split the extent */
4004         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4005                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4006                                          flags | EXT4_GET_BLOCKS_CONVERT);
4007                 if (ret <= 0)
4008                         goto out;
4009                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4010                 goto out;
4011         }
4012         /* IO end_io complete, convert the filled extent to written */
4013         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4014                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4015                         if (allocated > map->m_len)
4016                                 allocated = map->m_len;
4017                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4018                                                  allocated);
4019                         if (err < 0)
4020                                 goto out2;
4021                 }
4022                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4023                                                            ppath);
4024                 if (ret >= 0) {
4025                         ext4_update_inode_fsync_trans(handle, inode, 1);
4026                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4027                                                  path, map->m_len);
4028                 } else
4029                         err = ret;
4030                 map->m_flags |= EXT4_MAP_MAPPED;
4031                 map->m_pblk = newblock;
4032                 if (allocated > map->m_len)
4033                         allocated = map->m_len;
4034                 map->m_len = allocated;
4035                 goto out2;
4036         }
4037         /* buffered IO case */
4038         /*
4039          * repeat fallocate creation request
4040          * we already have an unwritten extent
4041          */
4042         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4043                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4044                 goto map_out;
4045         }
4046
4047         /* buffered READ or buffered write_begin() lookup */
4048         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4049                 /*
4050                  * We have blocks reserved already.  We
4051                  * return allocated blocks so that delalloc
4052                  * won't do block reservation for us.  But
4053                  * the buffer head will be unmapped so that
4054                  * a read from the block returns 0s.
4055                  */
4056                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4057                 goto out1;
4058         }
4059
4060         /* buffered write, writepage time, convert*/
4061         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4062         if (ret >= 0)
4063                 ext4_update_inode_fsync_trans(handle, inode, 1);
4064 out:
4065         if (ret <= 0) {
4066                 err = ret;
4067                 goto out2;
4068         } else
4069                 allocated = ret;
4070         map->m_flags |= EXT4_MAP_NEW;
4071         if (allocated > map->m_len)
4072                 allocated = map->m_len;
4073         map->m_len = allocated;
4074
4075 map_out:
4076         map->m_flags |= EXT4_MAP_MAPPED;
4077         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4078                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4079                                          map->m_len);
4080                 if (err < 0)
4081                         goto out2;
4082         }
4083 out1:
4084         if (allocated > map->m_len)
4085                 allocated = map->m_len;
4086         ext4_ext_show_leaf(inode, path);
4087         map->m_pblk = newblock;
4088         map->m_len = allocated;
4089 out2:
4090         return err ? err : allocated;
4091 }
4092
4093 /*
4094  * get_implied_cluster_alloc - check to see if the requested
4095  * allocation (in the map structure) overlaps with a cluster already
4096  * allocated in an extent.
4097  *      @sb     The filesystem superblock structure
4098  *      @map    The requested lblk->pblk mapping
4099  *      @ex     The extent structure which might contain an implied
4100  *                      cluster allocation
4101  *
4102  * This function is called by ext4_ext_map_blocks() after we failed to
4103  * find blocks that were already in the inode's extent tree.  Hence,
4104  * we know that the beginning of the requested region cannot overlap
4105  * the extent from the inode's extent tree.  There are three cases we
4106  * want to catch.  The first is this case:
4107  *
4108  *               |--- cluster # N--|
4109  *    |--- extent ---|  |---- requested region ---|
4110  *                      |==========|
4111  *
4112  * The second case that we need to test for is this one:
4113  *
4114  *   |--------- cluster # N ----------------|
4115  *         |--- requested region --|   |------- extent ----|
4116  *         |=======================|
4117  *
4118  * The third case is when the requested region lies between two extents
4119  * within the same cluster:
4120  *          |------------- cluster # N-------------|
4121  * |----- ex -----|                  |---- ex_right ----|
4122  *                  |------ requested region ------|
4123  *                  |================|
4124  *
4125  * In each of the above cases, we need to set the map->m_pblk and
4126  * map->m_len so it corresponds to the return the extent labelled as
4127  * "|====|" from cluster #N, since it is already in use for data in
4128  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4129  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4130  * as a new "allocated" block region.  Otherwise, we will return 0 and
4131  * ext4_ext_map_blocks() will then allocate one or more new clusters
4132  * by calling ext4_mb_new_blocks().
4133  */
4134 static int get_implied_cluster_alloc(struct super_block *sb,
4135                                      struct ext4_map_blocks *map,
4136                                      struct ext4_extent *ex,
4137                                      struct ext4_ext_path *path)
4138 {
4139         struct ext4_sb_info *sbi = EXT4_SB(sb);
4140         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4141         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4142         ext4_lblk_t rr_cluster_start;
4143         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4144         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4145         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4146
4147         /* The extent passed in that we are trying to match */
4148         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4149         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4150
4151         /* The requested region passed into ext4_map_blocks() */
4152         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4153
4154         if ((rr_cluster_start == ex_cluster_end) ||
4155             (rr_cluster_start == ex_cluster_start)) {
4156                 if (rr_cluster_start == ex_cluster_end)
4157                         ee_start += ee_len - 1;
4158                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4159                 map->m_len = min(map->m_len,
4160                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4161                 /*
4162                  * Check for and handle this case:
4163                  *
4164                  *   |--------- cluster # N-------------|
4165                  *                     |------- extent ----|
4166                  *         |--- requested region ---|
4167                  *         |===========|
4168                  */
4169
4170                 if (map->m_lblk < ee_block)
4171                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4172
4173                 /*
4174                  * Check for the case where there is already another allocated
4175                  * block to the right of 'ex' but before the end of the cluster.
4176                  *
4177                  *          |------------- cluster # N-------------|
4178                  * |----- ex -----|                  |---- ex_right ----|
4179                  *                  |------ requested region ------|
4180                  *                  |================|
4181                  */
4182                 if (map->m_lblk > ee_block) {
4183                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4184                         map->m_len = min(map->m_len, next - map->m_lblk);
4185                 }
4186
4187                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4188                 return 1;
4189         }
4190
4191         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4192         return 0;
4193 }
4194
4195
4196 /*
4197  * Block allocation/map/preallocation routine for extents based files
4198  *
4199  *
4200  * Need to be called with
4201  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4202  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4203  *
4204  * return > 0, number of of blocks already mapped/allocated
4205  *          if create == 0 and these are pre-allocated blocks
4206  *              buffer head is unmapped
4207  *          otherwise blocks are mapped
4208  *
4209  * return = 0, if plain look up failed (blocks have not been allocated)
4210  *          buffer head is unmapped
4211  *
4212  * return < 0, error case.
4213  */
4214 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4215                         struct ext4_map_blocks *map, int flags)
4216 {
4217         struct ext4_ext_path *path = NULL;
4218         struct ext4_extent newex, *ex, *ex2;
4219         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4220         ext4_fsblk_t newblock = 0;
4221         int free_on_err = 0, err = 0, depth, ret;
4222         unsigned int allocated = 0, offset = 0;
4223         unsigned int allocated_clusters = 0;
4224         struct ext4_allocation_request ar;
4225         ext4_lblk_t cluster_offset;
4226         bool map_from_cluster = false;
4227
4228         ext_debug("blocks %u/%u requested for inode %lu\n",
4229                   map->m_lblk, map->m_len, inode->i_ino);
4230         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4231
4232         /* find extent for this block */
4233         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4234         if (IS_ERR(path)) {
4235                 err = PTR_ERR(path);
4236                 path = NULL;
4237                 goto out2;
4238         }
4239
4240         depth = ext_depth(inode);
4241
4242         /*
4243          * consistent leaf must not be empty;
4244          * this situation is possible, though, _during_ tree modification;
4245          * this is why assert can't be put in ext4_find_extent()
4246          */
4247         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4248                 EXT4_ERROR_INODE(inode, "bad extent address "
4249                                  "lblock: %lu, depth: %d pblock %lld",
4250                                  (unsigned long) map->m_lblk, depth,
4251                                  path[depth].p_block);
4252                 err = -EFSCORRUPTED;
4253                 goto out2;
4254         }
4255
4256         ex = path[depth].p_ext;
4257         if (ex) {
4258                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4259                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4260                 unsigned short ee_len;
4261
4262
4263                 /*
4264                  * unwritten extents are treated as holes, except that
4265                  * we split out initialized portions during a write.
4266                  */
4267                 ee_len = ext4_ext_get_actual_len(ex);
4268
4269                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4270
4271                 /* if found extent covers block, simply return it */
4272                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4273                         newblock = map->m_lblk - ee_block + ee_start;
4274                         /* number of remaining blocks in the extent */
4275                         allocated = ee_len - (map->m_lblk - ee_block);
4276                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4277                                   ee_block, ee_len, newblock);
4278
4279                         /*
4280                          * If the extent is initialized check whether the
4281                          * caller wants to convert it to unwritten.
4282                          */
4283                         if ((!ext4_ext_is_unwritten(ex)) &&
4284                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4285                                 allocated = convert_initialized_extent(
4286                                                 handle, inode, map, &path,
4287                                                 allocated);
4288                                 goto out2;
4289                         } else if (!ext4_ext_is_unwritten(ex))
4290                                 goto out;
4291
4292                         ret = ext4_ext_handle_unwritten_extents(
4293                                 handle, inode, map, &path, flags,
4294                                 allocated, newblock);
4295                         if (ret < 0)
4296                                 err = ret;
4297                         else
4298                                 allocated = ret;
4299                         goto out2;
4300                 }
4301         }
4302
4303         /*
4304          * requested block isn't allocated yet;
4305          * we couldn't try to create block if create flag is zero
4306          */
4307         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4308                 ext4_lblk_t hole_start, hole_len;
4309
4310                 hole_start = map->m_lblk;
4311                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4312                 /*
4313                  * put just found gap into cache to speed up
4314                  * subsequent requests
4315                  */
4316                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4317
4318                 /* Update hole_len to reflect hole size after map->m_lblk */
4319                 if (hole_start != map->m_lblk)
4320                         hole_len -= map->m_lblk - hole_start;
4321                 map->m_pblk = 0;
4322                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4323
4324                 goto out2;
4325         }
4326
4327         /*
4328          * Okay, we need to do block allocation.
4329          */
4330         newex.ee_block = cpu_to_le32(map->m_lblk);
4331         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4332
4333         /*
4334          * If we are doing bigalloc, check to see if the extent returned
4335          * by ext4_find_extent() implies a cluster we can use.
4336          */
4337         if (cluster_offset && ex &&
4338             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4339                 ar.len = allocated = map->m_len;
4340                 newblock = map->m_pblk;
4341                 map_from_cluster = true;
4342                 goto got_allocated_blocks;
4343         }
4344
4345         /* find neighbour allocated blocks */
4346         ar.lleft = map->m_lblk;
4347         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4348         if (err)
4349                 goto out2;
4350         ar.lright = map->m_lblk;
4351         ex2 = NULL;
4352         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4353         if (err)
4354                 goto out2;
4355
4356         /* Check if the extent after searching to the right implies a
4357          * cluster we can use. */
4358         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4359             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4360                 ar.len = allocated = map->m_len;
4361                 newblock = map->m_pblk;
4362                 map_from_cluster = true;
4363                 goto got_allocated_blocks;
4364         }
4365
4366         /*
4367          * See if request is beyond maximum number of blocks we can have in
4368          * a single extent. For an initialized extent this limit is
4369          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4370          * EXT_UNWRITTEN_MAX_LEN.
4371          */
4372         if (map->m_len > EXT_INIT_MAX_LEN &&
4373             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4374                 map->m_len = EXT_INIT_MAX_LEN;
4375         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4376                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4377                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4378
4379         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4380         newex.ee_len = cpu_to_le16(map->m_len);
4381         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4382         if (err)
4383                 allocated = ext4_ext_get_actual_len(&newex);
4384         else
4385                 allocated = map->m_len;
4386
4387         /* allocate new block */
4388         ar.inode = inode;
4389         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4390         ar.logical = map->m_lblk;
4391         /*
4392          * We calculate the offset from the beginning of the cluster
4393          * for the logical block number, since when we allocate a
4394          * physical cluster, the physical block should start at the
4395          * same offset from the beginning of the cluster.  This is
4396          * needed so that future calls to get_implied_cluster_alloc()
4397          * work correctly.
4398          */
4399         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4400         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4401         ar.goal -= offset;
4402         ar.logical -= offset;
4403         if (S_ISREG(inode->i_mode))
4404                 ar.flags = EXT4_MB_HINT_DATA;
4405         else
4406                 /* disable in-core preallocation for non-regular files */
4407                 ar.flags = 0;
4408         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4409                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4410         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4411                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4412         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4413                 ar.flags |= EXT4_MB_USE_RESERVED;
4414         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4415         if (!newblock)
4416                 goto out2;
4417         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4418                   ar.goal, newblock, allocated);
4419         free_on_err = 1;
4420         allocated_clusters = ar.len;
4421         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4422         if (ar.len > allocated)
4423                 ar.len = allocated;
4424
4425 got_allocated_blocks:
4426         /* try to insert new extent into found leaf and return */
4427         ext4_ext_store_pblock(&newex, newblock + offset);
4428         newex.ee_len = cpu_to_le16(ar.len);
4429         /* Mark unwritten */
4430         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4431                 ext4_ext_mark_unwritten(&newex);
4432                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4433         }
4434
4435         err = 0;
4436         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4437                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4438                                          path, ar.len);
4439         if (!err)
4440                 err = ext4_ext_insert_extent(handle, inode, &path,
4441                                              &newex, flags);
4442
4443         if (err && free_on_err) {
4444                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4445                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4446                 /* free data blocks we just allocated */
4447                 /* not a good idea to call discard here directly,
4448                  * but otherwise we'd need to call it every free() */
4449                 ext4_discard_preallocations(inode);
4450                 ext4_free_blocks(handle, inode, NULL, newblock,
4451                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4452                 goto out2;
4453         }
4454
4455         /* previous routine could use block we allocated */
4456         newblock = ext4_ext_pblock(&newex);
4457         allocated = ext4_ext_get_actual_len(&newex);
4458         if (allocated > map->m_len)
4459                 allocated = map->m_len;
4460         map->m_flags |= EXT4_MAP_NEW;
4461
4462         /*
4463          * Reduce the reserved cluster count to reflect successful deferred
4464          * allocation of delayed allocated clusters or direct allocation of
4465          * clusters discovered to be delayed allocated.  Once allocated, a
4466          * cluster is not included in the reserved count.
4467          */
4468         if (test_opt(inode->i_sb, DELALLOC) && !map_from_cluster) {
4469                 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4470                         /*
4471                          * When allocating delayed allocated clusters, simply
4472                          * reduce the reserved cluster count and claim quota
4473                          */
4474                         ext4_da_update_reserve_space(inode, allocated_clusters,
4475                                                         1);
4476                 } else {
4477                         ext4_lblk_t lblk, len;
4478                         unsigned int n;
4479
4480                         /*
4481                          * When allocating non-delayed allocated clusters
4482                          * (from fallocate, filemap, DIO, or clusters
4483                          * allocated when delalloc has been disabled by
4484                          * ext4_nonda_switch), reduce the reserved cluster
4485                          * count by the number of allocated clusters that
4486                          * have previously been delayed allocated.  Quota
4487                          * has been claimed by ext4_mb_new_blocks() above,
4488                          * so release the quota reservations made for any
4489                          * previously delayed allocated clusters.
4490                          */
4491                         lblk = EXT4_LBLK_CMASK(sbi, map->m_lblk);
4492                         len = allocated_clusters << sbi->s_cluster_bits;
4493                         n = ext4_es_delayed_clu(inode, lblk, len);
4494                         if (n > 0)
4495                                 ext4_da_update_reserve_space(inode, (int) n, 0);
4496                 }
4497         }
4498
4499         /*
4500          * Cache the extent and update transaction to commit on fdatasync only
4501          * when it is _not_ an unwritten extent.
4502          */
4503         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4504                 ext4_update_inode_fsync_trans(handle, inode, 1);
4505         else
4506                 ext4_update_inode_fsync_trans(handle, inode, 0);
4507 out:
4508         if (allocated > map->m_len)
4509                 allocated = map->m_len;
4510         ext4_ext_show_leaf(inode, path);
4511         map->m_flags |= EXT4_MAP_MAPPED;
4512         map->m_pblk = newblock;
4513         map->m_len = allocated;
4514 out2:
4515         ext4_ext_drop_refs(path);
4516         kfree(path);
4517
4518         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4519                                        err ? err : allocated);
4520         return err ? err : allocated;
4521 }
4522
4523 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4524 {
4525         struct super_block *sb = inode->i_sb;
4526         ext4_lblk_t last_block;
4527         int err = 0;
4528
4529         /*
4530          * TODO: optimization is possible here.
4531          * Probably we need not scan at all,
4532          * because page truncation is enough.
4533          */
4534
4535         /* we have to know where to truncate from in crash case */
4536         EXT4_I(inode)->i_disksize = inode->i_size;
4537         err = ext4_mark_inode_dirty(handle, inode);
4538         if (err)
4539                 return err;
4540
4541         last_block = (inode->i_size + sb->s_blocksize - 1)
4542                         >> EXT4_BLOCK_SIZE_BITS(sb);
4543 retry:
4544         err = ext4_es_remove_extent(inode, last_block,
4545                                     EXT_MAX_BLOCKS - last_block);
4546         if (err == -ENOMEM) {
4547                 cond_resched();
4548                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4549                 goto retry;
4550         }
4551         if (err)
4552                 return err;
4553         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4554 }
4555
4556 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4557                                   ext4_lblk_t len, loff_t new_size,
4558                                   int flags)
4559 {
4560         struct inode *inode = file_inode(file);
4561         handle_t *handle;
4562         int ret = 0;
4563         int ret2 = 0;
4564         int retries = 0;
4565         int depth = 0;
4566         struct ext4_map_blocks map;
4567         unsigned int credits;
4568         loff_t epos;
4569
4570         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4571         map.m_lblk = offset;
4572         map.m_len = len;
4573         /*
4574          * Don't normalize the request if it can fit in one extent so
4575          * that it doesn't get unnecessarily split into multiple
4576          * extents.
4577          */
4578         if (len <= EXT_UNWRITTEN_MAX_LEN)
4579                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4580
4581         /*
4582          * credits to insert 1 extent into extent tree
4583          */
4584         credits = ext4_chunk_trans_blocks(inode, len);
4585         depth = ext_depth(inode);
4586
4587 retry:
4588         while (ret >= 0 && len) {
4589                 /*
4590                  * Recalculate credits when extent tree depth changes.
4591                  */
4592                 if (depth != ext_depth(inode)) {
4593                         credits = ext4_chunk_trans_blocks(inode, len);
4594                         depth = ext_depth(inode);
4595                 }
4596
4597                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4598                                             credits);
4599                 if (IS_ERR(handle)) {
4600                         ret = PTR_ERR(handle);
4601                         break;
4602                 }
4603                 ret = ext4_map_blocks(handle, inode, &map, flags);
4604                 if (ret <= 0) {
4605                         ext4_debug("inode #%lu: block %u: len %u: "
4606                                    "ext4_ext_map_blocks returned %d",
4607                                    inode->i_ino, map.m_lblk,
4608                                    map.m_len, ret);
4609                         ext4_mark_inode_dirty(handle, inode);
4610                         ret2 = ext4_journal_stop(handle);
4611                         break;
4612                 }
4613                 map.m_lblk += ret;
4614                 map.m_len = len = len - ret;
4615                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4616                 inode->i_ctime = current_time(inode);
4617                 if (new_size) {
4618                         if (epos > new_size)
4619                                 epos = new_size;
4620                         if (ext4_update_inode_size(inode, epos) & 0x1)
4621                                 inode->i_mtime = inode->i_ctime;
4622                 } else {
4623                         if (epos > inode->i_size)
4624                                 ext4_set_inode_flag(inode,
4625                                                     EXT4_INODE_EOFBLOCKS);
4626                 }
4627                 ext4_mark_inode_dirty(handle, inode);
4628                 ext4_update_inode_fsync_trans(handle, inode, 1);
4629                 ret2 = ext4_journal_stop(handle);
4630                 if (ret2)
4631                         break;
4632         }
4633         if (ret == -ENOSPC &&
4634                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4635                 ret = 0;
4636                 goto retry;
4637         }
4638
4639         return ret > 0 ? ret2 : ret;
4640 }
4641
4642 static long ext4_zero_range(struct file *file, loff_t offset,
4643                             loff_t len, int mode)
4644 {
4645         struct inode *inode = file_inode(file);
4646         handle_t *handle = NULL;
4647         unsigned int max_blocks;
4648         loff_t new_size = 0;
4649         int ret = 0;
4650         int flags;
4651         int credits;
4652         int partial_begin, partial_end;
4653         loff_t start, end;
4654         ext4_lblk_t lblk;
4655         unsigned int blkbits = inode->i_blkbits;
4656
4657         trace_ext4_zero_range(inode, offset, len, mode);
4658
4659         if (!S_ISREG(inode->i_mode))
4660                 return -EINVAL;
4661
4662         /* Call ext4_force_commit to flush all data in case of data=journal. */
4663         if (ext4_should_journal_data(inode)) {
4664                 ret = ext4_force_commit(inode->i_sb);
4665                 if (ret)
4666                         return ret;
4667         }
4668
4669         /*
4670          * Round up offset. This is not fallocate, we neet to zero out
4671          * blocks, so convert interior block aligned part of the range to
4672          * unwritten and possibly manually zero out unaligned parts of the
4673          * range.
4674          */
4675         start = round_up(offset, 1 << blkbits);
4676         end = round_down((offset + len), 1 << blkbits);
4677
4678         if (start < offset || end > offset + len)
4679                 return -EINVAL;
4680         partial_begin = offset & ((1 << blkbits) - 1);
4681         partial_end = (offset + len) & ((1 << blkbits) - 1);
4682
4683         lblk = start >> blkbits;
4684         max_blocks = (end >> blkbits);
4685         if (max_blocks < lblk)
4686                 max_blocks = 0;
4687         else
4688                 max_blocks -= lblk;
4689
4690         inode_lock(inode);
4691
4692         /*
4693          * Indirect files do not support unwritten extnets
4694          */
4695         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4696                 ret = -EOPNOTSUPP;
4697                 goto out_mutex;
4698         }
4699
4700         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4701             (offset + len > i_size_read(inode) ||
4702              offset + len > EXT4_I(inode)->i_disksize)) {
4703                 new_size = offset + len;
4704                 ret = inode_newsize_ok(inode, new_size);
4705                 if (ret)
4706                         goto out_mutex;
4707         }
4708
4709         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4710         if (mode & FALLOC_FL_KEEP_SIZE)
4711                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4712
4713         /* Wait all existing dio workers, newcomers will block on i_mutex */
4714         inode_dio_wait(inode);
4715
4716         /* Preallocate the range including the unaligned edges */
4717         if (partial_begin || partial_end) {
4718                 ret = ext4_alloc_file_blocks(file,
4719                                 round_down(offset, 1 << blkbits) >> blkbits,
4720                                 (round_up((offset + len), 1 << blkbits) -
4721                                  round_down(offset, 1 << blkbits)) >> blkbits,
4722                                 new_size, flags);
4723                 if (ret)
4724                         goto out_mutex;
4725
4726         }
4727
4728         /* Zero range excluding the unaligned edges */
4729         if (max_blocks > 0) {
4730                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4731                           EXT4_EX_NOCACHE);
4732
4733                 /*
4734                  * Prevent page faults from reinstantiating pages we have
4735                  * released from page cache.
4736                  */
4737                 down_write(&EXT4_I(inode)->i_mmap_sem);
4738
4739                 ret = ext4_break_layouts(inode);
4740                 if (ret) {
4741                         up_write(&EXT4_I(inode)->i_mmap_sem);
4742                         goto out_mutex;
4743                 }
4744
4745                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4746                 if (ret) {
4747                         up_write(&EXT4_I(inode)->i_mmap_sem);
4748                         goto out_mutex;
4749                 }
4750                 /* Now release the pages and zero block aligned part of pages */
4751                 truncate_pagecache_range(inode, start, end - 1);
4752                 inode->i_mtime = inode->i_ctime = current_time(inode);
4753
4754                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4755                                              flags);
4756                 up_write(&EXT4_I(inode)->i_mmap_sem);
4757                 if (ret)
4758                         goto out_mutex;
4759         }
4760         if (!partial_begin && !partial_end)
4761                 goto out_mutex;
4762
4763         /*
4764          * In worst case we have to writeout two nonadjacent unwritten
4765          * blocks and update the inode
4766          */
4767         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4768         if (ext4_should_journal_data(inode))
4769                 credits += 2;
4770         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4771         if (IS_ERR(handle)) {
4772                 ret = PTR_ERR(handle);
4773                 ext4_std_error(inode->i_sb, ret);
4774                 goto out_mutex;
4775         }
4776
4777         inode->i_mtime = inode->i_ctime = current_time(inode);
4778         if (new_size) {
4779                 ext4_update_inode_size(inode, new_size);
4780         } else {
4781                 /*
4782                 * Mark that we allocate beyond EOF so the subsequent truncate
4783                 * can proceed even if the new size is the same as i_size.
4784                 */
4785                 if ((offset + len) > i_size_read(inode))
4786                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4787         }
4788         ext4_mark_inode_dirty(handle, inode);
4789
4790         /* Zero out partial block at the edges of the range */
4791         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4792         if (ret >= 0)
4793                 ext4_update_inode_fsync_trans(handle, inode, 1);
4794
4795         if (file->f_flags & O_SYNC)
4796                 ext4_handle_sync(handle);
4797
4798         ext4_journal_stop(handle);
4799 out_mutex:
4800         inode_unlock(inode);
4801         return ret;
4802 }
4803
4804 /*
4805  * preallocate space for a file. This implements ext4's fallocate file
4806  * operation, which gets called from sys_fallocate system call.
4807  * For block-mapped files, posix_fallocate should fall back to the method
4808  * of writing zeroes to the required new blocks (the same behavior which is
4809  * expected for file systems which do not support fallocate() system call).
4810  */
4811 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4812 {
4813         struct inode *inode = file_inode(file);
4814         loff_t new_size = 0;
4815         unsigned int max_blocks;
4816         int ret = 0;
4817         int flags;
4818         ext4_lblk_t lblk;
4819         unsigned int blkbits = inode->i_blkbits;
4820
4821         /*
4822          * Encrypted inodes can't handle collapse range or insert
4823          * range since we would need to re-encrypt blocks with a
4824          * different IV or XTS tweak (which are based on the logical
4825          * block number).
4826          *
4827          * XXX It's not clear why zero range isn't working, but we'll
4828          * leave it disabled for encrypted inodes for now.  This is a
4829          * bug we should fix....
4830          */
4831         if (IS_ENCRYPTED(inode) &&
4832             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4833                      FALLOC_FL_ZERO_RANGE)))
4834                 return -EOPNOTSUPP;
4835
4836         /* Return error if mode is not supported */
4837         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4838                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4839                      FALLOC_FL_INSERT_RANGE))
4840                 return -EOPNOTSUPP;
4841
4842         if (mode & FALLOC_FL_PUNCH_HOLE)
4843                 return ext4_punch_hole(inode, offset, len);
4844
4845         ret = ext4_convert_inline_data(inode);
4846         if (ret)
4847                 return ret;
4848
4849         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4850                 return ext4_collapse_range(inode, offset, len);
4851
4852         if (mode & FALLOC_FL_INSERT_RANGE)
4853                 return ext4_insert_range(inode, offset, len);
4854
4855         if (mode & FALLOC_FL_ZERO_RANGE)
4856                 return ext4_zero_range(file, offset, len, mode);
4857
4858         trace_ext4_fallocate_enter(inode, offset, len, mode);
4859         lblk = offset >> blkbits;
4860
4861         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4862         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4863         if (mode & FALLOC_FL_KEEP_SIZE)
4864                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4865
4866         inode_lock(inode);
4867
4868         /*
4869          * We only support preallocation for extent-based files only
4870          */
4871         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4872                 ret = -EOPNOTSUPP;
4873                 goto out;
4874         }
4875
4876         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4877             (offset + len > i_size_read(inode) ||
4878              offset + len > EXT4_I(inode)->i_disksize)) {
4879                 new_size = offset + len;
4880                 ret = inode_newsize_ok(inode, new_size);
4881                 if (ret)
4882                         goto out;
4883         }
4884
4885         /* Wait all existing dio workers, newcomers will block on i_mutex */
4886         inode_dio_wait(inode);
4887
4888         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
4889         if (ret)
4890                 goto out;
4891
4892         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4893                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4894                                                 EXT4_I(inode)->i_sync_tid);
4895         }
4896 out:
4897         inode_unlock(inode);
4898         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4899         return ret;
4900 }
4901
4902 /*
4903  * This function convert a range of blocks to written extents
4904  * The caller of this function will pass the start offset and the size.
4905  * all unwritten extents within this range will be converted to
4906  * written extents.
4907  *
4908  * This function is called from the direct IO end io call back
4909  * function, to convert the fallocated extents after IO is completed.
4910  * Returns 0 on success.
4911  */
4912 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4913                                    loff_t offset, ssize_t len)
4914 {
4915         unsigned int max_blocks;
4916         int ret = 0;
4917         int ret2 = 0;
4918         struct ext4_map_blocks map;
4919         unsigned int credits, blkbits = inode->i_blkbits;
4920
4921         map.m_lblk = offset >> blkbits;
4922         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4923
4924         /*
4925          * This is somewhat ugly but the idea is clear: When transaction is
4926          * reserved, everything goes into it. Otherwise we rather start several
4927          * smaller transactions for conversion of each extent separately.
4928          */
4929         if (handle) {
4930                 handle = ext4_journal_start_reserved(handle,
4931                                                      EXT4_HT_EXT_CONVERT);
4932                 if (IS_ERR(handle))
4933                         return PTR_ERR(handle);
4934                 credits = 0;
4935         } else {
4936                 /*
4937                  * credits to insert 1 extent into extent tree
4938                  */
4939                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
4940         }
4941         while (ret >= 0 && ret < max_blocks) {
4942                 map.m_lblk += ret;
4943                 map.m_len = (max_blocks -= ret);
4944                 if (credits) {
4945                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4946                                                     credits);
4947                         if (IS_ERR(handle)) {
4948                                 ret = PTR_ERR(handle);
4949                                 break;
4950                         }
4951                 }
4952                 ret = ext4_map_blocks(handle, inode, &map,
4953                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4954                 if (ret <= 0)
4955                         ext4_warning(inode->i_sb,
4956                                      "inode #%lu: block %u: len %u: "
4957                                      "ext4_ext_map_blocks returned %d",
4958                                      inode->i_ino, map.m_lblk,
4959                                      map.m_len, ret);
4960                 ext4_mark_inode_dirty(handle, inode);
4961                 if (credits)
4962                         ret2 = ext4_journal_stop(handle);
4963                 if (ret <= 0 || ret2)
4964                         break;
4965         }
4966         if (!credits)
4967                 ret2 = ext4_journal_stop(handle);
4968         return ret > 0 ? ret2 : ret;
4969 }
4970
4971 /*
4972  * If newes is not existing extent (newes->ec_pblk equals zero) find
4973  * delayed extent at start of newes and update newes accordingly and
4974  * return start of the next delayed extent.
4975  *
4976  * If newes is existing extent (newes->ec_pblk is not equal zero)
4977  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4978  * extent found. Leave newes unmodified.
4979  */
4980 static int ext4_find_delayed_extent(struct inode *inode,
4981                                     struct extent_status *newes)
4982 {
4983         struct extent_status es;
4984         ext4_lblk_t block, next_del;
4985
4986         if (newes->es_pblk == 0) {
4987                 ext4_es_find_extent_range(inode, &ext4_es_is_delayed,
4988                                           newes->es_lblk,
4989                                           newes->es_lblk + newes->es_len - 1,
4990                                           &es);
4991
4992                 /*
4993                  * No extent in extent-tree contains block @newes->es_pblk,
4994                  * then the block may stay in 1)a hole or 2)delayed-extent.
4995                  */
4996                 if (es.es_len == 0)
4997                         /* A hole found. */
4998                         return 0;
4999
5000                 if (es.es_lblk > newes->es_lblk) {
5001                         /* A hole found. */
5002                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5003                                             newes->es_len);
5004                         return 0;
5005                 }
5006
5007                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5008         }
5009
5010         block = newes->es_lblk + newes->es_len;
5011         ext4_es_find_extent_range(inode, &ext4_es_is_delayed, block,
5012                                   EXT_MAX_BLOCKS, &es);
5013         if (es.es_len == 0)
5014                 next_del = EXT_MAX_BLOCKS;
5015         else
5016                 next_del = es.es_lblk;
5017
5018         return next_del;
5019 }
5020 /* fiemap flags we can handle specified here */
5021 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5022
5023 static int ext4_xattr_fiemap(struct inode *inode,
5024                                 struct fiemap_extent_info *fieinfo)
5025 {
5026         __u64 physical = 0;
5027         __u64 length;
5028         __u32 flags = FIEMAP_EXTENT_LAST;
5029         int blockbits = inode->i_sb->s_blocksize_bits;
5030         int error = 0;
5031
5032         /* in-inode? */
5033         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5034                 struct ext4_iloc iloc;
5035                 int offset;     /* offset of xattr in inode */
5036
5037                 error = ext4_get_inode_loc(inode, &iloc);
5038                 if (error)
5039                         return error;
5040                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5041                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5042                                 EXT4_I(inode)->i_extra_isize;
5043                 physical += offset;
5044                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5045                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5046                 brelse(iloc.bh);
5047         } else { /* external block */
5048                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5049                 length = inode->i_sb->s_blocksize;
5050         }
5051
5052         if (physical)
5053                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5054                                                 length, flags);
5055         return (error < 0 ? error : 0);
5056 }
5057
5058 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5059                 __u64 start, __u64 len)
5060 {
5061         ext4_lblk_t start_blk;
5062         int error = 0;
5063
5064         if (ext4_has_inline_data(inode)) {
5065                 int has_inline = 1;
5066
5067                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5068                                                 start, len);
5069
5070                 if (has_inline)
5071                         return error;
5072         }
5073
5074         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5075                 error = ext4_ext_precache(inode);
5076                 if (error)
5077                         return error;
5078         }
5079
5080         /* fallback to generic here if not in extents fmt */
5081         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5082                 return generic_block_fiemap(inode, fieinfo, start, len,
5083                         ext4_get_block);
5084
5085         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5086                 return -EBADR;
5087
5088         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5089                 error = ext4_xattr_fiemap(inode, fieinfo);
5090         } else {
5091                 ext4_lblk_t len_blks;
5092                 __u64 last_blk;
5093
5094                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5095                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5096                 if (last_blk >= EXT_MAX_BLOCKS)
5097                         last_blk = EXT_MAX_BLOCKS-1;
5098                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5099
5100                 /*
5101                  * Walk the extent tree gathering extent information
5102                  * and pushing extents back to the user.
5103                  */
5104                 error = ext4_fill_fiemap_extents(inode, start_blk,
5105                                                  len_blks, fieinfo);
5106         }
5107         return error;
5108 }
5109
5110 /*
5111  * ext4_access_path:
5112  * Function to access the path buffer for marking it dirty.
5113  * It also checks if there are sufficient credits left in the journal handle
5114  * to update path.
5115  */
5116 static int
5117 ext4_access_path(handle_t *handle, struct inode *inode,
5118                 struct ext4_ext_path *path)
5119 {
5120         int credits, err;
5121
5122         if (!ext4_handle_valid(handle))
5123                 return 0;
5124
5125         /*
5126          * Check if need to extend journal credits
5127          * 3 for leaf, sb, and inode plus 2 (bmap and group
5128          * descriptor) for each block group; assume two block
5129          * groups
5130          */
5131         if (handle->h_buffer_credits < 7) {
5132                 credits = ext4_writepage_trans_blocks(inode);
5133                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5134                 /* EAGAIN is success */
5135                 if (err && err != -EAGAIN)
5136                         return err;
5137         }
5138
5139         err = ext4_ext_get_access(handle, inode, path);
5140         return err;
5141 }
5142
5143 /*
5144  * ext4_ext_shift_path_extents:
5145  * Shift the extents of a path structure lying between path[depth].p_ext
5146  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5147  * if it is right shift or left shift operation.
5148  */
5149 static int
5150 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5151                             struct inode *inode, handle_t *handle,
5152                             enum SHIFT_DIRECTION SHIFT)
5153 {
5154         int depth, err = 0;
5155         struct ext4_extent *ex_start, *ex_last;
5156         bool update = 0;
5157         depth = path->p_depth;
5158
5159         while (depth >= 0) {
5160                 if (depth == path->p_depth) {
5161                         ex_start = path[depth].p_ext;
5162                         if (!ex_start)
5163                                 return -EFSCORRUPTED;
5164
5165                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5166
5167                         err = ext4_access_path(handle, inode, path + depth);
5168                         if (err)
5169                                 goto out;
5170
5171                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5172                                 update = 1;
5173
5174                         while (ex_start <= ex_last) {
5175                                 if (SHIFT == SHIFT_LEFT) {
5176                                         le32_add_cpu(&ex_start->ee_block,
5177                                                 -shift);
5178                                         /* Try to merge to the left. */
5179                                         if ((ex_start >
5180                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5181                                             &&
5182                                             ext4_ext_try_to_merge_right(inode,
5183                                             path, ex_start - 1))
5184                                                 ex_last--;
5185                                         else
5186                                                 ex_start++;
5187                                 } else {
5188                                         le32_add_cpu(&ex_last->ee_block, shift);
5189                                         ext4_ext_try_to_merge_right(inode, path,
5190                                                 ex_last);
5191                                         ex_last--;
5192                                 }
5193                         }
5194                         err = ext4_ext_dirty(handle, inode, path + depth);
5195                         if (err)
5196                                 goto out;
5197
5198                         if (--depth < 0 || !update)
5199                                 break;
5200                 }
5201
5202                 /* Update index too */
5203                 err = ext4_access_path(handle, inode, path + depth);
5204                 if (err)
5205                         goto out;
5206
5207                 if (SHIFT == SHIFT_LEFT)
5208                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5209                 else
5210                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5211                 err = ext4_ext_dirty(handle, inode, path + depth);
5212                 if (err)
5213                         goto out;
5214
5215                 /* we are done if current index is not a starting index */
5216                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5217                         break;
5218
5219                 depth--;
5220         }
5221
5222 out:
5223         return err;
5224 }
5225
5226 /*
5227  * ext4_ext_shift_extents:
5228  * All the extents which lies in the range from @start to the last allocated
5229  * block for the @inode are shifted either towards left or right (depending
5230  * upon @SHIFT) by @shift blocks.
5231  * On success, 0 is returned, error otherwise.
5232  */
5233 static int
5234 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5235                        ext4_lblk_t start, ext4_lblk_t shift,
5236                        enum SHIFT_DIRECTION SHIFT)
5237 {
5238         struct ext4_ext_path *path;
5239         int ret = 0, depth;
5240         struct ext4_extent *extent;
5241         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5242
5243         /* Let path point to the last extent */
5244         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5245                                 EXT4_EX_NOCACHE);
5246         if (IS_ERR(path))
5247                 return PTR_ERR(path);
5248
5249         depth = path->p_depth;
5250         extent = path[depth].p_ext;
5251         if (!extent)
5252                 goto out;
5253
5254         stop = le32_to_cpu(extent->ee_block);
5255
5256        /*
5257         * For left shifts, make sure the hole on the left is big enough to
5258         * accommodate the shift.  For right shifts, make sure the last extent
5259         * won't be shifted beyond EXT_MAX_BLOCKS.
5260         */
5261         if (SHIFT == SHIFT_LEFT) {
5262                 path = ext4_find_extent(inode, start - 1, &path,
5263                                         EXT4_EX_NOCACHE);
5264                 if (IS_ERR(path))
5265                         return PTR_ERR(path);
5266                 depth = path->p_depth;
5267                 extent =  path[depth].p_ext;
5268                 if (extent) {
5269                         ex_start = le32_to_cpu(extent->ee_block);
5270                         ex_end = le32_to_cpu(extent->ee_block) +
5271                                 ext4_ext_get_actual_len(extent);
5272                 } else {
5273                         ex_start = 0;
5274                         ex_end = 0;
5275                 }
5276
5277                 if ((start == ex_start && shift > ex_start) ||
5278                     (shift > start - ex_end)) {
5279                         ret = -EINVAL;
5280                         goto out;
5281                 }
5282         } else {
5283                 if (shift > EXT_MAX_BLOCKS -
5284                     (stop + ext4_ext_get_actual_len(extent))) {
5285                         ret = -EINVAL;
5286                         goto out;
5287                 }
5288         }
5289
5290         /*
5291          * In case of left shift, iterator points to start and it is increased
5292          * till we reach stop. In case of right shift, iterator points to stop
5293          * and it is decreased till we reach start.
5294          */
5295         if (SHIFT == SHIFT_LEFT)
5296                 iterator = &start;
5297         else
5298                 iterator = &stop;
5299
5300         /*
5301          * Its safe to start updating extents.  Start and stop are unsigned, so
5302          * in case of right shift if extent with 0 block is reached, iterator
5303          * becomes NULL to indicate the end of the loop.
5304          */
5305         while (iterator && start <= stop) {
5306                 path = ext4_find_extent(inode, *iterator, &path,
5307                                         EXT4_EX_NOCACHE);
5308                 if (IS_ERR(path))
5309                         return PTR_ERR(path);
5310                 depth = path->p_depth;
5311                 extent = path[depth].p_ext;
5312                 if (!extent) {
5313                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5314                                          (unsigned long) *iterator);
5315                         return -EFSCORRUPTED;
5316                 }
5317                 if (SHIFT == SHIFT_LEFT && *iterator >
5318                     le32_to_cpu(extent->ee_block)) {
5319                         /* Hole, move to the next extent */
5320                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5321                                 path[depth].p_ext++;
5322                         } else {
5323                                 *iterator = ext4_ext_next_allocated_block(path);
5324                                 continue;
5325                         }
5326                 }
5327
5328                 if (SHIFT == SHIFT_LEFT) {
5329                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5330                         *iterator = le32_to_cpu(extent->ee_block) +
5331                                         ext4_ext_get_actual_len(extent);
5332                 } else {
5333                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5334                         if (le32_to_cpu(extent->ee_block) > 0)
5335                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5336                         else
5337                                 /* Beginning is reached, end of the loop */
5338                                 iterator = NULL;
5339                         /* Update path extent in case we need to stop */
5340                         while (le32_to_cpu(extent->ee_block) < start)
5341                                 extent++;
5342                         path[depth].p_ext = extent;
5343                 }
5344                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5345                                 handle, SHIFT);
5346                 if (ret)
5347                         break;
5348         }
5349 out:
5350         ext4_ext_drop_refs(path);
5351         kfree(path);
5352         return ret;
5353 }
5354
5355 /*
5356  * ext4_collapse_range:
5357  * This implements the fallocate's collapse range functionality for ext4
5358  * Returns: 0 and non-zero on error.
5359  */
5360 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5361 {
5362         struct super_block *sb = inode->i_sb;
5363         ext4_lblk_t punch_start, punch_stop;
5364         handle_t *handle;
5365         unsigned int credits;
5366         loff_t new_size, ioffset;
5367         int ret;
5368
5369         /*
5370          * We need to test this early because xfstests assumes that a
5371          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5372          * system does not support collapse range.
5373          */
5374         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5375                 return -EOPNOTSUPP;
5376
5377         /* Collapse range works only on fs block size aligned offsets. */
5378         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5379             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5380                 return -EINVAL;
5381
5382         if (!S_ISREG(inode->i_mode))
5383                 return -EINVAL;
5384
5385         trace_ext4_collapse_range(inode, offset, len);
5386
5387         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5388         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5389
5390         /* Call ext4_force_commit to flush all data in case of data=journal. */
5391         if (ext4_should_journal_data(inode)) {
5392                 ret = ext4_force_commit(inode->i_sb);
5393                 if (ret)
5394                         return ret;
5395         }
5396
5397         inode_lock(inode);
5398         /*
5399          * There is no need to overlap collapse range with EOF, in which case
5400          * it is effectively a truncate operation
5401          */
5402         if (offset + len >= i_size_read(inode)) {
5403                 ret = -EINVAL;
5404                 goto out_mutex;
5405         }
5406
5407         /* Currently just for extent based files */
5408         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5409                 ret = -EOPNOTSUPP;
5410                 goto out_mutex;
5411         }
5412
5413         /* Wait for existing dio to complete */
5414         inode_dio_wait(inode);
5415
5416         /*
5417          * Prevent page faults from reinstantiating pages we have released from
5418          * page cache.
5419          */
5420         down_write(&EXT4_I(inode)->i_mmap_sem);
5421
5422         ret = ext4_break_layouts(inode);
5423         if (ret)
5424                 goto out_mmap;
5425
5426         /*
5427          * Need to round down offset to be aligned with page size boundary
5428          * for page size > block size.
5429          */
5430         ioffset = round_down(offset, PAGE_SIZE);
5431         /*
5432          * Write tail of the last page before removed range since it will get
5433          * removed from the page cache below.
5434          */
5435         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5436         if (ret)
5437                 goto out_mmap;
5438         /*
5439          * Write data that will be shifted to preserve them when discarding
5440          * page cache below. We are also protected from pages becoming dirty
5441          * by i_mmap_sem.
5442          */
5443         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5444                                            LLONG_MAX);
5445         if (ret)
5446                 goto out_mmap;
5447         truncate_pagecache(inode, ioffset);
5448
5449         credits = ext4_writepage_trans_blocks(inode);
5450         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5451         if (IS_ERR(handle)) {
5452                 ret = PTR_ERR(handle);
5453                 goto out_mmap;
5454         }
5455
5456         down_write(&EXT4_I(inode)->i_data_sem);
5457         ext4_discard_preallocations(inode);
5458
5459         ret = ext4_es_remove_extent(inode, punch_start,
5460                                     EXT_MAX_BLOCKS - punch_start);
5461         if (ret) {
5462                 up_write(&EXT4_I(inode)->i_data_sem);
5463                 goto out_stop;
5464         }
5465
5466         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5467         if (ret) {
5468                 up_write(&EXT4_I(inode)->i_data_sem);
5469                 goto out_stop;
5470         }
5471         ext4_discard_preallocations(inode);
5472
5473         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5474                                      punch_stop - punch_start, SHIFT_LEFT);
5475         if (ret) {
5476                 up_write(&EXT4_I(inode)->i_data_sem);
5477                 goto out_stop;
5478         }
5479
5480         new_size = i_size_read(inode) - len;
5481         i_size_write(inode, new_size);
5482         EXT4_I(inode)->i_disksize = new_size;
5483
5484         up_write(&EXT4_I(inode)->i_data_sem);
5485         if (IS_SYNC(inode))
5486                 ext4_handle_sync(handle);
5487         inode->i_mtime = inode->i_ctime = current_time(inode);
5488         ext4_mark_inode_dirty(handle, inode);
5489         ext4_update_inode_fsync_trans(handle, inode, 1);
5490
5491 out_stop:
5492         ext4_journal_stop(handle);
5493 out_mmap:
5494         up_write(&EXT4_I(inode)->i_mmap_sem);
5495 out_mutex:
5496         inode_unlock(inode);
5497         return ret;
5498 }
5499
5500 /*
5501  * ext4_insert_range:
5502  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5503  * The data blocks starting from @offset to the EOF are shifted by @len
5504  * towards right to create a hole in the @inode. Inode size is increased
5505  * by len bytes.
5506  * Returns 0 on success, error otherwise.
5507  */
5508 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5509 {
5510         struct super_block *sb = inode->i_sb;
5511         handle_t *handle;
5512         struct ext4_ext_path *path;
5513         struct ext4_extent *extent;
5514         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5515         unsigned int credits, ee_len;
5516         int ret = 0, depth, split_flag = 0;
5517         loff_t ioffset;
5518
5519         /*
5520          * We need to test this early because xfstests assumes that an
5521          * insert range of (0, 1) will return EOPNOTSUPP if the file
5522          * system does not support insert range.
5523          */
5524         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5525                 return -EOPNOTSUPP;
5526
5527         /* Insert range works only on fs block size aligned offsets. */
5528         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5529                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5530                 return -EINVAL;
5531
5532         if (!S_ISREG(inode->i_mode))
5533                 return -EOPNOTSUPP;
5534
5535         trace_ext4_insert_range(inode, offset, len);
5536
5537         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5538         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5539
5540         /* Call ext4_force_commit to flush all data in case of data=journal */
5541         if (ext4_should_journal_data(inode)) {
5542                 ret = ext4_force_commit(inode->i_sb);
5543                 if (ret)
5544                         return ret;
5545         }
5546
5547         inode_lock(inode);
5548         /* Currently just for extent based files */
5549         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5550                 ret = -EOPNOTSUPP;
5551                 goto out_mutex;
5552         }
5553
5554         /* Check for wrap through zero */
5555         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5556                 ret = -EFBIG;
5557                 goto out_mutex;
5558         }
5559
5560         /* Offset should be less than i_size */
5561         if (offset >= i_size_read(inode)) {
5562                 ret = -EINVAL;
5563                 goto out_mutex;
5564         }
5565
5566         /* Wait for existing dio to complete */
5567         inode_dio_wait(inode);
5568
5569         /*
5570          * Prevent page faults from reinstantiating pages we have released from
5571          * page cache.
5572          */
5573         down_write(&EXT4_I(inode)->i_mmap_sem);
5574
5575         ret = ext4_break_layouts(inode);
5576         if (ret)
5577                 goto out_mmap;
5578
5579         /*
5580          * Need to round down to align start offset to page size boundary
5581          * for page size > block size.
5582          */
5583         ioffset = round_down(offset, PAGE_SIZE);
5584         /* Write out all dirty pages */
5585         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5586                         LLONG_MAX);
5587         if (ret)
5588                 goto out_mmap;
5589         truncate_pagecache(inode, ioffset);
5590
5591         credits = ext4_writepage_trans_blocks(inode);
5592         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5593         if (IS_ERR(handle)) {
5594                 ret = PTR_ERR(handle);
5595                 goto out_mmap;
5596         }
5597
5598         /* Expand file to avoid data loss if there is error while shifting */
5599         inode->i_size += len;
5600         EXT4_I(inode)->i_disksize += len;
5601         inode->i_mtime = inode->i_ctime = current_time(inode);
5602         ret = ext4_mark_inode_dirty(handle, inode);
5603         if (ret)
5604                 goto out_stop;
5605
5606         down_write(&EXT4_I(inode)->i_data_sem);
5607         ext4_discard_preallocations(inode);
5608
5609         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5610         if (IS_ERR(path)) {
5611                 up_write(&EXT4_I(inode)->i_data_sem);
5612                 goto out_stop;
5613         }
5614
5615         depth = ext_depth(inode);
5616         extent = path[depth].p_ext;
5617         if (extent) {
5618                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5619                 ee_len = ext4_ext_get_actual_len(extent);
5620
5621                 /*
5622                  * If offset_lblk is not the starting block of extent, split
5623                  * the extent @offset_lblk
5624                  */
5625                 if ((offset_lblk > ee_start_lblk) &&
5626                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5627                         if (ext4_ext_is_unwritten(extent))
5628                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5629                                         EXT4_EXT_MARK_UNWRIT2;
5630                         ret = ext4_split_extent_at(handle, inode, &path,
5631                                         offset_lblk, split_flag,
5632                                         EXT4_EX_NOCACHE |
5633                                         EXT4_GET_BLOCKS_PRE_IO |
5634                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5635                 }
5636
5637                 ext4_ext_drop_refs(path);
5638                 kfree(path);
5639                 if (ret < 0) {
5640                         up_write(&EXT4_I(inode)->i_data_sem);
5641                         goto out_stop;
5642                 }
5643         } else {
5644                 ext4_ext_drop_refs(path);
5645                 kfree(path);
5646         }
5647
5648         ret = ext4_es_remove_extent(inode, offset_lblk,
5649                         EXT_MAX_BLOCKS - offset_lblk);
5650         if (ret) {
5651                 up_write(&EXT4_I(inode)->i_data_sem);
5652                 goto out_stop;
5653         }
5654
5655         /*
5656          * if offset_lblk lies in a hole which is at start of file, use
5657          * ee_start_lblk to shift extents
5658          */
5659         ret = ext4_ext_shift_extents(inode, handle,
5660                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5661                 len_lblk, SHIFT_RIGHT);
5662
5663         up_write(&EXT4_I(inode)->i_data_sem);
5664         if (IS_SYNC(inode))
5665                 ext4_handle_sync(handle);
5666         if (ret >= 0)
5667                 ext4_update_inode_fsync_trans(handle, inode, 1);
5668
5669 out_stop:
5670         ext4_journal_stop(handle);
5671 out_mmap:
5672         up_write(&EXT4_I(inode)->i_mmap_sem);
5673 out_mutex:
5674         inode_unlock(inode);
5675         return ret;
5676 }
5677
5678 /**
5679  * ext4_swap_extents - Swap extents between two inodes
5680  *
5681  * @inode1:     First inode
5682  * @inode2:     Second inode
5683  * @lblk1:      Start block for first inode
5684  * @lblk2:      Start block for second inode
5685  * @count:      Number of blocks to swap
5686  * @unwritten: Mark second inode's extents as unwritten after swap
5687  * @erp:        Pointer to save error value
5688  *
5689  * This helper routine does exactly what is promise "swap extents". All other
5690  * stuff such as page-cache locking consistency, bh mapping consistency or
5691  * extent's data copying must be performed by caller.
5692  * Locking:
5693  *              i_mutex is held for both inodes
5694  *              i_data_sem is locked for write for both inodes
5695  * Assumptions:
5696  *              All pages from requested range are locked for both inodes
5697  */
5698 int
5699 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5700                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5701                   ext4_lblk_t count, int unwritten, int *erp)
5702 {
5703         struct ext4_ext_path *path1 = NULL;
5704         struct ext4_ext_path *path2 = NULL;
5705         int replaced_count = 0;
5706
5707         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5708         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5709         BUG_ON(!inode_is_locked(inode1));
5710         BUG_ON(!inode_is_locked(inode2));
5711
5712         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5713         if (unlikely(*erp))
5714                 return 0;
5715         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5716         if (unlikely(*erp))
5717                 return 0;
5718
5719         while (count) {
5720                 struct ext4_extent *ex1, *ex2, tmp_ex;
5721                 ext4_lblk_t e1_blk, e2_blk;
5722                 int e1_len, e2_len, len;
5723                 int split = 0;
5724
5725                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5726                 if (IS_ERR(path1)) {
5727                         *erp = PTR_ERR(path1);
5728                         path1 = NULL;
5729                 finish:
5730                         count = 0;
5731                         goto repeat;
5732                 }
5733                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5734                 if (IS_ERR(path2)) {
5735                         *erp = PTR_ERR(path2);
5736                         path2 = NULL;
5737                         goto finish;
5738                 }
5739                 ex1 = path1[path1->p_depth].p_ext;
5740                 ex2 = path2[path2->p_depth].p_ext;
5741                 /* Do we have somthing to swap ? */
5742                 if (unlikely(!ex2 || !ex1))
5743                         goto finish;
5744
5745                 e1_blk = le32_to_cpu(ex1->ee_block);
5746                 e2_blk = le32_to_cpu(ex2->ee_block);
5747                 e1_len = ext4_ext_get_actual_len(ex1);
5748                 e2_len = ext4_ext_get_actual_len(ex2);
5749
5750                 /* Hole handling */
5751                 if (!in_range(lblk1, e1_blk, e1_len) ||
5752                     !in_range(lblk2, e2_blk, e2_len)) {
5753                         ext4_lblk_t next1, next2;
5754
5755                         /* if hole after extent, then go to next extent */
5756                         next1 = ext4_ext_next_allocated_block(path1);
5757                         next2 = ext4_ext_next_allocated_block(path2);
5758                         /* If hole before extent, then shift to that extent */
5759                         if (e1_blk > lblk1)
5760                                 next1 = e1_blk;
5761                         if (e2_blk > lblk2)
5762                                 next2 = e2_blk;
5763                         /* Do we have something to swap */
5764                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5765                                 goto finish;
5766                         /* Move to the rightest boundary */
5767                         len = next1 - lblk1;
5768                         if (len < next2 - lblk2)
5769                                 len = next2 - lblk2;
5770                         if (len > count)
5771                                 len = count;
5772                         lblk1 += len;
5773                         lblk2 += len;
5774                         count -= len;
5775                         goto repeat;
5776                 }
5777
5778                 /* Prepare left boundary */
5779                 if (e1_blk < lblk1) {
5780                         split = 1;
5781                         *erp = ext4_force_split_extent_at(handle, inode1,
5782                                                 &path1, lblk1, 0);
5783                         if (unlikely(*erp))
5784                                 goto finish;
5785                 }
5786                 if (e2_blk < lblk2) {
5787                         split = 1;
5788                         *erp = ext4_force_split_extent_at(handle, inode2,
5789                                                 &path2,  lblk2, 0);
5790                         if (unlikely(*erp))
5791                                 goto finish;
5792                 }
5793                 /* ext4_split_extent_at() may result in leaf extent split,
5794                  * path must to be revalidated. */
5795                 if (split)
5796                         goto repeat;
5797
5798                 /* Prepare right boundary */
5799                 len = count;
5800                 if (len > e1_blk + e1_len - lblk1)
5801                         len = e1_blk + e1_len - lblk1;
5802                 if (len > e2_blk + e2_len - lblk2)
5803                         len = e2_blk + e2_len - lblk2;
5804
5805                 if (len != e1_len) {
5806                         split = 1;
5807                         *erp = ext4_force_split_extent_at(handle, inode1,
5808                                                 &path1, lblk1 + len, 0);
5809                         if (unlikely(*erp))
5810                                 goto finish;
5811                 }
5812                 if (len != e2_len) {
5813                         split = 1;
5814                         *erp = ext4_force_split_extent_at(handle, inode2,
5815                                                 &path2, lblk2 + len, 0);
5816                         if (*erp)
5817                                 goto finish;
5818                 }
5819                 /* ext4_split_extent_at() may result in leaf extent split,
5820                  * path must to be revalidated. */
5821                 if (split)
5822                         goto repeat;
5823
5824                 BUG_ON(e2_len != e1_len);
5825                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5826                 if (unlikely(*erp))
5827                         goto finish;
5828                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5829                 if (unlikely(*erp))
5830                         goto finish;
5831
5832                 /* Both extents are fully inside boundaries. Swap it now */
5833                 tmp_ex = *ex1;
5834                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5835                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5836                 ex1->ee_len = cpu_to_le16(e2_len);
5837                 ex2->ee_len = cpu_to_le16(e1_len);
5838                 if (unwritten)
5839                         ext4_ext_mark_unwritten(ex2);
5840                 if (ext4_ext_is_unwritten(&tmp_ex))
5841                         ext4_ext_mark_unwritten(ex1);
5842
5843                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5844                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5845                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5846                                       path2->p_depth);
5847                 if (unlikely(*erp))
5848                         goto finish;
5849                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5850                                       path1->p_depth);
5851                 /*
5852                  * Looks scarry ah..? second inode already points to new blocks,
5853                  * and it was successfully dirtied. But luckily error may happen
5854                  * only due to journal error, so full transaction will be
5855                  * aborted anyway.
5856                  */
5857                 if (unlikely(*erp))
5858                         goto finish;
5859                 lblk1 += len;
5860                 lblk2 += len;
5861                 replaced_count += len;
5862                 count -= len;
5863
5864         repeat:
5865                 ext4_ext_drop_refs(path1);
5866                 kfree(path1);
5867                 ext4_ext_drop_refs(path2);
5868                 kfree(path2);
5869                 path1 = path2 = NULL;
5870         }
5871         return replaced_count;
5872 }
5873
5874 /*
5875  * ext4_clu_mapped - determine whether any block in a logical cluster has
5876  *                   been mapped to a physical cluster
5877  *
5878  * @inode - file containing the logical cluster
5879  * @lclu - logical cluster of interest
5880  *
5881  * Returns 1 if any block in the logical cluster is mapped, signifying
5882  * that a physical cluster has been allocated for it.  Otherwise,
5883  * returns 0.  Can also return negative error codes.  Derived from
5884  * ext4_ext_map_blocks().
5885  */
5886 int ext4_clu_mapped(struct inode *inode, ext4_lblk_t lclu)
5887 {
5888         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5889         struct ext4_ext_path *path;
5890         int depth, mapped = 0, err = 0;
5891         struct ext4_extent *extent;
5892         ext4_lblk_t first_lblk, first_lclu, last_lclu;
5893
5894         /* search for the extent closest to the first block in the cluster */
5895         path = ext4_find_extent(inode, EXT4_C2B(sbi, lclu), NULL, 0);
5896         if (IS_ERR(path)) {
5897                 err = PTR_ERR(path);
5898                 path = NULL;
5899                 goto out;
5900         }
5901
5902         depth = ext_depth(inode);
5903
5904         /*
5905          * A consistent leaf must not be empty.  This situation is possible,
5906          * though, _during_ tree modification, and it's why an assert can't
5907          * be put in ext4_find_extent().
5908          */
5909         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
5910                 EXT4_ERROR_INODE(inode,
5911                     "bad extent address - lblock: %lu, depth: %d, pblock: %lld",
5912                                  (unsigned long) EXT4_C2B(sbi, lclu),
5913                                  depth, path[depth].p_block);
5914                 err = -EFSCORRUPTED;
5915                 goto out;
5916         }
5917
5918         extent = path[depth].p_ext;
5919
5920         /* can't be mapped if the extent tree is empty */
5921         if (extent == NULL)
5922                 goto out;
5923
5924         first_lblk = le32_to_cpu(extent->ee_block);
5925         first_lclu = EXT4_B2C(sbi, first_lblk);
5926
5927         /*
5928          * Three possible outcomes at this point - found extent spanning
5929          * the target cluster, to the left of the target cluster, or to the
5930          * right of the target cluster.  The first two cases are handled here.
5931          * The last case indicates the target cluster is not mapped.
5932          */
5933         if (lclu >= first_lclu) {
5934                 last_lclu = EXT4_B2C(sbi, first_lblk +
5935                                      ext4_ext_get_actual_len(extent) - 1);
5936                 if (lclu <= last_lclu) {
5937                         mapped = 1;
5938                 } else {
5939                         first_lblk = ext4_ext_next_allocated_block(path);
5940                         first_lclu = EXT4_B2C(sbi, first_lblk);
5941                         if (lclu == first_lclu)
5942                                 mapped = 1;
5943                 }
5944         }
5945
5946 out:
5947         ext4_ext_drop_refs(path);
5948         kfree(path);
5949
5950         return err ? err : mapped;
5951 }