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         err = __ext4_ext_check(function, line, inode,
 522                                ext_block_hdr(bh), depth, pblk);
 523         if (err)
 524                 goto errout;
 525         set_buffer_verified(bh);
 526         /*
 527          * If this is a leaf block, cache all of its entries
 528          */
 529         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 530                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 531                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 532                 ext4_lblk_t prev = 0;
 533                 int i;
 534
 535                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 536                         unsigned int status = EXTENT_STATUS_WRITTEN;
 537                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 538                         int len = ext4_ext_get_actual_len(ex);
 539
 540                         if (prev && (prev != lblk))
 541                                 ext4_es_cache_extent(inode, prev,
 542                                                      lblk - prev, ~0,
 543                                                      EXTENT_STATUS_HOLE);
 544
 545                         if (ext4_ext_is_unwritten(ex))
 546                                 status = EXTENT_STATUS_UNWRITTEN;
 547                         ext4_es_cache_extent(inode, lblk, len,
 548                                              ext4_ext_pblock(ex), status);
 549                         prev = lblk + len;
 550                 }
 551         }
 552         return bh;
 553 errout:
 554         put_bh(bh);
 555         return ERR_PTR(err);
 556
 557 }
 558
 559 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 560         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 561                                  (depth), (flags))
 562
 563 /*
 564  * This function is called to cache a file's extent information in the
 565  * extent status tree
 566  */
 567 int ext4_ext_precache(struct inode *inode)
 568 {
 569         struct ext4_inode_info *ei = EXT4_I(inode);
 570         struct ext4_ext_path *path = NULL;
 571         struct buffer_head *bh;
 572         int i = 0, depth, ret = 0;
 573
 574         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 575                 return 0;       /* not an extent-mapped inode */
 576
 577         down_read(&ei->i_data_sem);
 578         depth = ext_depth(inode);
 579
 580         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 581                        GFP_NOFS);
 582         if (path == NULL) {
 583                 up_read(&ei->i_data_sem);
 584                 return -ENOMEM;
 585         }
 586
 587         /* Don't cache anything if there are no external extent blocks */
 588         if (depth == 0)
 589                 goto out;
 590         path[0].p_hdr = ext_inode_hdr(inode);
 591         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 592         if (ret)
 593                 goto out;
 594         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 595         while (i >= 0) {
 596                 /*
 597                  * If this is a leaf block or we've reached the end of
 598                  * the index block, go up
 599                  */
 600                 if ((i == depth) ||
 601                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 602                         brelse(path[i].p_bh);
 603                         path[i].p_bh = NULL;
 604                         i--;
 605                         continue;
 606                 }
 607                 bh = read_extent_tree_block(inode,
 608                                             ext4_idx_pblock(path[i].p_idx++),
 609                                             depth - i - 1,
 610                                             EXT4_EX_FORCE_CACHE);
 611                 if (IS_ERR(bh)) {
 612                         ret = PTR_ERR(bh);
 613                         break;
 614                 }
 615                 i++;
 616                 path[i].p_bh = bh;
 617                 path[i].p_hdr = ext_block_hdr(bh);
 618                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 619         }
 620         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 621 out:
 622         up_read(&ei->i_data_sem);
 623         ext4_ext_drop_refs(path);
 624         kfree(path);
 625         return ret;
 626 }
 627
 628 #ifdef EXT_DEBUG
 629 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 630 {
 631         int k, l = path->p_depth;
 632
 633         ext_debug("path:");
 634         for (k = 0; k <= l; k++, path++) {
 635                 if (path->p_idx) {
 636                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 637                             ext4_idx_pblock(path->p_idx));
 638                 } else if (path->p_ext) {
 639                         ext_debug("  %d:[%d]%d:%llu ",
 640                                   le32_to_cpu(path->p_ext->ee_block),
 641                                   ext4_ext_is_unwritten(path->p_ext),
 642                                   ext4_ext_get_actual_len(path->p_ext),
 643                                   ext4_ext_pblock(path->p_ext));
 644                 } else
 645                         ext_debug("  []");
 646         }
 647         ext_debug("\n");
 648 }
 649
 650 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 651 {
 652         int depth = ext_depth(inode);
 653         struct ext4_extent_header *eh;
 654         struct ext4_extent *ex;
 655         int i;
 656
 657         if (!path)
 658                 return;
 659
 660         eh = path[depth].p_hdr;
 661         ex = EXT_FIRST_EXTENT(eh);
 662
 663         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 664
 665         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 666                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 667                           ext4_ext_is_unwritten(ex),
 668                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 669         }
 670         ext_debug("\n");
 671 }
 672
 673 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 674                         ext4_fsblk_t newblock, int level)
 675 {
 676         int depth = ext_depth(inode);
 677         struct ext4_extent *ex;
 678
 679         if (depth != level) {
 680                 struct ext4_extent_idx *idx;
 681                 idx = path[level].p_idx;
 682                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 683                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 684                                         le32_to_cpu(idx->ei_block),
 685                                         ext4_idx_pblock(idx),
 686                                         newblock);
 687                         idx++;
 688                 }
 689
 690                 return;
 691         }
 692
 693         ex = path[depth].p_ext;
 694         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 695                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 696                                 le32_to_cpu(ex->ee_block),
 697                                 ext4_ext_pblock(ex),
 698                                 ext4_ext_is_unwritten(ex),
 699                                 ext4_ext_get_actual_len(ex),
 700                                 newblock);
 701                 ex++;
 702         }
 703 }
 704
 705 #else
 706 #define ext4_ext_show_path(inode, path)
 707 #define ext4_ext_show_leaf(inode, path)
 708 #define ext4_ext_show_move(inode, path, newblock, level)
 709 #endif
 710
 711 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 712 {
 713         int depth, i;
 714
 715         if (!path)
 716                 return;
 717         depth = path->p_depth;
 718         for (i = 0; i <= depth; i++, path++)
 719                 if (path->p_bh) {
 720                         brelse(path->p_bh);
 721                         path->p_bh = NULL;
 722                 }
 723 }
 724
 725 /*
 726  * ext4_ext_binsearch_idx:
 727  * binary search for the closest index of the given block
 728  * the header must be checked before calling this
 729  */
 730 static void
 731 ext4_ext_binsearch_idx(struct inode *inode,
 732                         struct ext4_ext_path *path, ext4_lblk_t block)
 733 {
 734         struct ext4_extent_header *eh = path->p_hdr;
 735         struct ext4_extent_idx *r, *l, *m;
 736
 737
 738         ext_debug("binsearch for %u(idx):  ", block);
 739
 740         l = EXT_FIRST_INDEX(eh) + 1;
 741         r = EXT_LAST_INDEX(eh);
 742         while (l <= r) {
 743                 m = l + (r - l) / 2;
 744                 if (block < le32_to_cpu(m->ei_block))
 745                         r = m - 1;
 746                 else
 747                         l = m + 1;
 748                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 749                                 m, le32_to_cpu(m->ei_block),
 750                                 r, le32_to_cpu(r->ei_block));
 751         }
 752
 753         path->p_idx = l - 1;
 754         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 755                   ext4_idx_pblock(path->p_idx));
 756
 757 #ifdef CHECK_BINSEARCH
 758         {
 759                 struct ext4_extent_idx *chix, *ix;
 760                 int k;
 761
 762                 chix = ix = EXT_FIRST_INDEX(eh);
 763                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 764                   if (k != 0 &&
 765                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 766                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 767                                        "first=0x%p\n", k,
 768                                        ix, EXT_FIRST_INDEX(eh));
 769                                 printk(KERN_DEBUG "%u <= %u\n",
 770                                        le32_to_cpu(ix->ei_block),
 771                                        le32_to_cpu(ix[-1].ei_block));
 772                         }
 773                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 774                                            <= le32_to_cpu(ix[-1].ei_block));
 775                         if (block < le32_to_cpu(ix->ei_block))
 776                                 break;
 777                         chix = ix;
 778                 }
 779                 BUG_ON(chix != path->p_idx);
 780         }
 781 #endif
 782
 783 }
 784
 785 /*
 786  * ext4_ext_binsearch:
 787  * binary search for closest extent of the given block
 788  * the header must be checked before calling this
 789  */
 790 static void
 791 ext4_ext_binsearch(struct inode *inode,
 792                 struct ext4_ext_path *path, ext4_lblk_t block)
 793 {
 794         struct ext4_extent_header *eh = path->p_hdr;
 795         struct ext4_extent *r, *l, *m;
 796
 797         if (eh->eh_entries == 0) {
 798                 /*
 799                  * this leaf is empty:
 800                  * we get such a leaf in split/add case
 801                  */
 802                 return;
 803         }
 804
 805         ext_debug("binsearch for %u:  ", block);
 806
 807         l = EXT_FIRST_EXTENT(eh) + 1;
 808         r = EXT_LAST_EXTENT(eh);
 809
 810         while (l <= r) {
 811                 m = l + (r - l) / 2;
 812                 if (block < le32_to_cpu(m->ee_block))
 813                         r = m - 1;
 814                 else
 815                         l = m + 1;
 816                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 817                                 m, le32_to_cpu(m->ee_block),
 818                                 r, le32_to_cpu(r->ee_block));
 819         }
 820
 821         path->p_ext = l - 1;
 822         ext_debug("  -> %d:%llu:[%d]%d ",
 823                         le32_to_cpu(path->p_ext->ee_block),
 824                         ext4_ext_pblock(path->p_ext),
 825                         ext4_ext_is_unwritten(path->p_ext),
 826                         ext4_ext_get_actual_len(path->p_ext));
 827
 828 #ifdef CHECK_BINSEARCH
 829         {
 830                 struct ext4_extent *chex, *ex;
 831                 int k;
 832
 833                 chex = ex = EXT_FIRST_EXTENT(eh);
 834                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 835                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 836                                           <= le32_to_cpu(ex[-1].ee_block));
 837                         if (block < le32_to_cpu(ex->ee_block))
 838                                 break;
 839                         chex = ex;
 840                 }
 841                 BUG_ON(chex != path->p_ext);
 842         }
 843 #endif
 844
 845 }
 846
 847 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 848 {
 849         struct ext4_extent_header *eh;
 850
 851         eh = ext_inode_hdr(inode);
 852         eh->eh_depth = 0;
 853         eh->eh_entries = 0;
 854         eh->eh_magic = EXT4_EXT_MAGIC;
 855         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 856         ext4_mark_inode_dirty(handle, inode);
 857         return 0;
 858 }
 859
 860 struct ext4_ext_path *
 861 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 862                  struct ext4_ext_path **orig_path, int flags)
 863 {
 864         struct ext4_extent_header *eh;
 865         struct buffer_head *bh;
 866         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 867         short int depth, i, ppos = 0;
 868         int ret;
 869
 870         eh = ext_inode_hdr(inode);
 871         depth = ext_depth(inode);
 872
 873         if (path) {
 874                 ext4_ext_drop_refs(path);
 875                 if (depth > path[0].p_maxdepth) {
 876                         kfree(path);
 877                         *orig_path = path = NULL;
 878                 }
 879         }
 880         if (!path) {
 881                 /* account possible depth increase */
 882                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 883                                 GFP_NOFS);
 884                 if (unlikely(!path))
 885                         return ERR_PTR(-ENOMEM);
 886                 path[0].p_maxdepth = depth + 1;
 887         }
 888         path[0].p_hdr = eh;
 889         path[0].p_bh = NULL;
 890
 891         i = depth;
 892         /* walk through the tree */
 893         while (i) {
 894                 ext_debug("depth %d: num %d, max %d\n",
 895                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 896
 897                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 898                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 899                 path[ppos].p_depth = i;
 900                 path[ppos].p_ext = NULL;
 901
 902                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 903                                             flags);
 904                 if (IS_ERR(bh)) {
 905                         ret = PTR_ERR(bh);
 906                         goto err;
 907                 }
 908
 909                 eh = ext_block_hdr(bh);
 910                 ppos++;
 911                 path[ppos].p_bh = bh;
 912                 path[ppos].p_hdr = eh;
 913         }
 914
 915         path[ppos].p_depth = i;
 916         path[ppos].p_ext = NULL;
 917         path[ppos].p_idx = NULL;
 918
 919         /* find extent */
 920         ext4_ext_binsearch(inode, path + ppos, block);
 921         /* if not an empty leaf */
 922         if (path[ppos].p_ext)
 923                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 924
 925         ext4_ext_show_path(inode, path);
 926
 927         return path;
 928
 929 err:
 930         ext4_ext_drop_refs(path);
 931         kfree(path);
 932         if (orig_path)
 933                 *orig_path = NULL;
 934         return ERR_PTR(ret);
 935 }
 936
 937 /*
 938  * ext4_ext_insert_index:
 939  * insert new index [@logical;@ptr] into the block at @curp;
 940  * check where to insert: before @curp or after @curp
 941  */
 942 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 943                                  struct ext4_ext_path *curp,
 944                                  int logical, ext4_fsblk_t ptr)
 945 {
 946         struct ext4_extent_idx *ix;
 947         int len, err;
 948
 949         err = ext4_ext_get_access(handle, inode, curp);
 950         if (err)
 951                 return err;
 952
 953         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 954                 EXT4_ERROR_INODE(inode,
 955                                  "logical %d == ei_block %d!",
 956                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 957                 return -EFSCORRUPTED;
 958         }
 959
 960         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 961                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 962                 EXT4_ERROR_INODE(inode,
 963                                  "eh_entries %d >= eh_max %d!",
 964                                  le16_to_cpu(curp->p_hdr->eh_entries),
 965                                  le16_to_cpu(curp->p_hdr->eh_max));
 966                 return -EFSCORRUPTED;
 967         }
 968
 969         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 970                 /* insert after */
 971                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 972                 ix = curp->p_idx + 1;
 973         } else {
 974                 /* insert before */
 975                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 976                 ix = curp->p_idx;
 977         }
 978
 979         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 980         BUG_ON(len < 0);
 981         if (len > 0) {
 982                 ext_debug("insert new index %d: "
 983                                 "move %d indices from 0x%p to 0x%p\n",
 984                                 logical, len, ix, ix + 1);
 985                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 986         }
 987
 988         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 989                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 990                 return -EFSCORRUPTED;
 991         }
 992
 993         ix->ei_block = cpu_to_le32(logical);
 994         ext4_idx_store_pblock(ix, ptr);
 995         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 996
 997         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 998                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 999                 return -EFSCORRUPTED;
1000         }
1001
1002         err = ext4_ext_dirty(handle, inode, curp);
1003         ext4_std_error(inode->i_sb, err);
1004
1005         return err;
1006 }
1007
1008 /*
1009  * ext4_ext_split:
1010  * inserts new subtree into the path, using free index entry
1011  * at depth @at:
1012  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1013  * - makes decision where to split
1014  * - moves remaining extents and index entries (right to the split point)
1015  *   into the newly allocated blocks
1016  * - initializes subtree
1017  */
1018 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1019                           unsigned int flags,
1020                           struct ext4_ext_path *path,
1021                           struct ext4_extent *newext, int at)
1022 {
1023         struct buffer_head *bh = NULL;
1024         int depth = ext_depth(inode);
1025         struct ext4_extent_header *neh;
1026         struct ext4_extent_idx *fidx;
1027         int i = at, k, m, a;
1028         ext4_fsblk_t newblock, oldblock;
1029         __le32 border;
1030         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1031         int err = 0;
1032
1033         /* make decision: where to split? */
1034         /* FIXME: now decision is simplest: at current extent */
1035
1036         /* if current leaf will be split, then we should use
1037          * border from split point */
1038         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1039                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1040                 return -EFSCORRUPTED;
1041         }
1042         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1043                 border = path[depth].p_ext[1].ee_block;
1044                 ext_debug("leaf will be split."
1045                                 " next leaf starts at %d\n",
1046                                   le32_to_cpu(border));
1047         } else {
1048                 border = newext->ee_block;
1049                 ext_debug("leaf will be added."
1050                                 " next leaf starts at %d\n",
1051                                 le32_to_cpu(border));
1052         }
1053
1054         /*
1055          * If error occurs, then we break processing
1056          * and mark filesystem read-only. index won't
1057          * be inserted and tree will be in consistent
1058          * state. Next mount will repair buffers too.
1059          */
1060
1061         /*
1062          * Get array to track all allocated blocks.
1063          * We need this to handle errors and free blocks
1064          * upon them.
1065          */
1066         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1067         if (!ablocks)
1068                 return -ENOMEM;
1069
1070         /* allocate all needed blocks */
1071         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1072         for (a = 0; a < depth - at; a++) {
1073                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1074                                                    newext, &err, flags);
1075                 if (newblock == 0)
1076                         goto cleanup;
1077                 ablocks[a] = newblock;
1078         }
1079
1080         /* initialize new leaf */
1081         newblock = ablocks[--a];
1082         if (unlikely(newblock == 0)) {
1083                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1084                 err = -EFSCORRUPTED;
1085                 goto cleanup;
1086         }
1087         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1088         if (unlikely(!bh)) {
1089                 err = -ENOMEM;
1090                 goto cleanup;
1091         }
1092         lock_buffer(bh);
1093
1094         err = ext4_journal_get_create_access(handle, bh);
1095         if (err)
1096                 goto cleanup;
1097
1098         neh = ext_block_hdr(bh);
1099         neh->eh_entries = 0;
1100         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1101         neh->eh_magic = EXT4_EXT_MAGIC;
1102         neh->eh_depth = 0;
1103
1104         /* move remainder of path[depth] to the new leaf */
1105         if (unlikely(path[depth].p_hdr->eh_entries !=
1106                      path[depth].p_hdr->eh_max)) {
1107                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1108                                  path[depth].p_hdr->eh_entries,
1109                                  path[depth].p_hdr->eh_max);
1110                 err = -EFSCORRUPTED;
1111                 goto cleanup;
1112         }
1113         /* start copy from next extent */
1114         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1115         ext4_ext_show_move(inode, path, newblock, depth);
1116         if (m) {
1117                 struct ext4_extent *ex;
1118                 ex = EXT_FIRST_EXTENT(neh);
1119                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1120                 le16_add_cpu(&neh->eh_entries, m);
1121         }
1122
1123         ext4_extent_block_csum_set(inode, neh);
1124         set_buffer_uptodate(bh);
1125         unlock_buffer(bh);
1126
1127         err = ext4_handle_dirty_metadata(handle, inode, bh);
1128         if (err)
1129                 goto cleanup;
1130         brelse(bh);
1131         bh = NULL;
1132
1133         /* correct old leaf */
1134         if (m) {
1135                 err = ext4_ext_get_access(handle, inode, path + depth);
1136                 if (err)
1137                         goto cleanup;
1138                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1139                 err = ext4_ext_dirty(handle, inode, path + depth);
1140                 if (err)
1141                         goto cleanup;
1142
1143         }
1144
1145         /* create intermediate indexes */
1146         k = depth - at - 1;
1147         if (unlikely(k < 0)) {
1148                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1149                 err = -EFSCORRUPTED;
1150                 goto cleanup;
1151         }
1152         if (k)
1153                 ext_debug("create %d intermediate indices\n", k);
1154         /* insert new index into current index block */
1155         /* current depth stored in i var */
1156         i = depth - 1;
1157         while (k--) {
1158                 oldblock = newblock;
1159                 newblock = ablocks[--a];
1160                 bh = sb_getblk(inode->i_sb, newblock);
1161                 if (unlikely(!bh)) {
1162                         err = -ENOMEM;
1163                         goto cleanup;
1164                 }
1165                 lock_buffer(bh);
1166
1167                 err = ext4_journal_get_create_access(handle, bh);
1168                 if (err)
1169                         goto cleanup;
1170
1171                 neh = ext_block_hdr(bh);
1172                 neh->eh_entries = cpu_to_le16(1);
1173                 neh->eh_magic = EXT4_EXT_MAGIC;
1174                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1175                 neh->eh_depth = cpu_to_le16(depth - i);
1176                 fidx = EXT_FIRST_INDEX(neh);
1177                 fidx->ei_block = border;
1178                 ext4_idx_store_pblock(fidx, oldblock);
1179
1180                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1181                                 i, newblock, le32_to_cpu(border), oldblock);
1182
1183                 /* move remainder of path[i] to the new index block */
1184                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1185                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1186                         EXT4_ERROR_INODE(inode,
1187                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1188                                          le32_to_cpu(path[i].p_ext->ee_block));
1189                         err = -EFSCORRUPTED;
1190                         goto cleanup;
1191                 }
1192                 /* start copy indexes */
1193                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1194                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1195                                 EXT_MAX_INDEX(path[i].p_hdr));
1196                 ext4_ext_show_move(inode, path, newblock, i);
1197                 if (m) {
1198                         memmove(++fidx, path[i].p_idx,
1199                                 sizeof(struct ext4_extent_idx) * m);
1200                         le16_add_cpu(&neh->eh_entries, m);
1201                 }
1202                 ext4_extent_block_csum_set(inode, neh);
1203                 set_buffer_uptodate(bh);
1204                 unlock_buffer(bh);
1205
1206                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1207                 if (err)
1208                         goto cleanup;
1209                 brelse(bh);
1210                 bh = NULL;
1211
1212                 /* correct old index */
1213                 if (m) {
1214                         err = ext4_ext_get_access(handle, inode, path + i);
1215                         if (err)
1216                                 goto cleanup;
1217                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1218                         err = ext4_ext_dirty(handle, inode, path + i);
1219                         if (err)
1220                                 goto cleanup;
1221                 }
1222
1223                 i--;
1224         }
1225
1226         /* insert new index */
1227         err = ext4_ext_insert_index(handle, inode, path + at,
1228                                     le32_to_cpu(border), newblock);
1229
1230 cleanup:
1231         if (bh) {
1232                 if (buffer_locked(bh))
1233                         unlock_buffer(bh);
1234                 brelse(bh);
1235         }
1236
1237         if (err) {
1238                 /* free all allocated blocks in error case */
1239                 for (i = 0; i < depth; i++) {
1240                         if (!ablocks[i])
1241                                 continue;
1242                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1243                                          EXT4_FREE_BLOCKS_METADATA);
1244                 }
1245         }
1246         kfree(ablocks);
1247
1248         return err;
1249 }
1250
1251 /*
1252  * ext4_ext_grow_indepth:
1253  * implements tree growing procedure:
1254  * - allocates new block
1255  * - moves top-level data (index block or leaf) into the new block
1256  * - initializes new top-level, creating index that points to the
1257  *   just created block
1258  */
1259 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1260                                  unsigned int flags)
1261 {
1262         struct ext4_extent_header *neh;
1263         struct buffer_head *bh;
1264         ext4_fsblk_t newblock, goal = 0;
1265         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1266         int err = 0;
1267
1268         /* Try to prepend new index to old one */
1269         if (ext_depth(inode))
1270                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1271         if (goal > le32_to_cpu(es->s_first_data_block)) {
1272                 flags |= EXT4_MB_HINT_TRY_GOAL;
1273                 goal--;
1274         } else
1275                 goal = ext4_inode_to_goal_block(inode);
1276         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1277                                         NULL, &err);
1278         if (newblock == 0)
1279                 return err;
1280
1281         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1282         if (unlikely(!bh))
1283                 return -ENOMEM;
1284         lock_buffer(bh);
1285
1286         err = ext4_journal_get_create_access(handle, bh);
1287         if (err) {
1288                 unlock_buffer(bh);
1289                 goto out;
1290         }
1291
1292         /* move top-level index/leaf into new block */
1293         memmove(bh->b_data, EXT4_I(inode)->i_data,
1294                 sizeof(EXT4_I(inode)->i_data));
1295
1296         /* set size of new block */
1297         neh = ext_block_hdr(bh);
1298         /* old root could have indexes or leaves
1299          * so calculate e_max right way */
1300         if (ext_depth(inode))
1301                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1302         else
1303                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1304         neh->eh_magic = EXT4_EXT_MAGIC;
1305         ext4_extent_block_csum_set(inode, neh);
1306         set_buffer_uptodate(bh);
1307         unlock_buffer(bh);
1308
1309         err = ext4_handle_dirty_metadata(handle, inode, bh);
1310         if (err)
1311                 goto out;
1312
1313         /* Update top-level index: num,max,pointer */
1314         neh = ext_inode_hdr(inode);
1315         neh->eh_entries = cpu_to_le16(1);
1316         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1317         if (neh->eh_depth == 0) {
1318                 /* Root extent block becomes index block */
1319                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1320                 EXT_FIRST_INDEX(neh)->ei_block =
1321                         EXT_FIRST_EXTENT(neh)->ee_block;
1322         }
1323         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1324                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1325                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1326                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1327
1328         le16_add_cpu(&neh->eh_depth, 1);
1329         ext4_mark_inode_dirty(handle, inode);
1330 out:
1331         brelse(bh);
1332
1333         return err;
1334 }
1335
1336 /*
1337  * ext4_ext_create_new_leaf:
1338  * finds empty index and adds new leaf.
1339  * if no free index is found, then it requests in-depth growing.
1340  */
1341 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1342                                     unsigned int mb_flags,
1343                                     unsigned int gb_flags,
1344                                     struct ext4_ext_path **ppath,
1345                                     struct ext4_extent *newext)
1346 {
1347         struct ext4_ext_path *path = *ppath;
1348         struct ext4_ext_path *curp;
1349         int depth, i, err = 0;
1350
1351 repeat:
1352         i = depth = ext_depth(inode);
1353
1354         /* walk up to the tree and look for free index entry */
1355         curp = path + depth;
1356         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1357                 i--;
1358                 curp--;
1359         }
1360
1361         /* we use already allocated block for index block,
1362          * so subsequent data blocks should be contiguous */
1363         if (EXT_HAS_FREE_INDEX(curp)) {
1364                 /* if we found index with free entry, then use that
1365                  * entry: create all needed subtree and add new leaf */
1366                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1367                 if (err)
1368                         goto out;
1369
1370                 /* refill path */
1371                 path = ext4_find_extent(inode,
1372                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1373                                     ppath, gb_flags);
1374                 if (IS_ERR(path))
1375                         err = PTR_ERR(path);
1376         } else {
1377                 /* tree is full, time to grow in depth */
1378                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1379                 if (err)
1380                         goto out;
1381
1382                 /* refill path */
1383                 path = ext4_find_extent(inode,
1384                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1385                                     ppath, gb_flags);
1386                 if (IS_ERR(path)) {
1387                         err = PTR_ERR(path);
1388                         goto out;
1389                 }
1390
1391                 /*
1392                  * only first (depth 0 -> 1) produces free space;
1393                  * in all other cases we have to split the grown tree
1394                  */
1395                 depth = ext_depth(inode);
1396                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1397                         /* now we need to split */
1398                         goto repeat;
1399                 }
1400         }
1401
1402 out:
1403         return err;
1404 }
1405
1406 /*
1407  * search the closest allocated block to the left for *logical
1408  * and returns it at @logical + it's physical address at @phys
1409  * if *logical is the smallest allocated block, the function
1410  * returns 0 at @phys
1411  * return value contains 0 (success) or error code
1412  */
1413 static int ext4_ext_search_left(struct inode *inode,
1414                                 struct ext4_ext_path *path,
1415                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1416 {
1417         struct ext4_extent_idx *ix;
1418         struct ext4_extent *ex;
1419         int depth, ee_len;
1420
1421         if (unlikely(path == NULL)) {
1422                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1423                 return -EFSCORRUPTED;
1424         }
1425         depth = path->p_depth;
1426         *phys = 0;
1427
1428         if (depth == 0 && path->p_ext == NULL)
1429                 return 0;
1430
1431         /* usually extent in the path covers blocks smaller
1432          * then *logical, but it can be that extent is the
1433          * first one in the file */
1434
1435         ex = path[depth].p_ext;
1436         ee_len = ext4_ext_get_actual_len(ex);
1437         if (*logical < le32_to_cpu(ex->ee_block)) {
1438                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1439                         EXT4_ERROR_INODE(inode,
1440                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1441                                          *logical, le32_to_cpu(ex->ee_block));
1442                         return -EFSCORRUPTED;
1443                 }
1444                 while (--depth >= 0) {
1445                         ix = path[depth].p_idx;
1446                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1447                                 EXT4_ERROR_INODE(inode,
1448                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1449                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1450                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1451                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1452                                   depth);
1453                                 return -EFSCORRUPTED;
1454                         }
1455                 }
1456                 return 0;
1457         }
1458
1459         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1460                 EXT4_ERROR_INODE(inode,
1461                                  "logical %d < ee_block %d + ee_len %d!",
1462                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1463                 return -EFSCORRUPTED;
1464         }
1465
1466         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1467         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1468         return 0;
1469 }
1470
1471 /*
1472  * search the closest allocated block to the right for *logical
1473  * and returns it at @logical + it's physical address at @phys
1474  * if *logical is the largest allocated block, the function
1475  * returns 0 at @phys
1476  * return value contains 0 (success) or error code
1477  */
1478 static int ext4_ext_search_right(struct inode *inode,
1479                                  struct ext4_ext_path *path,
1480                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1481                                  struct ext4_extent **ret_ex)
1482 {
1483         struct buffer_head *bh = NULL;
1484         struct ext4_extent_header *eh;
1485         struct ext4_extent_idx *ix;
1486         struct ext4_extent *ex;
1487         ext4_fsblk_t block;
1488         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1489         int ee_len;
1490
1491         if (unlikely(path == NULL)) {
1492                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1493                 return -EFSCORRUPTED;
1494         }
1495         depth = path->p_depth;
1496         *phys = 0;
1497
1498         if (depth == 0 && path->p_ext == NULL)
1499                 return 0;
1500
1501         /* usually extent in the path covers blocks smaller
1502          * then *logical, but it can be that extent is the
1503          * first one in the file */
1504
1505         ex = path[depth].p_ext;
1506         ee_len = ext4_ext_get_actual_len(ex);
1507         if (*logical < le32_to_cpu(ex->ee_block)) {
1508                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1509                         EXT4_ERROR_INODE(inode,
1510                                          "first_extent(path[%d].p_hdr) != ex",
1511                                          depth);
1512                         return -EFSCORRUPTED;
1513                 }
1514                 while (--depth >= 0) {
1515                         ix = path[depth].p_idx;
1516                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1517                                 EXT4_ERROR_INODE(inode,
1518                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1519                                                  *logical);
1520                                 return -EFSCORRUPTED;
1521                         }
1522                 }
1523                 goto found_extent;
1524         }
1525
1526         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1527                 EXT4_ERROR_INODE(inode,
1528                                  "logical %d < ee_block %d + ee_len %d!",
1529                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1530                 return -EFSCORRUPTED;
1531         }
1532
1533         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1534                 /* next allocated block in this leaf */
1535                 ex++;
1536                 goto found_extent;
1537         }
1538
1539         /* go up and search for index to the right */
1540         while (--depth >= 0) {
1541                 ix = path[depth].p_idx;
1542                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1543                         goto got_index;
1544         }
1545
1546         /* we've gone up to the root and found no index to the right */
1547         return 0;
1548
1549 got_index:
1550         /* we've found index to the right, let's
1551          * follow it and find the closest allocated
1552          * block to the right */
1553         ix++;
1554         block = ext4_idx_pblock(ix);
1555         while (++depth < path->p_depth) {
1556                 /* subtract from p_depth to get proper eh_depth */
1557                 bh = read_extent_tree_block(inode, block,
1558                                             path->p_depth - depth, 0);
1559                 if (IS_ERR(bh))
1560                         return PTR_ERR(bh);
1561                 eh = ext_block_hdr(bh);
1562                 ix = EXT_FIRST_INDEX(eh);
1563                 block = ext4_idx_pblock(ix);
1564                 put_bh(bh);
1565         }
1566
1567         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1568         if (IS_ERR(bh))
1569                 return PTR_ERR(bh);
1570         eh = ext_block_hdr(bh);
1571         ex = EXT_FIRST_EXTENT(eh);
1572 found_extent:
1573         *logical = le32_to_cpu(ex->ee_block);
1574         *phys = ext4_ext_pblock(ex);
1575         *ret_ex = ex;
1576         if (bh)
1577                 put_bh(bh);
1578         return 0;
1579 }
1580
1581 /*
1582  * ext4_ext_next_allocated_block:
1583  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1584  * NOTE: it considers block number from index entry as
1585  * allocated block. Thus, index entries have to be consistent
1586  * with leaves.
1587  */
1588 ext4_lblk_t
1589 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1590 {
1591         int depth;
1592
1593         BUG_ON(path == NULL);
1594         depth = path->p_depth;
1595
1596         if (depth == 0 && path->p_ext == NULL)
1597                 return EXT_MAX_BLOCKS;
1598
1599         while (depth >= 0) {
1600                 if (depth == path->p_depth) {
1601                         /* leaf */
1602                         if (path[depth].p_ext &&
1603                                 path[depth].p_ext !=
1604                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1605                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1606                 } else {
1607                         /* index */
1608                         if (path[depth].p_idx !=
1609                                         EXT_LAST_INDEX(path[depth].p_hdr))
1610                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1611                 }
1612                 depth--;
1613         }
1614
1615         return EXT_MAX_BLOCKS;
1616 }
1617
1618 /*
1619  * ext4_ext_next_leaf_block:
1620  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1621  */
1622 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1623 {
1624         int depth;
1625
1626         BUG_ON(path == NULL);
1627         depth = path->p_depth;
1628
1629         /* zero-tree has no leaf blocks at all */
1630         if (depth == 0)
1631                 return EXT_MAX_BLOCKS;
1632
1633         /* go to index block */
1634         depth--;
1635
1636         while (depth >= 0) {
1637                 if (path[depth].p_idx !=
1638                                 EXT_LAST_INDEX(path[depth].p_hdr))
1639                         return (ext4_lblk_t)
1640                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1641                 depth--;
1642         }
1643
1644         return EXT_MAX_BLOCKS;
1645 }
1646
1647 /*
1648  * ext4_ext_correct_indexes:
1649  * if leaf gets modified and modified extent is first in the leaf,
1650  * then we have to correct all indexes above.
1651  * TODO: do we need to correct tree in all cases?
1652  */
1653 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1654                                 struct ext4_ext_path *path)
1655 {
1656         struct ext4_extent_header *eh;
1657         int depth = ext_depth(inode);
1658         struct ext4_extent *ex;
1659         __le32 border;
1660         int k, err = 0;
1661
1662         eh = path[depth].p_hdr;
1663         ex = path[depth].p_ext;
1664
1665         if (unlikely(ex == NULL || eh == NULL)) {
1666                 EXT4_ERROR_INODE(inode,
1667                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1668                 return -EFSCORRUPTED;
1669         }
1670
1671         if (depth == 0) {
1672                 /* there is no tree at all */
1673                 return 0;
1674         }
1675
1676         if (ex != EXT_FIRST_EXTENT(eh)) {
1677                 /* we correct tree if first leaf got modified only */
1678                 return 0;
1679         }
1680
1681         /*
1682          * TODO: we need correction if border is smaller than current one
1683          */
1684         k = depth - 1;
1685         border = path[depth].p_ext->ee_block;
1686         err = ext4_ext_get_access(handle, inode, path + k);
1687         if (err)
1688                 return err;
1689         path[k].p_idx->ei_block = border;
1690         err = ext4_ext_dirty(handle, inode, path + k);
1691         if (err)
1692                 return err;
1693
1694         while (k--) {
1695                 /* change all left-side indexes */
1696                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1697                         break;
1698                 err = ext4_ext_get_access(handle, inode, path + k);
1699                 if (err)
1700                         break;
1701                 path[k].p_idx->ei_block = border;
1702                 err = ext4_ext_dirty(handle, inode, path + k);
1703                 if (err)
1704                         break;
1705         }
1706
1707         return err;
1708 }
1709
1710 int
1711 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1712                                 struct ext4_extent *ex2)
1713 {
1714         unsigned short ext1_ee_len, ext2_ee_len;
1715
1716         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1717                 return 0;
1718
1719         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1720         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1721
1722         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1723                         le32_to_cpu(ex2->ee_block))
1724                 return 0;
1725
1726         /*
1727          * To allow future support for preallocated extents to be added
1728          * as an RO_COMPAT feature, refuse to merge to extents if
1729          * this can result in the top bit of ee_len being set.
1730          */
1731         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1732                 return 0;
1733         /*
1734          * The check for IO to unwritten extent is somewhat racy as we
1735          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1736          * dropping i_data_sem. But reserved blocks should save us in that
1737          * case.
1738          */
1739         if (ext4_ext_is_unwritten(ex1) &&
1740             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1741              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1742              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1743                 return 0;
1744 #ifdef AGGRESSIVE_TEST
1745         if (ext1_ee_len >= 4)
1746                 return 0;
1747 #endif
1748
1749         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1750                 return 1;
1751         return 0;
1752 }
1753
1754 /*
1755  * This function tries to merge the "ex" extent to the next extent in the tree.
1756  * It always tries to merge towards right. If you want to merge towards
1757  * left, pass "ex - 1" as argument instead of "ex".
1758  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1759  * 1 if they got merged.
1760  */
1761 static int ext4_ext_try_to_merge_right(struct inode *inode,
1762                                  struct ext4_ext_path *path,
1763                                  struct ext4_extent *ex)
1764 {
1765         struct ext4_extent_header *eh;
1766         unsigned int depth, len;
1767         int merge_done = 0, unwritten;
1768
1769         depth = ext_depth(inode);
1770         BUG_ON(path[depth].p_hdr == NULL);
1771         eh = path[depth].p_hdr;
1772
1773         while (ex < EXT_LAST_EXTENT(eh)) {
1774                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1775                         break;
1776                 /* merge with next extent! */
1777                 unwritten = ext4_ext_is_unwritten(ex);
1778                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1779                                 + ext4_ext_get_actual_len(ex + 1));
1780                 if (unwritten)
1781                         ext4_ext_mark_unwritten(ex);
1782
1783                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1784                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1785                                 * sizeof(struct ext4_extent);
1786                         memmove(ex + 1, ex + 2, len);
1787                 }
1788                 le16_add_cpu(&eh->eh_entries, -1);
1789                 merge_done = 1;
1790                 WARN_ON(eh->eh_entries == 0);
1791                 if (!eh->eh_entries)
1792                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1793         }
1794
1795         return merge_done;
1796 }
1797
1798 /*
1799  * This function does a very simple check to see if we can collapse
1800  * an extent tree with a single extent tree leaf block into the inode.
1801  */
1802 static void ext4_ext_try_to_merge_up(handle_t *handle,
1803                                      struct inode *inode,
1804                                      struct ext4_ext_path *path)
1805 {
1806         size_t s;
1807         unsigned max_root = ext4_ext_space_root(inode, 0);
1808         ext4_fsblk_t blk;
1809
1810         if ((path[0].p_depth != 1) ||
1811             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1812             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1813                 return;
1814
1815         /*
1816          * We need to modify the block allocation bitmap and the block
1817          * group descriptor to release the extent tree block.  If we
1818          * can't get the journal credits, give up.
1819          */
1820         if (ext4_journal_extend(handle, 2))
1821                 return;
1822
1823         /*
1824          * Copy the extent data up to the inode
1825          */
1826         blk = ext4_idx_pblock(path[0].p_idx);
1827         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1828                 sizeof(struct ext4_extent_idx);
1829         s += sizeof(struct ext4_extent_header);
1830
1831         path[1].p_maxdepth = path[0].p_maxdepth;
1832         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1833         path[0].p_depth = 0;
1834         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1835                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1836         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1837
1838         brelse(path[1].p_bh);
1839         ext4_free_blocks(handle, inode, NULL, blk, 1,
1840                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1841 }
1842
1843 /*
1844  * This function tries to merge the @ex extent to neighbours in the tree.
1845  * return 1 if merge left else 0.
1846  */
1847 static void ext4_ext_try_to_merge(handle_t *handle,
1848                                   struct inode *inode,
1849                                   struct ext4_ext_path *path,
1850                                   struct ext4_extent *ex) {
1851         struct ext4_extent_header *eh;
1852         unsigned int depth;
1853         int merge_done = 0;
1854
1855         depth = ext_depth(inode);
1856         BUG_ON(path[depth].p_hdr == NULL);
1857         eh = path[depth].p_hdr;
1858
1859         if (ex > EXT_FIRST_EXTENT(eh))
1860                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1861
1862         if (!merge_done)
1863                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1864
1865         ext4_ext_try_to_merge_up(handle, inode, path);
1866 }
1867
1868 /*
1869  * check if a portion of the "newext" extent overlaps with an
1870  * existing extent.
1871  *
1872  * If there is an overlap discovered, it updates the length of the newext
1873  * such that there will be no overlap, and then returns 1.
1874  * If there is no overlap found, it returns 0.
1875  */
1876 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1877                                            struct inode *inode,
1878                                            struct ext4_extent *newext,
1879                                            struct ext4_ext_path *path)
1880 {
1881         ext4_lblk_t b1, b2;
1882         unsigned int depth, len1;
1883         unsigned int ret = 0;
1884
1885         b1 = le32_to_cpu(newext->ee_block);
1886         len1 = ext4_ext_get_actual_len(newext);
1887         depth = ext_depth(inode);
1888         if (!path[depth].p_ext)
1889                 goto out;
1890         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1891
1892         /*
1893          * get the next allocated block if the extent in the path
1894          * is before the requested block(s)
1895          */
1896         if (b2 < b1) {
1897                 b2 = ext4_ext_next_allocated_block(path);
1898                 if (b2 == EXT_MAX_BLOCKS)
1899                         goto out;
1900                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1901         }
1902
1903         /* check for wrap through zero on extent logical start block*/
1904         if (b1 + len1 < b1) {
1905                 len1 = EXT_MAX_BLOCKS - b1;
1906                 newext->ee_len = cpu_to_le16(len1);
1907                 ret = 1;
1908         }
1909
1910         /* check for overlap */
1911         if (b1 + len1 > b2) {
1912                 newext->ee_len = cpu_to_le16(b2 - b1);
1913                 ret = 1;
1914         }
1915 out:
1916         return ret;
1917 }
1918
1919 /*
1920  * ext4_ext_insert_extent:
1921  * tries to merge requsted extent into the existing extent or
1922  * inserts requested extent as new one into the tree,
1923  * creating new leaf in the no-space case.
1924  */
1925 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1926                                 struct ext4_ext_path **ppath,
1927                                 struct ext4_extent *newext, int gb_flags)
1928 {
1929         struct ext4_ext_path *path = *ppath;
1930         struct ext4_extent_header *eh;
1931         struct ext4_extent *ex, *fex;
1932         struct ext4_extent *nearex; /* nearest extent */
1933         struct ext4_ext_path *npath = NULL;
1934         int depth, len, err;
1935         ext4_lblk_t next;
1936         int mb_flags = 0, unwritten;
1937
1938         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1939                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1940         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1941                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1942                 return -EFSCORRUPTED;
1943         }
1944         depth = ext_depth(inode);
1945         ex = path[depth].p_ext;
1946         eh = path[depth].p_hdr;
1947         if (unlikely(path[depth].p_hdr == NULL)) {
1948                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1949                 return -EFSCORRUPTED;
1950         }
1951
1952         /* try to insert block into found extent and return */
1953         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1954
1955                 /*
1956                  * Try to see whether we should rather test the extent on
1957                  * right from ex, or from the left of ex. This is because
1958                  * ext4_find_extent() can return either extent on the
1959                  * left, or on the right from the searched position. This
1960                  * will make merging more effective.
1961                  */
1962                 if (ex < EXT_LAST_EXTENT(eh) &&
1963                     (le32_to_cpu(ex->ee_block) +
1964                     ext4_ext_get_actual_len(ex) <
1965                     le32_to_cpu(newext->ee_block))) {
1966                         ex += 1;
1967                         goto prepend;
1968                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1969                            (le32_to_cpu(newext->ee_block) +
1970                            ext4_ext_get_actual_len(newext) <
1971                            le32_to_cpu(ex->ee_block)))
1972                         ex -= 1;
1973
1974                 /* Try to append newex to the ex */
1975                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1976                         ext_debug("append [%d]%d block to %u:[%d]%d"
1977                                   "(from %llu)\n",
1978                                   ext4_ext_is_unwritten(newext),
1979                                   ext4_ext_get_actual_len(newext),
1980                                   le32_to_cpu(ex->ee_block),
1981                                   ext4_ext_is_unwritten(ex),
1982                                   ext4_ext_get_actual_len(ex),
1983                                   ext4_ext_pblock(ex));
1984                         err = ext4_ext_get_access(handle, inode,
1985                                                   path + depth);
1986                         if (err)
1987                                 return err;
1988                         unwritten = ext4_ext_is_unwritten(ex);
1989                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1990                                         + ext4_ext_get_actual_len(newext));
1991                         if (unwritten)
1992                                 ext4_ext_mark_unwritten(ex);
1993                         eh = path[depth].p_hdr;
1994                         nearex = ex;
1995                         goto merge;
1996                 }
1997
1998 prepend:
1999                 /* Try to prepend newex to the ex */
2000                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2001                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2002                                   "(from %llu)\n",
2003                                   le32_to_cpu(newext->ee_block),
2004                                   ext4_ext_is_unwritten(newext),
2005                                   ext4_ext_get_actual_len(newext),
2006                                   le32_to_cpu(ex->ee_block),
2007                                   ext4_ext_is_unwritten(ex),
2008                                   ext4_ext_get_actual_len(ex),
2009                                   ext4_ext_pblock(ex));
2010                         err = ext4_ext_get_access(handle, inode,
2011                                                   path + depth);
2012                         if (err)
2013                                 return err;
2014
2015                         unwritten = ext4_ext_is_unwritten(ex);
2016                         ex->ee_block = newext->ee_block;
2017                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2018                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2019                                         + ext4_ext_get_actual_len(newext));
2020                         if (unwritten)
2021                                 ext4_ext_mark_unwritten(ex);
2022                         eh = path[depth].p_hdr;
2023                         nearex = ex;
2024                         goto merge;
2025                 }
2026         }
2027
2028         depth = ext_depth(inode);
2029         eh = path[depth].p_hdr;
2030         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2031                 goto has_space;
2032
2033         /* probably next leaf has space for us? */
2034         fex = EXT_LAST_EXTENT(eh);
2035         next = EXT_MAX_BLOCKS;
2036         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2037                 next = ext4_ext_next_leaf_block(path);
2038         if (next != EXT_MAX_BLOCKS) {
2039                 ext_debug("next leaf block - %u\n", next);
2040                 BUG_ON(npath != NULL);
2041                 npath = ext4_find_extent(inode, next, NULL, 0);
2042                 if (IS_ERR(npath))
2043                         return PTR_ERR(npath);
2044                 BUG_ON(npath->p_depth != path->p_depth);
2045                 eh = npath[depth].p_hdr;
2046                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2047                         ext_debug("next leaf isn't full(%d)\n",
2048                                   le16_to_cpu(eh->eh_entries));
2049                         path = npath;
2050                         goto has_space;
2051                 }
2052                 ext_debug("next leaf has no free space(%d,%d)\n",
2053                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2054         }
2055
2056         /*
2057          * There is no free space in the found leaf.
2058          * We're gonna add a new leaf in the tree.
2059          */
2060         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2061                 mb_flags |= EXT4_MB_USE_RESERVED;
2062         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2063                                        ppath, newext);
2064         if (err)
2065                 goto cleanup;
2066         depth = ext_depth(inode);
2067         eh = path[depth].p_hdr;
2068
2069 has_space:
2070         nearex = path[depth].p_ext;
2071
2072         err = ext4_ext_get_access(handle, inode, path + depth);
2073         if (err)
2074                 goto cleanup;
2075
2076         if (!nearex) {
2077                 /* there is no extent in this leaf, create first one */
2078                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2079                                 le32_to_cpu(newext->ee_block),
2080                                 ext4_ext_pblock(newext),
2081                                 ext4_ext_is_unwritten(newext),
2082                                 ext4_ext_get_actual_len(newext));
2083                 nearex = EXT_FIRST_EXTENT(eh);
2084         } else {
2085                 if (le32_to_cpu(newext->ee_block)
2086                            > le32_to_cpu(nearex->ee_block)) {
2087                         /* Insert after */
2088                         ext_debug("insert %u:%llu:[%d]%d before: "
2089                                         "nearest %p\n",
2090                                         le32_to_cpu(newext->ee_block),
2091                                         ext4_ext_pblock(newext),
2092                                         ext4_ext_is_unwritten(newext),
2093                                         ext4_ext_get_actual_len(newext),
2094                                         nearex);
2095                         nearex++;
2096                 } else {
2097                         /* Insert before */
2098                         BUG_ON(newext->ee_block == nearex->ee_block);
2099                         ext_debug("insert %u:%llu:[%d]%d after: "
2100                                         "nearest %p\n",
2101                                         le32_to_cpu(newext->ee_block),
2102                                         ext4_ext_pblock(newext),
2103                                         ext4_ext_is_unwritten(newext),
2104                                         ext4_ext_get_actual_len(newext),
2105                                         nearex);
2106                 }
2107                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2108                 if (len > 0) {
2109                         ext_debug("insert %u:%llu:[%d]%d: "
2110                                         "move %d extents from 0x%p to 0x%p\n",
2111                                         le32_to_cpu(newext->ee_block),
2112                                         ext4_ext_pblock(newext),
2113                                         ext4_ext_is_unwritten(newext),
2114                                         ext4_ext_get_actual_len(newext),
2115                                         len, nearex, nearex + 1);
2116                         memmove(nearex + 1, nearex,
2117                                 len * sizeof(struct ext4_extent));
2118                 }
2119         }
2120
2121         le16_add_cpu(&eh->eh_entries, 1);
2122         path[depth].p_ext = nearex;
2123         nearex->ee_block = newext->ee_block;
2124         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2125         nearex->ee_len = newext->ee_len;
2126
2127 merge:
2128         /* try to merge extents */
2129         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2130                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2131
2132
2133         /* time to correct all indexes above */
2134         err = ext4_ext_correct_indexes(handle, inode, path);
2135         if (err)
2136                 goto cleanup;
2137
2138         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2139
2140 cleanup:
2141         ext4_ext_drop_refs(npath);
2142         kfree(npath);
2143         return err;
2144 }
2145
2146 static int ext4_fill_fiemap_extents(struct inode *inode,
2147                                     ext4_lblk_t block, ext4_lblk_t num,
2148                                     struct fiemap_extent_info *fieinfo)
2149 {
2150         struct ext4_ext_path *path = NULL;
2151         struct ext4_extent *ex;
2152         struct extent_status es;
2153         ext4_lblk_t next, next_del, start = 0, end = 0;
2154         ext4_lblk_t last = block + num;
2155         int exists, depth = 0, err = 0;
2156         unsigned int flags = 0;
2157         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2158
2159         while (block < last && block != EXT_MAX_BLOCKS) {
2160                 num = last - block;
2161                 /* find extent for this block */
2162                 down_read(&EXT4_I(inode)->i_data_sem);
2163
2164                 path = ext4_find_extent(inode, block, &path, 0);
2165                 if (IS_ERR(path)) {
2166                         up_read(&EXT4_I(inode)->i_data_sem);
2167                         err = PTR_ERR(path);
2168                         path = NULL;
2169                         break;
2170                 }
2171
2172                 depth = ext_depth(inode);
2173                 if (unlikely(path[depth].p_hdr == NULL)) {
2174                         up_read(&EXT4_I(inode)->i_data_sem);
2175                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2176                         err = -EFSCORRUPTED;
2177                         break;
2178                 }
2179                 ex = path[depth].p_ext;
2180                 next = ext4_ext_next_allocated_block(path);
2181
2182                 flags = 0;
2183                 exists = 0;
2184                 if (!ex) {
2185                         /* there is no extent yet, so try to allocate
2186                          * all requested space */
2187                         start = block;
2188                         end = block + num;
2189                 } else if (le32_to_cpu(ex->ee_block) > block) {
2190                         /* need to allocate space before found extent */
2191                         start = block;
2192                         end = le32_to_cpu(ex->ee_block);
2193                         if (block + num < end)
2194                                 end = block + num;
2195                 } else if (block >= le32_to_cpu(ex->ee_block)
2196                                         + ext4_ext_get_actual_len(ex)) {
2197                         /* need to allocate space after found extent */
2198                         start = block;
2199                         end = block + num;
2200                         if (end >= next)
2201                                 end = next;
2202                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2203                         /*
2204                          * some part of requested space is covered
2205                          * by found extent
2206                          */
2207                         start = block;
2208                         end = le32_to_cpu(ex->ee_block)
2209                                 + ext4_ext_get_actual_len(ex);
2210                         if (block + num < end)
2211                                 end = block + num;
2212                         exists = 1;
2213                 } else {
2214                         BUG();
2215                 }
2216                 BUG_ON(end <= start);
2217
2218                 if (!exists) {
2219                         es.es_lblk = start;
2220                         es.es_len = end - start;
2221                         es.es_pblk = 0;
2222                 } else {
2223                         es.es_lblk = le32_to_cpu(ex->ee_block);
2224                         es.es_len = ext4_ext_get_actual_len(ex);
2225                         es.es_pblk = ext4_ext_pblock(ex);
2226                         if (ext4_ext_is_unwritten(ex))
2227                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2228                 }
2229
2230                 /*
2231                  * Find delayed extent and update es accordingly. We call
2232                  * it even in !exists case to find out whether es is the
2233                  * last existing extent or not.
2234                  */
2235                 next_del = ext4_find_delayed_extent(inode, &es);
2236                 if (!exists && next_del) {
2237                         exists = 1;
2238                         flags |= (FIEMAP_EXTENT_DELALLOC |
2239                                   FIEMAP_EXTENT_UNKNOWN);
2240                 }
2241                 up_read(&EXT4_I(inode)->i_data_sem);
2242
2243                 if (unlikely(es.es_len == 0)) {
2244                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2245                         err = -EFSCORRUPTED;
2246                         break;
2247                 }
2248
2249                 /*
2250                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2251                  * we need to check next == EXT_MAX_BLOCKS because it is
2252                  * possible that an extent is with unwritten and delayed
2253                  * status due to when an extent is delayed allocated and
2254                  * is allocated by fallocate status tree will track both of
2255                  * them in a extent.
2256                  *
2257                  * So we could return a unwritten and delayed extent, and
2258                  * its block is equal to 'next'.
2259                  */
2260                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2261                         flags |= FIEMAP_EXTENT_LAST;
2262                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2263                                      next != EXT_MAX_BLOCKS)) {
2264                                 EXT4_ERROR_INODE(inode,
2265                                                  "next extent == %u, next "
2266                                                  "delalloc extent = %u",
2267                                                  next, next_del);
2268                                 err = -EFSCORRUPTED;
2269                                 break;
2270                         }
2271                 }
2272
2273                 if (exists) {
2274                         err = fiemap_fill_next_extent(fieinfo,
2275                                 (__u64)es.es_lblk << blksize_bits,
2276                                 (__u64)es.es_pblk << blksize_bits,
2277                                 (__u64)es.es_len << blksize_bits,
2278                                 flags);
2279                         if (err < 0)
2280                                 break;
2281                         if (err == 1) {
2282                                 err = 0;
2283                                 break;
2284                         }
2285                 }
2286
2287                 block = es.es_lblk + es.es_len;
2288         }
2289
2290         ext4_ext_drop_refs(path);
2291         kfree(path);
2292         return err;
2293 }
2294
2295 /*
2296  * ext4_ext_determine_hole - determine hole around given block
2297  * @inode:      inode we lookup in
2298  * @path:       path in extent tree to @lblk
2299  * @lblk:       pointer to logical block around which we want to determine hole
2300  *
2301  * Determine hole length (and start if easily possible) around given logical
2302  * block. We don't try too hard to find the beginning of the hole but @path
2303  * actually points to extent before @lblk, we provide it.
2304  *
2305  * The function returns the length of a hole starting at @lblk. We update @lblk
2306  * to the beginning of the hole if we managed to find it.
2307  */
2308 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2309                                            struct ext4_ext_path *path,
2310                                            ext4_lblk_t *lblk)
2311 {
2312         int depth = ext_depth(inode);
2313         struct ext4_extent *ex;
2314         ext4_lblk_t len;
2315
2316         ex = path[depth].p_ext;
2317         if (ex == NULL) {
2318                 /* there is no extent yet, so gap is [0;-] */
2319                 *lblk = 0;
2320                 len = EXT_MAX_BLOCKS;
2321         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2322                 len = le32_to_cpu(ex->ee_block) - *lblk;
2323         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2324                         + ext4_ext_get_actual_len(ex)) {
2325                 ext4_lblk_t next;
2326
2327                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2328                 next = ext4_ext_next_allocated_block(path);
2329                 BUG_ON(next == *lblk);
2330                 len = next - *lblk;
2331         } else {
2332                 BUG();
2333         }
2334         return len;
2335 }
2336
2337 /*
2338  * ext4_ext_put_gap_in_cache:
2339  * calculate boundaries of the gap that the requested block fits into
2340  * and cache this gap
2341  */
2342 static void
2343 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2344                           ext4_lblk_t hole_len)
2345 {
2346         struct extent_status es;
2347
2348         ext4_es_find_delayed_extent_range(inode, hole_start,
2349                                           hole_start + hole_len - 1, &es);
2350         if (es.es_len) {
2351                 /* There's delayed extent containing lblock? */
2352                 if (es.es_lblk <= hole_start)
2353                         return;
2354                 hole_len = min(es.es_lblk - hole_start, hole_len);
2355         }
2356         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2357         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2358                               EXTENT_STATUS_HOLE);
2359 }
2360
2361 /*
2362  * ext4_ext_rm_idx:
2363  * removes index from the index block.
2364  */
2365 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2366                         struct ext4_ext_path *path, int depth)
2367 {
2368         int err;
2369         ext4_fsblk_t leaf;
2370
2371         /* free index block */
2372         depth--;
2373         path = path + depth;
2374         leaf = ext4_idx_pblock(path->p_idx);
2375         if (unlikely(path->p_hdr->eh_entries == 0)) {
2376                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2377                 return -EFSCORRUPTED;
2378         }
2379         err = ext4_ext_get_access(handle, inode, path);
2380         if (err)
2381                 return err;
2382
2383         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2384                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2385                 len *= sizeof(struct ext4_extent_idx);
2386                 memmove(path->p_idx, path->p_idx + 1, len);
2387         }
2388
2389         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2390         err = ext4_ext_dirty(handle, inode, path);
2391         if (err)
2392                 return err;
2393         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2394         trace_ext4_ext_rm_idx(inode, leaf);
2395
2396         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2397                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2398
2399         while (--depth >= 0) {
2400                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2401                         break;
2402                 path--;
2403                 err = ext4_ext_get_access(handle, inode, path);
2404                 if (err)
2405                         break;
2406                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2407                 err = ext4_ext_dirty(handle, inode, path);
2408                 if (err)
2409                         break;
2410         }
2411         return err;
2412 }
2413
2414 /*
2415  * ext4_ext_calc_credits_for_single_extent:
2416  * This routine returns max. credits that needed to insert an extent
2417  * to the extent tree.
2418  * When pass the actual path, the caller should calculate credits
2419  * under i_data_sem.
2420  */
2421 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2422                                                 struct ext4_ext_path *path)
2423 {
2424         if (path) {
2425                 int depth = ext_depth(inode);
2426                 int ret = 0;
2427
2428                 /* probably there is space in leaf? */
2429                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2430                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2431
2432                         /*
2433                          *  There are some space in the leaf tree, no
2434                          *  need to account for leaf block credit
2435                          *
2436                          *  bitmaps and block group descriptor blocks
2437                          *  and other metadata blocks still need to be
2438                          *  accounted.
2439                          */
2440                         /* 1 bitmap, 1 block group descriptor */
2441                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2442                         return ret;
2443                 }
2444         }
2445
2446         return ext4_chunk_trans_blocks(inode, nrblocks);
2447 }
2448
2449 /*
2450  * How many index/leaf blocks need to change/allocate to add @extents extents?
2451  *
2452  * If we add a single extent, then in the worse case, each tree level
2453  * index/leaf need to be changed in case of the tree split.
2454  *
2455  * If more extents are inserted, they could cause the whole tree split more
2456  * than once, but this is really rare.
2457  */
2458 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2459 {
2460         int index;
2461         int depth;
2462
2463         /* If we are converting the inline data, only one is needed here. */
2464         if (ext4_has_inline_data(inode))
2465                 return 1;
2466
2467         depth = ext_depth(inode);
2468
2469         if (extents <= 1)
2470                 index = depth * 2;
2471         else
2472                 index = depth * 3;
2473
2474         return index;
2475 }
2476
2477 static inline int get_default_free_blocks_flags(struct inode *inode)
2478 {
2479         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2480             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2481                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2482         else if (ext4_should_journal_data(inode))
2483                 return EXT4_FREE_BLOCKS_FORGET;
2484         return 0;
2485 }
2486
2487 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2488                               struct ext4_extent *ex,
2489                               long long *partial_cluster,
2490                               ext4_lblk_t from, ext4_lblk_t to)
2491 {
2492         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2493         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2494         ext4_fsblk_t pblk;
2495         int flags = get_default_free_blocks_flags(inode);
2496
2497         /*
2498          * For bigalloc file systems, we never free a partial cluster
2499          * at the beginning of the extent.  Instead, we make a note
2500          * that we tried freeing the cluster, and check to see if we
2501          * need to free it on a subsequent call to ext4_remove_blocks,
2502          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2503          */
2504         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2505
2506         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2507         /*
2508          * If we have a partial cluster, and it's different from the
2509          * cluster of the last block, we need to explicitly free the
2510          * partial cluster here.
2511          */
2512         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2513         if (*partial_cluster > 0 &&
2514             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2515                 ext4_free_blocks(handle, inode, NULL,
2516                                  EXT4_C2B(sbi, *partial_cluster),
2517                                  sbi->s_cluster_ratio, flags);
2518                 *partial_cluster = 0;
2519         }
2520
2521 #ifdef EXTENTS_STATS
2522         {
2523                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2524                 spin_lock(&sbi->s_ext_stats_lock);
2525                 sbi->s_ext_blocks += ee_len;
2526                 sbi->s_ext_extents++;
2527                 if (ee_len < sbi->s_ext_min)
2528                         sbi->s_ext_min = ee_len;
2529                 if (ee_len > sbi->s_ext_max)
2530                         sbi->s_ext_max = ee_len;
2531                 if (ext_depth(inode) > sbi->s_depth_max)
2532                         sbi->s_depth_max = ext_depth(inode);
2533                 spin_unlock(&sbi->s_ext_stats_lock);
2534         }
2535 #endif
2536         if (from >= le32_to_cpu(ex->ee_block)
2537             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2538                 /* tail removal */
2539                 ext4_lblk_t num;
2540                 long long first_cluster;
2541
2542                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2543                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2544                 /*
2545                  * Usually we want to free partial cluster at the end of the
2546                  * extent, except for the situation when the cluster is still
2547                  * used by any other extent (partial_cluster is negative).
2548                  */
2549                 if (*partial_cluster < 0 &&
2550                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2551                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2552
2553                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2554                           num, pblk, *partial_cluster);
2555                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2556                 /*
2557                  * If the block range to be freed didn't start at the
2558                  * beginning of a cluster, and we removed the entire
2559                  * extent and the cluster is not used by any other extent,
2560                  * save the partial cluster here, since we might need to
2561                  * delete if we determine that the truncate or punch hole
2562                  * operation has removed all of the blocks in the cluster.
2563                  * If that cluster is used by another extent, preserve its
2564                  * negative value so it isn't freed later on.
2565                  *
2566                  * If the whole extent wasn't freed, we've reached the
2567                  * start of the truncated/punched region and have finished
2568                  * removing blocks.  If there's a partial cluster here it's
2569                  * shared with the remainder of the extent and is no longer
2570                  * a candidate for removal.
2571                  */
2572                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2573                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2574                         if (first_cluster != -*partial_cluster)
2575                                 *partial_cluster = first_cluster;
2576                 } else {
2577                         *partial_cluster = 0;
2578                 }
2579         } else
2580                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2581                            "%u-%u from %u:%u",
2582                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2583         return 0;
2584 }
2585
2586
2587 /*
2588  * ext4_ext_rm_leaf() Removes the extents associated with the
2589  * blocks appearing between "start" and "end".  Both "start"
2590  * and "end" must appear in the same extent or EIO is returned.
2591  *
2592  * @handle: The journal handle
2593  * @inode:  The files inode
2594  * @path:   The path to the leaf
2595  * @partial_cluster: The cluster which we'll have to free if all extents
2596  *                   has been released from it.  However, if this value is
2597  *                   negative, it's a cluster just to the right of the
2598  *                   punched region and it must not be freed.
2599  * @start:  The first block to remove
2600  * @end:   The last block to remove
2601  */
2602 static int
2603 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2604                  struct ext4_ext_path *path,
2605                  long long *partial_cluster,
2606                  ext4_lblk_t start, ext4_lblk_t end)
2607 {
2608         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2609         int err = 0, correct_index = 0;
2610         int depth = ext_depth(inode), credits;
2611         struct ext4_extent_header *eh;
2612         ext4_lblk_t a, b;
2613         unsigned num;
2614         ext4_lblk_t ex_ee_block;
2615         unsigned short ex_ee_len;
2616         unsigned unwritten = 0;
2617         struct ext4_extent *ex;
2618         ext4_fsblk_t pblk;
2619
2620         /* the header must be checked already in ext4_ext_remove_space() */
2621         ext_debug("truncate since %u in leaf to %u\n", start, end);
2622         if (!path[depth].p_hdr)
2623                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2624         eh = path[depth].p_hdr;
2625         if (unlikely(path[depth].p_hdr == NULL)) {
2626                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2627                 return -EFSCORRUPTED;
2628         }
2629         /* find where to start removing */
2630         ex = path[depth].p_ext;
2631         if (!ex)
2632                 ex = EXT_LAST_EXTENT(eh);
2633
2634         ex_ee_block = le32_to_cpu(ex->ee_block);
2635         ex_ee_len = ext4_ext_get_actual_len(ex);
2636
2637         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2638
2639         while (ex >= EXT_FIRST_EXTENT(eh) &&
2640                         ex_ee_block + ex_ee_len > start) {
2641
2642                 if (ext4_ext_is_unwritten(ex))
2643                         unwritten = 1;
2644                 else
2645                         unwritten = 0;
2646
2647                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2648                           unwritten, ex_ee_len);
2649                 path[depth].p_ext = ex;
2650
2651                 a = ex_ee_block > start ? ex_ee_block : start;
2652                 b = ex_ee_block+ex_ee_len - 1 < end ?
2653                         ex_ee_block+ex_ee_len - 1 : end;
2654
2655                 ext_debug("  border %u:%u\n", a, b);
2656
2657                 /* If this extent is beyond the end of the hole, skip it */
2658                 if (end < ex_ee_block) {
2659                         /*
2660                          * We're going to skip this extent and move to another,
2661                          * so note that its first cluster is in use to avoid
2662                          * freeing it when removing blocks.  Eventually, the
2663                          * right edge of the truncated/punched region will
2664                          * be just to the left.
2665                          */
2666                         if (sbi->s_cluster_ratio > 1) {
2667                                 pblk = ext4_ext_pblock(ex);
2668                                 *partial_cluster =
2669                                         -(long long) EXT4_B2C(sbi, pblk);
2670                         }
2671                         ex--;
2672                         ex_ee_block = le32_to_cpu(ex->ee_block);
2673                         ex_ee_len = ext4_ext_get_actual_len(ex);
2674                         continue;
2675                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2676                         EXT4_ERROR_INODE(inode,
2677                                          "can not handle truncate %u:%u "
2678                                          "on extent %u:%u",
2679                                          start, end, ex_ee_block,
2680                                          ex_ee_block + ex_ee_len - 1);
2681                         err = -EFSCORRUPTED;
2682                         goto out;
2683                 } else if (a != ex_ee_block) {
2684                         /* remove tail of the extent */
2685                         num = a - ex_ee_block;
2686                 } else {
2687                         /* remove whole extent: excellent! */
2688                         num = 0;
2689                 }
2690                 /*
2691                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2692                  * descriptor) for each block group; assume two block
2693                  * groups plus ex_ee_len/blocks_per_block_group for
2694                  * the worst case
2695                  */
2696                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2697                 if (ex == EXT_FIRST_EXTENT(eh)) {
2698                         correct_index = 1;
2699                         credits += (ext_depth(inode)) + 1;
2700                 }
2701                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2702
2703                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2704                 if (err)
2705                         goto out;
2706
2707                 err = ext4_ext_get_access(handle, inode, path + depth);
2708                 if (err)
2709                         goto out;
2710
2711                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2712                                          a, b);
2713                 if (err)
2714                         goto out;
2715
2716                 if (num == 0)
2717                         /* this extent is removed; mark slot entirely unused */
2718                         ext4_ext_store_pblock(ex, 0);
2719
2720                 ex->ee_len = cpu_to_le16(num);
2721                 /*
2722                  * Do not mark unwritten if all the blocks in the
2723                  * extent have been removed.
2724                  */
2725                 if (unwritten && num)
2726                         ext4_ext_mark_unwritten(ex);
2727                 /*
2728                  * If the extent was completely released,
2729                  * we need to remove it from the leaf
2730                  */
2731                 if (num == 0) {
2732                         if (end != EXT_MAX_BLOCKS - 1) {
2733                                 /*
2734                                  * For hole punching, we need to scoot all the
2735                                  * extents up when an extent is removed so that
2736                                  * we dont have blank extents in the middle
2737                                  */
2738                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2739                                         sizeof(struct ext4_extent));
2740
2741                                 /* Now get rid of the one at the end */
2742                                 memset(EXT_LAST_EXTENT(eh), 0,
2743                                         sizeof(struct ext4_extent));
2744                         }
2745                         le16_add_cpu(&eh->eh_entries, -1);
2746                 }
2747
2748                 err = ext4_ext_dirty(handle, inode, path + depth);
2749                 if (err)
2750                         goto out;
2751
2752                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2753                                 ext4_ext_pblock(ex));
2754                 ex--;
2755                 ex_ee_block = le32_to_cpu(ex->ee_block);
2756                 ex_ee_len = ext4_ext_get_actual_len(ex);
2757         }
2758
2759         if (correct_index && eh->eh_entries)
2760                 err = ext4_ext_correct_indexes(handle, inode, path);
2761
2762         /*
2763          * If there's a partial cluster and at least one extent remains in
2764          * the leaf, free the partial cluster if it isn't shared with the
2765          * current extent.  If it is shared with the current extent
2766          * we zero partial_cluster because we've reached the start of the
2767          * truncated/punched region and we're done removing blocks.
2768          */
2769         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2770                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2771                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2772                         ext4_free_blocks(handle, inode, NULL,
2773                                          EXT4_C2B(sbi, *partial_cluster),
2774                                          sbi->s_cluster_ratio,
2775                                          get_default_free_blocks_flags(inode));
2776                 }
2777                 *partial_cluster = 0;
2778         }
2779
2780         /* if this leaf is free, then we should
2781          * remove it from index block above */
2782         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2783                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2784
2785 out:
2786         return err;
2787 }
2788
2789 /*
2790  * ext4_ext_more_to_rm:
2791  * returns 1 if current index has to be freed (even partial)
2792  */
2793 static int
2794 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2795 {
2796         BUG_ON(path->p_idx == NULL);
2797
2798         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2799                 return 0;
2800
2801         /*
2802          * if truncate on deeper level happened, it wasn't partial,
2803          * so we have to consider current index for truncation
2804          */
2805         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2806                 return 0;
2807         return 1;
2808 }
2809
2810 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2811                           ext4_lblk_t end)
2812 {
2813         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2814         int depth = ext_depth(inode);
2815         struct ext4_ext_path *path = NULL;
2816         long long partial_cluster = 0;
2817         handle_t *handle;
2818         int i = 0, err = 0;
2819
2820         ext_debug("truncate since %u to %u\n", start, end);
2821
2822         /* probably first extent we're gonna free will be last in block */
2823         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2824         if (IS_ERR(handle))
2825                 return PTR_ERR(handle);
2826
2827 again:
2828         trace_ext4_ext_remove_space(inode, start, end, depth);
2829
2830         /*
2831          * Check if we are removing extents inside the extent tree. If that
2832          * is the case, we are going to punch a hole inside the extent tree
2833          * so we have to check whether we need to split the extent covering
2834          * the last block to remove so we can easily remove the part of it
2835          * in ext4_ext_rm_leaf().
2836          */
2837         if (end < EXT_MAX_BLOCKS - 1) {
2838                 struct ext4_extent *ex;
2839                 ext4_lblk_t ee_block, ex_end, lblk;
2840                 ext4_fsblk_t pblk;
2841
2842                 /* find extent for or closest extent to this block */
2843                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2844                 if (IS_ERR(path)) {
2845                         ext4_journal_stop(handle);
2846                         return PTR_ERR(path);
2847                 }
2848                 depth = ext_depth(inode);
2849                 /* Leaf not may not exist only if inode has no blocks at all */
2850                 ex = path[depth].p_ext;
2851                 if (!ex) {
2852                         if (depth) {
2853                                 EXT4_ERROR_INODE(inode,
2854                                                  "path[%d].p_hdr == NULL",
2855                                                  depth);
2856                                 err = -EFSCORRUPTED;
2857                         }
2858                         goto out;
2859                 }
2860
2861                 ee_block = le32_to_cpu(ex->ee_block);
2862                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2863
2864                 /*
2865                  * See if the last block is inside the extent, if so split
2866                  * the extent at 'end' block so we can easily remove the
2867                  * tail of the first part of the split extent in
2868                  * ext4_ext_rm_leaf().
2869                  */
2870                 if (end >= ee_block && end < ex_end) {
2871
2872                         /*
2873                          * If we're going to split the extent, note that
2874                          * the cluster containing the block after 'end' is
2875                          * in use to avoid freeing it when removing blocks.
2876                          */
2877                         if (sbi->s_cluster_ratio > 1) {
2878                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2879                                 partial_cluster =
2880                                         -(long long) EXT4_B2C(sbi, pblk);
2881                         }
2882
2883                         /*
2884                          * Split the extent in two so that 'end' is the last
2885                          * block in the first new extent. Also we should not
2886                          * fail removing space due to ENOSPC so try to use
2887                          * reserved block if that happens.
2888                          */
2889                         err = ext4_force_split_extent_at(handle, inode, &path,
2890                                                          end + 1, 1);
2891                         if (err < 0)
2892                                 goto out;
2893
2894                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2895                         /*
2896                          * If there's an extent to the right its first cluster
2897                          * contains the immediate right boundary of the
2898                          * truncated/punched region.  Set partial_cluster to
2899                          * its negative value so it won't be freed if shared
2900                          * with the current extent.  The end < ee_block case
2901                          * is handled in ext4_ext_rm_leaf().
2902                          */
2903                         lblk = ex_end + 1;
2904                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2905                                                     &ex);
2906                         if (err)
2907                                 goto out;
2908                         if (pblk)
2909                                 partial_cluster =
2910                                         -(long long) EXT4_B2C(sbi, pblk);
2911                 }
2912         }
2913         /*
2914          * We start scanning from right side, freeing all the blocks
2915          * after i_size and walking into the tree depth-wise.
2916          */
2917         depth = ext_depth(inode);
2918         if (path) {
2919                 int k = i = depth;
2920                 while (--k > 0)
2921                         path[k].p_block =
2922                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2923         } else {
2924                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2925                                GFP_NOFS);
2926                 if (path == NULL) {
2927                         ext4_journal_stop(handle);
2928                         return -ENOMEM;
2929                 }
2930                 path[0].p_maxdepth = path[0].p_depth = depth;
2931                 path[0].p_hdr = ext_inode_hdr(inode);
2932                 i = 0;
2933
2934                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2935                         err = -EFSCORRUPTED;
2936                         goto out;
2937                 }
2938         }
2939         err = 0;
2940
2941         while (i >= 0 && err == 0) {
2942                 if (i == depth) {
2943                         /* this is leaf block */
2944                         err = ext4_ext_rm_leaf(handle, inode, path,
2945                                                &partial_cluster, start,
2946                                                end);
2947                         /* root level has p_bh == NULL, brelse() eats this */
2948                         brelse(path[i].p_bh);
2949                         path[i].p_bh = NULL;
2950                         i--;
2951                         continue;
2952                 }
2953
2954                 /* this is index block */
2955                 if (!path[i].p_hdr) {
2956                         ext_debug("initialize header\n");
2957                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2958                 }
2959
2960                 if (!path[i].p_idx) {
2961                         /* this level hasn't been touched yet */
2962                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2963                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2964                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2965                                   path[i].p_hdr,
2966                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2967                 } else {
2968                         /* we were already here, see at next index */
2969                         path[i].p_idx--;
2970                 }
2971
2972                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2973                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2974                                 path[i].p_idx);
2975                 if (ext4_ext_more_to_rm(path + i)) {
2976                         struct buffer_head *bh;
2977                         /* go to the next level */
2978                         ext_debug("move to level %d (block %llu)\n",
2979                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2980                         memset(path + i + 1, 0, sizeof(*path));
2981                         bh = read_extent_tree_block(inode,
2982                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2983                                 EXT4_EX_NOCACHE);
2984                         if (IS_ERR(bh)) {
2985                                 /* should we reset i_size? */
2986                                 err = PTR_ERR(bh);
2987                                 break;
2988                         }
2989                         /* Yield here to deal with large extent trees.
2990                          * Should be a no-op if we did IO above. */
2991                         cond_resched();
2992                         if (WARN_ON(i + 1 > depth)) {
2993                                 err = -EFSCORRUPTED;
2994                                 break;
2995                         }
2996                         path[i + 1].p_bh = bh;
2997
2998                         /* save actual number of indexes since this
2999                          * number is changed at the next iteration */
3000                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3001                         i++;
3002                 } else {
3003                         /* we finished processing this index, go up */
3004                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3005                                 /* index is empty, remove it;
3006                                  * handle must be already prepared by the
3007                                  * truncatei_leaf() */
3008                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3009                         }
3010                         /* root level has p_bh == NULL, brelse() eats this */
3011                         brelse(path[i].p_bh);
3012                         path[i].p_bh = NULL;
3013                         i--;
3014                         ext_debug("return to level %d\n", i);
3015                 }
3016         }
3017
3018         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3019                         partial_cluster, path->p_hdr->eh_entries);
3020
3021         /*
3022          * If we still have something in the partial cluster and we have removed
3023          * even the first extent, then we should free the blocks in the partial
3024          * cluster as well.  (This code will only run when there are no leaves
3025          * to the immediate left of the truncated/punched region.)
3026          */
3027         if (partial_cluster > 0 && err == 0) {
3028                 /* don't zero partial_cluster since it's not used afterwards */
3029                 ext4_free_blocks(handle, inode, NULL,
3030                                  EXT4_C2B(sbi, partial_cluster),
3031                                  sbi->s_cluster_ratio,
3032                                  get_default_free_blocks_flags(inode));
3033         }
3034
3035         /* TODO: flexible tree reduction should be here */
3036         if (path->p_hdr->eh_entries == 0) {
3037                 /*
3038                  * truncate to zero freed all the tree,
3039                  * so we need to correct eh_depth
3040                  */
3041                 err = ext4_ext_get_access(handle, inode, path);
3042                 if (err == 0) {
3043                         ext_inode_hdr(inode)->eh_depth = 0;
3044                         ext_inode_hdr(inode)->eh_max =
3045                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3046                         err = ext4_ext_dirty(handle, inode, path);
3047                 }
3048         }
3049 out:
3050         ext4_ext_drop_refs(path);
3051         kfree(path);
3052         path = NULL;
3053         if (err == -EAGAIN)
3054                 goto again;
3055         ext4_journal_stop(handle);
3056
3057         return err;
3058 }
3059
3060 /*
3061  * called at mount time
3062  */
3063 void ext4_ext_init(struct super_block *sb)
3064 {
3065         /*
3066          * possible initialization would be here
3067          */
3068
3069         if (ext4_has_feature_extents(sb)) {
3070 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3071                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3072 #ifdef AGGRESSIVE_TEST
3073                        ", aggressive tests"
3074 #endif
3075 #ifdef CHECK_BINSEARCH
3076                        ", check binsearch"
3077 #endif
3078 #ifdef EXTENTS_STATS
3079                        ", stats"
3080 #endif
3081                        "\n");
3082 #endif
3083 #ifdef EXTENTS_STATS
3084                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3085                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3086                 EXT4_SB(sb)->s_ext_max = 0;
3087 #endif
3088         }
3089 }
3090
3091 /*
3092  * called at umount time
3093  */
3094 void ext4_ext_release(struct super_block *sb)
3095 {
3096         if (!ext4_has_feature_extents(sb))
3097                 return;
3098
3099 #ifdef EXTENTS_STATS
3100         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3101                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3102                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3103                         sbi->s_ext_blocks, sbi->s_ext_extents,
3104                         sbi->s_ext_blocks / sbi->s_ext_extents);
3105                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3106                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3107         }
3108 #endif
3109 }
3110
3111 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3112 {
3113         ext4_lblk_t  ee_block;
3114         ext4_fsblk_t ee_pblock;
3115         unsigned int ee_len;
3116
3117         ee_block  = le32_to_cpu(ex->ee_block);
3118         ee_len    = ext4_ext_get_actual_len(ex);
3119         ee_pblock = ext4_ext_pblock(ex);
3120
3121         if (ee_len == 0)
3122                 return 0;
3123
3124         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3125                                      EXTENT_STATUS_WRITTEN);
3126 }
3127
3128 /* FIXME!! we need to try to merge to left or right after zero-out  */
3129 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3130 {
3131         ext4_fsblk_t ee_pblock;
3132         unsigned int ee_len;
3133
3134         ee_len    = ext4_ext_get_actual_len(ex);
3135         ee_pblock = ext4_ext_pblock(ex);
3136         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3137                                   ee_len);
3138 }
3139
3140 /*
3141  * ext4_split_extent_at() splits an extent at given block.
3142  *
3143  * @handle: the journal handle
3144  * @inode: the file inode
3145  * @path: the path to the extent
3146  * @split: the logical block where the extent is splitted.
3147  * @split_flags: indicates if the extent could be zeroout if split fails, and
3148  *               the states(init or unwritten) of new extents.
3149  * @flags: flags used to insert new extent to extent tree.
3150  *
3151  *
3152  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3153  * of which are deterimined by split_flag.
3154  *
3155  * There are two cases:
3156  *  a> the extent are splitted into two extent.
3157  *  b> split is not needed, and just mark the extent.
3158  *
3159  * return 0 on success.
3160  */
3161 static int ext4_split_extent_at(handle_t *handle,
3162                              struct inode *inode,
3163                              struct ext4_ext_path **ppath,
3164                              ext4_lblk_t split,
3165                              int split_flag,
3166                              int flags)
3167 {
3168         struct ext4_ext_path *path = *ppath;
3169         ext4_fsblk_t newblock;
3170         ext4_lblk_t ee_block;
3171         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3172         struct ext4_extent *ex2 = NULL;
3173         unsigned int ee_len, depth;
3174         int err = 0;
3175
3176         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3177                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3178
3179         ext_debug("ext4_split_extents_at: inode %lu, logical"
3180                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3181
3182         ext4_ext_show_leaf(inode, path);
3183
3184         depth = ext_depth(inode);
3185         ex = path[depth].p_ext;
3186         ee_block = le32_to_cpu(ex->ee_block);
3187         ee_len = ext4_ext_get_actual_len(ex);
3188         newblock = split - ee_block + ext4_ext_pblock(ex);
3189
3190         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3191         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3192                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3193                              EXT4_EXT_MARK_UNWRIT1 |
3194                              EXT4_EXT_MARK_UNWRIT2));
3195
3196         err = ext4_ext_get_access(handle, inode, path + depth);
3197         if (err)
3198                 goto out;
3199
3200         if (split == ee_block) {
3201                 /*
3202                  * case b: block @split is the block that the extent begins with
3203                  * then we just change the state of the extent, and splitting
3204                  * is not needed.
3205                  */
3206                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3207                         ext4_ext_mark_unwritten(ex);
3208                 else
3209                         ext4_ext_mark_initialized(ex);
3210
3211                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3212                         ext4_ext_try_to_merge(handle, inode, path, ex);
3213
3214                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3215                 goto out;
3216         }
3217
3218         /* case a */
3219         memcpy(&orig_ex, ex, sizeof(orig_ex));
3220         ex->ee_len = cpu_to_le16(split - ee_block);
3221         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3222                 ext4_ext_mark_unwritten(ex);
3223
3224         /*
3225          * path may lead to new leaf, not to original leaf any more
3226          * after ext4_ext_insert_extent() returns,
3227          */
3228         err = ext4_ext_dirty(handle, inode, path + depth);
3229         if (err)
3230                 goto fix_extent_len;
3231
3232         ex2 = &newex;
3233         ex2->ee_block = cpu_to_le32(split);
3234         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3235         ext4_ext_store_pblock(ex2, newblock);
3236         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3237                 ext4_ext_mark_unwritten(ex2);
3238
3239         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3240         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3241                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3242                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3243                                 err = ext4_ext_zeroout(inode, ex2);
3244                                 zero_ex.ee_block = ex2->ee_block;
3245                                 zero_ex.ee_len = cpu_to_le16(
3246                                                 ext4_ext_get_actual_len(ex2));
3247                                 ext4_ext_store_pblock(&zero_ex,
3248                                                       ext4_ext_pblock(ex2));
3249                         } else {
3250                                 err = ext4_ext_zeroout(inode, ex);
3251                                 zero_ex.ee_block = ex->ee_block;
3252                                 zero_ex.ee_len = cpu_to_le16(
3253                                                 ext4_ext_get_actual_len(ex));
3254                                 ext4_ext_store_pblock(&zero_ex,
3255                                                       ext4_ext_pblock(ex));
3256                         }
3257                 } else {
3258                         err = ext4_ext_zeroout(inode, &orig_ex);
3259                         zero_ex.ee_block = orig_ex.ee_block;
3260                         zero_ex.ee_len = cpu_to_le16(
3261                                                 ext4_ext_get_actual_len(&orig_ex));
3262                         ext4_ext_store_pblock(&zero_ex,
3263                                               ext4_ext_pblock(&orig_ex));
3264                 }
3265
3266                 if (err)
3267                         goto fix_extent_len;
3268                 /* update the extent length and mark as initialized */
3269                 ex->ee_len = cpu_to_le16(ee_len);
3270                 ext4_ext_try_to_merge(handle, inode, path, ex);
3271                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3272                 if (err)
3273                         goto fix_extent_len;
3274
3275                 /* update extent status tree */
3276                 err = ext4_zeroout_es(inode, &zero_ex);
3277
3278                 goto out;
3279         } else if (err)
3280                 goto fix_extent_len;
3281
3282 out:
3283         ext4_ext_show_leaf(inode, path);
3284         return err;
3285
3286 fix_extent_len:
3287         ex->ee_len = orig_ex.ee_len;
3288         ext4_ext_dirty(handle, inode, path + path->p_depth);
3289         return err;
3290 }
3291
3292 /*
3293  * ext4_split_extents() splits an extent and mark extent which is covered
3294  * by @map as split_flags indicates
3295  *
3296  * It may result in splitting the extent into multiple extents (up to three)
3297  * There are three possibilities:
3298  *   a> There is no split required
3299  *   b> Splits in two extents: Split is happening at either end of the extent
3300  *   c> Splits in three extents: Somone is splitting in middle of the extent
3301  *
3302  */
3303 static int ext4_split_extent(handle_t *handle,
3304                               struct inode *inode,
3305                               struct ext4_ext_path **ppath,
3306                               struct ext4_map_blocks *map,
3307                               int split_flag,
3308                               int flags)
3309 {
3310         struct ext4_ext_path *path = *ppath;
3311         ext4_lblk_t ee_block;
3312         struct ext4_extent *ex;
3313         unsigned int ee_len, depth;
3314         int err = 0;
3315         int unwritten;
3316         int split_flag1, flags1;
3317         int allocated = map->m_len;
3318
3319         depth = ext_depth(inode);
3320         ex = path[depth].p_ext;
3321         ee_block = le32_to_cpu(ex->ee_block);
3322         ee_len = ext4_ext_get_actual_len(ex);
3323         unwritten = ext4_ext_is_unwritten(ex);
3324
3325         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3326                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3327                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3328                 if (unwritten)
3329                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3330                                        EXT4_EXT_MARK_UNWRIT2;
3331                 if (split_flag & EXT4_EXT_DATA_VALID2)
3332                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3333                 err = ext4_split_extent_at(handle, inode, ppath,
3334                                 map->m_lblk + map->m_len, split_flag1, flags1);
3335                 if (err)
3336                         goto out;
3337         } else {
3338                 allocated = ee_len - (map->m_lblk - ee_block);
3339         }
3340         /*
3341          * Update path is required because previous ext4_split_extent_at() may
3342          * result in split of original leaf or extent zeroout.
3343          */
3344         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3345         if (IS_ERR(path))
3346                 return PTR_ERR(path);
3347         depth = ext_depth(inode);
3348         ex = path[depth].p_ext;
3349         if (!ex) {
3350                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3351                                  (unsigned long) map->m_lblk);
3352                 return -EFSCORRUPTED;
3353         }
3354         unwritten = ext4_ext_is_unwritten(ex);
3355         split_flag1 = 0;
3356
3357         if (map->m_lblk >= ee_block) {
3358                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3359                 if (unwritten) {
3360                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3361                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3362                                                      EXT4_EXT_MARK_UNWRIT2);
3363                 }
3364                 err = ext4_split_extent_at(handle, inode, ppath,
3365                                 map->m_lblk, split_flag1, flags);
3366                 if (err)
3367                         goto out;
3368         }
3369
3370         ext4_ext_show_leaf(inode, path);
3371 out:
3372         return err ? err : allocated;
3373 }
3374
3375 /*
3376  * This function is called by ext4_ext_map_blocks() if someone tries to write
3377  * to an unwritten extent. It may result in splitting the unwritten
3378  * extent into multiple extents (up to three - one initialized and two
3379  * unwritten).
3380  * There are three possibilities:
3381  *   a> There is no split required: Entire extent should be initialized
3382  *   b> Splits in two extents: Write is happening at either end of the extent
3383  *   c> Splits in three extents: Somone is writing in middle of the extent
3384  *
3385  * Pre-conditions:
3386  *  - The extent pointed to by 'path' is unwritten.
3387  *  - The extent pointed to by 'path' contains a superset
3388  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3389  *
3390  * Post-conditions on success:
3391  *  - the returned value is the number of blocks beyond map->l_lblk
3392  *    that are allocated and initialized.
3393  *    It is guaranteed to be >= map->m_len.
3394  */
3395 static int ext4_ext_convert_to_initialized(handle_t *handle,
3396                                            struct inode *inode,
3397                                            struct ext4_map_blocks *map,
3398                                            struct ext4_ext_path **ppath,
3399                                            int flags)
3400 {
3401         struct ext4_ext_path *path = *ppath;
3402         struct ext4_sb_info *sbi;
3403         struct ext4_extent_header *eh;
3404         struct ext4_map_blocks split_map;
3405         struct ext4_extent zero_ex1, zero_ex2;
3406         struct ext4_extent *ex, *abut_ex;
3407         ext4_lblk_t ee_block, eof_block;
3408         unsigned int ee_len, depth, map_len = map->m_len;
3409         int allocated = 0, max_zeroout = 0;
3410         int err = 0;
3411         int split_flag = EXT4_EXT_DATA_VALID2;
3412
3413         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3414                 "block %llu, max_blocks %u\n", inode->i_ino,
3415                 (unsigned long long)map->m_lblk, map_len);
3416
3417         sbi = EXT4_SB(inode->i_sb);
3418         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3419                 inode->i_sb->s_blocksize_bits;
3420         if (eof_block < map->m_lblk + map_len)
3421                 eof_block = map->m_lblk + map_len;
3422
3423         depth = ext_depth(inode);
3424         eh = path[depth].p_hdr;
3425         ex = path[depth].p_ext;
3426         ee_block = le32_to_cpu(ex->ee_block);
3427         ee_len = ext4_ext_get_actual_len(ex);
3428         zero_ex1.ee_len = 0;
3429         zero_ex2.ee_len = 0;
3430
3431         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3432
3433         /* Pre-conditions */
3434         BUG_ON(!ext4_ext_is_unwritten(ex));
3435         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3436
3437         /*
3438          * Attempt to transfer newly initialized blocks from the currently
3439          * unwritten extent to its neighbor. This is much cheaper
3440          * than an insertion followed by a merge as those involve costly
3441          * memmove() calls. Transferring to the left is the common case in
3442          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3443          * followed by append writes.
3444          *
3445          * Limitations of the current logic:
3446          *  - L1: we do not deal with writes covering the whole extent.
3447          *    This would require removing the extent if the transfer
3448          *    is possible.
3449          *  - L2: we only attempt to merge with an extent stored in the
3450          *    same extent tree node.
3451          */
3452         if ((map->m_lblk == ee_block) &&
3453                 /* See if we can merge left */
3454                 (map_len < ee_len) &&           /*L1*/
3455                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3456                 ext4_lblk_t prev_lblk;
3457                 ext4_fsblk_t prev_pblk, ee_pblk;
3458                 unsigned int prev_len;
3459
3460                 abut_ex = ex - 1;
3461                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3462                 prev_len = ext4_ext_get_actual_len(abut_ex);
3463                 prev_pblk = ext4_ext_pblock(abut_ex);
3464                 ee_pblk = ext4_ext_pblock(ex);
3465
3466                 /*
3467                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3468                  * upon those conditions:
3469                  * - C1: abut_ex is initialized,
3470                  * - C2: abut_ex is logically abutting ex,
3471                  * - C3: abut_ex is physically abutting ex,
3472                  * - C4: abut_ex can receive the additional blocks without
3473                  *   overflowing the (initialized) length limit.
3474                  */
3475                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3476                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3477                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3478                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3479                         err = ext4_ext_get_access(handle, inode, path + depth);
3480                         if (err)
3481                                 goto out;
3482
3483                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3484                                 map, ex, abut_ex);
3485
3486                         /* Shift the start of ex by 'map_len' blocks */
3487                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3488                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3489                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3490                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3491
3492                         /* Extend abut_ex by 'map_len' blocks */
3493                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3494
3495                         /* Result: number of initialized blocks past m_lblk */
3496                         allocated = map_len;
3497                 }
3498         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3499                    (map_len < ee_len) &&        /*L1*/
3500                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3501                 /* See if we can merge right */
3502                 ext4_lblk_t next_lblk;
3503                 ext4_fsblk_t next_pblk, ee_pblk;
3504                 unsigned int next_len;
3505
3506                 abut_ex = ex + 1;
3507                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3508                 next_len = ext4_ext_get_actual_len(abut_ex);
3509                 next_pblk = ext4_ext_pblock(abut_ex);
3510                 ee_pblk = ext4_ext_pblock(ex);
3511
3512                 /*
3513                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3514                  * upon those conditions:
3515                  * - C1: abut_ex is initialized,
3516                  * - C2: abut_ex is logically abutting ex,
3517                  * - C3: abut_ex is physically abutting ex,
3518                  * - C4: abut_ex can receive the additional blocks without
3519                  *   overflowing the (initialized) length limit.
3520                  */
3521                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3522                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3523                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3524                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3525                         err = ext4_ext_get_access(handle, inode, path + depth);
3526                         if (err)
3527                                 goto out;
3528
3529                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3530                                 map, ex, abut_ex);
3531
3532                         /* Shift the start of abut_ex by 'map_len' blocks */
3533                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3534                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3535                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3536                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3537
3538                         /* Extend abut_ex by 'map_len' blocks */
3539                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3540
3541                         /* Result: number of initialized blocks past m_lblk */
3542                         allocated = map_len;
3543                 }
3544         }
3545         if (allocated) {
3546                 /* Mark the block containing both extents as dirty */
3547                 ext4_ext_dirty(handle, inode, path + depth);
3548
3549                 /* Update path to point to the right extent */
3550                 path[depth].p_ext = abut_ex;
3551                 goto out;
3552         } else
3553                 allocated = ee_len - (map->m_lblk - ee_block);
3554
3555         WARN_ON(map->m_lblk < ee_block);
3556         /*
3557          * It is safe to convert extent to initialized via explicit
3558          * zeroout only if extent is fully inside i_size or new_size.
3559          */
3560         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3561
3562         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3563                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3564                         (inode->i_sb->s_blocksize_bits - 10);
3565
3566         if (ext4_encrypted_inode(inode))
3567                 max_zeroout = 0;
3568
3569         /*
3570          * five cases:
3571          * 1. split the extent into three extents.
3572          * 2. split the extent into two extents, zeroout the head of the first
3573          *    extent.
3574          * 3. split the extent into two extents, zeroout the tail of the second
3575          *    extent.
3576          * 4. split the extent into two extents with out zeroout.
3577          * 5. no splitting needed, just possibly zeroout the head and / or the
3578          *    tail of the extent.
3579          */
3580         split_map.m_lblk = map->m_lblk;
3581         split_map.m_len = map->m_len;
3582
3583         if (max_zeroout && (allocated > split_map.m_len)) {
3584                 if (allocated <= max_zeroout) {
3585                         /* case 3 or 5 */
3586                         zero_ex1.ee_block =
3587                                  cpu_to_le32(split_map.m_lblk +
3588                                              split_map.m_len);
3589                         zero_ex1.ee_len =
3590                                 cpu_to_le16(allocated - split_map.m_len);
3591                         ext4_ext_store_pblock(&zero_ex1,
3592                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3593                                 split_map.m_len - ee_block);
3594                         err = ext4_ext_zeroout(inode, &zero_ex1);
3595                         if (err)
3596                                 goto out;
3597                         split_map.m_len = allocated;
3598                 }
3599                 if (split_map.m_lblk - ee_block + split_map.m_len <
3600                                                                 max_zeroout) {
3601                         /* case 2 or 5 */
3602                         if (split_map.m_lblk != ee_block) {
3603                                 zero_ex2.ee_block = ex->ee_block;
3604                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3605                                                         ee_block);
3606                                 ext4_ext_store_pblock(&zero_ex2,
3607                                                       ext4_ext_pblock(ex));
3608                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3609                                 if (err)
3610                                         goto out;
3611                         }
3612
3613                         split_map.m_len += split_map.m_lblk - ee_block;
3614                         split_map.m_lblk = ee_block;
3615                         allocated = map->m_len;
3616                 }
3617         }
3618
3619         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3620                                 flags);
3621         if (err > 0)
3622                 err = 0;
3623 out:
3624         /* If we have gotten a failure, don't zero out status tree */
3625         if (!err) {
3626                 err = ext4_zeroout_es(inode, &zero_ex1);
3627                 if (!err)
3628                         err = ext4_zeroout_es(inode, &zero_ex2);
3629         }
3630         return err ? err : allocated;
3631 }
3632
3633 /*
3634  * This function is called by ext4_ext_map_blocks() from
3635  * ext4_get_blocks_dio_write() when DIO to write
3636  * to an unwritten extent.
3637  *
3638  * Writing to an unwritten extent may result in splitting the unwritten
3639  * extent into multiple initialized/unwritten extents (up to three)
3640  * There are three possibilities:
3641  *   a> There is no split required: Entire extent should be unwritten
3642  *   b> Splits in two extents: Write is happening at either end of the extent
3643  *   c> Splits in three extents: Somone is writing in middle of the extent
3644  *
3645  * This works the same way in the case of initialized -> unwritten conversion.
3646  *
3647  * One of more index blocks maybe needed if the extent tree grow after
3648  * the unwritten extent split. To prevent ENOSPC occur at the IO
3649  * complete, we need to split the unwritten extent before DIO submit
3650  * the IO. The unwritten extent called at this time will be split
3651  * into three unwritten extent(at most). After IO complete, the part
3652  * being filled will be convert to initialized by the end_io callback function
3653  * via ext4_convert_unwritten_extents().
3654  *
3655  * Returns the size of unwritten extent to be written on success.
3656  */
3657 static int ext4_split_convert_extents(handle_t *handle,
3658                                         struct inode *inode,
3659                                         struct ext4_map_blocks *map,
3660                                         struct ext4_ext_path **ppath,
3661                                         int flags)
3662 {
3663         struct ext4_ext_path *path = *ppath;
3664         ext4_lblk_t eof_block;
3665         ext4_lblk_t ee_block;
3666         struct ext4_extent *ex;
3667         unsigned int ee_len;
3668         int split_flag = 0, depth;
3669
3670         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3671                   __func__, inode->i_ino,
3672                   (unsigned long long)map->m_lblk, map->m_len);
3673
3674         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3675                 inode->i_sb->s_blocksize_bits;
3676         if (eof_block < map->m_lblk + map->m_len)
3677                 eof_block = map->m_lblk + map->m_len;
3678         /*
3679          * It is safe to convert extent to initialized via explicit
3680          * zeroout only if extent is fully insde i_size or new_size.
3681          */
3682         depth = ext_depth(inode);
3683         ex = path[depth].p_ext;
3684         ee_block = le32_to_cpu(ex->ee_block);
3685         ee_len = ext4_ext_get_actual_len(ex);
3686
3687         /* Convert to unwritten */
3688         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3689                 split_flag |= EXT4_EXT_DATA_VALID1;
3690         /* Convert to initialized */
3691         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3692                 split_flag |= ee_block + ee_len <= eof_block ?
3693                               EXT4_EXT_MAY_ZEROOUT : 0;
3694                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3695         }
3696         flags |= EXT4_GET_BLOCKS_PRE_IO;
3697         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3698 }
3699
3700 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3701                                                 struct inode *inode,
3702                                                 struct ext4_map_blocks *map,
3703                                                 struct ext4_ext_path **ppath)
3704 {
3705         struct ext4_ext_path *path = *ppath;
3706         struct ext4_extent *ex;
3707         ext4_lblk_t ee_block;
3708         unsigned int ee_len;
3709         int depth;
3710         int err = 0;
3711
3712         depth = ext_depth(inode);
3713         ex = path[depth].p_ext;
3714         ee_block = le32_to_cpu(ex->ee_block);
3715         ee_len = ext4_ext_get_actual_len(ex);
3716
3717         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3718                 "block %llu, max_blocks %u\n", inode->i_ino,
3719                   (unsigned long long)ee_block, ee_len);
3720
3721         /* If extent is larger than requested it is a clear sign that we still
3722          * have some extent state machine issues left. So extent_split is still
3723          * required.
3724          * TODO: Once all related issues will be fixed this situation should be
3725          * illegal.
3726          */
3727         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3728 #ifdef EXT4_DEBUG
3729                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3730                              " len %u; IO logical block %llu, len %u",
3731                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3732                              (unsigned long long)map->m_lblk, map->m_len);
3733 #endif
3734                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3735                                                  EXT4_GET_BLOCKS_CONVERT);
3736                 if (err < 0)
3737                         return err;
3738                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3739                 if (IS_ERR(path))
3740                         return PTR_ERR(path);
3741                 depth = ext_depth(inode);
3742                 ex = path[depth].p_ext;
3743         }
3744
3745         err = ext4_ext_get_access(handle, inode, path + depth);
3746         if (err)
3747                 goto out;
3748         /* first mark the extent as initialized */
3749         ext4_ext_mark_initialized(ex);
3750
3751         /* note: ext4_ext_correct_indexes() isn't needed here because
3752          * borders are not changed
3753          */
3754         ext4_ext_try_to_merge(handle, inode, path, ex);
3755
3756         /* Mark modified extent as dirty */
3757         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3758 out:
3759         ext4_ext_show_leaf(inode, path);
3760         return err;
3761 }
3762
3763 /*
3764  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3765  */
3766 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3767                               ext4_lblk_t lblk,
3768                               struct ext4_ext_path *path,
3769                               unsigned int len)
3770 {
3771         int i, depth;
3772         struct ext4_extent_header *eh;
3773         struct ext4_extent *last_ex;
3774
3775         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3776                 return 0;
3777
3778         depth = ext_depth(inode);
3779         eh = path[depth].p_hdr;
3780
3781         /*
3782          * We're going to remove EOFBLOCKS_FL entirely in future so we
3783          * do not care for this case anymore. Simply remove the flag
3784          * if there are no extents.
3785          */
3786         if (unlikely(!eh->eh_entries))
3787                 goto out;
3788         last_ex = EXT_LAST_EXTENT(eh);
3789         /*
3790          * We should clear the EOFBLOCKS_FL flag if we are writing the
3791          * last block in the last extent in the file.  We test this by
3792          * first checking to see if the caller to
3793          * ext4_ext_get_blocks() was interested in the last block (or
3794          * a block beyond the last block) in the current extent.  If
3795          * this turns out to be false, we can bail out from this
3796          * function immediately.
3797          */
3798         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3799             ext4_ext_get_actual_len(last_ex))
3800                 return 0;
3801         /*
3802          * If the caller does appear to be planning to write at or
3803          * beyond the end of the current extent, we then test to see
3804          * if the current extent is the last extent in the file, by
3805          * checking to make sure it was reached via the rightmost node
3806          * at each level of the tree.
3807          */
3808         for (i = depth-1; i >= 0; i--)
3809                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3810                         return 0;
3811 out:
3812         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3813         return ext4_mark_inode_dirty(handle, inode);
3814 }
3815
3816 /**
3817  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3818  *
3819  * Return 1 if there is a delalloc block in the range, otherwise 0.
3820  */
3821 int ext4_find_delalloc_range(struct inode *inode,
3822                              ext4_lblk_t lblk_start,
3823                              ext4_lblk_t lblk_end)
3824 {
3825         struct extent_status es;
3826
3827         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3828         if (es.es_len == 0)
3829                 return 0; /* there is no delay extent in this tree */
3830         else if (es.es_lblk <= lblk_start &&
3831                  lblk_start < es.es_lblk + es.es_len)
3832                 return 1;
3833         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3834                 return 1;
3835         else
3836                 return 0;
3837 }
3838
3839 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3840 {
3841         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3842         ext4_lblk_t lblk_start, lblk_end;
3843         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3844         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3845
3846         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3847 }
3848
3849 /**
3850  * Determines how many complete clusters (out of those specified by the 'map')
3851  * are under delalloc and were reserved quota for.
3852  * This function is called when we are writing out the blocks that were
3853  * originally written with their allocation delayed, but then the space was
3854  * allocated using fallocate() before the delayed allocation could be resolved.
3855  * The cases to look for are:
3856  * ('=' indicated delayed allocated blocks
3857  *  '-' indicates non-delayed allocated blocks)
3858  * (a) partial clusters towards beginning and/or end outside of allocated range
3859  *     are not delalloc'ed.
3860  *      Ex:
3861  *      |----c---=|====c====|====c====|===-c----|
3862  *               |++++++ allocated ++++++|
3863  *      ==> 4 complete clusters in above example
3864  *
3865  * (b) partial cluster (outside of allocated range) towards either end is
3866  *     marked for delayed allocation. In this case, we will exclude that
3867  *     cluster.
3868  *      Ex:
3869  *      |----====c========|========c========|
3870  *           |++++++ allocated ++++++|
3871  *      ==> 1 complete clusters in above example
3872  *
3873  *      Ex:
3874  *      |================c================|
3875  *            |++++++ allocated ++++++|
3876  *      ==> 0 complete clusters in above example
3877  *
3878  * The ext4_da_update_reserve_space will be called only if we
3879  * determine here that there were some "entire" clusters that span
3880  * this 'allocated' range.
3881  * In the non-bigalloc case, this function will just end up returning num_blks
3882  * without ever calling ext4_find_delalloc_range.
3883  */
3884 static unsigned int
3885 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3886                            unsigned int num_blks)
3887 {
3888         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3889         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3890         ext4_lblk_t lblk_from, lblk_to, c_offset;
3891         unsigned int allocated_clusters = 0;
3892
3893         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3894         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3895
3896         /* max possible clusters for this allocation */
3897         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3898
3899         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3900
3901         /* Check towards left side */
3902         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3903         if (c_offset) {
3904                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3905                 lblk_to = lblk_from + c_offset - 1;
3906
3907                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3908                         allocated_clusters--;
3909         }
3910
3911         /* Now check towards right. */
3912         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3913         if (allocated_clusters && c_offset) {
3914                 lblk_from = lblk_start + num_blks;
3915                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3916
3917                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3918                         allocated_clusters--;
3919         }
3920
3921         return allocated_clusters;
3922 }
3923
3924 static int
3925 convert_initialized_extent(handle_t *handle, struct inode *inode,
3926                            struct ext4_map_blocks *map,
3927                            struct ext4_ext_path **ppath,
3928                            unsigned int allocated)
3929 {
3930         struct ext4_ext_path *path = *ppath;
3931         struct ext4_extent *ex;
3932         ext4_lblk_t ee_block;
3933         unsigned int ee_len;
3934         int depth;
3935         int err = 0;
3936
3937         /*
3938          * Make sure that the extent is no bigger than we support with
3939          * unwritten extent
3940          */
3941         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3942                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3943
3944         depth = ext_depth(inode);
3945         ex = path[depth].p_ext;
3946         ee_block = le32_to_cpu(ex->ee_block);
3947         ee_len = ext4_ext_get_actual_len(ex);
3948
3949         ext_debug("%s: inode %lu, logical"
3950                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3951                   (unsigned long long)ee_block, ee_len);
3952
3953         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3954                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3955                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3956                 if (err < 0)
3957                         return err;
3958                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3959                 if (IS_ERR(path))
3960                         return PTR_ERR(path);
3961                 depth = ext_depth(inode);
3962                 ex = path[depth].p_ext;
3963                 if (!ex) {
3964                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3965                                          (unsigned long) map->m_lblk);
3966                         return -EFSCORRUPTED;
3967                 }
3968         }
3969
3970         err = ext4_ext_get_access(handle, inode, path + depth);
3971         if (err)
3972                 return err;
3973         /* first mark the extent as unwritten */
3974         ext4_ext_mark_unwritten(ex);
3975
3976         /* note: ext4_ext_correct_indexes() isn't needed here because
3977          * borders are not changed
3978          */
3979         ext4_ext_try_to_merge(handle, inode, path, ex);
3980
3981         /* Mark modified extent as dirty */
3982         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3983         if (err)
3984                 return err;
3985         ext4_ext_show_leaf(inode, path);
3986
3987         ext4_update_inode_fsync_trans(handle, inode, 1);
3988         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3989         if (err)
3990                 return err;
3991         map->m_flags |= EXT4_MAP_UNWRITTEN;
3992         if (allocated > map->m_len)
3993                 allocated = map->m_len;
3994         map->m_len = allocated;
3995         return allocated;
3996 }
3997
3998 static int
3999 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4000                         struct ext4_map_blocks *map,
4001                         struct ext4_ext_path **ppath, int flags,
4002                         unsigned int allocated, ext4_fsblk_t newblock)
4003 {
4004         struct ext4_ext_path *path = *ppath;
4005         int ret = 0;
4006         int err = 0;
4007
4008         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4009                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4010                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4011                   flags, allocated);
4012         ext4_ext_show_leaf(inode, path);
4013
4014         /*
4015          * When writing into unwritten space, we should not fail to
4016          * allocate metadata blocks for the new extent block if needed.
4017          */
4018         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4019
4020         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4021                                                     allocated, newblock);
4022
4023         /* get_block() before submit the IO, split the extent */
4024         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4025                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4026                                          flags | EXT4_GET_BLOCKS_CONVERT);
4027                 if (ret <= 0)
4028                         goto out;
4029                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4030                 goto out;
4031         }
4032         /* IO end_io complete, convert the filled extent to written */
4033         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4034                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4035                         if (allocated > map->m_len)
4036                                 allocated = map->m_len;
4037                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4038                                                  allocated);
4039                         if (err < 0)
4040                                 goto out2;
4041                 }
4042                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4043                                                            ppath);
4044                 if (ret >= 0) {
4045                         ext4_update_inode_fsync_trans(handle, inode, 1);
4046                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4047                                                  path, map->m_len);
4048                 } else
4049                         err = ret;
4050                 map->m_flags |= EXT4_MAP_MAPPED;
4051                 map->m_pblk = newblock;
4052                 if (allocated > map->m_len)
4053                         allocated = map->m_len;
4054                 map->m_len = allocated;
4055                 goto out2;
4056         }
4057         /* buffered IO case */
4058         /*
4059          * repeat fallocate creation request
4060          * we already have an unwritten extent
4061          */
4062         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4063                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4064                 goto map_out;
4065         }
4066
4067         /* buffered READ or buffered write_begin() lookup */
4068         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4069                 /*
4070                  * We have blocks reserved already.  We
4071                  * return allocated blocks so that delalloc
4072                  * won't do block reservation for us.  But
4073                  * the buffer head will be unmapped so that
4074                  * a read from the block returns 0s.
4075                  */
4076                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4077                 goto out1;
4078         }
4079
4080         /* buffered write, writepage time, convert*/
4081         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4082         if (ret >= 0)
4083                 ext4_update_inode_fsync_trans(handle, inode, 1);
4084 out:
4085         if (ret <= 0) {
4086                 err = ret;
4087                 goto out2;
4088         } else
4089                 allocated = ret;
4090         map->m_flags |= EXT4_MAP_NEW;
4091         /*
4092          * if we allocated more blocks than requested
4093          * we need to make sure we unmap the extra block
4094          * allocated. The actual needed block will get
4095          * unmapped later when we find the buffer_head marked
4096          * new.
4097          */
4098         if (allocated > map->m_len) {
4099                 clean_bdev_aliases(inode->i_sb->s_bdev, newblock + map->m_len,
4100                                    allocated - map->m_len);
4101                 allocated = map->m_len;
4102         }
4103         map->m_len = allocated;
4104
4105         /*
4106          * If we have done fallocate with the offset that is already
4107          * delayed allocated, we would have block reservation
4108          * and quota reservation done in the delayed write path.
4109          * But fallocate would have already updated quota and block
4110          * count for this offset. So cancel these reservation
4111          */
4112         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4113                 unsigned int reserved_clusters;
4114                 reserved_clusters = get_reserved_cluster_alloc(inode,
4115                                 map->m_lblk, map->m_len);
4116                 if (reserved_clusters)
4117                         ext4_da_update_reserve_space(inode,
4118                                                      reserved_clusters,
4119                                                      0);
4120         }
4121
4122 map_out:
4123         map->m_flags |= EXT4_MAP_MAPPED;
4124         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4125                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4126                                          map->m_len);
4127                 if (err < 0)
4128                         goto out2;
4129         }
4130 out1:
4131         if (allocated > map->m_len)
4132                 allocated = map->m_len;
4133         ext4_ext_show_leaf(inode, path);
4134         map->m_pblk = newblock;
4135         map->m_len = allocated;
4136 out2:
4137         return err ? err : allocated;
4138 }
4139
4140 /*
4141  * get_implied_cluster_alloc - check to see if the requested
4142  * allocation (in the map structure) overlaps with a cluster already
4143  * allocated in an extent.
4144  *      @sb     The filesystem superblock structure
4145  *      @map    The requested lblk->pblk mapping
4146  *      @ex     The extent structure which might contain an implied
4147  *                      cluster allocation
4148  *
4149  * This function is called by ext4_ext_map_blocks() after we failed to
4150  * find blocks that were already in the inode's extent tree.  Hence,
4151  * we know that the beginning of the requested region cannot overlap
4152  * the extent from the inode's extent tree.  There are three cases we
4153  * want to catch.  The first is this case:
4154  *
4155  *               |--- cluster # N--|
4156  *    |--- extent ---|  |---- requested region ---|
4157  *                      |==========|
4158  *
4159  * The second case that we need to test for is this one:
4160  *
4161  *   |--------- cluster # N ----------------|
4162  *         |--- requested region --|   |------- extent ----|
4163  *         |=======================|
4164  *
4165  * The third case is when the requested region lies between two extents
4166  * within the same cluster:
4167  *          |------------- cluster # N-------------|
4168  * |----- ex -----|                  |---- ex_right ----|
4169  *                  |------ requested region ------|
4170  *                  |================|
4171  *
4172  * In each of the above cases, we need to set the map->m_pblk and
4173  * map->m_len so it corresponds to the return the extent labelled as
4174  * "|====|" from cluster #N, since it is already in use for data in
4175  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4176  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4177  * as a new "allocated" block region.  Otherwise, we will return 0 and
4178  * ext4_ext_map_blocks() will then allocate one or more new clusters
4179  * by calling ext4_mb_new_blocks().
4180  */
4181 static int get_implied_cluster_alloc(struct super_block *sb,
4182                                      struct ext4_map_blocks *map,
4183                                      struct ext4_extent *ex,
4184                                      struct ext4_ext_path *path)
4185 {
4186         struct ext4_sb_info *sbi = EXT4_SB(sb);
4187         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4188         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4189         ext4_lblk_t rr_cluster_start;
4190         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4191         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4192         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4193
4194         /* The extent passed in that we are trying to match */
4195         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4196         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4197
4198         /* The requested region passed into ext4_map_blocks() */
4199         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4200
4201         if ((rr_cluster_start == ex_cluster_end) ||
4202             (rr_cluster_start == ex_cluster_start)) {
4203                 if (rr_cluster_start == ex_cluster_end)
4204                         ee_start += ee_len - 1;
4205                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4206                 map->m_len = min(map->m_len,
4207                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4208                 /*
4209                  * Check for and handle this case:
4210                  *
4211                  *   |--------- cluster # N-------------|
4212                  *                     |------- extent ----|
4213                  *         |--- requested region ---|
4214                  *         |===========|
4215                  */
4216
4217                 if (map->m_lblk < ee_block)
4218                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4219
4220                 /*
4221                  * Check for the case where there is already another allocated
4222                  * block to the right of 'ex' but before the end of the cluster.
4223                  *
4224                  *          |------------- cluster # N-------------|
4225                  * |----- ex -----|                  |---- ex_right ----|
4226                  *                  |------ requested region ------|
4227                  *                  |================|
4228                  */
4229                 if (map->m_lblk > ee_block) {
4230                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4231                         map->m_len = min(map->m_len, next - map->m_lblk);
4232                 }
4233
4234                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4235                 return 1;
4236         }
4237
4238         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4239         return 0;
4240 }
4241
4242
4243 /*
4244  * Block allocation/map/preallocation routine for extents based files
4245  *
4246  *
4247  * Need to be called with
4248  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4249  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4250  *
4251  * return > 0, number of of blocks already mapped/allocated
4252  *          if create == 0 and these are pre-allocated blocks
4253  *              buffer head is unmapped
4254  *          otherwise blocks are mapped
4255  *
4256  * return = 0, if plain look up failed (blocks have not been allocated)
4257  *          buffer head is unmapped
4258  *
4259  * return < 0, error case.
4260  */
4261 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4262                         struct ext4_map_blocks *map, int flags)
4263 {
4264         struct ext4_ext_path *path = NULL;
4265         struct ext4_extent newex, *ex, *ex2;
4266         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4267         ext4_fsblk_t newblock = 0;
4268         int free_on_err = 0, err = 0, depth, ret;
4269         unsigned int allocated = 0, offset = 0;
4270         unsigned int allocated_clusters = 0;
4271         struct ext4_allocation_request ar;
4272         ext4_lblk_t cluster_offset;
4273         bool map_from_cluster = false;
4274
4275         ext_debug("blocks %u/%u requested for inode %lu\n",
4276                   map->m_lblk, map->m_len, inode->i_ino);
4277         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4278
4279         /* find extent for this block */
4280         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4281         if (IS_ERR(path)) {
4282                 err = PTR_ERR(path);
4283                 path = NULL;
4284                 goto out2;
4285         }
4286
4287         depth = ext_depth(inode);
4288
4289         /*
4290          * consistent leaf must not be empty;
4291          * this situation is possible, though, _during_ tree modification;
4292          * this is why assert can't be put in ext4_find_extent()
4293          */
4294         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4295                 EXT4_ERROR_INODE(inode, "bad extent address "
4296                                  "lblock: %lu, depth: %d pblock %lld",
4297                                  (unsigned long) map->m_lblk, depth,
4298                                  path[depth].p_block);
4299                 err = -EFSCORRUPTED;
4300                 goto out2;
4301         }
4302
4303         ex = path[depth].p_ext;
4304         if (ex) {
4305                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4306                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4307                 unsigned short ee_len;
4308
4309
4310                 /*
4311                  * unwritten extents are treated as holes, except that
4312                  * we split out initialized portions during a write.
4313                  */
4314                 ee_len = ext4_ext_get_actual_len(ex);
4315
4316                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4317
4318                 /* if found extent covers block, simply return it */
4319                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4320                         newblock = map->m_lblk - ee_block + ee_start;
4321                         /* number of remaining blocks in the extent */
4322                         allocated = ee_len - (map->m_lblk - ee_block);
4323                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4324                                   ee_block, ee_len, newblock);
4325
4326                         /*
4327                          * If the extent is initialized check whether the
4328                          * caller wants to convert it to unwritten.
4329                          */
4330                         if ((!ext4_ext_is_unwritten(ex)) &&
4331                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4332                                 allocated = convert_initialized_extent(
4333                                                 handle, inode, map, &path,
4334                                                 allocated);
4335                                 goto out2;
4336                         } else if (!ext4_ext_is_unwritten(ex))
4337                                 goto out;
4338
4339                         ret = ext4_ext_handle_unwritten_extents(
4340                                 handle, inode, map, &path, flags,
4341                                 allocated, newblock);
4342                         if (ret < 0)
4343                                 err = ret;
4344                         else
4345                                 allocated = ret;
4346                         goto out2;
4347                 }
4348         }
4349
4350         /*
4351          * requested block isn't allocated yet;
4352          * we couldn't try to create block if create flag is zero
4353          */
4354         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4355                 ext4_lblk_t hole_start, hole_len;
4356
4357                 hole_start = map->m_lblk;
4358                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4359                 /*
4360                  * put just found gap into cache to speed up
4361                  * subsequent requests
4362                  */
4363                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4364
4365                 /* Update hole_len to reflect hole size after map->m_lblk */
4366                 if (hole_start != map->m_lblk)
4367                         hole_len -= map->m_lblk - hole_start;
4368                 map->m_pblk = 0;
4369                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4370
4371                 goto out2;
4372         }
4373
4374         /*
4375          * Okay, we need to do block allocation.
4376          */
4377         newex.ee_block = cpu_to_le32(map->m_lblk);
4378         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4379
4380         /*
4381          * If we are doing bigalloc, check to see if the extent returned
4382          * by ext4_find_extent() implies a cluster we can use.
4383          */
4384         if (cluster_offset && ex &&
4385             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4386                 ar.len = allocated = map->m_len;
4387                 newblock = map->m_pblk;
4388                 map_from_cluster = true;
4389                 goto got_allocated_blocks;
4390         }
4391
4392         /* find neighbour allocated blocks */
4393         ar.lleft = map->m_lblk;
4394         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4395         if (err)
4396                 goto out2;
4397         ar.lright = map->m_lblk;
4398         ex2 = NULL;
4399         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4400         if (err)
4401                 goto out2;
4402
4403         /* Check if the extent after searching to the right implies a
4404          * cluster we can use. */
4405         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4406             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4407                 ar.len = allocated = map->m_len;
4408                 newblock = map->m_pblk;
4409                 map_from_cluster = true;
4410                 goto got_allocated_blocks;
4411         }
4412
4413         /*
4414          * See if request is beyond maximum number of blocks we can have in
4415          * a single extent. For an initialized extent this limit is
4416          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4417          * EXT_UNWRITTEN_MAX_LEN.
4418          */
4419         if (map->m_len > EXT_INIT_MAX_LEN &&
4420             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4421                 map->m_len = EXT_INIT_MAX_LEN;
4422         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4423                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4424                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4425
4426         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4427         newex.ee_len = cpu_to_le16(map->m_len);
4428         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4429         if (err)
4430                 allocated = ext4_ext_get_actual_len(&newex);
4431         else
4432                 allocated = map->m_len;
4433
4434         /* allocate new block */
4435         ar.inode = inode;
4436         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4437         ar.logical = map->m_lblk;
4438         /*
4439          * We calculate the offset from the beginning of the cluster
4440          * for the logical block number, since when we allocate a
4441          * physical cluster, the physical block should start at the
4442          * same offset from the beginning of the cluster.  This is
4443          * needed so that future calls to get_implied_cluster_alloc()
4444          * work correctly.
4445          */
4446         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4447         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4448         ar.goal -= offset;
4449         ar.logical -= offset;
4450         if (S_ISREG(inode->i_mode))
4451                 ar.flags = EXT4_MB_HINT_DATA;
4452         else
4453                 /* disable in-core preallocation for non-regular files */
4454                 ar.flags = 0;
4455         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4456                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4457         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4458                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4459         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4460                 ar.flags |= EXT4_MB_USE_RESERVED;
4461         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4462         if (!newblock)
4463                 goto out2;
4464         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4465                   ar.goal, newblock, allocated);
4466         free_on_err = 1;
4467         allocated_clusters = ar.len;
4468         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4469         if (ar.len > allocated)
4470                 ar.len = allocated;
4471
4472 got_allocated_blocks:
4473         /* try to insert new extent into found leaf and return */
4474         ext4_ext_store_pblock(&newex, newblock + offset);
4475         newex.ee_len = cpu_to_le16(ar.len);
4476         /* Mark unwritten */
4477         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4478                 ext4_ext_mark_unwritten(&newex);
4479                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4480         }
4481
4482         err = 0;
4483         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4484                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4485                                          path, ar.len);
4486         if (!err)
4487                 err = ext4_ext_insert_extent(handle, inode, &path,
4488                                              &newex, flags);
4489
4490         if (err && free_on_err) {
4491                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4492                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4493                 /* free data blocks we just allocated */
4494                 /* not a good idea to call discard here directly,
4495                  * but otherwise we'd need to call it every free() */
4496                 ext4_discard_preallocations(inode);
4497                 ext4_free_blocks(handle, inode, NULL, newblock,
4498                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4499                 goto out2;
4500         }
4501
4502         /* previous routine could use block we allocated */
4503         newblock = ext4_ext_pblock(&newex);
4504         allocated = ext4_ext_get_actual_len(&newex);
4505         if (allocated > map->m_len)
4506                 allocated = map->m_len;
4507         map->m_flags |= EXT4_MAP_NEW;
4508
4509         /*
4510          * Update reserved blocks/metadata blocks after successful
4511          * block allocation which had been deferred till now.
4512          */
4513         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4514                 unsigned int reserved_clusters;
4515                 /*
4516                  * Check how many clusters we had reserved this allocated range
4517                  */
4518                 reserved_clusters = get_reserved_cluster_alloc(inode,
4519                                                 map->m_lblk, allocated);
4520                 if (!map_from_cluster) {
4521                         BUG_ON(allocated_clusters < reserved_clusters);
4522                         if (reserved_clusters < allocated_clusters) {
4523                                 struct ext4_inode_info *ei = EXT4_I(inode);
4524                                 int reservation = allocated_clusters -
4525                                                   reserved_clusters;
4526                                 /*
4527                                  * It seems we claimed few clusters outside of
4528                                  * the range of this allocation. We should give
4529                                  * it back to the reservation pool. This can
4530                                  * happen in the following case:
4531                                  *
4532                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4533                                  *   cluster has 4 blocks. Thus, the clusters
4534                                  *   are [0-3],[4-7],[8-11]...
4535                                  * * First comes delayed allocation write for
4536                                  *   logical blocks 10 & 11. Since there were no
4537                                  *   previous delayed allocated blocks in the
4538                                  *   range [8-11], we would reserve 1 cluster
4539                                  *   for this write.
4540                                  * * Next comes write for logical blocks 3 to 8.
4541                                  *   In this case, we will reserve 2 clusters
4542                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4543                                  *   that range has a delayed allocated blocks.
4544                                  *   Thus total reserved clusters now becomes 3.
4545                                  * * Now, during the delayed allocation writeout
4546                                  *   time, we will first write blocks [3-8] and
4547                                  *   allocate 3 clusters for writing these
4548                                  *   blocks. Also, we would claim all these
4549                                  *   three clusters above.
4550                                  * * Now when we come here to writeout the
4551                                  *   blocks [10-11], we would expect to claim
4552                                  *   the reservation of 1 cluster we had made
4553                                  *   (and we would claim it since there are no
4554                                  *   more delayed allocated blocks in the range
4555                                  *   [8-11]. But our reserved cluster count had
4556                                  *   already gone to 0.
4557                                  *
4558                                  *   Thus, at the step 4 above when we determine
4559                                  *   that there are still some unwritten delayed
4560                                  *   allocated blocks outside of our current
4561                                  *   block range, we should increment the
4562                                  *   reserved clusters count so that when the
4563                                  *   remaining blocks finally gets written, we
4564                                  *   could claim them.
4565                                  */
4566                                 dquot_reserve_block(inode,
4567                                                 EXT4_C2B(sbi, reservation));
4568                                 spin_lock(&ei->i_block_reservation_lock);
4569                                 ei->i_reserved_data_blocks += reservation;
4570                                 spin_unlock(&ei->i_block_reservation_lock);
4571                         }
4572                         /*
4573                          * We will claim quota for all newly allocated blocks.
4574                          * We're updating the reserved space *after* the
4575                          * correction above so we do not accidentally free
4576                          * all the metadata reservation because we might
4577                          * actually need it later on.
4578                          */
4579                         ext4_da_update_reserve_space(inode, allocated_clusters,
4580                                                         1);
4581                 }
4582         }
4583
4584         /*
4585          * Cache the extent and update transaction to commit on fdatasync only
4586          * when it is _not_ an unwritten extent.
4587          */
4588         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4589                 ext4_update_inode_fsync_trans(handle, inode, 1);
4590         else
4591                 ext4_update_inode_fsync_trans(handle, inode, 0);
4592 out:
4593         if (allocated > map->m_len)
4594                 allocated = map->m_len;
4595         ext4_ext_show_leaf(inode, path);
4596         map->m_flags |= EXT4_MAP_MAPPED;
4597         map->m_pblk = newblock;
4598         map->m_len = allocated;
4599 out2:
4600         ext4_ext_drop_refs(path);
4601         kfree(path);
4602
4603         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4604                                        err ? err : allocated);
4605         return err ? err : allocated;
4606 }
4607
4608 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4609 {
4610         struct super_block *sb = inode->i_sb;
4611         ext4_lblk_t last_block;
4612         int err = 0;
4613
4614         /*
4615          * TODO: optimization is possible here.
4616          * Probably we need not scan at all,
4617          * because page truncation is enough.
4618          */
4619
4620         /* we have to know where to truncate from in crash case */
4621         EXT4_I(inode)->i_disksize = inode->i_size;
4622         err = ext4_mark_inode_dirty(handle, inode);
4623         if (err)
4624                 return err;
4625
4626         last_block = (inode->i_size + sb->s_blocksize - 1)
4627                         >> EXT4_BLOCK_SIZE_BITS(sb);
4628 retry:
4629         err = ext4_es_remove_extent(inode, last_block,
4630                                     EXT_MAX_BLOCKS - last_block);
4631         if (err == -ENOMEM) {
4632                 cond_resched();
4633                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4634                 goto retry;
4635         }
4636         if (err)
4637                 return err;
4638         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4639 }
4640
4641 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4642                                   ext4_lblk_t len, loff_t new_size,
4643                                   int flags)
4644 {
4645         struct inode *inode = file_inode(file);
4646         handle_t *handle;
4647         int ret = 0;
4648         int ret2 = 0;
4649         int retries = 0;
4650         int depth = 0;
4651         struct ext4_map_blocks map;
4652         unsigned int credits;
4653         loff_t epos;
4654
4655         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4656         map.m_lblk = offset;
4657         map.m_len = len;
4658         /*
4659          * Don't normalize the request if it can fit in one extent so
4660          * that it doesn't get unnecessarily split into multiple
4661          * extents.
4662          */
4663         if (len <= EXT_UNWRITTEN_MAX_LEN)
4664                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4665
4666         /*
4667          * credits to insert 1 extent into extent tree
4668          */
4669         credits = ext4_chunk_trans_blocks(inode, len);
4670         depth = ext_depth(inode);
4671
4672 retry:
4673         while (ret >= 0 && len) {
4674                 /*
4675                  * Recalculate credits when extent tree depth changes.
4676                  */
4677                 if (depth != ext_depth(inode)) {
4678                         credits = ext4_chunk_trans_blocks(inode, len);
4679                         depth = ext_depth(inode);
4680                 }
4681
4682                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4683                                             credits);
4684                 if (IS_ERR(handle)) {
4685                         ret = PTR_ERR(handle);
4686                         break;
4687                 }
4688                 ret = ext4_map_blocks(handle, inode, &map, flags);
4689                 if (ret <= 0) {
4690                         ext4_debug("inode #%lu: block %u: len %u: "
4691                                    "ext4_ext_map_blocks returned %d",
4692                                    inode->i_ino, map.m_lblk,
4693                                    map.m_len, ret);
4694                         ext4_mark_inode_dirty(handle, inode);
4695                         ret2 = ext4_journal_stop(handle);
4696                         break;
4697                 }
4698                 map.m_lblk += ret;
4699                 map.m_len = len = len - ret;
4700                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4701                 inode->i_ctime = current_time(inode);
4702                 if (new_size) {
4703                         if (epos > new_size)
4704                                 epos = new_size;
4705                         if (ext4_update_inode_size(inode, epos) & 0x1)
4706                                 inode->i_mtime = inode->i_ctime;
4707                 } else {
4708                         if (epos > inode->i_size)
4709                                 ext4_set_inode_flag(inode,
4710                                                     EXT4_INODE_EOFBLOCKS);
4711                 }
4712                 ext4_mark_inode_dirty(handle, inode);
4713                 ext4_update_inode_fsync_trans(handle, inode, 1);
4714                 ret2 = ext4_journal_stop(handle);
4715                 if (ret2)
4716                         break;
4717         }
4718         if (ret == -ENOSPC &&
4719                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4720                 ret = 0;
4721                 goto retry;
4722         }
4723
4724         return ret > 0 ? ret2 : ret;
4725 }
4726
4727 static long ext4_zero_range(struct file *file, loff_t offset,
4728                             loff_t len, int mode)
4729 {
4730         struct inode *inode = file_inode(file);
4731         handle_t *handle = NULL;
4732         unsigned int max_blocks;
4733         loff_t new_size = 0;
4734         int ret = 0;
4735         int flags;
4736         int credits;
4737         int partial_begin, partial_end;
4738         loff_t start, end;
4739         ext4_lblk_t lblk;
4740         unsigned int blkbits = inode->i_blkbits;
4741
4742         trace_ext4_zero_range(inode, offset, len, mode);
4743
4744         if (!S_ISREG(inode->i_mode))
4745                 return -EINVAL;
4746
4747         /* Call ext4_force_commit to flush all data in case of data=journal. */
4748         if (ext4_should_journal_data(inode)) {
4749                 ret = ext4_force_commit(inode->i_sb);
4750                 if (ret)
4751                         return ret;
4752         }
4753
4754         /*
4755          * Round up offset. This is not fallocate, we neet to zero out
4756          * blocks, so convert interior block aligned part of the range to
4757          * unwritten and possibly manually zero out unaligned parts of the
4758          * range.
4759          */
4760         start = round_up(offset, 1 << blkbits);
4761         end = round_down((offset + len), 1 << blkbits);
4762
4763         if (start < offset || end > offset + len)
4764                 return -EINVAL;
4765         partial_begin = offset & ((1 << blkbits) - 1);
4766         partial_end = (offset + len) & ((1 << blkbits) - 1);
4767
4768         lblk = start >> blkbits;
4769         max_blocks = (end >> blkbits);
4770         if (max_blocks < lblk)
4771                 max_blocks = 0;
4772         else
4773                 max_blocks -= lblk;
4774
4775         inode_lock(inode);
4776
4777         /*
4778          * Indirect files do not support unwritten extnets
4779          */
4780         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4781                 ret = -EOPNOTSUPP;
4782                 goto out_mutex;
4783         }
4784
4785         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4786             (offset + len > i_size_read(inode) ||
4787              offset + len > EXT4_I(inode)->i_disksize)) {
4788                 new_size = offset + len;
4789                 ret = inode_newsize_ok(inode, new_size);
4790                 if (ret)
4791                         goto out_mutex;
4792         }
4793
4794         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4795         if (mode & FALLOC_FL_KEEP_SIZE)
4796                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4797
4798         /* Wait all existing dio workers, newcomers will block on i_mutex */
4799         inode_dio_wait(inode);
4800
4801         /* Preallocate the range including the unaligned edges */
4802         if (partial_begin || partial_end) {
4803                 ret = ext4_alloc_file_blocks(file,
4804                                 round_down(offset, 1 << blkbits) >> blkbits,
4805                                 (round_up((offset + len), 1 << blkbits) -
4806                                  round_down(offset, 1 << blkbits)) >> blkbits,
4807                                 new_size, flags);
4808                 if (ret)
4809                         goto out_mutex;
4810
4811         }
4812
4813         /* Zero range excluding the unaligned edges */
4814         if (max_blocks > 0) {
4815                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4816                           EXT4_EX_NOCACHE);
4817
4818                 /*
4819                  * Prevent page faults from reinstantiating pages we have
4820                  * released from page cache.
4821                  */
4822                 down_write(&EXT4_I(inode)->i_mmap_sem);
4823                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4824                 if (ret) {
4825                         up_write(&EXT4_I(inode)->i_mmap_sem);
4826                         goto out_mutex;
4827                 }
4828                 /* Now release the pages and zero block aligned part of pages */
4829                 truncate_pagecache_range(inode, start, end - 1);
4830                 inode->i_mtime = inode->i_ctime = current_time(inode);
4831
4832                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4833                                              flags);
4834                 up_write(&EXT4_I(inode)->i_mmap_sem);
4835                 if (ret)
4836                         goto out_mutex;
4837         }
4838         if (!partial_begin && !partial_end)
4839                 goto out_mutex;
4840
4841         /*
4842          * In worst case we have to writeout two nonadjacent unwritten
4843          * blocks and update the inode
4844          */
4845         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4846         if (ext4_should_journal_data(inode))
4847                 credits += 2;
4848         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4849         if (IS_ERR(handle)) {
4850                 ret = PTR_ERR(handle);
4851                 ext4_std_error(inode->i_sb, ret);
4852                 goto out_mutex;
4853         }
4854
4855         inode->i_mtime = inode->i_ctime = current_time(inode);
4856         if (new_size) {
4857                 ext4_update_inode_size(inode, new_size);
4858         } else {
4859                 /*
4860                 * Mark that we allocate beyond EOF so the subsequent truncate
4861                 * can proceed even if the new size is the same as i_size.
4862                 */
4863                 if ((offset + len) > i_size_read(inode))
4864                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4865         }
4866         ext4_mark_inode_dirty(handle, inode);
4867
4868         /* Zero out partial block at the edges of the range */
4869         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4870         if (ret >= 0)
4871                 ext4_update_inode_fsync_trans(handle, inode, 1);
4872
4873         if (file->f_flags & O_SYNC)
4874                 ext4_handle_sync(handle);
4875
4876         ext4_journal_stop(handle);
4877 out_mutex:
4878         inode_unlock(inode);
4879         return ret;
4880 }
4881
4882 /*
4883  * preallocate space for a file. This implements ext4's fallocate file
4884  * operation, which gets called from sys_fallocate system call.
4885  * For block-mapped files, posix_fallocate should fall back to the method
4886  * of writing zeroes to the required new blocks (the same behavior which is
4887  * expected for file systems which do not support fallocate() system call).
4888  */
4889 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4890 {
4891         struct inode *inode = file_inode(file);
4892         loff_t new_size = 0;
4893         unsigned int max_blocks;
4894         int ret = 0;
4895         int flags;
4896         ext4_lblk_t lblk;
4897         unsigned int blkbits = inode->i_blkbits;
4898
4899         /*
4900          * Encrypted inodes can't handle collapse range or insert
4901          * range since we would need to re-encrypt blocks with a
4902          * different IV or XTS tweak (which are based on the logical
4903          * block number).
4904          *
4905          * XXX It's not clear why zero range isn't working, but we'll
4906          * leave it disabled for encrypted inodes for now.  This is a
4907          * bug we should fix....
4908          */
4909         if (ext4_encrypted_inode(inode) &&
4910             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4911                      FALLOC_FL_ZERO_RANGE)))
4912                 return -EOPNOTSUPP;
4913
4914         /* Return error if mode is not supported */
4915         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4916                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4917                      FALLOC_FL_INSERT_RANGE))
4918                 return -EOPNOTSUPP;
4919
4920         if (mode & FALLOC_FL_PUNCH_HOLE)
4921                 return ext4_punch_hole(inode, offset, len);
4922
4923         ret = ext4_convert_inline_data(inode);
4924         if (ret)
4925                 return ret;
4926
4927         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4928                 return ext4_collapse_range(inode, offset, len);
4929
4930         if (mode & FALLOC_FL_INSERT_RANGE)
4931                 return ext4_insert_range(inode, offset, len);
4932
4933         if (mode & FALLOC_FL_ZERO_RANGE)
4934                 return ext4_zero_range(file, offset, len, mode);
4935
4936         trace_ext4_fallocate_enter(inode, offset, len, mode);
4937         lblk = offset >> blkbits;
4938
4939         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4940         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4941         if (mode & FALLOC_FL_KEEP_SIZE)
4942                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4943
4944         inode_lock(inode);
4945
4946         /*
4947          * We only support preallocation for extent-based files only
4948          */
4949         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4950                 ret = -EOPNOTSUPP;
4951                 goto out;
4952         }
4953
4954         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4955             (offset + len > i_size_read(inode) ||
4956              offset + len > EXT4_I(inode)->i_disksize)) {
4957                 new_size = offset + len;
4958                 ret = inode_newsize_ok(inode, new_size);
4959                 if (ret)
4960                         goto out;
4961         }
4962
4963         /* Wait all existing dio workers, newcomers will block on i_mutex */
4964         inode_dio_wait(inode);
4965
4966         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
4967         if (ret)
4968                 goto out;
4969
4970         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4971                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4972                                                 EXT4_I(inode)->i_sync_tid);
4973         }
4974 out:
4975         inode_unlock(inode);
4976         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4977         return ret;
4978 }
4979
4980 /*
4981  * This function convert a range of blocks to written extents
4982  * The caller of this function will pass the start offset and the size.
4983  * all unwritten extents within this range will be converted to
4984  * written extents.
4985  *
4986  * This function is called from the direct IO end io call back
4987  * function, to convert the fallocated extents after IO is completed.
4988  * Returns 0 on success.
4989  */
4990 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4991                                    loff_t offset, ssize_t len)
4992 {
4993         unsigned int max_blocks;
4994         int ret = 0;
4995         int ret2 = 0;
4996         struct ext4_map_blocks map;
4997         unsigned int credits, blkbits = inode->i_blkbits;
4998
4999         map.m_lblk = offset >> blkbits;
5000         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
5001
5002         /*
5003          * This is somewhat ugly but the idea is clear: When transaction is
5004          * reserved, everything goes into it. Otherwise we rather start several
5005          * smaller transactions for conversion of each extent separately.
5006          */
5007         if (handle) {
5008                 handle = ext4_journal_start_reserved(handle,
5009                                                      EXT4_HT_EXT_CONVERT);
5010                 if (IS_ERR(handle))
5011                         return PTR_ERR(handle);
5012                 credits = 0;
5013         } else {
5014                 /*
5015                  * credits to insert 1 extent into extent tree
5016                  */
5017                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5018         }
5019         while (ret >= 0 && ret < max_blocks) {
5020                 map.m_lblk += ret;
5021                 map.m_len = (max_blocks -= ret);
5022                 if (credits) {
5023                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5024                                                     credits);
5025                         if (IS_ERR(handle)) {
5026                                 ret = PTR_ERR(handle);
5027                                 break;
5028                         }
5029                 }
5030                 ret = ext4_map_blocks(handle, inode, &map,
5031                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5032                 if (ret <= 0)
5033                         ext4_warning(inode->i_sb,
5034                                      "inode #%lu: block %u: len %u: "
5035                                      "ext4_ext_map_blocks returned %d",
5036                                      inode->i_ino, map.m_lblk,
5037                                      map.m_len, ret);
5038                 ext4_mark_inode_dirty(handle, inode);
5039                 if (credits)
5040                         ret2 = ext4_journal_stop(handle);
5041                 if (ret <= 0 || ret2)
5042                         break;
5043         }
5044         if (!credits)
5045                 ret2 = ext4_journal_stop(handle);
5046         return ret > 0 ? ret2 : ret;
5047 }
5048
5049 /*
5050  * If newes is not existing extent (newes->ec_pblk equals zero) find
5051  * delayed extent at start of newes and update newes accordingly and
5052  * return start of the next delayed extent.
5053  *
5054  * If newes is existing extent (newes->ec_pblk is not equal zero)
5055  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5056  * extent found. Leave newes unmodified.
5057  */
5058 static int ext4_find_delayed_extent(struct inode *inode,
5059                                     struct extent_status *newes)
5060 {
5061         struct extent_status es;
5062         ext4_lblk_t block, next_del;
5063
5064         if (newes->es_pblk == 0) {
5065                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5066                                 newes->es_lblk + newes->es_len - 1, &es);
5067
5068                 /*
5069                  * No extent in extent-tree contains block @newes->es_pblk,
5070                  * then the block may stay in 1)a hole or 2)delayed-extent.
5071                  */
5072                 if (es.es_len == 0)
5073                         /* A hole found. */
5074                         return 0;
5075
5076                 if (es.es_lblk > newes->es_lblk) {
5077                         /* A hole found. */
5078                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5079                                             newes->es_len);
5080                         return 0;
5081                 }
5082
5083                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5084         }
5085
5086         block = newes->es_lblk + newes->es_len;
5087         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5088         if (es.es_len == 0)
5089                 next_del = EXT_MAX_BLOCKS;
5090         else
5091                 next_del = es.es_lblk;
5092
5093         return next_del;
5094 }
5095 /* fiemap flags we can handle specified here */
5096 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5097
5098 static int ext4_xattr_fiemap(struct inode *inode,
5099                                 struct fiemap_extent_info *fieinfo)
5100 {
5101         __u64 physical = 0;
5102         __u64 length;
5103         __u32 flags = FIEMAP_EXTENT_LAST;
5104         int blockbits = inode->i_sb->s_blocksize_bits;
5105         int error = 0;
5106
5107         /* in-inode? */
5108         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5109                 struct ext4_iloc iloc;
5110                 int offset;     /* offset of xattr in inode */
5111
5112                 error = ext4_get_inode_loc(inode, &iloc);
5113                 if (error)
5114                         return error;
5115                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5116                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5117                                 EXT4_I(inode)->i_extra_isize;
5118                 physical += offset;
5119                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5120                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5121                 brelse(iloc.bh);
5122         } else { /* external block */
5123                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5124                 length = inode->i_sb->s_blocksize;
5125         }
5126
5127         if (physical)
5128                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5129                                                 length, flags);
5130         return (error < 0 ? error : 0);
5131 }
5132
5133 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5134                 __u64 start, __u64 len)
5135 {
5136         ext4_lblk_t start_blk;
5137         int error = 0;
5138
5139         if (ext4_has_inline_data(inode)) {
5140                 int has_inline = 1;
5141
5142                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5143                                                 start, len);
5144
5145                 if (has_inline)
5146                         return error;
5147         }
5148
5149         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5150                 error = ext4_ext_precache(inode);
5151                 if (error)
5152                         return error;
5153         }
5154
5155         /* fallback to generic here if not in extents fmt */
5156         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5157                 return generic_block_fiemap(inode, fieinfo, start, len,
5158                         ext4_get_block);
5159
5160         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5161                 return -EBADR;
5162
5163         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5164                 error = ext4_xattr_fiemap(inode, fieinfo);
5165         } else {
5166                 ext4_lblk_t len_blks;
5167                 __u64 last_blk;
5168
5169                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5170                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5171                 if (last_blk >= EXT_MAX_BLOCKS)
5172                         last_blk = EXT_MAX_BLOCKS-1;
5173                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5174
5175                 /*
5176                  * Walk the extent tree gathering extent information
5177                  * and pushing extents back to the user.
5178                  */
5179                 error = ext4_fill_fiemap_extents(inode, start_blk,
5180                                                  len_blks, fieinfo);
5181         }
5182         return error;
5183 }
5184
5185 /*
5186  * ext4_access_path:
5187  * Function to access the path buffer for marking it dirty.
5188  * It also checks if there are sufficient credits left in the journal handle
5189  * to update path.
5190  */
5191 static int
5192 ext4_access_path(handle_t *handle, struct inode *inode,
5193                 struct ext4_ext_path *path)
5194 {
5195         int credits, err;
5196
5197         if (!ext4_handle_valid(handle))
5198                 return 0;
5199
5200         /*
5201          * Check if need to extend journal credits
5202          * 3 for leaf, sb, and inode plus 2 (bmap and group
5203          * descriptor) for each block group; assume two block
5204          * groups
5205          */
5206         if (handle->h_buffer_credits < 7) {
5207                 credits = ext4_writepage_trans_blocks(inode);
5208                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5209                 /* EAGAIN is success */
5210                 if (err && err != -EAGAIN)
5211                         return err;
5212         }
5213
5214         err = ext4_ext_get_access(handle, inode, path);
5215         return err;
5216 }
5217
5218 /*
5219  * ext4_ext_shift_path_extents:
5220  * Shift the extents of a path structure lying between path[depth].p_ext
5221  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5222  * if it is right shift or left shift operation.
5223  */
5224 static int
5225 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5226                             struct inode *inode, handle_t *handle,
5227                             enum SHIFT_DIRECTION SHIFT)
5228 {
5229         int depth, err = 0;
5230         struct ext4_extent *ex_start, *ex_last;
5231         bool update = 0;
5232         depth = path->p_depth;
5233
5234         while (depth >= 0) {
5235                 if (depth == path->p_depth) {
5236                         ex_start = path[depth].p_ext;
5237                         if (!ex_start)
5238                                 return -EFSCORRUPTED;
5239
5240                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5241
5242                         err = ext4_access_path(handle, inode, path + depth);
5243                         if (err)
5244                                 goto out;
5245
5246                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5247                                 update = 1;
5248
5249                         while (ex_start <= ex_last) {
5250                                 if (SHIFT == SHIFT_LEFT) {
5251                                         le32_add_cpu(&ex_start->ee_block,
5252                                                 -shift);
5253                                         /* Try to merge to the left. */
5254                                         if ((ex_start >
5255                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5256                                             &&
5257                                             ext4_ext_try_to_merge_right(inode,
5258                                             path, ex_start - 1))
5259                                                 ex_last--;
5260                                         else
5261                                                 ex_start++;
5262                                 } else {
5263                                         le32_add_cpu(&ex_last->ee_block, shift);
5264                                         ext4_ext_try_to_merge_right(inode, path,
5265                                                 ex_last);
5266                                         ex_last--;
5267                                 }
5268                         }
5269                         err = ext4_ext_dirty(handle, inode, path + depth);
5270                         if (err)
5271                                 goto out;
5272
5273                         if (--depth < 0 || !update)
5274                                 break;
5275                 }
5276
5277                 /* Update index too */
5278                 err = ext4_access_path(handle, inode, path + depth);
5279                 if (err)
5280                         goto out;
5281
5282                 if (SHIFT == SHIFT_LEFT)
5283                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5284                 else
5285                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5286                 err = ext4_ext_dirty(handle, inode, path + depth);
5287                 if (err)
5288                         goto out;
5289
5290                 /* we are done if current index is not a starting index */
5291                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5292                         break;
5293
5294                 depth--;
5295         }
5296
5297 out:
5298         return err;
5299 }
5300
5301 /*
5302  * ext4_ext_shift_extents:
5303  * All the extents which lies in the range from @start to the last allocated
5304  * block for the @inode are shifted either towards left or right (depending
5305  * upon @SHIFT) by @shift blocks.
5306  * On success, 0 is returned, error otherwise.
5307  */
5308 static int
5309 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5310                        ext4_lblk_t start, ext4_lblk_t shift,
5311                        enum SHIFT_DIRECTION SHIFT)
5312 {
5313         struct ext4_ext_path *path;
5314         int ret = 0, depth;
5315         struct ext4_extent *extent;
5316         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5317
5318         /* Let path point to the last extent */
5319         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5320                                 EXT4_EX_NOCACHE);
5321         if (IS_ERR(path))
5322                 return PTR_ERR(path);
5323
5324         depth = path->p_depth;
5325         extent = path[depth].p_ext;
5326         if (!extent)
5327                 goto out;
5328
5329         stop = le32_to_cpu(extent->ee_block);
5330
5331        /*
5332         * For left shifts, make sure the hole on the left is big enough to
5333         * accommodate the shift.  For right shifts, make sure the last extent
5334         * won't be shifted beyond EXT_MAX_BLOCKS.
5335         */
5336         if (SHIFT == SHIFT_LEFT) {
5337                 path = ext4_find_extent(inode, start - 1, &path,
5338                                         EXT4_EX_NOCACHE);
5339                 if (IS_ERR(path))
5340                         return PTR_ERR(path);
5341                 depth = path->p_depth;
5342                 extent =  path[depth].p_ext;
5343                 if (extent) {
5344                         ex_start = le32_to_cpu(extent->ee_block);
5345                         ex_end = le32_to_cpu(extent->ee_block) +
5346                                 ext4_ext_get_actual_len(extent);
5347                 } else {
5348                         ex_start = 0;
5349                         ex_end = 0;
5350                 }
5351
5352                 if ((start == ex_start && shift > ex_start) ||
5353                     (shift > start - ex_end)) {
5354                         ret = -EINVAL;
5355                         goto out;
5356                 }
5357         } else {
5358                 if (shift > EXT_MAX_BLOCKS -
5359                     (stop + ext4_ext_get_actual_len(extent))) {
5360                         ret = -EINVAL;
5361                         goto out;
5362                 }
5363         }
5364
5365         /*
5366          * In case of left shift, iterator points to start and it is increased
5367          * till we reach stop. In case of right shift, iterator points to stop
5368          * and it is decreased till we reach start.
5369          */
5370         if (SHIFT == SHIFT_LEFT)
5371                 iterator = &start;
5372         else
5373                 iterator = &stop;
5374
5375         /*
5376          * Its safe to start updating extents.  Start and stop are unsigned, so
5377          * in case of right shift if extent with 0 block is reached, iterator
5378          * becomes NULL to indicate the end of the loop.
5379          */
5380         while (iterator && start <= stop) {
5381                 path = ext4_find_extent(inode, *iterator, &path,
5382                                         EXT4_EX_NOCACHE);
5383                 if (IS_ERR(path))
5384                         return PTR_ERR(path);
5385                 depth = path->p_depth;
5386                 extent = path[depth].p_ext;
5387                 if (!extent) {
5388                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5389                                          (unsigned long) *iterator);
5390                         return -EFSCORRUPTED;
5391                 }
5392                 if (SHIFT == SHIFT_LEFT && *iterator >
5393                     le32_to_cpu(extent->ee_block)) {
5394                         /* Hole, move to the next extent */
5395                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5396                                 path[depth].p_ext++;
5397                         } else {
5398                                 *iterator = ext4_ext_next_allocated_block(path);
5399                                 continue;
5400                         }
5401                 }
5402
5403                 if (SHIFT == SHIFT_LEFT) {
5404                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5405                         *iterator = le32_to_cpu(extent->ee_block) +
5406                                         ext4_ext_get_actual_len(extent);
5407                 } else {
5408                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5409                         if (le32_to_cpu(extent->ee_block) > 0)
5410                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5411                         else
5412                                 /* Beginning is reached, end of the loop */
5413                                 iterator = NULL;
5414                         /* Update path extent in case we need to stop */
5415                         while (le32_to_cpu(extent->ee_block) < start)
5416                                 extent++;
5417                         path[depth].p_ext = extent;
5418                 }
5419                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5420                                 handle, SHIFT);
5421                 if (ret)
5422                         break;
5423         }
5424 out:
5425         ext4_ext_drop_refs(path);
5426         kfree(path);
5427         return ret;
5428 }
5429
5430 /*
5431  * ext4_collapse_range:
5432  * This implements the fallocate's collapse range functionality for ext4
5433  * Returns: 0 and non-zero on error.
5434  */
5435 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5436 {
5437         struct super_block *sb = inode->i_sb;
5438         ext4_lblk_t punch_start, punch_stop;
5439         handle_t *handle;
5440         unsigned int credits;
5441         loff_t new_size, ioffset;
5442         int ret;
5443
5444         /*
5445          * We need to test this early because xfstests assumes that a
5446          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5447          * system does not support collapse range.
5448          */
5449         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5450                 return -EOPNOTSUPP;
5451
5452         /* Collapse range works only on fs block size aligned offsets. */
5453         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5454             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5455                 return -EINVAL;
5456
5457         if (!S_ISREG(inode->i_mode))
5458                 return -EINVAL;
5459
5460         trace_ext4_collapse_range(inode, offset, len);
5461
5462         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5463         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5464
5465         /* Call ext4_force_commit to flush all data in case of data=journal. */
5466         if (ext4_should_journal_data(inode)) {
5467                 ret = ext4_force_commit(inode->i_sb);
5468                 if (ret)
5469                         return ret;
5470         }
5471
5472         inode_lock(inode);
5473         /*
5474          * There is no need to overlap collapse range with EOF, in which case
5475          * it is effectively a truncate operation
5476          */
5477         if (offset + len >= i_size_read(inode)) {
5478                 ret = -EINVAL;
5479                 goto out_mutex;
5480         }
5481
5482         /* Currently just for extent based files */
5483         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5484                 ret = -EOPNOTSUPP;
5485                 goto out_mutex;
5486         }
5487
5488         /* Wait for existing dio to complete */
5489         inode_dio_wait(inode);
5490
5491         /*
5492          * Prevent page faults from reinstantiating pages we have released from
5493          * page cache.
5494          */
5495         down_write(&EXT4_I(inode)->i_mmap_sem);
5496         /*
5497          * Need to round down offset to be aligned with page size boundary
5498          * for page size > block size.
5499          */
5500         ioffset = round_down(offset, PAGE_SIZE);
5501         /*
5502          * Write tail of the last page before removed range since it will get
5503          * removed from the page cache below.
5504          */
5505         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5506         if (ret)
5507                 goto out_mmap;
5508         /*
5509          * Write data that will be shifted to preserve them when discarding
5510          * page cache below. We are also protected from pages becoming dirty
5511          * by i_mmap_sem.
5512          */
5513         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5514                                            LLONG_MAX);
5515         if (ret)
5516                 goto out_mmap;
5517         truncate_pagecache(inode, ioffset);
5518
5519         credits = ext4_writepage_trans_blocks(inode);
5520         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5521         if (IS_ERR(handle)) {
5522                 ret = PTR_ERR(handle);
5523                 goto out_mmap;
5524         }
5525
5526         down_write(&EXT4_I(inode)->i_data_sem);
5527         ext4_discard_preallocations(inode);
5528
5529         ret = ext4_es_remove_extent(inode, punch_start,
5530                                     EXT_MAX_BLOCKS - punch_start);
5531         if (ret) {
5532                 up_write(&EXT4_I(inode)->i_data_sem);
5533                 goto out_stop;
5534         }
5535
5536         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5537         if (ret) {
5538                 up_write(&EXT4_I(inode)->i_data_sem);
5539                 goto out_stop;
5540         }
5541         ext4_discard_preallocations(inode);
5542
5543         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5544                                      punch_stop - punch_start, SHIFT_LEFT);
5545         if (ret) {
5546                 up_write(&EXT4_I(inode)->i_data_sem);
5547                 goto out_stop;
5548         }
5549
5550         new_size = i_size_read(inode) - len;
5551         i_size_write(inode, new_size);
5552         EXT4_I(inode)->i_disksize = new_size;
5553
5554         up_write(&EXT4_I(inode)->i_data_sem);
5555         if (IS_SYNC(inode))
5556                 ext4_handle_sync(handle);
5557         inode->i_mtime = inode->i_ctime = current_time(inode);
5558         ext4_mark_inode_dirty(handle, inode);
5559         ext4_update_inode_fsync_trans(handle, inode, 1);
5560
5561 out_stop:
5562         ext4_journal_stop(handle);
5563 out_mmap:
5564         up_write(&EXT4_I(inode)->i_mmap_sem);
5565 out_mutex:
5566         inode_unlock(inode);
5567         return ret;
5568 }
5569
5570 /*
5571  * ext4_insert_range:
5572  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5573  * The data blocks starting from @offset to the EOF are shifted by @len
5574  * towards right to create a hole in the @inode. Inode size is increased
5575  * by len bytes.
5576  * Returns 0 on success, error otherwise.
5577  */
5578 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5579 {
5580         struct super_block *sb = inode->i_sb;
5581         handle_t *handle;
5582         struct ext4_ext_path *path;
5583         struct ext4_extent *extent;
5584         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5585         unsigned int credits, ee_len;
5586         int ret = 0, depth, split_flag = 0;
5587         loff_t ioffset;
5588
5589         /*
5590          * We need to test this early because xfstests assumes that an
5591          * insert range of (0, 1) will return EOPNOTSUPP if the file
5592          * system does not support insert range.
5593          */
5594         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5595                 return -EOPNOTSUPP;
5596
5597         /* Insert range works only on fs block size aligned offsets. */
5598         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5599                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5600                 return -EINVAL;
5601
5602         if (!S_ISREG(inode->i_mode))
5603                 return -EOPNOTSUPP;
5604
5605         trace_ext4_insert_range(inode, offset, len);
5606
5607         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5608         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5609
5610         /* Call ext4_force_commit to flush all data in case of data=journal */
5611         if (ext4_should_journal_data(inode)) {
5612                 ret = ext4_force_commit(inode->i_sb);
5613                 if (ret)
5614                         return ret;
5615         }
5616
5617         inode_lock(inode);
5618         /* Currently just for extent based files */
5619         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5620                 ret = -EOPNOTSUPP;
5621                 goto out_mutex;
5622         }
5623
5624         /* Check for wrap through zero */
5625         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5626                 ret = -EFBIG;
5627                 goto out_mutex;
5628         }
5629
5630         /* Offset should be less than i_size */
5631         if (offset >= i_size_read(inode)) {
5632                 ret = -EINVAL;
5633                 goto out_mutex;
5634         }
5635
5636         /* Wait for existing dio to complete */
5637         inode_dio_wait(inode);
5638
5639         /*
5640          * Prevent page faults from reinstantiating pages we have released from
5641          * page cache.
5642          */
5643         down_write(&EXT4_I(inode)->i_mmap_sem);
5644         /*
5645          * Need to round down to align start offset to page size boundary
5646          * for page size > block size.
5647          */
5648         ioffset = round_down(offset, PAGE_SIZE);
5649         /* Write out all dirty pages */
5650         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5651                         LLONG_MAX);
5652         if (ret)
5653                 goto out_mmap;
5654         truncate_pagecache(inode, ioffset);
5655
5656         credits = ext4_writepage_trans_blocks(inode);
5657         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5658         if (IS_ERR(handle)) {
5659                 ret = PTR_ERR(handle);
5660                 goto out_mmap;
5661         }
5662
5663         /* Expand file to avoid data loss if there is error while shifting */
5664         inode->i_size += len;
5665         EXT4_I(inode)->i_disksize += len;
5666         inode->i_mtime = inode->i_ctime = current_time(inode);
5667         ret = ext4_mark_inode_dirty(handle, inode);
5668         if (ret)
5669                 goto out_stop;
5670
5671         down_write(&EXT4_I(inode)->i_data_sem);
5672         ext4_discard_preallocations(inode);
5673
5674         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5675         if (IS_ERR(path)) {
5676                 up_write(&EXT4_I(inode)->i_data_sem);
5677                 goto out_stop;
5678         }
5679
5680         depth = ext_depth(inode);
5681         extent = path[depth].p_ext;
5682         if (extent) {
5683                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5684                 ee_len = ext4_ext_get_actual_len(extent);
5685
5686                 /*
5687                  * If offset_lblk is not the starting block of extent, split
5688                  * the extent @offset_lblk
5689                  */
5690                 if ((offset_lblk > ee_start_lblk) &&
5691                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5692                         if (ext4_ext_is_unwritten(extent))
5693                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5694                                         EXT4_EXT_MARK_UNWRIT2;
5695                         ret = ext4_split_extent_at(handle, inode, &path,
5696                                         offset_lblk, split_flag,
5697                                         EXT4_EX_NOCACHE |
5698                                         EXT4_GET_BLOCKS_PRE_IO |
5699                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5700                 }
5701
5702                 ext4_ext_drop_refs(path);
5703                 kfree(path);
5704                 if (ret < 0) {
5705                         up_write(&EXT4_I(inode)->i_data_sem);
5706                         goto out_stop;
5707                 }
5708         } else {
5709                 ext4_ext_drop_refs(path);
5710                 kfree(path);
5711         }
5712
5713         ret = ext4_es_remove_extent(inode, offset_lblk,
5714                         EXT_MAX_BLOCKS - offset_lblk);
5715         if (ret) {
5716                 up_write(&EXT4_I(inode)->i_data_sem);
5717                 goto out_stop;
5718         }
5719
5720         /*
5721          * if offset_lblk lies in a hole which is at start of file, use
5722          * ee_start_lblk to shift extents
5723          */
5724         ret = ext4_ext_shift_extents(inode, handle,
5725                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5726                 len_lblk, SHIFT_RIGHT);
5727
5728         up_write(&EXT4_I(inode)->i_data_sem);
5729         if (IS_SYNC(inode))
5730                 ext4_handle_sync(handle);
5731         if (ret >= 0)
5732                 ext4_update_inode_fsync_trans(handle, inode, 1);
5733
5734 out_stop:
5735         ext4_journal_stop(handle);
5736 out_mmap:
5737         up_write(&EXT4_I(inode)->i_mmap_sem);
5738 out_mutex:
5739         inode_unlock(inode);
5740         return ret;
5741 }
5742
5743 /**
5744  * ext4_swap_extents - Swap extents between two inodes
5745  *
5746  * @inode1:     First inode
5747  * @inode2:     Second inode
5748  * @lblk1:      Start block for first inode
5749  * @lblk2:      Start block for second inode
5750  * @count:      Number of blocks to swap
5751  * @unwritten: Mark second inode's extents as unwritten after swap
5752  * @erp:        Pointer to save error value
5753  *
5754  * This helper routine does exactly what is promise "swap extents". All other
5755  * stuff such as page-cache locking consistency, bh mapping consistency or
5756  * extent's data copying must be performed by caller.
5757  * Locking:
5758  *              i_mutex is held for both inodes
5759  *              i_data_sem is locked for write for both inodes
5760  * Assumptions:
5761  *              All pages from requested range are locked for both inodes
5762  */
5763 int
5764 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5765                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5766                   ext4_lblk_t count, int unwritten, int *erp)
5767 {
5768         struct ext4_ext_path *path1 = NULL;
5769         struct ext4_ext_path *path2 = NULL;
5770         int replaced_count = 0;
5771
5772         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5773         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5774         BUG_ON(!inode_is_locked(inode1));
5775         BUG_ON(!inode_is_locked(inode2));
5776
5777         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5778         if (unlikely(*erp))
5779                 return 0;
5780         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5781         if (unlikely(*erp))
5782                 return 0;
5783
5784         while (count) {
5785                 struct ext4_extent *ex1, *ex2, tmp_ex;
5786                 ext4_lblk_t e1_blk, e2_blk;
5787                 int e1_len, e2_len, len;
5788                 int split = 0;
5789
5790                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5791                 if (IS_ERR(path1)) {
5792                         *erp = PTR_ERR(path1);
5793                         path1 = NULL;
5794                 finish:
5795                         count = 0;
5796                         goto repeat;
5797                 }
5798                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5799                 if (IS_ERR(path2)) {
5800                         *erp = PTR_ERR(path2);
5801                         path2 = NULL;
5802                         goto finish;
5803                 }
5804                 ex1 = path1[path1->p_depth].p_ext;
5805                 ex2 = path2[path2->p_depth].p_ext;
5806                 /* Do we have somthing to swap ? */
5807                 if (unlikely(!ex2 || !ex1))
5808                         goto finish;
5809
5810                 e1_blk = le32_to_cpu(ex1->ee_block);
5811                 e2_blk = le32_to_cpu(ex2->ee_block);
5812                 e1_len = ext4_ext_get_actual_len(ex1);
5813                 e2_len = ext4_ext_get_actual_len(ex2);
5814
5815                 /* Hole handling */
5816                 if (!in_range(lblk1, e1_blk, e1_len) ||
5817                     !in_range(lblk2, e2_blk, e2_len)) {
5818                         ext4_lblk_t next1, next2;
5819
5820                         /* if hole after extent, then go to next extent */
5821                         next1 = ext4_ext_next_allocated_block(path1);
5822                         next2 = ext4_ext_next_allocated_block(path2);
5823                         /* If hole before extent, then shift to that extent */
5824                         if (e1_blk > lblk1)
5825                                 next1 = e1_blk;
5826                         if (e2_blk > lblk2)
5827                                 next2 = e2_blk;
5828                         /* Do we have something to swap */
5829                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5830                                 goto finish;
5831                         /* Move to the rightest boundary */
5832                         len = next1 - lblk1;
5833                         if (len < next2 - lblk2)
5834                                 len = next2 - lblk2;
5835                         if (len > count)
5836                                 len = count;
5837                         lblk1 += len;
5838                         lblk2 += len;
5839                         count -= len;
5840                         goto repeat;
5841                 }
5842
5843                 /* Prepare left boundary */
5844                 if (e1_blk < lblk1) {
5845                         split = 1;
5846                         *erp = ext4_force_split_extent_at(handle, inode1,
5847                                                 &path1, lblk1, 0);
5848                         if (unlikely(*erp))
5849                                 goto finish;
5850                 }
5851                 if (e2_blk < lblk2) {
5852                         split = 1;
5853                         *erp = ext4_force_split_extent_at(handle, inode2,
5854                                                 &path2,  lblk2, 0);
5855                         if (unlikely(*erp))
5856                                 goto finish;
5857                 }
5858                 /* ext4_split_extent_at() may result in leaf extent split,
5859                  * path must to be revalidated. */
5860                 if (split)
5861                         goto repeat;
5862
5863                 /* Prepare right boundary */
5864                 len = count;
5865                 if (len > e1_blk + e1_len - lblk1)
5866                         len = e1_blk + e1_len - lblk1;
5867                 if (len > e2_blk + e2_len - lblk2)
5868                         len = e2_blk + e2_len - lblk2;
5869
5870                 if (len != e1_len) {
5871                         split = 1;
5872                         *erp = ext4_force_split_extent_at(handle, inode1,
5873                                                 &path1, lblk1 + len, 0);
5874                         if (unlikely(*erp))
5875                                 goto finish;
5876                 }
5877                 if (len != e2_len) {
5878                         split = 1;
5879                         *erp = ext4_force_split_extent_at(handle, inode2,
5880                                                 &path2, lblk2 + len, 0);
5881                         if (*erp)
5882                                 goto finish;
5883                 }
5884                 /* ext4_split_extent_at() may result in leaf extent split,
5885                  * path must to be revalidated. */
5886                 if (split)
5887                         goto repeat;
5888
5889                 BUG_ON(e2_len != e1_len);
5890                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5891                 if (unlikely(*erp))
5892                         goto finish;
5893                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5894                 if (unlikely(*erp))
5895                         goto finish;
5896
5897                 /* Both extents are fully inside boundaries. Swap it now */
5898                 tmp_ex = *ex1;
5899                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5900                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5901                 ex1->ee_len = cpu_to_le16(e2_len);
5902                 ex2->ee_len = cpu_to_le16(e1_len);
5903                 if (unwritten)
5904                         ext4_ext_mark_unwritten(ex2);
5905                 if (ext4_ext_is_unwritten(&tmp_ex))
5906                         ext4_ext_mark_unwritten(ex1);
5907
5908                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5909                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5910                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5911                                       path2->p_depth);
5912                 if (unlikely(*erp))
5913                         goto finish;
5914                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5915                                       path1->p_depth);
5916                 /*
5917                  * Looks scarry ah..? second inode already points to new blocks,
5918                  * and it was successfully dirtied. But luckily error may happen
5919                  * only due to journal error, so full transaction will be
5920                  * aborted anyway.
5921                  */
5922                 if (unlikely(*erp))
5923                         goto finish;
5924                 lblk1 += len;
5925                 lblk2 += len;
5926                 replaced_count += len;
5927                 count -= len;
5928
5929         repeat:
5930                 ext4_ext_drop_refs(path1);
5931                 kfree(path1);
5932                 ext4_ext_drop_refs(path2);
5933                 kfree(path2);
5934                 path1 = path2 = NULL;
5935         }
5936         return replaced_count;
5937 }