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 = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
 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         if (depth < 0 || depth > EXT4_MAX_EXTENT_DEPTH) {
 873                 EXT4_ERROR_INODE(inode, "inode has invalid extent depth: %d",
 874                                  depth);
 875                 ret = -EFSCORRUPTED;
 876                 goto err;
 877         }
 878
 879         if (path) {
 880                 ext4_ext_drop_refs(path);
 881                 if (depth > path[0].p_maxdepth) {
 882                         kfree(path);
 883                         *orig_path = path = NULL;
 884                 }
 885         }
 886         if (!path) {
 887                 /* account possible depth increase */
 888                 path = kcalloc(depth + 2, sizeof(struct ext4_ext_path),
 889                                 GFP_NOFS);
 890                 if (unlikely(!path))
 891                         return ERR_PTR(-ENOMEM);
 892                 path[0].p_maxdepth = depth + 1;
 893         }
 894         path[0].p_hdr = eh;
 895         path[0].p_bh = NULL;
 896
 897         i = depth;
 898         /* walk through the tree */
 899         while (i) {
 900                 ext_debug("depth %d: num %d, max %d\n",
 901                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 902
 903                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 904                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 905                 path[ppos].p_depth = i;
 906                 path[ppos].p_ext = NULL;
 907
 908                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 909                                             flags);
 910                 if (IS_ERR(bh)) {
 911                         ret = PTR_ERR(bh);
 912                         goto err;
 913                 }
 914
 915                 eh = ext_block_hdr(bh);
 916                 ppos++;
 917                 path[ppos].p_bh = bh;
 918                 path[ppos].p_hdr = eh;
 919         }
 920
 921         path[ppos].p_depth = i;
 922         path[ppos].p_ext = NULL;
 923         path[ppos].p_idx = NULL;
 924
 925         /* find extent */
 926         ext4_ext_binsearch(inode, path + ppos, block);
 927         /* if not an empty leaf */
 928         if (path[ppos].p_ext)
 929                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 930
 931         ext4_ext_show_path(inode, path);
 932
 933         return path;
 934
 935 err:
 936         ext4_ext_drop_refs(path);
 937         kfree(path);
 938         if (orig_path)
 939                 *orig_path = NULL;
 940         return ERR_PTR(ret);
 941 }
 942
 943 /*
 944  * ext4_ext_insert_index:
 945  * insert new index [@logical;@ptr] into the block at @curp;
 946  * check where to insert: before @curp or after @curp
 947  */
 948 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 949                                  struct ext4_ext_path *curp,
 950                                  int logical, ext4_fsblk_t ptr)
 951 {
 952         struct ext4_extent_idx *ix;
 953         int len, err;
 954
 955         err = ext4_ext_get_access(handle, inode, curp);
 956         if (err)
 957                 return err;
 958
 959         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 960                 EXT4_ERROR_INODE(inode,
 961                                  "logical %d == ei_block %d!",
 962                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 963                 return -EFSCORRUPTED;
 964         }
 965
 966         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 967                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 968                 EXT4_ERROR_INODE(inode,
 969                                  "eh_entries %d >= eh_max %d!",
 970                                  le16_to_cpu(curp->p_hdr->eh_entries),
 971                                  le16_to_cpu(curp->p_hdr->eh_max));
 972                 return -EFSCORRUPTED;
 973         }
 974
 975         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 976                 /* insert after */
 977                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 978                 ix = curp->p_idx + 1;
 979         } else {
 980                 /* insert before */
 981                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 982                 ix = curp->p_idx;
 983         }
 984
 985         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 986         BUG_ON(len < 0);
 987         if (len > 0) {
 988                 ext_debug("insert new index %d: "
 989                                 "move %d indices from 0x%p to 0x%p\n",
 990                                 logical, len, ix, ix + 1);
 991                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 992         }
 993
 994         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 995                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 996                 return -EFSCORRUPTED;
 997         }
 998
 999         ix->ei_block = cpu_to_le32(logical);
1000         ext4_idx_store_pblock(ix, ptr);
1001         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1002
1003         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1004                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1005                 return -EFSCORRUPTED;
1006         }
1007
1008         err = ext4_ext_dirty(handle, inode, curp);
1009         ext4_std_error(inode->i_sb, err);
1010
1011         return err;
1012 }
1013
1014 /*
1015  * ext4_ext_split:
1016  * inserts new subtree into the path, using free index entry
1017  * at depth @at:
1018  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1019  * - makes decision where to split
1020  * - moves remaining extents and index entries (right to the split point)
1021  *   into the newly allocated blocks
1022  * - initializes subtree
1023  */
1024 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1025                           unsigned int flags,
1026                           struct ext4_ext_path *path,
1027                           struct ext4_extent *newext, int at)
1028 {
1029         struct buffer_head *bh = NULL;
1030         int depth = ext_depth(inode);
1031         struct ext4_extent_header *neh;
1032         struct ext4_extent_idx *fidx;
1033         int i = at, k, m, a;
1034         ext4_fsblk_t newblock, oldblock;
1035         __le32 border;
1036         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1037         int err = 0;
1038         size_t ext_size = 0;
1039
1040         /* make decision: where to split? */
1041         /* FIXME: now decision is simplest: at current extent */
1042
1043         /* if current leaf will be split, then we should use
1044          * border from split point */
1045         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1046                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1047                 return -EFSCORRUPTED;
1048         }
1049         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1050                 border = path[depth].p_ext[1].ee_block;
1051                 ext_debug("leaf will be split."
1052                                 " next leaf starts at %d\n",
1053                                   le32_to_cpu(border));
1054         } else {
1055                 border = newext->ee_block;
1056                 ext_debug("leaf will be added."
1057                                 " next leaf starts at %d\n",
1058                                 le32_to_cpu(border));
1059         }
1060
1061         /*
1062          * If error occurs, then we break processing
1063          * and mark filesystem read-only. index won't
1064          * be inserted and tree will be in consistent
1065          * state. Next mount will repair buffers too.
1066          */
1067
1068         /*
1069          * Get array to track all allocated blocks.
1070          * We need this to handle errors and free blocks
1071          * upon them.
1072          */
1073         ablocks = kcalloc(depth, sizeof(ext4_fsblk_t), GFP_NOFS);
1074         if (!ablocks)
1075                 return -ENOMEM;
1076
1077         /* allocate all needed blocks */
1078         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1079         for (a = 0; a < depth - at; a++) {
1080                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1081                                                    newext, &err, flags);
1082                 if (newblock == 0)
1083                         goto cleanup;
1084                 ablocks[a] = newblock;
1085         }
1086
1087         /* initialize new leaf */
1088         newblock = ablocks[--a];
1089         if (unlikely(newblock == 0)) {
1090                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1091                 err = -EFSCORRUPTED;
1092                 goto cleanup;
1093         }
1094         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1095         if (unlikely(!bh)) {
1096                 err = -ENOMEM;
1097                 goto cleanup;
1098         }
1099         lock_buffer(bh);
1100
1101         err = ext4_journal_get_create_access(handle, bh);
1102         if (err)
1103                 goto cleanup;
1104
1105         neh = ext_block_hdr(bh);
1106         neh->eh_entries = 0;
1107         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1108         neh->eh_magic = EXT4_EXT_MAGIC;
1109         neh->eh_depth = 0;
1110
1111         /* move remainder of path[depth] to the new leaf */
1112         if (unlikely(path[depth].p_hdr->eh_entries !=
1113                      path[depth].p_hdr->eh_max)) {
1114                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1115                                  path[depth].p_hdr->eh_entries,
1116                                  path[depth].p_hdr->eh_max);
1117                 err = -EFSCORRUPTED;
1118                 goto cleanup;
1119         }
1120         /* start copy from next extent */
1121         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1122         ext4_ext_show_move(inode, path, newblock, depth);
1123         if (m) {
1124                 struct ext4_extent *ex;
1125                 ex = EXT_FIRST_EXTENT(neh);
1126                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1127                 le16_add_cpu(&neh->eh_entries, m);
1128         }
1129
1130         /* zero out unused area in the extent block */
1131         ext_size = sizeof(struct ext4_extent_header) +
1132                 sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries);
1133         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1134         ext4_extent_block_csum_set(inode, neh);
1135         set_buffer_uptodate(bh);
1136         unlock_buffer(bh);
1137
1138         err = ext4_handle_dirty_metadata(handle, inode, bh);
1139         if (err)
1140                 goto cleanup;
1141         brelse(bh);
1142         bh = NULL;
1143
1144         /* correct old leaf */
1145         if (m) {
1146                 err = ext4_ext_get_access(handle, inode, path + depth);
1147                 if (err)
1148                         goto cleanup;
1149                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1150                 err = ext4_ext_dirty(handle, inode, path + depth);
1151                 if (err)
1152                         goto cleanup;
1153
1154         }
1155
1156         /* create intermediate indexes */
1157         k = depth - at - 1;
1158         if (unlikely(k < 0)) {
1159                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1160                 err = -EFSCORRUPTED;
1161                 goto cleanup;
1162         }
1163         if (k)
1164                 ext_debug("create %d intermediate indices\n", k);
1165         /* insert new index into current index block */
1166         /* current depth stored in i var */
1167         i = depth - 1;
1168         while (k--) {
1169                 oldblock = newblock;
1170                 newblock = ablocks[--a];
1171                 bh = sb_getblk(inode->i_sb, newblock);
1172                 if (unlikely(!bh)) {
1173                         err = -ENOMEM;
1174                         goto cleanup;
1175                 }
1176                 lock_buffer(bh);
1177
1178                 err = ext4_journal_get_create_access(handle, bh);
1179                 if (err)
1180                         goto cleanup;
1181
1182                 neh = ext_block_hdr(bh);
1183                 neh->eh_entries = cpu_to_le16(1);
1184                 neh->eh_magic = EXT4_EXT_MAGIC;
1185                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1186                 neh->eh_depth = cpu_to_le16(depth - i);
1187                 fidx = EXT_FIRST_INDEX(neh);
1188                 fidx->ei_block = border;
1189                 ext4_idx_store_pblock(fidx, oldblock);
1190
1191                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1192                                 i, newblock, le32_to_cpu(border), oldblock);
1193
1194                 /* move remainder of path[i] to the new index block */
1195                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1196                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1197                         EXT4_ERROR_INODE(inode,
1198                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1199                                          le32_to_cpu(path[i].p_ext->ee_block));
1200                         err = -EFSCORRUPTED;
1201                         goto cleanup;
1202                 }
1203                 /* start copy indexes */
1204                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1205                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1206                                 EXT_MAX_INDEX(path[i].p_hdr));
1207                 ext4_ext_show_move(inode, path, newblock, i);
1208                 if (m) {
1209                         memmove(++fidx, path[i].p_idx,
1210                                 sizeof(struct ext4_extent_idx) * m);
1211                         le16_add_cpu(&neh->eh_entries, m);
1212                 }
1213                 /* zero out unused area in the extent block */
1214                 ext_size = sizeof(struct ext4_extent_header) +
1215                    (sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries));
1216                 memset(bh->b_data + ext_size, 0,
1217                         inode->i_sb->s_blocksize - ext_size);
1218                 ext4_extent_block_csum_set(inode, neh);
1219                 set_buffer_uptodate(bh);
1220                 unlock_buffer(bh);
1221
1222                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1223                 if (err)
1224                         goto cleanup;
1225                 brelse(bh);
1226                 bh = NULL;
1227
1228                 /* correct old index */
1229                 if (m) {
1230                         err = ext4_ext_get_access(handle, inode, path + i);
1231                         if (err)
1232                                 goto cleanup;
1233                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1234                         err = ext4_ext_dirty(handle, inode, path + i);
1235                         if (err)
1236                                 goto cleanup;
1237                 }
1238
1239                 i--;
1240         }
1241
1242         /* insert new index */
1243         err = ext4_ext_insert_index(handle, inode, path + at,
1244                                     le32_to_cpu(border), newblock);
1245
1246 cleanup:
1247         if (bh) {
1248                 if (buffer_locked(bh))
1249                         unlock_buffer(bh);
1250                 brelse(bh);
1251         }
1252
1253         if (err) {
1254                 /* free all allocated blocks in error case */
1255                 for (i = 0; i < depth; i++) {
1256                         if (!ablocks[i])
1257                                 continue;
1258                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1259                                          EXT4_FREE_BLOCKS_METADATA);
1260                 }
1261         }
1262         kfree(ablocks);
1263
1264         return err;
1265 }
1266
1267 /*
1268  * ext4_ext_grow_indepth:
1269  * implements tree growing procedure:
1270  * - allocates new block
1271  * - moves top-level data (index block or leaf) into the new block
1272  * - initializes new top-level, creating index that points to the
1273  *   just created block
1274  */
1275 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1276                                  unsigned int flags)
1277 {
1278         struct ext4_extent_header *neh;
1279         struct buffer_head *bh;
1280         ext4_fsblk_t newblock, goal = 0;
1281         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1282         int err = 0;
1283         size_t ext_size = 0;
1284
1285         /* Try to prepend new index to old one */
1286         if (ext_depth(inode))
1287                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1288         if (goal > le32_to_cpu(es->s_first_data_block)) {
1289                 flags |= EXT4_MB_HINT_TRY_GOAL;
1290                 goal--;
1291         } else
1292                 goal = ext4_inode_to_goal_block(inode);
1293         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1294                                         NULL, &err);
1295         if (newblock == 0)
1296                 return err;
1297
1298         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1299         if (unlikely(!bh))
1300                 return -ENOMEM;
1301         lock_buffer(bh);
1302
1303         err = ext4_journal_get_create_access(handle, bh);
1304         if (err) {
1305                 unlock_buffer(bh);
1306                 goto out;
1307         }
1308
1309         ext_size = sizeof(EXT4_I(inode)->i_data);
1310         /* move top-level index/leaf into new block */
1311         memmove(bh->b_data, EXT4_I(inode)->i_data, ext_size);
1312         /* zero out unused area in the extent block */
1313         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1314
1315         /* set size of new block */
1316         neh = ext_block_hdr(bh);
1317         /* old root could have indexes or leaves
1318          * so calculate e_max right way */
1319         if (ext_depth(inode))
1320                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1321         else
1322                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1323         neh->eh_magic = EXT4_EXT_MAGIC;
1324         ext4_extent_block_csum_set(inode, neh);
1325         set_buffer_uptodate(bh);
1326         unlock_buffer(bh);
1327
1328         err = ext4_handle_dirty_metadata(handle, inode, bh);
1329         if (err)
1330                 goto out;
1331
1332         /* Update top-level index: num,max,pointer */
1333         neh = ext_inode_hdr(inode);
1334         neh->eh_entries = cpu_to_le16(1);
1335         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1336         if (neh->eh_depth == 0) {
1337                 /* Root extent block becomes index block */
1338                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1339                 EXT_FIRST_INDEX(neh)->ei_block =
1340                         EXT_FIRST_EXTENT(neh)->ee_block;
1341         }
1342         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1343                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1344                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1345                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1346
1347         le16_add_cpu(&neh->eh_depth, 1);
1348         ext4_mark_inode_dirty(handle, inode);
1349 out:
1350         brelse(bh);
1351
1352         return err;
1353 }
1354
1355 /*
1356  * ext4_ext_create_new_leaf:
1357  * finds empty index and adds new leaf.
1358  * if no free index is found, then it requests in-depth growing.
1359  */
1360 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1361                                     unsigned int mb_flags,
1362                                     unsigned int gb_flags,
1363                                     struct ext4_ext_path **ppath,
1364                                     struct ext4_extent *newext)
1365 {
1366         struct ext4_ext_path *path = *ppath;
1367         struct ext4_ext_path *curp;
1368         int depth, i, err = 0;
1369
1370 repeat:
1371         i = depth = ext_depth(inode);
1372
1373         /* walk up to the tree and look for free index entry */
1374         curp = path + depth;
1375         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1376                 i--;
1377                 curp--;
1378         }
1379
1380         /* we use already allocated block for index block,
1381          * so subsequent data blocks should be contiguous */
1382         if (EXT_HAS_FREE_INDEX(curp)) {
1383                 /* if we found index with free entry, then use that
1384                  * entry: create all needed subtree and add new leaf */
1385                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1386                 if (err)
1387                         goto out;
1388
1389                 /* refill path */
1390                 path = ext4_find_extent(inode,
1391                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1392                                     ppath, gb_flags);
1393                 if (IS_ERR(path))
1394                         err = PTR_ERR(path);
1395         } else {
1396                 /* tree is full, time to grow in depth */
1397                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1398                 if (err)
1399                         goto out;
1400
1401                 /* refill path */
1402                 path = ext4_find_extent(inode,
1403                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1404                                     ppath, gb_flags);
1405                 if (IS_ERR(path)) {
1406                         err = PTR_ERR(path);
1407                         goto out;
1408                 }
1409
1410                 /*
1411                  * only first (depth 0 -> 1) produces free space;
1412                  * in all other cases we have to split the grown tree
1413                  */
1414                 depth = ext_depth(inode);
1415                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1416                         /* now we need to split */
1417                         goto repeat;
1418                 }
1419         }
1420
1421 out:
1422         return err;
1423 }
1424
1425 /*
1426  * search the closest allocated block to the left for *logical
1427  * and returns it at @logical + it's physical address at @phys
1428  * if *logical is the smallest allocated block, the function
1429  * returns 0 at @phys
1430  * return value contains 0 (success) or error code
1431  */
1432 static int ext4_ext_search_left(struct inode *inode,
1433                                 struct ext4_ext_path *path,
1434                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1435 {
1436         struct ext4_extent_idx *ix;
1437         struct ext4_extent *ex;
1438         int depth, ee_len;
1439
1440         if (unlikely(path == NULL)) {
1441                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1442                 return -EFSCORRUPTED;
1443         }
1444         depth = path->p_depth;
1445         *phys = 0;
1446
1447         if (depth == 0 && path->p_ext == NULL)
1448                 return 0;
1449
1450         /* usually extent in the path covers blocks smaller
1451          * then *logical, but it can be that extent is the
1452          * first one in the file */
1453
1454         ex = path[depth].p_ext;
1455         ee_len = ext4_ext_get_actual_len(ex);
1456         if (*logical < le32_to_cpu(ex->ee_block)) {
1457                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1458                         EXT4_ERROR_INODE(inode,
1459                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1460                                          *logical, le32_to_cpu(ex->ee_block));
1461                         return -EFSCORRUPTED;
1462                 }
1463                 while (--depth >= 0) {
1464                         ix = path[depth].p_idx;
1465                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1466                                 EXT4_ERROR_INODE(inode,
1467                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1468                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1469                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1470                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1471                                   depth);
1472                                 return -EFSCORRUPTED;
1473                         }
1474                 }
1475                 return 0;
1476         }
1477
1478         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1479                 EXT4_ERROR_INODE(inode,
1480                                  "logical %d < ee_block %d + ee_len %d!",
1481                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1482                 return -EFSCORRUPTED;
1483         }
1484
1485         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1486         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1487         return 0;
1488 }
1489
1490 /*
1491  * search the closest allocated block to the right for *logical
1492  * and returns it at @logical + it's physical address at @phys
1493  * if *logical is the largest allocated block, the function
1494  * returns 0 at @phys
1495  * return value contains 0 (success) or error code
1496  */
1497 static int ext4_ext_search_right(struct inode *inode,
1498                                  struct ext4_ext_path *path,
1499                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1500                                  struct ext4_extent **ret_ex)
1501 {
1502         struct buffer_head *bh = NULL;
1503         struct ext4_extent_header *eh;
1504         struct ext4_extent_idx *ix;
1505         struct ext4_extent *ex;
1506         ext4_fsblk_t block;
1507         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1508         int ee_len;
1509
1510         if (unlikely(path == NULL)) {
1511                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1512                 return -EFSCORRUPTED;
1513         }
1514         depth = path->p_depth;
1515         *phys = 0;
1516
1517         if (depth == 0 && path->p_ext == NULL)
1518                 return 0;
1519
1520         /* usually extent in the path covers blocks smaller
1521          * then *logical, but it can be that extent is the
1522          * first one in the file */
1523
1524         ex = path[depth].p_ext;
1525         ee_len = ext4_ext_get_actual_len(ex);
1526         if (*logical < le32_to_cpu(ex->ee_block)) {
1527                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1528                         EXT4_ERROR_INODE(inode,
1529                                          "first_extent(path[%d].p_hdr) != ex",
1530                                          depth);
1531                         return -EFSCORRUPTED;
1532                 }
1533                 while (--depth >= 0) {
1534                         ix = path[depth].p_idx;
1535                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1536                                 EXT4_ERROR_INODE(inode,
1537                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1538                                                  *logical);
1539                                 return -EFSCORRUPTED;
1540                         }
1541                 }
1542                 goto found_extent;
1543         }
1544
1545         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1546                 EXT4_ERROR_INODE(inode,
1547                                  "logical %d < ee_block %d + ee_len %d!",
1548                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1549                 return -EFSCORRUPTED;
1550         }
1551
1552         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1553                 /* next allocated block in this leaf */
1554                 ex++;
1555                 goto found_extent;
1556         }
1557
1558         /* go up and search for index to the right */
1559         while (--depth >= 0) {
1560                 ix = path[depth].p_idx;
1561                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1562                         goto got_index;
1563         }
1564
1565         /* we've gone up to the root and found no index to the right */
1566         return 0;
1567
1568 got_index:
1569         /* we've found index to the right, let's
1570          * follow it and find the closest allocated
1571          * block to the right */
1572         ix++;
1573         block = ext4_idx_pblock(ix);
1574         while (++depth < path->p_depth) {
1575                 /* subtract from p_depth to get proper eh_depth */
1576                 bh = read_extent_tree_block(inode, block,
1577                                             path->p_depth - depth, 0);
1578                 if (IS_ERR(bh))
1579                         return PTR_ERR(bh);
1580                 eh = ext_block_hdr(bh);
1581                 ix = EXT_FIRST_INDEX(eh);
1582                 block = ext4_idx_pblock(ix);
1583                 put_bh(bh);
1584         }
1585
1586         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1587         if (IS_ERR(bh))
1588                 return PTR_ERR(bh);
1589         eh = ext_block_hdr(bh);
1590         ex = EXT_FIRST_EXTENT(eh);
1591 found_extent:
1592         *logical = le32_to_cpu(ex->ee_block);
1593         *phys = ext4_ext_pblock(ex);
1594         *ret_ex = ex;
1595         if (bh)
1596                 put_bh(bh);
1597         return 0;
1598 }
1599
1600 /*
1601  * ext4_ext_next_allocated_block:
1602  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1603  * NOTE: it considers block number from index entry as
1604  * allocated block. Thus, index entries have to be consistent
1605  * with leaves.
1606  */
1607 ext4_lblk_t
1608 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1609 {
1610         int depth;
1611
1612         BUG_ON(path == NULL);
1613         depth = path->p_depth;
1614
1615         if (depth == 0 && path->p_ext == NULL)
1616                 return EXT_MAX_BLOCKS;
1617
1618         while (depth >= 0) {
1619                 if (depth == path->p_depth) {
1620                         /* leaf */
1621                         if (path[depth].p_ext &&
1622                                 path[depth].p_ext !=
1623                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1624                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1625                 } else {
1626                         /* index */
1627                         if (path[depth].p_idx !=
1628                                         EXT_LAST_INDEX(path[depth].p_hdr))
1629                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1630                 }
1631                 depth--;
1632         }
1633
1634         return EXT_MAX_BLOCKS;
1635 }
1636
1637 /*
1638  * ext4_ext_next_leaf_block:
1639  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1640  */
1641 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1642 {
1643         int depth;
1644
1645         BUG_ON(path == NULL);
1646         depth = path->p_depth;
1647
1648         /* zero-tree has no leaf blocks at all */
1649         if (depth == 0)
1650                 return EXT_MAX_BLOCKS;
1651
1652         /* go to index block */
1653         depth--;
1654
1655         while (depth >= 0) {
1656                 if (path[depth].p_idx !=
1657                                 EXT_LAST_INDEX(path[depth].p_hdr))
1658                         return (ext4_lblk_t)
1659                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1660                 depth--;
1661         }
1662
1663         return EXT_MAX_BLOCKS;
1664 }
1665
1666 /*
1667  * ext4_ext_correct_indexes:
1668  * if leaf gets modified and modified extent is first in the leaf,
1669  * then we have to correct all indexes above.
1670  * TODO: do we need to correct tree in all cases?
1671  */
1672 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1673                                 struct ext4_ext_path *path)
1674 {
1675         struct ext4_extent_header *eh;
1676         int depth = ext_depth(inode);
1677         struct ext4_extent *ex;
1678         __le32 border;
1679         int k, err = 0;
1680
1681         eh = path[depth].p_hdr;
1682         ex = path[depth].p_ext;
1683
1684         if (unlikely(ex == NULL || eh == NULL)) {
1685                 EXT4_ERROR_INODE(inode,
1686                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1687                 return -EFSCORRUPTED;
1688         }
1689
1690         if (depth == 0) {
1691                 /* there is no tree at all */
1692                 return 0;
1693         }
1694
1695         if (ex != EXT_FIRST_EXTENT(eh)) {
1696                 /* we correct tree if first leaf got modified only */
1697                 return 0;
1698         }
1699
1700         /*
1701          * TODO: we need correction if border is smaller than current one
1702          */
1703         k = depth - 1;
1704         border = path[depth].p_ext->ee_block;
1705         err = ext4_ext_get_access(handle, inode, path + k);
1706         if (err)
1707                 return err;
1708         path[k].p_idx->ei_block = border;
1709         err = ext4_ext_dirty(handle, inode, path + k);
1710         if (err)
1711                 return err;
1712
1713         while (k--) {
1714                 /* change all left-side indexes */
1715                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1716                         break;
1717                 err = ext4_ext_get_access(handle, inode, path + k);
1718                 if (err)
1719                         break;
1720                 path[k].p_idx->ei_block = border;
1721                 err = ext4_ext_dirty(handle, inode, path + k);
1722                 if (err)
1723                         break;
1724         }
1725
1726         return err;
1727 }
1728
1729 int
1730 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1731                                 struct ext4_extent *ex2)
1732 {
1733         unsigned short ext1_ee_len, ext2_ee_len;
1734
1735         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1736                 return 0;
1737
1738         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1739         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1740
1741         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1742                         le32_to_cpu(ex2->ee_block))
1743                 return 0;
1744
1745         /*
1746          * To allow future support for preallocated extents to be added
1747          * as an RO_COMPAT feature, refuse to merge to extents if
1748          * this can result in the top bit of ee_len being set.
1749          */
1750         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1751                 return 0;
1752         /*
1753          * The check for IO to unwritten extent is somewhat racy as we
1754          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1755          * dropping i_data_sem. But reserved blocks should save us in that
1756          * case.
1757          */
1758         if (ext4_ext_is_unwritten(ex1) &&
1759             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1760              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1761              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1762                 return 0;
1763 #ifdef AGGRESSIVE_TEST
1764         if (ext1_ee_len >= 4)
1765                 return 0;
1766 #endif
1767
1768         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1769                 return 1;
1770         return 0;
1771 }
1772
1773 /*
1774  * This function tries to merge the "ex" extent to the next extent in the tree.
1775  * It always tries to merge towards right. If you want to merge towards
1776  * left, pass "ex - 1" as argument instead of "ex".
1777  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1778  * 1 if they got merged.
1779  */
1780 static int ext4_ext_try_to_merge_right(struct inode *inode,
1781                                  struct ext4_ext_path *path,
1782                                  struct ext4_extent *ex)
1783 {
1784         struct ext4_extent_header *eh;
1785         unsigned int depth, len;
1786         int merge_done = 0, unwritten;
1787
1788         depth = ext_depth(inode);
1789         BUG_ON(path[depth].p_hdr == NULL);
1790         eh = path[depth].p_hdr;
1791
1792         while (ex < EXT_LAST_EXTENT(eh)) {
1793                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1794                         break;
1795                 /* merge with next extent! */
1796                 unwritten = ext4_ext_is_unwritten(ex);
1797                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1798                                 + ext4_ext_get_actual_len(ex + 1));
1799                 if (unwritten)
1800                         ext4_ext_mark_unwritten(ex);
1801
1802                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1803                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1804                                 * sizeof(struct ext4_extent);
1805                         memmove(ex + 1, ex + 2, len);
1806                 }
1807                 le16_add_cpu(&eh->eh_entries, -1);
1808                 merge_done = 1;
1809                 WARN_ON(eh->eh_entries == 0);
1810                 if (!eh->eh_entries)
1811                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1812         }
1813
1814         return merge_done;
1815 }
1816
1817 /*
1818  * This function does a very simple check to see if we can collapse
1819  * an extent tree with a single extent tree leaf block into the inode.
1820  */
1821 static void ext4_ext_try_to_merge_up(handle_t *handle,
1822                                      struct inode *inode,
1823                                      struct ext4_ext_path *path)
1824 {
1825         size_t s;
1826         unsigned max_root = ext4_ext_space_root(inode, 0);
1827         ext4_fsblk_t blk;
1828
1829         if ((path[0].p_depth != 1) ||
1830             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1831             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1832                 return;
1833
1834         /*
1835          * We need to modify the block allocation bitmap and the block
1836          * group descriptor to release the extent tree block.  If we
1837          * can't get the journal credits, give up.
1838          */
1839         if (ext4_journal_extend(handle, 2))
1840                 return;
1841
1842         /*
1843          * Copy the extent data up to the inode
1844          */
1845         blk = ext4_idx_pblock(path[0].p_idx);
1846         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1847                 sizeof(struct ext4_extent_idx);
1848         s += sizeof(struct ext4_extent_header);
1849
1850         path[1].p_maxdepth = path[0].p_maxdepth;
1851         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1852         path[0].p_depth = 0;
1853         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1854                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1855         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1856
1857         brelse(path[1].p_bh);
1858         ext4_free_blocks(handle, inode, NULL, blk, 1,
1859                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1860 }
1861
1862 /*
1863  * This function tries to merge the @ex extent to neighbours in the tree.
1864  * return 1 if merge left else 0.
1865  */
1866 static void ext4_ext_try_to_merge(handle_t *handle,
1867                                   struct inode *inode,
1868                                   struct ext4_ext_path *path,
1869                                   struct ext4_extent *ex) {
1870         struct ext4_extent_header *eh;
1871         unsigned int depth;
1872         int merge_done = 0;
1873
1874         depth = ext_depth(inode);
1875         BUG_ON(path[depth].p_hdr == NULL);
1876         eh = path[depth].p_hdr;
1877
1878         if (ex > EXT_FIRST_EXTENT(eh))
1879                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1880
1881         if (!merge_done)
1882                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1883
1884         ext4_ext_try_to_merge_up(handle, inode, path);
1885 }
1886
1887 /*
1888  * check if a portion of the "newext" extent overlaps with an
1889  * existing extent.
1890  *
1891  * If there is an overlap discovered, it updates the length of the newext
1892  * such that there will be no overlap, and then returns 1.
1893  * If there is no overlap found, it returns 0.
1894  */
1895 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1896                                            struct inode *inode,
1897                                            struct ext4_extent *newext,
1898                                            struct ext4_ext_path *path)
1899 {
1900         ext4_lblk_t b1, b2;
1901         unsigned int depth, len1;
1902         unsigned int ret = 0;
1903
1904         b1 = le32_to_cpu(newext->ee_block);
1905         len1 = ext4_ext_get_actual_len(newext);
1906         depth = ext_depth(inode);
1907         if (!path[depth].p_ext)
1908                 goto out;
1909         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1910
1911         /*
1912          * get the next allocated block if the extent in the path
1913          * is before the requested block(s)
1914          */
1915         if (b2 < b1) {
1916                 b2 = ext4_ext_next_allocated_block(path);
1917                 if (b2 == EXT_MAX_BLOCKS)
1918                         goto out;
1919                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1920         }
1921
1922         /* check for wrap through zero on extent logical start block*/
1923         if (b1 + len1 < b1) {
1924                 len1 = EXT_MAX_BLOCKS - b1;
1925                 newext->ee_len = cpu_to_le16(len1);
1926                 ret = 1;
1927         }
1928
1929         /* check for overlap */
1930         if (b1 + len1 > b2) {
1931                 newext->ee_len = cpu_to_le16(b2 - b1);
1932                 ret = 1;
1933         }
1934 out:
1935         return ret;
1936 }
1937
1938 /*
1939  * ext4_ext_insert_extent:
1940  * tries to merge requsted extent into the existing extent or
1941  * inserts requested extent as new one into the tree,
1942  * creating new leaf in the no-space case.
1943  */
1944 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1945                                 struct ext4_ext_path **ppath,
1946                                 struct ext4_extent *newext, int gb_flags)
1947 {
1948         struct ext4_ext_path *path = *ppath;
1949         struct ext4_extent_header *eh;
1950         struct ext4_extent *ex, *fex;
1951         struct ext4_extent *nearex; /* nearest extent */
1952         struct ext4_ext_path *npath = NULL;
1953         int depth, len, err;
1954         ext4_lblk_t next;
1955         int mb_flags = 0, unwritten;
1956
1957         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1958                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1959         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1960                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1961                 return -EFSCORRUPTED;
1962         }
1963         depth = ext_depth(inode);
1964         ex = path[depth].p_ext;
1965         eh = path[depth].p_hdr;
1966         if (unlikely(path[depth].p_hdr == NULL)) {
1967                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1968                 return -EFSCORRUPTED;
1969         }
1970
1971         /* try to insert block into found extent and return */
1972         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1973
1974                 /*
1975                  * Try to see whether we should rather test the extent on
1976                  * right from ex, or from the left of ex. This is because
1977                  * ext4_find_extent() can return either extent on the
1978                  * left, or on the right from the searched position. This
1979                  * will make merging more effective.
1980                  */
1981                 if (ex < EXT_LAST_EXTENT(eh) &&
1982                     (le32_to_cpu(ex->ee_block) +
1983                     ext4_ext_get_actual_len(ex) <
1984                     le32_to_cpu(newext->ee_block))) {
1985                         ex += 1;
1986                         goto prepend;
1987                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1988                            (le32_to_cpu(newext->ee_block) +
1989                            ext4_ext_get_actual_len(newext) <
1990                            le32_to_cpu(ex->ee_block)))
1991                         ex -= 1;
1992
1993                 /* Try to append newex to the ex */
1994                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1995                         ext_debug("append [%d]%d block to %u:[%d]%d"
1996                                   "(from %llu)\n",
1997                                   ext4_ext_is_unwritten(newext),
1998                                   ext4_ext_get_actual_len(newext),
1999                                   le32_to_cpu(ex->ee_block),
2000                                   ext4_ext_is_unwritten(ex),
2001                                   ext4_ext_get_actual_len(ex),
2002                                   ext4_ext_pblock(ex));
2003                         err = ext4_ext_get_access(handle, inode,
2004                                                   path + depth);
2005                         if (err)
2006                                 return err;
2007                         unwritten = ext4_ext_is_unwritten(ex);
2008                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2009                                         + ext4_ext_get_actual_len(newext));
2010                         if (unwritten)
2011                                 ext4_ext_mark_unwritten(ex);
2012                         eh = path[depth].p_hdr;
2013                         nearex = ex;
2014                         goto merge;
2015                 }
2016
2017 prepend:
2018                 /* Try to prepend newex to the ex */
2019                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2020                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2021                                   "(from %llu)\n",
2022                                   le32_to_cpu(newext->ee_block),
2023                                   ext4_ext_is_unwritten(newext),
2024                                   ext4_ext_get_actual_len(newext),
2025                                   le32_to_cpu(ex->ee_block),
2026                                   ext4_ext_is_unwritten(ex),
2027                                   ext4_ext_get_actual_len(ex),
2028                                   ext4_ext_pblock(ex));
2029                         err = ext4_ext_get_access(handle, inode,
2030                                                   path + depth);
2031                         if (err)
2032                                 return err;
2033
2034                         unwritten = ext4_ext_is_unwritten(ex);
2035                         ex->ee_block = newext->ee_block;
2036                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2037                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2038                                         + ext4_ext_get_actual_len(newext));
2039                         if (unwritten)
2040                                 ext4_ext_mark_unwritten(ex);
2041                         eh = path[depth].p_hdr;
2042                         nearex = ex;
2043                         goto merge;
2044                 }
2045         }
2046
2047         depth = ext_depth(inode);
2048         eh = path[depth].p_hdr;
2049         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2050                 goto has_space;
2051
2052         /* probably next leaf has space for us? */
2053         fex = EXT_LAST_EXTENT(eh);
2054         next = EXT_MAX_BLOCKS;
2055         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2056                 next = ext4_ext_next_leaf_block(path);
2057         if (next != EXT_MAX_BLOCKS) {
2058                 ext_debug("next leaf block - %u\n", next);
2059                 BUG_ON(npath != NULL);
2060                 npath = ext4_find_extent(inode, next, NULL, 0);
2061                 if (IS_ERR(npath))
2062                         return PTR_ERR(npath);
2063                 BUG_ON(npath->p_depth != path->p_depth);
2064                 eh = npath[depth].p_hdr;
2065                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2066                         ext_debug("next leaf isn't full(%d)\n",
2067                                   le16_to_cpu(eh->eh_entries));
2068                         path = npath;
2069                         goto has_space;
2070                 }
2071                 ext_debug("next leaf has no free space(%d,%d)\n",
2072                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2073         }
2074
2075         /*
2076          * There is no free space in the found leaf.
2077          * We're gonna add a new leaf in the tree.
2078          */
2079         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2080                 mb_flags |= EXT4_MB_USE_RESERVED;
2081         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2082                                        ppath, newext);
2083         if (err)
2084                 goto cleanup;
2085         depth = ext_depth(inode);
2086         eh = path[depth].p_hdr;
2087
2088 has_space:
2089         nearex = path[depth].p_ext;
2090
2091         err = ext4_ext_get_access(handle, inode, path + depth);
2092         if (err)
2093                 goto cleanup;
2094
2095         if (!nearex) {
2096                 /* there is no extent in this leaf, create first one */
2097                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2098                                 le32_to_cpu(newext->ee_block),
2099                                 ext4_ext_pblock(newext),
2100                                 ext4_ext_is_unwritten(newext),
2101                                 ext4_ext_get_actual_len(newext));
2102                 nearex = EXT_FIRST_EXTENT(eh);
2103         } else {
2104                 if (le32_to_cpu(newext->ee_block)
2105                            > le32_to_cpu(nearex->ee_block)) {
2106                         /* Insert after */
2107                         ext_debug("insert %u:%llu:[%d]%d before: "
2108                                         "nearest %p\n",
2109                                         le32_to_cpu(newext->ee_block),
2110                                         ext4_ext_pblock(newext),
2111                                         ext4_ext_is_unwritten(newext),
2112                                         ext4_ext_get_actual_len(newext),
2113                                         nearex);
2114                         nearex++;
2115                 } else {
2116                         /* Insert before */
2117                         BUG_ON(newext->ee_block == nearex->ee_block);
2118                         ext_debug("insert %u:%llu:[%d]%d after: "
2119                                         "nearest %p\n",
2120                                         le32_to_cpu(newext->ee_block),
2121                                         ext4_ext_pblock(newext),
2122                                         ext4_ext_is_unwritten(newext),
2123                                         ext4_ext_get_actual_len(newext),
2124                                         nearex);
2125                 }
2126                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2127                 if (len > 0) {
2128                         ext_debug("insert %u:%llu:[%d]%d: "
2129                                         "move %d extents from 0x%p to 0x%p\n",
2130                                         le32_to_cpu(newext->ee_block),
2131                                         ext4_ext_pblock(newext),
2132                                         ext4_ext_is_unwritten(newext),
2133                                         ext4_ext_get_actual_len(newext),
2134                                         len, nearex, nearex + 1);
2135                         memmove(nearex + 1, nearex,
2136                                 len * sizeof(struct ext4_extent));
2137                 }
2138         }
2139
2140         le16_add_cpu(&eh->eh_entries, 1);
2141         path[depth].p_ext = nearex;
2142         nearex->ee_block = newext->ee_block;
2143         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2144         nearex->ee_len = newext->ee_len;
2145
2146 merge:
2147         /* try to merge extents */
2148         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2149                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2150
2151
2152         /* time to correct all indexes above */
2153         err = ext4_ext_correct_indexes(handle, inode, path);
2154         if (err)
2155                 goto cleanup;
2156
2157         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2158
2159 cleanup:
2160         ext4_ext_drop_refs(npath);
2161         kfree(npath);
2162         return err;
2163 }
2164
2165 static int ext4_fill_fiemap_extents(struct inode *inode,
2166                                     ext4_lblk_t block, ext4_lblk_t num,
2167                                     struct fiemap_extent_info *fieinfo)
2168 {
2169         struct ext4_ext_path *path = NULL;
2170         struct ext4_extent *ex;
2171         struct extent_status es;
2172         ext4_lblk_t next, next_del, start = 0, end = 0;
2173         ext4_lblk_t last = block + num;
2174         int exists, depth = 0, err = 0;
2175         unsigned int flags = 0;
2176         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2177
2178         while (block < last && block != EXT_MAX_BLOCKS) {
2179                 num = last - block;
2180                 /* find extent for this block */
2181                 down_read(&EXT4_I(inode)->i_data_sem);
2182
2183                 path = ext4_find_extent(inode, block, &path, 0);
2184                 if (IS_ERR(path)) {
2185                         up_read(&EXT4_I(inode)->i_data_sem);
2186                         err = PTR_ERR(path);
2187                         path = NULL;
2188                         break;
2189                 }
2190
2191                 depth = ext_depth(inode);
2192                 if (unlikely(path[depth].p_hdr == NULL)) {
2193                         up_read(&EXT4_I(inode)->i_data_sem);
2194                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2195                         err = -EFSCORRUPTED;
2196                         break;
2197                 }
2198                 ex = path[depth].p_ext;
2199                 next = ext4_ext_next_allocated_block(path);
2200
2201                 flags = 0;
2202                 exists = 0;
2203                 if (!ex) {
2204                         /* there is no extent yet, so try to allocate
2205                          * all requested space */
2206                         start = block;
2207                         end = block + num;
2208                 } else if (le32_to_cpu(ex->ee_block) > block) {
2209                         /* need to allocate space before found extent */
2210                         start = block;
2211                         end = le32_to_cpu(ex->ee_block);
2212                         if (block + num < end)
2213                                 end = block + num;
2214                 } else if (block >= le32_to_cpu(ex->ee_block)
2215                                         + ext4_ext_get_actual_len(ex)) {
2216                         /* need to allocate space after found extent */
2217                         start = block;
2218                         end = block + num;
2219                         if (end >= next)
2220                                 end = next;
2221                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2222                         /*
2223                          * some part of requested space is covered
2224                          * by found extent
2225                          */
2226                         start = block;
2227                         end = le32_to_cpu(ex->ee_block)
2228                                 + ext4_ext_get_actual_len(ex);
2229                         if (block + num < end)
2230                                 end = block + num;
2231                         exists = 1;
2232                 } else {
2233                         BUG();
2234                 }
2235                 BUG_ON(end <= start);
2236
2237                 if (!exists) {
2238                         es.es_lblk = start;
2239                         es.es_len = end - start;
2240                         es.es_pblk = 0;
2241                 } else {
2242                         es.es_lblk = le32_to_cpu(ex->ee_block);
2243                         es.es_len = ext4_ext_get_actual_len(ex);
2244                         es.es_pblk = ext4_ext_pblock(ex);
2245                         if (ext4_ext_is_unwritten(ex))
2246                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2247                 }
2248
2249                 /*
2250                  * Find delayed extent and update es accordingly. We call
2251                  * it even in !exists case to find out whether es is the
2252                  * last existing extent or not.
2253                  */
2254                 next_del = ext4_find_delayed_extent(inode, &es);
2255                 if (!exists && next_del) {
2256                         exists = 1;
2257                         flags |= (FIEMAP_EXTENT_DELALLOC |
2258                                   FIEMAP_EXTENT_UNKNOWN);
2259                 }
2260                 up_read(&EXT4_I(inode)->i_data_sem);
2261
2262                 if (unlikely(es.es_len == 0)) {
2263                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2264                         err = -EFSCORRUPTED;
2265                         break;
2266                 }
2267
2268                 /*
2269                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2270                  * we need to check next == EXT_MAX_BLOCKS because it is
2271                  * possible that an extent is with unwritten and delayed
2272                  * status due to when an extent is delayed allocated and
2273                  * is allocated by fallocate status tree will track both of
2274                  * them in a extent.
2275                  *
2276                  * So we could return a unwritten and delayed extent, and
2277                  * its block is equal to 'next'.
2278                  */
2279                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2280                         flags |= FIEMAP_EXTENT_LAST;
2281                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2282                                      next != EXT_MAX_BLOCKS)) {
2283                                 EXT4_ERROR_INODE(inode,
2284                                                  "next extent == %u, next "
2285                                                  "delalloc extent = %u",
2286                                                  next, next_del);
2287                                 err = -EFSCORRUPTED;
2288                                 break;
2289                         }
2290                 }
2291
2292                 if (exists) {
2293                         err = fiemap_fill_next_extent(fieinfo,
2294                                 (__u64)es.es_lblk << blksize_bits,
2295                                 (__u64)es.es_pblk << blksize_bits,
2296                                 (__u64)es.es_len << blksize_bits,
2297                                 flags);
2298                         if (err < 0)
2299                                 break;
2300                         if (err == 1) {
2301                                 err = 0;
2302                                 break;
2303                         }
2304                 }
2305
2306                 block = es.es_lblk + es.es_len;
2307         }
2308
2309         ext4_ext_drop_refs(path);
2310         kfree(path);
2311         return err;
2312 }
2313
2314 /*
2315  * ext4_ext_determine_hole - determine hole around given block
2316  * @inode:      inode we lookup in
2317  * @path:       path in extent tree to @lblk
2318  * @lblk:       pointer to logical block around which we want to determine hole
2319  *
2320  * Determine hole length (and start if easily possible) around given logical
2321  * block. We don't try too hard to find the beginning of the hole but @path
2322  * actually points to extent before @lblk, we provide it.
2323  *
2324  * The function returns the length of a hole starting at @lblk. We update @lblk
2325  * to the beginning of the hole if we managed to find it.
2326  */
2327 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2328                                            struct ext4_ext_path *path,
2329                                            ext4_lblk_t *lblk)
2330 {
2331         int depth = ext_depth(inode);
2332         struct ext4_extent *ex;
2333         ext4_lblk_t len;
2334
2335         ex = path[depth].p_ext;
2336         if (ex == NULL) {
2337                 /* there is no extent yet, so gap is [0;-] */
2338                 *lblk = 0;
2339                 len = EXT_MAX_BLOCKS;
2340         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2341                 len = le32_to_cpu(ex->ee_block) - *lblk;
2342         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2343                         + ext4_ext_get_actual_len(ex)) {
2344                 ext4_lblk_t next;
2345
2346                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2347                 next = ext4_ext_next_allocated_block(path);
2348                 BUG_ON(next == *lblk);
2349                 len = next - *lblk;
2350         } else {
2351                 BUG();
2352         }
2353         return len;
2354 }
2355
2356 /*
2357  * ext4_ext_put_gap_in_cache:
2358  * calculate boundaries of the gap that the requested block fits into
2359  * and cache this gap
2360  */
2361 static void
2362 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2363                           ext4_lblk_t hole_len)
2364 {
2365         struct extent_status es;
2366
2367         ext4_es_find_extent_range(inode, &ext4_es_is_delayed, hole_start,
2368                                   hole_start + hole_len - 1, &es);
2369         if (es.es_len) {
2370                 /* There's delayed extent containing lblock? */
2371                 if (es.es_lblk <= hole_start)
2372                         return;
2373                 hole_len = min(es.es_lblk - hole_start, hole_len);
2374         }
2375         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2376         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2377                               EXTENT_STATUS_HOLE);
2378 }
2379
2380 /*
2381  * ext4_ext_rm_idx:
2382  * removes index from the index block.
2383  */
2384 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2385                         struct ext4_ext_path *path, int depth)
2386 {
2387         int err;
2388         ext4_fsblk_t leaf;
2389
2390         /* free index block */
2391         depth--;
2392         path = path + depth;
2393         leaf = ext4_idx_pblock(path->p_idx);
2394         if (unlikely(path->p_hdr->eh_entries == 0)) {
2395                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2396                 return -EFSCORRUPTED;
2397         }
2398         err = ext4_ext_get_access(handle, inode, path);
2399         if (err)
2400                 return err;
2401
2402         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2403                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2404                 len *= sizeof(struct ext4_extent_idx);
2405                 memmove(path->p_idx, path->p_idx + 1, len);
2406         }
2407
2408         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2409         err = ext4_ext_dirty(handle, inode, path);
2410         if (err)
2411                 return err;
2412         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2413         trace_ext4_ext_rm_idx(inode, leaf);
2414
2415         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2416                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2417
2418         while (--depth >= 0) {
2419                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2420                         break;
2421                 path--;
2422                 err = ext4_ext_get_access(handle, inode, path);
2423                 if (err)
2424                         break;
2425                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2426                 err = ext4_ext_dirty(handle, inode, path);
2427                 if (err)
2428                         break;
2429         }
2430         return err;
2431 }
2432
2433 /*
2434  * ext4_ext_calc_credits_for_single_extent:
2435  * This routine returns max. credits that needed to insert an extent
2436  * to the extent tree.
2437  * When pass the actual path, the caller should calculate credits
2438  * under i_data_sem.
2439  */
2440 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2441                                                 struct ext4_ext_path *path)
2442 {
2443         if (path) {
2444                 int depth = ext_depth(inode);
2445                 int ret = 0;
2446
2447                 /* probably there is space in leaf? */
2448                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2449                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2450
2451                         /*
2452                          *  There are some space in the leaf tree, no
2453                          *  need to account for leaf block credit
2454                          *
2455                          *  bitmaps and block group descriptor blocks
2456                          *  and other metadata blocks still need to be
2457                          *  accounted.
2458                          */
2459                         /* 1 bitmap, 1 block group descriptor */
2460                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2461                         return ret;
2462                 }
2463         }
2464
2465         return ext4_chunk_trans_blocks(inode, nrblocks);
2466 }
2467
2468 /*
2469  * How many index/leaf blocks need to change/allocate to add @extents extents?
2470  *
2471  * If we add a single extent, then in the worse case, each tree level
2472  * index/leaf need to be changed in case of the tree split.
2473  *
2474  * If more extents are inserted, they could cause the whole tree split more
2475  * than once, but this is really rare.
2476  */
2477 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2478 {
2479         int index;
2480         int depth;
2481
2482         /* If we are converting the inline data, only one is needed here. */
2483         if (ext4_has_inline_data(inode))
2484                 return 1;
2485
2486         depth = ext_depth(inode);
2487
2488         if (extents <= 1)
2489                 index = depth * 2;
2490         else
2491                 index = depth * 3;
2492
2493         return index;
2494 }
2495
2496 static inline int get_default_free_blocks_flags(struct inode *inode)
2497 {
2498         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2499             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2500                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2501         else if (ext4_should_journal_data(inode))
2502                 return EXT4_FREE_BLOCKS_FORGET;
2503         return 0;
2504 }
2505
2506 /*
2507  * ext4_rereserve_cluster - increment the reserved cluster count when
2508  *                          freeing a cluster with a pending reservation
2509  *
2510  * @inode - file containing the cluster
2511  * @lblk - logical block in cluster to be reserved
2512  *
2513  * Increments the reserved cluster count and adjusts quota in a bigalloc
2514  * file system when freeing a partial cluster containing at least one
2515  * delayed and unwritten block.  A partial cluster meeting that
2516  * requirement will have a pending reservation.  If so, the
2517  * RERESERVE_CLUSTER flag is used when calling ext4_free_blocks() to
2518  * defer reserved and allocated space accounting to a subsequent call
2519  * to this function.
2520  */
2521 static void ext4_rereserve_cluster(struct inode *inode, ext4_lblk_t lblk)
2522 {
2523         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2524         struct ext4_inode_info *ei = EXT4_I(inode);
2525
2526         dquot_reclaim_block(inode, EXT4_C2B(sbi, 1));
2527
2528         spin_lock(&ei->i_block_reservation_lock);
2529         ei->i_reserved_data_blocks++;
2530         percpu_counter_add(&sbi->s_dirtyclusters_counter, 1);
2531         spin_unlock(&ei->i_block_reservation_lock);
2532
2533         percpu_counter_add(&sbi->s_freeclusters_counter, 1);
2534         ext4_remove_pending(inode, lblk);
2535 }
2536
2537 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2538                               struct ext4_extent *ex,
2539                               struct partial_cluster *partial,
2540                               ext4_lblk_t from, ext4_lblk_t to)
2541 {
2542         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2543         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2544         ext4_fsblk_t last_pblk, pblk;
2545         ext4_lblk_t num;
2546         int flags;
2547
2548         /* only extent tail removal is allowed */
2549         if (from < le32_to_cpu(ex->ee_block) ||
2550             to != le32_to_cpu(ex->ee_block) + ee_len - 1) {
2551                 ext4_error(sbi->s_sb,
2552                            "strange request: removal(2) %u-%u from %u:%u",
2553                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2554                 return 0;
2555         }
2556
2557 #ifdef EXTENTS_STATS
2558         spin_lock(&sbi->s_ext_stats_lock);
2559         sbi->s_ext_blocks += ee_len;
2560         sbi->s_ext_extents++;
2561         if (ee_len < sbi->s_ext_min)
2562                 sbi->s_ext_min = ee_len;
2563         if (ee_len > sbi->s_ext_max)
2564                 sbi->s_ext_max = ee_len;
2565         if (ext_depth(inode) > sbi->s_depth_max)
2566                 sbi->s_depth_max = ext_depth(inode);
2567         spin_unlock(&sbi->s_ext_stats_lock);
2568 #endif
2569
2570         trace_ext4_remove_blocks(inode, ex, from, to, partial);
2571
2572         /*
2573          * if we have a partial cluster, and it's different from the
2574          * cluster of the last block in the extent, we free it
2575          */
2576         last_pblk = ext4_ext_pblock(ex) + ee_len - 1;
2577
2578         if (partial->state != initial &&
2579             partial->pclu != EXT4_B2C(sbi, last_pblk)) {
2580                 if (partial->state == tofree) {
2581                         flags = get_default_free_blocks_flags(inode);
2582                         if (ext4_is_pending(inode, partial->lblk))
2583                                 flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2584                         ext4_free_blocks(handle, inode, NULL,
2585                                          EXT4_C2B(sbi, partial->pclu),
2586                                          sbi->s_cluster_ratio, flags);
2587                         if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2588                                 ext4_rereserve_cluster(inode, partial->lblk);
2589                 }
2590                 partial->state = initial;
2591         }
2592
2593         num = le32_to_cpu(ex->ee_block) + ee_len - from;
2594         pblk = ext4_ext_pblock(ex) + ee_len - num;
2595
2596         /*
2597          * We free the partial cluster at the end of the extent (if any),
2598          * unless the cluster is used by another extent (partial_cluster
2599          * state is nofree).  If a partial cluster exists here, it must be
2600          * shared with the last block in the extent.
2601          */
2602         flags = get_default_free_blocks_flags(inode);
2603
2604         /* partial, left end cluster aligned, right end unaligned */
2605         if ((EXT4_LBLK_COFF(sbi, to) != sbi->s_cluster_ratio - 1) &&
2606             (EXT4_LBLK_CMASK(sbi, to) >= from) &&
2607             (partial->state != nofree)) {
2608                 if (ext4_is_pending(inode, to))
2609                         flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2610                 ext4_free_blocks(handle, inode, NULL,
2611                                  EXT4_PBLK_CMASK(sbi, last_pblk),
2612                                  sbi->s_cluster_ratio, flags);
2613                 if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2614                         ext4_rereserve_cluster(inode, to);
2615                 partial->state = initial;
2616                 flags = get_default_free_blocks_flags(inode);
2617         }
2618
2619         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2620
2621         /*
2622          * For bigalloc file systems, we never free a partial cluster
2623          * at the beginning of the extent.  Instead, we check to see if we
2624          * need to free it on a subsequent call to ext4_remove_blocks,
2625          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2626          */
2627         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2628         ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2629
2630         /* reset the partial cluster if we've freed past it */
2631         if (partial->state != initial && partial->pclu != EXT4_B2C(sbi, pblk))
2632                 partial->state = initial;
2633
2634         /*
2635          * If we've freed the entire extent but the beginning is not left
2636          * cluster aligned and is not marked as ineligible for freeing we
2637          * record the partial cluster at the beginning of the extent.  It
2638          * wasn't freed by the preceding ext4_free_blocks() call, and we
2639          * need to look farther to the left to determine if it's to be freed
2640          * (not shared with another extent). Else, reset the partial
2641          * cluster - we're either  done freeing or the beginning of the
2642          * extent is left cluster aligned.
2643          */
2644         if (EXT4_LBLK_COFF(sbi, from) && num == ee_len) {
2645                 if (partial->state == initial) {
2646                         partial->pclu = EXT4_B2C(sbi, pblk);
2647                         partial->lblk = from;
2648                         partial->state = tofree;
2649                 }
2650         } else {
2651                 partial->state = initial;
2652         }
2653
2654         return 0;
2655 }
2656
2657 /*
2658  * ext4_ext_rm_leaf() Removes the extents associated with the
2659  * blocks appearing between "start" and "end".  Both "start"
2660  * and "end" must appear in the same extent or EIO is returned.
2661  *
2662  * @handle: The journal handle
2663  * @inode:  The files inode
2664  * @path:   The path to the leaf
2665  * @partial_cluster: The cluster which we'll have to free if all extents
2666  *                   has been released from it.  However, if this value is
2667  *                   negative, it's a cluster just to the right of the
2668  *                   punched region and it must not be freed.
2669  * @start:  The first block to remove
2670  * @end:   The last block to remove
2671  */
2672 static int
2673 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2674                  struct ext4_ext_path *path,
2675                  struct partial_cluster *partial,
2676                  ext4_lblk_t start, ext4_lblk_t end)
2677 {
2678         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2679         int err = 0, correct_index = 0;
2680         int depth = ext_depth(inode), credits;
2681         struct ext4_extent_header *eh;
2682         ext4_lblk_t a, b;
2683         unsigned num;
2684         ext4_lblk_t ex_ee_block;
2685         unsigned short ex_ee_len;
2686         unsigned unwritten = 0;
2687         struct ext4_extent *ex;
2688         ext4_fsblk_t pblk;
2689
2690         /* the header must be checked already in ext4_ext_remove_space() */
2691         ext_debug("truncate since %u in leaf to %u\n", start, end);
2692         if (!path[depth].p_hdr)
2693                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2694         eh = path[depth].p_hdr;
2695         if (unlikely(path[depth].p_hdr == NULL)) {
2696                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2697                 return -EFSCORRUPTED;
2698         }
2699         /* find where to start removing */
2700         ex = path[depth].p_ext;
2701         if (!ex)
2702                 ex = EXT_LAST_EXTENT(eh);
2703
2704         ex_ee_block = le32_to_cpu(ex->ee_block);
2705         ex_ee_len = ext4_ext_get_actual_len(ex);
2706
2707         trace_ext4_ext_rm_leaf(inode, start, ex, partial);
2708
2709         while (ex >= EXT_FIRST_EXTENT(eh) &&
2710                         ex_ee_block + ex_ee_len > start) {
2711
2712                 if (ext4_ext_is_unwritten(ex))
2713                         unwritten = 1;
2714                 else
2715                         unwritten = 0;
2716
2717                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2718                           unwritten, ex_ee_len);
2719                 path[depth].p_ext = ex;
2720
2721                 a = ex_ee_block > start ? ex_ee_block : start;
2722                 b = ex_ee_block+ex_ee_len - 1 < end ?
2723                         ex_ee_block+ex_ee_len - 1 : end;
2724
2725                 ext_debug("  border %u:%u\n", a, b);
2726
2727                 /* If this extent is beyond the end of the hole, skip it */
2728                 if (end < ex_ee_block) {
2729                         /*
2730                          * We're going to skip this extent and move to another,
2731                          * so note that its first cluster is in use to avoid
2732                          * freeing it when removing blocks.  Eventually, the
2733                          * right edge of the truncated/punched region will
2734                          * be just to the left.
2735                          */
2736                         if (sbi->s_cluster_ratio > 1) {
2737                                 pblk = ext4_ext_pblock(ex);
2738                                 partial->pclu = EXT4_B2C(sbi, pblk);
2739                                 partial->state = nofree;
2740                         }
2741                         ex--;
2742                         ex_ee_block = le32_to_cpu(ex->ee_block);
2743                         ex_ee_len = ext4_ext_get_actual_len(ex);
2744                         continue;
2745                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2746                         EXT4_ERROR_INODE(inode,
2747                                          "can not handle truncate %u:%u "
2748                                          "on extent %u:%u",
2749                                          start, end, ex_ee_block,
2750                                          ex_ee_block + ex_ee_len - 1);
2751                         err = -EFSCORRUPTED;
2752                         goto out;
2753                 } else if (a != ex_ee_block) {
2754                         /* remove tail of the extent */
2755                         num = a - ex_ee_block;
2756                 } else {
2757                         /* remove whole extent: excellent! */
2758                         num = 0;
2759                 }
2760                 /*
2761                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2762                  * descriptor) for each block group; assume two block
2763                  * groups plus ex_ee_len/blocks_per_block_group for
2764                  * the worst case
2765                  */
2766                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2767                 if (ex == EXT_FIRST_EXTENT(eh)) {
2768                         correct_index = 1;
2769                         credits += (ext_depth(inode)) + 1;
2770                 }
2771                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2772
2773                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2774                 if (err)
2775                         goto out;
2776
2777                 err = ext4_ext_get_access(handle, inode, path + depth);
2778                 if (err)
2779                         goto out;
2780
2781                 err = ext4_remove_blocks(handle, inode, ex, partial, a, b);
2782                 if (err)
2783                         goto out;
2784
2785                 if (num == 0)
2786                         /* this extent is removed; mark slot entirely unused */
2787                         ext4_ext_store_pblock(ex, 0);
2788
2789                 ex->ee_len = cpu_to_le16(num);
2790                 /*
2791                  * Do not mark unwritten if all the blocks in the
2792                  * extent have been removed.
2793                  */
2794                 if (unwritten && num)
2795                         ext4_ext_mark_unwritten(ex);
2796                 /*
2797                  * If the extent was completely released,
2798                  * we need to remove it from the leaf
2799                  */
2800                 if (num == 0) {
2801                         if (end != EXT_MAX_BLOCKS - 1) {
2802                                 /*
2803                                  * For hole punching, we need to scoot all the
2804                                  * extents up when an extent is removed so that
2805                                  * we dont have blank extents in the middle
2806                                  */
2807                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2808                                         sizeof(struct ext4_extent));
2809
2810                                 /* Now get rid of the one at the end */
2811                                 memset(EXT_LAST_EXTENT(eh), 0,
2812                                         sizeof(struct ext4_extent));
2813                         }
2814                         le16_add_cpu(&eh->eh_entries, -1);
2815                 }
2816
2817                 err = ext4_ext_dirty(handle, inode, path + depth);
2818                 if (err)
2819                         goto out;
2820
2821                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2822                                 ext4_ext_pblock(ex));
2823                 ex--;
2824                 ex_ee_block = le32_to_cpu(ex->ee_block);
2825                 ex_ee_len = ext4_ext_get_actual_len(ex);
2826         }
2827
2828         if (correct_index && eh->eh_entries)
2829                 err = ext4_ext_correct_indexes(handle, inode, path);
2830
2831         /*
2832          * If there's a partial cluster and at least one extent remains in
2833          * the leaf, free the partial cluster if it isn't shared with the
2834          * current extent.  If it is shared with the current extent
2835          * we reset the partial cluster because we've reached the start of the
2836          * truncated/punched region and we're done removing blocks.
2837          */
2838         if (partial->state == tofree && ex >= EXT_FIRST_EXTENT(eh)) {
2839                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2840                 if (partial->pclu != EXT4_B2C(sbi, pblk)) {
2841                         int flags = get_default_free_blocks_flags(inode);
2842
2843                         if (ext4_is_pending(inode, partial->lblk))
2844                                 flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
2845                         ext4_free_blocks(handle, inode, NULL,
2846                                          EXT4_C2B(sbi, partial->pclu),
2847                                          sbi->s_cluster_ratio, flags);
2848                         if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
2849                                 ext4_rereserve_cluster(inode, partial->lblk);
2850                 }
2851                 partial->state = initial;
2852         }
2853
2854         /* if this leaf is free, then we should
2855          * remove it from index block above */
2856         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2857                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2858
2859 out:
2860         return err;
2861 }
2862
2863 /*
2864  * ext4_ext_more_to_rm:
2865  * returns 1 if current index has to be freed (even partial)
2866  */
2867 static int
2868 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2869 {
2870         BUG_ON(path->p_idx == NULL);
2871
2872         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2873                 return 0;
2874
2875         /*
2876          * if truncate on deeper level happened, it wasn't partial,
2877          * so we have to consider current index for truncation
2878          */
2879         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2880                 return 0;
2881         return 1;
2882 }
2883
2884 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2885                           ext4_lblk_t end)
2886 {
2887         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2888         int depth = ext_depth(inode);
2889         struct ext4_ext_path *path = NULL;
2890         struct partial_cluster partial;
2891         handle_t *handle;
2892         int i = 0, err = 0;
2893
2894         partial.pclu = 0;
2895         partial.lblk = 0;
2896         partial.state = initial;
2897
2898         ext_debug("truncate since %u to %u\n", start, end);
2899
2900         /* probably first extent we're gonna free will be last in block */
2901         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2902         if (IS_ERR(handle))
2903                 return PTR_ERR(handle);
2904
2905 again:
2906         trace_ext4_ext_remove_space(inode, start, end, depth);
2907
2908         /*
2909          * Check if we are removing extents inside the extent tree. If that
2910          * is the case, we are going to punch a hole inside the extent tree
2911          * so we have to check whether we need to split the extent covering
2912          * the last block to remove so we can easily remove the part of it
2913          * in ext4_ext_rm_leaf().
2914          */
2915         if (end < EXT_MAX_BLOCKS - 1) {
2916                 struct ext4_extent *ex;
2917                 ext4_lblk_t ee_block, ex_end, lblk;
2918                 ext4_fsblk_t pblk;
2919
2920                 /* find extent for or closest extent to this block */
2921                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2922                 if (IS_ERR(path)) {
2923                         ext4_journal_stop(handle);
2924                         return PTR_ERR(path);
2925                 }
2926                 depth = ext_depth(inode);
2927                 /* Leaf not may not exist only if inode has no blocks at all */
2928                 ex = path[depth].p_ext;
2929                 if (!ex) {
2930                         if (depth) {
2931                                 EXT4_ERROR_INODE(inode,
2932                                                  "path[%d].p_hdr == NULL",
2933                                                  depth);
2934                                 err = -EFSCORRUPTED;
2935                         }
2936                         goto out;
2937                 }
2938
2939                 ee_block = le32_to_cpu(ex->ee_block);
2940                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2941
2942                 /*
2943                  * See if the last block is inside the extent, if so split
2944                  * the extent at 'end' block so we can easily remove the
2945                  * tail of the first part of the split extent in
2946                  * ext4_ext_rm_leaf().
2947                  */
2948                 if (end >= ee_block && end < ex_end) {
2949
2950                         /*
2951                          * If we're going to split the extent, note that
2952                          * the cluster containing the block after 'end' is
2953                          * in use to avoid freeing it when removing blocks.
2954                          */
2955                         if (sbi->s_cluster_ratio > 1) {
2956                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2957                                 partial.pclu = EXT4_B2C(sbi, pblk);
2958                                 partial.state = nofree;
2959                         }
2960
2961                         /*
2962                          * Split the extent in two so that 'end' is the last
2963                          * block in the first new extent. Also we should not
2964                          * fail removing space due to ENOSPC so try to use
2965                          * reserved block if that happens.
2966                          */
2967                         err = ext4_force_split_extent_at(handle, inode, &path,
2968                                                          end + 1, 1);
2969                         if (err < 0)
2970                                 goto out;
2971
2972                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end &&
2973                            partial.state == initial) {
2974                         /*
2975                          * If we're punching, there's an extent to the right.
2976                          * If the partial cluster hasn't been set, set it to
2977                          * that extent's first cluster and its state to nofree
2978                          * so it won't be freed should it contain blocks to be
2979                          * removed. If it's already set (tofree/nofree), we're
2980                          * retrying and keep the original partial cluster info
2981                          * so a cluster marked tofree as a result of earlier
2982                          * extent removal is not lost.
2983                          */
2984                         lblk = ex_end + 1;
2985                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2986                                                     &ex);
2987                         if (err)
2988                                 goto out;
2989                         if (pblk) {
2990                                 partial.pclu = EXT4_B2C(sbi, pblk);
2991                                 partial.state = nofree;
2992                         }
2993                 }
2994         }
2995         /*
2996          * We start scanning from right side, freeing all the blocks
2997          * after i_size and walking into the tree depth-wise.
2998          */
2999         depth = ext_depth(inode);
3000         if (path) {
3001                 int k = i = depth;
3002                 while (--k > 0)
3003                         path[k].p_block =
3004                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
3005         } else {
3006                 path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
3007                                GFP_NOFS);
3008                 if (path == NULL) {
3009                         ext4_journal_stop(handle);
3010                         return -ENOMEM;
3011                 }
3012                 path[0].p_maxdepth = path[0].p_depth = depth;
3013                 path[0].p_hdr = ext_inode_hdr(inode);
3014                 i = 0;
3015
3016                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
3017                         err = -EFSCORRUPTED;
3018                         goto out;
3019                 }
3020         }
3021         err = 0;
3022
3023         while (i >= 0 && err == 0) {
3024                 if (i == depth) {
3025                         /* this is leaf block */
3026                         err = ext4_ext_rm_leaf(handle, inode, path,
3027                                                &partial, start, end);
3028                         /* root level has p_bh == NULL, brelse() eats this */
3029                         brelse(path[i].p_bh);
3030                         path[i].p_bh = NULL;
3031                         i--;
3032                         continue;
3033                 }
3034
3035                 /* this is index block */
3036                 if (!path[i].p_hdr) {
3037                         ext_debug("initialize header\n");
3038                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
3039                 }
3040
3041                 if (!path[i].p_idx) {
3042                         /* this level hasn't been touched yet */
3043                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
3044                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
3045                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
3046                                   path[i].p_hdr,
3047                                   le16_to_cpu(path[i].p_hdr->eh_entries));
3048                 } else {
3049                         /* we were already here, see at next index */
3050                         path[i].p_idx--;
3051                 }
3052
3053                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
3054                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
3055                                 path[i].p_idx);
3056                 if (ext4_ext_more_to_rm(path + i)) {
3057                         struct buffer_head *bh;
3058                         /* go to the next level */
3059                         ext_debug("move to level %d (block %llu)\n",
3060                                   i + 1, ext4_idx_pblock(path[i].p_idx));
3061                         memset(path + i + 1, 0, sizeof(*path));
3062                         bh = read_extent_tree_block(inode,
3063                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
3064                                 EXT4_EX_NOCACHE);
3065                         if (IS_ERR(bh)) {
3066                                 /* should we reset i_size? */
3067                                 err = PTR_ERR(bh);
3068                                 break;
3069                         }
3070                         /* Yield here to deal with large extent trees.
3071                          * Should be a no-op if we did IO above. */
3072                         cond_resched();
3073                         if (WARN_ON(i + 1 > depth)) {
3074                                 err = -EFSCORRUPTED;
3075                                 break;
3076                         }
3077                         path[i + 1].p_bh = bh;
3078
3079                         /* save actual number of indexes since this
3080                          * number is changed at the next iteration */
3081                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3082                         i++;
3083                 } else {
3084                         /* we finished processing this index, go up */
3085                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3086                                 /* index is empty, remove it;
3087                                  * handle must be already prepared by the
3088                                  * truncatei_leaf() */
3089                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3090                         }
3091                         /* root level has p_bh == NULL, brelse() eats this */
3092                         brelse(path[i].p_bh);
3093                         path[i].p_bh = NULL;
3094                         i--;
3095                         ext_debug("return to level %d\n", i);
3096                 }
3097         }
3098
3099         trace_ext4_ext_remove_space_done(inode, start, end, depth, &partial,
3100                                          path->p_hdr->eh_entries);
3101
3102         /*
3103          * if there's a partial cluster and we have removed the first extent
3104          * in the file, then we also free the partial cluster, if any
3105          */
3106         if (partial.state == tofree && err == 0) {
3107                 int flags = get_default_free_blocks_flags(inode);
3108
3109                 if (ext4_is_pending(inode, partial.lblk))
3110                         flags |= EXT4_FREE_BLOCKS_RERESERVE_CLUSTER;
3111                 ext4_free_blocks(handle, inode, NULL,
3112                                  EXT4_C2B(sbi, partial.pclu),
3113                                  sbi->s_cluster_ratio, flags);
3114                 if (flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)
3115                         ext4_rereserve_cluster(inode, partial.lblk);
3116                 partial.state = initial;
3117         }
3118
3119         /* TODO: flexible tree reduction should be here */
3120         if (path->p_hdr->eh_entries == 0) {
3121                 /*
3122                  * truncate to zero freed all the tree,
3123                  * so we need to correct eh_depth
3124                  */
3125                 err = ext4_ext_get_access(handle, inode, path);
3126                 if (err == 0) {
3127                         ext_inode_hdr(inode)->eh_depth = 0;
3128                         ext_inode_hdr(inode)->eh_max =
3129                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3130                         err = ext4_ext_dirty(handle, inode, path);
3131                 }
3132         }
3133 out:
3134         ext4_ext_drop_refs(path);
3135         kfree(path);
3136         path = NULL;
3137         if (err == -EAGAIN)
3138                 goto again;
3139         ext4_journal_stop(handle);
3140
3141         return err;
3142 }
3143
3144 /*
3145  * called at mount time
3146  */
3147 void ext4_ext_init(struct super_block *sb)
3148 {
3149         /*
3150          * possible initialization would be here
3151          */
3152
3153         if (ext4_has_feature_extents(sb)) {
3154 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3155                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3156 #ifdef AGGRESSIVE_TEST
3157                        ", aggressive tests"
3158 #endif
3159 #ifdef CHECK_BINSEARCH
3160                        ", check binsearch"
3161 #endif
3162 #ifdef EXTENTS_STATS
3163                        ", stats"
3164 #endif
3165                        "\n");
3166 #endif
3167 #ifdef EXTENTS_STATS
3168                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3169                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3170                 EXT4_SB(sb)->s_ext_max = 0;
3171 #endif
3172         }
3173 }
3174
3175 /*
3176  * called at umount time
3177  */
3178 void ext4_ext_release(struct super_block *sb)
3179 {
3180         if (!ext4_has_feature_extents(sb))
3181                 return;
3182
3183 #ifdef EXTENTS_STATS
3184         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3185                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3186                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3187                         sbi->s_ext_blocks, sbi->s_ext_extents,
3188                         sbi->s_ext_blocks / sbi->s_ext_extents);
3189                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3190                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3191         }
3192 #endif
3193 }
3194
3195 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3196 {
3197         ext4_lblk_t  ee_block;
3198         ext4_fsblk_t ee_pblock;
3199         unsigned int ee_len;
3200
3201         ee_block  = le32_to_cpu(ex->ee_block);
3202         ee_len    = ext4_ext_get_actual_len(ex);
3203         ee_pblock = ext4_ext_pblock(ex);
3204
3205         if (ee_len == 0)
3206                 return 0;
3207
3208         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3209                                      EXTENT_STATUS_WRITTEN);
3210 }
3211
3212 /* FIXME!! we need to try to merge to left or right after zero-out  */
3213 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3214 {
3215         ext4_fsblk_t ee_pblock;
3216         unsigned int ee_len;
3217
3218         ee_len    = ext4_ext_get_actual_len(ex);
3219         ee_pblock = ext4_ext_pblock(ex);
3220         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3221                                   ee_len);
3222 }
3223
3224 /*
3225  * ext4_split_extent_at() splits an extent at given block.
3226  *
3227  * @handle: the journal handle
3228  * @inode: the file inode
3229  * @path: the path to the extent
3230  * @split: the logical block where the extent is splitted.
3231  * @split_flags: indicates if the extent could be zeroout if split fails, and
3232  *               the states(init or unwritten) of new extents.
3233  * @flags: flags used to insert new extent to extent tree.
3234  *
3235  *
3236  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3237  * of which are deterimined by split_flag.
3238  *
3239  * There are two cases:
3240  *  a> the extent are splitted into two extent.
3241  *  b> split is not needed, and just mark the extent.
3242  *
3243  * return 0 on success.
3244  */
3245 static int ext4_split_extent_at(handle_t *handle,
3246                              struct inode *inode,
3247                              struct ext4_ext_path **ppath,
3248                              ext4_lblk_t split,
3249                              int split_flag,
3250                              int flags)
3251 {
3252         struct ext4_ext_path *path = *ppath;
3253         ext4_fsblk_t newblock;
3254         ext4_lblk_t ee_block;
3255         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3256         struct ext4_extent *ex2 = NULL;
3257         unsigned int ee_len, depth;
3258         int err = 0;
3259
3260         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3261                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3262
3263         ext_debug("ext4_split_extents_at: inode %lu, logical"
3264                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3265
3266         ext4_ext_show_leaf(inode, path);
3267
3268         depth = ext_depth(inode);
3269         ex = path[depth].p_ext;
3270         ee_block = le32_to_cpu(ex->ee_block);
3271         ee_len = ext4_ext_get_actual_len(ex);
3272         newblock = split - ee_block + ext4_ext_pblock(ex);
3273
3274         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3275         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3276                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3277                              EXT4_EXT_MARK_UNWRIT1 |
3278                              EXT4_EXT_MARK_UNWRIT2));
3279
3280         err = ext4_ext_get_access(handle, inode, path + depth);
3281         if (err)
3282                 goto out;
3283
3284         if (split == ee_block) {
3285                 /*
3286                  * case b: block @split is the block that the extent begins with
3287                  * then we just change the state of the extent, and splitting
3288                  * is not needed.
3289                  */
3290                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3291                         ext4_ext_mark_unwritten(ex);
3292                 else
3293                         ext4_ext_mark_initialized(ex);
3294
3295                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3296                         ext4_ext_try_to_merge(handle, inode, path, ex);
3297
3298                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3299                 goto out;
3300         }
3301
3302         /* case a */
3303         memcpy(&orig_ex, ex, sizeof(orig_ex));
3304         ex->ee_len = cpu_to_le16(split - ee_block);
3305         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3306                 ext4_ext_mark_unwritten(ex);
3307
3308         /*
3309          * path may lead to new leaf, not to original leaf any more
3310          * after ext4_ext_insert_extent() returns,
3311          */
3312         err = ext4_ext_dirty(handle, inode, path + depth);
3313         if (err)
3314                 goto fix_extent_len;
3315
3316         ex2 = &newex;
3317         ex2->ee_block = cpu_to_le32(split);
3318         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3319         ext4_ext_store_pblock(ex2, newblock);
3320         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3321                 ext4_ext_mark_unwritten(ex2);
3322
3323         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3324         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3325                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3326                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3327                                 err = ext4_ext_zeroout(inode, ex2);
3328                                 zero_ex.ee_block = ex2->ee_block;
3329                                 zero_ex.ee_len = cpu_to_le16(
3330                                                 ext4_ext_get_actual_len(ex2));
3331                                 ext4_ext_store_pblock(&zero_ex,
3332                                                       ext4_ext_pblock(ex2));
3333                         } else {
3334                                 err = ext4_ext_zeroout(inode, ex);
3335                                 zero_ex.ee_block = ex->ee_block;
3336                                 zero_ex.ee_len = cpu_to_le16(
3337                                                 ext4_ext_get_actual_len(ex));
3338                                 ext4_ext_store_pblock(&zero_ex,
3339                                                       ext4_ext_pblock(ex));
3340                         }
3341                 } else {
3342                         err = ext4_ext_zeroout(inode, &orig_ex);
3343                         zero_ex.ee_block = orig_ex.ee_block;
3344                         zero_ex.ee_len = cpu_to_le16(
3345                                                 ext4_ext_get_actual_len(&orig_ex));
3346                         ext4_ext_store_pblock(&zero_ex,
3347                                               ext4_ext_pblock(&orig_ex));
3348                 }
3349
3350                 if (err)
3351                         goto fix_extent_len;
3352                 /* update the extent length and mark as initialized */
3353                 ex->ee_len = cpu_to_le16(ee_len);
3354                 ext4_ext_try_to_merge(handle, inode, path, ex);
3355                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3356                 if (err)
3357                         goto fix_extent_len;
3358
3359                 /* update extent status tree */
3360                 err = ext4_zeroout_es(inode, &zero_ex);
3361
3362                 goto out;
3363         } else if (err)
3364                 goto fix_extent_len;
3365
3366 out:
3367         ext4_ext_show_leaf(inode, path);
3368         return err;
3369
3370 fix_extent_len:
3371         ex->ee_len = orig_ex.ee_len;
3372         ext4_ext_dirty(handle, inode, path + path->p_depth);
3373         return err;
3374 }
3375
3376 /*
3377  * ext4_split_extents() splits an extent and mark extent which is covered
3378  * by @map as split_flags indicates
3379  *
3380  * It may result in splitting the extent into multiple extents (up to three)
3381  * There are three possibilities:
3382  *   a> There is no split required
3383  *   b> Splits in two extents: Split is happening at either end of the extent
3384  *   c> Splits in three extents: Somone is splitting in middle of the extent
3385  *
3386  */
3387 static int ext4_split_extent(handle_t *handle,
3388                               struct inode *inode,
3389                               struct ext4_ext_path **ppath,
3390                               struct ext4_map_blocks *map,
3391                               int split_flag,
3392                               int flags)
3393 {
3394         struct ext4_ext_path *path = *ppath;
3395         ext4_lblk_t ee_block;
3396         struct ext4_extent *ex;
3397         unsigned int ee_len, depth;
3398         int err = 0;
3399         int unwritten;
3400         int split_flag1, flags1;
3401         int allocated = map->m_len;
3402
3403         depth = ext_depth(inode);
3404         ex = path[depth].p_ext;
3405         ee_block = le32_to_cpu(ex->ee_block);
3406         ee_len = ext4_ext_get_actual_len(ex);
3407         unwritten = ext4_ext_is_unwritten(ex);
3408
3409         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3410                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3411                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3412                 if (unwritten)
3413                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3414                                        EXT4_EXT_MARK_UNWRIT2;
3415                 if (split_flag & EXT4_EXT_DATA_VALID2)
3416                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3417                 err = ext4_split_extent_at(handle, inode, ppath,
3418                                 map->m_lblk + map->m_len, split_flag1, flags1);
3419                 if (err)
3420                         goto out;
3421         } else {
3422                 allocated = ee_len - (map->m_lblk - ee_block);
3423         }
3424         /*
3425          * Update path is required because previous ext4_split_extent_at() may
3426          * result in split of original leaf or extent zeroout.
3427          */
3428         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3429         if (IS_ERR(path))
3430                 return PTR_ERR(path);
3431         depth = ext_depth(inode);
3432         ex = path[depth].p_ext;
3433         if (!ex) {
3434                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3435                                  (unsigned long) map->m_lblk);
3436                 return -EFSCORRUPTED;
3437         }
3438         unwritten = ext4_ext_is_unwritten(ex);
3439         split_flag1 = 0;
3440
3441         if (map->m_lblk >= ee_block) {
3442                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3443                 if (unwritten) {
3444                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3445                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3446                                                      EXT4_EXT_MARK_UNWRIT2);
3447                 }
3448                 err = ext4_split_extent_at(handle, inode, ppath,
3449                                 map->m_lblk, split_flag1, flags);
3450                 if (err)
3451                         goto out;
3452         }
3453
3454         ext4_ext_show_leaf(inode, path);
3455 out:
3456         return err ? err : allocated;
3457 }
3458
3459 /*
3460  * This function is called by ext4_ext_map_blocks() if someone tries to write
3461  * to an unwritten extent. It may result in splitting the unwritten
3462  * extent into multiple extents (up to three - one initialized and two
3463  * unwritten).
3464  * There are three possibilities:
3465  *   a> There is no split required: Entire extent should be initialized
3466  *   b> Splits in two extents: Write is happening at either end of the extent
3467  *   c> Splits in three extents: Somone is writing in middle of the extent
3468  *
3469  * Pre-conditions:
3470  *  - The extent pointed to by 'path' is unwritten.
3471  *  - The extent pointed to by 'path' contains a superset
3472  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3473  *
3474  * Post-conditions on success:
3475  *  - the returned value is the number of blocks beyond map->l_lblk
3476  *    that are allocated and initialized.
3477  *    It is guaranteed to be >= map->m_len.
3478  */
3479 static int ext4_ext_convert_to_initialized(handle_t *handle,
3480                                            struct inode *inode,
3481                                            struct ext4_map_blocks *map,
3482                                            struct ext4_ext_path **ppath,
3483                                            int flags)
3484 {
3485         struct ext4_ext_path *path = *ppath;
3486         struct ext4_sb_info *sbi;
3487         struct ext4_extent_header *eh;
3488         struct ext4_map_blocks split_map;
3489         struct ext4_extent zero_ex1, zero_ex2;
3490         struct ext4_extent *ex, *abut_ex;
3491         ext4_lblk_t ee_block, eof_block;
3492         unsigned int ee_len, depth, map_len = map->m_len;
3493         int allocated = 0, max_zeroout = 0;
3494         int err = 0;
3495         int split_flag = EXT4_EXT_DATA_VALID2;
3496
3497         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3498                 "block %llu, max_blocks %u\n", inode->i_ino,
3499                 (unsigned long long)map->m_lblk, map_len);
3500
3501         sbi = EXT4_SB(inode->i_sb);
3502         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3503                 inode->i_sb->s_blocksize_bits;
3504         if (eof_block < map->m_lblk + map_len)
3505                 eof_block = map->m_lblk + map_len;
3506
3507         depth = ext_depth(inode);
3508         eh = path[depth].p_hdr;
3509         ex = path[depth].p_ext;
3510         ee_block = le32_to_cpu(ex->ee_block);
3511         ee_len = ext4_ext_get_actual_len(ex);
3512         zero_ex1.ee_len = 0;
3513         zero_ex2.ee_len = 0;
3514
3515         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3516
3517         /* Pre-conditions */
3518         BUG_ON(!ext4_ext_is_unwritten(ex));
3519         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3520
3521         /*
3522          * Attempt to transfer newly initialized blocks from the currently
3523          * unwritten extent to its neighbor. This is much cheaper
3524          * than an insertion followed by a merge as those involve costly
3525          * memmove() calls. Transferring to the left is the common case in
3526          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3527          * followed by append writes.
3528          *
3529          * Limitations of the current logic:
3530          *  - L1: we do not deal with writes covering the whole extent.
3531          *    This would require removing the extent if the transfer
3532          *    is possible.
3533          *  - L2: we only attempt to merge with an extent stored in the
3534          *    same extent tree node.
3535          */
3536         if ((map->m_lblk == ee_block) &&
3537                 /* See if we can merge left */
3538                 (map_len < ee_len) &&           /*L1*/
3539                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3540                 ext4_lblk_t prev_lblk;
3541                 ext4_fsblk_t prev_pblk, ee_pblk;
3542                 unsigned int prev_len;
3543
3544                 abut_ex = ex - 1;
3545                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3546                 prev_len = ext4_ext_get_actual_len(abut_ex);
3547                 prev_pblk = ext4_ext_pblock(abut_ex);
3548                 ee_pblk = ext4_ext_pblock(ex);
3549
3550                 /*
3551                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3552                  * upon those conditions:
3553                  * - C1: abut_ex is initialized,
3554                  * - C2: abut_ex is logically abutting ex,
3555                  * - C3: abut_ex is physically abutting ex,
3556                  * - C4: abut_ex can receive the additional blocks without
3557                  *   overflowing the (initialized) length limit.
3558                  */
3559                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3560                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3561                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3562                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3563                         err = ext4_ext_get_access(handle, inode, path + depth);
3564                         if (err)
3565                                 goto out;
3566
3567                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3568                                 map, ex, abut_ex);
3569
3570                         /* Shift the start of ex by 'map_len' blocks */
3571                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3572                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3573                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3574                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3575
3576                         /* Extend abut_ex by 'map_len' blocks */
3577                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3578
3579                         /* Result: number of initialized blocks past m_lblk */
3580                         allocated = map_len;
3581                 }
3582         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3583                    (map_len < ee_len) &&        /*L1*/
3584                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3585                 /* See if we can merge right */
3586                 ext4_lblk_t next_lblk;
3587                 ext4_fsblk_t next_pblk, ee_pblk;
3588                 unsigned int next_len;
3589
3590                 abut_ex = ex + 1;
3591                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3592                 next_len = ext4_ext_get_actual_len(abut_ex);
3593                 next_pblk = ext4_ext_pblock(abut_ex);
3594                 ee_pblk = ext4_ext_pblock(ex);
3595
3596                 /*
3597                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3598                  * upon those conditions:
3599                  * - C1: abut_ex is initialized,
3600                  * - C2: abut_ex is logically abutting ex,
3601                  * - C3: abut_ex is physically abutting ex,
3602                  * - C4: abut_ex can receive the additional blocks without
3603                  *   overflowing the (initialized) length limit.
3604                  */
3605                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3606                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3607                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3608                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3609                         err = ext4_ext_get_access(handle, inode, path + depth);
3610                         if (err)
3611                                 goto out;
3612
3613                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3614                                 map, ex, abut_ex);
3615
3616                         /* Shift the start of abut_ex by 'map_len' blocks */
3617                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3618                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3619                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3620                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3621
3622                         /* Extend abut_ex by 'map_len' blocks */
3623                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3624
3625                         /* Result: number of initialized blocks past m_lblk */
3626                         allocated = map_len;
3627                 }
3628         }
3629         if (allocated) {
3630                 /* Mark the block containing both extents as dirty */
3631                 ext4_ext_dirty(handle, inode, path + depth);
3632
3633                 /* Update path to point to the right extent */
3634                 path[depth].p_ext = abut_ex;
3635                 goto out;
3636         } else
3637                 allocated = ee_len - (map->m_lblk - ee_block);
3638
3639         WARN_ON(map->m_lblk < ee_block);
3640         /*
3641          * It is safe to convert extent to initialized via explicit
3642          * zeroout only if extent is fully inside i_size or new_size.
3643          */
3644         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3645
3646         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3647                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3648                         (inode->i_sb->s_blocksize_bits - 10);
3649
3650         if (IS_ENCRYPTED(inode))
3651                 max_zeroout = 0;
3652
3653         /*
3654          * five cases:
3655          * 1. split the extent into three extents.
3656          * 2. split the extent into two extents, zeroout the head of the first
3657          *    extent.
3658          * 3. split the extent into two extents, zeroout the tail of the second
3659          *    extent.
3660          * 4. split the extent into two extents with out zeroout.
3661          * 5. no splitting needed, just possibly zeroout the head and / or the
3662          *    tail of the extent.
3663          */
3664         split_map.m_lblk = map->m_lblk;
3665         split_map.m_len = map->m_len;
3666
3667         if (max_zeroout && (allocated > split_map.m_len)) {
3668                 if (allocated <= max_zeroout) {
3669                         /* case 3 or 5 */
3670                         zero_ex1.ee_block =
3671                                  cpu_to_le32(split_map.m_lblk +
3672                                              split_map.m_len);
3673                         zero_ex1.ee_len =
3674                                 cpu_to_le16(allocated - split_map.m_len);
3675                         ext4_ext_store_pblock(&zero_ex1,
3676                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3677                                 split_map.m_len - ee_block);
3678                         err = ext4_ext_zeroout(inode, &zero_ex1);
3679                         if (err)
3680                                 goto out;
3681                         split_map.m_len = allocated;
3682                 }
3683                 if (split_map.m_lblk - ee_block + split_map.m_len <
3684                                                                 max_zeroout) {
3685                         /* case 2 or 5 */
3686                         if (split_map.m_lblk != ee_block) {
3687                                 zero_ex2.ee_block = ex->ee_block;
3688                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3689                                                         ee_block);
3690                                 ext4_ext_store_pblock(&zero_ex2,
3691                                                       ext4_ext_pblock(ex));
3692                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3693                                 if (err)
3694                                         goto out;
3695                         }
3696
3697                         split_map.m_len += split_map.m_lblk - ee_block;
3698                         split_map.m_lblk = ee_block;
3699                         allocated = map->m_len;
3700                 }
3701         }
3702
3703         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3704                                 flags);
3705         if (err > 0)
3706                 err = 0;
3707 out:
3708         /* If we have gotten a failure, don't zero out status tree */
3709         if (!err) {
3710                 err = ext4_zeroout_es(inode, &zero_ex1);
3711                 if (!err)
3712                         err = ext4_zeroout_es(inode, &zero_ex2);
3713         }
3714         return err ? err : allocated;
3715 }
3716
3717 /*
3718  * This function is called by ext4_ext_map_blocks() from
3719  * ext4_get_blocks_dio_write() when DIO to write
3720  * to an unwritten extent.
3721  *
3722  * Writing to an unwritten extent may result in splitting the unwritten
3723  * extent into multiple initialized/unwritten extents (up to three)
3724  * There are three possibilities:
3725  *   a> There is no split required: Entire extent should be unwritten
3726  *   b> Splits in two extents: Write is happening at either end of the extent
3727  *   c> Splits in three extents: Somone is writing in middle of the extent
3728  *
3729  * This works the same way in the case of initialized -> unwritten conversion.
3730  *
3731  * One of more index blocks maybe needed if the extent tree grow after
3732  * the unwritten extent split. To prevent ENOSPC occur at the IO
3733  * complete, we need to split the unwritten extent before DIO submit
3734  * the IO. The unwritten extent called at this time will be split
3735  * into three unwritten extent(at most). After IO complete, the part
3736  * being filled will be convert to initialized by the end_io callback function
3737  * via ext4_convert_unwritten_extents().
3738  *
3739  * Returns the size of unwritten extent to be written on success.
3740  */
3741 static int ext4_split_convert_extents(handle_t *handle,
3742                                         struct inode *inode,
3743                                         struct ext4_map_blocks *map,
3744                                         struct ext4_ext_path **ppath,
3745                                         int flags)
3746 {
3747         struct ext4_ext_path *path = *ppath;
3748         ext4_lblk_t eof_block;
3749         ext4_lblk_t ee_block;
3750         struct ext4_extent *ex;
3751         unsigned int ee_len;
3752         int split_flag = 0, depth;
3753
3754         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3755                   __func__, inode->i_ino,
3756                   (unsigned long long)map->m_lblk, map->m_len);
3757
3758         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3759                 inode->i_sb->s_blocksize_bits;
3760         if (eof_block < map->m_lblk + map->m_len)
3761                 eof_block = map->m_lblk + map->m_len;
3762         /*
3763          * It is safe to convert extent to initialized via explicit
3764          * zeroout only if extent is fully insde i_size or new_size.
3765          */
3766         depth = ext_depth(inode);
3767         ex = path[depth].p_ext;
3768         ee_block = le32_to_cpu(ex->ee_block);
3769         ee_len = ext4_ext_get_actual_len(ex);
3770
3771         /* Convert to unwritten */
3772         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3773                 split_flag |= EXT4_EXT_DATA_VALID1;
3774         /* Convert to initialized */
3775         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3776                 split_flag |= ee_block + ee_len <= eof_block ?
3777                               EXT4_EXT_MAY_ZEROOUT : 0;
3778                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3779         }
3780         flags |= EXT4_GET_BLOCKS_PRE_IO;
3781         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3782 }
3783
3784 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3785                                                 struct inode *inode,
3786                                                 struct ext4_map_blocks *map,
3787                                                 struct ext4_ext_path **ppath)
3788 {
3789         struct ext4_ext_path *path = *ppath;
3790         struct ext4_extent *ex;
3791         ext4_lblk_t ee_block;
3792         unsigned int ee_len;
3793         int depth;
3794         int err = 0;
3795
3796         depth = ext_depth(inode);
3797         ex = path[depth].p_ext;
3798         ee_block = le32_to_cpu(ex->ee_block);
3799         ee_len = ext4_ext_get_actual_len(ex);
3800
3801         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3802                 "block %llu, max_blocks %u\n", inode->i_ino,
3803                   (unsigned long long)ee_block, ee_len);
3804
3805         /* If extent is larger than requested it is a clear sign that we still
3806          * have some extent state machine issues left. So extent_split is still
3807          * required.
3808          * TODO: Once all related issues will be fixed this situation should be
3809          * illegal.
3810          */
3811         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3812 #ifdef EXT4_DEBUG
3813                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3814                              " len %u; IO logical block %llu, len %u",
3815                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3816                              (unsigned long long)map->m_lblk, map->m_len);
3817 #endif
3818                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3819                                                  EXT4_GET_BLOCKS_CONVERT);
3820                 if (err < 0)
3821                         return err;
3822                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3823                 if (IS_ERR(path))
3824                         return PTR_ERR(path);
3825                 depth = ext_depth(inode);
3826                 ex = path[depth].p_ext;
3827         }
3828
3829         err = ext4_ext_get_access(handle, inode, path + depth);
3830         if (err)
3831                 goto out;
3832         /* first mark the extent as initialized */
3833         ext4_ext_mark_initialized(ex);
3834
3835         /* note: ext4_ext_correct_indexes() isn't needed here because
3836          * borders are not changed
3837          */
3838         ext4_ext_try_to_merge(handle, inode, path, ex);
3839
3840         /* Mark modified extent as dirty */
3841         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3842 out:
3843         ext4_ext_show_leaf(inode, path);
3844         return err;
3845 }
3846
3847 /*
3848  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3849  */
3850 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3851                               ext4_lblk_t lblk,
3852                               struct ext4_ext_path *path,
3853                               unsigned int len)
3854 {
3855         int i, depth;
3856         struct ext4_extent_header *eh;
3857         struct ext4_extent *last_ex;
3858
3859         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3860                 return 0;
3861
3862         depth = ext_depth(inode);
3863         eh = path[depth].p_hdr;
3864
3865         /*
3866          * We're going to remove EOFBLOCKS_FL entirely in future so we
3867          * do not care for this case anymore. Simply remove the flag
3868          * if there are no extents.
3869          */
3870         if (unlikely(!eh->eh_entries))
3871                 goto out;
3872         last_ex = EXT_LAST_EXTENT(eh);
3873         /*
3874          * We should clear the EOFBLOCKS_FL flag if we are writing the
3875          * last block in the last extent in the file.  We test this by
3876          * first checking to see if the caller to
3877          * ext4_ext_get_blocks() was interested in the last block (or
3878          * a block beyond the last block) in the current extent.  If
3879          * this turns out to be false, we can bail out from this
3880          * function immediately.
3881          */
3882         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3883             ext4_ext_get_actual_len(last_ex))
3884                 return 0;
3885         /*
3886          * If the caller does appear to be planning to write at or
3887          * beyond the end of the current extent, we then test to see
3888          * if the current extent is the last extent in the file, by
3889          * checking to make sure it was reached via the rightmost node
3890          * at each level of the tree.
3891          */
3892         for (i = depth-1; i >= 0; i--)
3893                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3894                         return 0;
3895 out:
3896         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3897         return ext4_mark_inode_dirty(handle, inode);
3898 }
3899
3900 static int
3901 convert_initialized_extent(handle_t *handle, struct inode *inode,
3902                            struct ext4_map_blocks *map,
3903                            struct ext4_ext_path **ppath,
3904                            unsigned int allocated)
3905 {
3906         struct ext4_ext_path *path = *ppath;
3907         struct ext4_extent *ex;
3908         ext4_lblk_t ee_block;
3909         unsigned int ee_len;
3910         int depth;
3911         int err = 0;
3912
3913         /*
3914          * Make sure that the extent is no bigger than we support with
3915          * unwritten extent
3916          */
3917         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3918                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3919
3920         depth = ext_depth(inode);
3921         ex = path[depth].p_ext;
3922         ee_block = le32_to_cpu(ex->ee_block);
3923         ee_len = ext4_ext_get_actual_len(ex);
3924
3925         ext_debug("%s: inode %lu, logical"
3926                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3927                   (unsigned long long)ee_block, ee_len);
3928
3929         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3930                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3931                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3932                 if (err < 0)
3933                         return err;
3934                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3935                 if (IS_ERR(path))
3936                         return PTR_ERR(path);
3937                 depth = ext_depth(inode);
3938                 ex = path[depth].p_ext;
3939                 if (!ex) {
3940                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3941                                          (unsigned long) map->m_lblk);
3942                         return -EFSCORRUPTED;
3943                 }
3944         }
3945
3946         err = ext4_ext_get_access(handle, inode, path + depth);
3947         if (err)
3948                 return err;
3949         /* first mark the extent as unwritten */
3950         ext4_ext_mark_unwritten(ex);
3951
3952         /* note: ext4_ext_correct_indexes() isn't needed here because
3953          * borders are not changed
3954          */
3955         ext4_ext_try_to_merge(handle, inode, path, ex);
3956
3957         /* Mark modified extent as dirty */
3958         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3959         if (err)
3960                 return err;
3961         ext4_ext_show_leaf(inode, path);
3962
3963         ext4_update_inode_fsync_trans(handle, inode, 1);
3964         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3965         if (err)
3966                 return err;
3967         map->m_flags |= EXT4_MAP_UNWRITTEN;
3968         if (allocated > map->m_len)
3969                 allocated = map->m_len;
3970         map->m_len = allocated;
3971         return allocated;
3972 }
3973
3974 static int
3975 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
3976                         struct ext4_map_blocks *map,
3977                         struct ext4_ext_path **ppath, int flags,
3978                         unsigned int allocated, ext4_fsblk_t newblock)
3979 {
3980         struct ext4_ext_path *path = *ppath;
3981         int ret = 0;
3982         int err = 0;
3983
3984         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
3985                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3986                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3987                   flags, allocated);
3988         ext4_ext_show_leaf(inode, path);
3989
3990         /*
3991          * When writing into unwritten space, we should not fail to
3992          * allocate metadata blocks for the new extent block if needed.
3993          */
3994         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
3995
3996         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
3997                                                     allocated, newblock);
3998
3999         /* get_block() before submit the IO, split the extent */
4000         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4001                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4002                                          flags | EXT4_GET_BLOCKS_CONVERT);
4003                 if (ret <= 0)
4004                         goto out;
4005                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4006                 goto out;
4007         }
4008         /* IO end_io complete, convert the filled extent to written */
4009         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4010                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4011                         if (allocated > map->m_len)
4012                                 allocated = map->m_len;
4013                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4014                                                  allocated);
4015                         if (err < 0)
4016                                 goto out2;
4017                 }
4018                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4019                                                            ppath);
4020                 if (ret >= 0) {
4021                         ext4_update_inode_fsync_trans(handle, inode, 1);
4022                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4023                                                  path, map->m_len);
4024                 } else
4025                         err = ret;
4026                 map->m_flags |= EXT4_MAP_MAPPED;
4027                 map->m_pblk = newblock;
4028                 if (allocated > map->m_len)
4029                         allocated = map->m_len;
4030                 map->m_len = allocated;
4031                 goto out2;
4032         }
4033         /* buffered IO case */
4034         /*
4035          * repeat fallocate creation request
4036          * we already have an unwritten extent
4037          */
4038         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4039                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4040                 goto map_out;
4041         }
4042
4043         /* buffered READ or buffered write_begin() lookup */
4044         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4045                 /*
4046                  * We have blocks reserved already.  We
4047                  * return allocated blocks so that delalloc
4048                  * won't do block reservation for us.  But
4049                  * the buffer head will be unmapped so that
4050                  * a read from the block returns 0s.
4051                  */
4052                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4053                 goto out1;
4054         }
4055
4056         /* buffered write, writepage time, convert*/
4057         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4058         if (ret >= 0)
4059                 ext4_update_inode_fsync_trans(handle, inode, 1);
4060 out:
4061         if (ret <= 0) {
4062                 err = ret;
4063                 goto out2;
4064         } else
4065                 allocated = ret;
4066         map->m_flags |= EXT4_MAP_NEW;
4067         if (allocated > map->m_len)
4068                 allocated = map->m_len;
4069         map->m_len = allocated;
4070
4071 map_out:
4072         map->m_flags |= EXT4_MAP_MAPPED;
4073         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4074                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4075                                          map->m_len);
4076                 if (err < 0)
4077                         goto out2;
4078         }
4079 out1:
4080         if (allocated > map->m_len)
4081                 allocated = map->m_len;
4082         ext4_ext_show_leaf(inode, path);
4083         map->m_pblk = newblock;
4084         map->m_len = allocated;
4085 out2:
4086         return err ? err : allocated;
4087 }
4088
4089 /*
4090  * get_implied_cluster_alloc - check to see if the requested
4091  * allocation (in the map structure) overlaps with a cluster already
4092  * allocated in an extent.
4093  *      @sb     The filesystem superblock structure
4094  *      @map    The requested lblk->pblk mapping
4095  *      @ex     The extent structure which might contain an implied
4096  *                      cluster allocation
4097  *
4098  * This function is called by ext4_ext_map_blocks() after we failed to
4099  * find blocks that were already in the inode's extent tree.  Hence,
4100  * we know that the beginning of the requested region cannot overlap
4101  * the extent from the inode's extent tree.  There are three cases we
4102  * want to catch.  The first is this case:
4103  *
4104  *               |--- cluster # N--|
4105  *    |--- extent ---|  |---- requested region ---|
4106  *                      |==========|
4107  *
4108  * The second case that we need to test for is this one:
4109  *
4110  *   |--------- cluster # N ----------------|
4111  *         |--- requested region --|   |------- extent ----|
4112  *         |=======================|
4113  *
4114  * The third case is when the requested region lies between two extents
4115  * within the same cluster:
4116  *          |------------- cluster # N-------------|
4117  * |----- ex -----|                  |---- ex_right ----|
4118  *                  |------ requested region ------|
4119  *                  |================|
4120  *
4121  * In each of the above cases, we need to set the map->m_pblk and
4122  * map->m_len so it corresponds to the return the extent labelled as
4123  * "|====|" from cluster #N, since it is already in use for data in
4124  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4125  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4126  * as a new "allocated" block region.  Otherwise, we will return 0 and
4127  * ext4_ext_map_blocks() will then allocate one or more new clusters
4128  * by calling ext4_mb_new_blocks().
4129  */
4130 static int get_implied_cluster_alloc(struct super_block *sb,
4131                                      struct ext4_map_blocks *map,
4132                                      struct ext4_extent *ex,
4133                                      struct ext4_ext_path *path)
4134 {
4135         struct ext4_sb_info *sbi = EXT4_SB(sb);
4136         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4137         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4138         ext4_lblk_t rr_cluster_start;
4139         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4140         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4141         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4142
4143         /* The extent passed in that we are trying to match */
4144         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4145         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4146
4147         /* The requested region passed into ext4_map_blocks() */
4148         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4149
4150         if ((rr_cluster_start == ex_cluster_end) ||
4151             (rr_cluster_start == ex_cluster_start)) {
4152                 if (rr_cluster_start == ex_cluster_end)
4153                         ee_start += ee_len - 1;
4154                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4155                 map->m_len = min(map->m_len,
4156                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4157                 /*
4158                  * Check for and handle this case:
4159                  *
4160                  *   |--------- cluster # N-------------|
4161                  *                     |------- extent ----|
4162                  *         |--- requested region ---|
4163                  *         |===========|
4164                  */
4165
4166                 if (map->m_lblk < ee_block)
4167                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4168
4169                 /*
4170                  * Check for the case where there is already another allocated
4171                  * block to the right of 'ex' but before the end of the cluster.
4172                  *
4173                  *          |------------- cluster # N-------------|
4174                  * |----- ex -----|                  |---- ex_right ----|
4175                  *                  |------ requested region ------|
4176                  *                  |================|
4177                  */
4178                 if (map->m_lblk > ee_block) {
4179                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4180                         map->m_len = min(map->m_len, next - map->m_lblk);
4181                 }
4182
4183                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4184                 return 1;
4185         }
4186
4187         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4188         return 0;
4189 }
4190
4191
4192 /*
4193  * Block allocation/map/preallocation routine for extents based files
4194  *
4195  *
4196  * Need to be called with
4197  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4198  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4199  *
4200  * return > 0, number of of blocks already mapped/allocated
4201  *          if create == 0 and these are pre-allocated blocks
4202  *              buffer head is unmapped
4203  *          otherwise blocks are mapped
4204  *
4205  * return = 0, if plain look up failed (blocks have not been allocated)
4206  *          buffer head is unmapped
4207  *
4208  * return < 0, error case.
4209  */
4210 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4211                         struct ext4_map_blocks *map, int flags)
4212 {
4213         struct ext4_ext_path *path = NULL;
4214         struct ext4_extent newex, *ex, *ex2;
4215         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4216         ext4_fsblk_t newblock = 0;
4217         int free_on_err = 0, err = 0, depth, ret;
4218         unsigned int allocated = 0, offset = 0;
4219         unsigned int allocated_clusters = 0;
4220         struct ext4_allocation_request ar;
4221         ext4_lblk_t cluster_offset;
4222         bool map_from_cluster = false;
4223
4224         ext_debug("blocks %u/%u requested for inode %lu\n",
4225                   map->m_lblk, map->m_len, inode->i_ino);
4226         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4227
4228         /* find extent for this block */
4229         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4230         if (IS_ERR(path)) {
4231                 err = PTR_ERR(path);
4232                 path = NULL;
4233                 goto out2;
4234         }
4235
4236         depth = ext_depth(inode);
4237
4238         /*
4239          * consistent leaf must not be empty;
4240          * this situation is possible, though, _during_ tree modification;
4241          * this is why assert can't be put in ext4_find_extent()
4242          */
4243         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4244                 EXT4_ERROR_INODE(inode, "bad extent address "
4245                                  "lblock: %lu, depth: %d pblock %lld",
4246                                  (unsigned long) map->m_lblk, depth,
4247                                  path[depth].p_block);
4248                 err = -EFSCORRUPTED;
4249                 goto out2;
4250         }
4251
4252         ex = path[depth].p_ext;
4253         if (ex) {
4254                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4255                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4256                 unsigned short ee_len;
4257
4258
4259                 /*
4260                  * unwritten extents are treated as holes, except that
4261                  * we split out initialized portions during a write.
4262                  */
4263                 ee_len = ext4_ext_get_actual_len(ex);
4264
4265                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4266
4267                 /* if found extent covers block, simply return it */
4268                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4269                         newblock = map->m_lblk - ee_block + ee_start;
4270                         /* number of remaining blocks in the extent */
4271                         allocated = ee_len - (map->m_lblk - ee_block);
4272                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4273                                   ee_block, ee_len, newblock);
4274
4275                         /*
4276                          * If the extent is initialized check whether the
4277                          * caller wants to convert it to unwritten.
4278                          */
4279                         if ((!ext4_ext_is_unwritten(ex)) &&
4280                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4281                                 allocated = convert_initialized_extent(
4282                                                 handle, inode, map, &path,
4283                                                 allocated);
4284                                 goto out2;
4285                         } else if (!ext4_ext_is_unwritten(ex))
4286                                 goto out;
4287
4288                         ret = ext4_ext_handle_unwritten_extents(
4289                                 handle, inode, map, &path, flags,
4290                                 allocated, newblock);
4291                         if (ret < 0)
4292                                 err = ret;
4293                         else
4294                                 allocated = ret;
4295                         goto out2;
4296                 }
4297         }
4298
4299         /*
4300          * requested block isn't allocated yet;
4301          * we couldn't try to create block if create flag is zero
4302          */
4303         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4304                 ext4_lblk_t hole_start, hole_len;
4305
4306                 hole_start = map->m_lblk;
4307                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4308                 /*
4309                  * put just found gap into cache to speed up
4310                  * subsequent requests
4311                  */
4312                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4313
4314                 /* Update hole_len to reflect hole size after map->m_lblk */
4315                 if (hole_start != map->m_lblk)
4316                         hole_len -= map->m_lblk - hole_start;
4317                 map->m_pblk = 0;
4318                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4319
4320                 goto out2;
4321         }
4322
4323         /*
4324          * Okay, we need to do block allocation.
4325          */
4326         newex.ee_block = cpu_to_le32(map->m_lblk);
4327         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4328
4329         /*
4330          * If we are doing bigalloc, check to see if the extent returned
4331          * by ext4_find_extent() implies a cluster we can use.
4332          */
4333         if (cluster_offset && ex &&
4334             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4335                 ar.len = allocated = map->m_len;
4336                 newblock = map->m_pblk;
4337                 map_from_cluster = true;
4338                 goto got_allocated_blocks;
4339         }
4340
4341         /* find neighbour allocated blocks */
4342         ar.lleft = map->m_lblk;
4343         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4344         if (err)
4345                 goto out2;
4346         ar.lright = map->m_lblk;
4347         ex2 = NULL;
4348         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4349         if (err)
4350                 goto out2;
4351
4352         /* Check if the extent after searching to the right implies a
4353          * cluster we can use. */
4354         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4355             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4356                 ar.len = allocated = map->m_len;
4357                 newblock = map->m_pblk;
4358                 map_from_cluster = true;
4359                 goto got_allocated_blocks;
4360         }
4361
4362         /*
4363          * See if request is beyond maximum number of blocks we can have in
4364          * a single extent. For an initialized extent this limit is
4365          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4366          * EXT_UNWRITTEN_MAX_LEN.
4367          */
4368         if (map->m_len > EXT_INIT_MAX_LEN &&
4369             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4370                 map->m_len = EXT_INIT_MAX_LEN;
4371         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4372                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4373                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4374
4375         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4376         newex.ee_len = cpu_to_le16(map->m_len);
4377         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4378         if (err)
4379                 allocated = ext4_ext_get_actual_len(&newex);
4380         else
4381                 allocated = map->m_len;
4382
4383         /* allocate new block */
4384         ar.inode = inode;
4385         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4386         ar.logical = map->m_lblk;
4387         /*
4388          * We calculate the offset from the beginning of the cluster
4389          * for the logical block number, since when we allocate a
4390          * physical cluster, the physical block should start at the
4391          * same offset from the beginning of the cluster.  This is
4392          * needed so that future calls to get_implied_cluster_alloc()
4393          * work correctly.
4394          */
4395         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4396         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4397         ar.goal -= offset;
4398         ar.logical -= offset;
4399         if (S_ISREG(inode->i_mode))
4400                 ar.flags = EXT4_MB_HINT_DATA;
4401         else
4402                 /* disable in-core preallocation for non-regular files */
4403                 ar.flags = 0;
4404         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4405                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4406         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4407                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4408         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4409                 ar.flags |= EXT4_MB_USE_RESERVED;
4410         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4411         if (!newblock)
4412                 goto out2;
4413         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4414                   ar.goal, newblock, allocated);
4415         free_on_err = 1;
4416         allocated_clusters = ar.len;
4417         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4418         if (ar.len > allocated)
4419                 ar.len = allocated;
4420
4421 got_allocated_blocks:
4422         /* try to insert new extent into found leaf and return */
4423         ext4_ext_store_pblock(&newex, newblock + offset);
4424         newex.ee_len = cpu_to_le16(ar.len);
4425         /* Mark unwritten */
4426         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4427                 ext4_ext_mark_unwritten(&newex);
4428                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4429         }
4430
4431         err = 0;
4432         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4433                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4434                                          path, ar.len);
4435         if (!err)
4436                 err = ext4_ext_insert_extent(handle, inode, &path,
4437                                              &newex, flags);
4438
4439         if (err && free_on_err) {
4440                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4441                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4442                 /* free data blocks we just allocated */
4443                 /* not a good idea to call discard here directly,
4444                  * but otherwise we'd need to call it every free() */
4445                 ext4_discard_preallocations(inode);
4446                 ext4_free_blocks(handle, inode, NULL, newblock,
4447                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4448                 goto out2;
4449         }
4450
4451         /* previous routine could use block we allocated */
4452         newblock = ext4_ext_pblock(&newex);
4453         allocated = ext4_ext_get_actual_len(&newex);
4454         if (allocated > map->m_len)
4455                 allocated = map->m_len;
4456         map->m_flags |= EXT4_MAP_NEW;
4457
4458         /*
4459          * Reduce the reserved cluster count to reflect successful deferred
4460          * allocation of delayed allocated clusters or direct allocation of
4461          * clusters discovered to be delayed allocated.  Once allocated, a
4462          * cluster is not included in the reserved count.
4463          */
4464         if (test_opt(inode->i_sb, DELALLOC) && !map_from_cluster) {
4465                 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4466                         /*
4467                          * When allocating delayed allocated clusters, simply
4468                          * reduce the reserved cluster count and claim quota
4469                          */
4470                         ext4_da_update_reserve_space(inode, allocated_clusters,
4471                                                         1);
4472                 } else {
4473                         ext4_lblk_t lblk, len;
4474                         unsigned int n;
4475
4476                         /*
4477                          * When allocating non-delayed allocated clusters
4478                          * (from fallocate, filemap, DIO, or clusters
4479                          * allocated when delalloc has been disabled by
4480                          * ext4_nonda_switch), reduce the reserved cluster
4481                          * count by the number of allocated clusters that
4482                          * have previously been delayed allocated.  Quota
4483                          * has been claimed by ext4_mb_new_blocks() above,
4484                          * so release the quota reservations made for any
4485                          * previously delayed allocated clusters.
4486                          */
4487                         lblk = EXT4_LBLK_CMASK(sbi, map->m_lblk);
4488                         len = allocated_clusters << sbi->s_cluster_bits;
4489                         n = ext4_es_delayed_clu(inode, lblk, len);
4490                         if (n > 0)
4491                                 ext4_da_update_reserve_space(inode, (int) n, 0);
4492                 }
4493         }
4494
4495         /*
4496          * Cache the extent and update transaction to commit on fdatasync only
4497          * when it is _not_ an unwritten extent.
4498          */
4499         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4500                 ext4_update_inode_fsync_trans(handle, inode, 1);
4501         else
4502                 ext4_update_inode_fsync_trans(handle, inode, 0);
4503 out:
4504         if (allocated > map->m_len)
4505                 allocated = map->m_len;
4506         ext4_ext_show_leaf(inode, path);
4507         map->m_flags |= EXT4_MAP_MAPPED;
4508         map->m_pblk = newblock;
4509         map->m_len = allocated;
4510 out2:
4511         ext4_ext_drop_refs(path);
4512         kfree(path);
4513
4514         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4515                                        err ? err : allocated);
4516         return err ? err : allocated;
4517 }
4518
4519 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4520 {
4521         struct super_block *sb = inode->i_sb;
4522         ext4_lblk_t last_block;
4523         int err = 0;
4524
4525         /*
4526          * TODO: optimization is possible here.
4527          * Probably we need not scan at all,
4528          * because page truncation is enough.
4529          */
4530
4531         /* we have to know where to truncate from in crash case */
4532         EXT4_I(inode)->i_disksize = inode->i_size;
4533         err = ext4_mark_inode_dirty(handle, inode);
4534         if (err)
4535                 return err;
4536
4537         last_block = (inode->i_size + sb->s_blocksize - 1)
4538                         >> EXT4_BLOCK_SIZE_BITS(sb);
4539 retry:
4540         err = ext4_es_remove_extent(inode, last_block,
4541                                     EXT_MAX_BLOCKS - last_block);
4542         if (err == -ENOMEM) {
4543                 cond_resched();
4544                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4545                 goto retry;
4546         }
4547         if (err)
4548                 return err;
4549         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4550 }
4551
4552 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4553                                   ext4_lblk_t len, loff_t new_size,
4554                                   int flags)
4555 {
4556         struct inode *inode = file_inode(file);
4557         handle_t *handle;
4558         int ret = 0;
4559         int ret2 = 0;
4560         int retries = 0;
4561         int depth = 0;
4562         struct ext4_map_blocks map;
4563         unsigned int credits;
4564         loff_t epos;
4565
4566         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4567         map.m_lblk = offset;
4568         map.m_len = len;
4569         /*
4570          * Don't normalize the request if it can fit in one extent so
4571          * that it doesn't get unnecessarily split into multiple
4572          * extents.
4573          */
4574         if (len <= EXT_UNWRITTEN_MAX_LEN)
4575                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4576
4577         /*
4578          * credits to insert 1 extent into extent tree
4579          */
4580         credits = ext4_chunk_trans_blocks(inode, len);
4581         depth = ext_depth(inode);
4582
4583 retry:
4584         while (ret >= 0 && len) {
4585                 /*
4586                  * Recalculate credits when extent tree depth changes.
4587                  */
4588                 if (depth != ext_depth(inode)) {
4589                         credits = ext4_chunk_trans_blocks(inode, len);
4590                         depth = ext_depth(inode);
4591                 }
4592
4593                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4594                                             credits);
4595                 if (IS_ERR(handle)) {
4596                         ret = PTR_ERR(handle);
4597                         break;
4598                 }
4599                 ret = ext4_map_blocks(handle, inode, &map, flags);
4600                 if (ret <= 0) {
4601                         ext4_debug("inode #%lu: block %u: len %u: "
4602                                    "ext4_ext_map_blocks returned %d",
4603                                    inode->i_ino, map.m_lblk,
4604                                    map.m_len, ret);
4605                         ext4_mark_inode_dirty(handle, inode);
4606                         ret2 = ext4_journal_stop(handle);
4607                         break;
4608                 }
4609                 map.m_lblk += ret;
4610                 map.m_len = len = len - ret;
4611                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4612                 inode->i_ctime = current_time(inode);
4613                 if (new_size) {
4614                         if (epos > new_size)
4615                                 epos = new_size;
4616                         if (ext4_update_inode_size(inode, epos) & 0x1)
4617                                 inode->i_mtime = inode->i_ctime;
4618                 } else {
4619                         if (epos > inode->i_size)
4620                                 ext4_set_inode_flag(inode,
4621                                                     EXT4_INODE_EOFBLOCKS);
4622                 }
4623                 ext4_mark_inode_dirty(handle, inode);
4624                 ext4_update_inode_fsync_trans(handle, inode, 1);
4625                 ret2 = ext4_journal_stop(handle);
4626                 if (ret2)
4627                         break;
4628         }
4629         if (ret == -ENOSPC &&
4630                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4631                 ret = 0;
4632                 goto retry;
4633         }
4634
4635         return ret > 0 ? ret2 : ret;
4636 }
4637
4638 static long ext4_zero_range(struct file *file, loff_t offset,
4639                             loff_t len, int mode)
4640 {
4641         struct inode *inode = file_inode(file);
4642         handle_t *handle = NULL;
4643         unsigned int max_blocks;
4644         loff_t new_size = 0;
4645         int ret = 0;
4646         int flags;
4647         int credits;
4648         int partial_begin, partial_end;
4649         loff_t start, end;
4650         ext4_lblk_t lblk;
4651         unsigned int blkbits = inode->i_blkbits;
4652
4653         trace_ext4_zero_range(inode, offset, len, mode);
4654
4655         if (!S_ISREG(inode->i_mode))
4656                 return -EINVAL;
4657
4658         /* Call ext4_force_commit to flush all data in case of data=journal. */
4659         if (ext4_should_journal_data(inode)) {
4660                 ret = ext4_force_commit(inode->i_sb);
4661                 if (ret)
4662                         return ret;
4663         }
4664
4665         /*
4666          * Round up offset. This is not fallocate, we neet to zero out
4667          * blocks, so convert interior block aligned part of the range to
4668          * unwritten and possibly manually zero out unaligned parts of the
4669          * range.
4670          */
4671         start = round_up(offset, 1 << blkbits);
4672         end = round_down((offset + len), 1 << blkbits);
4673
4674         if (start < offset || end > offset + len)
4675                 return -EINVAL;
4676         partial_begin = offset & ((1 << blkbits) - 1);
4677         partial_end = (offset + len) & ((1 << blkbits) - 1);
4678
4679         lblk = start >> blkbits;
4680         max_blocks = (end >> blkbits);
4681         if (max_blocks < lblk)
4682                 max_blocks = 0;
4683         else
4684                 max_blocks -= lblk;
4685
4686         inode_lock(inode);
4687
4688         /*
4689          * Indirect files do not support unwritten extnets
4690          */
4691         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4692                 ret = -EOPNOTSUPP;
4693                 goto out_mutex;
4694         }
4695
4696         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4697             (offset + len > i_size_read(inode) ||
4698              offset + len > EXT4_I(inode)->i_disksize)) {
4699                 new_size = offset + len;
4700                 ret = inode_newsize_ok(inode, new_size);
4701                 if (ret)
4702                         goto out_mutex;
4703         }
4704
4705         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4706         if (mode & FALLOC_FL_KEEP_SIZE)
4707                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4708
4709         /* Wait all existing dio workers, newcomers will block on i_mutex */
4710         inode_dio_wait(inode);
4711
4712         /* Preallocate the range including the unaligned edges */
4713         if (partial_begin || partial_end) {
4714                 ret = ext4_alloc_file_blocks(file,
4715                                 round_down(offset, 1 << blkbits) >> blkbits,
4716                                 (round_up((offset + len), 1 << blkbits) -
4717                                  round_down(offset, 1 << blkbits)) >> blkbits,
4718                                 new_size, flags);
4719                 if (ret)
4720                         goto out_mutex;
4721
4722         }
4723
4724         /* Zero range excluding the unaligned edges */
4725         if (max_blocks > 0) {
4726                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4727                           EXT4_EX_NOCACHE);
4728
4729                 /*
4730                  * Prevent page faults from reinstantiating pages we have
4731                  * released from page cache.
4732                  */
4733                 down_write(&EXT4_I(inode)->i_mmap_sem);
4734
4735                 ret = ext4_break_layouts(inode);
4736                 if (ret) {
4737                         up_write(&EXT4_I(inode)->i_mmap_sem);
4738                         goto out_mutex;
4739                 }
4740
4741                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4742                 if (ret) {
4743                         up_write(&EXT4_I(inode)->i_mmap_sem);
4744                         goto out_mutex;
4745                 }
4746                 /* Now release the pages and zero block aligned part of pages */
4747                 truncate_pagecache_range(inode, start, end - 1);
4748                 inode->i_mtime = inode->i_ctime = current_time(inode);
4749
4750                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4751                                              flags);
4752                 up_write(&EXT4_I(inode)->i_mmap_sem);
4753                 if (ret)
4754                         goto out_mutex;
4755         }
4756         if (!partial_begin && !partial_end)
4757                 goto out_mutex;
4758
4759         /*
4760          * In worst case we have to writeout two nonadjacent unwritten
4761          * blocks and update the inode
4762          */
4763         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4764         if (ext4_should_journal_data(inode))
4765                 credits += 2;
4766         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4767         if (IS_ERR(handle)) {
4768                 ret = PTR_ERR(handle);
4769                 ext4_std_error(inode->i_sb, ret);
4770                 goto out_mutex;
4771         }
4772
4773         inode->i_mtime = inode->i_ctime = current_time(inode);
4774         if (new_size) {
4775                 ext4_update_inode_size(inode, new_size);
4776         } else {
4777                 /*
4778                 * Mark that we allocate beyond EOF so the subsequent truncate
4779                 * can proceed even if the new size is the same as i_size.
4780                 */
4781                 if ((offset + len) > i_size_read(inode))
4782                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4783         }
4784         ext4_mark_inode_dirty(handle, inode);
4785
4786         /* Zero out partial block at the edges of the range */
4787         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4788         if (ret >= 0)
4789                 ext4_update_inode_fsync_trans(handle, inode, 1);
4790
4791         if (file->f_flags & O_SYNC)
4792                 ext4_handle_sync(handle);
4793
4794         ext4_journal_stop(handle);
4795 out_mutex:
4796         inode_unlock(inode);
4797         return ret;
4798 }
4799
4800 /*
4801  * preallocate space for a file. This implements ext4's fallocate file
4802  * operation, which gets called from sys_fallocate system call.
4803  * For block-mapped files, posix_fallocate should fall back to the method
4804  * of writing zeroes to the required new blocks (the same behavior which is
4805  * expected for file systems which do not support fallocate() system call).
4806  */
4807 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4808 {
4809         struct inode *inode = file_inode(file);
4810         loff_t new_size = 0;
4811         unsigned int max_blocks;
4812         int ret = 0;
4813         int flags;
4814         ext4_lblk_t lblk;
4815         unsigned int blkbits = inode->i_blkbits;
4816
4817         /*
4818          * Encrypted inodes can't handle collapse range or insert
4819          * range since we would need to re-encrypt blocks with a
4820          * different IV or XTS tweak (which are based on the logical
4821          * block number).
4822          *
4823          * XXX It's not clear why zero range isn't working, but we'll
4824          * leave it disabled for encrypted inodes for now.  This is a
4825          * bug we should fix....
4826          */
4827         if (IS_ENCRYPTED(inode) &&
4828             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4829                      FALLOC_FL_ZERO_RANGE)))
4830                 return -EOPNOTSUPP;
4831
4832         /* Return error if mode is not supported */
4833         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4834                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4835                      FALLOC_FL_INSERT_RANGE))
4836                 return -EOPNOTSUPP;
4837
4838         if (mode & FALLOC_FL_PUNCH_HOLE)
4839                 return ext4_punch_hole(inode, offset, len);
4840
4841         ret = ext4_convert_inline_data(inode);
4842         if (ret)
4843                 return ret;
4844
4845         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4846                 return ext4_collapse_range(inode, offset, len);
4847
4848         if (mode & FALLOC_FL_INSERT_RANGE)
4849                 return ext4_insert_range(inode, offset, len);
4850
4851         if (mode & FALLOC_FL_ZERO_RANGE)
4852                 return ext4_zero_range(file, offset, len, mode);
4853
4854         trace_ext4_fallocate_enter(inode, offset, len, mode);
4855         lblk = offset >> blkbits;
4856
4857         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4858         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4859         if (mode & FALLOC_FL_KEEP_SIZE)
4860                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4861
4862         inode_lock(inode);
4863
4864         /*
4865          * We only support preallocation for extent-based files only
4866          */
4867         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4868                 ret = -EOPNOTSUPP;
4869                 goto out;
4870         }
4871
4872         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4873             (offset + len > i_size_read(inode) ||
4874              offset + len > EXT4_I(inode)->i_disksize)) {
4875                 new_size = offset + len;
4876                 ret = inode_newsize_ok(inode, new_size);
4877                 if (ret)
4878                         goto out;
4879         }
4880
4881         /* Wait all existing dio workers, newcomers will block on i_mutex */
4882         inode_dio_wait(inode);
4883
4884         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
4885         if (ret)
4886                 goto out;
4887
4888         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4889                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4890                                                 EXT4_I(inode)->i_sync_tid);
4891         }
4892 out:
4893         inode_unlock(inode);
4894         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4895         return ret;
4896 }
4897
4898 /*
4899  * This function convert a range of blocks to written extents
4900  * The caller of this function will pass the start offset and the size.
4901  * all unwritten extents within this range will be converted to
4902  * written extents.
4903  *
4904  * This function is called from the direct IO end io call back
4905  * function, to convert the fallocated extents after IO is completed.
4906  * Returns 0 on success.
4907  */
4908 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4909                                    loff_t offset, ssize_t len)
4910 {
4911         unsigned int max_blocks;
4912         int ret = 0;
4913         int ret2 = 0;
4914         struct ext4_map_blocks map;
4915         unsigned int credits, blkbits = inode->i_blkbits;
4916
4917         map.m_lblk = offset >> blkbits;
4918         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4919
4920         /*
4921          * This is somewhat ugly but the idea is clear: When transaction is
4922          * reserved, everything goes into it. Otherwise we rather start several
4923          * smaller transactions for conversion of each extent separately.
4924          */
4925         if (handle) {
4926                 handle = ext4_journal_start_reserved(handle,
4927                                                      EXT4_HT_EXT_CONVERT);
4928                 if (IS_ERR(handle))
4929                         return PTR_ERR(handle);
4930                 credits = 0;
4931         } else {
4932                 /*
4933                  * credits to insert 1 extent into extent tree
4934                  */
4935                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
4936         }
4937         while (ret >= 0 && ret < max_blocks) {
4938                 map.m_lblk += ret;
4939                 map.m_len = (max_blocks -= ret);
4940                 if (credits) {
4941                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4942                                                     credits);
4943                         if (IS_ERR(handle)) {
4944                                 ret = PTR_ERR(handle);
4945                                 break;
4946                         }
4947                 }
4948                 ret = ext4_map_blocks(handle, inode, &map,
4949                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4950                 if (ret <= 0)
4951                         ext4_warning(inode->i_sb,
4952                                      "inode #%lu: block %u: len %u: "
4953                                      "ext4_ext_map_blocks returned %d",
4954                                      inode->i_ino, map.m_lblk,
4955                                      map.m_len, ret);
4956                 ext4_mark_inode_dirty(handle, inode);
4957                 if (credits)
4958                         ret2 = ext4_journal_stop(handle);
4959                 if (ret <= 0 || ret2)
4960                         break;
4961         }
4962         if (!credits)
4963                 ret2 = ext4_journal_stop(handle);
4964         return ret > 0 ? ret2 : ret;
4965 }
4966
4967 /*
4968  * If newes is not existing extent (newes->ec_pblk equals zero) find
4969  * delayed extent at start of newes and update newes accordingly and
4970  * return start of the next delayed extent.
4971  *
4972  * If newes is existing extent (newes->ec_pblk is not equal zero)
4973  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4974  * extent found. Leave newes unmodified.
4975  */
4976 static int ext4_find_delayed_extent(struct inode *inode,
4977                                     struct extent_status *newes)
4978 {
4979         struct extent_status es;
4980         ext4_lblk_t block, next_del;
4981
4982         if (newes->es_pblk == 0) {
4983                 ext4_es_find_extent_range(inode, &ext4_es_is_delayed,
4984                                           newes->es_lblk,
4985                                           newes->es_lblk + newes->es_len - 1,
4986                                           &es);
4987
4988                 /*
4989                  * No extent in extent-tree contains block @newes->es_pblk,
4990                  * then the block may stay in 1)a hole or 2)delayed-extent.
4991                  */
4992                 if (es.es_len == 0)
4993                         /* A hole found. */
4994                         return 0;
4995
4996                 if (es.es_lblk > newes->es_lblk) {
4997                         /* A hole found. */
4998                         newes->es_len = min(es.es_lblk - newes->es_lblk,
4999                                             newes->es_len);
5000                         return 0;
5001                 }
5002
5003                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5004         }
5005
5006         block = newes->es_lblk + newes->es_len;
5007         ext4_es_find_extent_range(inode, &ext4_es_is_delayed, block,
5008                                   EXT_MAX_BLOCKS, &es);
5009         if (es.es_len == 0)
5010                 next_del = EXT_MAX_BLOCKS;
5011         else
5012                 next_del = es.es_lblk;
5013
5014         return next_del;
5015 }
5016 /* fiemap flags we can handle specified here */
5017 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5018
5019 static int ext4_xattr_fiemap(struct inode *inode,
5020                                 struct fiemap_extent_info *fieinfo)
5021 {
5022         __u64 physical = 0;
5023         __u64 length;
5024         __u32 flags = FIEMAP_EXTENT_LAST;
5025         int blockbits = inode->i_sb->s_blocksize_bits;
5026         int error = 0;
5027
5028         /* in-inode? */
5029         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5030                 struct ext4_iloc iloc;
5031                 int offset;     /* offset of xattr in inode */
5032
5033                 error = ext4_get_inode_loc(inode, &iloc);
5034                 if (error)
5035                         return error;
5036                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5037                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5038                                 EXT4_I(inode)->i_extra_isize;
5039                 physical += offset;
5040                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5041                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5042                 brelse(iloc.bh);
5043         } else { /* external block */
5044                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5045                 length = inode->i_sb->s_blocksize;
5046         }
5047
5048         if (physical)
5049                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5050                                                 length, flags);
5051         return (error < 0 ? error : 0);
5052 }
5053
5054 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5055                 __u64 start, __u64 len)
5056 {
5057         ext4_lblk_t start_blk;
5058         int error = 0;
5059
5060         if (ext4_has_inline_data(inode)) {
5061                 int has_inline = 1;
5062
5063                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5064                                                 start, len);
5065
5066                 if (has_inline)
5067                         return error;
5068         }
5069
5070         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5071                 error = ext4_ext_precache(inode);
5072                 if (error)
5073                         return error;
5074         }
5075
5076         /* fallback to generic here if not in extents fmt */
5077         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5078                 return generic_block_fiemap(inode, fieinfo, start, len,
5079                         ext4_get_block);
5080
5081         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5082                 return -EBADR;
5083
5084         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5085                 error = ext4_xattr_fiemap(inode, fieinfo);
5086         } else {
5087                 ext4_lblk_t len_blks;
5088                 __u64 last_blk;
5089
5090                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5091                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5092                 if (last_blk >= EXT_MAX_BLOCKS)
5093                         last_blk = EXT_MAX_BLOCKS-1;
5094                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5095
5096                 /*
5097                  * Walk the extent tree gathering extent information
5098                  * and pushing extents back to the user.
5099                  */
5100                 error = ext4_fill_fiemap_extents(inode, start_blk,
5101                                                  len_blks, fieinfo);
5102         }
5103         return error;
5104 }
5105
5106 /*
5107  * ext4_access_path:
5108  * Function to access the path buffer for marking it dirty.
5109  * It also checks if there are sufficient credits left in the journal handle
5110  * to update path.
5111  */
5112 static int
5113 ext4_access_path(handle_t *handle, struct inode *inode,
5114                 struct ext4_ext_path *path)
5115 {
5116         int credits, err;
5117
5118         if (!ext4_handle_valid(handle))
5119                 return 0;
5120
5121         /*
5122          * Check if need to extend journal credits
5123          * 3 for leaf, sb, and inode plus 2 (bmap and group
5124          * descriptor) for each block group; assume two block
5125          * groups
5126          */
5127         if (handle->h_buffer_credits < 7) {
5128                 credits = ext4_writepage_trans_blocks(inode);
5129                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5130                 /* EAGAIN is success */
5131                 if (err && err != -EAGAIN)
5132                         return err;
5133         }
5134
5135         err = ext4_ext_get_access(handle, inode, path);
5136         return err;
5137 }
5138
5139 /*
5140  * ext4_ext_shift_path_extents:
5141  * Shift the extents of a path structure lying between path[depth].p_ext
5142  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5143  * if it is right shift or left shift operation.
5144  */
5145 static int
5146 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5147                             struct inode *inode, handle_t *handle,
5148                             enum SHIFT_DIRECTION SHIFT)
5149 {
5150         int depth, err = 0;
5151         struct ext4_extent *ex_start, *ex_last;
5152         bool update = 0;
5153         depth = path->p_depth;
5154
5155         while (depth >= 0) {
5156                 if (depth == path->p_depth) {
5157                         ex_start = path[depth].p_ext;
5158                         if (!ex_start)
5159                                 return -EFSCORRUPTED;
5160
5161                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5162
5163                         err = ext4_access_path(handle, inode, path + depth);
5164                         if (err)
5165                                 goto out;
5166
5167                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5168                                 update = 1;
5169
5170                         while (ex_start <= ex_last) {
5171                                 if (SHIFT == SHIFT_LEFT) {
5172                                         le32_add_cpu(&ex_start->ee_block,
5173                                                 -shift);
5174                                         /* Try to merge to the left. */
5175                                         if ((ex_start >
5176                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5177                                             &&
5178                                             ext4_ext_try_to_merge_right(inode,
5179                                             path, ex_start - 1))
5180                                                 ex_last--;
5181                                         else
5182                                                 ex_start++;
5183                                 } else {
5184                                         le32_add_cpu(&ex_last->ee_block, shift);
5185                                         ext4_ext_try_to_merge_right(inode, path,
5186                                                 ex_last);
5187                                         ex_last--;
5188                                 }
5189                         }
5190                         err = ext4_ext_dirty(handle, inode, path + depth);
5191                         if (err)
5192                                 goto out;
5193
5194                         if (--depth < 0 || !update)
5195                                 break;
5196                 }
5197
5198                 /* Update index too */
5199                 err = ext4_access_path(handle, inode, path + depth);
5200                 if (err)
5201                         goto out;
5202
5203                 if (SHIFT == SHIFT_LEFT)
5204                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5205                 else
5206                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5207                 err = ext4_ext_dirty(handle, inode, path + depth);
5208                 if (err)
5209                         goto out;
5210
5211                 /* we are done if current index is not a starting index */
5212                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5213                         break;
5214
5215                 depth--;
5216         }
5217
5218 out:
5219         return err;
5220 }
5221
5222 /*
5223  * ext4_ext_shift_extents:
5224  * All the extents which lies in the range from @start to the last allocated
5225  * block for the @inode are shifted either towards left or right (depending
5226  * upon @SHIFT) by @shift blocks.
5227  * On success, 0 is returned, error otherwise.
5228  */
5229 static int
5230 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5231                        ext4_lblk_t start, ext4_lblk_t shift,
5232                        enum SHIFT_DIRECTION SHIFT)
5233 {
5234         struct ext4_ext_path *path;
5235         int ret = 0, depth;
5236         struct ext4_extent *extent;
5237         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5238
5239         /* Let path point to the last extent */
5240         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5241                                 EXT4_EX_NOCACHE);
5242         if (IS_ERR(path))
5243                 return PTR_ERR(path);
5244
5245         depth = path->p_depth;
5246         extent = path[depth].p_ext;
5247         if (!extent)
5248                 goto out;
5249
5250         stop = le32_to_cpu(extent->ee_block);
5251
5252        /*
5253         * For left shifts, make sure the hole on the left is big enough to
5254         * accommodate the shift.  For right shifts, make sure the last extent
5255         * won't be shifted beyond EXT_MAX_BLOCKS.
5256         */
5257         if (SHIFT == SHIFT_LEFT) {
5258                 path = ext4_find_extent(inode, start - 1, &path,
5259                                         EXT4_EX_NOCACHE);
5260                 if (IS_ERR(path))
5261                         return PTR_ERR(path);
5262                 depth = path->p_depth;
5263                 extent =  path[depth].p_ext;
5264                 if (extent) {
5265                         ex_start = le32_to_cpu(extent->ee_block);
5266                         ex_end = le32_to_cpu(extent->ee_block) +
5267                                 ext4_ext_get_actual_len(extent);
5268                 } else {
5269                         ex_start = 0;
5270                         ex_end = 0;
5271                 }
5272
5273                 if ((start == ex_start && shift > ex_start) ||
5274                     (shift > start - ex_end)) {
5275                         ret = -EINVAL;
5276                         goto out;
5277                 }
5278         } else {
5279                 if (shift > EXT_MAX_BLOCKS -
5280                     (stop + ext4_ext_get_actual_len(extent))) {
5281                         ret = -EINVAL;
5282                         goto out;
5283                 }
5284         }
5285
5286         /*
5287          * In case of left shift, iterator points to start and it is increased
5288          * till we reach stop. In case of right shift, iterator points to stop
5289          * and it is decreased till we reach start.
5290          */
5291         if (SHIFT == SHIFT_LEFT)
5292                 iterator = &start;
5293         else
5294                 iterator = &stop;
5295
5296         /*
5297          * Its safe to start updating extents.  Start and stop are unsigned, so
5298          * in case of right shift if extent with 0 block is reached, iterator
5299          * becomes NULL to indicate the end of the loop.
5300          */
5301         while (iterator && start <= stop) {
5302                 path = ext4_find_extent(inode, *iterator, &path,
5303                                         EXT4_EX_NOCACHE);
5304                 if (IS_ERR(path))
5305                         return PTR_ERR(path);
5306                 depth = path->p_depth;
5307                 extent = path[depth].p_ext;
5308                 if (!extent) {
5309                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5310                                          (unsigned long) *iterator);
5311                         return -EFSCORRUPTED;
5312                 }
5313                 if (SHIFT == SHIFT_LEFT && *iterator >
5314                     le32_to_cpu(extent->ee_block)) {
5315                         /* Hole, move to the next extent */
5316                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5317                                 path[depth].p_ext++;
5318                         } else {
5319                                 *iterator = ext4_ext_next_allocated_block(path);
5320                                 continue;
5321                         }
5322                 }
5323
5324                 if (SHIFT == SHIFT_LEFT) {
5325                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5326                         *iterator = le32_to_cpu(extent->ee_block) +
5327                                         ext4_ext_get_actual_len(extent);
5328                 } else {
5329                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5330                         if (le32_to_cpu(extent->ee_block) > 0)
5331                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5332                         else
5333                                 /* Beginning is reached, end of the loop */
5334                                 iterator = NULL;
5335                         /* Update path extent in case we need to stop */
5336                         while (le32_to_cpu(extent->ee_block) < start)
5337                                 extent++;
5338                         path[depth].p_ext = extent;
5339                 }
5340                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5341                                 handle, SHIFT);
5342                 if (ret)
5343                         break;
5344         }
5345 out:
5346         ext4_ext_drop_refs(path);
5347         kfree(path);
5348         return ret;
5349 }
5350
5351 /*
5352  * ext4_collapse_range:
5353  * This implements the fallocate's collapse range functionality for ext4
5354  * Returns: 0 and non-zero on error.
5355  */
5356 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5357 {
5358         struct super_block *sb = inode->i_sb;
5359         ext4_lblk_t punch_start, punch_stop;
5360         handle_t *handle;
5361         unsigned int credits;
5362         loff_t new_size, ioffset;
5363         int ret;
5364
5365         /*
5366          * We need to test this early because xfstests assumes that a
5367          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5368          * system does not support collapse range.
5369          */
5370         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5371                 return -EOPNOTSUPP;
5372
5373         /* Collapse range works only on fs block size aligned offsets. */
5374         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5375             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5376                 return -EINVAL;
5377
5378         if (!S_ISREG(inode->i_mode))
5379                 return -EINVAL;
5380
5381         trace_ext4_collapse_range(inode, offset, len);
5382
5383         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5384         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5385
5386         /* Call ext4_force_commit to flush all data in case of data=journal. */
5387         if (ext4_should_journal_data(inode)) {
5388                 ret = ext4_force_commit(inode->i_sb);
5389                 if (ret)
5390                         return ret;
5391         }
5392
5393         inode_lock(inode);
5394         /*
5395          * There is no need to overlap collapse range with EOF, in which case
5396          * it is effectively a truncate operation
5397          */
5398         if (offset + len >= i_size_read(inode)) {
5399                 ret = -EINVAL;
5400                 goto out_mutex;
5401         }
5402
5403         /* Currently just for extent based files */
5404         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5405                 ret = -EOPNOTSUPP;
5406                 goto out_mutex;
5407         }
5408
5409         /* Wait for existing dio to complete */
5410         inode_dio_wait(inode);
5411
5412         /*
5413          * Prevent page faults from reinstantiating pages we have released from
5414          * page cache.
5415          */
5416         down_write(&EXT4_I(inode)->i_mmap_sem);
5417
5418         ret = ext4_break_layouts(inode);
5419         if (ret)
5420                 goto out_mmap;
5421
5422         /*
5423          * Need to round down offset to be aligned with page size boundary
5424          * for page size > block size.
5425          */
5426         ioffset = round_down(offset, PAGE_SIZE);
5427         /*
5428          * Write tail of the last page before removed range since it will get
5429          * removed from the page cache below.
5430          */
5431         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5432         if (ret)
5433                 goto out_mmap;
5434         /*
5435          * Write data that will be shifted to preserve them when discarding
5436          * page cache below. We are also protected from pages becoming dirty
5437          * by i_mmap_sem.
5438          */
5439         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5440                                            LLONG_MAX);
5441         if (ret)
5442                 goto out_mmap;
5443         truncate_pagecache(inode, ioffset);
5444
5445         credits = ext4_writepage_trans_blocks(inode);
5446         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5447         if (IS_ERR(handle)) {
5448                 ret = PTR_ERR(handle);
5449                 goto out_mmap;
5450         }
5451
5452         down_write(&EXT4_I(inode)->i_data_sem);
5453         ext4_discard_preallocations(inode);
5454
5455         ret = ext4_es_remove_extent(inode, punch_start,
5456                                     EXT_MAX_BLOCKS - punch_start);
5457         if (ret) {
5458                 up_write(&EXT4_I(inode)->i_data_sem);
5459                 goto out_stop;
5460         }
5461
5462         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5463         if (ret) {
5464                 up_write(&EXT4_I(inode)->i_data_sem);
5465                 goto out_stop;
5466         }
5467         ext4_discard_preallocations(inode);
5468
5469         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5470                                      punch_stop - punch_start, SHIFT_LEFT);
5471         if (ret) {
5472                 up_write(&EXT4_I(inode)->i_data_sem);
5473                 goto out_stop;
5474         }
5475
5476         new_size = i_size_read(inode) - len;
5477         i_size_write(inode, new_size);
5478         EXT4_I(inode)->i_disksize = new_size;
5479
5480         up_write(&EXT4_I(inode)->i_data_sem);
5481         if (IS_SYNC(inode))
5482                 ext4_handle_sync(handle);
5483         inode->i_mtime = inode->i_ctime = current_time(inode);
5484         ext4_mark_inode_dirty(handle, inode);
5485         ext4_update_inode_fsync_trans(handle, inode, 1);
5486
5487 out_stop:
5488         ext4_journal_stop(handle);
5489 out_mmap:
5490         up_write(&EXT4_I(inode)->i_mmap_sem);
5491 out_mutex:
5492         inode_unlock(inode);
5493         return ret;
5494 }
5495
5496 /*
5497  * ext4_insert_range:
5498  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5499  * The data blocks starting from @offset to the EOF are shifted by @len
5500  * towards right to create a hole in the @inode. Inode size is increased
5501  * by len bytes.
5502  * Returns 0 on success, error otherwise.
5503  */
5504 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5505 {
5506         struct super_block *sb = inode->i_sb;
5507         handle_t *handle;
5508         struct ext4_ext_path *path;
5509         struct ext4_extent *extent;
5510         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5511         unsigned int credits, ee_len;
5512         int ret = 0, depth, split_flag = 0;
5513         loff_t ioffset;
5514
5515         /*
5516          * We need to test this early because xfstests assumes that an
5517          * insert range of (0, 1) will return EOPNOTSUPP if the file
5518          * system does not support insert range.
5519          */
5520         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5521                 return -EOPNOTSUPP;
5522
5523         /* Insert range works only on fs block size aligned offsets. */
5524         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5525                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5526                 return -EINVAL;
5527
5528         if (!S_ISREG(inode->i_mode))
5529                 return -EOPNOTSUPP;
5530
5531         trace_ext4_insert_range(inode, offset, len);
5532
5533         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5534         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5535
5536         /* Call ext4_force_commit to flush all data in case of data=journal */
5537         if (ext4_should_journal_data(inode)) {
5538                 ret = ext4_force_commit(inode->i_sb);
5539                 if (ret)
5540                         return ret;
5541         }
5542
5543         inode_lock(inode);
5544         /* Currently just for extent based files */
5545         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5546                 ret = -EOPNOTSUPP;
5547                 goto out_mutex;
5548         }
5549
5550         /* Check for wrap through zero */
5551         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5552                 ret = -EFBIG;
5553                 goto out_mutex;
5554         }
5555
5556         /* Offset should be less than i_size */
5557         if (offset >= i_size_read(inode)) {
5558                 ret = -EINVAL;
5559                 goto out_mutex;
5560         }
5561
5562         /* Wait for existing dio to complete */
5563         inode_dio_wait(inode);
5564
5565         /*
5566          * Prevent page faults from reinstantiating pages we have released from
5567          * page cache.
5568          */
5569         down_write(&EXT4_I(inode)->i_mmap_sem);
5570
5571         ret = ext4_break_layouts(inode);
5572         if (ret)
5573                 goto out_mmap;
5574
5575         /*
5576          * Need to round down to align start offset to page size boundary
5577          * for page size > block size.
5578          */
5579         ioffset = round_down(offset, PAGE_SIZE);
5580         /* Write out all dirty pages */
5581         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5582                         LLONG_MAX);
5583         if (ret)
5584                 goto out_mmap;
5585         truncate_pagecache(inode, ioffset);
5586
5587         credits = ext4_writepage_trans_blocks(inode);
5588         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5589         if (IS_ERR(handle)) {
5590                 ret = PTR_ERR(handle);
5591                 goto out_mmap;
5592         }
5593
5594         /* Expand file to avoid data loss if there is error while shifting */
5595         inode->i_size += len;
5596         EXT4_I(inode)->i_disksize += len;
5597         inode->i_mtime = inode->i_ctime = current_time(inode);
5598         ret = ext4_mark_inode_dirty(handle, inode);
5599         if (ret)
5600                 goto out_stop;
5601
5602         down_write(&EXT4_I(inode)->i_data_sem);
5603         ext4_discard_preallocations(inode);
5604
5605         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5606         if (IS_ERR(path)) {
5607                 up_write(&EXT4_I(inode)->i_data_sem);
5608                 goto out_stop;
5609         }
5610
5611         depth = ext_depth(inode);
5612         extent = path[depth].p_ext;
5613         if (extent) {
5614                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5615                 ee_len = ext4_ext_get_actual_len(extent);
5616
5617                 /*
5618                  * If offset_lblk is not the starting block of extent, split
5619                  * the extent @offset_lblk
5620                  */
5621                 if ((offset_lblk > ee_start_lblk) &&
5622                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5623                         if (ext4_ext_is_unwritten(extent))
5624                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5625                                         EXT4_EXT_MARK_UNWRIT2;
5626                         ret = ext4_split_extent_at(handle, inode, &path,
5627                                         offset_lblk, split_flag,
5628                                         EXT4_EX_NOCACHE |
5629                                         EXT4_GET_BLOCKS_PRE_IO |
5630                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5631                 }
5632
5633                 ext4_ext_drop_refs(path);
5634                 kfree(path);
5635                 if (ret < 0) {
5636                         up_write(&EXT4_I(inode)->i_data_sem);
5637                         goto out_stop;
5638                 }
5639         } else {
5640                 ext4_ext_drop_refs(path);
5641                 kfree(path);
5642         }
5643
5644         ret = ext4_es_remove_extent(inode, offset_lblk,
5645                         EXT_MAX_BLOCKS - offset_lblk);
5646         if (ret) {
5647                 up_write(&EXT4_I(inode)->i_data_sem);
5648                 goto out_stop;
5649         }
5650
5651         /*
5652          * if offset_lblk lies in a hole which is at start of file, use
5653          * ee_start_lblk to shift extents
5654          */
5655         ret = ext4_ext_shift_extents(inode, handle,
5656                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5657                 len_lblk, SHIFT_RIGHT);
5658
5659         up_write(&EXT4_I(inode)->i_data_sem);
5660         if (IS_SYNC(inode))
5661                 ext4_handle_sync(handle);
5662         if (ret >= 0)
5663                 ext4_update_inode_fsync_trans(handle, inode, 1);
5664
5665 out_stop:
5666         ext4_journal_stop(handle);
5667 out_mmap:
5668         up_write(&EXT4_I(inode)->i_mmap_sem);
5669 out_mutex:
5670         inode_unlock(inode);
5671         return ret;
5672 }
5673
5674 /**
5675  * ext4_swap_extents - Swap extents between two inodes
5676  *
5677  * @inode1:     First inode
5678  * @inode2:     Second inode
5679  * @lblk1:      Start block for first inode
5680  * @lblk2:      Start block for second inode
5681  * @count:      Number of blocks to swap
5682  * @unwritten: Mark second inode's extents as unwritten after swap
5683  * @erp:        Pointer to save error value
5684  *
5685  * This helper routine does exactly what is promise "swap extents". All other
5686  * stuff such as page-cache locking consistency, bh mapping consistency or
5687  * extent's data copying must be performed by caller.
5688  * Locking:
5689  *              i_mutex is held for both inodes
5690  *              i_data_sem is locked for write for both inodes
5691  * Assumptions:
5692  *              All pages from requested range are locked for both inodes
5693  */
5694 int
5695 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5696                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5697                   ext4_lblk_t count, int unwritten, int *erp)
5698 {
5699         struct ext4_ext_path *path1 = NULL;
5700         struct ext4_ext_path *path2 = NULL;
5701         int replaced_count = 0;
5702
5703         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5704         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5705         BUG_ON(!inode_is_locked(inode1));
5706         BUG_ON(!inode_is_locked(inode2));
5707
5708         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5709         if (unlikely(*erp))
5710                 return 0;
5711         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5712         if (unlikely(*erp))
5713                 return 0;
5714
5715         while (count) {
5716                 struct ext4_extent *ex1, *ex2, tmp_ex;
5717                 ext4_lblk_t e1_blk, e2_blk;
5718                 int e1_len, e2_len, len;
5719                 int split = 0;
5720
5721                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5722                 if (IS_ERR(path1)) {
5723                         *erp = PTR_ERR(path1);
5724                         path1 = NULL;
5725                 finish:
5726                         count = 0;
5727                         goto repeat;
5728                 }
5729                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5730                 if (IS_ERR(path2)) {
5731                         *erp = PTR_ERR(path2);
5732                         path2 = NULL;
5733                         goto finish;
5734                 }
5735                 ex1 = path1[path1->p_depth].p_ext;
5736                 ex2 = path2[path2->p_depth].p_ext;
5737                 /* Do we have somthing to swap ? */
5738                 if (unlikely(!ex2 || !ex1))
5739                         goto finish;
5740
5741                 e1_blk = le32_to_cpu(ex1->ee_block);
5742                 e2_blk = le32_to_cpu(ex2->ee_block);
5743                 e1_len = ext4_ext_get_actual_len(ex1);
5744                 e2_len = ext4_ext_get_actual_len(ex2);
5745
5746                 /* Hole handling */
5747                 if (!in_range(lblk1, e1_blk, e1_len) ||
5748                     !in_range(lblk2, e2_blk, e2_len)) {
5749                         ext4_lblk_t next1, next2;
5750
5751                         /* if hole after extent, then go to next extent */
5752                         next1 = ext4_ext_next_allocated_block(path1);
5753                         next2 = ext4_ext_next_allocated_block(path2);
5754                         /* If hole before extent, then shift to that extent */
5755                         if (e1_blk > lblk1)
5756                                 next1 = e1_blk;
5757                         if (e2_blk > lblk2)
5758                                 next2 = e2_blk;
5759                         /* Do we have something to swap */
5760                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5761                                 goto finish;
5762                         /* Move to the rightest boundary */
5763                         len = next1 - lblk1;
5764                         if (len < next2 - lblk2)
5765                                 len = next2 - lblk2;
5766                         if (len > count)
5767                                 len = count;
5768                         lblk1 += len;
5769                         lblk2 += len;
5770                         count -= len;
5771                         goto repeat;
5772                 }
5773
5774                 /* Prepare left boundary */
5775                 if (e1_blk < lblk1) {
5776                         split = 1;
5777                         *erp = ext4_force_split_extent_at(handle, inode1,
5778                                                 &path1, lblk1, 0);
5779                         if (unlikely(*erp))
5780                                 goto finish;
5781                 }
5782                 if (e2_blk < lblk2) {
5783                         split = 1;
5784                         *erp = ext4_force_split_extent_at(handle, inode2,
5785                                                 &path2,  lblk2, 0);
5786                         if (unlikely(*erp))
5787                                 goto finish;
5788                 }
5789                 /* ext4_split_extent_at() may result in leaf extent split,
5790                  * path must to be revalidated. */
5791                 if (split)
5792                         goto repeat;
5793
5794                 /* Prepare right boundary */
5795                 len = count;
5796                 if (len > e1_blk + e1_len - lblk1)
5797                         len = e1_blk + e1_len - lblk1;
5798                 if (len > e2_blk + e2_len - lblk2)
5799                         len = e2_blk + e2_len - lblk2;
5800
5801                 if (len != e1_len) {
5802                         split = 1;
5803                         *erp = ext4_force_split_extent_at(handle, inode1,
5804                                                 &path1, lblk1 + len, 0);
5805                         if (unlikely(*erp))
5806                                 goto finish;
5807                 }
5808                 if (len != e2_len) {
5809                         split = 1;
5810                         *erp = ext4_force_split_extent_at(handle, inode2,
5811                                                 &path2, lblk2 + len, 0);
5812                         if (*erp)
5813                                 goto finish;
5814                 }
5815                 /* ext4_split_extent_at() may result in leaf extent split,
5816                  * path must to be revalidated. */
5817                 if (split)
5818                         goto repeat;
5819
5820                 BUG_ON(e2_len != e1_len);
5821                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5822                 if (unlikely(*erp))
5823                         goto finish;
5824                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5825                 if (unlikely(*erp))
5826                         goto finish;
5827
5828                 /* Both extents are fully inside boundaries. Swap it now */
5829                 tmp_ex = *ex1;
5830                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5831                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5832                 ex1->ee_len = cpu_to_le16(e2_len);
5833                 ex2->ee_len = cpu_to_le16(e1_len);
5834                 if (unwritten)
5835                         ext4_ext_mark_unwritten(ex2);
5836                 if (ext4_ext_is_unwritten(&tmp_ex))
5837                         ext4_ext_mark_unwritten(ex1);
5838
5839                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5840                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5841                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5842                                       path2->p_depth);
5843                 if (unlikely(*erp))
5844                         goto finish;
5845                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5846                                       path1->p_depth);
5847                 /*
5848                  * Looks scarry ah..? second inode already points to new blocks,
5849                  * and it was successfully dirtied. But luckily error may happen
5850                  * only due to journal error, so full transaction will be
5851                  * aborted anyway.
5852                  */
5853                 if (unlikely(*erp))
5854                         goto finish;
5855                 lblk1 += len;
5856                 lblk2 += len;
5857                 replaced_count += len;
5858                 count -= len;
5859
5860         repeat:
5861                 ext4_ext_drop_refs(path1);
5862                 kfree(path1);
5863                 ext4_ext_drop_refs(path2);
5864                 kfree(path2);
5865                 path1 = path2 = NULL;
5866         }
5867         return replaced_count;
5868 }
5869
5870 /*
5871  * ext4_clu_mapped - determine whether any block in a logical cluster has
5872  *                   been mapped to a physical cluster
5873  *
5874  * @inode - file containing the logical cluster
5875  * @lclu - logical cluster of interest
5876  *
5877  * Returns 1 if any block in the logical cluster is mapped, signifying
5878  * that a physical cluster has been allocated for it.  Otherwise,
5879  * returns 0.  Can also return negative error codes.  Derived from
5880  * ext4_ext_map_blocks().
5881  */
5882 int ext4_clu_mapped(struct inode *inode, ext4_lblk_t lclu)
5883 {
5884         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5885         struct ext4_ext_path *path;
5886         int depth, mapped = 0, err = 0;
5887         struct ext4_extent *extent;
5888         ext4_lblk_t first_lblk, first_lclu, last_lclu;
5889
5890         /* search for the extent closest to the first block in the cluster */
5891         path = ext4_find_extent(inode, EXT4_C2B(sbi, lclu), NULL, 0);
5892         if (IS_ERR(path)) {
5893                 err = PTR_ERR(path);
5894                 path = NULL;
5895                 goto out;
5896         }
5897
5898         depth = ext_depth(inode);
5899
5900         /*
5901          * A consistent leaf must not be empty.  This situation is possible,
5902          * though, _during_ tree modification, and it's why an assert can't
5903          * be put in ext4_find_extent().
5904          */
5905         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
5906                 EXT4_ERROR_INODE(inode,
5907                     "bad extent address - lblock: %lu, depth: %d, pblock: %lld",
5908                                  (unsigned long) EXT4_C2B(sbi, lclu),
5909                                  depth, path[depth].p_block);
5910                 err = -EFSCORRUPTED;
5911                 goto out;
5912         }
5913
5914         extent = path[depth].p_ext;
5915
5916         /* can't be mapped if the extent tree is empty */
5917         if (extent == NULL)
5918                 goto out;
5919
5920         first_lblk = le32_to_cpu(extent->ee_block);
5921         first_lclu = EXT4_B2C(sbi, first_lblk);
5922
5923         /*
5924          * Three possible outcomes at this point - found extent spanning
5925          * the target cluster, to the left of the target cluster, or to the
5926          * right of the target cluster.  The first two cases are handled here.
5927          * The last case indicates the target cluster is not mapped.
5928          */
5929         if (lclu >= first_lclu) {
5930                 last_lclu = EXT4_B2C(sbi, first_lblk +
5931                                      ext4_ext_get_actual_len(extent) - 1);
5932                 if (lclu <= last_lclu) {
5933                         mapped = 1;
5934                 } else {
5935                         first_lblk = ext4_ext_next_allocated_block(path);
5936                         first_lclu = EXT4_B2C(sbi, first_lblk);
5937                         if (lclu == first_lclu)
5938                                 mapped = 1;
5939                 }
5940         }
5941
5942 out:
5943         ext4_ext_drop_refs(path);
5944         kfree(path);
5945
5946         return err ? err : mapped;
5947 }