fs/ext4/extents.c

   1 /*
   2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3  * Written by Alex Tomas <alex@clusterfs.com>
   4  *
   5  * Architecture independence:
   6  *   Copyright (c) 2005, Bull S.A.
   7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public Licens
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <linux/falloc.h>
  41 #include <asm/uaccess.h>
  42 #include <linux/fiemap.h>
  43 #include "ext4_jbd2.h"
  44 #include "ext4_extents.h"
  45 #include "xattr.h"
  46
  47 #include <trace/events/ext4.h>
  48
  49 /*
  50  * used by extent splitting.
  51  */
  52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  53                                         due to ENOSPC */
  54 #define EXT4_EXT_MARK_UNINIT1   0x2  /* mark first half uninitialized */
  55 #define EXT4_EXT_MARK_UNINIT2   0x4  /* mark second half uninitialized */
  56
  57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  59
  60 static __le32 ext4_extent_block_csum(struct inode *inode,
  61                                      struct ext4_extent_header *eh)
  62 {
  63         struct ext4_inode_info *ei = EXT4_I(inode);
  64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  65         __u32 csum;
  66
  67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  68                            EXT4_EXTENT_TAIL_OFFSET(eh));
  69         return cpu_to_le32(csum);
  70 }
  71
  72 static int ext4_extent_block_csum_verify(struct inode *inode,
  73                                          struct ext4_extent_header *eh)
  74 {
  75         struct ext4_extent_tail *et;
  76
  77         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  78                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  79                 return 1;
  80
  81         et = find_ext4_extent_tail(eh);
  82         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  83                 return 0;
  84         return 1;
  85 }
  86
  87 static void ext4_extent_block_csum_set(struct inode *inode,
  88                                        struct ext4_extent_header *eh)
  89 {
  90         struct ext4_extent_tail *et;
  91
  92         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  93                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  94                 return;
  95
  96         et = find_ext4_extent_tail(eh);
  97         et->et_checksum = ext4_extent_block_csum(inode, eh);
  98 }
  99
 100 static int ext4_split_extent(handle_t *handle,
 101                                 struct inode *inode,
 102                                 struct ext4_ext_path *path,
 103                                 struct ext4_map_blocks *map,
 104                                 int split_flag,
 105                                 int flags);
 106
 107 static int ext4_split_extent_at(handle_t *handle,
 108                              struct inode *inode,
 109                              struct ext4_ext_path *path,
 110                              ext4_lblk_t split,
 111                              int split_flag,
 112                              int flags);
 113
 114 static int ext4_find_delayed_extent(struct inode *inode,
 115                                     struct ext4_ext_cache *newex);
 116
 117 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 118                                             struct inode *inode,
 119                                             int needed)
 120 {
 121         int err;
 122
 123         if (!ext4_handle_valid(handle))
 124                 return 0;
 125         if (handle->h_buffer_credits > needed)
 126                 return 0;
 127         err = ext4_journal_extend(handle, needed);
 128         if (err <= 0)
 129                 return err;
 130         err = ext4_truncate_restart_trans(handle, inode, needed);
 131         if (err == 0)
 132                 err = -EAGAIN;
 133
 134         return err;
 135 }
 136
 137 /*
 138  * could return:
 139  *  - EROFS
 140  *  - ENOMEM
 141  */
 142 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 143                                 struct ext4_ext_path *path)
 144 {
 145         if (path->p_bh) {
 146                 /* path points to block */
 147                 return ext4_journal_get_write_access(handle, path->p_bh);
 148         }
 149         /* path points to leaf/index in inode body */
 150         /* we use in-core data, no need to protect them */
 151         return 0;
 152 }
 153
 154 /*
 155  * could return:
 156  *  - EROFS
 157  *  - ENOMEM
 158  *  - EIO
 159  */
 160 #define ext4_ext_dirty(handle, inode, path) \
 161                 __ext4_ext_dirty(__func__, __LINE__, (handle), (inode), (path))
 162 static int __ext4_ext_dirty(const char *where, unsigned int line,
 163                             handle_t *handle, struct inode *inode,
 164                             struct ext4_ext_path *path)
 165 {
 166         int err;
 167         if (path->p_bh) {
 168                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 169                 /* path points to block */
 170                 err = __ext4_handle_dirty_metadata(where, line, handle,
 171                                                    inode, path->p_bh);
 172         } else {
 173                 /* path points to leaf/index in inode body */
 174                 err = ext4_mark_inode_dirty(handle, inode);
 175         }
 176         return err;
 177 }
 178
 179 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 180                               struct ext4_ext_path *path,
 181                               ext4_lblk_t block)
 182 {
 183         if (path) {
 184                 int depth = path->p_depth;
 185                 struct ext4_extent *ex;
 186
 187                 /*
 188                  * Try to predict block placement assuming that we are
 189                  * filling in a file which will eventually be
 190                  * non-sparse --- i.e., in the case of libbfd writing
 191                  * an ELF object sections out-of-order but in a way
 192                  * the eventually results in a contiguous object or
 193                  * executable file, or some database extending a table
 194                  * space file.  However, this is actually somewhat
 195                  * non-ideal if we are writing a sparse file such as
 196                  * qemu or KVM writing a raw image file that is going
 197                  * to stay fairly sparse, since it will end up
 198                  * fragmenting the file system's free space.  Maybe we
 199                  * should have some hueristics or some way to allow
 200                  * userspace to pass a hint to file system,
 201                  * especially if the latter case turns out to be
 202                  * common.
 203                  */
 204                 ex = path[depth].p_ext;
 205                 if (ex) {
 206                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 207                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 208
 209                         if (block > ext_block)
 210                                 return ext_pblk + (block - ext_block);
 211                         else
 212                                 return ext_pblk - (ext_block - block);
 213                 }
 214
 215                 /* it looks like index is empty;
 216                  * try to find starting block from index itself */
 217                 if (path[depth].p_bh)
 218                         return path[depth].p_bh->b_blocknr;
 219         }
 220
 221         /* OK. use inode's group */
 222         return ext4_inode_to_goal_block(inode);
 223 }
 224
 225 /*
 226  * Allocation for a meta data block
 227  */
 228 static ext4_fsblk_t
 229 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 230                         struct ext4_ext_path *path,
 231                         struct ext4_extent *ex, int *err, unsigned int flags)
 232 {
 233         ext4_fsblk_t goal, newblock;
 234
 235         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 236         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 237                                         NULL, err);
 238         return newblock;
 239 }
 240
 241 static inline int ext4_ext_space_block(struct inode *inode, int check)
 242 {
 243         int size;
 244
 245         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 246                         / sizeof(struct ext4_extent);
 247 #ifdef AGGRESSIVE_TEST
 248         if (!check && size > 6)
 249                 size = 6;
 250 #endif
 251         return size;
 252 }
 253
 254 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 255 {
 256         int size;
 257
 258         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 259                         / sizeof(struct ext4_extent_idx);
 260 #ifdef AGGRESSIVE_TEST
 261         if (!check && size > 5)
 262                 size = 5;
 263 #endif
 264         return size;
 265 }
 266
 267 static inline int ext4_ext_space_root(struct inode *inode, int check)
 268 {
 269         int size;
 270
 271         size = sizeof(EXT4_I(inode)->i_data);
 272         size -= sizeof(struct ext4_extent_header);
 273         size /= sizeof(struct ext4_extent);
 274 #ifdef AGGRESSIVE_TEST
 275         if (!check && size > 3)
 276                 size = 3;
 277 #endif
 278         return size;
 279 }
 280
 281 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 282 {
 283         int size;
 284
 285         size = sizeof(EXT4_I(inode)->i_data);
 286         size -= sizeof(struct ext4_extent_header);
 287         size /= sizeof(struct ext4_extent_idx);
 288 #ifdef AGGRESSIVE_TEST
 289         if (!check && size > 4)
 290                 size = 4;
 291 #endif
 292         return size;
 293 }
 294
 295 /*
 296  * Calculate the number of metadata blocks needed
 297  * to allocate @blocks
 298  * Worse case is one block per extent
 299  */
 300 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 301 {
 302         struct ext4_inode_info *ei = EXT4_I(inode);
 303         int idxs;
 304
 305         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 306                 / sizeof(struct ext4_extent_idx));
 307
 308         /*
 309          * If the new delayed allocation block is contiguous with the
 310          * previous da block, it can share index blocks with the
 311          * previous block, so we only need to allocate a new index
 312          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 313          * an additional index block, and at ldxs**3 blocks, yet
 314          * another index blocks.
 315          */
 316         if (ei->i_da_metadata_calc_len &&
 317             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 318                 int num = 0;
 319
 320                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 321                         num++;
 322                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 323                         num++;
 324                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 325                         num++;
 326                         ei->i_da_metadata_calc_len = 0;
 327                 } else
 328                         ei->i_da_metadata_calc_len++;
 329                 ei->i_da_metadata_calc_last_lblock++;
 330                 return num;
 331         }
 332
 333         /*
 334          * In the worst case we need a new set of index blocks at
 335          * every level of the inode's extent tree.
 336          */
 337         ei->i_da_metadata_calc_len = 1;
 338         ei->i_da_metadata_calc_last_lblock = lblock;
 339         return ext_depth(inode) + 1;
 340 }
 341
 342 static int
 343 ext4_ext_max_entries(struct inode *inode, int depth)
 344 {
 345         int max;
 346
 347         if (depth == ext_depth(inode)) {
 348                 if (depth == 0)
 349                         max = ext4_ext_space_root(inode, 1);
 350                 else
 351                         max = ext4_ext_space_root_idx(inode, 1);
 352         } else {
 353                 if (depth == 0)
 354                         max = ext4_ext_space_block(inode, 1);
 355                 else
 356                         max = ext4_ext_space_block_idx(inode, 1);
 357         }
 358
 359         return max;
 360 }
 361
 362 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 363 {
 364         ext4_fsblk_t block = ext4_ext_pblock(ext);
 365         int len = ext4_ext_get_actual_len(ext);
 366
 367         if (len == 0)
 368                 return 0;
 369         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 370 }
 371
 372 static int ext4_valid_extent_idx(struct inode *inode,
 373                                 struct ext4_extent_idx *ext_idx)
 374 {
 375         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 376
 377         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 378 }
 379
 380 static int ext4_valid_extent_entries(struct inode *inode,
 381                                 struct ext4_extent_header *eh,
 382                                 int depth)
 383 {
 384         unsigned short entries;
 385         if (eh->eh_entries == 0)
 386                 return 1;
 387
 388         entries = le16_to_cpu(eh->eh_entries);
 389
 390         if (depth == 0) {
 391                 /* leaf entries */
 392                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 393                 while (entries) {
 394                         if (!ext4_valid_extent(inode, ext))
 395                                 return 0;
 396                         ext++;
 397                         entries--;
 398                 }
 399         } else {
 400                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 401                 while (entries) {
 402                         if (!ext4_valid_extent_idx(inode, ext_idx))
 403                                 return 0;
 404                         ext_idx++;
 405                         entries--;
 406                 }
 407         }
 408         return 1;
 409 }
 410
 411 static int __ext4_ext_check(const char *function, unsigned int line,
 412                             struct inode *inode, struct ext4_extent_header *eh,
 413                             int depth)
 414 {
 415         const char *error_msg;
 416         int max = 0;
 417
 418         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 419                 error_msg = "invalid magic";
 420                 goto corrupted;
 421         }
 422         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 423                 error_msg = "unexpected eh_depth";
 424                 goto corrupted;
 425         }
 426         if (unlikely(eh->eh_max == 0)) {
 427                 error_msg = "invalid eh_max";
 428                 goto corrupted;
 429         }
 430         max = ext4_ext_max_entries(inode, depth);
 431         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 432                 error_msg = "too large eh_max";
 433                 goto corrupted;
 434         }
 435         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 436                 error_msg = "invalid eh_entries";
 437                 goto corrupted;
 438         }
 439         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 440                 error_msg = "invalid extent entries";
 441                 goto corrupted;
 442         }
 443         /* Verify checksum on non-root extent tree nodes */
 444         if (ext_depth(inode) != depth &&
 445             !ext4_extent_block_csum_verify(inode, eh)) {
 446                 error_msg = "extent tree corrupted";
 447                 goto corrupted;
 448         }
 449         return 0;
 450
 451 corrupted:
 452         ext4_error_inode(inode, function, line, 0,
 453                         "bad header/extent: %s - magic %x, "
 454                         "entries %u, max %u(%u), depth %u(%u)",
 455                         error_msg, le16_to_cpu(eh->eh_magic),
 456                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 457                         max, le16_to_cpu(eh->eh_depth), depth);
 458
 459         return -EIO;
 460 }
 461
 462 #define ext4_ext_check(inode, eh, depth)        \
 463         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
 464
 465 int ext4_ext_check_inode(struct inode *inode)
 466 {
 467         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
 468 }
 469
 470 static int __ext4_ext_check_block(const char *function, unsigned int line,
 471                                   struct inode *inode,
 472                                   struct ext4_extent_header *eh,
 473                                   int depth,
 474                                   struct buffer_head *bh)
 475 {
 476         int ret;
 477
 478         if (buffer_verified(bh))
 479                 return 0;
 480         ret = ext4_ext_check(inode, eh, depth);
 481         if (ret)
 482                 return ret;
 483         set_buffer_verified(bh);
 484         return ret;
 485 }
 486
 487 #define ext4_ext_check_block(inode, eh, depth, bh)      \
 488         __ext4_ext_check_block(__func__, __LINE__, inode, eh, depth, bh)
 489
 490 #ifdef EXT_DEBUG
 491 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 492 {
 493         int k, l = path->p_depth;
 494
 495         ext_debug("path:");
 496         for (k = 0; k <= l; k++, path++) {
 497                 if (path->p_idx) {
 498                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 499                             ext4_idx_pblock(path->p_idx));
 500                 } else if (path->p_ext) {
 501                         ext_debug("  %d:[%d]%d:%llu ",
 502                                   le32_to_cpu(path->p_ext->ee_block),
 503                                   ext4_ext_is_uninitialized(path->p_ext),
 504                                   ext4_ext_get_actual_len(path->p_ext),
 505                                   ext4_ext_pblock(path->p_ext));
 506                 } else
 507                         ext_debug("  []");
 508         }
 509         ext_debug("\n");
 510 }
 511
 512 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 513 {
 514         int depth = ext_depth(inode);
 515         struct ext4_extent_header *eh;
 516         struct ext4_extent *ex;
 517         int i;
 518
 519         if (!path)
 520                 return;
 521
 522         eh = path[depth].p_hdr;
 523         ex = EXT_FIRST_EXTENT(eh);
 524
 525         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 526
 527         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 528                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 529                           ext4_ext_is_uninitialized(ex),
 530                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 531         }
 532         ext_debug("\n");
 533 }
 534
 535 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 536                         ext4_fsblk_t newblock, int level)
 537 {
 538         int depth = ext_depth(inode);
 539         struct ext4_extent *ex;
 540
 541         if (depth != level) {
 542                 struct ext4_extent_idx *idx;
 543                 idx = path[level].p_idx;
 544                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 545                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 546                                         le32_to_cpu(idx->ei_block),
 547                                         ext4_idx_pblock(idx),
 548                                         newblock);
 549                         idx++;
 550                 }
 551
 552                 return;
 553         }
 554
 555         ex = path[depth].p_ext;
 556         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 557                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 558                                 le32_to_cpu(ex->ee_block),
 559                                 ext4_ext_pblock(ex),
 560                                 ext4_ext_is_uninitialized(ex),
 561                                 ext4_ext_get_actual_len(ex),
 562                                 newblock);
 563                 ex++;
 564         }
 565 }
 566
 567 #else
 568 #define ext4_ext_show_path(inode, path)
 569 #define ext4_ext_show_leaf(inode, path)
 570 #define ext4_ext_show_move(inode, path, newblock, level)
 571 #endif
 572
 573 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 574 {
 575         int depth = path->p_depth;
 576         int i;
 577
 578         for (i = 0; i <= depth; i++, path++)
 579                 if (path->p_bh) {
 580                         brelse(path->p_bh);
 581                         path->p_bh = NULL;
 582                 }
 583 }
 584
 585 /*
 586  * ext4_ext_binsearch_idx:
 587  * binary search for the closest index of the given block
 588  * the header must be checked before calling this
 589  */
 590 static void
 591 ext4_ext_binsearch_idx(struct inode *inode,
 592                         struct ext4_ext_path *path, ext4_lblk_t block)
 593 {
 594         struct ext4_extent_header *eh = path->p_hdr;
 595         struct ext4_extent_idx *r, *l, *m;
 596
 597
 598         ext_debug("binsearch for %u(idx):  ", block);
 599
 600         l = EXT_FIRST_INDEX(eh) + 1;
 601         r = EXT_LAST_INDEX(eh);
 602         while (l <= r) {
 603                 m = l + (r - l) / 2;
 604                 if (block < le32_to_cpu(m->ei_block))
 605                         r = m - 1;
 606                 else
 607                         l = m + 1;
 608                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 609                                 m, le32_to_cpu(m->ei_block),
 610                                 r, le32_to_cpu(r->ei_block));
 611         }
 612
 613         path->p_idx = l - 1;
 614         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 615                   ext4_idx_pblock(path->p_idx));
 616
 617 #ifdef CHECK_BINSEARCH
 618         {
 619                 struct ext4_extent_idx *chix, *ix;
 620                 int k;
 621
 622                 chix = ix = EXT_FIRST_INDEX(eh);
 623                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 624                   if (k != 0 &&
 625                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 626                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 627                                        "first=0x%p\n", k,
 628                                        ix, EXT_FIRST_INDEX(eh));
 629                                 printk(KERN_DEBUG "%u <= %u\n",
 630                                        le32_to_cpu(ix->ei_block),
 631                                        le32_to_cpu(ix[-1].ei_block));
 632                         }
 633                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 634                                            <= le32_to_cpu(ix[-1].ei_block));
 635                         if (block < le32_to_cpu(ix->ei_block))
 636                                 break;
 637                         chix = ix;
 638                 }
 639                 BUG_ON(chix != path->p_idx);
 640         }
 641 #endif
 642
 643 }
 644
 645 /*
 646  * ext4_ext_binsearch:
 647  * binary search for closest extent of the given block
 648  * the header must be checked before calling this
 649  */
 650 static void
 651 ext4_ext_binsearch(struct inode *inode,
 652                 struct ext4_ext_path *path, ext4_lblk_t block)
 653 {
 654         struct ext4_extent_header *eh = path->p_hdr;
 655         struct ext4_extent *r, *l, *m;
 656
 657         if (eh->eh_entries == 0) {
 658                 /*
 659                  * this leaf is empty:
 660                  * we get such a leaf in split/add case
 661                  */
 662                 return;
 663         }
 664
 665         ext_debug("binsearch for %u:  ", block);
 666
 667         l = EXT_FIRST_EXTENT(eh) + 1;
 668         r = EXT_LAST_EXTENT(eh);
 669
 670         while (l <= r) {
 671                 m = l + (r - l) / 2;
 672                 if (block < le32_to_cpu(m->ee_block))
 673                         r = m - 1;
 674                 else
 675                         l = m + 1;
 676                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 677                                 m, le32_to_cpu(m->ee_block),
 678                                 r, le32_to_cpu(r->ee_block));
 679         }
 680
 681         path->p_ext = l - 1;
 682         ext_debug("  -> %d:%llu:[%d]%d ",
 683                         le32_to_cpu(path->p_ext->ee_block),
 684                         ext4_ext_pblock(path->p_ext),
 685                         ext4_ext_is_uninitialized(path->p_ext),
 686                         ext4_ext_get_actual_len(path->p_ext));
 687
 688 #ifdef CHECK_BINSEARCH
 689         {
 690                 struct ext4_extent *chex, *ex;
 691                 int k;
 692
 693                 chex = ex = EXT_FIRST_EXTENT(eh);
 694                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 695                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 696                                           <= le32_to_cpu(ex[-1].ee_block));
 697                         if (block < le32_to_cpu(ex->ee_block))
 698                                 break;
 699                         chex = ex;
 700                 }
 701                 BUG_ON(chex != path->p_ext);
 702         }
 703 #endif
 704
 705 }
 706
 707 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 708 {
 709         struct ext4_extent_header *eh;
 710
 711         eh = ext_inode_hdr(inode);
 712         eh->eh_depth = 0;
 713         eh->eh_entries = 0;
 714         eh->eh_magic = EXT4_EXT_MAGIC;
 715         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 716         ext4_mark_inode_dirty(handle, inode);
 717         ext4_ext_invalidate_cache(inode);
 718         return 0;
 719 }
 720
 721 struct ext4_ext_path *
 722 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
 723                                         struct ext4_ext_path *path)
 724 {
 725         struct ext4_extent_header *eh;
 726         struct buffer_head *bh;
 727         short int depth, i, ppos = 0, alloc = 0;
 728
 729         eh = ext_inode_hdr(inode);
 730         depth = ext_depth(inode);
 731
 732         /* account possible depth increase */
 733         if (!path) {
 734                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 735                                 GFP_NOFS);
 736                 if (!path)
 737                         return ERR_PTR(-ENOMEM);
 738                 alloc = 1;
 739         }
 740         path[0].p_hdr = eh;
 741         path[0].p_bh = NULL;
 742
 743         i = depth;
 744         /* walk through the tree */
 745         while (i) {
 746                 ext_debug("depth %d: num %d, max %d\n",
 747                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 748
 749                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 750                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 751                 path[ppos].p_depth = i;
 752                 path[ppos].p_ext = NULL;
 753
 754                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
 755                 if (unlikely(!bh))
 756                         goto err;
 757                 if (!bh_uptodate_or_lock(bh)) {
 758                         trace_ext4_ext_load_extent(inode, block,
 759                                                 path[ppos].p_block);
 760                         if (bh_submit_read(bh) < 0) {
 761                                 put_bh(bh);
 762                                 goto err;
 763                         }
 764                 }
 765                 eh = ext_block_hdr(bh);
 766                 ppos++;
 767                 if (unlikely(ppos > depth)) {
 768                         put_bh(bh);
 769                         EXT4_ERROR_INODE(inode,
 770                                          "ppos %d > depth %d", ppos, depth);
 771                         goto err;
 772                 }
 773                 path[ppos].p_bh = bh;
 774                 path[ppos].p_hdr = eh;
 775                 i--;
 776
 777                 if (ext4_ext_check_block(inode, eh, i, bh))
 778                         goto err;
 779         }
 780
 781         path[ppos].p_depth = i;
 782         path[ppos].p_ext = NULL;
 783         path[ppos].p_idx = NULL;
 784
 785         /* find extent */
 786         ext4_ext_binsearch(inode, path + ppos, block);
 787         /* if not an empty leaf */
 788         if (path[ppos].p_ext)
 789                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 790
 791         ext4_ext_show_path(inode, path);
 792
 793         return path;
 794
 795 err:
 796         ext4_ext_drop_refs(path);
 797         if (alloc)
 798                 kfree(path);
 799         return ERR_PTR(-EIO);
 800 }
 801
 802 /*
 803  * ext4_ext_insert_index:
 804  * insert new index [@logical;@ptr] into the block at @curp;
 805  * check where to insert: before @curp or after @curp
 806  */
 807 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 808                                  struct ext4_ext_path *curp,
 809                                  int logical, ext4_fsblk_t ptr)
 810 {
 811         struct ext4_extent_idx *ix;
 812         int len, err;
 813
 814         err = ext4_ext_get_access(handle, inode, curp);
 815         if (err)
 816                 return err;
 817
 818         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 819                 EXT4_ERROR_INODE(inode,
 820                                  "logical %d == ei_block %d!",
 821                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 822                 return -EIO;
 823         }
 824
 825         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 826                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 827                 EXT4_ERROR_INODE(inode,
 828                                  "eh_entries %d >= eh_max %d!",
 829                                  le16_to_cpu(curp->p_hdr->eh_entries),
 830                                  le16_to_cpu(curp->p_hdr->eh_max));
 831                 return -EIO;
 832         }
 833
 834         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 835                 /* insert after */
 836                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 837                 ix = curp->p_idx + 1;
 838         } else {
 839                 /* insert before */
 840                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 841                 ix = curp->p_idx;
 842         }
 843
 844         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 845         BUG_ON(len < 0);
 846         if (len > 0) {
 847                 ext_debug("insert new index %d: "
 848                                 "move %d indices from 0x%p to 0x%p\n",
 849                                 logical, len, ix, ix + 1);
 850                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 851         }
 852
 853         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 854                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 855                 return -EIO;
 856         }
 857
 858         ix->ei_block = cpu_to_le32(logical);
 859         ext4_idx_store_pblock(ix, ptr);
 860         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 861
 862         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 863                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 864                 return -EIO;
 865         }
 866
 867         err = ext4_ext_dirty(handle, inode, curp);
 868         ext4_std_error(inode->i_sb, err);
 869
 870         return err;
 871 }
 872
 873 /*
 874  * ext4_ext_split:
 875  * inserts new subtree into the path, using free index entry
 876  * at depth @at:
 877  * - allocates all needed blocks (new leaf and all intermediate index blocks)
 878  * - makes decision where to split
 879  * - moves remaining extents and index entries (right to the split point)
 880  *   into the newly allocated blocks
 881  * - initializes subtree
 882  */
 883 static int ext4_ext_split(handle_t *handle, struct inode *inode,
 884                           unsigned int flags,
 885                           struct ext4_ext_path *path,
 886                           struct ext4_extent *newext, int at)
 887 {
 888         struct buffer_head *bh = NULL;
 889         int depth = ext_depth(inode);
 890         struct ext4_extent_header *neh;
 891         struct ext4_extent_idx *fidx;
 892         int i = at, k, m, a;
 893         ext4_fsblk_t newblock, oldblock;
 894         __le32 border;
 895         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
 896         int err = 0;
 897
 898         /* make decision: where to split? */
 899         /* FIXME: now decision is simplest: at current extent */
 900
 901         /* if current leaf will be split, then we should use
 902          * border from split point */
 903         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
 904                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
 905                 return -EIO;
 906         }
 907         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
 908                 border = path[depth].p_ext[1].ee_block;
 909                 ext_debug("leaf will be split."
 910                                 " next leaf starts at %d\n",
 911                                   le32_to_cpu(border));
 912         } else {
 913                 border = newext->ee_block;
 914                 ext_debug("leaf will be added."
 915                                 " next leaf starts at %d\n",
 916                                 le32_to_cpu(border));
 917         }
 918
 919         /*
 920          * If error occurs, then we break processing
 921          * and mark filesystem read-only. index won't
 922          * be inserted and tree will be in consistent
 923          * state. Next mount will repair buffers too.
 924          */
 925
 926         /*
 927          * Get array to track all allocated blocks.
 928          * We need this to handle errors and free blocks
 929          * upon them.
 930          */
 931         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
 932         if (!ablocks)
 933                 return -ENOMEM;
 934
 935         /* allocate all needed blocks */
 936         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
 937         for (a = 0; a < depth - at; a++) {
 938                 newblock = ext4_ext_new_meta_block(handle, inode, path,
 939                                                    newext, &err, flags);
 940                 if (newblock == 0)
 941                         goto cleanup;
 942                 ablocks[a] = newblock;
 943         }
 944
 945         /* initialize new leaf */
 946         newblock = ablocks[--a];
 947         if (unlikely(newblock == 0)) {
 948                 EXT4_ERROR_INODE(inode, "newblock == 0!");
 949                 err = -EIO;
 950                 goto cleanup;
 951         }
 952         bh = sb_getblk(inode->i_sb, newblock);
 953         if (!bh) {
 954                 err = -EIO;
 955                 goto cleanup;
 956         }
 957         lock_buffer(bh);
 958
 959         err = ext4_journal_get_create_access(handle, bh);
 960         if (err)
 961                 goto cleanup;
 962
 963         neh = ext_block_hdr(bh);
 964         neh->eh_entries = 0;
 965         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
 966         neh->eh_magic = EXT4_EXT_MAGIC;
 967         neh->eh_depth = 0;
 968
 969         /* move remainder of path[depth] to the new leaf */
 970         if (unlikely(path[depth].p_hdr->eh_entries !=
 971                      path[depth].p_hdr->eh_max)) {
 972                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
 973                                  path[depth].p_hdr->eh_entries,
 974                                  path[depth].p_hdr->eh_max);
 975                 err = -EIO;
 976                 goto cleanup;
 977         }
 978         /* start copy from next extent */
 979         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
 980         ext4_ext_show_move(inode, path, newblock, depth);
 981         if (m) {
 982                 struct ext4_extent *ex;
 983                 ex = EXT_FIRST_EXTENT(neh);
 984                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
 985                 le16_add_cpu(&neh->eh_entries, m);
 986         }
 987
 988         ext4_extent_block_csum_set(inode, neh);
 989         set_buffer_uptodate(bh);
 990         unlock_buffer(bh);
 991
 992         err = ext4_handle_dirty_metadata(handle, inode, bh);
 993         if (err)
 994                 goto cleanup;
 995         brelse(bh);
 996         bh = NULL;
 997
 998         /* correct old leaf */
 999         if (m) {
1000                 err = ext4_ext_get_access(handle, inode, path + depth);
1001                 if (err)
1002                         goto cleanup;
1003                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1004                 err = ext4_ext_dirty(handle, inode, path + depth);
1005                 if (err)
1006                         goto cleanup;
1007
1008         }
1009
1010         /* create intermediate indexes */
1011         k = depth - at - 1;
1012         if (unlikely(k < 0)) {
1013                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1014                 err = -EIO;
1015                 goto cleanup;
1016         }
1017         if (k)
1018                 ext_debug("create %d intermediate indices\n", k);
1019         /* insert new index into current index block */
1020         /* current depth stored in i var */
1021         i = depth - 1;
1022         while (k--) {
1023                 oldblock = newblock;
1024                 newblock = ablocks[--a];
1025                 bh = sb_getblk(inode->i_sb, newblock);
1026                 if (!bh) {
1027                         err = -EIO;
1028                         goto cleanup;
1029                 }
1030                 lock_buffer(bh);
1031
1032                 err = ext4_journal_get_create_access(handle, bh);
1033                 if (err)
1034                         goto cleanup;
1035
1036                 neh = ext_block_hdr(bh);
1037                 neh->eh_entries = cpu_to_le16(1);
1038                 neh->eh_magic = EXT4_EXT_MAGIC;
1039                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1040                 neh->eh_depth = cpu_to_le16(depth - i);
1041                 fidx = EXT_FIRST_INDEX(neh);
1042                 fidx->ei_block = border;
1043                 ext4_idx_store_pblock(fidx, oldblock);
1044
1045                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1046                                 i, newblock, le32_to_cpu(border), oldblock);
1047
1048                 /* move remainder of path[i] to the new index block */
1049                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1050                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1051                         EXT4_ERROR_INODE(inode,
1052                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1053                                          le32_to_cpu(path[i].p_ext->ee_block));
1054                         err = -EIO;
1055                         goto cleanup;
1056                 }
1057                 /* start copy indexes */
1058                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1059                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1060                                 EXT_MAX_INDEX(path[i].p_hdr));
1061                 ext4_ext_show_move(inode, path, newblock, i);
1062                 if (m) {
1063                         memmove(++fidx, path[i].p_idx,
1064                                 sizeof(struct ext4_extent_idx) * m);
1065                         le16_add_cpu(&neh->eh_entries, m);
1066                 }
1067                 ext4_extent_block_csum_set(inode, neh);
1068                 set_buffer_uptodate(bh);
1069                 unlock_buffer(bh);
1070
1071                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1072                 if (err)
1073                         goto cleanup;
1074                 brelse(bh);
1075                 bh = NULL;
1076
1077                 /* correct old index */
1078                 if (m) {
1079                         err = ext4_ext_get_access(handle, inode, path + i);
1080                         if (err)
1081                                 goto cleanup;
1082                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1083                         err = ext4_ext_dirty(handle, inode, path + i);
1084                         if (err)
1085                                 goto cleanup;
1086                 }
1087
1088                 i--;
1089         }
1090
1091         /* insert new index */
1092         err = ext4_ext_insert_index(handle, inode, path + at,
1093                                     le32_to_cpu(border), newblock);
1094
1095 cleanup:
1096         if (bh) {
1097                 if (buffer_locked(bh))
1098                         unlock_buffer(bh);
1099                 brelse(bh);
1100         }
1101
1102         if (err) {
1103                 /* free all allocated blocks in error case */
1104                 for (i = 0; i < depth; i++) {
1105                         if (!ablocks[i])
1106                                 continue;
1107                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1108                                          EXT4_FREE_BLOCKS_METADATA);
1109                 }
1110         }
1111         kfree(ablocks);
1112
1113         return err;
1114 }
1115
1116 /*
1117  * ext4_ext_grow_indepth:
1118  * implements tree growing procedure:
1119  * - allocates new block
1120  * - moves top-level data (index block or leaf) into the new block
1121  * - initializes new top-level, creating index that points to the
1122  *   just created block
1123  */
1124 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1125                                  unsigned int flags,
1126                                  struct ext4_extent *newext)
1127 {
1128         struct ext4_extent_header *neh;
1129         struct buffer_head *bh;
1130         ext4_fsblk_t newblock;
1131         int err = 0;
1132
1133         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1134                 newext, &err, flags);
1135         if (newblock == 0)
1136                 return err;
1137
1138         bh = sb_getblk(inode->i_sb, newblock);
1139         if (!bh) {
1140                 err = -EIO;
1141                 ext4_std_error(inode->i_sb, err);
1142                 return err;
1143         }
1144         lock_buffer(bh);
1145
1146         err = ext4_journal_get_create_access(handle, bh);
1147         if (err) {
1148                 unlock_buffer(bh);
1149                 goto out;
1150         }
1151
1152         /* move top-level index/leaf into new block */
1153         memmove(bh->b_data, EXT4_I(inode)->i_data,
1154                 sizeof(EXT4_I(inode)->i_data));
1155
1156         /* set size of new block */
1157         neh = ext_block_hdr(bh);
1158         /* old root could have indexes or leaves
1159          * so calculate e_max right way */
1160         if (ext_depth(inode))
1161                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1162         else
1163                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1164         neh->eh_magic = EXT4_EXT_MAGIC;
1165         ext4_extent_block_csum_set(inode, neh);
1166         set_buffer_uptodate(bh);
1167         unlock_buffer(bh);
1168
1169         err = ext4_handle_dirty_metadata(handle, inode, bh);
1170         if (err)
1171                 goto out;
1172
1173         /* Update top-level index: num,max,pointer */
1174         neh = ext_inode_hdr(inode);
1175         neh->eh_entries = cpu_to_le16(1);
1176         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1177         if (neh->eh_depth == 0) {
1178                 /* Root extent block becomes index block */
1179                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1180                 EXT_FIRST_INDEX(neh)->ei_block =
1181                         EXT_FIRST_EXTENT(neh)->ee_block;
1182         }
1183         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1184                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1185                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1186                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1187
1188         le16_add_cpu(&neh->eh_depth, 1);
1189         ext4_mark_inode_dirty(handle, inode);
1190 out:
1191         brelse(bh);
1192
1193         return err;
1194 }
1195
1196 /*
1197  * ext4_ext_create_new_leaf:
1198  * finds empty index and adds new leaf.
1199  * if no free index is found, then it requests in-depth growing.
1200  */
1201 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1202                                     unsigned int flags,
1203                                     struct ext4_ext_path *path,
1204                                     struct ext4_extent *newext)
1205 {
1206         struct ext4_ext_path *curp;
1207         int depth, i, err = 0;
1208
1209 repeat:
1210         i = depth = ext_depth(inode);
1211
1212         /* walk up to the tree and look for free index entry */
1213         curp = path + depth;
1214         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1215                 i--;
1216                 curp--;
1217         }
1218
1219         /* we use already allocated block for index block,
1220          * so subsequent data blocks should be contiguous */
1221         if (EXT_HAS_FREE_INDEX(curp)) {
1222                 /* if we found index with free entry, then use that
1223                  * entry: create all needed subtree and add new leaf */
1224                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1225                 if (err)
1226                         goto out;
1227
1228                 /* refill path */
1229                 ext4_ext_drop_refs(path);
1230                 path = ext4_ext_find_extent(inode,
1231                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1232                                     path);
1233                 if (IS_ERR(path))
1234                         err = PTR_ERR(path);
1235         } else {
1236                 /* tree is full, time to grow in depth */
1237                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1238                 if (err)
1239                         goto out;
1240
1241                 /* refill path */
1242                 ext4_ext_drop_refs(path);
1243                 path = ext4_ext_find_extent(inode,
1244                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1245                                     path);
1246                 if (IS_ERR(path)) {
1247                         err = PTR_ERR(path);
1248                         goto out;
1249                 }
1250
1251                 /*
1252                  * only first (depth 0 -> 1) produces free space;
1253                  * in all other cases we have to split the grown tree
1254                  */
1255                 depth = ext_depth(inode);
1256                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1257                         /* now we need to split */
1258                         goto repeat;
1259                 }
1260         }
1261
1262 out:
1263         return err;
1264 }
1265
1266 /*
1267  * search the closest allocated block to the left for *logical
1268  * and returns it at @logical + it's physical address at @phys
1269  * if *logical is the smallest allocated block, the function
1270  * returns 0 at @phys
1271  * return value contains 0 (success) or error code
1272  */
1273 static int ext4_ext_search_left(struct inode *inode,
1274                                 struct ext4_ext_path *path,
1275                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1276 {
1277         struct ext4_extent_idx *ix;
1278         struct ext4_extent *ex;
1279         int depth, ee_len;
1280
1281         if (unlikely(path == NULL)) {
1282                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1283                 return -EIO;
1284         }
1285         depth = path->p_depth;
1286         *phys = 0;
1287
1288         if (depth == 0 && path->p_ext == NULL)
1289                 return 0;
1290
1291         /* usually extent in the path covers blocks smaller
1292          * then *logical, but it can be that extent is the
1293          * first one in the file */
1294
1295         ex = path[depth].p_ext;
1296         ee_len = ext4_ext_get_actual_len(ex);
1297         if (*logical < le32_to_cpu(ex->ee_block)) {
1298                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1299                         EXT4_ERROR_INODE(inode,
1300                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1301                                          *logical, le32_to_cpu(ex->ee_block));
1302                         return -EIO;
1303                 }
1304                 while (--depth >= 0) {
1305                         ix = path[depth].p_idx;
1306                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1307                                 EXT4_ERROR_INODE(inode,
1308                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1309                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1310                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1311                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1312                                   depth);
1313                                 return -EIO;
1314                         }
1315                 }
1316                 return 0;
1317         }
1318
1319         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1320                 EXT4_ERROR_INODE(inode,
1321                                  "logical %d < ee_block %d + ee_len %d!",
1322                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1323                 return -EIO;
1324         }
1325
1326         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1327         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1328         return 0;
1329 }
1330
1331 /*
1332  * search the closest allocated block to the right for *logical
1333  * and returns it at @logical + it's physical address at @phys
1334  * if *logical is the largest allocated block, the function
1335  * returns 0 at @phys
1336  * return value contains 0 (success) or error code
1337  */
1338 static int ext4_ext_search_right(struct inode *inode,
1339                                  struct ext4_ext_path *path,
1340                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1341                                  struct ext4_extent **ret_ex)
1342 {
1343         struct buffer_head *bh = NULL;
1344         struct ext4_extent_header *eh;
1345         struct ext4_extent_idx *ix;
1346         struct ext4_extent *ex;
1347         ext4_fsblk_t block;
1348         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1349         int ee_len;
1350
1351         if (unlikely(path == NULL)) {
1352                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1353                 return -EIO;
1354         }
1355         depth = path->p_depth;
1356         *phys = 0;
1357
1358         if (depth == 0 && path->p_ext == NULL)
1359                 return 0;
1360
1361         /* usually extent in the path covers blocks smaller
1362          * then *logical, but it can be that extent is the
1363          * first one in the file */
1364
1365         ex = path[depth].p_ext;
1366         ee_len = ext4_ext_get_actual_len(ex);
1367         if (*logical < le32_to_cpu(ex->ee_block)) {
1368                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1369                         EXT4_ERROR_INODE(inode,
1370                                          "first_extent(path[%d].p_hdr) != ex",
1371                                          depth);
1372                         return -EIO;
1373                 }
1374                 while (--depth >= 0) {
1375                         ix = path[depth].p_idx;
1376                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1377                                 EXT4_ERROR_INODE(inode,
1378                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1379                                                  *logical);
1380                                 return -EIO;
1381                         }
1382                 }
1383                 goto found_extent;
1384         }
1385
1386         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1387                 EXT4_ERROR_INODE(inode,
1388                                  "logical %d < ee_block %d + ee_len %d!",
1389                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1390                 return -EIO;
1391         }
1392
1393         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1394                 /* next allocated block in this leaf */
1395                 ex++;
1396                 goto found_extent;
1397         }
1398
1399         /* go up and search for index to the right */
1400         while (--depth >= 0) {
1401                 ix = path[depth].p_idx;
1402                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1403                         goto got_index;
1404         }
1405
1406         /* we've gone up to the root and found no index to the right */
1407         return 0;
1408
1409 got_index:
1410         /* we've found index to the right, let's
1411          * follow it and find the closest allocated
1412          * block to the right */
1413         ix++;
1414         block = ext4_idx_pblock(ix);
1415         while (++depth < path->p_depth) {
1416                 bh = sb_bread(inode->i_sb, block);
1417                 if (bh == NULL)
1418                         return -EIO;
1419                 eh = ext_block_hdr(bh);
1420                 /* subtract from p_depth to get proper eh_depth */
1421                 if (ext4_ext_check_block(inode, eh,
1422                                          path->p_depth - depth, bh)) {
1423                         put_bh(bh);
1424                         return -EIO;
1425                 }
1426                 ix = EXT_FIRST_INDEX(eh);
1427                 block = ext4_idx_pblock(ix);
1428                 put_bh(bh);
1429         }
1430
1431         bh = sb_bread(inode->i_sb, block);
1432         if (bh == NULL)
1433                 return -EIO;
1434         eh = ext_block_hdr(bh);
1435         if (ext4_ext_check_block(inode, eh, path->p_depth - depth, bh)) {
1436                 put_bh(bh);
1437                 return -EIO;
1438         }
1439         ex = EXT_FIRST_EXTENT(eh);
1440 found_extent:
1441         *logical = le32_to_cpu(ex->ee_block);
1442         *phys = ext4_ext_pblock(ex);
1443         *ret_ex = ex;
1444         if (bh)
1445                 put_bh(bh);
1446         return 0;
1447 }
1448
1449 /*
1450  * ext4_ext_next_allocated_block:
1451  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1452  * NOTE: it considers block number from index entry as
1453  * allocated block. Thus, index entries have to be consistent
1454  * with leaves.
1455  */
1456 static ext4_lblk_t
1457 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1458 {
1459         int depth;
1460
1461         BUG_ON(path == NULL);
1462         depth = path->p_depth;
1463
1464         if (depth == 0 && path->p_ext == NULL)
1465                 return EXT_MAX_BLOCKS;
1466
1467         while (depth >= 0) {
1468                 if (depth == path->p_depth) {
1469                         /* leaf */
1470                         if (path[depth].p_ext &&
1471                                 path[depth].p_ext !=
1472                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1473                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1474                 } else {
1475                         /* index */
1476                         if (path[depth].p_idx !=
1477                                         EXT_LAST_INDEX(path[depth].p_hdr))
1478                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1479                 }
1480                 depth--;
1481         }
1482
1483         return EXT_MAX_BLOCKS;
1484 }
1485
1486 /*
1487  * ext4_ext_next_leaf_block:
1488  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1489  */
1490 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1491 {
1492         int depth;
1493
1494         BUG_ON(path == NULL);
1495         depth = path->p_depth;
1496
1497         /* zero-tree has no leaf blocks at all */
1498         if (depth == 0)
1499                 return EXT_MAX_BLOCKS;
1500
1501         /* go to index block */
1502         depth--;
1503
1504         while (depth >= 0) {
1505                 if (path[depth].p_idx !=
1506                                 EXT_LAST_INDEX(path[depth].p_hdr))
1507                         return (ext4_lblk_t)
1508                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1509                 depth--;
1510         }
1511
1512         return EXT_MAX_BLOCKS;
1513 }
1514
1515 /*
1516  * ext4_ext_correct_indexes:
1517  * if leaf gets modified and modified extent is first in the leaf,
1518  * then we have to correct all indexes above.
1519  * TODO: do we need to correct tree in all cases?
1520  */
1521 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1522                                 struct ext4_ext_path *path)
1523 {
1524         struct ext4_extent_header *eh;
1525         int depth = ext_depth(inode);
1526         struct ext4_extent *ex;
1527         __le32 border;
1528         int k, err = 0;
1529
1530         eh = path[depth].p_hdr;
1531         ex = path[depth].p_ext;
1532
1533         if (unlikely(ex == NULL || eh == NULL)) {
1534                 EXT4_ERROR_INODE(inode,
1535                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1536                 return -EIO;
1537         }
1538
1539         if (depth == 0) {
1540                 /* there is no tree at all */
1541                 return 0;
1542         }
1543
1544         if (ex != EXT_FIRST_EXTENT(eh)) {
1545                 /* we correct tree if first leaf got modified only */
1546                 return 0;
1547         }
1548
1549         /*
1550          * TODO: we need correction if border is smaller than current one
1551          */
1552         k = depth - 1;
1553         border = path[depth].p_ext->ee_block;
1554         err = ext4_ext_get_access(handle, inode, path + k);
1555         if (err)
1556                 return err;
1557         path[k].p_idx->ei_block = border;
1558         err = ext4_ext_dirty(handle, inode, path + k);
1559         if (err)
1560                 return err;
1561
1562         while (k--) {
1563                 /* change all left-side indexes */
1564                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1565                         break;
1566                 err = ext4_ext_get_access(handle, inode, path + k);
1567                 if (err)
1568                         break;
1569                 path[k].p_idx->ei_block = border;
1570                 err = ext4_ext_dirty(handle, inode, path + k);
1571                 if (err)
1572                         break;
1573         }
1574
1575         return err;
1576 }
1577
1578 int
1579 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1580                                 struct ext4_extent *ex2)
1581 {
1582         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1583
1584         /*
1585          * Make sure that either both extents are uninitialized, or
1586          * both are _not_.
1587          */
1588         if (ext4_ext_is_uninitialized(ex1) ^ ext4_ext_is_uninitialized(ex2))
1589                 return 0;
1590
1591         if (ext4_ext_is_uninitialized(ex1))
1592                 max_len = EXT_UNINIT_MAX_LEN;
1593         else
1594                 max_len = EXT_INIT_MAX_LEN;
1595
1596         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1597         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1598
1599         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1600                         le32_to_cpu(ex2->ee_block))
1601                 return 0;
1602
1603         /*
1604          * To allow future support for preallocated extents to be added
1605          * as an RO_COMPAT feature, refuse to merge to extents if
1606          * this can result in the top bit of ee_len being set.
1607          */
1608         if (ext1_ee_len + ext2_ee_len > max_len)
1609                 return 0;
1610 #ifdef AGGRESSIVE_TEST
1611         if (ext1_ee_len >= 4)
1612                 return 0;
1613 #endif
1614
1615         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1616                 return 1;
1617         return 0;
1618 }
1619
1620 /*
1621  * This function tries to merge the "ex" extent to the next extent in the tree.
1622  * It always tries to merge towards right. If you want to merge towards
1623  * left, pass "ex - 1" as argument instead of "ex".
1624  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1625  * 1 if they got merged.
1626  */
1627 static int ext4_ext_try_to_merge_right(struct inode *inode,
1628                                  struct ext4_ext_path *path,
1629                                  struct ext4_extent *ex)
1630 {
1631         struct ext4_extent_header *eh;
1632         unsigned int depth, len;
1633         int merge_done = 0;
1634         int uninitialized = 0;
1635
1636         depth = ext_depth(inode);
1637         BUG_ON(path[depth].p_hdr == NULL);
1638         eh = path[depth].p_hdr;
1639
1640         while (ex < EXT_LAST_EXTENT(eh)) {
1641                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1642                         break;
1643                 /* merge with next extent! */
1644                 if (ext4_ext_is_uninitialized(ex))
1645                         uninitialized = 1;
1646                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1647                                 + ext4_ext_get_actual_len(ex + 1));
1648                 if (uninitialized)
1649                         ext4_ext_mark_uninitialized(ex);
1650
1651                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1652                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1653                                 * sizeof(struct ext4_extent);
1654                         memmove(ex + 1, ex + 2, len);
1655                 }
1656                 le16_add_cpu(&eh->eh_entries, -1);
1657                 merge_done = 1;
1658                 WARN_ON(eh->eh_entries == 0);
1659                 if (!eh->eh_entries)
1660                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1661         }
1662
1663         return merge_done;
1664 }
1665
1666 /*
1667  * This function does a very simple check to see if we can collapse
1668  * an extent tree with a single extent tree leaf block into the inode.
1669  */
1670 static void ext4_ext_try_to_merge_up(handle_t *handle,
1671                                      struct inode *inode,
1672                                      struct ext4_ext_path *path)
1673 {
1674         size_t s;
1675         unsigned max_root = ext4_ext_space_root(inode, 0);
1676         ext4_fsblk_t blk;
1677
1678         if ((path[0].p_depth != 1) ||
1679             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1680             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1681                 return;
1682
1683         /*
1684          * We need to modify the block allocation bitmap and the block
1685          * group descriptor to release the extent tree block.  If we
1686          * can't get the journal credits, give up.
1687          */
1688         if (ext4_journal_extend(handle, 2))
1689                 return;
1690
1691         /*
1692          * Copy the extent data up to the inode
1693          */
1694         blk = ext4_idx_pblock(path[0].p_idx);
1695         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1696                 sizeof(struct ext4_extent_idx);
1697         s += sizeof(struct ext4_extent_header);
1698
1699         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1700         path[0].p_depth = 0;
1701         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1702                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1703         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1704
1705         brelse(path[1].p_bh);
1706         ext4_free_blocks(handle, inode, NULL, blk, 1,
1707                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1708 }
1709
1710 /*
1711  * This function tries to merge the @ex extent to neighbours in the tree.
1712  * return 1 if merge left else 0.
1713  */
1714 static void ext4_ext_try_to_merge(handle_t *handle,
1715                                   struct inode *inode,
1716                                   struct ext4_ext_path *path,
1717                                   struct ext4_extent *ex) {
1718         struct ext4_extent_header *eh;
1719         unsigned int depth;
1720         int merge_done = 0;
1721
1722         depth = ext_depth(inode);
1723         BUG_ON(path[depth].p_hdr == NULL);
1724         eh = path[depth].p_hdr;
1725
1726         if (ex > EXT_FIRST_EXTENT(eh))
1727                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1728
1729         if (!merge_done)
1730                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1731
1732         ext4_ext_try_to_merge_up(handle, inode, path);
1733 }
1734
1735 /*
1736  * check if a portion of the "newext" extent overlaps with an
1737  * existing extent.
1738  *
1739  * If there is an overlap discovered, it updates the length of the newext
1740  * such that there will be no overlap, and then returns 1.
1741  * If there is no overlap found, it returns 0.
1742  */
1743 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1744                                            struct inode *inode,
1745                                            struct ext4_extent *newext,
1746                                            struct ext4_ext_path *path)
1747 {
1748         ext4_lblk_t b1, b2;
1749         unsigned int depth, len1;
1750         unsigned int ret = 0;
1751
1752         b1 = le32_to_cpu(newext->ee_block);
1753         len1 = ext4_ext_get_actual_len(newext);
1754         depth = ext_depth(inode);
1755         if (!path[depth].p_ext)
1756                 goto out;
1757         b2 = le32_to_cpu(path[depth].p_ext->ee_block);
1758         b2 &= ~(sbi->s_cluster_ratio - 1);
1759
1760         /*
1761          * get the next allocated block if the extent in the path
1762          * is before the requested block(s)
1763          */
1764         if (b2 < b1) {
1765                 b2 = ext4_ext_next_allocated_block(path);
1766                 if (b2 == EXT_MAX_BLOCKS)
1767                         goto out;
1768                 b2 &= ~(sbi->s_cluster_ratio - 1);
1769         }
1770
1771         /* check for wrap through zero on extent logical start block*/
1772         if (b1 + len1 < b1) {
1773                 len1 = EXT_MAX_BLOCKS - b1;
1774                 newext->ee_len = cpu_to_le16(len1);
1775                 ret = 1;
1776         }
1777
1778         /* check for overlap */
1779         if (b1 + len1 > b2) {
1780                 newext->ee_len = cpu_to_le16(b2 - b1);
1781                 ret = 1;
1782         }
1783 out:
1784         return ret;
1785 }
1786
1787 /*
1788  * ext4_ext_insert_extent:
1789  * tries to merge requsted extent into the existing extent or
1790  * inserts requested extent as new one into the tree,
1791  * creating new leaf in the no-space case.
1792  */
1793 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1794                                 struct ext4_ext_path *path,
1795                                 struct ext4_extent *newext, int flag)
1796 {
1797         struct ext4_extent_header *eh;
1798         struct ext4_extent *ex, *fex;
1799         struct ext4_extent *nearex; /* nearest extent */
1800         struct ext4_ext_path *npath = NULL;
1801         int depth, len, err;
1802         ext4_lblk_t next;
1803         unsigned uninitialized = 0;
1804         int flags = 0;
1805
1806         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1807                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1808                 return -EIO;
1809         }
1810         depth = ext_depth(inode);
1811         ex = path[depth].p_ext;
1812         if (unlikely(path[depth].p_hdr == NULL)) {
1813                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1814                 return -EIO;
1815         }
1816
1817         /* try to insert block into found extent and return */
1818         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)
1819                 && ext4_can_extents_be_merged(inode, ex, newext)) {
1820                 ext_debug("append [%d]%d block to %u:[%d]%d (from %llu)\n",
1821                           ext4_ext_is_uninitialized(newext),
1822                           ext4_ext_get_actual_len(newext),
1823                           le32_to_cpu(ex->ee_block),
1824                           ext4_ext_is_uninitialized(ex),
1825                           ext4_ext_get_actual_len(ex),
1826                           ext4_ext_pblock(ex));
1827                 err = ext4_ext_get_access(handle, inode, path + depth);
1828                 if (err)
1829                         return err;
1830
1831                 /*
1832                  * ext4_can_extents_be_merged should have checked that either
1833                  * both extents are uninitialized, or both aren't. Thus we
1834                  * need to check only one of them here.
1835                  */
1836                 if (ext4_ext_is_uninitialized(ex))
1837                         uninitialized = 1;
1838                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1839                                         + ext4_ext_get_actual_len(newext));
1840                 if (uninitialized)
1841                         ext4_ext_mark_uninitialized(ex);
1842                 eh = path[depth].p_hdr;
1843                 nearex = ex;
1844                 goto merge;
1845         }
1846
1847         depth = ext_depth(inode);
1848         eh = path[depth].p_hdr;
1849         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1850                 goto has_space;
1851
1852         /* probably next leaf has space for us? */
1853         fex = EXT_LAST_EXTENT(eh);
1854         next = EXT_MAX_BLOCKS;
1855         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1856                 next = ext4_ext_next_leaf_block(path);
1857         if (next != EXT_MAX_BLOCKS) {
1858                 ext_debug("next leaf block - %u\n", next);
1859                 BUG_ON(npath != NULL);
1860                 npath = ext4_ext_find_extent(inode, next, NULL);
1861                 if (IS_ERR(npath))
1862                         return PTR_ERR(npath);
1863                 BUG_ON(npath->p_depth != path->p_depth);
1864                 eh = npath[depth].p_hdr;
1865                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1866                         ext_debug("next leaf isn't full(%d)\n",
1867                                   le16_to_cpu(eh->eh_entries));
1868                         path = npath;
1869                         goto has_space;
1870                 }
1871                 ext_debug("next leaf has no free space(%d,%d)\n",
1872                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1873         }
1874
1875         /*
1876          * There is no free space in the found leaf.
1877          * We're gonna add a new leaf in the tree.
1878          */
1879         if (flag & EXT4_GET_BLOCKS_PUNCH_OUT_EXT)
1880                 flags = EXT4_MB_USE_ROOT_BLOCKS;
1881         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1882         if (err)
1883                 goto cleanup;
1884         depth = ext_depth(inode);
1885         eh = path[depth].p_hdr;
1886
1887 has_space:
1888         nearex = path[depth].p_ext;
1889
1890         err = ext4_ext_get_access(handle, inode, path + depth);
1891         if (err)
1892                 goto cleanup;
1893
1894         if (!nearex) {
1895                 /* there is no extent in this leaf, create first one */
1896                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1897                                 le32_to_cpu(newext->ee_block),
1898                                 ext4_ext_pblock(newext),
1899                                 ext4_ext_is_uninitialized(newext),
1900                                 ext4_ext_get_actual_len(newext));
1901                 nearex = EXT_FIRST_EXTENT(eh);
1902         } else {
1903                 if (le32_to_cpu(newext->ee_block)
1904                            > le32_to_cpu(nearex->ee_block)) {
1905                         /* Insert after */
1906                         ext_debug("insert %u:%llu:[%d]%d before: "
1907                                         "nearest %p\n",
1908                                         le32_to_cpu(newext->ee_block),
1909                                         ext4_ext_pblock(newext),
1910                                         ext4_ext_is_uninitialized(newext),
1911                                         ext4_ext_get_actual_len(newext),
1912                                         nearex);
1913                         nearex++;
1914                 } else {
1915                         /* Insert before */
1916                         BUG_ON(newext->ee_block == nearex->ee_block);
1917                         ext_debug("insert %u:%llu:[%d]%d after: "
1918                                         "nearest %p\n",
1919                                         le32_to_cpu(newext->ee_block),
1920                                         ext4_ext_pblock(newext),
1921                                         ext4_ext_is_uninitialized(newext),
1922                                         ext4_ext_get_actual_len(newext),
1923                                         nearex);
1924                 }
1925                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1926                 if (len > 0) {
1927                         ext_debug("insert %u:%llu:[%d]%d: "
1928                                         "move %d extents from 0x%p to 0x%p\n",
1929                                         le32_to_cpu(newext->ee_block),
1930                                         ext4_ext_pblock(newext),
1931                                         ext4_ext_is_uninitialized(newext),
1932                                         ext4_ext_get_actual_len(newext),
1933                                         len, nearex, nearex + 1);
1934                         memmove(nearex + 1, nearex,
1935                                 len * sizeof(struct ext4_extent));
1936                 }
1937         }
1938
1939         le16_add_cpu(&eh->eh_entries, 1);
1940         path[depth].p_ext = nearex;
1941         nearex->ee_block = newext->ee_block;
1942         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
1943         nearex->ee_len = newext->ee_len;
1944
1945 merge:
1946         /* try to merge extents */
1947         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
1948                 ext4_ext_try_to_merge(handle, inode, path, nearex);
1949
1950
1951         /* time to correct all indexes above */
1952         err = ext4_ext_correct_indexes(handle, inode, path);
1953         if (err)
1954                 goto cleanup;
1955
1956         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
1957
1958 cleanup:
1959         if (npath) {
1960                 ext4_ext_drop_refs(npath);
1961                 kfree(npath);
1962         }
1963         ext4_ext_invalidate_cache(inode);
1964         return err;
1965 }
1966
1967 static int ext4_fill_fiemap_extents(struct inode *inode,
1968                                     ext4_lblk_t block, ext4_lblk_t num,
1969                                     struct fiemap_extent_info *fieinfo)
1970 {
1971         struct ext4_ext_path *path = NULL;
1972         struct ext4_ext_cache newex;
1973         struct ext4_extent *ex;
1974         ext4_lblk_t next, next_del, start = 0, end = 0;
1975         ext4_lblk_t last = block + num;
1976         int exists, depth = 0, err = 0;
1977         unsigned int flags = 0;
1978         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
1979
1980         while (block < last && block != EXT_MAX_BLOCKS) {
1981                 num = last - block;
1982                 /* find extent for this block */
1983                 down_read(&EXT4_I(inode)->i_data_sem);
1984
1985                 if (path && ext_depth(inode) != depth) {
1986                         /* depth was changed. we have to realloc path */
1987                         kfree(path);
1988                         path = NULL;
1989                 }
1990
1991                 path = ext4_ext_find_extent(inode, block, path);
1992                 if (IS_ERR(path)) {
1993                         up_read(&EXT4_I(inode)->i_data_sem);
1994                         err = PTR_ERR(path);
1995                         path = NULL;
1996                         break;
1997                 }
1998
1999                 depth = ext_depth(inode);
2000                 if (unlikely(path[depth].p_hdr == NULL)) {
2001                         up_read(&EXT4_I(inode)->i_data_sem);
2002                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2003                         err = -EIO;
2004                         break;
2005                 }
2006                 ex = path[depth].p_ext;
2007                 next = ext4_ext_next_allocated_block(path);
2008                 ext4_ext_drop_refs(path);
2009
2010                 flags = 0;
2011                 exists = 0;
2012                 if (!ex) {
2013                         /* there is no extent yet, so try to allocate
2014                          * all requested space */
2015                         start = block;
2016                         end = block + num;
2017                 } else if (le32_to_cpu(ex->ee_block) > block) {
2018                         /* need to allocate space before found extent */
2019                         start = block;
2020                         end = le32_to_cpu(ex->ee_block);
2021                         if (block + num < end)
2022                                 end = block + num;
2023                 } else if (block >= le32_to_cpu(ex->ee_block)
2024                                         + ext4_ext_get_actual_len(ex)) {
2025                         /* need to allocate space after found extent */
2026                         start = block;
2027                         end = block + num;
2028                         if (end >= next)
2029                                 end = next;
2030                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2031                         /*
2032                          * some part of requested space is covered
2033                          * by found extent
2034                          */
2035                         start = block;
2036                         end = le32_to_cpu(ex->ee_block)
2037                                 + ext4_ext_get_actual_len(ex);
2038                         if (block + num < end)
2039                                 end = block + num;
2040                         exists = 1;
2041                 } else {
2042                         BUG();
2043                 }
2044                 BUG_ON(end <= start);
2045
2046                 if (!exists) {
2047                         newex.ec_block = start;
2048                         newex.ec_len = end - start;
2049                         newex.ec_start = 0;
2050                 } else {
2051                         newex.ec_block = le32_to_cpu(ex->ee_block);
2052                         newex.ec_len = ext4_ext_get_actual_len(ex);
2053                         newex.ec_start = ext4_ext_pblock(ex);
2054                         if (ext4_ext_is_uninitialized(ex))
2055                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2056                 }
2057
2058                 /*
2059                  * Find delayed extent and update newex accordingly. We call
2060                  * it even in !exists case to find out whether newex is the
2061                  * last existing extent or not.
2062                  */
2063                 next_del = ext4_find_delayed_extent(inode, &newex);
2064                 if (!exists && next_del) {
2065                         exists = 1;
2066                         flags |= FIEMAP_EXTENT_DELALLOC;
2067                 }
2068                 up_read(&EXT4_I(inode)->i_data_sem);
2069
2070                 if (unlikely(newex.ec_len == 0)) {
2071                         EXT4_ERROR_INODE(inode, "newex.ec_len == 0");
2072                         err = -EIO;
2073                         break;
2074                 }
2075
2076                 /* This is possible iff next == next_del == EXT_MAX_BLOCKS */
2077                 if (next == next_del) {
2078                         flags |= FIEMAP_EXTENT_LAST;
2079                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2080                                      next != EXT_MAX_BLOCKS)) {
2081                                 EXT4_ERROR_INODE(inode,
2082                                                  "next extent == %u, next "
2083                                                  "delalloc extent = %u",
2084                                                  next, next_del);
2085                                 err = -EIO;
2086                                 break;
2087                         }
2088                 }
2089
2090                 if (exists) {
2091                         err = fiemap_fill_next_extent(fieinfo,
2092                                 (__u64)newex.ec_block << blksize_bits,
2093                                 (__u64)newex.ec_start << blksize_bits,
2094                                 (__u64)newex.ec_len << blksize_bits,
2095                                 flags);
2096                         if (err < 0)
2097                                 break;
2098                         if (err == 1) {
2099                                 err = 0;
2100                                 break;
2101                         }
2102                 }
2103
2104                 block = newex.ec_block + newex.ec_len;
2105         }
2106
2107         if (path) {
2108                 ext4_ext_drop_refs(path);
2109                 kfree(path);
2110         }
2111
2112         return err;
2113 }
2114
2115 static void
2116 ext4_ext_put_in_cache(struct inode *inode, ext4_lblk_t block,
2117                         __u32 len, ext4_fsblk_t start)
2118 {
2119         struct ext4_ext_cache *cex;
2120         BUG_ON(len == 0);
2121         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
2122         trace_ext4_ext_put_in_cache(inode, block, len, start);
2123         cex = &EXT4_I(inode)->i_cached_extent;
2124         cex->ec_block = block;
2125         cex->ec_len = len;
2126         cex->ec_start = start;
2127         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
2128 }
2129
2130 /*
2131  * ext4_ext_put_gap_in_cache:
2132  * calculate boundaries of the gap that the requested block fits into
2133  * and cache this gap
2134  */
2135 static void
2136 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2137                                 ext4_lblk_t block)
2138 {
2139         int depth = ext_depth(inode);
2140         unsigned long len;
2141         ext4_lblk_t lblock;
2142         struct ext4_extent *ex;
2143
2144         ex = path[depth].p_ext;
2145         if (ex == NULL) {
2146                 /* there is no extent yet, so gap is [0;-] */
2147                 lblock = 0;
2148                 len = EXT_MAX_BLOCKS;
2149                 ext_debug("cache gap(whole file):");
2150         } else if (block < le32_to_cpu(ex->ee_block)) {
2151                 lblock = block;
2152                 len = le32_to_cpu(ex->ee_block) - block;
2153                 ext_debug("cache gap(before): %u [%u:%u]",
2154                                 block,
2155                                 le32_to_cpu(ex->ee_block),
2156                                  ext4_ext_get_actual_len(ex));
2157         } else if (block >= le32_to_cpu(ex->ee_block)
2158                         + ext4_ext_get_actual_len(ex)) {
2159                 ext4_lblk_t next;
2160                 lblock = le32_to_cpu(ex->ee_block)
2161                         + ext4_ext_get_actual_len(ex);
2162
2163                 next = ext4_ext_next_allocated_block(path);
2164                 ext_debug("cache gap(after): [%u:%u] %u",
2165                                 le32_to_cpu(ex->ee_block),
2166                                 ext4_ext_get_actual_len(ex),
2167                                 block);
2168                 BUG_ON(next == lblock);
2169                 len = next - lblock;
2170         } else {
2171                 lblock = len = 0;
2172                 BUG();
2173         }
2174
2175         ext_debug(" -> %u:%lu\n", lblock, len);
2176         ext4_ext_put_in_cache(inode, lblock, len, 0);
2177 }
2178
2179 /*
2180  * ext4_ext_in_cache()
2181  * Checks to see if the given block is in the cache.
2182  * If it is, the cached extent is stored in the given
2183  * cache extent pointer.
2184  *
2185  * @inode: The files inode
2186  * @block: The block to look for in the cache
2187  * @ex:    Pointer where the cached extent will be stored
2188  *         if it contains block
2189  *
2190  * Return 0 if cache is invalid; 1 if the cache is valid
2191  */
2192 static int
2193 ext4_ext_in_cache(struct inode *inode, ext4_lblk_t block,
2194                   struct ext4_extent *ex)
2195 {
2196         struct ext4_ext_cache *cex;
2197         int ret = 0;
2198
2199         /*
2200          * We borrow i_block_reservation_lock to protect i_cached_extent
2201          */
2202         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
2203         cex = &EXT4_I(inode)->i_cached_extent;
2204
2205         /* has cache valid data? */
2206         if (cex->ec_len == 0)
2207                 goto errout;
2208
2209         if (in_range(block, cex->ec_block, cex->ec_len)) {
2210                 ex->ee_block = cpu_to_le32(cex->ec_block);
2211                 ext4_ext_store_pblock(ex, cex->ec_start);
2212                 ex->ee_len = cpu_to_le16(cex->ec_len);
2213                 ext_debug("%u cached by %u:%u:%llu\n",
2214                                 block,
2215                                 cex->ec_block, cex->ec_len, cex->ec_start);
2216                 ret = 1;
2217         }
2218 errout:
2219         trace_ext4_ext_in_cache(inode, block, ret);
2220         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
2221         return ret;
2222 }
2223
2224 /*
2225  * ext4_ext_rm_idx:
2226  * removes index from the index block.
2227  */
2228 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2229                         struct ext4_ext_path *path)
2230 {
2231         int err;
2232         ext4_fsblk_t leaf;
2233
2234         /* free index block */
2235         path--;
2236         leaf = ext4_idx_pblock(path->p_idx);
2237         if (unlikely(path->p_hdr->eh_entries == 0)) {
2238                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2239                 return -EIO;
2240         }
2241         err = ext4_ext_get_access(handle, inode, path);
2242         if (err)
2243                 return err;
2244
2245         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2246                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2247                 len *= sizeof(struct ext4_extent_idx);
2248                 memmove(path->p_idx, path->p_idx + 1, len);
2249         }
2250
2251         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2252         err = ext4_ext_dirty(handle, inode, path);
2253         if (err)
2254                 return err;
2255         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2256         trace_ext4_ext_rm_idx(inode, leaf);
2257
2258         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2259                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2260         return err;
2261 }
2262
2263 /*
2264  * ext4_ext_calc_credits_for_single_extent:
2265  * This routine returns max. credits that needed to insert an extent
2266  * to the extent tree.
2267  * When pass the actual path, the caller should calculate credits
2268  * under i_data_sem.
2269  */
2270 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2271                                                 struct ext4_ext_path *path)
2272 {
2273         if (path) {
2274                 int depth = ext_depth(inode);
2275                 int ret = 0;
2276
2277                 /* probably there is space in leaf? */
2278                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2279                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2280
2281                         /*
2282                          *  There are some space in the leaf tree, no
2283                          *  need to account for leaf block credit
2284                          *
2285                          *  bitmaps and block group descriptor blocks
2286                          *  and other metadata blocks still need to be
2287                          *  accounted.
2288                          */
2289                         /* 1 bitmap, 1 block group descriptor */
2290                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2291                         return ret;
2292                 }
2293         }
2294
2295         return ext4_chunk_trans_blocks(inode, nrblocks);
2296 }
2297
2298 /*
2299  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2300  *
2301  * if nrblocks are fit in a single extent (chunk flag is 1), then
2302  * in the worse case, each tree level index/leaf need to be changed
2303  * if the tree split due to insert a new extent, then the old tree
2304  * index/leaf need to be updated too
2305  *
2306  * If the nrblocks are discontiguous, they could cause
2307  * the whole tree split more than once, but this is really rare.
2308  */
2309 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2310 {
2311         int index;
2312         int depth;
2313
2314         /* If we are converting the inline data, only one is needed here. */
2315         if (ext4_has_inline_data(inode))
2316                 return 1;
2317
2318         depth = ext_depth(inode);
2319
2320         if (chunk)
2321                 index = depth * 2;
2322         else
2323                 index = depth * 3;
2324
2325         return index;
2326 }
2327
2328 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2329                               struct ext4_extent *ex,
2330                               ext4_fsblk_t *partial_cluster,
2331                               ext4_lblk_t from, ext4_lblk_t to)
2332 {
2333         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2334         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2335         ext4_fsblk_t pblk;
2336         int flags = 0;
2337
2338         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2339                 flags |= EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2340         else if (ext4_should_journal_data(inode))
2341                 flags |= EXT4_FREE_BLOCKS_FORGET;
2342
2343         /*
2344          * For bigalloc file systems, we never free a partial cluster
2345          * at the beginning of the extent.  Instead, we make a note
2346          * that we tried freeing the cluster, and check to see if we
2347          * need to free it on a subsequent call to ext4_remove_blocks,
2348          * or at the end of the ext4_truncate() operation.
2349          */
2350         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2351
2352         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2353         /*
2354          * If we have a partial cluster, and it's different from the
2355          * cluster of the last block, we need to explicitly free the
2356          * partial cluster here.
2357          */
2358         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2359         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2360                 ext4_free_blocks(handle, inode, NULL,
2361                                  EXT4_C2B(sbi, *partial_cluster),
2362                                  sbi->s_cluster_ratio, flags);
2363                 *partial_cluster = 0;
2364         }
2365
2366 #ifdef EXTENTS_STATS
2367         {
2368                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2369                 spin_lock(&sbi->s_ext_stats_lock);
2370                 sbi->s_ext_blocks += ee_len;
2371                 sbi->s_ext_extents++;
2372                 if (ee_len < sbi->s_ext_min)
2373                         sbi->s_ext_min = ee_len;
2374                 if (ee_len > sbi->s_ext_max)
2375                         sbi->s_ext_max = ee_len;
2376                 if (ext_depth(inode) > sbi->s_depth_max)
2377                         sbi->s_depth_max = ext_depth(inode);
2378                 spin_unlock(&sbi->s_ext_stats_lock);
2379         }
2380 #endif
2381         if (from >= le32_to_cpu(ex->ee_block)
2382             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2383                 /* tail removal */
2384                 ext4_lblk_t num;
2385
2386                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2387                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2388                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2389                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2390                 /*
2391                  * If the block range to be freed didn't start at the
2392                  * beginning of a cluster, and we removed the entire
2393                  * extent, save the partial cluster here, since we
2394                  * might need to delete if we determine that the
2395                  * truncate operation has removed all of the blocks in
2396                  * the cluster.
2397                  */
2398                 if (pblk & (sbi->s_cluster_ratio - 1) &&
2399                     (ee_len == num))
2400                         *partial_cluster = EXT4_B2C(sbi, pblk);
2401                 else
2402                         *partial_cluster = 0;
2403         } else if (from == le32_to_cpu(ex->ee_block)
2404                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2405                 /* head removal */
2406                 ext4_lblk_t num;
2407                 ext4_fsblk_t start;
2408
2409                 num = to - from;
2410                 start = ext4_ext_pblock(ex);
2411
2412                 ext_debug("free first %u blocks starting %llu\n", num, start);
2413                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2414
2415         } else {
2416                 printk(KERN_INFO "strange request: removal(2) "
2417                                 "%u-%u from %u:%u\n",
2418                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2419         }
2420         return 0;
2421 }
2422
2423
2424 /*
2425  * ext4_ext_rm_leaf() Removes the extents associated with the
2426  * blocks appearing between "start" and "end", and splits the extents
2427  * if "start" and "end" appear in the same extent
2428  *
2429  * @handle: The journal handle
2430  * @inode:  The files inode
2431  * @path:   The path to the leaf
2432  * @start:  The first block to remove
2433  * @end:   The last block to remove
2434  */
2435 static int
2436 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2437                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2438                  ext4_lblk_t start, ext4_lblk_t end)
2439 {
2440         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2441         int err = 0, correct_index = 0;
2442         int depth = ext_depth(inode), credits;
2443         struct ext4_extent_header *eh;
2444         ext4_lblk_t a, b;
2445         unsigned num;
2446         ext4_lblk_t ex_ee_block;
2447         unsigned short ex_ee_len;
2448         unsigned uninitialized = 0;
2449         struct ext4_extent *ex;
2450
2451         /* the header must be checked already in ext4_ext_remove_space() */
2452         ext_debug("truncate since %u in leaf to %u\n", start, end);
2453         if (!path[depth].p_hdr)
2454                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2455         eh = path[depth].p_hdr;
2456         if (unlikely(path[depth].p_hdr == NULL)) {
2457                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2458                 return -EIO;
2459         }
2460         /* find where to start removing */
2461         ex = EXT_LAST_EXTENT(eh);
2462
2463         ex_ee_block = le32_to_cpu(ex->ee_block);
2464         ex_ee_len = ext4_ext_get_actual_len(ex);
2465
2466         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2467
2468         while (ex >= EXT_FIRST_EXTENT(eh) &&
2469                         ex_ee_block + ex_ee_len > start) {
2470
2471                 if (ext4_ext_is_uninitialized(ex))
2472                         uninitialized = 1;
2473                 else
2474                         uninitialized = 0;
2475
2476                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2477                          uninitialized, ex_ee_len);
2478                 path[depth].p_ext = ex;
2479
2480                 a = ex_ee_block > start ? ex_ee_block : start;
2481                 b = ex_ee_block+ex_ee_len - 1 < end ?
2482                         ex_ee_block+ex_ee_len - 1 : end;
2483
2484                 ext_debug("  border %u:%u\n", a, b);
2485
2486                 /* If this extent is beyond the end of the hole, skip it */
2487                 if (end < ex_ee_block) {
2488                         ex--;
2489                         ex_ee_block = le32_to_cpu(ex->ee_block);
2490                         ex_ee_len = ext4_ext_get_actual_len(ex);
2491                         continue;
2492                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2493                         EXT4_ERROR_INODE(inode,
2494                                          "can not handle truncate %u:%u "
2495                                          "on extent %u:%u",
2496                                          start, end, ex_ee_block,
2497                                          ex_ee_block + ex_ee_len - 1);
2498                         err = -EIO;
2499                         goto out;
2500                 } else if (a != ex_ee_block) {
2501                         /* remove tail of the extent */
2502                         num = a - ex_ee_block;
2503                 } else {
2504                         /* remove whole extent: excellent! */
2505                         num = 0;
2506                 }
2507                 /*
2508                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2509                  * descriptor) for each block group; assume two block
2510                  * groups plus ex_ee_len/blocks_per_block_group for
2511                  * the worst case
2512                  */
2513                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2514                 if (ex == EXT_FIRST_EXTENT(eh)) {
2515                         correct_index = 1;
2516                         credits += (ext_depth(inode)) + 1;
2517                 }
2518                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2519
2520                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2521                 if (err)
2522                         goto out;
2523
2524                 err = ext4_ext_get_access(handle, inode, path + depth);
2525                 if (err)
2526                         goto out;
2527
2528                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2529                                          a, b);
2530                 if (err)
2531                         goto out;
2532
2533                 if (num == 0)
2534                         /* this extent is removed; mark slot entirely unused */
2535                         ext4_ext_store_pblock(ex, 0);
2536
2537                 ex->ee_len = cpu_to_le16(num);
2538                 /*
2539                  * Do not mark uninitialized if all the blocks in the
2540                  * extent have been removed.
2541                  */
2542                 if (uninitialized && num)
2543                         ext4_ext_mark_uninitialized(ex);
2544                 /*
2545                  * If the extent was completely released,
2546                  * we need to remove it from the leaf
2547                  */
2548                 if (num == 0) {
2549                         if (end != EXT_MAX_BLOCKS - 1) {
2550                                 /*
2551                                  * For hole punching, we need to scoot all the
2552                                  * extents up when an extent is removed so that
2553                                  * we dont have blank extents in the middle
2554                                  */
2555                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2556                                         sizeof(struct ext4_extent));
2557
2558                                 /* Now get rid of the one at the end */
2559                                 memset(EXT_LAST_EXTENT(eh), 0,
2560                                         sizeof(struct ext4_extent));
2561                         }
2562                         le16_add_cpu(&eh->eh_entries, -1);
2563                 } else
2564                         *partial_cluster = 0;
2565
2566                 err = ext4_ext_dirty(handle, inode, path + depth);
2567                 if (err)
2568                         goto out;
2569
2570                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2571                                 ext4_ext_pblock(ex));
2572                 ex--;
2573                 ex_ee_block = le32_to_cpu(ex->ee_block);
2574                 ex_ee_len = ext4_ext_get_actual_len(ex);
2575         }
2576
2577         if (correct_index && eh->eh_entries)
2578                 err = ext4_ext_correct_indexes(handle, inode, path);
2579
2580         /*
2581          * If there is still a entry in the leaf node, check to see if
2582          * it references the partial cluster.  This is the only place
2583          * where it could; if it doesn't, we can free the cluster.
2584          */
2585         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2586             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2587              *partial_cluster)) {
2588                 int flags = EXT4_FREE_BLOCKS_FORGET;
2589
2590                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2591                         flags |= EXT4_FREE_BLOCKS_METADATA;
2592
2593                 ext4_free_blocks(handle, inode, NULL,
2594                                  EXT4_C2B(sbi, *partial_cluster),
2595                                  sbi->s_cluster_ratio, flags);
2596                 *partial_cluster = 0;
2597         }
2598
2599         /* if this leaf is free, then we should
2600          * remove it from index block above */
2601         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2602                 err = ext4_ext_rm_idx(handle, inode, path + depth);
2603
2604 out:
2605         return err;
2606 }
2607
2608 /*
2609  * ext4_ext_more_to_rm:
2610  * returns 1 if current index has to be freed (even partial)
2611  */
2612 static int
2613 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2614 {
2615         BUG_ON(path->p_idx == NULL);
2616
2617         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2618                 return 0;
2619
2620         /*
2621          * if truncate on deeper level happened, it wasn't partial,
2622          * so we have to consider current index for truncation
2623          */
2624         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2625                 return 0;
2626         return 1;
2627 }
2628
2629 static int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2630                                  ext4_lblk_t end)
2631 {
2632         struct super_block *sb = inode->i_sb;
2633         int depth = ext_depth(inode);
2634         struct ext4_ext_path *path = NULL;
2635         ext4_fsblk_t partial_cluster = 0;
2636         handle_t *handle;
2637         int i = 0, err = 0;
2638
2639         ext_debug("truncate since %u to %u\n", start, end);
2640
2641         /* probably first extent we're gonna free will be last in block */
2642         handle = ext4_journal_start(inode, depth + 1);
2643         if (IS_ERR(handle))
2644                 return PTR_ERR(handle);
2645
2646 again:
2647         ext4_ext_invalidate_cache(inode);
2648
2649         trace_ext4_ext_remove_space(inode, start, depth);
2650
2651         /*
2652          * Check if we are removing extents inside the extent tree. If that
2653          * is the case, we are going to punch a hole inside the extent tree
2654          * so we have to check whether we need to split the extent covering
2655          * the last block to remove so we can easily remove the part of it
2656          * in ext4_ext_rm_leaf().
2657          */
2658         if (end < EXT_MAX_BLOCKS - 1) {
2659                 struct ext4_extent *ex;
2660                 ext4_lblk_t ee_block;
2661
2662                 /* find extent for this block */
2663                 path = ext4_ext_find_extent(inode, end, NULL);
2664                 if (IS_ERR(path)) {
2665                         ext4_journal_stop(handle);
2666                         return PTR_ERR(path);
2667                 }
2668                 depth = ext_depth(inode);
2669                 /* Leaf not may not exist only if inode has no blocks at all */
2670                 ex = path[depth].p_ext;
2671                 if (!ex) {
2672                         if (depth) {
2673                                 EXT4_ERROR_INODE(inode,
2674                                                  "path[%d].p_hdr == NULL",
2675                                                  depth);
2676                                 err = -EIO;
2677                         }
2678                         goto out;
2679                 }
2680
2681                 ee_block = le32_to_cpu(ex->ee_block);
2682
2683                 /*
2684                  * See if the last block is inside the extent, if so split
2685                  * the extent at 'end' block so we can easily remove the
2686                  * tail of the first part of the split extent in
2687                  * ext4_ext_rm_leaf().
2688                  */
2689                 if (end >= ee_block &&
2690                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2691                         int split_flag = 0;
2692
2693                         if (ext4_ext_is_uninitialized(ex))
2694                                 split_flag = EXT4_EXT_MARK_UNINIT1 |
2695                                              EXT4_EXT_MARK_UNINIT2;
2696
2697                         /*
2698                          * Split the extent in two so that 'end' is the last
2699                          * block in the first new extent
2700                          */
2701                         err = ext4_split_extent_at(handle, inode, path,
2702                                                 end + 1, split_flag,
2703                                                 EXT4_GET_BLOCKS_PRE_IO |
2704                                                 EXT4_GET_BLOCKS_PUNCH_OUT_EXT);
2705
2706                         if (err < 0)
2707                                 goto out;
2708                 }
2709         }
2710         /*
2711          * We start scanning from right side, freeing all the blocks
2712          * after i_size and walking into the tree depth-wise.
2713          */
2714         depth = ext_depth(inode);
2715         if (path) {
2716                 int k = i = depth;
2717                 while (--k > 0)
2718                         path[k].p_block =
2719                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2720         } else {
2721                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2722                                GFP_NOFS);
2723                 if (path == NULL) {
2724                         ext4_journal_stop(handle);
2725                         return -ENOMEM;
2726                 }
2727                 path[0].p_depth = depth;
2728                 path[0].p_hdr = ext_inode_hdr(inode);
2729                 i = 0;
2730
2731                 if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2732                         err = -EIO;
2733                         goto out;
2734                 }
2735         }
2736         err = 0;
2737
2738         while (i >= 0 && err == 0) {
2739                 if (i == depth) {
2740                         /* this is leaf block */
2741                         err = ext4_ext_rm_leaf(handle, inode, path,
2742                                                &partial_cluster, start,
2743                                                end);
2744                         /* root level has p_bh == NULL, brelse() eats this */
2745                         brelse(path[i].p_bh);
2746                         path[i].p_bh = NULL;
2747                         i--;
2748                         continue;
2749                 }
2750
2751                 /* this is index block */
2752                 if (!path[i].p_hdr) {
2753                         ext_debug("initialize header\n");
2754                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2755                 }
2756
2757                 if (!path[i].p_idx) {
2758                         /* this level hasn't been touched yet */
2759                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2760                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2761                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2762                                   path[i].p_hdr,
2763                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2764                 } else {
2765                         /* we were already here, see at next index */
2766                         path[i].p_idx--;
2767                 }
2768
2769                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2770                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2771                                 path[i].p_idx);
2772                 if (ext4_ext_more_to_rm(path + i)) {
2773                         struct buffer_head *bh;
2774                         /* go to the next level */
2775                         ext_debug("move to level %d (block %llu)\n",
2776                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2777                         memset(path + i + 1, 0, sizeof(*path));
2778                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2779                         if (!bh) {
2780                                 /* should we reset i_size? */
2781                                 err = -EIO;
2782                                 break;
2783                         }
2784                         if (WARN_ON(i + 1 > depth)) {
2785                                 err = -EIO;
2786                                 break;
2787                         }
2788                         if (ext4_ext_check_block(inode, ext_block_hdr(bh),
2789                                                         depth - i - 1, bh)) {
2790                                 err = -EIO;
2791                                 break;
2792                         }
2793                         path[i + 1].p_bh = bh;
2794
2795                         /* save actual number of indexes since this
2796                          * number is changed at the next iteration */
2797                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2798                         i++;
2799                 } else {
2800                         /* we finished processing this index, go up */
2801                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2802                                 /* index is empty, remove it;
2803                                  * handle must be already prepared by the
2804                                  * truncatei_leaf() */
2805                                 err = ext4_ext_rm_idx(handle, inode, path + i);
2806                         }
2807                         /* root level has p_bh == NULL, brelse() eats this */
2808                         brelse(path[i].p_bh);
2809                         path[i].p_bh = NULL;
2810                         i--;
2811                         ext_debug("return to level %d\n", i);
2812                 }
2813         }
2814
2815         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2816                         path->p_hdr->eh_entries);
2817
2818         /* If we still have something in the partial cluster and we have removed
2819          * even the first extent, then we should free the blocks in the partial
2820          * cluster as well. */
2821         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2822                 int flags = EXT4_FREE_BLOCKS_FORGET;
2823
2824                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2825                         flags |= EXT4_FREE_BLOCKS_METADATA;
2826
2827                 ext4_free_blocks(handle, inode, NULL,
2828                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2829                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2830                 partial_cluster = 0;
2831         }
2832
2833         /* TODO: flexible tree reduction should be here */
2834         if (path->p_hdr->eh_entries == 0) {
2835                 /*
2836                  * truncate to zero freed all the tree,
2837                  * so we need to correct eh_depth
2838                  */
2839                 err = ext4_ext_get_access(handle, inode, path);
2840                 if (err == 0) {
2841                         ext_inode_hdr(inode)->eh_depth = 0;
2842                         ext_inode_hdr(inode)->eh_max =
2843                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2844                         err = ext4_ext_dirty(handle, inode, path);
2845                 }
2846         }
2847 out:
2848         ext4_ext_drop_refs(path);
2849         kfree(path);
2850         if (err == -EAGAIN) {
2851                 path = NULL;
2852                 goto again;
2853         }
2854         ext4_journal_stop(handle);
2855
2856         return err;
2857 }
2858
2859 /*
2860  * called at mount time
2861  */
2862 void ext4_ext_init(struct super_block *sb)
2863 {
2864         /*
2865          * possible initialization would be here
2866          */
2867
2868         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2869 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2870                 printk(KERN_INFO "EXT4-fs: file extents enabled"
2871 #ifdef AGGRESSIVE_TEST
2872                        ", aggressive tests"
2873 #endif
2874 #ifdef CHECK_BINSEARCH
2875                        ", check binsearch"
2876 #endif
2877 #ifdef EXTENTS_STATS
2878                        ", stats"
2879 #endif
2880                        "\n");
2881 #endif
2882 #ifdef EXTENTS_STATS
2883                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2884                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2885                 EXT4_SB(sb)->s_ext_max = 0;
2886 #endif
2887         }
2888 }
2889
2890 /*
2891  * called at umount time
2892  */
2893 void ext4_ext_release(struct super_block *sb)
2894 {
2895         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2896                 return;
2897
2898 #ifdef EXTENTS_STATS
2899         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2900                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2901                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2902                         sbi->s_ext_blocks, sbi->s_ext_extents,
2903                         sbi->s_ext_blocks / sbi->s_ext_extents);
2904                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2905                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2906         }
2907 #endif
2908 }
2909
2910 /* FIXME!! we need to try to merge to left or right after zero-out  */
2911 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2912 {
2913         ext4_fsblk_t ee_pblock;
2914         unsigned int ee_len;
2915         int ret;
2916
2917         ee_len    = ext4_ext_get_actual_len(ex);
2918         ee_pblock = ext4_ext_pblock(ex);
2919
2920         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2921         if (ret > 0)
2922                 ret = 0;
2923
2924         return ret;
2925 }
2926
2927 /*
2928  * ext4_split_extent_at() splits an extent at given block.
2929  *
2930  * @handle: the journal handle
2931  * @inode: the file inode
2932  * @path: the path to the extent
2933  * @split: the logical block where the extent is splitted.
2934  * @split_flags: indicates if the extent could be zeroout if split fails, and
2935  *               the states(init or uninit) of new extents.
2936  * @flags: flags used to insert new extent to extent tree.
2937  *
2938  *
2939  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2940  * of which are deterimined by split_flag.
2941  *
2942  * There are two cases:
2943  *  a> the extent are splitted into two extent.
2944  *  b> split is not needed, and just mark the extent.
2945  *
2946  * return 0 on success.
2947  */
2948 static int ext4_split_extent_at(handle_t *handle,
2949                              struct inode *inode,
2950                              struct ext4_ext_path *path,
2951                              ext4_lblk_t split,
2952                              int split_flag,
2953                              int flags)
2954 {
2955         ext4_fsblk_t newblock;
2956         ext4_lblk_t ee_block;
2957         struct ext4_extent *ex, newex, orig_ex;
2958         struct ext4_extent *ex2 = NULL;
2959         unsigned int ee_len, depth;
2960         int err = 0;
2961
2962         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
2963                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
2964
2965         ext_debug("ext4_split_extents_at: inode %lu, logical"
2966                 "block %llu\n", inode->i_ino, (unsigned long long)split);
2967
2968         ext4_ext_show_leaf(inode, path);
2969
2970         depth = ext_depth(inode);
2971         ex = path[depth].p_ext;
2972         ee_block = le32_to_cpu(ex->ee_block);
2973         ee_len = ext4_ext_get_actual_len(ex);
2974         newblock = split - ee_block + ext4_ext_pblock(ex);
2975
2976         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
2977
2978         err = ext4_ext_get_access(handle, inode, path + depth);
2979         if (err)
2980                 goto out;
2981
2982         if (split == ee_block) {
2983                 /*
2984                  * case b: block @split is the block that the extent begins with
2985                  * then we just change the state of the extent, and splitting
2986                  * is not needed.
2987                  */
2988                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
2989                         ext4_ext_mark_uninitialized(ex);
2990                 else
2991                         ext4_ext_mark_initialized(ex);
2992
2993                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
2994                         ext4_ext_try_to_merge(handle, inode, path, ex);
2995
2996                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2997                 goto out;
2998         }
2999
3000         /* case a */
3001         memcpy(&orig_ex, ex, sizeof(orig_ex));
3002         ex->ee_len = cpu_to_le16(split - ee_block);
3003         if (split_flag & EXT4_EXT_MARK_UNINIT1)
3004                 ext4_ext_mark_uninitialized(ex);
3005
3006         /*
3007          * path may lead to new leaf, not to original leaf any more
3008          * after ext4_ext_insert_extent() returns,
3009          */
3010         err = ext4_ext_dirty(handle, inode, path + depth);
3011         if (err)
3012                 goto fix_extent_len;
3013
3014         ex2 = &newex;
3015         ex2->ee_block = cpu_to_le32(split);
3016         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3017         ext4_ext_store_pblock(ex2, newblock);
3018         if (split_flag & EXT4_EXT_MARK_UNINIT2)
3019                 ext4_ext_mark_uninitialized(ex2);
3020
3021         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
3022         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3023                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3024                         if (split_flag & EXT4_EXT_DATA_VALID1)
3025                                 err = ext4_ext_zeroout(inode, ex2);
3026                         else
3027                                 err = ext4_ext_zeroout(inode, ex);
3028                 } else
3029                         err = ext4_ext_zeroout(inode, &orig_ex);
3030
3031                 if (err)
3032                         goto fix_extent_len;
3033                 /* update the extent length and mark as initialized */
3034                 ex->ee_len = cpu_to_le16(ee_len);
3035                 ext4_ext_try_to_merge(handle, inode, path, ex);
3036                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3037                 goto out;
3038         } else if (err)
3039                 goto fix_extent_len;
3040
3041 out:
3042         ext4_ext_show_leaf(inode, path);
3043         return err;
3044
3045 fix_extent_len:
3046         ex->ee_len = orig_ex.ee_len;
3047         ext4_ext_dirty(handle, inode, path + depth);
3048         return err;
3049 }
3050
3051 /*
3052  * ext4_split_extents() splits an extent and mark extent which is covered
3053  * by @map as split_flags indicates
3054  *
3055  * It may result in splitting the extent into multiple extents (upto three)
3056  * There are three possibilities:
3057  *   a> There is no split required
3058  *   b> Splits in two extents: Split is happening at either end of the extent
3059  *   c> Splits in three extents: Somone is splitting in middle of the extent
3060  *
3061  */
3062 static int ext4_split_extent(handle_t *handle,
3063                               struct inode *inode,
3064                               struct ext4_ext_path *path,
3065                               struct ext4_map_blocks *map,
3066                               int split_flag,
3067                               int flags)
3068 {
3069         ext4_lblk_t ee_block;
3070         struct ext4_extent *ex;
3071         unsigned int ee_len, depth;
3072         int err = 0;
3073         int uninitialized;
3074         int split_flag1, flags1;
3075
3076         depth = ext_depth(inode);
3077         ex = path[depth].p_ext;
3078         ee_block = le32_to_cpu(ex->ee_block);
3079         ee_len = ext4_ext_get_actual_len(ex);
3080         uninitialized = ext4_ext_is_uninitialized(ex);
3081
3082         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3083                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3084                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3085                 if (uninitialized)
3086                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
3087                                        EXT4_EXT_MARK_UNINIT2;
3088                 if (split_flag & EXT4_EXT_DATA_VALID2)
3089                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3090                 err = ext4_split_extent_at(handle, inode, path,
3091                                 map->m_lblk + map->m_len, split_flag1, flags1);
3092                 if (err)
3093                         goto out;
3094         }
3095
3096         ext4_ext_drop_refs(path);
3097         path = ext4_ext_find_extent(inode, map->m_lblk, path);
3098         if (IS_ERR(path))
3099                 return PTR_ERR(path);
3100
3101         if (map->m_lblk >= ee_block) {
3102                 split_flag1 = split_flag & (EXT4_EXT_MAY_ZEROOUT |
3103                                             EXT4_EXT_DATA_VALID2);
3104                 if (uninitialized)
3105                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
3106                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
3107                         split_flag1 |= EXT4_EXT_MARK_UNINIT2;
3108                 err = ext4_split_extent_at(handle, inode, path,
3109                                 map->m_lblk, split_flag1, flags);
3110                 if (err)
3111                         goto out;
3112         }
3113
3114         ext4_ext_show_leaf(inode, path);
3115 out:
3116         return err ? err : map->m_len;
3117 }
3118
3119 /*
3120  * This function is called by ext4_ext_map_blocks() if someone tries to write
3121  * to an uninitialized extent. It may result in splitting the uninitialized
3122  * extent into multiple extents (up to three - one initialized and two
3123  * uninitialized).
3124  * There are three possibilities:
3125  *   a> There is no split required: Entire extent should be initialized
3126  *   b> Splits in two extents: Write is happening at either end of the extent
3127  *   c> Splits in three extents: Somone is writing in middle of the extent
3128  *
3129  * Pre-conditions:
3130  *  - The extent pointed to by 'path' is uninitialized.
3131  *  - The extent pointed to by 'path' contains a superset
3132  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3133  *
3134  * Post-conditions on success:
3135  *  - the returned value is the number of blocks beyond map->l_lblk
3136  *    that are allocated and initialized.
3137  *    It is guaranteed to be >= map->m_len.
3138  */
3139 static int ext4_ext_convert_to_initialized(handle_t *handle,
3140                                            struct inode *inode,
3141                                            struct ext4_map_blocks *map,
3142                                            struct ext4_ext_path *path)
3143 {
3144         struct ext4_sb_info *sbi;
3145         struct ext4_extent_header *eh;
3146         struct ext4_map_blocks split_map;
3147         struct ext4_extent zero_ex;
3148         struct ext4_extent *ex;
3149         ext4_lblk_t ee_block, eof_block;
3150         unsigned int ee_len, depth;
3151         int allocated, max_zeroout = 0;
3152         int err = 0;
3153         int split_flag = 0;
3154
3155         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3156                 "block %llu, max_blocks %u\n", inode->i_ino,
3157                 (unsigned long long)map->m_lblk, map->m_len);
3158
3159         sbi = EXT4_SB(inode->i_sb);
3160         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3161                 inode->i_sb->s_blocksize_bits;
3162         if (eof_block < map->m_lblk + map->m_len)
3163                 eof_block = map->m_lblk + map->m_len;
3164
3165         depth = ext_depth(inode);
3166         eh = path[depth].p_hdr;
3167         ex = path[depth].p_ext;
3168         ee_block = le32_to_cpu(ex->ee_block);
3169         ee_len = ext4_ext_get_actual_len(ex);
3170         allocated = ee_len - (map->m_lblk - ee_block);
3171
3172         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3173
3174         /* Pre-conditions */
3175         BUG_ON(!ext4_ext_is_uninitialized(ex));
3176         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3177
3178         /*
3179          * Attempt to transfer newly initialized blocks from the currently
3180          * uninitialized extent to its left neighbor. This is much cheaper
3181          * than an insertion followed by a merge as those involve costly
3182          * memmove() calls. This is the common case in steady state for
3183          * workloads doing fallocate(FALLOC_FL_KEEP_SIZE) followed by append
3184          * writes.
3185          *
3186          * Limitations of the current logic:
3187          *  - L1: we only deal with writes at the start of the extent.
3188          *    The approach could be extended to writes at the end
3189          *    of the extent but this scenario was deemed less common.
3190          *  - L2: we do not deal with writes covering the whole extent.
3191          *    This would require removing the extent if the transfer
3192          *    is possible.
3193          *  - L3: we only attempt to merge with an extent stored in the
3194          *    same extent tree node.
3195          */
3196         if ((map->m_lblk == ee_block) &&        /*L1*/
3197                 (map->m_len < ee_len) &&        /*L2*/
3198                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L3*/
3199                 struct ext4_extent *prev_ex;
3200                 ext4_lblk_t prev_lblk;
3201                 ext4_fsblk_t prev_pblk, ee_pblk;
3202                 unsigned int prev_len, write_len;
3203
3204                 prev_ex = ex - 1;
3205                 prev_lblk = le32_to_cpu(prev_ex->ee_block);
3206                 prev_len = ext4_ext_get_actual_len(prev_ex);
3207                 prev_pblk = ext4_ext_pblock(prev_ex);
3208                 ee_pblk = ext4_ext_pblock(ex);
3209                 write_len = map->m_len;
3210
3211                 /*
3212                  * A transfer of blocks from 'ex' to 'prev_ex' is allowed
3213                  * upon those conditions:
3214                  * - C1: prev_ex is initialized,
3215                  * - C2: prev_ex is logically abutting ex,
3216                  * - C3: prev_ex is physically abutting ex,
3217                  * - C4: prev_ex can receive the additional blocks without
3218                  *   overflowing the (initialized) length limit.
3219                  */
3220                 if ((!ext4_ext_is_uninitialized(prev_ex)) &&            /*C1*/
3221                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3222                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3223                         (prev_len < (EXT_INIT_MAX_LEN - write_len))) {  /*C4*/
3224                         err = ext4_ext_get_access(handle, inode, path + depth);
3225                         if (err)
3226                                 goto out;
3227
3228                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3229                                 map, ex, prev_ex);
3230
3231                         /* Shift the start of ex by 'write_len' blocks */
3232                         ex->ee_block = cpu_to_le32(ee_block + write_len);
3233                         ext4_ext_store_pblock(ex, ee_pblk + write_len);
3234                         ex->ee_len = cpu_to_le16(ee_len - write_len);
3235                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3236
3237                         /* Extend prev_ex by 'write_len' blocks */
3238                         prev_ex->ee_len = cpu_to_le16(prev_len + write_len);
3239
3240                         /* Mark the block containing both extents as dirty */
3241                         ext4_ext_dirty(handle, inode, path + depth);
3242
3243                         /* Update path to point to the right extent */
3244                         path[depth].p_ext = prev_ex;
3245
3246                         /* Result: number of initialized blocks past m_lblk */
3247                         allocated = write_len;
3248                         goto out;
3249                 }
3250         }
3251
3252         WARN_ON(map->m_lblk < ee_block);
3253         /*
3254          * It is safe to convert extent to initialized via explicit
3255          * zeroout only if extent is fully insde i_size or new_size.
3256          */
3257         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3258
3259         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3260                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3261                         inode->i_sb->s_blocksize_bits;
3262
3263         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3264         if (max_zeroout && (ee_len <= max_zeroout)) {
3265                 err = ext4_ext_zeroout(inode, ex);
3266                 if (err)
3267                         goto out;
3268
3269                 err = ext4_ext_get_access(handle, inode, path + depth);
3270                 if (err)
3271                         goto out;
3272                 ext4_ext_mark_initialized(ex);
3273                 ext4_ext_try_to_merge(handle, inode, path, ex);
3274                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3275                 goto out;
3276         }
3277
3278         /*
3279          * four cases:
3280          * 1. split the extent into three extents.
3281          * 2. split the extent into two extents, zeroout the first half.
3282          * 3. split the extent into two extents, zeroout the second half.
3283          * 4. split the extent into two extents with out zeroout.
3284          */
3285         split_map.m_lblk = map->m_lblk;
3286         split_map.m_len = map->m_len;
3287
3288         if (max_zeroout && (allocated > map->m_len)) {
3289                 if (allocated <= max_zeroout) {
3290                         /* case 3 */
3291                         zero_ex.ee_block =
3292                                          cpu_to_le32(map->m_lblk);
3293                         zero_ex.ee_len = cpu_to_le16(allocated);
3294                         ext4_ext_store_pblock(&zero_ex,
3295                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3296                         err = ext4_ext_zeroout(inode, &zero_ex);
3297                         if (err)
3298                                 goto out;
3299                         split_map.m_lblk = map->m_lblk;
3300                         split_map.m_len = allocated;
3301                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3302                         /* case 2 */
3303                         if (map->m_lblk != ee_block) {
3304                                 zero_ex.ee_block = ex->ee_block;
3305                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3306                                                         ee_block);
3307                                 ext4_ext_store_pblock(&zero_ex,
3308                                                       ext4_ext_pblock(ex));
3309                                 err = ext4_ext_zeroout(inode, &zero_ex);
3310                                 if (err)
3311                                         goto out;
3312                         }
3313
3314                         split_map.m_lblk = ee_block;
3315                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3316                         allocated = map->m_len;
3317                 }
3318         }
3319
3320         allocated = ext4_split_extent(handle, inode, path,
3321                                       &split_map, split_flag, 0);
3322         if (allocated < 0)
3323                 err = allocated;
3324
3325 out:
3326         return err ? err : allocated;
3327 }
3328
3329 /*
3330  * This function is called by ext4_ext_map_blocks() from
3331  * ext4_get_blocks_dio_write() when DIO to write
3332  * to an uninitialized extent.
3333  *
3334  * Writing to an uninitialized extent may result in splitting the uninitialized
3335  * extent into multiple initialized/uninitialized extents (up to three)
3336  * There are three possibilities:
3337  *   a> There is no split required: Entire extent should be uninitialized
3338  *   b> Splits in two extents: Write is happening at either end of the extent
3339  *   c> Splits in three extents: Somone is writing in middle of the extent
3340  *
3341  * One of more index blocks maybe needed if the extent tree grow after
3342  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3343  * complete, we need to split the uninitialized extent before DIO submit
3344  * the IO. The uninitialized extent called at this time will be split
3345  * into three uninitialized extent(at most). After IO complete, the part
3346  * being filled will be convert to initialized by the end_io callback function
3347  * via ext4_convert_unwritten_extents().
3348  *
3349  * Returns the size of uninitialized extent to be written on success.
3350  */
3351 static int ext4_split_unwritten_extents(handle_t *handle,
3352                                         struct inode *inode,
3353                                         struct ext4_map_blocks *map,
3354                                         struct ext4_ext_path *path,
3355                                         int flags)
3356 {
3357         ext4_lblk_t eof_block;
3358         ext4_lblk_t ee_block;
3359         struct ext4_extent *ex;
3360         unsigned int ee_len;
3361         int split_flag = 0, depth;
3362
3363         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3364                 "block %llu, max_blocks %u\n", inode->i_ino,
3365                 (unsigned long long)map->m_lblk, map->m_len);
3366
3367         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3368                 inode->i_sb->s_blocksize_bits;
3369         if (eof_block < map->m_lblk + map->m_len)
3370                 eof_block = map->m_lblk + map->m_len;
3371         /*
3372          * It is safe to convert extent to initialized via explicit
3373          * zeroout only if extent is fully insde i_size or new_size.
3374          */
3375         depth = ext_depth(inode);
3376         ex = path[depth].p_ext;
3377         ee_block = le32_to_cpu(ex->ee_block);
3378         ee_len = ext4_ext_get_actual_len(ex);
3379
3380         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3381         split_flag |= EXT4_EXT_MARK_UNINIT2;
3382         if (flags & EXT4_GET_BLOCKS_CONVERT)
3383                 split_flag |= EXT4_EXT_DATA_VALID2;
3384         flags |= EXT4_GET_BLOCKS_PRE_IO;
3385         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3386 }
3387
3388 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3389                                                 struct inode *inode,
3390                                                 struct ext4_map_blocks *map,
3391                                                 struct ext4_ext_path *path)
3392 {
3393         struct ext4_extent *ex;
3394         ext4_lblk_t ee_block;
3395         unsigned int ee_len;
3396         int depth;
3397         int err = 0;
3398
3399         depth = ext_depth(inode);
3400         ex = path[depth].p_ext;
3401         ee_block = le32_to_cpu(ex->ee_block);
3402         ee_len = ext4_ext_get_actual_len(ex);
3403
3404         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3405                 "block %llu, max_blocks %u\n", inode->i_ino,
3406                   (unsigned long long)ee_block, ee_len);
3407
3408         /* If extent is larger than requested then split is required */
3409         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3410                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3411                                                    EXT4_GET_BLOCKS_CONVERT);
3412                 if (err < 0)
3413                         goto out;
3414                 ext4_ext_drop_refs(path);
3415                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3416                 if (IS_ERR(path)) {
3417                         err = PTR_ERR(path);
3418                         goto out;
3419                 }
3420                 depth = ext_depth(inode);
3421                 ex = path[depth].p_ext;
3422         }
3423
3424         err = ext4_ext_get_access(handle, inode, path + depth);
3425         if (err)
3426                 goto out;
3427         /* first mark the extent as initialized */
3428         ext4_ext_mark_initialized(ex);
3429
3430         /* note: ext4_ext_correct_indexes() isn't needed here because
3431          * borders are not changed
3432          */
3433         ext4_ext_try_to_merge(handle, inode, path, ex);
3434
3435         /* Mark modified extent as dirty */
3436         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3437 out:
3438         ext4_ext_show_leaf(inode, path);
3439         return err;
3440 }
3441
3442 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3443                         sector_t block, int count)
3444 {
3445         int i;
3446         for (i = 0; i < count; i++)
3447                 unmap_underlying_metadata(bdev, block + i);
3448 }
3449
3450 /*
3451  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3452  */
3453 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3454                               ext4_lblk_t lblk,
3455                               struct ext4_ext_path *path,
3456                               unsigned int len)
3457 {
3458         int i, depth;
3459         struct ext4_extent_header *eh;
3460         struct ext4_extent *last_ex;
3461
3462         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3463                 return 0;
3464
3465         depth = ext_depth(inode);
3466         eh = path[depth].p_hdr;
3467
3468         /*
3469          * We're going to remove EOFBLOCKS_FL entirely in future so we
3470          * do not care for this case anymore. Simply remove the flag
3471          * if there are no extents.
3472          */
3473         if (unlikely(!eh->eh_entries))
3474                 goto out;
3475         last_ex = EXT_LAST_EXTENT(eh);
3476         /*
3477          * We should clear the EOFBLOCKS_FL flag if we are writing the
3478          * last block in the last extent in the file.  We test this by
3479          * first checking to see if the caller to
3480          * ext4_ext_get_blocks() was interested in the last block (or
3481          * a block beyond the last block) in the current extent.  If
3482          * this turns out to be false, we can bail out from this
3483          * function immediately.
3484          */
3485         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3486             ext4_ext_get_actual_len(last_ex))
3487                 return 0;
3488         /*
3489          * If the caller does appear to be planning to write at or
3490          * beyond the end of the current extent, we then test to see
3491          * if the current extent is the last extent in the file, by
3492          * checking to make sure it was reached via the rightmost node
3493          * at each level of the tree.
3494          */
3495         for (i = depth-1; i >= 0; i--)
3496                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3497                         return 0;
3498 out:
3499         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3500         return ext4_mark_inode_dirty(handle, inode);
3501 }
3502
3503 /**
3504  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3505  *
3506  * Return 1 if there is a delalloc block in the range, otherwise 0.
3507  */
3508 static int ext4_find_delalloc_range(struct inode *inode,
3509                                     ext4_lblk_t lblk_start,
3510                                     ext4_lblk_t lblk_end)
3511 {
3512         struct extent_status es;
3513
3514         es.start = lblk_start;
3515         ext4_es_find_extent(inode, &es);
3516         if (es.len == 0)
3517                 return 0; /* there is no delay extent in this tree */
3518         else if (es.start <= lblk_start && lblk_start < es.start + es.len)
3519                 return 1;
3520         else if (lblk_start <= es.start && es.start <= lblk_end)
3521                 return 1;
3522         else
3523                 return 0;
3524 }
3525
3526 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3527 {
3528         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3529         ext4_lblk_t lblk_start, lblk_end;
3530         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3531         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3532
3533         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3534 }
3535
3536 /**
3537  * Determines how many complete clusters (out of those specified by the 'map')
3538  * are under delalloc and were reserved quota for.
3539  * This function is called when we are writing out the blocks that were
3540  * originally written with their allocation delayed, but then the space was
3541  * allocated using fallocate() before the delayed allocation could be resolved.
3542  * The cases to look for are:
3543  * ('=' indicated delayed allocated blocks
3544  *  '-' indicates non-delayed allocated blocks)
3545  * (a) partial clusters towards beginning and/or end outside of allocated range
3546  *     are not delalloc'ed.
3547  *      Ex:
3548  *      |----c---=|====c====|====c====|===-c----|
3549  *               |++++++ allocated ++++++|
3550  *      ==> 4 complete clusters in above example
3551  *
3552  * (b) partial cluster (outside of allocated range) towards either end is
3553  *     marked for delayed allocation. In this case, we will exclude that
3554  *     cluster.
3555  *      Ex:
3556  *      |----====c========|========c========|
3557  *           |++++++ allocated ++++++|
3558  *      ==> 1 complete clusters in above example
3559  *
3560  *      Ex:
3561  *      |================c================|
3562  *            |++++++ allocated ++++++|
3563  *      ==> 0 complete clusters in above example
3564  *
3565  * The ext4_da_update_reserve_space will be called only if we
3566  * determine here that there were some "entire" clusters that span
3567  * this 'allocated' range.
3568  * In the non-bigalloc case, this function will just end up returning num_blks
3569  * without ever calling ext4_find_delalloc_range.
3570  */
3571 static unsigned int
3572 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3573                            unsigned int num_blks)
3574 {
3575         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3576         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3577         ext4_lblk_t lblk_from, lblk_to, c_offset;
3578         unsigned int allocated_clusters = 0;
3579
3580         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3581         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3582
3583         /* max possible clusters for this allocation */
3584         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3585
3586         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3587
3588         /* Check towards left side */
3589         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3590         if (c_offset) {
3591                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3592                 lblk_to = lblk_from + c_offset - 1;
3593
3594                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3595                         allocated_clusters--;
3596         }
3597
3598         /* Now check towards right. */
3599         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3600         if (allocated_clusters && c_offset) {
3601                 lblk_from = lblk_start + num_blks;
3602                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3603
3604                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3605                         allocated_clusters--;
3606         }
3607
3608         return allocated_clusters;
3609 }
3610
3611 static int
3612 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3613                         struct ext4_map_blocks *map,
3614                         struct ext4_ext_path *path, int flags,
3615                         unsigned int allocated, ext4_fsblk_t newblock)
3616 {
3617         int ret = 0;
3618         int err = 0;
3619         ext4_io_end_t *io = ext4_inode_aio(inode);
3620
3621         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3622                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3623                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3624                   flags, allocated);
3625         ext4_ext_show_leaf(inode, path);
3626
3627         trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
3628                                                     allocated, newblock);
3629
3630         /* get_block() before submit the IO, split the extent */
3631         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3632                 ret = ext4_split_unwritten_extents(handle, inode, map,
3633                                                    path, flags);
3634                 if (ret <= 0)
3635                         goto out;
3636                 /*
3637                  * Flag the inode(non aio case) or end_io struct (aio case)
3638                  * that this IO needs to conversion to written when IO is
3639                  * completed
3640                  */
3641                 if (io)
3642                         ext4_set_io_unwritten_flag(inode, io);
3643                 else
3644                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3645                 if (ext4_should_dioread_nolock(inode))
3646                         map->m_flags |= EXT4_MAP_UNINIT;
3647                 goto out;
3648         }
3649         /* IO end_io complete, convert the filled extent to written */
3650         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3651                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3652                                                         path);
3653                 if (ret >= 0) {
3654                         ext4_update_inode_fsync_trans(handle, inode, 1);
3655                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3656                                                  path, map->m_len);
3657                 } else
3658                         err = ret;
3659                 goto out2;
3660         }
3661         /* buffered IO case */
3662         /*
3663          * repeat fallocate creation request
3664          * we already have an unwritten extent
3665          */
3666         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT)
3667                 goto map_out;
3668
3669         /* buffered READ or buffered write_begin() lookup */
3670         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3671                 /*
3672                  * We have blocks reserved already.  We
3673                  * return allocated blocks so that delalloc
3674                  * won't do block reservation for us.  But
3675                  * the buffer head will be unmapped so that
3676                  * a read from the block returns 0s.
3677                  */
3678                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3679                 goto out1;
3680         }
3681
3682         /* buffered write, writepage time, convert*/
3683         ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
3684         if (ret >= 0)
3685                 ext4_update_inode_fsync_trans(handle, inode, 1);
3686 out:
3687         if (ret <= 0) {
3688                 err = ret;
3689                 goto out2;
3690         } else
3691                 allocated = ret;
3692         map->m_flags |= EXT4_MAP_NEW;
3693         /*
3694          * if we allocated more blocks than requested
3695          * we need to make sure we unmap the extra block
3696          * allocated. The actual needed block will get
3697          * unmapped later when we find the buffer_head marked
3698          * new.
3699          */
3700         if (allocated > map->m_len) {
3701                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3702                                         newblock + map->m_len,
3703                                         allocated - map->m_len);
3704                 allocated = map->m_len;
3705         }
3706
3707         /*
3708          * If we have done fallocate with the offset that is already
3709          * delayed allocated, we would have block reservation
3710          * and quota reservation done in the delayed write path.
3711          * But fallocate would have already updated quota and block
3712          * count for this offset. So cancel these reservation
3713          */
3714         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3715                 unsigned int reserved_clusters;
3716                 reserved_clusters = get_reserved_cluster_alloc(inode,
3717                                 map->m_lblk, map->m_len);
3718                 if (reserved_clusters)
3719                         ext4_da_update_reserve_space(inode,
3720                                                      reserved_clusters,
3721                                                      0);
3722         }
3723
3724 map_out:
3725         map->m_flags |= EXT4_MAP_MAPPED;
3726         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3727                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3728                                          map->m_len);
3729                 if (err < 0)
3730                         goto out2;
3731         }
3732 out1:
3733         if (allocated > map->m_len)
3734                 allocated = map->m_len;
3735         ext4_ext_show_leaf(inode, path);
3736         map->m_pblk = newblock;
3737         map->m_len = allocated;
3738 out2:
3739         if (path) {
3740                 ext4_ext_drop_refs(path);
3741                 kfree(path);
3742         }
3743         return err ? err : allocated;
3744 }
3745
3746 /*
3747  * get_implied_cluster_alloc - check to see if the requested
3748  * allocation (in the map structure) overlaps with a cluster already
3749  * allocated in an extent.
3750  *      @sb     The filesystem superblock structure
3751  *      @map    The requested lblk->pblk mapping
3752  *      @ex     The extent structure which might contain an implied
3753  *                      cluster allocation
3754  *
3755  * This function is called by ext4_ext_map_blocks() after we failed to
3756  * find blocks that were already in the inode's extent tree.  Hence,
3757  * we know that the beginning of the requested region cannot overlap
3758  * the extent from the inode's extent tree.  There are three cases we
3759  * want to catch.  The first is this case:
3760  *
3761  *               |--- cluster # N--|
3762  *    |--- extent ---|  |---- requested region ---|
3763  *                      |==========|
3764  *
3765  * The second case that we need to test for is this one:
3766  *
3767  *   |--------- cluster # N ----------------|
3768  *         |--- requested region --|   |------- extent ----|
3769  *         |=======================|
3770  *
3771  * The third case is when the requested region lies between two extents
3772  * within the same cluster:
3773  *          |------------- cluster # N-------------|
3774  * |----- ex -----|                  |---- ex_right ----|
3775  *                  |------ requested region ------|
3776  *                  |================|
3777  *
3778  * In each of the above cases, we need to set the map->m_pblk and
3779  * map->m_len so it corresponds to the return the extent labelled as
3780  * "|====|" from cluster #N, since it is already in use for data in
3781  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3782  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3783  * as a new "allocated" block region.  Otherwise, we will return 0 and
3784  * ext4_ext_map_blocks() will then allocate one or more new clusters
3785  * by calling ext4_mb_new_blocks().
3786  */
3787 static int get_implied_cluster_alloc(struct super_block *sb,
3788                                      struct ext4_map_blocks *map,
3789                                      struct ext4_extent *ex,
3790                                      struct ext4_ext_path *path)
3791 {
3792         struct ext4_sb_info *sbi = EXT4_SB(sb);
3793         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3794         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3795         ext4_lblk_t rr_cluster_start;
3796         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3797         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3798         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3799
3800         /* The extent passed in that we are trying to match */
3801         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3802         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3803
3804         /* The requested region passed into ext4_map_blocks() */
3805         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3806
3807         if ((rr_cluster_start == ex_cluster_end) ||
3808             (rr_cluster_start == ex_cluster_start)) {
3809                 if (rr_cluster_start == ex_cluster_end)
3810                         ee_start += ee_len - 1;
3811                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3812                         c_offset;
3813                 map->m_len = min(map->m_len,
3814                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3815                 /*
3816                  * Check for and handle this case:
3817                  *
3818                  *   |--------- cluster # N-------------|
3819                  *                     |------- extent ----|
3820                  *         |--- requested region ---|
3821                  *         |===========|
3822                  */
3823
3824                 if (map->m_lblk < ee_block)
3825                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3826
3827                 /*
3828                  * Check for the case where there is already another allocated
3829                  * block to the right of 'ex' but before the end of the cluster.
3830                  *
3831                  *          |------------- cluster # N-------------|
3832                  * |----- ex -----|                  |---- ex_right ----|
3833                  *                  |------ requested region ------|
3834                  *                  |================|
3835                  */
3836                 if (map->m_lblk > ee_block) {
3837                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
3838                         map->m_len = min(map->m_len, next - map->m_lblk);
3839                 }
3840
3841                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
3842                 return 1;
3843         }
3844
3845         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
3846         return 0;
3847 }
3848
3849
3850 /*
3851  * Block allocation/map/preallocation routine for extents based files
3852  *
3853  *
3854  * Need to be called with
3855  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
3856  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
3857  *
3858  * return > 0, number of of blocks already mapped/allocated
3859  *          if create == 0 and these are pre-allocated blocks
3860  *              buffer head is unmapped
3861  *          otherwise blocks are mapped
3862  *
3863  * return = 0, if plain look up failed (blocks have not been allocated)
3864  *          buffer head is unmapped
3865  *
3866  * return < 0, error case.
3867  */
3868 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
3869                         struct ext4_map_blocks *map, int flags)
3870 {
3871         struct ext4_ext_path *path = NULL;
3872         struct ext4_extent newex, *ex, *ex2;
3873         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3874         ext4_fsblk_t newblock = 0;
3875         int free_on_err = 0, err = 0, depth;
3876         unsigned int allocated = 0, offset = 0;
3877         unsigned int allocated_clusters = 0;
3878         struct ext4_allocation_request ar;
3879         ext4_io_end_t *io = ext4_inode_aio(inode);
3880         ext4_lblk_t cluster_offset;
3881         int set_unwritten = 0;
3882
3883         ext_debug("blocks %u/%u requested for inode %lu\n",
3884                   map->m_lblk, map->m_len, inode->i_ino);
3885         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
3886
3887         /* check in cache */
3888         if (ext4_ext_in_cache(inode, map->m_lblk, &newex)) {
3889                 if (!newex.ee_start_lo && !newex.ee_start_hi) {
3890                         if ((sbi->s_cluster_ratio > 1) &&
3891                             ext4_find_delalloc_cluster(inode, map->m_lblk))
3892                                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3893
3894                         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3895                                 /*
3896                                  * block isn't allocated yet and
3897                                  * user doesn't want to allocate it
3898                                  */
3899                                 goto out2;
3900                         }
3901                         /* we should allocate requested block */
3902                 } else {
3903                         /* block is already allocated */
3904                         if (sbi->s_cluster_ratio > 1)
3905                                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3906                         newblock = map->m_lblk
3907                                    - le32_to_cpu(newex.ee_block)
3908                                    + ext4_ext_pblock(&newex);
3909                         /* number of remaining blocks in the extent */
3910                         allocated = ext4_ext_get_actual_len(&newex) -
3911                                 (map->m_lblk - le32_to_cpu(newex.ee_block));
3912                         goto out;
3913                 }
3914         }
3915
3916         /* find extent for this block */
3917         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
3918         if (IS_ERR(path)) {
3919                 err = PTR_ERR(path);
3920                 path = NULL;
3921                 goto out2;
3922         }
3923
3924         depth = ext_depth(inode);
3925
3926         /*
3927          * consistent leaf must not be empty;
3928          * this situation is possible, though, _during_ tree modification;
3929          * this is why assert can't be put in ext4_ext_find_extent()
3930          */
3931         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
3932                 EXT4_ERROR_INODE(inode, "bad extent address "
3933                                  "lblock: %lu, depth: %d pblock %lld",
3934                                  (unsigned long) map->m_lblk, depth,
3935                                  path[depth].p_block);
3936                 err = -EIO;
3937                 goto out2;
3938         }
3939
3940         ex = path[depth].p_ext;
3941         if (ex) {
3942                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3943                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3944                 unsigned short ee_len;
3945
3946                 /*
3947                  * Uninitialized extents are treated as holes, except that
3948                  * we split out initialized portions during a write.
3949                  */
3950                 ee_len = ext4_ext_get_actual_len(ex);
3951
3952                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
3953
3954                 /* if found extent covers block, simply return it */
3955                 if (in_range(map->m_lblk, ee_block, ee_len)) {
3956                         newblock = map->m_lblk - ee_block + ee_start;
3957                         /* number of remaining blocks in the extent */
3958                         allocated = ee_len - (map->m_lblk - ee_block);
3959                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
3960                                   ee_block, ee_len, newblock);
3961
3962                         /*
3963                          * Do not put uninitialized extent
3964                          * in the cache
3965                          */
3966                         if (!ext4_ext_is_uninitialized(ex)) {
3967                                 ext4_ext_put_in_cache(inode, ee_block,
3968                                         ee_len, ee_start);
3969                                 goto out;
3970                         }
3971                         allocated = ext4_ext_handle_uninitialized_extents(
3972                                 handle, inode, map, path, flags,
3973                                 allocated, newblock);
3974                         goto out3;
3975                 }
3976         }
3977
3978         if ((sbi->s_cluster_ratio > 1) &&
3979             ext4_find_delalloc_cluster(inode, map->m_lblk))
3980                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3981
3982         /*
3983          * requested block isn't allocated yet;
3984          * we couldn't try to create block if create flag is zero
3985          */
3986         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3987                 /*
3988                  * put just found gap into cache to speed up
3989                  * subsequent requests
3990                  */
3991                 ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
3992                 goto out2;
3993         }
3994
3995         /*
3996          * Okay, we need to do block allocation.
3997          */
3998         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
3999         newex.ee_block = cpu_to_le32(map->m_lblk);
4000         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
4001
4002         /*
4003          * If we are doing bigalloc, check to see if the extent returned
4004          * by ext4_ext_find_extent() implies a cluster we can use.
4005          */
4006         if (cluster_offset && ex &&
4007             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4008                 ar.len = allocated = map->m_len;
4009                 newblock = map->m_pblk;
4010                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4011                 goto got_allocated_blocks;
4012         }
4013
4014         /* find neighbour allocated blocks */
4015         ar.lleft = map->m_lblk;
4016         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4017         if (err)
4018                 goto out2;
4019         ar.lright = map->m_lblk;
4020         ex2 = NULL;
4021         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4022         if (err)
4023                 goto out2;
4024
4025         /* Check if the extent after searching to the right implies a
4026          * cluster we can use. */
4027         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4028             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4029                 ar.len = allocated = map->m_len;
4030                 newblock = map->m_pblk;
4031                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4032                 goto got_allocated_blocks;
4033         }
4034
4035         /*
4036          * See if request is beyond maximum number of blocks we can have in
4037          * a single extent. For an initialized extent this limit is
4038          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4039          * EXT_UNINIT_MAX_LEN.
4040          */
4041         if (map->m_len > EXT_INIT_MAX_LEN &&
4042             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4043                 map->m_len = EXT_INIT_MAX_LEN;
4044         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4045                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4046                 map->m_len = EXT_UNINIT_MAX_LEN;
4047
4048         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4049         newex.ee_len = cpu_to_le16(map->m_len);
4050         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4051         if (err)
4052                 allocated = ext4_ext_get_actual_len(&newex);
4053         else
4054                 allocated = map->m_len;
4055
4056         /* allocate new block */
4057         ar.inode = inode;
4058         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4059         ar.logical = map->m_lblk;
4060         /*
4061          * We calculate the offset from the beginning of the cluster
4062          * for the logical block number, since when we allocate a
4063          * physical cluster, the physical block should start at the
4064          * same offset from the beginning of the cluster.  This is
4065          * needed so that future calls to get_implied_cluster_alloc()
4066          * work correctly.
4067          */
4068         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
4069         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4070         ar.goal -= offset;
4071         ar.logical -= offset;
4072         if (S_ISREG(inode->i_mode))
4073                 ar.flags = EXT4_MB_HINT_DATA;
4074         else
4075                 /* disable in-core preallocation for non-regular files */
4076                 ar.flags = 0;
4077         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4078                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4079         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4080         if (!newblock)
4081                 goto out2;
4082         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4083                   ar.goal, newblock, allocated);
4084         free_on_err = 1;
4085         allocated_clusters = ar.len;
4086         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4087         if (ar.len > allocated)
4088                 ar.len = allocated;
4089
4090 got_allocated_blocks:
4091         /* try to insert new extent into found leaf and return */
4092         ext4_ext_store_pblock(&newex, newblock + offset);
4093         newex.ee_len = cpu_to_le16(ar.len);
4094         /* Mark uninitialized */
4095         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4096                 ext4_ext_mark_uninitialized(&newex);
4097                 /*
4098                  * io_end structure was created for every IO write to an
4099                  * uninitialized extent. To avoid unnecessary conversion,
4100                  * here we flag the IO that really needs the conversion.
4101                  * For non asycn direct IO case, flag the inode state
4102                  * that we need to perform conversion when IO is done.
4103                  */
4104                 if ((flags & EXT4_GET_BLOCKS_PRE_IO))
4105                         set_unwritten = 1;
4106                 if (ext4_should_dioread_nolock(inode))
4107                         map->m_flags |= EXT4_MAP_UNINIT;
4108         }
4109
4110         err = 0;
4111         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4112                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4113                                          path, ar.len);
4114         if (!err)
4115                 err = ext4_ext_insert_extent(handle, inode, path,
4116                                              &newex, flags);
4117
4118         if (!err && set_unwritten) {
4119                 if (io)
4120                         ext4_set_io_unwritten_flag(inode, io);
4121                 else
4122                         ext4_set_inode_state(inode,
4123                                              EXT4_STATE_DIO_UNWRITTEN);
4124         }
4125
4126         if (err && free_on_err) {
4127                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4128                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4129                 /* free data blocks we just allocated */
4130                 /* not a good idea to call discard here directly,
4131                  * but otherwise we'd need to call it every free() */
4132                 ext4_discard_preallocations(inode);
4133                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4134                                  ext4_ext_get_actual_len(&newex), fb_flags);
4135                 goto out2;
4136         }
4137
4138         /* previous routine could use block we allocated */
4139         newblock = ext4_ext_pblock(&newex);
4140         allocated = ext4_ext_get_actual_len(&newex);
4141         if (allocated > map->m_len)
4142                 allocated = map->m_len;
4143         map->m_flags |= EXT4_MAP_NEW;
4144
4145         /*
4146          * Update reserved blocks/metadata blocks after successful
4147          * block allocation which had been deferred till now.
4148          */
4149         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4150                 unsigned int reserved_clusters;
4151                 /*
4152                  * Check how many clusters we had reserved this allocated range
4153                  */
4154                 reserved_clusters = get_reserved_cluster_alloc(inode,
4155                                                 map->m_lblk, allocated);
4156                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4157                         if (reserved_clusters) {
4158                                 /*
4159                                  * We have clusters reserved for this range.
4160                                  * But since we are not doing actual allocation
4161                                  * and are simply using blocks from previously
4162                                  * allocated cluster, we should release the
4163                                  * reservation and not claim quota.
4164                                  */
4165                                 ext4_da_update_reserve_space(inode,
4166                                                 reserved_clusters, 0);
4167                         }
4168                 } else {
4169                         BUG_ON(allocated_clusters < reserved_clusters);
4170                         /* We will claim quota for all newly allocated blocks.*/
4171                         ext4_da_update_reserve_space(inode, allocated_clusters,
4172                                                         1);
4173                         if (reserved_clusters < allocated_clusters) {
4174                                 struct ext4_inode_info *ei = EXT4_I(inode);
4175                                 int reservation = allocated_clusters -
4176                                                   reserved_clusters;
4177                                 /*
4178                                  * It seems we claimed few clusters outside of
4179                                  * the range of this allocation. We should give
4180                                  * it back to the reservation pool. This can
4181                                  * happen in the following case:
4182                                  *
4183                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4184                                  *   cluster has 4 blocks. Thus, the clusters
4185                                  *   are [0-3],[4-7],[8-11]...
4186                                  * * First comes delayed allocation write for
4187                                  *   logical blocks 10 & 11. Since there were no
4188                                  *   previous delayed allocated blocks in the
4189                                  *   range [8-11], we would reserve 1 cluster
4190                                  *   for this write.
4191                                  * * Next comes write for logical blocks 3 to 8.
4192                                  *   In this case, we will reserve 2 clusters
4193                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4194                                  *   that range has a delayed allocated blocks.
4195                                  *   Thus total reserved clusters now becomes 3.
4196                                  * * Now, during the delayed allocation writeout
4197                                  *   time, we will first write blocks [3-8] and
4198                                  *   allocate 3 clusters for writing these
4199                                  *   blocks. Also, we would claim all these
4200                                  *   three clusters above.
4201                                  * * Now when we come here to writeout the
4202                                  *   blocks [10-11], we would expect to claim
4203                                  *   the reservation of 1 cluster we had made
4204                                  *   (and we would claim it since there are no
4205                                  *   more delayed allocated blocks in the range
4206                                  *   [8-11]. But our reserved cluster count had
4207                                  *   already gone to 0.
4208                                  *
4209                                  *   Thus, at the step 4 above when we determine
4210                                  *   that there are still some unwritten delayed
4211                                  *   allocated blocks outside of our current
4212                                  *   block range, we should increment the
4213                                  *   reserved clusters count so that when the
4214                                  *   remaining blocks finally gets written, we
4215                                  *   could claim them.
4216                                  */
4217                                 dquot_reserve_block(inode,
4218                                                 EXT4_C2B(sbi, reservation));
4219                                 spin_lock(&ei->i_block_reservation_lock);
4220                                 ei->i_reserved_data_blocks += reservation;
4221                                 spin_unlock(&ei->i_block_reservation_lock);
4222                         }
4223                 }
4224         }
4225
4226         /*
4227          * Cache the extent and update transaction to commit on fdatasync only
4228          * when it is _not_ an uninitialized extent.
4229          */
4230         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0) {
4231                 ext4_ext_put_in_cache(inode, map->m_lblk, allocated, newblock);
4232                 ext4_update_inode_fsync_trans(handle, inode, 1);
4233         } else
4234                 ext4_update_inode_fsync_trans(handle, inode, 0);
4235 out:
4236         if (allocated > map->m_len)
4237                 allocated = map->m_len;
4238         ext4_ext_show_leaf(inode, path);
4239         map->m_flags |= EXT4_MAP_MAPPED;
4240         map->m_pblk = newblock;
4241         map->m_len = allocated;
4242 out2:
4243         if (path) {
4244                 ext4_ext_drop_refs(path);
4245                 kfree(path);
4246         }
4247
4248 out3:
4249         trace_ext4_ext_map_blocks_exit(inode, map, err ? err : allocated);
4250
4251         return err ? err : allocated;
4252 }
4253
4254 void ext4_ext_truncate(struct inode *inode)
4255 {
4256         struct address_space *mapping = inode->i_mapping;
4257         struct super_block *sb = inode->i_sb;
4258         ext4_lblk_t last_block;
4259         handle_t *handle;
4260         loff_t page_len;
4261         int err = 0;
4262
4263         /*
4264          * finish any pending end_io work so we won't run the risk of
4265          * converting any truncated blocks to initialized later
4266          */
4267         ext4_flush_unwritten_io(inode);
4268
4269         /*
4270          * probably first extent we're gonna free will be last in block
4271          */
4272         err = ext4_writepage_trans_blocks(inode);
4273         handle = ext4_journal_start(inode, err);
4274         if (IS_ERR(handle))
4275                 return;
4276
4277         if (inode->i_size % PAGE_CACHE_SIZE != 0) {
4278                 page_len = PAGE_CACHE_SIZE -
4279                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4280
4281                 err = ext4_discard_partial_page_buffers(handle,
4282                         mapping, inode->i_size, page_len, 0);
4283
4284                 if (err)
4285                         goto out_stop;
4286         }
4287
4288         if (ext4_orphan_add(handle, inode))
4289                 goto out_stop;
4290
4291         down_write(&EXT4_I(inode)->i_data_sem);
4292         ext4_ext_invalidate_cache(inode);
4293
4294         ext4_discard_preallocations(inode);
4295
4296         /*
4297          * TODO: optimization is possible here.
4298          * Probably we need not scan at all,
4299          * because page truncation is enough.
4300          */
4301
4302         /* we have to know where to truncate from in crash case */
4303         EXT4_I(inode)->i_disksize = inode->i_size;
4304         ext4_mark_inode_dirty(handle, inode);
4305
4306         last_block = (inode->i_size + sb->s_blocksize - 1)
4307                         >> EXT4_BLOCK_SIZE_BITS(sb);
4308         err = ext4_es_remove_extent(inode, last_block,
4309                                     EXT_MAX_BLOCKS - last_block);
4310         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4311
4312         /* In a multi-transaction truncate, we only make the final
4313          * transaction synchronous.
4314          */
4315         if (IS_SYNC(inode))
4316                 ext4_handle_sync(handle);
4317
4318         up_write(&EXT4_I(inode)->i_data_sem);
4319
4320 out_stop:
4321         /*
4322          * If this was a simple ftruncate() and the file will remain alive,
4323          * then we need to clear up the orphan record which we created above.
4324          * However, if this was a real unlink then we were called by
4325          * ext4_delete_inode(), and we allow that function to clean up the
4326          * orphan info for us.
4327          */
4328         if (inode->i_nlink)
4329                 ext4_orphan_del(handle, inode);
4330
4331         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4332         ext4_mark_inode_dirty(handle, inode);
4333         ext4_journal_stop(handle);
4334 }
4335
4336 static void ext4_falloc_update_inode(struct inode *inode,
4337                                 int mode, loff_t new_size, int update_ctime)
4338 {
4339         struct timespec now;
4340
4341         if (update_ctime) {
4342                 now = current_fs_time(inode->i_sb);
4343                 if (!timespec_equal(&inode->i_ctime, &now))
4344                         inode->i_ctime = now;
4345         }
4346         /*
4347          * Update only when preallocation was requested beyond
4348          * the file size.
4349          */
4350         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4351                 if (new_size > i_size_read(inode))
4352                         i_size_write(inode, new_size);
4353                 if (new_size > EXT4_I(inode)->i_disksize)
4354                         ext4_update_i_disksize(inode, new_size);
4355         } else {
4356                 /*
4357                  * Mark that we allocate beyond EOF so the subsequent truncate
4358                  * can proceed even if the new size is the same as i_size.
4359                  */
4360                 if (new_size > i_size_read(inode))
4361                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4362         }
4363
4364 }
4365
4366 /*
4367  * preallocate space for a file. This implements ext4's fallocate file
4368  * operation, which gets called from sys_fallocate system call.
4369  * For block-mapped files, posix_fallocate should fall back to the method
4370  * of writing zeroes to the required new blocks (the same behavior which is
4371  * expected for file systems which do not support fallocate() system call).
4372  */
4373 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4374 {
4375         struct inode *inode = file->f_path.dentry->d_inode;
4376         handle_t *handle;
4377         loff_t new_size;
4378         unsigned int max_blocks;
4379         int ret = 0;
4380         int ret2 = 0;
4381         int retries = 0;
4382         int flags;
4383         struct ext4_map_blocks map;
4384         unsigned int credits, blkbits = inode->i_blkbits;
4385
4386         /*
4387          * currently supporting (pre)allocate mode for extent-based
4388          * files _only_
4389          */
4390         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4391                 return -EOPNOTSUPP;
4392
4393         /* Return error if mode is not supported */
4394         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4395                 return -EOPNOTSUPP;
4396
4397         if (mode & FALLOC_FL_PUNCH_HOLE)
4398                 return ext4_punch_hole(file, offset, len);
4399
4400         ret = ext4_convert_inline_data(inode);
4401         if (ret)
4402                 return ret;
4403
4404         trace_ext4_fallocate_enter(inode, offset, len, mode);
4405         map.m_lblk = offset >> blkbits;
4406         /*
4407          * We can't just convert len to max_blocks because
4408          * If blocksize = 4096 offset = 3072 and len = 2048
4409          */
4410         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4411                 - map.m_lblk;
4412         /*
4413          * credits to insert 1 extent into extent tree
4414          */
4415         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4416         mutex_lock(&inode->i_mutex);
4417         ret = inode_newsize_ok(inode, (len + offset));
4418         if (ret) {
4419                 mutex_unlock(&inode->i_mutex);
4420                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4421                 return ret;
4422         }
4423         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4424         if (mode & FALLOC_FL_KEEP_SIZE)
4425                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4426         /*
4427          * Don't normalize the request if it can fit in one extent so
4428          * that it doesn't get unnecessarily split into multiple
4429          * extents.
4430          */
4431         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4432                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4433
4434         /* Prevent race condition between unwritten */
4435         ext4_flush_unwritten_io(inode);
4436 retry:
4437         while (ret >= 0 && ret < max_blocks) {
4438                 map.m_lblk = map.m_lblk + ret;
4439                 map.m_len = max_blocks = max_blocks - ret;
4440                 handle = ext4_journal_start(inode, credits);
4441                 if (IS_ERR(handle)) {
4442                         ret = PTR_ERR(handle);
4443                         break;
4444                 }
4445                 ret = ext4_map_blocks(handle, inode, &map, flags);
4446                 if (ret <= 0) {
4447 #ifdef EXT4FS_DEBUG
4448                         WARN_ON(ret <= 0);
4449                         printk(KERN_ERR "%s: ext4_ext_map_blocks "
4450                                     "returned error inode#%lu, block=%u, "
4451                                     "max_blocks=%u", __func__,
4452                                     inode->i_ino, map.m_lblk, max_blocks);
4453 #endif
4454                         ext4_mark_inode_dirty(handle, inode);
4455                         ret2 = ext4_journal_stop(handle);
4456                         break;
4457                 }
4458                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4459                                                 blkbits) >> blkbits))
4460                         new_size = offset + len;
4461                 else
4462                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4463
4464                 ext4_falloc_update_inode(inode, mode, new_size,
4465                                          (map.m_flags & EXT4_MAP_NEW));
4466                 ext4_mark_inode_dirty(handle, inode);
4467                 if ((file->f_flags & O_SYNC) && ret >= max_blocks)
4468                         ext4_handle_sync(handle);
4469                 ret2 = ext4_journal_stop(handle);
4470                 if (ret2)
4471                         break;
4472         }
4473         if (ret == -ENOSPC &&
4474                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4475                 ret = 0;
4476                 goto retry;
4477         }
4478         mutex_unlock(&inode->i_mutex);
4479         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4480                                 ret > 0 ? ret2 : ret);
4481         return ret > 0 ? ret2 : ret;
4482 }
4483
4484 /*
4485  * This function convert a range of blocks to written extents
4486  * The caller of this function will pass the start offset and the size.
4487  * all unwritten extents within this range will be converted to
4488  * written extents.
4489  *
4490  * This function is called from the direct IO end io call back
4491  * function, to convert the fallocated extents after IO is completed.
4492  * Returns 0 on success.
4493  */
4494 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4495                                     ssize_t len)
4496 {
4497         handle_t *handle;
4498         unsigned int max_blocks;
4499         int ret = 0;
4500         int ret2 = 0;
4501         struct ext4_map_blocks map;
4502         unsigned int credits, blkbits = inode->i_blkbits;
4503
4504         map.m_lblk = offset >> blkbits;
4505         /*
4506          * We can't just convert len to max_blocks because
4507          * If blocksize = 4096 offset = 3072 and len = 2048
4508          */
4509         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4510                       map.m_lblk);
4511         /*
4512          * credits to insert 1 extent into extent tree
4513          */
4514         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4515         while (ret >= 0 && ret < max_blocks) {
4516                 map.m_lblk += ret;
4517                 map.m_len = (max_blocks -= ret);
4518                 handle = ext4_journal_start(inode, credits);
4519                 if (IS_ERR(handle)) {
4520                         ret = PTR_ERR(handle);
4521                         break;
4522                 }
4523                 ret = ext4_map_blocks(handle, inode, &map,
4524                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4525                 if (ret <= 0) {
4526                         WARN_ON(ret <= 0);
4527                         ext4_msg(inode->i_sb, KERN_ERR,
4528                                  "%s:%d: inode #%lu: block %u: len %u: "
4529                                  "ext4_ext_map_blocks returned %d",
4530                                  __func__, __LINE__, inode->i_ino, map.m_lblk,
4531                                  map.m_len, ret);
4532                 }
4533                 ext4_mark_inode_dirty(handle, inode);
4534                 ret2 = ext4_journal_stop(handle);
4535                 if (ret <= 0 || ret2 )
4536                         break;
4537         }
4538         return ret > 0 ? ret2 : ret;
4539 }
4540
4541 /*
4542  * If newex is not existing extent (newex->ec_start equals zero) find
4543  * delayed extent at start of newex and update newex accordingly and
4544  * return start of the next delayed extent.
4545  *
4546  * If newex is existing extent (newex->ec_start is not equal zero)
4547  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4548  * extent found. Leave newex unmodified.
4549  */
4550 static int ext4_find_delayed_extent(struct inode *inode,
4551                                     struct ext4_ext_cache *newex)
4552 {
4553         struct extent_status es;
4554         ext4_lblk_t next_del;
4555
4556         es.start = newex->ec_block;
4557         next_del = ext4_es_find_extent(inode, &es);
4558
4559         if (newex->ec_start == 0) {
4560                 /*
4561                  * No extent in extent-tree contains block @newex->ec_start,
4562                  * then the block may stay in 1)a hole or 2)delayed-extent.
4563                  */
4564                 if (es.len == 0)
4565                         /* A hole found. */
4566                         return 0;
4567
4568                 if (es.start > newex->ec_block) {
4569                         /* A hole found. */
4570                         newex->ec_len = min(es.start - newex->ec_block,
4571                                             newex->ec_len);
4572                         return 0;
4573                 }
4574
4575                 newex->ec_len = es.start + es.len - newex->ec_block;
4576         }
4577
4578         return next_del;
4579 }
4580 /* fiemap flags we can handle specified here */
4581 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4582
4583 static int ext4_xattr_fiemap(struct inode *inode,
4584                                 struct fiemap_extent_info *fieinfo)
4585 {
4586         __u64 physical = 0;
4587         __u64 length;
4588         __u32 flags = FIEMAP_EXTENT_LAST;
4589         int blockbits = inode->i_sb->s_blocksize_bits;
4590         int error = 0;
4591
4592         /* in-inode? */
4593         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4594                 struct ext4_iloc iloc;
4595                 int offset;     /* offset of xattr in inode */
4596
4597                 error = ext4_get_inode_loc(inode, &iloc);
4598                 if (error)
4599                         return error;
4600                 physical = iloc.bh->b_blocknr << blockbits;
4601                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4602                                 EXT4_I(inode)->i_extra_isize;
4603                 physical += offset;
4604                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4605                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4606                 brelse(iloc.bh);
4607         } else { /* external block */
4608                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4609                 length = inode->i_sb->s_blocksize;
4610         }
4611
4612         if (physical)
4613                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4614                                                 length, flags);
4615         return (error < 0 ? error : 0);
4616 }
4617
4618 /*
4619  * ext4_ext_punch_hole
4620  *
4621  * Punches a hole of "length" bytes in a file starting
4622  * at byte "offset"
4623  *
4624  * @inode:  The inode of the file to punch a hole in
4625  * @offset: The starting byte offset of the hole
4626  * @length: The length of the hole
4627  *
4628  * Returns the number of blocks removed or negative on err
4629  */
4630 int ext4_ext_punch_hole(struct file *file, loff_t offset, loff_t length)
4631 {
4632         struct inode *inode = file->f_path.dentry->d_inode;
4633         struct super_block *sb = inode->i_sb;
4634         ext4_lblk_t first_block, stop_block;
4635         struct address_space *mapping = inode->i_mapping;
4636         handle_t *handle;
4637         loff_t first_page, last_page, page_len;
4638         loff_t first_page_offset, last_page_offset;
4639         int credits, err = 0;
4640
4641         /*
4642          * Write out all dirty pages to avoid race conditions
4643          * Then release them.
4644          */
4645         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4646                 err = filemap_write_and_wait_range(mapping,
4647                         offset, offset + length - 1);
4648
4649                 if (err)
4650                         return err;
4651         }
4652
4653         mutex_lock(&inode->i_mutex);
4654         /* It's not possible punch hole on append only file */
4655         if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
4656                 err = -EPERM;
4657                 goto out_mutex;
4658         }
4659         if (IS_SWAPFILE(inode)) {
4660                 err = -ETXTBSY;
4661                 goto out_mutex;
4662         }
4663
4664         /* No need to punch hole beyond i_size */
4665         if (offset >= inode->i_size)
4666                 goto out_mutex;
4667
4668         /*
4669          * If the hole extends beyond i_size, set the hole
4670          * to end after the page that contains i_size
4671          */
4672         if (offset + length > inode->i_size) {
4673                 length = inode->i_size +
4674                    PAGE_CACHE_SIZE - (inode->i_size & (PAGE_CACHE_SIZE - 1)) -
4675                    offset;
4676         }
4677
4678         first_page = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
4679         last_page = (offset + length) >> PAGE_CACHE_SHIFT;
4680
4681         first_page_offset = first_page << PAGE_CACHE_SHIFT;
4682         last_page_offset = last_page << PAGE_CACHE_SHIFT;
4683
4684         /* Now release the pages */
4685         if (last_page_offset > first_page_offset) {
4686                 truncate_pagecache_range(inode, first_page_offset,
4687                                          last_page_offset - 1);
4688         }
4689
4690         /* Wait all existing dio workers, newcomers will block on i_mutex */
4691         ext4_inode_block_unlocked_dio(inode);
4692         err = ext4_flush_unwritten_io(inode);
4693         if (err)
4694                 goto out_dio;
4695         inode_dio_wait(inode);
4696
4697         credits = ext4_writepage_trans_blocks(inode);
4698         handle = ext4_journal_start(inode, credits);
4699         if (IS_ERR(handle)) {
4700                 err = PTR_ERR(handle);
4701                 goto out_dio;
4702         }
4703
4704
4705         /*
4706          * Now we need to zero out the non-page-aligned data in the
4707          * pages at the start and tail of the hole, and unmap the buffer
4708          * heads for the block aligned regions of the page that were
4709          * completely zeroed.
4710          */
4711         if (first_page > last_page) {
4712                 /*
4713                  * If the file space being truncated is contained within a page
4714                  * just zero out and unmap the middle of that page
4715                  */
4716                 err = ext4_discard_partial_page_buffers(handle,
4717                         mapping, offset, length, 0);
4718
4719                 if (err)
4720                         goto out;
4721         } else {
4722                 /*
4723                  * zero out and unmap the partial page that contains
4724                  * the start of the hole
4725                  */
4726                 page_len  = first_page_offset - offset;
4727                 if (page_len > 0) {
4728                         err = ext4_discard_partial_page_buffers(handle, mapping,
4729                                                    offset, page_len, 0);
4730                         if (err)
4731                                 goto out;
4732                 }
4733
4734                 /*
4735                  * zero out and unmap the partial page that contains
4736                  * the end of the hole
4737                  */
4738                 page_len = offset + length - last_page_offset;
4739                 if (page_len > 0) {
4740                         err = ext4_discard_partial_page_buffers(handle, mapping,
4741                                         last_page_offset, page_len, 0);
4742                         if (err)
4743                                 goto out;
4744                 }
4745         }
4746
4747         /*
4748          * If i_size is contained in the last page, we need to
4749          * unmap and zero the partial page after i_size
4750          */
4751         if (inode->i_size >> PAGE_CACHE_SHIFT == last_page &&
4752            inode->i_size % PAGE_CACHE_SIZE != 0) {
4753
4754                 page_len = PAGE_CACHE_SIZE -
4755                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4756
4757                 if (page_len > 0) {
4758                         err = ext4_discard_partial_page_buffers(handle,
4759                           mapping, inode->i_size, page_len, 0);
4760
4761                         if (err)
4762                                 goto out;
4763                 }
4764         }
4765
4766         first_block = (offset + sb->s_blocksize - 1) >>
4767                 EXT4_BLOCK_SIZE_BITS(sb);
4768         stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4769
4770         /* If there are no blocks to remove, return now */
4771         if (first_block >= stop_block)
4772                 goto out;
4773
4774         down_write(&EXT4_I(inode)->i_data_sem);
4775         ext4_ext_invalidate_cache(inode);
4776         ext4_discard_preallocations(inode);
4777
4778         err = ext4_es_remove_extent(inode, first_block,
4779                                     stop_block - first_block);
4780         err = ext4_ext_remove_space(inode, first_block, stop_block - 1);
4781
4782         ext4_ext_invalidate_cache(inode);
4783         ext4_discard_preallocations(inode);
4784
4785         if (IS_SYNC(inode))
4786                 ext4_handle_sync(handle);
4787
4788         up_write(&EXT4_I(inode)->i_data_sem);
4789
4790 out:
4791         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4792         ext4_mark_inode_dirty(handle, inode);
4793         ext4_journal_stop(handle);
4794 out_dio:
4795         ext4_inode_resume_unlocked_dio(inode);
4796 out_mutex:
4797         mutex_unlock(&inode->i_mutex);
4798         return err;
4799 }
4800
4801 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4802                 __u64 start, __u64 len)
4803 {
4804         ext4_lblk_t start_blk;
4805         int error = 0;
4806
4807         if (ext4_has_inline_data(inode)) {
4808                 int has_inline = 1;
4809
4810                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
4811
4812                 if (has_inline)
4813                         return error;
4814         }
4815
4816         /* fallback to generic here if not in extents fmt */
4817         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4818                 return generic_block_fiemap(inode, fieinfo, start, len,
4819                         ext4_get_block);
4820
4821         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4822                 return -EBADR;
4823
4824         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4825                 error = ext4_xattr_fiemap(inode, fieinfo);
4826         } else {
4827                 ext4_lblk_t len_blks;
4828                 __u64 last_blk;
4829
4830                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4831                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4832                 if (last_blk >= EXT_MAX_BLOCKS)
4833                         last_blk = EXT_MAX_BLOCKS-1;
4834                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4835
4836                 /*
4837                  * Walk the extent tree gathering extent information
4838                  * and pushing extents back to the user.
4839                  */
4840                 error = ext4_fill_fiemap_extents(inode, start_blk,
4841                                                  len_blks, fieinfo);
4842         }
4843
4844         return error;
4845 }