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/module.h>
  33 #include <linux/fs.h>
  34 #include <linux/time.h>
  35 #include <linux/jbd2.h>
  36 #include <linux/highuid.h>
  37 #include <linux/pagemap.h>
  38 #include <linux/quotaops.h>
  39 #include <linux/string.h>
  40 #include <linux/slab.h>
  41 #include <linux/falloc.h>
  42 #include <asm/uaccess.h>
  43 #include <linux/fiemap.h>
  44 #include "ext4_jbd2.h"
  45
  46 #include <trace/events/ext4.h>
  47
  48 static int ext4_split_extent(handle_t *handle,
  49                                 struct inode *inode,
  50                                 struct ext4_ext_path *path,
  51                                 struct ext4_map_blocks *map,
  52                                 int split_flag,
  53                                 int flags);
  54
  55 static int ext4_ext_truncate_extend_restart(handle_t *handle,
  56                                             struct inode *inode,
  57                                             int needed)
  58 {
  59         int err;
  60
  61         if (!ext4_handle_valid(handle))
  62                 return 0;
  63         if (handle->h_buffer_credits > needed)
  64                 return 0;
  65         err = ext4_journal_extend(handle, needed);
  66         if (err <= 0)
  67                 return err;
  68         err = ext4_truncate_restart_trans(handle, inode, needed);
  69         if (err == 0)
  70                 err = -EAGAIN;
  71
  72         return err;
  73 }
  74
  75 /*
  76  * could return:
  77  *  - EROFS
  78  *  - ENOMEM
  79  */
  80 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
  81                                 struct ext4_ext_path *path)
  82 {
  83         if (path->p_bh) {
  84                 /* path points to block */
  85                 return ext4_journal_get_write_access(handle, path->p_bh);
  86         }
  87         /* path points to leaf/index in inode body */
  88         /* we use in-core data, no need to protect them */
  89         return 0;
  90 }
  91
  92 /*
  93  * could return:
  94  *  - EROFS
  95  *  - ENOMEM
  96  *  - EIO
  97  */
  98 #define ext4_ext_dirty(handle, inode, path) \
  99                 __ext4_ext_dirty(__func__, __LINE__, (handle), (inode), (path))
 100 static int __ext4_ext_dirty(const char *where, unsigned int line,
 101                             handle_t *handle, struct inode *inode,
 102                             struct ext4_ext_path *path)
 103 {
 104         int err;
 105         if (path->p_bh) {
 106                 /* path points to block */
 107                 err = __ext4_handle_dirty_metadata(where, line, handle,
 108                                                    inode, path->p_bh);
 109         } else {
 110                 /* path points to leaf/index in inode body */
 111                 err = ext4_mark_inode_dirty(handle, inode);
 112         }
 113         return err;
 114 }
 115
 116 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 117                               struct ext4_ext_path *path,
 118                               ext4_lblk_t block)
 119 {
 120         if (path) {
 121                 int depth = path->p_depth;
 122                 struct ext4_extent *ex;
 123
 124                 /*
 125                  * Try to predict block placement assuming that we are
 126                  * filling in a file which will eventually be
 127                  * non-sparse --- i.e., in the case of libbfd writing
 128                  * an ELF object sections out-of-order but in a way
 129                  * the eventually results in a contiguous object or
 130                  * executable file, or some database extending a table
 131                  * space file.  However, this is actually somewhat
 132                  * non-ideal if we are writing a sparse file such as
 133                  * qemu or KVM writing a raw image file that is going
 134                  * to stay fairly sparse, since it will end up
 135                  * fragmenting the file system's free space.  Maybe we
 136                  * should have some hueristics or some way to allow
 137                  * userspace to pass a hint to file system,
 138                  * especially if the latter case turns out to be
 139                  * common.
 140                  */
 141                 ex = path[depth].p_ext;
 142                 if (ex) {
 143                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 144                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 145
 146                         if (block > ext_block)
 147                                 return ext_pblk + (block - ext_block);
 148                         else
 149                                 return ext_pblk - (ext_block - block);
 150                 }
 151
 152                 /* it looks like index is empty;
 153                  * try to find starting block from index itself */
 154                 if (path[depth].p_bh)
 155                         return path[depth].p_bh->b_blocknr;
 156         }
 157
 158         /* OK. use inode's group */
 159         return ext4_inode_to_goal_block(inode);
 160 }
 161
 162 /*
 163  * Allocation for a meta data block
 164  */
 165 static ext4_fsblk_t
 166 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 167                         struct ext4_ext_path *path,
 168                         struct ext4_extent *ex, int *err, unsigned int flags)
 169 {
 170         ext4_fsblk_t goal, newblock;
 171
 172         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 173         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 174                                         NULL, err);
 175         return newblock;
 176 }
 177
 178 static inline int ext4_ext_space_block(struct inode *inode, int check)
 179 {
 180         int size;
 181
 182         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 183                         / sizeof(struct ext4_extent);
 184 #ifdef AGGRESSIVE_TEST
 185         if (!check && size > 6)
 186                 size = 6;
 187 #endif
 188         return size;
 189 }
 190
 191 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 192 {
 193         int size;
 194
 195         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 196                         / sizeof(struct ext4_extent_idx);
 197 #ifdef AGGRESSIVE_TEST
 198         if (!check && size > 5)
 199                 size = 5;
 200 #endif
 201         return size;
 202 }
 203
 204 static inline int ext4_ext_space_root(struct inode *inode, int check)
 205 {
 206         int size;
 207
 208         size = sizeof(EXT4_I(inode)->i_data);
 209         size -= sizeof(struct ext4_extent_header);
 210         size /= sizeof(struct ext4_extent);
 211 #ifdef AGGRESSIVE_TEST
 212         if (!check && size > 3)
 213                 size = 3;
 214 #endif
 215         return size;
 216 }
 217
 218 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 219 {
 220         int size;
 221
 222         size = sizeof(EXT4_I(inode)->i_data);
 223         size -= sizeof(struct ext4_extent_header);
 224         size /= sizeof(struct ext4_extent_idx);
 225 #ifdef AGGRESSIVE_TEST
 226         if (!check && size > 4)
 227                 size = 4;
 228 #endif
 229         return size;
 230 }
 231
 232 /*
 233  * Calculate the number of metadata blocks needed
 234  * to allocate @blocks
 235  * Worse case is one block per extent
 236  */
 237 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 238 {
 239         struct ext4_inode_info *ei = EXT4_I(inode);
 240         int idxs;
 241
 242         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 243                 / sizeof(struct ext4_extent_idx));
 244
 245         /*
 246          * If the new delayed allocation block is contiguous with the
 247          * previous da block, it can share index blocks with the
 248          * previous block, so we only need to allocate a new index
 249          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 250          * an additional index block, and at ldxs**3 blocks, yet
 251          * another index blocks.
 252          */
 253         if (ei->i_da_metadata_calc_len &&
 254             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 255                 int num = 0;
 256
 257                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 258                         num++;
 259                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 260                         num++;
 261                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 262                         num++;
 263                         ei->i_da_metadata_calc_len = 0;
 264                 } else
 265                         ei->i_da_metadata_calc_len++;
 266                 ei->i_da_metadata_calc_last_lblock++;
 267                 return num;
 268         }
 269
 270         /*
 271          * In the worst case we need a new set of index blocks at
 272          * every level of the inode's extent tree.
 273          */
 274         ei->i_da_metadata_calc_len = 1;
 275         ei->i_da_metadata_calc_last_lblock = lblock;
 276         return ext_depth(inode) + 1;
 277 }
 278
 279 static int
 280 ext4_ext_max_entries(struct inode *inode, int depth)
 281 {
 282         int max;
 283
 284         if (depth == ext_depth(inode)) {
 285                 if (depth == 0)
 286                         max = ext4_ext_space_root(inode, 1);
 287                 else
 288                         max = ext4_ext_space_root_idx(inode, 1);
 289         } else {
 290                 if (depth == 0)
 291                         max = ext4_ext_space_block(inode, 1);
 292                 else
 293                         max = ext4_ext_space_block_idx(inode, 1);
 294         }
 295
 296         return max;
 297 }
 298
 299 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 300 {
 301         ext4_fsblk_t block = ext4_ext_pblock(ext);
 302         int len = ext4_ext_get_actual_len(ext);
 303
 304         if (len == 0)
 305                 return 0;
 306         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 307 }
 308
 309 static int ext4_valid_extent_idx(struct inode *inode,
 310                                 struct ext4_extent_idx *ext_idx)
 311 {
 312         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 313
 314         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 315 }
 316
 317 static int ext4_valid_extent_entries(struct inode *inode,
 318                                 struct ext4_extent_header *eh,
 319                                 int depth)
 320 {
 321         unsigned short entries;
 322         if (eh->eh_entries == 0)
 323                 return 1;
 324
 325         entries = le16_to_cpu(eh->eh_entries);
 326
 327         if (depth == 0) {
 328                 /* leaf entries */
 329                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 330                 while (entries) {
 331                         if (!ext4_valid_extent(inode, ext))
 332                                 return 0;
 333                         ext++;
 334                         entries--;
 335                 }
 336         } else {
 337                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 338                 while (entries) {
 339                         if (!ext4_valid_extent_idx(inode, ext_idx))
 340                                 return 0;
 341                         ext_idx++;
 342                         entries--;
 343                 }
 344         }
 345         return 1;
 346 }
 347
 348 static int __ext4_ext_check(const char *function, unsigned int line,
 349                             struct inode *inode, struct ext4_extent_header *eh,
 350                             int depth)
 351 {
 352         const char *error_msg;
 353         int max = 0;
 354
 355         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 356                 error_msg = "invalid magic";
 357                 goto corrupted;
 358         }
 359         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 360                 error_msg = "unexpected eh_depth";
 361                 goto corrupted;
 362         }
 363         if (unlikely(eh->eh_max == 0)) {
 364                 error_msg = "invalid eh_max";
 365                 goto corrupted;
 366         }
 367         max = ext4_ext_max_entries(inode, depth);
 368         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 369                 error_msg = "too large eh_max";
 370                 goto corrupted;
 371         }
 372         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 373                 error_msg = "invalid eh_entries";
 374                 goto corrupted;
 375         }
 376         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 377                 error_msg = "invalid extent entries";
 378                 goto corrupted;
 379         }
 380         return 0;
 381
 382 corrupted:
 383         ext4_error_inode(inode, function, line, 0,
 384                         "bad header/extent: %s - magic %x, "
 385                         "entries %u, max %u(%u), depth %u(%u)",
 386                         error_msg, le16_to_cpu(eh->eh_magic),
 387                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 388                         max, le16_to_cpu(eh->eh_depth), depth);
 389
 390         return -EIO;
 391 }
 392
 393 #define ext4_ext_check(inode, eh, depth)        \
 394         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
 395
 396 int ext4_ext_check_inode(struct inode *inode)
 397 {
 398         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
 399 }
 400
 401 #ifdef EXT_DEBUG
 402 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 403 {
 404         int k, l = path->p_depth;
 405
 406         ext_debug("path:");
 407         for (k = 0; k <= l; k++, path++) {
 408                 if (path->p_idx) {
 409                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 410                             ext4_idx_pblock(path->p_idx));
 411                 } else if (path->p_ext) {
 412                         ext_debug("  %d:[%d]%d:%llu ",
 413                                   le32_to_cpu(path->p_ext->ee_block),
 414                                   ext4_ext_is_uninitialized(path->p_ext),
 415                                   ext4_ext_get_actual_len(path->p_ext),
 416                                   ext4_ext_pblock(path->p_ext));
 417                 } else
 418                         ext_debug("  []");
 419         }
 420         ext_debug("\n");
 421 }
 422
 423 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 424 {
 425         int depth = ext_depth(inode);
 426         struct ext4_extent_header *eh;
 427         struct ext4_extent *ex;
 428         int i;
 429
 430         if (!path)
 431                 return;
 432
 433         eh = path[depth].p_hdr;
 434         ex = EXT_FIRST_EXTENT(eh);
 435
 436         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 437
 438         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 439                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 440                           ext4_ext_is_uninitialized(ex),
 441                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 442         }
 443         ext_debug("\n");
 444 }
 445
 446 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 447                         ext4_fsblk_t newblock, int level)
 448 {
 449         int depth = ext_depth(inode);
 450         struct ext4_extent *ex;
 451
 452         if (depth != level) {
 453                 struct ext4_extent_idx *idx;
 454                 idx = path[level].p_idx;
 455                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 456                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 457                                         le32_to_cpu(idx->ei_block),
 458                                         ext4_idx_pblock(idx),
 459                                         newblock);
 460                         idx++;
 461                 }
 462
 463                 return;
 464         }
 465
 466         ex = path[depth].p_ext;
 467         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 468                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 469                                 le32_to_cpu(ex->ee_block),
 470                                 ext4_ext_pblock(ex),
 471                                 ext4_ext_is_uninitialized(ex),
 472                                 ext4_ext_get_actual_len(ex),
 473                                 newblock);
 474                 ex++;
 475         }
 476 }
 477
 478 #else
 479 #define ext4_ext_show_path(inode, path)
 480 #define ext4_ext_show_leaf(inode, path)
 481 #define ext4_ext_show_move(inode, path, newblock, level)
 482 #endif
 483
 484 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 485 {
 486         int depth = path->p_depth;
 487         int i;
 488
 489         for (i = 0; i <= depth; i++, path++)
 490                 if (path->p_bh) {
 491                         brelse(path->p_bh);
 492                         path->p_bh = NULL;
 493                 }
 494 }
 495
 496 /*
 497  * ext4_ext_binsearch_idx:
 498  * binary search for the closest index of the given block
 499  * the header must be checked before calling this
 500  */
 501 static void
 502 ext4_ext_binsearch_idx(struct inode *inode,
 503                         struct ext4_ext_path *path, ext4_lblk_t block)
 504 {
 505         struct ext4_extent_header *eh = path->p_hdr;
 506         struct ext4_extent_idx *r, *l, *m;
 507
 508
 509         ext_debug("binsearch for %u(idx):  ", block);
 510
 511         l = EXT_FIRST_INDEX(eh) + 1;
 512         r = EXT_LAST_INDEX(eh);
 513         while (l <= r) {
 514                 m = l + (r - l) / 2;
 515                 if (block < le32_to_cpu(m->ei_block))
 516                         r = m - 1;
 517                 else
 518                         l = m + 1;
 519                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 520                                 m, le32_to_cpu(m->ei_block),
 521                                 r, le32_to_cpu(r->ei_block));
 522         }
 523
 524         path->p_idx = l - 1;
 525         ext_debug("  -> %d->%lld ", le32_to_cpu(path->p_idx->ei_block),
 526                   ext4_idx_pblock(path->p_idx));
 527
 528 #ifdef CHECK_BINSEARCH
 529         {
 530                 struct ext4_extent_idx *chix, *ix;
 531                 int k;
 532
 533                 chix = ix = EXT_FIRST_INDEX(eh);
 534                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 535                   if (k != 0 &&
 536                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 537                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 538                                        "first=0x%p\n", k,
 539                                        ix, EXT_FIRST_INDEX(eh));
 540                                 printk(KERN_DEBUG "%u <= %u\n",
 541                                        le32_to_cpu(ix->ei_block),
 542                                        le32_to_cpu(ix[-1].ei_block));
 543                         }
 544                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 545                                            <= le32_to_cpu(ix[-1].ei_block));
 546                         if (block < le32_to_cpu(ix->ei_block))
 547                                 break;
 548                         chix = ix;
 549                 }
 550                 BUG_ON(chix != path->p_idx);
 551         }
 552 #endif
 553
 554 }
 555
 556 /*
 557  * ext4_ext_binsearch:
 558  * binary search for closest extent of the given block
 559  * the header must be checked before calling this
 560  */
 561 static void
 562 ext4_ext_binsearch(struct inode *inode,
 563                 struct ext4_ext_path *path, ext4_lblk_t block)
 564 {
 565         struct ext4_extent_header *eh = path->p_hdr;
 566         struct ext4_extent *r, *l, *m;
 567
 568         if (eh->eh_entries == 0) {
 569                 /*
 570                  * this leaf is empty:
 571                  * we get such a leaf in split/add case
 572                  */
 573                 return;
 574         }
 575
 576         ext_debug("binsearch for %u:  ", block);
 577
 578         l = EXT_FIRST_EXTENT(eh) + 1;
 579         r = EXT_LAST_EXTENT(eh);
 580
 581         while (l <= r) {
 582                 m = l + (r - l) / 2;
 583                 if (block < le32_to_cpu(m->ee_block))
 584                         r = m - 1;
 585                 else
 586                         l = m + 1;
 587                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 588                                 m, le32_to_cpu(m->ee_block),
 589                                 r, le32_to_cpu(r->ee_block));
 590         }
 591
 592         path->p_ext = l - 1;
 593         ext_debug("  -> %d:%llu:[%d]%d ",
 594                         le32_to_cpu(path->p_ext->ee_block),
 595                         ext4_ext_pblock(path->p_ext),
 596                         ext4_ext_is_uninitialized(path->p_ext),
 597                         ext4_ext_get_actual_len(path->p_ext));
 598
 599 #ifdef CHECK_BINSEARCH
 600         {
 601                 struct ext4_extent *chex, *ex;
 602                 int k;
 603
 604                 chex = ex = EXT_FIRST_EXTENT(eh);
 605                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 606                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 607                                           <= le32_to_cpu(ex[-1].ee_block));
 608                         if (block < le32_to_cpu(ex->ee_block))
 609                                 break;
 610                         chex = ex;
 611                 }
 612                 BUG_ON(chex != path->p_ext);
 613         }
 614 #endif
 615
 616 }
 617
 618 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 619 {
 620         struct ext4_extent_header *eh;
 621
 622         eh = ext_inode_hdr(inode);
 623         eh->eh_depth = 0;
 624         eh->eh_entries = 0;
 625         eh->eh_magic = EXT4_EXT_MAGIC;
 626         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 627         ext4_mark_inode_dirty(handle, inode);
 628         ext4_ext_invalidate_cache(inode);
 629         return 0;
 630 }
 631
 632 struct ext4_ext_path *
 633 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
 634                                         struct ext4_ext_path *path)
 635 {
 636         struct ext4_extent_header *eh;
 637         struct buffer_head *bh;
 638         short int depth, i, ppos = 0, alloc = 0;
 639
 640         eh = ext_inode_hdr(inode);
 641         depth = ext_depth(inode);
 642
 643         /* account possible depth increase */
 644         if (!path) {
 645                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 646                                 GFP_NOFS);
 647                 if (!path)
 648                         return ERR_PTR(-ENOMEM);
 649                 alloc = 1;
 650         }
 651         path[0].p_hdr = eh;
 652         path[0].p_bh = NULL;
 653
 654         i = depth;
 655         /* walk through the tree */
 656         while (i) {
 657                 int need_to_validate = 0;
 658
 659                 ext_debug("depth %d: num %d, max %d\n",
 660                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 661
 662                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 663                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 664                 path[ppos].p_depth = i;
 665                 path[ppos].p_ext = NULL;
 666
 667                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
 668                 if (unlikely(!bh))
 669                         goto err;
 670                 if (!bh_uptodate_or_lock(bh)) {
 671                         trace_ext4_ext_load_extent(inode, block,
 672                                                 path[ppos].p_block);
 673                         if (bh_submit_read(bh) < 0) {
 674                                 put_bh(bh);
 675                                 goto err;
 676                         }
 677                         /* validate the extent entries */
 678                         need_to_validate = 1;
 679                 }
 680                 eh = ext_block_hdr(bh);
 681                 ppos++;
 682                 if (unlikely(ppos > depth)) {
 683                         put_bh(bh);
 684                         EXT4_ERROR_INODE(inode,
 685                                          "ppos %d > depth %d", ppos, depth);
 686                         goto err;
 687                 }
 688                 path[ppos].p_bh = bh;
 689                 path[ppos].p_hdr = eh;
 690                 i--;
 691
 692                 if (need_to_validate && ext4_ext_check(inode, eh, i))
 693                         goto err;
 694         }
 695
 696         path[ppos].p_depth = i;
 697         path[ppos].p_ext = NULL;
 698         path[ppos].p_idx = NULL;
 699
 700         /* find extent */
 701         ext4_ext_binsearch(inode, path + ppos, block);
 702         /* if not an empty leaf */
 703         if (path[ppos].p_ext)
 704                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 705
 706         ext4_ext_show_path(inode, path);
 707
 708         return path;
 709
 710 err:
 711         ext4_ext_drop_refs(path);
 712         if (alloc)
 713                 kfree(path);
 714         return ERR_PTR(-EIO);
 715 }
 716
 717 /*
 718  * ext4_ext_insert_index:
 719  * insert new index [@logical;@ptr] into the block at @curp;
 720  * check where to insert: before @curp or after @curp
 721  */
 722 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 723                                  struct ext4_ext_path *curp,
 724                                  int logical, ext4_fsblk_t ptr)
 725 {
 726         struct ext4_extent_idx *ix;
 727         int len, err;
 728
 729         err = ext4_ext_get_access(handle, inode, curp);
 730         if (err)
 731                 return err;
 732
 733         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 734                 EXT4_ERROR_INODE(inode,
 735                                  "logical %d == ei_block %d!",
 736                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 737                 return -EIO;
 738         }
 739
 740         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 741                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 742                 EXT4_ERROR_INODE(inode,
 743                                  "eh_entries %d >= eh_max %d!",
 744                                  le16_to_cpu(curp->p_hdr->eh_entries),
 745                                  le16_to_cpu(curp->p_hdr->eh_max));
 746                 return -EIO;
 747         }
 748
 749         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 750                 /* insert after */
 751                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 752                 ix = curp->p_idx + 1;
 753         } else {
 754                 /* insert before */
 755                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 756                 ix = curp->p_idx;
 757         }
 758
 759         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 760         BUG_ON(len < 0);
 761         if (len > 0) {
 762                 ext_debug("insert new index %d: "
 763                                 "move %d indices from 0x%p to 0x%p\n",
 764                                 logical, len, ix, ix + 1);
 765                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 766         }
 767
 768         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 769                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 770                 return -EIO;
 771         }
 772
 773         ix->ei_block = cpu_to_le32(logical);
 774         ext4_idx_store_pblock(ix, ptr);
 775         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 776
 777         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 778                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 779                 return -EIO;
 780         }
 781
 782         err = ext4_ext_dirty(handle, inode, curp);
 783         ext4_std_error(inode->i_sb, err);
 784
 785         return err;
 786 }
 787
 788 /*
 789  * ext4_ext_split:
 790  * inserts new subtree into the path, using free index entry
 791  * at depth @at:
 792  * - allocates all needed blocks (new leaf and all intermediate index blocks)
 793  * - makes decision where to split
 794  * - moves remaining extents and index entries (right to the split point)
 795  *   into the newly allocated blocks
 796  * - initializes subtree
 797  */
 798 static int ext4_ext_split(handle_t *handle, struct inode *inode,
 799                           unsigned int flags,
 800                           struct ext4_ext_path *path,
 801                           struct ext4_extent *newext, int at)
 802 {
 803         struct buffer_head *bh = NULL;
 804         int depth = ext_depth(inode);
 805         struct ext4_extent_header *neh;
 806         struct ext4_extent_idx *fidx;
 807         int i = at, k, m, a;
 808         ext4_fsblk_t newblock, oldblock;
 809         __le32 border;
 810         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
 811         int err = 0;
 812
 813         /* make decision: where to split? */
 814         /* FIXME: now decision is simplest: at current extent */
 815
 816         /* if current leaf will be split, then we should use
 817          * border from split point */
 818         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
 819                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
 820                 return -EIO;
 821         }
 822         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
 823                 border = path[depth].p_ext[1].ee_block;
 824                 ext_debug("leaf will be split."
 825                                 " next leaf starts at %d\n",
 826                                   le32_to_cpu(border));
 827         } else {
 828                 border = newext->ee_block;
 829                 ext_debug("leaf will be added."
 830                                 " next leaf starts at %d\n",
 831                                 le32_to_cpu(border));
 832         }
 833
 834         /*
 835          * If error occurs, then we break processing
 836          * and mark filesystem read-only. index won't
 837          * be inserted and tree will be in consistent
 838          * state. Next mount will repair buffers too.
 839          */
 840
 841         /*
 842          * Get array to track all allocated blocks.
 843          * We need this to handle errors and free blocks
 844          * upon them.
 845          */
 846         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
 847         if (!ablocks)
 848                 return -ENOMEM;
 849
 850         /* allocate all needed blocks */
 851         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
 852         for (a = 0; a < depth - at; a++) {
 853                 newblock = ext4_ext_new_meta_block(handle, inode, path,
 854                                                    newext, &err, flags);
 855                 if (newblock == 0)
 856                         goto cleanup;
 857                 ablocks[a] = newblock;
 858         }
 859
 860         /* initialize new leaf */
 861         newblock = ablocks[--a];
 862         if (unlikely(newblock == 0)) {
 863                 EXT4_ERROR_INODE(inode, "newblock == 0!");
 864                 err = -EIO;
 865                 goto cleanup;
 866         }
 867         bh = sb_getblk(inode->i_sb, newblock);
 868         if (!bh) {
 869                 err = -EIO;
 870                 goto cleanup;
 871         }
 872         lock_buffer(bh);
 873
 874         err = ext4_journal_get_create_access(handle, bh);
 875         if (err)
 876                 goto cleanup;
 877
 878         neh = ext_block_hdr(bh);
 879         neh->eh_entries = 0;
 880         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
 881         neh->eh_magic = EXT4_EXT_MAGIC;
 882         neh->eh_depth = 0;
 883
 884         /* move remainder of path[depth] to the new leaf */
 885         if (unlikely(path[depth].p_hdr->eh_entries !=
 886                      path[depth].p_hdr->eh_max)) {
 887                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
 888                                  path[depth].p_hdr->eh_entries,
 889                                  path[depth].p_hdr->eh_max);
 890                 err = -EIO;
 891                 goto cleanup;
 892         }
 893         /* start copy from next extent */
 894         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
 895         ext4_ext_show_move(inode, path, newblock, depth);
 896         if (m) {
 897                 struct ext4_extent *ex;
 898                 ex = EXT_FIRST_EXTENT(neh);
 899                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
 900                 le16_add_cpu(&neh->eh_entries, m);
 901         }
 902
 903         set_buffer_uptodate(bh);
 904         unlock_buffer(bh);
 905
 906         err = ext4_handle_dirty_metadata(handle, inode, bh);
 907         if (err)
 908                 goto cleanup;
 909         brelse(bh);
 910         bh = NULL;
 911
 912         /* correct old leaf */
 913         if (m) {
 914                 err = ext4_ext_get_access(handle, inode, path + depth);
 915                 if (err)
 916                         goto cleanup;
 917                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
 918                 err = ext4_ext_dirty(handle, inode, path + depth);
 919                 if (err)
 920                         goto cleanup;
 921
 922         }
 923
 924         /* create intermediate indexes */
 925         k = depth - at - 1;
 926         if (unlikely(k < 0)) {
 927                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
 928                 err = -EIO;
 929                 goto cleanup;
 930         }
 931         if (k)
 932                 ext_debug("create %d intermediate indices\n", k);
 933         /* insert new index into current index block */
 934         /* current depth stored in i var */
 935         i = depth - 1;
 936         while (k--) {
 937                 oldblock = newblock;
 938                 newblock = ablocks[--a];
 939                 bh = sb_getblk(inode->i_sb, newblock);
 940                 if (!bh) {
 941                         err = -EIO;
 942                         goto cleanup;
 943                 }
 944                 lock_buffer(bh);
 945
 946                 err = ext4_journal_get_create_access(handle, bh);
 947                 if (err)
 948                         goto cleanup;
 949
 950                 neh = ext_block_hdr(bh);
 951                 neh->eh_entries = cpu_to_le16(1);
 952                 neh->eh_magic = EXT4_EXT_MAGIC;
 953                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
 954                 neh->eh_depth = cpu_to_le16(depth - i);
 955                 fidx = EXT_FIRST_INDEX(neh);
 956                 fidx->ei_block = border;
 957                 ext4_idx_store_pblock(fidx, oldblock);
 958
 959                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
 960                                 i, newblock, le32_to_cpu(border), oldblock);
 961
 962                 /* move remainder of path[i] to the new index block */
 963                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
 964                                         EXT_LAST_INDEX(path[i].p_hdr))) {
 965                         EXT4_ERROR_INODE(inode,
 966                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
 967                                          le32_to_cpu(path[i].p_ext->ee_block));
 968                         err = -EIO;
 969                         goto cleanup;
 970                 }
 971                 /* start copy indexes */
 972                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
 973                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
 974                                 EXT_MAX_INDEX(path[i].p_hdr));
 975                 ext4_ext_show_move(inode, path, newblock, i);
 976                 if (m) {
 977                         memmove(++fidx, path[i].p_idx,
 978                                 sizeof(struct ext4_extent_idx) * m);
 979                         le16_add_cpu(&neh->eh_entries, m);
 980                 }
 981                 set_buffer_uptodate(bh);
 982                 unlock_buffer(bh);
 983
 984                 err = ext4_handle_dirty_metadata(handle, inode, bh);
 985                 if (err)
 986                         goto cleanup;
 987                 brelse(bh);
 988                 bh = NULL;
 989
 990                 /* correct old index */
 991                 if (m) {
 992                         err = ext4_ext_get_access(handle, inode, path + i);
 993                         if (err)
 994                                 goto cleanup;
 995                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
 996                         err = ext4_ext_dirty(handle, inode, path + i);
 997                         if (err)
 998                                 goto cleanup;
 999                 }
1000
1001                 i--;
1002         }
1003
1004         /* insert new index */
1005         err = ext4_ext_insert_index(handle, inode, path + at,
1006                                     le32_to_cpu(border), newblock);
1007
1008 cleanup:
1009         if (bh) {
1010                 if (buffer_locked(bh))
1011                         unlock_buffer(bh);
1012                 brelse(bh);
1013         }
1014
1015         if (err) {
1016                 /* free all allocated blocks in error case */
1017                 for (i = 0; i < depth; i++) {
1018                         if (!ablocks[i])
1019                                 continue;
1020                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1021                                          EXT4_FREE_BLOCKS_METADATA);
1022                 }
1023         }
1024         kfree(ablocks);
1025
1026         return err;
1027 }
1028
1029 /*
1030  * ext4_ext_grow_indepth:
1031  * implements tree growing procedure:
1032  * - allocates new block
1033  * - moves top-level data (index block or leaf) into the new block
1034  * - initializes new top-level, creating index that points to the
1035  *   just created block
1036  */
1037 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1038                                  unsigned int flags,
1039                                  struct ext4_extent *newext)
1040 {
1041         struct ext4_extent_header *neh;
1042         struct buffer_head *bh;
1043         ext4_fsblk_t newblock;
1044         int err = 0;
1045
1046         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1047                 newext, &err, flags);
1048         if (newblock == 0)
1049                 return err;
1050
1051         bh = sb_getblk(inode->i_sb, newblock);
1052         if (!bh) {
1053                 err = -EIO;
1054                 ext4_std_error(inode->i_sb, err);
1055                 return err;
1056         }
1057         lock_buffer(bh);
1058
1059         err = ext4_journal_get_create_access(handle, bh);
1060         if (err) {
1061                 unlock_buffer(bh);
1062                 goto out;
1063         }
1064
1065         /* move top-level index/leaf into new block */
1066         memmove(bh->b_data, EXT4_I(inode)->i_data,
1067                 sizeof(EXT4_I(inode)->i_data));
1068
1069         /* set size of new block */
1070         neh = ext_block_hdr(bh);
1071         /* old root could have indexes or leaves
1072          * so calculate e_max right way */
1073         if (ext_depth(inode))
1074                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1075         else
1076                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1077         neh->eh_magic = EXT4_EXT_MAGIC;
1078         set_buffer_uptodate(bh);
1079         unlock_buffer(bh);
1080
1081         err = ext4_handle_dirty_metadata(handle, inode, bh);
1082         if (err)
1083                 goto out;
1084
1085         /* Update top-level index: num,max,pointer */
1086         neh = ext_inode_hdr(inode);
1087         neh->eh_entries = cpu_to_le16(1);
1088         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1089         if (neh->eh_depth == 0) {
1090                 /* Root extent block becomes index block */
1091                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1092                 EXT_FIRST_INDEX(neh)->ei_block =
1093                         EXT_FIRST_EXTENT(neh)->ee_block;
1094         }
1095         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1096                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1097                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1098                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1099
1100         neh->eh_depth = cpu_to_le16(le16_to_cpu(neh->eh_depth) + 1);
1101         ext4_mark_inode_dirty(handle, inode);
1102 out:
1103         brelse(bh);
1104
1105         return err;
1106 }
1107
1108 /*
1109  * ext4_ext_create_new_leaf:
1110  * finds empty index and adds new leaf.
1111  * if no free index is found, then it requests in-depth growing.
1112  */
1113 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1114                                     unsigned int flags,
1115                                     struct ext4_ext_path *path,
1116                                     struct ext4_extent *newext)
1117 {
1118         struct ext4_ext_path *curp;
1119         int depth, i, err = 0;
1120
1121 repeat:
1122         i = depth = ext_depth(inode);
1123
1124         /* walk up to the tree and look for free index entry */
1125         curp = path + depth;
1126         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1127                 i--;
1128                 curp--;
1129         }
1130
1131         /* we use already allocated block for index block,
1132          * so subsequent data blocks should be contiguous */
1133         if (EXT_HAS_FREE_INDEX(curp)) {
1134                 /* if we found index with free entry, then use that
1135                  * entry: create all needed subtree and add new leaf */
1136                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1137                 if (err)
1138                         goto out;
1139
1140                 /* refill path */
1141                 ext4_ext_drop_refs(path);
1142                 path = ext4_ext_find_extent(inode,
1143                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1144                                     path);
1145                 if (IS_ERR(path))
1146                         err = PTR_ERR(path);
1147         } else {
1148                 /* tree is full, time to grow in depth */
1149                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1150                 if (err)
1151                         goto out;
1152
1153                 /* refill path */
1154                 ext4_ext_drop_refs(path);
1155                 path = ext4_ext_find_extent(inode,
1156                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1157                                     path);
1158                 if (IS_ERR(path)) {
1159                         err = PTR_ERR(path);
1160                         goto out;
1161                 }
1162
1163                 /*
1164                  * only first (depth 0 -> 1) produces free space;
1165                  * in all other cases we have to split the grown tree
1166                  */
1167                 depth = ext_depth(inode);
1168                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1169                         /* now we need to split */
1170                         goto repeat;
1171                 }
1172         }
1173
1174 out:
1175         return err;
1176 }
1177
1178 /*
1179  * search the closest allocated block to the left for *logical
1180  * and returns it at @logical + it's physical address at @phys
1181  * if *logical is the smallest allocated block, the function
1182  * returns 0 at @phys
1183  * return value contains 0 (success) or error code
1184  */
1185 static int ext4_ext_search_left(struct inode *inode,
1186                                 struct ext4_ext_path *path,
1187                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1188 {
1189         struct ext4_extent_idx *ix;
1190         struct ext4_extent *ex;
1191         int depth, ee_len;
1192
1193         if (unlikely(path == NULL)) {
1194                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1195                 return -EIO;
1196         }
1197         depth = path->p_depth;
1198         *phys = 0;
1199
1200         if (depth == 0 && path->p_ext == NULL)
1201                 return 0;
1202
1203         /* usually extent in the path covers blocks smaller
1204          * then *logical, but it can be that extent is the
1205          * first one in the file */
1206
1207         ex = path[depth].p_ext;
1208         ee_len = ext4_ext_get_actual_len(ex);
1209         if (*logical < le32_to_cpu(ex->ee_block)) {
1210                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1211                         EXT4_ERROR_INODE(inode,
1212                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1213                                          *logical, le32_to_cpu(ex->ee_block));
1214                         return -EIO;
1215                 }
1216                 while (--depth >= 0) {
1217                         ix = path[depth].p_idx;
1218                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1219                                 EXT4_ERROR_INODE(inode,
1220                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1221                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1222                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1223                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1224                                   depth);
1225                                 return -EIO;
1226                         }
1227                 }
1228                 return 0;
1229         }
1230
1231         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1232                 EXT4_ERROR_INODE(inode,
1233                                  "logical %d < ee_block %d + ee_len %d!",
1234                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1235                 return -EIO;
1236         }
1237
1238         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1239         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1240         return 0;
1241 }
1242
1243 /*
1244  * search the closest allocated block to the right for *logical
1245  * and returns it at @logical + it's physical address at @phys
1246  * if *logical is the largest allocated block, the function
1247  * returns 0 at @phys
1248  * return value contains 0 (success) or error code
1249  */
1250 static int ext4_ext_search_right(struct inode *inode,
1251                                  struct ext4_ext_path *path,
1252                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1253                                  struct ext4_extent **ret_ex)
1254 {
1255         struct buffer_head *bh = NULL;
1256         struct ext4_extent_header *eh;
1257         struct ext4_extent_idx *ix;
1258         struct ext4_extent *ex;
1259         ext4_fsblk_t block;
1260         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1261         int ee_len;
1262
1263         if (unlikely(path == NULL)) {
1264                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1265                 return -EIO;
1266         }
1267         depth = path->p_depth;
1268         *phys = 0;
1269
1270         if (depth == 0 && path->p_ext == NULL)
1271                 return 0;
1272
1273         /* usually extent in the path covers blocks smaller
1274          * then *logical, but it can be that extent is the
1275          * first one in the file */
1276
1277         ex = path[depth].p_ext;
1278         ee_len = ext4_ext_get_actual_len(ex);
1279         if (*logical < le32_to_cpu(ex->ee_block)) {
1280                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1281                         EXT4_ERROR_INODE(inode,
1282                                          "first_extent(path[%d].p_hdr) != ex",
1283                                          depth);
1284                         return -EIO;
1285                 }
1286                 while (--depth >= 0) {
1287                         ix = path[depth].p_idx;
1288                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1289                                 EXT4_ERROR_INODE(inode,
1290                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1291                                                  *logical);
1292                                 return -EIO;
1293                         }
1294                 }
1295                 goto found_extent;
1296         }
1297
1298         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1299                 EXT4_ERROR_INODE(inode,
1300                                  "logical %d < ee_block %d + ee_len %d!",
1301                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1302                 return -EIO;
1303         }
1304
1305         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1306                 /* next allocated block in this leaf */
1307                 ex++;
1308                 goto found_extent;
1309         }
1310
1311         /* go up and search for index to the right */
1312         while (--depth >= 0) {
1313                 ix = path[depth].p_idx;
1314                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1315                         goto got_index;
1316         }
1317
1318         /* we've gone up to the root and found no index to the right */
1319         return 0;
1320
1321 got_index:
1322         /* we've found index to the right, let's
1323          * follow it and find the closest allocated
1324          * block to the right */
1325         ix++;
1326         block = ext4_idx_pblock(ix);
1327         while (++depth < path->p_depth) {
1328                 bh = sb_bread(inode->i_sb, block);
1329                 if (bh == NULL)
1330                         return -EIO;
1331                 eh = ext_block_hdr(bh);
1332                 /* subtract from p_depth to get proper eh_depth */
1333                 if (ext4_ext_check(inode, eh, path->p_depth - depth)) {
1334                         put_bh(bh);
1335                         return -EIO;
1336                 }
1337                 ix = EXT_FIRST_INDEX(eh);
1338                 block = ext4_idx_pblock(ix);
1339                 put_bh(bh);
1340         }
1341
1342         bh = sb_bread(inode->i_sb, block);
1343         if (bh == NULL)
1344                 return -EIO;
1345         eh = ext_block_hdr(bh);
1346         if (ext4_ext_check(inode, eh, path->p_depth - depth)) {
1347                 put_bh(bh);
1348                 return -EIO;
1349         }
1350         ex = EXT_FIRST_EXTENT(eh);
1351 found_extent:
1352         *logical = le32_to_cpu(ex->ee_block);
1353         *phys = ext4_ext_pblock(ex);
1354         *ret_ex = ex;
1355         if (bh)
1356                 put_bh(bh);
1357         return 0;
1358 }
1359
1360 /*
1361  * ext4_ext_next_allocated_block:
1362  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1363  * NOTE: it considers block number from index entry as
1364  * allocated block. Thus, index entries have to be consistent
1365  * with leaves.
1366  */
1367 static ext4_lblk_t
1368 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1369 {
1370         int depth;
1371
1372         BUG_ON(path == NULL);
1373         depth = path->p_depth;
1374
1375         if (depth == 0 && path->p_ext == NULL)
1376                 return EXT_MAX_BLOCKS;
1377
1378         while (depth >= 0) {
1379                 if (depth == path->p_depth) {
1380                         /* leaf */
1381                         if (path[depth].p_ext &&
1382                                 path[depth].p_ext !=
1383                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1384                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1385                 } else {
1386                         /* index */
1387                         if (path[depth].p_idx !=
1388                                         EXT_LAST_INDEX(path[depth].p_hdr))
1389                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1390                 }
1391                 depth--;
1392         }
1393
1394         return EXT_MAX_BLOCKS;
1395 }
1396
1397 /*
1398  * ext4_ext_next_leaf_block:
1399  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1400  */
1401 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1402 {
1403         int depth;
1404
1405         BUG_ON(path == NULL);
1406         depth = path->p_depth;
1407
1408         /* zero-tree has no leaf blocks at all */
1409         if (depth == 0)
1410                 return EXT_MAX_BLOCKS;
1411
1412         /* go to index block */
1413         depth--;
1414
1415         while (depth >= 0) {
1416                 if (path[depth].p_idx !=
1417                                 EXT_LAST_INDEX(path[depth].p_hdr))
1418                         return (ext4_lblk_t)
1419                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1420                 depth--;
1421         }
1422
1423         return EXT_MAX_BLOCKS;
1424 }
1425
1426 /*
1427  * ext4_ext_correct_indexes:
1428  * if leaf gets modified and modified extent is first in the leaf,
1429  * then we have to correct all indexes above.
1430  * TODO: do we need to correct tree in all cases?
1431  */
1432 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1433                                 struct ext4_ext_path *path)
1434 {
1435         struct ext4_extent_header *eh;
1436         int depth = ext_depth(inode);
1437         struct ext4_extent *ex;
1438         __le32 border;
1439         int k, err = 0;
1440
1441         eh = path[depth].p_hdr;
1442         ex = path[depth].p_ext;
1443
1444         if (unlikely(ex == NULL || eh == NULL)) {
1445                 EXT4_ERROR_INODE(inode,
1446                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1447                 return -EIO;
1448         }
1449
1450         if (depth == 0) {
1451                 /* there is no tree at all */
1452                 return 0;
1453         }
1454
1455         if (ex != EXT_FIRST_EXTENT(eh)) {
1456                 /* we correct tree if first leaf got modified only */
1457                 return 0;
1458         }
1459
1460         /*
1461          * TODO: we need correction if border is smaller than current one
1462          */
1463         k = depth - 1;
1464         border = path[depth].p_ext->ee_block;
1465         err = ext4_ext_get_access(handle, inode, path + k);
1466         if (err)
1467                 return err;
1468         path[k].p_idx->ei_block = border;
1469         err = ext4_ext_dirty(handle, inode, path + k);
1470         if (err)
1471                 return err;
1472
1473         while (k--) {
1474                 /* change all left-side indexes */
1475                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1476                         break;
1477                 err = ext4_ext_get_access(handle, inode, path + k);
1478                 if (err)
1479                         break;
1480                 path[k].p_idx->ei_block = border;
1481                 err = ext4_ext_dirty(handle, inode, path + k);
1482                 if (err)
1483                         break;
1484         }
1485
1486         return err;
1487 }
1488
1489 int
1490 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1491                                 struct ext4_extent *ex2)
1492 {
1493         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1494
1495         /*
1496          * Make sure that either both extents are uninitialized, or
1497          * both are _not_.
1498          */
1499         if (ext4_ext_is_uninitialized(ex1) ^ ext4_ext_is_uninitialized(ex2))
1500                 return 0;
1501
1502         if (ext4_ext_is_uninitialized(ex1))
1503                 max_len = EXT_UNINIT_MAX_LEN;
1504         else
1505                 max_len = EXT_INIT_MAX_LEN;
1506
1507         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1508         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1509
1510         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1511                         le32_to_cpu(ex2->ee_block))
1512                 return 0;
1513
1514         /*
1515          * To allow future support for preallocated extents to be added
1516          * as an RO_COMPAT feature, refuse to merge to extents if
1517          * this can result in the top bit of ee_len being set.
1518          */
1519         if (ext1_ee_len + ext2_ee_len > max_len)
1520                 return 0;
1521 #ifdef AGGRESSIVE_TEST
1522         if (ext1_ee_len >= 4)
1523                 return 0;
1524 #endif
1525
1526         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1527                 return 1;
1528         return 0;
1529 }
1530
1531 /*
1532  * This function tries to merge the "ex" extent to the next extent in the tree.
1533  * It always tries to merge towards right. If you want to merge towards
1534  * left, pass "ex - 1" as argument instead of "ex".
1535  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1536  * 1 if they got merged.
1537  */
1538 static int ext4_ext_try_to_merge_right(struct inode *inode,
1539                                  struct ext4_ext_path *path,
1540                                  struct ext4_extent *ex)
1541 {
1542         struct ext4_extent_header *eh;
1543         unsigned int depth, len;
1544         int merge_done = 0;
1545         int uninitialized = 0;
1546
1547         depth = ext_depth(inode);
1548         BUG_ON(path[depth].p_hdr == NULL);
1549         eh = path[depth].p_hdr;
1550
1551         while (ex < EXT_LAST_EXTENT(eh)) {
1552                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1553                         break;
1554                 /* merge with next extent! */
1555                 if (ext4_ext_is_uninitialized(ex))
1556                         uninitialized = 1;
1557                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1558                                 + ext4_ext_get_actual_len(ex + 1));
1559                 if (uninitialized)
1560                         ext4_ext_mark_uninitialized(ex);
1561
1562                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1563                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1564                                 * sizeof(struct ext4_extent);
1565                         memmove(ex + 1, ex + 2, len);
1566                 }
1567                 le16_add_cpu(&eh->eh_entries, -1);
1568                 merge_done = 1;
1569                 WARN_ON(eh->eh_entries == 0);
1570                 if (!eh->eh_entries)
1571                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1572         }
1573
1574         return merge_done;
1575 }
1576
1577 /*
1578  * This function tries to merge the @ex extent to neighbours in the tree.
1579  * return 1 if merge left else 0.
1580  */
1581 static int ext4_ext_try_to_merge(struct inode *inode,
1582                                   struct ext4_ext_path *path,
1583                                   struct ext4_extent *ex) {
1584         struct ext4_extent_header *eh;
1585         unsigned int depth;
1586         int merge_done = 0;
1587         int ret = 0;
1588
1589         depth = ext_depth(inode);
1590         BUG_ON(path[depth].p_hdr == NULL);
1591         eh = path[depth].p_hdr;
1592
1593         if (ex > EXT_FIRST_EXTENT(eh))
1594                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1595
1596         if (!merge_done)
1597                 ret = ext4_ext_try_to_merge_right(inode, path, ex);
1598
1599         return ret;
1600 }
1601
1602 /*
1603  * check if a portion of the "newext" extent overlaps with an
1604  * existing extent.
1605  *
1606  * If there is an overlap discovered, it updates the length of the newext
1607  * such that there will be no overlap, and then returns 1.
1608  * If there is no overlap found, it returns 0.
1609  */
1610 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1611                                            struct inode *inode,
1612                                            struct ext4_extent *newext,
1613                                            struct ext4_ext_path *path)
1614 {
1615         ext4_lblk_t b1, b2;
1616         unsigned int depth, len1;
1617         unsigned int ret = 0;
1618
1619         b1 = le32_to_cpu(newext->ee_block);
1620         len1 = ext4_ext_get_actual_len(newext);
1621         depth = ext_depth(inode);
1622         if (!path[depth].p_ext)
1623                 goto out;
1624         b2 = le32_to_cpu(path[depth].p_ext->ee_block);
1625         b2 &= ~(sbi->s_cluster_ratio - 1);
1626
1627         /*
1628          * get the next allocated block if the extent in the path
1629          * is before the requested block(s)
1630          */
1631         if (b2 < b1) {
1632                 b2 = ext4_ext_next_allocated_block(path);
1633                 if (b2 == EXT_MAX_BLOCKS)
1634                         goto out;
1635                 b2 &= ~(sbi->s_cluster_ratio - 1);
1636         }
1637
1638         /* check for wrap through zero on extent logical start block*/
1639         if (b1 + len1 < b1) {
1640                 len1 = EXT_MAX_BLOCKS - b1;
1641                 newext->ee_len = cpu_to_le16(len1);
1642                 ret = 1;
1643         }
1644
1645         /* check for overlap */
1646         if (b1 + len1 > b2) {
1647                 newext->ee_len = cpu_to_le16(b2 - b1);
1648                 ret = 1;
1649         }
1650 out:
1651         return ret;
1652 }
1653
1654 /*
1655  * ext4_ext_insert_extent:
1656  * tries to merge requsted extent into the existing extent or
1657  * inserts requested extent as new one into the tree,
1658  * creating new leaf in the no-space case.
1659  */
1660 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1661                                 struct ext4_ext_path *path,
1662                                 struct ext4_extent *newext, int flag)
1663 {
1664         struct ext4_extent_header *eh;
1665         struct ext4_extent *ex, *fex;
1666         struct ext4_extent *nearex; /* nearest extent */
1667         struct ext4_ext_path *npath = NULL;
1668         int depth, len, err;
1669         ext4_lblk_t next;
1670         unsigned uninitialized = 0;
1671         int flags = 0;
1672
1673         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1674                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1675                 return -EIO;
1676         }
1677         depth = ext_depth(inode);
1678         ex = path[depth].p_ext;
1679         if (unlikely(path[depth].p_hdr == NULL)) {
1680                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1681                 return -EIO;
1682         }
1683
1684         /* try to insert block into found extent and return */
1685         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)
1686                 && ext4_can_extents_be_merged(inode, ex, newext)) {
1687                 ext_debug("append [%d]%d block to %u:[%d]%d (from %llu)\n",
1688                           ext4_ext_is_uninitialized(newext),
1689                           ext4_ext_get_actual_len(newext),
1690                           le32_to_cpu(ex->ee_block),
1691                           ext4_ext_is_uninitialized(ex),
1692                           ext4_ext_get_actual_len(ex),
1693                           ext4_ext_pblock(ex));
1694                 err = ext4_ext_get_access(handle, inode, path + depth);
1695                 if (err)
1696                         return err;
1697
1698                 /*
1699                  * ext4_can_extents_be_merged should have checked that either
1700                  * both extents are uninitialized, or both aren't. Thus we
1701                  * need to check only one of them here.
1702                  */
1703                 if (ext4_ext_is_uninitialized(ex))
1704                         uninitialized = 1;
1705                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1706                                         + ext4_ext_get_actual_len(newext));
1707                 if (uninitialized)
1708                         ext4_ext_mark_uninitialized(ex);
1709                 eh = path[depth].p_hdr;
1710                 nearex = ex;
1711                 goto merge;
1712         }
1713
1714         depth = ext_depth(inode);
1715         eh = path[depth].p_hdr;
1716         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1717                 goto has_space;
1718
1719         /* probably next leaf has space for us? */
1720         fex = EXT_LAST_EXTENT(eh);
1721         next = EXT_MAX_BLOCKS;
1722         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1723                 next = ext4_ext_next_leaf_block(path);
1724         if (next != EXT_MAX_BLOCKS) {
1725                 ext_debug("next leaf block - %u\n", next);
1726                 BUG_ON(npath != NULL);
1727                 npath = ext4_ext_find_extent(inode, next, NULL);
1728                 if (IS_ERR(npath))
1729                         return PTR_ERR(npath);
1730                 BUG_ON(npath->p_depth != path->p_depth);
1731                 eh = npath[depth].p_hdr;
1732                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1733                         ext_debug("next leaf isn't full(%d)\n",
1734                                   le16_to_cpu(eh->eh_entries));
1735                         path = npath;
1736                         goto has_space;
1737                 }
1738                 ext_debug("next leaf has no free space(%d,%d)\n",
1739                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1740         }
1741
1742         /*
1743          * There is no free space in the found leaf.
1744          * We're gonna add a new leaf in the tree.
1745          */
1746         if (flag & EXT4_GET_BLOCKS_PUNCH_OUT_EXT)
1747                 flags = EXT4_MB_USE_ROOT_BLOCKS;
1748         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1749         if (err)
1750                 goto cleanup;
1751         depth = ext_depth(inode);
1752         eh = path[depth].p_hdr;
1753
1754 has_space:
1755         nearex = path[depth].p_ext;
1756
1757         err = ext4_ext_get_access(handle, inode, path + depth);
1758         if (err)
1759                 goto cleanup;
1760
1761         if (!nearex) {
1762                 /* there is no extent in this leaf, create first one */
1763                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1764                                 le32_to_cpu(newext->ee_block),
1765                                 ext4_ext_pblock(newext),
1766                                 ext4_ext_is_uninitialized(newext),
1767                                 ext4_ext_get_actual_len(newext));
1768                 nearex = EXT_FIRST_EXTENT(eh);
1769         } else {
1770                 if (le32_to_cpu(newext->ee_block)
1771                            > le32_to_cpu(nearex->ee_block)) {
1772                         /* Insert after */
1773                         ext_debug("insert %u:%llu:[%d]%d before: "
1774                                         "nearest %p\n",
1775                                         le32_to_cpu(newext->ee_block),
1776                                         ext4_ext_pblock(newext),
1777                                         ext4_ext_is_uninitialized(newext),
1778                                         ext4_ext_get_actual_len(newext),
1779                                         nearex);
1780                         nearex++;
1781                 } else {
1782                         /* Insert before */
1783                         BUG_ON(newext->ee_block == nearex->ee_block);
1784                         ext_debug("insert %u:%llu:[%d]%d after: "
1785                                         "nearest %p\n",
1786                                         le32_to_cpu(newext->ee_block),
1787                                         ext4_ext_pblock(newext),
1788                                         ext4_ext_is_uninitialized(newext),
1789                                         ext4_ext_get_actual_len(newext),
1790                                         nearex);
1791                 }
1792                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1793                 if (len > 0) {
1794                         ext_debug("insert %u:%llu:[%d]%d: "
1795                                         "move %d extents from 0x%p to 0x%p\n",
1796                                         le32_to_cpu(newext->ee_block),
1797                                         ext4_ext_pblock(newext),
1798                                         ext4_ext_is_uninitialized(newext),
1799                                         ext4_ext_get_actual_len(newext),
1800                                         len, nearex, nearex + 1);
1801                         memmove(nearex + 1, nearex,
1802                                 len * sizeof(struct ext4_extent));
1803                 }
1804         }
1805
1806         le16_add_cpu(&eh->eh_entries, 1);
1807         path[depth].p_ext = nearex;
1808         nearex->ee_block = newext->ee_block;
1809         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
1810         nearex->ee_len = newext->ee_len;
1811
1812 merge:
1813         /* try to merge extents to the right */
1814         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
1815                 ext4_ext_try_to_merge(inode, path, nearex);
1816
1817         /* try to merge extents to the left */
1818
1819         /* time to correct all indexes above */
1820         err = ext4_ext_correct_indexes(handle, inode, path);
1821         if (err)
1822                 goto cleanup;
1823
1824         err = ext4_ext_dirty(handle, inode, path + depth);
1825
1826 cleanup:
1827         if (npath) {
1828                 ext4_ext_drop_refs(npath);
1829                 kfree(npath);
1830         }
1831         ext4_ext_invalidate_cache(inode);
1832         return err;
1833 }
1834
1835 static int ext4_ext_walk_space(struct inode *inode, ext4_lblk_t block,
1836                                ext4_lblk_t num, ext_prepare_callback func,
1837                                void *cbdata)
1838 {
1839         struct ext4_ext_path *path = NULL;
1840         struct ext4_ext_cache cbex;
1841         struct ext4_extent *ex;
1842         ext4_lblk_t next, start = 0, end = 0;
1843         ext4_lblk_t last = block + num;
1844         int depth, exists, err = 0;
1845
1846         BUG_ON(func == NULL);
1847         BUG_ON(inode == NULL);
1848
1849         while (block < last && block != EXT_MAX_BLOCKS) {
1850                 num = last - block;
1851                 /* find extent for this block */
1852                 down_read(&EXT4_I(inode)->i_data_sem);
1853                 path = ext4_ext_find_extent(inode, block, path);
1854                 up_read(&EXT4_I(inode)->i_data_sem);
1855                 if (IS_ERR(path)) {
1856                         err = PTR_ERR(path);
1857                         path = NULL;
1858                         break;
1859                 }
1860
1861                 depth = ext_depth(inode);
1862                 if (unlikely(path[depth].p_hdr == NULL)) {
1863                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1864                         err = -EIO;
1865                         break;
1866                 }
1867                 ex = path[depth].p_ext;
1868                 next = ext4_ext_next_allocated_block(path);
1869
1870                 exists = 0;
1871                 if (!ex) {
1872                         /* there is no extent yet, so try to allocate
1873                          * all requested space */
1874                         start = block;
1875                         end = block + num;
1876                 } else if (le32_to_cpu(ex->ee_block) > block) {
1877                         /* need to allocate space before found extent */
1878                         start = block;
1879                         end = le32_to_cpu(ex->ee_block);
1880                         if (block + num < end)
1881                                 end = block + num;
1882                 } else if (block >= le32_to_cpu(ex->ee_block)
1883                                         + ext4_ext_get_actual_len(ex)) {
1884                         /* need to allocate space after found extent */
1885                         start = block;
1886                         end = block + num;
1887                         if (end >= next)
1888                                 end = next;
1889                 } else if (block >= le32_to_cpu(ex->ee_block)) {
1890                         /*
1891                          * some part of requested space is covered
1892                          * by found extent
1893                          */
1894                         start = block;
1895                         end = le32_to_cpu(ex->ee_block)
1896                                 + ext4_ext_get_actual_len(ex);
1897                         if (block + num < end)
1898                                 end = block + num;
1899                         exists = 1;
1900                 } else {
1901                         BUG();
1902                 }
1903                 BUG_ON(end <= start);
1904
1905                 if (!exists) {
1906                         cbex.ec_block = start;
1907                         cbex.ec_len = end - start;
1908                         cbex.ec_start = 0;
1909                 } else {
1910                         cbex.ec_block = le32_to_cpu(ex->ee_block);
1911                         cbex.ec_len = ext4_ext_get_actual_len(ex);
1912                         cbex.ec_start = ext4_ext_pblock(ex);
1913                 }
1914
1915                 if (unlikely(cbex.ec_len == 0)) {
1916                         EXT4_ERROR_INODE(inode, "cbex.ec_len == 0");
1917                         err = -EIO;
1918                         break;
1919                 }
1920                 err = func(inode, next, &cbex, ex, cbdata);
1921                 ext4_ext_drop_refs(path);
1922
1923                 if (err < 0)
1924                         break;
1925
1926                 if (err == EXT_REPEAT)
1927                         continue;
1928                 else if (err == EXT_BREAK) {
1929                         err = 0;
1930                         break;
1931                 }
1932
1933                 if (ext_depth(inode) != depth) {
1934                         /* depth was changed. we have to realloc path */
1935                         kfree(path);
1936                         path = NULL;
1937                 }
1938
1939                 block = cbex.ec_block + cbex.ec_len;
1940         }
1941
1942         if (path) {
1943                 ext4_ext_drop_refs(path);
1944                 kfree(path);
1945         }
1946
1947         return err;
1948 }
1949
1950 static void
1951 ext4_ext_put_in_cache(struct inode *inode, ext4_lblk_t block,
1952                         __u32 len, ext4_fsblk_t start)
1953 {
1954         struct ext4_ext_cache *cex;
1955         BUG_ON(len == 0);
1956         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1957         trace_ext4_ext_put_in_cache(inode, block, len, start);
1958         cex = &EXT4_I(inode)->i_cached_extent;
1959         cex->ec_block = block;
1960         cex->ec_len = len;
1961         cex->ec_start = start;
1962         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1963 }
1964
1965 /*
1966  * ext4_ext_put_gap_in_cache:
1967  * calculate boundaries of the gap that the requested block fits into
1968  * and cache this gap
1969  */
1970 static void
1971 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
1972                                 ext4_lblk_t block)
1973 {
1974         int depth = ext_depth(inode);
1975         unsigned long len;
1976         ext4_lblk_t lblock;
1977         struct ext4_extent *ex;
1978
1979         ex = path[depth].p_ext;
1980         if (ex == NULL) {
1981                 /* there is no extent yet, so gap is [0;-] */
1982                 lblock = 0;
1983                 len = EXT_MAX_BLOCKS;
1984                 ext_debug("cache gap(whole file):");
1985         } else if (block < le32_to_cpu(ex->ee_block)) {
1986                 lblock = block;
1987                 len = le32_to_cpu(ex->ee_block) - block;
1988                 ext_debug("cache gap(before): %u [%u:%u]",
1989                                 block,
1990                                 le32_to_cpu(ex->ee_block),
1991                                  ext4_ext_get_actual_len(ex));
1992         } else if (block >= le32_to_cpu(ex->ee_block)
1993                         + ext4_ext_get_actual_len(ex)) {
1994                 ext4_lblk_t next;
1995                 lblock = le32_to_cpu(ex->ee_block)
1996                         + ext4_ext_get_actual_len(ex);
1997
1998                 next = ext4_ext_next_allocated_block(path);
1999                 ext_debug("cache gap(after): [%u:%u] %u",
2000                                 le32_to_cpu(ex->ee_block),
2001                                 ext4_ext_get_actual_len(ex),
2002                                 block);
2003                 BUG_ON(next == lblock);
2004                 len = next - lblock;
2005         } else {
2006                 lblock = len = 0;
2007                 BUG();
2008         }
2009
2010         ext_debug(" -> %u:%lu\n", lblock, len);
2011         ext4_ext_put_in_cache(inode, lblock, len, 0);
2012 }
2013
2014 /*
2015  * ext4_ext_check_cache()
2016  * Checks to see if the given block is in the cache.
2017  * If it is, the cached extent is stored in the given
2018  * cache extent pointer.  If the cached extent is a hole,
2019  * this routine should be used instead of
2020  * ext4_ext_in_cache if the calling function needs to
2021  * know the size of the hole.
2022  *
2023  * @inode: The files inode
2024  * @block: The block to look for in the cache
2025  * @ex:    Pointer where the cached extent will be stored
2026  *         if it contains block
2027  *
2028  * Return 0 if cache is invalid; 1 if the cache is valid
2029  */
2030 static int ext4_ext_check_cache(struct inode *inode, ext4_lblk_t block,
2031         struct ext4_ext_cache *ex){
2032         struct ext4_ext_cache *cex;
2033         struct ext4_sb_info *sbi;
2034         int ret = 0;
2035
2036         /*
2037          * We borrow i_block_reservation_lock to protect i_cached_extent
2038          */
2039         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
2040         cex = &EXT4_I(inode)->i_cached_extent;
2041         sbi = EXT4_SB(inode->i_sb);
2042
2043         /* has cache valid data? */
2044         if (cex->ec_len == 0)
2045                 goto errout;
2046
2047         if (in_range(block, cex->ec_block, cex->ec_len)) {
2048                 memcpy(ex, cex, sizeof(struct ext4_ext_cache));
2049                 ext_debug("%u cached by %u:%u:%llu\n",
2050                                 block,
2051                                 cex->ec_block, cex->ec_len, cex->ec_start);
2052                 ret = 1;
2053         }
2054 errout:
2055         trace_ext4_ext_in_cache(inode, block, ret);
2056         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
2057         return ret;
2058 }
2059
2060 /*
2061  * ext4_ext_in_cache()
2062  * Checks to see if the given block is in the cache.
2063  * If it is, the cached extent is stored in the given
2064  * extent pointer.
2065  *
2066  * @inode: The files inode
2067  * @block: The block to look for in the cache
2068  * @ex:    Pointer where the cached extent will be stored
2069  *         if it contains block
2070  *
2071  * Return 0 if cache is invalid; 1 if the cache is valid
2072  */
2073 static int
2074 ext4_ext_in_cache(struct inode *inode, ext4_lblk_t block,
2075                         struct ext4_extent *ex)
2076 {
2077         struct ext4_ext_cache cex;
2078         int ret = 0;
2079
2080         if (ext4_ext_check_cache(inode, block, &cex)) {
2081                 ex->ee_block = cpu_to_le32(cex.ec_block);
2082                 ext4_ext_store_pblock(ex, cex.ec_start);
2083                 ex->ee_len = cpu_to_le16(cex.ec_len);
2084                 ret = 1;
2085         }
2086
2087         return ret;
2088 }
2089
2090
2091 /*
2092  * ext4_ext_rm_idx:
2093  * removes index from the index block.
2094  */
2095 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2096                         struct ext4_ext_path *path, int depth)
2097 {
2098         int err;
2099         ext4_fsblk_t leaf;
2100
2101         /* free index block */
2102         depth--;
2103         path = path + depth;
2104         leaf = ext4_idx_pblock(path->p_idx);
2105         if (unlikely(path->p_hdr->eh_entries == 0)) {
2106                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2107                 return -EIO;
2108         }
2109         err = ext4_ext_get_access(handle, inode, path);
2110         if (err)
2111                 return err;
2112
2113         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2114                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2115                 len *= sizeof(struct ext4_extent_idx);
2116                 memmove(path->p_idx, path->p_idx + 1, len);
2117         }
2118
2119         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2120         err = ext4_ext_dirty(handle, inode, path);
2121         if (err)
2122                 return err;
2123         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2124         trace_ext4_ext_rm_idx(inode, leaf);
2125
2126         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2127                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2128
2129         while (--depth >= 0) {
2130                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2131                         break;
2132                 path--;
2133                 err = ext4_ext_get_access(handle, inode, path);
2134                 if (err)
2135                         break;
2136                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2137                 err = ext4_ext_dirty(handle, inode, path);
2138                 if (err)
2139                         break;
2140         }
2141         return err;
2142 }
2143
2144 /*
2145  * ext4_ext_calc_credits_for_single_extent:
2146  * This routine returns max. credits that needed to insert an extent
2147  * to the extent tree.
2148  * When pass the actual path, the caller should calculate credits
2149  * under i_data_sem.
2150  */
2151 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2152                                                 struct ext4_ext_path *path)
2153 {
2154         if (path) {
2155                 int depth = ext_depth(inode);
2156                 int ret = 0;
2157
2158                 /* probably there is space in leaf? */
2159                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2160                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2161
2162                         /*
2163                          *  There are some space in the leaf tree, no
2164                          *  need to account for leaf block credit
2165                          *
2166                          *  bitmaps and block group descriptor blocks
2167                          *  and other metadata blocks still need to be
2168                          *  accounted.
2169                          */
2170                         /* 1 bitmap, 1 block group descriptor */
2171                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2172                         return ret;
2173                 }
2174         }
2175
2176         return ext4_chunk_trans_blocks(inode, nrblocks);
2177 }
2178
2179 /*
2180  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2181  *
2182  * if nrblocks are fit in a single extent (chunk flag is 1), then
2183  * in the worse case, each tree level index/leaf need to be changed
2184  * if the tree split due to insert a new extent, then the old tree
2185  * index/leaf need to be updated too
2186  *
2187  * If the nrblocks are discontiguous, they could cause
2188  * the whole tree split more than once, but this is really rare.
2189  */
2190 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2191 {
2192         int index;
2193         int depth = ext_depth(inode);
2194
2195         if (chunk)
2196                 index = depth * 2;
2197         else
2198                 index = depth * 3;
2199
2200         return index;
2201 }
2202
2203 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2204                               struct ext4_extent *ex,
2205                               ext4_fsblk_t *partial_cluster,
2206                               ext4_lblk_t from, ext4_lblk_t to)
2207 {
2208         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2209         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2210         ext4_fsblk_t pblk;
2211         int flags = EXT4_FREE_BLOCKS_FORGET;
2212
2213         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2214                 flags |= EXT4_FREE_BLOCKS_METADATA;
2215         /*
2216          * For bigalloc file systems, we never free a partial cluster
2217          * at the beginning of the extent.  Instead, we make a note
2218          * that we tried freeing the cluster, and check to see if we
2219          * need to free it on a subsequent call to ext4_remove_blocks,
2220          * or at the end of the ext4_truncate() operation.
2221          */
2222         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2223
2224         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2225         /*
2226          * If we have a partial cluster, and it's different from the
2227          * cluster of the last block, we need to explicitly free the
2228          * partial cluster here.
2229          */
2230         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2231         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2232                 ext4_free_blocks(handle, inode, NULL,
2233                                  EXT4_C2B(sbi, *partial_cluster),
2234                                  sbi->s_cluster_ratio, flags);
2235                 *partial_cluster = 0;
2236         }
2237
2238 #ifdef EXTENTS_STATS
2239         {
2240                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2241                 spin_lock(&sbi->s_ext_stats_lock);
2242                 sbi->s_ext_blocks += ee_len;
2243                 sbi->s_ext_extents++;
2244                 if (ee_len < sbi->s_ext_min)
2245                         sbi->s_ext_min = ee_len;
2246                 if (ee_len > sbi->s_ext_max)
2247                         sbi->s_ext_max = ee_len;
2248                 if (ext_depth(inode) > sbi->s_depth_max)
2249                         sbi->s_depth_max = ext_depth(inode);
2250                 spin_unlock(&sbi->s_ext_stats_lock);
2251         }
2252 #endif
2253         if (from >= le32_to_cpu(ex->ee_block)
2254             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2255                 /* tail removal */
2256                 ext4_lblk_t num;
2257
2258                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2259                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2260                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2261                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2262                 /*
2263                  * If the block range to be freed didn't start at the
2264                  * beginning of a cluster, and we removed the entire
2265                  * extent, save the partial cluster here, since we
2266                  * might need to delete if we determine that the
2267                  * truncate operation has removed all of the blocks in
2268                  * the cluster.
2269                  */
2270                 if (pblk & (sbi->s_cluster_ratio - 1) &&
2271                     (ee_len == num))
2272                         *partial_cluster = EXT4_B2C(sbi, pblk);
2273                 else
2274                         *partial_cluster = 0;
2275         } else if (from == le32_to_cpu(ex->ee_block)
2276                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2277                 /* head removal */
2278                 ext4_lblk_t num;
2279                 ext4_fsblk_t start;
2280
2281                 num = to - from;
2282                 start = ext4_ext_pblock(ex);
2283
2284                 ext_debug("free first %u blocks starting %llu\n", num, start);
2285                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2286
2287         } else {
2288                 printk(KERN_INFO "strange request: removal(2) "
2289                                 "%u-%u from %u:%u\n",
2290                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2291         }
2292         return 0;
2293 }
2294
2295
2296 /*
2297  * ext4_ext_rm_leaf() Removes the extents associated with the
2298  * blocks appearing between "start" and "end", and splits the extents
2299  * if "start" and "end" appear in the same extent
2300  *
2301  * @handle: The journal handle
2302  * @inode:  The files inode
2303  * @path:   The path to the leaf
2304  * @start:  The first block to remove
2305  * @end:   The last block to remove
2306  */
2307 static int
2308 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2309                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2310                  ext4_lblk_t start, ext4_lblk_t end)
2311 {
2312         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2313         int err = 0, correct_index = 0;
2314         int depth = ext_depth(inode), credits;
2315         struct ext4_extent_header *eh;
2316         ext4_lblk_t a, b;
2317         unsigned num;
2318         ext4_lblk_t ex_ee_block;
2319         unsigned short ex_ee_len;
2320         unsigned uninitialized = 0;
2321         struct ext4_extent *ex;
2322
2323         /* the header must be checked already in ext4_ext_remove_space() */
2324         ext_debug("truncate since %u in leaf\n", start);
2325         if (!path[depth].p_hdr)
2326                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2327         eh = path[depth].p_hdr;
2328         if (unlikely(path[depth].p_hdr == NULL)) {
2329                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2330                 return -EIO;
2331         }
2332         /* find where to start removing */
2333         ex = EXT_LAST_EXTENT(eh);
2334
2335         ex_ee_block = le32_to_cpu(ex->ee_block);
2336         ex_ee_len = ext4_ext_get_actual_len(ex);
2337
2338         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2339
2340         while (ex >= EXT_FIRST_EXTENT(eh) &&
2341                         ex_ee_block + ex_ee_len > start) {
2342
2343                 if (ext4_ext_is_uninitialized(ex))
2344                         uninitialized = 1;
2345                 else
2346                         uninitialized = 0;
2347
2348                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2349                          uninitialized, ex_ee_len);
2350                 path[depth].p_ext = ex;
2351
2352                 a = ex_ee_block > start ? ex_ee_block : start;
2353                 b = ex_ee_block+ex_ee_len - 1 < end ?
2354                         ex_ee_block+ex_ee_len - 1 : end;
2355
2356                 ext_debug("  border %u:%u\n", a, b);
2357
2358                 /* If this extent is beyond the end of the hole, skip it */
2359                 if (end <= ex_ee_block) {
2360                         ex--;
2361                         ex_ee_block = le32_to_cpu(ex->ee_block);
2362                         ex_ee_len = ext4_ext_get_actual_len(ex);
2363                         continue;
2364                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2365                         EXT4_ERROR_INODE(inode,"  bad truncate %u:%u\n",
2366                                          start, end);
2367                         err = -EIO;
2368                         goto out;
2369                 } else if (a != ex_ee_block) {
2370                         /* remove tail of the extent */
2371                         num = a - ex_ee_block;
2372                 } else {
2373                         /* remove whole extent: excellent! */
2374                         num = 0;
2375                 }
2376                 /*
2377                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2378                  * descriptor) for each block group; assume two block
2379                  * groups plus ex_ee_len/blocks_per_block_group for
2380                  * the worst case
2381                  */
2382                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2383                 if (ex == EXT_FIRST_EXTENT(eh)) {
2384                         correct_index = 1;
2385                         credits += (ext_depth(inode)) + 1;
2386                 }
2387                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2388
2389                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2390                 if (err)
2391                         goto out;
2392
2393                 err = ext4_ext_get_access(handle, inode, path + depth);
2394                 if (err)
2395                         goto out;
2396
2397                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2398                                          a, b);
2399                 if (err)
2400                         goto out;
2401
2402                 if (num == 0)
2403                         /* this extent is removed; mark slot entirely unused */
2404                         ext4_ext_store_pblock(ex, 0);
2405
2406                 ex->ee_len = cpu_to_le16(num);
2407                 /*
2408                  * Do not mark uninitialized if all the blocks in the
2409                  * extent have been removed.
2410                  */
2411                 if (uninitialized && num)
2412                         ext4_ext_mark_uninitialized(ex);
2413                 /*
2414                  * If the extent was completely released,
2415                  * we need to remove it from the leaf
2416                  */
2417                 if (num == 0) {
2418                         if (end != EXT_MAX_BLOCKS - 1) {
2419                                 /*
2420                                  * For hole punching, we need to scoot all the
2421                                  * extents up when an extent is removed so that
2422                                  * we dont have blank extents in the middle
2423                                  */
2424                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2425                                         sizeof(struct ext4_extent));
2426
2427                                 /* Now get rid of the one at the end */
2428                                 memset(EXT_LAST_EXTENT(eh), 0,
2429                                         sizeof(struct ext4_extent));
2430                         }
2431                         le16_add_cpu(&eh->eh_entries, -1);
2432                 } else
2433                         *partial_cluster = 0;
2434
2435                 err = ext4_ext_dirty(handle, inode, path + depth);
2436                 if (err)
2437                         goto out;
2438
2439                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2440                                 ext4_ext_pblock(ex));
2441                 ex--;
2442                 ex_ee_block = le32_to_cpu(ex->ee_block);
2443                 ex_ee_len = ext4_ext_get_actual_len(ex);
2444         }
2445
2446         if (correct_index && eh->eh_entries)
2447                 err = ext4_ext_correct_indexes(handle, inode, path);
2448
2449         /*
2450          * If there is still a entry in the leaf node, check to see if
2451          * it references the partial cluster.  This is the only place
2452          * where it could; if it doesn't, we can free the cluster.
2453          */
2454         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2455             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2456              *partial_cluster)) {
2457                 int flags = EXT4_FREE_BLOCKS_FORGET;
2458
2459                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2460                         flags |= EXT4_FREE_BLOCKS_METADATA;
2461
2462                 ext4_free_blocks(handle, inode, NULL,
2463                                  EXT4_C2B(sbi, *partial_cluster),
2464                                  sbi->s_cluster_ratio, flags);
2465                 *partial_cluster = 0;
2466         }
2467
2468         /* if this leaf is free, then we should
2469          * remove it from index block above */
2470         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2471                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2472
2473 out:
2474         return err;
2475 }
2476
2477 /*
2478  * ext4_ext_more_to_rm:
2479  * returns 1 if current index has to be freed (even partial)
2480  */
2481 static int
2482 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2483 {
2484         BUG_ON(path->p_idx == NULL);
2485
2486         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2487                 return 0;
2488
2489         /*
2490          * if truncate on deeper level happened, it wasn't partial,
2491          * so we have to consider current index for truncation
2492          */
2493         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2494                 return 0;
2495         return 1;
2496 }
2497
2498 static int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start)
2499 {
2500         struct super_block *sb = inode->i_sb;
2501         int depth = ext_depth(inode);
2502         struct ext4_ext_path *path;
2503         ext4_fsblk_t partial_cluster = 0;
2504         handle_t *handle;
2505         int i, err;
2506
2507         ext_debug("truncate since %u\n", start);
2508
2509         /* probably first extent we're gonna free will be last in block */
2510         handle = ext4_journal_start(inode, depth + 1);
2511         if (IS_ERR(handle))
2512                 return PTR_ERR(handle);
2513
2514 again:
2515         ext4_ext_invalidate_cache(inode);
2516
2517         trace_ext4_ext_remove_space(inode, start, depth);
2518
2519         /*
2520          * We start scanning from right side, freeing all the blocks
2521          * after i_size and walking into the tree depth-wise.
2522          */
2523         depth = ext_depth(inode);
2524         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1), GFP_NOFS);
2525         if (path == NULL) {
2526                 ext4_journal_stop(handle);
2527                 return -ENOMEM;
2528         }
2529         path[0].p_depth = depth;
2530         path[0].p_hdr = ext_inode_hdr(inode);
2531         if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2532                 err = -EIO;
2533                 goto out;
2534         }
2535         i = err = 0;
2536
2537         while (i >= 0 && err == 0) {
2538                 if (i == depth) {
2539                         /* this is leaf block */
2540                         err = ext4_ext_rm_leaf(handle, inode, path,
2541                                                &partial_cluster, start,
2542                                                EXT_MAX_BLOCKS - 1);
2543                         /* root level has p_bh == NULL, brelse() eats this */
2544                         brelse(path[i].p_bh);
2545                         path[i].p_bh = NULL;
2546                         i--;
2547                         continue;
2548                 }
2549
2550                 /* this is index block */
2551                 if (!path[i].p_hdr) {
2552                         ext_debug("initialize header\n");
2553                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2554                 }
2555
2556                 if (!path[i].p_idx) {
2557                         /* this level hasn't been touched yet */
2558                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2559                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2560                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2561                                   path[i].p_hdr,
2562                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2563                 } else {
2564                         /* we were already here, see at next index */
2565                         path[i].p_idx--;
2566                 }
2567
2568                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2569                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2570                                 path[i].p_idx);
2571                 if (ext4_ext_more_to_rm(path + i)) {
2572                         struct buffer_head *bh;
2573                         /* go to the next level */
2574                         ext_debug("move to level %d (block %llu)\n",
2575                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2576                         memset(path + i + 1, 0, sizeof(*path));
2577                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2578                         if (!bh) {
2579                                 /* should we reset i_size? */
2580                                 err = -EIO;
2581                                 break;
2582                         }
2583                         if (WARN_ON(i + 1 > depth)) {
2584                                 err = -EIO;
2585                                 break;
2586                         }
2587                         if (ext4_ext_check(inode, ext_block_hdr(bh),
2588                                                         depth - i - 1)) {
2589                                 err = -EIO;
2590                                 break;
2591                         }
2592                         path[i + 1].p_bh = bh;
2593
2594                         /* save actual number of indexes since this
2595                          * number is changed at the next iteration */
2596                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2597                         i++;
2598                 } else {
2599                         /* we finished processing this index, go up */
2600                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2601                                 /* index is empty, remove it;
2602                                  * handle must be already prepared by the
2603                                  * truncatei_leaf() */
2604                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2605                         }
2606                         /* root level has p_bh == NULL, brelse() eats this */
2607                         brelse(path[i].p_bh);
2608                         path[i].p_bh = NULL;
2609                         i--;
2610                         ext_debug("return to level %d\n", i);
2611                 }
2612         }
2613
2614         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2615                         path->p_hdr->eh_entries);
2616
2617         /* If we still have something in the partial cluster and we have removed
2618          * even the first extent, then we should free the blocks in the partial
2619          * cluster as well. */
2620         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2621                 int flags = EXT4_FREE_BLOCKS_FORGET;
2622
2623                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2624                         flags |= EXT4_FREE_BLOCKS_METADATA;
2625
2626                 ext4_free_blocks(handle, inode, NULL,
2627                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2628                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2629                 partial_cluster = 0;
2630         }
2631
2632         /* TODO: flexible tree reduction should be here */
2633         if (path->p_hdr->eh_entries == 0) {
2634                 /*
2635                  * truncate to zero freed all the tree,
2636                  * so we need to correct eh_depth
2637                  */
2638                 err = ext4_ext_get_access(handle, inode, path);
2639                 if (err == 0) {
2640                         ext_inode_hdr(inode)->eh_depth = 0;
2641                         ext_inode_hdr(inode)->eh_max =
2642                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2643                         err = ext4_ext_dirty(handle, inode, path);
2644                 }
2645         }
2646 out:
2647         ext4_ext_drop_refs(path);
2648         kfree(path);
2649         if (err == -EAGAIN)
2650                 goto again;
2651         ext4_journal_stop(handle);
2652
2653         return err;
2654 }
2655
2656 /*
2657  * called at mount time
2658  */
2659 void ext4_ext_init(struct super_block *sb)
2660 {
2661         /*
2662          * possible initialization would be here
2663          */
2664
2665         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2666 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2667                 printk(KERN_INFO "EXT4-fs: file extents enabled");
2668 #ifdef AGGRESSIVE_TEST
2669                 printk(", aggressive tests");
2670 #endif
2671 #ifdef CHECK_BINSEARCH
2672                 printk(", check binsearch");
2673 #endif
2674 #ifdef EXTENTS_STATS
2675                 printk(", stats");
2676 #endif
2677                 printk("\n");
2678 #endif
2679 #ifdef EXTENTS_STATS
2680                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2681                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2682                 EXT4_SB(sb)->s_ext_max = 0;
2683 #endif
2684         }
2685 }
2686
2687 /*
2688  * called at umount time
2689  */
2690 void ext4_ext_release(struct super_block *sb)
2691 {
2692         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2693                 return;
2694
2695 #ifdef EXTENTS_STATS
2696         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2697                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2698                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2699                         sbi->s_ext_blocks, sbi->s_ext_extents,
2700                         sbi->s_ext_blocks / sbi->s_ext_extents);
2701                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2702                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2703         }
2704 #endif
2705 }
2706
2707 /* FIXME!! we need to try to merge to left or right after zero-out  */
2708 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2709 {
2710         ext4_fsblk_t ee_pblock;
2711         unsigned int ee_len;
2712         int ret;
2713
2714         ee_len    = ext4_ext_get_actual_len(ex);
2715         ee_pblock = ext4_ext_pblock(ex);
2716
2717         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2718         if (ret > 0)
2719                 ret = 0;
2720
2721         return ret;
2722 }
2723
2724 /*
2725  * used by extent splitting.
2726  */
2727 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
2728                                         due to ENOSPC */
2729 #define EXT4_EXT_MARK_UNINIT1   0x2  /* mark first half uninitialized */
2730 #define EXT4_EXT_MARK_UNINIT2   0x4  /* mark second half uninitialized */
2731
2732 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
2733 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
2734
2735 /*
2736  * ext4_split_extent_at() splits an extent at given block.
2737  *
2738  * @handle: the journal handle
2739  * @inode: the file inode
2740  * @path: the path to the extent
2741  * @split: the logical block where the extent is splitted.
2742  * @split_flags: indicates if the extent could be zeroout if split fails, and
2743  *               the states(init or uninit) of new extents.
2744  * @flags: flags used to insert new extent to extent tree.
2745  *
2746  *
2747  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2748  * of which are deterimined by split_flag.
2749  *
2750  * There are two cases:
2751  *  a> the extent are splitted into two extent.
2752  *  b> split is not needed, and just mark the extent.
2753  *
2754  * return 0 on success.
2755  */
2756 static int ext4_split_extent_at(handle_t *handle,
2757                              struct inode *inode,
2758                              struct ext4_ext_path *path,
2759                              ext4_lblk_t split,
2760                              int split_flag,
2761                              int flags)
2762 {
2763         ext4_fsblk_t newblock;
2764         ext4_lblk_t ee_block;
2765         struct ext4_extent *ex, newex, orig_ex;
2766         struct ext4_extent *ex2 = NULL;
2767         unsigned int ee_len, depth;
2768         int err = 0;
2769
2770         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
2771                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
2772
2773         ext_debug("ext4_split_extents_at: inode %lu, logical"
2774                 "block %llu\n", inode->i_ino, (unsigned long long)split);
2775
2776         ext4_ext_show_leaf(inode, path);
2777
2778         depth = ext_depth(inode);
2779         ex = path[depth].p_ext;
2780         ee_block = le32_to_cpu(ex->ee_block);
2781         ee_len = ext4_ext_get_actual_len(ex);
2782         newblock = split - ee_block + ext4_ext_pblock(ex);
2783
2784         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
2785
2786         err = ext4_ext_get_access(handle, inode, path + depth);
2787         if (err)
2788                 goto out;
2789
2790         if (split == ee_block) {
2791                 /*
2792                  * case b: block @split is the block that the extent begins with
2793                  * then we just change the state of the extent, and splitting
2794                  * is not needed.
2795                  */
2796                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
2797                         ext4_ext_mark_uninitialized(ex);
2798                 else
2799                         ext4_ext_mark_initialized(ex);
2800
2801                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
2802                         ext4_ext_try_to_merge(inode, path, ex);
2803
2804                 err = ext4_ext_dirty(handle, inode, path + depth);
2805                 goto out;
2806         }
2807
2808         /* case a */
2809         memcpy(&orig_ex, ex, sizeof(orig_ex));
2810         ex->ee_len = cpu_to_le16(split - ee_block);
2811         if (split_flag & EXT4_EXT_MARK_UNINIT1)
2812                 ext4_ext_mark_uninitialized(ex);
2813
2814         /*
2815          * path may lead to new leaf, not to original leaf any more
2816          * after ext4_ext_insert_extent() returns,
2817          */
2818         err = ext4_ext_dirty(handle, inode, path + depth);
2819         if (err)
2820                 goto fix_extent_len;
2821
2822         ex2 = &newex;
2823         ex2->ee_block = cpu_to_le32(split);
2824         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
2825         ext4_ext_store_pblock(ex2, newblock);
2826         if (split_flag & EXT4_EXT_MARK_UNINIT2)
2827                 ext4_ext_mark_uninitialized(ex2);
2828
2829         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
2830         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
2831                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
2832                         if (split_flag & EXT4_EXT_DATA_VALID1)
2833                                 err = ext4_ext_zeroout(inode, ex2);
2834                         else
2835                                 err = ext4_ext_zeroout(inode, ex);
2836                 } else
2837                         err = ext4_ext_zeroout(inode, &orig_ex);
2838
2839                 if (err)
2840                         goto fix_extent_len;
2841                 /* update the extent length and mark as initialized */
2842                 ex->ee_len = cpu_to_le16(ee_len);
2843                 ext4_ext_try_to_merge(inode, path, ex);
2844                 err = ext4_ext_dirty(handle, inode, path + depth);
2845                 goto out;
2846         } else if (err)
2847                 goto fix_extent_len;
2848
2849 out:
2850         ext4_ext_show_leaf(inode, path);
2851         return err;
2852
2853 fix_extent_len:
2854         ex->ee_len = orig_ex.ee_len;
2855         ext4_ext_dirty(handle, inode, path + depth);
2856         return err;
2857 }
2858
2859 /*
2860  * ext4_split_extents() splits an extent and mark extent which is covered
2861  * by @map as split_flags indicates
2862  *
2863  * It may result in splitting the extent into multiple extents (upto three)
2864  * There are three possibilities:
2865  *   a> There is no split required
2866  *   b> Splits in two extents: Split is happening at either end of the extent
2867  *   c> Splits in three extents: Somone is splitting in middle of the extent
2868  *
2869  */
2870 static int ext4_split_extent(handle_t *handle,
2871                               struct inode *inode,
2872                               struct ext4_ext_path *path,
2873                               struct ext4_map_blocks *map,
2874                               int split_flag,
2875                               int flags)
2876 {
2877         ext4_lblk_t ee_block;
2878         struct ext4_extent *ex;
2879         unsigned int ee_len, depth;
2880         int err = 0;
2881         int uninitialized;
2882         int split_flag1, flags1;
2883
2884         depth = ext_depth(inode);
2885         ex = path[depth].p_ext;
2886         ee_block = le32_to_cpu(ex->ee_block);
2887         ee_len = ext4_ext_get_actual_len(ex);
2888         uninitialized = ext4_ext_is_uninitialized(ex);
2889
2890         if (map->m_lblk + map->m_len < ee_block + ee_len) {
2891                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
2892                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
2893                 if (uninitialized)
2894                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
2895                                        EXT4_EXT_MARK_UNINIT2;
2896                 if (split_flag & EXT4_EXT_DATA_VALID2)
2897                         split_flag1 |= EXT4_EXT_DATA_VALID1;
2898                 err = ext4_split_extent_at(handle, inode, path,
2899                                 map->m_lblk + map->m_len, split_flag1, flags1);
2900                 if (err)
2901                         goto out;
2902         }
2903
2904         ext4_ext_drop_refs(path);
2905         path = ext4_ext_find_extent(inode, map->m_lblk, path);
2906         if (IS_ERR(path))
2907                 return PTR_ERR(path);
2908
2909         if (map->m_lblk >= ee_block) {
2910                 split_flag1 = split_flag & (EXT4_EXT_MAY_ZEROOUT |
2911                                             EXT4_EXT_DATA_VALID2);
2912                 if (uninitialized)
2913                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
2914                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
2915                         split_flag1 |= EXT4_EXT_MARK_UNINIT2;
2916                 err = ext4_split_extent_at(handle, inode, path,
2917                                 map->m_lblk, split_flag1, flags);
2918                 if (err)
2919                         goto out;
2920         }
2921
2922         ext4_ext_show_leaf(inode, path);
2923 out:
2924         return err ? err : map->m_len;
2925 }
2926
2927 #define EXT4_EXT_ZERO_LEN 7
2928 /*
2929  * This function is called by ext4_ext_map_blocks() if someone tries to write
2930  * to an uninitialized extent. It may result in splitting the uninitialized
2931  * extent into multiple extents (up to three - one initialized and two
2932  * uninitialized).
2933  * There are three possibilities:
2934  *   a> There is no split required: Entire extent should be initialized
2935  *   b> Splits in two extents: Write is happening at either end of the extent
2936  *   c> Splits in three extents: Somone is writing in middle of the extent
2937  *
2938  * Pre-conditions:
2939  *  - The extent pointed to by 'path' is uninitialized.
2940  *  - The extent pointed to by 'path' contains a superset
2941  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
2942  *
2943  * Post-conditions on success:
2944  *  - the returned value is the number of blocks beyond map->l_lblk
2945  *    that are allocated and initialized.
2946  *    It is guaranteed to be >= map->m_len.
2947  */
2948 static int ext4_ext_convert_to_initialized(handle_t *handle,
2949                                            struct inode *inode,
2950                                            struct ext4_map_blocks *map,
2951                                            struct ext4_ext_path *path)
2952 {
2953         struct ext4_extent_header *eh;
2954         struct ext4_map_blocks split_map;
2955         struct ext4_extent zero_ex;
2956         struct ext4_extent *ex;
2957         ext4_lblk_t ee_block, eof_block;
2958         unsigned int ee_len, depth;
2959         int allocated;
2960         int err = 0;
2961         int split_flag = 0;
2962
2963         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
2964                 "block %llu, max_blocks %u\n", inode->i_ino,
2965                 (unsigned long long)map->m_lblk, map->m_len);
2966
2967         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
2968                 inode->i_sb->s_blocksize_bits;
2969         if (eof_block < map->m_lblk + map->m_len)
2970                 eof_block = map->m_lblk + map->m_len;
2971
2972         depth = ext_depth(inode);
2973         eh = path[depth].p_hdr;
2974         ex = path[depth].p_ext;
2975         ee_block = le32_to_cpu(ex->ee_block);
2976         ee_len = ext4_ext_get_actual_len(ex);
2977         allocated = ee_len - (map->m_lblk - ee_block);
2978
2979         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
2980
2981         /* Pre-conditions */
2982         BUG_ON(!ext4_ext_is_uninitialized(ex));
2983         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
2984
2985         /*
2986          * Attempt to transfer newly initialized blocks from the currently
2987          * uninitialized extent to its left neighbor. This is much cheaper
2988          * than an insertion followed by a merge as those involve costly
2989          * memmove() calls. This is the common case in steady state for
2990          * workloads doing fallocate(FALLOC_FL_KEEP_SIZE) followed by append
2991          * writes.
2992          *
2993          * Limitations of the current logic:
2994          *  - L1: we only deal with writes at the start of the extent.
2995          *    The approach could be extended to writes at the end
2996          *    of the extent but this scenario was deemed less common.
2997          *  - L2: we do not deal with writes covering the whole extent.
2998          *    This would require removing the extent if the transfer
2999          *    is possible.
3000          *  - L3: we only attempt to merge with an extent stored in the
3001          *    same extent tree node.
3002          */
3003         if ((map->m_lblk == ee_block) &&        /*L1*/
3004                 (map->m_len < ee_len) &&        /*L2*/
3005                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L3*/
3006                 struct ext4_extent *prev_ex;
3007                 ext4_lblk_t prev_lblk;
3008                 ext4_fsblk_t prev_pblk, ee_pblk;
3009                 unsigned int prev_len, write_len;
3010
3011                 prev_ex = ex - 1;
3012                 prev_lblk = le32_to_cpu(prev_ex->ee_block);
3013                 prev_len = ext4_ext_get_actual_len(prev_ex);
3014                 prev_pblk = ext4_ext_pblock(prev_ex);
3015                 ee_pblk = ext4_ext_pblock(ex);
3016                 write_len = map->m_len;
3017
3018                 /*
3019                  * A transfer of blocks from 'ex' to 'prev_ex' is allowed
3020                  * upon those conditions:
3021                  * - C1: prev_ex is initialized,
3022                  * - C2: prev_ex is logically abutting ex,
3023                  * - C3: prev_ex is physically abutting ex,
3024                  * - C4: prev_ex can receive the additional blocks without
3025                  *   overflowing the (initialized) length limit.
3026                  */
3027                 if ((!ext4_ext_is_uninitialized(prev_ex)) &&            /*C1*/
3028                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3029                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3030                         (prev_len < (EXT_INIT_MAX_LEN - write_len))) {  /*C4*/
3031                         err = ext4_ext_get_access(handle, inode, path + depth);
3032                         if (err)
3033                                 goto out;
3034
3035                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3036                                 map, ex, prev_ex);
3037
3038                         /* Shift the start of ex by 'write_len' blocks */
3039                         ex->ee_block = cpu_to_le32(ee_block + write_len);
3040                         ext4_ext_store_pblock(ex, ee_pblk + write_len);
3041                         ex->ee_len = cpu_to_le16(ee_len - write_len);
3042                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3043
3044                         /* Extend prev_ex by 'write_len' blocks */
3045                         prev_ex->ee_len = cpu_to_le16(prev_len + write_len);
3046
3047                         /* Mark the block containing both extents as dirty */
3048                         ext4_ext_dirty(handle, inode, path + depth);
3049
3050                         /* Update path to point to the right extent */
3051                         path[depth].p_ext = prev_ex;
3052
3053                         /* Result: number of initialized blocks past m_lblk */
3054                         allocated = write_len;
3055                         goto out;
3056                 }
3057         }
3058
3059         WARN_ON(map->m_lblk < ee_block);
3060         /*
3061          * It is safe to convert extent to initialized via explicit
3062          * zeroout only if extent is fully insde i_size or new_size.
3063          */
3064         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3065
3066         /* If extent has less than 2*EXT4_EXT_ZERO_LEN zerout directly */
3067         if (ee_len <= 2*EXT4_EXT_ZERO_LEN &&
3068             (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3069                 err = ext4_ext_zeroout(inode, ex);
3070                 if (err)
3071                         goto out;
3072
3073                 err = ext4_ext_get_access(handle, inode, path + depth);
3074                 if (err)
3075                         goto out;
3076                 ext4_ext_mark_initialized(ex);
3077                 ext4_ext_try_to_merge(inode, path, ex);
3078                 err = ext4_ext_dirty(handle, inode, path + depth);
3079                 goto out;
3080         }
3081
3082         /*
3083          * four cases:
3084          * 1. split the extent into three extents.
3085          * 2. split the extent into two extents, zeroout the first half.
3086          * 3. split the extent into two extents, zeroout the second half.
3087          * 4. split the extent into two extents with out zeroout.
3088          */
3089         split_map.m_lblk = map->m_lblk;
3090         split_map.m_len = map->m_len;
3091
3092         if (allocated > map->m_len) {
3093                 if (allocated <= EXT4_EXT_ZERO_LEN &&
3094                     (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3095                         /* case 3 */
3096                         zero_ex.ee_block =
3097                                          cpu_to_le32(map->m_lblk);
3098                         zero_ex.ee_len = cpu_to_le16(allocated);
3099                         ext4_ext_store_pblock(&zero_ex,
3100                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3101                         err = ext4_ext_zeroout(inode, &zero_ex);
3102                         if (err)
3103                                 goto out;
3104                         split_map.m_lblk = map->m_lblk;
3105                         split_map.m_len = allocated;
3106                 } else if ((map->m_lblk - ee_block + map->m_len <
3107                            EXT4_EXT_ZERO_LEN) &&
3108                            (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3109                         /* case 2 */
3110                         if (map->m_lblk != ee_block) {
3111                                 zero_ex.ee_block = ex->ee_block;
3112                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3113                                                         ee_block);
3114                                 ext4_ext_store_pblock(&zero_ex,
3115                                                       ext4_ext_pblock(ex));
3116                                 err = ext4_ext_zeroout(inode, &zero_ex);
3117                                 if (err)
3118                                         goto out;
3119                         }
3120
3121                         split_map.m_lblk = ee_block;
3122                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3123                         allocated = map->m_len;
3124                 }
3125         }
3126
3127         allocated = ext4_split_extent(handle, inode, path,
3128                                        &split_map, split_flag, 0);
3129         if (allocated < 0)
3130                 err = allocated;
3131
3132 out:
3133         return err ? err : allocated;
3134 }
3135
3136 /*
3137  * This function is called by ext4_ext_map_blocks() from
3138  * ext4_get_blocks_dio_write() when DIO to write
3139  * to an uninitialized extent.
3140  *
3141  * Writing to an uninitialized extent may result in splitting the uninitialized
3142  * extent into multiple /initialized uninitialized extents (up to three)
3143  * There are three possibilities:
3144  *   a> There is no split required: Entire extent should be uninitialized
3145  *   b> Splits in two extents: Write is happening at either end of the extent
3146  *   c> Splits in three extents: Somone is writing in middle of the extent
3147  *
3148  * One of more index blocks maybe needed if the extent tree grow after
3149  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3150  * complete, we need to split the uninitialized extent before DIO submit
3151  * the IO. The uninitialized extent called at this time will be split
3152  * into three uninitialized extent(at most). After IO complete, the part
3153  * being filled will be convert to initialized by the end_io callback function
3154  * via ext4_convert_unwritten_extents().
3155  *
3156  * Returns the size of uninitialized extent to be written on success.
3157  */
3158 static int ext4_split_unwritten_extents(handle_t *handle,
3159                                         struct inode *inode,
3160                                         struct ext4_map_blocks *map,
3161                                         struct ext4_ext_path *path,
3162                                         int flags)
3163 {
3164         ext4_lblk_t eof_block;
3165         ext4_lblk_t ee_block;
3166         struct ext4_extent *ex;
3167         unsigned int ee_len;
3168         int split_flag = 0, depth;
3169
3170         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3171                 "block %llu, max_blocks %u\n", inode->i_ino,
3172                 (unsigned long long)map->m_lblk, map->m_len);
3173
3174         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3175                 inode->i_sb->s_blocksize_bits;
3176         if (eof_block < map->m_lblk + map->m_len)
3177                 eof_block = map->m_lblk + map->m_len;
3178         /*
3179          * It is safe to convert extent to initialized via explicit
3180          * zeroout only if extent is fully insde i_size or new_size.
3181          */
3182         depth = ext_depth(inode);
3183         ex = path[depth].p_ext;
3184         ee_block = le32_to_cpu(ex->ee_block);
3185         ee_len = ext4_ext_get_actual_len(ex);
3186
3187         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3188         split_flag |= EXT4_EXT_MARK_UNINIT2;
3189         if (flags & EXT4_GET_BLOCKS_CONVERT)
3190                 split_flag |= EXT4_EXT_DATA_VALID2;
3191         flags |= EXT4_GET_BLOCKS_PRE_IO;
3192         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3193 }
3194
3195 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3196                                                 struct inode *inode,
3197                                                 struct ext4_map_blocks *map,
3198                                                 struct ext4_ext_path *path)
3199 {
3200         struct ext4_extent *ex;
3201         ext4_lblk_t ee_block;
3202         unsigned int ee_len;
3203         int depth;
3204         int err = 0;
3205
3206         depth = ext_depth(inode);
3207         ex = path[depth].p_ext;
3208         ee_block = le32_to_cpu(ex->ee_block);
3209         ee_len = ext4_ext_get_actual_len(ex);
3210
3211         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3212                 "block %llu, max_blocks %u\n", inode->i_ino,
3213                   (unsigned long long)ee_block, ee_len);
3214
3215         /* If extent is larger than requested then split is required */
3216         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3217                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3218                                                    EXT4_GET_BLOCKS_CONVERT);
3219                 if (err < 0)
3220                         goto out;
3221                 ext4_ext_drop_refs(path);
3222                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3223                 if (IS_ERR(path)) {
3224                         err = PTR_ERR(path);
3225                         goto out;
3226                 }
3227                 depth = ext_depth(inode);
3228                 ex = path[depth].p_ext;
3229         }
3230
3231         err = ext4_ext_get_access(handle, inode, path + depth);
3232         if (err)
3233                 goto out;
3234         /* first mark the extent as initialized */
3235         ext4_ext_mark_initialized(ex);
3236
3237         /* note: ext4_ext_correct_indexes() isn't needed here because
3238          * borders are not changed
3239          */
3240         ext4_ext_try_to_merge(inode, path, ex);
3241
3242         /* Mark modified extent as dirty */
3243         err = ext4_ext_dirty(handle, inode, path + depth);
3244 out:
3245         ext4_ext_show_leaf(inode, path);
3246         return err;
3247 }
3248
3249 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3250                         sector_t block, int count)
3251 {
3252         int i;
3253         for (i = 0; i < count; i++)
3254                 unmap_underlying_metadata(bdev, block + i);
3255 }
3256
3257 /*
3258  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3259  */
3260 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3261                               ext4_lblk_t lblk,
3262                               struct ext4_ext_path *path,
3263                               unsigned int len)
3264 {
3265         int i, depth;
3266         struct ext4_extent_header *eh;
3267         struct ext4_extent *last_ex;
3268
3269         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3270                 return 0;
3271
3272         depth = ext_depth(inode);
3273         eh = path[depth].p_hdr;
3274
3275         if (unlikely(!eh->eh_entries)) {
3276                 EXT4_ERROR_INODE(inode, "eh->eh_entries == 0 and "
3277                                  "EOFBLOCKS_FL set");
3278                 return -EIO;
3279         }
3280         last_ex = EXT_LAST_EXTENT(eh);
3281         /*
3282          * We should clear the EOFBLOCKS_FL flag if we are writing the
3283          * last block in the last extent in the file.  We test this by
3284          * first checking to see if the caller to
3285          * ext4_ext_get_blocks() was interested in the last block (or
3286          * a block beyond the last block) in the current extent.  If
3287          * this turns out to be false, we can bail out from this
3288          * function immediately.
3289          */
3290         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3291             ext4_ext_get_actual_len(last_ex))
3292                 return 0;
3293         /*
3294          * If the caller does appear to be planning to write at or
3295          * beyond the end of the current extent, we then test to see
3296          * if the current extent is the last extent in the file, by
3297          * checking to make sure it was reached via the rightmost node
3298          * at each level of the tree.
3299          */
3300         for (i = depth-1; i >= 0; i--)
3301                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3302                         return 0;
3303         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3304         return ext4_mark_inode_dirty(handle, inode);
3305 }
3306
3307 /**
3308  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3309  *
3310  * Goes through the buffer heads in the range [lblk_start, lblk_end] and returns
3311  * whether there are any buffers marked for delayed allocation. It returns '1'
3312  * on the first delalloc'ed buffer head found. If no buffer head in the given
3313  * range is marked for delalloc, it returns 0.
3314  * lblk_start should always be <= lblk_end.
3315  * search_hint_reverse is to indicate that searching in reverse from lblk_end to
3316  * lblk_start might be more efficient (i.e., we will likely hit the delalloc'ed
3317  * block sooner). This is useful when blocks are truncated sequentially from
3318  * lblk_start towards lblk_end.
3319  */
3320 static int ext4_find_delalloc_range(struct inode *inode,
3321                                     ext4_lblk_t lblk_start,
3322                                     ext4_lblk_t lblk_end,
3323                                     int search_hint_reverse)
3324 {
3325         struct address_space *mapping = inode->i_mapping;
3326         struct buffer_head *head, *bh = NULL;
3327         struct page *page;
3328         ext4_lblk_t i, pg_lblk;
3329         pgoff_t index;
3330
3331         /* reverse search wont work if fs block size is less than page size */
3332         if (inode->i_blkbits < PAGE_CACHE_SHIFT)
3333                 search_hint_reverse = 0;
3334
3335         if (search_hint_reverse)
3336                 i = lblk_end;
3337         else
3338                 i = lblk_start;
3339
3340         index = i >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
3341
3342         while ((i >= lblk_start) && (i <= lblk_end)) {
3343                 page = find_get_page(mapping, index);
3344                 if (!page)
3345                         goto nextpage;
3346
3347                 if (!page_has_buffers(page))
3348                         goto nextpage;
3349
3350                 head = page_buffers(page);
3351                 if (!head)
3352                         goto nextpage;
3353
3354                 bh = head;
3355                 pg_lblk = index << (PAGE_CACHE_SHIFT -
3356                                                 inode->i_blkbits);
3357                 do {
3358                         if (unlikely(pg_lblk < lblk_start)) {
3359                                 /*
3360                                  * This is possible when fs block size is less
3361                                  * than page size and our cluster starts/ends in
3362                                  * middle of the page. So we need to skip the
3363                                  * initial few blocks till we reach the 'lblk'
3364                                  */
3365                                 pg_lblk++;
3366                                 continue;
3367                         }
3368
3369                         /* Check if the buffer is delayed allocated and that it
3370                          * is not yet mapped. (when da-buffers are mapped during
3371                          * their writeout, their da_mapped bit is set.)
3372                          */
3373                         if (buffer_delay(bh) && !buffer_da_mapped(bh)) {
3374                                 page_cache_release(page);
3375                                 trace_ext4_find_delalloc_range(inode,
3376                                                 lblk_start, lblk_end,
3377                                                 search_hint_reverse,
3378                                                 1, i);
3379                                 return 1;
3380                         }
3381                         if (search_hint_reverse)
3382                                 i--;
3383                         else
3384                                 i++;
3385                 } while ((i >= lblk_start) && (i <= lblk_end) &&
3386                                 ((bh = bh->b_this_page) != head));
3387 nextpage:
3388                 if (page)
3389                         page_cache_release(page);
3390                 /*
3391                  * Move to next page. 'i' will be the first lblk in the next
3392                  * page.
3393                  */
3394                 if (search_hint_reverse)
3395                         index--;
3396                 else
3397                         index++;
3398                 i = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
3399         }
3400
3401         trace_ext4_find_delalloc_range(inode, lblk_start, lblk_end,
3402                                         search_hint_reverse, 0, 0);
3403         return 0;
3404 }
3405
3406 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk,
3407                                int search_hint_reverse)
3408 {
3409         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3410         ext4_lblk_t lblk_start, lblk_end;
3411         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3412         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3413
3414         return ext4_find_delalloc_range(inode, lblk_start, lblk_end,
3415                                         search_hint_reverse);
3416 }
3417
3418 /**
3419  * Determines how many complete clusters (out of those specified by the 'map')
3420  * are under delalloc and were reserved quota for.
3421  * This function is called when we are writing out the blocks that were
3422  * originally written with their allocation delayed, but then the space was
3423  * allocated using fallocate() before the delayed allocation could be resolved.
3424  * The cases to look for are:
3425  * ('=' indicated delayed allocated blocks
3426  *  '-' indicates non-delayed allocated blocks)
3427  * (a) partial clusters towards beginning and/or end outside of allocated range
3428  *     are not delalloc'ed.
3429  *      Ex:
3430  *      |----c---=|====c====|====c====|===-c----|
3431  *               |++++++ allocated ++++++|
3432  *      ==> 4 complete clusters in above example
3433  *
3434  * (b) partial cluster (outside of allocated range) towards either end is
3435  *     marked for delayed allocation. In this case, we will exclude that
3436  *     cluster.
3437  *      Ex:
3438  *      |----====c========|========c========|
3439  *           |++++++ allocated ++++++|
3440  *      ==> 1 complete clusters in above example
3441  *
3442  *      Ex:
3443  *      |================c================|
3444  *            |++++++ allocated ++++++|
3445  *      ==> 0 complete clusters in above example
3446  *
3447  * The ext4_da_update_reserve_space will be called only if we
3448  * determine here that there were some "entire" clusters that span
3449  * this 'allocated' range.
3450  * In the non-bigalloc case, this function will just end up returning num_blks
3451  * without ever calling ext4_find_delalloc_range.
3452  */
3453 static unsigned int
3454 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3455                            unsigned int num_blks)
3456 {
3457         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3458         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3459         ext4_lblk_t lblk_from, lblk_to, c_offset;
3460         unsigned int allocated_clusters = 0;
3461
3462         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3463         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3464
3465         /* max possible clusters for this allocation */
3466         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3467
3468         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3469
3470         /* Check towards left side */
3471         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3472         if (c_offset) {
3473                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3474                 lblk_to = lblk_from + c_offset - 1;
3475
3476                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to, 0))
3477                         allocated_clusters--;
3478         }
3479
3480         /* Now check towards right. */
3481         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3482         if (allocated_clusters && c_offset) {
3483                 lblk_from = lblk_start + num_blks;
3484                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3485
3486                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to, 0))
3487                         allocated_clusters--;
3488         }
3489
3490         return allocated_clusters;
3491 }
3492
3493 static int
3494 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3495                         struct ext4_map_blocks *map,
3496                         struct ext4_ext_path *path, int flags,
3497                         unsigned int allocated, ext4_fsblk_t newblock)
3498 {
3499         int ret = 0;
3500         int err = 0;
3501         ext4_io_end_t *io = EXT4_I(inode)->cur_aio_dio;
3502
3503         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical"
3504                   "block %llu, max_blocks %u, flags %d, allocated %u",
3505                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3506                   flags, allocated);
3507         ext4_ext_show_leaf(inode, path);
3508
3509         trace_ext4_ext_handle_uninitialized_extents(inode, map, allocated,
3510                                                     newblock);
3511
3512         /* get_block() before submit the IO, split the extent */
3513         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3514                 ret = ext4_split_unwritten_extents(handle, inode, map,
3515                                                    path, flags);
3516                 /*
3517                  * Flag the inode(non aio case) or end_io struct (aio case)
3518                  * that this IO needs to conversion to written when IO is
3519                  * completed
3520                  */
3521                 if (io)
3522                         ext4_set_io_unwritten_flag(inode, io);
3523                 else
3524                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3525                 if (ext4_should_dioread_nolock(inode))
3526                         map->m_flags |= EXT4_MAP_UNINIT;
3527                 goto out;
3528         }
3529         /* IO end_io complete, convert the filled extent to written */
3530         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3531                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3532                                                         path);
3533                 if (ret >= 0) {
3534                         ext4_update_inode_fsync_trans(handle, inode, 1);
3535                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3536                                                  path, map->m_len);
3537                 } else
3538                         err = ret;
3539                 goto out2;
3540         }
3541         /* buffered IO case */
3542         /*
3543          * repeat fallocate creation request
3544          * we already have an unwritten extent
3545          */
3546         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT)
3547                 goto map_out;
3548
3549         /* buffered READ or buffered write_begin() lookup */
3550         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3551                 /*
3552                  * We have blocks reserved already.  We
3553                  * return allocated blocks so that delalloc
3554                  * won't do block reservation for us.  But
3555                  * the buffer head will be unmapped so that
3556                  * a read from the block returns 0s.
3557                  */
3558                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3559                 goto out1;
3560         }
3561
3562         /* buffered write, writepage time, convert*/
3563         ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
3564         if (ret >= 0)
3565                 ext4_update_inode_fsync_trans(handle, inode, 1);
3566 out:
3567         if (ret <= 0) {
3568                 err = ret;
3569                 goto out2;
3570         } else
3571                 allocated = ret;
3572         map->m_flags |= EXT4_MAP_NEW;
3573         /*
3574          * if we allocated more blocks than requested
3575          * we need to make sure we unmap the extra block
3576          * allocated. The actual needed block will get
3577          * unmapped later when we find the buffer_head marked
3578          * new.
3579          */
3580         if (allocated > map->m_len) {
3581                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3582                                         newblock + map->m_len,
3583                                         allocated - map->m_len);
3584                 allocated = map->m_len;
3585         }
3586
3587         /*
3588          * If we have done fallocate with the offset that is already
3589          * delayed allocated, we would have block reservation
3590          * and quota reservation done in the delayed write path.
3591          * But fallocate would have already updated quota and block
3592          * count for this offset. So cancel these reservation
3593          */
3594         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3595                 unsigned int reserved_clusters;
3596                 reserved_clusters = get_reserved_cluster_alloc(inode,
3597                                 map->m_lblk, map->m_len);
3598                 if (reserved_clusters)
3599                         ext4_da_update_reserve_space(inode,
3600                                                      reserved_clusters,
3601                                                      0);
3602         }
3603
3604 map_out:
3605         map->m_flags |= EXT4_MAP_MAPPED;
3606         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3607                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3608                                          map->m_len);
3609                 if (err < 0)
3610                         goto out2;
3611         }
3612 out1:
3613         if (allocated > map->m_len)
3614                 allocated = map->m_len;
3615         ext4_ext_show_leaf(inode, path);
3616         map->m_pblk = newblock;
3617         map->m_len = allocated;
3618 out2:
3619         if (path) {
3620                 ext4_ext_drop_refs(path);
3621                 kfree(path);
3622         }
3623         return err ? err : allocated;
3624 }
3625
3626 /*
3627  * get_implied_cluster_alloc - check to see if the requested
3628  * allocation (in the map structure) overlaps with a cluster already
3629  * allocated in an extent.
3630  *      @sb     The filesystem superblock structure
3631  *      @map    The requested lblk->pblk mapping
3632  *      @ex     The extent structure which might contain an implied
3633  *                      cluster allocation
3634  *
3635  * This function is called by ext4_ext_map_blocks() after we failed to
3636  * find blocks that were already in the inode's extent tree.  Hence,
3637  * we know that the beginning of the requested region cannot overlap
3638  * the extent from the inode's extent tree.  There are three cases we
3639  * want to catch.  The first is this case:
3640  *
3641  *               |--- cluster # N--|
3642  *    |--- extent ---|  |---- requested region ---|
3643  *                      |==========|
3644  *
3645  * The second case that we need to test for is this one:
3646  *
3647  *   |--------- cluster # N ----------------|
3648  *         |--- requested region --|   |------- extent ----|
3649  *         |=======================|
3650  *
3651  * The third case is when the requested region lies between two extents
3652  * within the same cluster:
3653  *          |------------- cluster # N-------------|
3654  * |----- ex -----|                  |---- ex_right ----|
3655  *                  |------ requested region ------|
3656  *                  |================|
3657  *
3658  * In each of the above cases, we need to set the map->m_pblk and
3659  * map->m_len so it corresponds to the return the extent labelled as
3660  * "|====|" from cluster #N, since it is already in use for data in
3661  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3662  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3663  * as a new "allocated" block region.  Otherwise, we will return 0 and
3664  * ext4_ext_map_blocks() will then allocate one or more new clusters
3665  * by calling ext4_mb_new_blocks().
3666  */
3667 static int get_implied_cluster_alloc(struct super_block *sb,
3668                                      struct ext4_map_blocks *map,
3669                                      struct ext4_extent *ex,
3670                                      struct ext4_ext_path *path)
3671 {
3672         struct ext4_sb_info *sbi = EXT4_SB(sb);
3673         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3674         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3675         ext4_lblk_t rr_cluster_start, rr_cluster_end;
3676         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3677         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3678         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3679
3680         /* The extent passed in that we are trying to match */
3681         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3682         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3683
3684         /* The requested region passed into ext4_map_blocks() */
3685         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3686         rr_cluster_end = EXT4_B2C(sbi, map->m_lblk + map->m_len - 1);
3687
3688         if ((rr_cluster_start == ex_cluster_end) ||
3689             (rr_cluster_start == ex_cluster_start)) {
3690                 if (rr_cluster_start == ex_cluster_end)
3691                         ee_start += ee_len - 1;
3692                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3693                         c_offset;
3694                 map->m_len = min(map->m_len,
3695                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3696                 /*
3697                  * Check for and handle this case:
3698                  *
3699                  *   |--------- cluster # N-------------|
3700                  *                     |------- extent ----|
3701                  *         |--- requested region ---|
3702                  *         |===========|
3703                  */
3704
3705                 if (map->m_lblk < ee_block)
3706                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3707
3708                 /*
3709                  * Check for the case where there is already another allocated
3710                  * block to the right of 'ex' but before the end of the cluster.
3711                  *
3712                  *          |------------- cluster # N-------------|
3713                  * |----- ex -----|                  |---- ex_right ----|
3714                  *                  |------ requested region ------|
3715                  *                  |================|
3716                  */
3717                 if (map->m_lblk > ee_block) {
3718                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
3719                         map->m_len = min(map->m_len, next - map->m_lblk);
3720                 }
3721
3722                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
3723                 return 1;
3724         }
3725
3726         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
3727         return 0;
3728 }
3729
3730
3731 /*
3732  * Block allocation/map/preallocation routine for extents based files
3733  *
3734  *
3735  * Need to be called with
3736  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
3737  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
3738  *
3739  * return > 0, number of of blocks already mapped/allocated
3740  *          if create == 0 and these are pre-allocated blocks
3741  *              buffer head is unmapped
3742  *          otherwise blocks are mapped
3743  *
3744  * return = 0, if plain look up failed (blocks have not been allocated)
3745  *          buffer head is unmapped
3746  *
3747  * return < 0, error case.
3748  */
3749 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
3750                         struct ext4_map_blocks *map, int flags)
3751 {
3752         struct ext4_ext_path *path = NULL;
3753         struct ext4_extent newex, *ex, *ex2;
3754         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3755         ext4_fsblk_t newblock = 0;
3756         int free_on_err = 0, err = 0, depth, ret;
3757         unsigned int allocated = 0, offset = 0;
3758         unsigned int allocated_clusters = 0;
3759         unsigned int punched_out = 0;
3760         unsigned int result = 0;
3761         struct ext4_allocation_request ar;
3762         ext4_io_end_t *io = EXT4_I(inode)->cur_aio_dio;
3763         ext4_lblk_t cluster_offset;
3764
3765         ext_debug("blocks %u/%u requested for inode %lu\n",
3766                   map->m_lblk, map->m_len, inode->i_ino);
3767         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
3768
3769         /* check in cache */
3770         if (!(flags & EXT4_GET_BLOCKS_PUNCH_OUT_EXT) &&
3771                 ext4_ext_in_cache(inode, map->m_lblk, &newex)) {
3772                 if (!newex.ee_start_lo && !newex.ee_start_hi) {
3773                         if ((sbi->s_cluster_ratio > 1) &&
3774                             ext4_find_delalloc_cluster(inode, map->m_lblk, 0))
3775                                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3776
3777                         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3778                                 /*
3779                                  * block isn't allocated yet and
3780                                  * user doesn't want to allocate it
3781                                  */
3782                                 goto out2;
3783                         }
3784                         /* we should allocate requested block */
3785                 } else {
3786                         /* block is already allocated */
3787                         if (sbi->s_cluster_ratio > 1)
3788                                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3789                         newblock = map->m_lblk
3790                                    - le32_to_cpu(newex.ee_block)
3791                                    + ext4_ext_pblock(&newex);
3792                         /* number of remaining blocks in the extent */
3793                         allocated = ext4_ext_get_actual_len(&newex) -
3794                                 (map->m_lblk - le32_to_cpu(newex.ee_block));
3795                         goto out;
3796                 }
3797         }
3798
3799         /* find extent for this block */
3800         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
3801         if (IS_ERR(path)) {
3802                 err = PTR_ERR(path);
3803                 path = NULL;
3804                 goto out2;
3805         }
3806
3807         depth = ext_depth(inode);
3808
3809         /*
3810          * consistent leaf must not be empty;
3811          * this situation is possible, though, _during_ tree modification;
3812          * this is why assert can't be put in ext4_ext_find_extent()
3813          */
3814         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
3815                 EXT4_ERROR_INODE(inode, "bad extent address "
3816                                  "lblock: %lu, depth: %d pblock %lld",
3817                                  (unsigned long) map->m_lblk, depth,
3818                                  path[depth].p_block);
3819                 err = -EIO;
3820                 goto out2;
3821         }
3822
3823         ex = path[depth].p_ext;
3824         if (ex) {
3825                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3826                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3827                 unsigned short ee_len;
3828
3829                 /*
3830                  * Uninitialized extents are treated as holes, except that
3831                  * we split out initialized portions during a write.
3832                  */
3833                 ee_len = ext4_ext_get_actual_len(ex);
3834
3835                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
3836
3837                 /* if found extent covers block, simply return it */
3838                 if (in_range(map->m_lblk, ee_block, ee_len)) {
3839                         struct ext4_map_blocks punch_map;
3840                         ext4_fsblk_t partial_cluster = 0;
3841
3842                         newblock = map->m_lblk - ee_block + ee_start;
3843                         /* number of remaining blocks in the extent */
3844                         allocated = ee_len - (map->m_lblk - ee_block);
3845                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
3846                                   ee_block, ee_len, newblock);
3847
3848                         if ((flags & EXT4_GET_BLOCKS_PUNCH_OUT_EXT) == 0) {
3849                                 /*
3850                                  * Do not put uninitialized extent
3851                                  * in the cache
3852                                  */
3853                                 if (!ext4_ext_is_uninitialized(ex)) {
3854                                         ext4_ext_put_in_cache(inode, ee_block,
3855                                                 ee_len, ee_start);
3856                                         goto out;
3857                                 }
3858                                 ret = ext4_ext_handle_uninitialized_extents(
3859                                         handle, inode, map, path, flags,
3860                                         allocated, newblock);
3861                                 return ret;
3862                         }
3863
3864                         /*
3865                          * Punch out the map length, but only to the
3866                          * end of the extent
3867                          */
3868                         punched_out = allocated < map->m_len ?
3869                                 allocated : map->m_len;
3870
3871                         /*
3872                          * Sense extents need to be converted to
3873                          * uninitialized, they must fit in an
3874                          * uninitialized extent
3875                          */
3876                         if (punched_out > EXT_UNINIT_MAX_LEN)
3877                                 punched_out = EXT_UNINIT_MAX_LEN;
3878
3879                         punch_map.m_lblk = map->m_lblk;
3880                         punch_map.m_pblk = newblock;
3881                         punch_map.m_len = punched_out;
3882                         punch_map.m_flags = 0;
3883
3884                         /* Check to see if the extent needs to be split */
3885                         if (punch_map.m_len != ee_len ||
3886                                 punch_map.m_lblk != ee_block) {
3887
3888                                 ret = ext4_split_extent(handle, inode,
3889                                 path, &punch_map, 0,
3890                                 EXT4_GET_BLOCKS_PUNCH_OUT_EXT |
3891                                 EXT4_GET_BLOCKS_PRE_IO);
3892
3893                                 if (ret < 0) {
3894                                         err = ret;
3895                                         goto out2;
3896                                 }
3897                                 /*
3898                                  * find extent for the block at
3899                                  * the start of the hole
3900                                  */
3901                                 ext4_ext_drop_refs(path);
3902                                 kfree(path);
3903
3904                                 path = ext4_ext_find_extent(inode,
3905                                 map->m_lblk, NULL);
3906                                 if (IS_ERR(path)) {
3907                                         err = PTR_ERR(path);
3908                                         path = NULL;
3909                                         goto out2;
3910                                 }
3911
3912                                 depth = ext_depth(inode);
3913                                 ex = path[depth].p_ext;
3914                                 ee_len = ext4_ext_get_actual_len(ex);
3915                                 ee_block = le32_to_cpu(ex->ee_block);
3916                                 ee_start = ext4_ext_pblock(ex);
3917
3918                         }
3919
3920                         ext4_ext_mark_uninitialized(ex);
3921
3922                         ext4_ext_invalidate_cache(inode);
3923
3924                         err = ext4_ext_rm_leaf(handle, inode, path,
3925                                                &partial_cluster, map->m_lblk,
3926                                                map->m_lblk + punched_out);
3927
3928                         if (!err && path->p_hdr->eh_entries == 0) {
3929                                 /*
3930                                  * Punch hole freed all of this sub tree,
3931                                  * so we need to correct eh_depth
3932                                  */
3933                                 err = ext4_ext_get_access(handle, inode, path);
3934                                 if (err == 0) {
3935                                         ext_inode_hdr(inode)->eh_depth = 0;
3936                                         ext_inode_hdr(inode)->eh_max =
3937                                         cpu_to_le16(ext4_ext_space_root(
3938                                                 inode, 0));
3939
3940                                         err = ext4_ext_dirty(
3941                                                 handle, inode, path);
3942                                 }
3943                         }
3944
3945                         goto out2;
3946                 }
3947         }
3948
3949         if ((sbi->s_cluster_ratio > 1) &&
3950             ext4_find_delalloc_cluster(inode, map->m_lblk, 0))
3951                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3952
3953         /*
3954          * requested block isn't allocated yet;
3955          * we couldn't try to create block if create flag is zero
3956          */
3957         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3958                 /*
3959                  * put just found gap into cache to speed up
3960                  * subsequent requests
3961                  */
3962                 ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
3963                 goto out2;
3964         }
3965
3966         /*
3967          * Okay, we need to do block allocation.
3968          */
3969         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
3970         newex.ee_block = cpu_to_le32(map->m_lblk);
3971         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3972
3973         /*
3974          * If we are doing bigalloc, check to see if the extent returned
3975          * by ext4_ext_find_extent() implies a cluster we can use.
3976          */
3977         if (cluster_offset && ex &&
3978             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
3979                 ar.len = allocated = map->m_len;
3980                 newblock = map->m_pblk;
3981                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3982                 goto got_allocated_blocks;
3983         }
3984
3985         /* find neighbour allocated blocks */
3986         ar.lleft = map->m_lblk;
3987         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
3988         if (err)
3989                 goto out2;
3990         ar.lright = map->m_lblk;
3991         ex2 = NULL;
3992         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
3993         if (err)
3994                 goto out2;
3995
3996         /* Check if the extent after searching to the right implies a
3997          * cluster we can use. */
3998         if ((sbi->s_cluster_ratio > 1) && ex2 &&
3999             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4000                 ar.len = allocated = map->m_len;
4001                 newblock = map->m_pblk;
4002                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4003                 goto got_allocated_blocks;
4004         }
4005
4006         /*
4007          * See if request is beyond maximum number of blocks we can have in
4008          * a single extent. For an initialized extent this limit is
4009          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4010          * EXT_UNINIT_MAX_LEN.
4011          */
4012         if (map->m_len > EXT_INIT_MAX_LEN &&
4013             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4014                 map->m_len = EXT_INIT_MAX_LEN;
4015         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4016                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4017                 map->m_len = EXT_UNINIT_MAX_LEN;
4018
4019         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4020         newex.ee_len = cpu_to_le16(map->m_len);
4021         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4022         if (err)
4023                 allocated = ext4_ext_get_actual_len(&newex);
4024         else
4025                 allocated = map->m_len;
4026
4027         /* allocate new block */
4028         ar.inode = inode;
4029         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4030         ar.logical = map->m_lblk;
4031         /*
4032          * We calculate the offset from the beginning of the cluster
4033          * for the logical block number, since when we allocate a
4034          * physical cluster, the physical block should start at the
4035          * same offset from the beginning of the cluster.  This is
4036          * needed so that future calls to get_implied_cluster_alloc()
4037          * work correctly.
4038          */
4039         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
4040         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4041         ar.goal -= offset;
4042         ar.logical -= offset;
4043         if (S_ISREG(inode->i_mode))
4044                 ar.flags = EXT4_MB_HINT_DATA;
4045         else
4046                 /* disable in-core preallocation for non-regular files */
4047                 ar.flags = 0;
4048         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4049                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4050         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4051         if (!newblock)
4052                 goto out2;
4053         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4054                   ar.goal, newblock, allocated);
4055         free_on_err = 1;
4056         allocated_clusters = ar.len;
4057         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4058         if (ar.len > allocated)
4059                 ar.len = allocated;
4060
4061 got_allocated_blocks:
4062         /* try to insert new extent into found leaf and return */
4063         ext4_ext_store_pblock(&newex, newblock + offset);
4064         newex.ee_len = cpu_to_le16(ar.len);
4065         /* Mark uninitialized */
4066         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4067                 ext4_ext_mark_uninitialized(&newex);
4068                 /*
4069                  * io_end structure was created for every IO write to an
4070                  * uninitialized extent. To avoid unnecessary conversion,
4071                  * here we flag the IO that really needs the conversion.
4072                  * For non asycn direct IO case, flag the inode state
4073                  * that we need to perform conversion when IO is done.
4074                  */
4075                 if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
4076                         if (io)
4077                                 ext4_set_io_unwritten_flag(inode, io);
4078                         else
4079                                 ext4_set_inode_state(inode,
4080                                                      EXT4_STATE_DIO_UNWRITTEN);
4081                 }
4082                 if (ext4_should_dioread_nolock(inode))
4083                         map->m_flags |= EXT4_MAP_UNINIT;
4084         }
4085
4086         err = 0;
4087         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4088                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4089                                          path, ar.len);
4090         if (!err)
4091                 err = ext4_ext_insert_extent(handle, inode, path,
4092                                              &newex, flags);
4093         if (err && free_on_err) {
4094                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4095                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4096                 /* free data blocks we just allocated */
4097                 /* not a good idea to call discard here directly,
4098                  * but otherwise we'd need to call it every free() */
4099                 ext4_discard_preallocations(inode);
4100                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4101                                  ext4_ext_get_actual_len(&newex), fb_flags);
4102                 goto out2;
4103         }
4104
4105         /* previous routine could use block we allocated */
4106         newblock = ext4_ext_pblock(&newex);
4107         allocated = ext4_ext_get_actual_len(&newex);
4108         if (allocated > map->m_len)
4109                 allocated = map->m_len;
4110         map->m_flags |= EXT4_MAP_NEW;
4111
4112         /*
4113          * Update reserved blocks/metadata blocks after successful
4114          * block allocation which had been deferred till now.
4115          */
4116         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4117                 unsigned int reserved_clusters;
4118                 /*
4119                  * Check how many clusters we had reserved this allocated range
4120                  */
4121                 reserved_clusters = get_reserved_cluster_alloc(inode,
4122                                                 map->m_lblk, allocated);
4123                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4124                         if (reserved_clusters) {
4125                                 /*
4126                                  * We have clusters reserved for this range.
4127                                  * But since we are not doing actual allocation
4128                                  * and are simply using blocks from previously
4129                                  * allocated cluster, we should release the
4130                                  * reservation and not claim quota.
4131                                  */
4132                                 ext4_da_update_reserve_space(inode,
4133                                                 reserved_clusters, 0);
4134                         }
4135                 } else {
4136                         BUG_ON(allocated_clusters < reserved_clusters);
4137                         /* We will claim quota for all newly allocated blocks.*/
4138                         ext4_da_update_reserve_space(inode, allocated_clusters,
4139                                                         1);
4140                         if (reserved_clusters < allocated_clusters) {
4141                                 struct ext4_inode_info *ei = EXT4_I(inode);
4142                                 int reservation = allocated_clusters -
4143                                                   reserved_clusters;
4144                                 /*
4145                                  * It seems we claimed few clusters outside of
4146                                  * the range of this allocation. We should give
4147                                  * it back to the reservation pool. This can
4148                                  * happen in the following case:
4149                                  *
4150                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4151                                  *   cluster has 4 blocks. Thus, the clusters
4152                                  *   are [0-3],[4-7],[8-11]...
4153                                  * * First comes delayed allocation write for
4154                                  *   logical blocks 10 & 11. Since there were no
4155                                  *   previous delayed allocated blocks in the
4156                                  *   range [8-11], we would reserve 1 cluster
4157                                  *   for this write.
4158                                  * * Next comes write for logical blocks 3 to 8.
4159                                  *   In this case, we will reserve 2 clusters
4160                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4161                                  *   that range has a delayed allocated blocks.
4162                                  *   Thus total reserved clusters now becomes 3.
4163                                  * * Now, during the delayed allocation writeout
4164                                  *   time, we will first write blocks [3-8] and
4165                                  *   allocate 3 clusters for writing these
4166                                  *   blocks. Also, we would claim all these
4167                                  *   three clusters above.
4168                                  * * Now when we come here to writeout the
4169                                  *   blocks [10-11], we would expect to claim
4170                                  *   the reservation of 1 cluster we had made
4171                                  *   (and we would claim it since there are no
4172                                  *   more delayed allocated blocks in the range
4173                                  *   [8-11]. But our reserved cluster count had
4174                                  *   already gone to 0.
4175                                  *
4176                                  *   Thus, at the step 4 above when we determine
4177                                  *   that there are still some unwritten delayed
4178                                  *   allocated blocks outside of our current
4179                                  *   block range, we should increment the
4180                                  *   reserved clusters count so that when the
4181                                  *   remaining blocks finally gets written, we
4182                                  *   could claim them.
4183                                  */
4184                                 dquot_reserve_block(inode,
4185                                                 EXT4_C2B(sbi, reservation));
4186                                 spin_lock(&ei->i_block_reservation_lock);
4187                                 ei->i_reserved_data_blocks += reservation;
4188                                 spin_unlock(&ei->i_block_reservation_lock);
4189                         }
4190                 }
4191         }
4192
4193         /*
4194          * Cache the extent and update transaction to commit on fdatasync only
4195          * when it is _not_ an uninitialized extent.
4196          */
4197         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0) {
4198                 ext4_ext_put_in_cache(inode, map->m_lblk, allocated, newblock);
4199                 ext4_update_inode_fsync_trans(handle, inode, 1);
4200         } else
4201                 ext4_update_inode_fsync_trans(handle, inode, 0);
4202 out:
4203         if (allocated > map->m_len)
4204                 allocated = map->m_len;
4205         ext4_ext_show_leaf(inode, path);
4206         map->m_flags |= EXT4_MAP_MAPPED;
4207         map->m_pblk = newblock;
4208         map->m_len = allocated;
4209 out2:
4210         if (path) {
4211                 ext4_ext_drop_refs(path);
4212                 kfree(path);
4213         }
4214         result = (flags & EXT4_GET_BLOCKS_PUNCH_OUT_EXT) ?
4215                         punched_out : allocated;
4216
4217         trace_ext4_ext_map_blocks_exit(inode, map->m_lblk,
4218                 newblock, map->m_len, err ? err : result);
4219
4220         return err ? err : result;
4221 }
4222
4223 void ext4_ext_truncate(struct inode *inode)
4224 {
4225         struct address_space *mapping = inode->i_mapping;
4226         struct super_block *sb = inode->i_sb;
4227         ext4_lblk_t last_block;
4228         handle_t *handle;
4229         loff_t page_len;
4230         int err = 0;
4231
4232         /*
4233          * finish any pending end_io work so we won't run the risk of
4234          * converting any truncated blocks to initialized later
4235          */
4236         ext4_flush_completed_IO(inode);
4237
4238         /*
4239          * probably first extent we're gonna free will be last in block
4240          */
4241         err = ext4_writepage_trans_blocks(inode);
4242         handle = ext4_journal_start(inode, err);
4243         if (IS_ERR(handle))
4244                 return;
4245
4246         if (inode->i_size % PAGE_CACHE_SIZE != 0) {
4247                 page_len = PAGE_CACHE_SIZE -
4248                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4249
4250                 err = ext4_discard_partial_page_buffers(handle,
4251                         mapping, inode->i_size, page_len, 0);
4252
4253                 if (err)
4254                         goto out_stop;
4255         }
4256
4257         if (ext4_orphan_add(handle, inode))
4258                 goto out_stop;
4259
4260         down_write(&EXT4_I(inode)->i_data_sem);
4261         ext4_ext_invalidate_cache(inode);
4262
4263         ext4_discard_preallocations(inode);
4264
4265         /*
4266          * TODO: optimization is possible here.
4267          * Probably we need not scan at all,
4268          * because page truncation is enough.
4269          */
4270
4271         /* we have to know where to truncate from in crash case */
4272         EXT4_I(inode)->i_disksize = inode->i_size;
4273         ext4_mark_inode_dirty(handle, inode);
4274
4275         last_block = (inode->i_size + sb->s_blocksize - 1)
4276                         >> EXT4_BLOCK_SIZE_BITS(sb);
4277         err = ext4_ext_remove_space(inode, last_block);
4278
4279         /* In a multi-transaction truncate, we only make the final
4280          * transaction synchronous.
4281          */
4282         if (IS_SYNC(inode))
4283                 ext4_handle_sync(handle);
4284
4285         up_write(&EXT4_I(inode)->i_data_sem);
4286
4287 out_stop:
4288         /*
4289          * If this was a simple ftruncate() and the file will remain alive,
4290          * then we need to clear up the orphan record which we created above.
4291          * However, if this was a real unlink then we were called by
4292          * ext4_delete_inode(), and we allow that function to clean up the
4293          * orphan info for us.
4294          */
4295         if (inode->i_nlink)
4296                 ext4_orphan_del(handle, inode);
4297
4298         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4299         ext4_mark_inode_dirty(handle, inode);
4300         ext4_journal_stop(handle);
4301 }
4302
4303 static void ext4_falloc_update_inode(struct inode *inode,
4304                                 int mode, loff_t new_size, int update_ctime)
4305 {
4306         struct timespec now;
4307
4308         if (update_ctime) {
4309                 now = current_fs_time(inode->i_sb);
4310                 if (!timespec_equal(&inode->i_ctime, &now))
4311                         inode->i_ctime = now;
4312         }
4313         /*
4314          * Update only when preallocation was requested beyond
4315          * the file size.
4316          */
4317         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4318                 if (new_size > i_size_read(inode))
4319                         i_size_write(inode, new_size);
4320                 if (new_size > EXT4_I(inode)->i_disksize)
4321                         ext4_update_i_disksize(inode, new_size);
4322         } else {
4323                 /*
4324                  * Mark that we allocate beyond EOF so the subsequent truncate
4325                  * can proceed even if the new size is the same as i_size.
4326                  */
4327                 if (new_size > i_size_read(inode))
4328                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4329         }
4330
4331 }
4332
4333 /*
4334  * preallocate space for a file. This implements ext4's fallocate file
4335  * operation, which gets called from sys_fallocate system call.
4336  * For block-mapped files, posix_fallocate should fall back to the method
4337  * of writing zeroes to the required new blocks (the same behavior which is
4338  * expected for file systems which do not support fallocate() system call).
4339  */
4340 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4341 {
4342         struct inode *inode = file->f_path.dentry->d_inode;
4343         handle_t *handle;
4344         loff_t new_size;
4345         unsigned int max_blocks;
4346         int ret = 0;
4347         int ret2 = 0;
4348         int retries = 0;
4349         int flags;
4350         struct ext4_map_blocks map;
4351         unsigned int credits, blkbits = inode->i_blkbits;
4352
4353         /*
4354          * currently supporting (pre)allocate mode for extent-based
4355          * files _only_
4356          */
4357         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4358                 return -EOPNOTSUPP;
4359
4360         /* Return error if mode is not supported */
4361         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4362                 return -EOPNOTSUPP;
4363
4364         if (mode & FALLOC_FL_PUNCH_HOLE)
4365                 return ext4_punch_hole(file, offset, len);
4366
4367         trace_ext4_fallocate_enter(inode, offset, len, mode);
4368         map.m_lblk = offset >> blkbits;
4369         /*
4370          * We can't just convert len to max_blocks because
4371          * If blocksize = 4096 offset = 3072 and len = 2048
4372          */
4373         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4374                 - map.m_lblk;
4375         /*
4376          * credits to insert 1 extent into extent tree
4377          */
4378         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4379         mutex_lock(&inode->i_mutex);
4380         ret = inode_newsize_ok(inode, (len + offset));
4381         if (ret) {
4382                 mutex_unlock(&inode->i_mutex);
4383                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4384                 return ret;
4385         }
4386         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4387         if (mode & FALLOC_FL_KEEP_SIZE)
4388                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4389         /*
4390          * Don't normalize the request if it can fit in one extent so
4391          * that it doesn't get unnecessarily split into multiple
4392          * extents.
4393          */
4394         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4395                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4396 retry:
4397         while (ret >= 0 && ret < max_blocks) {
4398                 map.m_lblk = map.m_lblk + ret;
4399                 map.m_len = max_blocks = max_blocks - ret;
4400                 handle = ext4_journal_start(inode, credits);
4401                 if (IS_ERR(handle)) {
4402                         ret = PTR_ERR(handle);
4403                         break;
4404                 }
4405                 ret = ext4_map_blocks(handle, inode, &map, flags);
4406                 if (ret <= 0) {
4407 #ifdef EXT4FS_DEBUG
4408                         WARN_ON(ret <= 0);
4409                         printk(KERN_ERR "%s: ext4_ext_map_blocks "
4410                                     "returned error inode#%lu, block=%u, "
4411                                     "max_blocks=%u", __func__,
4412                                     inode->i_ino, map.m_lblk, max_blocks);
4413 #endif
4414                         ext4_mark_inode_dirty(handle, inode);
4415                         ret2 = ext4_journal_stop(handle);
4416                         break;
4417                 }
4418                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4419                                                 blkbits) >> blkbits))
4420                         new_size = offset + len;
4421                 else
4422                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4423
4424                 ext4_falloc_update_inode(inode, mode, new_size,
4425                                          (map.m_flags & EXT4_MAP_NEW));
4426                 ext4_mark_inode_dirty(handle, inode);
4427                 ret2 = ext4_journal_stop(handle);
4428                 if (ret2)
4429                         break;
4430         }
4431         if (ret == -ENOSPC &&
4432                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4433                 ret = 0;
4434                 goto retry;
4435         }
4436         mutex_unlock(&inode->i_mutex);
4437         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4438                                 ret > 0 ? ret2 : ret);
4439         return ret > 0 ? ret2 : ret;
4440 }
4441
4442 /*
4443  * This function convert a range of blocks to written extents
4444  * The caller of this function will pass the start offset and the size.
4445  * all unwritten extents within this range will be converted to
4446  * written extents.
4447  *
4448  * This function is called from the direct IO end io call back
4449  * function, to convert the fallocated extents after IO is completed.
4450  * Returns 0 on success.
4451  */
4452 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4453                                     ssize_t len)
4454 {
4455         handle_t *handle;
4456         unsigned int max_blocks;
4457         int ret = 0;
4458         int ret2 = 0;
4459         struct ext4_map_blocks map;
4460         unsigned int credits, blkbits = inode->i_blkbits;
4461
4462         map.m_lblk = offset >> blkbits;
4463         /*
4464          * We can't just convert len to max_blocks because
4465          * If blocksize = 4096 offset = 3072 and len = 2048
4466          */
4467         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4468                       map.m_lblk);
4469         /*
4470          * credits to insert 1 extent into extent tree
4471          */
4472         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4473         while (ret >= 0 && ret < max_blocks) {
4474                 map.m_lblk += ret;
4475                 map.m_len = (max_blocks -= ret);
4476                 handle = ext4_journal_start(inode, credits);
4477                 if (IS_ERR(handle)) {
4478                         ret = PTR_ERR(handle);
4479                         break;
4480                 }
4481                 ret = ext4_map_blocks(handle, inode, &map,
4482                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4483                 if (ret <= 0) {
4484                         WARN_ON(ret <= 0);
4485                         printk(KERN_ERR "%s: ext4_ext_map_blocks "
4486                                     "returned error inode#%lu, block=%u, "
4487                                     "max_blocks=%u", __func__,
4488                                     inode->i_ino, map.m_lblk, map.m_len);
4489                 }
4490                 ext4_mark_inode_dirty(handle, inode);
4491                 ret2 = ext4_journal_stop(handle);
4492                 if (ret <= 0 || ret2 )
4493                         break;
4494         }
4495         return ret > 0 ? ret2 : ret;
4496 }
4497
4498 /*
4499  * Callback function called for each extent to gather FIEMAP information.
4500  */
4501 static int ext4_ext_fiemap_cb(struct inode *inode, ext4_lblk_t next,
4502                        struct ext4_ext_cache *newex, struct ext4_extent *ex,
4503                        void *data)
4504 {
4505         __u64   logical;
4506         __u64   physical;
4507         __u64   length;
4508         __u32   flags = 0;
4509         int             ret = 0;
4510         struct fiemap_extent_info *fieinfo = data;
4511         unsigned char blksize_bits;
4512
4513         blksize_bits = inode->i_sb->s_blocksize_bits;
4514         logical = (__u64)newex->ec_block << blksize_bits;
4515
4516         if (newex->ec_start == 0) {
4517                 /*
4518                  * No extent in extent-tree contains block @newex->ec_start,
4519                  * then the block may stay in 1)a hole or 2)delayed-extent.
4520                  *
4521                  * Holes or delayed-extents are processed as follows.
4522                  * 1. lookup dirty pages with specified range in pagecache.
4523                  *    If no page is got, then there is no delayed-extent and
4524                  *    return with EXT_CONTINUE.
4525                  * 2. find the 1st mapped buffer,
4526                  * 3. check if the mapped buffer is both in the request range
4527                  *    and a delayed buffer. If not, there is no delayed-extent,
4528                  *    then return.
4529                  * 4. a delayed-extent is found, the extent will be collected.
4530                  */
4531                 ext4_lblk_t     end = 0;
4532                 pgoff_t         last_offset;
4533                 pgoff_t         offset;
4534                 pgoff_t         index;
4535                 pgoff_t         start_index = 0;
4536                 struct page     **pages = NULL;
4537                 struct buffer_head *bh = NULL;
4538                 struct buffer_head *head = NULL;
4539                 unsigned int nr_pages = PAGE_SIZE / sizeof(struct page *);
4540
4541                 pages = kmalloc(PAGE_SIZE, GFP_KERNEL);
4542                 if (pages == NULL)
4543                         return -ENOMEM;
4544
4545                 offset = logical >> PAGE_SHIFT;
4546 repeat:
4547                 last_offset = offset;
4548                 head = NULL;
4549                 ret = find_get_pages_tag(inode->i_mapping, &offset,
4550                                         PAGECACHE_TAG_DIRTY, nr_pages, pages);
4551
4552                 if (!(flags & FIEMAP_EXTENT_DELALLOC)) {
4553                         /* First time, try to find a mapped buffer. */
4554                         if (ret == 0) {
4555 out:
4556                                 for (index = 0; index < ret; index++)
4557                                         page_cache_release(pages[index]);
4558                                 /* just a hole. */
4559                                 kfree(pages);
4560                                 return EXT_CONTINUE;
4561                         }
4562                         index = 0;
4563
4564 next_page:
4565                         /* Try to find the 1st mapped buffer. */
4566                         end = ((__u64)pages[index]->index << PAGE_SHIFT) >>
4567                                   blksize_bits;
4568                         if (!page_has_buffers(pages[index]))
4569                                 goto out;
4570                         head = page_buffers(pages[index]);
4571                         if (!head)
4572                                 goto out;
4573
4574                         index++;
4575                         bh = head;
4576                         do {
4577                                 if (end >= newex->ec_block +
4578                                         newex->ec_len)
4579                                         /* The buffer is out of
4580                                          * the request range.
4581                                          */
4582                                         goto out;
4583
4584                                 if (buffer_mapped(bh) &&
4585                                     end >= newex->ec_block) {
4586                                         start_index = index - 1;
4587                                         /* get the 1st mapped buffer. */
4588                                         goto found_mapped_buffer;
4589                                 }
4590
4591                                 bh = bh->b_this_page;
4592                                 end++;
4593                         } while (bh != head);
4594
4595                         /* No mapped buffer in the range found in this page,
4596                          * We need to look up next page.
4597                          */
4598                         if (index >= ret) {
4599                                 /* There is no page left, but we need to limit
4600                                  * newex->ec_len.
4601                                  */
4602                                 newex->ec_len = end - newex->ec_block;
4603                                 goto out;
4604                         }
4605                         goto next_page;
4606                 } else {
4607                         /*Find contiguous delayed buffers. */
4608                         if (ret > 0 && pages[0]->index == last_offset)
4609                                 head = page_buffers(pages[0]);
4610                         bh = head;
4611                         index = 1;
4612                         start_index = 0;
4613                 }
4614
4615 found_mapped_buffer:
4616                 if (bh != NULL && buffer_delay(bh)) {
4617                         /* 1st or contiguous delayed buffer found. */
4618                         if (!(flags & FIEMAP_EXTENT_DELALLOC)) {
4619                                 /*
4620                                  * 1st delayed buffer found, record
4621                                  * the start of extent.
4622                                  */
4623                                 flags |= FIEMAP_EXTENT_DELALLOC;
4624                                 newex->ec_block = end;
4625                                 logical = (__u64)end << blksize_bits;
4626                         }
4627                         /* Find contiguous delayed buffers. */
4628                         do {
4629                                 if (!buffer_delay(bh))
4630                                         goto found_delayed_extent;
4631                                 bh = bh->b_this_page;
4632                                 end++;
4633                         } while (bh != head);
4634
4635                         for (; index < ret; index++) {
4636                                 if (!page_has_buffers(pages[index])) {
4637                                         bh = NULL;
4638                                         break;
4639                                 }
4640                                 head = page_buffers(pages[index]);
4641                                 if (!head) {
4642                                         bh = NULL;
4643                                         break;
4644                                 }
4645
4646                                 if (pages[index]->index !=
4647                                     pages[start_index]->index + index
4648                                     - start_index) {
4649                                         /* Blocks are not contiguous. */
4650                                         bh = NULL;
4651                                         break;
4652                                 }
4653                                 bh = head;
4654                                 do {
4655                                         if (!buffer_delay(bh))
4656                                                 /* Delayed-extent ends. */
4657                                                 goto found_delayed_extent;
4658                                         bh = bh->b_this_page;
4659                                         end++;
4660                                 } while (bh != head);
4661                         }
4662                 } else if (!(flags & FIEMAP_EXTENT_DELALLOC))
4663                         /* a hole found. */
4664                         goto out;
4665
4666 found_delayed_extent:
4667                 newex->ec_len = min(end - newex->ec_block,
4668                                                 (ext4_lblk_t)EXT_INIT_MAX_LEN);
4669                 if (ret == nr_pages && bh != NULL &&
4670                         newex->ec_len < EXT_INIT_MAX_LEN &&
4671                         buffer_delay(bh)) {
4672                         /* Have not collected an extent and continue. */
4673                         for (index = 0; index < ret; index++)
4674                                 page_cache_release(pages[index]);
4675                         goto repeat;
4676                 }
4677
4678                 for (index = 0; index < ret; index++)
4679                         page_cache_release(pages[index]);
4680                 kfree(pages);
4681         }
4682
4683         physical = (__u64)newex->ec_start << blksize_bits;
4684         length =   (__u64)newex->ec_len << blksize_bits;
4685
4686         if (ex && ext4_ext_is_uninitialized(ex))
4687                 flags |= FIEMAP_EXTENT_UNWRITTEN;
4688
4689         if (next == EXT_MAX_BLOCKS)
4690                 flags |= FIEMAP_EXTENT_LAST;
4691
4692         ret = fiemap_fill_next_extent(fieinfo, logical, physical,
4693                                         length, flags);
4694         if (ret < 0)
4695                 return ret;
4696         if (ret == 1)
4697                 return EXT_BREAK;
4698         return EXT_CONTINUE;
4699 }
4700 /* fiemap flags we can handle specified here */
4701 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4702
4703 static int ext4_xattr_fiemap(struct inode *inode,
4704                                 struct fiemap_extent_info *fieinfo)
4705 {
4706         __u64 physical = 0;
4707         __u64 length;
4708         __u32 flags = FIEMAP_EXTENT_LAST;
4709         int blockbits = inode->i_sb->s_blocksize_bits;
4710         int error = 0;
4711
4712         /* in-inode? */
4713         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4714                 struct ext4_iloc iloc;
4715                 int offset;     /* offset of xattr in inode */
4716
4717                 error = ext4_get_inode_loc(inode, &iloc);
4718                 if (error)
4719                         return error;
4720                 physical = iloc.bh->b_blocknr << blockbits;
4721                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4722                                 EXT4_I(inode)->i_extra_isize;
4723                 physical += offset;
4724                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4725                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4726                 brelse(iloc.bh);
4727         } else { /* external block */
4728                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4729                 length = inode->i_sb->s_blocksize;
4730         }
4731
4732         if (physical)
4733                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4734                                                 length, flags);
4735         return (error < 0 ? error : 0);
4736 }
4737
4738 /*
4739  * ext4_ext_punch_hole
4740  *
4741  * Punches a hole of "length" bytes in a file starting
4742  * at byte "offset"
4743  *
4744  * @inode:  The inode of the file to punch a hole in
4745  * @offset: The starting byte offset of the hole
4746  * @length: The length of the hole
4747  *
4748  * Returns the number of blocks removed or negative on err
4749  */
4750 int ext4_ext_punch_hole(struct file *file, loff_t offset, loff_t length)
4751 {
4752         struct inode *inode = file->f_path.dentry->d_inode;
4753         struct super_block *sb = inode->i_sb;
4754         struct ext4_ext_cache cache_ex;
4755         ext4_lblk_t first_block, last_block, num_blocks, iblock, max_blocks;
4756         struct address_space *mapping = inode->i_mapping;
4757         struct ext4_map_blocks map;
4758         handle_t *handle;
4759         loff_t first_page, last_page, page_len;
4760         loff_t first_page_offset, last_page_offset;
4761         int ret, credits, blocks_released, err = 0;
4762
4763         /* No need to punch hole beyond i_size */
4764         if (offset >= inode->i_size)
4765                 return 0;
4766
4767         /*
4768          * If the hole extends beyond i_size, set the hole
4769          * to end after the page that contains i_size
4770          */
4771         if (offset + length > inode->i_size) {
4772                 length = inode->i_size +
4773                    PAGE_CACHE_SIZE - (inode->i_size & (PAGE_CACHE_SIZE - 1)) -
4774                    offset;
4775         }
4776
4777         first_block = (offset + sb->s_blocksize - 1) >>
4778                 EXT4_BLOCK_SIZE_BITS(sb);
4779         last_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4780
4781         first_page = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
4782         last_page = (offset + length) >> PAGE_CACHE_SHIFT;
4783
4784         first_page_offset = first_page << PAGE_CACHE_SHIFT;
4785         last_page_offset = last_page << PAGE_CACHE_SHIFT;
4786
4787         /*
4788          * Write out all dirty pages to avoid race conditions
4789          * Then release them.
4790          */
4791         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4792                 err = filemap_write_and_wait_range(mapping,
4793                         offset, offset + length - 1);
4794
4795                 if (err)
4796                         return err;
4797         }
4798
4799         /* Now release the pages */
4800         if (last_page_offset > first_page_offset) {
4801                 truncate_inode_pages_range(mapping, first_page_offset,
4802                                            last_page_offset-1);
4803         }
4804
4805         /* finish any pending end_io work */
4806         ext4_flush_completed_IO(inode);
4807
4808         credits = ext4_writepage_trans_blocks(inode);
4809         handle = ext4_journal_start(inode, credits);
4810         if (IS_ERR(handle))
4811                 return PTR_ERR(handle);
4812
4813         err = ext4_orphan_add(handle, inode);
4814         if (err)
4815                 goto out;
4816
4817         /*
4818          * Now we need to zero out the non-page-aligned data in the
4819          * pages at the start and tail of the hole, and unmap the buffer
4820          * heads for the block aligned regions of the page that were
4821          * completely zeroed.
4822          */
4823         if (first_page > last_page) {
4824                 /*
4825                  * If the file space being truncated is contained within a page
4826                  * just zero out and unmap the middle of that page
4827                  */
4828                 err = ext4_discard_partial_page_buffers(handle,
4829                         mapping, offset, length, 0);
4830
4831                 if (err)
4832                         goto out;
4833         } else {
4834                 /*
4835                  * zero out and unmap the partial page that contains
4836                  * the start of the hole
4837                  */
4838                 page_len  = first_page_offset - offset;
4839                 if (page_len > 0) {
4840                         err = ext4_discard_partial_page_buffers(handle, mapping,
4841                                                    offset, page_len, 0);
4842                         if (err)
4843                                 goto out;
4844                 }
4845
4846                 /*
4847                  * zero out and unmap the partial page that contains
4848                  * the end of the hole
4849                  */
4850                 page_len = offset + length - last_page_offset;
4851                 if (page_len > 0) {
4852                         err = ext4_discard_partial_page_buffers(handle, mapping,
4853                                         last_page_offset, page_len, 0);
4854                         if (err)
4855                                 goto out;
4856                 }
4857         }
4858
4859
4860         /*
4861          * If i_size is contained in the last page, we need to
4862          * unmap and zero the partial page after i_size
4863          */
4864         if (inode->i_size >> PAGE_CACHE_SHIFT == last_page &&
4865            inode->i_size % PAGE_CACHE_SIZE != 0) {
4866
4867                 page_len = PAGE_CACHE_SIZE -
4868                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4869
4870                 if (page_len > 0) {
4871                         err = ext4_discard_partial_page_buffers(handle,
4872                           mapping, inode->i_size, page_len, 0);
4873
4874                         if (err)
4875                                 goto out;
4876                 }
4877         }
4878
4879         /* If there are no blocks to remove, return now */
4880         if (first_block >= last_block)
4881                 goto out;
4882
4883         down_write(&EXT4_I(inode)->i_data_sem);
4884         ext4_ext_invalidate_cache(inode);
4885         ext4_discard_preallocations(inode);
4886
4887         /*
4888          * Loop over all the blocks and identify blocks
4889          * that need to be punched out
4890          */
4891         iblock = first_block;
4892         blocks_released = 0;
4893         while (iblock < last_block) {
4894                 max_blocks = last_block - iblock;
4895                 num_blocks = 1;
4896                 memset(&map, 0, sizeof(map));
4897                 map.m_lblk = iblock;
4898                 map.m_len = max_blocks;
4899                 ret = ext4_ext_map_blocks(handle, inode, &map,
4900                         EXT4_GET_BLOCKS_PUNCH_OUT_EXT);
4901
4902                 if (ret > 0) {
4903                         blocks_released += ret;
4904                         num_blocks = ret;
4905                 } else if (ret == 0) {
4906                         /*
4907                          * If map blocks could not find the block,
4908                          * then it is in a hole.  If the hole was
4909                          * not already cached, then map blocks should
4910                          * put it in the cache.  So we can get the hole
4911                          * out of the cache
4912                          */
4913                         memset(&cache_ex, 0, sizeof(cache_ex));
4914                         if ((ext4_ext_check_cache(inode, iblock, &cache_ex)) &&
4915                                 !cache_ex.ec_start) {
4916
4917                                 /* The hole is cached */
4918                                 num_blocks = cache_ex.ec_block +
4919                                 cache_ex.ec_len - iblock;
4920
4921                         } else {
4922                                 /* The block could not be identified */
4923                                 err = -EIO;
4924                                 break;
4925                         }
4926                 } else {
4927                         /* Map blocks error */
4928                         err = ret;
4929                         break;
4930                 }
4931
4932                 if (num_blocks == 0) {
4933                         /* This condition should never happen */
4934                         ext_debug("Block lookup failed");
4935                         err = -EIO;
4936                         break;
4937                 }
4938
4939                 iblock += num_blocks;
4940         }
4941
4942         if (blocks_released > 0) {
4943                 ext4_ext_invalidate_cache(inode);
4944                 ext4_discard_preallocations(inode);
4945         }
4946
4947         if (IS_SYNC(inode))
4948                 ext4_handle_sync(handle);
4949
4950         up_write(&EXT4_I(inode)->i_data_sem);
4951
4952 out:
4953         ext4_orphan_del(handle, inode);
4954         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4955         ext4_mark_inode_dirty(handle, inode);
4956         ext4_journal_stop(handle);
4957         return err;
4958 }
4959 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4960                 __u64 start, __u64 len)
4961 {
4962         ext4_lblk_t start_blk;
4963         int error = 0;
4964
4965         /* fallback to generic here if not in extents fmt */
4966         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4967                 return generic_block_fiemap(inode, fieinfo, start, len,
4968                         ext4_get_block);
4969
4970         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4971                 return -EBADR;
4972
4973         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4974                 error = ext4_xattr_fiemap(inode, fieinfo);
4975         } else {
4976                 ext4_lblk_t len_blks;
4977                 __u64 last_blk;
4978
4979                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4980                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4981                 if (last_blk >= EXT_MAX_BLOCKS)
4982                         last_blk = EXT_MAX_BLOCKS-1;
4983                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4984
4985                 /*
4986                  * Walk the extent tree gathering extent information.
4987                  * ext4_ext_fiemap_cb will push extents back to user.
4988                  */
4989                 error = ext4_ext_walk_space(inode, start_blk, len_blks,
4990                                           ext4_ext_fiemap_cb, fieinfo);
4991         }
4992
4993         return error;
4994 }