fs/ext4/extents.c

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