fs/ext4/extents.c

   1 /*
   2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3  * Written by Alex Tomas <alex@clusterfs.com>
   4  *
   5  * Architecture independence:
   6  *   Copyright (c) 2005, Bull S.A.
   7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public Licens
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <linux/falloc.h>
  41 #include <asm/uaccess.h>
  42 #include <linux/fiemap.h>
  43 #include "ext4_jbd2.h"
  44 #include "ext4_extents.h"
  45 #include "xattr.h"
  46
  47 #include <trace/events/ext4.h>
  48
  49 /*
  50  * used by extent splitting.
  51  */
  52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  53                                         due to ENOSPC */
  54 #define EXT4_EXT_MARK_UNINIT1   0x2  /* mark first half uninitialized */
  55 #define EXT4_EXT_MARK_UNINIT2   0x4  /* mark second half uninitialized */
  56
  57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  59
  60 static __le32 ext4_extent_block_csum(struct inode *inode,
  61                                      struct ext4_extent_header *eh)
  62 {
  63         struct ext4_inode_info *ei = EXT4_I(inode);
  64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  65         __u32 csum;
  66
  67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  68                            EXT4_EXTENT_TAIL_OFFSET(eh));
  69         return cpu_to_le32(csum);
  70 }
  71
  72 static int ext4_extent_block_csum_verify(struct inode *inode,
  73                                          struct ext4_extent_header *eh)
  74 {
  75         struct ext4_extent_tail *et;
  76
  77         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  78                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  79                 return 1;
  80
  81         et = find_ext4_extent_tail(eh);
  82         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  83                 return 0;
  84         return 1;
  85 }
  86
  87 static void ext4_extent_block_csum_set(struct inode *inode,
  88                                        struct ext4_extent_header *eh)
  89 {
  90         struct ext4_extent_tail *et;
  91
  92         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  93                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  94                 return;
  95
  96         et = find_ext4_extent_tail(eh);
  97         et->et_checksum = ext4_extent_block_csum(inode, eh);
  98 }
  99
 100 static int ext4_split_extent(handle_t *handle,
 101                                 struct inode *inode,
 102                                 struct ext4_ext_path *path,
 103                                 struct ext4_map_blocks *map,
 104                                 int split_flag,
 105                                 int flags);
 106
 107 static int ext4_split_extent_at(handle_t *handle,
 108                              struct inode *inode,
 109                              struct ext4_ext_path *path,
 110                              ext4_lblk_t split,
 111                              int split_flag,
 112                              int flags);
 113
 114 static int ext4_find_delayed_extent(struct inode *inode,
 115                                     struct extent_status *newes);
 116
 117 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 118                                             struct inode *inode,
 119                                             int needed)
 120 {
 121         int err;
 122
 123         if (!ext4_handle_valid(handle))
 124                 return 0;
 125         if (handle->h_buffer_credits > needed)
 126                 return 0;
 127         err = ext4_journal_extend(handle, needed);
 128         if (err <= 0)
 129                 return err;
 130         err = ext4_truncate_restart_trans(handle, inode, needed);
 131         if (err == 0)
 132                 err = -EAGAIN;
 133
 134         return err;
 135 }
 136
 137 /*
 138  * could return:
 139  *  - EROFS
 140  *  - ENOMEM
 141  */
 142 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 143                                 struct ext4_ext_path *path)
 144 {
 145         if (path->p_bh) {
 146                 /* path points to block */
 147                 return ext4_journal_get_write_access(handle, path->p_bh);
 148         }
 149         /* path points to leaf/index in inode body */
 150         /* we use in-core data, no need to protect them */
 151         return 0;
 152 }
 153
 154 /*
 155  * could return:
 156  *  - EROFS
 157  *  - ENOMEM
 158  *  - EIO
 159  */
 160 #define ext4_ext_dirty(handle, inode, path) \
 161                 __ext4_ext_dirty(__func__, __LINE__, (handle), (inode), (path))
 162 static int __ext4_ext_dirty(const char *where, unsigned int line,
 163                             handle_t *handle, struct inode *inode,
 164                             struct ext4_ext_path *path)
 165 {
 166         int err;
 167         if (path->p_bh) {
 168                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 169                 /* path points to block */
 170                 err = __ext4_handle_dirty_metadata(where, line, handle,
 171                                                    inode, path->p_bh);
 172         } else {
 173                 /* path points to leaf/index in inode body */
 174                 err = ext4_mark_inode_dirty(handle, inode);
 175         }
 176         return err;
 177 }
 178
 179 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 180                               struct ext4_ext_path *path,
 181                               ext4_lblk_t block)
 182 {
 183         if (path) {
 184                 int depth = path->p_depth;
 185                 struct ext4_extent *ex;
 186
 187                 /*
 188                  * Try to predict block placement assuming that we are
 189                  * filling in a file which will eventually be
 190                  * non-sparse --- i.e., in the case of libbfd writing
 191                  * an ELF object sections out-of-order but in a way
 192                  * the eventually results in a contiguous object or
 193                  * executable file, or some database extending a table
 194                  * space file.  However, this is actually somewhat
 195                  * non-ideal if we are writing a sparse file such as
 196                  * qemu or KVM writing a raw image file that is going
 197                  * to stay fairly sparse, since it will end up
 198                  * fragmenting the file system's free space.  Maybe we
 199                  * should have some hueristics or some way to allow
 200                  * userspace to pass a hint to file system,
 201                  * especially if the latter case turns out to be
 202                  * common.
 203                  */
 204                 ex = path[depth].p_ext;
 205                 if (ex) {
 206                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 207                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 208
 209                         if (block > ext_block)
 210                                 return ext_pblk + (block - ext_block);
 211                         else
 212                                 return ext_pblk - (ext_block - block);
 213                 }
 214
 215                 /* it looks like index is empty;
 216                  * try to find starting block from index itself */
 217                 if (path[depth].p_bh)
 218                         return path[depth].p_bh->b_blocknr;
 219         }
 220
 221         /* OK. use inode's group */
 222         return ext4_inode_to_goal_block(inode);
 223 }
 224
 225 /*
 226  * Allocation for a meta data block
 227  */
 228 static ext4_fsblk_t
 229 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 230                         struct ext4_ext_path *path,
 231                         struct ext4_extent *ex, int *err, unsigned int flags)
 232 {
 233         ext4_fsblk_t goal, newblock;
 234
 235         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 236         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 237                                         NULL, err);
 238         return newblock;
 239 }
 240
 241 static inline int ext4_ext_space_block(struct inode *inode, int check)
 242 {
 243         int size;
 244
 245         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 246                         / sizeof(struct ext4_extent);
 247 #ifdef AGGRESSIVE_TEST
 248         if (!check && size > 6)
 249                 size = 6;
 250 #endif
 251         return size;
 252 }
 253
 254 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 255 {
 256         int size;
 257
 258         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 259                         / sizeof(struct ext4_extent_idx);
 260 #ifdef AGGRESSIVE_TEST
 261         if (!check && size > 5)
 262                 size = 5;
 263 #endif
 264         return size;
 265 }
 266
 267 static inline int ext4_ext_space_root(struct inode *inode, int check)
 268 {
 269         int size;
 270
 271         size = sizeof(EXT4_I(inode)->i_data);
 272         size -= sizeof(struct ext4_extent_header);
 273         size /= sizeof(struct ext4_extent);
 274 #ifdef AGGRESSIVE_TEST
 275         if (!check && size > 3)
 276                 size = 3;
 277 #endif
 278         return size;
 279 }
 280
 281 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 282 {
 283         int size;
 284
 285         size = sizeof(EXT4_I(inode)->i_data);
 286         size -= sizeof(struct ext4_extent_header);
 287         size /= sizeof(struct ext4_extent_idx);
 288 #ifdef AGGRESSIVE_TEST
 289         if (!check && size > 4)
 290                 size = 4;
 291 #endif
 292         return size;
 293 }
 294
 295 /*
 296  * Calculate the number of metadata blocks needed
 297  * to allocate @blocks
 298  * Worse case is one block per extent
 299  */
 300 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 301 {
 302         struct ext4_inode_info *ei = EXT4_I(inode);
 303         int idxs;
 304
 305         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 306                 / sizeof(struct ext4_extent_idx));
 307
 308         /*
 309          * If the new delayed allocation block is contiguous with the
 310          * previous da block, it can share index blocks with the
 311          * previous block, so we only need to allocate a new index
 312          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 313          * an additional index block, and at ldxs**3 blocks, yet
 314          * another index blocks.
 315          */
 316         if (ei->i_da_metadata_calc_len &&
 317             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 318                 int num = 0;
 319
 320                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 321                         num++;
 322                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 323                         num++;
 324                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 325                         num++;
 326                         ei->i_da_metadata_calc_len = 0;
 327                 } else
 328                         ei->i_da_metadata_calc_len++;
 329                 ei->i_da_metadata_calc_last_lblock++;
 330                 return num;
 331         }
 332
 333         /*
 334          * In the worst case we need a new set of index blocks at
 335          * every level of the inode's extent tree.
 336          */
 337         ei->i_da_metadata_calc_len = 1;
 338         ei->i_da_metadata_calc_last_lblock = lblock;
 339         return ext_depth(inode) + 1;
 340 }
 341
 342 static int
 343 ext4_ext_max_entries(struct inode *inode, int depth)
 344 {
 345         int max;
 346
 347         if (depth == ext_depth(inode)) {
 348                 if (depth == 0)
 349                         max = ext4_ext_space_root(inode, 1);
 350                 else
 351                         max = ext4_ext_space_root_idx(inode, 1);
 352         } else {
 353                 if (depth == 0)
 354                         max = ext4_ext_space_block(inode, 1);
 355                 else
 356                         max = ext4_ext_space_block_idx(inode, 1);
 357         }
 358
 359         return max;
 360 }
 361
 362 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 363 {
 364         ext4_fsblk_t block = ext4_ext_pblock(ext);
 365         int len = ext4_ext_get_actual_len(ext);
 366
 367         if (len == 0)
 368                 return 0;
 369         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 370 }
 371
 372 static int ext4_valid_extent_idx(struct inode *inode,
 373                                 struct ext4_extent_idx *ext_idx)
 374 {
 375         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 376
 377         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 378 }
 379
 380 static int ext4_valid_extent_entries(struct inode *inode,
 381                                 struct ext4_extent_header *eh,
 382                                 int depth)
 383 {
 384         unsigned short entries;
 385         if (eh->eh_entries == 0)
 386                 return 1;
 387
 388         entries = le16_to_cpu(eh->eh_entries);
 389
 390         if (depth == 0) {
 391                 /* leaf entries */
 392                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 393                 while (entries) {
 394                         if (!ext4_valid_extent(inode, ext))
 395                                 return 0;
 396                         ext++;
 397                         entries--;
 398                 }
 399         } else {
 400                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 401                 while (entries) {
 402                         if (!ext4_valid_extent_idx(inode, ext_idx))
 403                                 return 0;
 404                         ext_idx++;
 405                         entries--;
 406                 }
 407         }
 408         return 1;
 409 }
 410
 411 static int __ext4_ext_check(const char *function, unsigned int line,
 412                             struct inode *inode, struct ext4_extent_header *eh,
 413                             int depth)
 414 {
 415         const char *error_msg;
 416         int max = 0;
 417
 418         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 419                 error_msg = "invalid magic";
 420                 goto corrupted;
 421         }
 422         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 423                 error_msg = "unexpected eh_depth";
 424                 goto corrupted;
 425         }
 426         if (unlikely(eh->eh_max == 0)) {
 427                 error_msg = "invalid eh_max";
 428                 goto corrupted;
 429         }
 430         max = ext4_ext_max_entries(inode, depth);
 431         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 432                 error_msg = "too large eh_max";
 433                 goto corrupted;
 434         }
 435         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 436                 error_msg = "invalid eh_entries";
 437                 goto corrupted;
 438         }
 439         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 440                 error_msg = "invalid extent entries";
 441                 goto corrupted;
 442         }
 443         /* Verify checksum on non-root extent tree nodes */
 444         if (ext_depth(inode) != depth &&
 445             !ext4_extent_block_csum_verify(inode, eh)) {
 446                 error_msg = "extent tree corrupted";
 447                 goto corrupted;
 448         }
 449         return 0;
 450
 451 corrupted:
 452         ext4_error_inode(inode, function, line, 0,
 453                         "bad header/extent: %s - magic %x, "
 454                         "entries %u, max %u(%u), depth %u(%u)",
 455                         error_msg, le16_to_cpu(eh->eh_magic),
 456                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 457                         max, le16_to_cpu(eh->eh_depth), depth);
 458
 459         return -EIO;
 460 }
 461
 462 #define ext4_ext_check(inode, eh, depth)        \
 463         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
 464
 465 int ext4_ext_check_inode(struct inode *inode)
 466 {
 467         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
 468 }
 469
 470 static int __ext4_ext_check_block(const char *function, unsigned int line,
 471                                   struct inode *inode,
 472                                   struct ext4_extent_header *eh,
 473                                   int depth,
 474                                   struct buffer_head *bh)
 475 {
 476         int ret;
 477
 478         if (buffer_verified(bh))
 479                 return 0;
 480         ret = ext4_ext_check(inode, eh, depth);
 481         if (ret)
 482                 return ret;
 483         set_buffer_verified(bh);
 484         return ret;
 485 }
 486
 487 #define ext4_ext_check_block(inode, eh, depth, bh)      \
 488         __ext4_ext_check_block(__func__, __LINE__, inode, eh, depth, bh)
 489
 490 #ifdef EXT_DEBUG
 491 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 492 {
 493         int k, l = path->p_depth;
 494
 495         ext_debug("path:");
 496         for (k = 0; k <= l; k++, path++) {
 497                 if (path->p_idx) {
 498                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 499                             ext4_idx_pblock(path->p_idx));
 500                 } else if (path->p_ext) {
 501                         ext_debug("  %d:[%d]%d:%llu ",
 502                                   le32_to_cpu(path->p_ext->ee_block),
 503                                   ext4_ext_is_uninitialized(path->p_ext),
 504                                   ext4_ext_get_actual_len(path->p_ext),
 505                                   ext4_ext_pblock(path->p_ext));
 506                 } else
 507                         ext_debug("  []");
 508         }
 509         ext_debug("\n");
 510 }
 511
 512 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 513 {
 514         int depth = ext_depth(inode);
 515         struct ext4_extent_header *eh;
 516         struct ext4_extent *ex;
 517         int i;
 518
 519         if (!path)
 520                 return;
 521
 522         eh = path[depth].p_hdr;
 523         ex = EXT_FIRST_EXTENT(eh);
 524
 525         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 526
 527         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 528                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 529                           ext4_ext_is_uninitialized(ex),
 530                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 531         }
 532         ext_debug("\n");
 533 }
 534
 535 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 536                         ext4_fsblk_t newblock, int level)
 537 {
 538         int depth = ext_depth(inode);
 539         struct ext4_extent *ex;
 540
 541         if (depth != level) {
 542                 struct ext4_extent_idx *idx;
 543                 idx = path[level].p_idx;
 544                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 545                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 546                                         le32_to_cpu(idx->ei_block),
 547                                         ext4_idx_pblock(idx),
 548                                         newblock);
 549                         idx++;
 550                 }
 551
 552                 return;
 553         }
 554
 555         ex = path[depth].p_ext;
 556         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 557                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 558                                 le32_to_cpu(ex->ee_block),
 559                                 ext4_ext_pblock(ex),
 560                                 ext4_ext_is_uninitialized(ex),
 561                                 ext4_ext_get_actual_len(ex),
 562                                 newblock);
 563                 ex++;
 564         }
 565 }
 566
 567 #else
 568 #define ext4_ext_show_path(inode, path)
 569 #define ext4_ext_show_leaf(inode, path)
 570 #define ext4_ext_show_move(inode, path, newblock, level)
 571 #endif
 572
 573 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 574 {
 575         int depth = path->p_depth;
 576         int i;
 577
 578         for (i = 0; i <= depth; i++, path++)
 579                 if (path->p_bh) {
 580                         brelse(path->p_bh);
 581                         path->p_bh = NULL;
 582                 }
 583 }
 584
 585 /*
 586  * ext4_ext_binsearch_idx:
 587  * binary search for the closest index of the given block
 588  * the header must be checked before calling this
 589  */
 590 static void
 591 ext4_ext_binsearch_idx(struct inode *inode,
 592                         struct ext4_ext_path *path, ext4_lblk_t block)
 593 {
 594         struct ext4_extent_header *eh = path->p_hdr;
 595         struct ext4_extent_idx *r, *l, *m;
 596
 597
 598         ext_debug("binsearch for %u(idx):  ", block);
 599
 600         l = EXT_FIRST_INDEX(eh) + 1;
 601         r = EXT_LAST_INDEX(eh);
 602         while (l <= r) {
 603                 m = l + (r - l) / 2;
 604                 if (block < le32_to_cpu(m->ei_block))
 605                         r = m - 1;
 606                 else
 607                         l = m + 1;
 608                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 609                                 m, le32_to_cpu(m->ei_block),
 610                                 r, le32_to_cpu(r->ei_block));
 611         }
 612
 613         path->p_idx = l - 1;
 614         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 615                   ext4_idx_pblock(path->p_idx));
 616
 617 #ifdef CHECK_BINSEARCH
 618         {
 619                 struct ext4_extent_idx *chix, *ix;
 620                 int k;
 621
 622                 chix = ix = EXT_FIRST_INDEX(eh);
 623                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 624                   if (k != 0 &&
 625                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 626                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 627                                        "first=0x%p\n", k,
 628                                        ix, EXT_FIRST_INDEX(eh));
 629                                 printk(KERN_DEBUG "%u <= %u\n",
 630                                        le32_to_cpu(ix->ei_block),
 631                                        le32_to_cpu(ix[-1].ei_block));
 632                         }
 633                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 634                                            <= le32_to_cpu(ix[-1].ei_block));
 635                         if (block < le32_to_cpu(ix->ei_block))
 636                                 break;
 637                         chix = ix;
 638                 }
 639                 BUG_ON(chix != path->p_idx);
 640         }
 641 #endif
 642
 643 }
 644
 645 /*
 646  * ext4_ext_binsearch:
 647  * binary search for closest extent of the given block
 648  * the header must be checked before calling this
 649  */
 650 static void
 651 ext4_ext_binsearch(struct inode *inode,
 652                 struct ext4_ext_path *path, ext4_lblk_t block)
 653 {
 654         struct ext4_extent_header *eh = path->p_hdr;
 655         struct ext4_extent *r, *l, *m;
 656
 657         if (eh->eh_entries == 0) {
 658                 /*
 659                  * this leaf is empty:
 660                  * we get such a leaf in split/add case
 661                  */
 662                 return;
 663         }
 664
 665         ext_debug("binsearch for %u:  ", block);
 666
 667         l = EXT_FIRST_EXTENT(eh) + 1;
 668         r = EXT_LAST_EXTENT(eh);
 669
 670         while (l <= r) {
 671                 m = l + (r - l) / 2;
 672                 if (block < le32_to_cpu(m->ee_block))
 673                         r = m - 1;
 674                 else
 675                         l = m + 1;
 676                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 677                                 m, le32_to_cpu(m->ee_block),
 678                                 r, le32_to_cpu(r->ee_block));
 679         }
 680
 681         path->p_ext = l - 1;
 682         ext_debug("  -> %d:%llu:[%d]%d ",
 683                         le32_to_cpu(path->p_ext->ee_block),
 684                         ext4_ext_pblock(path->p_ext),
 685                         ext4_ext_is_uninitialized(path->p_ext),
 686                         ext4_ext_get_actual_len(path->p_ext));
 687
 688 #ifdef CHECK_BINSEARCH
 689         {
 690                 struct ext4_extent *chex, *ex;
 691                 int k;
 692
 693                 chex = ex = EXT_FIRST_EXTENT(eh);
 694                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 695                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 696                                           <= le32_to_cpu(ex[-1].ee_block));
 697                         if (block < le32_to_cpu(ex->ee_block))
 698                                 break;
 699                         chex = ex;
 700                 }
 701                 BUG_ON(chex != path->p_ext);
 702         }
 703 #endif
 704
 705 }
 706
 707 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 708 {
 709         struct ext4_extent_header *eh;
 710
 711         eh = ext_inode_hdr(inode);
 712         eh->eh_depth = 0;
 713         eh->eh_entries = 0;
 714         eh->eh_magic = EXT4_EXT_MAGIC;
 715         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 716         ext4_mark_inode_dirty(handle, inode);
 717         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         int ret;
 728
 729         eh = ext_inode_hdr(inode);
 730         depth = ext_depth(inode);
 731
 732         /* account possible depth increase */
 733         if (!path) {
 734                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 735                                 GFP_NOFS);
 736                 if (!path)
 737                         return ERR_PTR(-ENOMEM);
 738                 alloc = 1;
 739         }
 740         path[0].p_hdr = eh;
 741         path[0].p_bh = NULL;
 742
 743         i = depth;
 744         /* walk through the tree */
 745         while (i) {
 746                 ext_debug("depth %d: num %d, max %d\n",
 747                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 748
 749                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 750                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 751                 path[ppos].p_depth = i;
 752                 path[ppos].p_ext = NULL;
 753
 754                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
 755                 if (unlikely(!bh)) {
 756                         ret = -ENOMEM;
 757                         goto err;
 758                 }
 759                 if (!bh_uptodate_or_lock(bh)) {
 760                         trace_ext4_ext_load_extent(inode, block,
 761                                                 path[ppos].p_block);
 762                         ret = bh_submit_read(bh);
 763                         if (ret < 0) {
 764                                 put_bh(bh);
 765                                 goto err;
 766                         }
 767                 }
 768                 eh = ext_block_hdr(bh);
 769                 ppos++;
 770                 if (unlikely(ppos > depth)) {
 771                         put_bh(bh);
 772                         EXT4_ERROR_INODE(inode,
 773                                          "ppos %d > depth %d", ppos, depth);
 774                         ret = -EIO;
 775                         goto err;
 776                 }
 777                 path[ppos].p_bh = bh;
 778                 path[ppos].p_hdr = eh;
 779                 i--;
 780
 781                 ret = ext4_ext_check_block(inode, eh, i, bh);
 782                 if (ret < 0)
 783                         goto err;
 784         }
 785
 786         path[ppos].p_depth = i;
 787         path[ppos].p_ext = NULL;
 788         path[ppos].p_idx = NULL;
 789
 790         /* find extent */
 791         ext4_ext_binsearch(inode, path + ppos, block);
 792         /* if not an empty leaf */
 793         if (path[ppos].p_ext)
 794                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 795
 796         ext4_ext_show_path(inode, path);
 797
 798         return path;
 799
 800 err:
 801         ext4_ext_drop_refs(path);
 802         if (alloc)
 803                 kfree(path);
 804         return ERR_PTR(ret);
 805 }
 806
 807 /*
 808  * ext4_ext_insert_index:
 809  * insert new index [@logical;@ptr] into the block at @curp;
 810  * check where to insert: before @curp or after @curp
 811  */
 812 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 813                                  struct ext4_ext_path *curp,
 814                                  int logical, ext4_fsblk_t ptr)
 815 {
 816         struct ext4_extent_idx *ix;
 817         int len, err;
 818
 819         err = ext4_ext_get_access(handle, inode, curp);
 820         if (err)
 821                 return err;
 822
 823         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 824                 EXT4_ERROR_INODE(inode,
 825                                  "logical %d == ei_block %d!",
 826                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 827                 return -EIO;
 828         }
 829
 830         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 831                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 832                 EXT4_ERROR_INODE(inode,
 833                                  "eh_entries %d >= eh_max %d!",
 834                                  le16_to_cpu(curp->p_hdr->eh_entries),
 835                                  le16_to_cpu(curp->p_hdr->eh_max));
 836                 return -EIO;
 837         }
 838
 839         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 840                 /* insert after */
 841                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 842                 ix = curp->p_idx + 1;
 843         } else {
 844                 /* insert before */
 845                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 846                 ix = curp->p_idx;
 847         }
 848
 849         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 850         BUG_ON(len < 0);
 851         if (len > 0) {
 852                 ext_debug("insert new index %d: "
 853                                 "move %d indices from 0x%p to 0x%p\n",
 854                                 logical, len, ix, ix + 1);
 855                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 856         }
 857
 858         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 859                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 860                 return -EIO;
 861         }
 862
 863         ix->ei_block = cpu_to_le32(logical);
 864         ext4_idx_store_pblock(ix, ptr);
 865         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 866
 867         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 868                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 869                 return -EIO;
 870         }
 871
 872         err = ext4_ext_dirty(handle, inode, curp);
 873         ext4_std_error(inode->i_sb, err);
 874
 875         return err;
 876 }
 877
 878 /*
 879  * ext4_ext_split:
 880  * inserts new subtree into the path, using free index entry
 881  * at depth @at:
 882  * - allocates all needed blocks (new leaf and all intermediate index blocks)
 883  * - makes decision where to split
 884  * - moves remaining extents and index entries (right to the split point)
 885  *   into the newly allocated blocks
 886  * - initializes subtree
 887  */
 888 static int ext4_ext_split(handle_t *handle, struct inode *inode,
 889                           unsigned int flags,
 890                           struct ext4_ext_path *path,
 891                           struct ext4_extent *newext, int at)
 892 {
 893         struct buffer_head *bh = NULL;
 894         int depth = ext_depth(inode);
 895         struct ext4_extent_header *neh;
 896         struct ext4_extent_idx *fidx;
 897         int i = at, k, m, a;
 898         ext4_fsblk_t newblock, oldblock;
 899         __le32 border;
 900         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
 901         int err = 0;
 902
 903         /* make decision: where to split? */
 904         /* FIXME: now decision is simplest: at current extent */
 905
 906         /* if current leaf will be split, then we should use
 907          * border from split point */
 908         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
 909                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
 910                 return -EIO;
 911         }
 912         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
 913                 border = path[depth].p_ext[1].ee_block;
 914                 ext_debug("leaf will be split."
 915                                 " next leaf starts at %d\n",
 916                                   le32_to_cpu(border));
 917         } else {
 918                 border = newext->ee_block;
 919                 ext_debug("leaf will be added."
 920                                 " next leaf starts at %d\n",
 921                                 le32_to_cpu(border));
 922         }
 923
 924         /*
 925          * If error occurs, then we break processing
 926          * and mark filesystem read-only. index won't
 927          * be inserted and tree will be in consistent
 928          * state. Next mount will repair buffers too.
 929          */
 930
 931         /*
 932          * Get array to track all allocated blocks.
 933          * We need this to handle errors and free blocks
 934          * upon them.
 935          */
 936         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
 937         if (!ablocks)
 938                 return -ENOMEM;
 939
 940         /* allocate all needed blocks */
 941         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
 942         for (a = 0; a < depth - at; a++) {
 943                 newblock = ext4_ext_new_meta_block(handle, inode, path,
 944                                                    newext, &err, flags);
 945                 if (newblock == 0)
 946                         goto cleanup;
 947                 ablocks[a] = newblock;
 948         }
 949
 950         /* initialize new leaf */
 951         newblock = ablocks[--a];
 952         if (unlikely(newblock == 0)) {
 953                 EXT4_ERROR_INODE(inode, "newblock == 0!");
 954                 err = -EIO;
 955                 goto cleanup;
 956         }
 957         bh = sb_getblk(inode->i_sb, newblock);
 958         if (unlikely(!bh)) {
 959                 err = -ENOMEM;
 960                 goto cleanup;
 961         }
 962         lock_buffer(bh);
 963
 964         err = ext4_journal_get_create_access(handle, bh);
 965         if (err)
 966                 goto cleanup;
 967
 968         neh = ext_block_hdr(bh);
 969         neh->eh_entries = 0;
 970         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
 971         neh->eh_magic = EXT4_EXT_MAGIC;
 972         neh->eh_depth = 0;
 973
 974         /* move remainder of path[depth] to the new leaf */
 975         if (unlikely(path[depth].p_hdr->eh_entries !=
 976                      path[depth].p_hdr->eh_max)) {
 977                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
 978                                  path[depth].p_hdr->eh_entries,
 979                                  path[depth].p_hdr->eh_max);
 980                 err = -EIO;
 981                 goto cleanup;
 982         }
 983         /* start copy from next extent */
 984         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
 985         ext4_ext_show_move(inode, path, newblock, depth);
 986         if (m) {
 987                 struct ext4_extent *ex;
 988                 ex = EXT_FIRST_EXTENT(neh);
 989                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
 990                 le16_add_cpu(&neh->eh_entries, m);
 991         }
 992
 993         ext4_extent_block_csum_set(inode, neh);
 994         set_buffer_uptodate(bh);
 995         unlock_buffer(bh);
 996
 997         err = ext4_handle_dirty_metadata(handle, inode, bh);
 998         if (err)
 999                 goto cleanup;
1000         brelse(bh);
1001         bh = NULL;
1002
1003         /* correct old leaf */
1004         if (m) {
1005                 err = ext4_ext_get_access(handle, inode, path + depth);
1006                 if (err)
1007                         goto cleanup;
1008                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1009                 err = ext4_ext_dirty(handle, inode, path + depth);
1010                 if (err)
1011                         goto cleanup;
1012
1013         }
1014
1015         /* create intermediate indexes */
1016         k = depth - at - 1;
1017         if (unlikely(k < 0)) {
1018                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1019                 err = -EIO;
1020                 goto cleanup;
1021         }
1022         if (k)
1023                 ext_debug("create %d intermediate indices\n", k);
1024         /* insert new index into current index block */
1025         /* current depth stored in i var */
1026         i = depth - 1;
1027         while (k--) {
1028                 oldblock = newblock;
1029                 newblock = ablocks[--a];
1030                 bh = sb_getblk(inode->i_sb, newblock);
1031                 if (unlikely(!bh)) {
1032                         err = -ENOMEM;
1033                         goto cleanup;
1034                 }
1035                 lock_buffer(bh);
1036
1037                 err = ext4_journal_get_create_access(handle, bh);
1038                 if (err)
1039                         goto cleanup;
1040
1041                 neh = ext_block_hdr(bh);
1042                 neh->eh_entries = cpu_to_le16(1);
1043                 neh->eh_magic = EXT4_EXT_MAGIC;
1044                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1045                 neh->eh_depth = cpu_to_le16(depth - i);
1046                 fidx = EXT_FIRST_INDEX(neh);
1047                 fidx->ei_block = border;
1048                 ext4_idx_store_pblock(fidx, oldblock);
1049
1050                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1051                                 i, newblock, le32_to_cpu(border), oldblock);
1052
1053                 /* move remainder of path[i] to the new index block */
1054                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1055                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1056                         EXT4_ERROR_INODE(inode,
1057                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1058                                          le32_to_cpu(path[i].p_ext->ee_block));
1059                         err = -EIO;
1060                         goto cleanup;
1061                 }
1062                 /* start copy indexes */
1063                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1064                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1065                                 EXT_MAX_INDEX(path[i].p_hdr));
1066                 ext4_ext_show_move(inode, path, newblock, i);
1067                 if (m) {
1068                         memmove(++fidx, path[i].p_idx,
1069                                 sizeof(struct ext4_extent_idx) * m);
1070                         le16_add_cpu(&neh->eh_entries, m);
1071                 }
1072                 ext4_extent_block_csum_set(inode, neh);
1073                 set_buffer_uptodate(bh);
1074                 unlock_buffer(bh);
1075
1076                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1077                 if (err)
1078                         goto cleanup;
1079                 brelse(bh);
1080                 bh = NULL;
1081
1082                 /* correct old index */
1083                 if (m) {
1084                         err = ext4_ext_get_access(handle, inode, path + i);
1085                         if (err)
1086                                 goto cleanup;
1087                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1088                         err = ext4_ext_dirty(handle, inode, path + i);
1089                         if (err)
1090                                 goto cleanup;
1091                 }
1092
1093                 i--;
1094         }
1095
1096         /* insert new index */
1097         err = ext4_ext_insert_index(handle, inode, path + at,
1098                                     le32_to_cpu(border), newblock);
1099
1100 cleanup:
1101         if (bh) {
1102                 if (buffer_locked(bh))
1103                         unlock_buffer(bh);
1104                 brelse(bh);
1105         }
1106
1107         if (err) {
1108                 /* free all allocated blocks in error case */
1109                 for (i = 0; i < depth; i++) {
1110                         if (!ablocks[i])
1111                                 continue;
1112                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1113                                          EXT4_FREE_BLOCKS_METADATA);
1114                 }
1115         }
1116         kfree(ablocks);
1117
1118         return err;
1119 }
1120
1121 /*
1122  * ext4_ext_grow_indepth:
1123  * implements tree growing procedure:
1124  * - allocates new block
1125  * - moves top-level data (index block or leaf) into the new block
1126  * - initializes new top-level, creating index that points to the
1127  *   just created block
1128  */
1129 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1130                                  unsigned int flags,
1131                                  struct ext4_extent *newext)
1132 {
1133         struct ext4_extent_header *neh;
1134         struct buffer_head *bh;
1135         ext4_fsblk_t newblock;
1136         int err = 0;
1137
1138         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1139                 newext, &err, flags);
1140         if (newblock == 0)
1141                 return err;
1142
1143         bh = sb_getblk(inode->i_sb, newblock);
1144         if (unlikely(!bh))
1145                 return -ENOMEM;
1146         lock_buffer(bh);
1147
1148         err = ext4_journal_get_create_access(handle, bh);
1149         if (err) {
1150                 unlock_buffer(bh);
1151                 goto out;
1152         }
1153
1154         /* move top-level index/leaf into new block */
1155         memmove(bh->b_data, EXT4_I(inode)->i_data,
1156                 sizeof(EXT4_I(inode)->i_data));
1157
1158         /* set size of new block */
1159         neh = ext_block_hdr(bh);
1160         /* old root could have indexes or leaves
1161          * so calculate e_max right way */
1162         if (ext_depth(inode))
1163                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1164         else
1165                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1166         neh->eh_magic = EXT4_EXT_MAGIC;
1167         ext4_extent_block_csum_set(inode, neh);
1168         set_buffer_uptodate(bh);
1169         unlock_buffer(bh);
1170
1171         err = ext4_handle_dirty_metadata(handle, inode, bh);
1172         if (err)
1173                 goto out;
1174
1175         /* Update top-level index: num,max,pointer */
1176         neh = ext_inode_hdr(inode);
1177         neh->eh_entries = cpu_to_le16(1);
1178         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1179         if (neh->eh_depth == 0) {
1180                 /* Root extent block becomes index block */
1181                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1182                 EXT_FIRST_INDEX(neh)->ei_block =
1183                         EXT_FIRST_EXTENT(neh)->ee_block;
1184         }
1185         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1186                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1187                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1188                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1189
1190         le16_add_cpu(&neh->eh_depth, 1);
1191         ext4_mark_inode_dirty(handle, inode);
1192 out:
1193         brelse(bh);
1194
1195         return err;
1196 }
1197
1198 /*
1199  * ext4_ext_create_new_leaf:
1200  * finds empty index and adds new leaf.
1201  * if no free index is found, then it requests in-depth growing.
1202  */
1203 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1204                                     unsigned int flags,
1205                                     struct ext4_ext_path *path,
1206                                     struct ext4_extent *newext)
1207 {
1208         struct ext4_ext_path *curp;
1209         int depth, i, err = 0;
1210
1211 repeat:
1212         i = depth = ext_depth(inode);
1213
1214         /* walk up to the tree and look for free index entry */
1215         curp = path + depth;
1216         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1217                 i--;
1218                 curp--;
1219         }
1220
1221         /* we use already allocated block for index block,
1222          * so subsequent data blocks should be contiguous */
1223         if (EXT_HAS_FREE_INDEX(curp)) {
1224                 /* if we found index with free entry, then use that
1225                  * entry: create all needed subtree and add new leaf */
1226                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1227                 if (err)
1228                         goto out;
1229
1230                 /* refill path */
1231                 ext4_ext_drop_refs(path);
1232                 path = ext4_ext_find_extent(inode,
1233                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1234                                     path);
1235                 if (IS_ERR(path))
1236                         err = PTR_ERR(path);
1237         } else {
1238                 /* tree is full, time to grow in depth */
1239                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1240                 if (err)
1241                         goto out;
1242
1243                 /* refill path */
1244                 ext4_ext_drop_refs(path);
1245                 path = ext4_ext_find_extent(inode,
1246                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1247                                     path);
1248                 if (IS_ERR(path)) {
1249                         err = PTR_ERR(path);
1250                         goto out;
1251                 }
1252
1253                 /*
1254                  * only first (depth 0 -> 1) produces free space;
1255                  * in all other cases we have to split the grown tree
1256                  */
1257                 depth = ext_depth(inode);
1258                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1259                         /* now we need to split */
1260                         goto repeat;
1261                 }
1262         }
1263
1264 out:
1265         return err;
1266 }
1267
1268 /*
1269  * search the closest allocated block to the left for *logical
1270  * and returns it at @logical + it's physical address at @phys
1271  * if *logical is the smallest allocated block, the function
1272  * returns 0 at @phys
1273  * return value contains 0 (success) or error code
1274  */
1275 static int ext4_ext_search_left(struct inode *inode,
1276                                 struct ext4_ext_path *path,
1277                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1278 {
1279         struct ext4_extent_idx *ix;
1280         struct ext4_extent *ex;
1281         int depth, ee_len;
1282
1283         if (unlikely(path == NULL)) {
1284                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1285                 return -EIO;
1286         }
1287         depth = path->p_depth;
1288         *phys = 0;
1289
1290         if (depth == 0 && path->p_ext == NULL)
1291                 return 0;
1292
1293         /* usually extent in the path covers blocks smaller
1294          * then *logical, but it can be that extent is the
1295          * first one in the file */
1296
1297         ex = path[depth].p_ext;
1298         ee_len = ext4_ext_get_actual_len(ex);
1299         if (*logical < le32_to_cpu(ex->ee_block)) {
1300                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1301                         EXT4_ERROR_INODE(inode,
1302                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1303                                          *logical, le32_to_cpu(ex->ee_block));
1304                         return -EIO;
1305                 }
1306                 while (--depth >= 0) {
1307                         ix = path[depth].p_idx;
1308                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1309                                 EXT4_ERROR_INODE(inode,
1310                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1311                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1312                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1313                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1314                                   depth);
1315                                 return -EIO;
1316                         }
1317                 }
1318                 return 0;
1319         }
1320
1321         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1322                 EXT4_ERROR_INODE(inode,
1323                                  "logical %d < ee_block %d + ee_len %d!",
1324                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1325                 return -EIO;
1326         }
1327
1328         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1329         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1330         return 0;
1331 }
1332
1333 /*
1334  * search the closest allocated block to the right for *logical
1335  * and returns it at @logical + it's physical address at @phys
1336  * if *logical is the largest allocated block, the function
1337  * returns 0 at @phys
1338  * return value contains 0 (success) or error code
1339  */
1340 static int ext4_ext_search_right(struct inode *inode,
1341                                  struct ext4_ext_path *path,
1342                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1343                                  struct ext4_extent **ret_ex)
1344 {
1345         struct buffer_head *bh = NULL;
1346         struct ext4_extent_header *eh;
1347         struct ext4_extent_idx *ix;
1348         struct ext4_extent *ex;
1349         ext4_fsblk_t block;
1350         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1351         int ee_len;
1352
1353         if (unlikely(path == NULL)) {
1354                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1355                 return -EIO;
1356         }
1357         depth = path->p_depth;
1358         *phys = 0;
1359
1360         if (depth == 0 && path->p_ext == NULL)
1361                 return 0;
1362
1363         /* usually extent in the path covers blocks smaller
1364          * then *logical, but it can be that extent is the
1365          * first one in the file */
1366
1367         ex = path[depth].p_ext;
1368         ee_len = ext4_ext_get_actual_len(ex);
1369         if (*logical < le32_to_cpu(ex->ee_block)) {
1370                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1371                         EXT4_ERROR_INODE(inode,
1372                                          "first_extent(path[%d].p_hdr) != ex",
1373                                          depth);
1374                         return -EIO;
1375                 }
1376                 while (--depth >= 0) {
1377                         ix = path[depth].p_idx;
1378                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1379                                 EXT4_ERROR_INODE(inode,
1380                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1381                                                  *logical);
1382                                 return -EIO;
1383                         }
1384                 }
1385                 goto found_extent;
1386         }
1387
1388         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1389                 EXT4_ERROR_INODE(inode,
1390                                  "logical %d < ee_block %d + ee_len %d!",
1391                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1392                 return -EIO;
1393         }
1394
1395         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1396                 /* next allocated block in this leaf */
1397                 ex++;
1398                 goto found_extent;
1399         }
1400
1401         /* go up and search for index to the right */
1402         while (--depth >= 0) {
1403                 ix = path[depth].p_idx;
1404                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1405                         goto got_index;
1406         }
1407
1408         /* we've gone up to the root and found no index to the right */
1409         return 0;
1410
1411 got_index:
1412         /* we've found index to the right, let's
1413          * follow it and find the closest allocated
1414          * block to the right */
1415         ix++;
1416         block = ext4_idx_pblock(ix);
1417         while (++depth < path->p_depth) {
1418                 bh = sb_bread(inode->i_sb, block);
1419                 if (bh == NULL)
1420                         return -EIO;
1421                 eh = ext_block_hdr(bh);
1422                 /* subtract from p_depth to get proper eh_depth */
1423                 if (ext4_ext_check_block(inode, eh,
1424                                          path->p_depth - depth, bh)) {
1425                         put_bh(bh);
1426                         return -EIO;
1427                 }
1428                 ix = EXT_FIRST_INDEX(eh);
1429                 block = ext4_idx_pblock(ix);
1430                 put_bh(bh);
1431         }
1432
1433         bh = sb_bread(inode->i_sb, block);
1434         if (bh == NULL)
1435                 return -EIO;
1436         eh = ext_block_hdr(bh);
1437         if (ext4_ext_check_block(inode, eh, path->p_depth - depth, bh)) {
1438                 put_bh(bh);
1439                 return -EIO;
1440         }
1441         ex = EXT_FIRST_EXTENT(eh);
1442 found_extent:
1443         *logical = le32_to_cpu(ex->ee_block);
1444         *phys = ext4_ext_pblock(ex);
1445         *ret_ex = ex;
1446         if (bh)
1447                 put_bh(bh);
1448         return 0;
1449 }
1450
1451 /*
1452  * ext4_ext_next_allocated_block:
1453  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1454  * NOTE: it considers block number from index entry as
1455  * allocated block. Thus, index entries have to be consistent
1456  * with leaves.
1457  */
1458 static ext4_lblk_t
1459 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1460 {
1461         int depth;
1462
1463         BUG_ON(path == NULL);
1464         depth = path->p_depth;
1465
1466         if (depth == 0 && path->p_ext == NULL)
1467                 return EXT_MAX_BLOCKS;
1468
1469         while (depth >= 0) {
1470                 if (depth == path->p_depth) {
1471                         /* leaf */
1472                         if (path[depth].p_ext &&
1473                                 path[depth].p_ext !=
1474                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1475                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1476                 } else {
1477                         /* index */
1478                         if (path[depth].p_idx !=
1479                                         EXT_LAST_INDEX(path[depth].p_hdr))
1480                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1481                 }
1482                 depth--;
1483         }
1484
1485         return EXT_MAX_BLOCKS;
1486 }
1487
1488 /*
1489  * ext4_ext_next_leaf_block:
1490  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1491  */
1492 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1493 {
1494         int depth;
1495
1496         BUG_ON(path == NULL);
1497         depth = path->p_depth;
1498
1499         /* zero-tree has no leaf blocks at all */
1500         if (depth == 0)
1501                 return EXT_MAX_BLOCKS;
1502
1503         /* go to index block */
1504         depth--;
1505
1506         while (depth >= 0) {
1507                 if (path[depth].p_idx !=
1508                                 EXT_LAST_INDEX(path[depth].p_hdr))
1509                         return (ext4_lblk_t)
1510                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1511                 depth--;
1512         }
1513
1514         return EXT_MAX_BLOCKS;
1515 }
1516
1517 /*
1518  * ext4_ext_correct_indexes:
1519  * if leaf gets modified and modified extent is first in the leaf,
1520  * then we have to correct all indexes above.
1521  * TODO: do we need to correct tree in all cases?
1522  */
1523 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1524                                 struct ext4_ext_path *path)
1525 {
1526         struct ext4_extent_header *eh;
1527         int depth = ext_depth(inode);
1528         struct ext4_extent *ex;
1529         __le32 border;
1530         int k, err = 0;
1531
1532         eh = path[depth].p_hdr;
1533         ex = path[depth].p_ext;
1534
1535         if (unlikely(ex == NULL || eh == NULL)) {
1536                 EXT4_ERROR_INODE(inode,
1537                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1538                 return -EIO;
1539         }
1540
1541         if (depth == 0) {
1542                 /* there is no tree at all */
1543                 return 0;
1544         }
1545
1546         if (ex != EXT_FIRST_EXTENT(eh)) {
1547                 /* we correct tree if first leaf got modified only */
1548                 return 0;
1549         }
1550
1551         /*
1552          * TODO: we need correction if border is smaller than current one
1553          */
1554         k = depth - 1;
1555         border = path[depth].p_ext->ee_block;
1556         err = ext4_ext_get_access(handle, inode, path + k);
1557         if (err)
1558                 return err;
1559         path[k].p_idx->ei_block = border;
1560         err = ext4_ext_dirty(handle, inode, path + k);
1561         if (err)
1562                 return err;
1563
1564         while (k--) {
1565                 /* change all left-side indexes */
1566                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1567                         break;
1568                 err = ext4_ext_get_access(handle, inode, path + k);
1569                 if (err)
1570                         break;
1571                 path[k].p_idx->ei_block = border;
1572                 err = ext4_ext_dirty(handle, inode, path + k);
1573                 if (err)
1574                         break;
1575         }
1576
1577         return err;
1578 }
1579
1580 int
1581 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1582                                 struct ext4_extent *ex2)
1583 {
1584         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1585
1586         /*
1587          * Make sure that both extents are initialized. We don't merge
1588          * uninitialized extents so that we can be sure that end_io code has
1589          * the extent that was written properly split out and conversion to
1590          * initialized is trivial.
1591          */
1592         if (ext4_ext_is_uninitialized(ex1) || ext4_ext_is_uninitialized(ex2))
1593                 return 0;
1594
1595         if (ext4_ext_is_uninitialized(ex1))
1596                 max_len = EXT_UNINIT_MAX_LEN;
1597         else
1598                 max_len = EXT_INIT_MAX_LEN;
1599
1600         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1601         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1602
1603         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1604                         le32_to_cpu(ex2->ee_block))
1605                 return 0;
1606
1607         /*
1608          * To allow future support for preallocated extents to be added
1609          * as an RO_COMPAT feature, refuse to merge to extents if
1610          * this can result in the top bit of ee_len being set.
1611          */
1612         if (ext1_ee_len + ext2_ee_len > max_len)
1613                 return 0;
1614 #ifdef AGGRESSIVE_TEST
1615         if (ext1_ee_len >= 4)
1616                 return 0;
1617 #endif
1618
1619         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1620                 return 1;
1621         return 0;
1622 }
1623
1624 /*
1625  * This function tries to merge the "ex" extent to the next extent in the tree.
1626  * It always tries to merge towards right. If you want to merge towards
1627  * left, pass "ex - 1" as argument instead of "ex".
1628  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1629  * 1 if they got merged.
1630  */
1631 static int ext4_ext_try_to_merge_right(struct inode *inode,
1632                                  struct ext4_ext_path *path,
1633                                  struct ext4_extent *ex)
1634 {
1635         struct ext4_extent_header *eh;
1636         unsigned int depth, len;
1637         int merge_done = 0;
1638         int uninitialized = 0;
1639
1640         depth = ext_depth(inode);
1641         BUG_ON(path[depth].p_hdr == NULL);
1642         eh = path[depth].p_hdr;
1643
1644         while (ex < EXT_LAST_EXTENT(eh)) {
1645                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1646                         break;
1647                 /* merge with next extent! */
1648                 if (ext4_ext_is_uninitialized(ex))
1649                         uninitialized = 1;
1650                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1651                                 + ext4_ext_get_actual_len(ex + 1));
1652                 if (uninitialized)
1653                         ext4_ext_mark_uninitialized(ex);
1654
1655                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1656                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1657                                 * sizeof(struct ext4_extent);
1658                         memmove(ex + 1, ex + 2, len);
1659                 }
1660                 le16_add_cpu(&eh->eh_entries, -1);
1661                 merge_done = 1;
1662                 WARN_ON(eh->eh_entries == 0);
1663                 if (!eh->eh_entries)
1664                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1665         }
1666
1667         return merge_done;
1668 }
1669
1670 /*
1671  * This function does a very simple check to see if we can collapse
1672  * an extent tree with a single extent tree leaf block into the inode.
1673  */
1674 static void ext4_ext_try_to_merge_up(handle_t *handle,
1675                                      struct inode *inode,
1676                                      struct ext4_ext_path *path)
1677 {
1678         size_t s;
1679         unsigned max_root = ext4_ext_space_root(inode, 0);
1680         ext4_fsblk_t blk;
1681
1682         if ((path[0].p_depth != 1) ||
1683             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1684             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1685                 return;
1686
1687         /*
1688          * We need to modify the block allocation bitmap and the block
1689          * group descriptor to release the extent tree block.  If we
1690          * can't get the journal credits, give up.
1691          */
1692         if (ext4_journal_extend(handle, 2))
1693                 return;
1694
1695         /*
1696          * Copy the extent data up to the inode
1697          */
1698         blk = ext4_idx_pblock(path[0].p_idx);
1699         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1700                 sizeof(struct ext4_extent_idx);
1701         s += sizeof(struct ext4_extent_header);
1702
1703         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1704         path[0].p_depth = 0;
1705         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1706                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1707         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1708
1709         brelse(path[1].p_bh);
1710         ext4_free_blocks(handle, inode, NULL, blk, 1,
1711                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1712 }
1713
1714 /*
1715  * This function tries to merge the @ex extent to neighbours in the tree.
1716  * return 1 if merge left else 0.
1717  */
1718 static void ext4_ext_try_to_merge(handle_t *handle,
1719                                   struct inode *inode,
1720                                   struct ext4_ext_path *path,
1721                                   struct ext4_extent *ex) {
1722         struct ext4_extent_header *eh;
1723         unsigned int depth;
1724         int merge_done = 0;
1725
1726         depth = ext_depth(inode);
1727         BUG_ON(path[depth].p_hdr == NULL);
1728         eh = path[depth].p_hdr;
1729
1730         if (ex > EXT_FIRST_EXTENT(eh))
1731                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1732
1733         if (!merge_done)
1734                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1735
1736         ext4_ext_try_to_merge_up(handle, inode, path);
1737 }
1738
1739 /*
1740  * check if a portion of the "newext" extent overlaps with an
1741  * existing extent.
1742  *
1743  * If there is an overlap discovered, it updates the length of the newext
1744  * such that there will be no overlap, and then returns 1.
1745  * If there is no overlap found, it returns 0.
1746  */
1747 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1748                                            struct inode *inode,
1749                                            struct ext4_extent *newext,
1750                                            struct ext4_ext_path *path)
1751 {
1752         ext4_lblk_t b1, b2;
1753         unsigned int depth, len1;
1754         unsigned int ret = 0;
1755
1756         b1 = le32_to_cpu(newext->ee_block);
1757         len1 = ext4_ext_get_actual_len(newext);
1758         depth = ext_depth(inode);
1759         if (!path[depth].p_ext)
1760                 goto out;
1761         b2 = le32_to_cpu(path[depth].p_ext->ee_block);
1762         b2 &= ~(sbi->s_cluster_ratio - 1);
1763
1764         /*
1765          * get the next allocated block if the extent in the path
1766          * is before the requested block(s)
1767          */
1768         if (b2 < b1) {
1769                 b2 = ext4_ext_next_allocated_block(path);
1770                 if (b2 == EXT_MAX_BLOCKS)
1771                         goto out;
1772                 b2 &= ~(sbi->s_cluster_ratio - 1);
1773         }
1774
1775         /* check for wrap through zero on extent logical start block*/
1776         if (b1 + len1 < b1) {
1777                 len1 = EXT_MAX_BLOCKS - b1;
1778                 newext->ee_len = cpu_to_le16(len1);
1779                 ret = 1;
1780         }
1781
1782         /* check for overlap */
1783         if (b1 + len1 > b2) {
1784                 newext->ee_len = cpu_to_le16(b2 - b1);
1785                 ret = 1;
1786         }
1787 out:
1788         return ret;
1789 }
1790
1791 /*
1792  * ext4_ext_insert_extent:
1793  * tries to merge requsted extent into the existing extent or
1794  * inserts requested extent as new one into the tree,
1795  * creating new leaf in the no-space case.
1796  */
1797 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1798                                 struct ext4_ext_path *path,
1799                                 struct ext4_extent *newext, int flag)
1800 {
1801         struct ext4_extent_header *eh;
1802         struct ext4_extent *ex, *fex;
1803         struct ext4_extent *nearex; /* nearest extent */
1804         struct ext4_ext_path *npath = NULL;
1805         int depth, len, err;
1806         ext4_lblk_t next;
1807         unsigned uninitialized = 0;
1808         int flags = 0;
1809
1810         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1811                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1812                 return -EIO;
1813         }
1814         depth = ext_depth(inode);
1815         ex = path[depth].p_ext;
1816         if (unlikely(path[depth].p_hdr == NULL)) {
1817                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1818                 return -EIO;
1819         }
1820
1821         /* try to insert block into found extent and return */
1822         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)
1823                 && ext4_can_extents_be_merged(inode, ex, newext)) {
1824                 ext_debug("append [%d]%d block to %u:[%d]%d (from %llu)\n",
1825                           ext4_ext_is_uninitialized(newext),
1826                           ext4_ext_get_actual_len(newext),
1827                           le32_to_cpu(ex->ee_block),
1828                           ext4_ext_is_uninitialized(ex),
1829                           ext4_ext_get_actual_len(ex),
1830                           ext4_ext_pblock(ex));
1831                 err = ext4_ext_get_access(handle, inode, path + depth);
1832                 if (err)
1833                         return err;
1834
1835                 /*
1836                  * ext4_can_extents_be_merged should have checked that either
1837                  * both extents are uninitialized, or both aren't. Thus we
1838                  * need to check only one of them here.
1839                  */
1840                 if (ext4_ext_is_uninitialized(ex))
1841                         uninitialized = 1;
1842                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1843                                         + ext4_ext_get_actual_len(newext));
1844                 if (uninitialized)
1845                         ext4_ext_mark_uninitialized(ex);
1846                 eh = path[depth].p_hdr;
1847                 nearex = ex;
1848                 goto merge;
1849         }
1850
1851         depth = ext_depth(inode);
1852         eh = path[depth].p_hdr;
1853         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1854                 goto has_space;
1855
1856         /* probably next leaf has space for us? */
1857         fex = EXT_LAST_EXTENT(eh);
1858         next = EXT_MAX_BLOCKS;
1859         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1860                 next = ext4_ext_next_leaf_block(path);
1861         if (next != EXT_MAX_BLOCKS) {
1862                 ext_debug("next leaf block - %u\n", next);
1863                 BUG_ON(npath != NULL);
1864                 npath = ext4_ext_find_extent(inode, next, NULL);
1865                 if (IS_ERR(npath))
1866                         return PTR_ERR(npath);
1867                 BUG_ON(npath->p_depth != path->p_depth);
1868                 eh = npath[depth].p_hdr;
1869                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1870                         ext_debug("next leaf isn't full(%d)\n",
1871                                   le16_to_cpu(eh->eh_entries));
1872                         path = npath;
1873                         goto has_space;
1874                 }
1875                 ext_debug("next leaf has no free space(%d,%d)\n",
1876                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1877         }
1878
1879         /*
1880          * There is no free space in the found leaf.
1881          * We're gonna add a new leaf in the tree.
1882          */
1883         if (flag & EXT4_GET_BLOCKS_PUNCH_OUT_EXT)
1884                 flags = EXT4_MB_USE_ROOT_BLOCKS;
1885         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1886         if (err)
1887                 goto cleanup;
1888         depth = ext_depth(inode);
1889         eh = path[depth].p_hdr;
1890
1891 has_space:
1892         nearex = path[depth].p_ext;
1893
1894         err = ext4_ext_get_access(handle, inode, path + depth);
1895         if (err)
1896                 goto cleanup;
1897
1898         if (!nearex) {
1899                 /* there is no extent in this leaf, create first one */
1900                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1901                                 le32_to_cpu(newext->ee_block),
1902                                 ext4_ext_pblock(newext),
1903                                 ext4_ext_is_uninitialized(newext),
1904                                 ext4_ext_get_actual_len(newext));
1905                 nearex = EXT_FIRST_EXTENT(eh);
1906         } else {
1907                 if (le32_to_cpu(newext->ee_block)
1908                            > le32_to_cpu(nearex->ee_block)) {
1909                         /* Insert after */
1910                         ext_debug("insert %u:%llu:[%d]%d before: "
1911                                         "nearest %p\n",
1912                                         le32_to_cpu(newext->ee_block),
1913                                         ext4_ext_pblock(newext),
1914                                         ext4_ext_is_uninitialized(newext),
1915                                         ext4_ext_get_actual_len(newext),
1916                                         nearex);
1917                         nearex++;
1918                 } else {
1919                         /* Insert before */
1920                         BUG_ON(newext->ee_block == nearex->ee_block);
1921                         ext_debug("insert %u:%llu:[%d]%d after: "
1922                                         "nearest %p\n",
1923                                         le32_to_cpu(newext->ee_block),
1924                                         ext4_ext_pblock(newext),
1925                                         ext4_ext_is_uninitialized(newext),
1926                                         ext4_ext_get_actual_len(newext),
1927                                         nearex);
1928                 }
1929                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1930                 if (len > 0) {
1931                         ext_debug("insert %u:%llu:[%d]%d: "
1932                                         "move %d extents from 0x%p to 0x%p\n",
1933                                         le32_to_cpu(newext->ee_block),
1934                                         ext4_ext_pblock(newext),
1935                                         ext4_ext_is_uninitialized(newext),
1936                                         ext4_ext_get_actual_len(newext),
1937                                         len, nearex, nearex + 1);
1938                         memmove(nearex + 1, nearex,
1939                                 len * sizeof(struct ext4_extent));
1940                 }
1941         }
1942
1943         le16_add_cpu(&eh->eh_entries, 1);
1944         path[depth].p_ext = nearex;
1945         nearex->ee_block = newext->ee_block;
1946         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
1947         nearex->ee_len = newext->ee_len;
1948
1949 merge:
1950         /* try to merge extents */
1951         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
1952                 ext4_ext_try_to_merge(handle, inode, path, nearex);
1953
1954
1955         /* time to correct all indexes above */
1956         err = ext4_ext_correct_indexes(handle, inode, path);
1957         if (err)
1958                 goto cleanup;
1959
1960         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
1961
1962 cleanup:
1963         if (npath) {
1964                 ext4_ext_drop_refs(npath);
1965                 kfree(npath);
1966         }
1967         return err;
1968 }
1969
1970 static int ext4_fill_fiemap_extents(struct inode *inode,
1971                                     ext4_lblk_t block, ext4_lblk_t num,
1972                                     struct fiemap_extent_info *fieinfo)
1973 {
1974         struct ext4_ext_path *path = NULL;
1975         struct ext4_extent *ex;
1976         struct extent_status es;
1977         ext4_lblk_t next, next_del, start = 0, end = 0;
1978         ext4_lblk_t last = block + num;
1979         int exists, depth = 0, err = 0;
1980         unsigned int flags = 0;
1981         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
1982
1983         while (block < last && block != EXT_MAX_BLOCKS) {
1984                 num = last - block;
1985                 /* find extent for this block */
1986                 down_read(&EXT4_I(inode)->i_data_sem);
1987
1988                 if (path && ext_depth(inode) != depth) {
1989                         /* depth was changed. we have to realloc path */
1990                         kfree(path);
1991                         path = NULL;
1992                 }
1993
1994                 path = ext4_ext_find_extent(inode, block, path);
1995                 if (IS_ERR(path)) {
1996                         up_read(&EXT4_I(inode)->i_data_sem);
1997                         err = PTR_ERR(path);
1998                         path = NULL;
1999                         break;
2000                 }
2001
2002                 depth = ext_depth(inode);
2003                 if (unlikely(path[depth].p_hdr == NULL)) {
2004                         up_read(&EXT4_I(inode)->i_data_sem);
2005                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2006                         err = -EIO;
2007                         break;
2008                 }
2009                 ex = path[depth].p_ext;
2010                 next = ext4_ext_next_allocated_block(path);
2011                 ext4_ext_drop_refs(path);
2012
2013                 flags = 0;
2014                 exists = 0;
2015                 if (!ex) {
2016                         /* there is no extent yet, so try to allocate
2017                          * all requested space */
2018                         start = block;
2019                         end = block + num;
2020                 } else if (le32_to_cpu(ex->ee_block) > block) {
2021                         /* need to allocate space before found extent */
2022                         start = block;
2023                         end = le32_to_cpu(ex->ee_block);
2024                         if (block + num < end)
2025                                 end = block + num;
2026                 } else if (block >= le32_to_cpu(ex->ee_block)
2027                                         + ext4_ext_get_actual_len(ex)) {
2028                         /* need to allocate space after found extent */
2029                         start = block;
2030                         end = block + num;
2031                         if (end >= next)
2032                                 end = next;
2033                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2034                         /*
2035                          * some part of requested space is covered
2036                          * by found extent
2037                          */
2038                         start = block;
2039                         end = le32_to_cpu(ex->ee_block)
2040                                 + ext4_ext_get_actual_len(ex);
2041                         if (block + num < end)
2042                                 end = block + num;
2043                         exists = 1;
2044                 } else {
2045                         BUG();
2046                 }
2047                 BUG_ON(end <= start);
2048
2049                 if (!exists) {
2050                         es.es_lblk = start;
2051                         es.es_len = end - start;
2052                         es.es_pblk = 0;
2053                 } else {
2054                         es.es_lblk = le32_to_cpu(ex->ee_block);
2055                         es.es_len = ext4_ext_get_actual_len(ex);
2056                         es.es_pblk = ext4_ext_pblock(ex);
2057                         if (ext4_ext_is_uninitialized(ex))
2058                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2059                 }
2060
2061                 /*
2062                  * Find delayed extent and update es accordingly. We call
2063                  * it even in !exists case to find out whether es is the
2064                  * last existing extent or not.
2065                  */
2066                 next_del = ext4_find_delayed_extent(inode, &es);
2067                 if (!exists && next_del) {
2068                         exists = 1;
2069                         flags |= FIEMAP_EXTENT_DELALLOC;
2070                 }
2071                 up_read(&EXT4_I(inode)->i_data_sem);
2072
2073                 if (unlikely(es.es_len == 0)) {
2074                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2075                         err = -EIO;
2076                         break;
2077                 }
2078
2079                 /*
2080                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2081                  * we need to check next == EXT_MAX_BLOCKS because it is
2082                  * possible that an extent is with unwritten and delayed
2083                  * status due to when an extent is delayed allocated and
2084                  * is allocated by fallocate status tree will track both of
2085                  * them in a extent.
2086                  *
2087                  * So we could return a unwritten and delayed extent, and
2088                  * its block is equal to 'next'.
2089                  */
2090                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2091                         flags |= FIEMAP_EXTENT_LAST;
2092                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2093                                      next != EXT_MAX_BLOCKS)) {
2094                                 EXT4_ERROR_INODE(inode,
2095                                                  "next extent == %u, next "
2096                                                  "delalloc extent = %u",
2097                                                  next, next_del);
2098                                 err = -EIO;
2099                                 break;
2100                         }
2101                 }
2102
2103                 if (exists) {
2104                         err = fiemap_fill_next_extent(fieinfo,
2105                                 (__u64)es.es_lblk << blksize_bits,
2106                                 (__u64)es.es_pblk << blksize_bits,
2107                                 (__u64)es.es_len << blksize_bits,
2108                                 flags);
2109                         if (err < 0)
2110                                 break;
2111                         if (err == 1) {
2112                                 err = 0;
2113                                 break;
2114                         }
2115                 }
2116
2117                 block = es.es_lblk + es.es_len;
2118         }
2119
2120         if (path) {
2121                 ext4_ext_drop_refs(path);
2122                 kfree(path);
2123         }
2124
2125         return err;
2126 }
2127
2128 /*
2129  * ext4_ext_put_gap_in_cache:
2130  * calculate boundaries of the gap that the requested block fits into
2131  * and cache this gap
2132  */
2133 static void
2134 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2135                                 ext4_lblk_t block)
2136 {
2137         int depth = ext_depth(inode);
2138         unsigned long len;
2139         ext4_lblk_t lblock;
2140         struct ext4_extent *ex;
2141
2142         ex = path[depth].p_ext;
2143         if (ex == NULL) {
2144                 /*
2145                  * there is no extent yet, so gap is [0;-] and we
2146                  * don't cache it
2147                  */
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                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2157                         ext4_es_insert_extent(inode, lblock, len, ~0,
2158                                               EXTENT_STATUS_HOLE);
2159         } else if (block >= le32_to_cpu(ex->ee_block)
2160                         + ext4_ext_get_actual_len(ex)) {
2161                 ext4_lblk_t next;
2162                 lblock = le32_to_cpu(ex->ee_block)
2163                         + ext4_ext_get_actual_len(ex);
2164
2165                 next = ext4_ext_next_allocated_block(path);
2166                 ext_debug("cache gap(after): [%u:%u] %u",
2167                                 le32_to_cpu(ex->ee_block),
2168                                 ext4_ext_get_actual_len(ex),
2169                                 block);
2170                 BUG_ON(next == lblock);
2171                 len = next - lblock;
2172                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2173                         ext4_es_insert_extent(inode, lblock, len, ~0,
2174                                               EXTENT_STATUS_HOLE);
2175         } else {
2176                 lblock = len = 0;
2177                 BUG();
2178         }
2179
2180         ext_debug(" -> %u:%lu\n", lblock, len);
2181 }
2182
2183 /*
2184  * ext4_ext_rm_idx:
2185  * removes index from the index block.
2186  */
2187 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2188                         struct ext4_ext_path *path, int depth)
2189 {
2190         int err;
2191         ext4_fsblk_t leaf;
2192
2193         /* free index block */
2194         depth--;
2195         path = path + depth;
2196         leaf = ext4_idx_pblock(path->p_idx);
2197         if (unlikely(path->p_hdr->eh_entries == 0)) {
2198                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2199                 return -EIO;
2200         }
2201         err = ext4_ext_get_access(handle, inode, path);
2202         if (err)
2203                 return err;
2204
2205         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2206                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2207                 len *= sizeof(struct ext4_extent_idx);
2208                 memmove(path->p_idx, path->p_idx + 1, len);
2209         }
2210
2211         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2212         err = ext4_ext_dirty(handle, inode, path);
2213         if (err)
2214                 return err;
2215         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2216         trace_ext4_ext_rm_idx(inode, leaf);
2217
2218         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2219                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2220
2221         while (--depth >= 0) {
2222                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2223                         break;
2224                 path--;
2225                 err = ext4_ext_get_access(handle, inode, path);
2226                 if (err)
2227                         break;
2228                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2229                 err = ext4_ext_dirty(handle, inode, path);
2230                 if (err)
2231                         break;
2232         }
2233         return err;
2234 }
2235
2236 /*
2237  * ext4_ext_calc_credits_for_single_extent:
2238  * This routine returns max. credits that needed to insert an extent
2239  * to the extent tree.
2240  * When pass the actual path, the caller should calculate credits
2241  * under i_data_sem.
2242  */
2243 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2244                                                 struct ext4_ext_path *path)
2245 {
2246         if (path) {
2247                 int depth = ext_depth(inode);
2248                 int ret = 0;
2249
2250                 /* probably there is space in leaf? */
2251                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2252                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2253
2254                         /*
2255                          *  There are some space in the leaf tree, no
2256                          *  need to account for leaf block credit
2257                          *
2258                          *  bitmaps and block group descriptor blocks
2259                          *  and other metadata blocks still need to be
2260                          *  accounted.
2261                          */
2262                         /* 1 bitmap, 1 block group descriptor */
2263                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2264                         return ret;
2265                 }
2266         }
2267
2268         return ext4_chunk_trans_blocks(inode, nrblocks);
2269 }
2270
2271 /*
2272  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2273  *
2274  * if nrblocks are fit in a single extent (chunk flag is 1), then
2275  * in the worse case, each tree level index/leaf need to be changed
2276  * if the tree split due to insert a new extent, then the old tree
2277  * index/leaf need to be updated too
2278  *
2279  * If the nrblocks are discontiguous, they could cause
2280  * the whole tree split more than once, but this is really rare.
2281  */
2282 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2283 {
2284         int index;
2285         int depth;
2286
2287         /* If we are converting the inline data, only one is needed here. */
2288         if (ext4_has_inline_data(inode))
2289                 return 1;
2290
2291         depth = ext_depth(inode);
2292
2293         if (chunk)
2294                 index = depth * 2;
2295         else
2296                 index = depth * 3;
2297
2298         return index;
2299 }
2300
2301 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2302                               struct ext4_extent *ex,
2303                               ext4_fsblk_t *partial_cluster,
2304                               ext4_lblk_t from, ext4_lblk_t to)
2305 {
2306         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2307         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2308         ext4_fsblk_t pblk;
2309         int flags = 0;
2310
2311         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2312                 flags |= EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2313         else if (ext4_should_journal_data(inode))
2314                 flags |= EXT4_FREE_BLOCKS_FORGET;
2315
2316         /*
2317          * For bigalloc file systems, we never free a partial cluster
2318          * at the beginning of the extent.  Instead, we make a note
2319          * that we tried freeing the cluster, and check to see if we
2320          * need to free it on a subsequent call to ext4_remove_blocks,
2321          * or at the end of the ext4_truncate() operation.
2322          */
2323         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2324
2325         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2326         /*
2327          * If we have a partial cluster, and it's different from the
2328          * cluster of the last block, we need to explicitly free the
2329          * partial cluster here.
2330          */
2331         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2332         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2333                 ext4_free_blocks(handle, inode, NULL,
2334                                  EXT4_C2B(sbi, *partial_cluster),
2335                                  sbi->s_cluster_ratio, flags);
2336                 *partial_cluster = 0;
2337         }
2338
2339 #ifdef EXTENTS_STATS
2340         {
2341                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2342                 spin_lock(&sbi->s_ext_stats_lock);
2343                 sbi->s_ext_blocks += ee_len;
2344                 sbi->s_ext_extents++;
2345                 if (ee_len < sbi->s_ext_min)
2346                         sbi->s_ext_min = ee_len;
2347                 if (ee_len > sbi->s_ext_max)
2348                         sbi->s_ext_max = ee_len;
2349                 if (ext_depth(inode) > sbi->s_depth_max)
2350                         sbi->s_depth_max = ext_depth(inode);
2351                 spin_unlock(&sbi->s_ext_stats_lock);
2352         }
2353 #endif
2354         if (from >= le32_to_cpu(ex->ee_block)
2355             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2356                 /* tail removal */
2357                 ext4_lblk_t num;
2358
2359                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2360                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2361                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2362                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2363                 /*
2364                  * If the block range to be freed didn't start at the
2365                  * beginning of a cluster, and we removed the entire
2366                  * extent, save the partial cluster here, since we
2367                  * might need to delete if we determine that the
2368                  * truncate operation has removed all of the blocks in
2369                  * the cluster.
2370                  */
2371                 if (pblk & (sbi->s_cluster_ratio - 1) &&
2372                     (ee_len == num))
2373                         *partial_cluster = EXT4_B2C(sbi, pblk);
2374                 else
2375                         *partial_cluster = 0;
2376         } else if (from == le32_to_cpu(ex->ee_block)
2377                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2378                 /* head removal */
2379                 ext4_lblk_t num;
2380                 ext4_fsblk_t start;
2381
2382                 num = to - from;
2383                 start = ext4_ext_pblock(ex);
2384
2385                 ext_debug("free first %u blocks starting %llu\n", num, start);
2386                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2387
2388         } else {
2389                 printk(KERN_INFO "strange request: removal(2) "
2390                                 "%u-%u from %u:%u\n",
2391                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2392         }
2393         return 0;
2394 }
2395
2396
2397 /*
2398  * ext4_ext_rm_leaf() Removes the extents associated with the
2399  * blocks appearing between "start" and "end", and splits the extents
2400  * if "start" and "end" appear in the same extent
2401  *
2402  * @handle: The journal handle
2403  * @inode:  The files inode
2404  * @path:   The path to the leaf
2405  * @start:  The first block to remove
2406  * @end:   The last block to remove
2407  */
2408 static int
2409 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2410                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2411                  ext4_lblk_t start, ext4_lblk_t end)
2412 {
2413         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2414         int err = 0, correct_index = 0;
2415         int depth = ext_depth(inode), credits;
2416         struct ext4_extent_header *eh;
2417         ext4_lblk_t a, b;
2418         unsigned num;
2419         ext4_lblk_t ex_ee_block;
2420         unsigned short ex_ee_len;
2421         unsigned uninitialized = 0;
2422         struct ext4_extent *ex;
2423
2424         /* the header must be checked already in ext4_ext_remove_space() */
2425         ext_debug("truncate since %u in leaf to %u\n", start, end);
2426         if (!path[depth].p_hdr)
2427                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2428         eh = path[depth].p_hdr;
2429         if (unlikely(path[depth].p_hdr == NULL)) {
2430                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2431                 return -EIO;
2432         }
2433         /* find where to start removing */
2434         ex = EXT_LAST_EXTENT(eh);
2435
2436         ex_ee_block = le32_to_cpu(ex->ee_block);
2437         ex_ee_len = ext4_ext_get_actual_len(ex);
2438
2439         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2440
2441         while (ex >= EXT_FIRST_EXTENT(eh) &&
2442                         ex_ee_block + ex_ee_len > start) {
2443
2444                 if (ext4_ext_is_uninitialized(ex))
2445                         uninitialized = 1;
2446                 else
2447                         uninitialized = 0;
2448
2449                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2450                          uninitialized, ex_ee_len);
2451                 path[depth].p_ext = ex;
2452
2453                 a = ex_ee_block > start ? ex_ee_block : start;
2454                 b = ex_ee_block+ex_ee_len - 1 < end ?
2455                         ex_ee_block+ex_ee_len - 1 : end;
2456
2457                 ext_debug("  border %u:%u\n", a, b);
2458
2459                 /* If this extent is beyond the end of the hole, skip it */
2460                 if (end < ex_ee_block) {
2461                         ex--;
2462                         ex_ee_block = le32_to_cpu(ex->ee_block);
2463                         ex_ee_len = ext4_ext_get_actual_len(ex);
2464                         continue;
2465                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2466                         EXT4_ERROR_INODE(inode,
2467                                          "can not handle truncate %u:%u "
2468                                          "on extent %u:%u",
2469                                          start, end, ex_ee_block,
2470                                          ex_ee_block + ex_ee_len - 1);
2471                         err = -EIO;
2472                         goto out;
2473                 } else if (a != ex_ee_block) {
2474                         /* remove tail of the extent */
2475                         num = a - ex_ee_block;
2476                 } else {
2477                         /* remove whole extent: excellent! */
2478                         num = 0;
2479                 }
2480                 /*
2481                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2482                  * descriptor) for each block group; assume two block
2483                  * groups plus ex_ee_len/blocks_per_block_group for
2484                  * the worst case
2485                  */
2486                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2487                 if (ex == EXT_FIRST_EXTENT(eh)) {
2488                         correct_index = 1;
2489                         credits += (ext_depth(inode)) + 1;
2490                 }
2491                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2492
2493                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2494                 if (err)
2495                         goto out;
2496
2497                 err = ext4_ext_get_access(handle, inode, path + depth);
2498                 if (err)
2499                         goto out;
2500
2501                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2502                                          a, b);
2503                 if (err)
2504                         goto out;
2505
2506                 if (num == 0)
2507                         /* this extent is removed; mark slot entirely unused */
2508                         ext4_ext_store_pblock(ex, 0);
2509
2510                 ex->ee_len = cpu_to_le16(num);
2511                 /*
2512                  * Do not mark uninitialized if all the blocks in the
2513                  * extent have been removed.
2514                  */
2515                 if (uninitialized && num)
2516                         ext4_ext_mark_uninitialized(ex);
2517                 /*
2518                  * If the extent was completely released,
2519                  * we need to remove it from the leaf
2520                  */
2521                 if (num == 0) {
2522                         if (end != EXT_MAX_BLOCKS - 1) {
2523                                 /*
2524                                  * For hole punching, we need to scoot all the
2525                                  * extents up when an extent is removed so that
2526                                  * we dont have blank extents in the middle
2527                                  */
2528                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2529                                         sizeof(struct ext4_extent));
2530
2531                                 /* Now get rid of the one at the end */
2532                                 memset(EXT_LAST_EXTENT(eh), 0,
2533                                         sizeof(struct ext4_extent));
2534                         }
2535                         le16_add_cpu(&eh->eh_entries, -1);
2536                 } else
2537                         *partial_cluster = 0;
2538
2539                 err = ext4_ext_dirty(handle, inode, path + depth);
2540                 if (err)
2541                         goto out;
2542
2543                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2544                                 ext4_ext_pblock(ex));
2545                 ex--;
2546                 ex_ee_block = le32_to_cpu(ex->ee_block);
2547                 ex_ee_len = ext4_ext_get_actual_len(ex);
2548         }
2549
2550         if (correct_index && eh->eh_entries)
2551                 err = ext4_ext_correct_indexes(handle, inode, path);
2552
2553         /*
2554          * If there is still a entry in the leaf node, check to see if
2555          * it references the partial cluster.  This is the only place
2556          * where it could; if it doesn't, we can free the cluster.
2557          */
2558         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2559             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2560              *partial_cluster)) {
2561                 int flags = EXT4_FREE_BLOCKS_FORGET;
2562
2563                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2564                         flags |= EXT4_FREE_BLOCKS_METADATA;
2565
2566                 ext4_free_blocks(handle, inode, NULL,
2567                                  EXT4_C2B(sbi, *partial_cluster),
2568                                  sbi->s_cluster_ratio, flags);
2569                 *partial_cluster = 0;
2570         }
2571
2572         /* if this leaf is free, then we should
2573          * remove it from index block above */
2574         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2575                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2576
2577 out:
2578         return err;
2579 }
2580
2581 /*
2582  * ext4_ext_more_to_rm:
2583  * returns 1 if current index has to be freed (even partial)
2584  */
2585 static int
2586 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2587 {
2588         BUG_ON(path->p_idx == NULL);
2589
2590         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2591                 return 0;
2592
2593         /*
2594          * if truncate on deeper level happened, it wasn't partial,
2595          * so we have to consider current index for truncation
2596          */
2597         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2598                 return 0;
2599         return 1;
2600 }
2601
2602 static int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2603                                  ext4_lblk_t end)
2604 {
2605         struct super_block *sb = inode->i_sb;
2606         int depth = ext_depth(inode);
2607         struct ext4_ext_path *path = NULL;
2608         ext4_fsblk_t partial_cluster = 0;
2609         handle_t *handle;
2610         int i = 0, err = 0;
2611
2612         ext_debug("truncate since %u to %u\n", start, end);
2613
2614         /* probably first extent we're gonna free will be last in block */
2615         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2616         if (IS_ERR(handle))
2617                 return PTR_ERR(handle);
2618
2619 again:
2620         trace_ext4_ext_remove_space(inode, start, depth);
2621
2622         /*
2623          * Check if we are removing extents inside the extent tree. If that
2624          * is the case, we are going to punch a hole inside the extent tree
2625          * so we have to check whether we need to split the extent covering
2626          * the last block to remove so we can easily remove the part of it
2627          * in ext4_ext_rm_leaf().
2628          */
2629         if (end < EXT_MAX_BLOCKS - 1) {
2630                 struct ext4_extent *ex;
2631                 ext4_lblk_t ee_block;
2632
2633                 /* find extent for this block */
2634                 path = ext4_ext_find_extent(inode, end, NULL);
2635                 if (IS_ERR(path)) {
2636                         ext4_journal_stop(handle);
2637                         return PTR_ERR(path);
2638                 }
2639                 depth = ext_depth(inode);
2640                 /* Leaf not may not exist only if inode has no blocks at all */
2641                 ex = path[depth].p_ext;
2642                 if (!ex) {
2643                         if (depth) {
2644                                 EXT4_ERROR_INODE(inode,
2645                                                  "path[%d].p_hdr == NULL",
2646                                                  depth);
2647                                 err = -EIO;
2648                         }
2649                         goto out;
2650                 }
2651
2652                 ee_block = le32_to_cpu(ex->ee_block);
2653
2654                 /*
2655                  * See if the last block is inside the extent, if so split
2656                  * the extent at 'end' block so we can easily remove the
2657                  * tail of the first part of the split extent in
2658                  * ext4_ext_rm_leaf().
2659                  */
2660                 if (end >= ee_block &&
2661                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2662                         int split_flag = 0;
2663
2664                         if (ext4_ext_is_uninitialized(ex))
2665                                 split_flag = EXT4_EXT_MARK_UNINIT1 |
2666                                              EXT4_EXT_MARK_UNINIT2;
2667
2668                         /*
2669                          * Split the extent in two so that 'end' is the last
2670                          * block in the first new extent
2671                          */
2672                         err = ext4_split_extent_at(handle, inode, path,
2673                                                 end + 1, split_flag,
2674                                                 EXT4_GET_BLOCKS_PRE_IO |
2675                                                 EXT4_GET_BLOCKS_PUNCH_OUT_EXT);
2676
2677                         if (err < 0)
2678                                 goto out;
2679                 }
2680         }
2681         /*
2682          * We start scanning from right side, freeing all the blocks
2683          * after i_size and walking into the tree depth-wise.
2684          */
2685         depth = ext_depth(inode);
2686         if (path) {
2687                 int k = i = depth;
2688                 while (--k > 0)
2689                         path[k].p_block =
2690                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2691         } else {
2692                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2693                                GFP_NOFS);
2694                 if (path == NULL) {
2695                         ext4_journal_stop(handle);
2696                         return -ENOMEM;
2697                 }
2698                 path[0].p_depth = depth;
2699                 path[0].p_hdr = ext_inode_hdr(inode);
2700                 i = 0;
2701
2702                 if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2703                         err = -EIO;
2704                         goto out;
2705                 }
2706         }
2707         err = 0;
2708
2709         while (i >= 0 && err == 0) {
2710                 if (i == depth) {
2711                         /* this is leaf block */
2712                         err = ext4_ext_rm_leaf(handle, inode, path,
2713                                                &partial_cluster, start,
2714                                                end);
2715                         /* root level has p_bh == NULL, brelse() eats this */
2716                         brelse(path[i].p_bh);
2717                         path[i].p_bh = NULL;
2718                         i--;
2719                         continue;
2720                 }
2721
2722                 /* this is index block */
2723                 if (!path[i].p_hdr) {
2724                         ext_debug("initialize header\n");
2725                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2726                 }
2727
2728                 if (!path[i].p_idx) {
2729                         /* this level hasn't been touched yet */
2730                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2731                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2732                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2733                                   path[i].p_hdr,
2734                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2735                 } else {
2736                         /* we were already here, see at next index */
2737                         path[i].p_idx--;
2738                 }
2739
2740                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2741                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2742                                 path[i].p_idx);
2743                 if (ext4_ext_more_to_rm(path + i)) {
2744                         struct buffer_head *bh;
2745                         /* go to the next level */
2746                         ext_debug("move to level %d (block %llu)\n",
2747                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2748                         memset(path + i + 1, 0, sizeof(*path));
2749                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2750                         if (!bh) {
2751                                 /* should we reset i_size? */
2752                                 err = -EIO;
2753                                 break;
2754                         }
2755                         if (WARN_ON(i + 1 > depth)) {
2756                                 err = -EIO;
2757                                 break;
2758                         }
2759                         if (ext4_ext_check_block(inode, ext_block_hdr(bh),
2760                                                         depth - i - 1, bh)) {
2761                                 err = -EIO;
2762                                 break;
2763                         }
2764                         path[i + 1].p_bh = bh;
2765
2766                         /* save actual number of indexes since this
2767                          * number is changed at the next iteration */
2768                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2769                         i++;
2770                 } else {
2771                         /* we finished processing this index, go up */
2772                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2773                                 /* index is empty, remove it;
2774                                  * handle must be already prepared by the
2775                                  * truncatei_leaf() */
2776                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2777                         }
2778                         /* root level has p_bh == NULL, brelse() eats this */
2779                         brelse(path[i].p_bh);
2780                         path[i].p_bh = NULL;
2781                         i--;
2782                         ext_debug("return to level %d\n", i);
2783                 }
2784         }
2785
2786         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2787                         path->p_hdr->eh_entries);
2788
2789         /* If we still have something in the partial cluster and we have removed
2790          * even the first extent, then we should free the blocks in the partial
2791          * cluster as well. */
2792         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2793                 int flags = EXT4_FREE_BLOCKS_FORGET;
2794
2795                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2796                         flags |= EXT4_FREE_BLOCKS_METADATA;
2797
2798                 ext4_free_blocks(handle, inode, NULL,
2799                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2800                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2801                 partial_cluster = 0;
2802         }
2803
2804         /* TODO: flexible tree reduction should be here */
2805         if (path->p_hdr->eh_entries == 0) {
2806                 /*
2807                  * truncate to zero freed all the tree,
2808                  * so we need to correct eh_depth
2809                  */
2810                 err = ext4_ext_get_access(handle, inode, path);
2811                 if (err == 0) {
2812                         ext_inode_hdr(inode)->eh_depth = 0;
2813                         ext_inode_hdr(inode)->eh_max =
2814                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2815                         err = ext4_ext_dirty(handle, inode, path);
2816                 }
2817         }
2818 out:
2819         ext4_ext_drop_refs(path);
2820         kfree(path);
2821         if (err == -EAGAIN) {
2822                 path = NULL;
2823                 goto again;
2824         }
2825         ext4_journal_stop(handle);
2826
2827         return err;
2828 }
2829
2830 /*
2831  * called at mount time
2832  */
2833 void ext4_ext_init(struct super_block *sb)
2834 {
2835         /*
2836          * possible initialization would be here
2837          */
2838
2839         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2840 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2841                 printk(KERN_INFO "EXT4-fs: file extents enabled"
2842 #ifdef AGGRESSIVE_TEST
2843                        ", aggressive tests"
2844 #endif
2845 #ifdef CHECK_BINSEARCH
2846                        ", check binsearch"
2847 #endif
2848 #ifdef EXTENTS_STATS
2849                        ", stats"
2850 #endif
2851                        "\n");
2852 #endif
2853 #ifdef EXTENTS_STATS
2854                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2855                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2856                 EXT4_SB(sb)->s_ext_max = 0;
2857 #endif
2858         }
2859 }
2860
2861 /*
2862  * called at umount time
2863  */
2864 void ext4_ext_release(struct super_block *sb)
2865 {
2866         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2867                 return;
2868
2869 #ifdef EXTENTS_STATS
2870         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2871                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2872                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2873                         sbi->s_ext_blocks, sbi->s_ext_extents,
2874                         sbi->s_ext_blocks / sbi->s_ext_extents);
2875                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2876                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2877         }
2878 #endif
2879 }
2880
2881 /* FIXME!! we need to try to merge to left or right after zero-out  */
2882 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2883 {
2884         ext4_fsblk_t ee_pblock;
2885         unsigned int ee_len;
2886         int ret;
2887
2888         ee_len    = ext4_ext_get_actual_len(ex);
2889         ee_pblock = ext4_ext_pblock(ex);
2890
2891         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2892         if (ret > 0)
2893                 ret = 0;
2894
2895         return ret;
2896 }
2897
2898 /*
2899  * ext4_split_extent_at() splits an extent at given block.
2900  *
2901  * @handle: the journal handle
2902  * @inode: the file inode
2903  * @path: the path to the extent
2904  * @split: the logical block where the extent is splitted.
2905  * @split_flags: indicates if the extent could be zeroout if split fails, and
2906  *               the states(init or uninit) of new extents.
2907  * @flags: flags used to insert new extent to extent tree.
2908  *
2909  *
2910  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2911  * of which are deterimined by split_flag.
2912  *
2913  * There are two cases:
2914  *  a> the extent are splitted into two extent.
2915  *  b> split is not needed, and just mark the extent.
2916  *
2917  * return 0 on success.
2918  */
2919 static int ext4_split_extent_at(handle_t *handle,
2920                              struct inode *inode,
2921                              struct ext4_ext_path *path,
2922                              ext4_lblk_t split,
2923                              int split_flag,
2924                              int flags)
2925 {
2926         ext4_fsblk_t newblock;
2927         ext4_lblk_t ee_block;
2928         struct ext4_extent *ex, newex, orig_ex, zero_ex;
2929         struct ext4_extent *ex2 = NULL;
2930         unsigned int ee_len, depth;
2931         int err = 0;
2932
2933         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
2934                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
2935
2936         ext_debug("ext4_split_extents_at: inode %lu, logical"
2937                 "block %llu\n", inode->i_ino, (unsigned long long)split);
2938
2939         ext4_ext_show_leaf(inode, path);
2940
2941         depth = ext_depth(inode);
2942         ex = path[depth].p_ext;
2943         ee_block = le32_to_cpu(ex->ee_block);
2944         ee_len = ext4_ext_get_actual_len(ex);
2945         newblock = split - ee_block + ext4_ext_pblock(ex);
2946
2947         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
2948         BUG_ON(!ext4_ext_is_uninitialized(ex) &&
2949                split_flag & (EXT4_EXT_MAY_ZEROOUT |
2950                              EXT4_EXT_MARK_UNINIT1 |
2951                              EXT4_EXT_MARK_UNINIT2));
2952
2953         err = ext4_ext_get_access(handle, inode, path + depth);
2954         if (err)
2955                 goto out;
2956
2957         if (split == ee_block) {
2958                 /*
2959                  * case b: block @split is the block that the extent begins with
2960                  * then we just change the state of the extent, and splitting
2961                  * is not needed.
2962                  */
2963                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
2964                         ext4_ext_mark_uninitialized(ex);
2965                 else
2966                         ext4_ext_mark_initialized(ex);
2967
2968                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
2969                         ext4_ext_try_to_merge(handle, inode, path, ex);
2970
2971                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2972                 goto out;
2973         }
2974
2975         /* case a */
2976         memcpy(&orig_ex, ex, sizeof(orig_ex));
2977         ex->ee_len = cpu_to_le16(split - ee_block);
2978         if (split_flag & EXT4_EXT_MARK_UNINIT1)
2979                 ext4_ext_mark_uninitialized(ex);
2980
2981         /*
2982          * path may lead to new leaf, not to original leaf any more
2983          * after ext4_ext_insert_extent() returns,
2984          */
2985         err = ext4_ext_dirty(handle, inode, path + depth);
2986         if (err)
2987                 goto fix_extent_len;
2988
2989         ex2 = &newex;
2990         ex2->ee_block = cpu_to_le32(split);
2991         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
2992         ext4_ext_store_pblock(ex2, newblock);
2993         if (split_flag & EXT4_EXT_MARK_UNINIT2)
2994                 ext4_ext_mark_uninitialized(ex2);
2995
2996         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
2997         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
2998                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
2999                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3000                                 err = ext4_ext_zeroout(inode, ex2);
3001                                 zero_ex.ee_block = ex2->ee_block;
3002                                 zero_ex.ee_len = ext4_ext_get_actual_len(ex2);
3003                                 ext4_ext_store_pblock(&zero_ex,
3004                                                       ext4_ext_pblock(ex2));
3005                         } else {
3006                                 err = ext4_ext_zeroout(inode, ex);
3007                                 zero_ex.ee_block = ex->ee_block;
3008                                 zero_ex.ee_len = ext4_ext_get_actual_len(ex);
3009                                 ext4_ext_store_pblock(&zero_ex,
3010                                                       ext4_ext_pblock(ex));
3011                         }
3012                 } else {
3013                         err = ext4_ext_zeroout(inode, &orig_ex);
3014                         zero_ex.ee_block = orig_ex.ee_block;
3015                         zero_ex.ee_len = ext4_ext_get_actual_len(&orig_ex);
3016                         ext4_ext_store_pblock(&zero_ex,
3017                                               ext4_ext_pblock(&orig_ex));
3018                 }
3019
3020                 if (err)
3021                         goto fix_extent_len;
3022                 /* update the extent length and mark as initialized */
3023                 ex->ee_len = cpu_to_le16(ee_len);
3024                 ext4_ext_try_to_merge(handle, inode, path, ex);
3025                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3026                 if (err)
3027                         goto fix_extent_len;
3028
3029                 /* update extent status tree */
3030                 err = ext4_es_zeroout(inode, &zero_ex);
3031
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         int allocated = map->m_len;
3071
3072         depth = ext_depth(inode);
3073         ex = path[depth].p_ext;
3074         ee_block = le32_to_cpu(ex->ee_block);
3075         ee_len = ext4_ext_get_actual_len(ex);
3076         uninitialized = ext4_ext_is_uninitialized(ex);
3077
3078         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3079                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3080                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3081                 if (uninitialized)
3082                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
3083                                        EXT4_EXT_MARK_UNINIT2;
3084                 if (split_flag & EXT4_EXT_DATA_VALID2)
3085                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3086                 err = ext4_split_extent_at(handle, inode, path,
3087                                 map->m_lblk + map->m_len, split_flag1, flags1);
3088                 if (err)
3089                         goto out;
3090         } else {
3091                 allocated = ee_len - (map->m_lblk - ee_block);
3092         }
3093         /*
3094          * Update path is required because previous ext4_split_extent_at() may
3095          * result in split of original leaf or extent zeroout.
3096          */
3097         ext4_ext_drop_refs(path);
3098         path = ext4_ext_find_extent(inode, map->m_lblk, path);
3099         if (IS_ERR(path))
3100                 return PTR_ERR(path);
3101         depth = ext_depth(inode);
3102         ex = path[depth].p_ext;
3103         uninitialized = ext4_ext_is_uninitialized(ex);
3104         split_flag1 = 0;
3105
3106         if (map->m_lblk >= ee_block) {
3107                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3108                 if (uninitialized) {
3109                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
3110                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3111                                                      EXT4_EXT_MARK_UNINIT2);
3112                 }
3113                 err = ext4_split_extent_at(handle, inode, path,
3114                                 map->m_lblk, split_flag1, flags);
3115                 if (err)
3116                         goto out;
3117         }
3118
3119         ext4_ext_show_leaf(inode, path);
3120 out:
3121         return err ? err : allocated;
3122 }
3123
3124 /*
3125  * This function is called by ext4_ext_map_blocks() if someone tries to write
3126  * to an uninitialized extent. It may result in splitting the uninitialized
3127  * extent into multiple extents (up to three - one initialized and two
3128  * uninitialized).
3129  * There are three possibilities:
3130  *   a> There is no split required: Entire extent should be initialized
3131  *   b> Splits in two extents: Write is happening at either end of the extent
3132  *   c> Splits in three extents: Somone is writing in middle of the extent
3133  *
3134  * Pre-conditions:
3135  *  - The extent pointed to by 'path' is uninitialized.
3136  *  - The extent pointed to by 'path' contains a superset
3137  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3138  *
3139  * Post-conditions on success:
3140  *  - the returned value is the number of blocks beyond map->l_lblk
3141  *    that are allocated and initialized.
3142  *    It is guaranteed to be >= map->m_len.
3143  */
3144 static int ext4_ext_convert_to_initialized(handle_t *handle,
3145                                            struct inode *inode,
3146                                            struct ext4_map_blocks *map,
3147                                            struct ext4_ext_path *path)
3148 {
3149         struct ext4_sb_info *sbi;
3150         struct ext4_extent_header *eh;
3151         struct ext4_map_blocks split_map;
3152         struct ext4_extent zero_ex;
3153         struct ext4_extent *ex;
3154         ext4_lblk_t ee_block, eof_block;
3155         unsigned int ee_len, depth;
3156         int allocated, max_zeroout = 0;
3157         int err = 0;
3158         int split_flag = 0;
3159
3160         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3161                 "block %llu, max_blocks %u\n", inode->i_ino,
3162                 (unsigned long long)map->m_lblk, map->m_len);
3163
3164         sbi = EXT4_SB(inode->i_sb);
3165         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3166                 inode->i_sb->s_blocksize_bits;
3167         if (eof_block < map->m_lblk + map->m_len)
3168                 eof_block = map->m_lblk + map->m_len;
3169
3170         depth = ext_depth(inode);
3171         eh = path[depth].p_hdr;
3172         ex = path[depth].p_ext;
3173         ee_block = le32_to_cpu(ex->ee_block);
3174         ee_len = ext4_ext_get_actual_len(ex);
3175         allocated = ee_len - (map->m_lblk - ee_block);
3176         zero_ex.ee_len = 0;
3177
3178         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3179
3180         /* Pre-conditions */
3181         BUG_ON(!ext4_ext_is_uninitialized(ex));
3182         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3183
3184         /*
3185          * Attempt to transfer newly initialized blocks from the currently
3186          * uninitialized extent to its left neighbor. This is much cheaper
3187          * than an insertion followed by a merge as those involve costly
3188          * memmove() calls. This is the common case in steady state for
3189          * workloads doing fallocate(FALLOC_FL_KEEP_SIZE) followed by append
3190          * writes.
3191          *
3192          * Limitations of the current logic:
3193          *  - L1: we only deal with writes at the start of the extent.
3194          *    The approach could be extended to writes at the end
3195          *    of the extent but this scenario was deemed less common.
3196          *  - L2: we do not deal with writes covering the whole extent.
3197          *    This would require removing the extent if the transfer
3198          *    is possible.
3199          *  - L3: we only attempt to merge with an extent stored in the
3200          *    same extent tree node.
3201          */
3202         if ((map->m_lblk == ee_block) &&        /*L1*/
3203                 (map->m_len < ee_len) &&        /*L2*/
3204                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L3*/
3205                 struct ext4_extent *prev_ex;
3206                 ext4_lblk_t prev_lblk;
3207                 ext4_fsblk_t prev_pblk, ee_pblk;
3208                 unsigned int prev_len, write_len;
3209
3210                 prev_ex = ex - 1;
3211                 prev_lblk = le32_to_cpu(prev_ex->ee_block);
3212                 prev_len = ext4_ext_get_actual_len(prev_ex);
3213                 prev_pblk = ext4_ext_pblock(prev_ex);
3214                 ee_pblk = ext4_ext_pblock(ex);
3215                 write_len = map->m_len;
3216
3217                 /*
3218                  * A transfer of blocks from 'ex' to 'prev_ex' is allowed
3219                  * upon those conditions:
3220                  * - C1: prev_ex is initialized,
3221                  * - C2: prev_ex is logically abutting ex,
3222                  * - C3: prev_ex is physically abutting ex,
3223                  * - C4: prev_ex can receive the additional blocks without
3224                  *   overflowing the (initialized) length limit.
3225                  */
3226                 if ((!ext4_ext_is_uninitialized(prev_ex)) &&            /*C1*/
3227                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3228                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3229                         (prev_len < (EXT_INIT_MAX_LEN - write_len))) {  /*C4*/
3230                         err = ext4_ext_get_access(handle, inode, path + depth);
3231                         if (err)
3232                                 goto out;
3233
3234                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3235                                 map, ex, prev_ex);
3236
3237                         /* Shift the start of ex by 'write_len' blocks */
3238                         ex->ee_block = cpu_to_le32(ee_block + write_len);
3239                         ext4_ext_store_pblock(ex, ee_pblk + write_len);
3240                         ex->ee_len = cpu_to_le16(ee_len - write_len);
3241                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3242
3243                         /* Extend prev_ex by 'write_len' blocks */
3244                         prev_ex->ee_len = cpu_to_le16(prev_len + write_len);
3245
3246                         /* Mark the block containing both extents as dirty */
3247                         ext4_ext_dirty(handle, inode, path + depth);
3248
3249                         /* Update path to point to the right extent */
3250                         path[depth].p_ext = prev_ex;
3251
3252                         /* Result: number of initialized blocks past m_lblk */
3253                         allocated = write_len;
3254                         goto out;
3255                 }
3256         }
3257
3258         WARN_ON(map->m_lblk < ee_block);
3259         /*
3260          * It is safe to convert extent to initialized via explicit
3261          * zeroout only if extent is fully insde i_size or new_size.
3262          */
3263         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3264
3265         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3266                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3267                         (inode->i_sb->s_blocksize_bits - 10);
3268
3269         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3270         if (max_zeroout && (ee_len <= max_zeroout)) {
3271                 err = ext4_ext_zeroout(inode, ex);
3272                 if (err)
3273                         goto out;
3274                 zero_ex.ee_block = ex->ee_block;
3275                 zero_ex.ee_len = ext4_ext_get_actual_len(ex);
3276                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3277
3278                 err = ext4_ext_get_access(handle, inode, path + depth);
3279                 if (err)
3280                         goto out;
3281                 ext4_ext_mark_initialized(ex);
3282                 ext4_ext_try_to_merge(handle, inode, path, ex);
3283                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3284                 goto out;
3285         }
3286
3287         /*
3288          * four cases:
3289          * 1. split the extent into three extents.
3290          * 2. split the extent into two extents, zeroout the first half.
3291          * 3. split the extent into two extents, zeroout the second half.
3292          * 4. split the extent into two extents with out zeroout.
3293          */
3294         split_map.m_lblk = map->m_lblk;
3295         split_map.m_len = map->m_len;
3296
3297         if (max_zeroout && (allocated > map->m_len)) {
3298                 if (allocated <= max_zeroout) {
3299                         /* case 3 */
3300                         zero_ex.ee_block =
3301                                          cpu_to_le32(map->m_lblk);
3302                         zero_ex.ee_len = cpu_to_le16(allocated);
3303                         ext4_ext_store_pblock(&zero_ex,
3304                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3305                         err = ext4_ext_zeroout(inode, &zero_ex);
3306                         if (err)
3307                                 goto out;
3308                         split_map.m_lblk = map->m_lblk;
3309                         split_map.m_len = allocated;
3310                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3311                         /* case 2 */
3312                         if (map->m_lblk != ee_block) {
3313                                 zero_ex.ee_block = ex->ee_block;
3314                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3315                                                         ee_block);
3316                                 ext4_ext_store_pblock(&zero_ex,
3317                                                       ext4_ext_pblock(ex));
3318                                 err = ext4_ext_zeroout(inode, &zero_ex);
3319                                 if (err)
3320                                         goto out;
3321                         }
3322
3323                         split_map.m_lblk = ee_block;
3324                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3325                         allocated = map->m_len;
3326                 }
3327         }
3328
3329         allocated = ext4_split_extent(handle, inode, path,
3330                                       &split_map, split_flag, 0);
3331         if (allocated < 0)
3332                 err = allocated;
3333
3334 out:
3335         /* If we have gotten a failure, don't zero out status tree */
3336         if (!err)
3337                 err = ext4_es_zeroout(inode, &zero_ex);
3338         return err ? err : allocated;
3339 }
3340
3341 /*
3342  * This function is called by ext4_ext_map_blocks() from
3343  * ext4_get_blocks_dio_write() when DIO to write
3344  * to an uninitialized extent.
3345  *
3346  * Writing to an uninitialized extent may result in splitting the uninitialized
3347  * extent into multiple initialized/uninitialized extents (up to three)
3348  * There are three possibilities:
3349  *   a> There is no split required: Entire extent should be uninitialized
3350  *   b> Splits in two extents: Write is happening at either end of the extent
3351  *   c> Splits in three extents: Somone is writing in middle of the extent
3352  *
3353  * One of more index blocks maybe needed if the extent tree grow after
3354  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3355  * complete, we need to split the uninitialized extent before DIO submit
3356  * the IO. The uninitialized extent called at this time will be split
3357  * into three uninitialized extent(at most). After IO complete, the part
3358  * being filled will be convert to initialized by the end_io callback function
3359  * via ext4_convert_unwritten_extents().
3360  *
3361  * Returns the size of uninitialized extent to be written on success.
3362  */
3363 static int ext4_split_unwritten_extents(handle_t *handle,
3364                                         struct inode *inode,
3365                                         struct ext4_map_blocks *map,
3366                                         struct ext4_ext_path *path,
3367                                         int flags)
3368 {
3369         ext4_lblk_t eof_block;
3370         ext4_lblk_t ee_block;
3371         struct ext4_extent *ex;
3372         unsigned int ee_len;
3373         int split_flag = 0, depth;
3374
3375         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3376                 "block %llu, max_blocks %u\n", inode->i_ino,
3377                 (unsigned long long)map->m_lblk, map->m_len);
3378
3379         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3380                 inode->i_sb->s_blocksize_bits;
3381         if (eof_block < map->m_lblk + map->m_len)
3382                 eof_block = map->m_lblk + map->m_len;
3383         /*
3384          * It is safe to convert extent to initialized via explicit
3385          * zeroout only if extent is fully insde i_size or new_size.
3386          */
3387         depth = ext_depth(inode);
3388         ex = path[depth].p_ext;
3389         ee_block = le32_to_cpu(ex->ee_block);
3390         ee_len = ext4_ext_get_actual_len(ex);
3391
3392         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3393         split_flag |= EXT4_EXT_MARK_UNINIT2;
3394         if (flags & EXT4_GET_BLOCKS_CONVERT)
3395                 split_flag |= EXT4_EXT_DATA_VALID2;
3396         flags |= EXT4_GET_BLOCKS_PRE_IO;
3397         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3398 }
3399
3400 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3401                                                 struct inode *inode,
3402                                                 struct ext4_map_blocks *map,
3403                                                 struct ext4_ext_path *path)
3404 {
3405         struct ext4_extent *ex;
3406         ext4_lblk_t ee_block;
3407         unsigned int ee_len;
3408         int depth;
3409         int err = 0;
3410
3411         depth = ext_depth(inode);
3412         ex = path[depth].p_ext;
3413         ee_block = le32_to_cpu(ex->ee_block);
3414         ee_len = ext4_ext_get_actual_len(ex);
3415
3416         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3417                 "block %llu, max_blocks %u\n", inode->i_ino,
3418                   (unsigned long long)ee_block, ee_len);
3419
3420         /* If extent is larger than requested it is a clear sign that we still
3421          * have some extent state machine issues left. So extent_split is still
3422          * required.
3423          * TODO: Once all related issues will be fixed this situation should be
3424          * illegal.
3425          */
3426         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3427 #ifdef EXT4_DEBUG
3428                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3429                              " len %u; IO logical block %llu, len %u\n",
3430                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3431                              (unsigned long long)map->m_lblk, map->m_len);
3432 #endif
3433                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3434                                                    EXT4_GET_BLOCKS_CONVERT);
3435                 if (err < 0)
3436                         goto out;
3437                 ext4_ext_drop_refs(path);
3438                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3439                 if (IS_ERR(path)) {
3440                         err = PTR_ERR(path);
3441                         goto out;
3442                 }
3443                 depth = ext_depth(inode);
3444                 ex = path[depth].p_ext;
3445         }
3446
3447         err = ext4_ext_get_access(handle, inode, path + depth);
3448         if (err)
3449                 goto out;
3450         /* first mark the extent as initialized */
3451         ext4_ext_mark_initialized(ex);
3452
3453         /* note: ext4_ext_correct_indexes() isn't needed here because
3454          * borders are not changed
3455          */
3456         ext4_ext_try_to_merge(handle, inode, path, ex);
3457
3458         /* Mark modified extent as dirty */
3459         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3460 out:
3461         ext4_ext_show_leaf(inode, path);
3462         return err;
3463 }
3464
3465 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3466                         sector_t block, int count)
3467 {
3468         int i;
3469         for (i = 0; i < count; i++)
3470                 unmap_underlying_metadata(bdev, block + i);
3471 }
3472
3473 /*
3474  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3475  */
3476 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3477                               ext4_lblk_t lblk,
3478                               struct ext4_ext_path *path,
3479                               unsigned int len)
3480 {
3481         int i, depth;
3482         struct ext4_extent_header *eh;
3483         struct ext4_extent *last_ex;
3484
3485         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3486                 return 0;
3487
3488         depth = ext_depth(inode);
3489         eh = path[depth].p_hdr;
3490
3491         /*
3492          * We're going to remove EOFBLOCKS_FL entirely in future so we
3493          * do not care for this case anymore. Simply remove the flag
3494          * if there are no extents.
3495          */
3496         if (unlikely(!eh->eh_entries))
3497                 goto out;
3498         last_ex = EXT_LAST_EXTENT(eh);
3499         /*
3500          * We should clear the EOFBLOCKS_FL flag if we are writing the
3501          * last block in the last extent in the file.  We test this by
3502          * first checking to see if the caller to
3503          * ext4_ext_get_blocks() was interested in the last block (or
3504          * a block beyond the last block) in the current extent.  If
3505          * this turns out to be false, we can bail out from this
3506          * function immediately.
3507          */
3508         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3509             ext4_ext_get_actual_len(last_ex))
3510                 return 0;
3511         /*
3512          * If the caller does appear to be planning to write at or
3513          * beyond the end of the current extent, we then test to see
3514          * if the current extent is the last extent in the file, by
3515          * checking to make sure it was reached via the rightmost node
3516          * at each level of the tree.
3517          */
3518         for (i = depth-1; i >= 0; i--)
3519                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3520                         return 0;
3521 out:
3522         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3523         return ext4_mark_inode_dirty(handle, inode);
3524 }
3525
3526 /**
3527  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3528  *
3529  * Return 1 if there is a delalloc block in the range, otherwise 0.
3530  */
3531 int ext4_find_delalloc_range(struct inode *inode,
3532                              ext4_lblk_t lblk_start,
3533                              ext4_lblk_t lblk_end)
3534 {
3535         struct extent_status es;
3536
3537         ext4_es_find_delayed_extent(inode, lblk_start, &es);
3538         if (es.es_len == 0)
3539                 return 0; /* there is no delay extent in this tree */
3540         else if (es.es_lblk <= lblk_start &&
3541                  lblk_start < es.es_lblk + es.es_len)
3542                 return 1;
3543         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3544                 return 1;
3545         else
3546                 return 0;
3547 }
3548
3549 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3550 {
3551         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3552         ext4_lblk_t lblk_start, lblk_end;
3553         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3554         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3555
3556         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3557 }
3558
3559 /**
3560  * Determines how many complete clusters (out of those specified by the 'map')
3561  * are under delalloc and were reserved quota for.
3562  * This function is called when we are writing out the blocks that were
3563  * originally written with their allocation delayed, but then the space was
3564  * allocated using fallocate() before the delayed allocation could be resolved.
3565  * The cases to look for are:
3566  * ('=' indicated delayed allocated blocks
3567  *  '-' indicates non-delayed allocated blocks)
3568  * (a) partial clusters towards beginning and/or end outside of allocated range
3569  *     are not delalloc'ed.
3570  *      Ex:
3571  *      |----c---=|====c====|====c====|===-c----|
3572  *               |++++++ allocated ++++++|
3573  *      ==> 4 complete clusters in above example
3574  *
3575  * (b) partial cluster (outside of allocated range) towards either end is
3576  *     marked for delayed allocation. In this case, we will exclude that
3577  *     cluster.
3578  *      Ex:
3579  *      |----====c========|========c========|
3580  *           |++++++ allocated ++++++|
3581  *      ==> 1 complete clusters in above example
3582  *
3583  *      Ex:
3584  *      |================c================|
3585  *            |++++++ allocated ++++++|
3586  *      ==> 0 complete clusters in above example
3587  *
3588  * The ext4_da_update_reserve_space will be called only if we
3589  * determine here that there were some "entire" clusters that span
3590  * this 'allocated' range.
3591  * In the non-bigalloc case, this function will just end up returning num_blks
3592  * without ever calling ext4_find_delalloc_range.
3593  */
3594 static unsigned int
3595 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3596                            unsigned int num_blks)
3597 {
3598         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3599         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3600         ext4_lblk_t lblk_from, lblk_to, c_offset;
3601         unsigned int allocated_clusters = 0;
3602
3603         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3604         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3605
3606         /* max possible clusters for this allocation */
3607         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3608
3609         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3610
3611         /* Check towards left side */
3612         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3613         if (c_offset) {
3614                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3615                 lblk_to = lblk_from + c_offset - 1;
3616
3617                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3618                         allocated_clusters--;
3619         }
3620
3621         /* Now check towards right. */
3622         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3623         if (allocated_clusters && c_offset) {
3624                 lblk_from = lblk_start + num_blks;
3625                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3626
3627                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3628                         allocated_clusters--;
3629         }
3630
3631         return allocated_clusters;
3632 }
3633
3634 static int
3635 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3636                         struct ext4_map_blocks *map,
3637                         struct ext4_ext_path *path, int flags,
3638                         unsigned int allocated, ext4_fsblk_t newblock)
3639 {
3640         int ret = 0;
3641         int err = 0;
3642         ext4_io_end_t *io = ext4_inode_aio(inode);
3643
3644         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3645                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3646                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3647                   flags, allocated);
3648         ext4_ext_show_leaf(inode, path);
3649
3650         trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
3651                                                     allocated, newblock);
3652
3653         /* get_block() before submit the IO, split the extent */
3654         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3655                 ret = ext4_split_unwritten_extents(handle, inode, map,
3656                                                    path, flags);
3657                 if (ret <= 0)
3658                         goto out;
3659                 /*
3660                  * Flag the inode(non aio case) or end_io struct (aio case)
3661                  * that this IO needs to conversion to written when IO is
3662                  * completed
3663                  */
3664                 if (io)
3665                         ext4_set_io_unwritten_flag(inode, io);
3666                 else
3667                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3668                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3669                 if (ext4_should_dioread_nolock(inode))
3670                         map->m_flags |= EXT4_MAP_UNINIT;
3671                 goto out;
3672         }
3673         /* IO end_io complete, convert the filled extent to written */
3674         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3675                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3676                                                         path);
3677                 if (ret >= 0) {
3678                         ext4_update_inode_fsync_trans(handle, inode, 1);
3679                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3680                                                  path, map->m_len);
3681                 } else
3682                         err = ret;
3683                 map->m_flags |= EXT4_MAP_MAPPED;
3684                 if (allocated > map->m_len)
3685                         allocated = map->m_len;
3686                 map->m_len = allocated;
3687                 goto out2;
3688         }
3689         /* buffered IO case */
3690         /*
3691          * repeat fallocate creation request
3692          * we already have an unwritten extent
3693          */
3694         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT) {
3695                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3696                 goto map_out;
3697         }
3698
3699         /* buffered READ or buffered write_begin() lookup */
3700         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3701                 /*
3702                  * We have blocks reserved already.  We
3703                  * return allocated blocks so that delalloc
3704                  * won't do block reservation for us.  But
3705                  * the buffer head will be unmapped so that
3706                  * a read from the block returns 0s.
3707                  */
3708                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3709                 goto out1;
3710         }
3711
3712         /* buffered write, writepage time, convert*/
3713         ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
3714         if (ret >= 0)
3715                 ext4_update_inode_fsync_trans(handle, inode, 1);
3716 out:
3717         if (ret <= 0) {
3718                 err = ret;
3719                 goto out2;
3720         } else
3721                 allocated = ret;
3722         map->m_flags |= EXT4_MAP_NEW;
3723         /*
3724          * if we allocated more blocks than requested
3725          * we need to make sure we unmap the extra block
3726          * allocated. The actual needed block will get
3727          * unmapped later when we find the buffer_head marked
3728          * new.
3729          */
3730         if (allocated > map->m_len) {
3731                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3732                                         newblock + map->m_len,
3733                                         allocated - map->m_len);
3734                 allocated = map->m_len;
3735         }
3736         map->m_len = allocated;
3737
3738         /*
3739          * If we have done fallocate with the offset that is already
3740          * delayed allocated, we would have block reservation
3741          * and quota reservation done in the delayed write path.
3742          * But fallocate would have already updated quota and block
3743          * count for this offset. So cancel these reservation
3744          */
3745         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3746                 unsigned int reserved_clusters;
3747                 reserved_clusters = get_reserved_cluster_alloc(inode,
3748                                 map->m_lblk, map->m_len);
3749                 if (reserved_clusters)
3750                         ext4_da_update_reserve_space(inode,
3751                                                      reserved_clusters,
3752                                                      0);
3753         }
3754
3755 map_out:
3756         map->m_flags |= EXT4_MAP_MAPPED;
3757         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3758                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3759                                          map->m_len);
3760                 if (err < 0)
3761                         goto out2;
3762         }
3763 out1:
3764         if (allocated > map->m_len)
3765                 allocated = map->m_len;
3766         ext4_ext_show_leaf(inode, path);
3767         map->m_pblk = newblock;
3768         map->m_len = allocated;
3769 out2:
3770         if (path) {
3771                 ext4_ext_drop_refs(path);
3772                 kfree(path);
3773         }
3774         return err ? err : allocated;
3775 }
3776
3777 /*
3778  * get_implied_cluster_alloc - check to see if the requested
3779  * allocation (in the map structure) overlaps with a cluster already
3780  * allocated in an extent.
3781  *      @sb     The filesystem superblock structure
3782  *      @map    The requested lblk->pblk mapping
3783  *      @ex     The extent structure which might contain an implied
3784  *                      cluster allocation
3785  *
3786  * This function is called by ext4_ext_map_blocks() after we failed to
3787  * find blocks that were already in the inode's extent tree.  Hence,
3788  * we know that the beginning of the requested region cannot overlap
3789  * the extent from the inode's extent tree.  There are three cases we
3790  * want to catch.  The first is this case:
3791  *
3792  *               |--- cluster # N--|
3793  *    |--- extent ---|  |---- requested region ---|
3794  *                      |==========|
3795  *
3796  * The second case that we need to test for is this one:
3797  *
3798  *   |--------- cluster # N ----------------|
3799  *         |--- requested region --|   |------- extent ----|
3800  *         |=======================|
3801  *
3802  * The third case is when the requested region lies between two extents
3803  * within the same cluster:
3804  *          |------------- cluster # N-------------|
3805  * |----- ex -----|                  |---- ex_right ----|
3806  *                  |------ requested region ------|
3807  *                  |================|
3808  *
3809  * In each of the above cases, we need to set the map->m_pblk and
3810  * map->m_len so it corresponds to the return the extent labelled as
3811  * "|====|" from cluster #N, since it is already in use for data in
3812  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3813  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3814  * as a new "allocated" block region.  Otherwise, we will return 0 and
3815  * ext4_ext_map_blocks() will then allocate one or more new clusters
3816  * by calling ext4_mb_new_blocks().
3817  */
3818 static int get_implied_cluster_alloc(struct super_block *sb,
3819                                      struct ext4_map_blocks *map,
3820                                      struct ext4_extent *ex,
3821                                      struct ext4_ext_path *path)
3822 {
3823         struct ext4_sb_info *sbi = EXT4_SB(sb);
3824         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3825         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3826         ext4_lblk_t rr_cluster_start;
3827         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3828         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3829         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3830
3831         /* The extent passed in that we are trying to match */
3832         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3833         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3834
3835         /* The requested region passed into ext4_map_blocks() */
3836         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3837
3838         if ((rr_cluster_start == ex_cluster_end) ||
3839             (rr_cluster_start == ex_cluster_start)) {
3840                 if (rr_cluster_start == ex_cluster_end)
3841                         ee_start += ee_len - 1;
3842                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3843                         c_offset;
3844                 map->m_len = min(map->m_len,
3845                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3846                 /*
3847                  * Check for and handle this case:
3848                  *
3849                  *   |--------- cluster # N-------------|
3850                  *                     |------- extent ----|
3851                  *         |--- requested region ---|
3852                  *         |===========|
3853                  */
3854
3855                 if (map->m_lblk < ee_block)
3856                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3857
3858                 /*
3859                  * Check for the case where there is already another allocated
3860                  * block to the right of 'ex' but before the end of the cluster.
3861                  *
3862                  *          |------------- cluster # N-------------|
3863                  * |----- ex -----|                  |---- ex_right ----|
3864                  *                  |------ requested region ------|
3865                  *                  |================|
3866                  */
3867                 if (map->m_lblk > ee_block) {
3868                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
3869                         map->m_len = min(map->m_len, next - map->m_lblk);
3870                 }
3871
3872                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
3873                 return 1;
3874         }
3875
3876         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
3877         return 0;
3878 }
3879
3880
3881 /*
3882  * Block allocation/map/preallocation routine for extents based files
3883  *
3884  *
3885  * Need to be called with
3886  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
3887  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
3888  *
3889  * return > 0, number of of blocks already mapped/allocated
3890  *          if create == 0 and these are pre-allocated blocks
3891  *              buffer head is unmapped
3892  *          otherwise blocks are mapped
3893  *
3894  * return = 0, if plain look up failed (blocks have not been allocated)
3895  *          buffer head is unmapped
3896  *
3897  * return < 0, error case.
3898  */
3899 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
3900                         struct ext4_map_blocks *map, int flags)
3901 {
3902         struct ext4_ext_path *path = NULL;
3903         struct ext4_extent newex, *ex, *ex2;
3904         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3905         ext4_fsblk_t newblock = 0;
3906         int free_on_err = 0, err = 0, depth;
3907         unsigned int allocated = 0, offset = 0;
3908         unsigned int allocated_clusters = 0;
3909         struct ext4_allocation_request ar;
3910         ext4_io_end_t *io = ext4_inode_aio(inode);
3911         ext4_lblk_t cluster_offset;
3912         int set_unwritten = 0;
3913
3914         ext_debug("blocks %u/%u requested for inode %lu\n",
3915                   map->m_lblk, map->m_len, inode->i_ino);
3916         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
3917
3918         /* find extent for this block */
3919         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
3920         if (IS_ERR(path)) {
3921                 err = PTR_ERR(path);
3922                 path = NULL;
3923                 goto out2;
3924         }
3925
3926         depth = ext_depth(inode);
3927
3928         /*
3929          * consistent leaf must not be empty;
3930          * this situation is possible, though, _during_ tree modification;
3931          * this is why assert can't be put in ext4_ext_find_extent()
3932          */
3933         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
3934                 EXT4_ERROR_INODE(inode, "bad extent address "
3935                                  "lblock: %lu, depth: %d pblock %lld",
3936                                  (unsigned long) map->m_lblk, depth,
3937                                  path[depth].p_block);
3938                 err = -EIO;
3939                 goto out2;
3940         }
3941
3942         ex = path[depth].p_ext;
3943         if (ex) {
3944                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3945                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3946                 unsigned short ee_len;
3947
3948                 /*
3949                  * Uninitialized extents are treated as holes, except that
3950                  * we split out initialized portions during a write.
3951                  */
3952                 ee_len = ext4_ext_get_actual_len(ex);
3953
3954                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
3955
3956                 /* if found extent covers block, simply return it */
3957                 if (in_range(map->m_lblk, ee_block, ee_len)) {
3958                         newblock = map->m_lblk - ee_block + ee_start;
3959                         /* number of remaining blocks in the extent */
3960                         allocated = ee_len - (map->m_lblk - ee_block);
3961                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
3962                                   ee_block, ee_len, newblock);
3963
3964                         if (!ext4_ext_is_uninitialized(ex))
3965                                 goto out;
3966
3967                         allocated = ext4_ext_handle_uninitialized_extents(
3968                                 handle, inode, map, path, flags,
3969                                 allocated, newblock);
3970                         goto out3;
3971                 }
3972         }
3973
3974         if ((sbi->s_cluster_ratio > 1) &&
3975             ext4_find_delalloc_cluster(inode, map->m_lblk))
3976                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3977
3978         /*
3979          * requested block isn't allocated yet;
3980          * we couldn't try to create block if create flag is zero
3981          */
3982         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3983                 /*
3984                  * put just found gap into cache to speed up
3985                  * subsequent requests
3986                  */
3987                 if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
3988                         ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
3989                 goto out2;
3990         }
3991
3992         /*
3993          * Okay, we need to do block allocation.
3994          */
3995         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
3996         newex.ee_block = cpu_to_le32(map->m_lblk);
3997         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3998
3999         /*
4000          * If we are doing bigalloc, check to see if the extent returned
4001          * by ext4_ext_find_extent() implies a cluster we can use.
4002          */
4003         if (cluster_offset && ex &&
4004             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4005                 ar.len = allocated = map->m_len;
4006                 newblock = map->m_pblk;
4007                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4008                 goto got_allocated_blocks;
4009         }
4010
4011         /* find neighbour allocated blocks */
4012         ar.lleft = map->m_lblk;
4013         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4014         if (err)
4015                 goto out2;
4016         ar.lright = map->m_lblk;
4017         ex2 = NULL;
4018         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4019         if (err)
4020                 goto out2;
4021
4022         /* Check if the extent after searching to the right implies a
4023          * cluster we can use. */
4024         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4025             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4026                 ar.len = allocated = map->m_len;
4027                 newblock = map->m_pblk;
4028                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4029                 goto got_allocated_blocks;
4030         }
4031
4032         /*
4033          * See if request is beyond maximum number of blocks we can have in
4034          * a single extent. For an initialized extent this limit is
4035          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4036          * EXT_UNINIT_MAX_LEN.
4037          */
4038         if (map->m_len > EXT_INIT_MAX_LEN &&
4039             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4040                 map->m_len = EXT_INIT_MAX_LEN;
4041         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4042                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4043                 map->m_len = EXT_UNINIT_MAX_LEN;
4044
4045         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4046         newex.ee_len = cpu_to_le16(map->m_len);
4047         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4048         if (err)
4049                 allocated = ext4_ext_get_actual_len(&newex);
4050         else
4051                 allocated = map->m_len;
4052
4053         /* allocate new block */
4054         ar.inode = inode;
4055         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4056         ar.logical = map->m_lblk;
4057         /*
4058          * We calculate the offset from the beginning of the cluster
4059          * for the logical block number, since when we allocate a
4060          * physical cluster, the physical block should start at the
4061          * same offset from the beginning of the cluster.  This is
4062          * needed so that future calls to get_implied_cluster_alloc()
4063          * work correctly.
4064          */
4065         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
4066         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4067         ar.goal -= offset;
4068         ar.logical -= offset;
4069         if (S_ISREG(inode->i_mode))
4070                 ar.flags = EXT4_MB_HINT_DATA;
4071         else
4072                 /* disable in-core preallocation for non-regular files */
4073                 ar.flags = 0;
4074         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4075                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4076         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4077         if (!newblock)
4078                 goto out2;
4079         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4080                   ar.goal, newblock, allocated);
4081         free_on_err = 1;
4082         allocated_clusters = ar.len;
4083         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4084         if (ar.len > allocated)
4085                 ar.len = allocated;
4086
4087 got_allocated_blocks:
4088         /* try to insert new extent into found leaf and return */
4089         ext4_ext_store_pblock(&newex, newblock + offset);
4090         newex.ee_len = cpu_to_le16(ar.len);
4091         /* Mark uninitialized */
4092         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4093                 ext4_ext_mark_uninitialized(&newex);
4094                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4095                 /*
4096                  * io_end structure was created for every IO write to an
4097                  * uninitialized extent. To avoid unnecessary conversion,
4098                  * here we flag the IO that really needs the conversion.
4099                  * For non asycn direct IO case, flag the inode state
4100                  * that we need to perform conversion when IO is done.
4101                  */
4102                 if ((flags & EXT4_GET_BLOCKS_PRE_IO))
4103                         set_unwritten = 1;
4104                 if (ext4_should_dioread_nolock(inode))
4105                         map->m_flags |= EXT4_MAP_UNINIT;
4106         }
4107
4108         err = 0;
4109         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4110                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4111                                          path, ar.len);
4112         if (!err)
4113                 err = ext4_ext_insert_extent(handle, inode, path,
4114                                              &newex, flags);
4115
4116         if (!err && set_unwritten) {
4117                 if (io)
4118                         ext4_set_io_unwritten_flag(inode, io);
4119                 else
4120                         ext4_set_inode_state(inode,
4121                                              EXT4_STATE_DIO_UNWRITTEN);
4122         }
4123
4124         if (err && free_on_err) {
4125                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4126                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4127                 /* free data blocks we just allocated */
4128                 /* not a good idea to call discard here directly,
4129                  * but otherwise we'd need to call it every free() */
4130                 ext4_discard_preallocations(inode);
4131                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4132                                  ext4_ext_get_actual_len(&newex), fb_flags);
4133                 goto out2;
4134         }
4135
4136         /* previous routine could use block we allocated */
4137         newblock = ext4_ext_pblock(&newex);
4138         allocated = ext4_ext_get_actual_len(&newex);
4139         if (allocated > map->m_len)
4140                 allocated = map->m_len;
4141         map->m_flags |= EXT4_MAP_NEW;
4142
4143         /*
4144          * Update reserved blocks/metadata blocks after successful
4145          * block allocation which had been deferred till now.
4146          */
4147         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4148                 unsigned int reserved_clusters;
4149                 /*
4150                  * Check how many clusters we had reserved this allocated range
4151                  */
4152                 reserved_clusters = get_reserved_cluster_alloc(inode,
4153                                                 map->m_lblk, allocated);
4154                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4155                         if (reserved_clusters) {
4156                                 /*
4157                                  * We have clusters reserved for this range.
4158                                  * But since we are not doing actual allocation
4159                                  * and are simply using blocks from previously
4160                                  * allocated cluster, we should release the
4161                                  * reservation and not claim quota.
4162                                  */
4163                                 ext4_da_update_reserve_space(inode,
4164                                                 reserved_clusters, 0);
4165                         }
4166                 } else {
4167                         BUG_ON(allocated_clusters < reserved_clusters);
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                          * We will claim quota for all newly allocated blocks.
4220                          * We're updating the reserved space *after* the
4221                          * correction above so we do not accidentally free
4222                          * all the metadata reservation because we might
4223                          * actually need it later on.
4224                          */
4225                         ext4_da_update_reserve_space(inode, allocated_clusters,
4226                                                         1);
4227                 }
4228         }
4229
4230         /*
4231          * Cache the extent and update transaction to commit on fdatasync only
4232          * when it is _not_ an uninitialized extent.
4233          */
4234         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0)
4235                 ext4_update_inode_fsync_trans(handle, inode, 1);
4236         else
4237                 ext4_update_inode_fsync_trans(handle, inode, 0);
4238 out:
4239         if (allocated > map->m_len)
4240                 allocated = map->m_len;
4241         ext4_ext_show_leaf(inode, path);
4242         map->m_flags |= EXT4_MAP_MAPPED;
4243         map->m_pblk = newblock;
4244         map->m_len = allocated;
4245 out2:
4246         if (path) {
4247                 ext4_ext_drop_refs(path);
4248                 kfree(path);
4249         }
4250
4251 out3:
4252         trace_ext4_ext_map_blocks_exit(inode, map, err ? err : allocated);
4253
4254         return err ? err : allocated;
4255 }
4256
4257 void ext4_ext_truncate(struct inode *inode)
4258 {
4259         struct address_space *mapping = inode->i_mapping;
4260         struct super_block *sb = inode->i_sb;
4261         ext4_lblk_t last_block;
4262         handle_t *handle;
4263         loff_t page_len;
4264         int err = 0;
4265
4266         /*
4267          * finish any pending end_io work so we won't run the risk of
4268          * converting any truncated blocks to initialized later
4269          */
4270         ext4_flush_unwritten_io(inode);
4271
4272         /*
4273          * probably first extent we're gonna free will be last in block
4274          */
4275         err = ext4_writepage_trans_blocks(inode);
4276         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, err);
4277         if (IS_ERR(handle))
4278                 return;
4279
4280         if (inode->i_size % PAGE_CACHE_SIZE != 0) {
4281                 page_len = PAGE_CACHE_SIZE -
4282                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4283
4284                 err = ext4_discard_partial_page_buffers(handle,
4285                         mapping, inode->i_size, page_len, 0);
4286
4287                 if (err)
4288                         goto out_stop;
4289         }
4290
4291         if (ext4_orphan_add(handle, inode))
4292                 goto out_stop;
4293
4294         down_write(&EXT4_I(inode)->i_data_sem);
4295
4296         ext4_discard_preallocations(inode);
4297
4298         /*
4299          * TODO: optimization is possible here.
4300          * Probably we need not scan at all,
4301          * because page truncation is enough.
4302          */
4303
4304         /* we have to know where to truncate from in crash case */
4305         EXT4_I(inode)->i_disksize = inode->i_size;
4306         ext4_mark_inode_dirty(handle, inode);
4307
4308         last_block = (inode->i_size + sb->s_blocksize - 1)
4309                         >> EXT4_BLOCK_SIZE_BITS(sb);
4310         err = ext4_es_remove_extent(inode, last_block,
4311                                     EXT_MAX_BLOCKS - last_block);
4312         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4313
4314         /* In a multi-transaction truncate, we only make the final
4315          * transaction synchronous.
4316          */
4317         if (IS_SYNC(inode))
4318                 ext4_handle_sync(handle);
4319
4320         up_write(&EXT4_I(inode)->i_data_sem);
4321
4322 out_stop:
4323         /*
4324          * If this was a simple ftruncate() and the file will remain alive,
4325          * then we need to clear up the orphan record which we created above.
4326          * However, if this was a real unlink then we were called by
4327          * ext4_delete_inode(), and we allow that function to clean up the
4328          * orphan info for us.
4329          */
4330         if (inode->i_nlink)
4331                 ext4_orphan_del(handle, inode);
4332
4333         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4334         ext4_mark_inode_dirty(handle, inode);
4335         ext4_journal_stop(handle);
4336 }
4337
4338 static void ext4_falloc_update_inode(struct inode *inode,
4339                                 int mode, loff_t new_size, int update_ctime)
4340 {
4341         struct timespec now;
4342
4343         if (update_ctime) {
4344                 now = current_fs_time(inode->i_sb);
4345                 if (!timespec_equal(&inode->i_ctime, &now))
4346                         inode->i_ctime = now;
4347         }
4348         /*
4349          * Update only when preallocation was requested beyond
4350          * the file size.
4351          */
4352         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4353                 if (new_size > i_size_read(inode))
4354                         i_size_write(inode, new_size);
4355                 if (new_size > EXT4_I(inode)->i_disksize)
4356                         ext4_update_i_disksize(inode, new_size);
4357         } else {
4358                 /*
4359                  * Mark that we allocate beyond EOF so the subsequent truncate
4360                  * can proceed even if the new size is the same as i_size.
4361                  */
4362                 if (new_size > i_size_read(inode))
4363                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4364         }
4365
4366 }
4367
4368 /*
4369  * preallocate space for a file. This implements ext4's fallocate file
4370  * operation, which gets called from sys_fallocate system call.
4371  * For block-mapped files, posix_fallocate should fall back to the method
4372  * of writing zeroes to the required new blocks (the same behavior which is
4373  * expected for file systems which do not support fallocate() system call).
4374  */
4375 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4376 {
4377         struct inode *inode = file_inode(file);
4378         handle_t *handle;
4379         loff_t new_size;
4380         unsigned int max_blocks;
4381         int ret = 0;
4382         int ret2 = 0;
4383         int retries = 0;
4384         int flags;
4385         struct ext4_map_blocks map;
4386         unsigned int credits, blkbits = inode->i_blkbits;
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         ret = ext4_convert_inline_data(inode);
4396         if (ret)
4397                 return ret;
4398
4399         /*
4400          * currently supporting (pre)allocate mode for extent-based
4401          * files _only_
4402          */
4403         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4404                 return -EOPNOTSUPP;
4405
4406         trace_ext4_fallocate_enter(inode, offset, len, mode);
4407         map.m_lblk = offset >> blkbits;
4408         /*
4409          * We can't just convert len to max_blocks because
4410          * If blocksize = 4096 offset = 3072 and len = 2048
4411          */
4412         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4413                 - map.m_lblk;
4414         /*
4415          * credits to insert 1 extent into extent tree
4416          */
4417         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4418         mutex_lock(&inode->i_mutex);
4419         ret = inode_newsize_ok(inode, (len + offset));
4420         if (ret) {
4421                 mutex_unlock(&inode->i_mutex);
4422                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4423                 return ret;
4424         }
4425         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4426         if (mode & FALLOC_FL_KEEP_SIZE)
4427                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4428         /*
4429          * Don't normalize the request if it can fit in one extent so
4430          * that it doesn't get unnecessarily split into multiple
4431          * extents.
4432          */
4433         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4434                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4435
4436 retry:
4437         while (ret >= 0 && ret < max_blocks) {
4438                 map.m_lblk = map.m_lblk + ret;
4439                 map.m_len = max_blocks = max_blocks - ret;
4440                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4441                                             credits);
4442                 if (IS_ERR(handle)) {
4443                         ret = PTR_ERR(handle);
4444                         break;
4445                 }
4446                 ret = ext4_map_blocks(handle, inode, &map, flags);
4447                 if (ret <= 0) {
4448 #ifdef EXT4FS_DEBUG
4449                         ext4_warning(inode->i_sb,
4450                                      "inode #%lu: block %u: len %u: "
4451                                      "ext4_ext_map_blocks returned %d",
4452                                      inode->i_ino, map.m_lblk,
4453                                      map.m_len, ret);
4454 #endif
4455                         ext4_mark_inode_dirty(handle, inode);
4456                         ret2 = ext4_journal_stop(handle);
4457                         break;
4458                 }
4459                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4460                                                 blkbits) >> blkbits))
4461                         new_size = offset + len;
4462                 else
4463                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4464
4465                 ext4_falloc_update_inode(inode, mode, new_size,
4466                                          (map.m_flags & EXT4_MAP_NEW));
4467                 ext4_mark_inode_dirty(handle, inode);
4468                 if ((file->f_flags & O_SYNC) && ret >= max_blocks)
4469                         ext4_handle_sync(handle);
4470                 ret2 = ext4_journal_stop(handle);
4471                 if (ret2)
4472                         break;
4473         }
4474         if (ret == -ENOSPC &&
4475                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4476                 ret = 0;
4477                 goto retry;
4478         }
4479         mutex_unlock(&inode->i_mutex);
4480         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4481                                 ret > 0 ? ret2 : ret);
4482         return ret > 0 ? ret2 : ret;
4483 }
4484
4485 /*
4486  * This function convert a range of blocks to written extents
4487  * The caller of this function will pass the start offset and the size.
4488  * all unwritten extents within this range will be converted to
4489  * written extents.
4490  *
4491  * This function is called from the direct IO end io call back
4492  * function, to convert the fallocated extents after IO is completed.
4493  * Returns 0 on success.
4494  */
4495 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4496                                     ssize_t len)
4497 {
4498         handle_t *handle;
4499         unsigned int max_blocks;
4500         int ret = 0;
4501         int ret2 = 0;
4502         struct ext4_map_blocks map;
4503         unsigned int credits, blkbits = inode->i_blkbits;
4504
4505         map.m_lblk = offset >> blkbits;
4506         /*
4507          * We can't just convert len to max_blocks because
4508          * If blocksize = 4096 offset = 3072 and len = 2048
4509          */
4510         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4511                       map.m_lblk);
4512         /*
4513          * credits to insert 1 extent into extent tree
4514          */
4515         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4516         while (ret >= 0 && ret < max_blocks) {
4517                 map.m_lblk += ret;
4518                 map.m_len = (max_blocks -= ret);
4519                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
4520                 if (IS_ERR(handle)) {
4521                         ret = PTR_ERR(handle);
4522                         break;
4523                 }
4524                 ret = ext4_map_blocks(handle, inode, &map,
4525                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4526                 if (ret <= 0)
4527                         ext4_warning(inode->i_sb,
4528                                      "inode #%lu: block %u: len %u: "
4529                                      "ext4_ext_map_blocks returned %d",
4530                                      inode->i_ino, map.m_lblk,
4531                                      map.m_len, ret);
4532                 ext4_mark_inode_dirty(handle, inode);
4533                 ret2 = ext4_journal_stop(handle);
4534                 if (ret <= 0 || ret2 )
4535                         break;
4536         }
4537         return ret > 0 ? ret2 : ret;
4538 }
4539
4540 /*
4541  * If newes is not existing extent (newes->ec_pblk equals zero) find
4542  * delayed extent at start of newes and update newes accordingly and
4543  * return start of the next delayed extent.
4544  *
4545  * If newes is existing extent (newes->ec_pblk is not equal zero)
4546  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4547  * extent found. Leave newes unmodified.
4548  */
4549 static int ext4_find_delayed_extent(struct inode *inode,
4550                                     struct extent_status *newes)
4551 {
4552         struct extent_status es;
4553         ext4_lblk_t block, next_del;
4554
4555         ext4_es_find_delayed_extent(inode, newes->es_lblk, &es);
4556
4557         if (newes->es_pblk == 0) {
4558                 /*
4559                  * No extent in extent-tree contains block @newes->es_pblk,
4560                  * then the block may stay in 1)a hole or 2)delayed-extent.
4561                  */
4562                 if (es.es_len == 0)
4563                         /* A hole found. */
4564                         return 0;
4565
4566                 if (es.es_lblk > newes->es_lblk) {
4567                         /* A hole found. */
4568                         newes->es_len = min(es.es_lblk - newes->es_lblk,
4569                                             newes->es_len);
4570                         return 0;
4571                 }
4572
4573                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
4574         }
4575
4576         block = newes->es_lblk + newes->es_len;
4577         ext4_es_find_delayed_extent(inode, block, &es);
4578         if (es.es_len == 0)
4579                 next_del = EXT_MAX_BLOCKS;
4580         else
4581                 next_del = es.es_lblk;
4582
4583         return next_del;
4584 }
4585 /* fiemap flags we can handle specified here */
4586 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4587
4588 static int ext4_xattr_fiemap(struct inode *inode,
4589                                 struct fiemap_extent_info *fieinfo)
4590 {
4591         __u64 physical = 0;
4592         __u64 length;
4593         __u32 flags = FIEMAP_EXTENT_LAST;
4594         int blockbits = inode->i_sb->s_blocksize_bits;
4595         int error = 0;
4596
4597         /* in-inode? */
4598         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4599                 struct ext4_iloc iloc;
4600                 int offset;     /* offset of xattr in inode */
4601
4602                 error = ext4_get_inode_loc(inode, &iloc);
4603                 if (error)
4604                         return error;
4605                 physical = iloc.bh->b_blocknr << blockbits;
4606                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4607                                 EXT4_I(inode)->i_extra_isize;
4608                 physical += offset;
4609                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4610                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4611                 brelse(iloc.bh);
4612         } else { /* external block */
4613                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4614                 length = inode->i_sb->s_blocksize;
4615         }
4616
4617         if (physical)
4618                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4619                                                 length, flags);
4620         return (error < 0 ? error : 0);
4621 }
4622
4623 /*
4624  * ext4_ext_punch_hole
4625  *
4626  * Punches a hole of "length" bytes in a file starting
4627  * at byte "offset"
4628  *
4629  * @inode:  The inode of the file to punch a hole in
4630  * @offset: The starting byte offset of the hole
4631  * @length: The length of the hole
4632  *
4633  * Returns the number of blocks removed or negative on err
4634  */
4635 int ext4_ext_punch_hole(struct file *file, loff_t offset, loff_t length)
4636 {
4637         struct inode *inode = file_inode(file);
4638         struct super_block *sb = inode->i_sb;
4639         ext4_lblk_t first_block, stop_block;
4640         struct address_space *mapping = inode->i_mapping;
4641         handle_t *handle;
4642         loff_t first_page, last_page, page_len;
4643         loff_t first_page_offset, last_page_offset;
4644         int credits, err = 0;
4645
4646         /*
4647          * Write out all dirty pages to avoid race conditions
4648          * Then release them.
4649          */
4650         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4651                 err = filemap_write_and_wait_range(mapping,
4652                         offset, offset + length - 1);
4653
4654                 if (err)
4655                         return err;
4656         }
4657
4658         mutex_lock(&inode->i_mutex);
4659         /* It's not possible punch hole on append only file */
4660         if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) {
4661                 err = -EPERM;
4662                 goto out_mutex;
4663         }
4664         if (IS_SWAPFILE(inode)) {
4665                 err = -ETXTBSY;
4666                 goto out_mutex;
4667         }
4668
4669         /* No need to punch hole beyond i_size */
4670         if (offset >= inode->i_size)
4671                 goto out_mutex;
4672
4673         /*
4674          * If the hole extends beyond i_size, set the hole
4675          * to end after the page that contains i_size
4676          */
4677         if (offset + length > inode->i_size) {
4678                 length = inode->i_size +
4679                    PAGE_CACHE_SIZE - (inode->i_size & (PAGE_CACHE_SIZE - 1)) -
4680                    offset;
4681         }
4682
4683         first_page = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
4684         last_page = (offset + length) >> PAGE_CACHE_SHIFT;
4685
4686         first_page_offset = first_page << PAGE_CACHE_SHIFT;
4687         last_page_offset = last_page << PAGE_CACHE_SHIFT;
4688
4689         /* Now release the pages */
4690         if (last_page_offset > first_page_offset) {
4691                 truncate_pagecache_range(inode, first_page_offset,
4692                                          last_page_offset - 1);
4693         }
4694
4695         /* Wait all existing dio workers, newcomers will block on i_mutex */
4696         ext4_inode_block_unlocked_dio(inode);
4697         err = ext4_flush_unwritten_io(inode);
4698         if (err)
4699                 goto out_dio;
4700         inode_dio_wait(inode);
4701
4702         credits = ext4_writepage_trans_blocks(inode);
4703         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
4704         if (IS_ERR(handle)) {
4705                 err = PTR_ERR(handle);
4706                 goto out_dio;
4707         }
4708
4709
4710         /*
4711          * Now we need to zero out the non-page-aligned data in the
4712          * pages at the start and tail of the hole, and unmap the buffer
4713          * heads for the block aligned regions of the page that were
4714          * completely zeroed.
4715          */
4716         if (first_page > last_page) {
4717                 /*
4718                  * If the file space being truncated is contained within a page
4719                  * just zero out and unmap the middle of that page
4720                  */
4721                 err = ext4_discard_partial_page_buffers(handle,
4722                         mapping, offset, length, 0);
4723
4724                 if (err)
4725                         goto out;
4726         } else {
4727                 /*
4728                  * zero out and unmap the partial page that contains
4729                  * the start of the hole
4730                  */
4731                 page_len  = first_page_offset - offset;
4732                 if (page_len > 0) {
4733                         err = ext4_discard_partial_page_buffers(handle, mapping,
4734                                                    offset, page_len, 0);
4735                         if (err)
4736                                 goto out;
4737                 }
4738
4739                 /*
4740                  * zero out and unmap the partial page that contains
4741                  * the end of the hole
4742                  */
4743                 page_len = offset + length - last_page_offset;
4744                 if (page_len > 0) {
4745                         err = ext4_discard_partial_page_buffers(handle, mapping,
4746                                         last_page_offset, page_len, 0);
4747                         if (err)
4748                                 goto out;
4749                 }
4750         }
4751
4752         /*
4753          * If i_size is contained in the last page, we need to
4754          * unmap and zero the partial page after i_size
4755          */
4756         if (inode->i_size >> PAGE_CACHE_SHIFT == last_page &&
4757            inode->i_size % PAGE_CACHE_SIZE != 0) {
4758
4759                 page_len = PAGE_CACHE_SIZE -
4760                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4761
4762                 if (page_len > 0) {
4763                         err = ext4_discard_partial_page_buffers(handle,
4764                           mapping, inode->i_size, page_len, 0);
4765
4766                         if (err)
4767                                 goto out;
4768                 }
4769         }
4770
4771         first_block = (offset + sb->s_blocksize - 1) >>
4772                 EXT4_BLOCK_SIZE_BITS(sb);
4773         stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4774
4775         /* If there are no blocks to remove, return now */
4776         if (first_block >= stop_block)
4777                 goto out;
4778
4779         down_write(&EXT4_I(inode)->i_data_sem);
4780         ext4_discard_preallocations(inode);
4781
4782         err = ext4_es_remove_extent(inode, first_block,
4783                                     stop_block - first_block);
4784         err = ext4_ext_remove_space(inode, first_block, stop_block - 1);
4785
4786         ext4_discard_preallocations(inode);
4787
4788         if (IS_SYNC(inode))
4789                 ext4_handle_sync(handle);
4790
4791         up_write(&EXT4_I(inode)->i_data_sem);
4792
4793 out:
4794         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4795         ext4_mark_inode_dirty(handle, inode);
4796         ext4_journal_stop(handle);
4797 out_dio:
4798         ext4_inode_resume_unlocked_dio(inode);
4799 out_mutex:
4800         mutex_unlock(&inode->i_mutex);
4801         return err;
4802 }
4803
4804 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4805                 __u64 start, __u64 len)
4806 {
4807         ext4_lblk_t start_blk;
4808         int error = 0;
4809
4810         if (ext4_has_inline_data(inode)) {
4811                 int has_inline = 1;
4812
4813                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
4814
4815                 if (has_inline)
4816                         return error;
4817         }
4818
4819         /* fallback to generic here if not in extents fmt */
4820         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4821                 return generic_block_fiemap(inode, fieinfo, start, len,
4822                         ext4_get_block);
4823
4824         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4825                 return -EBADR;
4826
4827         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4828                 error = ext4_xattr_fiemap(inode, fieinfo);
4829         } else {
4830                 ext4_lblk_t len_blks;
4831                 __u64 last_blk;
4832
4833                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4834                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4835                 if (last_blk >= EXT_MAX_BLOCKS)
4836                         last_blk = EXT_MAX_BLOCKS-1;
4837                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4838
4839                 /*
4840                  * Walk the extent tree gathering extent information
4841                  * and pushing extents back to the user.
4842                  */
4843                 error = ext4_fill_fiemap_extents(inode, start_blk,
4844                                                  len_blks, fieinfo);
4845         }
4846
4847         return error;
4848 }