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         eh = path[depth].p_hdr;
1817         if (unlikely(path[depth].p_hdr == NULL)) {
1818                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1819                 return -EIO;
1820         }
1821
1822         /* try to insert block into found extent and return */
1823         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)) {
1824
1825                 /*
1826                  * Try to see whether we should rather test the extent on
1827                  * right from ex, or from the left of ex. This is because
1828                  * ext4_ext_find_extent() can return either extent on the
1829                  * left, or on the right from the searched position. This
1830                  * will make merging more effective.
1831                  */
1832                 if (ex < EXT_LAST_EXTENT(eh) &&
1833                     (le32_to_cpu(ex->ee_block) +
1834                     ext4_ext_get_actual_len(ex) <
1835                     le32_to_cpu(newext->ee_block))) {
1836                         ex += 1;
1837                         goto prepend;
1838                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1839                            (le32_to_cpu(newext->ee_block) +
1840                            ext4_ext_get_actual_len(newext) <
1841                            le32_to_cpu(ex->ee_block)))
1842                         ex -= 1;
1843
1844                 /* Try to append newex to the ex */
1845                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1846                         ext_debug("append [%d]%d block to %u:[%d]%d"
1847                                   "(from %llu)\n",
1848                                   ext4_ext_is_uninitialized(newext),
1849                                   ext4_ext_get_actual_len(newext),
1850                                   le32_to_cpu(ex->ee_block),
1851                                   ext4_ext_is_uninitialized(ex),
1852                                   ext4_ext_get_actual_len(ex),
1853                                   ext4_ext_pblock(ex));
1854                         err = ext4_ext_get_access(handle, inode,
1855                                                   path + depth);
1856                         if (err)
1857                                 return err;
1858
1859                         /*
1860                          * ext4_can_extents_be_merged should have checked
1861                          * that either both extents are uninitialized, or
1862                          * both aren't. Thus we need to check only one of
1863                          * them here.
1864                          */
1865                         if (ext4_ext_is_uninitialized(ex))
1866                                 uninitialized = 1;
1867                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1868                                         + ext4_ext_get_actual_len(newext));
1869                         if (uninitialized)
1870                                 ext4_ext_mark_uninitialized(ex);
1871                         eh = path[depth].p_hdr;
1872                         nearex = ex;
1873                         goto merge;
1874                 }
1875
1876 prepend:
1877                 /* Try to prepend newex to the ex */
1878                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
1879                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
1880                                   "(from %llu)\n",
1881                                   le32_to_cpu(newext->ee_block),
1882                                   ext4_ext_is_uninitialized(newext),
1883                                   ext4_ext_get_actual_len(newext),
1884                                   le32_to_cpu(ex->ee_block),
1885                                   ext4_ext_is_uninitialized(ex),
1886                                   ext4_ext_get_actual_len(ex),
1887                                   ext4_ext_pblock(ex));
1888                         err = ext4_ext_get_access(handle, inode,
1889                                                   path + depth);
1890                         if (err)
1891                                 return err;
1892
1893                         /*
1894                          * ext4_can_extents_be_merged should have checked
1895                          * that either both extents are uninitialized, or
1896                          * both aren't. Thus we need to check only one of
1897                          * them here.
1898                          */
1899                         if (ext4_ext_is_uninitialized(ex))
1900                                 uninitialized = 1;
1901                         ex->ee_block = newext->ee_block;
1902                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
1903                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1904                                         + ext4_ext_get_actual_len(newext));
1905                         if (uninitialized)
1906                                 ext4_ext_mark_uninitialized(ex);
1907                         eh = path[depth].p_hdr;
1908                         nearex = ex;
1909                         goto merge;
1910                 }
1911         }
1912
1913         depth = ext_depth(inode);
1914         eh = path[depth].p_hdr;
1915         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1916                 goto has_space;
1917
1918         /* probably next leaf has space for us? */
1919         fex = EXT_LAST_EXTENT(eh);
1920         next = EXT_MAX_BLOCKS;
1921         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1922                 next = ext4_ext_next_leaf_block(path);
1923         if (next != EXT_MAX_BLOCKS) {
1924                 ext_debug("next leaf block - %u\n", next);
1925                 BUG_ON(npath != NULL);
1926                 npath = ext4_ext_find_extent(inode, next, NULL);
1927                 if (IS_ERR(npath))
1928                         return PTR_ERR(npath);
1929                 BUG_ON(npath->p_depth != path->p_depth);
1930                 eh = npath[depth].p_hdr;
1931                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1932                         ext_debug("next leaf isn't full(%d)\n",
1933                                   le16_to_cpu(eh->eh_entries));
1934                         path = npath;
1935                         goto has_space;
1936                 }
1937                 ext_debug("next leaf has no free space(%d,%d)\n",
1938                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1939         }
1940
1941         /*
1942          * There is no free space in the found leaf.
1943          * We're gonna add a new leaf in the tree.
1944          */
1945         if (flag & EXT4_GET_BLOCKS_PUNCH_OUT_EXT)
1946                 flags = EXT4_MB_USE_ROOT_BLOCKS;
1947         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1948         if (err)
1949                 goto cleanup;
1950         depth = ext_depth(inode);
1951         eh = path[depth].p_hdr;
1952
1953 has_space:
1954         nearex = path[depth].p_ext;
1955
1956         err = ext4_ext_get_access(handle, inode, path + depth);
1957         if (err)
1958                 goto cleanup;
1959
1960         if (!nearex) {
1961                 /* there is no extent in this leaf, create first one */
1962                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1963                                 le32_to_cpu(newext->ee_block),
1964                                 ext4_ext_pblock(newext),
1965                                 ext4_ext_is_uninitialized(newext),
1966                                 ext4_ext_get_actual_len(newext));
1967                 nearex = EXT_FIRST_EXTENT(eh);
1968         } else {
1969                 if (le32_to_cpu(newext->ee_block)
1970                            > le32_to_cpu(nearex->ee_block)) {
1971                         /* Insert after */
1972                         ext_debug("insert %u:%llu:[%d]%d before: "
1973                                         "nearest %p\n",
1974                                         le32_to_cpu(newext->ee_block),
1975                                         ext4_ext_pblock(newext),
1976                                         ext4_ext_is_uninitialized(newext),
1977                                         ext4_ext_get_actual_len(newext),
1978                                         nearex);
1979                         nearex++;
1980                 } else {
1981                         /* Insert before */
1982                         BUG_ON(newext->ee_block == nearex->ee_block);
1983                         ext_debug("insert %u:%llu:[%d]%d after: "
1984                                         "nearest %p\n",
1985                                         le32_to_cpu(newext->ee_block),
1986                                         ext4_ext_pblock(newext),
1987                                         ext4_ext_is_uninitialized(newext),
1988                                         ext4_ext_get_actual_len(newext),
1989                                         nearex);
1990                 }
1991                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1992                 if (len > 0) {
1993                         ext_debug("insert %u:%llu:[%d]%d: "
1994                                         "move %d extents from 0x%p to 0x%p\n",
1995                                         le32_to_cpu(newext->ee_block),
1996                                         ext4_ext_pblock(newext),
1997                                         ext4_ext_is_uninitialized(newext),
1998                                         ext4_ext_get_actual_len(newext),
1999                                         len, nearex, nearex + 1);
2000                         memmove(nearex + 1, nearex,
2001                                 len * sizeof(struct ext4_extent));
2002                 }
2003         }
2004
2005         le16_add_cpu(&eh->eh_entries, 1);
2006         path[depth].p_ext = nearex;
2007         nearex->ee_block = newext->ee_block;
2008         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2009         nearex->ee_len = newext->ee_len;
2010
2011 merge:
2012         /* try to merge extents */
2013         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
2014                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2015
2016
2017         /* time to correct all indexes above */
2018         err = ext4_ext_correct_indexes(handle, inode, path);
2019         if (err)
2020                 goto cleanup;
2021
2022         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2023
2024 cleanup:
2025         if (npath) {
2026                 ext4_ext_drop_refs(npath);
2027                 kfree(npath);
2028         }
2029         return err;
2030 }
2031
2032 static int ext4_fill_fiemap_extents(struct inode *inode,
2033                                     ext4_lblk_t block, ext4_lblk_t num,
2034                                     struct fiemap_extent_info *fieinfo)
2035 {
2036         struct ext4_ext_path *path = NULL;
2037         struct ext4_extent *ex;
2038         struct extent_status es;
2039         ext4_lblk_t next, next_del, start = 0, end = 0;
2040         ext4_lblk_t last = block + num;
2041         int exists, depth = 0, err = 0;
2042         unsigned int flags = 0;
2043         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2044
2045         while (block < last && block != EXT_MAX_BLOCKS) {
2046                 num = last - block;
2047                 /* find extent for this block */
2048                 down_read(&EXT4_I(inode)->i_data_sem);
2049
2050                 if (path && ext_depth(inode) != depth) {
2051                         /* depth was changed. we have to realloc path */
2052                         kfree(path);
2053                         path = NULL;
2054                 }
2055
2056                 path = ext4_ext_find_extent(inode, block, path);
2057                 if (IS_ERR(path)) {
2058                         up_read(&EXT4_I(inode)->i_data_sem);
2059                         err = PTR_ERR(path);
2060                         path = NULL;
2061                         break;
2062                 }
2063
2064                 depth = ext_depth(inode);
2065                 if (unlikely(path[depth].p_hdr == NULL)) {
2066                         up_read(&EXT4_I(inode)->i_data_sem);
2067                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2068                         err = -EIO;
2069                         break;
2070                 }
2071                 ex = path[depth].p_ext;
2072                 next = ext4_ext_next_allocated_block(path);
2073                 ext4_ext_drop_refs(path);
2074
2075                 flags = 0;
2076                 exists = 0;
2077                 if (!ex) {
2078                         /* there is no extent yet, so try to allocate
2079                          * all requested space */
2080                         start = block;
2081                         end = block + num;
2082                 } else if (le32_to_cpu(ex->ee_block) > block) {
2083                         /* need to allocate space before found extent */
2084                         start = block;
2085                         end = le32_to_cpu(ex->ee_block);
2086                         if (block + num < end)
2087                                 end = block + num;
2088                 } else if (block >= le32_to_cpu(ex->ee_block)
2089                                         + ext4_ext_get_actual_len(ex)) {
2090                         /* need to allocate space after found extent */
2091                         start = block;
2092                         end = block + num;
2093                         if (end >= next)
2094                                 end = next;
2095                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2096                         /*
2097                          * some part of requested space is covered
2098                          * by found extent
2099                          */
2100                         start = block;
2101                         end = le32_to_cpu(ex->ee_block)
2102                                 + ext4_ext_get_actual_len(ex);
2103                         if (block + num < end)
2104                                 end = block + num;
2105                         exists = 1;
2106                 } else {
2107                         BUG();
2108                 }
2109                 BUG_ON(end <= start);
2110
2111                 if (!exists) {
2112                         es.es_lblk = start;
2113                         es.es_len = end - start;
2114                         es.es_pblk = 0;
2115                 } else {
2116                         es.es_lblk = le32_to_cpu(ex->ee_block);
2117                         es.es_len = ext4_ext_get_actual_len(ex);
2118                         es.es_pblk = ext4_ext_pblock(ex);
2119                         if (ext4_ext_is_uninitialized(ex))
2120                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2121                 }
2122
2123                 /*
2124                  * Find delayed extent and update es accordingly. We call
2125                  * it even in !exists case to find out whether es is the
2126                  * last existing extent or not.
2127                  */
2128                 next_del = ext4_find_delayed_extent(inode, &es);
2129                 if (!exists && next_del) {
2130                         exists = 1;
2131                         flags |= FIEMAP_EXTENT_DELALLOC;
2132                 }
2133                 up_read(&EXT4_I(inode)->i_data_sem);
2134
2135                 if (unlikely(es.es_len == 0)) {
2136                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2137                         err = -EIO;
2138                         break;
2139                 }
2140
2141                 /*
2142                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2143                  * we need to check next == EXT_MAX_BLOCKS because it is
2144                  * possible that an extent is with unwritten and delayed
2145                  * status due to when an extent is delayed allocated and
2146                  * is allocated by fallocate status tree will track both of
2147                  * them in a extent.
2148                  *
2149                  * So we could return a unwritten and delayed extent, and
2150                  * its block is equal to 'next'.
2151                  */
2152                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2153                         flags |= FIEMAP_EXTENT_LAST;
2154                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2155                                      next != EXT_MAX_BLOCKS)) {
2156                                 EXT4_ERROR_INODE(inode,
2157                                                  "next extent == %u, next "
2158                                                  "delalloc extent = %u",
2159                                                  next, next_del);
2160                                 err = -EIO;
2161                                 break;
2162                         }
2163                 }
2164
2165                 if (exists) {
2166                         err = fiemap_fill_next_extent(fieinfo,
2167                                 (__u64)es.es_lblk << blksize_bits,
2168                                 (__u64)es.es_pblk << blksize_bits,
2169                                 (__u64)es.es_len << blksize_bits,
2170                                 flags);
2171                         if (err < 0)
2172                                 break;
2173                         if (err == 1) {
2174                                 err = 0;
2175                                 break;
2176                         }
2177                 }
2178
2179                 block = es.es_lblk + es.es_len;
2180         }
2181
2182         if (path) {
2183                 ext4_ext_drop_refs(path);
2184                 kfree(path);
2185         }
2186
2187         return err;
2188 }
2189
2190 /*
2191  * ext4_ext_put_gap_in_cache:
2192  * calculate boundaries of the gap that the requested block fits into
2193  * and cache this gap
2194  */
2195 static void
2196 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2197                                 ext4_lblk_t block)
2198 {
2199         int depth = ext_depth(inode);
2200         unsigned long len;
2201         ext4_lblk_t lblock;
2202         struct ext4_extent *ex;
2203
2204         ex = path[depth].p_ext;
2205         if (ex == NULL) {
2206                 /*
2207                  * there is no extent yet, so gap is [0;-] and we
2208                  * don't cache it
2209                  */
2210                 ext_debug("cache gap(whole file):");
2211         } else if (block < le32_to_cpu(ex->ee_block)) {
2212                 lblock = block;
2213                 len = le32_to_cpu(ex->ee_block) - block;
2214                 ext_debug("cache gap(before): %u [%u:%u]",
2215                                 block,
2216                                 le32_to_cpu(ex->ee_block),
2217                                  ext4_ext_get_actual_len(ex));
2218                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2219                         ext4_es_insert_extent(inode, lblock, len, ~0,
2220                                               EXTENT_STATUS_HOLE);
2221         } else if (block >= le32_to_cpu(ex->ee_block)
2222                         + ext4_ext_get_actual_len(ex)) {
2223                 ext4_lblk_t next;
2224                 lblock = le32_to_cpu(ex->ee_block)
2225                         + ext4_ext_get_actual_len(ex);
2226
2227                 next = ext4_ext_next_allocated_block(path);
2228                 ext_debug("cache gap(after): [%u:%u] %u",
2229                                 le32_to_cpu(ex->ee_block),
2230                                 ext4_ext_get_actual_len(ex),
2231                                 block);
2232                 BUG_ON(next == lblock);
2233                 len = next - lblock;
2234                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2235                         ext4_es_insert_extent(inode, lblock, len, ~0,
2236                                               EXTENT_STATUS_HOLE);
2237         } else {
2238                 lblock = len = 0;
2239                 BUG();
2240         }
2241
2242         ext_debug(" -> %u:%lu\n", lblock, len);
2243 }
2244
2245 /*
2246  * ext4_ext_rm_idx:
2247  * removes index from the index block.
2248  */
2249 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2250                         struct ext4_ext_path *path, int depth)
2251 {
2252         int err;
2253         ext4_fsblk_t leaf;
2254
2255         /* free index block */
2256         depth--;
2257         path = path + depth;
2258         leaf = ext4_idx_pblock(path->p_idx);
2259         if (unlikely(path->p_hdr->eh_entries == 0)) {
2260                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2261                 return -EIO;
2262         }
2263         err = ext4_ext_get_access(handle, inode, path);
2264         if (err)
2265                 return err;
2266
2267         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2268                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2269                 len *= sizeof(struct ext4_extent_idx);
2270                 memmove(path->p_idx, path->p_idx + 1, len);
2271         }
2272
2273         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2274         err = ext4_ext_dirty(handle, inode, path);
2275         if (err)
2276                 return err;
2277         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2278         trace_ext4_ext_rm_idx(inode, leaf);
2279
2280         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2281                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2282
2283         while (--depth >= 0) {
2284                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2285                         break;
2286                 path--;
2287                 err = ext4_ext_get_access(handle, inode, path);
2288                 if (err)
2289                         break;
2290                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2291                 err = ext4_ext_dirty(handle, inode, path);
2292                 if (err)
2293                         break;
2294         }
2295         return err;
2296 }
2297
2298 /*
2299  * ext4_ext_calc_credits_for_single_extent:
2300  * This routine returns max. credits that needed to insert an extent
2301  * to the extent tree.
2302  * When pass the actual path, the caller should calculate credits
2303  * under i_data_sem.
2304  */
2305 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2306                                                 struct ext4_ext_path *path)
2307 {
2308         if (path) {
2309                 int depth = ext_depth(inode);
2310                 int ret = 0;
2311
2312                 /* probably there is space in leaf? */
2313                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2314                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2315
2316                         /*
2317                          *  There are some space in the leaf tree, no
2318                          *  need to account for leaf block credit
2319                          *
2320                          *  bitmaps and block group descriptor blocks
2321                          *  and other metadata blocks still need to be
2322                          *  accounted.
2323                          */
2324                         /* 1 bitmap, 1 block group descriptor */
2325                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2326                         return ret;
2327                 }
2328         }
2329
2330         return ext4_chunk_trans_blocks(inode, nrblocks);
2331 }
2332
2333 /*
2334  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2335  *
2336  * if nrblocks are fit in a single extent (chunk flag is 1), then
2337  * in the worse case, each tree level index/leaf need to be changed
2338  * if the tree split due to insert a new extent, then the old tree
2339  * index/leaf need to be updated too
2340  *
2341  * If the nrblocks are discontiguous, they could cause
2342  * the whole tree split more than once, but this is really rare.
2343  */
2344 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2345 {
2346         int index;
2347         int depth;
2348
2349         /* If we are converting the inline data, only one is needed here. */
2350         if (ext4_has_inline_data(inode))
2351                 return 1;
2352
2353         depth = ext_depth(inode);
2354
2355         if (chunk)
2356                 index = depth * 2;
2357         else
2358                 index = depth * 3;
2359
2360         return index;
2361 }
2362
2363 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2364                               struct ext4_extent *ex,
2365                               ext4_fsblk_t *partial_cluster,
2366                               ext4_lblk_t from, ext4_lblk_t to)
2367 {
2368         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2369         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2370         ext4_fsblk_t pblk;
2371         int flags = 0;
2372
2373         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2374                 flags |= EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2375         else if (ext4_should_journal_data(inode))
2376                 flags |= EXT4_FREE_BLOCKS_FORGET;
2377
2378         /*
2379          * For bigalloc file systems, we never free a partial cluster
2380          * at the beginning of the extent.  Instead, we make a note
2381          * that we tried freeing the cluster, and check to see if we
2382          * need to free it on a subsequent call to ext4_remove_blocks,
2383          * or at the end of the ext4_truncate() operation.
2384          */
2385         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2386
2387         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2388         /*
2389          * If we have a partial cluster, and it's different from the
2390          * cluster of the last block, we need to explicitly free the
2391          * partial cluster here.
2392          */
2393         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2394         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2395                 ext4_free_blocks(handle, inode, NULL,
2396                                  EXT4_C2B(sbi, *partial_cluster),
2397                                  sbi->s_cluster_ratio, flags);
2398                 *partial_cluster = 0;
2399         }
2400
2401 #ifdef EXTENTS_STATS
2402         {
2403                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2404                 spin_lock(&sbi->s_ext_stats_lock);
2405                 sbi->s_ext_blocks += ee_len;
2406                 sbi->s_ext_extents++;
2407                 if (ee_len < sbi->s_ext_min)
2408                         sbi->s_ext_min = ee_len;
2409                 if (ee_len > sbi->s_ext_max)
2410                         sbi->s_ext_max = ee_len;
2411                 if (ext_depth(inode) > sbi->s_depth_max)
2412                         sbi->s_depth_max = ext_depth(inode);
2413                 spin_unlock(&sbi->s_ext_stats_lock);
2414         }
2415 #endif
2416         if (from >= le32_to_cpu(ex->ee_block)
2417             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2418                 /* tail removal */
2419                 ext4_lblk_t num;
2420
2421                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2422                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2423                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2424                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2425                 /*
2426                  * If the block range to be freed didn't start at the
2427                  * beginning of a cluster, and we removed the entire
2428                  * extent, save the partial cluster here, since we
2429                  * might need to delete if we determine that the
2430                  * truncate operation has removed all of the blocks in
2431                  * the cluster.
2432                  */
2433                 if (pblk & (sbi->s_cluster_ratio - 1) &&
2434                     (ee_len == num))
2435                         *partial_cluster = EXT4_B2C(sbi, pblk);
2436                 else
2437                         *partial_cluster = 0;
2438         } else if (from == le32_to_cpu(ex->ee_block)
2439                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2440                 /* head removal */
2441                 ext4_lblk_t num;
2442                 ext4_fsblk_t start;
2443
2444                 num = to - from;
2445                 start = ext4_ext_pblock(ex);
2446
2447                 ext_debug("free first %u blocks starting %llu\n", num, start);
2448                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2449
2450         } else {
2451                 printk(KERN_INFO "strange request: removal(2) "
2452                                 "%u-%u from %u:%u\n",
2453                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2454         }
2455         return 0;
2456 }
2457
2458
2459 /*
2460  * ext4_ext_rm_leaf() Removes the extents associated with the
2461  * blocks appearing between "start" and "end", and splits the extents
2462  * if "start" and "end" appear in the same extent
2463  *
2464  * @handle: The journal handle
2465  * @inode:  The files inode
2466  * @path:   The path to the leaf
2467  * @start:  The first block to remove
2468  * @end:   The last block to remove
2469  */
2470 static int
2471 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2472                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2473                  ext4_lblk_t start, ext4_lblk_t end)
2474 {
2475         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2476         int err = 0, correct_index = 0;
2477         int depth = ext_depth(inode), credits;
2478         struct ext4_extent_header *eh;
2479         ext4_lblk_t a, b;
2480         unsigned num;
2481         ext4_lblk_t ex_ee_block;
2482         unsigned short ex_ee_len;
2483         unsigned uninitialized = 0;
2484         struct ext4_extent *ex;
2485
2486         /* the header must be checked already in ext4_ext_remove_space() */
2487         ext_debug("truncate since %u in leaf to %u\n", start, end);
2488         if (!path[depth].p_hdr)
2489                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2490         eh = path[depth].p_hdr;
2491         if (unlikely(path[depth].p_hdr == NULL)) {
2492                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2493                 return -EIO;
2494         }
2495         /* find where to start removing */
2496         ex = EXT_LAST_EXTENT(eh);
2497
2498         ex_ee_block = le32_to_cpu(ex->ee_block);
2499         ex_ee_len = ext4_ext_get_actual_len(ex);
2500
2501         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2502
2503         while (ex >= EXT_FIRST_EXTENT(eh) &&
2504                         ex_ee_block + ex_ee_len > start) {
2505
2506                 if (ext4_ext_is_uninitialized(ex))
2507                         uninitialized = 1;
2508                 else
2509                         uninitialized = 0;
2510
2511                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2512                          uninitialized, ex_ee_len);
2513                 path[depth].p_ext = ex;
2514
2515                 a = ex_ee_block > start ? ex_ee_block : start;
2516                 b = ex_ee_block+ex_ee_len - 1 < end ?
2517                         ex_ee_block+ex_ee_len - 1 : end;
2518
2519                 ext_debug("  border %u:%u\n", a, b);
2520
2521                 /* If this extent is beyond the end of the hole, skip it */
2522                 if (end < ex_ee_block) {
2523                         ex--;
2524                         ex_ee_block = le32_to_cpu(ex->ee_block);
2525                         ex_ee_len = ext4_ext_get_actual_len(ex);
2526                         continue;
2527                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2528                         EXT4_ERROR_INODE(inode,
2529                                          "can not handle truncate %u:%u "
2530                                          "on extent %u:%u",
2531                                          start, end, ex_ee_block,
2532                                          ex_ee_block + ex_ee_len - 1);
2533                         err = -EIO;
2534                         goto out;
2535                 } else if (a != ex_ee_block) {
2536                         /* remove tail of the extent */
2537                         num = a - ex_ee_block;
2538                 } else {
2539                         /* remove whole extent: excellent! */
2540                         num = 0;
2541                 }
2542                 /*
2543                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2544                  * descriptor) for each block group; assume two block
2545                  * groups plus ex_ee_len/blocks_per_block_group for
2546                  * the worst case
2547                  */
2548                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2549                 if (ex == EXT_FIRST_EXTENT(eh)) {
2550                         correct_index = 1;
2551                         credits += (ext_depth(inode)) + 1;
2552                 }
2553                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2554
2555                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2556                 if (err)
2557                         goto out;
2558
2559                 err = ext4_ext_get_access(handle, inode, path + depth);
2560                 if (err)
2561                         goto out;
2562
2563                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2564                                          a, b);
2565                 if (err)
2566                         goto out;
2567
2568                 if (num == 0)
2569                         /* this extent is removed; mark slot entirely unused */
2570                         ext4_ext_store_pblock(ex, 0);
2571
2572                 ex->ee_len = cpu_to_le16(num);
2573                 /*
2574                  * Do not mark uninitialized if all the blocks in the
2575                  * extent have been removed.
2576                  */
2577                 if (uninitialized && num)
2578                         ext4_ext_mark_uninitialized(ex);
2579                 /*
2580                  * If the extent was completely released,
2581                  * we need to remove it from the leaf
2582                  */
2583                 if (num == 0) {
2584                         if (end != EXT_MAX_BLOCKS - 1) {
2585                                 /*
2586                                  * For hole punching, we need to scoot all the
2587                                  * extents up when an extent is removed so that
2588                                  * we dont have blank extents in the middle
2589                                  */
2590                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2591                                         sizeof(struct ext4_extent));
2592
2593                                 /* Now get rid of the one at the end */
2594                                 memset(EXT_LAST_EXTENT(eh), 0,
2595                                         sizeof(struct ext4_extent));
2596                         }
2597                         le16_add_cpu(&eh->eh_entries, -1);
2598                 } else
2599                         *partial_cluster = 0;
2600
2601                 err = ext4_ext_dirty(handle, inode, path + depth);
2602                 if (err)
2603                         goto out;
2604
2605                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2606                                 ext4_ext_pblock(ex));
2607                 ex--;
2608                 ex_ee_block = le32_to_cpu(ex->ee_block);
2609                 ex_ee_len = ext4_ext_get_actual_len(ex);
2610         }
2611
2612         if (correct_index && eh->eh_entries)
2613                 err = ext4_ext_correct_indexes(handle, inode, path);
2614
2615         /*
2616          * If there is still a entry in the leaf node, check to see if
2617          * it references the partial cluster.  This is the only place
2618          * where it could; if it doesn't, we can free the cluster.
2619          */
2620         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2621             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2622              *partial_cluster)) {
2623                 int flags = EXT4_FREE_BLOCKS_FORGET;
2624
2625                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2626                         flags |= EXT4_FREE_BLOCKS_METADATA;
2627
2628                 ext4_free_blocks(handle, inode, NULL,
2629                                  EXT4_C2B(sbi, *partial_cluster),
2630                                  sbi->s_cluster_ratio, flags);
2631                 *partial_cluster = 0;
2632         }
2633
2634         /* if this leaf is free, then we should
2635          * remove it from index block above */
2636         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2637                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2638
2639 out:
2640         return err;
2641 }
2642
2643 /*
2644  * ext4_ext_more_to_rm:
2645  * returns 1 if current index has to be freed (even partial)
2646  */
2647 static int
2648 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2649 {
2650         BUG_ON(path->p_idx == NULL);
2651
2652         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2653                 return 0;
2654
2655         /*
2656          * if truncate on deeper level happened, it wasn't partial,
2657          * so we have to consider current index for truncation
2658          */
2659         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2660                 return 0;
2661         return 1;
2662 }
2663
2664 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2665                           ext4_lblk_t end)
2666 {
2667         struct super_block *sb = inode->i_sb;
2668         int depth = ext_depth(inode);
2669         struct ext4_ext_path *path = NULL;
2670         ext4_fsblk_t partial_cluster = 0;
2671         handle_t *handle;
2672         int i = 0, err = 0;
2673
2674         ext_debug("truncate since %u to %u\n", start, end);
2675
2676         /* probably first extent we're gonna free will be last in block */
2677         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2678         if (IS_ERR(handle))
2679                 return PTR_ERR(handle);
2680
2681 again:
2682         trace_ext4_ext_remove_space(inode, start, depth);
2683
2684         /*
2685          * Check if we are removing extents inside the extent tree. If that
2686          * is the case, we are going to punch a hole inside the extent tree
2687          * so we have to check whether we need to split the extent covering
2688          * the last block to remove so we can easily remove the part of it
2689          * in ext4_ext_rm_leaf().
2690          */
2691         if (end < EXT_MAX_BLOCKS - 1) {
2692                 struct ext4_extent *ex;
2693                 ext4_lblk_t ee_block;
2694
2695                 /* find extent for this block */
2696                 path = ext4_ext_find_extent(inode, end, NULL);
2697                 if (IS_ERR(path)) {
2698                         ext4_journal_stop(handle);
2699                         return PTR_ERR(path);
2700                 }
2701                 depth = ext_depth(inode);
2702                 /* Leaf not may not exist only if inode has no blocks at all */
2703                 ex = path[depth].p_ext;
2704                 if (!ex) {
2705                         if (depth) {
2706                                 EXT4_ERROR_INODE(inode,
2707                                                  "path[%d].p_hdr == NULL",
2708                                                  depth);
2709                                 err = -EIO;
2710                         }
2711                         goto out;
2712                 }
2713
2714                 ee_block = le32_to_cpu(ex->ee_block);
2715
2716                 /*
2717                  * See if the last block is inside the extent, if so split
2718                  * the extent at 'end' block so we can easily remove the
2719                  * tail of the first part of the split extent in
2720                  * ext4_ext_rm_leaf().
2721                  */
2722                 if (end >= ee_block &&
2723                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2724                         int split_flag = 0;
2725
2726                         if (ext4_ext_is_uninitialized(ex))
2727                                 split_flag = EXT4_EXT_MARK_UNINIT1 |
2728                                              EXT4_EXT_MARK_UNINIT2;
2729
2730                         /*
2731                          * Split the extent in two so that 'end' is the last
2732                          * block in the first new extent
2733                          */
2734                         err = ext4_split_extent_at(handle, inode, path,
2735                                                 end + 1, split_flag,
2736                                                 EXT4_GET_BLOCKS_PRE_IO |
2737                                                 EXT4_GET_BLOCKS_PUNCH_OUT_EXT);
2738
2739                         if (err < 0)
2740                                 goto out;
2741                 }
2742         }
2743         /*
2744          * We start scanning from right side, freeing all the blocks
2745          * after i_size and walking into the tree depth-wise.
2746          */
2747         depth = ext_depth(inode);
2748         if (path) {
2749                 int k = i = depth;
2750                 while (--k > 0)
2751                         path[k].p_block =
2752                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2753         } else {
2754                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2755                                GFP_NOFS);
2756                 if (path == NULL) {
2757                         ext4_journal_stop(handle);
2758                         return -ENOMEM;
2759                 }
2760                 path[0].p_depth = depth;
2761                 path[0].p_hdr = ext_inode_hdr(inode);
2762                 i = 0;
2763
2764                 if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2765                         err = -EIO;
2766                         goto out;
2767                 }
2768         }
2769         err = 0;
2770
2771         while (i >= 0 && err == 0) {
2772                 if (i == depth) {
2773                         /* this is leaf block */
2774                         err = ext4_ext_rm_leaf(handle, inode, path,
2775                                                &partial_cluster, start,
2776                                                end);
2777                         /* root level has p_bh == NULL, brelse() eats this */
2778                         brelse(path[i].p_bh);
2779                         path[i].p_bh = NULL;
2780                         i--;
2781                         continue;
2782                 }
2783
2784                 /* this is index block */
2785                 if (!path[i].p_hdr) {
2786                         ext_debug("initialize header\n");
2787                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2788                 }
2789
2790                 if (!path[i].p_idx) {
2791                         /* this level hasn't been touched yet */
2792                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2793                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2794                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2795                                   path[i].p_hdr,
2796                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2797                 } else {
2798                         /* we were already here, see at next index */
2799                         path[i].p_idx--;
2800                 }
2801
2802                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2803                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2804                                 path[i].p_idx);
2805                 if (ext4_ext_more_to_rm(path + i)) {
2806                         struct buffer_head *bh;
2807                         /* go to the next level */
2808                         ext_debug("move to level %d (block %llu)\n",
2809                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2810                         memset(path + i + 1, 0, sizeof(*path));
2811                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2812                         if (!bh) {
2813                                 /* should we reset i_size? */
2814                                 err = -EIO;
2815                                 break;
2816                         }
2817                         if (WARN_ON(i + 1 > depth)) {
2818                                 err = -EIO;
2819                                 break;
2820                         }
2821                         if (ext4_ext_check_block(inode, ext_block_hdr(bh),
2822                                                         depth - i - 1, bh)) {
2823                                 err = -EIO;
2824                                 break;
2825                         }
2826                         path[i + 1].p_bh = bh;
2827
2828                         /* save actual number of indexes since this
2829                          * number is changed at the next iteration */
2830                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2831                         i++;
2832                 } else {
2833                         /* we finished processing this index, go up */
2834                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2835                                 /* index is empty, remove it;
2836                                  * handle must be already prepared by the
2837                                  * truncatei_leaf() */
2838                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2839                         }
2840                         /* root level has p_bh == NULL, brelse() eats this */
2841                         brelse(path[i].p_bh);
2842                         path[i].p_bh = NULL;
2843                         i--;
2844                         ext_debug("return to level %d\n", i);
2845                 }
2846         }
2847
2848         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2849                         path->p_hdr->eh_entries);
2850
2851         /* If we still have something in the partial cluster and we have removed
2852          * even the first extent, then we should free the blocks in the partial
2853          * cluster as well. */
2854         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2855                 int flags = EXT4_FREE_BLOCKS_FORGET;
2856
2857                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2858                         flags |= EXT4_FREE_BLOCKS_METADATA;
2859
2860                 ext4_free_blocks(handle, inode, NULL,
2861                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2862                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2863                 partial_cluster = 0;
2864         }
2865
2866         /* TODO: flexible tree reduction should be here */
2867         if (path->p_hdr->eh_entries == 0) {
2868                 /*
2869                  * truncate to zero freed all the tree,
2870                  * so we need to correct eh_depth
2871                  */
2872                 err = ext4_ext_get_access(handle, inode, path);
2873                 if (err == 0) {
2874                         ext_inode_hdr(inode)->eh_depth = 0;
2875                         ext_inode_hdr(inode)->eh_max =
2876                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2877                         err = ext4_ext_dirty(handle, inode, path);
2878                 }
2879         }
2880 out:
2881         ext4_ext_drop_refs(path);
2882         kfree(path);
2883         if (err == -EAGAIN) {
2884                 path = NULL;
2885                 goto again;
2886         }
2887         ext4_journal_stop(handle);
2888
2889         return err;
2890 }
2891
2892 /*
2893  * called at mount time
2894  */
2895 void ext4_ext_init(struct super_block *sb)
2896 {
2897         /*
2898          * possible initialization would be here
2899          */
2900
2901         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2902 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2903                 printk(KERN_INFO "EXT4-fs: file extents enabled"
2904 #ifdef AGGRESSIVE_TEST
2905                        ", aggressive tests"
2906 #endif
2907 #ifdef CHECK_BINSEARCH
2908                        ", check binsearch"
2909 #endif
2910 #ifdef EXTENTS_STATS
2911                        ", stats"
2912 #endif
2913                        "\n");
2914 #endif
2915 #ifdef EXTENTS_STATS
2916                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2917                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2918                 EXT4_SB(sb)->s_ext_max = 0;
2919 #endif
2920         }
2921 }
2922
2923 /*
2924  * called at umount time
2925  */
2926 void ext4_ext_release(struct super_block *sb)
2927 {
2928         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2929                 return;
2930
2931 #ifdef EXTENTS_STATS
2932         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2933                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2934                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2935                         sbi->s_ext_blocks, sbi->s_ext_extents,
2936                         sbi->s_ext_blocks / sbi->s_ext_extents);
2937                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2938                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2939         }
2940 #endif
2941 }
2942
2943 /* FIXME!! we need to try to merge to left or right after zero-out  */
2944 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2945 {
2946         ext4_fsblk_t ee_pblock;
2947         unsigned int ee_len;
2948         int ret;
2949
2950         ee_len    = ext4_ext_get_actual_len(ex);
2951         ee_pblock = ext4_ext_pblock(ex);
2952
2953         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2954         if (ret > 0)
2955                 ret = 0;
2956
2957         return ret;
2958 }
2959
2960 /*
2961  * ext4_split_extent_at() splits an extent at given block.
2962  *
2963  * @handle: the journal handle
2964  * @inode: the file inode
2965  * @path: the path to the extent
2966  * @split: the logical block where the extent is splitted.
2967  * @split_flags: indicates if the extent could be zeroout if split fails, and
2968  *               the states(init or uninit) of new extents.
2969  * @flags: flags used to insert new extent to extent tree.
2970  *
2971  *
2972  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2973  * of which are deterimined by split_flag.
2974  *
2975  * There are two cases:
2976  *  a> the extent are splitted into two extent.
2977  *  b> split is not needed, and just mark the extent.
2978  *
2979  * return 0 on success.
2980  */
2981 static int ext4_split_extent_at(handle_t *handle,
2982                              struct inode *inode,
2983                              struct ext4_ext_path *path,
2984                              ext4_lblk_t split,
2985                              int split_flag,
2986                              int flags)
2987 {
2988         ext4_fsblk_t newblock;
2989         ext4_lblk_t ee_block;
2990         struct ext4_extent *ex, newex, orig_ex, zero_ex;
2991         struct ext4_extent *ex2 = NULL;
2992         unsigned int ee_len, depth;
2993         int err = 0;
2994
2995         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
2996                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
2997
2998         ext_debug("ext4_split_extents_at: inode %lu, logical"
2999                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3000
3001         ext4_ext_show_leaf(inode, path);
3002
3003         depth = ext_depth(inode);
3004         ex = path[depth].p_ext;
3005         ee_block = le32_to_cpu(ex->ee_block);
3006         ee_len = ext4_ext_get_actual_len(ex);
3007         newblock = split - ee_block + ext4_ext_pblock(ex);
3008
3009         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3010         BUG_ON(!ext4_ext_is_uninitialized(ex) &&
3011                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3012                              EXT4_EXT_MARK_UNINIT1 |
3013                              EXT4_EXT_MARK_UNINIT2));
3014
3015         err = ext4_ext_get_access(handle, inode, path + depth);
3016         if (err)
3017                 goto out;
3018
3019         if (split == ee_block) {
3020                 /*
3021                  * case b: block @split is the block that the extent begins with
3022                  * then we just change the state of the extent, and splitting
3023                  * is not needed.
3024                  */
3025                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
3026                         ext4_ext_mark_uninitialized(ex);
3027                 else
3028                         ext4_ext_mark_initialized(ex);
3029
3030                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3031                         ext4_ext_try_to_merge(handle, inode, path, ex);
3032
3033                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3034                 goto out;
3035         }
3036
3037         /* case a */
3038         memcpy(&orig_ex, ex, sizeof(orig_ex));
3039         ex->ee_len = cpu_to_le16(split - ee_block);
3040         if (split_flag & EXT4_EXT_MARK_UNINIT1)
3041                 ext4_ext_mark_uninitialized(ex);
3042
3043         /*
3044          * path may lead to new leaf, not to original leaf any more
3045          * after ext4_ext_insert_extent() returns,
3046          */
3047         err = ext4_ext_dirty(handle, inode, path + depth);
3048         if (err)
3049                 goto fix_extent_len;
3050
3051         ex2 = &newex;
3052         ex2->ee_block = cpu_to_le32(split);
3053         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3054         ext4_ext_store_pblock(ex2, newblock);
3055         if (split_flag & EXT4_EXT_MARK_UNINIT2)
3056                 ext4_ext_mark_uninitialized(ex2);
3057
3058         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
3059         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3060                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3061                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3062                                 err = ext4_ext_zeroout(inode, ex2);
3063                                 zero_ex.ee_block = ex2->ee_block;
3064                                 zero_ex.ee_len = cpu_to_le16(
3065                                                 ext4_ext_get_actual_len(ex2));
3066                                 ext4_ext_store_pblock(&zero_ex,
3067                                                       ext4_ext_pblock(ex2));
3068                         } else {
3069                                 err = ext4_ext_zeroout(inode, ex);
3070                                 zero_ex.ee_block = ex->ee_block;
3071                                 zero_ex.ee_len = cpu_to_le16(
3072                                                 ext4_ext_get_actual_len(ex));
3073                                 ext4_ext_store_pblock(&zero_ex,
3074                                                       ext4_ext_pblock(ex));
3075                         }
3076                 } else {
3077                         err = ext4_ext_zeroout(inode, &orig_ex);
3078                         zero_ex.ee_block = orig_ex.ee_block;
3079                         zero_ex.ee_len = cpu_to_le16(
3080                                                 ext4_ext_get_actual_len(&orig_ex));
3081                         ext4_ext_store_pblock(&zero_ex,
3082                                               ext4_ext_pblock(&orig_ex));
3083                 }
3084
3085                 if (err)
3086                         goto fix_extent_len;
3087                 /* update the extent length and mark as initialized */
3088                 ex->ee_len = cpu_to_le16(ee_len);
3089                 ext4_ext_try_to_merge(handle, inode, path, ex);
3090                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3091                 if (err)
3092                         goto fix_extent_len;
3093
3094                 /* update extent status tree */
3095                 err = ext4_es_zeroout(inode, &zero_ex);
3096
3097                 goto out;
3098         } else if (err)
3099                 goto fix_extent_len;
3100
3101 out:
3102         ext4_ext_show_leaf(inode, path);
3103         return err;
3104
3105 fix_extent_len:
3106         ex->ee_len = orig_ex.ee_len;
3107         ext4_ext_dirty(handle, inode, path + depth);
3108         return err;
3109 }
3110
3111 /*
3112  * ext4_split_extents() splits an extent and mark extent which is covered
3113  * by @map as split_flags indicates
3114  *
3115  * It may result in splitting the extent into multiple extents (upto three)
3116  * There are three possibilities:
3117  *   a> There is no split required
3118  *   b> Splits in two extents: Split is happening at either end of the extent
3119  *   c> Splits in three extents: Somone is splitting in middle of the extent
3120  *
3121  */
3122 static int ext4_split_extent(handle_t *handle,
3123                               struct inode *inode,
3124                               struct ext4_ext_path *path,
3125                               struct ext4_map_blocks *map,
3126                               int split_flag,
3127                               int flags)
3128 {
3129         ext4_lblk_t ee_block;
3130         struct ext4_extent *ex;
3131         unsigned int ee_len, depth;
3132         int err = 0;
3133         int uninitialized;
3134         int split_flag1, flags1;
3135         int allocated = map->m_len;
3136
3137         depth = ext_depth(inode);
3138         ex = path[depth].p_ext;
3139         ee_block = le32_to_cpu(ex->ee_block);
3140         ee_len = ext4_ext_get_actual_len(ex);
3141         uninitialized = ext4_ext_is_uninitialized(ex);
3142
3143         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3144                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3145                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3146                 if (uninitialized)
3147                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
3148                                        EXT4_EXT_MARK_UNINIT2;
3149                 if (split_flag & EXT4_EXT_DATA_VALID2)
3150                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3151                 err = ext4_split_extent_at(handle, inode, path,
3152                                 map->m_lblk + map->m_len, split_flag1, flags1);
3153                 if (err)
3154                         goto out;
3155         } else {
3156                 allocated = ee_len - (map->m_lblk - ee_block);
3157         }
3158         /*
3159          * Update path is required because previous ext4_split_extent_at() may
3160          * result in split of original leaf or extent zeroout.
3161          */
3162         ext4_ext_drop_refs(path);
3163         path = ext4_ext_find_extent(inode, map->m_lblk, path);
3164         if (IS_ERR(path))
3165                 return PTR_ERR(path);
3166         depth = ext_depth(inode);
3167         ex = path[depth].p_ext;
3168         uninitialized = ext4_ext_is_uninitialized(ex);
3169         split_flag1 = 0;
3170
3171         if (map->m_lblk >= ee_block) {
3172                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3173                 if (uninitialized) {
3174                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
3175                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3176                                                      EXT4_EXT_MARK_UNINIT2);
3177                 }
3178                 err = ext4_split_extent_at(handle, inode, path,
3179                                 map->m_lblk, split_flag1, flags);
3180                 if (err)
3181                         goto out;
3182         }
3183
3184         ext4_ext_show_leaf(inode, path);
3185 out:
3186         return err ? err : allocated;
3187 }
3188
3189 /*
3190  * This function is called by ext4_ext_map_blocks() if someone tries to write
3191  * to an uninitialized extent. It may result in splitting the uninitialized
3192  * extent into multiple extents (up to three - one initialized and two
3193  * uninitialized).
3194  * There are three possibilities:
3195  *   a> There is no split required: Entire extent should be initialized
3196  *   b> Splits in two extents: Write is happening at either end of the extent
3197  *   c> Splits in three extents: Somone is writing in middle of the extent
3198  *
3199  * Pre-conditions:
3200  *  - The extent pointed to by 'path' is uninitialized.
3201  *  - The extent pointed to by 'path' contains a superset
3202  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3203  *
3204  * Post-conditions on success:
3205  *  - the returned value is the number of blocks beyond map->l_lblk
3206  *    that are allocated and initialized.
3207  *    It is guaranteed to be >= map->m_len.
3208  */
3209 static int ext4_ext_convert_to_initialized(handle_t *handle,
3210                                            struct inode *inode,
3211                                            struct ext4_map_blocks *map,
3212                                            struct ext4_ext_path *path)
3213 {
3214         struct ext4_sb_info *sbi;
3215         struct ext4_extent_header *eh;
3216         struct ext4_map_blocks split_map;
3217         struct ext4_extent zero_ex;
3218         struct ext4_extent *ex, *abut_ex;
3219         ext4_lblk_t ee_block, eof_block;
3220         unsigned int ee_len, depth, map_len = map->m_len;
3221         int allocated = 0, max_zeroout = 0;
3222         int err = 0;
3223         int split_flag = 0;
3224
3225         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3226                 "block %llu, max_blocks %u\n", inode->i_ino,
3227                 (unsigned long long)map->m_lblk, map_len);
3228
3229         sbi = EXT4_SB(inode->i_sb);
3230         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3231                 inode->i_sb->s_blocksize_bits;
3232         if (eof_block < map->m_lblk + map_len)
3233                 eof_block = map->m_lblk + map_len;
3234
3235         depth = ext_depth(inode);
3236         eh = path[depth].p_hdr;
3237         ex = path[depth].p_ext;
3238         ee_block = le32_to_cpu(ex->ee_block);
3239         ee_len = ext4_ext_get_actual_len(ex);
3240         zero_ex.ee_len = 0;
3241
3242         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3243
3244         /* Pre-conditions */
3245         BUG_ON(!ext4_ext_is_uninitialized(ex));
3246         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3247
3248         /*
3249          * Attempt to transfer newly initialized blocks from the currently
3250          * uninitialized extent to its neighbor. This is much cheaper
3251          * than an insertion followed by a merge as those involve costly
3252          * memmove() calls. Transferring to the left is the common case in
3253          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3254          * followed by append writes.
3255          *
3256          * Limitations of the current logic:
3257          *  - L1: we do not deal with writes covering the whole extent.
3258          *    This would require removing the extent if the transfer
3259          *    is possible.
3260          *  - L2: we only attempt to merge with an extent stored in the
3261          *    same extent tree node.
3262          */
3263         if ((map->m_lblk == ee_block) &&
3264                 /* See if we can merge left */
3265                 (map_len < ee_len) &&           /*L1*/
3266                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3267                 ext4_lblk_t prev_lblk;
3268                 ext4_fsblk_t prev_pblk, ee_pblk;
3269                 unsigned int prev_len;
3270
3271                 abut_ex = ex - 1;
3272                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3273                 prev_len = ext4_ext_get_actual_len(abut_ex);
3274                 prev_pblk = ext4_ext_pblock(abut_ex);
3275                 ee_pblk = ext4_ext_pblock(ex);
3276
3277                 /*
3278                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3279                  * upon those conditions:
3280                  * - C1: abut_ex is initialized,
3281                  * - C2: abut_ex is logically abutting ex,
3282                  * - C3: abut_ex is physically abutting ex,
3283                  * - C4: abut_ex can receive the additional blocks without
3284                  *   overflowing the (initialized) length limit.
3285                  */
3286                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3287                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3288                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3289                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3290                         err = ext4_ext_get_access(handle, inode, path + depth);
3291                         if (err)
3292                                 goto out;
3293
3294                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3295                                 map, ex, abut_ex);
3296
3297                         /* Shift the start of ex by 'map_len' blocks */
3298                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3299                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3300                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3301                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3302
3303                         /* Extend abut_ex by 'map_len' blocks */
3304                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3305
3306                         /* Result: number of initialized blocks past m_lblk */
3307                         allocated = map_len;
3308                 }
3309         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3310                    (map_len < ee_len) &&        /*L1*/
3311                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3312                 /* See if we can merge right */
3313                 ext4_lblk_t next_lblk;
3314                 ext4_fsblk_t next_pblk, ee_pblk;
3315                 unsigned int next_len;
3316
3317                 abut_ex = ex + 1;
3318                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3319                 next_len = ext4_ext_get_actual_len(abut_ex);
3320                 next_pblk = ext4_ext_pblock(abut_ex);
3321                 ee_pblk = ext4_ext_pblock(ex);
3322
3323                 /*
3324                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3325                  * upon those conditions:
3326                  * - C1: abut_ex is initialized,
3327                  * - C2: abut_ex is logically abutting ex,
3328                  * - C3: abut_ex is physically abutting ex,
3329                  * - C4: abut_ex can receive the additional blocks without
3330                  *   overflowing the (initialized) length limit.
3331                  */
3332                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3333                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3334                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3335                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3336                         err = ext4_ext_get_access(handle, inode, path + depth);
3337                         if (err)
3338                                 goto out;
3339
3340                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3341                                 map, ex, abut_ex);
3342
3343                         /* Shift the start of abut_ex by 'map_len' blocks */
3344                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3345                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3346                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3347                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3348
3349                         /* Extend abut_ex by 'map_len' blocks */
3350                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3351
3352                         /* Result: number of initialized blocks past m_lblk */
3353                         allocated = map_len;
3354                 }
3355         }
3356         if (allocated) {
3357                 /* Mark the block containing both extents as dirty */
3358                 ext4_ext_dirty(handle, inode, path + depth);
3359
3360                 /* Update path to point to the right extent */
3361                 path[depth].p_ext = abut_ex;
3362                 goto out;
3363         } else
3364                 allocated = ee_len - (map->m_lblk - ee_block);
3365
3366         WARN_ON(map->m_lblk < ee_block);
3367         /*
3368          * It is safe to convert extent to initialized via explicit
3369          * zeroout only if extent is fully insde i_size or new_size.
3370          */
3371         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3372
3373         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3374                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3375                         (inode->i_sb->s_blocksize_bits - 10);
3376
3377         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3378         if (max_zeroout && (ee_len <= max_zeroout)) {
3379                 err = ext4_ext_zeroout(inode, ex);
3380                 if (err)
3381                         goto out;
3382                 zero_ex.ee_block = ex->ee_block;
3383                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3384                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3385
3386                 err = ext4_ext_get_access(handle, inode, path + depth);
3387                 if (err)
3388                         goto out;
3389                 ext4_ext_mark_initialized(ex);
3390                 ext4_ext_try_to_merge(handle, inode, path, ex);
3391                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3392                 goto out;
3393         }
3394
3395         /*
3396          * four cases:
3397          * 1. split the extent into three extents.
3398          * 2. split the extent into two extents, zeroout the first half.
3399          * 3. split the extent into two extents, zeroout the second half.
3400          * 4. split the extent into two extents with out zeroout.
3401          */
3402         split_map.m_lblk = map->m_lblk;
3403         split_map.m_len = map->m_len;
3404
3405         if (max_zeroout && (allocated > map->m_len)) {
3406                 if (allocated <= max_zeroout) {
3407                         /* case 3 */
3408                         zero_ex.ee_block =
3409                                          cpu_to_le32(map->m_lblk);
3410                         zero_ex.ee_len = cpu_to_le16(allocated);
3411                         ext4_ext_store_pblock(&zero_ex,
3412                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3413                         err = ext4_ext_zeroout(inode, &zero_ex);
3414                         if (err)
3415                                 goto out;
3416                         split_map.m_lblk = map->m_lblk;
3417                         split_map.m_len = allocated;
3418                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3419                         /* case 2 */
3420                         if (map->m_lblk != ee_block) {
3421                                 zero_ex.ee_block = ex->ee_block;
3422                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3423                                                         ee_block);
3424                                 ext4_ext_store_pblock(&zero_ex,
3425                                                       ext4_ext_pblock(ex));
3426                                 err = ext4_ext_zeroout(inode, &zero_ex);
3427                                 if (err)
3428                                         goto out;
3429                         }
3430
3431                         split_map.m_lblk = ee_block;
3432                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3433                         allocated = map->m_len;
3434                 }
3435         }
3436
3437         allocated = ext4_split_extent(handle, inode, path,
3438                                       &split_map, split_flag, 0);
3439         if (allocated < 0)
3440                 err = allocated;
3441
3442 out:
3443         /* If we have gotten a failure, don't zero out status tree */
3444         if (!err)
3445                 err = ext4_es_zeroout(inode, &zero_ex);
3446         return err ? err : allocated;
3447 }
3448
3449 /*
3450  * This function is called by ext4_ext_map_blocks() from
3451  * ext4_get_blocks_dio_write() when DIO to write
3452  * to an uninitialized extent.
3453  *
3454  * Writing to an uninitialized extent may result in splitting the uninitialized
3455  * extent into multiple initialized/uninitialized extents (up to three)
3456  * There are three possibilities:
3457  *   a> There is no split required: Entire extent should be uninitialized
3458  *   b> Splits in two extents: Write is happening at either end of the extent
3459  *   c> Splits in three extents: Somone is writing in middle of the extent
3460  *
3461  * One of more index blocks maybe needed if the extent tree grow after
3462  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3463  * complete, we need to split the uninitialized extent before DIO submit
3464  * the IO. The uninitialized extent called at this time will be split
3465  * into three uninitialized extent(at most). After IO complete, the part
3466  * being filled will be convert to initialized by the end_io callback function
3467  * via ext4_convert_unwritten_extents().
3468  *
3469  * Returns the size of uninitialized extent to be written on success.
3470  */
3471 static int ext4_split_unwritten_extents(handle_t *handle,
3472                                         struct inode *inode,
3473                                         struct ext4_map_blocks *map,
3474                                         struct ext4_ext_path *path,
3475                                         int flags)
3476 {
3477         ext4_lblk_t eof_block;
3478         ext4_lblk_t ee_block;
3479         struct ext4_extent *ex;
3480         unsigned int ee_len;
3481         int split_flag = 0, depth;
3482
3483         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3484                 "block %llu, max_blocks %u\n", inode->i_ino,
3485                 (unsigned long long)map->m_lblk, map->m_len);
3486
3487         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3488                 inode->i_sb->s_blocksize_bits;
3489         if (eof_block < map->m_lblk + map->m_len)
3490                 eof_block = map->m_lblk + map->m_len;
3491         /*
3492          * It is safe to convert extent to initialized via explicit
3493          * zeroout only if extent is fully insde i_size or new_size.
3494          */
3495         depth = ext_depth(inode);
3496         ex = path[depth].p_ext;
3497         ee_block = le32_to_cpu(ex->ee_block);
3498         ee_len = ext4_ext_get_actual_len(ex);
3499
3500         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3501         split_flag |= EXT4_EXT_MARK_UNINIT2;
3502         if (flags & EXT4_GET_BLOCKS_CONVERT)
3503                 split_flag |= EXT4_EXT_DATA_VALID2;
3504         flags |= EXT4_GET_BLOCKS_PRE_IO;
3505         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3506 }
3507
3508 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3509                                                 struct inode *inode,
3510                                                 struct ext4_map_blocks *map,
3511                                                 struct ext4_ext_path *path)
3512 {
3513         struct ext4_extent *ex;
3514         ext4_lblk_t ee_block;
3515         unsigned int ee_len;
3516         int depth;
3517         int err = 0;
3518
3519         depth = ext_depth(inode);
3520         ex = path[depth].p_ext;
3521         ee_block = le32_to_cpu(ex->ee_block);
3522         ee_len = ext4_ext_get_actual_len(ex);
3523
3524         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3525                 "block %llu, max_blocks %u\n", inode->i_ino,
3526                   (unsigned long long)ee_block, ee_len);
3527
3528         /* If extent is larger than requested it is a clear sign that we still
3529          * have some extent state machine issues left. So extent_split is still
3530          * required.
3531          * TODO: Once all related issues will be fixed this situation should be
3532          * illegal.
3533          */
3534         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3535 #ifdef EXT4_DEBUG
3536                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3537                              " len %u; IO logical block %llu, len %u\n",
3538                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3539                              (unsigned long long)map->m_lblk, map->m_len);
3540 #endif
3541                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3542                                                    EXT4_GET_BLOCKS_CONVERT);
3543                 if (err < 0)
3544                         goto out;
3545                 ext4_ext_drop_refs(path);
3546                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3547                 if (IS_ERR(path)) {
3548                         err = PTR_ERR(path);
3549                         goto out;
3550                 }
3551                 depth = ext_depth(inode);
3552                 ex = path[depth].p_ext;
3553         }
3554
3555         err = ext4_ext_get_access(handle, inode, path + depth);
3556         if (err)
3557                 goto out;
3558         /* first mark the extent as initialized */
3559         ext4_ext_mark_initialized(ex);
3560
3561         /* note: ext4_ext_correct_indexes() isn't needed here because
3562          * borders are not changed
3563          */
3564         ext4_ext_try_to_merge(handle, inode, path, ex);
3565
3566         /* Mark modified extent as dirty */
3567         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3568 out:
3569         ext4_ext_show_leaf(inode, path);
3570         return err;
3571 }
3572
3573 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3574                         sector_t block, int count)
3575 {
3576         int i;
3577         for (i = 0; i < count; i++)
3578                 unmap_underlying_metadata(bdev, block + i);
3579 }
3580
3581 /*
3582  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3583  */
3584 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3585                               ext4_lblk_t lblk,
3586                               struct ext4_ext_path *path,
3587                               unsigned int len)
3588 {
3589         int i, depth;
3590         struct ext4_extent_header *eh;
3591         struct ext4_extent *last_ex;
3592
3593         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3594                 return 0;
3595
3596         depth = ext_depth(inode);
3597         eh = path[depth].p_hdr;
3598
3599         /*
3600          * We're going to remove EOFBLOCKS_FL entirely in future so we
3601          * do not care for this case anymore. Simply remove the flag
3602          * if there are no extents.
3603          */
3604         if (unlikely(!eh->eh_entries))
3605                 goto out;
3606         last_ex = EXT_LAST_EXTENT(eh);
3607         /*
3608          * We should clear the EOFBLOCKS_FL flag if we are writing the
3609          * last block in the last extent in the file.  We test this by
3610          * first checking to see if the caller to
3611          * ext4_ext_get_blocks() was interested in the last block (or
3612          * a block beyond the last block) in the current extent.  If
3613          * this turns out to be false, we can bail out from this
3614          * function immediately.
3615          */
3616         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3617             ext4_ext_get_actual_len(last_ex))
3618                 return 0;
3619         /*
3620          * If the caller does appear to be planning to write at or
3621          * beyond the end of the current extent, we then test to see
3622          * if the current extent is the last extent in the file, by
3623          * checking to make sure it was reached via the rightmost node
3624          * at each level of the tree.
3625          */
3626         for (i = depth-1; i >= 0; i--)
3627                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3628                         return 0;
3629 out:
3630         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3631         return ext4_mark_inode_dirty(handle, inode);
3632 }
3633
3634 /**
3635  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3636  *
3637  * Return 1 if there is a delalloc block in the range, otherwise 0.
3638  */
3639 int ext4_find_delalloc_range(struct inode *inode,
3640                              ext4_lblk_t lblk_start,
3641                              ext4_lblk_t lblk_end)
3642 {
3643         struct extent_status es;
3644
3645         ext4_es_find_delayed_extent(inode, lblk_start, &es);
3646         if (es.es_len == 0)
3647                 return 0; /* there is no delay extent in this tree */
3648         else if (es.es_lblk <= lblk_start &&
3649                  lblk_start < es.es_lblk + es.es_len)
3650                 return 1;
3651         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3652                 return 1;
3653         else
3654                 return 0;
3655 }
3656
3657 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3658 {
3659         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3660         ext4_lblk_t lblk_start, lblk_end;
3661         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3662         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3663
3664         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3665 }
3666
3667 /**
3668  * Determines how many complete clusters (out of those specified by the 'map')
3669  * are under delalloc and were reserved quota for.
3670  * This function is called when we are writing out the blocks that were
3671  * originally written with their allocation delayed, but then the space was
3672  * allocated using fallocate() before the delayed allocation could be resolved.
3673  * The cases to look for are:
3674  * ('=' indicated delayed allocated blocks
3675  *  '-' indicates non-delayed allocated blocks)
3676  * (a) partial clusters towards beginning and/or end outside of allocated range
3677  *     are not delalloc'ed.
3678  *      Ex:
3679  *      |----c---=|====c====|====c====|===-c----|
3680  *               |++++++ allocated ++++++|
3681  *      ==> 4 complete clusters in above example
3682  *
3683  * (b) partial cluster (outside of allocated range) towards either end is
3684  *     marked for delayed allocation. In this case, we will exclude that
3685  *     cluster.
3686  *      Ex:
3687  *      |----====c========|========c========|
3688  *           |++++++ allocated ++++++|
3689  *      ==> 1 complete clusters in above example
3690  *
3691  *      Ex:
3692  *      |================c================|
3693  *            |++++++ allocated ++++++|
3694  *      ==> 0 complete clusters in above example
3695  *
3696  * The ext4_da_update_reserve_space will be called only if we
3697  * determine here that there were some "entire" clusters that span
3698  * this 'allocated' range.
3699  * In the non-bigalloc case, this function will just end up returning num_blks
3700  * without ever calling ext4_find_delalloc_range.
3701  */
3702 static unsigned int
3703 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3704                            unsigned int num_blks)
3705 {
3706         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3707         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3708         ext4_lblk_t lblk_from, lblk_to, c_offset;
3709         unsigned int allocated_clusters = 0;
3710
3711         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3712         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3713
3714         /* max possible clusters for this allocation */
3715         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3716
3717         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3718
3719         /* Check towards left side */
3720         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3721         if (c_offset) {
3722                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3723                 lblk_to = lblk_from + c_offset - 1;
3724
3725                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3726                         allocated_clusters--;
3727         }
3728
3729         /* Now check towards right. */
3730         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3731         if (allocated_clusters && c_offset) {
3732                 lblk_from = lblk_start + num_blks;
3733                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3734
3735                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3736                         allocated_clusters--;
3737         }
3738
3739         return allocated_clusters;
3740 }
3741
3742 static int
3743 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3744                         struct ext4_map_blocks *map,
3745                         struct ext4_ext_path *path, int flags,
3746                         unsigned int allocated, ext4_fsblk_t newblock)
3747 {
3748         int ret = 0;
3749         int err = 0;
3750         ext4_io_end_t *io = ext4_inode_aio(inode);
3751
3752         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3753                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3754                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3755                   flags, allocated);
3756         ext4_ext_show_leaf(inode, path);
3757
3758         trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
3759                                                     allocated, newblock);
3760
3761         /* get_block() before submit the IO, split the extent */
3762         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3763                 ret = ext4_split_unwritten_extents(handle, inode, map,
3764                                                    path, flags);
3765                 if (ret <= 0)
3766                         goto out;
3767                 /*
3768                  * Flag the inode(non aio case) or end_io struct (aio case)
3769                  * that this IO needs to conversion to written when IO is
3770                  * completed
3771                  */
3772                 if (io)
3773                         ext4_set_io_unwritten_flag(inode, io);
3774                 else
3775                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3776                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3777                 if (ext4_should_dioread_nolock(inode))
3778                         map->m_flags |= EXT4_MAP_UNINIT;
3779                 goto out;
3780         }
3781         /* IO end_io complete, convert the filled extent to written */
3782         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3783                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3784                                                         path);
3785                 if (ret >= 0) {
3786                         ext4_update_inode_fsync_trans(handle, inode, 1);
3787                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3788                                                  path, map->m_len);
3789                 } else
3790                         err = ret;
3791                 map->m_flags |= EXT4_MAP_MAPPED;
3792                 if (allocated > map->m_len)
3793                         allocated = map->m_len;
3794                 map->m_len = allocated;
3795                 goto out2;
3796         }
3797         /* buffered IO case */
3798         /*
3799          * repeat fallocate creation request
3800          * we already have an unwritten extent
3801          */
3802         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT) {
3803                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3804                 goto map_out;
3805         }
3806
3807         /* buffered READ or buffered write_begin() lookup */
3808         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3809                 /*
3810                  * We have blocks reserved already.  We
3811                  * return allocated blocks so that delalloc
3812                  * won't do block reservation for us.  But
3813                  * the buffer head will be unmapped so that
3814                  * a read from the block returns 0s.
3815                  */
3816                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3817                 goto out1;
3818         }
3819
3820         /* buffered write, writepage time, convert*/
3821         ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
3822         if (ret >= 0)
3823                 ext4_update_inode_fsync_trans(handle, inode, 1);
3824 out:
3825         if (ret <= 0) {
3826                 err = ret;
3827                 goto out2;
3828         } else
3829                 allocated = ret;
3830         map->m_flags |= EXT4_MAP_NEW;
3831         /*
3832          * if we allocated more blocks than requested
3833          * we need to make sure we unmap the extra block
3834          * allocated. The actual needed block will get
3835          * unmapped later when we find the buffer_head marked
3836          * new.
3837          */
3838         if (allocated > map->m_len) {
3839                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3840                                         newblock + map->m_len,
3841                                         allocated - map->m_len);
3842                 allocated = map->m_len;
3843         }
3844         map->m_len = allocated;
3845
3846         /*
3847          * If we have done fallocate with the offset that is already
3848          * delayed allocated, we would have block reservation
3849          * and quota reservation done in the delayed write path.
3850          * But fallocate would have already updated quota and block
3851          * count for this offset. So cancel these reservation
3852          */
3853         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3854                 unsigned int reserved_clusters;
3855                 reserved_clusters = get_reserved_cluster_alloc(inode,
3856                                 map->m_lblk, map->m_len);
3857                 if (reserved_clusters)
3858                         ext4_da_update_reserve_space(inode,
3859                                                      reserved_clusters,
3860                                                      0);
3861         }
3862
3863 map_out:
3864         map->m_flags |= EXT4_MAP_MAPPED;
3865         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3866                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3867                                          map->m_len);
3868                 if (err < 0)
3869                         goto out2;
3870         }
3871 out1:
3872         if (allocated > map->m_len)
3873                 allocated = map->m_len;
3874         ext4_ext_show_leaf(inode, path);
3875         map->m_pblk = newblock;
3876         map->m_len = allocated;
3877 out2:
3878         if (path) {
3879                 ext4_ext_drop_refs(path);
3880                 kfree(path);
3881         }
3882         return err ? err : allocated;
3883 }
3884
3885 /*
3886  * get_implied_cluster_alloc - check to see if the requested
3887  * allocation (in the map structure) overlaps with a cluster already
3888  * allocated in an extent.
3889  *      @sb     The filesystem superblock structure
3890  *      @map    The requested lblk->pblk mapping
3891  *      @ex     The extent structure which might contain an implied
3892  *                      cluster allocation
3893  *
3894  * This function is called by ext4_ext_map_blocks() after we failed to
3895  * find blocks that were already in the inode's extent tree.  Hence,
3896  * we know that the beginning of the requested region cannot overlap
3897  * the extent from the inode's extent tree.  There are three cases we
3898  * want to catch.  The first is this case:
3899  *
3900  *               |--- cluster # N--|
3901  *    |--- extent ---|  |---- requested region ---|
3902  *                      |==========|
3903  *
3904  * The second case that we need to test for is this one:
3905  *
3906  *   |--------- cluster # N ----------------|
3907  *         |--- requested region --|   |------- extent ----|
3908  *         |=======================|
3909  *
3910  * The third case is when the requested region lies between two extents
3911  * within the same cluster:
3912  *          |------------- cluster # N-------------|
3913  * |----- ex -----|                  |---- ex_right ----|
3914  *                  |------ requested region ------|
3915  *                  |================|
3916  *
3917  * In each of the above cases, we need to set the map->m_pblk and
3918  * map->m_len so it corresponds to the return the extent labelled as
3919  * "|====|" from cluster #N, since it is already in use for data in
3920  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3921  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3922  * as a new "allocated" block region.  Otherwise, we will return 0 and
3923  * ext4_ext_map_blocks() will then allocate one or more new clusters
3924  * by calling ext4_mb_new_blocks().
3925  */
3926 static int get_implied_cluster_alloc(struct super_block *sb,
3927                                      struct ext4_map_blocks *map,
3928                                      struct ext4_extent *ex,
3929                                      struct ext4_ext_path *path)
3930 {
3931         struct ext4_sb_info *sbi = EXT4_SB(sb);
3932         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3933         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3934         ext4_lblk_t rr_cluster_start;
3935         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3936         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3937         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3938
3939         /* The extent passed in that we are trying to match */
3940         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3941         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3942
3943         /* The requested region passed into ext4_map_blocks() */
3944         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3945
3946         if ((rr_cluster_start == ex_cluster_end) ||
3947             (rr_cluster_start == ex_cluster_start)) {
3948                 if (rr_cluster_start == ex_cluster_end)
3949                         ee_start += ee_len - 1;
3950                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3951                         c_offset;
3952                 map->m_len = min(map->m_len,
3953                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3954                 /*
3955                  * Check for and handle this case:
3956                  *
3957                  *   |--------- cluster # N-------------|
3958                  *                     |------- extent ----|
3959                  *         |--- requested region ---|
3960                  *         |===========|
3961                  */
3962
3963                 if (map->m_lblk < ee_block)
3964                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3965
3966                 /*
3967                  * Check for the case where there is already another allocated
3968                  * block to the right of 'ex' but before the end of the cluster.
3969                  *
3970                  *          |------------- cluster # N-------------|
3971                  * |----- ex -----|                  |---- ex_right ----|
3972                  *                  |------ requested region ------|
3973                  *                  |================|
3974                  */
3975                 if (map->m_lblk > ee_block) {
3976                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
3977                         map->m_len = min(map->m_len, next - map->m_lblk);
3978                 }
3979
3980                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
3981                 return 1;
3982         }
3983
3984         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
3985         return 0;
3986 }
3987
3988
3989 /*
3990  * Block allocation/map/preallocation routine for extents based files
3991  *
3992  *
3993  * Need to be called with
3994  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
3995  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
3996  *
3997  * return > 0, number of of blocks already mapped/allocated
3998  *          if create == 0 and these are pre-allocated blocks
3999  *              buffer head is unmapped
4000  *          otherwise blocks are mapped
4001  *
4002  * return = 0, if plain look up failed (blocks have not been allocated)
4003  *          buffer head is unmapped
4004  *
4005  * return < 0, error case.
4006  */
4007 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4008                         struct ext4_map_blocks *map, int flags)
4009 {
4010         struct ext4_ext_path *path = NULL;
4011         struct ext4_extent newex, *ex, *ex2;
4012         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4013         ext4_fsblk_t newblock = 0;
4014         int free_on_err = 0, err = 0, depth;
4015         unsigned int allocated = 0, offset = 0;
4016         unsigned int allocated_clusters = 0;
4017         struct ext4_allocation_request ar;
4018         ext4_io_end_t *io = ext4_inode_aio(inode);
4019         ext4_lblk_t cluster_offset;
4020         int set_unwritten = 0;
4021
4022         ext_debug("blocks %u/%u requested for inode %lu\n",
4023                   map->m_lblk, map->m_len, inode->i_ino);
4024         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4025
4026         /* find extent for this block */
4027         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
4028         if (IS_ERR(path)) {
4029                 err = PTR_ERR(path);
4030                 path = NULL;
4031                 goto out2;
4032         }
4033
4034         depth = ext_depth(inode);
4035
4036         /*
4037          * consistent leaf must not be empty;
4038          * this situation is possible, though, _during_ tree modification;
4039          * this is why assert can't be put in ext4_ext_find_extent()
4040          */
4041         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4042                 EXT4_ERROR_INODE(inode, "bad extent address "
4043                                  "lblock: %lu, depth: %d pblock %lld",
4044                                  (unsigned long) map->m_lblk, depth,
4045                                  path[depth].p_block);
4046                 err = -EIO;
4047                 goto out2;
4048         }
4049
4050         ex = path[depth].p_ext;
4051         if (ex) {
4052                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4053                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4054                 unsigned short ee_len;
4055
4056                 /*
4057                  * Uninitialized extents are treated as holes, except that
4058                  * we split out initialized portions during a write.
4059                  */
4060                 ee_len = ext4_ext_get_actual_len(ex);
4061
4062                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4063
4064                 /* if found extent covers block, simply return it */
4065                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4066                         newblock = map->m_lblk - ee_block + ee_start;
4067                         /* number of remaining blocks in the extent */
4068                         allocated = ee_len - (map->m_lblk - ee_block);
4069                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4070                                   ee_block, ee_len, newblock);
4071
4072                         if (!ext4_ext_is_uninitialized(ex))
4073                                 goto out;
4074
4075                         allocated = ext4_ext_handle_uninitialized_extents(
4076                                 handle, inode, map, path, flags,
4077                                 allocated, newblock);
4078                         goto out3;
4079                 }
4080         }
4081
4082         if ((sbi->s_cluster_ratio > 1) &&
4083             ext4_find_delalloc_cluster(inode, map->m_lblk))
4084                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4085
4086         /*
4087          * requested block isn't allocated yet;
4088          * we couldn't try to create block if create flag is zero
4089          */
4090         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4091                 /*
4092                  * put just found gap into cache to speed up
4093                  * subsequent requests
4094                  */
4095                 if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
4096                         ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4097                 goto out2;
4098         }
4099
4100         /*
4101          * Okay, we need to do block allocation.
4102          */
4103         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
4104         newex.ee_block = cpu_to_le32(map->m_lblk);
4105         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
4106
4107         /*
4108          * If we are doing bigalloc, check to see if the extent returned
4109          * by ext4_ext_find_extent() implies a cluster we can use.
4110          */
4111         if (cluster_offset && ex &&
4112             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4113                 ar.len = allocated = map->m_len;
4114                 newblock = map->m_pblk;
4115                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4116                 goto got_allocated_blocks;
4117         }
4118
4119         /* find neighbour allocated blocks */
4120         ar.lleft = map->m_lblk;
4121         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4122         if (err)
4123                 goto out2;
4124         ar.lright = map->m_lblk;
4125         ex2 = NULL;
4126         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4127         if (err)
4128                 goto out2;
4129
4130         /* Check if the extent after searching to the right implies a
4131          * cluster we can use. */
4132         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4133             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4134                 ar.len = allocated = map->m_len;
4135                 newblock = map->m_pblk;
4136                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4137                 goto got_allocated_blocks;
4138         }
4139
4140         /*
4141          * See if request is beyond maximum number of blocks we can have in
4142          * a single extent. For an initialized extent this limit is
4143          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4144          * EXT_UNINIT_MAX_LEN.
4145          */
4146         if (map->m_len > EXT_INIT_MAX_LEN &&
4147             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4148                 map->m_len = EXT_INIT_MAX_LEN;
4149         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4150                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4151                 map->m_len = EXT_UNINIT_MAX_LEN;
4152
4153         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4154         newex.ee_len = cpu_to_le16(map->m_len);
4155         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4156         if (err)
4157                 allocated = ext4_ext_get_actual_len(&newex);
4158         else
4159                 allocated = map->m_len;
4160
4161         /* allocate new block */
4162         ar.inode = inode;
4163         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4164         ar.logical = map->m_lblk;
4165         /*
4166          * We calculate the offset from the beginning of the cluster
4167          * for the logical block number, since when we allocate a
4168          * physical cluster, the physical block should start at the
4169          * same offset from the beginning of the cluster.  This is
4170          * needed so that future calls to get_implied_cluster_alloc()
4171          * work correctly.
4172          */
4173         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
4174         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4175         ar.goal -= offset;
4176         ar.logical -= offset;
4177         if (S_ISREG(inode->i_mode))
4178                 ar.flags = EXT4_MB_HINT_DATA;
4179         else
4180                 /* disable in-core preallocation for non-regular files */
4181                 ar.flags = 0;
4182         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4183                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4184         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4185         if (!newblock)
4186                 goto out2;
4187         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4188                   ar.goal, newblock, allocated);
4189         free_on_err = 1;
4190         allocated_clusters = ar.len;
4191         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4192         if (ar.len > allocated)
4193                 ar.len = allocated;
4194
4195 got_allocated_blocks:
4196         /* try to insert new extent into found leaf and return */
4197         ext4_ext_store_pblock(&newex, newblock + offset);
4198         newex.ee_len = cpu_to_le16(ar.len);
4199         /* Mark uninitialized */
4200         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4201                 ext4_ext_mark_uninitialized(&newex);
4202                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4203                 /*
4204                  * io_end structure was created for every IO write to an
4205                  * uninitialized extent. To avoid unnecessary conversion,
4206                  * here we flag the IO that really needs the conversion.
4207                  * For non asycn direct IO case, flag the inode state
4208                  * that we need to perform conversion when IO is done.
4209                  */
4210                 if ((flags & EXT4_GET_BLOCKS_PRE_IO))
4211                         set_unwritten = 1;
4212                 if (ext4_should_dioread_nolock(inode))
4213                         map->m_flags |= EXT4_MAP_UNINIT;
4214         }
4215
4216         err = 0;
4217         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4218                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4219                                          path, ar.len);
4220         if (!err)
4221                 err = ext4_ext_insert_extent(handle, inode, path,
4222                                              &newex, flags);
4223
4224         if (!err && set_unwritten) {
4225                 if (io)
4226                         ext4_set_io_unwritten_flag(inode, io);
4227                 else
4228                         ext4_set_inode_state(inode,
4229                                              EXT4_STATE_DIO_UNWRITTEN);
4230         }
4231
4232         if (err && free_on_err) {
4233                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4234                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4235                 /* free data blocks we just allocated */
4236                 /* not a good idea to call discard here directly,
4237                  * but otherwise we'd need to call it every free() */
4238                 ext4_discard_preallocations(inode);
4239                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4240                                  ext4_ext_get_actual_len(&newex), fb_flags);
4241                 goto out2;
4242         }
4243
4244         /* previous routine could use block we allocated */
4245         newblock = ext4_ext_pblock(&newex);
4246         allocated = ext4_ext_get_actual_len(&newex);
4247         if (allocated > map->m_len)
4248                 allocated = map->m_len;
4249         map->m_flags |= EXT4_MAP_NEW;
4250
4251         /*
4252          * Update reserved blocks/metadata blocks after successful
4253          * block allocation which had been deferred till now.
4254          */
4255         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4256                 unsigned int reserved_clusters;
4257                 /*
4258                  * Check how many clusters we had reserved this allocated range
4259                  */
4260                 reserved_clusters = get_reserved_cluster_alloc(inode,
4261                                                 map->m_lblk, allocated);
4262                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4263                         if (reserved_clusters) {
4264                                 /*
4265                                  * We have clusters reserved for this range.
4266                                  * But since we are not doing actual allocation
4267                                  * and are simply using blocks from previously
4268                                  * allocated cluster, we should release the
4269                                  * reservation and not claim quota.
4270                                  */
4271                                 ext4_da_update_reserve_space(inode,
4272                                                 reserved_clusters, 0);
4273                         }
4274                 } else {
4275                         BUG_ON(allocated_clusters < reserved_clusters);
4276                         if (reserved_clusters < allocated_clusters) {
4277                                 struct ext4_inode_info *ei = EXT4_I(inode);
4278                                 int reservation = allocated_clusters -
4279                                                   reserved_clusters;
4280                                 /*
4281                                  * It seems we claimed few clusters outside of
4282                                  * the range of this allocation. We should give
4283                                  * it back to the reservation pool. This can
4284                                  * happen in the following case:
4285                                  *
4286                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4287                                  *   cluster has 4 blocks. Thus, the clusters
4288                                  *   are [0-3],[4-7],[8-11]...
4289                                  * * First comes delayed allocation write for
4290                                  *   logical blocks 10 & 11. Since there were no
4291                                  *   previous delayed allocated blocks in the
4292                                  *   range [8-11], we would reserve 1 cluster
4293                                  *   for this write.
4294                                  * * Next comes write for logical blocks 3 to 8.
4295                                  *   In this case, we will reserve 2 clusters
4296                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4297                                  *   that range has a delayed allocated blocks.
4298                                  *   Thus total reserved clusters now becomes 3.
4299                                  * * Now, during the delayed allocation writeout
4300                                  *   time, we will first write blocks [3-8] and
4301                                  *   allocate 3 clusters for writing these
4302                                  *   blocks. Also, we would claim all these
4303                                  *   three clusters above.
4304                                  * * Now when we come here to writeout the
4305                                  *   blocks [10-11], we would expect to claim
4306                                  *   the reservation of 1 cluster we had made
4307                                  *   (and we would claim it since there are no
4308                                  *   more delayed allocated blocks in the range
4309                                  *   [8-11]. But our reserved cluster count had
4310                                  *   already gone to 0.
4311                                  *
4312                                  *   Thus, at the step 4 above when we determine
4313                                  *   that there are still some unwritten delayed
4314                                  *   allocated blocks outside of our current
4315                                  *   block range, we should increment the
4316                                  *   reserved clusters count so that when the
4317                                  *   remaining blocks finally gets written, we
4318                                  *   could claim them.
4319                                  */
4320                                 dquot_reserve_block(inode,
4321                                                 EXT4_C2B(sbi, reservation));
4322                                 spin_lock(&ei->i_block_reservation_lock);
4323                                 ei->i_reserved_data_blocks += reservation;
4324                                 spin_unlock(&ei->i_block_reservation_lock);
4325                         }
4326                         /*
4327                          * We will claim quota for all newly allocated blocks.
4328                          * We're updating the reserved space *after* the
4329                          * correction above so we do not accidentally free
4330                          * all the metadata reservation because we might
4331                          * actually need it later on.
4332                          */
4333                         ext4_da_update_reserve_space(inode, allocated_clusters,
4334                                                         1);
4335                 }
4336         }
4337
4338         /*
4339          * Cache the extent and update transaction to commit on fdatasync only
4340          * when it is _not_ an uninitialized extent.
4341          */
4342         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0)
4343                 ext4_update_inode_fsync_trans(handle, inode, 1);
4344         else
4345                 ext4_update_inode_fsync_trans(handle, inode, 0);
4346 out:
4347         if (allocated > map->m_len)
4348                 allocated = map->m_len;
4349         ext4_ext_show_leaf(inode, path);
4350         map->m_flags |= EXT4_MAP_MAPPED;
4351         map->m_pblk = newblock;
4352         map->m_len = allocated;
4353 out2:
4354         if (path) {
4355                 ext4_ext_drop_refs(path);
4356                 kfree(path);
4357         }
4358
4359 out3:
4360         trace_ext4_ext_map_blocks_exit(inode, map, err ? err : allocated);
4361
4362         return err ? err : allocated;
4363 }
4364
4365 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4366 {
4367         struct super_block *sb = inode->i_sb;
4368         ext4_lblk_t last_block;
4369         int err = 0;
4370
4371         /*
4372          * TODO: optimization is possible here.
4373          * Probably we need not scan at all,
4374          * because page truncation is enough.
4375          */
4376
4377         /* we have to know where to truncate from in crash case */
4378         EXT4_I(inode)->i_disksize = inode->i_size;
4379         ext4_mark_inode_dirty(handle, inode);
4380
4381         last_block = (inode->i_size + sb->s_blocksize - 1)
4382                         >> EXT4_BLOCK_SIZE_BITS(sb);
4383         err = ext4_es_remove_extent(inode, last_block,
4384                                     EXT_MAX_BLOCKS - last_block);
4385         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4386 }
4387
4388 static void ext4_falloc_update_inode(struct inode *inode,
4389                                 int mode, loff_t new_size, int update_ctime)
4390 {
4391         struct timespec now;
4392
4393         if (update_ctime) {
4394                 now = current_fs_time(inode->i_sb);
4395                 if (!timespec_equal(&inode->i_ctime, &now))
4396                         inode->i_ctime = now;
4397         }
4398         /*
4399          * Update only when preallocation was requested beyond
4400          * the file size.
4401          */
4402         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4403                 if (new_size > i_size_read(inode))
4404                         i_size_write(inode, new_size);
4405                 if (new_size > EXT4_I(inode)->i_disksize)
4406                         ext4_update_i_disksize(inode, new_size);
4407         } else {
4408                 /*
4409                  * Mark that we allocate beyond EOF so the subsequent truncate
4410                  * can proceed even if the new size is the same as i_size.
4411                  */
4412                 if (new_size > i_size_read(inode))
4413                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4414         }
4415
4416 }
4417
4418 /*
4419  * preallocate space for a file. This implements ext4's fallocate file
4420  * operation, which gets called from sys_fallocate system call.
4421  * For block-mapped files, posix_fallocate should fall back to the method
4422  * of writing zeroes to the required new blocks (the same behavior which is
4423  * expected for file systems which do not support fallocate() system call).
4424  */
4425 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4426 {
4427         struct inode *inode = file_inode(file);
4428         handle_t *handle;
4429         loff_t new_size;
4430         unsigned int max_blocks;
4431         int ret = 0;
4432         int ret2 = 0;
4433         int retries = 0;
4434         int flags;
4435         struct ext4_map_blocks map;
4436         unsigned int credits, blkbits = inode->i_blkbits;
4437
4438         /* Return error if mode is not supported */
4439         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4440                 return -EOPNOTSUPP;
4441
4442         if (mode & FALLOC_FL_PUNCH_HOLE)
4443                 return ext4_punch_hole(file, offset, len);
4444
4445         ret = ext4_convert_inline_data(inode);
4446         if (ret)
4447                 return ret;
4448
4449         /*
4450          * currently supporting (pre)allocate mode for extent-based
4451          * files _only_
4452          */
4453         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4454                 return -EOPNOTSUPP;
4455
4456         trace_ext4_fallocate_enter(inode, offset, len, mode);
4457         map.m_lblk = offset >> blkbits;
4458         /*
4459          * We can't just convert len to max_blocks because
4460          * If blocksize = 4096 offset = 3072 and len = 2048
4461          */
4462         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4463                 - map.m_lblk;
4464         /*
4465          * credits to insert 1 extent into extent tree
4466          */
4467         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4468         mutex_lock(&inode->i_mutex);
4469         ret = inode_newsize_ok(inode, (len + offset));
4470         if (ret) {
4471                 mutex_unlock(&inode->i_mutex);
4472                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4473                 return ret;
4474         }
4475         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4476         if (mode & FALLOC_FL_KEEP_SIZE)
4477                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4478         /*
4479          * Don't normalize the request if it can fit in one extent so
4480          * that it doesn't get unnecessarily split into multiple
4481          * extents.
4482          */
4483         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4484                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4485
4486 retry:
4487         while (ret >= 0 && ret < max_blocks) {
4488                 map.m_lblk = map.m_lblk + ret;
4489                 map.m_len = max_blocks = max_blocks - ret;
4490                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4491                                             credits);
4492                 if (IS_ERR(handle)) {
4493                         ret = PTR_ERR(handle);
4494                         break;
4495                 }
4496                 ret = ext4_map_blocks(handle, inode, &map, flags);
4497                 if (ret <= 0) {
4498 #ifdef EXT4FS_DEBUG
4499                         ext4_warning(inode->i_sb,
4500                                      "inode #%lu: block %u: len %u: "
4501                                      "ext4_ext_map_blocks returned %d",
4502                                      inode->i_ino, map.m_lblk,
4503                                      map.m_len, ret);
4504 #endif
4505                         ext4_mark_inode_dirty(handle, inode);
4506                         ret2 = ext4_journal_stop(handle);
4507                         break;
4508                 }
4509                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4510                                                 blkbits) >> blkbits))
4511                         new_size = offset + len;
4512                 else
4513                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4514
4515                 ext4_falloc_update_inode(inode, mode, new_size,
4516                                          (map.m_flags & EXT4_MAP_NEW));
4517                 ext4_mark_inode_dirty(handle, inode);
4518                 if ((file->f_flags & O_SYNC) && ret >= max_blocks)
4519                         ext4_handle_sync(handle);
4520                 ret2 = ext4_journal_stop(handle);
4521                 if (ret2)
4522                         break;
4523         }
4524         if (ret == -ENOSPC &&
4525                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4526                 ret = 0;
4527                 goto retry;
4528         }
4529         mutex_unlock(&inode->i_mutex);
4530         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4531                                 ret > 0 ? ret2 : ret);
4532         return ret > 0 ? ret2 : ret;
4533 }
4534
4535 /*
4536  * This function convert a range of blocks to written extents
4537  * The caller of this function will pass the start offset and the size.
4538  * all unwritten extents within this range will be converted to
4539  * written extents.
4540  *
4541  * This function is called from the direct IO end io call back
4542  * function, to convert the fallocated extents after IO is completed.
4543  * Returns 0 on success.
4544  */
4545 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4546                                     ssize_t len)
4547 {
4548         handle_t *handle;
4549         unsigned int max_blocks;
4550         int ret = 0;
4551         int ret2 = 0;
4552         struct ext4_map_blocks map;
4553         unsigned int credits, blkbits = inode->i_blkbits;
4554
4555         map.m_lblk = offset >> blkbits;
4556         /*
4557          * We can't just convert len to max_blocks because
4558          * If blocksize = 4096 offset = 3072 and len = 2048
4559          */
4560         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4561                       map.m_lblk);
4562         /*
4563          * credits to insert 1 extent into extent tree
4564          */
4565         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4566         while (ret >= 0 && ret < max_blocks) {
4567                 map.m_lblk += ret;
4568                 map.m_len = (max_blocks -= ret);
4569                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
4570                 if (IS_ERR(handle)) {
4571                         ret = PTR_ERR(handle);
4572                         break;
4573                 }
4574                 ret = ext4_map_blocks(handle, inode, &map,
4575                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4576                 if (ret <= 0)
4577                         ext4_warning(inode->i_sb,
4578                                      "inode #%lu: block %u: len %u: "
4579                                      "ext4_ext_map_blocks returned %d",
4580                                      inode->i_ino, map.m_lblk,
4581                                      map.m_len, ret);
4582                 ext4_mark_inode_dirty(handle, inode);
4583                 ret2 = ext4_journal_stop(handle);
4584                 if (ret <= 0 || ret2 )
4585                         break;
4586         }
4587         return ret > 0 ? ret2 : ret;
4588 }
4589
4590 /*
4591  * If newes is not existing extent (newes->ec_pblk equals zero) find
4592  * delayed extent at start of newes and update newes accordingly and
4593  * return start of the next delayed extent.
4594  *
4595  * If newes is existing extent (newes->ec_pblk is not equal zero)
4596  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4597  * extent found. Leave newes unmodified.
4598  */
4599 static int ext4_find_delayed_extent(struct inode *inode,
4600                                     struct extent_status *newes)
4601 {
4602         struct extent_status es;
4603         ext4_lblk_t block, next_del;
4604
4605         ext4_es_find_delayed_extent(inode, newes->es_lblk, &es);
4606
4607         if (newes->es_pblk == 0) {
4608                 /*
4609                  * No extent in extent-tree contains block @newes->es_pblk,
4610                  * then the block may stay in 1)a hole or 2)delayed-extent.
4611                  */
4612                 if (es.es_len == 0)
4613                         /* A hole found. */
4614                         return 0;
4615
4616                 if (es.es_lblk > newes->es_lblk) {
4617                         /* A hole found. */
4618                         newes->es_len = min(es.es_lblk - newes->es_lblk,
4619                                             newes->es_len);
4620                         return 0;
4621                 }
4622
4623                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
4624         }
4625
4626         block = newes->es_lblk + newes->es_len;
4627         ext4_es_find_delayed_extent(inode, block, &es);
4628         if (es.es_len == 0)
4629                 next_del = EXT_MAX_BLOCKS;
4630         else
4631                 next_del = es.es_lblk;
4632
4633         return next_del;
4634 }
4635 /* fiemap flags we can handle specified here */
4636 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4637
4638 static int ext4_xattr_fiemap(struct inode *inode,
4639                                 struct fiemap_extent_info *fieinfo)
4640 {
4641         __u64 physical = 0;
4642         __u64 length;
4643         __u32 flags = FIEMAP_EXTENT_LAST;
4644         int blockbits = inode->i_sb->s_blocksize_bits;
4645         int error = 0;
4646
4647         /* in-inode? */
4648         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4649                 struct ext4_iloc iloc;
4650                 int offset;     /* offset of xattr in inode */
4651
4652                 error = ext4_get_inode_loc(inode, &iloc);
4653                 if (error)
4654                         return error;
4655                 physical = iloc.bh->b_blocknr << blockbits;
4656                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4657                                 EXT4_I(inode)->i_extra_isize;
4658                 physical += offset;
4659                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4660                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4661                 brelse(iloc.bh);
4662         } else { /* external block */
4663                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4664                 length = inode->i_sb->s_blocksize;
4665         }
4666
4667         if (physical)
4668                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4669                                                 length, flags);
4670         return (error < 0 ? error : 0);
4671 }
4672
4673 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4674                 __u64 start, __u64 len)
4675 {
4676         ext4_lblk_t start_blk;
4677         int error = 0;
4678
4679         if (ext4_has_inline_data(inode)) {
4680                 int has_inline = 1;
4681
4682                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
4683
4684                 if (has_inline)
4685                         return error;
4686         }
4687
4688         /* fallback to generic here if not in extents fmt */
4689         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4690                 return generic_block_fiemap(inode, fieinfo, start, len,
4691                         ext4_get_block);
4692
4693         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4694                 return -EBADR;
4695
4696         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4697                 error = ext4_xattr_fiemap(inode, fieinfo);
4698         } else {
4699                 ext4_lblk_t len_blks;
4700                 __u64 last_blk;
4701
4702                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4703                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4704                 if (last_blk >= EXT_MAX_BLOCKS)
4705                         last_blk = EXT_MAX_BLOCKS-1;
4706                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4707
4708                 /*
4709                  * Walk the extent tree gathering extent information
4710                  * and pushing extents back to the user.
4711                  */
4712                 error = ext4_fill_fiemap_extents(inode, start_blk,
4713                                                  len_blks, fieinfo);
4714         }
4715
4716         return error;
4717 }
4718
4719 /*
4720  * Migrate a simple extent-based inode to use the i_blocks[] array
4721  */
4722 int ext4_ind_migrate(struct inode *inode)
4723 {
4724         struct ext4_extent_header       *eh;
4725         struct ext4_super_block         *es = EXT4_SB(inode->i_sb)->s_es;
4726         struct ext4_inode_info          *ei = EXT4_I(inode);
4727         struct ext4_extent              *ex;
4728         unsigned int                    i, len;
4729         ext4_fsblk_t                    blk;
4730         handle_t                        *handle;
4731         int                             ret;
4732
4733         if (!EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
4734                                        EXT4_FEATURE_INCOMPAT_EXTENTS) ||
4735             (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4736                 return -EINVAL;
4737
4738         down_write(&EXT4_I(inode)->i_data_sem);
4739         ret = ext4_ext_check_inode(inode);
4740         if (ret)
4741                 goto errout;
4742
4743         eh = ext_inode_hdr(inode);
4744         ex  = EXT_FIRST_EXTENT(eh);
4745         if (ext4_blocks_count(es) > EXT4_MAX_BLOCK_FILE_PHYS ||
4746             eh->eh_depth != 0 || eh->eh_entries > 1) {
4747                 ret = -EOPNOTSUPP;
4748                 goto errout;
4749         }
4750         if (eh->eh_entries == 0)
4751                 blk = len = 0;
4752         else {
4753                 len = le16_to_cpu(ex->ee_len);
4754                 blk = ext4_ext_pblock(ex);
4755                 if (len > EXT4_NDIR_BLOCKS) {
4756                         ret = -EOPNOTSUPP;
4757                         goto errout;
4758                 }
4759         }
4760
4761         handle = ext4_journal_start(inode, EXT4_HT_MIGRATE, 1);
4762         if (IS_ERR(handle)) {
4763                 ret = PTR_ERR(handle);
4764                 goto errout;
4765         }
4766
4767         ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS);
4768         memset(ei->i_data, 0, sizeof(ei->i_data));
4769         for (i=0; i < len; i++)
4770                 ei->i_data[i] = cpu_to_le32(blk++);
4771         ext4_mark_inode_dirty(handle, inode);
4772         ext4_journal_stop(handle);
4773 errout:
4774         up_write(&EXT4_I(inode)->i_data_sem);
4775         return ret;
4776 }