Pileus Git - ~andy/linux/blob - 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 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 161                      struct inode *inode, struct ext4_ext_path *path)
 162 {
 163         int err;
 164         if (path->p_bh) {
 165                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 166                 /* path points to block */
 167                 err = __ext4_handle_dirty_metadata(where, line, handle,
 168                                                    inode, path->p_bh);
 169         } else {
 170                 /* path points to leaf/index in inode body */
 171                 err = ext4_mark_inode_dirty(handle, inode);
 172         }
 173         return err;
 174 }
 175
 176 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 177                               struct ext4_ext_path *path,
 178                               ext4_lblk_t block)
 179 {
 180         if (path) {
 181                 int depth = path->p_depth;
 182                 struct ext4_extent *ex;
 183
 184                 /*
 185                  * Try to predict block placement assuming that we are
 186                  * filling in a file which will eventually be
 187                  * non-sparse --- i.e., in the case of libbfd writing
 188                  * an ELF object sections out-of-order but in a way
 189                  * the eventually results in a contiguous object or
 190                  * executable file, or some database extending a table
 191                  * space file.  However, this is actually somewhat
 192                  * non-ideal if we are writing a sparse file such as
 193                  * qemu or KVM writing a raw image file that is going
 194                  * to stay fairly sparse, since it will end up
 195                  * fragmenting the file system's free space.  Maybe we
 196                  * should have some hueristics or some way to allow
 197                  * userspace to pass a hint to file system,
 198                  * especially if the latter case turns out to be
 199                  * common.
 200                  */
 201                 ex = path[depth].p_ext;
 202                 if (ex) {
 203                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 204                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 205
 206                         if (block > ext_block)
 207                                 return ext_pblk + (block - ext_block);
 208                         else
 209                                 return ext_pblk - (ext_block - block);
 210                 }
 211
 212                 /* it looks like index is empty;
 213                  * try to find starting block from index itself */
 214                 if (path[depth].p_bh)
 215                         return path[depth].p_bh->b_blocknr;
 216         }
 217
 218         /* OK. use inode's group */
 219         return ext4_inode_to_goal_block(inode);
 220 }
 221
 222 /*
 223  * Allocation for a meta data block
 224  */
 225 static ext4_fsblk_t
 226 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 227                         struct ext4_ext_path *path,
 228                         struct ext4_extent *ex, int *err, unsigned int flags)
 229 {
 230         ext4_fsblk_t goal, newblock;
 231
 232         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 233         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 234                                         NULL, err);
 235         return newblock;
 236 }
 237
 238 static inline int ext4_ext_space_block(struct inode *inode, int check)
 239 {
 240         int size;
 241
 242         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 243                         / sizeof(struct ext4_extent);
 244 #ifdef AGGRESSIVE_TEST
 245         if (!check && size > 6)
 246                 size = 6;
 247 #endif
 248         return size;
 249 }
 250
 251 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 252 {
 253         int size;
 254
 255         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 256                         / sizeof(struct ext4_extent_idx);
 257 #ifdef AGGRESSIVE_TEST
 258         if (!check && size > 5)
 259                 size = 5;
 260 #endif
 261         return size;
 262 }
 263
 264 static inline int ext4_ext_space_root(struct inode *inode, int check)
 265 {
 266         int size;
 267
 268         size = sizeof(EXT4_I(inode)->i_data);
 269         size -= sizeof(struct ext4_extent_header);
 270         size /= sizeof(struct ext4_extent);
 271 #ifdef AGGRESSIVE_TEST
 272         if (!check && size > 3)
 273                 size = 3;
 274 #endif
 275         return size;
 276 }
 277
 278 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 279 {
 280         int size;
 281
 282         size = sizeof(EXT4_I(inode)->i_data);
 283         size -= sizeof(struct ext4_extent_header);
 284         size /= sizeof(struct ext4_extent_idx);
 285 #ifdef AGGRESSIVE_TEST
 286         if (!check && size > 4)
 287                 size = 4;
 288 #endif
 289         return size;
 290 }
 291
 292 /*
 293  * Calculate the number of metadata blocks needed
 294  * to allocate @blocks
 295  * Worse case is one block per extent
 296  */
 297 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 298 {
 299         struct ext4_inode_info *ei = EXT4_I(inode);
 300         int idxs;
 301
 302         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 303                 / sizeof(struct ext4_extent_idx));
 304
 305         /*
 306          * If the new delayed allocation block is contiguous with the
 307          * previous da block, it can share index blocks with the
 308          * previous block, so we only need to allocate a new index
 309          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 310          * an additional index block, and at ldxs**3 blocks, yet
 311          * another index blocks.
 312          */
 313         if (ei->i_da_metadata_calc_len &&
 314             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 315                 int num = 0;
 316
 317                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 318                         num++;
 319                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 320                         num++;
 321                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 322                         num++;
 323                         ei->i_da_metadata_calc_len = 0;
 324                 } else
 325                         ei->i_da_metadata_calc_len++;
 326                 ei->i_da_metadata_calc_last_lblock++;
 327                 return num;
 328         }
 329
 330         /*
 331          * In the worst case we need a new set of index blocks at
 332          * every level of the inode's extent tree.
 333          */
 334         ei->i_da_metadata_calc_len = 1;
 335         ei->i_da_metadata_calc_last_lblock = lblock;
 336         return ext_depth(inode) + 1;
 337 }
 338
 339 static int
 340 ext4_ext_max_entries(struct inode *inode, int depth)
 341 {
 342         int max;
 343
 344         if (depth == ext_depth(inode)) {
 345                 if (depth == 0)
 346                         max = ext4_ext_space_root(inode, 1);
 347                 else
 348                         max = ext4_ext_space_root_idx(inode, 1);
 349         } else {
 350                 if (depth == 0)
 351                         max = ext4_ext_space_block(inode, 1);
 352                 else
 353                         max = ext4_ext_space_block_idx(inode, 1);
 354         }
 355
 356         return max;
 357 }
 358
 359 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 360 {
 361         ext4_fsblk_t block = ext4_ext_pblock(ext);
 362         int len = ext4_ext_get_actual_len(ext);
 363
 364         if (len == 0)
 365                 return 0;
 366         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 367 }
 368
 369 static int ext4_valid_extent_idx(struct inode *inode,
 370                                 struct ext4_extent_idx *ext_idx)
 371 {
 372         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 373
 374         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 375 }
 376
 377 static int ext4_valid_extent_entries(struct inode *inode,
 378                                 struct ext4_extent_header *eh,
 379                                 int depth)
 380 {
 381         unsigned short entries;
 382         if (eh->eh_entries == 0)
 383                 return 1;
 384
 385         entries = le16_to_cpu(eh->eh_entries);
 386
 387         if (depth == 0) {
 388                 /* leaf entries */
 389                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 390                 while (entries) {
 391                         if (!ext4_valid_extent(inode, ext))
 392                                 return 0;
 393                         ext++;
 394                         entries--;
 395                 }
 396         } else {
 397                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 398                 while (entries) {
 399                         if (!ext4_valid_extent_idx(inode, ext_idx))
 400                                 return 0;
 401                         ext_idx++;
 402                         entries--;
 403                 }
 404         }
 405         return 1;
 406 }
 407
 408 static int __ext4_ext_check(const char *function, unsigned int line,
 409                             struct inode *inode, struct ext4_extent_header *eh,
 410                             int depth)
 411 {
 412         const char *error_msg;
 413         int max = 0;
 414
 415         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 416                 error_msg = "invalid magic";
 417                 goto corrupted;
 418         }
 419         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 420                 error_msg = "unexpected eh_depth";
 421                 goto corrupted;
 422         }
 423         if (unlikely(eh->eh_max == 0)) {
 424                 error_msg = "invalid eh_max";
 425                 goto corrupted;
 426         }
 427         max = ext4_ext_max_entries(inode, depth);
 428         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 429                 error_msg = "too large eh_max";
 430                 goto corrupted;
 431         }
 432         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 433                 error_msg = "invalid eh_entries";
 434                 goto corrupted;
 435         }
 436         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 437                 error_msg = "invalid extent entries";
 438                 goto corrupted;
 439         }
 440         /* Verify checksum on non-root extent tree nodes */
 441         if (ext_depth(inode) != depth &&
 442             !ext4_extent_block_csum_verify(inode, eh)) {
 443                 error_msg = "extent tree corrupted";
 444                 goto corrupted;
 445         }
 446         return 0;
 447
 448 corrupted:
 449         ext4_error_inode(inode, function, line, 0,
 450                         "bad header/extent: %s - magic %x, "
 451                         "entries %u, max %u(%u), depth %u(%u)",
 452                         error_msg, le16_to_cpu(eh->eh_magic),
 453                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 454                         max, le16_to_cpu(eh->eh_depth), depth);
 455
 456         return -EIO;
 457 }
 458
 459 #define ext4_ext_check(inode, eh, depth)        \
 460         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
 461
 462 int ext4_ext_check_inode(struct inode *inode)
 463 {
 464         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
 465 }
 466
 467 static int __ext4_ext_check_block(const char *function, unsigned int line,
 468                                   struct inode *inode,
 469                                   struct ext4_extent_header *eh,
 470                                   int depth,
 471                                   struct buffer_head *bh)
 472 {
 473         int ret;
 474
 475         if (buffer_verified(bh))
 476                 return 0;
 477         ret = ext4_ext_check(inode, eh, depth);
 478         if (ret)
 479                 return ret;
 480         set_buffer_verified(bh);
 481         return ret;
 482 }
 483
 484 #define ext4_ext_check_block(inode, eh, depth, bh)      \
 485         __ext4_ext_check_block(__func__, __LINE__, inode, eh, depth, bh)
 486
 487 #ifdef EXT_DEBUG
 488 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 489 {
 490         int k, l = path->p_depth;
 491
 492         ext_debug("path:");
 493         for (k = 0; k <= l; k++, path++) {
 494                 if (path->p_idx) {
 495                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 496                             ext4_idx_pblock(path->p_idx));
 497                 } else if (path->p_ext) {
 498                         ext_debug("  %d:[%d]%d:%llu ",
 499                                   le32_to_cpu(path->p_ext->ee_block),
 500                                   ext4_ext_is_uninitialized(path->p_ext),
 501                                   ext4_ext_get_actual_len(path->p_ext),
 502                                   ext4_ext_pblock(path->p_ext));
 503                 } else
 504                         ext_debug("  []");
 505         }
 506         ext_debug("\n");
 507 }
 508
 509 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 510 {
 511         int depth = ext_depth(inode);
 512         struct ext4_extent_header *eh;
 513         struct ext4_extent *ex;
 514         int i;
 515
 516         if (!path)
 517                 return;
 518
 519         eh = path[depth].p_hdr;
 520         ex = EXT_FIRST_EXTENT(eh);
 521
 522         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 523
 524         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 525                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 526                           ext4_ext_is_uninitialized(ex),
 527                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 528         }
 529         ext_debug("\n");
 530 }
 531
 532 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 533                         ext4_fsblk_t newblock, int level)
 534 {
 535         int depth = ext_depth(inode);
 536         struct ext4_extent *ex;
 537
 538         if (depth != level) {
 539                 struct ext4_extent_idx *idx;
 540                 idx = path[level].p_idx;
 541                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 542                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 543                                         le32_to_cpu(idx->ei_block),
 544                                         ext4_idx_pblock(idx),
 545                                         newblock);
 546                         idx++;
 547                 }
 548
 549                 return;
 550         }
 551
 552         ex = path[depth].p_ext;
 553         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 554                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 555                                 le32_to_cpu(ex->ee_block),
 556                                 ext4_ext_pblock(ex),
 557                                 ext4_ext_is_uninitialized(ex),
 558                                 ext4_ext_get_actual_len(ex),
 559                                 newblock);
 560                 ex++;
 561         }
 562 }
 563
 564 #else
 565 #define ext4_ext_show_path(inode, path)
 566 #define ext4_ext_show_leaf(inode, path)
 567 #define ext4_ext_show_move(inode, path, newblock, level)
 568 #endif
 569
 570 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 571 {
 572         int depth = path->p_depth;
 573         int i;
 574
 575         for (i = 0; i <= depth; i++, path++)
 576                 if (path->p_bh) {
 577                         brelse(path->p_bh);
 578                         path->p_bh = NULL;
 579                 }
 580 }
 581
 582 /*
 583  * ext4_ext_binsearch_idx:
 584  * binary search for the closest index of the given block
 585  * the header must be checked before calling this
 586  */
 587 static void
 588 ext4_ext_binsearch_idx(struct inode *inode,
 589                         struct ext4_ext_path *path, ext4_lblk_t block)
 590 {
 591         struct ext4_extent_header *eh = path->p_hdr;
 592         struct ext4_extent_idx *r, *l, *m;
 593
 594
 595         ext_debug("binsearch for %u(idx):  ", block);
 596
 597         l = EXT_FIRST_INDEX(eh) + 1;
 598         r = EXT_LAST_INDEX(eh);
 599         while (l <= r) {
 600                 m = l + (r - l) / 2;
 601                 if (block < le32_to_cpu(m->ei_block))
 602                         r = m - 1;
 603                 else
 604                         l = m + 1;
 605                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 606                                 m, le32_to_cpu(m->ei_block),
 607                                 r, le32_to_cpu(r->ei_block));
 608         }
 609
 610         path->p_idx = l - 1;
 611         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 612                   ext4_idx_pblock(path->p_idx));
 613
 614 #ifdef CHECK_BINSEARCH
 615         {
 616                 struct ext4_extent_idx *chix, *ix;
 617                 int k;
 618
 619                 chix = ix = EXT_FIRST_INDEX(eh);
 620                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 621                   if (k != 0 &&
 622                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 623                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 624                                        "first=0x%p\n", k,
 625                                        ix, EXT_FIRST_INDEX(eh));
 626                                 printk(KERN_DEBUG "%u <= %u\n",
 627                                        le32_to_cpu(ix->ei_block),
 628                                        le32_to_cpu(ix[-1].ei_block));
 629                         }
 630                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 631                                            <= le32_to_cpu(ix[-1].ei_block));
 632                         if (block < le32_to_cpu(ix->ei_block))
 633                                 break;
 634                         chix = ix;
 635                 }
 636                 BUG_ON(chix != path->p_idx);
 637         }
 638 #endif
 639
 640 }
 641
 642 /*
 643  * ext4_ext_binsearch:
 644  * binary search for closest extent of the given block
 645  * the header must be checked before calling this
 646  */
 647 static void
 648 ext4_ext_binsearch(struct inode *inode,
 649                 struct ext4_ext_path *path, ext4_lblk_t block)
 650 {
 651         struct ext4_extent_header *eh = path->p_hdr;
 652         struct ext4_extent *r, *l, *m;
 653
 654         if (eh->eh_entries == 0) {
 655                 /*
 656                  * this leaf is empty:
 657                  * we get such a leaf in split/add case
 658                  */
 659                 return;
 660         }
 661
 662         ext_debug("binsearch for %u:  ", block);
 663
 664         l = EXT_FIRST_EXTENT(eh) + 1;
 665         r = EXT_LAST_EXTENT(eh);
 666
 667         while (l <= r) {
 668                 m = l + (r - l) / 2;
 669                 if (block < le32_to_cpu(m->ee_block))
 670                         r = m - 1;
 671                 else
 672                         l = m + 1;
 673                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 674                                 m, le32_to_cpu(m->ee_block),
 675                                 r, le32_to_cpu(r->ee_block));
 676         }
 677
 678         path->p_ext = l - 1;
 679         ext_debug("  -> %d:%llu:[%d]%d ",
 680                         le32_to_cpu(path->p_ext->ee_block),
 681                         ext4_ext_pblock(path->p_ext),
 682                         ext4_ext_is_uninitialized(path->p_ext),
 683                         ext4_ext_get_actual_len(path->p_ext));
 684
 685 #ifdef CHECK_BINSEARCH
 686         {
 687                 struct ext4_extent *chex, *ex;
 688                 int k;
 689
 690                 chex = ex = EXT_FIRST_EXTENT(eh);
 691                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 692                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 693                                           <= le32_to_cpu(ex[-1].ee_block));
 694                         if (block < le32_to_cpu(ex->ee_block))
 695                                 break;
 696                         chex = ex;
 697                 }
 698                 BUG_ON(chex != path->p_ext);
 699         }
 700 #endif
 701
 702 }
 703
 704 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 705 {
 706         struct ext4_extent_header *eh;
 707
 708         eh = ext_inode_hdr(inode);
 709         eh->eh_depth = 0;
 710         eh->eh_entries = 0;
 711         eh->eh_magic = EXT4_EXT_MAGIC;
 712         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 713         ext4_mark_inode_dirty(handle, inode);
 714         return 0;
 715 }
 716
 717 struct ext4_ext_path *
 718 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
 719                                         struct ext4_ext_path *path)
 720 {
 721         struct ext4_extent_header *eh;
 722         struct buffer_head *bh;
 723         short int depth, i, ppos = 0, alloc = 0;
 724         int ret;
 725
 726         eh = ext_inode_hdr(inode);
 727         depth = ext_depth(inode);
 728
 729         /* account possible depth increase */
 730         if (!path) {
 731                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 732                                 GFP_NOFS);
 733                 if (!path)
 734                         return ERR_PTR(-ENOMEM);
 735                 alloc = 1;
 736         }
 737         path[0].p_hdr = eh;
 738         path[0].p_bh = NULL;
 739
 740         i = depth;
 741         /* walk through the tree */
 742         while (i) {
 743                 ext_debug("depth %d: num %d, max %d\n",
 744                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 745
 746                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 747                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 748                 path[ppos].p_depth = i;
 749                 path[ppos].p_ext = NULL;
 750
 751                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
 752                 if (unlikely(!bh)) {
 753                         ret = -ENOMEM;
 754                         goto err;
 755                 }
 756                 if (!bh_uptodate_or_lock(bh)) {
 757                         trace_ext4_ext_load_extent(inode, block,
 758                                                 path[ppos].p_block);
 759                         ret = bh_submit_read(bh);
 760                         if (ret < 0) {
 761                                 put_bh(bh);
 762                                 goto err;
 763                         }
 764                 }
 765                 eh = ext_block_hdr(bh);
 766                 ppos++;
 767                 if (unlikely(ppos > depth)) {
 768                         put_bh(bh);
 769                         EXT4_ERROR_INODE(inode,
 770                                          "ppos %d > depth %d", ppos, depth);
 771                         ret = -EIO;
 772                         goto err;
 773                 }
 774                 path[ppos].p_bh = bh;
 775                 path[ppos].p_hdr = eh;
 776                 i--;
 777
 778                 ret = ext4_ext_check_block(inode, eh, i, bh);
 779                 if (ret < 0)
 780                         goto err;
 781         }
 782
 783         path[ppos].p_depth = i;
 784         path[ppos].p_ext = NULL;
 785         path[ppos].p_idx = NULL;
 786
 787         /* find extent */
 788         ext4_ext_binsearch(inode, path + ppos, block);
 789         /* if not an empty leaf */
 790         if (path[ppos].p_ext)
 791                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 792
 793         ext4_ext_show_path(inode, path);
 794
 795         return path;
 796
 797 err:
 798         ext4_ext_drop_refs(path);
 799         if (alloc)
 800                 kfree(path);
 801         return ERR_PTR(ret);
 802 }
 803
 804 /*
 805  * ext4_ext_insert_index:
 806  * insert new index [@logical;@ptr] into the block at @curp;
 807  * check where to insert: before @curp or after @curp
 808  */
 809 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 810                                  struct ext4_ext_path *curp,
 811                                  int logical, ext4_fsblk_t ptr)
 812 {
 813         struct ext4_extent_idx *ix;
 814         int len, err;
 815
 816         err = ext4_ext_get_access(handle, inode, curp);
 817         if (err)
 818                 return err;
 819
 820         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 821                 EXT4_ERROR_INODE(inode,
 822                                  "logical %d == ei_block %d!",
 823                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 824                 return -EIO;
 825         }
 826
 827         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 828                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 829                 EXT4_ERROR_INODE(inode,
 830                                  "eh_entries %d >= eh_max %d!",
 831                                  le16_to_cpu(curp->p_hdr->eh_entries),
 832                                  le16_to_cpu(curp->p_hdr->eh_max));
 833                 return -EIO;
 834         }
 835
 836         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 837                 /* insert after */
 838                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 839                 ix = curp->p_idx + 1;
 840         } else {
 841                 /* insert before */
 842                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 843                 ix = curp->p_idx;
 844         }
 845
 846         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 847         BUG_ON(len < 0);
 848         if (len > 0) {
 849                 ext_debug("insert new index %d: "
 850                                 "move %d indices from 0x%p to 0x%p\n",
 851                                 logical, len, ix, ix + 1);
 852                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 853         }
 854
 855         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 856                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 857                 return -EIO;
 858         }
 859
 860         ix->ei_block = cpu_to_le32(logical);
 861         ext4_idx_store_pblock(ix, ptr);
 862         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 863
 864         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 865                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 866                 return -EIO;
 867         }
 868
 869         err = ext4_ext_dirty(handle, inode, curp);
 870         ext4_std_error(inode->i_sb, err);
 871
 872         return err;
 873 }
 874
 875 /*
 876  * ext4_ext_split:
 877  * inserts new subtree into the path, using free index entry
 878  * at depth @at:
 879  * - allocates all needed blocks (new leaf and all intermediate index blocks)
 880  * - makes decision where to split
 881  * - moves remaining extents and index entries (right to the split point)
 882  *   into the newly allocated blocks
 883  * - initializes subtree
 884  */
 885 static int ext4_ext_split(handle_t *handle, struct inode *inode,
 886                           unsigned int flags,
 887                           struct ext4_ext_path *path,
 888                           struct ext4_extent *newext, int at)
 889 {
 890         struct buffer_head *bh = NULL;
 891         int depth = ext_depth(inode);
 892         struct ext4_extent_header *neh;
 893         struct ext4_extent_idx *fidx;
 894         int i = at, k, m, a;
 895         ext4_fsblk_t newblock, oldblock;
 896         __le32 border;
 897         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
 898         int err = 0;
 899
 900         /* make decision: where to split? */
 901         /* FIXME: now decision is simplest: at current extent */
 902
 903         /* if current leaf will be split, then we should use
 904          * border from split point */
 905         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
 906                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
 907                 return -EIO;
 908         }
 909         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
 910                 border = path[depth].p_ext[1].ee_block;
 911                 ext_debug("leaf will be split."
 912                                 " next leaf starts at %d\n",
 913                                   le32_to_cpu(border));
 914         } else {
 915                 border = newext->ee_block;
 916                 ext_debug("leaf will be added."
 917                                 " next leaf starts at %d\n",
 918                                 le32_to_cpu(border));
 919         }
 920
 921         /*
 922          * If error occurs, then we break processing
 923          * and mark filesystem read-only. index won't
 924          * be inserted and tree will be in consistent
 925          * state. Next mount will repair buffers too.
 926          */
 927
 928         /*
 929          * Get array to track all allocated blocks.
 930          * We need this to handle errors and free blocks
 931          * upon them.
 932          */
 933         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
 934         if (!ablocks)
 935                 return -ENOMEM;
 936
 937         /* allocate all needed blocks */
 938         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
 939         for (a = 0; a < depth - at; a++) {
 940                 newblock = ext4_ext_new_meta_block(handle, inode, path,
 941                                                    newext, &err, flags);
 942                 if (newblock == 0)
 943                         goto cleanup;
 944                 ablocks[a] = newblock;
 945         }
 946
 947         /* initialize new leaf */
 948         newblock = ablocks[--a];
 949         if (unlikely(newblock == 0)) {
 950                 EXT4_ERROR_INODE(inode, "newblock == 0!");
 951                 err = -EIO;
 952                 goto cleanup;
 953         }
 954         bh = sb_getblk(inode->i_sb, newblock);
 955         if (unlikely(!bh)) {
 956                 err = -ENOMEM;
 957                 goto cleanup;
 958         }
 959         lock_buffer(bh);
 960
 961         err = ext4_journal_get_create_access(handle, bh);
 962         if (err)
 963                 goto cleanup;
 964
 965         neh = ext_block_hdr(bh);
 966         neh->eh_entries = 0;
 967         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
 968         neh->eh_magic = EXT4_EXT_MAGIC;
 969         neh->eh_depth = 0;
 970
 971         /* move remainder of path[depth] to the new leaf */
 972         if (unlikely(path[depth].p_hdr->eh_entries !=
 973                      path[depth].p_hdr->eh_max)) {
 974                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
 975                                  path[depth].p_hdr->eh_entries,
 976                                  path[depth].p_hdr->eh_max);
 977                 err = -EIO;
 978                 goto cleanup;
 979         }
 980         /* start copy from next extent */
 981         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
 982         ext4_ext_show_move(inode, path, newblock, depth);
 983         if (m) {
 984                 struct ext4_extent *ex;
 985                 ex = EXT_FIRST_EXTENT(neh);
 986                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
 987                 le16_add_cpu(&neh->eh_entries, m);
 988         }
 989
 990         ext4_extent_block_csum_set(inode, neh);
 991         set_buffer_uptodate(bh);
 992         unlock_buffer(bh);
 993
 994         err = ext4_handle_dirty_metadata(handle, inode, bh);
 995         if (err)
 996                 goto cleanup;
 997         brelse(bh);
 998         bh = NULL;
 999
1000         /* correct old leaf */
1001         if (m) {
1002                 err = ext4_ext_get_access(handle, inode, path + depth);
1003                 if (err)
1004                         goto cleanup;
1005                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1006                 err = ext4_ext_dirty(handle, inode, path + depth);
1007                 if (err)
1008                         goto cleanup;
1009
1010         }
1011
1012         /* create intermediate indexes */
1013         k = depth - at - 1;
1014         if (unlikely(k < 0)) {
1015                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1016                 err = -EIO;
1017                 goto cleanup;
1018         }
1019         if (k)
1020                 ext_debug("create %d intermediate indices\n", k);
1021         /* insert new index into current index block */
1022         /* current depth stored in i var */
1023         i = depth - 1;
1024         while (k--) {
1025                 oldblock = newblock;
1026                 newblock = ablocks[--a];
1027                 bh = sb_getblk(inode->i_sb, newblock);
1028                 if (unlikely(!bh)) {
1029                         err = -ENOMEM;
1030                         goto cleanup;
1031                 }
1032                 lock_buffer(bh);
1033
1034                 err = ext4_journal_get_create_access(handle, bh);
1035                 if (err)
1036                         goto cleanup;
1037
1038                 neh = ext_block_hdr(bh);
1039                 neh->eh_entries = cpu_to_le16(1);
1040                 neh->eh_magic = EXT4_EXT_MAGIC;
1041                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1042                 neh->eh_depth = cpu_to_le16(depth - i);
1043                 fidx = EXT_FIRST_INDEX(neh);
1044                 fidx->ei_block = border;
1045                 ext4_idx_store_pblock(fidx, oldblock);
1046
1047                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1048                                 i, newblock, le32_to_cpu(border), oldblock);
1049
1050                 /* move remainder of path[i] to the new index block */
1051                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1052                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1053                         EXT4_ERROR_INODE(inode,
1054                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1055                                          le32_to_cpu(path[i].p_ext->ee_block));
1056                         err = -EIO;
1057                         goto cleanup;
1058                 }
1059                 /* start copy indexes */
1060                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1061                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1062                                 EXT_MAX_INDEX(path[i].p_hdr));
1063                 ext4_ext_show_move(inode, path, newblock, i);
1064                 if (m) {
1065                         memmove(++fidx, path[i].p_idx,
1066                                 sizeof(struct ext4_extent_idx) * m);
1067                         le16_add_cpu(&neh->eh_entries, m);
1068                 }
1069                 ext4_extent_block_csum_set(inode, neh);
1070                 set_buffer_uptodate(bh);
1071                 unlock_buffer(bh);
1072
1073                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1074                 if (err)
1075                         goto cleanup;
1076                 brelse(bh);
1077                 bh = NULL;
1078
1079                 /* correct old index */
1080                 if (m) {
1081                         err = ext4_ext_get_access(handle, inode, path + i);
1082                         if (err)
1083                                 goto cleanup;
1084                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1085                         err = ext4_ext_dirty(handle, inode, path + i);
1086                         if (err)
1087                                 goto cleanup;
1088                 }
1089
1090                 i--;
1091         }
1092
1093         /* insert new index */
1094         err = ext4_ext_insert_index(handle, inode, path + at,
1095                                     le32_to_cpu(border), newblock);
1096
1097 cleanup:
1098         if (bh) {
1099                 if (buffer_locked(bh))
1100                         unlock_buffer(bh);
1101                 brelse(bh);
1102         }
1103
1104         if (err) {
1105                 /* free all allocated blocks in error case */
1106                 for (i = 0; i < depth; i++) {
1107                         if (!ablocks[i])
1108                                 continue;
1109                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1110                                          EXT4_FREE_BLOCKS_METADATA);
1111                 }
1112         }
1113         kfree(ablocks);
1114
1115         return err;
1116 }
1117
1118 /*
1119  * ext4_ext_grow_indepth:
1120  * implements tree growing procedure:
1121  * - allocates new block
1122  * - moves top-level data (index block or leaf) into the new block
1123  * - initializes new top-level, creating index that points to the
1124  *   just created block
1125  */
1126 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1127                                  unsigned int flags,
1128                                  struct ext4_extent *newext)
1129 {
1130         struct ext4_extent_header *neh;
1131         struct buffer_head *bh;
1132         ext4_fsblk_t newblock;
1133         int err = 0;
1134
1135         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1136                 newext, &err, flags);
1137         if (newblock == 0)
1138                 return err;
1139
1140         bh = sb_getblk(inode->i_sb, newblock);
1141         if (unlikely(!bh))
1142                 return -ENOMEM;
1143         lock_buffer(bh);
1144
1145         err = ext4_journal_get_create_access(handle, bh);
1146         if (err) {
1147                 unlock_buffer(bh);
1148                 goto out;
1149         }
1150
1151         /* move top-level index/leaf into new block */
1152         memmove(bh->b_data, EXT4_I(inode)->i_data,
1153                 sizeof(EXT4_I(inode)->i_data));
1154
1155         /* set size of new block */
1156         neh = ext_block_hdr(bh);
1157         /* old root could have indexes or leaves
1158          * so calculate e_max right way */
1159         if (ext_depth(inode))
1160                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1161         else
1162                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1163         neh->eh_magic = EXT4_EXT_MAGIC;
1164         ext4_extent_block_csum_set(inode, neh);
1165         set_buffer_uptodate(bh);
1166         unlock_buffer(bh);
1167
1168         err = ext4_handle_dirty_metadata(handle, inode, bh);
1169         if (err)
1170                 goto out;
1171
1172         /* Update top-level index: num,max,pointer */
1173         neh = ext_inode_hdr(inode);
1174         neh->eh_entries = cpu_to_le16(1);
1175         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1176         if (neh->eh_depth == 0) {
1177                 /* Root extent block becomes index block */
1178                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1179                 EXT_FIRST_INDEX(neh)->ei_block =
1180                         EXT_FIRST_EXTENT(neh)->ee_block;
1181         }
1182         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1183                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1184                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1185                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1186
1187         le16_add_cpu(&neh->eh_depth, 1);
1188         ext4_mark_inode_dirty(handle, inode);
1189 out:
1190         brelse(bh);
1191
1192         return err;
1193 }
1194
1195 /*
1196  * ext4_ext_create_new_leaf:
1197  * finds empty index and adds new leaf.
1198  * if no free index is found, then it requests in-depth growing.
1199  */
1200 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1201                                     unsigned int flags,
1202                                     struct ext4_ext_path *path,
1203                                     struct ext4_extent *newext)
1204 {
1205         struct ext4_ext_path *curp;
1206         int depth, i, err = 0;
1207
1208 repeat:
1209         i = depth = ext_depth(inode);
1210
1211         /* walk up to the tree and look for free index entry */
1212         curp = path + depth;
1213         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1214                 i--;
1215                 curp--;
1216         }
1217
1218         /* we use already allocated block for index block,
1219          * so subsequent data blocks should be contiguous */
1220         if (EXT_HAS_FREE_INDEX(curp)) {
1221                 /* if we found index with free entry, then use that
1222                  * entry: create all needed subtree and add new leaf */
1223                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1224                 if (err)
1225                         goto out;
1226
1227                 /* refill path */
1228                 ext4_ext_drop_refs(path);
1229                 path = ext4_ext_find_extent(inode,
1230                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1231                                     path);
1232                 if (IS_ERR(path))
1233                         err = PTR_ERR(path);
1234         } else {
1235                 /* tree is full, time to grow in depth */
1236                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1237                 if (err)
1238                         goto out;
1239
1240                 /* refill path */
1241                 ext4_ext_drop_refs(path);
1242                 path = ext4_ext_find_extent(inode,
1243                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1244                                     path);
1245                 if (IS_ERR(path)) {
1246                         err = PTR_ERR(path);
1247                         goto out;
1248                 }
1249
1250                 /*
1251                  * only first (depth 0 -> 1) produces free space;
1252                  * in all other cases we have to split the grown tree
1253                  */
1254                 depth = ext_depth(inode);
1255                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1256                         /* now we need to split */
1257                         goto repeat;
1258                 }
1259         }
1260
1261 out:
1262         return err;
1263 }
1264
1265 /*
1266  * search the closest allocated block to the left for *logical
1267  * and returns it at @logical + it's physical address at @phys
1268  * if *logical is the smallest allocated block, the function
1269  * returns 0 at @phys
1270  * return value contains 0 (success) or error code
1271  */
1272 static int ext4_ext_search_left(struct inode *inode,
1273                                 struct ext4_ext_path *path,
1274                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1275 {
1276         struct ext4_extent_idx *ix;
1277         struct ext4_extent *ex;
1278         int depth, ee_len;
1279
1280         if (unlikely(path == NULL)) {
1281                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1282                 return -EIO;
1283         }
1284         depth = path->p_depth;
1285         *phys = 0;
1286
1287         if (depth == 0 && path->p_ext == NULL)
1288                 return 0;
1289
1290         /* usually extent in the path covers blocks smaller
1291          * then *logical, but it can be that extent is the
1292          * first one in the file */
1293
1294         ex = path[depth].p_ext;
1295         ee_len = ext4_ext_get_actual_len(ex);
1296         if (*logical < le32_to_cpu(ex->ee_block)) {
1297                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1298                         EXT4_ERROR_INODE(inode,
1299                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1300                                          *logical, le32_to_cpu(ex->ee_block));
1301                         return -EIO;
1302                 }
1303                 while (--depth >= 0) {
1304                         ix = path[depth].p_idx;
1305                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1306                                 EXT4_ERROR_INODE(inode,
1307                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1308                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1309                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1310                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1311                                   depth);
1312                                 return -EIO;
1313                         }
1314                 }
1315                 return 0;
1316         }
1317
1318         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1319                 EXT4_ERROR_INODE(inode,
1320                                  "logical %d < ee_block %d + ee_len %d!",
1321                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1322                 return -EIO;
1323         }
1324
1325         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1326         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1327         return 0;
1328 }
1329
1330 /*
1331  * search the closest allocated block to the right for *logical
1332  * and returns it at @logical + it's physical address at @phys
1333  * if *logical is the largest allocated block, the function
1334  * returns 0 at @phys
1335  * return value contains 0 (success) or error code
1336  */
1337 static int ext4_ext_search_right(struct inode *inode,
1338                                  struct ext4_ext_path *path,
1339                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1340                                  struct ext4_extent **ret_ex)
1341 {
1342         struct buffer_head *bh = NULL;
1343         struct ext4_extent_header *eh;
1344         struct ext4_extent_idx *ix;
1345         struct ext4_extent *ex;
1346         ext4_fsblk_t block;
1347         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1348         int ee_len;
1349
1350         if (unlikely(path == NULL)) {
1351                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1352                 return -EIO;
1353         }
1354         depth = path->p_depth;
1355         *phys = 0;
1356
1357         if (depth == 0 && path->p_ext == NULL)
1358                 return 0;
1359
1360         /* usually extent in the path covers blocks smaller
1361          * then *logical, but it can be that extent is the
1362          * first one in the file */
1363
1364         ex = path[depth].p_ext;
1365         ee_len = ext4_ext_get_actual_len(ex);
1366         if (*logical < le32_to_cpu(ex->ee_block)) {
1367                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1368                         EXT4_ERROR_INODE(inode,
1369                                          "first_extent(path[%d].p_hdr) != ex",
1370                                          depth);
1371                         return -EIO;
1372                 }
1373                 while (--depth >= 0) {
1374                         ix = path[depth].p_idx;
1375                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1376                                 EXT4_ERROR_INODE(inode,
1377                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1378                                                  *logical);
1379                                 return -EIO;
1380                         }
1381                 }
1382                 goto found_extent;
1383         }
1384
1385         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1386                 EXT4_ERROR_INODE(inode,
1387                                  "logical %d < ee_block %d + ee_len %d!",
1388                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1389                 return -EIO;
1390         }
1391
1392         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1393                 /* next allocated block in this leaf */
1394                 ex++;
1395                 goto found_extent;
1396         }
1397
1398         /* go up and search for index to the right */
1399         while (--depth >= 0) {
1400                 ix = path[depth].p_idx;
1401                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1402                         goto got_index;
1403         }
1404
1405         /* we've gone up to the root and found no index to the right */
1406         return 0;
1407
1408 got_index:
1409         /* we've found index to the right, let's
1410          * follow it and find the closest allocated
1411          * block to the right */
1412         ix++;
1413         block = ext4_idx_pblock(ix);
1414         while (++depth < path->p_depth) {
1415                 bh = sb_bread(inode->i_sb, block);
1416                 if (bh == NULL)
1417                         return -EIO;
1418                 eh = ext_block_hdr(bh);
1419                 /* subtract from p_depth to get proper eh_depth */
1420                 if (ext4_ext_check_block(inode, eh,
1421                                          path->p_depth - depth, bh)) {
1422                         put_bh(bh);
1423                         return -EIO;
1424                 }
1425                 ix = EXT_FIRST_INDEX(eh);
1426                 block = ext4_idx_pblock(ix);
1427                 put_bh(bh);
1428         }
1429
1430         bh = sb_bread(inode->i_sb, block);
1431         if (bh == NULL)
1432                 return -EIO;
1433         eh = ext_block_hdr(bh);
1434         if (ext4_ext_check_block(inode, eh, path->p_depth - depth, bh)) {
1435                 put_bh(bh);
1436                 return -EIO;
1437         }
1438         ex = EXT_FIRST_EXTENT(eh);
1439 found_extent:
1440         *logical = le32_to_cpu(ex->ee_block);
1441         *phys = ext4_ext_pblock(ex);
1442         *ret_ex = ex;
1443         if (bh)
1444                 put_bh(bh);
1445         return 0;
1446 }
1447
1448 /*
1449  * ext4_ext_next_allocated_block:
1450  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1451  * NOTE: it considers block number from index entry as
1452  * allocated block. Thus, index entries have to be consistent
1453  * with leaves.
1454  */
1455 static ext4_lblk_t
1456 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1457 {
1458         int depth;
1459
1460         BUG_ON(path == NULL);
1461         depth = path->p_depth;
1462
1463         if (depth == 0 && path->p_ext == NULL)
1464                 return EXT_MAX_BLOCKS;
1465
1466         while (depth >= 0) {
1467                 if (depth == path->p_depth) {
1468                         /* leaf */
1469                         if (path[depth].p_ext &&
1470                                 path[depth].p_ext !=
1471                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1472                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1473                 } else {
1474                         /* index */
1475                         if (path[depth].p_idx !=
1476                                         EXT_LAST_INDEX(path[depth].p_hdr))
1477                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1478                 }
1479                 depth--;
1480         }
1481
1482         return EXT_MAX_BLOCKS;
1483 }
1484
1485 /*
1486  * ext4_ext_next_leaf_block:
1487  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1488  */
1489 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1490 {
1491         int depth;
1492
1493         BUG_ON(path == NULL);
1494         depth = path->p_depth;
1495
1496         /* zero-tree has no leaf blocks at all */
1497         if (depth == 0)
1498                 return EXT_MAX_BLOCKS;
1499
1500         /* go to index block */
1501         depth--;
1502
1503         while (depth >= 0) {
1504                 if (path[depth].p_idx !=
1505                                 EXT_LAST_INDEX(path[depth].p_hdr))
1506                         return (ext4_lblk_t)
1507                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1508                 depth--;
1509         }
1510
1511         return EXT_MAX_BLOCKS;
1512 }
1513
1514 /*
1515  * ext4_ext_correct_indexes:
1516  * if leaf gets modified and modified extent is first in the leaf,
1517  * then we have to correct all indexes above.
1518  * TODO: do we need to correct tree in all cases?
1519  */
1520 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1521                                 struct ext4_ext_path *path)
1522 {
1523         struct ext4_extent_header *eh;
1524         int depth = ext_depth(inode);
1525         struct ext4_extent *ex;
1526         __le32 border;
1527         int k, err = 0;
1528
1529         eh = path[depth].p_hdr;
1530         ex = path[depth].p_ext;
1531
1532         if (unlikely(ex == NULL || eh == NULL)) {
1533                 EXT4_ERROR_INODE(inode,
1534                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1535                 return -EIO;
1536         }
1537
1538         if (depth == 0) {
1539                 /* there is no tree at all */
1540                 return 0;
1541         }
1542
1543         if (ex != EXT_FIRST_EXTENT(eh)) {
1544                 /* we correct tree if first leaf got modified only */
1545                 return 0;
1546         }
1547
1548         /*
1549          * TODO: we need correction if border is smaller than current one
1550          */
1551         k = depth - 1;
1552         border = path[depth].p_ext->ee_block;
1553         err = ext4_ext_get_access(handle, inode, path + k);
1554         if (err)
1555                 return err;
1556         path[k].p_idx->ei_block = border;
1557         err = ext4_ext_dirty(handle, inode, path + k);
1558         if (err)
1559                 return err;
1560
1561         while (k--) {
1562                 /* change all left-side indexes */
1563                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1564                         break;
1565                 err = ext4_ext_get_access(handle, inode, path + k);
1566                 if (err)
1567                         break;
1568                 path[k].p_idx->ei_block = border;
1569                 err = ext4_ext_dirty(handle, inode, path + k);
1570                 if (err)
1571                         break;
1572         }
1573
1574         return err;
1575 }
1576
1577 int
1578 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1579                                 struct ext4_extent *ex2)
1580 {
1581         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1582
1583         /*
1584          * Make sure that both extents are initialized. We don't merge
1585          * uninitialized extents so that we can be sure that end_io code has
1586          * the extent that was written properly split out and conversion to
1587          * initialized is trivial.
1588          */
1589         if (ext4_ext_is_uninitialized(ex1) || ext4_ext_is_uninitialized(ex2))
1590                 return 0;
1591
1592         if (ext4_ext_is_uninitialized(ex1))
1593                 max_len = EXT_UNINIT_MAX_LEN;
1594         else
1595                 max_len = EXT_INIT_MAX_LEN;
1596
1597         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1598         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1599
1600         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1601                         le32_to_cpu(ex2->ee_block))
1602                 return 0;
1603
1604         /*
1605          * To allow future support for preallocated extents to be added
1606          * as an RO_COMPAT feature, refuse to merge to extents if
1607          * this can result in the top bit of ee_len being set.
1608          */
1609         if (ext1_ee_len + ext2_ee_len > max_len)
1610                 return 0;
1611 #ifdef AGGRESSIVE_TEST
1612         if (ext1_ee_len >= 4)
1613                 return 0;
1614 #endif
1615
1616         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1617                 return 1;
1618         return 0;
1619 }
1620
1621 /*
1622  * This function tries to merge the "ex" extent to the next extent in the tree.
1623  * It always tries to merge towards right. If you want to merge towards
1624  * left, pass "ex - 1" as argument instead of "ex".
1625  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1626  * 1 if they got merged.
1627  */
1628 static int ext4_ext_try_to_merge_right(struct inode *inode,
1629                                  struct ext4_ext_path *path,
1630                                  struct ext4_extent *ex)
1631 {
1632         struct ext4_extent_header *eh;
1633         unsigned int depth, len;
1634         int merge_done = 0;
1635         int uninitialized = 0;
1636
1637         depth = ext_depth(inode);
1638         BUG_ON(path[depth].p_hdr == NULL);
1639         eh = path[depth].p_hdr;
1640
1641         while (ex < EXT_LAST_EXTENT(eh)) {
1642                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1643                         break;
1644                 /* merge with next extent! */
1645                 if (ext4_ext_is_uninitialized(ex))
1646                         uninitialized = 1;
1647                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1648                                 + ext4_ext_get_actual_len(ex + 1));
1649                 if (uninitialized)
1650                         ext4_ext_mark_uninitialized(ex);
1651
1652                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1653                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1654                                 * sizeof(struct ext4_extent);
1655                         memmove(ex + 1, ex + 2, len);
1656                 }
1657                 le16_add_cpu(&eh->eh_entries, -1);
1658                 merge_done = 1;
1659                 WARN_ON(eh->eh_entries == 0);
1660                 if (!eh->eh_entries)
1661                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1662         }
1663
1664         return merge_done;
1665 }
1666
1667 /*
1668  * This function does a very simple check to see if we can collapse
1669  * an extent tree with a single extent tree leaf block into the inode.
1670  */
1671 static void ext4_ext_try_to_merge_up(handle_t *handle,
1672                                      struct inode *inode,
1673                                      struct ext4_ext_path *path)
1674 {
1675         size_t s;
1676         unsigned max_root = ext4_ext_space_root(inode, 0);
1677         ext4_fsblk_t blk;
1678
1679         if ((path[0].p_depth != 1) ||
1680             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1681             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1682                 return;
1683
1684         /*
1685          * We need to modify the block allocation bitmap and the block
1686          * group descriptor to release the extent tree block.  If we
1687          * can't get the journal credits, give up.
1688          */
1689         if (ext4_journal_extend(handle, 2))
1690                 return;
1691
1692         /*
1693          * Copy the extent data up to the inode
1694          */
1695         blk = ext4_idx_pblock(path[0].p_idx);
1696         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1697                 sizeof(struct ext4_extent_idx);
1698         s += sizeof(struct ext4_extent_header);
1699
1700         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1701         path[0].p_depth = 0;
1702         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1703                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1704         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1705
1706         brelse(path[1].p_bh);
1707         ext4_free_blocks(handle, inode, NULL, blk, 1,
1708                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1709 }
1710
1711 /*
1712  * This function tries to merge the @ex extent to neighbours in the tree.
1713  * return 1 if merge left else 0.
1714  */
1715 static void ext4_ext_try_to_merge(handle_t *handle,
1716                                   struct inode *inode,
1717                                   struct ext4_ext_path *path,
1718                                   struct ext4_extent *ex) {
1719         struct ext4_extent_header *eh;
1720         unsigned int depth;
1721         int merge_done = 0;
1722
1723         depth = ext_depth(inode);
1724         BUG_ON(path[depth].p_hdr == NULL);
1725         eh = path[depth].p_hdr;
1726
1727         if (ex > EXT_FIRST_EXTENT(eh))
1728                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1729
1730         if (!merge_done)
1731                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1732
1733         ext4_ext_try_to_merge_up(handle, inode, path);
1734 }
1735
1736 /*
1737  * check if a portion of the "newext" extent overlaps with an
1738  * existing extent.
1739  *
1740  * If there is an overlap discovered, it updates the length of the newext
1741  * such that there will be no overlap, and then returns 1.
1742  * If there is no overlap found, it returns 0.
1743  */
1744 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1745                                            struct inode *inode,
1746                                            struct ext4_extent *newext,
1747                                            struct ext4_ext_path *path)
1748 {
1749         ext4_lblk_t b1, b2;
1750         unsigned int depth, len1;
1751         unsigned int ret = 0;
1752
1753         b1 = le32_to_cpu(newext->ee_block);
1754         len1 = ext4_ext_get_actual_len(newext);
1755         depth = ext_depth(inode);
1756         if (!path[depth].p_ext)
1757                 goto out;
1758         b2 = le32_to_cpu(path[depth].p_ext->ee_block);
1759         b2 &= ~(sbi->s_cluster_ratio - 1);
1760
1761         /*
1762          * get the next allocated block if the extent in the path
1763          * is before the requested block(s)
1764          */
1765         if (b2 < b1) {
1766                 b2 = ext4_ext_next_allocated_block(path);
1767                 if (b2 == EXT_MAX_BLOCKS)
1768                         goto out;
1769                 b2 &= ~(sbi->s_cluster_ratio - 1);
1770         }
1771
1772         /* check for wrap through zero on extent logical start block*/
1773         if (b1 + len1 < b1) {
1774                 len1 = EXT_MAX_BLOCKS - b1;
1775                 newext->ee_len = cpu_to_le16(len1);
1776                 ret = 1;
1777         }
1778
1779         /* check for overlap */
1780         if (b1 + len1 > b2) {
1781                 newext->ee_len = cpu_to_le16(b2 - b1);
1782                 ret = 1;
1783         }
1784 out:
1785         return ret;
1786 }
1787
1788 /*
1789  * ext4_ext_insert_extent:
1790  * tries to merge requsted extent into the existing extent or
1791  * inserts requested extent as new one into the tree,
1792  * creating new leaf in the no-space case.
1793  */
1794 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1795                                 struct ext4_ext_path *path,
1796                                 struct ext4_extent *newext, int flag)
1797 {
1798         struct ext4_extent_header *eh;
1799         struct ext4_extent *ex, *fex;
1800         struct ext4_extent *nearex; /* nearest extent */
1801         struct ext4_ext_path *npath = NULL;
1802         int depth, len, err;
1803         ext4_lblk_t next;
1804         unsigned uninitialized = 0;
1805         int flags = 0;
1806
1807         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1808                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1809                 return -EIO;
1810         }
1811         depth = ext_depth(inode);
1812         ex = path[depth].p_ext;
1813         eh = path[depth].p_hdr;
1814         if (unlikely(path[depth].p_hdr == NULL)) {
1815                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1816                 return -EIO;
1817         }
1818
1819         /* try to insert block into found extent and return */
1820         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)) {
1821
1822                 /*
1823                  * Try to see whether we should rather test the extent on
1824                  * right from ex, or from the left of ex. This is because
1825                  * ext4_ext_find_extent() can return either extent on the
1826                  * left, or on the right from the searched position. This
1827                  * will make merging more effective.
1828                  */
1829                 if (ex < EXT_LAST_EXTENT(eh) &&
1830                     (le32_to_cpu(ex->ee_block) +
1831                     ext4_ext_get_actual_len(ex) <
1832                     le32_to_cpu(newext->ee_block))) {
1833                         ex += 1;
1834                         goto prepend;
1835                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1836                            (le32_to_cpu(newext->ee_block) +
1837                            ext4_ext_get_actual_len(newext) <
1838                            le32_to_cpu(ex->ee_block)))
1839                         ex -= 1;
1840
1841                 /* Try to append newex to the ex */
1842                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1843                         ext_debug("append [%d]%d block to %u:[%d]%d"
1844                                   "(from %llu)\n",
1845                                   ext4_ext_is_uninitialized(newext),
1846                                   ext4_ext_get_actual_len(newext),
1847                                   le32_to_cpu(ex->ee_block),
1848                                   ext4_ext_is_uninitialized(ex),
1849                                   ext4_ext_get_actual_len(ex),
1850                                   ext4_ext_pblock(ex));
1851                         err = ext4_ext_get_access(handle, inode,
1852                                                   path + depth);
1853                         if (err)
1854                                 return err;
1855
1856                         /*
1857                          * ext4_can_extents_be_merged should have checked
1858                          * that either both extents are uninitialized, or
1859                          * both aren't. Thus we need to check only one of
1860                          * them here.
1861                          */
1862                         if (ext4_ext_is_uninitialized(ex))
1863                                 uninitialized = 1;
1864                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1865                                         + ext4_ext_get_actual_len(newext));
1866                         if (uninitialized)
1867                                 ext4_ext_mark_uninitialized(ex);
1868                         eh = path[depth].p_hdr;
1869                         nearex = ex;
1870                         goto merge;
1871                 }
1872
1873 prepend:
1874                 /* Try to prepend newex to the ex */
1875                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
1876                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
1877                                   "(from %llu)\n",
1878                                   le32_to_cpu(newext->ee_block),
1879                                   ext4_ext_is_uninitialized(newext),
1880                                   ext4_ext_get_actual_len(newext),
1881                                   le32_to_cpu(ex->ee_block),
1882                                   ext4_ext_is_uninitialized(ex),
1883                                   ext4_ext_get_actual_len(ex),
1884                                   ext4_ext_pblock(ex));
1885                         err = ext4_ext_get_access(handle, inode,
1886                                                   path + depth);
1887                         if (err)
1888                                 return err;
1889
1890                         /*
1891                          * ext4_can_extents_be_merged should have checked
1892                          * that either both extents are uninitialized, or
1893                          * both aren't. Thus we need to check only one of
1894                          * them here.
1895                          */
1896                         if (ext4_ext_is_uninitialized(ex))
1897                                 uninitialized = 1;
1898                         ex->ee_block = newext->ee_block;
1899                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
1900                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1901                                         + ext4_ext_get_actual_len(newext));
1902                         if (uninitialized)
1903                                 ext4_ext_mark_uninitialized(ex);
1904                         eh = path[depth].p_hdr;
1905                         nearex = ex;
1906                         goto merge;
1907                 }
1908         }
1909
1910         depth = ext_depth(inode);
1911         eh = path[depth].p_hdr;
1912         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1913                 goto has_space;
1914
1915         /* probably next leaf has space for us? */
1916         fex = EXT_LAST_EXTENT(eh);
1917         next = EXT_MAX_BLOCKS;
1918         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1919                 next = ext4_ext_next_leaf_block(path);
1920         if (next != EXT_MAX_BLOCKS) {
1921                 ext_debug("next leaf block - %u\n", next);
1922                 BUG_ON(npath != NULL);
1923                 npath = ext4_ext_find_extent(inode, next, NULL);
1924                 if (IS_ERR(npath))
1925                         return PTR_ERR(npath);
1926                 BUG_ON(npath->p_depth != path->p_depth);
1927                 eh = npath[depth].p_hdr;
1928                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1929                         ext_debug("next leaf isn't full(%d)\n",
1930                                   le16_to_cpu(eh->eh_entries));
1931                         path = npath;
1932                         goto has_space;
1933                 }
1934                 ext_debug("next leaf has no free space(%d,%d)\n",
1935                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1936         }
1937
1938         /*
1939          * There is no free space in the found leaf.
1940          * We're gonna add a new leaf in the tree.
1941          */
1942         if (flag & EXT4_GET_BLOCKS_METADATA_NOFAIL)
1943                 flags = EXT4_MB_USE_RESERVED;
1944         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1945         if (err)
1946                 goto cleanup;
1947         depth = ext_depth(inode);
1948         eh = path[depth].p_hdr;
1949
1950 has_space:
1951         nearex = path[depth].p_ext;
1952
1953         err = ext4_ext_get_access(handle, inode, path + depth);
1954         if (err)
1955                 goto cleanup;
1956
1957         if (!nearex) {
1958                 /* there is no extent in this leaf, create first one */
1959                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1960                                 le32_to_cpu(newext->ee_block),
1961                                 ext4_ext_pblock(newext),
1962                                 ext4_ext_is_uninitialized(newext),
1963                                 ext4_ext_get_actual_len(newext));
1964                 nearex = EXT_FIRST_EXTENT(eh);
1965         } else {
1966                 if (le32_to_cpu(newext->ee_block)
1967                            > le32_to_cpu(nearex->ee_block)) {
1968                         /* Insert after */
1969                         ext_debug("insert %u:%llu:[%d]%d before: "
1970                                         "nearest %p\n",
1971                                         le32_to_cpu(newext->ee_block),
1972                                         ext4_ext_pblock(newext),
1973                                         ext4_ext_is_uninitialized(newext),
1974                                         ext4_ext_get_actual_len(newext),
1975                                         nearex);
1976                         nearex++;
1977                 } else {
1978                         /* Insert before */
1979                         BUG_ON(newext->ee_block == nearex->ee_block);
1980                         ext_debug("insert %u:%llu:[%d]%d after: "
1981                                         "nearest %p\n",
1982                                         le32_to_cpu(newext->ee_block),
1983                                         ext4_ext_pblock(newext),
1984                                         ext4_ext_is_uninitialized(newext),
1985                                         ext4_ext_get_actual_len(newext),
1986                                         nearex);
1987                 }
1988                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1989                 if (len > 0) {
1990                         ext_debug("insert %u:%llu:[%d]%d: "
1991                                         "move %d extents from 0x%p to 0x%p\n",
1992                                         le32_to_cpu(newext->ee_block),
1993                                         ext4_ext_pblock(newext),
1994                                         ext4_ext_is_uninitialized(newext),
1995                                         ext4_ext_get_actual_len(newext),
1996                                         len, nearex, nearex + 1);
1997                         memmove(nearex + 1, nearex,
1998                                 len * sizeof(struct ext4_extent));
1999                 }
2000         }
2001
2002         le16_add_cpu(&eh->eh_entries, 1);
2003         path[depth].p_ext = nearex;
2004         nearex->ee_block = newext->ee_block;
2005         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2006         nearex->ee_len = newext->ee_len;
2007
2008 merge:
2009         /* try to merge extents */
2010         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
2011                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2012
2013
2014         /* time to correct all indexes above */
2015         err = ext4_ext_correct_indexes(handle, inode, path);
2016         if (err)
2017                 goto cleanup;
2018
2019         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2020
2021 cleanup:
2022         if (npath) {
2023                 ext4_ext_drop_refs(npath);
2024                 kfree(npath);
2025         }
2026         return err;
2027 }
2028
2029 static int ext4_fill_fiemap_extents(struct inode *inode,
2030                                     ext4_lblk_t block, ext4_lblk_t num,
2031                                     struct fiemap_extent_info *fieinfo)
2032 {
2033         struct ext4_ext_path *path = NULL;
2034         struct ext4_extent *ex;
2035         struct extent_status es;
2036         ext4_lblk_t next, next_del, start = 0, end = 0;
2037         ext4_lblk_t last = block + num;
2038         int exists, depth = 0, err = 0;
2039         unsigned int flags = 0;
2040         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2041
2042         while (block < last && block != EXT_MAX_BLOCKS) {
2043                 num = last - block;
2044                 /* find extent for this block */
2045                 down_read(&EXT4_I(inode)->i_data_sem);
2046
2047                 if (path && ext_depth(inode) != depth) {
2048                         /* depth was changed. we have to realloc path */
2049                         kfree(path);
2050                         path = NULL;
2051                 }
2052
2053                 path = ext4_ext_find_extent(inode, block, path);
2054                 if (IS_ERR(path)) {
2055                         up_read(&EXT4_I(inode)->i_data_sem);
2056                         err = PTR_ERR(path);
2057                         path = NULL;
2058                         break;
2059                 }
2060
2061                 depth = ext_depth(inode);
2062                 if (unlikely(path[depth].p_hdr == NULL)) {
2063                         up_read(&EXT4_I(inode)->i_data_sem);
2064                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2065                         err = -EIO;
2066                         break;
2067                 }
2068                 ex = path[depth].p_ext;
2069                 next = ext4_ext_next_allocated_block(path);
2070                 ext4_ext_drop_refs(path);
2071
2072                 flags = 0;
2073                 exists = 0;
2074                 if (!ex) {
2075                         /* there is no extent yet, so try to allocate
2076                          * all requested space */
2077                         start = block;
2078                         end = block + num;
2079                 } else if (le32_to_cpu(ex->ee_block) > block) {
2080                         /* need to allocate space before found extent */
2081                         start = block;
2082                         end = le32_to_cpu(ex->ee_block);
2083                         if (block + num < end)
2084                                 end = block + num;
2085                 } else if (block >= le32_to_cpu(ex->ee_block)
2086                                         + ext4_ext_get_actual_len(ex)) {
2087                         /* need to allocate space after found extent */
2088                         start = block;
2089                         end = block + num;
2090                         if (end >= next)
2091                                 end = next;
2092                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2093                         /*
2094                          * some part of requested space is covered
2095                          * by found extent
2096                          */
2097                         start = block;
2098                         end = le32_to_cpu(ex->ee_block)
2099                                 + ext4_ext_get_actual_len(ex);
2100                         if (block + num < end)
2101                                 end = block + num;
2102                         exists = 1;
2103                 } else {
2104                         BUG();
2105                 }
2106                 BUG_ON(end <= start);
2107
2108                 if (!exists) {
2109                         es.es_lblk = start;
2110                         es.es_len = end - start;
2111                         es.es_pblk = 0;
2112                 } else {
2113                         es.es_lblk = le32_to_cpu(ex->ee_block);
2114                         es.es_len = ext4_ext_get_actual_len(ex);
2115                         es.es_pblk = ext4_ext_pblock(ex);
2116                         if (ext4_ext_is_uninitialized(ex))
2117                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2118                 }
2119
2120                 /*
2121                  * Find delayed extent and update es accordingly. We call
2122                  * it even in !exists case to find out whether es is the
2123                  * last existing extent or not.
2124                  */
2125                 next_del = ext4_find_delayed_extent(inode, &es);
2126                 if (!exists && next_del) {
2127                         exists = 1;
2128                         flags |= FIEMAP_EXTENT_DELALLOC;
2129                 }
2130                 up_read(&EXT4_I(inode)->i_data_sem);
2131
2132                 if (unlikely(es.es_len == 0)) {
2133                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2134                         err = -EIO;
2135                         break;
2136                 }
2137
2138                 /*
2139                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2140                  * we need to check next == EXT_MAX_BLOCKS because it is
2141                  * possible that an extent is with unwritten and delayed
2142                  * status due to when an extent is delayed allocated and
2143                  * is allocated by fallocate status tree will track both of
2144                  * them in a extent.
2145                  *
2146                  * So we could return a unwritten and delayed extent, and
2147                  * its block is equal to 'next'.
2148                  */
2149                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2150                         flags |= FIEMAP_EXTENT_LAST;
2151                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2152                                      next != EXT_MAX_BLOCKS)) {
2153                                 EXT4_ERROR_INODE(inode,
2154                                                  "next extent == %u, next "
2155                                                  "delalloc extent = %u",
2156                                                  next, next_del);
2157                                 err = -EIO;
2158                                 break;
2159                         }
2160                 }
2161
2162                 if (exists) {
2163                         err = fiemap_fill_next_extent(fieinfo,
2164                                 (__u64)es.es_lblk << blksize_bits,
2165                                 (__u64)es.es_pblk << blksize_bits,
2166                                 (__u64)es.es_len << blksize_bits,
2167                                 flags);
2168                         if (err < 0)
2169                                 break;
2170                         if (err == 1) {
2171                                 err = 0;
2172                                 break;
2173                         }
2174                 }
2175
2176                 block = es.es_lblk + es.es_len;
2177         }
2178
2179         if (path) {
2180                 ext4_ext_drop_refs(path);
2181                 kfree(path);
2182         }
2183
2184         return err;
2185 }
2186
2187 /*
2188  * ext4_ext_put_gap_in_cache:
2189  * calculate boundaries of the gap that the requested block fits into
2190  * and cache this gap
2191  */
2192 static void
2193 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2194                                 ext4_lblk_t block)
2195 {
2196         int depth = ext_depth(inode);
2197         unsigned long len;
2198         ext4_lblk_t lblock;
2199         struct ext4_extent *ex;
2200
2201         ex = path[depth].p_ext;
2202         if (ex == NULL) {
2203                 /*
2204                  * there is no extent yet, so gap is [0;-] and we
2205                  * don't cache it
2206                  */
2207                 ext_debug("cache gap(whole file):");
2208         } else if (block < le32_to_cpu(ex->ee_block)) {
2209                 lblock = block;
2210                 len = le32_to_cpu(ex->ee_block) - block;
2211                 ext_debug("cache gap(before): %u [%u:%u]",
2212                                 block,
2213                                 le32_to_cpu(ex->ee_block),
2214                                  ext4_ext_get_actual_len(ex));
2215                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2216                         ext4_es_insert_extent(inode, lblock, len, ~0,
2217                                               EXTENT_STATUS_HOLE);
2218         } else if (block >= le32_to_cpu(ex->ee_block)
2219                         + ext4_ext_get_actual_len(ex)) {
2220                 ext4_lblk_t next;
2221                 lblock = le32_to_cpu(ex->ee_block)
2222                         + ext4_ext_get_actual_len(ex);
2223
2224                 next = ext4_ext_next_allocated_block(path);
2225                 ext_debug("cache gap(after): [%u:%u] %u",
2226                                 le32_to_cpu(ex->ee_block),
2227                                 ext4_ext_get_actual_len(ex),
2228                                 block);
2229                 BUG_ON(next == lblock);
2230                 len = next - lblock;
2231                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2232                         ext4_es_insert_extent(inode, lblock, len, ~0,
2233                                               EXTENT_STATUS_HOLE);
2234         } else {
2235                 lblock = len = 0;
2236                 BUG();
2237         }
2238
2239         ext_debug(" -> %u:%lu\n", lblock, len);
2240 }
2241
2242 /*
2243  * ext4_ext_rm_idx:
2244  * removes index from the index block.
2245  */
2246 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2247                         struct ext4_ext_path *path, int depth)
2248 {
2249         int err;
2250         ext4_fsblk_t leaf;
2251
2252         /* free index block */
2253         depth--;
2254         path = path + depth;
2255         leaf = ext4_idx_pblock(path->p_idx);
2256         if (unlikely(path->p_hdr->eh_entries == 0)) {
2257                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2258                 return -EIO;
2259         }
2260         err = ext4_ext_get_access(handle, inode, path);
2261         if (err)
2262                 return err;
2263
2264         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2265                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2266                 len *= sizeof(struct ext4_extent_idx);
2267                 memmove(path->p_idx, path->p_idx + 1, len);
2268         }
2269
2270         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2271         err = ext4_ext_dirty(handle, inode, path);
2272         if (err)
2273                 return err;
2274         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2275         trace_ext4_ext_rm_idx(inode, leaf);
2276
2277         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2278                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2279
2280         while (--depth >= 0) {
2281                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2282                         break;
2283                 path--;
2284                 err = ext4_ext_get_access(handle, inode, path);
2285                 if (err)
2286                         break;
2287                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2288                 err = ext4_ext_dirty(handle, inode, path);
2289                 if (err)
2290                         break;
2291         }
2292         return err;
2293 }
2294
2295 /*
2296  * ext4_ext_calc_credits_for_single_extent:
2297  * This routine returns max. credits that needed to insert an extent
2298  * to the extent tree.
2299  * When pass the actual path, the caller should calculate credits
2300  * under i_data_sem.
2301  */
2302 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2303                                                 struct ext4_ext_path *path)
2304 {
2305         if (path) {
2306                 int depth = ext_depth(inode);
2307                 int ret = 0;
2308
2309                 /* probably there is space in leaf? */
2310                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2311                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2312
2313                         /*
2314                          *  There are some space in the leaf tree, no
2315                          *  need to account for leaf block credit
2316                          *
2317                          *  bitmaps and block group descriptor blocks
2318                          *  and other metadata blocks still need to be
2319                          *  accounted.
2320                          */
2321                         /* 1 bitmap, 1 block group descriptor */
2322                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2323                         return ret;
2324                 }
2325         }
2326
2327         return ext4_chunk_trans_blocks(inode, nrblocks);
2328 }
2329
2330 /*
2331  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2332  *
2333  * if nrblocks are fit in a single extent (chunk flag is 1), then
2334  * in the worse case, each tree level index/leaf need to be changed
2335  * if the tree split due to insert a new extent, then the old tree
2336  * index/leaf need to be updated too
2337  *
2338  * If the nrblocks are discontiguous, they could cause
2339  * the whole tree split more than once, but this is really rare.
2340  */
2341 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2342 {
2343         int index;
2344         int depth;
2345
2346         /* If we are converting the inline data, only one is needed here. */
2347         if (ext4_has_inline_data(inode))
2348                 return 1;
2349
2350         depth = ext_depth(inode);
2351
2352         if (chunk)
2353                 index = depth * 2;
2354         else
2355                 index = depth * 3;
2356
2357         return index;
2358 }
2359
2360 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2361                               struct ext4_extent *ex,
2362                               long long *partial_cluster,
2363                               ext4_lblk_t from, ext4_lblk_t to)
2364 {
2365         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2366         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2367         ext4_fsblk_t pblk;
2368         int flags = 0;
2369
2370         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2371                 flags |= EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2372         else if (ext4_should_journal_data(inode))
2373                 flags |= EXT4_FREE_BLOCKS_FORGET;
2374
2375         /*
2376          * For bigalloc file systems, we never free a partial cluster
2377          * at the beginning of the extent.  Instead, we make a note
2378          * that we tried freeing the cluster, and check to see if we
2379          * need to free it on a subsequent call to ext4_remove_blocks,
2380          * or at the end of the ext4_truncate() operation.
2381          */
2382         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2383
2384         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2385         /*
2386          * If we have a partial cluster, and it's different from the
2387          * cluster of the last block, we need to explicitly free the
2388          * partial cluster here.
2389          */
2390         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2391         if ((*partial_cluster > 0) &&
2392             (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2393                 ext4_free_blocks(handle, inode, NULL,
2394                                  EXT4_C2B(sbi, *partial_cluster),
2395                                  sbi->s_cluster_ratio, flags);
2396                 *partial_cluster = 0;
2397         }
2398
2399 #ifdef EXTENTS_STATS
2400         {
2401                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2402                 spin_lock(&sbi->s_ext_stats_lock);
2403                 sbi->s_ext_blocks += ee_len;
2404                 sbi->s_ext_extents++;
2405                 if (ee_len < sbi->s_ext_min)
2406                         sbi->s_ext_min = ee_len;
2407                 if (ee_len > sbi->s_ext_max)
2408                         sbi->s_ext_max = ee_len;
2409                 if (ext_depth(inode) > sbi->s_depth_max)
2410                         sbi->s_depth_max = ext_depth(inode);
2411                 spin_unlock(&sbi->s_ext_stats_lock);
2412         }
2413 #endif
2414         if (from >= le32_to_cpu(ex->ee_block)
2415             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2416                 /* tail removal */
2417                 ext4_lblk_t num;
2418                 unsigned int unaligned;
2419
2420                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2421                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2422                 /*
2423                  * Usually we want to free partial cluster at the end of the
2424                  * extent, except for the situation when the cluster is still
2425                  * used by any other extent (partial_cluster is negative).
2426                  */
2427                 if (*partial_cluster < 0 &&
2428                     -(*partial_cluster) == EXT4_B2C(sbi, pblk + num - 1))
2429                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2430
2431                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2432                           num, pblk, *partial_cluster);
2433                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2434                 /*
2435                  * If the block range to be freed didn't start at the
2436                  * beginning of a cluster, and we removed the entire
2437                  * extent and the cluster is not used by any other extent,
2438                  * save the partial cluster here, since we might need to
2439                  * delete if we determine that the truncate operation has
2440                  * removed all of the blocks in the cluster.
2441                  *
2442                  * On the other hand, if we did not manage to free the whole
2443                  * extent, we have to mark the cluster as used (store negative
2444                  * cluster number in partial_cluster).
2445                  */
2446                 unaligned = pblk & (sbi->s_cluster_ratio - 1);
2447                 if (unaligned && (ee_len == num) &&
2448                     (*partial_cluster != -((long long)EXT4_B2C(sbi, pblk))))
2449                         *partial_cluster = EXT4_B2C(sbi, pblk);
2450                 else if (unaligned)
2451                         *partial_cluster = -((long long)EXT4_B2C(sbi, pblk));
2452                 else if (*partial_cluster > 0)
2453                         *partial_cluster = 0;
2454         } else
2455                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2456                            "%u-%u from %u:%u\n",
2457                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2458         return 0;
2459 }
2460
2461
2462 /*
2463  * ext4_ext_rm_leaf() Removes the extents associated with the
2464  * blocks appearing between "start" and "end", and splits the extents
2465  * if "start" and "end" appear in the same extent
2466  *
2467  * @handle: The journal handle
2468  * @inode:  The files inode
2469  * @path:   The path to the leaf
2470  * @partial_cluster: The cluster which we'll have to free if all extents
2471  *                   has been released from it. It gets negative in case
2472  *                   that the cluster is still used.
2473  * @start:  The first block to remove
2474  * @end:   The last block to remove
2475  */
2476 static int
2477 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2478                  struct ext4_ext_path *path,
2479                  long long *partial_cluster,
2480                  ext4_lblk_t start, ext4_lblk_t end)
2481 {
2482         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2483         int err = 0, correct_index = 0;
2484         int depth = ext_depth(inode), credits;
2485         struct ext4_extent_header *eh;
2486         ext4_lblk_t a, b;
2487         unsigned num;
2488         ext4_lblk_t ex_ee_block;
2489         unsigned short ex_ee_len;
2490         unsigned uninitialized = 0;
2491         struct ext4_extent *ex;
2492         ext4_fsblk_t pblk;
2493
2494         /* the header must be checked already in ext4_ext_remove_space() */
2495         ext_debug("truncate since %u in leaf to %u\n", start, end);
2496         if (!path[depth].p_hdr)
2497                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2498         eh = path[depth].p_hdr;
2499         if (unlikely(path[depth].p_hdr == NULL)) {
2500                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2501                 return -EIO;
2502         }
2503         /* find where to start removing */
2504         ex = EXT_LAST_EXTENT(eh);
2505
2506         ex_ee_block = le32_to_cpu(ex->ee_block);
2507         ex_ee_len = ext4_ext_get_actual_len(ex);
2508
2509         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2510
2511         while (ex >= EXT_FIRST_EXTENT(eh) &&
2512                         ex_ee_block + ex_ee_len > start) {
2513
2514                 if (ext4_ext_is_uninitialized(ex))
2515                         uninitialized = 1;
2516                 else
2517                         uninitialized = 0;
2518
2519                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2520                          uninitialized, ex_ee_len);
2521                 path[depth].p_ext = ex;
2522
2523                 a = ex_ee_block > start ? ex_ee_block : start;
2524                 b = ex_ee_block+ex_ee_len - 1 < end ?
2525                         ex_ee_block+ex_ee_len - 1 : end;
2526
2527                 ext_debug("  border %u:%u\n", a, b);
2528
2529                 /* If this extent is beyond the end of the hole, skip it */
2530                 if (end < ex_ee_block) {
2531                         /*
2532                          * We're going to skip this extent and move to another,
2533                          * so if this extent is not cluster aligned we have
2534                          * to mark the current cluster as used to avoid
2535                          * accidentally freeing it later on
2536                          */
2537                         pblk = ext4_ext_pblock(ex);
2538                         if (pblk & (sbi->s_cluster_ratio - 1))
2539                                 *partial_cluster =
2540                                         -((long long)EXT4_B2C(sbi, pblk));
2541                         ex--;
2542                         ex_ee_block = le32_to_cpu(ex->ee_block);
2543                         ex_ee_len = ext4_ext_get_actual_len(ex);
2544                         continue;
2545                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2546                         EXT4_ERROR_INODE(inode,
2547                                          "can not handle truncate %u:%u "
2548                                          "on extent %u:%u",
2549                                          start, end, ex_ee_block,
2550                                          ex_ee_block + ex_ee_len - 1);
2551                         err = -EIO;
2552                         goto out;
2553                 } else if (a != ex_ee_block) {
2554                         /* remove tail of the extent */
2555                         num = a - ex_ee_block;
2556                 } else {
2557                         /* remove whole extent: excellent! */
2558                         num = 0;
2559                 }
2560                 /*
2561                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2562                  * descriptor) for each block group; assume two block
2563                  * groups plus ex_ee_len/blocks_per_block_group for
2564                  * the worst case
2565                  */
2566                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2567                 if (ex == EXT_FIRST_EXTENT(eh)) {
2568                         correct_index = 1;
2569                         credits += (ext_depth(inode)) + 1;
2570                 }
2571                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2572
2573                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2574                 if (err)
2575                         goto out;
2576
2577                 err = ext4_ext_get_access(handle, inode, path + depth);
2578                 if (err)
2579                         goto out;
2580
2581                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2582                                          a, b);
2583                 if (err)
2584                         goto out;
2585
2586                 if (num == 0)
2587                         /* this extent is removed; mark slot entirely unused */
2588                         ext4_ext_store_pblock(ex, 0);
2589
2590                 ex->ee_len = cpu_to_le16(num);
2591                 /*
2592                  * Do not mark uninitialized if all the blocks in the
2593                  * extent have been removed.
2594                  */
2595                 if (uninitialized && num)
2596                         ext4_ext_mark_uninitialized(ex);
2597                 /*
2598                  * If the extent was completely released,
2599                  * we need to remove it from the leaf
2600                  */
2601                 if (num == 0) {
2602                         if (end != EXT_MAX_BLOCKS - 1) {
2603                                 /*
2604                                  * For hole punching, we need to scoot all the
2605                                  * extents up when an extent is removed so that
2606                                  * we dont have blank extents in the middle
2607                                  */
2608                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2609                                         sizeof(struct ext4_extent));
2610
2611                                 /* Now get rid of the one at the end */
2612                                 memset(EXT_LAST_EXTENT(eh), 0,
2613                                         sizeof(struct ext4_extent));
2614                         }
2615                         le16_add_cpu(&eh->eh_entries, -1);
2616                 } else if (*partial_cluster > 0)
2617                         *partial_cluster = 0;
2618
2619                 err = ext4_ext_dirty(handle, inode, path + depth);
2620                 if (err)
2621                         goto out;
2622
2623                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2624                                 ext4_ext_pblock(ex));
2625                 ex--;
2626                 ex_ee_block = le32_to_cpu(ex->ee_block);
2627                 ex_ee_len = ext4_ext_get_actual_len(ex);
2628         }
2629
2630         if (correct_index && eh->eh_entries)
2631                 err = ext4_ext_correct_indexes(handle, inode, path);
2632
2633         /*
2634          * Free the partial cluster only if the current extent does not
2635          * reference it. Otherwise we might free used cluster.
2636          */
2637         if (*partial_cluster > 0 &&
2638             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2639              *partial_cluster)) {
2640                 int flags = EXT4_FREE_BLOCKS_FORGET;
2641
2642                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2643                         flags |= EXT4_FREE_BLOCKS_METADATA;
2644
2645                 ext4_free_blocks(handle, inode, NULL,
2646                                  EXT4_C2B(sbi, *partial_cluster),
2647                                  sbi->s_cluster_ratio, flags);
2648                 *partial_cluster = 0;
2649         }
2650
2651         /* if this leaf is free, then we should
2652          * remove it from index block above */
2653         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2654                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2655
2656 out:
2657         return err;
2658 }
2659
2660 /*
2661  * ext4_ext_more_to_rm:
2662  * returns 1 if current index has to be freed (even partial)
2663  */
2664 static int
2665 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2666 {
2667         BUG_ON(path->p_idx == NULL);
2668
2669         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2670                 return 0;
2671
2672         /*
2673          * if truncate on deeper level happened, it wasn't partial,
2674          * so we have to consider current index for truncation
2675          */
2676         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2677                 return 0;
2678         return 1;
2679 }
2680
2681 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2682                           ext4_lblk_t end)
2683 {
2684         struct super_block *sb = inode->i_sb;
2685         int depth = ext_depth(inode);
2686         struct ext4_ext_path *path = NULL;
2687         long long partial_cluster = 0;
2688         handle_t *handle;
2689         int i = 0, err = 0;
2690
2691         ext_debug("truncate since %u to %u\n", start, end);
2692
2693         /* probably first extent we're gonna free will be last in block */
2694         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2695         if (IS_ERR(handle))
2696                 return PTR_ERR(handle);
2697
2698 again:
2699         trace_ext4_ext_remove_space(inode, start, end, depth);
2700
2701         /*
2702          * Check if we are removing extents inside the extent tree. If that
2703          * is the case, we are going to punch a hole inside the extent tree
2704          * so we have to check whether we need to split the extent covering
2705          * the last block to remove so we can easily remove the part of it
2706          * in ext4_ext_rm_leaf().
2707          */
2708         if (end < EXT_MAX_BLOCKS - 1) {
2709                 struct ext4_extent *ex;
2710                 ext4_lblk_t ee_block;
2711
2712                 /* find extent for this block */
2713                 path = ext4_ext_find_extent(inode, end, NULL);
2714                 if (IS_ERR(path)) {
2715                         ext4_journal_stop(handle);
2716                         return PTR_ERR(path);
2717                 }
2718                 depth = ext_depth(inode);
2719                 /* Leaf not may not exist only if inode has no blocks at all */
2720                 ex = path[depth].p_ext;
2721                 if (!ex) {
2722                         if (depth) {
2723                                 EXT4_ERROR_INODE(inode,
2724                                                  "path[%d].p_hdr == NULL",
2725                                                  depth);
2726                                 err = -EIO;
2727                         }
2728                         goto out;
2729                 }
2730
2731                 ee_block = le32_to_cpu(ex->ee_block);
2732
2733                 /*
2734                  * See if the last block is inside the extent, if so split
2735                  * the extent at 'end' block so we can easily remove the
2736                  * tail of the first part of the split extent in
2737                  * ext4_ext_rm_leaf().
2738                  */
2739                 if (end >= ee_block &&
2740                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2741                         int split_flag = 0;
2742
2743                         if (ext4_ext_is_uninitialized(ex))
2744                                 split_flag = EXT4_EXT_MARK_UNINIT1 |
2745                                              EXT4_EXT_MARK_UNINIT2;
2746
2747                         /*
2748                          * Split the extent in two so that 'end' is the last
2749                          * block in the first new extent. Also we should not
2750                          * fail removing space due to ENOSPC so try to use
2751                          * reserved block if that happens.
2752                          */
2753                         err = ext4_split_extent_at(handle, inode, path,
2754                                         end + 1, split_flag,
2755                                         EXT4_GET_BLOCKS_PRE_IO |
2756                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
2757
2758                         if (err < 0)
2759                                 goto out;
2760                 }
2761         }
2762         /*
2763          * We start scanning from right side, freeing all the blocks
2764          * after i_size and walking into the tree depth-wise.
2765          */
2766         depth = ext_depth(inode);
2767         if (path) {
2768                 int k = i = depth;
2769                 while (--k > 0)
2770                         path[k].p_block =
2771                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2772         } else {
2773                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2774                                GFP_NOFS);
2775                 if (path == NULL) {
2776                         ext4_journal_stop(handle);
2777                         return -ENOMEM;
2778                 }
2779                 path[0].p_depth = depth;
2780                 path[0].p_hdr = ext_inode_hdr(inode);
2781                 i = 0;
2782
2783                 if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2784                         err = -EIO;
2785                         goto out;
2786                 }
2787         }
2788         err = 0;
2789
2790         while (i >= 0 && err == 0) {
2791                 if (i == depth) {
2792                         /* this is leaf block */
2793                         err = ext4_ext_rm_leaf(handle, inode, path,
2794                                                &partial_cluster, start,
2795                                                end);
2796                         /* root level has p_bh == NULL, brelse() eats this */
2797                         brelse(path[i].p_bh);
2798                         path[i].p_bh = NULL;
2799                         i--;
2800                         continue;
2801                 }
2802
2803                 /* this is index block */
2804                 if (!path[i].p_hdr) {
2805                         ext_debug("initialize header\n");
2806                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2807                 }
2808
2809                 if (!path[i].p_idx) {
2810                         /* this level hasn't been touched yet */
2811                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2812                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2813                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2814                                   path[i].p_hdr,
2815                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2816                 } else {
2817                         /* we were already here, see at next index */
2818                         path[i].p_idx--;
2819                 }
2820
2821                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2822                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2823                                 path[i].p_idx);
2824                 if (ext4_ext_more_to_rm(path + i)) {
2825                         struct buffer_head *bh;
2826                         /* go to the next level */
2827                         ext_debug("move to level %d (block %llu)\n",
2828                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2829                         memset(path + i + 1, 0, sizeof(*path));
2830                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2831                         if (!bh) {
2832                                 /* should we reset i_size? */
2833                                 err = -EIO;
2834                                 break;
2835                         }
2836                         if (WARN_ON(i + 1 > depth)) {
2837                                 err = -EIO;
2838                                 break;
2839                         }
2840                         if (ext4_ext_check_block(inode, ext_block_hdr(bh),
2841                                                         depth - i - 1, bh)) {
2842                                 err = -EIO;
2843                                 break;
2844                         }
2845                         path[i + 1].p_bh = bh;
2846
2847                         /* save actual number of indexes since this
2848                          * number is changed at the next iteration */
2849                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2850                         i++;
2851                 } else {
2852                         /* we finished processing this index, go up */
2853                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2854                                 /* index is empty, remove it;
2855                                  * handle must be already prepared by the
2856                                  * truncatei_leaf() */
2857                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2858                         }
2859                         /* root level has p_bh == NULL, brelse() eats this */
2860                         brelse(path[i].p_bh);
2861                         path[i].p_bh = NULL;
2862                         i--;
2863                         ext_debug("return to level %d\n", i);
2864                 }
2865         }
2866
2867         trace_ext4_ext_remove_space_done(inode, start, end, depth,
2868                         partial_cluster, path->p_hdr->eh_entries);
2869
2870         /* If we still have something in the partial cluster and we have removed
2871          * even the first extent, then we should free the blocks in the partial
2872          * cluster as well. */
2873         if (partial_cluster > 0 && path->p_hdr->eh_entries == 0) {
2874                 int flags = EXT4_FREE_BLOCKS_FORGET;
2875
2876                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2877                         flags |= EXT4_FREE_BLOCKS_METADATA;
2878
2879                 ext4_free_blocks(handle, inode, NULL,
2880                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2881                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2882                 partial_cluster = 0;
2883         }
2884
2885         /* TODO: flexible tree reduction should be here */
2886         if (path->p_hdr->eh_entries == 0) {
2887                 /*
2888                  * truncate to zero freed all the tree,
2889                  * so we need to correct eh_depth
2890                  */
2891                 err = ext4_ext_get_access(handle, inode, path);
2892                 if (err == 0) {
2893                         ext_inode_hdr(inode)->eh_depth = 0;
2894                         ext_inode_hdr(inode)->eh_max =
2895                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2896                         err = ext4_ext_dirty(handle, inode, path);
2897                 }
2898         }
2899 out:
2900         ext4_ext_drop_refs(path);
2901         kfree(path);
2902         if (err == -EAGAIN) {
2903                 path = NULL;
2904                 goto again;
2905         }
2906         ext4_journal_stop(handle);
2907
2908         return err;
2909 }
2910
2911 /*
2912  * called at mount time
2913  */
2914 void ext4_ext_init(struct super_block *sb)
2915 {
2916         /*
2917          * possible initialization would be here
2918          */
2919
2920         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2921 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2922                 printk(KERN_INFO "EXT4-fs: file extents enabled"
2923 #ifdef AGGRESSIVE_TEST
2924                        ", aggressive tests"
2925 #endif
2926 #ifdef CHECK_BINSEARCH
2927                        ", check binsearch"
2928 #endif
2929 #ifdef EXTENTS_STATS
2930                        ", stats"
2931 #endif
2932                        "\n");
2933 #endif
2934 #ifdef EXTENTS_STATS
2935                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2936                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2937                 EXT4_SB(sb)->s_ext_max = 0;
2938 #endif
2939         }
2940 }
2941
2942 /*
2943  * called at umount time
2944  */
2945 void ext4_ext_release(struct super_block *sb)
2946 {
2947         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2948                 return;
2949
2950 #ifdef EXTENTS_STATS
2951         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2952                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2953                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2954                         sbi->s_ext_blocks, sbi->s_ext_extents,
2955                         sbi->s_ext_blocks / sbi->s_ext_extents);
2956                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2957                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2958         }
2959 #endif
2960 }
2961
2962 /* FIXME!! we need to try to merge to left or right after zero-out  */
2963 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2964 {
2965         ext4_fsblk_t ee_pblock;
2966         unsigned int ee_len;
2967         int ret;
2968
2969         ee_len    = ext4_ext_get_actual_len(ex);
2970         ee_pblock = ext4_ext_pblock(ex);
2971
2972         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2973         if (ret > 0)
2974                 ret = 0;
2975
2976         return ret;
2977 }
2978
2979 /*
2980  * ext4_split_extent_at() splits an extent at given block.
2981  *
2982  * @handle: the journal handle
2983  * @inode: the file inode
2984  * @path: the path to the extent
2985  * @split: the logical block where the extent is splitted.
2986  * @split_flags: indicates if the extent could be zeroout if split fails, and
2987  *               the states(init or uninit) of new extents.
2988  * @flags: flags used to insert new extent to extent tree.
2989  *
2990  *
2991  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2992  * of which are deterimined by split_flag.
2993  *
2994  * There are two cases:
2995  *  a> the extent are splitted into two extent.
2996  *  b> split is not needed, and just mark the extent.
2997  *
2998  * return 0 on success.
2999  */
3000 static int ext4_split_extent_at(handle_t *handle,
3001                              struct inode *inode,
3002                              struct ext4_ext_path *path,
3003                              ext4_lblk_t split,
3004                              int split_flag,
3005                              int flags)
3006 {
3007         ext4_fsblk_t newblock;
3008         ext4_lblk_t ee_block;
3009         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3010         struct ext4_extent *ex2 = NULL;
3011         unsigned int ee_len, depth;
3012         int err = 0;
3013
3014         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3015                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3016
3017         ext_debug("ext4_split_extents_at: inode %lu, logical"
3018                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3019
3020         ext4_ext_show_leaf(inode, path);
3021
3022         depth = ext_depth(inode);
3023         ex = path[depth].p_ext;
3024         ee_block = le32_to_cpu(ex->ee_block);
3025         ee_len = ext4_ext_get_actual_len(ex);
3026         newblock = split - ee_block + ext4_ext_pblock(ex);
3027
3028         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3029         BUG_ON(!ext4_ext_is_uninitialized(ex) &&
3030                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3031                              EXT4_EXT_MARK_UNINIT1 |
3032                              EXT4_EXT_MARK_UNINIT2));
3033
3034         err = ext4_ext_get_access(handle, inode, path + depth);
3035         if (err)
3036                 goto out;
3037
3038         if (split == ee_block) {
3039                 /*
3040                  * case b: block @split is the block that the extent begins with
3041                  * then we just change the state of the extent, and splitting
3042                  * is not needed.
3043                  */
3044                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
3045                         ext4_ext_mark_uninitialized(ex);
3046                 else
3047                         ext4_ext_mark_initialized(ex);
3048
3049                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3050                         ext4_ext_try_to_merge(handle, inode, path, ex);
3051
3052                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3053                 goto out;
3054         }
3055
3056         /* case a */
3057         memcpy(&orig_ex, ex, sizeof(orig_ex));
3058         ex->ee_len = cpu_to_le16(split - ee_block);
3059         if (split_flag & EXT4_EXT_MARK_UNINIT1)
3060                 ext4_ext_mark_uninitialized(ex);
3061
3062         /*
3063          * path may lead to new leaf, not to original leaf any more
3064          * after ext4_ext_insert_extent() returns,
3065          */
3066         err = ext4_ext_dirty(handle, inode, path + depth);
3067         if (err)
3068                 goto fix_extent_len;
3069
3070         ex2 = &newex;
3071         ex2->ee_block = cpu_to_le32(split);
3072         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3073         ext4_ext_store_pblock(ex2, newblock);
3074         if (split_flag & EXT4_EXT_MARK_UNINIT2)
3075                 ext4_ext_mark_uninitialized(ex2);
3076
3077         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
3078         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3079                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3080                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3081                                 err = ext4_ext_zeroout(inode, ex2);
3082                                 zero_ex.ee_block = ex2->ee_block;
3083                                 zero_ex.ee_len = cpu_to_le16(
3084                                                 ext4_ext_get_actual_len(ex2));
3085                                 ext4_ext_store_pblock(&zero_ex,
3086                                                       ext4_ext_pblock(ex2));
3087                         } else {
3088                                 err = ext4_ext_zeroout(inode, ex);
3089                                 zero_ex.ee_block = ex->ee_block;
3090                                 zero_ex.ee_len = cpu_to_le16(
3091                                                 ext4_ext_get_actual_len(ex));
3092                                 ext4_ext_store_pblock(&zero_ex,
3093                                                       ext4_ext_pblock(ex));
3094                         }
3095                 } else {
3096                         err = ext4_ext_zeroout(inode, &orig_ex);
3097                         zero_ex.ee_block = orig_ex.ee_block;
3098                         zero_ex.ee_len = cpu_to_le16(
3099                                                 ext4_ext_get_actual_len(&orig_ex));
3100                         ext4_ext_store_pblock(&zero_ex,
3101                                               ext4_ext_pblock(&orig_ex));
3102                 }
3103
3104                 if (err)
3105                         goto fix_extent_len;
3106                 /* update the extent length and mark as initialized */
3107                 ex->ee_len = cpu_to_le16(ee_len);
3108                 ext4_ext_try_to_merge(handle, inode, path, ex);
3109                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3110                 if (err)
3111                         goto fix_extent_len;
3112
3113                 /* update extent status tree */
3114                 err = ext4_es_zeroout(inode, &zero_ex);
3115
3116                 goto out;
3117         } else if (err)
3118                 goto fix_extent_len;
3119
3120 out:
3121         ext4_ext_show_leaf(inode, path);
3122         return err;
3123
3124 fix_extent_len:
3125         ex->ee_len = orig_ex.ee_len;
3126         ext4_ext_dirty(handle, inode, path + depth);
3127         return err;
3128 }
3129
3130 /*
3131  * ext4_split_extents() splits an extent and mark extent which is covered
3132  * by @map as split_flags indicates
3133  *
3134  * It may result in splitting the extent into multiple extents (upto three)
3135  * There are three possibilities:
3136  *   a> There is no split required
3137  *   b> Splits in two extents: Split is happening at either end of the extent
3138  *   c> Splits in three extents: Somone is splitting in middle of the extent
3139  *
3140  */
3141 static int ext4_split_extent(handle_t *handle,
3142                               struct inode *inode,
3143                               struct ext4_ext_path *path,
3144                               struct ext4_map_blocks *map,
3145                               int split_flag,
3146                               int flags)
3147 {
3148         ext4_lblk_t ee_block;
3149         struct ext4_extent *ex;
3150         unsigned int ee_len, depth;
3151         int err = 0;
3152         int uninitialized;
3153         int split_flag1, flags1;
3154         int allocated = map->m_len;
3155
3156         depth = ext_depth(inode);
3157         ex = path[depth].p_ext;
3158         ee_block = le32_to_cpu(ex->ee_block);
3159         ee_len = ext4_ext_get_actual_len(ex);
3160         uninitialized = ext4_ext_is_uninitialized(ex);
3161
3162         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3163                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3164                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3165                 if (uninitialized)
3166                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
3167                                        EXT4_EXT_MARK_UNINIT2;
3168                 if (split_flag & EXT4_EXT_DATA_VALID2)
3169                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3170                 err = ext4_split_extent_at(handle, inode, path,
3171                                 map->m_lblk + map->m_len, split_flag1, flags1);
3172                 if (err)
3173                         goto out;
3174         } else {
3175                 allocated = ee_len - (map->m_lblk - ee_block);
3176         }
3177         /*
3178          * Update path is required because previous ext4_split_extent_at() may
3179          * result in split of original leaf or extent zeroout.
3180          */
3181         ext4_ext_drop_refs(path);
3182         path = ext4_ext_find_extent(inode, map->m_lblk, path);
3183         if (IS_ERR(path))
3184                 return PTR_ERR(path);
3185         depth = ext_depth(inode);
3186         ex = path[depth].p_ext;
3187         uninitialized = ext4_ext_is_uninitialized(ex);
3188         split_flag1 = 0;
3189
3190         if (map->m_lblk >= ee_block) {
3191                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3192                 if (uninitialized) {
3193                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
3194                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3195                                                      EXT4_EXT_MARK_UNINIT2);
3196                 }
3197                 err = ext4_split_extent_at(handle, inode, path,
3198                                 map->m_lblk, split_flag1, flags);
3199                 if (err)
3200                         goto out;
3201         }
3202
3203         ext4_ext_show_leaf(inode, path);
3204 out:
3205         return err ? err : allocated;
3206 }
3207
3208 /*
3209  * This function is called by ext4_ext_map_blocks() if someone tries to write
3210  * to an uninitialized extent. It may result in splitting the uninitialized
3211  * extent into multiple extents (up to three - one initialized and two
3212  * uninitialized).
3213  * There are three possibilities:
3214  *   a> There is no split required: Entire extent should be initialized
3215  *   b> Splits in two extents: Write is happening at either end of the extent
3216  *   c> Splits in three extents: Somone is writing in middle of the extent
3217  *
3218  * Pre-conditions:
3219  *  - The extent pointed to by 'path' is uninitialized.
3220  *  - The extent pointed to by 'path' contains a superset
3221  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3222  *
3223  * Post-conditions on success:
3224  *  - the returned value is the number of blocks beyond map->l_lblk
3225  *    that are allocated and initialized.
3226  *    It is guaranteed to be >= map->m_len.
3227  */
3228 static int ext4_ext_convert_to_initialized(handle_t *handle,
3229                                            struct inode *inode,
3230                                            struct ext4_map_blocks *map,
3231                                            struct ext4_ext_path *path,
3232                                            int flags)
3233 {
3234         struct ext4_sb_info *sbi;
3235         struct ext4_extent_header *eh;
3236         struct ext4_map_blocks split_map;
3237         struct ext4_extent zero_ex;
3238         struct ext4_extent *ex, *abut_ex;
3239         ext4_lblk_t ee_block, eof_block;
3240         unsigned int ee_len, depth, map_len = map->m_len;
3241         int allocated = 0, max_zeroout = 0;
3242         int err = 0;
3243         int split_flag = 0;
3244
3245         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3246                 "block %llu, max_blocks %u\n", inode->i_ino,
3247                 (unsigned long long)map->m_lblk, map_len);
3248
3249         sbi = EXT4_SB(inode->i_sb);
3250         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3251                 inode->i_sb->s_blocksize_bits;
3252         if (eof_block < map->m_lblk + map_len)
3253                 eof_block = map->m_lblk + map_len;
3254
3255         depth = ext_depth(inode);
3256         eh = path[depth].p_hdr;
3257         ex = path[depth].p_ext;
3258         ee_block = le32_to_cpu(ex->ee_block);
3259         ee_len = ext4_ext_get_actual_len(ex);
3260         zero_ex.ee_len = 0;
3261
3262         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3263
3264         /* Pre-conditions */
3265         BUG_ON(!ext4_ext_is_uninitialized(ex));
3266         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3267
3268         /*
3269          * Attempt to transfer newly initialized blocks from the currently
3270          * uninitialized extent to its neighbor. This is much cheaper
3271          * than an insertion followed by a merge as those involve costly
3272          * memmove() calls. Transferring to the left is the common case in
3273          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3274          * followed by append writes.
3275          *
3276          * Limitations of the current logic:
3277          *  - L1: we do not deal with writes covering the whole extent.
3278          *    This would require removing the extent if the transfer
3279          *    is possible.
3280          *  - L2: we only attempt to merge with an extent stored in the
3281          *    same extent tree node.
3282          */
3283         if ((map->m_lblk == ee_block) &&
3284                 /* See if we can merge left */
3285                 (map_len < ee_len) &&           /*L1*/
3286                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3287                 ext4_lblk_t prev_lblk;
3288                 ext4_fsblk_t prev_pblk, ee_pblk;
3289                 unsigned int prev_len;
3290
3291                 abut_ex = ex - 1;
3292                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3293                 prev_len = ext4_ext_get_actual_len(abut_ex);
3294                 prev_pblk = ext4_ext_pblock(abut_ex);
3295                 ee_pblk = ext4_ext_pblock(ex);
3296
3297                 /*
3298                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3299                  * upon those conditions:
3300                  * - C1: abut_ex is initialized,
3301                  * - C2: abut_ex is logically abutting ex,
3302                  * - C3: abut_ex is physically abutting ex,
3303                  * - C4: abut_ex can receive the additional blocks without
3304                  *   overflowing the (initialized) length limit.
3305                  */
3306                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3307                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3308                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3309                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3310                         err = ext4_ext_get_access(handle, inode, path + depth);
3311                         if (err)
3312                                 goto out;
3313
3314                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3315                                 map, ex, abut_ex);
3316
3317                         /* Shift the start of ex by 'map_len' blocks */
3318                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3319                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3320                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3321                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3322
3323                         /* Extend abut_ex by 'map_len' blocks */
3324                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3325
3326                         /* Result: number of initialized blocks past m_lblk */
3327                         allocated = map_len;
3328                 }
3329         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3330                    (map_len < ee_len) &&        /*L1*/
3331                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3332                 /* See if we can merge right */
3333                 ext4_lblk_t next_lblk;
3334                 ext4_fsblk_t next_pblk, ee_pblk;
3335                 unsigned int next_len;
3336
3337                 abut_ex = ex + 1;
3338                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3339                 next_len = ext4_ext_get_actual_len(abut_ex);
3340                 next_pblk = ext4_ext_pblock(abut_ex);
3341                 ee_pblk = ext4_ext_pblock(ex);
3342
3343                 /*
3344                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3345                  * upon those conditions:
3346                  * - C1: abut_ex is initialized,
3347                  * - C2: abut_ex is logically abutting ex,
3348                  * - C3: abut_ex is physically abutting ex,
3349                  * - C4: abut_ex can receive the additional blocks without
3350                  *   overflowing the (initialized) length limit.
3351                  */
3352                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3353                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3354                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3355                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3356                         err = ext4_ext_get_access(handle, inode, path + depth);
3357                         if (err)
3358                                 goto out;
3359
3360                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3361                                 map, ex, abut_ex);
3362
3363                         /* Shift the start of abut_ex by 'map_len' blocks */
3364                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3365                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3366                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3367                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3368
3369                         /* Extend abut_ex by 'map_len' blocks */
3370                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3371
3372                         /* Result: number of initialized blocks past m_lblk */
3373                         allocated = map_len;
3374                 }
3375         }
3376         if (allocated) {
3377                 /* Mark the block containing both extents as dirty */
3378                 ext4_ext_dirty(handle, inode, path + depth);
3379
3380                 /* Update path to point to the right extent */
3381                 path[depth].p_ext = abut_ex;
3382                 goto out;
3383         } else
3384                 allocated = ee_len - (map->m_lblk - ee_block);
3385
3386         WARN_ON(map->m_lblk < ee_block);
3387         /*
3388          * It is safe to convert extent to initialized via explicit
3389          * zeroout only if extent is fully insde i_size or new_size.
3390          */
3391         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3392
3393         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3394                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3395                         (inode->i_sb->s_blocksize_bits - 10);
3396
3397         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3398         if (max_zeroout && (ee_len <= max_zeroout)) {
3399                 err = ext4_ext_zeroout(inode, ex);
3400                 if (err)
3401                         goto out;
3402                 zero_ex.ee_block = ex->ee_block;
3403                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3404                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3405
3406                 err = ext4_ext_get_access(handle, inode, path + depth);
3407                 if (err)
3408                         goto out;
3409                 ext4_ext_mark_initialized(ex);
3410                 ext4_ext_try_to_merge(handle, inode, path, ex);
3411                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3412                 goto out;
3413         }
3414
3415         /*
3416          * four cases:
3417          * 1. split the extent into three extents.
3418          * 2. split the extent into two extents, zeroout the first half.
3419          * 3. split the extent into two extents, zeroout the second half.
3420          * 4. split the extent into two extents with out zeroout.
3421          */
3422         split_map.m_lblk = map->m_lblk;
3423         split_map.m_len = map->m_len;
3424
3425         if (max_zeroout && (allocated > map->m_len)) {
3426                 if (allocated <= max_zeroout) {
3427                         /* case 3 */
3428                         zero_ex.ee_block =
3429                                          cpu_to_le32(map->m_lblk);
3430                         zero_ex.ee_len = cpu_to_le16(allocated);
3431                         ext4_ext_store_pblock(&zero_ex,
3432                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3433                         err = ext4_ext_zeroout(inode, &zero_ex);
3434                         if (err)
3435                                 goto out;
3436                         split_map.m_lblk = map->m_lblk;
3437                         split_map.m_len = allocated;
3438                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3439                         /* case 2 */
3440                         if (map->m_lblk != ee_block) {
3441                                 zero_ex.ee_block = ex->ee_block;
3442                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3443                                                         ee_block);
3444                                 ext4_ext_store_pblock(&zero_ex,
3445                                                       ext4_ext_pblock(ex));
3446                                 err = ext4_ext_zeroout(inode, &zero_ex);
3447                                 if (err)
3448                                         goto out;
3449                         }
3450
3451                         split_map.m_lblk = ee_block;
3452                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3453                         allocated = map->m_len;
3454                 }
3455         }
3456
3457         allocated = ext4_split_extent(handle, inode, path,
3458                                       &split_map, split_flag, flags);
3459         if (allocated < 0)
3460                 err = allocated;
3461
3462 out:
3463         /* If we have gotten a failure, don't zero out status tree */
3464         if (!err)
3465                 err = ext4_es_zeroout(inode, &zero_ex);
3466         return err ? err : allocated;
3467 }
3468
3469 /*
3470  * This function is called by ext4_ext_map_blocks() from
3471  * ext4_get_blocks_dio_write() when DIO to write
3472  * to an uninitialized extent.
3473  *
3474  * Writing to an uninitialized extent may result in splitting the uninitialized
3475  * extent into multiple initialized/uninitialized extents (up to three)
3476  * There are three possibilities:
3477  *   a> There is no split required: Entire extent should be uninitialized
3478  *   b> Splits in two extents: Write is happening at either end of the extent
3479  *   c> Splits in three extents: Somone is writing in middle of the extent
3480  *
3481  * One of more index blocks maybe needed if the extent tree grow after
3482  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3483  * complete, we need to split the uninitialized extent before DIO submit
3484  * the IO. The uninitialized extent called at this time will be split
3485  * into three uninitialized extent(at most). After IO complete, the part
3486  * being filled will be convert to initialized by the end_io callback function
3487  * via ext4_convert_unwritten_extents().
3488  *
3489  * Returns the size of uninitialized extent to be written on success.
3490  */
3491 static int ext4_split_unwritten_extents(handle_t *handle,
3492                                         struct inode *inode,
3493                                         struct ext4_map_blocks *map,
3494                                         struct ext4_ext_path *path,
3495                                         int flags)
3496 {
3497         ext4_lblk_t eof_block;
3498         ext4_lblk_t ee_block;
3499         struct ext4_extent *ex;
3500         unsigned int ee_len;
3501         int split_flag = 0, depth;
3502
3503         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3504                 "block %llu, max_blocks %u\n", inode->i_ino,
3505                 (unsigned long long)map->m_lblk, map->m_len);
3506
3507         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3508                 inode->i_sb->s_blocksize_bits;
3509         if (eof_block < map->m_lblk + map->m_len)
3510                 eof_block = map->m_lblk + map->m_len;
3511         /*
3512          * It is safe to convert extent to initialized via explicit
3513          * zeroout only if extent is fully insde i_size or new_size.
3514          */
3515         depth = ext_depth(inode);
3516         ex = path[depth].p_ext;
3517         ee_block = le32_to_cpu(ex->ee_block);
3518         ee_len = ext4_ext_get_actual_len(ex);
3519
3520         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3521         split_flag |= EXT4_EXT_MARK_UNINIT2;
3522         if (flags & EXT4_GET_BLOCKS_CONVERT)
3523                 split_flag |= EXT4_EXT_DATA_VALID2;
3524         flags |= EXT4_GET_BLOCKS_PRE_IO;
3525         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3526 }
3527
3528 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3529                                                 struct inode *inode,
3530                                                 struct ext4_map_blocks *map,
3531                                                 struct ext4_ext_path *path)
3532 {
3533         struct ext4_extent *ex;
3534         ext4_lblk_t ee_block;
3535         unsigned int ee_len;
3536         int depth;
3537         int err = 0;
3538
3539         depth = ext_depth(inode);
3540         ex = path[depth].p_ext;
3541         ee_block = le32_to_cpu(ex->ee_block);
3542         ee_len = ext4_ext_get_actual_len(ex);
3543
3544         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3545                 "block %llu, max_blocks %u\n", inode->i_ino,
3546                   (unsigned long long)ee_block, ee_len);
3547
3548         /* If extent is larger than requested it is a clear sign that we still
3549          * have some extent state machine issues left. So extent_split is still
3550          * required.
3551          * TODO: Once all related issues will be fixed this situation should be
3552          * illegal.
3553          */
3554         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3555 #ifdef EXT4_DEBUG
3556                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3557                              " len %u; IO logical block %llu, len %u\n",
3558                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3559                              (unsigned long long)map->m_lblk, map->m_len);
3560 #endif
3561                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3562                                                    EXT4_GET_BLOCKS_CONVERT);
3563                 if (err < 0)
3564                         goto out;
3565                 ext4_ext_drop_refs(path);
3566                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3567                 if (IS_ERR(path)) {
3568                         err = PTR_ERR(path);
3569                         goto out;
3570                 }
3571                 depth = ext_depth(inode);
3572                 ex = path[depth].p_ext;
3573         }
3574
3575         err = ext4_ext_get_access(handle, inode, path + depth);
3576         if (err)
3577                 goto out;
3578         /* first mark the extent as initialized */
3579         ext4_ext_mark_initialized(ex);
3580
3581         /* note: ext4_ext_correct_indexes() isn't needed here because
3582          * borders are not changed
3583          */
3584         ext4_ext_try_to_merge(handle, inode, path, ex);
3585
3586         /* Mark modified extent as dirty */
3587         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3588 out:
3589         ext4_ext_show_leaf(inode, path);
3590         return err;
3591 }
3592
3593 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3594                         sector_t block, int count)
3595 {
3596         int i;
3597         for (i = 0; i < count; i++)
3598                 unmap_underlying_metadata(bdev, block + i);
3599 }
3600
3601 /*
3602  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3603  */
3604 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3605                               ext4_lblk_t lblk,
3606                               struct ext4_ext_path *path,
3607                               unsigned int len)
3608 {
3609         int i, depth;
3610         struct ext4_extent_header *eh;
3611         struct ext4_extent *last_ex;
3612
3613         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3614                 return 0;
3615
3616         depth = ext_depth(inode);
3617         eh = path[depth].p_hdr;
3618
3619         /*
3620          * We're going to remove EOFBLOCKS_FL entirely in future so we
3621          * do not care for this case anymore. Simply remove the flag
3622          * if there are no extents.
3623          */
3624         if (unlikely(!eh->eh_entries))
3625                 goto out;
3626         last_ex = EXT_LAST_EXTENT(eh);
3627         /*
3628          * We should clear the EOFBLOCKS_FL flag if we are writing the
3629          * last block in the last extent in the file.  We test this by
3630          * first checking to see if the caller to
3631          * ext4_ext_get_blocks() was interested in the last block (or
3632          * a block beyond the last block) in the current extent.  If
3633          * this turns out to be false, we can bail out from this
3634          * function immediately.
3635          */
3636         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3637             ext4_ext_get_actual_len(last_ex))
3638                 return 0;
3639         /*
3640          * If the caller does appear to be planning to write at or
3641          * beyond the end of the current extent, we then test to see
3642          * if the current extent is the last extent in the file, by
3643          * checking to make sure it was reached via the rightmost node
3644          * at each level of the tree.
3645          */
3646         for (i = depth-1; i >= 0; i--)
3647                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3648                         return 0;
3649 out:
3650         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3651         return ext4_mark_inode_dirty(handle, inode);
3652 }
3653
3654 /**
3655  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3656  *
3657  * Return 1 if there is a delalloc block in the range, otherwise 0.
3658  */
3659 int ext4_find_delalloc_range(struct inode *inode,
3660                              ext4_lblk_t lblk_start,
3661                              ext4_lblk_t lblk_end)
3662 {
3663         struct extent_status es;
3664
3665         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3666         if (es.es_len == 0)
3667                 return 0; /* there is no delay extent in this tree */
3668         else if (es.es_lblk <= lblk_start &&
3669                  lblk_start < es.es_lblk + es.es_len)
3670                 return 1;
3671         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3672                 return 1;
3673         else
3674                 return 0;
3675 }
3676
3677 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3678 {
3679         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3680         ext4_lblk_t lblk_start, lblk_end;
3681         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3682         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3683
3684         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3685 }
3686
3687 /**
3688  * Determines how many complete clusters (out of those specified by the 'map')
3689  * are under delalloc and were reserved quota for.
3690  * This function is called when we are writing out the blocks that were
3691  * originally written with their allocation delayed, but then the space was
3692  * allocated using fallocate() before the delayed allocation could be resolved.
3693  * The cases to look for are:
3694  * ('=' indicated delayed allocated blocks
3695  *  '-' indicates non-delayed allocated blocks)
3696  * (a) partial clusters towards beginning and/or end outside of allocated range
3697  *     are not delalloc'ed.
3698  *      Ex:
3699  *      |----c---=|====c====|====c====|===-c----|
3700  *               |++++++ allocated ++++++|
3701  *      ==> 4 complete clusters in above example
3702  *
3703  * (b) partial cluster (outside of allocated range) towards either end is
3704  *     marked for delayed allocation. In this case, we will exclude that
3705  *     cluster.
3706  *      Ex:
3707  *      |----====c========|========c========|
3708  *           |++++++ allocated ++++++|
3709  *      ==> 1 complete clusters in above example
3710  *
3711  *      Ex:
3712  *      |================c================|
3713  *            |++++++ allocated ++++++|
3714  *      ==> 0 complete clusters in above example
3715  *
3716  * The ext4_da_update_reserve_space will be called only if we
3717  * determine here that there were some "entire" clusters that span
3718  * this 'allocated' range.
3719  * In the non-bigalloc case, this function will just end up returning num_blks
3720  * without ever calling ext4_find_delalloc_range.
3721  */
3722 static unsigned int
3723 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3724                            unsigned int num_blks)
3725 {
3726         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3727         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3728         ext4_lblk_t lblk_from, lblk_to, c_offset;
3729         unsigned int allocated_clusters = 0;
3730
3731         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3732         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3733
3734         /* max possible clusters for this allocation */
3735         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3736
3737         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3738
3739         /* Check towards left side */
3740         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3741         if (c_offset) {
3742                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3743                 lblk_to = lblk_from + c_offset - 1;
3744
3745                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3746                         allocated_clusters--;
3747         }
3748
3749         /* Now check towards right. */
3750         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3751         if (allocated_clusters && c_offset) {
3752                 lblk_from = lblk_start + num_blks;
3753                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3754
3755                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3756                         allocated_clusters--;
3757         }
3758
3759         return allocated_clusters;
3760 }
3761
3762 static int
3763 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3764                         struct ext4_map_blocks *map,
3765                         struct ext4_ext_path *path, int flags,
3766                         unsigned int allocated, ext4_fsblk_t newblock)
3767 {
3768         int ret = 0;
3769         int err = 0;
3770         ext4_io_end_t *io = ext4_inode_aio(inode);
3771
3772         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3773                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3774                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3775                   flags, allocated);
3776         ext4_ext_show_leaf(inode, path);
3777
3778         /*
3779          * When writing into uninitialized space, we should not fail to
3780          * allocate metadata blocks for the new extent block if needed.
3781          */
3782         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
3783
3784         trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
3785                                                     allocated, newblock);
3786
3787         /* get_block() before submit the IO, split the extent */
3788         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3789                 ret = ext4_split_unwritten_extents(handle, inode, map,
3790                                                    path, flags);
3791                 if (ret <= 0)
3792                         goto out;
3793                 /*
3794                  * Flag the inode(non aio case) or end_io struct (aio case)
3795                  * that this IO needs to conversion to written when IO is
3796                  * completed
3797                  */
3798                 if (io)
3799                         ext4_set_io_unwritten_flag(inode, io);
3800                 else
3801                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3802                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3803                 if (ext4_should_dioread_nolock(inode))
3804                         map->m_flags |= EXT4_MAP_UNINIT;
3805                 goto out;
3806         }
3807         /* IO end_io complete, convert the filled extent to written */
3808         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3809                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3810                                                         path);
3811                 if (ret >= 0) {
3812                         ext4_update_inode_fsync_trans(handle, inode, 1);
3813                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3814                                                  path, map->m_len);
3815                 } else
3816                         err = ret;
3817                 map->m_flags |= EXT4_MAP_MAPPED;
3818                 if (allocated > map->m_len)
3819                         allocated = map->m_len;
3820                 map->m_len = allocated;
3821                 goto out2;
3822         }
3823         /* buffered IO case */
3824         /*
3825          * repeat fallocate creation request
3826          * we already have an unwritten extent
3827          */
3828         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT) {
3829                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3830                 goto map_out;
3831         }
3832
3833         /* buffered READ or buffered write_begin() lookup */
3834         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3835                 /*
3836                  * We have blocks reserved already.  We
3837                  * return allocated blocks so that delalloc
3838                  * won't do block reservation for us.  But
3839                  * the buffer head will be unmapped so that
3840                  * a read from the block returns 0s.
3841                  */
3842                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3843                 goto out1;
3844         }
3845
3846         /* buffered write, writepage time, convert*/
3847         ret = ext4_ext_convert_to_initialized(handle, inode, map, path, flags);
3848         if (ret >= 0)
3849                 ext4_update_inode_fsync_trans(handle, inode, 1);
3850 out:
3851         if (ret <= 0) {
3852                 err = ret;
3853                 goto out2;
3854         } else
3855                 allocated = ret;
3856         map->m_flags |= EXT4_MAP_NEW;
3857         /*
3858          * if we allocated more blocks than requested
3859          * we need to make sure we unmap the extra block
3860          * allocated. The actual needed block will get
3861          * unmapped later when we find the buffer_head marked
3862          * new.
3863          */
3864         if (allocated > map->m_len) {
3865                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3866                                         newblock + map->m_len,
3867                                         allocated - map->m_len);
3868                 allocated = map->m_len;
3869         }
3870         map->m_len = allocated;
3871
3872         /*
3873          * If we have done fallocate with the offset that is already
3874          * delayed allocated, we would have block reservation
3875          * and quota reservation done in the delayed write path.
3876          * But fallocate would have already updated quota and block
3877          * count for this offset. So cancel these reservation
3878          */
3879         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3880                 unsigned int reserved_clusters;
3881                 reserved_clusters = get_reserved_cluster_alloc(inode,
3882                                 map->m_lblk, map->m_len);
3883                 if (reserved_clusters)
3884                         ext4_da_update_reserve_space(inode,
3885                                                      reserved_clusters,
3886                                                      0);
3887         }
3888
3889 map_out:
3890         map->m_flags |= EXT4_MAP_MAPPED;
3891         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3892                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3893                                          map->m_len);
3894                 if (err < 0)
3895                         goto out2;
3896         }
3897 out1:
3898         if (allocated > map->m_len)
3899                 allocated = map->m_len;
3900         ext4_ext_show_leaf(inode, path);
3901         map->m_pblk = newblock;
3902         map->m_len = allocated;
3903 out2:
3904         if (path) {
3905                 ext4_ext_drop_refs(path);
3906                 kfree(path);
3907         }
3908         return err ? err : allocated;
3909 }
3910
3911 /*
3912  * get_implied_cluster_alloc - check to see if the requested
3913  * allocation (in the map structure) overlaps with a cluster already
3914  * allocated in an extent.
3915  *      @sb     The filesystem superblock structure
3916  *      @map    The requested lblk->pblk mapping
3917  *      @ex     The extent structure which might contain an implied
3918  *                      cluster allocation
3919  *
3920  * This function is called by ext4_ext_map_blocks() after we failed to
3921  * find blocks that were already in the inode's extent tree.  Hence,
3922  * we know that the beginning of the requested region cannot overlap
3923  * the extent from the inode's extent tree.  There are three cases we
3924  * want to catch.  The first is this case:
3925  *
3926  *               |--- cluster # N--|
3927  *    |--- extent ---|  |---- requested region ---|
3928  *                      |==========|
3929  *
3930  * The second case that we need to test for is this one:
3931  *
3932  *   |--------- cluster # N ----------------|
3933  *         |--- requested region --|   |------- extent ----|
3934  *         |=======================|
3935  *
3936  * The third case is when the requested region lies between two extents
3937  * within the same cluster:
3938  *          |------------- cluster # N-------------|
3939  * |----- ex -----|                  |---- ex_right ----|
3940  *                  |------ requested region ------|
3941  *                  |================|
3942  *
3943  * In each of the above cases, we need to set the map->m_pblk and
3944  * map->m_len so it corresponds to the return the extent labelled as
3945  * "|====|" from cluster #N, since it is already in use for data in
3946  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3947  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3948  * as a new "allocated" block region.  Otherwise, we will return 0 and
3949  * ext4_ext_map_blocks() will then allocate one or more new clusters
3950  * by calling ext4_mb_new_blocks().
3951  */
3952 static int get_implied_cluster_alloc(struct super_block *sb,
3953                                      struct ext4_map_blocks *map,
3954                                      struct ext4_extent *ex,
3955                                      struct ext4_ext_path *path)
3956 {
3957         struct ext4_sb_info *sbi = EXT4_SB(sb);
3958         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3959         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3960         ext4_lblk_t rr_cluster_start;
3961         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3962         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3963         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3964
3965         /* The extent passed in that we are trying to match */
3966         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3967         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3968
3969         /* The requested region passed into ext4_map_blocks() */
3970         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3971
3972         if ((rr_cluster_start == ex_cluster_end) ||
3973             (rr_cluster_start == ex_cluster_start)) {
3974                 if (rr_cluster_start == ex_cluster_end)
3975                         ee_start += ee_len - 1;
3976                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3977                         c_offset;
3978                 map->m_len = min(map->m_len,
3979                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3980                 /*
3981                  * Check for and handle this case:
3982                  *
3983                  *   |--------- cluster # N-------------|
3984                  *                     |------- extent ----|
3985                  *         |--- requested region ---|
3986                  *         |===========|
3987                  */
3988
3989                 if (map->m_lblk < ee_block)
3990                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3991
3992                 /*
3993                  * Check for the case where there is already another allocated
3994                  * block to the right of 'ex' but before the end of the cluster.
3995                  *
3996                  *          |------------- cluster # N-------------|
3997                  * |----- ex -----|                  |---- ex_right ----|
3998                  *                  |------ requested region ------|
3999                  *                  |================|
4000                  */
4001                 if (map->m_lblk > ee_block) {
4002                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4003                         map->m_len = min(map->m_len, next - map->m_lblk);
4004                 }
4005
4006                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4007                 return 1;
4008         }
4009
4010         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4011         return 0;
4012 }
4013
4014
4015 /*
4016  * Block allocation/map/preallocation routine for extents based files
4017  *
4018  *
4019  * Need to be called with
4020  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4021  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4022  *
4023  * return > 0, number of of blocks already mapped/allocated
4024  *          if create == 0 and these are pre-allocated blocks
4025  *              buffer head is unmapped
4026  *          otherwise blocks are mapped
4027  *
4028  * return = 0, if plain look up failed (blocks have not been allocated)
4029  *          buffer head is unmapped
4030  *
4031  * return < 0, error case.
4032  */
4033 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4034                         struct ext4_map_blocks *map, int flags)
4035 {
4036         struct ext4_ext_path *path = NULL;
4037         struct ext4_extent newex, *ex, *ex2;
4038         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4039         ext4_fsblk_t newblock = 0;
4040         int free_on_err = 0, err = 0, depth;
4041         unsigned int allocated = 0, offset = 0;
4042         unsigned int allocated_clusters = 0;
4043         struct ext4_allocation_request ar;
4044         ext4_io_end_t *io = ext4_inode_aio(inode);
4045         ext4_lblk_t cluster_offset;
4046         int set_unwritten = 0;
4047
4048         ext_debug("blocks %u/%u requested for inode %lu\n",
4049                   map->m_lblk, map->m_len, inode->i_ino);
4050         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4051
4052         /* find extent for this block */
4053         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
4054         if (IS_ERR(path)) {
4055                 err = PTR_ERR(path);
4056                 path = NULL;
4057                 goto out2;
4058         }
4059
4060         depth = ext_depth(inode);
4061
4062         /*
4063          * consistent leaf must not be empty;
4064          * this situation is possible, though, _during_ tree modification;
4065          * this is why assert can't be put in ext4_ext_find_extent()
4066          */
4067         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4068                 EXT4_ERROR_INODE(inode, "bad extent address "
4069                                  "lblock: %lu, depth: %d pblock %lld",
4070                                  (unsigned long) map->m_lblk, depth,
4071                                  path[depth].p_block);
4072                 err = -EIO;
4073                 goto out2;
4074         }
4075
4076         ex = path[depth].p_ext;
4077         if (ex) {
4078                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4079                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4080                 unsigned short ee_len;
4081
4082                 /*
4083                  * Uninitialized extents are treated as holes, except that
4084                  * we split out initialized portions during a write.
4085                  */
4086                 ee_len = ext4_ext_get_actual_len(ex);
4087
4088                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4089
4090                 /* if found extent covers block, simply return it */
4091                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4092                         newblock = map->m_lblk - ee_block + ee_start;
4093                         /* number of remaining blocks in the extent */
4094                         allocated = ee_len - (map->m_lblk - ee_block);
4095                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4096                                   ee_block, ee_len, newblock);
4097
4098                         if (!ext4_ext_is_uninitialized(ex))
4099                                 goto out;
4100
4101                         allocated = ext4_ext_handle_uninitialized_extents(
4102                                 handle, inode, map, path, flags,
4103                                 allocated, newblock);
4104                         goto out3;
4105                 }
4106         }
4107
4108         if ((sbi->s_cluster_ratio > 1) &&
4109             ext4_find_delalloc_cluster(inode, map->m_lblk))
4110                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4111
4112         /*
4113          * requested block isn't allocated yet;
4114          * we couldn't try to create block if create flag is zero
4115          */
4116         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4117                 /*
4118                  * put just found gap into cache to speed up
4119                  * subsequent requests
4120                  */
4121                 if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
4122                         ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4123                 goto out2;
4124         }
4125
4126         /*
4127          * Okay, we need to do block allocation.
4128          */
4129         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
4130         newex.ee_block = cpu_to_le32(map->m_lblk);
4131         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
4132
4133         /*
4134          * If we are doing bigalloc, check to see if the extent returned
4135          * by ext4_ext_find_extent() implies a cluster we can use.
4136          */
4137         if (cluster_offset && ex &&
4138             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4139                 ar.len = allocated = map->m_len;
4140                 newblock = map->m_pblk;
4141                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4142                 goto got_allocated_blocks;
4143         }
4144
4145         /* find neighbour allocated blocks */
4146         ar.lleft = map->m_lblk;
4147         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4148         if (err)
4149                 goto out2;
4150         ar.lright = map->m_lblk;
4151         ex2 = NULL;
4152         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4153         if (err)
4154                 goto out2;
4155
4156         /* Check if the extent after searching to the right implies a
4157          * cluster we can use. */
4158         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4159             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4160                 ar.len = allocated = map->m_len;
4161                 newblock = map->m_pblk;
4162                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4163                 goto got_allocated_blocks;
4164         }
4165
4166         /*
4167          * See if request is beyond maximum number of blocks we can have in
4168          * a single extent. For an initialized extent this limit is
4169          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4170          * EXT_UNINIT_MAX_LEN.
4171          */
4172         if (map->m_len > EXT_INIT_MAX_LEN &&
4173             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4174                 map->m_len = EXT_INIT_MAX_LEN;
4175         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4176                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4177                 map->m_len = EXT_UNINIT_MAX_LEN;
4178
4179         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4180         newex.ee_len = cpu_to_le16(map->m_len);
4181         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4182         if (err)
4183                 allocated = ext4_ext_get_actual_len(&newex);
4184         else
4185                 allocated = map->m_len;
4186
4187         /* allocate new block */
4188         ar.inode = inode;
4189         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4190         ar.logical = map->m_lblk;
4191         /*
4192          * We calculate the offset from the beginning of the cluster
4193          * for the logical block number, since when we allocate a
4194          * physical cluster, the physical block should start at the
4195          * same offset from the beginning of the cluster.  This is
4196          * needed so that future calls to get_implied_cluster_alloc()
4197          * work correctly.
4198          */
4199         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
4200         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4201         ar.goal -= offset;
4202         ar.logical -= offset;
4203         if (S_ISREG(inode->i_mode))
4204                 ar.flags = EXT4_MB_HINT_DATA;
4205         else
4206                 /* disable in-core preallocation for non-regular files */
4207                 ar.flags = 0;
4208         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4209                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4210         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4211         if (!newblock)
4212                 goto out2;
4213         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4214                   ar.goal, newblock, allocated);
4215         free_on_err = 1;
4216         allocated_clusters = ar.len;
4217         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4218         if (ar.len > allocated)
4219                 ar.len = allocated;
4220
4221 got_allocated_blocks:
4222         /* try to insert new extent into found leaf and return */
4223         ext4_ext_store_pblock(&newex, newblock + offset);
4224         newex.ee_len = cpu_to_le16(ar.len);
4225         /* Mark uninitialized */
4226         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4227                 ext4_ext_mark_uninitialized(&newex);
4228                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4229                 /*
4230                  * io_end structure was created for every IO write to an
4231                  * uninitialized extent. To avoid unnecessary conversion,
4232                  * here we flag the IO that really needs the conversion.
4233                  * For non asycn direct IO case, flag the inode state
4234                  * that we need to perform conversion when IO is done.
4235                  */
4236                 if ((flags & EXT4_GET_BLOCKS_PRE_IO))
4237                         set_unwritten = 1;
4238                 if (ext4_should_dioread_nolock(inode))
4239                         map->m_flags |= EXT4_MAP_UNINIT;
4240         }
4241
4242         err = 0;
4243         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4244                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4245                                          path, ar.len);
4246         if (!err)
4247                 err = ext4_ext_insert_extent(handle, inode, path,
4248                                              &newex, flags);
4249
4250         if (!err && set_unwritten) {
4251                 if (io)
4252                         ext4_set_io_unwritten_flag(inode, io);
4253                 else
4254                         ext4_set_inode_state(inode,
4255                                              EXT4_STATE_DIO_UNWRITTEN);
4256         }
4257
4258         if (err && free_on_err) {
4259                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4260                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4261                 /* free data blocks we just allocated */
4262                 /* not a good idea to call discard here directly,
4263                  * but otherwise we'd need to call it every free() */
4264                 ext4_discard_preallocations(inode);
4265                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4266                                  ext4_ext_get_actual_len(&newex), fb_flags);
4267                 goto out2;
4268         }
4269
4270         /* previous routine could use block we allocated */
4271         newblock = ext4_ext_pblock(&newex);
4272         allocated = ext4_ext_get_actual_len(&newex);
4273         if (allocated > map->m_len)
4274                 allocated = map->m_len;
4275         map->m_flags |= EXT4_MAP_NEW;
4276
4277         /*
4278          * Update reserved blocks/metadata blocks after successful
4279          * block allocation which had been deferred till now.
4280          */
4281         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4282                 unsigned int reserved_clusters;
4283                 /*
4284                  * Check how many clusters we had reserved this allocated range
4285                  */
4286                 reserved_clusters = get_reserved_cluster_alloc(inode,
4287                                                 map->m_lblk, allocated);
4288                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4289                         if (reserved_clusters) {
4290                                 /*
4291                                  * We have clusters reserved for this range.
4292                                  * But since we are not doing actual allocation
4293                                  * and are simply using blocks from previously
4294                                  * allocated cluster, we should release the
4295                                  * reservation and not claim quota.
4296                                  */
4297                                 ext4_da_update_reserve_space(inode,
4298                                                 reserved_clusters, 0);
4299                         }
4300                 } else {
4301                         BUG_ON(allocated_clusters < reserved_clusters);
4302                         if (reserved_clusters < allocated_clusters) {
4303                                 struct ext4_inode_info *ei = EXT4_I(inode);
4304                                 int reservation = allocated_clusters -
4305                                                   reserved_clusters;
4306                                 /*
4307                                  * It seems we claimed few clusters outside of
4308                                  * the range of this allocation. We should give
4309                                  * it back to the reservation pool. This can
4310                                  * happen in the following case:
4311                                  *
4312                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4313                                  *   cluster has 4 blocks. Thus, the clusters
4314                                  *   are [0-3],[4-7],[8-11]...
4315                                  * * First comes delayed allocation write for
4316                                  *   logical blocks 10 & 11. Since there were no
4317                                  *   previous delayed allocated blocks in the
4318                                  *   range [8-11], we would reserve 1 cluster
4319                                  *   for this write.
4320                                  * * Next comes write for logical blocks 3 to 8.
4321                                  *   In this case, we will reserve 2 clusters
4322                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4323                                  *   that range has a delayed allocated blocks.
4324                                  *   Thus total reserved clusters now becomes 3.
4325                                  * * Now, during the delayed allocation writeout
4326                                  *   time, we will first write blocks [3-8] and
4327                                  *   allocate 3 clusters for writing these
4328                                  *   blocks. Also, we would claim all these
4329                                  *   three clusters above.
4330                                  * * Now when we come here to writeout the
4331                                  *   blocks [10-11], we would expect to claim
4332                                  *   the reservation of 1 cluster we had made
4333                                  *   (and we would claim it since there are no
4334                                  *   more delayed allocated blocks in the range
4335                                  *   [8-11]. But our reserved cluster count had
4336                                  *   already gone to 0.
4337                                  *
4338                                  *   Thus, at the step 4 above when we determine
4339                                  *   that there are still some unwritten delayed
4340                                  *   allocated blocks outside of our current
4341                                  *   block range, we should increment the
4342                                  *   reserved clusters count so that when the
4343                                  *   remaining blocks finally gets written, we
4344                                  *   could claim them.
4345                                  */
4346                                 dquot_reserve_block(inode,
4347                                                 EXT4_C2B(sbi, reservation));
4348                                 spin_lock(&ei->i_block_reservation_lock);
4349                                 ei->i_reserved_data_blocks += reservation;
4350                                 spin_unlock(&ei->i_block_reservation_lock);
4351                         }
4352                         /*
4353                          * We will claim quota for all newly allocated blocks.
4354                          * We're updating the reserved space *after* the
4355                          * correction above so we do not accidentally free
4356                          * all the metadata reservation because we might
4357                          * actually need it later on.
4358                          */
4359                         ext4_da_update_reserve_space(inode, allocated_clusters,
4360                                                         1);
4361                 }
4362         }
4363
4364         /*
4365          * Cache the extent and update transaction to commit on fdatasync only
4366          * when it is _not_ an uninitialized extent.
4367          */
4368         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0)
4369                 ext4_update_inode_fsync_trans(handle, inode, 1);
4370         else
4371                 ext4_update_inode_fsync_trans(handle, inode, 0);
4372 out:
4373         if (allocated > map->m_len)
4374                 allocated = map->m_len;
4375         ext4_ext_show_leaf(inode, path);
4376         map->m_flags |= EXT4_MAP_MAPPED;
4377         map->m_pblk = newblock;
4378         map->m_len = allocated;
4379 out2:
4380         if (path) {
4381                 ext4_ext_drop_refs(path);
4382                 kfree(path);
4383         }
4384
4385 out3:
4386         trace_ext4_ext_map_blocks_exit(inode, map, err ? err : allocated);
4387
4388         return err ? err : allocated;
4389 }
4390
4391 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4392 {
4393         struct super_block *sb = inode->i_sb;
4394         ext4_lblk_t last_block;
4395         int err = 0;
4396
4397         /*
4398          * TODO: optimization is possible here.
4399          * Probably we need not scan at all,
4400          * because page truncation is enough.
4401          */
4402
4403         /* we have to know where to truncate from in crash case */
4404         EXT4_I(inode)->i_disksize = inode->i_size;
4405         ext4_mark_inode_dirty(handle, inode);
4406
4407         last_block = (inode->i_size + sb->s_blocksize - 1)
4408                         >> EXT4_BLOCK_SIZE_BITS(sb);
4409         err = ext4_es_remove_extent(inode, last_block,
4410                                     EXT_MAX_BLOCKS - last_block);
4411         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4412 }
4413
4414 static void ext4_falloc_update_inode(struct inode *inode,
4415                                 int mode, loff_t new_size, int update_ctime)
4416 {
4417         struct timespec now;
4418
4419         if (update_ctime) {
4420                 now = current_fs_time(inode->i_sb);
4421                 if (!timespec_equal(&inode->i_ctime, &now))
4422                         inode->i_ctime = now;
4423         }
4424         /*
4425          * Update only when preallocation was requested beyond
4426          * the file size.
4427          */
4428         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4429                 if (new_size > i_size_read(inode))
4430                         i_size_write(inode, new_size);
4431                 if (new_size > EXT4_I(inode)->i_disksize)
4432                         ext4_update_i_disksize(inode, new_size);
4433         } else {
4434                 /*
4435                  * Mark that we allocate beyond EOF so the subsequent truncate
4436                  * can proceed even if the new size is the same as i_size.
4437                  */
4438                 if (new_size > i_size_read(inode))
4439                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4440         }
4441
4442 }
4443
4444 /*
4445  * preallocate space for a file. This implements ext4's fallocate file
4446  * operation, which gets called from sys_fallocate system call.
4447  * For block-mapped files, posix_fallocate should fall back to the method
4448  * of writing zeroes to the required new blocks (the same behavior which is
4449  * expected for file systems which do not support fallocate() system call).
4450  */
4451 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4452 {
4453         struct inode *inode = file_inode(file);
4454         handle_t *handle;
4455         loff_t new_size;
4456         unsigned int max_blocks;
4457         int ret = 0;
4458         int ret2 = 0;
4459         int retries = 0;
4460         int flags;
4461         struct ext4_map_blocks map;
4462         unsigned int credits, blkbits = inode->i_blkbits;
4463
4464         /* Return error if mode is not supported */
4465         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4466                 return -EOPNOTSUPP;
4467
4468         if (mode & FALLOC_FL_PUNCH_HOLE)
4469                 return ext4_punch_hole(file, offset, len);
4470
4471         ret = ext4_convert_inline_data(inode);
4472         if (ret)
4473                 return ret;
4474
4475         /*
4476          * currently supporting (pre)allocate mode for extent-based
4477          * files _only_
4478          */
4479         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4480                 return -EOPNOTSUPP;
4481
4482         trace_ext4_fallocate_enter(inode, offset, len, mode);
4483         map.m_lblk = offset >> blkbits;
4484         /*
4485          * We can't just convert len to max_blocks because
4486          * If blocksize = 4096 offset = 3072 and len = 2048
4487          */
4488         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4489                 - map.m_lblk;
4490         /*
4491          * credits to insert 1 extent into extent tree
4492          */
4493         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4494         mutex_lock(&inode->i_mutex);
4495         ret = inode_newsize_ok(inode, (len + offset));
4496         if (ret) {
4497                 mutex_unlock(&inode->i_mutex);
4498                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4499                 return ret;
4500         }
4501         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4502         if (mode & FALLOC_FL_KEEP_SIZE)
4503                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4504         /*
4505          * Don't normalize the request if it can fit in one extent so
4506          * that it doesn't get unnecessarily split into multiple
4507          * extents.
4508          */
4509         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4510                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4511
4512 retry:
4513         while (ret >= 0 && ret < max_blocks) {
4514                 map.m_lblk = map.m_lblk + ret;
4515                 map.m_len = max_blocks = max_blocks - ret;
4516                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4517                                             credits);
4518                 if (IS_ERR(handle)) {
4519                         ret = PTR_ERR(handle);
4520                         break;
4521                 }
4522                 ret = ext4_map_blocks(handle, inode, &map, flags);
4523                 if (ret <= 0) {
4524 #ifdef EXT4FS_DEBUG
4525                         ext4_warning(inode->i_sb,
4526                                      "inode #%lu: block %u: len %u: "
4527                                      "ext4_ext_map_blocks returned %d",
4528                                      inode->i_ino, map.m_lblk,
4529                                      map.m_len, ret);
4530 #endif
4531                         ext4_mark_inode_dirty(handle, inode);
4532                         ret2 = ext4_journal_stop(handle);
4533                         break;
4534                 }
4535                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4536                                                 blkbits) >> blkbits))
4537                         new_size = offset + len;
4538                 else
4539                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4540
4541                 ext4_falloc_update_inode(inode, mode, new_size,
4542                                          (map.m_flags & EXT4_MAP_NEW));
4543                 ext4_mark_inode_dirty(handle, inode);
4544                 if ((file->f_flags & O_SYNC) && ret >= max_blocks)
4545                         ext4_handle_sync(handle);
4546                 ret2 = ext4_journal_stop(handle);
4547                 if (ret2)
4548                         break;
4549         }
4550         if (ret == -ENOSPC &&
4551                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4552                 ret = 0;
4553                 goto retry;
4554         }
4555         mutex_unlock(&inode->i_mutex);
4556         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4557                                 ret > 0 ? ret2 : ret);
4558         return ret > 0 ? ret2 : ret;
4559 }
4560
4561 /*
4562  * This function convert a range of blocks to written extents
4563  * The caller of this function will pass the start offset and the size.
4564  * all unwritten extents within this range will be converted to
4565  * written extents.
4566  *
4567  * This function is called from the direct IO end io call back
4568  * function, to convert the fallocated extents after IO is completed.
4569  * Returns 0 on success.
4570  */
4571 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4572                                     ssize_t len)
4573 {
4574         handle_t *handle;
4575         unsigned int max_blocks;
4576         int ret = 0;
4577         int ret2 = 0;
4578         struct ext4_map_blocks map;
4579         unsigned int credits, blkbits = inode->i_blkbits;
4580
4581         map.m_lblk = offset >> blkbits;
4582         /*
4583          * We can't just convert len to max_blocks because
4584          * If blocksize = 4096 offset = 3072 and len = 2048
4585          */
4586         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4587                       map.m_lblk);
4588         /*
4589          * credits to insert 1 extent into extent tree
4590          */
4591         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4592         while (ret >= 0 && ret < max_blocks) {
4593                 map.m_lblk += ret;
4594                 map.m_len = (max_blocks -= ret);
4595                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
4596                 if (IS_ERR(handle)) {
4597                         ret = PTR_ERR(handle);
4598                         break;
4599                 }
4600                 ret = ext4_map_blocks(handle, inode, &map,
4601                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4602                 if (ret <= 0)
4603                         ext4_warning(inode->i_sb,
4604                                      "inode #%lu: block %u: len %u: "
4605                                      "ext4_ext_map_blocks returned %d",
4606                                      inode->i_ino, map.m_lblk,
4607                                      map.m_len, ret);
4608                 ext4_mark_inode_dirty(handle, inode);
4609                 ret2 = ext4_journal_stop(handle);
4610                 if (ret <= 0 || ret2 )
4611                         break;
4612         }
4613         return ret > 0 ? ret2 : ret;
4614 }
4615
4616 /*
4617  * If newes is not existing extent (newes->ec_pblk equals zero) find
4618  * delayed extent at start of newes and update newes accordingly and
4619  * return start of the next delayed extent.
4620  *
4621  * If newes is existing extent (newes->ec_pblk is not equal zero)
4622  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4623  * extent found. Leave newes unmodified.
4624  */
4625 static int ext4_find_delayed_extent(struct inode *inode,
4626                                     struct extent_status *newes)
4627 {
4628         struct extent_status es;
4629         ext4_lblk_t block, next_del;
4630
4631         if (newes->es_pblk == 0) {
4632                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
4633                                 newes->es_lblk + newes->es_len - 1, &es);
4634
4635                 /*
4636                  * No extent in extent-tree contains block @newes->es_pblk,
4637                  * then the block may stay in 1)a hole or 2)delayed-extent.
4638                  */
4639                 if (es.es_len == 0)
4640                         /* A hole found. */
4641                         return 0;
4642
4643                 if (es.es_lblk > newes->es_lblk) {
4644                         /* A hole found. */
4645                         newes->es_len = min(es.es_lblk - newes->es_lblk,
4646                                             newes->es_len);
4647                         return 0;
4648                 }
4649
4650                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
4651         }
4652
4653         block = newes->es_lblk + newes->es_len;
4654         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
4655         if (es.es_len == 0)
4656                 next_del = EXT_MAX_BLOCKS;
4657         else
4658                 next_del = es.es_lblk;
4659
4660         return next_del;
4661 }
4662 /* fiemap flags we can handle specified here */
4663 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4664
4665 static int ext4_xattr_fiemap(struct inode *inode,
4666                                 struct fiemap_extent_info *fieinfo)
4667 {
4668         __u64 physical = 0;
4669         __u64 length;
4670         __u32 flags = FIEMAP_EXTENT_LAST;
4671         int blockbits = inode->i_sb->s_blocksize_bits;
4672         int error = 0;
4673
4674         /* in-inode? */
4675         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4676                 struct ext4_iloc iloc;
4677                 int offset;     /* offset of xattr in inode */
4678
4679                 error = ext4_get_inode_loc(inode, &iloc);
4680                 if (error)
4681                         return error;
4682                 physical = iloc.bh->b_blocknr << blockbits;
4683                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4684                                 EXT4_I(inode)->i_extra_isize;
4685                 physical += offset;
4686                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4687                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4688                 brelse(iloc.bh);
4689         } else { /* external block */
4690                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4691                 length = inode->i_sb->s_blocksize;
4692         }
4693
4694         if (physical)
4695                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4696                                                 length, flags);
4697         return (error < 0 ? error : 0);
4698 }
4699
4700 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4701                 __u64 start, __u64 len)
4702 {
4703         ext4_lblk_t start_blk;
4704         int error = 0;
4705
4706         if (ext4_has_inline_data(inode)) {
4707                 int has_inline = 1;
4708
4709                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
4710
4711                 if (has_inline)
4712                         return error;
4713         }
4714
4715         /* fallback to generic here if not in extents fmt */
4716         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4717                 return generic_block_fiemap(inode, fieinfo, start, len,
4718                         ext4_get_block);
4719
4720         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4721                 return -EBADR;
4722
4723         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4724                 error = ext4_xattr_fiemap(inode, fieinfo);
4725         } else {
4726                 ext4_lblk_t len_blks;
4727                 __u64 last_blk;
4728
4729                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4730                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4731                 if (last_blk >= EXT_MAX_BLOCKS)
4732                         last_blk = EXT_MAX_BLOCKS-1;
4733                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4734
4735                 /*
4736                  * Walk the extent tree gathering extent information
4737                  * and pushing extents back to the user.
4738                  */
4739                 error = ext4_fill_fiemap_extents(inode, start_blk,
4740                                                  len_blks, fieinfo);
4741         }
4742
4743         return error;
4744 }