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 <asm/uaccess.h>
  41 #include <linux/fiemap.h>
  42 #include "ext4_jbd2.h"
  43 #include "ext4_extents.h"
  44 #include "xattr.h"
  45
  46 #include <trace/events/ext4.h>
  47
  48 /*
  49  * used by extent splitting.
  50  */
  51 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  52                                         due to ENOSPC */
  53 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  54 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  55
  56 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  57 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  58
  59 static __le32 ext4_extent_block_csum(struct inode *inode,
  60                                      struct ext4_extent_header *eh)
  61 {
  62         struct ext4_inode_info *ei = EXT4_I(inode);
  63         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  64         __u32 csum;
  65
  66         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  67                            EXT4_EXTENT_TAIL_OFFSET(eh));
  68         return cpu_to_le32(csum);
  69 }
  70
  71 static int ext4_extent_block_csum_verify(struct inode *inode,
  72                                          struct ext4_extent_header *eh)
  73 {
  74         struct ext4_extent_tail *et;
  75
  76         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  77                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  78                 return 1;
  79
  80         et = find_ext4_extent_tail(eh);
  81         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  82                 return 0;
  83         return 1;
  84 }
  85
  86 static void ext4_extent_block_csum_set(struct inode *inode,
  87                                        struct ext4_extent_header *eh)
  88 {
  89         struct ext4_extent_tail *et;
  90
  91         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  92                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  93                 return;
  94
  95         et = find_ext4_extent_tail(eh);
  96         et->et_checksum = ext4_extent_block_csum(inode, eh);
  97 }
  98
  99 static int ext4_split_extent(handle_t *handle,
 100                                 struct inode *inode,
 101                                 struct ext4_ext_path *path,
 102                                 struct ext4_map_blocks *map,
 103                                 int split_flag,
 104                                 int flags);
 105
 106 static int ext4_split_extent_at(handle_t *handle,
 107                              struct inode *inode,
 108                              struct ext4_ext_path *path,
 109                              ext4_lblk_t split,
 110                              int split_flag,
 111                              int flags);
 112
 113 static int ext4_find_delayed_extent(struct inode *inode,
 114                                     struct extent_status *newes);
 115
 116 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 117                                             struct inode *inode,
 118                                             int needed)
 119 {
 120         int err;
 121
 122         if (!ext4_handle_valid(handle))
 123                 return 0;
 124         if (handle->h_buffer_credits > needed)
 125                 return 0;
 126         err = ext4_journal_extend(handle, needed);
 127         if (err <= 0)
 128                 return err;
 129         err = ext4_truncate_restart_trans(handle, inode, needed);
 130         if (err == 0)
 131                 err = -EAGAIN;
 132
 133         return err;
 134 }
 135
 136 /*
 137  * could return:
 138  *  - EROFS
 139  *  - ENOMEM
 140  */
 141 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 142                                 struct ext4_ext_path *path)
 143 {
 144         if (path->p_bh) {
 145                 /* path points to block */
 146                 BUFFER_TRACE(path->p_bh, "get_write_access");
 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
 165         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 166         if (path->p_bh) {
 167                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 168                 /* path points to block */
 169                 err = __ext4_handle_dirty_metadata(where, line, handle,
 170                                                    inode, path->p_bh);
 171         } else {
 172                 /* path points to leaf/index in inode body */
 173                 err = ext4_mark_inode_dirty(handle, inode);
 174         }
 175         return err;
 176 }
 177
 178 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 179                               struct ext4_ext_path *path,
 180                               ext4_lblk_t block)
 181 {
 182         if (path) {
 183                 int depth = path->p_depth;
 184                 struct ext4_extent *ex;
 185
 186                 /*
 187                  * Try to predict block placement assuming that we are
 188                  * filling in a file which will eventually be
 189                  * non-sparse --- i.e., in the case of libbfd writing
 190                  * an ELF object sections out-of-order but in a way
 191                  * the eventually results in a contiguous object or
 192                  * executable file, or some database extending a table
 193                  * space file.  However, this is actually somewhat
 194                  * non-ideal if we are writing a sparse file such as
 195                  * qemu or KVM writing a raw image file that is going
 196                  * to stay fairly sparse, since it will end up
 197                  * fragmenting the file system's free space.  Maybe we
 198                  * should have some hueristics or some way to allow
 199                  * userspace to pass a hint to file system,
 200                  * especially if the latter case turns out to be
 201                  * common.
 202                  */
 203                 ex = path[depth].p_ext;
 204                 if (ex) {
 205                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 206                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 207
 208                         if (block > ext_block)
 209                                 return ext_pblk + (block - ext_block);
 210                         else
 211                                 return ext_pblk - (ext_block - block);
 212                 }
 213
 214                 /* it looks like index is empty;
 215                  * try to find starting block from index itself */
 216                 if (path[depth].p_bh)
 217                         return path[depth].p_bh->b_blocknr;
 218         }
 219
 220         /* OK. use inode's group */
 221         return ext4_inode_to_goal_block(inode);
 222 }
 223
 224 /*
 225  * Allocation for a meta data block
 226  */
 227 static ext4_fsblk_t
 228 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 229                         struct ext4_ext_path *path,
 230                         struct ext4_extent *ex, int *err, unsigned int flags)
 231 {
 232         ext4_fsblk_t goal, newblock;
 233
 234         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 235         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 236                                         NULL, err);
 237         return newblock;
 238 }
 239
 240 static inline int ext4_ext_space_block(struct inode *inode, int check)
 241 {
 242         int size;
 243
 244         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 245                         / sizeof(struct ext4_extent);
 246 #ifdef AGGRESSIVE_TEST
 247         if (!check && size > 6)
 248                 size = 6;
 249 #endif
 250         return size;
 251 }
 252
 253 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 254 {
 255         int size;
 256
 257         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 258                         / sizeof(struct ext4_extent_idx);
 259 #ifdef AGGRESSIVE_TEST
 260         if (!check && size > 5)
 261                 size = 5;
 262 #endif
 263         return size;
 264 }
 265
 266 static inline int ext4_ext_space_root(struct inode *inode, int check)
 267 {
 268         int size;
 269
 270         size = sizeof(EXT4_I(inode)->i_data);
 271         size -= sizeof(struct ext4_extent_header);
 272         size /= sizeof(struct ext4_extent);
 273 #ifdef AGGRESSIVE_TEST
 274         if (!check && size > 3)
 275                 size = 3;
 276 #endif
 277         return size;
 278 }
 279
 280 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 281 {
 282         int size;
 283
 284         size = sizeof(EXT4_I(inode)->i_data);
 285         size -= sizeof(struct ext4_extent_header);
 286         size /= sizeof(struct ext4_extent_idx);
 287 #ifdef AGGRESSIVE_TEST
 288         if (!check && size > 4)
 289                 size = 4;
 290 #endif
 291         return size;
 292 }
 293
 294 static inline int
 295 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 296                            struct ext4_ext_path *path, ext4_lblk_t lblk,
 297                            int nofail)
 298 {
 299         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 300
 301         return ext4_split_extent_at(handle, inode, path, lblk, unwritten ?
 302                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 303                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 304                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 305 }
 306
 307 /*
 308  * Calculate the number of metadata blocks needed
 309  * to allocate @blocks
 310  * Worse case is one block per extent
 311  */
 312 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 313 {
 314         struct ext4_inode_info *ei = EXT4_I(inode);
 315         int idxs;
 316
 317         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 318                 / sizeof(struct ext4_extent_idx));
 319
 320         /*
 321          * If the new delayed allocation block is contiguous with the
 322          * previous da block, it can share index blocks with the
 323          * previous block, so we only need to allocate a new index
 324          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 325          * an additional index block, and at ldxs**3 blocks, yet
 326          * another index blocks.
 327          */
 328         if (ei->i_da_metadata_calc_len &&
 329             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 330                 int num = 0;
 331
 332                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 333                         num++;
 334                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 335                         num++;
 336                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 337                         num++;
 338                         ei->i_da_metadata_calc_len = 0;
 339                 } else
 340                         ei->i_da_metadata_calc_len++;
 341                 ei->i_da_metadata_calc_last_lblock++;
 342                 return num;
 343         }
 344
 345         /*
 346          * In the worst case we need a new set of index blocks at
 347          * every level of the inode's extent tree.
 348          */
 349         ei->i_da_metadata_calc_len = 1;
 350         ei->i_da_metadata_calc_last_lblock = lblock;
 351         return ext_depth(inode) + 1;
 352 }
 353
 354 static int
 355 ext4_ext_max_entries(struct inode *inode, int depth)
 356 {
 357         int max;
 358
 359         if (depth == ext_depth(inode)) {
 360                 if (depth == 0)
 361                         max = ext4_ext_space_root(inode, 1);
 362                 else
 363                         max = ext4_ext_space_root_idx(inode, 1);
 364         } else {
 365                 if (depth == 0)
 366                         max = ext4_ext_space_block(inode, 1);
 367                 else
 368                         max = ext4_ext_space_block_idx(inode, 1);
 369         }
 370
 371         return max;
 372 }
 373
 374 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 375 {
 376         ext4_fsblk_t block = ext4_ext_pblock(ext);
 377         int len = ext4_ext_get_actual_len(ext);
 378         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 379         ext4_lblk_t last = lblock + len - 1;
 380
 381         if (lblock > last)
 382                 return 0;
 383         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 384 }
 385
 386 static int ext4_valid_extent_idx(struct inode *inode,
 387                                 struct ext4_extent_idx *ext_idx)
 388 {
 389         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 390
 391         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 392 }
 393
 394 static int ext4_valid_extent_entries(struct inode *inode,
 395                                 struct ext4_extent_header *eh,
 396                                 int depth)
 397 {
 398         unsigned short entries;
 399         if (eh->eh_entries == 0)
 400                 return 1;
 401
 402         entries = le16_to_cpu(eh->eh_entries);
 403
 404         if (depth == 0) {
 405                 /* leaf entries */
 406                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 407                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 408                 ext4_fsblk_t pblock = 0;
 409                 ext4_lblk_t lblock = 0;
 410                 ext4_lblk_t prev = 0;
 411                 int len = 0;
 412                 while (entries) {
 413                         if (!ext4_valid_extent(inode, ext))
 414                                 return 0;
 415
 416                         /* Check for overlapping extents */
 417                         lblock = le32_to_cpu(ext->ee_block);
 418                         len = ext4_ext_get_actual_len(ext);
 419                         if ((lblock <= prev) && prev) {
 420                                 pblock = ext4_ext_pblock(ext);
 421                                 es->s_last_error_block = cpu_to_le64(pblock);
 422                                 return 0;
 423                         }
 424                         ext++;
 425                         entries--;
 426                         prev = lblock + len - 1;
 427                 }
 428         } else {
 429                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 430                 while (entries) {
 431                         if (!ext4_valid_extent_idx(inode, ext_idx))
 432                                 return 0;
 433                         ext_idx++;
 434                         entries--;
 435                 }
 436         }
 437         return 1;
 438 }
 439
 440 static int __ext4_ext_check(const char *function, unsigned int line,
 441                             struct inode *inode, struct ext4_extent_header *eh,
 442                             int depth, ext4_fsblk_t pblk)
 443 {
 444         const char *error_msg;
 445         int max = 0;
 446
 447         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 448                 error_msg = "invalid magic";
 449                 goto corrupted;
 450         }
 451         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 452                 error_msg = "unexpected eh_depth";
 453                 goto corrupted;
 454         }
 455         if (unlikely(eh->eh_max == 0)) {
 456                 error_msg = "invalid eh_max";
 457                 goto corrupted;
 458         }
 459         max = ext4_ext_max_entries(inode, depth);
 460         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 461                 error_msg = "too large eh_max";
 462                 goto corrupted;
 463         }
 464         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 465                 error_msg = "invalid eh_entries";
 466                 goto corrupted;
 467         }
 468         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 469                 error_msg = "invalid extent entries";
 470                 goto corrupted;
 471         }
 472         /* Verify checksum on non-root extent tree nodes */
 473         if (ext_depth(inode) != depth &&
 474             !ext4_extent_block_csum_verify(inode, eh)) {
 475                 error_msg = "extent tree corrupted";
 476                 goto corrupted;
 477         }
 478         return 0;
 479
 480 corrupted:
 481         ext4_error_inode(inode, function, line, 0,
 482                          "pblk %llu bad header/extent: %s - magic %x, "
 483                          "entries %u, max %u(%u), depth %u(%u)",
 484                          (unsigned long long) pblk, error_msg,
 485                          le16_to_cpu(eh->eh_magic),
 486                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 487                          max, le16_to_cpu(eh->eh_depth), depth);
 488         return -EIO;
 489 }
 490
 491 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 492         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 493
 494 int ext4_ext_check_inode(struct inode *inode)
 495 {
 496         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 497 }
 498
 499 static struct buffer_head *
 500 __read_extent_tree_block(const char *function, unsigned int line,
 501                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 502                          int flags)
 503 {
 504         struct buffer_head              *bh;
 505         int                             err;
 506
 507         bh = sb_getblk(inode->i_sb, pblk);
 508         if (unlikely(!bh))
 509                 return ERR_PTR(-ENOMEM);
 510
 511         if (!bh_uptodate_or_lock(bh)) {
 512                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 513                 err = bh_submit_read(bh);
 514                 if (err < 0)
 515                         goto errout;
 516         }
 517         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 518                 return bh;
 519         err = __ext4_ext_check(function, line, inode,
 520                                ext_block_hdr(bh), depth, pblk);
 521         if (err)
 522                 goto errout;
 523         set_buffer_verified(bh);
 524         /*
 525          * If this is a leaf block, cache all of its entries
 526          */
 527         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 528                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 529                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 530                 ext4_lblk_t prev = 0;
 531                 int i;
 532
 533                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 534                         unsigned int status = EXTENT_STATUS_WRITTEN;
 535                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 536                         int len = ext4_ext_get_actual_len(ex);
 537
 538                         if (prev && (prev != lblk))
 539                                 ext4_es_cache_extent(inode, prev,
 540                                                      lblk - prev, ~0,
 541                                                      EXTENT_STATUS_HOLE);
 542
 543                         if (ext4_ext_is_unwritten(ex))
 544                                 status = EXTENT_STATUS_UNWRITTEN;
 545                         ext4_es_cache_extent(inode, lblk, len,
 546                                              ext4_ext_pblock(ex), status);
 547                         prev = lblk + len;
 548                 }
 549         }
 550         return bh;
 551 errout:
 552         put_bh(bh);
 553         return ERR_PTR(err);
 554
 555 }
 556
 557 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 558         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 559                                  (depth), (flags))
 560
 561 /*
 562  * This function is called to cache a file's extent information in the
 563  * extent status tree
 564  */
 565 int ext4_ext_precache(struct inode *inode)
 566 {
 567         struct ext4_inode_info *ei = EXT4_I(inode);
 568         struct ext4_ext_path *path = NULL;
 569         struct buffer_head *bh;
 570         int i = 0, depth, ret = 0;
 571
 572         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 573                 return 0;       /* not an extent-mapped inode */
 574
 575         down_read(&ei->i_data_sem);
 576         depth = ext_depth(inode);
 577
 578         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 579                        GFP_NOFS);
 580         if (path == NULL) {
 581                 up_read(&ei->i_data_sem);
 582                 return -ENOMEM;
 583         }
 584
 585         /* Don't cache anything if there are no external extent blocks */
 586         if (depth == 0)
 587                 goto out;
 588         path[0].p_hdr = ext_inode_hdr(inode);
 589         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 590         if (ret)
 591                 goto out;
 592         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 593         while (i >= 0) {
 594                 /*
 595                  * If this is a leaf block or we've reached the end of
 596                  * the index block, go up
 597                  */
 598                 if ((i == depth) ||
 599                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 600                         brelse(path[i].p_bh);
 601                         path[i].p_bh = NULL;
 602                         i--;
 603                         continue;
 604                 }
 605                 bh = read_extent_tree_block(inode,
 606                                             ext4_idx_pblock(path[i].p_idx++),
 607                                             depth - i - 1,
 608                                             EXT4_EX_FORCE_CACHE);
 609                 if (IS_ERR(bh)) {
 610                         ret = PTR_ERR(bh);
 611                         break;
 612                 }
 613                 i++;
 614                 path[i].p_bh = bh;
 615                 path[i].p_hdr = ext_block_hdr(bh);
 616                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 617         }
 618         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 619 out:
 620         up_read(&ei->i_data_sem);
 621         ext4_ext_drop_refs(path);
 622         kfree(path);
 623         return ret;
 624 }
 625
 626 #ifdef EXT_DEBUG
 627 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 628 {
 629         int k, l = path->p_depth;
 630
 631         ext_debug("path:");
 632         for (k = 0; k <= l; k++, path++) {
 633                 if (path->p_idx) {
 634                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 635                             ext4_idx_pblock(path->p_idx));
 636                 } else if (path->p_ext) {
 637                         ext_debug("  %d:[%d]%d:%llu ",
 638                                   le32_to_cpu(path->p_ext->ee_block),
 639                                   ext4_ext_is_unwritten(path->p_ext),
 640                                   ext4_ext_get_actual_len(path->p_ext),
 641                                   ext4_ext_pblock(path->p_ext));
 642                 } else
 643                         ext_debug("  []");
 644         }
 645         ext_debug("\n");
 646 }
 647
 648 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 649 {
 650         int depth = ext_depth(inode);
 651         struct ext4_extent_header *eh;
 652         struct ext4_extent *ex;
 653         int i;
 654
 655         if (!path)
 656                 return;
 657
 658         eh = path[depth].p_hdr;
 659         ex = EXT_FIRST_EXTENT(eh);
 660
 661         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 662
 663         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 664                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 665                           ext4_ext_is_unwritten(ex),
 666                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 667         }
 668         ext_debug("\n");
 669 }
 670
 671 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 672                         ext4_fsblk_t newblock, int level)
 673 {
 674         int depth = ext_depth(inode);
 675         struct ext4_extent *ex;
 676
 677         if (depth != level) {
 678                 struct ext4_extent_idx *idx;
 679                 idx = path[level].p_idx;
 680                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 681                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 682                                         le32_to_cpu(idx->ei_block),
 683                                         ext4_idx_pblock(idx),
 684                                         newblock);
 685                         idx++;
 686                 }
 687
 688                 return;
 689         }
 690
 691         ex = path[depth].p_ext;
 692         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 693                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 694                                 le32_to_cpu(ex->ee_block),
 695                                 ext4_ext_pblock(ex),
 696                                 ext4_ext_is_unwritten(ex),
 697                                 ext4_ext_get_actual_len(ex),
 698                                 newblock);
 699                 ex++;
 700         }
 701 }
 702
 703 #else
 704 #define ext4_ext_show_path(inode, path)
 705 #define ext4_ext_show_leaf(inode, path)
 706 #define ext4_ext_show_move(inode, path, newblock, level)
 707 #endif
 708
 709 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 710 {
 711         int depth = path->p_depth;
 712         int i;
 713
 714         for (i = 0; i <= depth; i++, path++)
 715                 if (path->p_bh) {
 716                         brelse(path->p_bh);
 717                         path->p_bh = NULL;
 718                 }
 719 }
 720
 721 /*
 722  * ext4_ext_binsearch_idx:
 723  * binary search for the closest index of the given block
 724  * the header must be checked before calling this
 725  */
 726 static void
 727 ext4_ext_binsearch_idx(struct inode *inode,
 728                         struct ext4_ext_path *path, ext4_lblk_t block)
 729 {
 730         struct ext4_extent_header *eh = path->p_hdr;
 731         struct ext4_extent_idx *r, *l, *m;
 732
 733
 734         ext_debug("binsearch for %u(idx):  ", block);
 735
 736         l = EXT_FIRST_INDEX(eh) + 1;
 737         r = EXT_LAST_INDEX(eh);
 738         while (l <= r) {
 739                 m = l + (r - l) / 2;
 740                 if (block < le32_to_cpu(m->ei_block))
 741                         r = m - 1;
 742                 else
 743                         l = m + 1;
 744                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 745                                 m, le32_to_cpu(m->ei_block),
 746                                 r, le32_to_cpu(r->ei_block));
 747         }
 748
 749         path->p_idx = l - 1;
 750         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 751                   ext4_idx_pblock(path->p_idx));
 752
 753 #ifdef CHECK_BINSEARCH
 754         {
 755                 struct ext4_extent_idx *chix, *ix;
 756                 int k;
 757
 758                 chix = ix = EXT_FIRST_INDEX(eh);
 759                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 760                   if (k != 0 &&
 761                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 762                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 763                                        "first=0x%p\n", k,
 764                                        ix, EXT_FIRST_INDEX(eh));
 765                                 printk(KERN_DEBUG "%u <= %u\n",
 766                                        le32_to_cpu(ix->ei_block),
 767                                        le32_to_cpu(ix[-1].ei_block));
 768                         }
 769                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 770                                            <= le32_to_cpu(ix[-1].ei_block));
 771                         if (block < le32_to_cpu(ix->ei_block))
 772                                 break;
 773                         chix = ix;
 774                 }
 775                 BUG_ON(chix != path->p_idx);
 776         }
 777 #endif
 778
 779 }
 780
 781 /*
 782  * ext4_ext_binsearch:
 783  * binary search for closest extent of the given block
 784  * the header must be checked before calling this
 785  */
 786 static void
 787 ext4_ext_binsearch(struct inode *inode,
 788                 struct ext4_ext_path *path, ext4_lblk_t block)
 789 {
 790         struct ext4_extent_header *eh = path->p_hdr;
 791         struct ext4_extent *r, *l, *m;
 792
 793         if (eh->eh_entries == 0) {
 794                 /*
 795                  * this leaf is empty:
 796                  * we get such a leaf in split/add case
 797                  */
 798                 return;
 799         }
 800
 801         ext_debug("binsearch for %u:  ", block);
 802
 803         l = EXT_FIRST_EXTENT(eh) + 1;
 804         r = EXT_LAST_EXTENT(eh);
 805
 806         while (l <= r) {
 807                 m = l + (r - l) / 2;
 808                 if (block < le32_to_cpu(m->ee_block))
 809                         r = m - 1;
 810                 else
 811                         l = m + 1;
 812                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 813                                 m, le32_to_cpu(m->ee_block),
 814                                 r, le32_to_cpu(r->ee_block));
 815         }
 816
 817         path->p_ext = l - 1;
 818         ext_debug("  -> %d:%llu:[%d]%d ",
 819                         le32_to_cpu(path->p_ext->ee_block),
 820                         ext4_ext_pblock(path->p_ext),
 821                         ext4_ext_is_unwritten(path->p_ext),
 822                         ext4_ext_get_actual_len(path->p_ext));
 823
 824 #ifdef CHECK_BINSEARCH
 825         {
 826                 struct ext4_extent *chex, *ex;
 827                 int k;
 828
 829                 chex = ex = EXT_FIRST_EXTENT(eh);
 830                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 831                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 832                                           <= le32_to_cpu(ex[-1].ee_block));
 833                         if (block < le32_to_cpu(ex->ee_block))
 834                                 break;
 835                         chex = ex;
 836                 }
 837                 BUG_ON(chex != path->p_ext);
 838         }
 839 #endif
 840
 841 }
 842
 843 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 844 {
 845         struct ext4_extent_header *eh;
 846
 847         eh = ext_inode_hdr(inode);
 848         eh->eh_depth = 0;
 849         eh->eh_entries = 0;
 850         eh->eh_magic = EXT4_EXT_MAGIC;
 851         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 852         ext4_mark_inode_dirty(handle, inode);
 853         return 0;
 854 }
 855
 856 struct ext4_ext_path *
 857 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
 858                      struct ext4_ext_path *path, int flags)
 859 {
 860         struct ext4_extent_header *eh;
 861         struct buffer_head *bh;
 862         short int depth, i, ppos = 0, alloc = 0;
 863         int ret;
 864
 865         eh = ext_inode_hdr(inode);
 866         depth = ext_depth(inode);
 867
 868         /* account possible depth increase */
 869         if (!path) {
 870                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 871                                 GFP_NOFS);
 872                 if (unlikely(!path))
 873                         return ERR_PTR(-ENOMEM);
 874                 alloc = 1;
 875         }
 876         path[0].p_hdr = eh;
 877         path[0].p_bh = NULL;
 878
 879         i = depth;
 880         /* walk through the tree */
 881         while (i) {
 882                 ext_debug("depth %d: num %d, max %d\n",
 883                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 884
 885                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 886                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 887                 path[ppos].p_depth = i;
 888                 path[ppos].p_ext = NULL;
 889
 890                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 891                                             flags);
 892                 if (unlikely(IS_ERR(bh))) {
 893                         ret = PTR_ERR(bh);
 894                         goto err;
 895                 }
 896
 897                 eh = ext_block_hdr(bh);
 898                 ppos++;
 899                 if (unlikely(ppos > depth)) {
 900                         put_bh(bh);
 901                         EXT4_ERROR_INODE(inode,
 902                                          "ppos %d > depth %d", ppos, depth);
 903                         ret = -EIO;
 904                         goto err;
 905                 }
 906                 path[ppos].p_bh = bh;
 907                 path[ppos].p_hdr = eh;
 908         }
 909
 910         path[ppos].p_depth = i;
 911         path[ppos].p_ext = NULL;
 912         path[ppos].p_idx = NULL;
 913
 914         /* find extent */
 915         ext4_ext_binsearch(inode, path + ppos, block);
 916         /* if not an empty leaf */
 917         if (path[ppos].p_ext)
 918                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 919
 920         ext4_ext_show_path(inode, path);
 921
 922         return path;
 923
 924 err:
 925         ext4_ext_drop_refs(path);
 926         if (alloc)
 927                 kfree(path);
 928         return ERR_PTR(ret);
 929 }
 930
 931 /*
 932  * ext4_ext_insert_index:
 933  * insert new index [@logical;@ptr] into the block at @curp;
 934  * check where to insert: before @curp or after @curp
 935  */
 936 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 937                                  struct ext4_ext_path *curp,
 938                                  int logical, ext4_fsblk_t ptr)
 939 {
 940         struct ext4_extent_idx *ix;
 941         int len, err;
 942
 943         err = ext4_ext_get_access(handle, inode, curp);
 944         if (err)
 945                 return err;
 946
 947         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 948                 EXT4_ERROR_INODE(inode,
 949                                  "logical %d == ei_block %d!",
 950                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 951                 return -EIO;
 952         }
 953
 954         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 955                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 956                 EXT4_ERROR_INODE(inode,
 957                                  "eh_entries %d >= eh_max %d!",
 958                                  le16_to_cpu(curp->p_hdr->eh_entries),
 959                                  le16_to_cpu(curp->p_hdr->eh_max));
 960                 return -EIO;
 961         }
 962
 963         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 964                 /* insert after */
 965                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 966                 ix = curp->p_idx + 1;
 967         } else {
 968                 /* insert before */
 969                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 970                 ix = curp->p_idx;
 971         }
 972
 973         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 974         BUG_ON(len < 0);
 975         if (len > 0) {
 976                 ext_debug("insert new index %d: "
 977                                 "move %d indices from 0x%p to 0x%p\n",
 978                                 logical, len, ix, ix + 1);
 979                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 980         }
 981
 982         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 983                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 984                 return -EIO;
 985         }
 986
 987         ix->ei_block = cpu_to_le32(logical);
 988         ext4_idx_store_pblock(ix, ptr);
 989         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 990
 991         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 992                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 993                 return -EIO;
 994         }
 995
 996         err = ext4_ext_dirty(handle, inode, curp);
 997         ext4_std_error(inode->i_sb, err);
 998
 999         return err;
1000 }
1001
1002 /*
1003  * ext4_ext_split:
1004  * inserts new subtree into the path, using free index entry
1005  * at depth @at:
1006  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1007  * - makes decision where to split
1008  * - moves remaining extents and index entries (right to the split point)
1009  *   into the newly allocated blocks
1010  * - initializes subtree
1011  */
1012 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1013                           unsigned int flags,
1014                           struct ext4_ext_path *path,
1015                           struct ext4_extent *newext, int at)
1016 {
1017         struct buffer_head *bh = NULL;
1018         int depth = ext_depth(inode);
1019         struct ext4_extent_header *neh;
1020         struct ext4_extent_idx *fidx;
1021         int i = at, k, m, a;
1022         ext4_fsblk_t newblock, oldblock;
1023         __le32 border;
1024         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1025         int err = 0;
1026
1027         /* make decision: where to split? */
1028         /* FIXME: now decision is simplest: at current extent */
1029
1030         /* if current leaf will be split, then we should use
1031          * border from split point */
1032         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1033                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1034                 return -EIO;
1035         }
1036         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1037                 border = path[depth].p_ext[1].ee_block;
1038                 ext_debug("leaf will be split."
1039                                 " next leaf starts at %d\n",
1040                                   le32_to_cpu(border));
1041         } else {
1042                 border = newext->ee_block;
1043                 ext_debug("leaf will be added."
1044                                 " next leaf starts at %d\n",
1045                                 le32_to_cpu(border));
1046         }
1047
1048         /*
1049          * If error occurs, then we break processing
1050          * and mark filesystem read-only. index won't
1051          * be inserted and tree will be in consistent
1052          * state. Next mount will repair buffers too.
1053          */
1054
1055         /*
1056          * Get array to track all allocated blocks.
1057          * We need this to handle errors and free blocks
1058          * upon them.
1059          */
1060         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1061         if (!ablocks)
1062                 return -ENOMEM;
1063
1064         /* allocate all needed blocks */
1065         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1066         for (a = 0; a < depth - at; a++) {
1067                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1068                                                    newext, &err, flags);
1069                 if (newblock == 0)
1070                         goto cleanup;
1071                 ablocks[a] = newblock;
1072         }
1073
1074         /* initialize new leaf */
1075         newblock = ablocks[--a];
1076         if (unlikely(newblock == 0)) {
1077                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1078                 err = -EIO;
1079                 goto cleanup;
1080         }
1081         bh = sb_getblk(inode->i_sb, newblock);
1082         if (unlikely(!bh)) {
1083                 err = -ENOMEM;
1084                 goto cleanup;
1085         }
1086         lock_buffer(bh);
1087
1088         err = ext4_journal_get_create_access(handle, bh);
1089         if (err)
1090                 goto cleanup;
1091
1092         neh = ext_block_hdr(bh);
1093         neh->eh_entries = 0;
1094         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1095         neh->eh_magic = EXT4_EXT_MAGIC;
1096         neh->eh_depth = 0;
1097
1098         /* move remainder of path[depth] to the new leaf */
1099         if (unlikely(path[depth].p_hdr->eh_entries !=
1100                      path[depth].p_hdr->eh_max)) {
1101                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1102                                  path[depth].p_hdr->eh_entries,
1103                                  path[depth].p_hdr->eh_max);
1104                 err = -EIO;
1105                 goto cleanup;
1106         }
1107         /* start copy from next extent */
1108         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1109         ext4_ext_show_move(inode, path, newblock, depth);
1110         if (m) {
1111                 struct ext4_extent *ex;
1112                 ex = EXT_FIRST_EXTENT(neh);
1113                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1114                 le16_add_cpu(&neh->eh_entries, m);
1115         }
1116
1117         ext4_extent_block_csum_set(inode, neh);
1118         set_buffer_uptodate(bh);
1119         unlock_buffer(bh);
1120
1121         err = ext4_handle_dirty_metadata(handle, inode, bh);
1122         if (err)
1123                 goto cleanup;
1124         brelse(bh);
1125         bh = NULL;
1126
1127         /* correct old leaf */
1128         if (m) {
1129                 err = ext4_ext_get_access(handle, inode, path + depth);
1130                 if (err)
1131                         goto cleanup;
1132                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1133                 err = ext4_ext_dirty(handle, inode, path + depth);
1134                 if (err)
1135                         goto cleanup;
1136
1137         }
1138
1139         /* create intermediate indexes */
1140         k = depth - at - 1;
1141         if (unlikely(k < 0)) {
1142                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1143                 err = -EIO;
1144                 goto cleanup;
1145         }
1146         if (k)
1147                 ext_debug("create %d intermediate indices\n", k);
1148         /* insert new index into current index block */
1149         /* current depth stored in i var */
1150         i = depth - 1;
1151         while (k--) {
1152                 oldblock = newblock;
1153                 newblock = ablocks[--a];
1154                 bh = sb_getblk(inode->i_sb, newblock);
1155                 if (unlikely(!bh)) {
1156                         err = -ENOMEM;
1157                         goto cleanup;
1158                 }
1159                 lock_buffer(bh);
1160
1161                 err = ext4_journal_get_create_access(handle, bh);
1162                 if (err)
1163                         goto cleanup;
1164
1165                 neh = ext_block_hdr(bh);
1166                 neh->eh_entries = cpu_to_le16(1);
1167                 neh->eh_magic = EXT4_EXT_MAGIC;
1168                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1169                 neh->eh_depth = cpu_to_le16(depth - i);
1170                 fidx = EXT_FIRST_INDEX(neh);
1171                 fidx->ei_block = border;
1172                 ext4_idx_store_pblock(fidx, oldblock);
1173
1174                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1175                                 i, newblock, le32_to_cpu(border), oldblock);
1176
1177                 /* move remainder of path[i] to the new index block */
1178                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1179                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1180                         EXT4_ERROR_INODE(inode,
1181                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1182                                          le32_to_cpu(path[i].p_ext->ee_block));
1183                         err = -EIO;
1184                         goto cleanup;
1185                 }
1186                 /* start copy indexes */
1187                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1188                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1189                                 EXT_MAX_INDEX(path[i].p_hdr));
1190                 ext4_ext_show_move(inode, path, newblock, i);
1191                 if (m) {
1192                         memmove(++fidx, path[i].p_idx,
1193                                 sizeof(struct ext4_extent_idx) * m);
1194                         le16_add_cpu(&neh->eh_entries, m);
1195                 }
1196                 ext4_extent_block_csum_set(inode, neh);
1197                 set_buffer_uptodate(bh);
1198                 unlock_buffer(bh);
1199
1200                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1201                 if (err)
1202                         goto cleanup;
1203                 brelse(bh);
1204                 bh = NULL;
1205
1206                 /* correct old index */
1207                 if (m) {
1208                         err = ext4_ext_get_access(handle, inode, path + i);
1209                         if (err)
1210                                 goto cleanup;
1211                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1212                         err = ext4_ext_dirty(handle, inode, path + i);
1213                         if (err)
1214                                 goto cleanup;
1215                 }
1216
1217                 i--;
1218         }
1219
1220         /* insert new index */
1221         err = ext4_ext_insert_index(handle, inode, path + at,
1222                                     le32_to_cpu(border), newblock);
1223
1224 cleanup:
1225         if (bh) {
1226                 if (buffer_locked(bh))
1227                         unlock_buffer(bh);
1228                 brelse(bh);
1229         }
1230
1231         if (err) {
1232                 /* free all allocated blocks in error case */
1233                 for (i = 0; i < depth; i++) {
1234                         if (!ablocks[i])
1235                                 continue;
1236                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1237                                          EXT4_FREE_BLOCKS_METADATA);
1238                 }
1239         }
1240         kfree(ablocks);
1241
1242         return err;
1243 }
1244
1245 /*
1246  * ext4_ext_grow_indepth:
1247  * implements tree growing procedure:
1248  * - allocates new block
1249  * - moves top-level data (index block or leaf) into the new block
1250  * - initializes new top-level, creating index that points to the
1251  *   just created block
1252  */
1253 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1254                                  unsigned int flags,
1255                                  struct ext4_extent *newext)
1256 {
1257         struct ext4_extent_header *neh;
1258         struct buffer_head *bh;
1259         ext4_fsblk_t newblock;
1260         int err = 0;
1261
1262         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1263                 newext, &err, flags);
1264         if (newblock == 0)
1265                 return err;
1266
1267         bh = sb_getblk(inode->i_sb, newblock);
1268         if (unlikely(!bh))
1269                 return -ENOMEM;
1270         lock_buffer(bh);
1271
1272         err = ext4_journal_get_create_access(handle, bh);
1273         if (err) {
1274                 unlock_buffer(bh);
1275                 goto out;
1276         }
1277
1278         /* move top-level index/leaf into new block */
1279         memmove(bh->b_data, EXT4_I(inode)->i_data,
1280                 sizeof(EXT4_I(inode)->i_data));
1281
1282         /* set size of new block */
1283         neh = ext_block_hdr(bh);
1284         /* old root could have indexes or leaves
1285          * so calculate e_max right way */
1286         if (ext_depth(inode))
1287                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1288         else
1289                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1290         neh->eh_magic = EXT4_EXT_MAGIC;
1291         ext4_extent_block_csum_set(inode, neh);
1292         set_buffer_uptodate(bh);
1293         unlock_buffer(bh);
1294
1295         err = ext4_handle_dirty_metadata(handle, inode, bh);
1296         if (err)
1297                 goto out;
1298
1299         /* Update top-level index: num,max,pointer */
1300         neh = ext_inode_hdr(inode);
1301         neh->eh_entries = cpu_to_le16(1);
1302         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1303         if (neh->eh_depth == 0) {
1304                 /* Root extent block becomes index block */
1305                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1306                 EXT_FIRST_INDEX(neh)->ei_block =
1307                         EXT_FIRST_EXTENT(neh)->ee_block;
1308         }
1309         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1310                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1311                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1312                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1313
1314         le16_add_cpu(&neh->eh_depth, 1);
1315         ext4_mark_inode_dirty(handle, inode);
1316 out:
1317         brelse(bh);
1318
1319         return err;
1320 }
1321
1322 /*
1323  * ext4_ext_create_new_leaf:
1324  * finds empty index and adds new leaf.
1325  * if no free index is found, then it requests in-depth growing.
1326  */
1327 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1328                                     unsigned int mb_flags,
1329                                     unsigned int gb_flags,
1330                                     struct ext4_ext_path *path,
1331                                     struct ext4_extent *newext)
1332 {
1333         struct ext4_ext_path *curp;
1334         int depth, i, err = 0;
1335
1336 repeat:
1337         i = depth = ext_depth(inode);
1338
1339         /* walk up to the tree and look for free index entry */
1340         curp = path + depth;
1341         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1342                 i--;
1343                 curp--;
1344         }
1345
1346         /* we use already allocated block for index block,
1347          * so subsequent data blocks should be contiguous */
1348         if (EXT_HAS_FREE_INDEX(curp)) {
1349                 /* if we found index with free entry, then use that
1350                  * entry: create all needed subtree and add new leaf */
1351                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1352                 if (err)
1353                         goto out;
1354
1355                 /* refill path */
1356                 ext4_ext_drop_refs(path);
1357                 path = ext4_ext_find_extent(inode,
1358                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1359                                     path, gb_flags);
1360                 if (IS_ERR(path))
1361                         err = PTR_ERR(path);
1362         } else {
1363                 /* tree is full, time to grow in depth */
1364                 err = ext4_ext_grow_indepth(handle, inode, mb_flags, newext);
1365                 if (err)
1366                         goto out;
1367
1368                 /* refill path */
1369                 ext4_ext_drop_refs(path);
1370                 path = ext4_ext_find_extent(inode,
1371                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1372                                     path, gb_flags);
1373                 if (IS_ERR(path)) {
1374                         err = PTR_ERR(path);
1375                         goto out;
1376                 }
1377
1378                 /*
1379                  * only first (depth 0 -> 1) produces free space;
1380                  * in all other cases we have to split the grown tree
1381                  */
1382                 depth = ext_depth(inode);
1383                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1384                         /* now we need to split */
1385                         goto repeat;
1386                 }
1387         }
1388
1389 out:
1390         return err;
1391 }
1392
1393 /*
1394  * search the closest allocated block to the left for *logical
1395  * and returns it at @logical + it's physical address at @phys
1396  * if *logical is the smallest allocated block, the function
1397  * returns 0 at @phys
1398  * return value contains 0 (success) or error code
1399  */
1400 static int ext4_ext_search_left(struct inode *inode,
1401                                 struct ext4_ext_path *path,
1402                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1403 {
1404         struct ext4_extent_idx *ix;
1405         struct ext4_extent *ex;
1406         int depth, ee_len;
1407
1408         if (unlikely(path == NULL)) {
1409                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1410                 return -EIO;
1411         }
1412         depth = path->p_depth;
1413         *phys = 0;
1414
1415         if (depth == 0 && path->p_ext == NULL)
1416                 return 0;
1417
1418         /* usually extent in the path covers blocks smaller
1419          * then *logical, but it can be that extent is the
1420          * first one in the file */
1421
1422         ex = path[depth].p_ext;
1423         ee_len = ext4_ext_get_actual_len(ex);
1424         if (*logical < le32_to_cpu(ex->ee_block)) {
1425                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1426                         EXT4_ERROR_INODE(inode,
1427                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1428                                          *logical, le32_to_cpu(ex->ee_block));
1429                         return -EIO;
1430                 }
1431                 while (--depth >= 0) {
1432                         ix = path[depth].p_idx;
1433                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1434                                 EXT4_ERROR_INODE(inode,
1435                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1436                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1437                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1438                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1439                                   depth);
1440                                 return -EIO;
1441                         }
1442                 }
1443                 return 0;
1444         }
1445
1446         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1447                 EXT4_ERROR_INODE(inode,
1448                                  "logical %d < ee_block %d + ee_len %d!",
1449                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1450                 return -EIO;
1451         }
1452
1453         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1454         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1455         return 0;
1456 }
1457
1458 /*
1459  * search the closest allocated block to the right for *logical
1460  * and returns it at @logical + it's physical address at @phys
1461  * if *logical is the largest allocated block, the function
1462  * returns 0 at @phys
1463  * return value contains 0 (success) or error code
1464  */
1465 static int ext4_ext_search_right(struct inode *inode,
1466                                  struct ext4_ext_path *path,
1467                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1468                                  struct ext4_extent **ret_ex)
1469 {
1470         struct buffer_head *bh = NULL;
1471         struct ext4_extent_header *eh;
1472         struct ext4_extent_idx *ix;
1473         struct ext4_extent *ex;
1474         ext4_fsblk_t block;
1475         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1476         int ee_len;
1477
1478         if (unlikely(path == NULL)) {
1479                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1480                 return -EIO;
1481         }
1482         depth = path->p_depth;
1483         *phys = 0;
1484
1485         if (depth == 0 && path->p_ext == NULL)
1486                 return 0;
1487
1488         /* usually extent in the path covers blocks smaller
1489          * then *logical, but it can be that extent is the
1490          * first one in the file */
1491
1492         ex = path[depth].p_ext;
1493         ee_len = ext4_ext_get_actual_len(ex);
1494         if (*logical < le32_to_cpu(ex->ee_block)) {
1495                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1496                         EXT4_ERROR_INODE(inode,
1497                                          "first_extent(path[%d].p_hdr) != ex",
1498                                          depth);
1499                         return -EIO;
1500                 }
1501                 while (--depth >= 0) {
1502                         ix = path[depth].p_idx;
1503                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1504                                 EXT4_ERROR_INODE(inode,
1505                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1506                                                  *logical);
1507                                 return -EIO;
1508                         }
1509                 }
1510                 goto found_extent;
1511         }
1512
1513         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1514                 EXT4_ERROR_INODE(inode,
1515                                  "logical %d < ee_block %d + ee_len %d!",
1516                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1517                 return -EIO;
1518         }
1519
1520         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1521                 /* next allocated block in this leaf */
1522                 ex++;
1523                 goto found_extent;
1524         }
1525
1526         /* go up and search for index to the right */
1527         while (--depth >= 0) {
1528                 ix = path[depth].p_idx;
1529                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1530                         goto got_index;
1531         }
1532
1533         /* we've gone up to the root and found no index to the right */
1534         return 0;
1535
1536 got_index:
1537         /* we've found index to the right, let's
1538          * follow it and find the closest allocated
1539          * block to the right */
1540         ix++;
1541         block = ext4_idx_pblock(ix);
1542         while (++depth < path->p_depth) {
1543                 /* subtract from p_depth to get proper eh_depth */
1544                 bh = read_extent_tree_block(inode, block,
1545                                             path->p_depth - depth, 0);
1546                 if (IS_ERR(bh))
1547                         return PTR_ERR(bh);
1548                 eh = ext_block_hdr(bh);
1549                 ix = EXT_FIRST_INDEX(eh);
1550                 block = ext4_idx_pblock(ix);
1551                 put_bh(bh);
1552         }
1553
1554         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1555         if (IS_ERR(bh))
1556                 return PTR_ERR(bh);
1557         eh = ext_block_hdr(bh);
1558         ex = EXT_FIRST_EXTENT(eh);
1559 found_extent:
1560         *logical = le32_to_cpu(ex->ee_block);
1561         *phys = ext4_ext_pblock(ex);
1562         *ret_ex = ex;
1563         if (bh)
1564                 put_bh(bh);
1565         return 0;
1566 }
1567
1568 /*
1569  * ext4_ext_next_allocated_block:
1570  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1571  * NOTE: it considers block number from index entry as
1572  * allocated block. Thus, index entries have to be consistent
1573  * with leaves.
1574  */
1575 ext4_lblk_t
1576 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1577 {
1578         int depth;
1579
1580         BUG_ON(path == NULL);
1581         depth = path->p_depth;
1582
1583         if (depth == 0 && path->p_ext == NULL)
1584                 return EXT_MAX_BLOCKS;
1585
1586         while (depth >= 0) {
1587                 if (depth == path->p_depth) {
1588                         /* leaf */
1589                         if (path[depth].p_ext &&
1590                                 path[depth].p_ext !=
1591                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1592                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1593                 } else {
1594                         /* index */
1595                         if (path[depth].p_idx !=
1596                                         EXT_LAST_INDEX(path[depth].p_hdr))
1597                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1598                 }
1599                 depth--;
1600         }
1601
1602         return EXT_MAX_BLOCKS;
1603 }
1604
1605 /*
1606  * ext4_ext_next_leaf_block:
1607  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1608  */
1609 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1610 {
1611         int depth;
1612
1613         BUG_ON(path == NULL);
1614         depth = path->p_depth;
1615
1616         /* zero-tree has no leaf blocks at all */
1617         if (depth == 0)
1618                 return EXT_MAX_BLOCKS;
1619
1620         /* go to index block */
1621         depth--;
1622
1623         while (depth >= 0) {
1624                 if (path[depth].p_idx !=
1625                                 EXT_LAST_INDEX(path[depth].p_hdr))
1626                         return (ext4_lblk_t)
1627                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1628                 depth--;
1629         }
1630
1631         return EXT_MAX_BLOCKS;
1632 }
1633
1634 /*
1635  * ext4_ext_correct_indexes:
1636  * if leaf gets modified and modified extent is first in the leaf,
1637  * then we have to correct all indexes above.
1638  * TODO: do we need to correct tree in all cases?
1639  */
1640 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1641                                 struct ext4_ext_path *path)
1642 {
1643         struct ext4_extent_header *eh;
1644         int depth = ext_depth(inode);
1645         struct ext4_extent *ex;
1646         __le32 border;
1647         int k, err = 0;
1648
1649         eh = path[depth].p_hdr;
1650         ex = path[depth].p_ext;
1651
1652         if (unlikely(ex == NULL || eh == NULL)) {
1653                 EXT4_ERROR_INODE(inode,
1654                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1655                 return -EIO;
1656         }
1657
1658         if (depth == 0) {
1659                 /* there is no tree at all */
1660                 return 0;
1661         }
1662
1663         if (ex != EXT_FIRST_EXTENT(eh)) {
1664                 /* we correct tree if first leaf got modified only */
1665                 return 0;
1666         }
1667
1668         /*
1669          * TODO: we need correction if border is smaller than current one
1670          */
1671         k = depth - 1;
1672         border = path[depth].p_ext->ee_block;
1673         err = ext4_ext_get_access(handle, inode, path + k);
1674         if (err)
1675                 return err;
1676         path[k].p_idx->ei_block = border;
1677         err = ext4_ext_dirty(handle, inode, path + k);
1678         if (err)
1679                 return err;
1680
1681         while (k--) {
1682                 /* change all left-side indexes */
1683                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1684                         break;
1685                 err = ext4_ext_get_access(handle, inode, path + k);
1686                 if (err)
1687                         break;
1688                 path[k].p_idx->ei_block = border;
1689                 err = ext4_ext_dirty(handle, inode, path + k);
1690                 if (err)
1691                         break;
1692         }
1693
1694         return err;
1695 }
1696
1697 int
1698 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1699                                 struct ext4_extent *ex2)
1700 {
1701         unsigned short ext1_ee_len, ext2_ee_len;
1702
1703         /*
1704          * Make sure that both extents are initialized. We don't merge
1705          * unwritten extents so that we can be sure that end_io code has
1706          * the extent that was written properly split out and conversion to
1707          * initialized is trivial.
1708          */
1709         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1710                 return 0;
1711
1712         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1713         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1714
1715         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1716                         le32_to_cpu(ex2->ee_block))
1717                 return 0;
1718
1719         /*
1720          * To allow future support for preallocated extents to be added
1721          * as an RO_COMPAT feature, refuse to merge to extents if
1722          * this can result in the top bit of ee_len being set.
1723          */
1724         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1725                 return 0;
1726         if (ext4_ext_is_unwritten(ex1) &&
1727             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1728              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1729              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1730                 return 0;
1731 #ifdef AGGRESSIVE_TEST
1732         if (ext1_ee_len >= 4)
1733                 return 0;
1734 #endif
1735
1736         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1737                 return 1;
1738         return 0;
1739 }
1740
1741 /*
1742  * This function tries to merge the "ex" extent to the next extent in the tree.
1743  * It always tries to merge towards right. If you want to merge towards
1744  * left, pass "ex - 1" as argument instead of "ex".
1745  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1746  * 1 if they got merged.
1747  */
1748 static int ext4_ext_try_to_merge_right(struct inode *inode,
1749                                  struct ext4_ext_path *path,
1750                                  struct ext4_extent *ex)
1751 {
1752         struct ext4_extent_header *eh;
1753         unsigned int depth, len;
1754         int merge_done = 0, unwritten;
1755
1756         depth = ext_depth(inode);
1757         BUG_ON(path[depth].p_hdr == NULL);
1758         eh = path[depth].p_hdr;
1759
1760         while (ex < EXT_LAST_EXTENT(eh)) {
1761                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1762                         break;
1763                 /* merge with next extent! */
1764                 unwritten = ext4_ext_is_unwritten(ex);
1765                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1766                                 + ext4_ext_get_actual_len(ex + 1));
1767                 if (unwritten)
1768                         ext4_ext_mark_unwritten(ex);
1769
1770                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1771                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1772                                 * sizeof(struct ext4_extent);
1773                         memmove(ex + 1, ex + 2, len);
1774                 }
1775                 le16_add_cpu(&eh->eh_entries, -1);
1776                 merge_done = 1;
1777                 WARN_ON(eh->eh_entries == 0);
1778                 if (!eh->eh_entries)
1779                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1780         }
1781
1782         return merge_done;
1783 }
1784
1785 /*
1786  * This function does a very simple check to see if we can collapse
1787  * an extent tree with a single extent tree leaf block into the inode.
1788  */
1789 static void ext4_ext_try_to_merge_up(handle_t *handle,
1790                                      struct inode *inode,
1791                                      struct ext4_ext_path *path)
1792 {
1793         size_t s;
1794         unsigned max_root = ext4_ext_space_root(inode, 0);
1795         ext4_fsblk_t blk;
1796
1797         if ((path[0].p_depth != 1) ||
1798             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1799             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1800                 return;
1801
1802         /*
1803          * We need to modify the block allocation bitmap and the block
1804          * group descriptor to release the extent tree block.  If we
1805          * can't get the journal credits, give up.
1806          */
1807         if (ext4_journal_extend(handle, 2))
1808                 return;
1809
1810         /*
1811          * Copy the extent data up to the inode
1812          */
1813         blk = ext4_idx_pblock(path[0].p_idx);
1814         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1815                 sizeof(struct ext4_extent_idx);
1816         s += sizeof(struct ext4_extent_header);
1817
1818         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1819         path[0].p_depth = 0;
1820         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1821                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1822         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1823
1824         brelse(path[1].p_bh);
1825         ext4_free_blocks(handle, inode, NULL, blk, 1,
1826                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1827 }
1828
1829 /*
1830  * This function tries to merge the @ex extent to neighbours in the tree.
1831  * return 1 if merge left else 0.
1832  */
1833 static void ext4_ext_try_to_merge(handle_t *handle,
1834                                   struct inode *inode,
1835                                   struct ext4_ext_path *path,
1836                                   struct ext4_extent *ex) {
1837         struct ext4_extent_header *eh;
1838         unsigned int depth;
1839         int merge_done = 0;
1840
1841         depth = ext_depth(inode);
1842         BUG_ON(path[depth].p_hdr == NULL);
1843         eh = path[depth].p_hdr;
1844
1845         if (ex > EXT_FIRST_EXTENT(eh))
1846                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1847
1848         if (!merge_done)
1849                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1850
1851         ext4_ext_try_to_merge_up(handle, inode, path);
1852 }
1853
1854 /*
1855  * check if a portion of the "newext" extent overlaps with an
1856  * existing extent.
1857  *
1858  * If there is an overlap discovered, it updates the length of the newext
1859  * such that there will be no overlap, and then returns 1.
1860  * If there is no overlap found, it returns 0.
1861  */
1862 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1863                                            struct inode *inode,
1864                                            struct ext4_extent *newext,
1865                                            struct ext4_ext_path *path)
1866 {
1867         ext4_lblk_t b1, b2;
1868         unsigned int depth, len1;
1869         unsigned int ret = 0;
1870
1871         b1 = le32_to_cpu(newext->ee_block);
1872         len1 = ext4_ext_get_actual_len(newext);
1873         depth = ext_depth(inode);
1874         if (!path[depth].p_ext)
1875                 goto out;
1876         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1877
1878         /*
1879          * get the next allocated block if the extent in the path
1880          * is before the requested block(s)
1881          */
1882         if (b2 < b1) {
1883                 b2 = ext4_ext_next_allocated_block(path);
1884                 if (b2 == EXT_MAX_BLOCKS)
1885                         goto out;
1886                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1887         }
1888
1889         /* check for wrap through zero on extent logical start block*/
1890         if (b1 + len1 < b1) {
1891                 len1 = EXT_MAX_BLOCKS - b1;
1892                 newext->ee_len = cpu_to_le16(len1);
1893                 ret = 1;
1894         }
1895
1896         /* check for overlap */
1897         if (b1 + len1 > b2) {
1898                 newext->ee_len = cpu_to_le16(b2 - b1);
1899                 ret = 1;
1900         }
1901 out:
1902         return ret;
1903 }
1904
1905 /*
1906  * ext4_ext_insert_extent:
1907  * tries to merge requsted extent into the existing extent or
1908  * inserts requested extent as new one into the tree,
1909  * creating new leaf in the no-space case.
1910  */
1911 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1912                                 struct ext4_ext_path *path,
1913                                 struct ext4_extent *newext, int gb_flags)
1914 {
1915         struct ext4_extent_header *eh;
1916         struct ext4_extent *ex, *fex;
1917         struct ext4_extent *nearex; /* nearest extent */
1918         struct ext4_ext_path *npath = NULL;
1919         int depth, len, err;
1920         ext4_lblk_t next;
1921         int mb_flags = 0, unwritten;
1922
1923         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1924                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1925                 return -EIO;
1926         }
1927         depth = ext_depth(inode);
1928         ex = path[depth].p_ext;
1929         eh = path[depth].p_hdr;
1930         if (unlikely(path[depth].p_hdr == NULL)) {
1931                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1932                 return -EIO;
1933         }
1934
1935         /* try to insert block into found extent and return */
1936         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1937
1938                 /*
1939                  * Try to see whether we should rather test the extent on
1940                  * right from ex, or from the left of ex. This is because
1941                  * ext4_ext_find_extent() can return either extent on the
1942                  * left, or on the right from the searched position. This
1943                  * will make merging more effective.
1944                  */
1945                 if (ex < EXT_LAST_EXTENT(eh) &&
1946                     (le32_to_cpu(ex->ee_block) +
1947                     ext4_ext_get_actual_len(ex) <
1948                     le32_to_cpu(newext->ee_block))) {
1949                         ex += 1;
1950                         goto prepend;
1951                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1952                            (le32_to_cpu(newext->ee_block) +
1953                            ext4_ext_get_actual_len(newext) <
1954                            le32_to_cpu(ex->ee_block)))
1955                         ex -= 1;
1956
1957                 /* Try to append newex to the ex */
1958                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1959                         ext_debug("append [%d]%d block to %u:[%d]%d"
1960                                   "(from %llu)\n",
1961                                   ext4_ext_is_unwritten(newext),
1962                                   ext4_ext_get_actual_len(newext),
1963                                   le32_to_cpu(ex->ee_block),
1964                                   ext4_ext_is_unwritten(ex),
1965                                   ext4_ext_get_actual_len(ex),
1966                                   ext4_ext_pblock(ex));
1967                         err = ext4_ext_get_access(handle, inode,
1968                                                   path + depth);
1969                         if (err)
1970                                 return err;
1971                         unwritten = ext4_ext_is_unwritten(ex);
1972                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1973                                         + ext4_ext_get_actual_len(newext));
1974                         if (unwritten)
1975                                 ext4_ext_mark_unwritten(ex);
1976                         eh = path[depth].p_hdr;
1977                         nearex = ex;
1978                         goto merge;
1979                 }
1980
1981 prepend:
1982                 /* Try to prepend newex to the ex */
1983                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
1984                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
1985                                   "(from %llu)\n",
1986                                   le32_to_cpu(newext->ee_block),
1987                                   ext4_ext_is_unwritten(newext),
1988                                   ext4_ext_get_actual_len(newext),
1989                                   le32_to_cpu(ex->ee_block),
1990                                   ext4_ext_is_unwritten(ex),
1991                                   ext4_ext_get_actual_len(ex),
1992                                   ext4_ext_pblock(ex));
1993                         err = ext4_ext_get_access(handle, inode,
1994                                                   path + depth);
1995                         if (err)
1996                                 return err;
1997
1998                         unwritten = ext4_ext_is_unwritten(ex);
1999                         ex->ee_block = newext->ee_block;
2000                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2001                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2002                                         + ext4_ext_get_actual_len(newext));
2003                         if (unwritten)
2004                                 ext4_ext_mark_unwritten(ex);
2005                         eh = path[depth].p_hdr;
2006                         nearex = ex;
2007                         goto merge;
2008                 }
2009         }
2010
2011         depth = ext_depth(inode);
2012         eh = path[depth].p_hdr;
2013         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2014                 goto has_space;
2015
2016         /* probably next leaf has space for us? */
2017         fex = EXT_LAST_EXTENT(eh);
2018         next = EXT_MAX_BLOCKS;
2019         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2020                 next = ext4_ext_next_leaf_block(path);
2021         if (next != EXT_MAX_BLOCKS) {
2022                 ext_debug("next leaf block - %u\n", next);
2023                 BUG_ON(npath != NULL);
2024                 npath = ext4_ext_find_extent(inode, next, NULL, 0);
2025                 if (IS_ERR(npath))
2026                         return PTR_ERR(npath);
2027                 BUG_ON(npath->p_depth != path->p_depth);
2028                 eh = npath[depth].p_hdr;
2029                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2030                         ext_debug("next leaf isn't full(%d)\n",
2031                                   le16_to_cpu(eh->eh_entries));
2032                         path = npath;
2033                         goto has_space;
2034                 }
2035                 ext_debug("next leaf has no free space(%d,%d)\n",
2036                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2037         }
2038
2039         /*
2040          * There is no free space in the found leaf.
2041          * We're gonna add a new leaf in the tree.
2042          */
2043         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2044                 mb_flags = EXT4_MB_USE_RESERVED;
2045         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2046                                        path, newext);
2047         if (err)
2048                 goto cleanup;
2049         depth = ext_depth(inode);
2050         eh = path[depth].p_hdr;
2051
2052 has_space:
2053         nearex = path[depth].p_ext;
2054
2055         err = ext4_ext_get_access(handle, inode, path + depth);
2056         if (err)
2057                 goto cleanup;
2058
2059         if (!nearex) {
2060                 /* there is no extent in this leaf, create first one */
2061                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2062                                 le32_to_cpu(newext->ee_block),
2063                                 ext4_ext_pblock(newext),
2064                                 ext4_ext_is_unwritten(newext),
2065                                 ext4_ext_get_actual_len(newext));
2066                 nearex = EXT_FIRST_EXTENT(eh);
2067         } else {
2068                 if (le32_to_cpu(newext->ee_block)
2069                            > le32_to_cpu(nearex->ee_block)) {
2070                         /* Insert after */
2071                         ext_debug("insert %u:%llu:[%d]%d before: "
2072                                         "nearest %p\n",
2073                                         le32_to_cpu(newext->ee_block),
2074                                         ext4_ext_pblock(newext),
2075                                         ext4_ext_is_unwritten(newext),
2076                                         ext4_ext_get_actual_len(newext),
2077                                         nearex);
2078                         nearex++;
2079                 } else {
2080                         /* Insert before */
2081                         BUG_ON(newext->ee_block == nearex->ee_block);
2082                         ext_debug("insert %u:%llu:[%d]%d after: "
2083                                         "nearest %p\n",
2084                                         le32_to_cpu(newext->ee_block),
2085                                         ext4_ext_pblock(newext),
2086                                         ext4_ext_is_unwritten(newext),
2087                                         ext4_ext_get_actual_len(newext),
2088                                         nearex);
2089                 }
2090                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2091                 if (len > 0) {
2092                         ext_debug("insert %u:%llu:[%d]%d: "
2093                                         "move %d extents from 0x%p to 0x%p\n",
2094                                         le32_to_cpu(newext->ee_block),
2095                                         ext4_ext_pblock(newext),
2096                                         ext4_ext_is_unwritten(newext),
2097                                         ext4_ext_get_actual_len(newext),
2098                                         len, nearex, nearex + 1);
2099                         memmove(nearex + 1, nearex,
2100                                 len * sizeof(struct ext4_extent));
2101                 }
2102         }
2103
2104         le16_add_cpu(&eh->eh_entries, 1);
2105         path[depth].p_ext = nearex;
2106         nearex->ee_block = newext->ee_block;
2107         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2108         nearex->ee_len = newext->ee_len;
2109
2110 merge:
2111         /* try to merge extents */
2112         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2113                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2114
2115
2116         /* time to correct all indexes above */
2117         err = ext4_ext_correct_indexes(handle, inode, path);
2118         if (err)
2119                 goto cleanup;
2120
2121         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2122
2123 cleanup:
2124         if (npath) {
2125                 ext4_ext_drop_refs(npath);
2126                 kfree(npath);
2127         }
2128         return err;
2129 }
2130
2131 static int ext4_fill_fiemap_extents(struct inode *inode,
2132                                     ext4_lblk_t block, ext4_lblk_t num,
2133                                     struct fiemap_extent_info *fieinfo)
2134 {
2135         struct ext4_ext_path *path = NULL;
2136         struct ext4_extent *ex;
2137         struct extent_status es;
2138         ext4_lblk_t next, next_del, start = 0, end = 0;
2139         ext4_lblk_t last = block + num;
2140         int exists, depth = 0, err = 0;
2141         unsigned int flags = 0;
2142         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2143
2144         while (block < last && block != EXT_MAX_BLOCKS) {
2145                 num = last - block;
2146                 /* find extent for this block */
2147                 down_read(&EXT4_I(inode)->i_data_sem);
2148
2149                 if (path && ext_depth(inode) != depth) {
2150                         /* depth was changed. we have to realloc path */
2151                         kfree(path);
2152                         path = NULL;
2153                 }
2154
2155                 path = ext4_ext_find_extent(inode, block, path, 0);
2156                 if (IS_ERR(path)) {
2157                         up_read(&EXT4_I(inode)->i_data_sem);
2158                         err = PTR_ERR(path);
2159                         path = NULL;
2160                         break;
2161                 }
2162
2163                 depth = ext_depth(inode);
2164                 if (unlikely(path[depth].p_hdr == NULL)) {
2165                         up_read(&EXT4_I(inode)->i_data_sem);
2166                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2167                         err = -EIO;
2168                         break;
2169                 }
2170                 ex = path[depth].p_ext;
2171                 next = ext4_ext_next_allocated_block(path);
2172                 ext4_ext_drop_refs(path);
2173
2174                 flags = 0;
2175                 exists = 0;
2176                 if (!ex) {
2177                         /* there is no extent yet, so try to allocate
2178                          * all requested space */
2179                         start = block;
2180                         end = block + num;
2181                 } else if (le32_to_cpu(ex->ee_block) > block) {
2182                         /* need to allocate space before found extent */
2183                         start = block;
2184                         end = le32_to_cpu(ex->ee_block);
2185                         if (block + num < end)
2186                                 end = block + num;
2187                 } else if (block >= le32_to_cpu(ex->ee_block)
2188                                         + ext4_ext_get_actual_len(ex)) {
2189                         /* need to allocate space after found extent */
2190                         start = block;
2191                         end = block + num;
2192                         if (end >= next)
2193                                 end = next;
2194                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2195                         /*
2196                          * some part of requested space is covered
2197                          * by found extent
2198                          */
2199                         start = block;
2200                         end = le32_to_cpu(ex->ee_block)
2201                                 + ext4_ext_get_actual_len(ex);
2202                         if (block + num < end)
2203                                 end = block + num;
2204                         exists = 1;
2205                 } else {
2206                         BUG();
2207                 }
2208                 BUG_ON(end <= start);
2209
2210                 if (!exists) {
2211                         es.es_lblk = start;
2212                         es.es_len = end - start;
2213                         es.es_pblk = 0;
2214                 } else {
2215                         es.es_lblk = le32_to_cpu(ex->ee_block);
2216                         es.es_len = ext4_ext_get_actual_len(ex);
2217                         es.es_pblk = ext4_ext_pblock(ex);
2218                         if (ext4_ext_is_unwritten(ex))
2219                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2220                 }
2221
2222                 /*
2223                  * Find delayed extent and update es accordingly. We call
2224                  * it even in !exists case to find out whether es is the
2225                  * last existing extent or not.
2226                  */
2227                 next_del = ext4_find_delayed_extent(inode, &es);
2228                 if (!exists && next_del) {
2229                         exists = 1;
2230                         flags |= (FIEMAP_EXTENT_DELALLOC |
2231                                   FIEMAP_EXTENT_UNKNOWN);
2232                 }
2233                 up_read(&EXT4_I(inode)->i_data_sem);
2234
2235                 if (unlikely(es.es_len == 0)) {
2236                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2237                         err = -EIO;
2238                         break;
2239                 }
2240
2241                 /*
2242                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2243                  * we need to check next == EXT_MAX_BLOCKS because it is
2244                  * possible that an extent is with unwritten and delayed
2245                  * status due to when an extent is delayed allocated and
2246                  * is allocated by fallocate status tree will track both of
2247                  * them in a extent.
2248                  *
2249                  * So we could return a unwritten and delayed extent, and
2250                  * its block is equal to 'next'.
2251                  */
2252                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2253                         flags |= FIEMAP_EXTENT_LAST;
2254                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2255                                      next != EXT_MAX_BLOCKS)) {
2256                                 EXT4_ERROR_INODE(inode,
2257                                                  "next extent == %u, next "
2258                                                  "delalloc extent = %u",
2259                                                  next, next_del);
2260                                 err = -EIO;
2261                                 break;
2262                         }
2263                 }
2264
2265                 if (exists) {
2266                         err = fiemap_fill_next_extent(fieinfo,
2267                                 (__u64)es.es_lblk << blksize_bits,
2268                                 (__u64)es.es_pblk << blksize_bits,
2269                                 (__u64)es.es_len << blksize_bits,
2270                                 flags);
2271                         if (err < 0)
2272                                 break;
2273                         if (err == 1) {
2274                                 err = 0;
2275                                 break;
2276                         }
2277                 }
2278
2279                 block = es.es_lblk + es.es_len;
2280         }
2281
2282         if (path) {
2283                 ext4_ext_drop_refs(path);
2284                 kfree(path);
2285         }
2286
2287         return err;
2288 }
2289
2290 /*
2291  * ext4_ext_put_gap_in_cache:
2292  * calculate boundaries of the gap that the requested block fits into
2293  * and cache this gap
2294  */
2295 static void
2296 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2297                                 ext4_lblk_t block)
2298 {
2299         int depth = ext_depth(inode);
2300         unsigned long len = 0;
2301         ext4_lblk_t lblock = 0;
2302         struct ext4_extent *ex;
2303
2304         ex = path[depth].p_ext;
2305         if (ex == NULL) {
2306                 /*
2307                  * there is no extent yet, so gap is [0;-] and we
2308                  * don't cache it
2309                  */
2310                 ext_debug("cache gap(whole file):");
2311         } else if (block < le32_to_cpu(ex->ee_block)) {
2312                 lblock = block;
2313                 len = le32_to_cpu(ex->ee_block) - block;
2314                 ext_debug("cache gap(before): %u [%u:%u]",
2315                                 block,
2316                                 le32_to_cpu(ex->ee_block),
2317                                  ext4_ext_get_actual_len(ex));
2318                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2319                         ext4_es_insert_extent(inode, lblock, len, ~0,
2320                                               EXTENT_STATUS_HOLE);
2321         } else if (block >= le32_to_cpu(ex->ee_block)
2322                         + ext4_ext_get_actual_len(ex)) {
2323                 ext4_lblk_t next;
2324                 lblock = le32_to_cpu(ex->ee_block)
2325                         + ext4_ext_get_actual_len(ex);
2326
2327                 next = ext4_ext_next_allocated_block(path);
2328                 ext_debug("cache gap(after): [%u:%u] %u",
2329                                 le32_to_cpu(ex->ee_block),
2330                                 ext4_ext_get_actual_len(ex),
2331                                 block);
2332                 BUG_ON(next == lblock);
2333                 len = next - lblock;
2334                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2335                         ext4_es_insert_extent(inode, lblock, len, ~0,
2336                                               EXTENT_STATUS_HOLE);
2337         } else {
2338                 BUG();
2339         }
2340
2341         ext_debug(" -> %u:%lu\n", lblock, len);
2342 }
2343
2344 /*
2345  * ext4_ext_rm_idx:
2346  * removes index from the index block.
2347  */
2348 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2349                         struct ext4_ext_path *path, int depth)
2350 {
2351         int err;
2352         ext4_fsblk_t leaf;
2353
2354         /* free index block */
2355         depth--;
2356         path = path + depth;
2357         leaf = ext4_idx_pblock(path->p_idx);
2358         if (unlikely(path->p_hdr->eh_entries == 0)) {
2359                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2360                 return -EIO;
2361         }
2362         err = ext4_ext_get_access(handle, inode, path);
2363         if (err)
2364                 return err;
2365
2366         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2367                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2368                 len *= sizeof(struct ext4_extent_idx);
2369                 memmove(path->p_idx, path->p_idx + 1, len);
2370         }
2371
2372         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2373         err = ext4_ext_dirty(handle, inode, path);
2374         if (err)
2375                 return err;
2376         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2377         trace_ext4_ext_rm_idx(inode, leaf);
2378
2379         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2380                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2381
2382         while (--depth >= 0) {
2383                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2384                         break;
2385                 path--;
2386                 err = ext4_ext_get_access(handle, inode, path);
2387                 if (err)
2388                         break;
2389                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2390                 err = ext4_ext_dirty(handle, inode, path);
2391                 if (err)
2392                         break;
2393         }
2394         return err;
2395 }
2396
2397 /*
2398  * ext4_ext_calc_credits_for_single_extent:
2399  * This routine returns max. credits that needed to insert an extent
2400  * to the extent tree.
2401  * When pass the actual path, the caller should calculate credits
2402  * under i_data_sem.
2403  */
2404 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2405                                                 struct ext4_ext_path *path)
2406 {
2407         if (path) {
2408                 int depth = ext_depth(inode);
2409                 int ret = 0;
2410
2411                 /* probably there is space in leaf? */
2412                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2413                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2414
2415                         /*
2416                          *  There are some space in the leaf tree, no
2417                          *  need to account for leaf block credit
2418                          *
2419                          *  bitmaps and block group descriptor blocks
2420                          *  and other metadata blocks still need to be
2421                          *  accounted.
2422                          */
2423                         /* 1 bitmap, 1 block group descriptor */
2424                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2425                         return ret;
2426                 }
2427         }
2428
2429         return ext4_chunk_trans_blocks(inode, nrblocks);
2430 }
2431
2432 /*
2433  * How many index/leaf blocks need to change/allocate to add @extents extents?
2434  *
2435  * If we add a single extent, then in the worse case, each tree level
2436  * index/leaf need to be changed in case of the tree split.
2437  *
2438  * If more extents are inserted, they could cause the whole tree split more
2439  * than once, but this is really rare.
2440  */
2441 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2442 {
2443         int index;
2444         int depth;
2445
2446         /* If we are converting the inline data, only one is needed here. */
2447         if (ext4_has_inline_data(inode))
2448                 return 1;
2449
2450         depth = ext_depth(inode);
2451
2452         if (extents <= 1)
2453                 index = depth * 2;
2454         else
2455                 index = depth * 3;
2456
2457         return index;
2458 }
2459
2460 static inline int get_default_free_blocks_flags(struct inode *inode)
2461 {
2462         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2463                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2464         else if (ext4_should_journal_data(inode))
2465                 return EXT4_FREE_BLOCKS_FORGET;
2466         return 0;
2467 }
2468
2469 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2470                               struct ext4_extent *ex,
2471                               long long *partial_cluster,
2472                               ext4_lblk_t from, ext4_lblk_t to)
2473 {
2474         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2475         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2476         ext4_fsblk_t pblk;
2477         int flags = get_default_free_blocks_flags(inode);
2478
2479         /*
2480          * For bigalloc file systems, we never free a partial cluster
2481          * at the beginning of the extent.  Instead, we make a note
2482          * that we tried freeing the cluster, and check to see if we
2483          * need to free it on a subsequent call to ext4_remove_blocks,
2484          * or at the end of the ext4_truncate() operation.
2485          */
2486         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2487
2488         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2489         /*
2490          * If we have a partial cluster, and it's different from the
2491          * cluster of the last block, we need to explicitly free the
2492          * partial cluster here.
2493          */
2494         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2495         if ((*partial_cluster > 0) &&
2496             (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2497                 ext4_free_blocks(handle, inode, NULL,
2498                                  EXT4_C2B(sbi, *partial_cluster),
2499                                  sbi->s_cluster_ratio, flags);
2500                 *partial_cluster = 0;
2501         }
2502
2503 #ifdef EXTENTS_STATS
2504         {
2505                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2506                 spin_lock(&sbi->s_ext_stats_lock);
2507                 sbi->s_ext_blocks += ee_len;
2508                 sbi->s_ext_extents++;
2509                 if (ee_len < sbi->s_ext_min)
2510                         sbi->s_ext_min = ee_len;
2511                 if (ee_len > sbi->s_ext_max)
2512                         sbi->s_ext_max = ee_len;
2513                 if (ext_depth(inode) > sbi->s_depth_max)
2514                         sbi->s_depth_max = ext_depth(inode);
2515                 spin_unlock(&sbi->s_ext_stats_lock);
2516         }
2517 #endif
2518         if (from >= le32_to_cpu(ex->ee_block)
2519             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2520                 /* tail removal */
2521                 ext4_lblk_t num;
2522                 unsigned int unaligned;
2523
2524                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2525                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2526                 /*
2527                  * Usually we want to free partial cluster at the end of the
2528                  * extent, except for the situation when the cluster is still
2529                  * used by any other extent (partial_cluster is negative).
2530                  */
2531                 if (*partial_cluster < 0 &&
2532                     -(*partial_cluster) == EXT4_B2C(sbi, pblk + num - 1))
2533                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2534
2535                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2536                           num, pblk, *partial_cluster);
2537                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2538                 /*
2539                  * If the block range to be freed didn't start at the
2540                  * beginning of a cluster, and we removed the entire
2541                  * extent and the cluster is not used by any other extent,
2542                  * save the partial cluster here, since we might need to
2543                  * delete if we determine that the truncate operation has
2544                  * removed all of the blocks in the cluster.
2545                  *
2546                  * On the other hand, if we did not manage to free the whole
2547                  * extent, we have to mark the cluster as used (store negative
2548                  * cluster number in partial_cluster).
2549                  */
2550                 unaligned = EXT4_PBLK_COFF(sbi, pblk);
2551                 if (unaligned && (ee_len == num) &&
2552                     (*partial_cluster != -((long long)EXT4_B2C(sbi, pblk))))
2553                         *partial_cluster = EXT4_B2C(sbi, pblk);
2554                 else if (unaligned)
2555                         *partial_cluster = -((long long)EXT4_B2C(sbi, pblk));
2556                 else if (*partial_cluster > 0)
2557                         *partial_cluster = 0;
2558         } else
2559                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2560                            "%u-%u from %u:%u\n",
2561                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2562         return 0;
2563 }
2564
2565
2566 /*
2567  * ext4_ext_rm_leaf() Removes the extents associated with the
2568  * blocks appearing between "start" and "end", and splits the extents
2569  * if "start" and "end" appear in the same extent
2570  *
2571  * @handle: The journal handle
2572  * @inode:  The files inode
2573  * @path:   The path to the leaf
2574  * @partial_cluster: The cluster which we'll have to free if all extents
2575  *                   has been released from it. It gets negative in case
2576  *                   that the cluster is still used.
2577  * @start:  The first block to remove
2578  * @end:   The last block to remove
2579  */
2580 static int
2581 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2582                  struct ext4_ext_path *path,
2583                  long long *partial_cluster,
2584                  ext4_lblk_t start, ext4_lblk_t end)
2585 {
2586         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2587         int err = 0, correct_index = 0;
2588         int depth = ext_depth(inode), credits;
2589         struct ext4_extent_header *eh;
2590         ext4_lblk_t a, b;
2591         unsigned num;
2592         ext4_lblk_t ex_ee_block;
2593         unsigned short ex_ee_len;
2594         unsigned unwritten = 0;
2595         struct ext4_extent *ex;
2596         ext4_fsblk_t pblk;
2597
2598         /* the header must be checked already in ext4_ext_remove_space() */
2599         ext_debug("truncate since %u in leaf to %u\n", start, end);
2600         if (!path[depth].p_hdr)
2601                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2602         eh = path[depth].p_hdr;
2603         if (unlikely(path[depth].p_hdr == NULL)) {
2604                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2605                 return -EIO;
2606         }
2607         /* find where to start removing */
2608         ex = path[depth].p_ext;
2609         if (!ex)
2610                 ex = EXT_LAST_EXTENT(eh);
2611
2612         ex_ee_block = le32_to_cpu(ex->ee_block);
2613         ex_ee_len = ext4_ext_get_actual_len(ex);
2614
2615         /*
2616          * If we're starting with an extent other than the last one in the
2617          * node, we need to see if it shares a cluster with the extent to
2618          * the right (towards the end of the file). If its leftmost cluster
2619          * is this extent's rightmost cluster and it is not cluster aligned,
2620          * we'll mark it as a partial that is not to be deallocated.
2621          */
2622
2623         if (ex != EXT_LAST_EXTENT(eh)) {
2624                 ext4_fsblk_t current_pblk, right_pblk;
2625                 long long current_cluster, right_cluster;
2626
2627                 current_pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2628                 current_cluster = (long long)EXT4_B2C(sbi, current_pblk);
2629                 right_pblk = ext4_ext_pblock(ex + 1);
2630                 right_cluster = (long long)EXT4_B2C(sbi, right_pblk);
2631                 if (current_cluster == right_cluster &&
2632                         EXT4_PBLK_COFF(sbi, right_pblk))
2633                         *partial_cluster = -right_cluster;
2634         }
2635
2636         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2637
2638         while (ex >= EXT_FIRST_EXTENT(eh) &&
2639                         ex_ee_block + ex_ee_len > start) {
2640
2641                 if (ext4_ext_is_unwritten(ex))
2642                         unwritten = 1;
2643                 else
2644                         unwritten = 0;
2645
2646                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2647                           unwritten, ex_ee_len);
2648                 path[depth].p_ext = ex;
2649
2650                 a = ex_ee_block > start ? ex_ee_block : start;
2651                 b = ex_ee_block+ex_ee_len - 1 < end ?
2652                         ex_ee_block+ex_ee_len - 1 : end;
2653
2654                 ext_debug("  border %u:%u\n", a, b);
2655
2656                 /* If this extent is beyond the end of the hole, skip it */
2657                 if (end < ex_ee_block) {
2658                         /*
2659                          * We're going to skip this extent and move to another,
2660                          * so if this extent is not cluster aligned we have
2661                          * to mark the current cluster as used to avoid
2662                          * accidentally freeing it later on
2663                          */
2664                         pblk = ext4_ext_pblock(ex);
2665                         if (EXT4_PBLK_COFF(sbi, pblk))
2666                                 *partial_cluster =
2667                                         -((long long)EXT4_B2C(sbi, pblk));
2668                         ex--;
2669                         ex_ee_block = le32_to_cpu(ex->ee_block);
2670                         ex_ee_len = ext4_ext_get_actual_len(ex);
2671                         continue;
2672                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2673                         EXT4_ERROR_INODE(inode,
2674                                          "can not handle truncate %u:%u "
2675                                          "on extent %u:%u",
2676                                          start, end, ex_ee_block,
2677                                          ex_ee_block + ex_ee_len - 1);
2678                         err = -EIO;
2679                         goto out;
2680                 } else if (a != ex_ee_block) {
2681                         /* remove tail of the extent */
2682                         num = a - ex_ee_block;
2683                 } else {
2684                         /* remove whole extent: excellent! */
2685                         num = 0;
2686                 }
2687                 /*
2688                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2689                  * descriptor) for each block group; assume two block
2690                  * groups plus ex_ee_len/blocks_per_block_group for
2691                  * the worst case
2692                  */
2693                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2694                 if (ex == EXT_FIRST_EXTENT(eh)) {
2695                         correct_index = 1;
2696                         credits += (ext_depth(inode)) + 1;
2697                 }
2698                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2699
2700                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2701                 if (err)
2702                         goto out;
2703
2704                 err = ext4_ext_get_access(handle, inode, path + depth);
2705                 if (err)
2706                         goto out;
2707
2708                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2709                                          a, b);
2710                 if (err)
2711                         goto out;
2712
2713                 if (num == 0)
2714                         /* this extent is removed; mark slot entirely unused */
2715                         ext4_ext_store_pblock(ex, 0);
2716
2717                 ex->ee_len = cpu_to_le16(num);
2718                 /*
2719                  * Do not mark unwritten if all the blocks in the
2720                  * extent have been removed.
2721                  */
2722                 if (unwritten && num)
2723                         ext4_ext_mark_unwritten(ex);
2724                 /*
2725                  * If the extent was completely released,
2726                  * we need to remove it from the leaf
2727                  */
2728                 if (num == 0) {
2729                         if (end != EXT_MAX_BLOCKS - 1) {
2730                                 /*
2731                                  * For hole punching, we need to scoot all the
2732                                  * extents up when an extent is removed so that
2733                                  * we dont have blank extents in the middle
2734                                  */
2735                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2736                                         sizeof(struct ext4_extent));
2737
2738                                 /* Now get rid of the one at the end */
2739                                 memset(EXT_LAST_EXTENT(eh), 0,
2740                                         sizeof(struct ext4_extent));
2741                         }
2742                         le16_add_cpu(&eh->eh_entries, -1);
2743                 } else if (*partial_cluster > 0)
2744                         *partial_cluster = 0;
2745
2746                 err = ext4_ext_dirty(handle, inode, path + depth);
2747                 if (err)
2748                         goto out;
2749
2750                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2751                                 ext4_ext_pblock(ex));
2752                 ex--;
2753                 ex_ee_block = le32_to_cpu(ex->ee_block);
2754                 ex_ee_len = ext4_ext_get_actual_len(ex);
2755         }
2756
2757         if (correct_index && eh->eh_entries)
2758                 err = ext4_ext_correct_indexes(handle, inode, path);
2759
2760         /*
2761          * If there's a partial cluster and at least one extent remains in
2762          * the leaf, free the partial cluster if it isn't shared with the
2763          * current extent.  If there's a partial cluster and no extents
2764          * remain in the leaf, it can't be freed here.  It can only be
2765          * freed when it's possible to determine if it's not shared with
2766          * any other extent - when the next leaf is processed or when space
2767          * removal is complete.
2768          */
2769         if (*partial_cluster > 0 && eh->eh_entries &&
2770             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2771              *partial_cluster)) {
2772                 int flags = get_default_free_blocks_flags(inode);
2773
2774                 ext4_free_blocks(handle, inode, NULL,
2775                                  EXT4_C2B(sbi, *partial_cluster),
2776                                  sbi->s_cluster_ratio, flags);
2777                 *partial_cluster = 0;
2778         }
2779
2780         /* if this leaf is free, then we should
2781          * remove it from index block above */
2782         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2783                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2784
2785 out:
2786         return err;
2787 }
2788
2789 /*
2790  * ext4_ext_more_to_rm:
2791  * returns 1 if current index has to be freed (even partial)
2792  */
2793 static int
2794 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2795 {
2796         BUG_ON(path->p_idx == NULL);
2797
2798         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2799                 return 0;
2800
2801         /*
2802          * if truncate on deeper level happened, it wasn't partial,
2803          * so we have to consider current index for truncation
2804          */
2805         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2806                 return 0;
2807         return 1;
2808 }
2809
2810 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2811                           ext4_lblk_t end)
2812 {
2813         struct super_block *sb = inode->i_sb;
2814         int depth = ext_depth(inode);
2815         struct ext4_ext_path *path = NULL;
2816         long long partial_cluster = 0;
2817         handle_t *handle;
2818         int i = 0, err = 0;
2819
2820         ext_debug("truncate since %u to %u\n", start, end);
2821
2822         /* probably first extent we're gonna free will be last in block */
2823         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2824         if (IS_ERR(handle))
2825                 return PTR_ERR(handle);
2826
2827 again:
2828         trace_ext4_ext_remove_space(inode, start, end, depth);
2829
2830         /*
2831          * Check if we are removing extents inside the extent tree. If that
2832          * is the case, we are going to punch a hole inside the extent tree
2833          * so we have to check whether we need to split the extent covering
2834          * the last block to remove so we can easily remove the part of it
2835          * in ext4_ext_rm_leaf().
2836          */
2837         if (end < EXT_MAX_BLOCKS - 1) {
2838                 struct ext4_extent *ex;
2839                 ext4_lblk_t ee_block;
2840
2841                 /* find extent for this block */
2842                 path = ext4_ext_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2843                 if (IS_ERR(path)) {
2844                         ext4_journal_stop(handle);
2845                         return PTR_ERR(path);
2846                 }
2847                 depth = ext_depth(inode);
2848                 /* Leaf not may not exist only if inode has no blocks at all */
2849                 ex = path[depth].p_ext;
2850                 if (!ex) {
2851                         if (depth) {
2852                                 EXT4_ERROR_INODE(inode,
2853                                                  "path[%d].p_hdr == NULL",
2854                                                  depth);
2855                                 err = -EIO;
2856                         }
2857                         goto out;
2858                 }
2859
2860                 ee_block = le32_to_cpu(ex->ee_block);
2861
2862                 /*
2863                  * See if the last block is inside the extent, if so split
2864                  * the extent at 'end' block so we can easily remove the
2865                  * tail of the first part of the split extent in
2866                  * ext4_ext_rm_leaf().
2867                  */
2868                 if (end >= ee_block &&
2869                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2870                         /*
2871                          * Split the extent in two so that 'end' is the last
2872                          * block in the first new extent. Also we should not
2873                          * fail removing space due to ENOSPC so try to use
2874                          * reserved block if that happens.
2875                          */
2876                         err = ext4_force_split_extent_at(handle, inode, path,
2877                                                          end + 1, 1);
2878                         if (err < 0)
2879                                 goto out;
2880                 }
2881         }
2882         /*
2883          * We start scanning from right side, freeing all the blocks
2884          * after i_size and walking into the tree depth-wise.
2885          */
2886         depth = ext_depth(inode);
2887         if (path) {
2888                 int k = i = depth;
2889                 while (--k > 0)
2890                         path[k].p_block =
2891                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2892         } else {
2893                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2894                                GFP_NOFS);
2895                 if (path == NULL) {
2896                         ext4_journal_stop(handle);
2897                         return -ENOMEM;
2898                 }
2899                 path[0].p_depth = depth;
2900                 path[0].p_hdr = ext_inode_hdr(inode);
2901                 i = 0;
2902
2903                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2904                         err = -EIO;
2905                         goto out;
2906                 }
2907         }
2908         err = 0;
2909
2910         while (i >= 0 && err == 0) {
2911                 if (i == depth) {
2912                         /* this is leaf block */
2913                         err = ext4_ext_rm_leaf(handle, inode, path,
2914                                                &partial_cluster, start,
2915                                                end);
2916                         /* root level has p_bh == NULL, brelse() eats this */
2917                         brelse(path[i].p_bh);
2918                         path[i].p_bh = NULL;
2919                         i--;
2920                         continue;
2921                 }
2922
2923                 /* this is index block */
2924                 if (!path[i].p_hdr) {
2925                         ext_debug("initialize header\n");
2926                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2927                 }
2928
2929                 if (!path[i].p_idx) {
2930                         /* this level hasn't been touched yet */
2931                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2932                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2933                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2934                                   path[i].p_hdr,
2935                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2936                 } else {
2937                         /* we were already here, see at next index */
2938                         path[i].p_idx--;
2939                 }
2940
2941                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2942                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2943                                 path[i].p_idx);
2944                 if (ext4_ext_more_to_rm(path + i)) {
2945                         struct buffer_head *bh;
2946                         /* go to the next level */
2947                         ext_debug("move to level %d (block %llu)\n",
2948                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2949                         memset(path + i + 1, 0, sizeof(*path));
2950                         bh = read_extent_tree_block(inode,
2951                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2952                                 EXT4_EX_NOCACHE);
2953                         if (IS_ERR(bh)) {
2954                                 /* should we reset i_size? */
2955                                 err = PTR_ERR(bh);
2956                                 break;
2957                         }
2958                         /* Yield here to deal with large extent trees.
2959                          * Should be a no-op if we did IO above. */
2960                         cond_resched();
2961                         if (WARN_ON(i + 1 > depth)) {
2962                                 err = -EIO;
2963                                 break;
2964                         }
2965                         path[i + 1].p_bh = bh;
2966
2967                         /* save actual number of indexes since this
2968                          * number is changed at the next iteration */
2969                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2970                         i++;
2971                 } else {
2972                         /* we finished processing this index, go up */
2973                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2974                                 /* index is empty, remove it;
2975                                  * handle must be already prepared by the
2976                                  * truncatei_leaf() */
2977                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2978                         }
2979                         /* root level has p_bh == NULL, brelse() eats this */
2980                         brelse(path[i].p_bh);
2981                         path[i].p_bh = NULL;
2982                         i--;
2983                         ext_debug("return to level %d\n", i);
2984                 }
2985         }
2986
2987         trace_ext4_ext_remove_space_done(inode, start, end, depth,
2988                         partial_cluster, path->p_hdr->eh_entries);
2989
2990         /* If we still have something in the partial cluster and we have removed
2991          * even the first extent, then we should free the blocks in the partial
2992          * cluster as well. */
2993         if (partial_cluster > 0 && path->p_hdr->eh_entries == 0) {
2994                 int flags = get_default_free_blocks_flags(inode);
2995
2996                 ext4_free_blocks(handle, inode, NULL,
2997                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2998                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2999                 partial_cluster = 0;
3000         }
3001
3002         /* TODO: flexible tree reduction should be here */
3003         if (path->p_hdr->eh_entries == 0) {
3004                 /*
3005                  * truncate to zero freed all the tree,
3006                  * so we need to correct eh_depth
3007                  */
3008                 err = ext4_ext_get_access(handle, inode, path);
3009                 if (err == 0) {
3010                         ext_inode_hdr(inode)->eh_depth = 0;
3011                         ext_inode_hdr(inode)->eh_max =
3012                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3013                         err = ext4_ext_dirty(handle, inode, path);
3014                 }
3015         }
3016 out:
3017         ext4_ext_drop_refs(path);
3018         kfree(path);
3019         if (err == -EAGAIN) {
3020                 path = NULL;
3021                 goto again;
3022         }
3023         ext4_journal_stop(handle);
3024
3025         return err;
3026 }
3027
3028 /*
3029  * called at mount time
3030  */
3031 void ext4_ext_init(struct super_block *sb)
3032 {
3033         /*
3034          * possible initialization would be here
3035          */
3036
3037         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
3038 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3039                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3040 #ifdef AGGRESSIVE_TEST
3041                        ", aggressive tests"
3042 #endif
3043 #ifdef CHECK_BINSEARCH
3044                        ", check binsearch"
3045 #endif
3046 #ifdef EXTENTS_STATS
3047                        ", stats"
3048 #endif
3049                        "\n");
3050 #endif
3051 #ifdef EXTENTS_STATS
3052                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3053                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3054                 EXT4_SB(sb)->s_ext_max = 0;
3055 #endif
3056         }
3057 }
3058
3059 /*
3060  * called at umount time
3061  */
3062 void ext4_ext_release(struct super_block *sb)
3063 {
3064         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
3065                 return;
3066
3067 #ifdef EXTENTS_STATS
3068         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3069                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3070                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3071                         sbi->s_ext_blocks, sbi->s_ext_extents,
3072                         sbi->s_ext_blocks / sbi->s_ext_extents);
3073                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3074                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3075         }
3076 #endif
3077 }
3078
3079 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3080 {
3081         ext4_lblk_t  ee_block;
3082         ext4_fsblk_t ee_pblock;
3083         unsigned int ee_len;
3084
3085         ee_block  = le32_to_cpu(ex->ee_block);
3086         ee_len    = ext4_ext_get_actual_len(ex);
3087         ee_pblock = ext4_ext_pblock(ex);
3088
3089         if (ee_len == 0)
3090                 return 0;
3091
3092         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3093                                      EXTENT_STATUS_WRITTEN);
3094 }
3095
3096 /* FIXME!! we need to try to merge to left or right after zero-out  */
3097 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3098 {
3099         ext4_fsblk_t ee_pblock;
3100         unsigned int ee_len;
3101         int ret;
3102
3103         ee_len    = ext4_ext_get_actual_len(ex);
3104         ee_pblock = ext4_ext_pblock(ex);
3105
3106         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
3107         if (ret > 0)
3108                 ret = 0;
3109
3110         return ret;
3111 }
3112
3113 /*
3114  * ext4_split_extent_at() splits an extent at given block.
3115  *
3116  * @handle: the journal handle
3117  * @inode: the file inode
3118  * @path: the path to the extent
3119  * @split: the logical block where the extent is splitted.
3120  * @split_flags: indicates if the extent could be zeroout if split fails, and
3121  *               the states(init or unwritten) of new extents.
3122  * @flags: flags used to insert new extent to extent tree.
3123  *
3124  *
3125  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3126  * of which are deterimined by split_flag.
3127  *
3128  * There are two cases:
3129  *  a> the extent are splitted into two extent.
3130  *  b> split is not needed, and just mark the extent.
3131  *
3132  * return 0 on success.
3133  */
3134 static int ext4_split_extent_at(handle_t *handle,
3135                              struct inode *inode,
3136                              struct ext4_ext_path *path,
3137                              ext4_lblk_t split,
3138                              int split_flag,
3139                              int flags)
3140 {
3141         ext4_fsblk_t newblock;
3142         ext4_lblk_t ee_block;
3143         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3144         struct ext4_extent *ex2 = NULL;
3145         unsigned int ee_len, depth;
3146         int err = 0;
3147
3148         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3149                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3150
3151         ext_debug("ext4_split_extents_at: inode %lu, logical"
3152                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3153
3154         ext4_ext_show_leaf(inode, path);
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         newblock = split - ee_block + ext4_ext_pblock(ex);
3161
3162         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3163         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3164                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3165                              EXT4_EXT_MARK_UNWRIT1 |
3166                              EXT4_EXT_MARK_UNWRIT2));
3167
3168         err = ext4_ext_get_access(handle, inode, path + depth);
3169         if (err)
3170                 goto out;
3171
3172         if (split == ee_block) {
3173                 /*
3174                  * case b: block @split is the block that the extent begins with
3175                  * then we just change the state of the extent, and splitting
3176                  * is not needed.
3177                  */
3178                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3179                         ext4_ext_mark_unwritten(ex);
3180                 else
3181                         ext4_ext_mark_initialized(ex);
3182
3183                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3184                         ext4_ext_try_to_merge(handle, inode, path, ex);
3185
3186                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3187                 goto out;
3188         }
3189
3190         /* case a */
3191         memcpy(&orig_ex, ex, sizeof(orig_ex));
3192         ex->ee_len = cpu_to_le16(split - ee_block);
3193         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3194                 ext4_ext_mark_unwritten(ex);
3195
3196         /*
3197          * path may lead to new leaf, not to original leaf any more
3198          * after ext4_ext_insert_extent() returns,
3199          */
3200         err = ext4_ext_dirty(handle, inode, path + depth);
3201         if (err)
3202                 goto fix_extent_len;
3203
3204         ex2 = &newex;
3205         ex2->ee_block = cpu_to_le32(split);
3206         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3207         ext4_ext_store_pblock(ex2, newblock);
3208         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3209                 ext4_ext_mark_unwritten(ex2);
3210
3211         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
3212         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3213                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3214                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3215                                 err = ext4_ext_zeroout(inode, ex2);
3216                                 zero_ex.ee_block = ex2->ee_block;
3217                                 zero_ex.ee_len = cpu_to_le16(
3218                                                 ext4_ext_get_actual_len(ex2));
3219                                 ext4_ext_store_pblock(&zero_ex,
3220                                                       ext4_ext_pblock(ex2));
3221                         } else {
3222                                 err = ext4_ext_zeroout(inode, ex);
3223                                 zero_ex.ee_block = ex->ee_block;
3224                                 zero_ex.ee_len = cpu_to_le16(
3225                                                 ext4_ext_get_actual_len(ex));
3226                                 ext4_ext_store_pblock(&zero_ex,
3227                                                       ext4_ext_pblock(ex));
3228                         }
3229                 } else {
3230                         err = ext4_ext_zeroout(inode, &orig_ex);
3231                         zero_ex.ee_block = orig_ex.ee_block;
3232                         zero_ex.ee_len = cpu_to_le16(
3233                                                 ext4_ext_get_actual_len(&orig_ex));
3234                         ext4_ext_store_pblock(&zero_ex,
3235                                               ext4_ext_pblock(&orig_ex));
3236                 }
3237
3238                 if (err)
3239                         goto fix_extent_len;
3240                 /* update the extent length and mark as initialized */
3241                 ex->ee_len = cpu_to_le16(ee_len);
3242                 ext4_ext_try_to_merge(handle, inode, path, ex);
3243                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3244                 if (err)
3245                         goto fix_extent_len;
3246
3247                 /* update extent status tree */
3248                 err = ext4_zeroout_es(inode, &zero_ex);
3249
3250                 goto out;
3251         } else if (err)
3252                 goto fix_extent_len;
3253
3254 out:
3255         ext4_ext_show_leaf(inode, path);
3256         return err;
3257
3258 fix_extent_len:
3259         ex->ee_len = orig_ex.ee_len;
3260         ext4_ext_dirty(handle, inode, path + path->p_depth);
3261         return err;
3262 }
3263
3264 /*
3265  * ext4_split_extents() splits an extent and mark extent which is covered
3266  * by @map as split_flags indicates
3267  *
3268  * It may result in splitting the extent into multiple extents (up to three)
3269  * There are three possibilities:
3270  *   a> There is no split required
3271  *   b> Splits in two extents: Split is happening at either end of the extent
3272  *   c> Splits in three extents: Somone is splitting in middle of the extent
3273  *
3274  */
3275 static int ext4_split_extent(handle_t *handle,
3276                               struct inode *inode,
3277                               struct ext4_ext_path *path,
3278                               struct ext4_map_blocks *map,
3279                               int split_flag,
3280                               int flags)
3281 {
3282         ext4_lblk_t ee_block;
3283         struct ext4_extent *ex;
3284         unsigned int ee_len, depth;
3285         int err = 0;
3286         int unwritten;
3287         int split_flag1, flags1;
3288         int allocated = map->m_len;
3289
3290         depth = ext_depth(inode);
3291         ex = path[depth].p_ext;
3292         ee_block = le32_to_cpu(ex->ee_block);
3293         ee_len = ext4_ext_get_actual_len(ex);
3294         unwritten = ext4_ext_is_unwritten(ex);
3295
3296         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3297                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3298                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3299                 if (unwritten)
3300                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3301                                        EXT4_EXT_MARK_UNWRIT2;
3302                 if (split_flag & EXT4_EXT_DATA_VALID2)
3303                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3304                 err = ext4_split_extent_at(handle, inode, path,
3305                                 map->m_lblk + map->m_len, split_flag1, flags1);
3306                 if (err)
3307                         goto out;
3308         } else {
3309                 allocated = ee_len - (map->m_lblk - ee_block);
3310         }
3311         /*
3312          * Update path is required because previous ext4_split_extent_at() may
3313          * result in split of original leaf or extent zeroout.
3314          */
3315         ext4_ext_drop_refs(path);
3316         path = ext4_ext_find_extent(inode, map->m_lblk, path, 0);
3317         if (IS_ERR(path))
3318                 return PTR_ERR(path);
3319         depth = ext_depth(inode);
3320         ex = path[depth].p_ext;
3321         if (!ex) {
3322                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3323                                  (unsigned long) map->m_lblk);
3324                 return -EIO;
3325         }
3326         unwritten = ext4_ext_is_unwritten(ex);
3327         split_flag1 = 0;
3328
3329         if (map->m_lblk >= ee_block) {
3330                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3331                 if (unwritten) {
3332                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3333                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3334                                                      EXT4_EXT_MARK_UNWRIT2);
3335                 }
3336                 err = ext4_split_extent_at(handle, inode, path,
3337                                 map->m_lblk, split_flag1, flags);
3338                 if (err)
3339                         goto out;
3340         }
3341
3342         ext4_ext_show_leaf(inode, path);
3343 out:
3344         return err ? err : allocated;
3345 }
3346
3347 /*
3348  * This function is called by ext4_ext_map_blocks() if someone tries to write
3349  * to an unwritten extent. It may result in splitting the unwritten
3350  * extent into multiple extents (up to three - one initialized and two
3351  * unwritten).
3352  * There are three possibilities:
3353  *   a> There is no split required: Entire extent should be initialized
3354  *   b> Splits in two extents: Write is happening at either end of the extent
3355  *   c> Splits in three extents: Somone is writing in middle of the extent
3356  *
3357  * Pre-conditions:
3358  *  - The extent pointed to by 'path' is unwritten.
3359  *  - The extent pointed to by 'path' contains a superset
3360  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3361  *
3362  * Post-conditions on success:
3363  *  - the returned value is the number of blocks beyond map->l_lblk
3364  *    that are allocated and initialized.
3365  *    It is guaranteed to be >= map->m_len.
3366  */
3367 static int ext4_ext_convert_to_initialized(handle_t *handle,
3368                                            struct inode *inode,
3369                                            struct ext4_map_blocks *map,
3370                                            struct ext4_ext_path *path,
3371                                            int flags)
3372 {
3373         struct ext4_sb_info *sbi;
3374         struct ext4_extent_header *eh;
3375         struct ext4_map_blocks split_map;
3376         struct ext4_extent zero_ex;
3377         struct ext4_extent *ex, *abut_ex;
3378         ext4_lblk_t ee_block, eof_block;
3379         unsigned int ee_len, depth, map_len = map->m_len;
3380         int allocated = 0, max_zeroout = 0;
3381         int err = 0;
3382         int split_flag = 0;
3383
3384         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3385                 "block %llu, max_blocks %u\n", inode->i_ino,
3386                 (unsigned long long)map->m_lblk, map_len);
3387
3388         sbi = EXT4_SB(inode->i_sb);
3389         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3390                 inode->i_sb->s_blocksize_bits;
3391         if (eof_block < map->m_lblk + map_len)
3392                 eof_block = map->m_lblk + map_len;
3393
3394         depth = ext_depth(inode);
3395         eh = path[depth].p_hdr;
3396         ex = path[depth].p_ext;
3397         ee_block = le32_to_cpu(ex->ee_block);
3398         ee_len = ext4_ext_get_actual_len(ex);
3399         zero_ex.ee_len = 0;
3400
3401         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3402
3403         /* Pre-conditions */
3404         BUG_ON(!ext4_ext_is_unwritten(ex));
3405         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3406
3407         /*
3408          * Attempt to transfer newly initialized blocks from the currently
3409          * unwritten extent to its neighbor. This is much cheaper
3410          * than an insertion followed by a merge as those involve costly
3411          * memmove() calls. Transferring to the left is the common case in
3412          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3413          * followed by append writes.
3414          *
3415          * Limitations of the current logic:
3416          *  - L1: we do not deal with writes covering the whole extent.
3417          *    This would require removing the extent if the transfer
3418          *    is possible.
3419          *  - L2: we only attempt to merge with an extent stored in the
3420          *    same extent tree node.
3421          */
3422         if ((map->m_lblk == ee_block) &&
3423                 /* See if we can merge left */
3424                 (map_len < ee_len) &&           /*L1*/
3425                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3426                 ext4_lblk_t prev_lblk;
3427                 ext4_fsblk_t prev_pblk, ee_pblk;
3428                 unsigned int prev_len;
3429
3430                 abut_ex = ex - 1;
3431                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3432                 prev_len = ext4_ext_get_actual_len(abut_ex);
3433                 prev_pblk = ext4_ext_pblock(abut_ex);
3434                 ee_pblk = ext4_ext_pblock(ex);
3435
3436                 /*
3437                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3438                  * upon those conditions:
3439                  * - C1: abut_ex is initialized,
3440                  * - C2: abut_ex is logically abutting ex,
3441                  * - C3: abut_ex is physically abutting ex,
3442                  * - C4: abut_ex can receive the additional blocks without
3443                  *   overflowing the (initialized) length limit.
3444                  */
3445                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3446                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3447                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3448                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3449                         err = ext4_ext_get_access(handle, inode, path + depth);
3450                         if (err)
3451                                 goto out;
3452
3453                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3454                                 map, ex, abut_ex);
3455
3456                         /* Shift the start of ex by 'map_len' blocks */
3457                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3458                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3459                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3460                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3461
3462                         /* Extend abut_ex by 'map_len' blocks */
3463                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3464
3465                         /* Result: number of initialized blocks past m_lblk */
3466                         allocated = map_len;
3467                 }
3468         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3469                    (map_len < ee_len) &&        /*L1*/
3470                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3471                 /* See if we can merge right */
3472                 ext4_lblk_t next_lblk;
3473                 ext4_fsblk_t next_pblk, ee_pblk;
3474                 unsigned int next_len;
3475
3476                 abut_ex = ex + 1;
3477                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3478                 next_len = ext4_ext_get_actual_len(abut_ex);
3479                 next_pblk = ext4_ext_pblock(abut_ex);
3480                 ee_pblk = ext4_ext_pblock(ex);
3481
3482                 /*
3483                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3484                  * upon those conditions:
3485                  * - C1: abut_ex is initialized,
3486                  * - C2: abut_ex is logically abutting ex,
3487                  * - C3: abut_ex is physically abutting ex,
3488                  * - C4: abut_ex can receive the additional blocks without
3489                  *   overflowing the (initialized) length limit.
3490                  */
3491                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3492                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3493                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3494                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3495                         err = ext4_ext_get_access(handle, inode, path + depth);
3496                         if (err)
3497                                 goto out;
3498
3499                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3500                                 map, ex, abut_ex);
3501
3502                         /* Shift the start of abut_ex by 'map_len' blocks */
3503                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3504                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3505                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3506                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3507
3508                         /* Extend abut_ex by 'map_len' blocks */
3509                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3510
3511                         /* Result: number of initialized blocks past m_lblk */
3512                         allocated = map_len;
3513                 }
3514         }
3515         if (allocated) {
3516                 /* Mark the block containing both extents as dirty */
3517                 ext4_ext_dirty(handle, inode, path + depth);
3518
3519                 /* Update path to point to the right extent */
3520                 path[depth].p_ext = abut_ex;
3521                 goto out;
3522         } else
3523                 allocated = ee_len - (map->m_lblk - ee_block);
3524
3525         WARN_ON(map->m_lblk < ee_block);
3526         /*
3527          * It is safe to convert extent to initialized via explicit
3528          * zeroout only if extent is fully inside i_size or new_size.
3529          */
3530         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3531
3532         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3533                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3534                         (inode->i_sb->s_blocksize_bits - 10);
3535
3536         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3537         if (max_zeroout && (ee_len <= max_zeroout)) {
3538                 err = ext4_ext_zeroout(inode, ex);
3539                 if (err)
3540                         goto out;
3541                 zero_ex.ee_block = ex->ee_block;
3542                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3543                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3544
3545                 err = ext4_ext_get_access(handle, inode, path + depth);
3546                 if (err)
3547                         goto out;
3548                 ext4_ext_mark_initialized(ex);
3549                 ext4_ext_try_to_merge(handle, inode, path, ex);
3550                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3551                 goto out;
3552         }
3553
3554         /*
3555          * four cases:
3556          * 1. split the extent into three extents.
3557          * 2. split the extent into two extents, zeroout the first half.
3558          * 3. split the extent into two extents, zeroout the second half.
3559          * 4. split the extent into two extents with out zeroout.
3560          */
3561         split_map.m_lblk = map->m_lblk;
3562         split_map.m_len = map->m_len;
3563
3564         if (max_zeroout && (allocated > map->m_len)) {
3565                 if (allocated <= max_zeroout) {
3566                         /* case 3 */
3567                         zero_ex.ee_block =
3568                                          cpu_to_le32(map->m_lblk);
3569                         zero_ex.ee_len = cpu_to_le16(allocated);
3570                         ext4_ext_store_pblock(&zero_ex,
3571                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3572                         err = ext4_ext_zeroout(inode, &zero_ex);
3573                         if (err)
3574                                 goto out;
3575                         split_map.m_lblk = map->m_lblk;
3576                         split_map.m_len = allocated;
3577                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3578                         /* case 2 */
3579                         if (map->m_lblk != ee_block) {
3580                                 zero_ex.ee_block = ex->ee_block;
3581                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3582                                                         ee_block);
3583                                 ext4_ext_store_pblock(&zero_ex,
3584                                                       ext4_ext_pblock(ex));
3585                                 err = ext4_ext_zeroout(inode, &zero_ex);
3586                                 if (err)
3587                                         goto out;
3588                         }
3589
3590                         split_map.m_lblk = ee_block;
3591                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3592                         allocated = map->m_len;
3593                 }
3594         }
3595
3596         allocated = ext4_split_extent(handle, inode, path,
3597                                       &split_map, split_flag, flags);
3598         if (allocated < 0)
3599                 err = allocated;
3600
3601 out:
3602         /* If we have gotten a failure, don't zero out status tree */
3603         if (!err)
3604                 err = ext4_zeroout_es(inode, &zero_ex);
3605         return err ? err : allocated;
3606 }
3607
3608 /*
3609  * This function is called by ext4_ext_map_blocks() from
3610  * ext4_get_blocks_dio_write() when DIO to write
3611  * to an unwritten extent.
3612  *
3613  * Writing to an unwritten extent may result in splitting the unwritten
3614  * extent into multiple initialized/unwritten extents (up to three)
3615  * There are three possibilities:
3616  *   a> There is no split required: Entire extent should be unwritten
3617  *   b> Splits in two extents: Write is happening at either end of the extent
3618  *   c> Splits in three extents: Somone is writing in middle of the extent
3619  *
3620  * This works the same way in the case of initialized -> unwritten conversion.
3621  *
3622  * One of more index blocks maybe needed if the extent tree grow after
3623  * the unwritten extent split. To prevent ENOSPC occur at the IO
3624  * complete, we need to split the unwritten extent before DIO submit
3625  * the IO. The unwritten extent called at this time will be split
3626  * into three unwritten extent(at most). After IO complete, the part
3627  * being filled will be convert to initialized by the end_io callback function
3628  * via ext4_convert_unwritten_extents().
3629  *
3630  * Returns the size of unwritten extent to be written on success.
3631  */
3632 static int ext4_split_convert_extents(handle_t *handle,
3633                                         struct inode *inode,
3634                                         struct ext4_map_blocks *map,
3635                                         struct ext4_ext_path *path,
3636                                         int flags)
3637 {
3638         ext4_lblk_t eof_block;
3639         ext4_lblk_t ee_block;
3640         struct ext4_extent *ex;
3641         unsigned int ee_len;
3642         int split_flag = 0, depth;
3643
3644         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3645                   __func__, inode->i_ino,
3646                   (unsigned long long)map->m_lblk, map->m_len);
3647
3648         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3649                 inode->i_sb->s_blocksize_bits;
3650         if (eof_block < map->m_lblk + map->m_len)
3651                 eof_block = map->m_lblk + map->m_len;
3652         /*
3653          * It is safe to convert extent to initialized via explicit
3654          * zeroout only if extent is fully insde i_size or new_size.
3655          */
3656         depth = ext_depth(inode);
3657         ex = path[depth].p_ext;
3658         ee_block = le32_to_cpu(ex->ee_block);
3659         ee_len = ext4_ext_get_actual_len(ex);
3660
3661         /* Convert to unwritten */
3662         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3663                 split_flag |= EXT4_EXT_DATA_VALID1;
3664         /* Convert to initialized */
3665         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3666                 split_flag |= ee_block + ee_len <= eof_block ?
3667                               EXT4_EXT_MAY_ZEROOUT : 0;
3668                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3669         }
3670         flags |= EXT4_GET_BLOCKS_PRE_IO;
3671         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3672 }
3673
3674 static int ext4_convert_initialized_extents(handle_t *handle,
3675                                             struct inode *inode,
3676                                             struct ext4_map_blocks *map,
3677                                             struct ext4_ext_path *path)
3678 {
3679         struct ext4_extent *ex;
3680         ext4_lblk_t ee_block;
3681         unsigned int ee_len;
3682         int depth;
3683         int err = 0;
3684
3685         depth = ext_depth(inode);
3686         ex = path[depth].p_ext;
3687         ee_block = le32_to_cpu(ex->ee_block);
3688         ee_len = ext4_ext_get_actual_len(ex);
3689
3690         ext_debug("%s: inode %lu, logical"
3691                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3692                   (unsigned long long)ee_block, ee_len);
3693
3694         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3695                 err = ext4_split_convert_extents(handle, inode, map, path,
3696                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3697                 if (err < 0)
3698                         goto out;
3699                 ext4_ext_drop_refs(path);
3700                 path = ext4_ext_find_extent(inode, map->m_lblk, path, 0);
3701                 if (IS_ERR(path)) {
3702                         err = PTR_ERR(path);
3703                         goto out;
3704                 }
3705                 depth = ext_depth(inode);
3706                 ex = path[depth].p_ext;
3707                 if (!ex) {
3708                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3709                                          (unsigned long) map->m_lblk);
3710                         err = -EIO;
3711                         goto out;
3712                 }
3713         }
3714
3715         err = ext4_ext_get_access(handle, inode, path + depth);
3716         if (err)
3717                 goto out;
3718         /* first mark the extent as unwritten */
3719         ext4_ext_mark_unwritten(ex);
3720
3721         /* note: ext4_ext_correct_indexes() isn't needed here because
3722          * borders are not changed
3723          */
3724         ext4_ext_try_to_merge(handle, inode, path, ex);
3725
3726         /* Mark modified extent as dirty */
3727         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3728 out:
3729         ext4_ext_show_leaf(inode, path);
3730         return err;
3731 }
3732
3733
3734 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3735                                                 struct inode *inode,
3736                                                 struct ext4_map_blocks *map,
3737                                                 struct ext4_ext_path *path)
3738 {
3739         struct ext4_extent *ex;
3740         ext4_lblk_t ee_block;
3741         unsigned int ee_len;
3742         int depth;
3743         int err = 0;
3744
3745         depth = ext_depth(inode);
3746         ex = path[depth].p_ext;
3747         ee_block = le32_to_cpu(ex->ee_block);
3748         ee_len = ext4_ext_get_actual_len(ex);
3749
3750         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3751                 "block %llu, max_blocks %u\n", inode->i_ino,
3752                   (unsigned long long)ee_block, ee_len);
3753
3754         /* If extent is larger than requested it is a clear sign that we still
3755          * have some extent state machine issues left. So extent_split is still
3756          * required.
3757          * TODO: Once all related issues will be fixed this situation should be
3758          * illegal.
3759          */
3760         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3761 #ifdef EXT4_DEBUG
3762                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3763                              " len %u; IO logical block %llu, len %u\n",
3764                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3765                              (unsigned long long)map->m_lblk, map->m_len);
3766 #endif
3767                 err = ext4_split_convert_extents(handle, inode, map, path,
3768                                                  EXT4_GET_BLOCKS_CONVERT);
3769                 if (err < 0)
3770                         goto out;
3771                 ext4_ext_drop_refs(path);
3772                 path = ext4_ext_find_extent(inode, map->m_lblk, path, 0);
3773                 if (IS_ERR(path)) {
3774                         err = PTR_ERR(path);
3775                         goto out;
3776                 }
3777                 depth = ext_depth(inode);
3778                 ex = path[depth].p_ext;
3779         }
3780
3781         err = ext4_ext_get_access(handle, inode, path + depth);
3782         if (err)
3783                 goto out;
3784         /* first mark the extent as initialized */
3785         ext4_ext_mark_initialized(ex);
3786
3787         /* note: ext4_ext_correct_indexes() isn't needed here because
3788          * borders are not changed
3789          */
3790         ext4_ext_try_to_merge(handle, inode, path, ex);
3791
3792         /* Mark modified extent as dirty */
3793         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3794 out:
3795         ext4_ext_show_leaf(inode, path);
3796         return err;
3797 }
3798
3799 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3800                         sector_t block, int count)
3801 {
3802         int i;
3803         for (i = 0; i < count; i++)
3804                 unmap_underlying_metadata(bdev, block + i);
3805 }
3806
3807 /*
3808  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3809  */
3810 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3811                               ext4_lblk_t lblk,
3812                               struct ext4_ext_path *path,
3813                               unsigned int len)
3814 {
3815         int i, depth;
3816         struct ext4_extent_header *eh;
3817         struct ext4_extent *last_ex;
3818
3819         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3820                 return 0;
3821
3822         depth = ext_depth(inode);
3823         eh = path[depth].p_hdr;
3824
3825         /*
3826          * We're going to remove EOFBLOCKS_FL entirely in future so we
3827          * do not care for this case anymore. Simply remove the flag
3828          * if there are no extents.
3829          */
3830         if (unlikely(!eh->eh_entries))
3831                 goto out;
3832         last_ex = EXT_LAST_EXTENT(eh);
3833         /*
3834          * We should clear the EOFBLOCKS_FL flag if we are writing the
3835          * last block in the last extent in the file.  We test this by
3836          * first checking to see if the caller to
3837          * ext4_ext_get_blocks() was interested in the last block (or
3838          * a block beyond the last block) in the current extent.  If
3839          * this turns out to be false, we can bail out from this
3840          * function immediately.
3841          */
3842         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3843             ext4_ext_get_actual_len(last_ex))
3844                 return 0;
3845         /*
3846          * If the caller does appear to be planning to write at or
3847          * beyond the end of the current extent, we then test to see
3848          * if the current extent is the last extent in the file, by
3849          * checking to make sure it was reached via the rightmost node
3850          * at each level of the tree.
3851          */
3852         for (i = depth-1; i >= 0; i--)
3853                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3854                         return 0;
3855 out:
3856         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3857         return ext4_mark_inode_dirty(handle, inode);
3858 }
3859
3860 /**
3861  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3862  *
3863  * Return 1 if there is a delalloc block in the range, otherwise 0.
3864  */
3865 int ext4_find_delalloc_range(struct inode *inode,
3866                              ext4_lblk_t lblk_start,
3867                              ext4_lblk_t lblk_end)
3868 {
3869         struct extent_status es;
3870
3871         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3872         if (es.es_len == 0)
3873                 return 0; /* there is no delay extent in this tree */
3874         else if (es.es_lblk <= lblk_start &&
3875                  lblk_start < es.es_lblk + es.es_len)
3876                 return 1;
3877         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3878                 return 1;
3879         else
3880                 return 0;
3881 }
3882
3883 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3884 {
3885         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3886         ext4_lblk_t lblk_start, lblk_end;
3887         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3888         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3889
3890         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3891 }
3892
3893 /**
3894  * Determines how many complete clusters (out of those specified by the 'map')
3895  * are under delalloc and were reserved quota for.
3896  * This function is called when we are writing out the blocks that were
3897  * originally written with their allocation delayed, but then the space was
3898  * allocated using fallocate() before the delayed allocation could be resolved.
3899  * The cases to look for are:
3900  * ('=' indicated delayed allocated blocks
3901  *  '-' indicates non-delayed allocated blocks)
3902  * (a) partial clusters towards beginning and/or end outside of allocated range
3903  *     are not delalloc'ed.
3904  *      Ex:
3905  *      |----c---=|====c====|====c====|===-c----|
3906  *               |++++++ allocated ++++++|
3907  *      ==> 4 complete clusters in above example
3908  *
3909  * (b) partial cluster (outside of allocated range) towards either end is
3910  *     marked for delayed allocation. In this case, we will exclude that
3911  *     cluster.
3912  *      Ex:
3913  *      |----====c========|========c========|
3914  *           |++++++ allocated ++++++|
3915  *      ==> 1 complete clusters in above example
3916  *
3917  *      Ex:
3918  *      |================c================|
3919  *            |++++++ allocated ++++++|
3920  *      ==> 0 complete clusters in above example
3921  *
3922  * The ext4_da_update_reserve_space will be called only if we
3923  * determine here that there were some "entire" clusters that span
3924  * this 'allocated' range.
3925  * In the non-bigalloc case, this function will just end up returning num_blks
3926  * without ever calling ext4_find_delalloc_range.
3927  */
3928 static unsigned int
3929 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3930                            unsigned int num_blks)
3931 {
3932         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3933         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3934         ext4_lblk_t lblk_from, lblk_to, c_offset;
3935         unsigned int allocated_clusters = 0;
3936
3937         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3938         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3939
3940         /* max possible clusters for this allocation */
3941         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3942
3943         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3944
3945         /* Check towards left side */
3946         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3947         if (c_offset) {
3948                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3949                 lblk_to = lblk_from + c_offset - 1;
3950
3951                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3952                         allocated_clusters--;
3953         }
3954
3955         /* Now check towards right. */
3956         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3957         if (allocated_clusters && c_offset) {
3958                 lblk_from = lblk_start + num_blks;
3959                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3960
3961                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3962                         allocated_clusters--;
3963         }
3964
3965         return allocated_clusters;
3966 }
3967
3968 static int
3969 ext4_ext_convert_initialized_extent(handle_t *handle, struct inode *inode,
3970                         struct ext4_map_blocks *map,
3971                         struct ext4_ext_path *path, int flags,
3972                         unsigned int allocated, ext4_fsblk_t newblock)
3973 {
3974         int ret = 0;
3975         int err = 0;
3976
3977         /*
3978          * Make sure that the extent is no bigger than we support with
3979          * unwritten extent
3980          */
3981         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3982                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3983
3984         ret = ext4_convert_initialized_extents(handle, inode, map,
3985                                                 path);
3986         if (ret >= 0) {
3987                 ext4_update_inode_fsync_trans(handle, inode, 1);
3988                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
3989                                          path, map->m_len);
3990         } else
3991                 err = ret;
3992         map->m_flags |= EXT4_MAP_UNWRITTEN;
3993         if (allocated > map->m_len)
3994                 allocated = map->m_len;
3995         map->m_len = allocated;
3996
3997         return err ? err : allocated;
3998 }
3999
4000 static int
4001 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4002                         struct ext4_map_blocks *map,
4003                         struct ext4_ext_path *path, int flags,
4004                         unsigned int allocated, ext4_fsblk_t newblock)
4005 {
4006         int ret = 0;
4007         int err = 0;
4008         ext4_io_end_t *io = ext4_inode_aio(inode);
4009
4010         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4011                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4012                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4013                   flags, allocated);
4014         ext4_ext_show_leaf(inode, path);
4015
4016         /*
4017          * When writing into unwritten space, we should not fail to
4018          * allocate metadata blocks for the new extent block if needed.
4019          */
4020         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4021
4022         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4023                                                     allocated, newblock);
4024
4025         /* get_block() before submit the IO, split the extent */
4026         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4027                 ret = ext4_split_convert_extents(handle, inode, map,
4028                                          path, flags | EXT4_GET_BLOCKS_CONVERT);
4029                 if (ret <= 0)
4030                         goto out;
4031                 /*
4032                  * Flag the inode(non aio case) or end_io struct (aio case)
4033                  * that this IO needs to conversion to written when IO is
4034                  * completed
4035                  */
4036                 if (io)
4037                         ext4_set_io_unwritten_flag(inode, io);
4038                 else
4039                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4040                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4041                 goto out;
4042         }
4043         /* IO end_io complete, convert the filled extent to written */
4044         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4045                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4046                                                         path);
4047                 if (ret >= 0) {
4048                         ext4_update_inode_fsync_trans(handle, inode, 1);
4049                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4050                                                  path, map->m_len);
4051                 } else
4052                         err = ret;
4053                 map->m_flags |= EXT4_MAP_MAPPED;
4054                 map->m_pblk = newblock;
4055                 if (allocated > map->m_len)
4056                         allocated = map->m_len;
4057                 map->m_len = allocated;
4058                 goto out2;
4059         }
4060         /* buffered IO case */
4061         /*
4062          * repeat fallocate creation request
4063          * we already have an unwritten extent
4064          */
4065         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4066                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4067                 goto map_out;
4068         }
4069
4070         /* buffered READ or buffered write_begin() lookup */
4071         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4072                 /*
4073                  * We have blocks reserved already.  We
4074                  * return allocated blocks so that delalloc
4075                  * won't do block reservation for us.  But
4076                  * the buffer head will be unmapped so that
4077                  * a read from the block returns 0s.
4078                  */
4079                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4080                 goto out1;
4081         }
4082
4083         /* buffered write, writepage time, convert*/
4084         ret = ext4_ext_convert_to_initialized(handle, inode, map, path, flags);
4085         if (ret >= 0)
4086                 ext4_update_inode_fsync_trans(handle, inode, 1);
4087 out:
4088         if (ret <= 0) {
4089                 err = ret;
4090                 goto out2;
4091         } else
4092                 allocated = ret;
4093         map->m_flags |= EXT4_MAP_NEW;
4094         /*
4095          * if we allocated more blocks than requested
4096          * we need to make sure we unmap the extra block
4097          * allocated. The actual needed block will get
4098          * unmapped later when we find the buffer_head marked
4099          * new.
4100          */
4101         if (allocated > map->m_len) {
4102                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
4103                                         newblock + map->m_len,
4104                                         allocated - map->m_len);
4105                 allocated = map->m_len;
4106         }
4107         map->m_len = allocated;
4108
4109         /*
4110          * If we have done fallocate with the offset that is already
4111          * delayed allocated, we would have block reservation
4112          * and quota reservation done in the delayed write path.
4113          * But fallocate would have already updated quota and block
4114          * count for this offset. So cancel these reservation
4115          */
4116         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4117                 unsigned int reserved_clusters;
4118                 reserved_clusters = get_reserved_cluster_alloc(inode,
4119                                 map->m_lblk, map->m_len);
4120                 if (reserved_clusters)
4121                         ext4_da_update_reserve_space(inode,
4122                                                      reserved_clusters,
4123                                                      0);
4124         }
4125
4126 map_out:
4127         map->m_flags |= EXT4_MAP_MAPPED;
4128         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4129                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4130                                          map->m_len);
4131                 if (err < 0)
4132                         goto out2;
4133         }
4134 out1:
4135         if (allocated > map->m_len)
4136                 allocated = map->m_len;
4137         ext4_ext_show_leaf(inode, path);
4138         map->m_pblk = newblock;
4139         map->m_len = allocated;
4140 out2:
4141         return err ? err : allocated;
4142 }
4143
4144 /*
4145  * get_implied_cluster_alloc - check to see if the requested
4146  * allocation (in the map structure) overlaps with a cluster already
4147  * allocated in an extent.
4148  *      @sb     The filesystem superblock structure
4149  *      @map    The requested lblk->pblk mapping
4150  *      @ex     The extent structure which might contain an implied
4151  *                      cluster allocation
4152  *
4153  * This function is called by ext4_ext_map_blocks() after we failed to
4154  * find blocks that were already in the inode's extent tree.  Hence,
4155  * we know that the beginning of the requested region cannot overlap
4156  * the extent from the inode's extent tree.  There are three cases we
4157  * want to catch.  The first is this case:
4158  *
4159  *               |--- cluster # N--|
4160  *    |--- extent ---|  |---- requested region ---|
4161  *                      |==========|
4162  *
4163  * The second case that we need to test for is this one:
4164  *
4165  *   |--------- cluster # N ----------------|
4166  *         |--- requested region --|   |------- extent ----|
4167  *         |=======================|
4168  *
4169  * The third case is when the requested region lies between two extents
4170  * within the same cluster:
4171  *          |------------- cluster # N-------------|
4172  * |----- ex -----|                  |---- ex_right ----|
4173  *                  |------ requested region ------|
4174  *                  |================|
4175  *
4176  * In each of the above cases, we need to set the map->m_pblk and
4177  * map->m_len so it corresponds to the return the extent labelled as
4178  * "|====|" from cluster #N, since it is already in use for data in
4179  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4180  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4181  * as a new "allocated" block region.  Otherwise, we will return 0 and
4182  * ext4_ext_map_blocks() will then allocate one or more new clusters
4183  * by calling ext4_mb_new_blocks().
4184  */
4185 static int get_implied_cluster_alloc(struct super_block *sb,
4186                                      struct ext4_map_blocks *map,
4187                                      struct ext4_extent *ex,
4188                                      struct ext4_ext_path *path)
4189 {
4190         struct ext4_sb_info *sbi = EXT4_SB(sb);
4191         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4192         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4193         ext4_lblk_t rr_cluster_start;
4194         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4195         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4196         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4197
4198         /* The extent passed in that we are trying to match */
4199         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4200         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4201
4202         /* The requested region passed into ext4_map_blocks() */
4203         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4204
4205         if ((rr_cluster_start == ex_cluster_end) ||
4206             (rr_cluster_start == ex_cluster_start)) {
4207                 if (rr_cluster_start == ex_cluster_end)
4208                         ee_start += ee_len - 1;
4209                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4210                 map->m_len = min(map->m_len,
4211                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4212                 /*
4213                  * Check for and handle this case:
4214                  *
4215                  *   |--------- cluster # N-------------|
4216                  *                     |------- extent ----|
4217                  *         |--- requested region ---|
4218                  *         |===========|
4219                  */
4220
4221                 if (map->m_lblk < ee_block)
4222                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4223
4224                 /*
4225                  * Check for the case where there is already another allocated
4226                  * block to the right of 'ex' but before the end of the cluster.
4227                  *
4228                  *          |------------- cluster # N-------------|
4229                  * |----- ex -----|                  |---- ex_right ----|
4230                  *                  |------ requested region ------|
4231                  *                  |================|
4232                  */
4233                 if (map->m_lblk > ee_block) {
4234                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4235                         map->m_len = min(map->m_len, next - map->m_lblk);
4236                 }
4237
4238                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4239                 return 1;
4240         }
4241
4242         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4243         return 0;
4244 }
4245
4246
4247 /*
4248  * Block allocation/map/preallocation routine for extents based files
4249  *
4250  *
4251  * Need to be called with
4252  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4253  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4254  *
4255  * return > 0, number of of blocks already mapped/allocated
4256  *          if create == 0 and these are pre-allocated blocks
4257  *              buffer head is unmapped
4258  *          otherwise blocks are mapped
4259  *
4260  * return = 0, if plain look up failed (blocks have not been allocated)
4261  *          buffer head is unmapped
4262  *
4263  * return < 0, error case.
4264  */
4265 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4266                         struct ext4_map_blocks *map, int flags)
4267 {
4268         struct ext4_ext_path *path = NULL;
4269         struct ext4_extent newex, *ex, *ex2;
4270         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4271         ext4_fsblk_t newblock = 0;
4272         int free_on_err = 0, err = 0, depth, ret;
4273         unsigned int allocated = 0, offset = 0;
4274         unsigned int allocated_clusters = 0;
4275         struct ext4_allocation_request ar;
4276         ext4_io_end_t *io = ext4_inode_aio(inode);
4277         ext4_lblk_t cluster_offset;
4278         int set_unwritten = 0;
4279
4280         ext_debug("blocks %u/%u requested for inode %lu\n",
4281                   map->m_lblk, map->m_len, inode->i_ino);
4282         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4283
4284         /* find extent for this block */
4285         path = ext4_ext_find_extent(inode, map->m_lblk, NULL, 0);
4286         if (IS_ERR(path)) {
4287                 err = PTR_ERR(path);
4288                 path = NULL;
4289                 goto out2;
4290         }
4291
4292         depth = ext_depth(inode);
4293
4294         /*
4295          * consistent leaf must not be empty;
4296          * this situation is possible, though, _during_ tree modification;
4297          * this is why assert can't be put in ext4_ext_find_extent()
4298          */
4299         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4300                 EXT4_ERROR_INODE(inode, "bad extent address "
4301                                  "lblock: %lu, depth: %d pblock %lld",
4302                                  (unsigned long) map->m_lblk, depth,
4303                                  path[depth].p_block);
4304                 err = -EIO;
4305                 goto out2;
4306         }
4307
4308         ex = path[depth].p_ext;
4309         if (ex) {
4310                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4311                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4312                 unsigned short ee_len;
4313
4314
4315                 /*
4316                  * unwritten extents are treated as holes, except that
4317                  * we split out initialized portions during a write.
4318                  */
4319                 ee_len = ext4_ext_get_actual_len(ex);
4320
4321                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4322
4323                 /* if found extent covers block, simply return it */
4324                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4325                         newblock = map->m_lblk - ee_block + ee_start;
4326                         /* number of remaining blocks in the extent */
4327                         allocated = ee_len - (map->m_lblk - ee_block);
4328                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4329                                   ee_block, ee_len, newblock);
4330
4331                         /*
4332                          * If the extent is initialized check whether the
4333                          * caller wants to convert it to unwritten.
4334                          */
4335                         if ((!ext4_ext_is_unwritten(ex)) &&
4336                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4337                                 allocated = ext4_ext_convert_initialized_extent(
4338                                                 handle, inode, map, path, flags,
4339                                                 allocated, newblock);
4340                                 goto out2;
4341                         } else if (!ext4_ext_is_unwritten(ex))
4342                                 goto out;
4343
4344                         ret = ext4_ext_handle_unwritten_extents(
4345                                 handle, inode, map, path, flags,
4346                                 allocated, newblock);
4347                         if (ret < 0)
4348                                 err = ret;
4349                         else
4350                                 allocated = ret;
4351                         goto out2;
4352                 }
4353         }
4354
4355         if ((sbi->s_cluster_ratio > 1) &&
4356             ext4_find_delalloc_cluster(inode, map->m_lblk))
4357                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4358
4359         /*
4360          * requested block isn't allocated yet;
4361          * we couldn't try to create block if create flag is zero
4362          */
4363         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4364                 /*
4365                  * put just found gap into cache to speed up
4366                  * subsequent requests
4367                  */
4368                 if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
4369                         ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4370                 goto out2;
4371         }
4372
4373         /*
4374          * Okay, we need to do block allocation.
4375          */
4376         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
4377         newex.ee_block = cpu_to_le32(map->m_lblk);
4378         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4379
4380         /*
4381          * If we are doing bigalloc, check to see if the extent returned
4382          * by ext4_ext_find_extent() implies a cluster we can use.
4383          */
4384         if (cluster_offset && ex &&
4385             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4386                 ar.len = allocated = map->m_len;
4387                 newblock = map->m_pblk;
4388                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4389                 goto got_allocated_blocks;
4390         }
4391
4392         /* find neighbour allocated blocks */
4393         ar.lleft = map->m_lblk;
4394         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4395         if (err)
4396                 goto out2;
4397         ar.lright = map->m_lblk;
4398         ex2 = NULL;
4399         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4400         if (err)
4401                 goto out2;
4402
4403         /* Check if the extent after searching to the right implies a
4404          * cluster we can use. */
4405         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4406             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4407                 ar.len = allocated = map->m_len;
4408                 newblock = map->m_pblk;
4409                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4410                 goto got_allocated_blocks;
4411         }
4412
4413         /*
4414          * See if request is beyond maximum number of blocks we can have in
4415          * a single extent. For an initialized extent this limit is
4416          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4417          * EXT_UNWRITTEN_MAX_LEN.
4418          */
4419         if (map->m_len > EXT_INIT_MAX_LEN &&
4420             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4421                 map->m_len = EXT_INIT_MAX_LEN;
4422         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4423                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4424                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4425
4426         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4427         newex.ee_len = cpu_to_le16(map->m_len);
4428         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4429         if (err)
4430                 allocated = ext4_ext_get_actual_len(&newex);
4431         else
4432                 allocated = map->m_len;
4433
4434         /* allocate new block */
4435         ar.inode = inode;
4436         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4437         ar.logical = map->m_lblk;
4438         /*
4439          * We calculate the offset from the beginning of the cluster
4440          * for the logical block number, since when we allocate a
4441          * physical cluster, the physical block should start at the
4442          * same offset from the beginning of the cluster.  This is
4443          * needed so that future calls to get_implied_cluster_alloc()
4444          * work correctly.
4445          */
4446         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4447         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4448         ar.goal -= offset;
4449         ar.logical -= offset;
4450         if (S_ISREG(inode->i_mode))
4451                 ar.flags = EXT4_MB_HINT_DATA;
4452         else
4453                 /* disable in-core preallocation for non-regular files */
4454                 ar.flags = 0;
4455         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4456                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4457         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4458         if (!newblock)
4459                 goto out2;
4460         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4461                   ar.goal, newblock, allocated);
4462         free_on_err = 1;
4463         allocated_clusters = ar.len;
4464         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4465         if (ar.len > allocated)
4466                 ar.len = allocated;
4467
4468 got_allocated_blocks:
4469         /* try to insert new extent into found leaf and return */
4470         ext4_ext_store_pblock(&newex, newblock + offset);
4471         newex.ee_len = cpu_to_le16(ar.len);
4472         /* Mark unwritten */
4473         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4474                 ext4_ext_mark_unwritten(&newex);
4475                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4476                 /*
4477                  * io_end structure was created for every IO write to an
4478                  * unwritten extent. To avoid unnecessary conversion,
4479                  * here we flag the IO that really needs the conversion.
4480                  * For non asycn direct IO case, flag the inode state
4481                  * that we need to perform conversion when IO is done.
4482                  */
4483                 if (flags & EXT4_GET_BLOCKS_PRE_IO)
4484                         set_unwritten = 1;
4485         }
4486
4487         err = 0;
4488         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4489                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4490                                          path, ar.len);
4491         if (!err)
4492                 err = ext4_ext_insert_extent(handle, inode, path,
4493                                              &newex, flags);
4494
4495         if (!err && set_unwritten) {
4496                 if (io)
4497                         ext4_set_io_unwritten_flag(inode, io);
4498                 else
4499                         ext4_set_inode_state(inode,
4500                                              EXT4_STATE_DIO_UNWRITTEN);
4501         }
4502
4503         if (err && free_on_err) {
4504                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4505                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4506                 /* free data blocks we just allocated */
4507                 /* not a good idea to call discard here directly,
4508                  * but otherwise we'd need to call it every free() */
4509                 ext4_discard_preallocations(inode);
4510                 ext4_free_blocks(handle, inode, NULL, newblock,
4511                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4512                 goto out2;
4513         }
4514
4515         /* previous routine could use block we allocated */
4516         newblock = ext4_ext_pblock(&newex);
4517         allocated = ext4_ext_get_actual_len(&newex);
4518         if (allocated > map->m_len)
4519                 allocated = map->m_len;
4520         map->m_flags |= EXT4_MAP_NEW;
4521
4522         /*
4523          * Update reserved blocks/metadata blocks after successful
4524          * block allocation which had been deferred till now.
4525          */
4526         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4527                 unsigned int reserved_clusters;
4528                 /*
4529                  * Check how many clusters we had reserved this allocated range
4530                  */
4531                 reserved_clusters = get_reserved_cluster_alloc(inode,
4532                                                 map->m_lblk, allocated);
4533                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4534                         if (reserved_clusters) {
4535                                 /*
4536                                  * We have clusters reserved for this range.
4537                                  * But since we are not doing actual allocation
4538                                  * and are simply using blocks from previously
4539                                  * allocated cluster, we should release the
4540                                  * reservation and not claim quota.
4541                                  */
4542                                 ext4_da_update_reserve_space(inode,
4543                                                 reserved_clusters, 0);
4544                         }
4545                 } else {
4546                         BUG_ON(allocated_clusters < reserved_clusters);
4547                         if (reserved_clusters < allocated_clusters) {
4548                                 struct ext4_inode_info *ei = EXT4_I(inode);
4549                                 int reservation = allocated_clusters -
4550                                                   reserved_clusters;
4551                                 /*
4552                                  * It seems we claimed few clusters outside of
4553                                  * the range of this allocation. We should give
4554                                  * it back to the reservation pool. This can
4555                                  * happen in the following case:
4556                                  *
4557                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4558                                  *   cluster has 4 blocks. Thus, the clusters
4559                                  *   are [0-3],[4-7],[8-11]...
4560                                  * * First comes delayed allocation write for
4561                                  *   logical blocks 10 & 11. Since there were no
4562                                  *   previous delayed allocated blocks in the
4563                                  *   range [8-11], we would reserve 1 cluster
4564                                  *   for this write.
4565                                  * * Next comes write for logical blocks 3 to 8.
4566                                  *   In this case, we will reserve 2 clusters
4567                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4568                                  *   that range has a delayed allocated blocks.
4569                                  *   Thus total reserved clusters now becomes 3.
4570                                  * * Now, during the delayed allocation writeout
4571                                  *   time, we will first write blocks [3-8] and
4572                                  *   allocate 3 clusters for writing these
4573                                  *   blocks. Also, we would claim all these
4574                                  *   three clusters above.
4575                                  * * Now when we come here to writeout the
4576                                  *   blocks [10-11], we would expect to claim
4577                                  *   the reservation of 1 cluster we had made
4578                                  *   (and we would claim it since there are no
4579                                  *   more delayed allocated blocks in the range
4580                                  *   [8-11]. But our reserved cluster count had
4581                                  *   already gone to 0.
4582                                  *
4583                                  *   Thus, at the step 4 above when we determine
4584                                  *   that there are still some unwritten delayed
4585                                  *   allocated blocks outside of our current
4586                                  *   block range, we should increment the
4587                                  *   reserved clusters count so that when the
4588                                  *   remaining blocks finally gets written, we
4589                                  *   could claim them.
4590                                  */
4591                                 dquot_reserve_block(inode,
4592                                                 EXT4_C2B(sbi, reservation));
4593                                 spin_lock(&ei->i_block_reservation_lock);
4594                                 ei->i_reserved_data_blocks += reservation;
4595                                 spin_unlock(&ei->i_block_reservation_lock);
4596                         }
4597                         /*
4598                          * We will claim quota for all newly allocated blocks.
4599                          * We're updating the reserved space *after* the
4600                          * correction above so we do not accidentally free
4601                          * all the metadata reservation because we might
4602                          * actually need it later on.
4603                          */
4604                         ext4_da_update_reserve_space(inode, allocated_clusters,
4605                                                         1);
4606                 }
4607         }
4608
4609         /*
4610          * Cache the extent and update transaction to commit on fdatasync only
4611          * when it is _not_ an unwritten extent.
4612          */
4613         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4614                 ext4_update_inode_fsync_trans(handle, inode, 1);
4615         else
4616                 ext4_update_inode_fsync_trans(handle, inode, 0);
4617 out:
4618         if (allocated > map->m_len)
4619                 allocated = map->m_len;
4620         ext4_ext_show_leaf(inode, path);
4621         map->m_flags |= EXT4_MAP_MAPPED;
4622         map->m_pblk = newblock;
4623         map->m_len = allocated;
4624 out2:
4625         if (path) {
4626                 ext4_ext_drop_refs(path);
4627                 kfree(path);
4628         }
4629
4630         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4631                                        err ? err : allocated);
4632         ext4_es_lru_add(inode);
4633         return err ? err : allocated;
4634 }
4635
4636 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4637 {
4638         struct super_block *sb = inode->i_sb;
4639         ext4_lblk_t last_block;
4640         int err = 0;
4641
4642         /*
4643          * TODO: optimization is possible here.
4644          * Probably we need not scan at all,
4645          * because page truncation is enough.
4646          */
4647
4648         /* we have to know where to truncate from in crash case */
4649         EXT4_I(inode)->i_disksize = inode->i_size;
4650         ext4_mark_inode_dirty(handle, inode);
4651
4652         last_block = (inode->i_size + sb->s_blocksize - 1)
4653                         >> EXT4_BLOCK_SIZE_BITS(sb);
4654 retry:
4655         err = ext4_es_remove_extent(inode, last_block,
4656                                     EXT_MAX_BLOCKS - last_block);
4657         if (err == -ENOMEM) {
4658                 cond_resched();
4659                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4660                 goto retry;
4661         }
4662         if (err) {
4663                 ext4_std_error(inode->i_sb, err);
4664                 return;
4665         }
4666         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4667         ext4_std_error(inode->i_sb, err);
4668 }
4669
4670 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4671                                   ext4_lblk_t len, loff_t new_size,
4672                                   int flags, int mode)
4673 {
4674         struct inode *inode = file_inode(file);
4675         handle_t *handle;
4676         int ret = 0;
4677         int ret2 = 0;
4678         int retries = 0;
4679         struct ext4_map_blocks map;
4680         unsigned int credits;
4681         loff_t epos;
4682
4683         map.m_lblk = offset;
4684         map.m_len = len;
4685         /*
4686          * Don't normalize the request if it can fit in one extent so
4687          * that it doesn't get unnecessarily split into multiple
4688          * extents.
4689          */
4690         if (len <= EXT_UNWRITTEN_MAX_LEN)
4691                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4692
4693         /*
4694          * credits to insert 1 extent into extent tree
4695          */
4696         credits = ext4_chunk_trans_blocks(inode, len);
4697
4698 retry:
4699         while (ret >= 0 && len) {
4700                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4701                                             credits);
4702                 if (IS_ERR(handle)) {
4703                         ret = PTR_ERR(handle);
4704                         break;
4705                 }
4706                 ret = ext4_map_blocks(handle, inode, &map, flags);
4707                 if (ret <= 0) {
4708                         ext4_debug("inode #%lu: block %u: len %u: "
4709                                    "ext4_ext_map_blocks returned %d",
4710                                    inode->i_ino, map.m_lblk,
4711                                    map.m_len, ret);
4712                         ext4_mark_inode_dirty(handle, inode);
4713                         ret2 = ext4_journal_stop(handle);
4714                         break;
4715                 }
4716                 map.m_lblk += ret;
4717                 map.m_len = len = len - ret;
4718                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4719                 inode->i_ctime = ext4_current_time(inode);
4720                 if (new_size) {
4721                         if (epos > new_size)
4722                                 epos = new_size;
4723                         if (ext4_update_inode_size(inode, epos) & 0x1)
4724                                 inode->i_mtime = inode->i_ctime;
4725                 } else {
4726                         if (epos > inode->i_size)
4727                                 ext4_set_inode_flag(inode,
4728                                                     EXT4_INODE_EOFBLOCKS);
4729                 }
4730                 ext4_mark_inode_dirty(handle, inode);
4731                 ret2 = ext4_journal_stop(handle);
4732                 if (ret2)
4733                         break;
4734         }
4735         if (ret == -ENOSPC &&
4736                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4737                 ret = 0;
4738                 goto retry;
4739         }
4740
4741         return ret > 0 ? ret2 : ret;
4742 }
4743
4744 static long ext4_zero_range(struct file *file, loff_t offset,
4745                             loff_t len, int mode)
4746 {
4747         struct inode *inode = file_inode(file);
4748         handle_t *handle = NULL;
4749         unsigned int max_blocks;
4750         loff_t new_size = 0;
4751         int ret = 0;
4752         int flags;
4753         int credits;
4754         int partial_begin, partial_end;
4755         loff_t start, end;
4756         ext4_lblk_t lblk;
4757         struct address_space *mapping = inode->i_mapping;
4758         unsigned int blkbits = inode->i_blkbits;
4759
4760         trace_ext4_zero_range(inode, offset, len, mode);
4761
4762         if (!S_ISREG(inode->i_mode))
4763                 return -EINVAL;
4764
4765         /* Call ext4_force_commit to flush all data in case of data=journal. */
4766         if (ext4_should_journal_data(inode)) {
4767                 ret = ext4_force_commit(inode->i_sb);
4768                 if (ret)
4769                         return ret;
4770         }
4771
4772         /*
4773          * Write out all dirty pages to avoid race conditions
4774          * Then release them.
4775          */
4776         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4777                 ret = filemap_write_and_wait_range(mapping, offset,
4778                                                    offset + len - 1);
4779                 if (ret)
4780                         return ret;
4781         }
4782
4783         /*
4784          * Round up offset. This is not fallocate, we neet to zero out
4785          * blocks, so convert interior block aligned part of the range to
4786          * unwritten and possibly manually zero out unaligned parts of the
4787          * range.
4788          */
4789         start = round_up(offset, 1 << blkbits);
4790         end = round_down((offset + len), 1 << blkbits);
4791
4792         if (start < offset || end > offset + len)
4793                 return -EINVAL;
4794         partial_begin = offset & ((1 << blkbits) - 1);
4795         partial_end = (offset + len) & ((1 << blkbits) - 1);
4796
4797         lblk = start >> blkbits;
4798         max_blocks = (end >> blkbits);
4799         if (max_blocks < lblk)
4800                 max_blocks = 0;
4801         else
4802                 max_blocks -= lblk;
4803
4804         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT |
4805                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4806                 EXT4_EX_NOCACHE;
4807         if (mode & FALLOC_FL_KEEP_SIZE)
4808                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4809
4810         mutex_lock(&inode->i_mutex);
4811
4812         /*
4813          * Indirect files do not support unwritten extnets
4814          */
4815         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4816                 ret = -EOPNOTSUPP;
4817                 goto out_mutex;
4818         }
4819
4820         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4821              offset + len > i_size_read(inode)) {
4822                 new_size = offset + len;
4823                 ret = inode_newsize_ok(inode, new_size);
4824                 if (ret)
4825                         goto out_mutex;
4826                 /*
4827                  * If we have a partial block after EOF we have to allocate
4828                  * the entire block.
4829                  */
4830                 if (partial_end)
4831                         max_blocks += 1;
4832         }
4833
4834         if (max_blocks > 0) {
4835
4836                 /* Now release the pages and zero block aligned part of pages*/
4837                 truncate_pagecache_range(inode, start, end - 1);
4838                 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4839
4840                 /* Wait all existing dio workers, newcomers will block on i_mutex */
4841                 ext4_inode_block_unlocked_dio(inode);
4842                 inode_dio_wait(inode);
4843
4844                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4845                                              flags, mode);
4846                 if (ret)
4847                         goto out_dio;
4848                 /*
4849                  * Remove entire range from the extent status tree.
4850                  *
4851                  * ext4_es_remove_extent(inode, lblk, max_blocks) is
4852                  * NOT sufficient.  I'm not sure why this is the case,
4853                  * but let's be conservative and remove the extent
4854                  * status tree for the entire inode.  There should be
4855                  * no outstanding delalloc extents thanks to the
4856                  * filemap_write_and_wait_range() call above.
4857                  */
4858                 ret = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
4859                 if (ret)
4860                         goto out_dio;
4861         }
4862         if (!partial_begin && !partial_end)
4863                 goto out_dio;
4864
4865         /*
4866          * In worst case we have to writeout two nonadjacent unwritten
4867          * blocks and update the inode
4868          */
4869         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4870         if (ext4_should_journal_data(inode))
4871                 credits += 2;
4872         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4873         if (IS_ERR(handle)) {
4874                 ret = PTR_ERR(handle);
4875                 ext4_std_error(inode->i_sb, ret);
4876                 goto out_dio;
4877         }
4878
4879         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4880         if (new_size) {
4881                 ext4_update_inode_size(inode, new_size);
4882         } else {
4883                 /*
4884                 * Mark that we allocate beyond EOF so the subsequent truncate
4885                 * can proceed even if the new size is the same as i_size.
4886                 */
4887                 if ((offset + len) > i_size_read(inode))
4888                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4889         }
4890         ext4_mark_inode_dirty(handle, inode);
4891
4892         /* Zero out partial block at the edges of the range */
4893         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4894
4895         if (file->f_flags & O_SYNC)
4896                 ext4_handle_sync(handle);
4897
4898         ext4_journal_stop(handle);
4899 out_dio:
4900         ext4_inode_resume_unlocked_dio(inode);
4901 out_mutex:
4902         mutex_unlock(&inode->i_mutex);
4903         return ret;
4904 }
4905
4906 /*
4907  * preallocate space for a file. This implements ext4's fallocate file
4908  * operation, which gets called from sys_fallocate system call.
4909  * For block-mapped files, posix_fallocate should fall back to the method
4910  * of writing zeroes to the required new blocks (the same behavior which is
4911  * expected for file systems which do not support fallocate() system call).
4912  */
4913 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4914 {
4915         struct inode *inode = file_inode(file);
4916         loff_t new_size = 0;
4917         unsigned int max_blocks;
4918         int ret = 0;
4919         int flags;
4920         ext4_lblk_t lblk;
4921         unsigned int blkbits = inode->i_blkbits;
4922
4923         /* Return error if mode is not supported */
4924         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4925                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
4926                 return -EOPNOTSUPP;
4927
4928         if (mode & FALLOC_FL_PUNCH_HOLE)
4929                 return ext4_punch_hole(inode, offset, len);
4930
4931         ret = ext4_convert_inline_data(inode);
4932         if (ret)
4933                 return ret;
4934
4935         /*
4936          * currently supporting (pre)allocate mode for extent-based
4937          * files _only_
4938          */
4939         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4940                 return -EOPNOTSUPP;
4941
4942         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4943                 return ext4_collapse_range(inode, offset, len);
4944
4945         if (mode & FALLOC_FL_ZERO_RANGE)
4946                 return ext4_zero_range(file, offset, len, mode);
4947
4948         trace_ext4_fallocate_enter(inode, offset, len, mode);
4949         lblk = offset >> blkbits;
4950         /*
4951          * We can't just convert len to max_blocks because
4952          * If blocksize = 4096 offset = 3072 and len = 2048
4953          */
4954         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4955                 - lblk;
4956
4957         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4958         if (mode & FALLOC_FL_KEEP_SIZE)
4959                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4960
4961         mutex_lock(&inode->i_mutex);
4962
4963         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4964              offset + len > i_size_read(inode)) {
4965                 new_size = offset + len;
4966                 ret = inode_newsize_ok(inode, new_size);
4967                 if (ret)
4968                         goto out;
4969         }
4970
4971         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4972                                      flags, mode);
4973         if (ret)
4974                 goto out;
4975
4976         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4977                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4978                                                 EXT4_I(inode)->i_sync_tid);
4979         }
4980 out:
4981         mutex_unlock(&inode->i_mutex);
4982         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4983         return ret;
4984 }
4985
4986 /*
4987  * This function convert a range of blocks to written extents
4988  * The caller of this function will pass the start offset and the size.
4989  * all unwritten extents within this range will be converted to
4990  * written extents.
4991  *
4992  * This function is called from the direct IO end io call back
4993  * function, to convert the fallocated extents after IO is completed.
4994  * Returns 0 on success.
4995  */
4996 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4997                                    loff_t offset, ssize_t len)
4998 {
4999         unsigned int max_blocks;
5000         int ret = 0;
5001         int ret2 = 0;
5002         struct ext4_map_blocks map;
5003         unsigned int credits, blkbits = inode->i_blkbits;
5004
5005         map.m_lblk = offset >> blkbits;
5006         /*
5007          * We can't just convert len to max_blocks because
5008          * If blocksize = 4096 offset = 3072 and len = 2048
5009          */
5010         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
5011                       map.m_lblk);
5012         /*
5013          * This is somewhat ugly but the idea is clear: When transaction is
5014          * reserved, everything goes into it. Otherwise we rather start several
5015          * smaller transactions for conversion of each extent separately.
5016          */
5017         if (handle) {
5018                 handle = ext4_journal_start_reserved(handle,
5019                                                      EXT4_HT_EXT_CONVERT);
5020                 if (IS_ERR(handle))
5021                         return PTR_ERR(handle);
5022                 credits = 0;
5023         } else {
5024                 /*
5025                  * credits to insert 1 extent into extent tree
5026                  */
5027                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5028         }
5029         while (ret >= 0 && ret < max_blocks) {
5030                 map.m_lblk += ret;
5031                 map.m_len = (max_blocks -= ret);
5032                 if (credits) {
5033                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5034                                                     credits);
5035                         if (IS_ERR(handle)) {
5036                                 ret = PTR_ERR(handle);
5037                                 break;
5038                         }
5039                 }
5040                 ret = ext4_map_blocks(handle, inode, &map,
5041                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5042                 if (ret <= 0)
5043                         ext4_warning(inode->i_sb,
5044                                      "inode #%lu: block %u: len %u: "
5045                                      "ext4_ext_map_blocks returned %d",
5046                                      inode->i_ino, map.m_lblk,
5047                                      map.m_len, ret);
5048                 ext4_mark_inode_dirty(handle, inode);
5049                 if (credits)
5050                         ret2 = ext4_journal_stop(handle);
5051                 if (ret <= 0 || ret2)
5052                         break;
5053         }
5054         if (!credits)
5055                 ret2 = ext4_journal_stop(handle);
5056         return ret > 0 ? ret2 : ret;
5057 }
5058
5059 /*
5060  * If newes is not existing extent (newes->ec_pblk equals zero) find
5061  * delayed extent at start of newes and update newes accordingly and
5062  * return start of the next delayed extent.
5063  *
5064  * If newes is existing extent (newes->ec_pblk is not equal zero)
5065  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5066  * extent found. Leave newes unmodified.
5067  */
5068 static int ext4_find_delayed_extent(struct inode *inode,
5069                                     struct extent_status *newes)
5070 {
5071         struct extent_status es;
5072         ext4_lblk_t block, next_del;
5073
5074         if (newes->es_pblk == 0) {
5075                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5076                                 newes->es_lblk + newes->es_len - 1, &es);
5077
5078                 /*
5079                  * No extent in extent-tree contains block @newes->es_pblk,
5080                  * then the block may stay in 1)a hole or 2)delayed-extent.
5081                  */
5082                 if (es.es_len == 0)
5083                         /* A hole found. */
5084                         return 0;
5085
5086                 if (es.es_lblk > newes->es_lblk) {
5087                         /* A hole found. */
5088                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5089                                             newes->es_len);
5090                         return 0;
5091                 }
5092
5093                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5094         }
5095
5096         block = newes->es_lblk + newes->es_len;
5097         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5098         if (es.es_len == 0)
5099                 next_del = EXT_MAX_BLOCKS;
5100         else
5101                 next_del = es.es_lblk;
5102
5103         return next_del;
5104 }
5105 /* fiemap flags we can handle specified here */
5106 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5107
5108 static int ext4_xattr_fiemap(struct inode *inode,
5109                                 struct fiemap_extent_info *fieinfo)
5110 {
5111         __u64 physical = 0;
5112         __u64 length;
5113         __u32 flags = FIEMAP_EXTENT_LAST;
5114         int blockbits = inode->i_sb->s_blocksize_bits;
5115         int error = 0;
5116
5117         /* in-inode? */
5118         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5119                 struct ext4_iloc iloc;
5120                 int offset;     /* offset of xattr in inode */
5121
5122                 error = ext4_get_inode_loc(inode, &iloc);
5123                 if (error)
5124                         return error;
5125                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5126                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5127                                 EXT4_I(inode)->i_extra_isize;
5128                 physical += offset;
5129                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5130                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5131                 brelse(iloc.bh);
5132         } else { /* external block */
5133                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5134                 length = inode->i_sb->s_blocksize;
5135         }
5136
5137         if (physical)
5138                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5139                                                 length, flags);
5140         return (error < 0 ? error : 0);
5141 }
5142
5143 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5144                 __u64 start, __u64 len)
5145 {
5146         ext4_lblk_t start_blk;
5147         int error = 0;
5148
5149         if (ext4_has_inline_data(inode)) {
5150                 int has_inline = 1;
5151
5152                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
5153
5154                 if (has_inline)
5155                         return error;
5156         }
5157
5158         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5159                 error = ext4_ext_precache(inode);
5160                 if (error)
5161                         return error;
5162         }
5163
5164         /* fallback to generic here if not in extents fmt */
5165         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5166                 return generic_block_fiemap(inode, fieinfo, start, len,
5167                         ext4_get_block);
5168
5169         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5170                 return -EBADR;
5171
5172         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5173                 error = ext4_xattr_fiemap(inode, fieinfo);
5174         } else {
5175                 ext4_lblk_t len_blks;
5176                 __u64 last_blk;
5177
5178                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5179                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5180                 if (last_blk >= EXT_MAX_BLOCKS)
5181                         last_blk = EXT_MAX_BLOCKS-1;
5182                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5183
5184                 /*
5185                  * Walk the extent tree gathering extent information
5186                  * and pushing extents back to the user.
5187                  */
5188                 error = ext4_fill_fiemap_extents(inode, start_blk,
5189                                                  len_blks, fieinfo);
5190         }
5191         ext4_es_lru_add(inode);
5192         return error;
5193 }
5194
5195 /*
5196  * ext4_access_path:
5197  * Function to access the path buffer for marking it dirty.
5198  * It also checks if there are sufficient credits left in the journal handle
5199  * to update path.
5200  */
5201 static int
5202 ext4_access_path(handle_t *handle, struct inode *inode,
5203                 struct ext4_ext_path *path)
5204 {
5205         int credits, err;
5206
5207         if (!ext4_handle_valid(handle))
5208                 return 0;
5209
5210         /*
5211          * Check if need to extend journal credits
5212          * 3 for leaf, sb, and inode plus 2 (bmap and group
5213          * descriptor) for each block group; assume two block
5214          * groups
5215          */
5216         if (handle->h_buffer_credits < 7) {
5217                 credits = ext4_writepage_trans_blocks(inode);
5218                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5219                 /* EAGAIN is success */
5220                 if (err && err != -EAGAIN)
5221                         return err;
5222         }
5223
5224         err = ext4_ext_get_access(handle, inode, path);
5225         return err;
5226 }
5227
5228 /*
5229  * ext4_ext_shift_path_extents:
5230  * Shift the extents of a path structure lying between path[depth].p_ext
5231  * and EXT_LAST_EXTENT(path[depth].p_hdr) downwards, by subtracting shift
5232  * from starting block for each extent.
5233  */
5234 static int
5235 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5236                             struct inode *inode, handle_t *handle,
5237                             ext4_lblk_t *start)
5238 {
5239         int depth, err = 0;
5240         struct ext4_extent *ex_start, *ex_last;
5241         bool update = 0;
5242         depth = path->p_depth;
5243
5244         while (depth >= 0) {
5245                 if (depth == path->p_depth) {
5246                         ex_start = path[depth].p_ext;
5247                         if (!ex_start)
5248                                 return -EIO;
5249
5250                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5251                         if (!ex_last)
5252                                 return -EIO;
5253
5254                         err = ext4_access_path(handle, inode, path + depth);
5255                         if (err)
5256                                 goto out;
5257
5258                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5259                                 update = 1;
5260
5261                         *start = le32_to_cpu(ex_last->ee_block) +
5262                                 ext4_ext_get_actual_len(ex_last);
5263
5264                         while (ex_start <= ex_last) {
5265                                 le32_add_cpu(&ex_start->ee_block, -shift);
5266                                 /* Try to merge to the left. */
5267                                 if ((ex_start >
5268                                      EXT_FIRST_EXTENT(path[depth].p_hdr)) &&
5269                                     ext4_ext_try_to_merge_right(inode,
5270                                                         path, ex_start - 1))
5271                                         ex_last--;
5272                                 else
5273                                         ex_start++;
5274                         }
5275                         err = ext4_ext_dirty(handle, inode, path + depth);
5276                         if (err)
5277                                 goto out;
5278
5279                         if (--depth < 0 || !update)
5280                                 break;
5281                 }
5282
5283                 /* Update index too */
5284                 err = ext4_access_path(handle, inode, path + depth);
5285                 if (err)
5286                         goto out;
5287
5288                 le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5289                 err = ext4_ext_dirty(handle, inode, path + depth);
5290                 if (err)
5291                         goto out;
5292
5293                 /* we are done if current index is not a starting index */
5294                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5295                         break;
5296
5297                 depth--;
5298         }
5299
5300 out:
5301         return err;
5302 }
5303
5304 /*
5305  * ext4_ext_shift_extents:
5306  * All the extents which lies in the range from start to the last allocated
5307  * block for the file are shifted downwards by shift blocks.
5308  * On success, 0 is returned, error otherwise.
5309  */
5310 static int
5311 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5312                        ext4_lblk_t start, ext4_lblk_t shift)
5313 {
5314         struct ext4_ext_path *path;
5315         int ret = 0, depth;
5316         struct ext4_extent *extent;
5317         ext4_lblk_t stop_block;
5318         ext4_lblk_t ex_start, ex_end;
5319
5320         /* Let path point to the last extent */
5321         path = ext4_ext_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 0);
5322         if (IS_ERR(path))
5323                 return PTR_ERR(path);
5324
5325         depth = path->p_depth;
5326         extent = path[depth].p_ext;
5327         if (!extent) {
5328                 ext4_ext_drop_refs(path);
5329                 kfree(path);
5330                 return ret;
5331         }
5332
5333         stop_block = le32_to_cpu(extent->ee_block) +
5334                         ext4_ext_get_actual_len(extent);
5335         ext4_ext_drop_refs(path);
5336         kfree(path);
5337
5338         /* Nothing to shift, if hole is at the end of file */
5339         if (start >= stop_block)
5340                 return ret;
5341
5342         /*
5343          * Don't start shifting extents until we make sure the hole is big
5344          * enough to accomodate the shift.
5345          */
5346         path = ext4_ext_find_extent(inode, start - 1, NULL, 0);
5347         if (IS_ERR(path))
5348                 return PTR_ERR(path);
5349         depth = path->p_depth;
5350         extent =  path[depth].p_ext;
5351         if (extent) {
5352                 ex_start = le32_to_cpu(extent->ee_block);
5353                 ex_end = le32_to_cpu(extent->ee_block) +
5354                         ext4_ext_get_actual_len(extent);
5355         } else {
5356                 ex_start = 0;
5357                 ex_end = 0;
5358         }
5359         ext4_ext_drop_refs(path);
5360         kfree(path);
5361
5362         if ((start == ex_start && shift > ex_start) ||
5363             (shift > start - ex_end))
5364                 return -EINVAL;
5365
5366         /* Its safe to start updating extents */
5367         while (start < stop_block) {
5368                 path = ext4_ext_find_extent(inode, start, NULL, 0);
5369                 if (IS_ERR(path))
5370                         return PTR_ERR(path);
5371                 depth = path->p_depth;
5372                 extent = path[depth].p_ext;
5373                 if (!extent) {
5374                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5375                                          (unsigned long) start);
5376                         return -EIO;
5377                 }
5378                 if (start > le32_to_cpu(extent->ee_block)) {
5379                         /* Hole, move to the next extent */
5380                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5381                                 path[depth].p_ext++;
5382                         } else {
5383                                 start = ext4_ext_next_allocated_block(path);
5384                                 ext4_ext_drop_refs(path);
5385                                 kfree(path);
5386                                 continue;
5387                         }
5388                 }
5389                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5390                                 handle, &start);
5391                 ext4_ext_drop_refs(path);
5392                 kfree(path);
5393                 if (ret)
5394                         break;
5395         }
5396
5397         return ret;
5398 }
5399
5400 /*
5401  * ext4_collapse_range:
5402  * This implements the fallocate's collapse range functionality for ext4
5403  * Returns: 0 and non-zero on error.
5404  */
5405 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5406 {
5407         struct super_block *sb = inode->i_sb;
5408         ext4_lblk_t punch_start, punch_stop;
5409         handle_t *handle;
5410         unsigned int credits;
5411         loff_t new_size, ioffset;
5412         int ret;
5413
5414         /* Collapse range works only on fs block size aligned offsets. */
5415         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5416             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5417                 return -EINVAL;
5418
5419         if (!S_ISREG(inode->i_mode))
5420                 return -EINVAL;
5421
5422         trace_ext4_collapse_range(inode, offset, len);
5423
5424         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5425         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5426
5427         /* Call ext4_force_commit to flush all data in case of data=journal. */
5428         if (ext4_should_journal_data(inode)) {
5429                 ret = ext4_force_commit(inode->i_sb);
5430                 if (ret)
5431                         return ret;
5432         }
5433
5434         /*
5435          * Need to round down offset to be aligned with page size boundary
5436          * for page size > block size.
5437          */
5438         ioffset = round_down(offset, PAGE_SIZE);
5439
5440         /* Write out all dirty pages */
5441         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5442                                            LLONG_MAX);
5443         if (ret)
5444                 return ret;
5445
5446         /* Take mutex lock */
5447         mutex_lock(&inode->i_mutex);
5448
5449         /*
5450          * There is no need to overlap collapse range with EOF, in which case
5451          * it is effectively a truncate operation
5452          */
5453         if (offset + len >= i_size_read(inode)) {
5454                 ret = -EINVAL;
5455                 goto out_mutex;
5456         }
5457
5458         /* Currently just for extent based files */
5459         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5460                 ret = -EOPNOTSUPP;
5461                 goto out_mutex;
5462         }
5463
5464         truncate_pagecache(inode, ioffset);
5465
5466         /* Wait for existing dio to complete */
5467         ext4_inode_block_unlocked_dio(inode);
5468         inode_dio_wait(inode);
5469
5470         credits = ext4_writepage_trans_blocks(inode);
5471         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5472         if (IS_ERR(handle)) {
5473                 ret = PTR_ERR(handle);
5474                 goto out_dio;
5475         }
5476
5477         down_write(&EXT4_I(inode)->i_data_sem);
5478         ext4_discard_preallocations(inode);
5479
5480         ret = ext4_es_remove_extent(inode, punch_start,
5481                                     EXT_MAX_BLOCKS - punch_start);
5482         if (ret) {
5483                 up_write(&EXT4_I(inode)->i_data_sem);
5484                 goto out_stop;
5485         }
5486
5487         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5488         if (ret) {
5489                 up_write(&EXT4_I(inode)->i_data_sem);
5490                 goto out_stop;
5491         }
5492         ext4_discard_preallocations(inode);
5493
5494         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5495                                      punch_stop - punch_start);
5496         if (ret) {
5497                 up_write(&EXT4_I(inode)->i_data_sem);
5498                 goto out_stop;
5499         }
5500
5501         new_size = i_size_read(inode) - len;
5502         i_size_write(inode, new_size);
5503         EXT4_I(inode)->i_disksize = new_size;
5504
5505         up_write(&EXT4_I(inode)->i_data_sem);
5506         if (IS_SYNC(inode))
5507                 ext4_handle_sync(handle);
5508         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5509         ext4_mark_inode_dirty(handle, inode);
5510
5511 out_stop:
5512         ext4_journal_stop(handle);
5513 out_dio:
5514         ext4_inode_resume_unlocked_dio(inode);
5515 out_mutex:
5516         mutex_unlock(&inode->i_mutex);
5517         return ret;
5518 }
5519
5520 /**
5521  * ext4_swap_extents - Swap extents between two inodes
5522  *
5523  * @inode1:     First inode
5524  * @inode2:     Second inode
5525  * @lblk1:      Start block for first inode
5526  * @lblk2:      Start block for second inode
5527  * @count:      Number of blocks to swap
5528  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5529  * @erp:        Pointer to save error value
5530  *
5531  * This helper routine does exactly what is promise "swap extents". All other
5532  * stuff such as page-cache locking consistency, bh mapping consistency or
5533  * extent's data copying must be performed by caller.
5534  * Locking:
5535  *              i_mutex is held for both inodes
5536  *              i_data_sem is locked for write for both inodes
5537  * Assumptions:
5538  *              All pages from requested range are locked for both inodes
5539  */
5540 int
5541 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5542                      struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5543                   ext4_lblk_t count, int unwritten, int *erp)
5544 {
5545         struct ext4_ext_path *path1 = NULL;
5546         struct ext4_ext_path *path2 = NULL;
5547         int replaced_count = 0;
5548
5549         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5550         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5551         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5552         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5553
5554         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5555         if (unlikely(*erp))
5556                 return 0;
5557         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5558         if (unlikely(*erp))
5559                 return 0;
5560
5561         while (count) {
5562                 struct ext4_extent *ex1, *ex2, tmp_ex;
5563                 ext4_lblk_t e1_blk, e2_blk;
5564                 int e1_len, e2_len, len;
5565                 int split = 0;
5566
5567                 path1 = ext4_ext_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5568                 if (unlikely(IS_ERR(path1))) {
5569                         *erp = PTR_ERR(path1);
5570                         path1 = NULL;
5571                 finish:
5572                         count = 0;
5573                         goto repeat;
5574                 }
5575                 path2 = ext4_ext_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5576                 if (unlikely(IS_ERR(path2))) {
5577                         *erp = PTR_ERR(path2);
5578                         path2 = NULL;
5579                         goto finish;
5580                 }
5581                 ex1 = path1[path1->p_depth].p_ext;
5582                 ex2 = path2[path2->p_depth].p_ext;
5583                 /* Do we have somthing to swap ? */
5584                 if (unlikely(!ex2 || !ex1))
5585                         goto finish;
5586
5587                 e1_blk = le32_to_cpu(ex1->ee_block);
5588                 e2_blk = le32_to_cpu(ex2->ee_block);
5589                 e1_len = ext4_ext_get_actual_len(ex1);
5590                 e2_len = ext4_ext_get_actual_len(ex2);
5591
5592                 /* Hole handling */
5593                 if (!in_range(lblk1, e1_blk, e1_len) ||
5594                     !in_range(lblk2, e2_blk, e2_len)) {
5595                         ext4_lblk_t next1, next2;
5596
5597                         /* if hole after extent, then go to next extent */
5598                         next1 = ext4_ext_next_allocated_block(path1);
5599                         next2 = ext4_ext_next_allocated_block(path2);
5600                         /* If hole before extent, then shift to that extent */
5601                         if (e1_blk > lblk1)
5602                                 next1 = e1_blk;
5603                         if (e2_blk > lblk2)
5604                                 next2 = e1_blk;
5605                         /* Do we have something to swap */
5606                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5607                                 goto finish;
5608                         /* Move to the rightest boundary */
5609                         len = next1 - lblk1;
5610                         if (len < next2 - lblk2)
5611                                 len = next2 - lblk2;
5612                         if (len > count)
5613                                 len = count;
5614                         lblk1 += len;
5615                         lblk2 += len;
5616                         count -= len;
5617                         goto repeat;
5618                 }
5619
5620                 /* Prepare left boundary */
5621                 if (e1_blk < lblk1) {
5622                         split = 1;
5623                         *erp = ext4_force_split_extent_at(handle, inode1,
5624                                                 path1, lblk1, 0);
5625                         if (unlikely(*erp))
5626                                 goto finish;
5627                 }
5628                 if (e2_blk < lblk2) {
5629                         split = 1;
5630                         *erp = ext4_force_split_extent_at(handle, inode2,
5631                                                 path2,  lblk2, 0);
5632                         if (unlikely(*erp))
5633                                 goto finish;
5634                 }
5635                 /* ext4_split_extent_at() may retult in leaf extent split,
5636                  * path must to be revalidated. */
5637                 if (split)
5638                         goto repeat;
5639
5640                 /* Prepare right boundary */
5641                 len = count;
5642                 if (len > e1_blk + e1_len - lblk1)
5643                         len = e1_blk + e1_len - lblk1;
5644                 if (len > e2_blk + e2_len - lblk2)
5645                         len = e2_blk + e2_len - lblk2;
5646
5647                 if (len != e1_len) {
5648                         split = 1;
5649                         *erp = ext4_force_split_extent_at(handle, inode1,
5650                                                 path1, lblk1 + len, 0);
5651                         if (unlikely(*erp))
5652                                 goto finish;
5653                 }
5654                 if (len != e2_len) {
5655                         split = 1;
5656                         *erp = ext4_force_split_extent_at(handle, inode2,
5657                                                 path2, lblk2 + len, 0);
5658                         if (*erp)
5659                                 goto finish;
5660                 }
5661                 /* ext4_split_extent_at() may retult in leaf extent split,
5662                  * path must to be revalidated. */
5663                 if (split)
5664                         goto repeat;
5665
5666                 BUG_ON(e2_len != e1_len);
5667                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5668                 if (unlikely(*erp))
5669                         goto finish;
5670                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5671                 if (unlikely(*erp))
5672                         goto finish;
5673
5674                 /* Both extents are fully inside boundaries. Swap it now */
5675                 tmp_ex = *ex1;
5676                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5677                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5678                 ex1->ee_len = cpu_to_le16(e2_len);
5679                 ex2->ee_len = cpu_to_le16(e1_len);
5680                 if (unwritten)
5681                         ext4_ext_mark_unwritten(ex2);
5682                 if (ext4_ext_is_unwritten(&tmp_ex))
5683                         ext4_ext_mark_unwritten(ex1);
5684
5685                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5686                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5687                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5688                                       path2->p_depth);
5689                 if (unlikely(*erp))
5690                         goto finish;
5691                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5692                                       path1->p_depth);
5693                 /*
5694                  * Looks scarry ah..? second inode already points to new blocks,
5695                  * and it was successfully dirtied. But luckily error may happen
5696                  * only due to journal error, so full transaction will be
5697                  * aborted anyway.
5698                  */
5699                 if (unlikely(*erp))
5700                         goto finish;
5701                 lblk1 += len;
5702                 lblk2 += len;
5703                 replaced_count += len;
5704                 count -= len;
5705
5706         repeat:
5707                 if (path1) {
5708                         ext4_ext_drop_refs(path1);
5709                         kfree(path1);
5710                         path1 = NULL;
5711                 }
5712                 if (path2) {
5713                         ext4_ext_drop_refs(path2);
5714                         kfree(path2);
5715                         path2 = NULL;
5716                 }
5717         }
5718         return replaced_count;
5719 }