kernel/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_metadata_csum(inode->i_sb))
  77                 return 1;
  78
  79         et = find_ext4_extent_tail(eh);
  80         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  81                 return 0;
  82         return 1;
  83 }
  84
  85 static void ext4_extent_block_csum_set(struct inode *inode,
  86                                        struct ext4_extent_header *eh)
  87 {
  88         struct ext4_extent_tail *et;
  89
  90         if (!ext4_has_metadata_csum(inode->i_sb))
  91                 return;
  92
  93         et = find_ext4_extent_tail(eh);
  94         et->et_checksum = ext4_extent_block_csum(inode, eh);
  95 }
  96
  97 static int ext4_split_extent(handle_t *handle,
  98                                 struct inode *inode,
  99                                 struct ext4_ext_path **ppath,
 100                                 struct ext4_map_blocks *map,
 101                                 int split_flag,
 102                                 int flags);
 103
 104 static int ext4_split_extent_at(handle_t *handle,
 105                              struct inode *inode,
 106                              struct ext4_ext_path **ppath,
 107                              ext4_lblk_t split,
 108                              int split_flag,
 109                              int flags);
 110
 111 static int ext4_find_delayed_extent(struct inode *inode,
 112                                     struct extent_status *newes);
 113
 114 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 115                                             struct inode *inode,
 116                                             int needed)
 117 {
 118         int err;
 119
 120         if (!ext4_handle_valid(handle))
 121                 return 0;
 122         if (handle->h_buffer_credits > needed)
 123                 return 0;
 124         err = ext4_journal_extend(handle, needed);
 125         if (err <= 0)
 126                 return err;
 127         err = ext4_truncate_restart_trans(handle, inode, needed);
 128         if (err == 0)
 129                 err = -EAGAIN;
 130
 131         return err;
 132 }
 133
 134 /*
 135  * could return:
 136  *  - EROFS
 137  *  - ENOMEM
 138  */
 139 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 140                                 struct ext4_ext_path *path)
 141 {
 142         if (path->p_bh) {
 143                 /* path points to block */
 144                 BUFFER_TRACE(path->p_bh, "get_write_access");
 145                 return ext4_journal_get_write_access(handle, path->p_bh);
 146         }
 147         /* path points to leaf/index in inode body */
 148         /* we use in-core data, no need to protect them */
 149         return 0;
 150 }
 151
 152 /*
 153  * could return:
 154  *  - EROFS
 155  *  - ENOMEM
 156  *  - EIO
 157  */
 158 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 159                      struct inode *inode, struct ext4_ext_path *path)
 160 {
 161         int err;
 162
 163         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 164         if (path->p_bh) {
 165                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 166                 /* path points to block */
 167                 err = __ext4_handle_dirty_metadata(where, line, handle,
 168                                                    inode, path->p_bh);
 169         } else {
 170                 /* path points to leaf/index in inode body */
 171                 err = ext4_mark_inode_dirty(handle, inode);
 172         }
 173         return err;
 174 }
 175
 176 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 177                               struct ext4_ext_path *path,
 178                               ext4_lblk_t block)
 179 {
 180         if (path) {
 181                 int depth = path->p_depth;
 182                 struct ext4_extent *ex;
 183
 184                 /*
 185                  * Try to predict block placement assuming that we are
 186                  * filling in a file which will eventually be
 187                  * non-sparse --- i.e., in the case of libbfd writing
 188                  * an ELF object sections out-of-order but in a way
 189                  * the eventually results in a contiguous object or
 190                  * executable file, or some database extending a table
 191                  * space file.  However, this is actually somewhat
 192                  * non-ideal if we are writing a sparse file such as
 193                  * qemu or KVM writing a raw image file that is going
 194                  * to stay fairly sparse, since it will end up
 195                  * fragmenting the file system's free space.  Maybe we
 196                  * should have some hueristics or some way to allow
 197                  * userspace to pass a hint to file system,
 198                  * especially if the latter case turns out to be
 199                  * common.
 200                  */
 201                 ex = path[depth].p_ext;
 202                 if (ex) {
 203                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 204                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 205
 206                         if (block > ext_block)
 207                                 return ext_pblk + (block - ext_block);
 208                         else
 209                                 return ext_pblk - (ext_block - block);
 210                 }
 211
 212                 /* it looks like index is empty;
 213                  * try to find starting block from index itself */
 214                 if (path[depth].p_bh)
 215                         return path[depth].p_bh->b_blocknr;
 216         }
 217
 218         /* OK. use inode's group */
 219         return ext4_inode_to_goal_block(inode);
 220 }
 221
 222 /*
 223  * Allocation for a meta data block
 224  */
 225 static ext4_fsblk_t
 226 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 227                         struct ext4_ext_path *path,
 228                         struct ext4_extent *ex, int *err, unsigned int flags)
 229 {
 230         ext4_fsblk_t goal, newblock;
 231
 232         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 233         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 234                                         NULL, err);
 235         return newblock;
 236 }
 237
 238 static inline int ext4_ext_space_block(struct inode *inode, int check)
 239 {
 240         int size;
 241
 242         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 243                         / sizeof(struct ext4_extent);
 244 #ifdef AGGRESSIVE_TEST
 245         if (!check && size > 6)
 246                 size = 6;
 247 #endif
 248         return size;
 249 }
 250
 251 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 252 {
 253         int size;
 254
 255         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 256                         / sizeof(struct ext4_extent_idx);
 257 #ifdef AGGRESSIVE_TEST
 258         if (!check && size > 5)
 259                 size = 5;
 260 #endif
 261         return size;
 262 }
 263
 264 static inline int ext4_ext_space_root(struct inode *inode, int check)
 265 {
 266         int size;
 267
 268         size = sizeof(EXT4_I(inode)->i_data);
 269         size -= sizeof(struct ext4_extent_header);
 270         size /= sizeof(struct ext4_extent);
 271 #ifdef AGGRESSIVE_TEST
 272         if (!check && size > 3)
 273                 size = 3;
 274 #endif
 275         return size;
 276 }
 277
 278 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 279 {
 280         int size;
 281
 282         size = sizeof(EXT4_I(inode)->i_data);
 283         size -= sizeof(struct ext4_extent_header);
 284         size /= sizeof(struct ext4_extent_idx);
 285 #ifdef AGGRESSIVE_TEST
 286         if (!check && size > 4)
 287                 size = 4;
 288 #endif
 289         return size;
 290 }
 291
 292 static inline int
 293 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 294                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 295                            int nofail)
 296 {
 297         struct ext4_ext_path *path = *ppath;
 298         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 299
 300         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 301                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 302                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 303                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 304 }
 305
 306 /*
 307  * Calculate the number of metadata blocks needed
 308  * to allocate @blocks
 309  * Worse case is one block per extent
 310  */
 311 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 312 {
 313         struct ext4_inode_info *ei = EXT4_I(inode);
 314         int idxs;
 315
 316         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 317                 / sizeof(struct ext4_extent_idx));
 318
 319         /*
 320          * If the new delayed allocation block is contiguous with the
 321          * previous da block, it can share index blocks with the
 322          * previous block, so we only need to allocate a new index
 323          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 324          * an additional index block, and at ldxs**3 blocks, yet
 325          * another index blocks.
 326          */
 327         if (ei->i_da_metadata_calc_len &&
 328             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 329                 int num = 0;
 330
 331                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 332                         num++;
 333                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 334                         num++;
 335                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 336                         num++;
 337                         ei->i_da_metadata_calc_len = 0;
 338                 } else
 339                         ei->i_da_metadata_calc_len++;
 340                 ei->i_da_metadata_calc_last_lblock++;
 341                 return num;
 342         }
 343
 344         /*
 345          * In the worst case we need a new set of index blocks at
 346          * every level of the inode's extent tree.
 347          */
 348         ei->i_da_metadata_calc_len = 1;
 349         ei->i_da_metadata_calc_last_lblock = lblock;
 350         return ext_depth(inode) + 1;
 351 }
 352
 353 static int
 354 ext4_ext_max_entries(struct inode *inode, int depth)
 355 {
 356         int max;
 357
 358         if (depth == ext_depth(inode)) {
 359                 if (depth == 0)
 360                         max = ext4_ext_space_root(inode, 1);
 361                 else
 362                         max = ext4_ext_space_root_idx(inode, 1);
 363         } else {
 364                 if (depth == 0)
 365                         max = ext4_ext_space_block(inode, 1);
 366                 else
 367                         max = ext4_ext_space_block_idx(inode, 1);
 368         }
 369
 370         return max;
 371 }
 372
 373 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 374 {
 375         ext4_fsblk_t block = ext4_ext_pblock(ext);
 376         int len = ext4_ext_get_actual_len(ext);
 377         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 378         ext4_lblk_t last = lblock + len - 1;
 379
 380         if (len == 0 || lblock > last)
 381                 return 0;
 382         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 383 }
 384
 385 static int ext4_valid_extent_idx(struct inode *inode,
 386                                 struct ext4_extent_idx *ext_idx)
 387 {
 388         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 389
 390         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 391 }
 392
 393 static int ext4_valid_extent_entries(struct inode *inode,
 394                                 struct ext4_extent_header *eh,
 395                                 int depth)
 396 {
 397         unsigned short entries;
 398         if (eh->eh_entries == 0)
 399                 return 1;
 400
 401         entries = le16_to_cpu(eh->eh_entries);
 402
 403         if (depth == 0) {
 404                 /* leaf entries */
 405                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 406                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 407                 ext4_fsblk_t pblock = 0;
 408                 ext4_lblk_t lblock = 0;
 409                 ext4_lblk_t prev = 0;
 410                 int len = 0;
 411                 while (entries) {
 412                         if (!ext4_valid_extent(inode, ext))
 413                                 return 0;
 414
 415                         /* Check for overlapping extents */
 416                         lblock = le32_to_cpu(ext->ee_block);
 417                         len = ext4_ext_get_actual_len(ext);
 418                         if ((lblock <= prev) && prev) {
 419                                 pblock = ext4_ext_pblock(ext);
 420                                 es->s_last_error_block = cpu_to_le64(pblock);
 421                                 return 0;
 422                         }
 423                         ext++;
 424                         entries--;
 425                         prev = lblock + len - 1;
 426                 }
 427         } else {
 428                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 429                 while (entries) {
 430                         if (!ext4_valid_extent_idx(inode, ext_idx))
 431                                 return 0;
 432                         ext_idx++;
 433                         entries--;
 434                 }
 435         }
 436         return 1;
 437 }
 438
 439 static int __ext4_ext_check(const char *function, unsigned int line,
 440                             struct inode *inode, struct ext4_extent_header *eh,
 441                             int depth, ext4_fsblk_t pblk)
 442 {
 443         const char *error_msg;
 444         int max = 0;
 445
 446         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 447                 error_msg = "invalid magic";
 448                 goto corrupted;
 449         }
 450         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 451                 error_msg = "unexpected eh_depth";
 452                 goto corrupted;
 453         }
 454         if (unlikely(eh->eh_max == 0)) {
 455                 error_msg = "invalid eh_max";
 456                 goto corrupted;
 457         }
 458         max = ext4_ext_max_entries(inode, depth);
 459         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 460                 error_msg = "too large eh_max";
 461                 goto corrupted;
 462         }
 463         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 464                 error_msg = "invalid eh_entries";
 465                 goto corrupted;
 466         }
 467         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 468                 error_msg = "invalid extent entries";
 469                 goto corrupted;
 470         }
 471         /* Verify checksum on non-root extent tree nodes */
 472         if (ext_depth(inode) != depth &&
 473             !ext4_extent_block_csum_verify(inode, eh)) {
 474                 error_msg = "extent tree corrupted";
 475                 goto corrupted;
 476         }
 477         return 0;
 478
 479 corrupted:
 480         ext4_error_inode(inode, function, line, 0,
 481                          "pblk %llu bad header/extent: %s - magic %x, "
 482                          "entries %u, max %u(%u), depth %u(%u)",
 483                          (unsigned long long) pblk, error_msg,
 484                          le16_to_cpu(eh->eh_magic),
 485                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 486                          max, le16_to_cpu(eh->eh_depth), depth);
 487         return -EIO;
 488 }
 489
 490 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 491         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 492
 493 int ext4_ext_check_inode(struct inode *inode)
 494 {
 495         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 496 }
 497
 498 static struct buffer_head *
 499 __read_extent_tree_block(const char *function, unsigned int line,
 500                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 501                          int flags)
 502 {
 503         struct buffer_head              *bh;
 504         int                             err;
 505
 506         bh = sb_getblk(inode->i_sb, pblk);
 507         if (unlikely(!bh))
 508                 return ERR_PTR(-ENOMEM);
 509
 510         if (!bh_uptodate_or_lock(bh)) {
 511                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 512                 err = bh_submit_read(bh);
 513                 if (err < 0)
 514                         goto errout;
 515         }
 516         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 517                 return bh;
 518         err = __ext4_ext_check(function, line, inode,
 519                                ext_block_hdr(bh), depth, pblk);
 520         if (err)
 521                 goto errout;
 522         set_buffer_verified(bh);
 523         /*
 524          * If this is a leaf block, cache all of its entries
 525          */
 526         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 527                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 528                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 529                 ext4_lblk_t prev = 0;
 530                 int i;
 531
 532                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 533                         unsigned int status = EXTENT_STATUS_WRITTEN;
 534                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 535                         int len = ext4_ext_get_actual_len(ex);
 536
 537                         if (prev && (prev != lblk))
 538                                 ext4_es_cache_extent(inode, prev,
 539                                                      lblk - prev, ~0,
 540                                                      EXTENT_STATUS_HOLE);
 541
 542                         if (ext4_ext_is_unwritten(ex))
 543                                 status = EXTENT_STATUS_UNWRITTEN;
 544                         ext4_es_cache_extent(inode, lblk, len,
 545                                              ext4_ext_pblock(ex), status);
 546                         prev = lblk + len;
 547                 }
 548         }
 549         return bh;
 550 errout:
 551         put_bh(bh);
 552         return ERR_PTR(err);
 553
 554 }
 555
 556 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 557         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 558                                  (depth), (flags))
 559
 560 /*
 561  * This function is called to cache a file's extent information in the
 562  * extent status tree
 563  */
 564 int ext4_ext_precache(struct inode *inode)
 565 {
 566         struct ext4_inode_info *ei = EXT4_I(inode);
 567         struct ext4_ext_path *path = NULL;
 568         struct buffer_head *bh;
 569         int i = 0, depth, ret = 0;
 570
 571         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 572                 return 0;       /* not an extent-mapped inode */
 573
 574         down_read(&ei->i_data_sem);
 575         depth = ext_depth(inode);
 576
 577         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 578                        GFP_NOFS);
 579         if (path == NULL) {
 580                 up_read(&ei->i_data_sem);
 581                 return -ENOMEM;
 582         }
 583
 584         /* Don't cache anything if there are no external extent blocks */
 585         if (depth == 0)
 586                 goto out;
 587         path[0].p_hdr = ext_inode_hdr(inode);
 588         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 589         if (ret)
 590                 goto out;
 591         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 592         while (i >= 0) {
 593                 /*
 594                  * If this is a leaf block or we've reached the end of
 595                  * the index block, go up
 596                  */
 597                 if ((i == depth) ||
 598                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 599                         brelse(path[i].p_bh);
 600                         path[i].p_bh = NULL;
 601                         i--;
 602                         continue;
 603                 }
 604                 bh = read_extent_tree_block(inode,
 605                                             ext4_idx_pblock(path[i].p_idx++),
 606                                             depth - i - 1,
 607                                             EXT4_EX_FORCE_CACHE);
 608                 if (IS_ERR(bh)) {
 609                         ret = PTR_ERR(bh);
 610                         break;
 611                 }
 612                 i++;
 613                 path[i].p_bh = bh;
 614                 path[i].p_hdr = ext_block_hdr(bh);
 615                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 616         }
 617         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 618 out:
 619         up_read(&ei->i_data_sem);
 620         ext4_ext_drop_refs(path);
 621         kfree(path);
 622         return ret;
 623 }
 624
 625 #ifdef EXT_DEBUG
 626 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 627 {
 628         int k, l = path->p_depth;
 629
 630         ext_debug("path:");
 631         for (k = 0; k <= l; k++, path++) {
 632                 if (path->p_idx) {
 633                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 634                             ext4_idx_pblock(path->p_idx));
 635                 } else if (path->p_ext) {
 636                         ext_debug("  %d:[%d]%d:%llu ",
 637                                   le32_to_cpu(path->p_ext->ee_block),
 638                                   ext4_ext_is_unwritten(path->p_ext),
 639                                   ext4_ext_get_actual_len(path->p_ext),
 640                                   ext4_ext_pblock(path->p_ext));
 641                 } else
 642                         ext_debug("  []");
 643         }
 644         ext_debug("\n");
 645 }
 646
 647 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 648 {
 649         int depth = ext_depth(inode);
 650         struct ext4_extent_header *eh;
 651         struct ext4_extent *ex;
 652         int i;
 653
 654         if (!path)
 655                 return;
 656
 657         eh = path[depth].p_hdr;
 658         ex = EXT_FIRST_EXTENT(eh);
 659
 660         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 661
 662         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 663                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 664                           ext4_ext_is_unwritten(ex),
 665                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 666         }
 667         ext_debug("\n");
 668 }
 669
 670 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 671                         ext4_fsblk_t newblock, int level)
 672 {
 673         int depth = ext_depth(inode);
 674         struct ext4_extent *ex;
 675
 676         if (depth != level) {
 677                 struct ext4_extent_idx *idx;
 678                 idx = path[level].p_idx;
 679                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 680                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 681                                         le32_to_cpu(idx->ei_block),
 682                                         ext4_idx_pblock(idx),
 683                                         newblock);
 684                         idx++;
 685                 }
 686
 687                 return;
 688         }
 689
 690         ex = path[depth].p_ext;
 691         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 692                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 693                                 le32_to_cpu(ex->ee_block),
 694                                 ext4_ext_pblock(ex),
 695                                 ext4_ext_is_unwritten(ex),
 696                                 ext4_ext_get_actual_len(ex),
 697                                 newblock);
 698                 ex++;
 699         }
 700 }
 701
 702 #else
 703 #define ext4_ext_show_path(inode, path)
 704 #define ext4_ext_show_leaf(inode, path)
 705 #define ext4_ext_show_move(inode, path, newblock, level)
 706 #endif
 707
 708 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 709 {
 710         int depth, i;
 711
 712         if (!path)
 713                 return;
 714         depth = path->p_depth;
 715         for (i = 0; i <= depth; i++, path++)
 716                 if (path->p_bh) {
 717                         brelse(path->p_bh);
 718                         path->p_bh = NULL;
 719                 }
 720 }
 721
 722 /*
 723  * ext4_ext_binsearch_idx:
 724  * binary search for the closest index of the given block
 725  * the header must be checked before calling this
 726  */
 727 static void
 728 ext4_ext_binsearch_idx(struct inode *inode,
 729                         struct ext4_ext_path *path, ext4_lblk_t block)
 730 {
 731         struct ext4_extent_header *eh = path->p_hdr;
 732         struct ext4_extent_idx *r, *l, *m;
 733
 734
 735         ext_debug("binsearch for %u(idx):  ", block);
 736
 737         l = EXT_FIRST_INDEX(eh) + 1;
 738         r = EXT_LAST_INDEX(eh);
 739         while (l <= r) {
 740                 m = l + (r - l) / 2;
 741                 if (block < le32_to_cpu(m->ei_block))
 742                         r = m - 1;
 743                 else
 744                         l = m + 1;
 745                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 746                                 m, le32_to_cpu(m->ei_block),
 747                                 r, le32_to_cpu(r->ei_block));
 748         }
 749
 750         path->p_idx = l - 1;
 751         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 752                   ext4_idx_pblock(path->p_idx));
 753
 754 #ifdef CHECK_BINSEARCH
 755         {
 756                 struct ext4_extent_idx *chix, *ix;
 757                 int k;
 758
 759                 chix = ix = EXT_FIRST_INDEX(eh);
 760                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 761                   if (k != 0 &&
 762                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 763                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 764                                        "first=0x%p\n", k,
 765                                        ix, EXT_FIRST_INDEX(eh));
 766                                 printk(KERN_DEBUG "%u <= %u\n",
 767                                        le32_to_cpu(ix->ei_block),
 768                                        le32_to_cpu(ix[-1].ei_block));
 769                         }
 770                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 771                                            <= le32_to_cpu(ix[-1].ei_block));
 772                         if (block < le32_to_cpu(ix->ei_block))
 773                                 break;
 774                         chix = ix;
 775                 }
 776                 BUG_ON(chix != path->p_idx);
 777         }
 778 #endif
 779
 780 }
 781
 782 /*
 783  * ext4_ext_binsearch:
 784  * binary search for closest extent of the given block
 785  * the header must be checked before calling this
 786  */
 787 static void
 788 ext4_ext_binsearch(struct inode *inode,
 789                 struct ext4_ext_path *path, ext4_lblk_t block)
 790 {
 791         struct ext4_extent_header *eh = path->p_hdr;
 792         struct ext4_extent *r, *l, *m;
 793
 794         if (eh->eh_entries == 0) {
 795                 /*
 796                  * this leaf is empty:
 797                  * we get such a leaf in split/add case
 798                  */
 799                 return;
 800         }
 801
 802         ext_debug("binsearch for %u:  ", block);
 803
 804         l = EXT_FIRST_EXTENT(eh) + 1;
 805         r = EXT_LAST_EXTENT(eh);
 806
 807         while (l <= r) {
 808                 m = l + (r - l) / 2;
 809                 if (block < le32_to_cpu(m->ee_block))
 810                         r = m - 1;
 811                 else
 812                         l = m + 1;
 813                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 814                                 m, le32_to_cpu(m->ee_block),
 815                                 r, le32_to_cpu(r->ee_block));
 816         }
 817
 818         path->p_ext = l - 1;
 819         ext_debug("  -> %d:%llu:[%d]%d ",
 820                         le32_to_cpu(path->p_ext->ee_block),
 821                         ext4_ext_pblock(path->p_ext),
 822                         ext4_ext_is_unwritten(path->p_ext),
 823                         ext4_ext_get_actual_len(path->p_ext));
 824
 825 #ifdef CHECK_BINSEARCH
 826         {
 827                 struct ext4_extent *chex, *ex;
 828                 int k;
 829
 830                 chex = ex = EXT_FIRST_EXTENT(eh);
 831                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 832                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 833                                           <= le32_to_cpu(ex[-1].ee_block));
 834                         if (block < le32_to_cpu(ex->ee_block))
 835                                 break;
 836                         chex = ex;
 837                 }
 838                 BUG_ON(chex != path->p_ext);
 839         }
 840 #endif
 841
 842 }
 843
 844 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 845 {
 846         struct ext4_extent_header *eh;
 847
 848         eh = ext_inode_hdr(inode);
 849         eh->eh_depth = 0;
 850         eh->eh_entries = 0;
 851         eh->eh_magic = EXT4_EXT_MAGIC;
 852         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 853         ext4_mark_inode_dirty(handle, inode);
 854         return 0;
 855 }
 856
 857 struct ext4_ext_path *
 858 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 859                  struct ext4_ext_path **orig_path, int flags)
 860 {
 861         struct ext4_extent_header *eh;
 862         struct buffer_head *bh;
 863         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 864         short int depth, i, ppos = 0;
 865         int ret;
 866
 867         eh = ext_inode_hdr(inode);
 868         depth = ext_depth(inode);
 869
 870         if (path) {
 871                 ext4_ext_drop_refs(path);
 872                 if (depth > path[0].p_maxdepth) {
 873                         kfree(path);
 874                         *orig_path = path = NULL;
 875                 }
 876         }
 877         if (!path) {
 878                 /* account possible depth increase */
 879                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 880                                 GFP_NOFS);
 881                 if (unlikely(!path))
 882                         return ERR_PTR(-ENOMEM);
 883                 path[0].p_maxdepth = depth + 1;
 884         }
 885         path[0].p_hdr = eh;
 886         path[0].p_bh = NULL;
 887
 888         i = depth;
 889         /* walk through the tree */
 890         while (i) {
 891                 ext_debug("depth %d: num %d, max %d\n",
 892                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 893
 894                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 895                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 896                 path[ppos].p_depth = i;
 897                 path[ppos].p_ext = NULL;
 898
 899                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 900                                             flags);
 901                 if (unlikely(IS_ERR(bh))) {
 902                         ret = PTR_ERR(bh);
 903                         goto err;
 904                 }
 905
 906                 eh = ext_block_hdr(bh);
 907                 ppos++;
 908                 if (unlikely(ppos > depth)) {
 909                         put_bh(bh);
 910                         EXT4_ERROR_INODE(inode,
 911                                          "ppos %d > depth %d", ppos, depth);
 912                         ret = -EIO;
 913                         goto err;
 914                 }
 915                 path[ppos].p_bh = bh;
 916                 path[ppos].p_hdr = eh;
 917         }
 918
 919         path[ppos].p_depth = i;
 920         path[ppos].p_ext = NULL;
 921         path[ppos].p_idx = NULL;
 922
 923         /* find extent */
 924         ext4_ext_binsearch(inode, path + ppos, block);
 925         /* if not an empty leaf */
 926         if (path[ppos].p_ext)
 927                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 928
 929         ext4_ext_show_path(inode, path);
 930
 931         return path;
 932
 933 err:
 934         ext4_ext_drop_refs(path);
 935         kfree(path);
 936         if (orig_path)
 937                 *orig_path = NULL;
 938         return ERR_PTR(ret);
 939 }
 940
 941 /*
 942  * ext4_ext_insert_index:
 943  * insert new index [@logical;@ptr] into the block at @curp;
 944  * check where to insert: before @curp or after @curp
 945  */
 946 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 947                                  struct ext4_ext_path *curp,
 948                                  int logical, ext4_fsblk_t ptr)
 949 {
 950         struct ext4_extent_idx *ix;
 951         int len, err;
 952
 953         err = ext4_ext_get_access(handle, inode, curp);
 954         if (err)
 955                 return err;
 956
 957         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 958                 EXT4_ERROR_INODE(inode,
 959                                  "logical %d == ei_block %d!",
 960                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 961                 return -EIO;
 962         }
 963
 964         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 965                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 966                 EXT4_ERROR_INODE(inode,
 967                                  "eh_entries %d >= eh_max %d!",
 968                                  le16_to_cpu(curp->p_hdr->eh_entries),
 969                                  le16_to_cpu(curp->p_hdr->eh_max));
 970                 return -EIO;
 971         }
 972
 973         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 974                 /* insert after */
 975                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 976                 ix = curp->p_idx + 1;
 977         } else {
 978                 /* insert before */
 979                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 980                 ix = curp->p_idx;
 981         }
 982
 983         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 984         BUG_ON(len < 0);
 985         if (len > 0) {
 986                 ext_debug("insert new index %d: "
 987                                 "move %d indices from 0x%p to 0x%p\n",
 988                                 logical, len, ix, ix + 1);
 989                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 990         }
 991
 992         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 993                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 994                 return -EIO;
 995         }
 996
 997         ix->ei_block = cpu_to_le32(logical);
 998         ext4_idx_store_pblock(ix, ptr);
 999         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1000
1001         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1002                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1003                 return -EIO;
1004         }
1005
1006         err = ext4_ext_dirty(handle, inode, curp);
1007         ext4_std_error(inode->i_sb, err);
1008
1009         return err;
1010 }
1011
1012 /*
1013  * ext4_ext_split:
1014  * inserts new subtree into the path, using free index entry
1015  * at depth @at:
1016  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1017  * - makes decision where to split
1018  * - moves remaining extents and index entries (right to the split point)
1019  *   into the newly allocated blocks
1020  * - initializes subtree
1021  */
1022 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1023                           unsigned int flags,
1024                           struct ext4_ext_path *path,
1025                           struct ext4_extent *newext, int at)
1026 {
1027         struct buffer_head *bh = NULL;
1028         int depth = ext_depth(inode);
1029         struct ext4_extent_header *neh;
1030         struct ext4_extent_idx *fidx;
1031         int i = at, k, m, a;
1032         ext4_fsblk_t newblock, oldblock;
1033         __le32 border;
1034         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1035         int err = 0;
1036
1037         /* make decision: where to split? */
1038         /* FIXME: now decision is simplest: at current extent */
1039
1040         /* if current leaf will be split, then we should use
1041          * border from split point */
1042         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1043                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1044                 return -EIO;
1045         }
1046         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1047                 border = path[depth].p_ext[1].ee_block;
1048                 ext_debug("leaf will be split."
1049                                 " next leaf starts at %d\n",
1050                                   le32_to_cpu(border));
1051         } else {
1052                 border = newext->ee_block;
1053                 ext_debug("leaf will be added."
1054                                 " next leaf starts at %d\n",
1055                                 le32_to_cpu(border));
1056         }
1057
1058         /*
1059          * If error occurs, then we break processing
1060          * and mark filesystem read-only. index won't
1061          * be inserted and tree will be in consistent
1062          * state. Next mount will repair buffers too.
1063          */
1064
1065         /*
1066          * Get array to track all allocated blocks.
1067          * We need this to handle errors and free blocks
1068          * upon them.
1069          */
1070         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1071         if (!ablocks)
1072                 return -ENOMEM;
1073
1074         /* allocate all needed blocks */
1075         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1076         for (a = 0; a < depth - at; a++) {
1077                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1078                                                    newext, &err, flags);
1079                 if (newblock == 0)
1080                         goto cleanup;
1081                 ablocks[a] = newblock;
1082         }
1083
1084         /* initialize new leaf */
1085         newblock = ablocks[--a];
1086         if (unlikely(newblock == 0)) {
1087                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1088                 err = -EIO;
1089                 goto cleanup;
1090         }
1091         bh = sb_getblk(inode->i_sb, newblock);
1092         if (unlikely(!bh)) {
1093                 err = -ENOMEM;
1094                 goto cleanup;
1095         }
1096         lock_buffer(bh);
1097
1098         err = ext4_journal_get_create_access(handle, bh);
1099         if (err)
1100                 goto cleanup;
1101
1102         neh = ext_block_hdr(bh);
1103         neh->eh_entries = 0;
1104         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1105         neh->eh_magic = EXT4_EXT_MAGIC;
1106         neh->eh_depth = 0;
1107
1108         /* move remainder of path[depth] to the new leaf */
1109         if (unlikely(path[depth].p_hdr->eh_entries !=
1110                      path[depth].p_hdr->eh_max)) {
1111                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1112                                  path[depth].p_hdr->eh_entries,
1113                                  path[depth].p_hdr->eh_max);
1114                 err = -EIO;
1115                 goto cleanup;
1116         }
1117         /* start copy from next extent */
1118         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1119         ext4_ext_show_move(inode, path, newblock, depth);
1120         if (m) {
1121                 struct ext4_extent *ex;
1122                 ex = EXT_FIRST_EXTENT(neh);
1123                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1124                 le16_add_cpu(&neh->eh_entries, m);
1125         }
1126
1127         ext4_extent_block_csum_set(inode, neh);
1128         set_buffer_uptodate(bh);
1129         unlock_buffer(bh);
1130
1131         err = ext4_handle_dirty_metadata(handle, inode, bh);
1132         if (err)
1133                 goto cleanup;
1134         brelse(bh);
1135         bh = NULL;
1136
1137         /* correct old leaf */
1138         if (m) {
1139                 err = ext4_ext_get_access(handle, inode, path + depth);
1140                 if (err)
1141                         goto cleanup;
1142                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1143                 err = ext4_ext_dirty(handle, inode, path + depth);
1144                 if (err)
1145                         goto cleanup;
1146
1147         }
1148
1149         /* create intermediate indexes */
1150         k = depth - at - 1;
1151         if (unlikely(k < 0)) {
1152                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1153                 err = -EIO;
1154                 goto cleanup;
1155         }
1156         if (k)
1157                 ext_debug("create %d intermediate indices\n", k);
1158         /* insert new index into current index block */
1159         /* current depth stored in i var */
1160         i = depth - 1;
1161         while (k--) {
1162                 oldblock = newblock;
1163                 newblock = ablocks[--a];
1164                 bh = sb_getblk(inode->i_sb, newblock);
1165                 if (unlikely(!bh)) {
1166                         err = -ENOMEM;
1167                         goto cleanup;
1168                 }
1169                 lock_buffer(bh);
1170
1171                 err = ext4_journal_get_create_access(handle, bh);
1172                 if (err)
1173                         goto cleanup;
1174
1175                 neh = ext_block_hdr(bh);
1176                 neh->eh_entries = cpu_to_le16(1);
1177                 neh->eh_magic = EXT4_EXT_MAGIC;
1178                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1179                 neh->eh_depth = cpu_to_le16(depth - i);
1180                 fidx = EXT_FIRST_INDEX(neh);
1181                 fidx->ei_block = border;
1182                 ext4_idx_store_pblock(fidx, oldblock);
1183
1184                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1185                                 i, newblock, le32_to_cpu(border), oldblock);
1186
1187                 /* move remainder of path[i] to the new index block */
1188                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1189                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1190                         EXT4_ERROR_INODE(inode,
1191                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1192                                          le32_to_cpu(path[i].p_ext->ee_block));
1193                         err = -EIO;
1194                         goto cleanup;
1195                 }
1196                 /* start copy indexes */
1197                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1198                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1199                                 EXT_MAX_INDEX(path[i].p_hdr));
1200                 ext4_ext_show_move(inode, path, newblock, i);
1201                 if (m) {
1202                         memmove(++fidx, path[i].p_idx,
1203                                 sizeof(struct ext4_extent_idx) * m);
1204                         le16_add_cpu(&neh->eh_entries, m);
1205                 }
1206                 ext4_extent_block_csum_set(inode, neh);
1207                 set_buffer_uptodate(bh);
1208                 unlock_buffer(bh);
1209
1210                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1211                 if (err)
1212                         goto cleanup;
1213                 brelse(bh);
1214                 bh = NULL;
1215
1216                 /* correct old index */
1217                 if (m) {
1218                         err = ext4_ext_get_access(handle, inode, path + i);
1219                         if (err)
1220                                 goto cleanup;
1221                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1222                         err = ext4_ext_dirty(handle, inode, path + i);
1223                         if (err)
1224                                 goto cleanup;
1225                 }
1226
1227                 i--;
1228         }
1229
1230         /* insert new index */
1231         err = ext4_ext_insert_index(handle, inode, path + at,
1232                                     le32_to_cpu(border), newblock);
1233
1234 cleanup:
1235         if (bh) {
1236                 if (buffer_locked(bh))
1237                         unlock_buffer(bh);
1238                 brelse(bh);
1239         }
1240
1241         if (err) {
1242                 /* free all allocated blocks in error case */
1243                 for (i = 0; i < depth; i++) {
1244                         if (!ablocks[i])
1245                                 continue;
1246                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1247                                          EXT4_FREE_BLOCKS_METADATA);
1248                 }
1249         }
1250         kfree(ablocks);
1251
1252         return err;
1253 }
1254
1255 /*
1256  * ext4_ext_grow_indepth:
1257  * implements tree growing procedure:
1258  * - allocates new block
1259  * - moves top-level data (index block or leaf) into the new block
1260  * - initializes new top-level, creating index that points to the
1261  *   just created block
1262  */
1263 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1264                                  unsigned int flags)
1265 {
1266         struct ext4_extent_header *neh;
1267         struct buffer_head *bh;
1268         ext4_fsblk_t newblock, goal = 0;
1269         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1270         int err = 0;
1271
1272         /* Try to prepend new index to old one */
1273         if (ext_depth(inode))
1274                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1275         if (goal > le32_to_cpu(es->s_first_data_block)) {
1276                 flags |= EXT4_MB_HINT_TRY_GOAL;
1277                 goal--;
1278         } else
1279                 goal = ext4_inode_to_goal_block(inode);
1280         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1281                                         NULL, &err);
1282         if (newblock == 0)
1283                 return err;
1284
1285         bh = sb_getblk(inode->i_sb, newblock);
1286         if (unlikely(!bh))
1287                 return -ENOMEM;
1288         lock_buffer(bh);
1289
1290         err = ext4_journal_get_create_access(handle, bh);
1291         if (err) {
1292                 unlock_buffer(bh);
1293                 goto out;
1294         }
1295
1296         /* move top-level index/leaf into new block */
1297         memmove(bh->b_data, EXT4_I(inode)->i_data,
1298                 sizeof(EXT4_I(inode)->i_data));
1299
1300         /* set size of new block */
1301         neh = ext_block_hdr(bh);
1302         /* old root could have indexes or leaves
1303          * so calculate e_max right way */
1304         if (ext_depth(inode))
1305                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1306         else
1307                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1308         neh->eh_magic = EXT4_EXT_MAGIC;
1309         ext4_extent_block_csum_set(inode, neh);
1310         set_buffer_uptodate(bh);
1311         unlock_buffer(bh);
1312
1313         err = ext4_handle_dirty_metadata(handle, inode, bh);
1314         if (err)
1315                 goto out;
1316
1317         /* Update top-level index: num,max,pointer */
1318         neh = ext_inode_hdr(inode);
1319         neh->eh_entries = cpu_to_le16(1);
1320         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1321         if (neh->eh_depth == 0) {
1322                 /* Root extent block becomes index block */
1323                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1324                 EXT_FIRST_INDEX(neh)->ei_block =
1325                         EXT_FIRST_EXTENT(neh)->ee_block;
1326         }
1327         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1328                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1329                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1330                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1331
1332         le16_add_cpu(&neh->eh_depth, 1);
1333         ext4_mark_inode_dirty(handle, inode);
1334 out:
1335         brelse(bh);
1336
1337         return err;
1338 }
1339
1340 /*
1341  * ext4_ext_create_new_leaf:
1342  * finds empty index and adds new leaf.
1343  * if no free index is found, then it requests in-depth growing.
1344  */
1345 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1346                                     unsigned int mb_flags,
1347                                     unsigned int gb_flags,
1348                                     struct ext4_ext_path **ppath,
1349                                     struct ext4_extent *newext)
1350 {
1351         struct ext4_ext_path *path = *ppath;
1352         struct ext4_ext_path *curp;
1353         int depth, i, err = 0;
1354
1355 repeat:
1356         i = depth = ext_depth(inode);
1357
1358         /* walk up to the tree and look for free index entry */
1359         curp = path + depth;
1360         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1361                 i--;
1362                 curp--;
1363         }
1364
1365         /* we use already allocated block for index block,
1366          * so subsequent data blocks should be contiguous */
1367         if (EXT_HAS_FREE_INDEX(curp)) {
1368                 /* if we found index with free entry, then use that
1369                  * entry: create all needed subtree and add new leaf */
1370                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1371                 if (err)
1372                         goto out;
1373
1374                 /* refill path */
1375                 path = ext4_find_extent(inode,
1376                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1377                                     ppath, gb_flags);
1378                 if (IS_ERR(path))
1379                         err = PTR_ERR(path);
1380         } else {
1381                 /* tree is full, time to grow in depth */
1382                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1383                 if (err)
1384                         goto out;
1385
1386                 /* refill path */
1387                 path = ext4_find_extent(inode,
1388                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1389                                     ppath, gb_flags);
1390                 if (IS_ERR(path)) {
1391                         err = PTR_ERR(path);
1392                         goto out;
1393                 }
1394
1395                 /*
1396                  * only first (depth 0 -> 1) produces free space;
1397                  * in all other cases we have to split the grown tree
1398                  */
1399                 depth = ext_depth(inode);
1400                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1401                         /* now we need to split */
1402                         goto repeat;
1403                 }
1404         }
1405
1406 out:
1407         return err;
1408 }
1409
1410 /*
1411  * search the closest allocated block to the left for *logical
1412  * and returns it at @logical + it's physical address at @phys
1413  * if *logical is the smallest allocated block, the function
1414  * returns 0 at @phys
1415  * return value contains 0 (success) or error code
1416  */
1417 static int ext4_ext_search_left(struct inode *inode,
1418                                 struct ext4_ext_path *path,
1419                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1420 {
1421         struct ext4_extent_idx *ix;
1422         struct ext4_extent *ex;
1423         int depth, ee_len;
1424
1425         if (unlikely(path == NULL)) {
1426                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1427                 return -EIO;
1428         }
1429         depth = path->p_depth;
1430         *phys = 0;
1431
1432         if (depth == 0 && path->p_ext == NULL)
1433                 return 0;
1434
1435         /* usually extent in the path covers blocks smaller
1436          * then *logical, but it can be that extent is the
1437          * first one in the file */
1438
1439         ex = path[depth].p_ext;
1440         ee_len = ext4_ext_get_actual_len(ex);
1441         if (*logical < le32_to_cpu(ex->ee_block)) {
1442                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1443                         EXT4_ERROR_INODE(inode,
1444                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1445                                          *logical, le32_to_cpu(ex->ee_block));
1446                         return -EIO;
1447                 }
1448                 while (--depth >= 0) {
1449                         ix = path[depth].p_idx;
1450                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1451                                 EXT4_ERROR_INODE(inode,
1452                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1453                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1454                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1455                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1456                                   depth);
1457                                 return -EIO;
1458                         }
1459                 }
1460                 return 0;
1461         }
1462
1463         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1464                 EXT4_ERROR_INODE(inode,
1465                                  "logical %d < ee_block %d + ee_len %d!",
1466                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1467                 return -EIO;
1468         }
1469
1470         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1471         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1472         return 0;
1473 }
1474
1475 /*
1476  * search the closest allocated block to the right for *logical
1477  * and returns it at @logical + it's physical address at @phys
1478  * if *logical is the largest allocated block, the function
1479  * returns 0 at @phys
1480  * return value contains 0 (success) or error code
1481  */
1482 static int ext4_ext_search_right(struct inode *inode,
1483                                  struct ext4_ext_path *path,
1484                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1485                                  struct ext4_extent **ret_ex)
1486 {
1487         struct buffer_head *bh = NULL;
1488         struct ext4_extent_header *eh;
1489         struct ext4_extent_idx *ix;
1490         struct ext4_extent *ex;
1491         ext4_fsblk_t block;
1492         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1493         int ee_len;
1494
1495         if (unlikely(path == NULL)) {
1496                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1497                 return -EIO;
1498         }
1499         depth = path->p_depth;
1500         *phys = 0;
1501
1502         if (depth == 0 && path->p_ext == NULL)
1503                 return 0;
1504
1505         /* usually extent in the path covers blocks smaller
1506          * then *logical, but it can be that extent is the
1507          * first one in the file */
1508
1509         ex = path[depth].p_ext;
1510         ee_len = ext4_ext_get_actual_len(ex);
1511         if (*logical < le32_to_cpu(ex->ee_block)) {
1512                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1513                         EXT4_ERROR_INODE(inode,
1514                                          "first_extent(path[%d].p_hdr) != ex",
1515                                          depth);
1516                         return -EIO;
1517                 }
1518                 while (--depth >= 0) {
1519                         ix = path[depth].p_idx;
1520                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1521                                 EXT4_ERROR_INODE(inode,
1522                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1523                                                  *logical);
1524                                 return -EIO;
1525                         }
1526                 }
1527                 goto found_extent;
1528         }
1529
1530         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1531                 EXT4_ERROR_INODE(inode,
1532                                  "logical %d < ee_block %d + ee_len %d!",
1533                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1534                 return -EIO;
1535         }
1536
1537         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1538                 /* next allocated block in this leaf */
1539                 ex++;
1540                 goto found_extent;
1541         }
1542
1543         /* go up and search for index to the right */
1544         while (--depth >= 0) {
1545                 ix = path[depth].p_idx;
1546                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1547                         goto got_index;
1548         }
1549
1550         /* we've gone up to the root and found no index to the right */
1551         return 0;
1552
1553 got_index:
1554         /* we've found index to the right, let's
1555          * follow it and find the closest allocated
1556          * block to the right */
1557         ix++;
1558         block = ext4_idx_pblock(ix);
1559         while (++depth < path->p_depth) {
1560                 /* subtract from p_depth to get proper eh_depth */
1561                 bh = read_extent_tree_block(inode, block,
1562                                             path->p_depth - depth, 0);
1563                 if (IS_ERR(bh))
1564                         return PTR_ERR(bh);
1565                 eh = ext_block_hdr(bh);
1566                 ix = EXT_FIRST_INDEX(eh);
1567                 block = ext4_idx_pblock(ix);
1568                 put_bh(bh);
1569         }
1570
1571         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1572         if (IS_ERR(bh))
1573                 return PTR_ERR(bh);
1574         eh = ext_block_hdr(bh);
1575         ex = EXT_FIRST_EXTENT(eh);
1576 found_extent:
1577         *logical = le32_to_cpu(ex->ee_block);
1578         *phys = ext4_ext_pblock(ex);
1579         *ret_ex = ex;
1580         if (bh)
1581                 put_bh(bh);
1582         return 0;
1583 }
1584
1585 /*
1586  * ext4_ext_next_allocated_block:
1587  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1588  * NOTE: it considers block number from index entry as
1589  * allocated block. Thus, index entries have to be consistent
1590  * with leaves.
1591  */
1592 ext4_lblk_t
1593 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1594 {
1595         int depth;
1596
1597         BUG_ON(path == NULL);
1598         depth = path->p_depth;
1599
1600         if (depth == 0 && path->p_ext == NULL)
1601                 return EXT_MAX_BLOCKS;
1602
1603         while (depth >= 0) {
1604                 if (depth == path->p_depth) {
1605                         /* leaf */
1606                         if (path[depth].p_ext &&
1607                                 path[depth].p_ext !=
1608                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1609                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1610                 } else {
1611                         /* index */
1612                         if (path[depth].p_idx !=
1613                                         EXT_LAST_INDEX(path[depth].p_hdr))
1614                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1615                 }
1616                 depth--;
1617         }
1618
1619         return EXT_MAX_BLOCKS;
1620 }
1621
1622 /*
1623  * ext4_ext_next_leaf_block:
1624  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1625  */
1626 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1627 {
1628         int depth;
1629
1630         BUG_ON(path == NULL);
1631         depth = path->p_depth;
1632
1633         /* zero-tree has no leaf blocks at all */
1634         if (depth == 0)
1635                 return EXT_MAX_BLOCKS;
1636
1637         /* go to index block */
1638         depth--;
1639
1640         while (depth >= 0) {
1641                 if (path[depth].p_idx !=
1642                                 EXT_LAST_INDEX(path[depth].p_hdr))
1643                         return (ext4_lblk_t)
1644                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1645                 depth--;
1646         }
1647
1648         return EXT_MAX_BLOCKS;
1649 }
1650
1651 /*
1652  * ext4_ext_correct_indexes:
1653  * if leaf gets modified and modified extent is first in the leaf,
1654  * then we have to correct all indexes above.
1655  * TODO: do we need to correct tree in all cases?
1656  */
1657 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1658                                 struct ext4_ext_path *path)
1659 {
1660         struct ext4_extent_header *eh;
1661         int depth = ext_depth(inode);
1662         struct ext4_extent *ex;
1663         __le32 border;
1664         int k, err = 0;
1665
1666         eh = path[depth].p_hdr;
1667         ex = path[depth].p_ext;
1668
1669         if (unlikely(ex == NULL || eh == NULL)) {
1670                 EXT4_ERROR_INODE(inode,
1671                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1672                 return -EIO;
1673         }
1674
1675         if (depth == 0) {
1676                 /* there is no tree at all */
1677                 return 0;
1678         }
1679
1680         if (ex != EXT_FIRST_EXTENT(eh)) {
1681                 /* we correct tree if first leaf got modified only */
1682                 return 0;
1683         }
1684
1685         /*
1686          * TODO: we need correction if border is smaller than current one
1687          */
1688         k = depth - 1;
1689         border = path[depth].p_ext->ee_block;
1690         err = ext4_ext_get_access(handle, inode, path + k);
1691         if (err)
1692                 return err;
1693         path[k].p_idx->ei_block = border;
1694         err = ext4_ext_dirty(handle, inode, path + k);
1695         if (err)
1696                 return err;
1697
1698         while (k--) {
1699                 /* change all left-side indexes */
1700                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1701                         break;
1702                 err = ext4_ext_get_access(handle, inode, path + k);
1703                 if (err)
1704                         break;
1705                 path[k].p_idx->ei_block = border;
1706                 err = ext4_ext_dirty(handle, inode, path + k);
1707                 if (err)
1708                         break;
1709         }
1710
1711         return err;
1712 }
1713
1714 int
1715 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1716                                 struct ext4_extent *ex2)
1717 {
1718         unsigned short ext1_ee_len, ext2_ee_len;
1719
1720         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1721                 return 0;
1722
1723         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1724         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1725
1726         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1727                         le32_to_cpu(ex2->ee_block))
1728                 return 0;
1729
1730         /*
1731          * To allow future support for preallocated extents to be added
1732          * as an RO_COMPAT feature, refuse to merge to extents if
1733          * this can result in the top bit of ee_len being set.
1734          */
1735         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1736                 return 0;
1737         if (ext4_ext_is_unwritten(ex1) &&
1738             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1739              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1740              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1741                 return 0;
1742 #ifdef AGGRESSIVE_TEST
1743         if (ext1_ee_len >= 4)
1744                 return 0;
1745 #endif
1746
1747         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1748                 return 1;
1749         return 0;
1750 }
1751
1752 /*
1753  * This function tries to merge the "ex" extent to the next extent in the tree.
1754  * It always tries to merge towards right. If you want to merge towards
1755  * left, pass "ex - 1" as argument instead of "ex".
1756  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1757  * 1 if they got merged.
1758  */
1759 static int ext4_ext_try_to_merge_right(struct inode *inode,
1760                                  struct ext4_ext_path *path,
1761                                  struct ext4_extent *ex)
1762 {
1763         struct ext4_extent_header *eh;
1764         unsigned int depth, len;
1765         int merge_done = 0, unwritten;
1766
1767         depth = ext_depth(inode);
1768         BUG_ON(path[depth].p_hdr == NULL);
1769         eh = path[depth].p_hdr;
1770
1771         while (ex < EXT_LAST_EXTENT(eh)) {
1772                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1773                         break;
1774                 /* merge with next extent! */
1775                 unwritten = ext4_ext_is_unwritten(ex);
1776                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1777                                 + ext4_ext_get_actual_len(ex + 1));
1778                 if (unwritten)
1779                         ext4_ext_mark_unwritten(ex);
1780
1781                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1782                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1783                                 * sizeof(struct ext4_extent);
1784                         memmove(ex + 1, ex + 2, len);
1785                 }
1786                 le16_add_cpu(&eh->eh_entries, -1);
1787                 merge_done = 1;
1788                 WARN_ON(eh->eh_entries == 0);
1789                 if (!eh->eh_entries)
1790                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1791         }
1792
1793         return merge_done;
1794 }
1795
1796 /*
1797  * This function does a very simple check to see if we can collapse
1798  * an extent tree with a single extent tree leaf block into the inode.
1799  */
1800 static void ext4_ext_try_to_merge_up(handle_t *handle,
1801                                      struct inode *inode,
1802                                      struct ext4_ext_path *path)
1803 {
1804         size_t s;
1805         unsigned max_root = ext4_ext_space_root(inode, 0);
1806         ext4_fsblk_t blk;
1807
1808         if ((path[0].p_depth != 1) ||
1809             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1810             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1811                 return;
1812
1813         /*
1814          * We need to modify the block allocation bitmap and the block
1815          * group descriptor to release the extent tree block.  If we
1816          * can't get the journal credits, give up.
1817          */
1818         if (ext4_journal_extend(handle, 2))
1819                 return;
1820
1821         /*
1822          * Copy the extent data up to the inode
1823          */
1824         blk = ext4_idx_pblock(path[0].p_idx);
1825         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1826                 sizeof(struct ext4_extent_idx);
1827         s += sizeof(struct ext4_extent_header);
1828
1829         path[1].p_maxdepth = path[0].p_maxdepth;
1830         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1831         path[0].p_depth = 0;
1832         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1833                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1834         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1835
1836         brelse(path[1].p_bh);
1837         ext4_free_blocks(handle, inode, NULL, blk, 1,
1838                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1839 }
1840
1841 /*
1842  * This function tries to merge the @ex extent to neighbours in the tree.
1843  * return 1 if merge left else 0.
1844  */
1845 static void ext4_ext_try_to_merge(handle_t *handle,
1846                                   struct inode *inode,
1847                                   struct ext4_ext_path *path,
1848                                   struct ext4_extent *ex) {
1849         struct ext4_extent_header *eh;
1850         unsigned int depth;
1851         int merge_done = 0;
1852
1853         depth = ext_depth(inode);
1854         BUG_ON(path[depth].p_hdr == NULL);
1855         eh = path[depth].p_hdr;
1856
1857         if (ex > EXT_FIRST_EXTENT(eh))
1858                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1859
1860         if (!merge_done)
1861                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1862
1863         ext4_ext_try_to_merge_up(handle, inode, path);
1864 }
1865
1866 /*
1867  * check if a portion of the "newext" extent overlaps with an
1868  * existing extent.
1869  *
1870  * If there is an overlap discovered, it updates the length of the newext
1871  * such that there will be no overlap, and then returns 1.
1872  * If there is no overlap found, it returns 0.
1873  */
1874 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1875                                            struct inode *inode,
1876                                            struct ext4_extent *newext,
1877                                            struct ext4_ext_path *path)
1878 {
1879         ext4_lblk_t b1, b2;
1880         unsigned int depth, len1;
1881         unsigned int ret = 0;
1882
1883         b1 = le32_to_cpu(newext->ee_block);
1884         len1 = ext4_ext_get_actual_len(newext);
1885         depth = ext_depth(inode);
1886         if (!path[depth].p_ext)
1887                 goto out;
1888         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1889
1890         /*
1891          * get the next allocated block if the extent in the path
1892          * is before the requested block(s)
1893          */
1894         if (b2 < b1) {
1895                 b2 = ext4_ext_next_allocated_block(path);
1896                 if (b2 == EXT_MAX_BLOCKS)
1897                         goto out;
1898                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1899         }
1900
1901         /* check for wrap through zero on extent logical start block*/
1902         if (b1 + len1 < b1) {
1903                 len1 = EXT_MAX_BLOCKS - b1;
1904                 newext->ee_len = cpu_to_le16(len1);
1905                 ret = 1;
1906         }
1907
1908         /* check for overlap */
1909         if (b1 + len1 > b2) {
1910                 newext->ee_len = cpu_to_le16(b2 - b1);
1911                 ret = 1;
1912         }
1913 out:
1914         return ret;
1915 }
1916
1917 /*
1918  * ext4_ext_insert_extent:
1919  * tries to merge requsted extent into the existing extent or
1920  * inserts requested extent as new one into the tree,
1921  * creating new leaf in the no-space case.
1922  */
1923 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1924                                 struct ext4_ext_path **ppath,
1925                                 struct ext4_extent *newext, int gb_flags)
1926 {
1927         struct ext4_ext_path *path = *ppath;
1928         struct ext4_extent_header *eh;
1929         struct ext4_extent *ex, *fex;
1930         struct ext4_extent *nearex; /* nearest extent */
1931         struct ext4_ext_path *npath = NULL;
1932         int depth, len, err;
1933         ext4_lblk_t next;
1934         int mb_flags = 0, unwritten;
1935
1936         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1937                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1938         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1939                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1940                 return -EIO;
1941         }
1942         depth = ext_depth(inode);
1943         ex = path[depth].p_ext;
1944         eh = path[depth].p_hdr;
1945         if (unlikely(path[depth].p_hdr == NULL)) {
1946                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1947                 return -EIO;
1948         }
1949
1950         /* try to insert block into found extent and return */
1951         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1952
1953                 /*
1954                  * Try to see whether we should rather test the extent on
1955                  * right from ex, or from the left of ex. This is because
1956                  * ext4_find_extent() can return either extent on the
1957                  * left, or on the right from the searched position. This
1958                  * will make merging more effective.
1959                  */
1960                 if (ex < EXT_LAST_EXTENT(eh) &&
1961                     (le32_to_cpu(ex->ee_block) +
1962                     ext4_ext_get_actual_len(ex) <
1963                     le32_to_cpu(newext->ee_block))) {
1964                         ex += 1;
1965                         goto prepend;
1966                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1967                            (le32_to_cpu(newext->ee_block) +
1968                            ext4_ext_get_actual_len(newext) <
1969                            le32_to_cpu(ex->ee_block)))
1970                         ex -= 1;
1971
1972                 /* Try to append newex to the ex */
1973                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1974                         ext_debug("append [%d]%d block to %u:[%d]%d"
1975                                   "(from %llu)\n",
1976                                   ext4_ext_is_unwritten(newext),
1977                                   ext4_ext_get_actual_len(newext),
1978                                   le32_to_cpu(ex->ee_block),
1979                                   ext4_ext_is_unwritten(ex),
1980                                   ext4_ext_get_actual_len(ex),
1981                                   ext4_ext_pblock(ex));
1982                         err = ext4_ext_get_access(handle, inode,
1983                                                   path + depth);
1984                         if (err)
1985                                 return err;
1986                         unwritten = ext4_ext_is_unwritten(ex);
1987                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1988                                         + ext4_ext_get_actual_len(newext));
1989                         if (unwritten)
1990                                 ext4_ext_mark_unwritten(ex);
1991                         eh = path[depth].p_hdr;
1992                         nearex = ex;
1993                         goto merge;
1994                 }
1995
1996 prepend:
1997                 /* Try to prepend newex to the ex */
1998                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
1999                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2000                                   "(from %llu)\n",
2001                                   le32_to_cpu(newext->ee_block),
2002                                   ext4_ext_is_unwritten(newext),
2003                                   ext4_ext_get_actual_len(newext),
2004                                   le32_to_cpu(ex->ee_block),
2005                                   ext4_ext_is_unwritten(ex),
2006                                   ext4_ext_get_actual_len(ex),
2007                                   ext4_ext_pblock(ex));
2008                         err = ext4_ext_get_access(handle, inode,
2009                                                   path + depth);
2010                         if (err)
2011                                 return err;
2012
2013                         unwritten = ext4_ext_is_unwritten(ex);
2014                         ex->ee_block = newext->ee_block;
2015                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2016                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2017                                         + ext4_ext_get_actual_len(newext));
2018                         if (unwritten)
2019                                 ext4_ext_mark_unwritten(ex);
2020                         eh = path[depth].p_hdr;
2021                         nearex = ex;
2022                         goto merge;
2023                 }
2024         }
2025
2026         depth = ext_depth(inode);
2027         eh = path[depth].p_hdr;
2028         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2029                 goto has_space;
2030
2031         /* probably next leaf has space for us? */
2032         fex = EXT_LAST_EXTENT(eh);
2033         next = EXT_MAX_BLOCKS;
2034         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2035                 next = ext4_ext_next_leaf_block(path);
2036         if (next != EXT_MAX_BLOCKS) {
2037                 ext_debug("next leaf block - %u\n", next);
2038                 BUG_ON(npath != NULL);
2039                 npath = ext4_find_extent(inode, next, NULL, 0);
2040                 if (IS_ERR(npath))
2041                         return PTR_ERR(npath);
2042                 BUG_ON(npath->p_depth != path->p_depth);
2043                 eh = npath[depth].p_hdr;
2044                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2045                         ext_debug("next leaf isn't full(%d)\n",
2046                                   le16_to_cpu(eh->eh_entries));
2047                         path = npath;
2048                         goto has_space;
2049                 }
2050                 ext_debug("next leaf has no free space(%d,%d)\n",
2051                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2052         }
2053
2054         /*
2055          * There is no free space in the found leaf.
2056          * We're gonna add a new leaf in the tree.
2057          */
2058         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2059                 mb_flags |= EXT4_MB_USE_RESERVED;
2060         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2061                                        ppath, newext);
2062         if (err)
2063                 goto cleanup;
2064         depth = ext_depth(inode);
2065         eh = path[depth].p_hdr;
2066
2067 has_space:
2068         nearex = path[depth].p_ext;
2069
2070         err = ext4_ext_get_access(handle, inode, path + depth);
2071         if (err)
2072                 goto cleanup;
2073
2074         if (!nearex) {
2075                 /* there is no extent in this leaf, create first one */
2076                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2077                                 le32_to_cpu(newext->ee_block),
2078                                 ext4_ext_pblock(newext),
2079                                 ext4_ext_is_unwritten(newext),
2080                                 ext4_ext_get_actual_len(newext));
2081                 nearex = EXT_FIRST_EXTENT(eh);
2082         } else {
2083                 if (le32_to_cpu(newext->ee_block)
2084                            > le32_to_cpu(nearex->ee_block)) {
2085                         /* Insert after */
2086                         ext_debug("insert %u:%llu:[%d]%d before: "
2087                                         "nearest %p\n",
2088                                         le32_to_cpu(newext->ee_block),
2089                                         ext4_ext_pblock(newext),
2090                                         ext4_ext_is_unwritten(newext),
2091                                         ext4_ext_get_actual_len(newext),
2092                                         nearex);
2093                         nearex++;
2094                 } else {
2095                         /* Insert before */
2096                         BUG_ON(newext->ee_block == nearex->ee_block);
2097                         ext_debug("insert %u:%llu:[%d]%d after: "
2098                                         "nearest %p\n",
2099                                         le32_to_cpu(newext->ee_block),
2100                                         ext4_ext_pblock(newext),
2101                                         ext4_ext_is_unwritten(newext),
2102                                         ext4_ext_get_actual_len(newext),
2103                                         nearex);
2104                 }
2105                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2106                 if (len > 0) {
2107                         ext_debug("insert %u:%llu:[%d]%d: "
2108                                         "move %d extents from 0x%p to 0x%p\n",
2109                                         le32_to_cpu(newext->ee_block),
2110                                         ext4_ext_pblock(newext),
2111                                         ext4_ext_is_unwritten(newext),
2112                                         ext4_ext_get_actual_len(newext),
2113                                         len, nearex, nearex + 1);
2114                         memmove(nearex + 1, nearex,
2115                                 len * sizeof(struct ext4_extent));
2116                 }
2117         }
2118
2119         le16_add_cpu(&eh->eh_entries, 1);
2120         path[depth].p_ext = nearex;
2121         nearex->ee_block = newext->ee_block;
2122         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2123         nearex->ee_len = newext->ee_len;
2124
2125 merge:
2126         /* try to merge extents */
2127         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2128                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2129
2130
2131         /* time to correct all indexes above */
2132         err = ext4_ext_correct_indexes(handle, inode, path);
2133         if (err)
2134                 goto cleanup;
2135
2136         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2137
2138 cleanup:
2139         ext4_ext_drop_refs(npath);
2140         kfree(npath);
2141         return err;
2142 }
2143
2144 static int ext4_fill_fiemap_extents(struct inode *inode,
2145                                     ext4_lblk_t block, ext4_lblk_t num,
2146                                     struct fiemap_extent_info *fieinfo)
2147 {
2148         struct ext4_ext_path *path = NULL;
2149         struct ext4_extent *ex;
2150         struct extent_status es;
2151         ext4_lblk_t next, next_del, start = 0, end = 0;
2152         ext4_lblk_t last = block + num;
2153         int exists, depth = 0, err = 0;
2154         unsigned int flags = 0;
2155         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2156
2157         while (block < last && block != EXT_MAX_BLOCKS) {
2158                 num = last - block;
2159                 /* find extent for this block */
2160                 down_read(&EXT4_I(inode)->i_data_sem);
2161
2162                 path = ext4_find_extent(inode, block, &path, 0);
2163                 if (IS_ERR(path)) {
2164                         up_read(&EXT4_I(inode)->i_data_sem);
2165                         err = PTR_ERR(path);
2166                         path = NULL;
2167                         break;
2168                 }
2169
2170                 depth = ext_depth(inode);
2171                 if (unlikely(path[depth].p_hdr == NULL)) {
2172                         up_read(&EXT4_I(inode)->i_data_sem);
2173                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2174                         err = -EIO;
2175                         break;
2176                 }
2177                 ex = path[depth].p_ext;
2178                 next = ext4_ext_next_allocated_block(path);
2179
2180                 flags = 0;
2181                 exists = 0;
2182                 if (!ex) {
2183                         /* there is no extent yet, so try to allocate
2184                          * all requested space */
2185                         start = block;
2186                         end = block + num;
2187                 } else if (le32_to_cpu(ex->ee_block) > block) {
2188                         /* need to allocate space before found extent */
2189                         start = block;
2190                         end = le32_to_cpu(ex->ee_block);
2191                         if (block + num < end)
2192                                 end = block + num;
2193                 } else if (block >= le32_to_cpu(ex->ee_block)
2194                                         + ext4_ext_get_actual_len(ex)) {
2195                         /* need to allocate space after found extent */
2196                         start = block;
2197                         end = block + num;
2198                         if (end >= next)
2199                                 end = next;
2200                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2201                         /*
2202                          * some part of requested space is covered
2203                          * by found extent
2204                          */
2205                         start = block;
2206                         end = le32_to_cpu(ex->ee_block)
2207                                 + ext4_ext_get_actual_len(ex);
2208                         if (block + num < end)
2209                                 end = block + num;
2210                         exists = 1;
2211                 } else {
2212                         BUG();
2213                 }
2214                 BUG_ON(end <= start);
2215
2216                 if (!exists) {
2217                         es.es_lblk = start;
2218                         es.es_len = end - start;
2219                         es.es_pblk = 0;
2220                 } else {
2221                         es.es_lblk = le32_to_cpu(ex->ee_block);
2222                         es.es_len = ext4_ext_get_actual_len(ex);
2223                         es.es_pblk = ext4_ext_pblock(ex);
2224                         if (ext4_ext_is_unwritten(ex))
2225                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2226                 }
2227
2228                 /*
2229                  * Find delayed extent and update es accordingly. We call
2230                  * it even in !exists case to find out whether es is the
2231                  * last existing extent or not.
2232                  */
2233                 next_del = ext4_find_delayed_extent(inode, &es);
2234                 if (!exists && next_del) {
2235                         exists = 1;
2236                         flags |= (FIEMAP_EXTENT_DELALLOC |
2237                                   FIEMAP_EXTENT_UNKNOWN);
2238                 }
2239                 up_read(&EXT4_I(inode)->i_data_sem);
2240
2241                 if (unlikely(es.es_len == 0)) {
2242                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2243                         err = -EIO;
2244                         break;
2245                 }
2246
2247                 /*
2248                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2249                  * we need to check next == EXT_MAX_BLOCKS because it is
2250                  * possible that an extent is with unwritten and delayed
2251                  * status due to when an extent is delayed allocated and
2252                  * is allocated by fallocate status tree will track both of
2253                  * them in a extent.
2254                  *
2255                  * So we could return a unwritten and delayed extent, and
2256                  * its block is equal to 'next'.
2257                  */
2258                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2259                         flags |= FIEMAP_EXTENT_LAST;
2260                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2261                                      next != EXT_MAX_BLOCKS)) {
2262                                 EXT4_ERROR_INODE(inode,
2263                                                  "next extent == %u, next "
2264                                                  "delalloc extent = %u",
2265                                                  next, next_del);
2266                                 err = -EIO;
2267                                 break;
2268                         }
2269                 }
2270
2271                 if (exists) {
2272                         err = fiemap_fill_next_extent(fieinfo,
2273                                 (__u64)es.es_lblk << blksize_bits,
2274                                 (__u64)es.es_pblk << blksize_bits,
2275                                 (__u64)es.es_len << blksize_bits,
2276                                 flags);
2277                         if (err < 0)
2278                                 break;
2279                         if (err == 1) {
2280                                 err = 0;
2281                                 break;
2282                         }
2283                 }
2284
2285                 block = es.es_lblk + es.es_len;
2286         }
2287
2288         ext4_ext_drop_refs(path);
2289         kfree(path);
2290         return err;
2291 }
2292
2293 /*
2294  * ext4_ext_put_gap_in_cache:
2295  * calculate boundaries of the gap that the requested block fits into
2296  * and cache this gap
2297  */
2298 static void
2299 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2300                                 ext4_lblk_t block)
2301 {
2302         int depth = ext_depth(inode);
2303         ext4_lblk_t len;
2304         ext4_lblk_t lblock;
2305         struct ext4_extent *ex;
2306         struct extent_status es;
2307
2308         ex = path[depth].p_ext;
2309         if (ex == NULL) {
2310                 /* there is no extent yet, so gap is [0;-] */
2311                 lblock = 0;
2312                 len = EXT_MAX_BLOCKS;
2313                 ext_debug("cache gap(whole file):");
2314         } else if (block < le32_to_cpu(ex->ee_block)) {
2315                 lblock = block;
2316                 len = le32_to_cpu(ex->ee_block) - block;
2317                 ext_debug("cache gap(before): %u [%u:%u]",
2318                                 block,
2319                                 le32_to_cpu(ex->ee_block),
2320                                  ext4_ext_get_actual_len(ex));
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         } else {
2335                 BUG();
2336         }
2337
2338         ext4_es_find_delayed_extent_range(inode, lblock, lblock + len - 1, &es);
2339         if (es.es_len) {
2340                 /* There's delayed extent containing lblock? */
2341                 if (es.es_lblk <= lblock)
2342                         return;
2343                 len = min(es.es_lblk - lblock, len);
2344         }
2345         ext_debug(" -> %u:%u\n", lblock, len);
2346         ext4_es_insert_extent(inode, lblock, len, ~0, EXTENT_STATUS_HOLE);
2347 }
2348
2349 /*
2350  * ext4_ext_rm_idx:
2351  * removes index from the index block.
2352  */
2353 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2354                         struct ext4_ext_path *path, int depth)
2355 {
2356         int err;
2357         ext4_fsblk_t leaf;
2358
2359         /* free index block */
2360         depth--;
2361         path = path + depth;
2362         leaf = ext4_idx_pblock(path->p_idx);
2363         if (unlikely(path->p_hdr->eh_entries == 0)) {
2364                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2365                 return -EIO;
2366         }
2367         err = ext4_ext_get_access(handle, inode, path);
2368         if (err)
2369                 return err;
2370
2371         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2372                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2373                 len *= sizeof(struct ext4_extent_idx);
2374                 memmove(path->p_idx, path->p_idx + 1, len);
2375         }
2376
2377         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2378         err = ext4_ext_dirty(handle, inode, path);
2379         if (err)
2380                 return err;
2381         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2382         trace_ext4_ext_rm_idx(inode, leaf);
2383
2384         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2385                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2386
2387         while (--depth >= 0) {
2388                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2389                         break;
2390                 path--;
2391                 err = ext4_ext_get_access(handle, inode, path);
2392                 if (err)
2393                         break;
2394                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2395                 err = ext4_ext_dirty(handle, inode, path);
2396                 if (err)
2397                         break;
2398         }
2399         return err;
2400 }
2401
2402 /*
2403  * ext4_ext_calc_credits_for_single_extent:
2404  * This routine returns max. credits that needed to insert an extent
2405  * to the extent tree.
2406  * When pass the actual path, the caller should calculate credits
2407  * under i_data_sem.
2408  */
2409 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2410                                                 struct ext4_ext_path *path)
2411 {
2412         if (path) {
2413                 int depth = ext_depth(inode);
2414                 int ret = 0;
2415
2416                 /* probably there is space in leaf? */
2417                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2418                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2419
2420                         /*
2421                          *  There are some space in the leaf tree, no
2422                          *  need to account for leaf block credit
2423                          *
2424                          *  bitmaps and block group descriptor blocks
2425                          *  and other metadata blocks still need to be
2426                          *  accounted.
2427                          */
2428                         /* 1 bitmap, 1 block group descriptor */
2429                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2430                         return ret;
2431                 }
2432         }
2433
2434         return ext4_chunk_trans_blocks(inode, nrblocks);
2435 }
2436
2437 /*
2438  * How many index/leaf blocks need to change/allocate to add @extents extents?
2439  *
2440  * If we add a single extent, then in the worse case, each tree level
2441  * index/leaf need to be changed in case of the tree split.
2442  *
2443  * If more extents are inserted, they could cause the whole tree split more
2444  * than once, but this is really rare.
2445  */
2446 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2447 {
2448         int index;
2449         int depth;
2450
2451         /* If we are converting the inline data, only one is needed here. */
2452         if (ext4_has_inline_data(inode))
2453                 return 1;
2454
2455         depth = ext_depth(inode);
2456
2457         if (extents <= 1)
2458                 index = depth * 2;
2459         else
2460                 index = depth * 3;
2461
2462         return index;
2463 }
2464
2465 static inline int get_default_free_blocks_flags(struct inode *inode)
2466 {
2467         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2468                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2469         else if (ext4_should_journal_data(inode))
2470                 return EXT4_FREE_BLOCKS_FORGET;
2471         return 0;
2472 }
2473
2474 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2475                               struct ext4_extent *ex,
2476                               long long *partial_cluster,
2477                               ext4_lblk_t from, ext4_lblk_t to)
2478 {
2479         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2480         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2481         ext4_fsblk_t pblk;
2482         int flags = get_default_free_blocks_flags(inode);
2483
2484         /*
2485          * For bigalloc file systems, we never free a partial cluster
2486          * at the beginning of the extent.  Instead, we make a note
2487          * that we tried freeing the cluster, and check to see if we
2488          * need to free it on a subsequent call to ext4_remove_blocks,
2489          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2490          */
2491         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2492
2493         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2494         /*
2495          * If we have a partial cluster, and it's different from the
2496          * cluster of the last block, we need to explicitly free the
2497          * partial cluster here.
2498          */
2499         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2500         if (*partial_cluster > 0 &&
2501             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2502                 ext4_free_blocks(handle, inode, NULL,
2503                                  EXT4_C2B(sbi, *partial_cluster),
2504                                  sbi->s_cluster_ratio, flags);
2505                 *partial_cluster = 0;
2506         }
2507
2508 #ifdef EXTENTS_STATS
2509         {
2510                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2511                 spin_lock(&sbi->s_ext_stats_lock);
2512                 sbi->s_ext_blocks += ee_len;
2513                 sbi->s_ext_extents++;
2514                 if (ee_len < sbi->s_ext_min)
2515                         sbi->s_ext_min = ee_len;
2516                 if (ee_len > sbi->s_ext_max)
2517                         sbi->s_ext_max = ee_len;
2518                 if (ext_depth(inode) > sbi->s_depth_max)
2519                         sbi->s_depth_max = ext_depth(inode);
2520                 spin_unlock(&sbi->s_ext_stats_lock);
2521         }
2522 #endif
2523         if (from >= le32_to_cpu(ex->ee_block)
2524             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2525                 /* tail removal */
2526                 ext4_lblk_t num;
2527                 long long first_cluster;
2528
2529                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2530                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2531                 /*
2532                  * Usually we want to free partial cluster at the end of the
2533                  * extent, except for the situation when the cluster is still
2534                  * used by any other extent (partial_cluster is negative).
2535                  */
2536                 if (*partial_cluster < 0 &&
2537                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2538                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2539
2540                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2541                           num, pblk, *partial_cluster);
2542                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2543                 /*
2544                  * If the block range to be freed didn't start at the
2545                  * beginning of a cluster, and we removed the entire
2546                  * extent and the cluster is not used by any other extent,
2547                  * save the partial cluster here, since we might need to
2548                  * delete if we determine that the truncate or punch hole
2549                  * operation has removed all of the blocks in the cluster.
2550                  * If that cluster is used by another extent, preserve its
2551                  * negative value so it isn't freed later on.
2552                  *
2553                  * If the whole extent wasn't freed, we've reached the
2554                  * start of the truncated/punched region and have finished
2555                  * removing blocks.  If there's a partial cluster here it's
2556                  * shared with the remainder of the extent and is no longer
2557                  * a candidate for removal.
2558                  */
2559                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2560                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2561                         if (first_cluster != -*partial_cluster)
2562                                 *partial_cluster = first_cluster;
2563                 } else {
2564                         *partial_cluster = 0;
2565                 }
2566         } else
2567                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2568                            "%u-%u from %u:%u\n",
2569                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2570         return 0;
2571 }
2572
2573
2574 /*
2575  * ext4_ext_rm_leaf() Removes the extents associated with the
2576  * blocks appearing between "start" and "end".  Both "start"
2577  * and "end" must appear in the same extent or EIO is returned.
2578  *
2579  * @handle: The journal handle
2580  * @inode:  The files inode
2581  * @path:   The path to the leaf
2582  * @partial_cluster: The cluster which we'll have to free if all extents
2583  *                   has been released from it.  However, if this value is
2584  *                   negative, it's a cluster just to the right of the
2585  *                   punched region and it must not be freed.
2586  * @start:  The first block to remove
2587  * @end:   The last block to remove
2588  */
2589 static int
2590 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2591                  struct ext4_ext_path *path,
2592                  long long *partial_cluster,
2593                  ext4_lblk_t start, ext4_lblk_t end)
2594 {
2595         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2596         int err = 0, correct_index = 0;
2597         int depth = ext_depth(inode), credits;
2598         struct ext4_extent_header *eh;
2599         ext4_lblk_t a, b;
2600         unsigned num;
2601         ext4_lblk_t ex_ee_block;
2602         unsigned short ex_ee_len;
2603         unsigned unwritten = 0;
2604         struct ext4_extent *ex;
2605         ext4_fsblk_t pblk;
2606
2607         /* the header must be checked already in ext4_ext_remove_space() */
2608         ext_debug("truncate since %u in leaf to %u\n", start, end);
2609         if (!path[depth].p_hdr)
2610                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2611         eh = path[depth].p_hdr;
2612         if (unlikely(path[depth].p_hdr == NULL)) {
2613                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2614                 return -EIO;
2615         }
2616         /* find where to start removing */
2617         ex = path[depth].p_ext;
2618         if (!ex)
2619                 ex = EXT_LAST_EXTENT(eh);
2620
2621         ex_ee_block = le32_to_cpu(ex->ee_block);
2622         ex_ee_len = ext4_ext_get_actual_len(ex);
2623
2624         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2625
2626         while (ex >= EXT_FIRST_EXTENT(eh) &&
2627                         ex_ee_block + ex_ee_len > start) {
2628
2629                 if (ext4_ext_is_unwritten(ex))
2630                         unwritten = 1;
2631                 else
2632                         unwritten = 0;
2633
2634                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2635                           unwritten, ex_ee_len);
2636                 path[depth].p_ext = ex;
2637
2638                 a = ex_ee_block > start ? ex_ee_block : start;
2639                 b = ex_ee_block+ex_ee_len - 1 < end ?
2640                         ex_ee_block+ex_ee_len - 1 : end;
2641
2642                 ext_debug("  border %u:%u\n", a, b);
2643
2644                 /* If this extent is beyond the end of the hole, skip it */
2645                 if (end < ex_ee_block) {
2646                         /*
2647                          * We're going to skip this extent and move to another,
2648                          * so note that its first cluster is in use to avoid
2649                          * freeing it when removing blocks.  Eventually, the
2650                          * right edge of the truncated/punched region will
2651                          * be just to the left.
2652                          */
2653                         if (sbi->s_cluster_ratio > 1) {
2654                                 pblk = ext4_ext_pblock(ex);
2655                                 *partial_cluster =
2656                                         -(long long) EXT4_B2C(sbi, pblk);
2657                         }
2658                         ex--;
2659                         ex_ee_block = le32_to_cpu(ex->ee_block);
2660                         ex_ee_len = ext4_ext_get_actual_len(ex);
2661                         continue;
2662                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2663                         EXT4_ERROR_INODE(inode,
2664                                          "can not handle truncate %u:%u "
2665                                          "on extent %u:%u",
2666                                          start, end, ex_ee_block,
2667                                          ex_ee_block + ex_ee_len - 1);
2668                         err = -EIO;
2669                         goto out;
2670                 } else if (a != ex_ee_block) {
2671                         /* remove tail of the extent */
2672                         num = a - ex_ee_block;
2673                 } else {
2674                         /* remove whole extent: excellent! */
2675                         num = 0;
2676                 }
2677                 /*
2678                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2679                  * descriptor) for each block group; assume two block
2680                  * groups plus ex_ee_len/blocks_per_block_group for
2681                  * the worst case
2682                  */
2683                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2684                 if (ex == EXT_FIRST_EXTENT(eh)) {
2685                         correct_index = 1;
2686                         credits += (ext_depth(inode)) + 1;
2687                 }
2688                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2689
2690                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2691                 if (err)
2692                         goto out;
2693
2694                 err = ext4_ext_get_access(handle, inode, path + depth);
2695                 if (err)
2696                         goto out;
2697
2698                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2699                                          a, b);
2700                 if (err)
2701                         goto out;
2702
2703                 if (num == 0)
2704                         /* this extent is removed; mark slot entirely unused */
2705                         ext4_ext_store_pblock(ex, 0);
2706
2707                 ex->ee_len = cpu_to_le16(num);
2708                 /*
2709                  * Do not mark unwritten if all the blocks in the
2710                  * extent have been removed.
2711                  */
2712                 if (unwritten && num)
2713                         ext4_ext_mark_unwritten(ex);
2714                 /*
2715                  * If the extent was completely released,
2716                  * we need to remove it from the leaf
2717                  */
2718                 if (num == 0) {
2719                         if (end != EXT_MAX_BLOCKS - 1) {
2720                                 /*
2721                                  * For hole punching, we need to scoot all the
2722                                  * extents up when an extent is removed so that
2723                                  * we dont have blank extents in the middle
2724                                  */
2725                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2726                                         sizeof(struct ext4_extent));
2727
2728                                 /* Now get rid of the one at the end */
2729                                 memset(EXT_LAST_EXTENT(eh), 0,
2730                                         sizeof(struct ext4_extent));
2731                         }
2732                         le16_add_cpu(&eh->eh_entries, -1);
2733                 }
2734
2735                 err = ext4_ext_dirty(handle, inode, path + depth);
2736                 if (err)
2737                         goto out;
2738
2739                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2740                                 ext4_ext_pblock(ex));
2741                 ex--;
2742                 ex_ee_block = le32_to_cpu(ex->ee_block);
2743                 ex_ee_len = ext4_ext_get_actual_len(ex);
2744         }
2745
2746         if (correct_index && eh->eh_entries)
2747                 err = ext4_ext_correct_indexes(handle, inode, path);
2748
2749         /*
2750          * If there's a partial cluster and at least one extent remains in
2751          * the leaf, free the partial cluster if it isn't shared with the
2752          * current extent.  If it is shared with the current extent
2753          * we zero partial_cluster because we've reached the start of the
2754          * truncated/punched region and we're done removing blocks.
2755          */
2756         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2757                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2758                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2759                         ext4_free_blocks(handle, inode, NULL,
2760                                          EXT4_C2B(sbi, *partial_cluster),
2761                                          sbi->s_cluster_ratio,
2762                                          get_default_free_blocks_flags(inode));
2763                 }
2764                 *partial_cluster = 0;
2765         }
2766
2767         /* if this leaf is free, then we should
2768          * remove it from index block above */
2769         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2770                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2771
2772 out:
2773         return err;
2774 }
2775
2776 /*
2777  * ext4_ext_more_to_rm:
2778  * returns 1 if current index has to be freed (even partial)
2779  */
2780 static int
2781 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2782 {
2783         BUG_ON(path->p_idx == NULL);
2784
2785         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2786                 return 0;
2787
2788         /*
2789          * if truncate on deeper level happened, it wasn't partial,
2790          * so we have to consider current index for truncation
2791          */
2792         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2793                 return 0;
2794         return 1;
2795 }
2796
2797 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2798                           ext4_lblk_t end)
2799 {
2800         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2801         int depth = ext_depth(inode);
2802         struct ext4_ext_path *path = NULL;
2803         long long partial_cluster = 0;
2804         handle_t *handle;
2805         int i = 0, err = 0;
2806
2807         ext_debug("truncate since %u to %u\n", start, end);
2808
2809         /* probably first extent we're gonna free will be last in block */
2810         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2811         if (IS_ERR(handle))
2812                 return PTR_ERR(handle);
2813
2814 again:
2815         trace_ext4_ext_remove_space(inode, start, end, depth);
2816
2817         /*
2818          * Check if we are removing extents inside the extent tree. If that
2819          * is the case, we are going to punch a hole inside the extent tree
2820          * so we have to check whether we need to split the extent covering
2821          * the last block to remove so we can easily remove the part of it
2822          * in ext4_ext_rm_leaf().
2823          */
2824         if (end < EXT_MAX_BLOCKS - 1) {
2825                 struct ext4_extent *ex;
2826                 ext4_lblk_t ee_block, ex_end, lblk;
2827                 ext4_fsblk_t pblk;
2828
2829                 /* find extent for or closest extent to this block */
2830                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2831                 if (IS_ERR(path)) {
2832                         ext4_journal_stop(handle);
2833                         return PTR_ERR(path);
2834                 }
2835                 depth = ext_depth(inode);
2836                 /* Leaf not may not exist only if inode has no blocks at all */
2837                 ex = path[depth].p_ext;
2838                 if (!ex) {
2839                         if (depth) {
2840                                 EXT4_ERROR_INODE(inode,
2841                                                  "path[%d].p_hdr == NULL",
2842                                                  depth);
2843                                 err = -EIO;
2844                         }
2845                         goto out;
2846                 }
2847
2848                 ee_block = le32_to_cpu(ex->ee_block);
2849                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2850
2851                 /*
2852                  * See if the last block is inside the extent, if so split
2853                  * the extent at 'end' block so we can easily remove the
2854                  * tail of the first part of the split extent in
2855                  * ext4_ext_rm_leaf().
2856                  */
2857                 if (end >= ee_block && end < ex_end) {
2858
2859                         /*
2860                          * If we're going to split the extent, note that
2861                          * the cluster containing the block after 'end' is
2862                          * in use to avoid freeing it when removing blocks.
2863                          */
2864                         if (sbi->s_cluster_ratio > 1) {
2865                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2866                                 partial_cluster =
2867                                         -(long long) EXT4_B2C(sbi, pblk);
2868                         }
2869
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                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2882                         /*
2883                          * If there's an extent to the right its first cluster
2884                          * contains the immediate right boundary of the
2885                          * truncated/punched region.  Set partial_cluster to
2886                          * its negative value so it won't be freed if shared
2887                          * with the current extent.  The end < ee_block case
2888                          * is handled in ext4_ext_rm_leaf().
2889                          */
2890                         lblk = ex_end + 1;
2891                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2892                                                     &ex);
2893                         if (err)
2894                                 goto out;
2895                         if (pblk)
2896                                 partial_cluster =
2897                                         -(long long) EXT4_B2C(sbi, pblk);
2898                 }
2899         }
2900         /*
2901          * We start scanning from right side, freeing all the blocks
2902          * after i_size and walking into the tree depth-wise.
2903          */
2904         depth = ext_depth(inode);
2905         if (path) {
2906                 int k = i = depth;
2907                 while (--k > 0)
2908                         path[k].p_block =
2909                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2910         } else {
2911                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2912                                GFP_NOFS);
2913                 if (path == NULL) {
2914                         ext4_journal_stop(handle);
2915                         return -ENOMEM;
2916                 }
2917                 path[0].p_maxdepth = path[0].p_depth = depth;
2918                 path[0].p_hdr = ext_inode_hdr(inode);
2919                 i = 0;
2920
2921                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2922                         err = -EIO;
2923                         goto out;
2924                 }
2925         }
2926         err = 0;
2927
2928         while (i >= 0 && err == 0) {
2929                 if (i == depth) {
2930                         /* this is leaf block */
2931                         err = ext4_ext_rm_leaf(handle, inode, path,
2932                                                &partial_cluster, start,
2933                                                end);
2934                         /* root level has p_bh == NULL, brelse() eats this */
2935                         brelse(path[i].p_bh);
2936                         path[i].p_bh = NULL;
2937                         i--;
2938                         continue;
2939                 }
2940
2941                 /* this is index block */
2942                 if (!path[i].p_hdr) {
2943                         ext_debug("initialize header\n");
2944                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2945                 }
2946
2947                 if (!path[i].p_idx) {
2948                         /* this level hasn't been touched yet */
2949                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2950                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2951                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2952                                   path[i].p_hdr,
2953                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2954                 } else {
2955                         /* we were already here, see at next index */
2956                         path[i].p_idx--;
2957                 }
2958
2959                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2960                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2961                                 path[i].p_idx);
2962                 if (ext4_ext_more_to_rm(path + i)) {
2963                         struct buffer_head *bh;
2964                         /* go to the next level */
2965                         ext_debug("move to level %d (block %llu)\n",
2966                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2967                         memset(path + i + 1, 0, sizeof(*path));
2968                         bh = read_extent_tree_block(inode,
2969                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2970                                 EXT4_EX_NOCACHE);
2971                         if (IS_ERR(bh)) {
2972                                 /* should we reset i_size? */
2973                                 err = PTR_ERR(bh);
2974                                 break;
2975                         }
2976                         /* Yield here to deal with large extent trees.
2977                          * Should be a no-op if we did IO above. */
2978                         cond_resched();
2979                         if (WARN_ON(i + 1 > depth)) {
2980                                 err = -EIO;
2981                                 break;
2982                         }
2983                         path[i + 1].p_bh = bh;
2984
2985                         /* save actual number of indexes since this
2986                          * number is changed at the next iteration */
2987                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2988                         i++;
2989                 } else {
2990                         /* we finished processing this index, go up */
2991                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2992                                 /* index is empty, remove it;
2993                                  * handle must be already prepared by the
2994                                  * truncatei_leaf() */
2995                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2996                         }
2997                         /* root level has p_bh == NULL, brelse() eats this */
2998                         brelse(path[i].p_bh);
2999                         path[i].p_bh = NULL;
3000                         i--;
3001                         ext_debug("return to level %d\n", i);
3002                 }
3003         }
3004
3005         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3006                         partial_cluster, path->p_hdr->eh_entries);
3007
3008         /*
3009          * If we still have something in the partial cluster and we have removed
3010          * even the first extent, then we should free the blocks in the partial
3011          * cluster as well.  (This code will only run when there are no leaves
3012          * to the immediate left of the truncated/punched region.)
3013          */
3014         if (partial_cluster > 0 && err == 0) {
3015                 /* don't zero partial_cluster since it's not used afterwards */
3016                 ext4_free_blocks(handle, inode, NULL,
3017                                  EXT4_C2B(sbi, partial_cluster),
3018                                  sbi->s_cluster_ratio,
3019                                  get_default_free_blocks_flags(inode));
3020         }
3021
3022         /* TODO: flexible tree reduction should be here */
3023         if (path->p_hdr->eh_entries == 0) {
3024                 /*
3025                  * truncate to zero freed all the tree,
3026                  * so we need to correct eh_depth
3027                  */
3028                 err = ext4_ext_get_access(handle, inode, path);
3029                 if (err == 0) {
3030                         ext_inode_hdr(inode)->eh_depth = 0;
3031                         ext_inode_hdr(inode)->eh_max =
3032                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3033                         err = ext4_ext_dirty(handle, inode, path);
3034                 }
3035         }
3036 out:
3037         ext4_ext_drop_refs(path);
3038         kfree(path);
3039         path = NULL;
3040         if (err == -EAGAIN)
3041                 goto again;
3042         ext4_journal_stop(handle);
3043
3044         return err;
3045 }
3046
3047 /*
3048  * called at mount time
3049  */
3050 void ext4_ext_init(struct super_block *sb)
3051 {
3052         /*
3053          * possible initialization would be here
3054          */
3055
3056         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
3057 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3058                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3059 #ifdef AGGRESSIVE_TEST
3060                        ", aggressive tests"
3061 #endif
3062 #ifdef CHECK_BINSEARCH
3063                        ", check binsearch"
3064 #endif
3065 #ifdef EXTENTS_STATS
3066                        ", stats"
3067 #endif
3068                        "\n");
3069 #endif
3070 #ifdef EXTENTS_STATS
3071                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3072                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3073                 EXT4_SB(sb)->s_ext_max = 0;
3074 #endif
3075         }
3076 }
3077
3078 /*
3079  * called at umount time
3080  */
3081 void ext4_ext_release(struct super_block *sb)
3082 {
3083         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
3084                 return;
3085
3086 #ifdef EXTENTS_STATS
3087         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3088                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3089                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3090                         sbi->s_ext_blocks, sbi->s_ext_extents,
3091                         sbi->s_ext_blocks / sbi->s_ext_extents);
3092                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3093                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3094         }
3095 #endif
3096 }
3097
3098 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3099 {
3100         ext4_lblk_t  ee_block;
3101         ext4_fsblk_t ee_pblock;
3102         unsigned int ee_len;
3103
3104         ee_block  = le32_to_cpu(ex->ee_block);
3105         ee_len    = ext4_ext_get_actual_len(ex);
3106         ee_pblock = ext4_ext_pblock(ex);
3107
3108         if (ee_len == 0)
3109                 return 0;
3110
3111         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3112                                      EXTENT_STATUS_WRITTEN);
3113 }
3114
3115 /* FIXME!! we need to try to merge to left or right after zero-out  */
3116 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3117 {
3118         ext4_fsblk_t ee_pblock;
3119         unsigned int ee_len;
3120         int ret;
3121
3122         ee_len    = ext4_ext_get_actual_len(ex);
3123         ee_pblock = ext4_ext_pblock(ex);
3124
3125         if (ext4_encrypted_inode(inode))
3126                 return ext4_encrypted_zeroout(inode, ex);
3127
3128         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
3129         if (ret > 0)
3130                 ret = 0;
3131
3132         return ret;
3133 }
3134
3135 /*
3136  * ext4_split_extent_at() splits an extent at given block.
3137  *
3138  * @handle: the journal handle
3139  * @inode: the file inode
3140  * @path: the path to the extent
3141  * @split: the logical block where the extent is splitted.
3142  * @split_flags: indicates if the extent could be zeroout if split fails, and
3143  *               the states(init or unwritten) of new extents.
3144  * @flags: flags used to insert new extent to extent tree.
3145  *
3146  *
3147  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3148  * of which are deterimined by split_flag.
3149  *
3150  * There are two cases:
3151  *  a> the extent are splitted into two extent.
3152  *  b> split is not needed, and just mark the extent.
3153  *
3154  * return 0 on success.
3155  */
3156 static int ext4_split_extent_at(handle_t *handle,
3157                              struct inode *inode,
3158                              struct ext4_ext_path **ppath,
3159                              ext4_lblk_t split,
3160                              int split_flag,
3161                              int flags)
3162 {
3163         struct ext4_ext_path *path = *ppath;
3164         ext4_fsblk_t newblock;
3165         ext4_lblk_t ee_block;
3166         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3167         struct ext4_extent *ex2 = NULL;
3168         unsigned int ee_len, depth;
3169         int err = 0;
3170
3171         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3172                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3173
3174         ext_debug("ext4_split_extents_at: inode %lu, logical"
3175                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3176
3177         ext4_ext_show_leaf(inode, path);
3178
3179         depth = ext_depth(inode);
3180         ex = path[depth].p_ext;
3181         ee_block = le32_to_cpu(ex->ee_block);
3182         ee_len = ext4_ext_get_actual_len(ex);
3183         newblock = split - ee_block + ext4_ext_pblock(ex);
3184
3185         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3186         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3187                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3188                              EXT4_EXT_MARK_UNWRIT1 |
3189                              EXT4_EXT_MARK_UNWRIT2));
3190
3191         err = ext4_ext_get_access(handle, inode, path + depth);
3192         if (err)
3193                 goto out;
3194
3195         if (split == ee_block) {
3196                 /*
3197                  * case b: block @split is the block that the extent begins with
3198                  * then we just change the state of the extent, and splitting
3199                  * is not needed.
3200                  */
3201                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3202                         ext4_ext_mark_unwritten(ex);
3203                 else
3204                         ext4_ext_mark_initialized(ex);
3205
3206                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3207                         ext4_ext_try_to_merge(handle, inode, path, ex);
3208
3209                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3210                 goto out;
3211         }
3212
3213         /* case a */
3214         memcpy(&orig_ex, ex, sizeof(orig_ex));
3215         ex->ee_len = cpu_to_le16(split - ee_block);
3216         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3217                 ext4_ext_mark_unwritten(ex);
3218
3219         /*
3220          * path may lead to new leaf, not to original leaf any more
3221          * after ext4_ext_insert_extent() returns,
3222          */
3223         err = ext4_ext_dirty(handle, inode, path + depth);
3224         if (err)
3225                 goto fix_extent_len;
3226
3227         ex2 = &newex;
3228         ex2->ee_block = cpu_to_le32(split);
3229         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3230         ext4_ext_store_pblock(ex2, newblock);
3231         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3232                 ext4_ext_mark_unwritten(ex2);
3233
3234         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3235         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3236                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3237                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3238                                 err = ext4_ext_zeroout(inode, ex2);
3239                                 zero_ex.ee_block = ex2->ee_block;
3240                                 zero_ex.ee_len = cpu_to_le16(
3241                                                 ext4_ext_get_actual_len(ex2));
3242                                 ext4_ext_store_pblock(&zero_ex,
3243                                                       ext4_ext_pblock(ex2));
3244                         } else {
3245                                 err = ext4_ext_zeroout(inode, ex);
3246                                 zero_ex.ee_block = ex->ee_block;
3247                                 zero_ex.ee_len = cpu_to_le16(
3248                                                 ext4_ext_get_actual_len(ex));
3249                                 ext4_ext_store_pblock(&zero_ex,
3250                                                       ext4_ext_pblock(ex));
3251                         }
3252                 } else {
3253                         err = ext4_ext_zeroout(inode, &orig_ex);
3254                         zero_ex.ee_block = orig_ex.ee_block;
3255                         zero_ex.ee_len = cpu_to_le16(
3256                                                 ext4_ext_get_actual_len(&orig_ex));
3257                         ext4_ext_store_pblock(&zero_ex,
3258                                               ext4_ext_pblock(&orig_ex));
3259                 }
3260
3261                 if (err)
3262                         goto fix_extent_len;
3263                 /* update the extent length and mark as initialized */
3264                 ex->ee_len = cpu_to_le16(ee_len);
3265                 ext4_ext_try_to_merge(handle, inode, path, ex);
3266                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3267                 if (err)
3268                         goto fix_extent_len;
3269
3270                 /* update extent status tree */
3271                 err = ext4_zeroout_es(inode, &zero_ex);
3272
3273                 goto out;
3274         } else if (err)
3275                 goto fix_extent_len;
3276
3277 out:
3278         ext4_ext_show_leaf(inode, path);
3279         return err;
3280
3281 fix_extent_len:
3282         ex->ee_len = orig_ex.ee_len;
3283         ext4_ext_dirty(handle, inode, path + path->p_depth);
3284         return err;
3285 }
3286
3287 /*
3288  * ext4_split_extents() splits an extent and mark extent which is covered
3289  * by @map as split_flags indicates
3290  *
3291  * It may result in splitting the extent into multiple extents (up to three)
3292  * There are three possibilities:
3293  *   a> There is no split required
3294  *   b> Splits in two extents: Split is happening at either end of the extent
3295  *   c> Splits in three extents: Somone is splitting in middle of the extent
3296  *
3297  */
3298 static int ext4_split_extent(handle_t *handle,
3299                               struct inode *inode,
3300                               struct ext4_ext_path **ppath,
3301                               struct ext4_map_blocks *map,
3302                               int split_flag,
3303                               int flags)
3304 {
3305         struct ext4_ext_path *path = *ppath;
3306         ext4_lblk_t ee_block;
3307         struct ext4_extent *ex;
3308         unsigned int ee_len, depth;
3309         int err = 0;
3310         int unwritten;
3311         int split_flag1, flags1;
3312         int allocated = map->m_len;
3313
3314         depth = ext_depth(inode);
3315         ex = path[depth].p_ext;
3316         ee_block = le32_to_cpu(ex->ee_block);
3317         ee_len = ext4_ext_get_actual_len(ex);
3318         unwritten = ext4_ext_is_unwritten(ex);
3319
3320         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3321                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3322                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3323                 if (unwritten)
3324                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3325                                        EXT4_EXT_MARK_UNWRIT2;
3326                 if (split_flag & EXT4_EXT_DATA_VALID2)
3327                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3328                 err = ext4_split_extent_at(handle, inode, ppath,
3329                                 map->m_lblk + map->m_len, split_flag1, flags1);
3330                 if (err)
3331                         goto out;
3332         } else {
3333                 allocated = ee_len - (map->m_lblk - ee_block);
3334         }
3335         /*
3336          * Update path is required because previous ext4_split_extent_at() may
3337          * result in split of original leaf or extent zeroout.
3338          */
3339         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3340         if (IS_ERR(path))
3341                 return PTR_ERR(path);
3342         depth = ext_depth(inode);
3343         ex = path[depth].p_ext;
3344         if (!ex) {
3345                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3346                                  (unsigned long) map->m_lblk);
3347                 return -EIO;
3348         }
3349         unwritten = ext4_ext_is_unwritten(ex);
3350         split_flag1 = 0;
3351
3352         if (map->m_lblk >= ee_block) {
3353                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3354                 if (unwritten) {
3355                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3356                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3357                                                      EXT4_EXT_MARK_UNWRIT2);
3358                 }
3359                 err = ext4_split_extent_at(handle, inode, ppath,
3360                                 map->m_lblk, split_flag1, flags);
3361                 if (err)
3362                         goto out;
3363         }
3364
3365         ext4_ext_show_leaf(inode, path);
3366 out:
3367         return err ? err : allocated;
3368 }
3369
3370 /*
3371  * This function is called by ext4_ext_map_blocks() if someone tries to write
3372  * to an unwritten extent. It may result in splitting the unwritten
3373  * extent into multiple extents (up to three - one initialized and two
3374  * unwritten).
3375  * There are three possibilities:
3376  *   a> There is no split required: Entire extent should be initialized
3377  *   b> Splits in two extents: Write is happening at either end of the extent
3378  *   c> Splits in three extents: Somone is writing in middle of the extent
3379  *
3380  * Pre-conditions:
3381  *  - The extent pointed to by 'path' is unwritten.
3382  *  - The extent pointed to by 'path' contains a superset
3383  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3384  *
3385  * Post-conditions on success:
3386  *  - the returned value is the number of blocks beyond map->l_lblk
3387  *    that are allocated and initialized.
3388  *    It is guaranteed to be >= map->m_len.
3389  */
3390 static int ext4_ext_convert_to_initialized(handle_t *handle,
3391                                            struct inode *inode,
3392                                            struct ext4_map_blocks *map,
3393                                            struct ext4_ext_path **ppath,
3394                                            int flags)
3395 {
3396         struct ext4_ext_path *path = *ppath;
3397         struct ext4_sb_info *sbi;
3398         struct ext4_extent_header *eh;
3399         struct ext4_map_blocks split_map;
3400         struct ext4_extent zero_ex;
3401         struct ext4_extent *ex, *abut_ex;
3402         ext4_lblk_t ee_block, eof_block;
3403         unsigned int ee_len, depth, map_len = map->m_len;
3404         int allocated = 0, max_zeroout = 0;
3405         int err = 0;
3406         int split_flag = 0;
3407
3408         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3409                 "block %llu, max_blocks %u\n", inode->i_ino,
3410                 (unsigned long long)map->m_lblk, map_len);
3411
3412         sbi = EXT4_SB(inode->i_sb);
3413         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3414                 inode->i_sb->s_blocksize_bits;
3415         if (eof_block < map->m_lblk + map_len)
3416                 eof_block = map->m_lblk + map_len;
3417
3418         depth = ext_depth(inode);
3419         eh = path[depth].p_hdr;
3420         ex = path[depth].p_ext;
3421         ee_block = le32_to_cpu(ex->ee_block);
3422         ee_len = ext4_ext_get_actual_len(ex);
3423         zero_ex.ee_len = 0;
3424
3425         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3426
3427         /* Pre-conditions */
3428         BUG_ON(!ext4_ext_is_unwritten(ex));
3429         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3430
3431         /*
3432          * Attempt to transfer newly initialized blocks from the currently
3433          * unwritten extent to its neighbor. This is much cheaper
3434          * than an insertion followed by a merge as those involve costly
3435          * memmove() calls. Transferring to the left is the common case in
3436          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3437          * followed by append writes.
3438          *
3439          * Limitations of the current logic:
3440          *  - L1: we do not deal with writes covering the whole extent.
3441          *    This would require removing the extent if the transfer
3442          *    is possible.
3443          *  - L2: we only attempt to merge with an extent stored in the
3444          *    same extent tree node.
3445          */
3446         if ((map->m_lblk == ee_block) &&
3447                 /* See if we can merge left */
3448                 (map_len < ee_len) &&           /*L1*/
3449                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3450                 ext4_lblk_t prev_lblk;
3451                 ext4_fsblk_t prev_pblk, ee_pblk;
3452                 unsigned int prev_len;
3453
3454                 abut_ex = ex - 1;
3455                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3456                 prev_len = ext4_ext_get_actual_len(abut_ex);
3457                 prev_pblk = ext4_ext_pblock(abut_ex);
3458                 ee_pblk = ext4_ext_pblock(ex);
3459
3460                 /*
3461                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3462                  * upon those conditions:
3463                  * - C1: abut_ex is initialized,
3464                  * - C2: abut_ex is logically abutting ex,
3465                  * - C3: abut_ex is physically abutting ex,
3466                  * - C4: abut_ex can receive the additional blocks without
3467                  *   overflowing the (initialized) length limit.
3468                  */
3469                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3470                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3471                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3472                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3473                         err = ext4_ext_get_access(handle, inode, path + depth);
3474                         if (err)
3475                                 goto out;
3476
3477                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3478                                 map, ex, abut_ex);
3479
3480                         /* Shift the start of ex by 'map_len' blocks */
3481                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3482                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3483                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3484                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3485
3486                         /* Extend abut_ex by 'map_len' blocks */
3487                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3488
3489                         /* Result: number of initialized blocks past m_lblk */
3490                         allocated = map_len;
3491                 }
3492         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3493                    (map_len < ee_len) &&        /*L1*/
3494                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3495                 /* See if we can merge right */
3496                 ext4_lblk_t next_lblk;
3497                 ext4_fsblk_t next_pblk, ee_pblk;
3498                 unsigned int next_len;
3499
3500                 abut_ex = ex + 1;
3501                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3502                 next_len = ext4_ext_get_actual_len(abut_ex);
3503                 next_pblk = ext4_ext_pblock(abut_ex);
3504                 ee_pblk = ext4_ext_pblock(ex);
3505
3506                 /*
3507                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3508                  * upon those conditions:
3509                  * - C1: abut_ex is initialized,
3510                  * - C2: abut_ex is logically abutting ex,
3511                  * - C3: abut_ex is physically abutting ex,
3512                  * - C4: abut_ex can receive the additional blocks without
3513                  *   overflowing the (initialized) length limit.
3514                  */
3515                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3516                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3517                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3518                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3519                         err = ext4_ext_get_access(handle, inode, path + depth);
3520                         if (err)
3521                                 goto out;
3522
3523                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3524                                 map, ex, abut_ex);
3525
3526                         /* Shift the start of abut_ex by 'map_len' blocks */
3527                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3528                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3529                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3530                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3531
3532                         /* Extend abut_ex by 'map_len' blocks */
3533                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3534
3535                         /* Result: number of initialized blocks past m_lblk */
3536                         allocated = map_len;
3537                 }
3538         }
3539         if (allocated) {
3540                 /* Mark the block containing both extents as dirty */
3541                 ext4_ext_dirty(handle, inode, path + depth);
3542
3543                 /* Update path to point to the right extent */
3544                 path[depth].p_ext = abut_ex;
3545                 goto out;
3546         } else
3547                 allocated = ee_len - (map->m_lblk - ee_block);
3548
3549         WARN_ON(map->m_lblk < ee_block);
3550         /*
3551          * It is safe to convert extent to initialized via explicit
3552          * zeroout only if extent is fully inside i_size or new_size.
3553          */
3554         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3555
3556         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3557                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3558                         (inode->i_sb->s_blocksize_bits - 10);
3559
3560         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3561         if (max_zeroout && (ee_len <= max_zeroout)) {
3562                 err = ext4_ext_zeroout(inode, ex);
3563                 if (err)
3564                         goto out;
3565                 zero_ex.ee_block = ex->ee_block;
3566                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3567                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3568
3569                 err = ext4_ext_get_access(handle, inode, path + depth);
3570                 if (err)
3571                         goto out;
3572                 ext4_ext_mark_initialized(ex);
3573                 ext4_ext_try_to_merge(handle, inode, path, ex);
3574                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3575                 goto out;
3576         }
3577
3578         /*
3579          * four cases:
3580          * 1. split the extent into three extents.
3581          * 2. split the extent into two extents, zeroout the first half.
3582          * 3. split the extent into two extents, zeroout the second half.
3583          * 4. split the extent into two extents with out zeroout.
3584          */
3585         split_map.m_lblk = map->m_lblk;
3586         split_map.m_len = map->m_len;
3587
3588         if (max_zeroout && (allocated > map->m_len)) {
3589                 if (allocated <= max_zeroout) {
3590                         /* case 3 */
3591                         zero_ex.ee_block =
3592                                          cpu_to_le32(map->m_lblk);
3593                         zero_ex.ee_len = cpu_to_le16(allocated);
3594                         ext4_ext_store_pblock(&zero_ex,
3595                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3596                         err = ext4_ext_zeroout(inode, &zero_ex);
3597                         if (err)
3598                                 goto out;
3599                         split_map.m_lblk = map->m_lblk;
3600                         split_map.m_len = allocated;
3601                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3602                         /* case 2 */
3603                         if (map->m_lblk != ee_block) {
3604                                 zero_ex.ee_block = ex->ee_block;
3605                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3606                                                         ee_block);
3607                                 ext4_ext_store_pblock(&zero_ex,
3608                                                       ext4_ext_pblock(ex));
3609                                 err = ext4_ext_zeroout(inode, &zero_ex);
3610                                 if (err)
3611                                         goto out;
3612                         }
3613
3614                         split_map.m_lblk = ee_block;
3615                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3616                         allocated = map->m_len;
3617                 }
3618         }
3619
3620         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3621                                 flags);
3622         if (err > 0)
3623                 err = 0;
3624 out:
3625         /* If we have gotten a failure, don't zero out status tree */
3626         if (!err)
3627                 err = ext4_zeroout_es(inode, &zero_ex);
3628         return err ? err : allocated;
3629 }
3630
3631 /*
3632  * This function is called by ext4_ext_map_blocks() from
3633  * ext4_get_blocks_dio_write() when DIO to write
3634  * to an unwritten extent.
3635  *
3636  * Writing to an unwritten extent may result in splitting the unwritten
3637  * extent into multiple initialized/unwritten extents (up to three)
3638  * There are three possibilities:
3639  *   a> There is no split required: Entire extent should be unwritten
3640  *   b> Splits in two extents: Write is happening at either end of the extent
3641  *   c> Splits in three extents: Somone is writing in middle of the extent
3642  *
3643  * This works the same way in the case of initialized -> unwritten conversion.
3644  *
3645  * One of more index blocks maybe needed if the extent tree grow after
3646  * the unwritten extent split. To prevent ENOSPC occur at the IO
3647  * complete, we need to split the unwritten extent before DIO submit
3648  * the IO. The unwritten extent called at this time will be split
3649  * into three unwritten extent(at most). After IO complete, the part
3650  * being filled will be convert to initialized by the end_io callback function
3651  * via ext4_convert_unwritten_extents().
3652  *
3653  * Returns the size of unwritten extent to be written on success.
3654  */
3655 static int ext4_split_convert_extents(handle_t *handle,
3656                                         struct inode *inode,
3657                                         struct ext4_map_blocks *map,
3658                                         struct ext4_ext_path **ppath,
3659                                         int flags)
3660 {
3661         struct ext4_ext_path *path = *ppath;
3662         ext4_lblk_t eof_block;
3663         ext4_lblk_t ee_block;
3664         struct ext4_extent *ex;
3665         unsigned int ee_len;
3666         int split_flag = 0, depth;
3667
3668         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3669                   __func__, inode->i_ino,
3670                   (unsigned long long)map->m_lblk, map->m_len);
3671
3672         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3673                 inode->i_sb->s_blocksize_bits;
3674         if (eof_block < map->m_lblk + map->m_len)
3675                 eof_block = map->m_lblk + map->m_len;
3676         /*
3677          * It is safe to convert extent to initialized via explicit
3678          * zeroout only if extent is fully insde i_size or new_size.
3679          */
3680         depth = ext_depth(inode);
3681         ex = path[depth].p_ext;
3682         ee_block = le32_to_cpu(ex->ee_block);
3683         ee_len = ext4_ext_get_actual_len(ex);
3684
3685         /* Convert to unwritten */
3686         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3687                 split_flag |= EXT4_EXT_DATA_VALID1;
3688         /* Convert to initialized */
3689         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3690                 split_flag |= ee_block + ee_len <= eof_block ?
3691                               EXT4_EXT_MAY_ZEROOUT : 0;
3692                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3693         }
3694         flags |= EXT4_GET_BLOCKS_PRE_IO;
3695         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3696 }
3697
3698 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3699                                                 struct inode *inode,
3700                                                 struct ext4_map_blocks *map,
3701                                                 struct ext4_ext_path **ppath)
3702 {
3703         struct ext4_ext_path *path = *ppath;
3704         struct ext4_extent *ex;
3705         ext4_lblk_t ee_block;
3706         unsigned int ee_len;
3707         int depth;
3708         int err = 0;
3709
3710         depth = ext_depth(inode);
3711         ex = path[depth].p_ext;
3712         ee_block = le32_to_cpu(ex->ee_block);
3713         ee_len = ext4_ext_get_actual_len(ex);
3714
3715         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3716                 "block %llu, max_blocks %u\n", inode->i_ino,
3717                   (unsigned long long)ee_block, ee_len);
3718
3719         /* If extent is larger than requested it is a clear sign that we still
3720          * have some extent state machine issues left. So extent_split is still
3721          * required.
3722          * TODO: Once all related issues will be fixed this situation should be
3723          * illegal.
3724          */
3725         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3726 #ifdef EXT4_DEBUG
3727                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3728                              " len %u; IO logical block %llu, len %u\n",
3729                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3730                              (unsigned long long)map->m_lblk, map->m_len);
3731 #endif
3732                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3733                                                  EXT4_GET_BLOCKS_CONVERT);
3734                 if (err < 0)
3735                         return err;
3736                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3737                 if (IS_ERR(path))
3738                         return PTR_ERR(path);
3739                 depth = ext_depth(inode);
3740                 ex = path[depth].p_ext;
3741         }
3742
3743         err = ext4_ext_get_access(handle, inode, path + depth);
3744         if (err)
3745                 goto out;
3746         /* first mark the extent as initialized */
3747         ext4_ext_mark_initialized(ex);
3748
3749         /* note: ext4_ext_correct_indexes() isn't needed here because
3750          * borders are not changed
3751          */
3752         ext4_ext_try_to_merge(handle, inode, path, ex);
3753
3754         /* Mark modified extent as dirty */
3755         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3756 out:
3757         ext4_ext_show_leaf(inode, path);
3758         return err;
3759 }
3760
3761 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3762                         sector_t block, int count)
3763 {
3764         int i;
3765         for (i = 0; i < count; i++)
3766                 unmap_underlying_metadata(bdev, block + i);
3767 }
3768
3769 /*
3770  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3771  */
3772 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3773                               ext4_lblk_t lblk,
3774                               struct ext4_ext_path *path,
3775                               unsigned int len)
3776 {
3777         int i, depth;
3778         struct ext4_extent_header *eh;
3779         struct ext4_extent *last_ex;
3780
3781         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3782                 return 0;
3783
3784         depth = ext_depth(inode);
3785         eh = path[depth].p_hdr;
3786
3787         /*
3788          * We're going to remove EOFBLOCKS_FL entirely in future so we
3789          * do not care for this case anymore. Simply remove the flag
3790          * if there are no extents.
3791          */
3792         if (unlikely(!eh->eh_entries))
3793                 goto out;
3794         last_ex = EXT_LAST_EXTENT(eh);
3795         /*
3796          * We should clear the EOFBLOCKS_FL flag if we are writing the
3797          * last block in the last extent in the file.  We test this by
3798          * first checking to see if the caller to
3799          * ext4_ext_get_blocks() was interested in the last block (or
3800          * a block beyond the last block) in the current extent.  If
3801          * this turns out to be false, we can bail out from this
3802          * function immediately.
3803          */
3804         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3805             ext4_ext_get_actual_len(last_ex))
3806                 return 0;
3807         /*
3808          * If the caller does appear to be planning to write at or
3809          * beyond the end of the current extent, we then test to see
3810          * if the current extent is the last extent in the file, by
3811          * checking to make sure it was reached via the rightmost node
3812          * at each level of the tree.
3813          */
3814         for (i = depth-1; i >= 0; i--)
3815                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3816                         return 0;
3817 out:
3818         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3819         return ext4_mark_inode_dirty(handle, inode);
3820 }
3821
3822 /**
3823  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3824  *
3825  * Return 1 if there is a delalloc block in the range, otherwise 0.
3826  */
3827 int ext4_find_delalloc_range(struct inode *inode,
3828                              ext4_lblk_t lblk_start,
3829                              ext4_lblk_t lblk_end)
3830 {
3831         struct extent_status es;
3832
3833         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3834         if (es.es_len == 0)
3835                 return 0; /* there is no delay extent in this tree */
3836         else if (es.es_lblk <= lblk_start &&
3837                  lblk_start < es.es_lblk + es.es_len)
3838                 return 1;
3839         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3840                 return 1;
3841         else
3842                 return 0;
3843 }
3844
3845 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3846 {
3847         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3848         ext4_lblk_t lblk_start, lblk_end;
3849         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3850         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3851
3852         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3853 }
3854
3855 /**
3856  * Determines how many complete clusters (out of those specified by the 'map')
3857  * are under delalloc and were reserved quota for.
3858  * This function is called when we are writing out the blocks that were
3859  * originally written with their allocation delayed, but then the space was
3860  * allocated using fallocate() before the delayed allocation could be resolved.
3861  * The cases to look for are:
3862  * ('=' indicated delayed allocated blocks
3863  *  '-' indicates non-delayed allocated blocks)
3864  * (a) partial clusters towards beginning and/or end outside of allocated range
3865  *     are not delalloc'ed.
3866  *      Ex:
3867  *      |----c---=|====c====|====c====|===-c----|
3868  *               |++++++ allocated ++++++|
3869  *      ==> 4 complete clusters in above example
3870  *
3871  * (b) partial cluster (outside of allocated range) towards either end is
3872  *     marked for delayed allocation. In this case, we will exclude that
3873  *     cluster.
3874  *      Ex:
3875  *      |----====c========|========c========|
3876  *           |++++++ allocated ++++++|
3877  *      ==> 1 complete clusters in above example
3878  *
3879  *      Ex:
3880  *      |================c================|
3881  *            |++++++ allocated ++++++|
3882  *      ==> 0 complete clusters in above example
3883  *
3884  * The ext4_da_update_reserve_space will be called only if we
3885  * determine here that there were some "entire" clusters that span
3886  * this 'allocated' range.
3887  * In the non-bigalloc case, this function will just end up returning num_blks
3888  * without ever calling ext4_find_delalloc_range.
3889  */
3890 static unsigned int
3891 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3892                            unsigned int num_blks)
3893 {
3894         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3895         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3896         ext4_lblk_t lblk_from, lblk_to, c_offset;
3897         unsigned int allocated_clusters = 0;
3898
3899         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3900         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3901
3902         /* max possible clusters for this allocation */
3903         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3904
3905         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3906
3907         /* Check towards left side */
3908         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3909         if (c_offset) {
3910                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3911                 lblk_to = lblk_from + c_offset - 1;
3912
3913                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3914                         allocated_clusters--;
3915         }
3916
3917         /* Now check towards right. */
3918         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3919         if (allocated_clusters && c_offset) {
3920                 lblk_from = lblk_start + num_blks;
3921                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3922
3923                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3924                         allocated_clusters--;
3925         }
3926
3927         return allocated_clusters;
3928 }
3929
3930 static int
3931 convert_initialized_extent(handle_t *handle, struct inode *inode,
3932                            struct ext4_map_blocks *map,
3933                            struct ext4_ext_path **ppath, int flags,
3934                            unsigned int allocated, ext4_fsblk_t newblock)
3935 {
3936         struct ext4_ext_path *path = *ppath;
3937         struct ext4_extent *ex;
3938         ext4_lblk_t ee_block;
3939         unsigned int ee_len;
3940         int depth;
3941         int err = 0;
3942
3943         /*
3944          * Make sure that the extent is no bigger than we support with
3945          * unwritten extent
3946          */
3947         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3948                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3949
3950         depth = ext_depth(inode);
3951         ex = path[depth].p_ext;
3952         ee_block = le32_to_cpu(ex->ee_block);
3953         ee_len = ext4_ext_get_actual_len(ex);
3954
3955         ext_debug("%s: inode %lu, logical"
3956                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3957                   (unsigned long long)ee_block, ee_len);
3958
3959         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3960                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3961                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3962                 if (err < 0)
3963                         return err;
3964                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3965                 if (IS_ERR(path))
3966                         return PTR_ERR(path);
3967                 depth = ext_depth(inode);
3968                 ex = path[depth].p_ext;
3969                 if (!ex) {
3970                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3971                                          (unsigned long) map->m_lblk);
3972                         return -EIO;
3973                 }
3974         }
3975
3976         err = ext4_ext_get_access(handle, inode, path + depth);
3977         if (err)
3978                 return err;
3979         /* first mark the extent as unwritten */
3980         ext4_ext_mark_unwritten(ex);
3981
3982         /* note: ext4_ext_correct_indexes() isn't needed here because
3983          * borders are not changed
3984          */
3985         ext4_ext_try_to_merge(handle, inode, path, ex);
3986
3987         /* Mark modified extent as dirty */
3988         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3989         if (err)
3990                 return err;
3991         ext4_ext_show_leaf(inode, path);
3992
3993         ext4_update_inode_fsync_trans(handle, inode, 1);
3994         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3995         if (err)
3996                 return err;
3997         map->m_flags |= EXT4_MAP_UNWRITTEN;
3998         if (allocated > map->m_len)
3999                 allocated = map->m_len;
4000         map->m_len = allocated;
4001         return allocated;
4002 }
4003
4004 static int
4005 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4006                         struct ext4_map_blocks *map,
4007                         struct ext4_ext_path **ppath, int flags,
4008                         unsigned int allocated, ext4_fsblk_t newblock)
4009 {
4010         struct ext4_ext_path *path = *ppath;
4011         int ret = 0;
4012         int err = 0;
4013         ext4_io_end_t *io = ext4_inode_aio(inode);
4014
4015         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4016                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4017                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4018                   flags, allocated);
4019         ext4_ext_show_leaf(inode, path);
4020
4021         /*
4022          * When writing into unwritten space, we should not fail to
4023          * allocate metadata blocks for the new extent block if needed.
4024          */
4025         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4026
4027         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4028                                                     allocated, newblock);
4029
4030         /* get_block() before submit the IO, split the extent */
4031         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4032                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4033                                          flags | EXT4_GET_BLOCKS_CONVERT);
4034                 if (ret <= 0)
4035                         goto out;
4036                 /*
4037                  * Flag the inode(non aio case) or end_io struct (aio case)
4038                  * that this IO needs to conversion to written when IO is
4039                  * completed
4040                  */
4041                 if (io)
4042                         ext4_set_io_unwritten_flag(inode, io);
4043                 else
4044                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4045                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4046                 goto out;
4047         }
4048         /* IO end_io complete, convert the filled extent to written */
4049         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4050                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4051                                                            ppath);
4052                 if (ret >= 0) {
4053                         ext4_update_inode_fsync_trans(handle, inode, 1);
4054                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4055                                                  path, map->m_len);
4056                 } else
4057                         err = ret;
4058                 map->m_flags |= EXT4_MAP_MAPPED;
4059                 map->m_pblk = newblock;
4060                 if (allocated > map->m_len)
4061                         allocated = map->m_len;
4062                 map->m_len = allocated;
4063                 goto out2;
4064         }
4065         /* buffered IO case */
4066         /*
4067          * repeat fallocate creation request
4068          * we already have an unwritten extent
4069          */
4070         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4071                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4072                 goto map_out;
4073         }
4074
4075         /* buffered READ or buffered write_begin() lookup */
4076         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4077                 /*
4078                  * We have blocks reserved already.  We
4079                  * return allocated blocks so that delalloc
4080                  * won't do block reservation for us.  But
4081                  * the buffer head will be unmapped so that
4082                  * a read from the block returns 0s.
4083                  */
4084                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4085                 goto out1;
4086         }
4087
4088         /* buffered write, writepage time, convert*/
4089         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4090         if (ret >= 0)
4091                 ext4_update_inode_fsync_trans(handle, inode, 1);
4092 out:
4093         if (ret <= 0) {
4094                 err = ret;
4095                 goto out2;
4096         } else
4097                 allocated = ret;
4098         map->m_flags |= EXT4_MAP_NEW;
4099         /*
4100          * if we allocated more blocks than requested
4101          * we need to make sure we unmap the extra block
4102          * allocated. The actual needed block will get
4103          * unmapped later when we find the buffer_head marked
4104          * new.
4105          */
4106         if (allocated > map->m_len) {
4107                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
4108                                         newblock + map->m_len,
4109                                         allocated - map->m_len);
4110                 allocated = map->m_len;
4111         }
4112         map->m_len = allocated;
4113
4114         /*
4115          * If we have done fallocate with the offset that is already
4116          * delayed allocated, we would have block reservation
4117          * and quota reservation done in the delayed write path.
4118          * But fallocate would have already updated quota and block
4119          * count for this offset. So cancel these reservation
4120          */
4121         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4122                 unsigned int reserved_clusters;
4123                 reserved_clusters = get_reserved_cluster_alloc(inode,
4124                                 map->m_lblk, map->m_len);
4125                 if (reserved_clusters)
4126                         ext4_da_update_reserve_space(inode,
4127                                                      reserved_clusters,
4128                                                      0);
4129         }
4130
4131 map_out:
4132         map->m_flags |= EXT4_MAP_MAPPED;
4133         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4134                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4135                                          map->m_len);
4136                 if (err < 0)
4137                         goto out2;
4138         }
4139 out1:
4140         if (allocated > map->m_len)
4141                 allocated = map->m_len;
4142         ext4_ext_show_leaf(inode, path);
4143         map->m_pblk = newblock;
4144         map->m_len = allocated;
4145 out2:
4146         return err ? err : allocated;
4147 }
4148
4149 /*
4150  * get_implied_cluster_alloc - check to see if the requested
4151  * allocation (in the map structure) overlaps with a cluster already
4152  * allocated in an extent.
4153  *      @sb     The filesystem superblock structure
4154  *      @map    The requested lblk->pblk mapping
4155  *      @ex     The extent structure which might contain an implied
4156  *                      cluster allocation
4157  *
4158  * This function is called by ext4_ext_map_blocks() after we failed to
4159  * find blocks that were already in the inode's extent tree.  Hence,
4160  * we know that the beginning of the requested region cannot overlap
4161  * the extent from the inode's extent tree.  There are three cases we
4162  * want to catch.  The first is this case:
4163  *
4164  *               |--- cluster # N--|
4165  *    |--- extent ---|  |---- requested region ---|
4166  *                      |==========|
4167  *
4168  * The second case that we need to test for is this one:
4169  *
4170  *   |--------- cluster # N ----------------|
4171  *         |--- requested region --|   |------- extent ----|
4172  *         |=======================|
4173  *
4174  * The third case is when the requested region lies between two extents
4175  * within the same cluster:
4176  *          |------------- cluster # N-------------|
4177  * |----- ex -----|                  |---- ex_right ----|
4178  *                  |------ requested region ------|
4179  *                  |================|
4180  *
4181  * In each of the above cases, we need to set the map->m_pblk and
4182  * map->m_len so it corresponds to the return the extent labelled as
4183  * "|====|" from cluster #N, since it is already in use for data in
4184  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4185  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4186  * as a new "allocated" block region.  Otherwise, we will return 0 and
4187  * ext4_ext_map_blocks() will then allocate one or more new clusters
4188  * by calling ext4_mb_new_blocks().
4189  */
4190 static int get_implied_cluster_alloc(struct super_block *sb,
4191                                      struct ext4_map_blocks *map,
4192                                      struct ext4_extent *ex,
4193                                      struct ext4_ext_path *path)
4194 {
4195         struct ext4_sb_info *sbi = EXT4_SB(sb);
4196         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4197         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4198         ext4_lblk_t rr_cluster_start;
4199         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4200         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4201         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4202
4203         /* The extent passed in that we are trying to match */
4204         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4205         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4206
4207         /* The requested region passed into ext4_map_blocks() */
4208         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4209
4210         if ((rr_cluster_start == ex_cluster_end) ||
4211             (rr_cluster_start == ex_cluster_start)) {
4212                 if (rr_cluster_start == ex_cluster_end)
4213                         ee_start += ee_len - 1;
4214                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4215                 map->m_len = min(map->m_len,
4216                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4217                 /*
4218                  * Check for and handle this case:
4219                  *
4220                  *   |--------- cluster # N-------------|
4221                  *                     |------- extent ----|
4222                  *         |--- requested region ---|
4223                  *         |===========|
4224                  */
4225
4226                 if (map->m_lblk < ee_block)
4227                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4228
4229                 /*
4230                  * Check for the case where there is already another allocated
4231                  * block to the right of 'ex' but before the end of the cluster.
4232                  *
4233                  *          |------------- cluster # N-------------|
4234                  * |----- ex -----|                  |---- ex_right ----|
4235                  *                  |------ requested region ------|
4236                  *                  |================|
4237                  */
4238                 if (map->m_lblk > ee_block) {
4239                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4240                         map->m_len = min(map->m_len, next - map->m_lblk);
4241                 }
4242
4243                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4244                 return 1;
4245         }
4246
4247         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4248         return 0;
4249 }
4250
4251
4252 /*
4253  * Block allocation/map/preallocation routine for extents based files
4254  *
4255  *
4256  * Need to be called with
4257  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4258  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4259  *
4260  * return > 0, number of of blocks already mapped/allocated
4261  *          if create == 0 and these are pre-allocated blocks
4262  *              buffer head is unmapped
4263  *          otherwise blocks are mapped
4264  *
4265  * return = 0, if plain look up failed (blocks have not been allocated)
4266  *          buffer head is unmapped
4267  *
4268  * return < 0, error case.
4269  */
4270 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4271                         struct ext4_map_blocks *map, int flags)
4272 {
4273         struct ext4_ext_path *path = NULL;
4274         struct ext4_extent newex, *ex, *ex2;
4275         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4276         ext4_fsblk_t newblock = 0;
4277         int free_on_err = 0, err = 0, depth, ret;
4278         unsigned int allocated = 0, offset = 0;
4279         unsigned int allocated_clusters = 0;
4280         struct ext4_allocation_request ar;
4281         ext4_io_end_t *io = ext4_inode_aio(inode);
4282         ext4_lblk_t cluster_offset;
4283         int set_unwritten = 0;
4284         bool map_from_cluster = false;
4285
4286         ext_debug("blocks %u/%u requested for inode %lu\n",
4287                   map->m_lblk, map->m_len, inode->i_ino);
4288         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4289
4290         /* find extent for this block */
4291         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4292         if (IS_ERR(path)) {
4293                 err = PTR_ERR(path);
4294                 path = NULL;
4295                 goto out2;
4296         }
4297
4298         depth = ext_depth(inode);
4299
4300         /*
4301          * consistent leaf must not be empty;
4302          * this situation is possible, though, _during_ tree modification;
4303          * this is why assert can't be put in ext4_find_extent()
4304          */
4305         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4306                 EXT4_ERROR_INODE(inode, "bad extent address "
4307                                  "lblock: %lu, depth: %d pblock %lld",
4308                                  (unsigned long) map->m_lblk, depth,
4309                                  path[depth].p_block);
4310                 err = -EIO;
4311                 goto out2;
4312         }
4313
4314         ex = path[depth].p_ext;
4315         if (ex) {
4316                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4317                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4318                 unsigned short ee_len;
4319
4320
4321                 /*
4322                  * unwritten extents are treated as holes, except that
4323                  * we split out initialized portions during a write.
4324                  */
4325                 ee_len = ext4_ext_get_actual_len(ex);
4326
4327                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4328
4329                 /* if found extent covers block, simply return it */
4330                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4331                         newblock = map->m_lblk - ee_block + ee_start;
4332                         /* number of remaining blocks in the extent */
4333                         allocated = ee_len - (map->m_lblk - ee_block);
4334                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4335                                   ee_block, ee_len, newblock);
4336
4337                         /*
4338                          * If the extent is initialized check whether the
4339                          * caller wants to convert it to unwritten.
4340                          */
4341                         if ((!ext4_ext_is_unwritten(ex)) &&
4342                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4343                                 allocated = convert_initialized_extent(
4344                                                 handle, inode, map, &path,
4345                                                 flags, allocated, newblock);
4346                                 goto out2;
4347                         } else if (!ext4_ext_is_unwritten(ex))
4348                                 goto out;
4349
4350                         ret = ext4_ext_handle_unwritten_extents(
4351                                 handle, inode, map, &path, flags,
4352                                 allocated, newblock);
4353                         if (ret < 0)
4354                                 err = ret;
4355                         else
4356                                 allocated = ret;
4357                         goto out2;
4358                 }
4359         }
4360
4361         /*
4362          * requested block isn't allocated yet;
4363          * we couldn't try to create block if create flag is zero
4364          */
4365         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4366                 /*
4367                  * put just found gap into cache to speed up
4368                  * subsequent requests
4369                  */
4370                 ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4371                 goto out2;
4372         }
4373
4374         /*
4375          * Okay, we need to do block allocation.
4376          */
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_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_from_cluster = true;
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_from_cluster = true;
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         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4458                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4459         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4460         if (!newblock)
4461                 goto out2;
4462         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4463                   ar.goal, newblock, allocated);
4464         free_on_err = 1;
4465         allocated_clusters = ar.len;
4466         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4467         if (ar.len > allocated)
4468                 ar.len = allocated;
4469
4470 got_allocated_blocks:
4471         /* try to insert new extent into found leaf and return */
4472         ext4_ext_store_pblock(&newex, newblock + offset);
4473         newex.ee_len = cpu_to_le16(ar.len);
4474         /* Mark unwritten */
4475         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4476                 ext4_ext_mark_unwritten(&newex);
4477                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4478                 /*
4479                  * io_end structure was created for every IO write to an
4480                  * unwritten extent. To avoid unnecessary conversion,
4481                  * here we flag the IO that really needs the conversion.
4482                  * For non asycn direct IO case, flag the inode state
4483                  * that we need to perform conversion when IO is done.
4484                  */
4485                 if (flags & EXT4_GET_BLOCKS_PRE_IO)
4486                         set_unwritten = 1;
4487         }
4488
4489         err = 0;
4490         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4491                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4492                                          path, ar.len);
4493         if (!err)
4494                 err = ext4_ext_insert_extent(handle, inode, &path,
4495                                              &newex, flags);
4496
4497         if (!err && set_unwritten) {
4498                 if (io)
4499                         ext4_set_io_unwritten_flag(inode, io);
4500                 else
4501                         ext4_set_inode_state(inode,
4502                                              EXT4_STATE_DIO_UNWRITTEN);
4503         }
4504
4505         if (err && free_on_err) {
4506                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4507                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4508                 /* free data blocks we just allocated */
4509                 /* not a good idea to call discard here directly,
4510                  * but otherwise we'd need to call it every free() */
4511                 ext4_discard_preallocations(inode);
4512                 ext4_free_blocks(handle, inode, NULL, newblock,
4513                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4514                 goto out2;
4515         }
4516
4517         /* previous routine could use block we allocated */
4518         newblock = ext4_ext_pblock(&newex);
4519         allocated = ext4_ext_get_actual_len(&newex);
4520         if (allocated > map->m_len)
4521                 allocated = map->m_len;
4522         map->m_flags |= EXT4_MAP_NEW;
4523
4524         /*
4525          * Update reserved blocks/metadata blocks after successful
4526          * block allocation which had been deferred till now.
4527          */
4528         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4529                 unsigned int reserved_clusters;
4530                 /*
4531                  * Check how many clusters we had reserved this allocated range
4532                  */
4533                 reserved_clusters = get_reserved_cluster_alloc(inode,
4534                                                 map->m_lblk, allocated);
4535                 if (!map_from_cluster) {
4536                         BUG_ON(allocated_clusters < reserved_clusters);
4537                         if (reserved_clusters < allocated_clusters) {
4538                                 struct ext4_inode_info *ei = EXT4_I(inode);
4539                                 int reservation = allocated_clusters -
4540                                                   reserved_clusters;
4541                                 /*
4542                                  * It seems we claimed few clusters outside of
4543                                  * the range of this allocation. We should give
4544                                  * it back to the reservation pool. This can
4545                                  * happen in the following case:
4546                                  *
4547                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4548                                  *   cluster has 4 blocks. Thus, the clusters
4549                                  *   are [0-3],[4-7],[8-11]...
4550                                  * * First comes delayed allocation write for
4551                                  *   logical blocks 10 & 11. Since there were no
4552                                  *   previous delayed allocated blocks in the
4553                                  *   range [8-11], we would reserve 1 cluster
4554                                  *   for this write.
4555                                  * * Next comes write for logical blocks 3 to 8.
4556                                  *   In this case, we will reserve 2 clusters
4557                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4558                                  *   that range has a delayed allocated blocks.
4559                                  *   Thus total reserved clusters now becomes 3.
4560                                  * * Now, during the delayed allocation writeout
4561                                  *   time, we will first write blocks [3-8] and
4562                                  *   allocate 3 clusters for writing these
4563                                  *   blocks. Also, we would claim all these
4564                                  *   three clusters above.
4565                                  * * Now when we come here to writeout the
4566                                  *   blocks [10-11], we would expect to claim
4567                                  *   the reservation of 1 cluster we had made
4568                                  *   (and we would claim it since there are no
4569                                  *   more delayed allocated blocks in the range
4570                                  *   [8-11]. But our reserved cluster count had
4571                                  *   already gone to 0.
4572                                  *
4573                                  *   Thus, at the step 4 above when we determine
4574                                  *   that there are still some unwritten delayed
4575                                  *   allocated blocks outside of our current
4576                                  *   block range, we should increment the
4577                                  *   reserved clusters count so that when the
4578                                  *   remaining blocks finally gets written, we
4579                                  *   could claim them.
4580                                  */
4581                                 dquot_reserve_block(inode,
4582                                                 EXT4_C2B(sbi, reservation));
4583                                 spin_lock(&ei->i_block_reservation_lock);
4584                                 ei->i_reserved_data_blocks += reservation;
4585                                 spin_unlock(&ei->i_block_reservation_lock);
4586                         }
4587                         /*
4588                          * We will claim quota for all newly allocated blocks.
4589                          * We're updating the reserved space *after* the
4590                          * correction above so we do not accidentally free
4591                          * all the metadata reservation because we might
4592                          * actually need it later on.
4593                          */
4594                         ext4_da_update_reserve_space(inode, allocated_clusters,
4595                                                         1);
4596                 }
4597         }
4598
4599         /*
4600          * Cache the extent and update transaction to commit on fdatasync only
4601          * when it is _not_ an unwritten extent.
4602          */
4603         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4604                 ext4_update_inode_fsync_trans(handle, inode, 1);
4605         else
4606                 ext4_update_inode_fsync_trans(handle, inode, 0);
4607 out:
4608         if (allocated > map->m_len)
4609                 allocated = map->m_len;
4610         ext4_ext_show_leaf(inode, path);
4611         map->m_flags |= EXT4_MAP_MAPPED;
4612         map->m_pblk = newblock;
4613         map->m_len = allocated;
4614 out2:
4615         ext4_ext_drop_refs(path);
4616         kfree(path);
4617
4618         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4619                                        err ? err : allocated);
4620         return err ? err : allocated;
4621 }
4622
4623 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4624 {
4625         struct super_block *sb = inode->i_sb;
4626         ext4_lblk_t last_block;
4627         int err = 0;
4628
4629         /*
4630          * TODO: optimization is possible here.
4631          * Probably we need not scan at all,
4632          * because page truncation is enough.
4633          */
4634
4635         /* we have to know where to truncate from in crash case */
4636         EXT4_I(inode)->i_disksize = inode->i_size;
4637         ext4_mark_inode_dirty(handle, inode);
4638
4639         last_block = (inode->i_size + sb->s_blocksize - 1)
4640                         >> EXT4_BLOCK_SIZE_BITS(sb);
4641 retry:
4642         err = ext4_es_remove_extent(inode, last_block,
4643                                     EXT_MAX_BLOCKS - last_block);
4644         if (err == -ENOMEM) {
4645                 cond_resched();
4646                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4647                 goto retry;
4648         }
4649         if (err) {
4650                 ext4_std_error(inode->i_sb, err);
4651                 return;
4652         }
4653         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4654         ext4_std_error(inode->i_sb, err);
4655 }
4656
4657 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4658                                   ext4_lblk_t len, loff_t new_size,
4659                                   int flags, int mode)
4660 {
4661         struct inode *inode = file_inode(file);
4662         handle_t *handle;
4663         int ret = 0;
4664         int ret2 = 0;
4665         int retries = 0;
4666         struct ext4_map_blocks map;
4667         unsigned int credits;
4668         loff_t epos;
4669
4670         map.m_lblk = offset;
4671         map.m_len = len;
4672         /*
4673          * Don't normalize the request if it can fit in one extent so
4674          * that it doesn't get unnecessarily split into multiple
4675          * extents.
4676          */
4677         if (len <= EXT_UNWRITTEN_MAX_LEN)
4678                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4679
4680         /*
4681          * credits to insert 1 extent into extent tree
4682          */
4683         credits = ext4_chunk_trans_blocks(inode, len);
4684
4685 retry:
4686         while (ret >= 0 && len) {
4687                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4688                                             credits);
4689                 if (IS_ERR(handle)) {
4690                         ret = PTR_ERR(handle);
4691                         break;
4692                 }
4693                 ret = ext4_map_blocks(handle, inode, &map, flags);
4694                 if (ret <= 0) {
4695                         ext4_debug("inode #%lu: block %u: len %u: "
4696                                    "ext4_ext_map_blocks returned %d",
4697                                    inode->i_ino, map.m_lblk,
4698                                    map.m_len, ret);
4699                         ext4_mark_inode_dirty(handle, inode);
4700                         ret2 = ext4_journal_stop(handle);
4701                         break;
4702                 }
4703                 map.m_lblk += ret;
4704                 map.m_len = len = len - ret;
4705                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4706                 inode->i_ctime = ext4_current_time(inode);
4707                 if (new_size) {
4708                         if (epos > new_size)
4709                                 epos = new_size;
4710                         if (ext4_update_inode_size(inode, epos) & 0x1)
4711                                 inode->i_mtime = inode->i_ctime;
4712                 } else {
4713                         if (epos > inode->i_size)
4714                                 ext4_set_inode_flag(inode,
4715                                                     EXT4_INODE_EOFBLOCKS);
4716                 }
4717                 ext4_mark_inode_dirty(handle, inode);
4718                 ret2 = ext4_journal_stop(handle);
4719                 if (ret2)
4720                         break;
4721         }
4722         if (ret == -ENOSPC &&
4723                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4724                 ret = 0;
4725                 goto retry;
4726         }
4727
4728         return ret > 0 ? ret2 : ret;
4729 }
4730
4731 static long ext4_zero_range(struct file *file, loff_t offset,
4732                             loff_t len, int mode)
4733 {
4734         struct inode *inode = file_inode(file);
4735         handle_t *handle = NULL;
4736         unsigned int max_blocks;
4737         loff_t new_size = 0;
4738         int ret = 0;
4739         int flags;
4740         int credits;
4741         int partial_begin, partial_end;
4742         loff_t start, end;
4743         ext4_lblk_t lblk;
4744         struct address_space *mapping = inode->i_mapping;
4745         unsigned int blkbits = inode->i_blkbits;
4746
4747         trace_ext4_zero_range(inode, offset, len, mode);
4748
4749         if (!S_ISREG(inode->i_mode))
4750                 return -EINVAL;
4751
4752         /* Call ext4_force_commit to flush all data in case of data=journal. */
4753         if (ext4_should_journal_data(inode)) {
4754                 ret = ext4_force_commit(inode->i_sb);
4755                 if (ret)
4756                         return ret;
4757         }
4758
4759         /*
4760          * Write out all dirty pages to avoid race conditions
4761          * Then release them.
4762          */
4763         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4764                 ret = filemap_write_and_wait_range(mapping, offset,
4765                                                    offset + len - 1);
4766                 if (ret)
4767                         return ret;
4768         }
4769
4770         /*
4771          * Round up offset. This is not fallocate, we neet to zero out
4772          * blocks, so convert interior block aligned part of the range to
4773          * unwritten and possibly manually zero out unaligned parts of the
4774          * range.
4775          */
4776         start = round_up(offset, 1 << blkbits);
4777         end = round_down((offset + len), 1 << blkbits);
4778
4779         if (start < offset || end > offset + len)
4780                 return -EINVAL;
4781         partial_begin = offset & ((1 << blkbits) - 1);
4782         partial_end = (offset + len) & ((1 << blkbits) - 1);
4783
4784         lblk = start >> blkbits;
4785         max_blocks = (end >> blkbits);
4786         if (max_blocks < lblk)
4787                 max_blocks = 0;
4788         else
4789                 max_blocks -= lblk;
4790
4791         mutex_lock(&inode->i_mutex);
4792
4793         /*
4794          * Indirect files do not support unwritten extnets
4795          */
4796         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4797                 ret = -EOPNOTSUPP;
4798                 goto out_mutex;
4799         }
4800
4801         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4802              offset + len > i_size_read(inode)) {
4803                 new_size = offset + len;
4804                 ret = inode_newsize_ok(inode, new_size);
4805                 if (ret)
4806                         goto out_mutex;
4807         }
4808
4809         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4810         if (mode & FALLOC_FL_KEEP_SIZE)
4811                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4812
4813         /* Preallocate the range including the unaligned edges */
4814         if (partial_begin || partial_end) {
4815                 ret = ext4_alloc_file_blocks(file,
4816                                 round_down(offset, 1 << blkbits) >> blkbits,
4817                                 (round_up((offset + len), 1 << blkbits) -
4818                                  round_down(offset, 1 << blkbits)) >> blkbits,
4819                                 new_size, flags, mode);
4820                 if (ret)
4821                         goto out_mutex;
4822
4823         }
4824
4825         /* Zero range excluding the unaligned edges */
4826         if (max_blocks > 0) {
4827                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4828                           EXT4_EX_NOCACHE);
4829
4830                 /* Now release the pages and zero block aligned part of pages*/
4831                 truncate_pagecache_range(inode, start, end - 1);
4832                 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4833
4834                 /* Wait all existing dio workers, newcomers will block on i_mutex */
4835                 ext4_inode_block_unlocked_dio(inode);
4836                 inode_dio_wait(inode);
4837
4838                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4839                                              flags, mode);
4840                 if (ret)
4841                         goto out_dio;
4842         }
4843         if (!partial_begin && !partial_end)
4844                 goto out_dio;
4845
4846         /*
4847          * In worst case we have to writeout two nonadjacent unwritten
4848          * blocks and update the inode
4849          */
4850         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4851         if (ext4_should_journal_data(inode))
4852                 credits += 2;
4853         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4854         if (IS_ERR(handle)) {
4855                 ret = PTR_ERR(handle);
4856                 ext4_std_error(inode->i_sb, ret);
4857                 goto out_dio;
4858         }
4859
4860         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4861         if (new_size) {
4862                 ext4_update_inode_size(inode, new_size);
4863         } else {
4864                 /*
4865                 * Mark that we allocate beyond EOF so the subsequent truncate
4866                 * can proceed even if the new size is the same as i_size.
4867                 */
4868                 if ((offset + len) > i_size_read(inode))
4869                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4870         }
4871         ext4_mark_inode_dirty(handle, inode);
4872
4873         /* Zero out partial block at the edges of the range */
4874         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4875
4876         if (file->f_flags & O_SYNC)
4877                 ext4_handle_sync(handle);
4878
4879         ext4_journal_stop(handle);
4880 out_dio:
4881         ext4_inode_resume_unlocked_dio(inode);
4882 out_mutex:
4883         mutex_unlock(&inode->i_mutex);
4884         return ret;
4885 }
4886
4887 /*
4888  * preallocate space for a file. This implements ext4's fallocate file
4889  * operation, which gets called from sys_fallocate system call.
4890  * For block-mapped files, posix_fallocate should fall back to the method
4891  * of writing zeroes to the required new blocks (the same behavior which is
4892  * expected for file systems which do not support fallocate() system call).
4893  */
4894 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4895 {
4896         struct inode *inode = file_inode(file);
4897         loff_t new_size = 0;
4898         unsigned int max_blocks;
4899         int ret = 0;
4900         int flags;
4901         ext4_lblk_t lblk;
4902         unsigned int blkbits = inode->i_blkbits;
4903
4904         /*
4905          * Encrypted inodes can't handle collapse range or insert
4906          * range since we would need to re-encrypt blocks with a
4907          * different IV or XTS tweak (which are based on the logical
4908          * block number).
4909          *
4910          * XXX It's not clear why zero range isn't working, but we'll
4911          * leave it disabled for encrypted inodes for now.  This is a
4912          * bug we should fix....
4913          */
4914         if (ext4_encrypted_inode(inode) &&
4915             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE)))
4916                 return -EOPNOTSUPP;
4917
4918         /* Return error if mode is not supported */
4919         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4920                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
4921                 return -EOPNOTSUPP;
4922
4923         if (mode & FALLOC_FL_PUNCH_HOLE)
4924                 return ext4_punch_hole(inode, offset, len);
4925
4926         ret = ext4_convert_inline_data(inode);
4927         if (ret)
4928                 return ret;
4929
4930         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4931                 return ext4_collapse_range(inode, offset, len);
4932
4933         if (mode & FALLOC_FL_ZERO_RANGE)
4934                 return ext4_zero_range(file, offset, len, mode);
4935
4936         trace_ext4_fallocate_enter(inode, offset, len, mode);
4937         lblk = offset >> blkbits;
4938         /*
4939          * We can't just convert len to max_blocks because
4940          * If blocksize = 4096 offset = 3072 and len = 2048
4941          */
4942         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4943                 - lblk;
4944
4945         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4946         if (mode & FALLOC_FL_KEEP_SIZE)
4947                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4948
4949         mutex_lock(&inode->i_mutex);
4950
4951         /*
4952          * We only support preallocation for extent-based files only
4953          */
4954         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4955                 ret = -EOPNOTSUPP;
4956                 goto out;
4957         }
4958
4959         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4960              offset + len > i_size_read(inode)) {
4961                 new_size = offset + len;
4962                 ret = inode_newsize_ok(inode, new_size);
4963                 if (ret)
4964                         goto out;
4965         }
4966
4967         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4968                                      flags, mode);
4969         if (ret)
4970                 goto out;
4971
4972         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4973                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4974                                                 EXT4_I(inode)->i_sync_tid);
4975         }
4976 out:
4977         mutex_unlock(&inode->i_mutex);
4978         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4979         return ret;
4980 }
4981
4982 /*
4983  * This function convert a range of blocks to written extents
4984  * The caller of this function will pass the start offset and the size.
4985  * all unwritten extents within this range will be converted to
4986  * written extents.
4987  *
4988  * This function is called from the direct IO end io call back
4989  * function, to convert the fallocated extents after IO is completed.
4990  * Returns 0 on success.
4991  */
4992 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4993                                    loff_t offset, ssize_t len)
4994 {
4995         unsigned int max_blocks;
4996         int ret = 0;
4997         int ret2 = 0;
4998         struct ext4_map_blocks map;
4999         unsigned int credits, blkbits = inode->i_blkbits;
5000
5001         map.m_lblk = offset >> blkbits;
5002         /*
5003          * We can't just convert len to max_blocks because
5004          * If blocksize = 4096 offset = 3072 and len = 2048
5005          */
5006         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
5007                       map.m_lblk);
5008         /*
5009          * This is somewhat ugly but the idea is clear: When transaction is
5010          * reserved, everything goes into it. Otherwise we rather start several
5011          * smaller transactions for conversion of each extent separately.
5012          */
5013         if (handle) {
5014                 handle = ext4_journal_start_reserved(handle,
5015                                                      EXT4_HT_EXT_CONVERT);
5016                 if (IS_ERR(handle))
5017                         return PTR_ERR(handle);
5018                 credits = 0;
5019         } else {
5020                 /*
5021                  * credits to insert 1 extent into extent tree
5022                  */
5023                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5024         }
5025         while (ret >= 0 && ret < max_blocks) {
5026                 map.m_lblk += ret;
5027                 map.m_len = (max_blocks -= ret);
5028                 if (credits) {
5029                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5030                                                     credits);
5031                         if (IS_ERR(handle)) {
5032                                 ret = PTR_ERR(handle);
5033                                 break;
5034                         }
5035                 }
5036                 ret = ext4_map_blocks(handle, inode, &map,
5037                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5038                 if (ret <= 0)
5039                         ext4_warning(inode->i_sb,
5040                                      "inode #%lu: block %u: len %u: "
5041                                      "ext4_ext_map_blocks returned %d",
5042                                      inode->i_ino, map.m_lblk,
5043                                      map.m_len, ret);
5044                 ext4_mark_inode_dirty(handle, inode);
5045                 if (credits)
5046                         ret2 = ext4_journal_stop(handle);
5047                 if (ret <= 0 || ret2)
5048                         break;
5049         }
5050         if (!credits)
5051                 ret2 = ext4_journal_stop(handle);
5052         return ret > 0 ? ret2 : ret;
5053 }
5054
5055 /*
5056  * If newes is not existing extent (newes->ec_pblk equals zero) find
5057  * delayed extent at start of newes and update newes accordingly and
5058  * return start of the next delayed extent.
5059  *
5060  * If newes is existing extent (newes->ec_pblk is not equal zero)
5061  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5062  * extent found. Leave newes unmodified.
5063  */
5064 static int ext4_find_delayed_extent(struct inode *inode,
5065                                     struct extent_status *newes)
5066 {
5067         struct extent_status es;
5068         ext4_lblk_t block, next_del;
5069
5070         if (newes->es_pblk == 0) {
5071                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5072                                 newes->es_lblk + newes->es_len - 1, &es);
5073
5074                 /*
5075                  * No extent in extent-tree contains block @newes->es_pblk,
5076                  * then the block may stay in 1)a hole or 2)delayed-extent.
5077                  */
5078                 if (es.es_len == 0)
5079                         /* A hole found. */
5080                         return 0;
5081
5082                 if (es.es_lblk > newes->es_lblk) {
5083                         /* A hole found. */
5084                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5085                                             newes->es_len);
5086                         return 0;
5087                 }
5088
5089                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5090         }
5091
5092         block = newes->es_lblk + newes->es_len;
5093         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5094         if (es.es_len == 0)
5095                 next_del = EXT_MAX_BLOCKS;
5096         else
5097                 next_del = es.es_lblk;
5098
5099         return next_del;
5100 }
5101 /* fiemap flags we can handle specified here */
5102 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5103
5104 static int ext4_xattr_fiemap(struct inode *inode,
5105                                 struct fiemap_extent_info *fieinfo)
5106 {
5107         __u64 physical = 0;
5108         __u64 length;
5109         __u32 flags = FIEMAP_EXTENT_LAST;
5110         int blockbits = inode->i_sb->s_blocksize_bits;
5111         int error = 0;
5112
5113         /* in-inode? */
5114         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5115                 struct ext4_iloc iloc;
5116                 int offset;     /* offset of xattr in inode */
5117
5118                 error = ext4_get_inode_loc(inode, &iloc);
5119                 if (error)
5120                         return error;
5121                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5122                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5123                                 EXT4_I(inode)->i_extra_isize;
5124                 physical += offset;
5125                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5126                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5127                 brelse(iloc.bh);
5128         } else { /* external block */
5129                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5130                 length = inode->i_sb->s_blocksize;
5131         }
5132
5133         if (physical)
5134                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5135                                                 length, flags);
5136         return (error < 0 ? error : 0);
5137 }
5138
5139 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5140                 __u64 start, __u64 len)
5141 {
5142         ext4_lblk_t start_blk;
5143         int error = 0;
5144
5145         if (ext4_has_inline_data(inode)) {
5146                 int has_inline = 1;
5147
5148                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5149                                                 start, len);
5150
5151                 if (has_inline)
5152                         return error;
5153         }
5154
5155         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5156                 error = ext4_ext_precache(inode);
5157                 if (error)
5158                         return error;
5159         }
5160
5161         /* fallback to generic here if not in extents fmt */
5162         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5163                 return generic_block_fiemap(inode, fieinfo, start, len,
5164                         ext4_get_block);
5165
5166         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5167                 return -EBADR;
5168
5169         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5170                 error = ext4_xattr_fiemap(inode, fieinfo);
5171         } else {
5172                 ext4_lblk_t len_blks;
5173                 __u64 last_blk;
5174
5175                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5176                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5177                 if (last_blk >= EXT_MAX_BLOCKS)
5178                         last_blk = EXT_MAX_BLOCKS-1;
5179                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5180
5181                 /*
5182                  * Walk the extent tree gathering extent information
5183                  * and pushing extents back to the user.
5184                  */
5185                 error = ext4_fill_fiemap_extents(inode, start_blk,
5186                                                  len_blks, fieinfo);
5187         }
5188         return error;
5189 }
5190
5191 /*
5192  * ext4_access_path:
5193  * Function to access the path buffer for marking it dirty.
5194  * It also checks if there are sufficient credits left in the journal handle
5195  * to update path.
5196  */
5197 static int
5198 ext4_access_path(handle_t *handle, struct inode *inode,
5199                 struct ext4_ext_path *path)
5200 {
5201         int credits, err;
5202
5203         if (!ext4_handle_valid(handle))
5204                 return 0;
5205
5206         /*
5207          * Check if need to extend journal credits
5208          * 3 for leaf, sb, and inode plus 2 (bmap and group
5209          * descriptor) for each block group; assume two block
5210          * groups
5211          */
5212         if (handle->h_buffer_credits < 7) {
5213                 credits = ext4_writepage_trans_blocks(inode);
5214                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5215                 /* EAGAIN is success */
5216                 if (err && err != -EAGAIN)
5217                         return err;
5218         }
5219
5220         err = ext4_ext_get_access(handle, inode, path);
5221         return err;
5222 }
5223
5224 /*
5225  * ext4_ext_shift_path_extents:
5226  * Shift the extents of a path structure lying between path[depth].p_ext
5227  * and EXT_LAST_EXTENT(path[depth].p_hdr) downwards, by subtracting shift
5228  * from starting block for each extent.
5229  */
5230 static int
5231 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5232                             struct inode *inode, handle_t *handle,
5233                             ext4_lblk_t *start)
5234 {
5235         int depth, err = 0;
5236         struct ext4_extent *ex_start, *ex_last;
5237         bool update = 0;
5238         depth = path->p_depth;
5239
5240         while (depth >= 0) {
5241                 if (depth == path->p_depth) {
5242                         ex_start = path[depth].p_ext;
5243                         if (!ex_start)
5244                                 return -EIO;
5245
5246                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5247
5248                         err = ext4_access_path(handle, inode, path + depth);
5249                         if (err)
5250                                 goto out;
5251
5252                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5253                                 update = 1;
5254
5255                         *start = le32_to_cpu(ex_last->ee_block) +
5256                                 ext4_ext_get_actual_len(ex_last);
5257
5258                         while (ex_start <= ex_last) {
5259                                 le32_add_cpu(&ex_start->ee_block, -shift);
5260                                 /* Try to merge to the left. */
5261                                 if ((ex_start >
5262                                      EXT_FIRST_EXTENT(path[depth].p_hdr)) &&
5263                                     ext4_ext_try_to_merge_right(inode,
5264                                                         path, ex_start - 1))
5265                                         ex_last--;
5266                                 else
5267                                         ex_start++;
5268                         }
5269                         err = ext4_ext_dirty(handle, inode, path + depth);
5270                         if (err)
5271                                 goto out;
5272
5273                         if (--depth < 0 || !update)
5274                                 break;
5275                 }
5276
5277                 /* Update index too */
5278                 err = ext4_access_path(handle, inode, path + depth);
5279                 if (err)
5280                         goto out;
5281
5282                 le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5283                 err = ext4_ext_dirty(handle, inode, path + depth);
5284                 if (err)
5285                         goto out;
5286
5287                 /* we are done if current index is not a starting index */
5288                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5289                         break;
5290
5291                 depth--;
5292         }
5293
5294 out:
5295         return err;
5296 }
5297
5298 /*
5299  * ext4_ext_shift_extents:
5300  * All the extents which lies in the range from start to the last allocated
5301  * block for the file are shifted downwards by shift blocks.
5302  * On success, 0 is returned, error otherwise.
5303  */
5304 static int
5305 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5306                        ext4_lblk_t start, ext4_lblk_t shift)
5307 {
5308         struct ext4_ext_path *path;
5309         int ret = 0, depth;
5310         struct ext4_extent *extent;
5311         ext4_lblk_t stop_block;
5312         ext4_lblk_t ex_start, ex_end;
5313
5314         /* Let path point to the last extent */
5315         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 0);
5316         if (IS_ERR(path))
5317                 return PTR_ERR(path);
5318
5319         depth = path->p_depth;
5320         extent = path[depth].p_ext;
5321         if (!extent)
5322                 goto out;
5323
5324         stop_block = le32_to_cpu(extent->ee_block) +
5325                         ext4_ext_get_actual_len(extent);
5326
5327         /* Nothing to shift, if hole is at the end of file */
5328         if (start >= stop_block)
5329                 goto out;
5330
5331         /*
5332          * Don't start shifting extents until we make sure the hole is big
5333          * enough to accomodate the shift.
5334          */
5335         path = ext4_find_extent(inode, start - 1, &path, 0);
5336         if (IS_ERR(path))
5337                 return PTR_ERR(path);
5338         depth = path->p_depth;
5339         extent =  path[depth].p_ext;
5340         if (extent) {
5341                 ex_start = le32_to_cpu(extent->ee_block);
5342                 ex_end = le32_to_cpu(extent->ee_block) +
5343                         ext4_ext_get_actual_len(extent);
5344         } else {
5345                 ex_start = 0;
5346                 ex_end = 0;
5347         }
5348
5349         if ((start == ex_start && shift > ex_start) ||
5350             (shift > start - ex_end))
5351                 return -EINVAL;
5352
5353         /* Its safe to start updating extents */
5354         while (start < stop_block) {
5355                 path = ext4_find_extent(inode, start, &path, 0);
5356                 if (IS_ERR(path))
5357                         return PTR_ERR(path);
5358                 depth = path->p_depth;
5359                 extent = path[depth].p_ext;
5360                 if (!extent) {
5361                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5362                                          (unsigned long) start);
5363                         return -EIO;
5364                 }
5365                 if (start > le32_to_cpu(extent->ee_block)) {
5366                         /* Hole, move to the next extent */
5367                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5368                                 path[depth].p_ext++;
5369                         } else {
5370                                 start = ext4_ext_next_allocated_block(path);
5371                                 continue;
5372                         }
5373                 }
5374                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5375                                 handle, &start);
5376                 if (ret)
5377                         break;
5378         }
5379 out:
5380         ext4_ext_drop_refs(path);
5381         kfree(path);
5382         return ret;
5383 }
5384
5385 /*
5386  * ext4_collapse_range:
5387  * This implements the fallocate's collapse range functionality for ext4
5388  * Returns: 0 and non-zero on error.
5389  */
5390 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5391 {
5392         struct super_block *sb = inode->i_sb;
5393         ext4_lblk_t punch_start, punch_stop;
5394         handle_t *handle;
5395         unsigned int credits;
5396         loff_t new_size, ioffset;
5397         int ret;
5398
5399         /*
5400          * We need to test this early because xfstests assumes that a
5401          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5402          * system does not support collapse range.
5403          */
5404         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5405                 return -EOPNOTSUPP;
5406
5407         /* Collapse range works only on fs block size aligned offsets. */
5408         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5409             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5410                 return -EINVAL;
5411
5412         if (!S_ISREG(inode->i_mode))
5413                 return -EINVAL;
5414
5415         trace_ext4_collapse_range(inode, offset, len);
5416
5417         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5418         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5419
5420         /* Call ext4_force_commit to flush all data in case of data=journal. */
5421         if (ext4_should_journal_data(inode)) {
5422                 ret = ext4_force_commit(inode->i_sb);
5423                 if (ret)
5424                         return ret;
5425         }
5426
5427         /*
5428          * Need to round down offset to be aligned with page size boundary
5429          * for page size > block size.
5430          */
5431         ioffset = round_down(offset, PAGE_SIZE);
5432
5433         /* Write out all dirty pages */
5434         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5435                                            LLONG_MAX);
5436         if (ret)
5437                 return ret;
5438
5439         /* Take mutex lock */
5440         mutex_lock(&inode->i_mutex);
5441
5442         /*
5443          * There is no need to overlap collapse range with EOF, in which case
5444          * it is effectively a truncate operation
5445          */
5446         if (offset + len >= i_size_read(inode)) {
5447                 ret = -EINVAL;
5448                 goto out_mutex;
5449         }
5450
5451         /* Currently just for extent based files */
5452         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5453                 ret = -EOPNOTSUPP;
5454                 goto out_mutex;
5455         }
5456
5457         truncate_pagecache(inode, ioffset);
5458
5459         /* Wait for existing dio to complete */
5460         ext4_inode_block_unlocked_dio(inode);
5461         inode_dio_wait(inode);
5462
5463         credits = ext4_writepage_trans_blocks(inode);
5464         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5465         if (IS_ERR(handle)) {
5466                 ret = PTR_ERR(handle);
5467                 goto out_dio;
5468         }
5469
5470         down_write(&EXT4_I(inode)->i_data_sem);
5471         ext4_discard_preallocations(inode);
5472
5473         ret = ext4_es_remove_extent(inode, punch_start,
5474                                     EXT_MAX_BLOCKS - punch_start);
5475         if (ret) {
5476                 up_write(&EXT4_I(inode)->i_data_sem);
5477                 goto out_stop;
5478         }
5479
5480         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5481         if (ret) {
5482                 up_write(&EXT4_I(inode)->i_data_sem);
5483                 goto out_stop;
5484         }
5485         ext4_discard_preallocations(inode);
5486
5487         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5488                                      punch_stop - punch_start);
5489         if (ret) {
5490                 up_write(&EXT4_I(inode)->i_data_sem);
5491                 goto out_stop;
5492         }
5493
5494         new_size = i_size_read(inode) - len;
5495         i_size_write(inode, new_size);
5496         EXT4_I(inode)->i_disksize = new_size;
5497
5498         up_write(&EXT4_I(inode)->i_data_sem);
5499         if (IS_SYNC(inode))
5500                 ext4_handle_sync(handle);
5501         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5502         ext4_mark_inode_dirty(handle, inode);
5503
5504 out_stop:
5505         ext4_journal_stop(handle);
5506 out_dio:
5507         ext4_inode_resume_unlocked_dio(inode);
5508 out_mutex:
5509         mutex_unlock(&inode->i_mutex);
5510         return ret;
5511 }
5512
5513 /**
5514  * ext4_swap_extents - Swap extents between two inodes
5515  *
5516  * @inode1:     First inode
5517  * @inode2:     Second inode
5518  * @lblk1:      Start block for first inode
5519  * @lblk2:      Start block for second inode
5520  * @count:      Number of blocks to swap
5521  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5522  * @erp:        Pointer to save error value
5523  *
5524  * This helper routine does exactly what is promise "swap extents". All other
5525  * stuff such as page-cache locking consistency, bh mapping consistency or
5526  * extent's data copying must be performed by caller.
5527  * Locking:
5528  *              i_mutex is held for both inodes
5529  *              i_data_sem is locked for write for both inodes
5530  * Assumptions:
5531  *              All pages from requested range are locked for both inodes
5532  */
5533 int
5534 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5535                      struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5536                   ext4_lblk_t count, int unwritten, int *erp)
5537 {
5538         struct ext4_ext_path *path1 = NULL;
5539         struct ext4_ext_path *path2 = NULL;
5540         int replaced_count = 0;
5541
5542         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5543         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5544         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5545         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5546
5547         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5548         if (unlikely(*erp))
5549                 return 0;
5550         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5551         if (unlikely(*erp))
5552                 return 0;
5553
5554         while (count) {
5555                 struct ext4_extent *ex1, *ex2, tmp_ex;
5556                 ext4_lblk_t e1_blk, e2_blk;
5557                 int e1_len, e2_len, len;
5558                 int split = 0;
5559
5560                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5561                 if (unlikely(IS_ERR(path1))) {
5562                         *erp = PTR_ERR(path1);
5563                         path1 = NULL;
5564                 finish:
5565                         count = 0;
5566                         goto repeat;
5567                 }
5568                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5569                 if (unlikely(IS_ERR(path2))) {
5570                         *erp = PTR_ERR(path2);
5571                         path2 = NULL;
5572                         goto finish;
5573                 }
5574                 ex1 = path1[path1->p_depth].p_ext;
5575                 ex2 = path2[path2->p_depth].p_ext;
5576                 /* Do we have somthing to swap ? */
5577                 if (unlikely(!ex2 || !ex1))
5578                         goto finish;
5579
5580                 e1_blk = le32_to_cpu(ex1->ee_block);
5581                 e2_blk = le32_to_cpu(ex2->ee_block);
5582                 e1_len = ext4_ext_get_actual_len(ex1);
5583                 e2_len = ext4_ext_get_actual_len(ex2);
5584
5585                 /* Hole handling */
5586                 if (!in_range(lblk1, e1_blk, e1_len) ||
5587                     !in_range(lblk2, e2_blk, e2_len)) {
5588                         ext4_lblk_t next1, next2;
5589
5590                         /* if hole after extent, then go to next extent */
5591                         next1 = ext4_ext_next_allocated_block(path1);
5592                         next2 = ext4_ext_next_allocated_block(path2);
5593                         /* If hole before extent, then shift to that extent */
5594                         if (e1_blk > lblk1)
5595                                 next1 = e1_blk;
5596                         if (e2_blk > lblk2)
5597                                 next2 = e1_blk;
5598                         /* Do we have something to swap */
5599                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5600                                 goto finish;
5601                         /* Move to the rightest boundary */
5602                         len = next1 - lblk1;
5603                         if (len < next2 - lblk2)
5604                                 len = next2 - lblk2;
5605                         if (len > count)
5606                                 len = count;
5607                         lblk1 += len;
5608                         lblk2 += len;
5609                         count -= len;
5610                         goto repeat;
5611                 }
5612
5613                 /* Prepare left boundary */
5614                 if (e1_blk < lblk1) {
5615                         split = 1;
5616                         *erp = ext4_force_split_extent_at(handle, inode1,
5617                                                 &path1, lblk1, 0);
5618                         if (unlikely(*erp))
5619                                 goto finish;
5620                 }
5621                 if (e2_blk < lblk2) {
5622                         split = 1;
5623                         *erp = ext4_force_split_extent_at(handle, inode2,
5624                                                 &path2,  lblk2, 0);
5625                         if (unlikely(*erp))
5626                                 goto finish;
5627                 }
5628                 /* ext4_split_extent_at() may result in leaf extent split,
5629                  * path must to be revalidated. */
5630                 if (split)
5631                         goto repeat;
5632
5633                 /* Prepare right boundary */
5634                 len = count;
5635                 if (len > e1_blk + e1_len - lblk1)
5636                         len = e1_blk + e1_len - lblk1;
5637                 if (len > e2_blk + e2_len - lblk2)
5638                         len = e2_blk + e2_len - lblk2;
5639
5640                 if (len != e1_len) {
5641                         split = 1;
5642                         *erp = ext4_force_split_extent_at(handle, inode1,
5643                                                 &path1, lblk1 + len, 0);
5644                         if (unlikely(*erp))
5645                                 goto finish;
5646                 }
5647                 if (len != e2_len) {
5648                         split = 1;
5649                         *erp = ext4_force_split_extent_at(handle, inode2,
5650                                                 &path2, lblk2 + len, 0);
5651                         if (*erp)
5652                                 goto finish;
5653                 }
5654                 /* ext4_split_extent_at() may result in leaf extent split,
5655                  * path must to be revalidated. */
5656                 if (split)
5657                         goto repeat;
5658
5659                 BUG_ON(e2_len != e1_len);
5660                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5661                 if (unlikely(*erp))
5662                         goto finish;
5663                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5664                 if (unlikely(*erp))
5665                         goto finish;
5666
5667                 /* Both extents are fully inside boundaries. Swap it now */
5668                 tmp_ex = *ex1;
5669                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5670                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5671                 ex1->ee_len = cpu_to_le16(e2_len);
5672                 ex2->ee_len = cpu_to_le16(e1_len);
5673                 if (unwritten)
5674                         ext4_ext_mark_unwritten(ex2);
5675                 if (ext4_ext_is_unwritten(&tmp_ex))
5676                         ext4_ext_mark_unwritten(ex1);
5677
5678                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5679                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5680                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5681                                       path2->p_depth);
5682                 if (unlikely(*erp))
5683                         goto finish;
5684                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5685                                       path1->p_depth);
5686                 /*
5687                  * Looks scarry ah..? second inode already points to new blocks,
5688                  * and it was successfully dirtied. But luckily error may happen
5689                  * only due to journal error, so full transaction will be
5690                  * aborted anyway.
5691                  */
5692                 if (unlikely(*erp))
5693                         goto finish;
5694                 lblk1 += len;
5695                 lblk2 += len;
5696                 replaced_count += len;
5697                 count -= len;
5698
5699         repeat:
5700                 ext4_ext_drop_refs(path1);
5701                 kfree(path1);
5702                 ext4_ext_drop_refs(path2);
5703                 kfree(path2);
5704                 path1 = path2 = NULL;
5705         }
5706         return replaced_count;
5707 }