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 License
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <linux/uaccess.h>
  41 #include <linux/fiemap.h>
  42 #include <linux/backing-dev.h>
  43 #include "ext4_jbd2.h"
  44 #include "ext4_extents.h"
  45 #include "xattr.h"
  46
  47 #include <trace/events/ext4.h>
  48
  49 /*
  50  * used by extent splitting.
  51  */
  52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  53                                         due to ENOSPC */
  54 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  55 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  56
  57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  59
  60 static __le32 ext4_extent_block_csum(struct inode *inode,
  61                                      struct ext4_extent_header *eh)
  62 {
  63         struct ext4_inode_info *ei = EXT4_I(inode);
  64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  65         __u32 csum;
  66
  67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  68                            EXT4_EXTENT_TAIL_OFFSET(eh));
  69         return cpu_to_le32(csum);
  70 }
  71
  72 static int ext4_extent_block_csum_verify(struct inode *inode,
  73                                          struct ext4_extent_header *eh)
  74 {
  75         struct ext4_extent_tail *et;
  76
  77         if (!ext4_has_metadata_csum(inode->i_sb))
  78                 return 1;
  79
  80         et = find_ext4_extent_tail(eh);
  81         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  82                 return 0;
  83         return 1;
  84 }
  85
  86 static void ext4_extent_block_csum_set(struct inode *inode,
  87                                        struct ext4_extent_header *eh)
  88 {
  89         struct ext4_extent_tail *et;
  90
  91         if (!ext4_has_metadata_csum(inode->i_sb))
  92                 return;
  93
  94         et = find_ext4_extent_tail(eh);
  95         et->et_checksum = ext4_extent_block_csum(inode, eh);
  96 }
  97
  98 static int ext4_split_extent(handle_t *handle,
  99                                 struct inode *inode,
 100                                 struct ext4_ext_path **ppath,
 101                                 struct ext4_map_blocks *map,
 102                                 int split_flag,
 103                                 int flags);
 104
 105 static int ext4_split_extent_at(handle_t *handle,
 106                              struct inode *inode,
 107                              struct ext4_ext_path **ppath,
 108                              ext4_lblk_t split,
 109                              int split_flag,
 110                              int flags);
 111
 112 static int ext4_find_delayed_extent(struct inode *inode,
 113                                     struct extent_status *newes);
 114
 115 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 116                                             struct inode *inode,
 117                                             int needed)
 118 {
 119         int err;
 120
 121         if (!ext4_handle_valid(handle))
 122                 return 0;
 123         if (handle->h_buffer_credits >= needed)
 124                 return 0;
 125         /*
 126          * If we need to extend the journal get a few extra blocks
 127          * while we're at it for efficiency's sake.
 128          */
 129         needed += 3;
 130         err = ext4_journal_extend(handle, needed - handle->h_buffer_credits);
 131         if (err <= 0)
 132                 return err;
 133         err = ext4_truncate_restart_trans(handle, inode, needed);
 134         if (err == 0)
 135                 err = -EAGAIN;
 136
 137         return err;
 138 }
 139
 140 /*
 141  * could return:
 142  *  - EROFS
 143  *  - ENOMEM
 144  */
 145 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 146                                 struct ext4_ext_path *path)
 147 {
 148         if (path->p_bh) {
 149                 /* path points to block */
 150                 BUFFER_TRACE(path->p_bh, "get_write_access");
 151                 return ext4_journal_get_write_access(handle, path->p_bh);
 152         }
 153         /* path points to leaf/index in inode body */
 154         /* we use in-core data, no need to protect them */
 155         return 0;
 156 }
 157
 158 /*
 159  * could return:
 160  *  - EROFS
 161  *  - ENOMEM
 162  *  - EIO
 163  */
 164 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 165                      struct inode *inode, struct ext4_ext_path *path)
 166 {
 167         int err;
 168
 169         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 170         if (path->p_bh) {
 171                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 172                 /* path points to block */
 173                 err = __ext4_handle_dirty_metadata(where, line, handle,
 174                                                    inode, path->p_bh);
 175         } else {
 176                 /* path points to leaf/index in inode body */
 177                 err = ext4_mark_inode_dirty(handle, inode);
 178         }
 179         return err;
 180 }
 181
 182 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 183                               struct ext4_ext_path *path,
 184                               ext4_lblk_t block)
 185 {
 186         if (path) {
 187                 int depth = path->p_depth;
 188                 struct ext4_extent *ex;
 189
 190                 /*
 191                  * Try to predict block placement assuming that we are
 192                  * filling in a file which will eventually be
 193                  * non-sparse --- i.e., in the case of libbfd writing
 194                  * an ELF object sections out-of-order but in a way
 195                  * the eventually results in a contiguous object or
 196                  * executable file, or some database extending a table
 197                  * space file.  However, this is actually somewhat
 198                  * non-ideal if we are writing a sparse file such as
 199                  * qemu or KVM writing a raw image file that is going
 200                  * to stay fairly sparse, since it will end up
 201                  * fragmenting the file system's free space.  Maybe we
 202                  * should have some hueristics or some way to allow
 203                  * userspace to pass a hint to file system,
 204                  * especially if the latter case turns out to be
 205                  * common.
 206                  */
 207                 ex = path[depth].p_ext;
 208                 if (ex) {
 209                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 210                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 211
 212                         if (block > ext_block)
 213                                 return ext_pblk + (block - ext_block);
 214                         else
 215                                 return ext_pblk - (ext_block - block);
 216                 }
 217
 218                 /* it looks like index is empty;
 219                  * try to find starting block from index itself */
 220                 if (path[depth].p_bh)
 221                         return path[depth].p_bh->b_blocknr;
 222         }
 223
 224         /* OK. use inode's group */
 225         return ext4_inode_to_goal_block(inode);
 226 }
 227
 228 /*
 229  * Allocation for a meta data block
 230  */
 231 static ext4_fsblk_t
 232 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 233                         struct ext4_ext_path *path,
 234                         struct ext4_extent *ex, int *err, unsigned int flags)
 235 {
 236         ext4_fsblk_t goal, newblock;
 237
 238         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 239         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 240                                         NULL, err);
 241         return newblock;
 242 }
 243
 244 static inline int ext4_ext_space_block(struct inode *inode, int check)
 245 {
 246         int size;
 247
 248         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 249                         / sizeof(struct ext4_extent);
 250 #ifdef AGGRESSIVE_TEST
 251         if (!check && size > 6)
 252                 size = 6;
 253 #endif
 254         return size;
 255 }
 256
 257 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 258 {
 259         int size;
 260
 261         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 262                         / sizeof(struct ext4_extent_idx);
 263 #ifdef AGGRESSIVE_TEST
 264         if (!check && size > 5)
 265                 size = 5;
 266 #endif
 267         return size;
 268 }
 269
 270 static inline int ext4_ext_space_root(struct inode *inode, int check)
 271 {
 272         int size;
 273
 274         size = sizeof(EXT4_I(inode)->i_data);
 275         size -= sizeof(struct ext4_extent_header);
 276         size /= sizeof(struct ext4_extent);
 277 #ifdef AGGRESSIVE_TEST
 278         if (!check && size > 3)
 279                 size = 3;
 280 #endif
 281         return size;
 282 }
 283
 284 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 285 {
 286         int size;
 287
 288         size = sizeof(EXT4_I(inode)->i_data);
 289         size -= sizeof(struct ext4_extent_header);
 290         size /= sizeof(struct ext4_extent_idx);
 291 #ifdef AGGRESSIVE_TEST
 292         if (!check && size > 4)
 293                 size = 4;
 294 #endif
 295         return size;
 296 }
 297
 298 static inline int
 299 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 300                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 301                            int nofail)
 302 {
 303         struct ext4_ext_path *path = *ppath;
 304         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 305
 306         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 307                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 308                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 309                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 310 }
 311
 312 /*
 313  * Calculate the number of metadata blocks needed
 314  * to allocate @blocks
 315  * Worse case is one block per extent
 316  */
 317 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 318 {
 319         struct ext4_inode_info *ei = EXT4_I(inode);
 320         int idxs;
 321
 322         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 323                 / sizeof(struct ext4_extent_idx));
 324
 325         /*
 326          * If the new delayed allocation block is contiguous with the
 327          * previous da block, it can share index blocks with the
 328          * previous block, so we only need to allocate a new index
 329          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 330          * an additional index block, and at ldxs**3 blocks, yet
 331          * another index blocks.
 332          */
 333         if (ei->i_da_metadata_calc_len &&
 334             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 335                 int num = 0;
 336
 337                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 338                         num++;
 339                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 340                         num++;
 341                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 342                         num++;
 343                         ei->i_da_metadata_calc_len = 0;
 344                 } else
 345                         ei->i_da_metadata_calc_len++;
 346                 ei->i_da_metadata_calc_last_lblock++;
 347                 return num;
 348         }
 349
 350         /*
 351          * In the worst case we need a new set of index blocks at
 352          * every level of the inode's extent tree.
 353          */
 354         ei->i_da_metadata_calc_len = 1;
 355         ei->i_da_metadata_calc_last_lblock = lblock;
 356         return ext_depth(inode) + 1;
 357 }
 358
 359 static int
 360 ext4_ext_max_entries(struct inode *inode, int depth)
 361 {
 362         int max;
 363
 364         if (depth == ext_depth(inode)) {
 365                 if (depth == 0)
 366                         max = ext4_ext_space_root(inode, 1);
 367                 else
 368                         max = ext4_ext_space_root_idx(inode, 1);
 369         } else {
 370                 if (depth == 0)
 371                         max = ext4_ext_space_block(inode, 1);
 372                 else
 373                         max = ext4_ext_space_block_idx(inode, 1);
 374         }
 375
 376         return max;
 377 }
 378
 379 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 380 {
 381         ext4_fsblk_t block = ext4_ext_pblock(ext);
 382         int len = ext4_ext_get_actual_len(ext);
 383         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 384
 385         /*
 386          * We allow neither:
 387          *  - zero length
 388          *  - overflow/wrap-around
 389          */
 390         if (lblock + len <= lblock)
 391                 return 0;
 392         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 393 }
 394
 395 static int ext4_valid_extent_idx(struct inode *inode,
 396                                 struct ext4_extent_idx *ext_idx)
 397 {
 398         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 399
 400         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 401 }
 402
 403 static int ext4_valid_extent_entries(struct inode *inode,
 404                                 struct ext4_extent_header *eh,
 405                                 int depth)
 406 {
 407         unsigned short entries;
 408         if (eh->eh_entries == 0)
 409                 return 1;
 410
 411         entries = le16_to_cpu(eh->eh_entries);
 412
 413         if (depth == 0) {
 414                 /* leaf entries */
 415                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 416                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 417                 ext4_fsblk_t pblock = 0;
 418                 ext4_lblk_t lblock = 0;
 419                 ext4_lblk_t prev = 0;
 420                 int len = 0;
 421                 while (entries) {
 422                         if (!ext4_valid_extent(inode, ext))
 423                                 return 0;
 424
 425                         /* Check for overlapping extents */
 426                         lblock = le32_to_cpu(ext->ee_block);
 427                         len = ext4_ext_get_actual_len(ext);
 428                         if ((lblock <= prev) && prev) {
 429                                 pblock = ext4_ext_pblock(ext);
 430                                 es->s_last_error_block = cpu_to_le64(pblock);
 431                                 return 0;
 432                         }
 433                         ext++;
 434                         entries--;
 435                         prev = lblock + len - 1;
 436                 }
 437         } else {
 438                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 439                 while (entries) {
 440                         if (!ext4_valid_extent_idx(inode, ext_idx))
 441                                 return 0;
 442                         ext_idx++;
 443                         entries--;
 444                 }
 445         }
 446         return 1;
 447 }
 448
 449 static int __ext4_ext_check(const char *function, unsigned int line,
 450                             struct inode *inode, struct ext4_extent_header *eh,
 451                             int depth, ext4_fsblk_t pblk)
 452 {
 453         const char *error_msg;
 454         int max = 0, err = -EFSCORRUPTED;
 455
 456         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 457                 error_msg = "invalid magic";
 458                 goto corrupted;
 459         }
 460         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 461                 error_msg = "unexpected eh_depth";
 462                 goto corrupted;
 463         }
 464         if (unlikely(eh->eh_max == 0)) {
 465                 error_msg = "invalid eh_max";
 466                 goto corrupted;
 467         }
 468         max = ext4_ext_max_entries(inode, depth);
 469         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 470                 error_msg = "too large eh_max";
 471                 goto corrupted;
 472         }
 473         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 474                 error_msg = "invalid eh_entries";
 475                 goto corrupted;
 476         }
 477         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 478                 error_msg = "invalid extent entries";
 479                 goto corrupted;
 480         }
 481         if (unlikely(depth > 32)) {
 482                 error_msg = "too large eh_depth";
 483                 goto corrupted;
 484         }
 485         /* Verify checksum on non-root extent tree nodes */
 486         if (ext_depth(inode) != depth &&
 487             !ext4_extent_block_csum_verify(inode, eh)) {
 488                 error_msg = "extent tree corrupted";
 489                 err = -EFSBADCRC;
 490                 goto corrupted;
 491         }
 492         return 0;
 493
 494 corrupted:
 495         ext4_error_inode(inode, function, line, 0,
 496                          "pblk %llu bad header/extent: %s - magic %x, "
 497                          "entries %u, max %u(%u), depth %u(%u)",
 498                          (unsigned long long) pblk, error_msg,
 499                          le16_to_cpu(eh->eh_magic),
 500                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 501                          max, le16_to_cpu(eh->eh_depth), depth);
 502         return err;
 503 }
 504
 505 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 506         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 507
 508 int ext4_ext_check_inode(struct inode *inode)
 509 {
 510         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 511 }
 512
 513 static struct buffer_head *
 514 __read_extent_tree_block(const char *function, unsigned int line,
 515                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 516                          int flags)
 517 {
 518         struct buffer_head              *bh;
 519         int                             err;
 520
 521         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
 522         if (unlikely(!bh))
 523                 return ERR_PTR(-ENOMEM);
 524
 525         if (!bh_uptodate_or_lock(bh)) {
 526                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 527                 err = bh_submit_read(bh);
 528                 if (err < 0)
 529                         goto errout;
 530         }
 531         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 532                 return bh;
 533         err = __ext4_ext_check(function, line, inode,
 534                                ext_block_hdr(bh), depth, pblk);
 535         if (err)
 536                 goto errout;
 537         set_buffer_verified(bh);
 538         /*
 539          * If this is a leaf block, cache all of its entries
 540          */
 541         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 542                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 543                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 544                 ext4_lblk_t prev = 0;
 545                 int i;
 546
 547                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 548                         unsigned int status = EXTENT_STATUS_WRITTEN;
 549                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 550                         int len = ext4_ext_get_actual_len(ex);
 551
 552                         if (prev && (prev != lblk))
 553                                 ext4_es_cache_extent(inode, prev,
 554                                                      lblk - prev, ~0,
 555                                                      EXTENT_STATUS_HOLE);
 556
 557                         if (ext4_ext_is_unwritten(ex))
 558                                 status = EXTENT_STATUS_UNWRITTEN;
 559                         ext4_es_cache_extent(inode, lblk, len,
 560                                              ext4_ext_pblock(ex), status);
 561                         prev = lblk + len;
 562                 }
 563         }
 564         return bh;
 565 errout:
 566         put_bh(bh);
 567         return ERR_PTR(err);
 568
 569 }
 570
 571 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 572         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 573                                  (depth), (flags))
 574
 575 /*
 576  * This function is called to cache a file's extent information in the
 577  * extent status tree
 578  */
 579 int ext4_ext_precache(struct inode *inode)
 580 {
 581         struct ext4_inode_info *ei = EXT4_I(inode);
 582         struct ext4_ext_path *path = NULL;
 583         struct buffer_head *bh;
 584         int i = 0, depth, ret = 0;
 585
 586         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 587                 return 0;       /* not an extent-mapped inode */
 588
 589         down_read(&ei->i_data_sem);
 590         depth = ext_depth(inode);
 591
 592         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 593                        GFP_NOFS);
 594         if (path == NULL) {
 595                 up_read(&ei->i_data_sem);
 596                 return -ENOMEM;
 597         }
 598
 599         /* Don't cache anything if there are no external extent blocks */
 600         if (depth == 0)
 601                 goto out;
 602         path[0].p_hdr = ext_inode_hdr(inode);
 603         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 604         if (ret)
 605                 goto out;
 606         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 607         while (i >= 0) {
 608                 /*
 609                  * If this is a leaf block or we've reached the end of
 610                  * the index block, go up
 611                  */
 612                 if ((i == depth) ||
 613                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 614                         brelse(path[i].p_bh);
 615                         path[i].p_bh = NULL;
 616                         i--;
 617                         continue;
 618                 }
 619                 bh = read_extent_tree_block(inode,
 620                                             ext4_idx_pblock(path[i].p_idx++),
 621                                             depth - i - 1,
 622                                             EXT4_EX_FORCE_CACHE);
 623                 if (IS_ERR(bh)) {
 624                         ret = PTR_ERR(bh);
 625                         break;
 626                 }
 627                 i++;
 628                 path[i].p_bh = bh;
 629                 path[i].p_hdr = ext_block_hdr(bh);
 630                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 631         }
 632         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 633 out:
 634         up_read(&ei->i_data_sem);
 635         ext4_ext_drop_refs(path);
 636         kfree(path);
 637         return ret;
 638 }
 639
 640 #ifdef EXT_DEBUG
 641 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 642 {
 643         int k, l = path->p_depth;
 644
 645         ext_debug("path:");
 646         for (k = 0; k <= l; k++, path++) {
 647                 if (path->p_idx) {
 648                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 649                             ext4_idx_pblock(path->p_idx));
 650                 } else if (path->p_ext) {
 651                         ext_debug("  %d:[%d]%d:%llu ",
 652                                   le32_to_cpu(path->p_ext->ee_block),
 653                                   ext4_ext_is_unwritten(path->p_ext),
 654                                   ext4_ext_get_actual_len(path->p_ext),
 655                                   ext4_ext_pblock(path->p_ext));
 656                 } else
 657                         ext_debug("  []");
 658         }
 659         ext_debug("\n");
 660 }
 661
 662 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 663 {
 664         int depth = ext_depth(inode);
 665         struct ext4_extent_header *eh;
 666         struct ext4_extent *ex;
 667         int i;
 668
 669         if (!path)
 670                 return;
 671
 672         eh = path[depth].p_hdr;
 673         ex = EXT_FIRST_EXTENT(eh);
 674
 675         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 676
 677         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 678                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 679                           ext4_ext_is_unwritten(ex),
 680                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 681         }
 682         ext_debug("\n");
 683 }
 684
 685 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 686                         ext4_fsblk_t newblock, int level)
 687 {
 688         int depth = ext_depth(inode);
 689         struct ext4_extent *ex;
 690
 691         if (depth != level) {
 692                 struct ext4_extent_idx *idx;
 693                 idx = path[level].p_idx;
 694                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 695                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 696                                         le32_to_cpu(idx->ei_block),
 697                                         ext4_idx_pblock(idx),
 698                                         newblock);
 699                         idx++;
 700                 }
 701
 702                 return;
 703         }
 704
 705         ex = path[depth].p_ext;
 706         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 707                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 708                                 le32_to_cpu(ex->ee_block),
 709                                 ext4_ext_pblock(ex),
 710                                 ext4_ext_is_unwritten(ex),
 711                                 ext4_ext_get_actual_len(ex),
 712                                 newblock);
 713                 ex++;
 714         }
 715 }
 716
 717 #else
 718 #define ext4_ext_show_path(inode, path)
 719 #define ext4_ext_show_leaf(inode, path)
 720 #define ext4_ext_show_move(inode, path, newblock, level)
 721 #endif
 722
 723 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 724 {
 725         int depth, i;
 726
 727         if (!path)
 728                 return;
 729         depth = path->p_depth;
 730         for (i = 0; i <= depth; i++, path++)
 731                 if (path->p_bh) {
 732                         brelse(path->p_bh);
 733                         path->p_bh = NULL;
 734                 }
 735 }
 736
 737 /*
 738  * ext4_ext_binsearch_idx:
 739  * binary search for the closest index of the given block
 740  * the header must be checked before calling this
 741  */
 742 static void
 743 ext4_ext_binsearch_idx(struct inode *inode,
 744                         struct ext4_ext_path *path, ext4_lblk_t block)
 745 {
 746         struct ext4_extent_header *eh = path->p_hdr;
 747         struct ext4_extent_idx *r, *l, *m;
 748
 749
 750         ext_debug("binsearch for %u(idx):  ", block);
 751
 752         l = EXT_FIRST_INDEX(eh) + 1;
 753         r = EXT_LAST_INDEX(eh);
 754         while (l <= r) {
 755                 m = l + (r - l) / 2;
 756                 if (block < le32_to_cpu(m->ei_block))
 757                         r = m - 1;
 758                 else
 759                         l = m + 1;
 760                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 761                                 m, le32_to_cpu(m->ei_block),
 762                                 r, le32_to_cpu(r->ei_block));
 763         }
 764
 765         path->p_idx = l - 1;
 766         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 767                   ext4_idx_pblock(path->p_idx));
 768
 769 #ifdef CHECK_BINSEARCH
 770         {
 771                 struct ext4_extent_idx *chix, *ix;
 772                 int k;
 773
 774                 chix = ix = EXT_FIRST_INDEX(eh);
 775                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 776                   if (k != 0 &&
 777                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 778                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 779                                        "first=0x%p\n", k,
 780                                        ix, EXT_FIRST_INDEX(eh));
 781                                 printk(KERN_DEBUG "%u <= %u\n",
 782                                        le32_to_cpu(ix->ei_block),
 783                                        le32_to_cpu(ix[-1].ei_block));
 784                         }
 785                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 786                                            <= le32_to_cpu(ix[-1].ei_block));
 787                         if (block < le32_to_cpu(ix->ei_block))
 788                                 break;
 789                         chix = ix;
 790                 }
 791                 BUG_ON(chix != path->p_idx);
 792         }
 793 #endif
 794
 795 }
 796
 797 /*
 798  * ext4_ext_binsearch:
 799  * binary search for closest extent of the given block
 800  * the header must be checked before calling this
 801  */
 802 static void
 803 ext4_ext_binsearch(struct inode *inode,
 804                 struct ext4_ext_path *path, ext4_lblk_t block)
 805 {
 806         struct ext4_extent_header *eh = path->p_hdr;
 807         struct ext4_extent *r, *l, *m;
 808
 809         if (eh->eh_entries == 0) {
 810                 /*
 811                  * this leaf is empty:
 812                  * we get such a leaf in split/add case
 813                  */
 814                 return;
 815         }
 816
 817         ext_debug("binsearch for %u:  ", block);
 818
 819         l = EXT_FIRST_EXTENT(eh) + 1;
 820         r = EXT_LAST_EXTENT(eh);
 821
 822         while (l <= r) {
 823                 m = l + (r - l) / 2;
 824                 if (block < le32_to_cpu(m->ee_block))
 825                         r = m - 1;
 826                 else
 827                         l = m + 1;
 828                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 829                                 m, le32_to_cpu(m->ee_block),
 830                                 r, le32_to_cpu(r->ee_block));
 831         }
 832
 833         path->p_ext = l - 1;
 834         ext_debug("  -> %d:%llu:[%d]%d ",
 835                         le32_to_cpu(path->p_ext->ee_block),
 836                         ext4_ext_pblock(path->p_ext),
 837                         ext4_ext_is_unwritten(path->p_ext),
 838                         ext4_ext_get_actual_len(path->p_ext));
 839
 840 #ifdef CHECK_BINSEARCH
 841         {
 842                 struct ext4_extent *chex, *ex;
 843                 int k;
 844
 845                 chex = ex = EXT_FIRST_EXTENT(eh);
 846                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 847                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 848                                           <= le32_to_cpu(ex[-1].ee_block));
 849                         if (block < le32_to_cpu(ex->ee_block))
 850                                 break;
 851                         chex = ex;
 852                 }
 853                 BUG_ON(chex != path->p_ext);
 854         }
 855 #endif
 856
 857 }
 858
 859 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 860 {
 861         struct ext4_extent_header *eh;
 862
 863         eh = ext_inode_hdr(inode);
 864         eh->eh_depth = 0;
 865         eh->eh_entries = 0;
 866         eh->eh_magic = EXT4_EXT_MAGIC;
 867         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 868         ext4_mark_inode_dirty(handle, inode);
 869         return 0;
 870 }
 871
 872 struct ext4_ext_path *
 873 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 874                  struct ext4_ext_path **orig_path, int flags)
 875 {
 876         struct ext4_extent_header *eh;
 877         struct buffer_head *bh;
 878         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 879         short int depth, i, ppos = 0;
 880         int ret;
 881
 882         eh = ext_inode_hdr(inode);
 883         depth = ext_depth(inode);
 884
 885         if (path) {
 886                 ext4_ext_drop_refs(path);
 887                 if (depth > path[0].p_maxdepth) {
 888                         kfree(path);
 889                         *orig_path = path = NULL;
 890                 }
 891         }
 892         if (!path) {
 893                 /* account possible depth increase */
 894                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 895                                 GFP_NOFS);
 896                 if (unlikely(!path))
 897                         return ERR_PTR(-ENOMEM);
 898                 path[0].p_maxdepth = depth + 1;
 899         }
 900         path[0].p_hdr = eh;
 901         path[0].p_bh = NULL;
 902
 903         i = depth;
 904         /* walk through the tree */
 905         while (i) {
 906                 ext_debug("depth %d: num %d, max %d\n",
 907                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 908
 909                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 910                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 911                 path[ppos].p_depth = i;
 912                 path[ppos].p_ext = NULL;
 913
 914                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 915                                             flags);
 916                 if (IS_ERR(bh)) {
 917                         ret = PTR_ERR(bh);
 918                         goto err;
 919                 }
 920
 921                 eh = ext_block_hdr(bh);
 922                 ppos++;
 923                 path[ppos].p_bh = bh;
 924                 path[ppos].p_hdr = eh;
 925         }
 926
 927         path[ppos].p_depth = i;
 928         path[ppos].p_ext = NULL;
 929         path[ppos].p_idx = NULL;
 930
 931         /* find extent */
 932         ext4_ext_binsearch(inode, path + ppos, block);
 933         /* if not an empty leaf */
 934         if (path[ppos].p_ext)
 935                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 936
 937         ext4_ext_show_path(inode, path);
 938
 939         return path;
 940
 941 err:
 942         ext4_ext_drop_refs(path);
 943         kfree(path);
 944         if (orig_path)
 945                 *orig_path = NULL;
 946         return ERR_PTR(ret);
 947 }
 948
 949 /*
 950  * ext4_ext_insert_index:
 951  * insert new index [@logical;@ptr] into the block at @curp;
 952  * check where to insert: before @curp or after @curp
 953  */
 954 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 955                                  struct ext4_ext_path *curp,
 956                                  int logical, ext4_fsblk_t ptr)
 957 {
 958         struct ext4_extent_idx *ix;
 959         int len, err;
 960
 961         err = ext4_ext_get_access(handle, inode, curp);
 962         if (err)
 963                 return err;
 964
 965         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 966                 EXT4_ERROR_INODE(inode,
 967                                  "logical %d == ei_block %d!",
 968                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 969                 return -EFSCORRUPTED;
 970         }
 971
 972         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 973                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 974                 EXT4_ERROR_INODE(inode,
 975                                  "eh_entries %d >= eh_max %d!",
 976                                  le16_to_cpu(curp->p_hdr->eh_entries),
 977                                  le16_to_cpu(curp->p_hdr->eh_max));
 978                 return -EFSCORRUPTED;
 979         }
 980
 981         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 982                 /* insert after */
 983                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 984                 ix = curp->p_idx + 1;
 985         } else {
 986                 /* insert before */
 987                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 988                 ix = curp->p_idx;
 989         }
 990
 991         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 992         BUG_ON(len < 0);
 993         if (len > 0) {
 994                 ext_debug("insert new index %d: "
 995                                 "move %d indices from 0x%p to 0x%p\n",
 996                                 logical, len, ix, ix + 1);
 997                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 998         }
 999
1000         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
1001                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
1002                 return -EFSCORRUPTED;
1003         }
1004
1005         ix->ei_block = cpu_to_le32(logical);
1006         ext4_idx_store_pblock(ix, ptr);
1007         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1008
1009         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1010                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1011                 return -EFSCORRUPTED;
1012         }
1013
1014         err = ext4_ext_dirty(handle, inode, curp);
1015         ext4_std_error(inode->i_sb, err);
1016
1017         return err;
1018 }
1019
1020 /*
1021  * ext4_ext_split:
1022  * inserts new subtree into the path, using free index entry
1023  * at depth @at:
1024  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1025  * - makes decision where to split
1026  * - moves remaining extents and index entries (right to the split point)
1027  *   into the newly allocated blocks
1028  * - initializes subtree
1029  */
1030 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1031                           unsigned int flags,
1032                           struct ext4_ext_path *path,
1033                           struct ext4_extent *newext, int at)
1034 {
1035         struct buffer_head *bh = NULL;
1036         int depth = ext_depth(inode);
1037         struct ext4_extent_header *neh;
1038         struct ext4_extent_idx *fidx;
1039         int i = at, k, m, a;
1040         ext4_fsblk_t newblock, oldblock;
1041         __le32 border;
1042         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1043         int err = 0;
1044
1045         /* make decision: where to split? */
1046         /* FIXME: now decision is simplest: at current extent */
1047
1048         /* if current leaf will be split, then we should use
1049          * border from split point */
1050         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1051                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1052                 return -EFSCORRUPTED;
1053         }
1054         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1055                 border = path[depth].p_ext[1].ee_block;
1056                 ext_debug("leaf will be split."
1057                                 " next leaf starts at %d\n",
1058                                   le32_to_cpu(border));
1059         } else {
1060                 border = newext->ee_block;
1061                 ext_debug("leaf will be added."
1062                                 " next leaf starts at %d\n",
1063                                 le32_to_cpu(border));
1064         }
1065
1066         /*
1067          * If error occurs, then we break processing
1068          * and mark filesystem read-only. index won't
1069          * be inserted and tree will be in consistent
1070          * state. Next mount will repair buffers too.
1071          */
1072
1073         /*
1074          * Get array to track all allocated blocks.
1075          * We need this to handle errors and free blocks
1076          * upon them.
1077          */
1078         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1079         if (!ablocks)
1080                 return -ENOMEM;
1081
1082         /* allocate all needed blocks */
1083         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1084         for (a = 0; a < depth - at; a++) {
1085                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1086                                                    newext, &err, flags);
1087                 if (newblock == 0)
1088                         goto cleanup;
1089                 ablocks[a] = newblock;
1090         }
1091
1092         /* initialize new leaf */
1093         newblock = ablocks[--a];
1094         if (unlikely(newblock == 0)) {
1095                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1096                 err = -EFSCORRUPTED;
1097                 goto cleanup;
1098         }
1099         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1100         if (unlikely(!bh)) {
1101                 err = -ENOMEM;
1102                 goto cleanup;
1103         }
1104         lock_buffer(bh);
1105
1106         err = ext4_journal_get_create_access(handle, bh);
1107         if (err)
1108                 goto cleanup;
1109
1110         neh = ext_block_hdr(bh);
1111         neh->eh_entries = 0;
1112         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1113         neh->eh_magic = EXT4_EXT_MAGIC;
1114         neh->eh_depth = 0;
1115
1116         /* move remainder of path[depth] to the new leaf */
1117         if (unlikely(path[depth].p_hdr->eh_entries !=
1118                      path[depth].p_hdr->eh_max)) {
1119                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1120                                  path[depth].p_hdr->eh_entries,
1121                                  path[depth].p_hdr->eh_max);
1122                 err = -EFSCORRUPTED;
1123                 goto cleanup;
1124         }
1125         /* start copy from next extent */
1126         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1127         ext4_ext_show_move(inode, path, newblock, depth);
1128         if (m) {
1129                 struct ext4_extent *ex;
1130                 ex = EXT_FIRST_EXTENT(neh);
1131                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1132                 le16_add_cpu(&neh->eh_entries, m);
1133         }
1134
1135         ext4_extent_block_csum_set(inode, neh);
1136         set_buffer_uptodate(bh);
1137         unlock_buffer(bh);
1138
1139         err = ext4_handle_dirty_metadata(handle, inode, bh);
1140         if (err)
1141                 goto cleanup;
1142         brelse(bh);
1143         bh = NULL;
1144
1145         /* correct old leaf */
1146         if (m) {
1147                 err = ext4_ext_get_access(handle, inode, path + depth);
1148                 if (err)
1149                         goto cleanup;
1150                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1151                 err = ext4_ext_dirty(handle, inode, path + depth);
1152                 if (err)
1153                         goto cleanup;
1154
1155         }
1156
1157         /* create intermediate indexes */
1158         k = depth - at - 1;
1159         if (unlikely(k < 0)) {
1160                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1161                 err = -EFSCORRUPTED;
1162                 goto cleanup;
1163         }
1164         if (k)
1165                 ext_debug("create %d intermediate indices\n", k);
1166         /* insert new index into current index block */
1167         /* current depth stored in i var */
1168         i = depth - 1;
1169         while (k--) {
1170                 oldblock = newblock;
1171                 newblock = ablocks[--a];
1172                 bh = sb_getblk(inode->i_sb, newblock);
1173                 if (unlikely(!bh)) {
1174                         err = -ENOMEM;
1175                         goto cleanup;
1176                 }
1177                 lock_buffer(bh);
1178
1179                 err = ext4_journal_get_create_access(handle, bh);
1180                 if (err)
1181                         goto cleanup;
1182
1183                 neh = ext_block_hdr(bh);
1184                 neh->eh_entries = cpu_to_le16(1);
1185                 neh->eh_magic = EXT4_EXT_MAGIC;
1186                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1187                 neh->eh_depth = cpu_to_le16(depth - i);
1188                 fidx = EXT_FIRST_INDEX(neh);
1189                 fidx->ei_block = border;
1190                 ext4_idx_store_pblock(fidx, oldblock);
1191
1192                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1193                                 i, newblock, le32_to_cpu(border), oldblock);
1194
1195                 /* move remainder of path[i] to the new index block */
1196                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1197                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1198                         EXT4_ERROR_INODE(inode,
1199                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1200                                          le32_to_cpu(path[i].p_ext->ee_block));
1201                         err = -EFSCORRUPTED;
1202                         goto cleanup;
1203                 }
1204                 /* start copy indexes */
1205                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1206                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1207                                 EXT_MAX_INDEX(path[i].p_hdr));
1208                 ext4_ext_show_move(inode, path, newblock, i);
1209                 if (m) {
1210                         memmove(++fidx, path[i].p_idx,
1211                                 sizeof(struct ext4_extent_idx) * m);
1212                         le16_add_cpu(&neh->eh_entries, m);
1213                 }
1214                 ext4_extent_block_csum_set(inode, neh);
1215                 set_buffer_uptodate(bh);
1216                 unlock_buffer(bh);
1217
1218                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1219                 if (err)
1220                         goto cleanup;
1221                 brelse(bh);
1222                 bh = NULL;
1223
1224                 /* correct old index */
1225                 if (m) {
1226                         err = ext4_ext_get_access(handle, inode, path + i);
1227                         if (err)
1228                                 goto cleanup;
1229                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1230                         err = ext4_ext_dirty(handle, inode, path + i);
1231                         if (err)
1232                                 goto cleanup;
1233                 }
1234
1235                 i--;
1236         }
1237
1238         /* insert new index */
1239         err = ext4_ext_insert_index(handle, inode, path + at,
1240                                     le32_to_cpu(border), newblock);
1241
1242 cleanup:
1243         if (bh) {
1244                 if (buffer_locked(bh))
1245                         unlock_buffer(bh);
1246                 brelse(bh);
1247         }
1248
1249         if (err) {
1250                 /* free all allocated blocks in error case */
1251                 for (i = 0; i < depth; i++) {
1252                         if (!ablocks[i])
1253                                 continue;
1254                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1255                                          EXT4_FREE_BLOCKS_METADATA);
1256                 }
1257         }
1258         kfree(ablocks);
1259
1260         return err;
1261 }
1262
1263 /*
1264  * ext4_ext_grow_indepth:
1265  * implements tree growing procedure:
1266  * - allocates new block
1267  * - moves top-level data (index block or leaf) into the new block
1268  * - initializes new top-level, creating index that points to the
1269  *   just created block
1270  */
1271 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1272                                  unsigned int flags)
1273 {
1274         struct ext4_extent_header *neh;
1275         struct buffer_head *bh;
1276         ext4_fsblk_t newblock, goal = 0;
1277         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1278         int err = 0;
1279
1280         /* Try to prepend new index to old one */
1281         if (ext_depth(inode))
1282                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1283         if (goal > le32_to_cpu(es->s_first_data_block)) {
1284                 flags |= EXT4_MB_HINT_TRY_GOAL;
1285                 goal--;
1286         } else
1287                 goal = ext4_inode_to_goal_block(inode);
1288         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1289                                         NULL, &err);
1290         if (newblock == 0)
1291                 return err;
1292
1293         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1294         if (unlikely(!bh))
1295                 return -ENOMEM;
1296         lock_buffer(bh);
1297
1298         err = ext4_journal_get_create_access(handle, bh);
1299         if (err) {
1300                 unlock_buffer(bh);
1301                 goto out;
1302         }
1303
1304         /* move top-level index/leaf into new block */
1305         memmove(bh->b_data, EXT4_I(inode)->i_data,
1306                 sizeof(EXT4_I(inode)->i_data));
1307
1308         /* set size of new block */
1309         neh = ext_block_hdr(bh);
1310         /* old root could have indexes or leaves
1311          * so calculate e_max right way */
1312         if (ext_depth(inode))
1313                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1314         else
1315                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1316         neh->eh_magic = EXT4_EXT_MAGIC;
1317         ext4_extent_block_csum_set(inode, neh);
1318         set_buffer_uptodate(bh);
1319         unlock_buffer(bh);
1320
1321         err = ext4_handle_dirty_metadata(handle, inode, bh);
1322         if (err)
1323                 goto out;
1324
1325         /* Update top-level index: num,max,pointer */
1326         neh = ext_inode_hdr(inode);
1327         neh->eh_entries = cpu_to_le16(1);
1328         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1329         if (neh->eh_depth == 0) {
1330                 /* Root extent block becomes index block */
1331                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1332                 EXT_FIRST_INDEX(neh)->ei_block =
1333                         EXT_FIRST_EXTENT(neh)->ee_block;
1334         }
1335         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1336                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1337                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1338                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1339
1340         le16_add_cpu(&neh->eh_depth, 1);
1341         ext4_mark_inode_dirty(handle, inode);
1342 out:
1343         brelse(bh);
1344
1345         return err;
1346 }
1347
1348 /*
1349  * ext4_ext_create_new_leaf:
1350  * finds empty index and adds new leaf.
1351  * if no free index is found, then it requests in-depth growing.
1352  */
1353 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1354                                     unsigned int mb_flags,
1355                                     unsigned int gb_flags,
1356                                     struct ext4_ext_path **ppath,
1357                                     struct ext4_extent *newext)
1358 {
1359         struct ext4_ext_path *path = *ppath;
1360         struct ext4_ext_path *curp;
1361         int depth, i, err = 0;
1362
1363 repeat:
1364         i = depth = ext_depth(inode);
1365
1366         /* walk up to the tree and look for free index entry */
1367         curp = path + depth;
1368         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1369                 i--;
1370                 curp--;
1371         }
1372
1373         /* we use already allocated block for index block,
1374          * so subsequent data blocks should be contiguous */
1375         if (EXT_HAS_FREE_INDEX(curp)) {
1376                 /* if we found index with free entry, then use that
1377                  * entry: create all needed subtree and add new leaf */
1378                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1379                 if (err)
1380                         goto out;
1381
1382                 /* refill path */
1383                 path = ext4_find_extent(inode,
1384                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1385                                     ppath, gb_flags);
1386                 if (IS_ERR(path))
1387                         err = PTR_ERR(path);
1388         } else {
1389                 /* tree is full, time to grow in depth */
1390                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1391                 if (err)
1392                         goto out;
1393
1394                 /* refill path */
1395                 path = ext4_find_extent(inode,
1396                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1397                                     ppath, gb_flags);
1398                 if (IS_ERR(path)) {
1399                         err = PTR_ERR(path);
1400                         goto out;
1401                 }
1402
1403                 /*
1404                  * only first (depth 0 -> 1) produces free space;
1405                  * in all other cases we have to split the grown tree
1406                  */
1407                 depth = ext_depth(inode);
1408                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1409                         /* now we need to split */
1410                         goto repeat;
1411                 }
1412         }
1413
1414 out:
1415         return err;
1416 }
1417
1418 /*
1419  * search the closest allocated block to the left for *logical
1420  * and returns it at @logical + it's physical address at @phys
1421  * if *logical is the smallest allocated block, the function
1422  * returns 0 at @phys
1423  * return value contains 0 (success) or error code
1424  */
1425 static int ext4_ext_search_left(struct inode *inode,
1426                                 struct ext4_ext_path *path,
1427                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1428 {
1429         struct ext4_extent_idx *ix;
1430         struct ext4_extent *ex;
1431         int depth, ee_len;
1432
1433         if (unlikely(path == NULL)) {
1434                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1435                 return -EFSCORRUPTED;
1436         }
1437         depth = path->p_depth;
1438         *phys = 0;
1439
1440         if (depth == 0 && path->p_ext == NULL)
1441                 return 0;
1442
1443         /* usually extent in the path covers blocks smaller
1444          * then *logical, but it can be that extent is the
1445          * first one in the file */
1446
1447         ex = path[depth].p_ext;
1448         ee_len = ext4_ext_get_actual_len(ex);
1449         if (*logical < le32_to_cpu(ex->ee_block)) {
1450                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1451                         EXT4_ERROR_INODE(inode,
1452                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1453                                          *logical, le32_to_cpu(ex->ee_block));
1454                         return -EFSCORRUPTED;
1455                 }
1456                 while (--depth >= 0) {
1457                         ix = path[depth].p_idx;
1458                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1459                                 EXT4_ERROR_INODE(inode,
1460                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1461                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1462                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1463                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1464                                   depth);
1465                                 return -EFSCORRUPTED;
1466                         }
1467                 }
1468                 return 0;
1469         }
1470
1471         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1472                 EXT4_ERROR_INODE(inode,
1473                                  "logical %d < ee_block %d + ee_len %d!",
1474                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1475                 return -EFSCORRUPTED;
1476         }
1477
1478         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1479         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1480         return 0;
1481 }
1482
1483 /*
1484  * search the closest allocated block to the right for *logical
1485  * and returns it at @logical + it's physical address at @phys
1486  * if *logical is the largest allocated block, the function
1487  * returns 0 at @phys
1488  * return value contains 0 (success) or error code
1489  */
1490 static int ext4_ext_search_right(struct inode *inode,
1491                                  struct ext4_ext_path *path,
1492                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1493                                  struct ext4_extent **ret_ex)
1494 {
1495         struct buffer_head *bh = NULL;
1496         struct ext4_extent_header *eh;
1497         struct ext4_extent_idx *ix;
1498         struct ext4_extent *ex;
1499         ext4_fsblk_t block;
1500         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1501         int ee_len;
1502
1503         if (unlikely(path == NULL)) {
1504                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1505                 return -EFSCORRUPTED;
1506         }
1507         depth = path->p_depth;
1508         *phys = 0;
1509
1510         if (depth == 0 && path->p_ext == NULL)
1511                 return 0;
1512
1513         /* usually extent in the path covers blocks smaller
1514          * then *logical, but it can be that extent is the
1515          * first one in the file */
1516
1517         ex = path[depth].p_ext;
1518         ee_len = ext4_ext_get_actual_len(ex);
1519         if (*logical < le32_to_cpu(ex->ee_block)) {
1520                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1521                         EXT4_ERROR_INODE(inode,
1522                                          "first_extent(path[%d].p_hdr) != ex",
1523                                          depth);
1524                         return -EFSCORRUPTED;
1525                 }
1526                 while (--depth >= 0) {
1527                         ix = path[depth].p_idx;
1528                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1529                                 EXT4_ERROR_INODE(inode,
1530                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1531                                                  *logical);
1532                                 return -EFSCORRUPTED;
1533                         }
1534                 }
1535                 goto found_extent;
1536         }
1537
1538         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1539                 EXT4_ERROR_INODE(inode,
1540                                  "logical %d < ee_block %d + ee_len %d!",
1541                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1542                 return -EFSCORRUPTED;
1543         }
1544
1545         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1546                 /* next allocated block in this leaf */
1547                 ex++;
1548                 goto found_extent;
1549         }
1550
1551         /* go up and search for index to the right */
1552         while (--depth >= 0) {
1553                 ix = path[depth].p_idx;
1554                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1555                         goto got_index;
1556         }
1557
1558         /* we've gone up to the root and found no index to the right */
1559         return 0;
1560
1561 got_index:
1562         /* we've found index to the right, let's
1563          * follow it and find the closest allocated
1564          * block to the right */
1565         ix++;
1566         block = ext4_idx_pblock(ix);
1567         while (++depth < path->p_depth) {
1568                 /* subtract from p_depth to get proper eh_depth */
1569                 bh = read_extent_tree_block(inode, block,
1570                                             path->p_depth - depth, 0);
1571                 if (IS_ERR(bh))
1572                         return PTR_ERR(bh);
1573                 eh = ext_block_hdr(bh);
1574                 ix = EXT_FIRST_INDEX(eh);
1575                 block = ext4_idx_pblock(ix);
1576                 put_bh(bh);
1577         }
1578
1579         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1580         if (IS_ERR(bh))
1581                 return PTR_ERR(bh);
1582         eh = ext_block_hdr(bh);
1583         ex = EXT_FIRST_EXTENT(eh);
1584 found_extent:
1585         *logical = le32_to_cpu(ex->ee_block);
1586         *phys = ext4_ext_pblock(ex);
1587         *ret_ex = ex;
1588         if (bh)
1589                 put_bh(bh);
1590         return 0;
1591 }
1592
1593 /*
1594  * ext4_ext_next_allocated_block:
1595  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1596  * NOTE: it considers block number from index entry as
1597  * allocated block. Thus, index entries have to be consistent
1598  * with leaves.
1599  */
1600 ext4_lblk_t
1601 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1602 {
1603         int depth;
1604
1605         BUG_ON(path == NULL);
1606         depth = path->p_depth;
1607
1608         if (depth == 0 && path->p_ext == NULL)
1609                 return EXT_MAX_BLOCKS;
1610
1611         while (depth >= 0) {
1612                 if (depth == path->p_depth) {
1613                         /* leaf */
1614                         if (path[depth].p_ext &&
1615                                 path[depth].p_ext !=
1616                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1617                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1618                 } else {
1619                         /* index */
1620                         if (path[depth].p_idx !=
1621                                         EXT_LAST_INDEX(path[depth].p_hdr))
1622                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1623                 }
1624                 depth--;
1625         }
1626
1627         return EXT_MAX_BLOCKS;
1628 }
1629
1630 /*
1631  * ext4_ext_next_leaf_block:
1632  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1633  */
1634 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1635 {
1636         int depth;
1637
1638         BUG_ON(path == NULL);
1639         depth = path->p_depth;
1640
1641         /* zero-tree has no leaf blocks at all */
1642         if (depth == 0)
1643                 return EXT_MAX_BLOCKS;
1644
1645         /* go to index block */
1646         depth--;
1647
1648         while (depth >= 0) {
1649                 if (path[depth].p_idx !=
1650                                 EXT_LAST_INDEX(path[depth].p_hdr))
1651                         return (ext4_lblk_t)
1652                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1653                 depth--;
1654         }
1655
1656         return EXT_MAX_BLOCKS;
1657 }
1658
1659 /*
1660  * ext4_ext_correct_indexes:
1661  * if leaf gets modified and modified extent is first in the leaf,
1662  * then we have to correct all indexes above.
1663  * TODO: do we need to correct tree in all cases?
1664  */
1665 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1666                                 struct ext4_ext_path *path)
1667 {
1668         struct ext4_extent_header *eh;
1669         int depth = ext_depth(inode);
1670         struct ext4_extent *ex;
1671         __le32 border;
1672         int k, err = 0;
1673
1674         eh = path[depth].p_hdr;
1675         ex = path[depth].p_ext;
1676
1677         if (unlikely(ex == NULL || eh == NULL)) {
1678                 EXT4_ERROR_INODE(inode,
1679                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1680                 return -EFSCORRUPTED;
1681         }
1682
1683         if (depth == 0) {
1684                 /* there is no tree at all */
1685                 return 0;
1686         }
1687
1688         if (ex != EXT_FIRST_EXTENT(eh)) {
1689                 /* we correct tree if first leaf got modified only */
1690                 return 0;
1691         }
1692
1693         /*
1694          * TODO: we need correction if border is smaller than current one
1695          */
1696         k = depth - 1;
1697         border = path[depth].p_ext->ee_block;
1698         err = ext4_ext_get_access(handle, inode, path + k);
1699         if (err)
1700                 return err;
1701         path[k].p_idx->ei_block = border;
1702         err = ext4_ext_dirty(handle, inode, path + k);
1703         if (err)
1704                 return err;
1705
1706         while (k--) {
1707                 /* change all left-side indexes */
1708                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1709                         break;
1710                 err = ext4_ext_get_access(handle, inode, path + k);
1711                 if (err)
1712                         break;
1713                 path[k].p_idx->ei_block = border;
1714                 err = ext4_ext_dirty(handle, inode, path + k);
1715                 if (err)
1716                         break;
1717         }
1718
1719         return err;
1720 }
1721
1722 int
1723 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1724                                 struct ext4_extent *ex2)
1725 {
1726         unsigned short ext1_ee_len, ext2_ee_len;
1727
1728         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1729                 return 0;
1730
1731         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1732         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1733
1734         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1735                         le32_to_cpu(ex2->ee_block))
1736                 return 0;
1737
1738         /*
1739          * To allow future support for preallocated extents to be added
1740          * as an RO_COMPAT feature, refuse to merge to extents if
1741          * this can result in the top bit of ee_len being set.
1742          */
1743         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1744                 return 0;
1745         /*
1746          * The check for IO to unwritten extent is somewhat racy as we
1747          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1748          * dropping i_data_sem. But reserved blocks should save us in that
1749          * case.
1750          */
1751         if (ext4_ext_is_unwritten(ex1) &&
1752             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1753              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1754              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1755                 return 0;
1756 #ifdef AGGRESSIVE_TEST
1757         if (ext1_ee_len >= 4)
1758                 return 0;
1759 #endif
1760
1761         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1762                 return 1;
1763         return 0;
1764 }
1765
1766 /*
1767  * This function tries to merge the "ex" extent to the next extent in the tree.
1768  * It always tries to merge towards right. If you want to merge towards
1769  * left, pass "ex - 1" as argument instead of "ex".
1770  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1771  * 1 if they got merged.
1772  */
1773 static int ext4_ext_try_to_merge_right(struct inode *inode,
1774                                  struct ext4_ext_path *path,
1775                                  struct ext4_extent *ex)
1776 {
1777         struct ext4_extent_header *eh;
1778         unsigned int depth, len;
1779         int merge_done = 0, unwritten;
1780
1781         depth = ext_depth(inode);
1782         BUG_ON(path[depth].p_hdr == NULL);
1783         eh = path[depth].p_hdr;
1784
1785         while (ex < EXT_LAST_EXTENT(eh)) {
1786                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1787                         break;
1788                 /* merge with next extent! */
1789                 unwritten = ext4_ext_is_unwritten(ex);
1790                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1791                                 + ext4_ext_get_actual_len(ex + 1));
1792                 if (unwritten)
1793                         ext4_ext_mark_unwritten(ex);
1794
1795                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1796                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1797                                 * sizeof(struct ext4_extent);
1798                         memmove(ex + 1, ex + 2, len);
1799                 }
1800                 le16_add_cpu(&eh->eh_entries, -1);
1801                 merge_done = 1;
1802                 WARN_ON(eh->eh_entries == 0);
1803                 if (!eh->eh_entries)
1804                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1805         }
1806
1807         return merge_done;
1808 }
1809
1810 /*
1811  * This function does a very simple check to see if we can collapse
1812  * an extent tree with a single extent tree leaf block into the inode.
1813  */
1814 static void ext4_ext_try_to_merge_up(handle_t *handle,
1815                                      struct inode *inode,
1816                                      struct ext4_ext_path *path)
1817 {
1818         size_t s;
1819         unsigned max_root = ext4_ext_space_root(inode, 0);
1820         ext4_fsblk_t blk;
1821
1822         if ((path[0].p_depth != 1) ||
1823             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1824             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1825                 return;
1826
1827         /*
1828          * We need to modify the block allocation bitmap and the block
1829          * group descriptor to release the extent tree block.  If we
1830          * can't get the journal credits, give up.
1831          */
1832         if (ext4_journal_extend(handle, 2))
1833                 return;
1834
1835         /*
1836          * Copy the extent data up to the inode
1837          */
1838         blk = ext4_idx_pblock(path[0].p_idx);
1839         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1840                 sizeof(struct ext4_extent_idx);
1841         s += sizeof(struct ext4_extent_header);
1842
1843         path[1].p_maxdepth = path[0].p_maxdepth;
1844         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1845         path[0].p_depth = 0;
1846         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1847                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1848         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1849
1850         brelse(path[1].p_bh);
1851         ext4_free_blocks(handle, inode, NULL, blk, 1,
1852                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1853 }
1854
1855 /*
1856  * This function tries to merge the @ex extent to neighbours in the tree.
1857  * return 1 if merge left else 0.
1858  */
1859 static void ext4_ext_try_to_merge(handle_t *handle,
1860                                   struct inode *inode,
1861                                   struct ext4_ext_path *path,
1862                                   struct ext4_extent *ex) {
1863         struct ext4_extent_header *eh;
1864         unsigned int depth;
1865         int merge_done = 0;
1866
1867         depth = ext_depth(inode);
1868         BUG_ON(path[depth].p_hdr == NULL);
1869         eh = path[depth].p_hdr;
1870
1871         if (ex > EXT_FIRST_EXTENT(eh))
1872                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1873
1874         if (!merge_done)
1875                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1876
1877         ext4_ext_try_to_merge_up(handle, inode, path);
1878 }
1879
1880 /*
1881  * check if a portion of the "newext" extent overlaps with an
1882  * existing extent.
1883  *
1884  * If there is an overlap discovered, it updates the length of the newext
1885  * such that there will be no overlap, and then returns 1.
1886  * If there is no overlap found, it returns 0.
1887  */
1888 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1889                                            struct inode *inode,
1890                                            struct ext4_extent *newext,
1891                                            struct ext4_ext_path *path)
1892 {
1893         ext4_lblk_t b1, b2;
1894         unsigned int depth, len1;
1895         unsigned int ret = 0;
1896
1897         b1 = le32_to_cpu(newext->ee_block);
1898         len1 = ext4_ext_get_actual_len(newext);
1899         depth = ext_depth(inode);
1900         if (!path[depth].p_ext)
1901                 goto out;
1902         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1903
1904         /*
1905          * get the next allocated block if the extent in the path
1906          * is before the requested block(s)
1907          */
1908         if (b2 < b1) {
1909                 b2 = ext4_ext_next_allocated_block(path);
1910                 if (b2 == EXT_MAX_BLOCKS)
1911                         goto out;
1912                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1913         }
1914
1915         /* check for wrap through zero on extent logical start block*/
1916         if (b1 + len1 < b1) {
1917                 len1 = EXT_MAX_BLOCKS - b1;
1918                 newext->ee_len = cpu_to_le16(len1);
1919                 ret = 1;
1920         }
1921
1922         /* check for overlap */
1923         if (b1 + len1 > b2) {
1924                 newext->ee_len = cpu_to_le16(b2 - b1);
1925                 ret = 1;
1926         }
1927 out:
1928         return ret;
1929 }
1930
1931 /*
1932  * ext4_ext_insert_extent:
1933  * tries to merge requsted extent into the existing extent or
1934  * inserts requested extent as new one into the tree,
1935  * creating new leaf in the no-space case.
1936  */
1937 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1938                                 struct ext4_ext_path **ppath,
1939                                 struct ext4_extent *newext, int gb_flags)
1940 {
1941         struct ext4_ext_path *path = *ppath;
1942         struct ext4_extent_header *eh;
1943         struct ext4_extent *ex, *fex;
1944         struct ext4_extent *nearex; /* nearest extent */
1945         struct ext4_ext_path *npath = NULL;
1946         int depth, len, err;
1947         ext4_lblk_t next;
1948         int mb_flags = 0, unwritten;
1949
1950         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1951                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1952         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1953                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1954                 return -EFSCORRUPTED;
1955         }
1956         depth = ext_depth(inode);
1957         ex = path[depth].p_ext;
1958         eh = path[depth].p_hdr;
1959         if (unlikely(path[depth].p_hdr == NULL)) {
1960                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1961                 return -EFSCORRUPTED;
1962         }
1963
1964         /* try to insert block into found extent and return */
1965         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1966
1967                 /*
1968                  * Try to see whether we should rather test the extent on
1969                  * right from ex, or from the left of ex. This is because
1970                  * ext4_find_extent() can return either extent on the
1971                  * left, or on the right from the searched position. This
1972                  * will make merging more effective.
1973                  */
1974                 if (ex < EXT_LAST_EXTENT(eh) &&
1975                     (le32_to_cpu(ex->ee_block) +
1976                     ext4_ext_get_actual_len(ex) <
1977                     le32_to_cpu(newext->ee_block))) {
1978                         ex += 1;
1979                         goto prepend;
1980                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1981                            (le32_to_cpu(newext->ee_block) +
1982                            ext4_ext_get_actual_len(newext) <
1983                            le32_to_cpu(ex->ee_block)))
1984                         ex -= 1;
1985
1986                 /* Try to append newex to the ex */
1987                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1988                         ext_debug("append [%d]%d block to %u:[%d]%d"
1989                                   "(from %llu)\n",
1990                                   ext4_ext_is_unwritten(newext),
1991                                   ext4_ext_get_actual_len(newext),
1992                                   le32_to_cpu(ex->ee_block),
1993                                   ext4_ext_is_unwritten(ex),
1994                                   ext4_ext_get_actual_len(ex),
1995                                   ext4_ext_pblock(ex));
1996                         err = ext4_ext_get_access(handle, inode,
1997                                                   path + depth);
1998                         if (err)
1999                                 return err;
2000                         unwritten = ext4_ext_is_unwritten(ex);
2001                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2002                                         + ext4_ext_get_actual_len(newext));
2003                         if (unwritten)
2004                                 ext4_ext_mark_unwritten(ex);
2005                         eh = path[depth].p_hdr;
2006                         nearex = ex;
2007                         goto merge;
2008                 }
2009
2010 prepend:
2011                 /* Try to prepend newex to the ex */
2012                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2013                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2014                                   "(from %llu)\n",
2015                                   le32_to_cpu(newext->ee_block),
2016                                   ext4_ext_is_unwritten(newext),
2017                                   ext4_ext_get_actual_len(newext),
2018                                   le32_to_cpu(ex->ee_block),
2019                                   ext4_ext_is_unwritten(ex),
2020                                   ext4_ext_get_actual_len(ex),
2021                                   ext4_ext_pblock(ex));
2022                         err = ext4_ext_get_access(handle, inode,
2023                                                   path + depth);
2024                         if (err)
2025                                 return err;
2026
2027                         unwritten = ext4_ext_is_unwritten(ex);
2028                         ex->ee_block = newext->ee_block;
2029                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2030                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2031                                         + ext4_ext_get_actual_len(newext));
2032                         if (unwritten)
2033                                 ext4_ext_mark_unwritten(ex);
2034                         eh = path[depth].p_hdr;
2035                         nearex = ex;
2036                         goto merge;
2037                 }
2038         }
2039
2040         depth = ext_depth(inode);
2041         eh = path[depth].p_hdr;
2042         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2043                 goto has_space;
2044
2045         /* probably next leaf has space for us? */
2046         fex = EXT_LAST_EXTENT(eh);
2047         next = EXT_MAX_BLOCKS;
2048         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2049                 next = ext4_ext_next_leaf_block(path);
2050         if (next != EXT_MAX_BLOCKS) {
2051                 ext_debug("next leaf block - %u\n", next);
2052                 BUG_ON(npath != NULL);
2053                 npath = ext4_find_extent(inode, next, NULL, 0);
2054                 if (IS_ERR(npath))
2055                         return PTR_ERR(npath);
2056                 BUG_ON(npath->p_depth != path->p_depth);
2057                 eh = npath[depth].p_hdr;
2058                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2059                         ext_debug("next leaf isn't full(%d)\n",
2060                                   le16_to_cpu(eh->eh_entries));
2061                         path = npath;
2062                         goto has_space;
2063                 }
2064                 ext_debug("next leaf has no free space(%d,%d)\n",
2065                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2066         }
2067
2068         /*
2069          * There is no free space in the found leaf.
2070          * We're gonna add a new leaf in the tree.
2071          */
2072         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2073                 mb_flags |= EXT4_MB_USE_RESERVED;
2074         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2075                                        ppath, newext);
2076         if (err)
2077                 goto cleanup;
2078         depth = ext_depth(inode);
2079         eh = path[depth].p_hdr;
2080
2081 has_space:
2082         nearex = path[depth].p_ext;
2083
2084         err = ext4_ext_get_access(handle, inode, path + depth);
2085         if (err)
2086                 goto cleanup;
2087
2088         if (!nearex) {
2089                 /* there is no extent in this leaf, create first one */
2090                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2091                                 le32_to_cpu(newext->ee_block),
2092                                 ext4_ext_pblock(newext),
2093                                 ext4_ext_is_unwritten(newext),
2094                                 ext4_ext_get_actual_len(newext));
2095                 nearex = EXT_FIRST_EXTENT(eh);
2096         } else {
2097                 if (le32_to_cpu(newext->ee_block)
2098                            > le32_to_cpu(nearex->ee_block)) {
2099                         /* Insert after */
2100                         ext_debug("insert %u:%llu:[%d]%d before: "
2101                                         "nearest %p\n",
2102                                         le32_to_cpu(newext->ee_block),
2103                                         ext4_ext_pblock(newext),
2104                                         ext4_ext_is_unwritten(newext),
2105                                         ext4_ext_get_actual_len(newext),
2106                                         nearex);
2107                         nearex++;
2108                 } else {
2109                         /* Insert before */
2110                         BUG_ON(newext->ee_block == nearex->ee_block);
2111                         ext_debug("insert %u:%llu:[%d]%d after: "
2112                                         "nearest %p\n",
2113                                         le32_to_cpu(newext->ee_block),
2114                                         ext4_ext_pblock(newext),
2115                                         ext4_ext_is_unwritten(newext),
2116                                         ext4_ext_get_actual_len(newext),
2117                                         nearex);
2118                 }
2119                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2120                 if (len > 0) {
2121                         ext_debug("insert %u:%llu:[%d]%d: "
2122                                         "move %d extents from 0x%p to 0x%p\n",
2123                                         le32_to_cpu(newext->ee_block),
2124                                         ext4_ext_pblock(newext),
2125                                         ext4_ext_is_unwritten(newext),
2126                                         ext4_ext_get_actual_len(newext),
2127                                         len, nearex, nearex + 1);
2128                         memmove(nearex + 1, nearex,
2129                                 len * sizeof(struct ext4_extent));
2130                 }
2131         }
2132
2133         le16_add_cpu(&eh->eh_entries, 1);
2134         path[depth].p_ext = nearex;
2135         nearex->ee_block = newext->ee_block;
2136         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2137         nearex->ee_len = newext->ee_len;
2138
2139 merge:
2140         /* try to merge extents */
2141         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2142                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2143
2144
2145         /* time to correct all indexes above */
2146         err = ext4_ext_correct_indexes(handle, inode, path);
2147         if (err)
2148                 goto cleanup;
2149
2150         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2151
2152 cleanup:
2153         ext4_ext_drop_refs(npath);
2154         kfree(npath);
2155         return err;
2156 }
2157
2158 static int ext4_fill_fiemap_extents(struct inode *inode,
2159                                     ext4_lblk_t block, ext4_lblk_t num,
2160                                     struct fiemap_extent_info *fieinfo)
2161 {
2162         struct ext4_ext_path *path = NULL;
2163         struct ext4_extent *ex;
2164         struct extent_status es;
2165         ext4_lblk_t next, next_del, start = 0, end = 0;
2166         ext4_lblk_t last = block + num;
2167         int exists, depth = 0, err = 0;
2168         unsigned int flags = 0;
2169         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2170
2171         while (block < last && block != EXT_MAX_BLOCKS) {
2172                 num = last - block;
2173                 /* find extent for this block */
2174                 down_read(&EXT4_I(inode)->i_data_sem);
2175
2176                 path = ext4_find_extent(inode, block, &path, 0);
2177                 if (IS_ERR(path)) {
2178                         up_read(&EXT4_I(inode)->i_data_sem);
2179                         err = PTR_ERR(path);
2180                         path = NULL;
2181                         break;
2182                 }
2183
2184                 depth = ext_depth(inode);
2185                 if (unlikely(path[depth].p_hdr == NULL)) {
2186                         up_read(&EXT4_I(inode)->i_data_sem);
2187                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2188                         err = -EFSCORRUPTED;
2189                         break;
2190                 }
2191                 ex = path[depth].p_ext;
2192                 next = ext4_ext_next_allocated_block(path);
2193
2194                 flags = 0;
2195                 exists = 0;
2196                 if (!ex) {
2197                         /* there is no extent yet, so try to allocate
2198                          * all requested space */
2199                         start = block;
2200                         end = block + num;
2201                 } else if (le32_to_cpu(ex->ee_block) > block) {
2202                         /* need to allocate space before found extent */
2203                         start = block;
2204                         end = le32_to_cpu(ex->ee_block);
2205                         if (block + num < end)
2206                                 end = block + num;
2207                 } else if (block >= le32_to_cpu(ex->ee_block)
2208                                         + ext4_ext_get_actual_len(ex)) {
2209                         /* need to allocate space after found extent */
2210                         start = block;
2211                         end = block + num;
2212                         if (end >= next)
2213                                 end = next;
2214                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2215                         /*
2216                          * some part of requested space is covered
2217                          * by found extent
2218                          */
2219                         start = block;
2220                         end = le32_to_cpu(ex->ee_block)
2221                                 + ext4_ext_get_actual_len(ex);
2222                         if (block + num < end)
2223                                 end = block + num;
2224                         exists = 1;
2225                 } else {
2226                         BUG();
2227                 }
2228                 BUG_ON(end <= start);
2229
2230                 if (!exists) {
2231                         es.es_lblk = start;
2232                         es.es_len = end - start;
2233                         es.es_pblk = 0;
2234                 } else {
2235                         es.es_lblk = le32_to_cpu(ex->ee_block);
2236                         es.es_len = ext4_ext_get_actual_len(ex);
2237                         es.es_pblk = ext4_ext_pblock(ex);
2238                         if (ext4_ext_is_unwritten(ex))
2239                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2240                 }
2241
2242                 /*
2243                  * Find delayed extent and update es accordingly. We call
2244                  * it even in !exists case to find out whether es is the
2245                  * last existing extent or not.
2246                  */
2247                 next_del = ext4_find_delayed_extent(inode, &es);
2248                 if (!exists && next_del) {
2249                         exists = 1;
2250                         flags |= (FIEMAP_EXTENT_DELALLOC |
2251                                   FIEMAP_EXTENT_UNKNOWN);
2252                 }
2253                 up_read(&EXT4_I(inode)->i_data_sem);
2254
2255                 if (unlikely(es.es_len == 0)) {
2256                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2257                         err = -EFSCORRUPTED;
2258                         break;
2259                 }
2260
2261                 /*
2262                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2263                  * we need to check next == EXT_MAX_BLOCKS because it is
2264                  * possible that an extent is with unwritten and delayed
2265                  * status due to when an extent is delayed allocated and
2266                  * is allocated by fallocate status tree will track both of
2267                  * them in a extent.
2268                  *
2269                  * So we could return a unwritten and delayed extent, and
2270                  * its block is equal to 'next'.
2271                  */
2272                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2273                         flags |= FIEMAP_EXTENT_LAST;
2274                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2275                                      next != EXT_MAX_BLOCKS)) {
2276                                 EXT4_ERROR_INODE(inode,
2277                                                  "next extent == %u, next "
2278                                                  "delalloc extent = %u",
2279                                                  next, next_del);
2280                                 err = -EFSCORRUPTED;
2281                                 break;
2282                         }
2283                 }
2284
2285                 if (exists) {
2286                         err = fiemap_fill_next_extent(fieinfo,
2287                                 (__u64)es.es_lblk << blksize_bits,
2288                                 (__u64)es.es_pblk << blksize_bits,
2289                                 (__u64)es.es_len << blksize_bits,
2290                                 flags);
2291                         if (err < 0)
2292                                 break;
2293                         if (err == 1) {
2294                                 err = 0;
2295                                 break;
2296                         }
2297                 }
2298
2299                 block = es.es_lblk + es.es_len;
2300         }
2301
2302         ext4_ext_drop_refs(path);
2303         kfree(path);
2304         return err;
2305 }
2306
2307 /*
2308  * ext4_ext_determine_hole - determine hole around given block
2309  * @inode:      inode we lookup in
2310  * @path:       path in extent tree to @lblk
2311  * @lblk:       pointer to logical block around which we want to determine hole
2312  *
2313  * Determine hole length (and start if easily possible) around given logical
2314  * block. We don't try too hard to find the beginning of the hole but @path
2315  * actually points to extent before @lblk, we provide it.
2316  *
2317  * The function returns the length of a hole starting at @lblk. We update @lblk
2318  * to the beginning of the hole if we managed to find it.
2319  */
2320 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2321                                            struct ext4_ext_path *path,
2322                                            ext4_lblk_t *lblk)
2323 {
2324         int depth = ext_depth(inode);
2325         struct ext4_extent *ex;
2326         ext4_lblk_t len;
2327
2328         ex = path[depth].p_ext;
2329         if (ex == NULL) {
2330                 /* there is no extent yet, so gap is [0;-] */
2331                 *lblk = 0;
2332                 len = EXT_MAX_BLOCKS;
2333         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2334                 len = le32_to_cpu(ex->ee_block) - *lblk;
2335         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2336                         + ext4_ext_get_actual_len(ex)) {
2337                 ext4_lblk_t next;
2338
2339                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2340                 next = ext4_ext_next_allocated_block(path);
2341                 BUG_ON(next == *lblk);
2342                 len = next - *lblk;
2343         } else {
2344                 BUG();
2345         }
2346         return len;
2347 }
2348
2349 /*
2350  * ext4_ext_put_gap_in_cache:
2351  * calculate boundaries of the gap that the requested block fits into
2352  * and cache this gap
2353  */
2354 static void
2355 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2356                           ext4_lblk_t hole_len)
2357 {
2358         struct extent_status es;
2359
2360         ext4_es_find_delayed_extent_range(inode, hole_start,
2361                                           hole_start + hole_len - 1, &es);
2362         if (es.es_len) {
2363                 /* There's delayed extent containing lblock? */
2364                 if (es.es_lblk <= hole_start)
2365                         return;
2366                 hole_len = min(es.es_lblk - hole_start, hole_len);
2367         }
2368         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2369         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2370                               EXTENT_STATUS_HOLE);
2371 }
2372
2373 /*
2374  * ext4_ext_rm_idx:
2375  * removes index from the index block.
2376  */
2377 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2378                         struct ext4_ext_path *path, int depth)
2379 {
2380         int err;
2381         ext4_fsblk_t leaf;
2382
2383         /* free index block */
2384         depth--;
2385         path = path + depth;
2386         leaf = ext4_idx_pblock(path->p_idx);
2387         if (unlikely(path->p_hdr->eh_entries == 0)) {
2388                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2389                 return -EFSCORRUPTED;
2390         }
2391         err = ext4_ext_get_access(handle, inode, path);
2392         if (err)
2393                 return err;
2394
2395         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2396                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2397                 len *= sizeof(struct ext4_extent_idx);
2398                 memmove(path->p_idx, path->p_idx + 1, len);
2399         }
2400
2401         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2402         err = ext4_ext_dirty(handle, inode, path);
2403         if (err)
2404                 return err;
2405         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2406         trace_ext4_ext_rm_idx(inode, leaf);
2407
2408         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2409                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2410
2411         while (--depth >= 0) {
2412                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2413                         break;
2414                 path--;
2415                 err = ext4_ext_get_access(handle, inode, path);
2416                 if (err)
2417                         break;
2418                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2419                 err = ext4_ext_dirty(handle, inode, path);
2420                 if (err)
2421                         break;
2422         }
2423         return err;
2424 }
2425
2426 /*
2427  * ext4_ext_calc_credits_for_single_extent:
2428  * This routine returns max. credits that needed to insert an extent
2429  * to the extent tree.
2430  * When pass the actual path, the caller should calculate credits
2431  * under i_data_sem.
2432  */
2433 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2434                                                 struct ext4_ext_path *path)
2435 {
2436         if (path) {
2437                 int depth = ext_depth(inode);
2438                 int ret = 0;
2439
2440                 /* probably there is space in leaf? */
2441                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2442                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2443
2444                         /*
2445                          *  There are some space in the leaf tree, no
2446                          *  need to account for leaf block credit
2447                          *
2448                          *  bitmaps and block group descriptor blocks
2449                          *  and other metadata blocks still need to be
2450                          *  accounted.
2451                          */
2452                         /* 1 bitmap, 1 block group descriptor */
2453                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2454                         return ret;
2455                 }
2456         }
2457
2458         return ext4_chunk_trans_blocks(inode, nrblocks);
2459 }
2460
2461 /*
2462  * How many index/leaf blocks need to change/allocate to add @extents extents?
2463  *
2464  * If we add a single extent, then in the worse case, each tree level
2465  * index/leaf need to be changed in case of the tree split.
2466  *
2467  * If more extents are inserted, they could cause the whole tree split more
2468  * than once, but this is really rare.
2469  */
2470 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2471 {
2472         int index;
2473         int depth;
2474
2475         /* If we are converting the inline data, only one is needed here. */
2476         if (ext4_has_inline_data(inode))
2477                 return 1;
2478
2479         depth = ext_depth(inode);
2480
2481         if (extents <= 1)
2482                 index = depth * 2;
2483         else
2484                 index = depth * 3;
2485
2486         return index;
2487 }
2488
2489 static inline int get_default_free_blocks_flags(struct inode *inode)
2490 {
2491         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2492             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2493                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2494         else if (ext4_should_journal_data(inode))
2495                 return EXT4_FREE_BLOCKS_FORGET;
2496         return 0;
2497 }
2498
2499 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2500                               struct ext4_extent *ex,
2501                               long long *partial_cluster,
2502                               ext4_lblk_t from, ext4_lblk_t to)
2503 {
2504         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2505         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2506         ext4_fsblk_t pblk;
2507         int flags = get_default_free_blocks_flags(inode);
2508
2509         /*
2510          * For bigalloc file systems, we never free a partial cluster
2511          * at the beginning of the extent.  Instead, we make a note
2512          * that we tried freeing the cluster, and check to see if we
2513          * need to free it on a subsequent call to ext4_remove_blocks,
2514          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2515          */
2516         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2517
2518         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2519         /*
2520          * If we have a partial cluster, and it's different from the
2521          * cluster of the last block, we need to explicitly free the
2522          * partial cluster here.
2523          */
2524         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2525         if (*partial_cluster > 0 &&
2526             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2527                 ext4_free_blocks(handle, inode, NULL,
2528                                  EXT4_C2B(sbi, *partial_cluster),
2529                                  sbi->s_cluster_ratio, flags);
2530                 *partial_cluster = 0;
2531         }
2532
2533 #ifdef EXTENTS_STATS
2534         {
2535                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2536                 spin_lock(&sbi->s_ext_stats_lock);
2537                 sbi->s_ext_blocks += ee_len;
2538                 sbi->s_ext_extents++;
2539                 if (ee_len < sbi->s_ext_min)
2540                         sbi->s_ext_min = ee_len;
2541                 if (ee_len > sbi->s_ext_max)
2542                         sbi->s_ext_max = ee_len;
2543                 if (ext_depth(inode) > sbi->s_depth_max)
2544                         sbi->s_depth_max = ext_depth(inode);
2545                 spin_unlock(&sbi->s_ext_stats_lock);
2546         }
2547 #endif
2548         if (from >= le32_to_cpu(ex->ee_block)
2549             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2550                 /* tail removal */
2551                 ext4_lblk_t num;
2552                 long long first_cluster;
2553
2554                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2555                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2556                 /*
2557                  * Usually we want to free partial cluster at the end of the
2558                  * extent, except for the situation when the cluster is still
2559                  * used by any other extent (partial_cluster is negative).
2560                  */
2561                 if (*partial_cluster < 0 &&
2562                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2563                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2564
2565                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2566                           num, pblk, *partial_cluster);
2567                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2568                 /*
2569                  * If the block range to be freed didn't start at the
2570                  * beginning of a cluster, and we removed the entire
2571                  * extent and the cluster is not used by any other extent,
2572                  * save the partial cluster here, since we might need to
2573                  * delete if we determine that the truncate or punch hole
2574                  * operation has removed all of the blocks in the cluster.
2575                  * If that cluster is used by another extent, preserve its
2576                  * negative value so it isn't freed later on.
2577                  *
2578                  * If the whole extent wasn't freed, we've reached the
2579                  * start of the truncated/punched region and have finished
2580                  * removing blocks.  If there's a partial cluster here it's
2581                  * shared with the remainder of the extent and is no longer
2582                  * a candidate for removal.
2583                  */
2584                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2585                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2586                         if (first_cluster != -*partial_cluster)
2587                                 *partial_cluster = first_cluster;
2588                 } else {
2589                         *partial_cluster = 0;
2590                 }
2591         } else
2592                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2593                            "%u-%u from %u:%u",
2594                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2595         return 0;
2596 }
2597
2598
2599 /*
2600  * ext4_ext_rm_leaf() Removes the extents associated with the
2601  * blocks appearing between "start" and "end".  Both "start"
2602  * and "end" must appear in the same extent or EIO is returned.
2603  *
2604  * @handle: The journal handle
2605  * @inode:  The files inode
2606  * @path:   The path to the leaf
2607  * @partial_cluster: The cluster which we'll have to free if all extents
2608  *                   has been released from it.  However, if this value is
2609  *                   negative, it's a cluster just to the right of the
2610  *                   punched region and it must not be freed.
2611  * @start:  The first block to remove
2612  * @end:   The last block to remove
2613  */
2614 static int
2615 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2616                  struct ext4_ext_path *path,
2617                  long long *partial_cluster,
2618                  ext4_lblk_t start, ext4_lblk_t end)
2619 {
2620         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2621         int err = 0, correct_index = 0;
2622         int depth = ext_depth(inode), credits;
2623         struct ext4_extent_header *eh;
2624         ext4_lblk_t a, b;
2625         unsigned num;
2626         ext4_lblk_t ex_ee_block;
2627         unsigned short ex_ee_len;
2628         unsigned unwritten = 0;
2629         struct ext4_extent *ex;
2630         ext4_fsblk_t pblk;
2631
2632         /* the header must be checked already in ext4_ext_remove_space() */
2633         ext_debug("truncate since %u in leaf to %u\n", start, end);
2634         if (!path[depth].p_hdr)
2635                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2636         eh = path[depth].p_hdr;
2637         if (unlikely(path[depth].p_hdr == NULL)) {
2638                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2639                 return -EFSCORRUPTED;
2640         }
2641         /* find where to start removing */
2642         ex = path[depth].p_ext;
2643         if (!ex)
2644                 ex = EXT_LAST_EXTENT(eh);
2645
2646         ex_ee_block = le32_to_cpu(ex->ee_block);
2647         ex_ee_len = ext4_ext_get_actual_len(ex);
2648
2649         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2650
2651         while (ex >= EXT_FIRST_EXTENT(eh) &&
2652                         ex_ee_block + ex_ee_len > start) {
2653
2654                 if (ext4_ext_is_unwritten(ex))
2655                         unwritten = 1;
2656                 else
2657                         unwritten = 0;
2658
2659                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2660                           unwritten, ex_ee_len);
2661                 path[depth].p_ext = ex;
2662
2663                 a = ex_ee_block > start ? ex_ee_block : start;
2664                 b = ex_ee_block+ex_ee_len - 1 < end ?
2665                         ex_ee_block+ex_ee_len - 1 : end;
2666
2667                 ext_debug("  border %u:%u\n", a, b);
2668
2669                 /* If this extent is beyond the end of the hole, skip it */
2670                 if (end < ex_ee_block) {
2671                         /*
2672                          * We're going to skip this extent and move to another,
2673                          * so note that its first cluster is in use to avoid
2674                          * freeing it when removing blocks.  Eventually, the
2675                          * right edge of the truncated/punched region will
2676                          * be just to the left.
2677                          */
2678                         if (sbi->s_cluster_ratio > 1) {
2679                                 pblk = ext4_ext_pblock(ex);
2680                                 *partial_cluster =
2681                                         -(long long) EXT4_B2C(sbi, pblk);
2682                         }
2683                         ex--;
2684                         ex_ee_block = le32_to_cpu(ex->ee_block);
2685                         ex_ee_len = ext4_ext_get_actual_len(ex);
2686                         continue;
2687                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2688                         EXT4_ERROR_INODE(inode,
2689                                          "can not handle truncate %u:%u "
2690                                          "on extent %u:%u",
2691                                          start, end, ex_ee_block,
2692                                          ex_ee_block + ex_ee_len - 1);
2693                         err = -EFSCORRUPTED;
2694                         goto out;
2695                 } else if (a != ex_ee_block) {
2696                         /* remove tail of the extent */
2697                         num = a - ex_ee_block;
2698                 } else {
2699                         /* remove whole extent: excellent! */
2700                         num = 0;
2701                 }
2702                 /*
2703                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2704                  * descriptor) for each block group; assume two block
2705                  * groups plus ex_ee_len/blocks_per_block_group for
2706                  * the worst case
2707                  */
2708                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2709                 if (ex == EXT_FIRST_EXTENT(eh)) {
2710                         correct_index = 1;
2711                         credits += (ext_depth(inode)) + 1;
2712                 }
2713                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2714
2715                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2716                 if (err)
2717                         goto out;
2718
2719                 err = ext4_ext_get_access(handle, inode, path + depth);
2720                 if (err)
2721                         goto out;
2722
2723                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2724                                          a, b);
2725                 if (err)
2726                         goto out;
2727
2728                 if (num == 0)
2729                         /* this extent is removed; mark slot entirely unused */
2730                         ext4_ext_store_pblock(ex, 0);
2731
2732                 ex->ee_len = cpu_to_le16(num);
2733                 /*
2734                  * Do not mark unwritten if all the blocks in the
2735                  * extent have been removed.
2736                  */
2737                 if (unwritten && num)
2738                         ext4_ext_mark_unwritten(ex);
2739                 /*
2740                  * If the extent was completely released,
2741                  * we need to remove it from the leaf
2742                  */
2743                 if (num == 0) {
2744                         if (end != EXT_MAX_BLOCKS - 1) {
2745                                 /*
2746                                  * For hole punching, we need to scoot all the
2747                                  * extents up when an extent is removed so that
2748                                  * we dont have blank extents in the middle
2749                                  */
2750                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2751                                         sizeof(struct ext4_extent));
2752
2753                                 /* Now get rid of the one at the end */
2754                                 memset(EXT_LAST_EXTENT(eh), 0,
2755                                         sizeof(struct ext4_extent));
2756                         }
2757                         le16_add_cpu(&eh->eh_entries, -1);
2758                 }
2759
2760                 err = ext4_ext_dirty(handle, inode, path + depth);
2761                 if (err)
2762                         goto out;
2763
2764                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2765                                 ext4_ext_pblock(ex));
2766                 ex--;
2767                 ex_ee_block = le32_to_cpu(ex->ee_block);
2768                 ex_ee_len = ext4_ext_get_actual_len(ex);
2769         }
2770
2771         if (correct_index && eh->eh_entries)
2772                 err = ext4_ext_correct_indexes(handle, inode, path);
2773
2774         /*
2775          * If there's a partial cluster and at least one extent remains in
2776          * the leaf, free the partial cluster if it isn't shared with the
2777          * current extent.  If it is shared with the current extent
2778          * we zero partial_cluster because we've reached the start of the
2779          * truncated/punched region and we're done removing blocks.
2780          */
2781         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2782                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2783                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2784                         ext4_free_blocks(handle, inode, NULL,
2785                                          EXT4_C2B(sbi, *partial_cluster),
2786                                          sbi->s_cluster_ratio,
2787                                          get_default_free_blocks_flags(inode));
2788                 }
2789                 *partial_cluster = 0;
2790         }
2791
2792         /* if this leaf is free, then we should
2793          * remove it from index block above */
2794         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2795                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2796
2797 out:
2798         return err;
2799 }
2800
2801 /*
2802  * ext4_ext_more_to_rm:
2803  * returns 1 if current index has to be freed (even partial)
2804  */
2805 static int
2806 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2807 {
2808         BUG_ON(path->p_idx == NULL);
2809
2810         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2811                 return 0;
2812
2813         /*
2814          * if truncate on deeper level happened, it wasn't partial,
2815          * so we have to consider current index for truncation
2816          */
2817         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2818                 return 0;
2819         return 1;
2820 }
2821
2822 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2823                           ext4_lblk_t end)
2824 {
2825         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2826         int depth = ext_depth(inode);
2827         struct ext4_ext_path *path = NULL;
2828         long long partial_cluster = 0;
2829         handle_t *handle;
2830         int i = 0, err = 0;
2831
2832         ext_debug("truncate since %u to %u\n", start, end);
2833
2834         /* probably first extent we're gonna free will be last in block */
2835         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2836         if (IS_ERR(handle))
2837                 return PTR_ERR(handle);
2838
2839 again:
2840         trace_ext4_ext_remove_space(inode, start, end, depth);
2841
2842         /*
2843          * Check if we are removing extents inside the extent tree. If that
2844          * is the case, we are going to punch a hole inside the extent tree
2845          * so we have to check whether we need to split the extent covering
2846          * the last block to remove so we can easily remove the part of it
2847          * in ext4_ext_rm_leaf().
2848          */
2849         if (end < EXT_MAX_BLOCKS - 1) {
2850                 struct ext4_extent *ex;
2851                 ext4_lblk_t ee_block, ex_end, lblk;
2852                 ext4_fsblk_t pblk;
2853
2854                 /* find extent for or closest extent to this block */
2855                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2856                 if (IS_ERR(path)) {
2857                         ext4_journal_stop(handle);
2858                         return PTR_ERR(path);
2859                 }
2860                 depth = ext_depth(inode);
2861                 /* Leaf not may not exist only if inode has no blocks at all */
2862                 ex = path[depth].p_ext;
2863                 if (!ex) {
2864                         if (depth) {
2865                                 EXT4_ERROR_INODE(inode,
2866                                                  "path[%d].p_hdr == NULL",
2867                                                  depth);
2868                                 err = -EFSCORRUPTED;
2869                         }
2870                         goto out;
2871                 }
2872
2873                 ee_block = le32_to_cpu(ex->ee_block);
2874                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2875
2876                 /*
2877                  * See if the last block is inside the extent, if so split
2878                  * the extent at 'end' block so we can easily remove the
2879                  * tail of the first part of the split extent in
2880                  * ext4_ext_rm_leaf().
2881                  */
2882                 if (end >= ee_block && end < ex_end) {
2883
2884                         /*
2885                          * If we're going to split the extent, note that
2886                          * the cluster containing the block after 'end' is
2887                          * in use to avoid freeing it when removing blocks.
2888                          */
2889                         if (sbi->s_cluster_ratio > 1) {
2890                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2891                                 partial_cluster =
2892                                         -(long long) EXT4_B2C(sbi, pblk);
2893                         }
2894
2895                         /*
2896                          * Split the extent in two so that 'end' is the last
2897                          * block in the first new extent. Also we should not
2898                          * fail removing space due to ENOSPC so try to use
2899                          * reserved block if that happens.
2900                          */
2901                         err = ext4_force_split_extent_at(handle, inode, &path,
2902                                                          end + 1, 1);
2903                         if (err < 0)
2904                                 goto out;
2905
2906                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2907                         /*
2908                          * If there's an extent to the right its first cluster
2909                          * contains the immediate right boundary of the
2910                          * truncated/punched region.  Set partial_cluster to
2911                          * its negative value so it won't be freed if shared
2912                          * with the current extent.  The end < ee_block case
2913                          * is handled in ext4_ext_rm_leaf().
2914                          */
2915                         lblk = ex_end + 1;
2916                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2917                                                     &ex);
2918                         if (err)
2919                                 goto out;
2920                         if (pblk)
2921                                 partial_cluster =
2922                                         -(long long) EXT4_B2C(sbi, pblk);
2923                 }
2924         }
2925         /*
2926          * We start scanning from right side, freeing all the blocks
2927          * after i_size and walking into the tree depth-wise.
2928          */
2929         depth = ext_depth(inode);
2930         if (path) {
2931                 int k = i = depth;
2932                 while (--k > 0)
2933                         path[k].p_block =
2934                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2935         } else {
2936                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2937                                GFP_NOFS);
2938                 if (path == NULL) {
2939                         ext4_journal_stop(handle);
2940                         return -ENOMEM;
2941                 }
2942                 path[0].p_maxdepth = path[0].p_depth = depth;
2943                 path[0].p_hdr = ext_inode_hdr(inode);
2944                 i = 0;
2945
2946                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2947                         err = -EFSCORRUPTED;
2948                         goto out;
2949                 }
2950         }
2951         err = 0;
2952
2953         while (i >= 0 && err == 0) {
2954                 if (i == depth) {
2955                         /* this is leaf block */
2956                         err = ext4_ext_rm_leaf(handle, inode, path,
2957                                                &partial_cluster, start,
2958                                                end);
2959                         /* root level has p_bh == NULL, brelse() eats this */
2960                         brelse(path[i].p_bh);
2961                         path[i].p_bh = NULL;
2962                         i--;
2963                         continue;
2964                 }
2965
2966                 /* this is index block */
2967                 if (!path[i].p_hdr) {
2968                         ext_debug("initialize header\n");
2969                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2970                 }
2971
2972                 if (!path[i].p_idx) {
2973                         /* this level hasn't been touched yet */
2974                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2975                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2976                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2977                                   path[i].p_hdr,
2978                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2979                 } else {
2980                         /* we were already here, see at next index */
2981                         path[i].p_idx--;
2982                 }
2983
2984                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2985                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2986                                 path[i].p_idx);
2987                 if (ext4_ext_more_to_rm(path + i)) {
2988                         struct buffer_head *bh;
2989                         /* go to the next level */
2990                         ext_debug("move to level %d (block %llu)\n",
2991                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2992                         memset(path + i + 1, 0, sizeof(*path));
2993                         bh = read_extent_tree_block(inode,
2994                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2995                                 EXT4_EX_NOCACHE);
2996                         if (IS_ERR(bh)) {
2997                                 /* should we reset i_size? */
2998                                 err = PTR_ERR(bh);
2999                                 break;
3000                         }
3001                         /* Yield here to deal with large extent trees.
3002                          * Should be a no-op if we did IO above. */
3003                         cond_resched();
3004                         if (WARN_ON(i + 1 > depth)) {
3005                                 err = -EFSCORRUPTED;
3006                                 break;
3007                         }
3008                         path[i + 1].p_bh = bh;
3009
3010                         /* save actual number of indexes since this
3011                          * number is changed at the next iteration */
3012                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3013                         i++;
3014                 } else {
3015                         /* we finished processing this index, go up */
3016                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3017                                 /* index is empty, remove it;
3018                                  * handle must be already prepared by the
3019                                  * truncatei_leaf() */
3020                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3021                         }
3022                         /* root level has p_bh == NULL, brelse() eats this */
3023                         brelse(path[i].p_bh);
3024                         path[i].p_bh = NULL;
3025                         i--;
3026                         ext_debug("return to level %d\n", i);
3027                 }
3028         }
3029
3030         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3031                         partial_cluster, path->p_hdr->eh_entries);
3032
3033         /*
3034          * If we still have something in the partial cluster and we have removed
3035          * even the first extent, then we should free the blocks in the partial
3036          * cluster as well.  (This code will only run when there are no leaves
3037          * to the immediate left of the truncated/punched region.)
3038          */
3039         if (partial_cluster > 0 && err == 0) {
3040                 /* don't zero partial_cluster since it's not used afterwards */
3041                 ext4_free_blocks(handle, inode, NULL,
3042                                  EXT4_C2B(sbi, partial_cluster),
3043                                  sbi->s_cluster_ratio,
3044                                  get_default_free_blocks_flags(inode));
3045         }
3046
3047         /* TODO: flexible tree reduction should be here */
3048         if (path->p_hdr->eh_entries == 0) {
3049                 /*
3050                  * truncate to zero freed all the tree,
3051                  * so we need to correct eh_depth
3052                  */
3053                 err = ext4_ext_get_access(handle, inode, path);
3054                 if (err == 0) {
3055                         ext_inode_hdr(inode)->eh_depth = 0;
3056                         ext_inode_hdr(inode)->eh_max =
3057                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3058                         err = ext4_ext_dirty(handle, inode, path);
3059                 }
3060         }
3061 out:
3062         ext4_ext_drop_refs(path);
3063         kfree(path);
3064         path = NULL;
3065         if (err == -EAGAIN)
3066                 goto again;
3067         ext4_journal_stop(handle);
3068
3069         return err;
3070 }
3071
3072 /*
3073  * called at mount time
3074  */
3075 void ext4_ext_init(struct super_block *sb)
3076 {
3077         /*
3078          * possible initialization would be here
3079          */
3080
3081         if (ext4_has_feature_extents(sb)) {
3082 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3083                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3084 #ifdef AGGRESSIVE_TEST
3085                        ", aggressive tests"
3086 #endif
3087 #ifdef CHECK_BINSEARCH
3088                        ", check binsearch"
3089 #endif
3090 #ifdef EXTENTS_STATS
3091                        ", stats"
3092 #endif
3093                        "\n");
3094 #endif
3095 #ifdef EXTENTS_STATS
3096                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3097                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3098                 EXT4_SB(sb)->s_ext_max = 0;
3099 #endif
3100         }
3101 }
3102
3103 /*
3104  * called at umount time
3105  */
3106 void ext4_ext_release(struct super_block *sb)
3107 {
3108         if (!ext4_has_feature_extents(sb))
3109                 return;
3110
3111 #ifdef EXTENTS_STATS
3112         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3113                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3114                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3115                         sbi->s_ext_blocks, sbi->s_ext_extents,
3116                         sbi->s_ext_blocks / sbi->s_ext_extents);
3117                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3118                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3119         }
3120 #endif
3121 }
3122
3123 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3124 {
3125         ext4_lblk_t  ee_block;
3126         ext4_fsblk_t ee_pblock;
3127         unsigned int ee_len;
3128
3129         ee_block  = le32_to_cpu(ex->ee_block);
3130         ee_len    = ext4_ext_get_actual_len(ex);
3131         ee_pblock = ext4_ext_pblock(ex);
3132
3133         if (ee_len == 0)
3134                 return 0;
3135
3136         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3137                                      EXTENT_STATUS_WRITTEN);
3138 }
3139
3140 /* FIXME!! we need to try to merge to left or right after zero-out  */
3141 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3142 {
3143         ext4_fsblk_t ee_pblock;
3144         unsigned int ee_len;
3145
3146         ee_len    = ext4_ext_get_actual_len(ex);
3147         ee_pblock = ext4_ext_pblock(ex);
3148         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3149                                   ee_len);
3150 }
3151
3152 /*
3153  * ext4_split_extent_at() splits an extent at given block.
3154  *
3155  * @handle: the journal handle
3156  * @inode: the file inode
3157  * @path: the path to the extent
3158  * @split: the logical block where the extent is splitted.
3159  * @split_flags: indicates if the extent could be zeroout if split fails, and
3160  *               the states(init or unwritten) of new extents.
3161  * @flags: flags used to insert new extent to extent tree.
3162  *
3163  *
3164  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3165  * of which are deterimined by split_flag.
3166  *
3167  * There are two cases:
3168  *  a> the extent are splitted into two extent.
3169  *  b> split is not needed, and just mark the extent.
3170  *
3171  * return 0 on success.
3172  */
3173 static int ext4_split_extent_at(handle_t *handle,
3174                              struct inode *inode,
3175                              struct ext4_ext_path **ppath,
3176                              ext4_lblk_t split,
3177                              int split_flag,
3178                              int flags)
3179 {
3180         struct ext4_ext_path *path = *ppath;
3181         ext4_fsblk_t newblock;
3182         ext4_lblk_t ee_block;
3183         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3184         struct ext4_extent *ex2 = NULL;
3185         unsigned int ee_len, depth;
3186         int err = 0;
3187
3188         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3189                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3190
3191         ext_debug("ext4_split_extents_at: inode %lu, logical"
3192                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3193
3194         ext4_ext_show_leaf(inode, path);
3195
3196         depth = ext_depth(inode);
3197         ex = path[depth].p_ext;
3198         ee_block = le32_to_cpu(ex->ee_block);
3199         ee_len = ext4_ext_get_actual_len(ex);
3200         newblock = split - ee_block + ext4_ext_pblock(ex);
3201
3202         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3203         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3204                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3205                              EXT4_EXT_MARK_UNWRIT1 |
3206                              EXT4_EXT_MARK_UNWRIT2));
3207
3208         err = ext4_ext_get_access(handle, inode, path + depth);
3209         if (err)
3210                 goto out;
3211
3212         if (split == ee_block) {
3213                 /*
3214                  * case b: block @split is the block that the extent begins with
3215                  * then we just change the state of the extent, and splitting
3216                  * is not needed.
3217                  */
3218                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3219                         ext4_ext_mark_unwritten(ex);
3220                 else
3221                         ext4_ext_mark_initialized(ex);
3222
3223                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3224                         ext4_ext_try_to_merge(handle, inode, path, ex);
3225
3226                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3227                 goto out;
3228         }
3229
3230         /* case a */
3231         memcpy(&orig_ex, ex, sizeof(orig_ex));
3232         ex->ee_len = cpu_to_le16(split - ee_block);
3233         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3234                 ext4_ext_mark_unwritten(ex);
3235
3236         /*
3237          * path may lead to new leaf, not to original leaf any more
3238          * after ext4_ext_insert_extent() returns,
3239          */
3240         err = ext4_ext_dirty(handle, inode, path + depth);
3241         if (err)
3242                 goto fix_extent_len;
3243
3244         ex2 = &newex;
3245         ex2->ee_block = cpu_to_le32(split);
3246         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3247         ext4_ext_store_pblock(ex2, newblock);
3248         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3249                 ext4_ext_mark_unwritten(ex2);
3250
3251         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3252         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3253                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3254                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3255                                 err = ext4_ext_zeroout(inode, ex2);
3256                                 zero_ex.ee_block = ex2->ee_block;
3257                                 zero_ex.ee_len = cpu_to_le16(
3258                                                 ext4_ext_get_actual_len(ex2));
3259                                 ext4_ext_store_pblock(&zero_ex,
3260                                                       ext4_ext_pblock(ex2));
3261                         } else {
3262                                 err = ext4_ext_zeroout(inode, ex);
3263                                 zero_ex.ee_block = ex->ee_block;
3264                                 zero_ex.ee_len = cpu_to_le16(
3265                                                 ext4_ext_get_actual_len(ex));
3266                                 ext4_ext_store_pblock(&zero_ex,
3267                                                       ext4_ext_pblock(ex));
3268                         }
3269                 } else {
3270                         err = ext4_ext_zeroout(inode, &orig_ex);
3271                         zero_ex.ee_block = orig_ex.ee_block;
3272                         zero_ex.ee_len = cpu_to_le16(
3273                                                 ext4_ext_get_actual_len(&orig_ex));
3274                         ext4_ext_store_pblock(&zero_ex,
3275                                               ext4_ext_pblock(&orig_ex));
3276                 }
3277
3278                 if (err)
3279                         goto fix_extent_len;
3280                 /* update the extent length and mark as initialized */
3281                 ex->ee_len = cpu_to_le16(ee_len);
3282                 ext4_ext_try_to_merge(handle, inode, path, ex);
3283                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3284                 if (err)
3285                         goto fix_extent_len;
3286
3287                 /* update extent status tree */
3288                 err = ext4_zeroout_es(inode, &zero_ex);
3289
3290                 goto out;
3291         } else if (err)
3292                 goto fix_extent_len;
3293
3294 out:
3295         ext4_ext_show_leaf(inode, path);
3296         return err;
3297
3298 fix_extent_len:
3299         ex->ee_len = orig_ex.ee_len;
3300         ext4_ext_dirty(handle, inode, path + path->p_depth);
3301         return err;
3302 }
3303
3304 /*
3305  * ext4_split_extents() splits an extent and mark extent which is covered
3306  * by @map as split_flags indicates
3307  *
3308  * It may result in splitting the extent into multiple extents (up to three)
3309  * There are three possibilities:
3310  *   a> There is no split required
3311  *   b> Splits in two extents: Split is happening at either end of the extent
3312  *   c> Splits in three extents: Somone is splitting in middle of the extent
3313  *
3314  */
3315 static int ext4_split_extent(handle_t *handle,
3316                               struct inode *inode,
3317                               struct ext4_ext_path **ppath,
3318                               struct ext4_map_blocks *map,
3319                               int split_flag,
3320                               int flags)
3321 {
3322         struct ext4_ext_path *path = *ppath;
3323         ext4_lblk_t ee_block;
3324         struct ext4_extent *ex;
3325         unsigned int ee_len, depth;
3326         int err = 0;
3327         int unwritten;
3328         int split_flag1, flags1;
3329         int allocated = map->m_len;
3330
3331         depth = ext_depth(inode);
3332         ex = path[depth].p_ext;
3333         ee_block = le32_to_cpu(ex->ee_block);
3334         ee_len = ext4_ext_get_actual_len(ex);
3335         unwritten = ext4_ext_is_unwritten(ex);
3336
3337         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3338                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3339                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3340                 if (unwritten)
3341                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3342                                        EXT4_EXT_MARK_UNWRIT2;
3343                 if (split_flag & EXT4_EXT_DATA_VALID2)
3344                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3345                 err = ext4_split_extent_at(handle, inode, ppath,
3346                                 map->m_lblk + map->m_len, split_flag1, flags1);
3347                 if (err)
3348                         goto out;
3349         } else {
3350                 allocated = ee_len - (map->m_lblk - ee_block);
3351         }
3352         /*
3353          * Update path is required because previous ext4_split_extent_at() may
3354          * result in split of original leaf or extent zeroout.
3355          */
3356         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3357         if (IS_ERR(path))
3358                 return PTR_ERR(path);
3359         depth = ext_depth(inode);
3360         ex = path[depth].p_ext;
3361         if (!ex) {
3362                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3363                                  (unsigned long) map->m_lblk);
3364                 return -EFSCORRUPTED;
3365         }
3366         unwritten = ext4_ext_is_unwritten(ex);
3367         split_flag1 = 0;
3368
3369         if (map->m_lblk >= ee_block) {
3370                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3371                 if (unwritten) {
3372                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3373                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3374                                                      EXT4_EXT_MARK_UNWRIT2);
3375                 }
3376                 err = ext4_split_extent_at(handle, inode, ppath,
3377                                 map->m_lblk, split_flag1, flags);
3378                 if (err)
3379                         goto out;
3380         }
3381
3382         ext4_ext_show_leaf(inode, path);
3383 out:
3384         return err ? err : allocated;
3385 }
3386
3387 /*
3388  * This function is called by ext4_ext_map_blocks() if someone tries to write
3389  * to an unwritten extent. It may result in splitting the unwritten
3390  * extent into multiple extents (up to three - one initialized and two
3391  * unwritten).
3392  * There are three possibilities:
3393  *   a> There is no split required: Entire extent should be initialized
3394  *   b> Splits in two extents: Write is happening at either end of the extent
3395  *   c> Splits in three extents: Somone is writing in middle of the extent
3396  *
3397  * Pre-conditions:
3398  *  - The extent pointed to by 'path' is unwritten.
3399  *  - The extent pointed to by 'path' contains a superset
3400  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3401  *
3402  * Post-conditions on success:
3403  *  - the returned value is the number of blocks beyond map->l_lblk
3404  *    that are allocated and initialized.
3405  *    It is guaranteed to be >= map->m_len.
3406  */
3407 static int ext4_ext_convert_to_initialized(handle_t *handle,
3408                                            struct inode *inode,
3409                                            struct ext4_map_blocks *map,
3410                                            struct ext4_ext_path **ppath,
3411                                            int flags)
3412 {
3413         struct ext4_ext_path *path = *ppath;
3414         struct ext4_sb_info *sbi;
3415         struct ext4_extent_header *eh;
3416         struct ext4_map_blocks split_map;
3417         struct ext4_extent zero_ex1, zero_ex2;
3418         struct ext4_extent *ex, *abut_ex;
3419         ext4_lblk_t ee_block, eof_block;
3420         unsigned int ee_len, depth, map_len = map->m_len;
3421         int allocated = 0, max_zeroout = 0;
3422         int err = 0;
3423         int split_flag = EXT4_EXT_DATA_VALID2;
3424
3425         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3426                 "block %llu, max_blocks %u\n", inode->i_ino,
3427                 (unsigned long long)map->m_lblk, map_len);
3428
3429         sbi = EXT4_SB(inode->i_sb);
3430         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3431                 inode->i_sb->s_blocksize_bits;
3432         if (eof_block < map->m_lblk + map_len)
3433                 eof_block = map->m_lblk + map_len;
3434
3435         depth = ext_depth(inode);
3436         eh = path[depth].p_hdr;
3437         ex = path[depth].p_ext;
3438         ee_block = le32_to_cpu(ex->ee_block);
3439         ee_len = ext4_ext_get_actual_len(ex);
3440         zero_ex1.ee_len = 0;
3441         zero_ex2.ee_len = 0;
3442
3443         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3444
3445         /* Pre-conditions */
3446         BUG_ON(!ext4_ext_is_unwritten(ex));
3447         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3448
3449         /*
3450          * Attempt to transfer newly initialized blocks from the currently
3451          * unwritten extent to its neighbor. This is much cheaper
3452          * than an insertion followed by a merge as those involve costly
3453          * memmove() calls. Transferring to the left is the common case in
3454          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3455          * followed by append writes.
3456          *
3457          * Limitations of the current logic:
3458          *  - L1: we do not deal with writes covering the whole extent.
3459          *    This would require removing the extent if the transfer
3460          *    is possible.
3461          *  - L2: we only attempt to merge with an extent stored in the
3462          *    same extent tree node.
3463          */
3464         if ((map->m_lblk == ee_block) &&
3465                 /* See if we can merge left */
3466                 (map_len < ee_len) &&           /*L1*/
3467                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3468                 ext4_lblk_t prev_lblk;
3469                 ext4_fsblk_t prev_pblk, ee_pblk;
3470                 unsigned int prev_len;
3471
3472                 abut_ex = ex - 1;
3473                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3474                 prev_len = ext4_ext_get_actual_len(abut_ex);
3475                 prev_pblk = ext4_ext_pblock(abut_ex);
3476                 ee_pblk = ext4_ext_pblock(ex);
3477
3478                 /*
3479                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3480                  * upon those conditions:
3481                  * - C1: abut_ex is initialized,
3482                  * - C2: abut_ex is logically abutting ex,
3483                  * - C3: abut_ex is physically abutting ex,
3484                  * - C4: abut_ex can receive the additional blocks without
3485                  *   overflowing the (initialized) length limit.
3486                  */
3487                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3488                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3489                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3490                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3491                         err = ext4_ext_get_access(handle, inode, path + depth);
3492                         if (err)
3493                                 goto out;
3494
3495                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3496                                 map, ex, abut_ex);
3497
3498                         /* Shift the start of ex by 'map_len' blocks */
3499                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3500                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3501                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3502                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3503
3504                         /* Extend abut_ex by 'map_len' blocks */
3505                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3506
3507                         /* Result: number of initialized blocks past m_lblk */
3508                         allocated = map_len;
3509                 }
3510         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3511                    (map_len < ee_len) &&        /*L1*/
3512                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3513                 /* See if we can merge right */
3514                 ext4_lblk_t next_lblk;
3515                 ext4_fsblk_t next_pblk, ee_pblk;
3516                 unsigned int next_len;
3517
3518                 abut_ex = ex + 1;
3519                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3520                 next_len = ext4_ext_get_actual_len(abut_ex);
3521                 next_pblk = ext4_ext_pblock(abut_ex);
3522                 ee_pblk = ext4_ext_pblock(ex);
3523
3524                 /*
3525                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3526                  * upon those conditions:
3527                  * - C1: abut_ex is initialized,
3528                  * - C2: abut_ex is logically abutting ex,
3529                  * - C3: abut_ex is physically abutting ex,
3530                  * - C4: abut_ex can receive the additional blocks without
3531                  *   overflowing the (initialized) length limit.
3532                  */
3533                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3534                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3535                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3536                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3537                         err = ext4_ext_get_access(handle, inode, path + depth);
3538                         if (err)
3539                                 goto out;
3540
3541                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3542                                 map, ex, abut_ex);
3543
3544                         /* Shift the start of abut_ex by 'map_len' blocks */
3545                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3546                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3547                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3548                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3549
3550                         /* Extend abut_ex by 'map_len' blocks */
3551                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3552
3553                         /* Result: number of initialized blocks past m_lblk */
3554                         allocated = map_len;
3555                 }
3556         }
3557         if (allocated) {
3558                 /* Mark the block containing both extents as dirty */
3559                 ext4_ext_dirty(handle, inode, path + depth);
3560
3561                 /* Update path to point to the right extent */
3562                 path[depth].p_ext = abut_ex;
3563                 goto out;
3564         } else
3565                 allocated = ee_len - (map->m_lblk - ee_block);
3566
3567         WARN_ON(map->m_lblk < ee_block);
3568         /*
3569          * It is safe to convert extent to initialized via explicit
3570          * zeroout only if extent is fully inside i_size or new_size.
3571          */
3572         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3573
3574         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3575                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3576                         (inode->i_sb->s_blocksize_bits - 10);
3577
3578         if (ext4_encrypted_inode(inode))
3579                 max_zeroout = 0;
3580
3581         /*
3582          * five cases:
3583          * 1. split the extent into three extents.
3584          * 2. split the extent into two extents, zeroout the head of the first
3585          *    extent.
3586          * 3. split the extent into two extents, zeroout the tail of the second
3587          *    extent.
3588          * 4. split the extent into two extents with out zeroout.
3589          * 5. no splitting needed, just possibly zeroout the head and / or the
3590          *    tail of the extent.
3591          */
3592         split_map.m_lblk = map->m_lblk;
3593         split_map.m_len = map->m_len;
3594
3595         if (max_zeroout && (allocated > split_map.m_len)) {
3596                 if (allocated <= max_zeroout) {
3597                         /* case 3 or 5 */
3598                         zero_ex1.ee_block =
3599                                  cpu_to_le32(split_map.m_lblk +
3600                                              split_map.m_len);
3601                         zero_ex1.ee_len =
3602                                 cpu_to_le16(allocated - split_map.m_len);
3603                         ext4_ext_store_pblock(&zero_ex1,
3604                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3605                                 split_map.m_len - ee_block);
3606                         err = ext4_ext_zeroout(inode, &zero_ex1);
3607                         if (err)
3608                                 goto out;
3609                         split_map.m_len = allocated;
3610                 }
3611                 if (split_map.m_lblk - ee_block + split_map.m_len <
3612                                                                 max_zeroout) {
3613                         /* case 2 or 5 */
3614                         if (split_map.m_lblk != ee_block) {
3615                                 zero_ex2.ee_block = ex->ee_block;
3616                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3617                                                         ee_block);
3618                                 ext4_ext_store_pblock(&zero_ex2,
3619                                                       ext4_ext_pblock(ex));
3620                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3621                                 if (err)
3622                                         goto out;
3623                         }
3624
3625                         split_map.m_len += split_map.m_lblk - ee_block;
3626                         split_map.m_lblk = ee_block;
3627                         allocated = map->m_len;
3628                 }
3629         }
3630
3631         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3632                                 flags);
3633         if (err > 0)
3634                 err = 0;
3635 out:
3636         /* If we have gotten a failure, don't zero out status tree */
3637         if (!err) {
3638                 err = ext4_zeroout_es(inode, &zero_ex1);
3639                 if (!err)
3640                         err = ext4_zeroout_es(inode, &zero_ex2);
3641         }
3642         return err ? err : allocated;
3643 }
3644
3645 /*
3646  * This function is called by ext4_ext_map_blocks() from
3647  * ext4_get_blocks_dio_write() when DIO to write
3648  * to an unwritten extent.
3649  *
3650  * Writing to an unwritten extent may result in splitting the unwritten
3651  * extent into multiple initialized/unwritten extents (up to three)
3652  * There are three possibilities:
3653  *   a> There is no split required: Entire extent should be unwritten
3654  *   b> Splits in two extents: Write is happening at either end of the extent
3655  *   c> Splits in three extents: Somone is writing in middle of the extent
3656  *
3657  * This works the same way in the case of initialized -> unwritten conversion.
3658  *
3659  * One of more index blocks maybe needed if the extent tree grow after
3660  * the unwritten extent split. To prevent ENOSPC occur at the IO
3661  * complete, we need to split the unwritten extent before DIO submit
3662  * the IO. The unwritten extent called at this time will be split
3663  * into three unwritten extent(at most). After IO complete, the part
3664  * being filled will be convert to initialized by the end_io callback function
3665  * via ext4_convert_unwritten_extents().
3666  *
3667  * Returns the size of unwritten extent to be written on success.
3668  */
3669 static int ext4_split_convert_extents(handle_t *handle,
3670                                         struct inode *inode,
3671                                         struct ext4_map_blocks *map,
3672                                         struct ext4_ext_path **ppath,
3673                                         int flags)
3674 {
3675         struct ext4_ext_path *path = *ppath;
3676         ext4_lblk_t eof_block;
3677         ext4_lblk_t ee_block;
3678         struct ext4_extent *ex;
3679         unsigned int ee_len;
3680         int split_flag = 0, depth;
3681
3682         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3683                   __func__, inode->i_ino,
3684                   (unsigned long long)map->m_lblk, map->m_len);
3685
3686         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3687                 inode->i_sb->s_blocksize_bits;
3688         if (eof_block < map->m_lblk + map->m_len)
3689                 eof_block = map->m_lblk + map->m_len;
3690         /*
3691          * It is safe to convert extent to initialized via explicit
3692          * zeroout only if extent is fully insde i_size or new_size.
3693          */
3694         depth = ext_depth(inode);
3695         ex = path[depth].p_ext;
3696         ee_block = le32_to_cpu(ex->ee_block);
3697         ee_len = ext4_ext_get_actual_len(ex);
3698
3699         /* Convert to unwritten */
3700         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3701                 split_flag |= EXT4_EXT_DATA_VALID1;
3702         /* Convert to initialized */
3703         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3704                 split_flag |= ee_block + ee_len <= eof_block ?
3705                               EXT4_EXT_MAY_ZEROOUT : 0;
3706                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3707         }
3708         flags |= EXT4_GET_BLOCKS_PRE_IO;
3709         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3710 }
3711
3712 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3713                                                 struct inode *inode,
3714                                                 struct ext4_map_blocks *map,
3715                                                 struct ext4_ext_path **ppath)
3716 {
3717         struct ext4_ext_path *path = *ppath;
3718         struct ext4_extent *ex;
3719         ext4_lblk_t ee_block;
3720         unsigned int ee_len;
3721         int depth;
3722         int err = 0;
3723
3724         depth = ext_depth(inode);
3725         ex = path[depth].p_ext;
3726         ee_block = le32_to_cpu(ex->ee_block);
3727         ee_len = ext4_ext_get_actual_len(ex);
3728
3729         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3730                 "block %llu, max_blocks %u\n", inode->i_ino,
3731                   (unsigned long long)ee_block, ee_len);
3732
3733         /* If extent is larger than requested it is a clear sign that we still
3734          * have some extent state machine issues left. So extent_split is still
3735          * required.
3736          * TODO: Once all related issues will be fixed this situation should be
3737          * illegal.
3738          */
3739         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3740 #ifdef EXT4_DEBUG
3741                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3742                              " len %u; IO logical block %llu, len %u",
3743                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3744                              (unsigned long long)map->m_lblk, map->m_len);
3745 #endif
3746                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3747                                                  EXT4_GET_BLOCKS_CONVERT);
3748                 if (err < 0)
3749                         return err;
3750                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3751                 if (IS_ERR(path))
3752                         return PTR_ERR(path);
3753                 depth = ext_depth(inode);
3754                 ex = path[depth].p_ext;
3755         }
3756
3757         err = ext4_ext_get_access(handle, inode, path + depth);
3758         if (err)
3759                 goto out;
3760         /* first mark the extent as initialized */
3761         ext4_ext_mark_initialized(ex);
3762
3763         /* note: ext4_ext_correct_indexes() isn't needed here because
3764          * borders are not changed
3765          */
3766         ext4_ext_try_to_merge(handle, inode, path, ex);
3767
3768         /* Mark modified extent as dirty */
3769         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3770 out:
3771         ext4_ext_show_leaf(inode, path);
3772         return err;
3773 }
3774
3775 /*
3776  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3777  */
3778 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3779                               ext4_lblk_t lblk,
3780                               struct ext4_ext_path *path,
3781                               unsigned int len)
3782 {
3783         int i, depth;
3784         struct ext4_extent_header *eh;
3785         struct ext4_extent *last_ex;
3786
3787         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3788                 return 0;
3789
3790         depth = ext_depth(inode);
3791         eh = path[depth].p_hdr;
3792
3793         /*
3794          * We're going to remove EOFBLOCKS_FL entirely in future so we
3795          * do not care for this case anymore. Simply remove the flag
3796          * if there are no extents.
3797          */
3798         if (unlikely(!eh->eh_entries))
3799                 goto out;
3800         last_ex = EXT_LAST_EXTENT(eh);
3801         /*
3802          * We should clear the EOFBLOCKS_FL flag if we are writing the
3803          * last block in the last extent in the file.  We test this by
3804          * first checking to see if the caller to
3805          * ext4_ext_get_blocks() was interested in the last block (or
3806          * a block beyond the last block) in the current extent.  If
3807          * this turns out to be false, we can bail out from this
3808          * function immediately.
3809          */
3810         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3811             ext4_ext_get_actual_len(last_ex))
3812                 return 0;
3813         /*
3814          * If the caller does appear to be planning to write at or
3815          * beyond the end of the current extent, we then test to see
3816          * if the current extent is the last extent in the file, by
3817          * checking to make sure it was reached via the rightmost node
3818          * at each level of the tree.
3819          */
3820         for (i = depth-1; i >= 0; i--)
3821                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3822                         return 0;
3823 out:
3824         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3825         return ext4_mark_inode_dirty(handle, inode);
3826 }
3827
3828 /**
3829  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3830  *
3831  * Return 1 if there is a delalloc block in the range, otherwise 0.
3832  */
3833 int ext4_find_delalloc_range(struct inode *inode,
3834                              ext4_lblk_t lblk_start,
3835                              ext4_lblk_t lblk_end)
3836 {
3837         struct extent_status es;
3838
3839         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3840         if (es.es_len == 0)
3841                 return 0; /* there is no delay extent in this tree */
3842         else if (es.es_lblk <= lblk_start &&
3843                  lblk_start < es.es_lblk + es.es_len)
3844                 return 1;
3845         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3846                 return 1;
3847         else
3848                 return 0;
3849 }
3850
3851 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3852 {
3853         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3854         ext4_lblk_t lblk_start, lblk_end;
3855         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3856         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3857
3858         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3859 }
3860
3861 /**
3862  * Determines how many complete clusters (out of those specified by the 'map')
3863  * are under delalloc and were reserved quota for.
3864  * This function is called when we are writing out the blocks that were
3865  * originally written with their allocation delayed, but then the space was
3866  * allocated using fallocate() before the delayed allocation could be resolved.
3867  * The cases to look for are:
3868  * ('=' indicated delayed allocated blocks
3869  *  '-' indicates non-delayed allocated blocks)
3870  * (a) partial clusters towards beginning and/or end outside of allocated range
3871  *     are not delalloc'ed.
3872  *      Ex:
3873  *      |----c---=|====c====|====c====|===-c----|
3874  *               |++++++ allocated ++++++|
3875  *      ==> 4 complete clusters in above example
3876  *
3877  * (b) partial cluster (outside of allocated range) towards either end is
3878  *     marked for delayed allocation. In this case, we will exclude that
3879  *     cluster.
3880  *      Ex:
3881  *      |----====c========|========c========|
3882  *           |++++++ allocated ++++++|
3883  *      ==> 1 complete clusters in above example
3884  *
3885  *      Ex:
3886  *      |================c================|
3887  *            |++++++ allocated ++++++|
3888  *      ==> 0 complete clusters in above example
3889  *
3890  * The ext4_da_update_reserve_space will be called only if we
3891  * determine here that there were some "entire" clusters that span
3892  * this 'allocated' range.
3893  * In the non-bigalloc case, this function will just end up returning num_blks
3894  * without ever calling ext4_find_delalloc_range.
3895  */
3896 static unsigned int
3897 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3898                            unsigned int num_blks)
3899 {
3900         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3901         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3902         ext4_lblk_t lblk_from, lblk_to, c_offset;
3903         unsigned int allocated_clusters = 0;
3904
3905         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3906         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3907
3908         /* max possible clusters for this allocation */
3909         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3910
3911         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3912
3913         /* Check towards left side */
3914         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3915         if (c_offset) {
3916                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3917                 lblk_to = lblk_from + c_offset - 1;
3918
3919                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3920                         allocated_clusters--;
3921         }
3922
3923         /* Now check towards right. */
3924         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3925         if (allocated_clusters && c_offset) {
3926                 lblk_from = lblk_start + num_blks;
3927                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3928
3929                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3930                         allocated_clusters--;
3931         }
3932
3933         return allocated_clusters;
3934 }
3935
3936 static int
3937 convert_initialized_extent(handle_t *handle, struct inode *inode,
3938                            struct ext4_map_blocks *map,
3939                            struct ext4_ext_path **ppath,
3940                            unsigned int allocated)
3941 {
3942         struct ext4_ext_path *path = *ppath;
3943         struct ext4_extent *ex;
3944         ext4_lblk_t ee_block;
3945         unsigned int ee_len;
3946         int depth;
3947         int err = 0;
3948
3949         /*
3950          * Make sure that the extent is no bigger than we support with
3951          * unwritten extent
3952          */
3953         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3954                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3955
3956         depth = ext_depth(inode);
3957         ex = path[depth].p_ext;
3958         ee_block = le32_to_cpu(ex->ee_block);
3959         ee_len = ext4_ext_get_actual_len(ex);
3960
3961         ext_debug("%s: inode %lu, logical"
3962                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3963                   (unsigned long long)ee_block, ee_len);
3964
3965         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3966                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3967                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3968                 if (err < 0)
3969                         return err;
3970                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3971                 if (IS_ERR(path))
3972                         return PTR_ERR(path);
3973                 depth = ext_depth(inode);
3974                 ex = path[depth].p_ext;
3975                 if (!ex) {
3976                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3977                                          (unsigned long) map->m_lblk);
3978                         return -EFSCORRUPTED;
3979                 }
3980         }
3981
3982         err = ext4_ext_get_access(handle, inode, path + depth);
3983         if (err)
3984                 return err;
3985         /* first mark the extent as unwritten */
3986         ext4_ext_mark_unwritten(ex);
3987
3988         /* note: ext4_ext_correct_indexes() isn't needed here because
3989          * borders are not changed
3990          */
3991         ext4_ext_try_to_merge(handle, inode, path, ex);
3992
3993         /* Mark modified extent as dirty */
3994         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3995         if (err)
3996                 return err;
3997         ext4_ext_show_leaf(inode, path);
3998
3999         ext4_update_inode_fsync_trans(handle, inode, 1);
4000         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
4001         if (err)
4002                 return err;
4003         map->m_flags |= EXT4_MAP_UNWRITTEN;
4004         if (allocated > map->m_len)
4005                 allocated = map->m_len;
4006         map->m_len = allocated;
4007         return allocated;
4008 }
4009
4010 static int
4011 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4012                         struct ext4_map_blocks *map,
4013                         struct ext4_ext_path **ppath, int flags,
4014                         unsigned int allocated, ext4_fsblk_t newblock)
4015 {
4016         struct ext4_ext_path *path = *ppath;
4017         int ret = 0;
4018         int err = 0;
4019
4020         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4021                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4022                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4023                   flags, allocated);
4024         ext4_ext_show_leaf(inode, path);
4025
4026         /*
4027          * When writing into unwritten space, we should not fail to
4028          * allocate metadata blocks for the new extent block if needed.
4029          */
4030         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4031
4032         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4033                                                     allocated, newblock);
4034
4035         /* get_block() before submit the IO, split the extent */
4036         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4037                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4038                                          flags | EXT4_GET_BLOCKS_CONVERT);
4039                 if (ret <= 0)
4040                         goto out;
4041                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4042                 goto out;
4043         }
4044         /* IO end_io complete, convert the filled extent to written */
4045         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4046                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4047                         if (allocated > map->m_len)
4048                                 allocated = map->m_len;
4049                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4050                                                  allocated);
4051                         if (err < 0)
4052                                 goto out2;
4053                 }
4054                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4055                                                            ppath);
4056                 if (ret >= 0) {
4057                         ext4_update_inode_fsync_trans(handle, inode, 1);
4058                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4059                                                  path, map->m_len);
4060                 } else
4061                         err = ret;
4062                 map->m_flags |= EXT4_MAP_MAPPED;
4063                 map->m_pblk = newblock;
4064                 if (allocated > map->m_len)
4065                         allocated = map->m_len;
4066                 map->m_len = allocated;
4067                 goto out2;
4068         }
4069         /* buffered IO case */
4070         /*
4071          * repeat fallocate creation request
4072          * we already have an unwritten extent
4073          */
4074         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4075                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4076                 goto map_out;
4077         }
4078
4079         /* buffered READ or buffered write_begin() lookup */
4080         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4081                 /*
4082                  * We have blocks reserved already.  We
4083                  * return allocated blocks so that delalloc
4084                  * won't do block reservation for us.  But
4085                  * the buffer head will be unmapped so that
4086                  * a read from the block returns 0s.
4087                  */
4088                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4089                 goto out1;
4090         }
4091
4092         /* buffered write, writepage time, convert*/
4093         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4094         if (ret >= 0)
4095                 ext4_update_inode_fsync_trans(handle, inode, 1);
4096 out:
4097         if (ret <= 0) {
4098                 err = ret;
4099                 goto out2;
4100         } else
4101                 allocated = ret;
4102         map->m_flags |= EXT4_MAP_NEW;
4103         /*
4104          * if we allocated more blocks than requested
4105          * we need to make sure we unmap the extra block
4106          * allocated. The actual needed block will get
4107          * unmapped later when we find the buffer_head marked
4108          * new.
4109          */
4110         if (allocated > map->m_len) {
4111                 clean_bdev_aliases(inode->i_sb->s_bdev, newblock + map->m_len,
4112                                    allocated - map->m_len);
4113                 allocated = map->m_len;
4114         }
4115         map->m_len = allocated;
4116
4117         /*
4118          * If we have done fallocate with the offset that is already
4119          * delayed allocated, we would have block reservation
4120          * and quota reservation done in the delayed write path.
4121          * But fallocate would have already updated quota and block
4122          * count for this offset. So cancel these reservation
4123          */
4124         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4125                 unsigned int reserved_clusters;
4126                 reserved_clusters = get_reserved_cluster_alloc(inode,
4127                                 map->m_lblk, map->m_len);
4128                 if (reserved_clusters)
4129                         ext4_da_update_reserve_space(inode,
4130                                                      reserved_clusters,
4131                                                      0);
4132         }
4133
4134 map_out:
4135         map->m_flags |= EXT4_MAP_MAPPED;
4136         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4137                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4138                                          map->m_len);
4139                 if (err < 0)
4140                         goto out2;
4141         }
4142 out1:
4143         if (allocated > map->m_len)
4144                 allocated = map->m_len;
4145         ext4_ext_show_leaf(inode, path);
4146         map->m_pblk = newblock;
4147         map->m_len = allocated;
4148 out2:
4149         return err ? err : allocated;
4150 }
4151
4152 /*
4153  * get_implied_cluster_alloc - check to see if the requested
4154  * allocation (in the map structure) overlaps with a cluster already
4155  * allocated in an extent.
4156  *      @sb     The filesystem superblock structure
4157  *      @map    The requested lblk->pblk mapping
4158  *      @ex     The extent structure which might contain an implied
4159  *                      cluster allocation
4160  *
4161  * This function is called by ext4_ext_map_blocks() after we failed to
4162  * find blocks that were already in the inode's extent tree.  Hence,
4163  * we know that the beginning of the requested region cannot overlap
4164  * the extent from the inode's extent tree.  There are three cases we
4165  * want to catch.  The first is this case:
4166  *
4167  *               |--- cluster # N--|
4168  *    |--- extent ---|  |---- requested region ---|
4169  *                      |==========|
4170  *
4171  * The second case that we need to test for is this one:
4172  *
4173  *   |--------- cluster # N ----------------|
4174  *         |--- requested region --|   |------- extent ----|
4175  *         |=======================|
4176  *
4177  * The third case is when the requested region lies between two extents
4178  * within the same cluster:
4179  *          |------------- cluster # N-------------|
4180  * |----- ex -----|                  |---- ex_right ----|
4181  *                  |------ requested region ------|
4182  *                  |================|
4183  *
4184  * In each of the above cases, we need to set the map->m_pblk and
4185  * map->m_len so it corresponds to the return the extent labelled as
4186  * "|====|" from cluster #N, since it is already in use for data in
4187  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4188  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4189  * as a new "allocated" block region.  Otherwise, we will return 0 and
4190  * ext4_ext_map_blocks() will then allocate one or more new clusters
4191  * by calling ext4_mb_new_blocks().
4192  */
4193 static int get_implied_cluster_alloc(struct super_block *sb,
4194                                      struct ext4_map_blocks *map,
4195                                      struct ext4_extent *ex,
4196                                      struct ext4_ext_path *path)
4197 {
4198         struct ext4_sb_info *sbi = EXT4_SB(sb);
4199         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4200         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4201         ext4_lblk_t rr_cluster_start;
4202         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4203         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4204         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4205
4206         /* The extent passed in that we are trying to match */
4207         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4208         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4209
4210         /* The requested region passed into ext4_map_blocks() */
4211         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4212
4213         if ((rr_cluster_start == ex_cluster_end) ||
4214             (rr_cluster_start == ex_cluster_start)) {
4215                 if (rr_cluster_start == ex_cluster_end)
4216                         ee_start += ee_len - 1;
4217                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4218                 map->m_len = min(map->m_len,
4219                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4220                 /*
4221                  * Check for and handle this case:
4222                  *
4223                  *   |--------- cluster # N-------------|
4224                  *                     |------- extent ----|
4225                  *         |--- requested region ---|
4226                  *         |===========|
4227                  */
4228
4229                 if (map->m_lblk < ee_block)
4230                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4231
4232                 /*
4233                  * Check for the case where there is already another allocated
4234                  * block to the right of 'ex' but before the end of the cluster.
4235                  *
4236                  *          |------------- cluster # N-------------|
4237                  * |----- ex -----|                  |---- ex_right ----|
4238                  *                  |------ requested region ------|
4239                  *                  |================|
4240                  */
4241                 if (map->m_lblk > ee_block) {
4242                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4243                         map->m_len = min(map->m_len, next - map->m_lblk);
4244                 }
4245
4246                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4247                 return 1;
4248         }
4249
4250         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4251         return 0;
4252 }
4253
4254
4255 /*
4256  * Block allocation/map/preallocation routine for extents based files
4257  *
4258  *
4259  * Need to be called with
4260  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4261  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4262  *
4263  * return > 0, number of of blocks already mapped/allocated
4264  *          if create == 0 and these are pre-allocated blocks
4265  *              buffer head is unmapped
4266  *          otherwise blocks are mapped
4267  *
4268  * return = 0, if plain look up failed (blocks have not been allocated)
4269  *          buffer head is unmapped
4270  *
4271  * return < 0, error case.
4272  */
4273 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4274                         struct ext4_map_blocks *map, int flags)
4275 {
4276         struct ext4_ext_path *path = NULL;
4277         struct ext4_extent newex, *ex, *ex2;
4278         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4279         ext4_fsblk_t newblock = 0;
4280         int free_on_err = 0, err = 0, depth, ret;
4281         unsigned int allocated = 0, offset = 0;
4282         unsigned int allocated_clusters = 0;
4283         struct ext4_allocation_request ar;
4284         ext4_lblk_t cluster_offset;
4285         bool map_from_cluster = false;
4286
4287         ext_debug("blocks %u/%u requested for inode %lu\n",
4288                   map->m_lblk, map->m_len, inode->i_ino);
4289         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4290
4291         /* find extent for this block */
4292         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4293         if (IS_ERR(path)) {
4294                 err = PTR_ERR(path);
4295                 path = NULL;
4296                 goto out2;
4297         }
4298
4299         depth = ext_depth(inode);
4300
4301         /*
4302          * consistent leaf must not be empty;
4303          * this situation is possible, though, _during_ tree modification;
4304          * this is why assert can't be put in ext4_find_extent()
4305          */
4306         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4307                 EXT4_ERROR_INODE(inode, "bad extent address "
4308                                  "lblock: %lu, depth: %d pblock %lld",
4309                                  (unsigned long) map->m_lblk, depth,
4310                                  path[depth].p_block);
4311                 err = -EFSCORRUPTED;
4312                 goto out2;
4313         }
4314
4315         ex = path[depth].p_ext;
4316         if (ex) {
4317                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4318                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4319                 unsigned short ee_len;
4320
4321
4322                 /*
4323                  * unwritten extents are treated as holes, except that
4324                  * we split out initialized portions during a write.
4325                  */
4326                 ee_len = ext4_ext_get_actual_len(ex);
4327
4328                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4329
4330                 /* if found extent covers block, simply return it */
4331                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4332                         newblock = map->m_lblk - ee_block + ee_start;
4333                         /* number of remaining blocks in the extent */
4334                         allocated = ee_len - (map->m_lblk - ee_block);
4335                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4336                                   ee_block, ee_len, newblock);
4337
4338                         /*
4339                          * If the extent is initialized check whether the
4340                          * caller wants to convert it to unwritten.
4341                          */
4342                         if ((!ext4_ext_is_unwritten(ex)) &&
4343                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4344                                 allocated = convert_initialized_extent(
4345                                                 handle, inode, map, &path,
4346                                                 allocated);
4347                                 goto out2;
4348                         } else if (!ext4_ext_is_unwritten(ex))
4349                                 goto out;
4350
4351                         ret = ext4_ext_handle_unwritten_extents(
4352                                 handle, inode, map, &path, flags,
4353                                 allocated, newblock);
4354                         if (ret < 0)
4355                                 err = ret;
4356                         else
4357                                 allocated = ret;
4358                         goto out2;
4359                 }
4360         }
4361
4362         /*
4363          * requested block isn't allocated yet;
4364          * we couldn't try to create block if create flag is zero
4365          */
4366         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4367                 ext4_lblk_t hole_start, hole_len;
4368
4369                 hole_start = map->m_lblk;
4370                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4371                 /*
4372                  * put just found gap into cache to speed up
4373                  * subsequent requests
4374                  */
4375                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4376
4377                 /* Update hole_len to reflect hole size after map->m_lblk */
4378                 if (hole_start != map->m_lblk)
4379                         hole_len -= map->m_lblk - hole_start;
4380                 map->m_pblk = 0;
4381                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4382
4383                 goto out2;
4384         }
4385
4386         /*
4387          * Okay, we need to do block allocation.
4388          */
4389         newex.ee_block = cpu_to_le32(map->m_lblk);
4390         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4391
4392         /*
4393          * If we are doing bigalloc, check to see if the extent returned
4394          * by ext4_find_extent() implies a cluster we can use.
4395          */
4396         if (cluster_offset && ex &&
4397             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4398                 ar.len = allocated = map->m_len;
4399                 newblock = map->m_pblk;
4400                 map_from_cluster = true;
4401                 goto got_allocated_blocks;
4402         }
4403
4404         /* find neighbour allocated blocks */
4405         ar.lleft = map->m_lblk;
4406         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4407         if (err)
4408                 goto out2;
4409         ar.lright = map->m_lblk;
4410         ex2 = NULL;
4411         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4412         if (err)
4413                 goto out2;
4414
4415         /* Check if the extent after searching to the right implies a
4416          * cluster we can use. */
4417         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4418             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4419                 ar.len = allocated = map->m_len;
4420                 newblock = map->m_pblk;
4421                 map_from_cluster = true;
4422                 goto got_allocated_blocks;
4423         }
4424
4425         /*
4426          * See if request is beyond maximum number of blocks we can have in
4427          * a single extent. For an initialized extent this limit is
4428          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4429          * EXT_UNWRITTEN_MAX_LEN.
4430          */
4431         if (map->m_len > EXT_INIT_MAX_LEN &&
4432             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4433                 map->m_len = EXT_INIT_MAX_LEN;
4434         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4435                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4436                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4437
4438         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4439         newex.ee_len = cpu_to_le16(map->m_len);
4440         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4441         if (err)
4442                 allocated = ext4_ext_get_actual_len(&newex);
4443         else
4444                 allocated = map->m_len;
4445
4446         /* allocate new block */
4447         ar.inode = inode;
4448         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4449         ar.logical = map->m_lblk;
4450         /*
4451          * We calculate the offset from the beginning of the cluster
4452          * for the logical block number, since when we allocate a
4453          * physical cluster, the physical block should start at the
4454          * same offset from the beginning of the cluster.  This is
4455          * needed so that future calls to get_implied_cluster_alloc()
4456          * work correctly.
4457          */
4458         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4459         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4460         ar.goal -= offset;
4461         ar.logical -= offset;
4462         if (S_ISREG(inode->i_mode))
4463                 ar.flags = EXT4_MB_HINT_DATA;
4464         else
4465                 /* disable in-core preallocation for non-regular files */
4466                 ar.flags = 0;
4467         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4468                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4469         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4470                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4471         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4472                 ar.flags |= EXT4_MB_USE_RESERVED;
4473         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4474         if (!newblock)
4475                 goto out2;
4476         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4477                   ar.goal, newblock, allocated);
4478         free_on_err = 1;
4479         allocated_clusters = ar.len;
4480         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4481         if (ar.len > allocated)
4482                 ar.len = allocated;
4483
4484 got_allocated_blocks:
4485         /* try to insert new extent into found leaf and return */
4486         ext4_ext_store_pblock(&newex, newblock + offset);
4487         newex.ee_len = cpu_to_le16(ar.len);
4488         /* Mark unwritten */
4489         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4490                 ext4_ext_mark_unwritten(&newex);
4491                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4492         }
4493
4494         err = 0;
4495         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4496                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4497                                          path, ar.len);
4498         if (!err)
4499                 err = ext4_ext_insert_extent(handle, inode, &path,
4500                                              &newex, flags);
4501
4502         if (err && free_on_err) {
4503                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4504                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4505                 /* free data blocks we just allocated */
4506                 /* not a good idea to call discard here directly,
4507                  * but otherwise we'd need to call it every free() */
4508                 ext4_discard_preallocations(inode);
4509                 ext4_free_blocks(handle, inode, NULL, newblock,
4510                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4511                 goto out2;
4512         }
4513
4514         /* previous routine could use block we allocated */
4515         newblock = ext4_ext_pblock(&newex);
4516         allocated = ext4_ext_get_actual_len(&newex);
4517         if (allocated > map->m_len)
4518                 allocated = map->m_len;
4519         map->m_flags |= EXT4_MAP_NEW;
4520
4521         /*
4522          * Update reserved blocks/metadata blocks after successful
4523          * block allocation which had been deferred till now.
4524          */
4525         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4526                 unsigned int reserved_clusters;
4527                 /*
4528                  * Check how many clusters we had reserved this allocated range
4529                  */
4530                 reserved_clusters = get_reserved_cluster_alloc(inode,
4531                                                 map->m_lblk, allocated);
4532                 if (!map_from_cluster) {
4533                         BUG_ON(allocated_clusters < reserved_clusters);
4534                         if (reserved_clusters < allocated_clusters) {
4535                                 struct ext4_inode_info *ei = EXT4_I(inode);
4536                                 int reservation = allocated_clusters -
4537                                                   reserved_clusters;
4538                                 /*
4539                                  * It seems we claimed few clusters outside of
4540                                  * the range of this allocation. We should give
4541                                  * it back to the reservation pool. This can
4542                                  * happen in the following case:
4543                                  *
4544                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4545                                  *   cluster has 4 blocks. Thus, the clusters
4546                                  *   are [0-3],[4-7],[8-11]...
4547                                  * * First comes delayed allocation write for
4548                                  *   logical blocks 10 & 11. Since there were no
4549                                  *   previous delayed allocated blocks in the
4550                                  *   range [8-11], we would reserve 1 cluster
4551                                  *   for this write.
4552                                  * * Next comes write for logical blocks 3 to 8.
4553                                  *   In this case, we will reserve 2 clusters
4554                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4555                                  *   that range has a delayed allocated blocks.
4556                                  *   Thus total reserved clusters now becomes 3.
4557                                  * * Now, during the delayed allocation writeout
4558                                  *   time, we will first write blocks [3-8] and
4559                                  *   allocate 3 clusters for writing these
4560                                  *   blocks. Also, we would claim all these
4561                                  *   three clusters above.
4562                                  * * Now when we come here to writeout the
4563                                  *   blocks [10-11], we would expect to claim
4564                                  *   the reservation of 1 cluster we had made
4565                                  *   (and we would claim it since there are no
4566                                  *   more delayed allocated blocks in the range
4567                                  *   [8-11]. But our reserved cluster count had
4568                                  *   already gone to 0.
4569                                  *
4570                                  *   Thus, at the step 4 above when we determine
4571                                  *   that there are still some unwritten delayed
4572                                  *   allocated blocks outside of our current
4573                                  *   block range, we should increment the
4574                                  *   reserved clusters count so that when the
4575                                  *   remaining blocks finally gets written, we
4576                                  *   could claim them.
4577                                  */
4578                                 dquot_reserve_block(inode,
4579                                                 EXT4_C2B(sbi, reservation));
4580                                 spin_lock(&ei->i_block_reservation_lock);
4581                                 ei->i_reserved_data_blocks += reservation;
4582                                 spin_unlock(&ei->i_block_reservation_lock);
4583                         }
4584                         /*
4585                          * We will claim quota for all newly allocated blocks.
4586                          * We're updating the reserved space *after* the
4587                          * correction above so we do not accidentally free
4588                          * all the metadata reservation because we might
4589                          * actually need it later on.
4590                          */
4591                         ext4_da_update_reserve_space(inode, allocated_clusters,
4592                                                         1);
4593                 }
4594         }
4595
4596         /*
4597          * Cache the extent and update transaction to commit on fdatasync only
4598          * when it is _not_ an unwritten extent.
4599          */
4600         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4601                 ext4_update_inode_fsync_trans(handle, inode, 1);
4602         else
4603                 ext4_update_inode_fsync_trans(handle, inode, 0);
4604 out:
4605         if (allocated > map->m_len)
4606                 allocated = map->m_len;
4607         ext4_ext_show_leaf(inode, path);
4608         map->m_flags |= EXT4_MAP_MAPPED;
4609         map->m_pblk = newblock;
4610         map->m_len = allocated;
4611 out2:
4612         ext4_ext_drop_refs(path);
4613         kfree(path);
4614
4615         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4616                                        err ? err : allocated);
4617         return err ? err : allocated;
4618 }
4619
4620 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4621 {
4622         struct super_block *sb = inode->i_sb;
4623         ext4_lblk_t last_block;
4624         int err = 0;
4625
4626         /*
4627          * TODO: optimization is possible here.
4628          * Probably we need not scan at all,
4629          * because page truncation is enough.
4630          */
4631
4632         /* we have to know where to truncate from in crash case */
4633         EXT4_I(inode)->i_disksize = inode->i_size;
4634         err = ext4_mark_inode_dirty(handle, inode);
4635         if (err)
4636                 return err;
4637
4638         last_block = (inode->i_size + sb->s_blocksize - 1)
4639                         >> EXT4_BLOCK_SIZE_BITS(sb);
4640 retry:
4641         err = ext4_es_remove_extent(inode, last_block,
4642                                     EXT_MAX_BLOCKS - last_block);
4643         if (err == -ENOMEM) {
4644                 cond_resched();
4645                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4646                 goto retry;
4647         }
4648         if (err)
4649                 return err;
4650         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4651 }
4652
4653 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4654                                   ext4_lblk_t len, loff_t new_size,
4655                                   int flags)
4656 {
4657         struct inode *inode = file_inode(file);
4658         handle_t *handle;
4659         int ret = 0;
4660         int ret2 = 0;
4661         int retries = 0;
4662         int depth = 0;
4663         struct ext4_map_blocks map;
4664         unsigned int credits;
4665         loff_t epos;
4666
4667         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4668         map.m_lblk = offset;
4669         map.m_len = len;
4670         /*
4671          * Don't normalize the request if it can fit in one extent so
4672          * that it doesn't get unnecessarily split into multiple
4673          * extents.
4674          */
4675         if (len <= EXT_UNWRITTEN_MAX_LEN)
4676                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4677
4678         /*
4679          * credits to insert 1 extent into extent tree
4680          */
4681         credits = ext4_chunk_trans_blocks(inode, len);
4682         depth = ext_depth(inode);
4683
4684 retry:
4685         while (ret >= 0 && len) {
4686                 /*
4687                  * Recalculate credits when extent tree depth changes.
4688                  */
4689                 if (depth != ext_depth(inode)) {
4690                         credits = ext4_chunk_trans_blocks(inode, len);
4691                         depth = ext_depth(inode);
4692                 }
4693
4694                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4695                                             credits);
4696                 if (IS_ERR(handle)) {
4697                         ret = PTR_ERR(handle);
4698                         break;
4699                 }
4700                 ret = ext4_map_blocks(handle, inode, &map, flags);
4701                 if (ret <= 0) {
4702                         ext4_debug("inode #%lu: block %u: len %u: "
4703                                    "ext4_ext_map_blocks returned %d",
4704                                    inode->i_ino, map.m_lblk,
4705                                    map.m_len, ret);
4706                         ext4_mark_inode_dirty(handle, inode);
4707                         ret2 = ext4_journal_stop(handle);
4708                         break;
4709                 }
4710                 map.m_lblk += ret;
4711                 map.m_len = len = len - ret;
4712                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4713                 inode->i_ctime = current_time(inode);
4714                 if (new_size) {
4715                         if (epos > new_size)
4716                                 epos = new_size;
4717                         if (ext4_update_inode_size(inode, epos) & 0x1)
4718                                 inode->i_mtime = inode->i_ctime;
4719                 } else {
4720                         if (epos > inode->i_size)
4721                                 ext4_set_inode_flag(inode,
4722                                                     EXT4_INODE_EOFBLOCKS);
4723                 }
4724                 ext4_mark_inode_dirty(handle, inode);
4725                 ret2 = ext4_journal_stop(handle);
4726                 if (ret2)
4727                         break;
4728         }
4729         if (ret == -ENOSPC &&
4730                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4731                 ret = 0;
4732                 goto retry;
4733         }
4734
4735         return ret > 0 ? ret2 : ret;
4736 }
4737
4738 static long ext4_zero_range(struct file *file, loff_t offset,
4739                             loff_t len, int mode)
4740 {
4741         struct inode *inode = file_inode(file);
4742         handle_t *handle = NULL;
4743         unsigned int max_blocks;
4744         loff_t new_size = 0;
4745         int ret = 0;
4746         int flags;
4747         int credits;
4748         int partial_begin, partial_end;
4749         loff_t start, end;
4750         ext4_lblk_t lblk;
4751         unsigned int blkbits = inode->i_blkbits;
4752
4753         trace_ext4_zero_range(inode, offset, len, mode);
4754
4755         if (!S_ISREG(inode->i_mode))
4756                 return -EINVAL;
4757
4758         /* Call ext4_force_commit to flush all data in case of data=journal. */
4759         if (ext4_should_journal_data(inode)) {
4760                 ret = ext4_force_commit(inode->i_sb);
4761                 if (ret)
4762                         return ret;
4763         }
4764
4765         /*
4766          * Round up offset. This is not fallocate, we neet to zero out
4767          * blocks, so convert interior block aligned part of the range to
4768          * unwritten and possibly manually zero out unaligned parts of the
4769          * range.
4770          */
4771         start = round_up(offset, 1 << blkbits);
4772         end = round_down((offset + len), 1 << blkbits);
4773
4774         if (start < offset || end > offset + len)
4775                 return -EINVAL;
4776         partial_begin = offset & ((1 << blkbits) - 1);
4777         partial_end = (offset + len) & ((1 << blkbits) - 1);
4778
4779         lblk = start >> blkbits;
4780         max_blocks = (end >> blkbits);
4781         if (max_blocks < lblk)
4782                 max_blocks = 0;
4783         else
4784                 max_blocks -= lblk;
4785
4786         inode_lock(inode);
4787
4788         /*
4789          * Indirect files do not support unwritten extnets
4790          */
4791         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4792                 ret = -EOPNOTSUPP;
4793                 goto out_mutex;
4794         }
4795
4796         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4797              offset + len > i_size_read(inode)) {
4798                 new_size = offset + len;
4799                 ret = inode_newsize_ok(inode, new_size);
4800                 if (ret)
4801                         goto out_mutex;
4802         }
4803
4804         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4805         if (mode & FALLOC_FL_KEEP_SIZE)
4806                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4807
4808         /* Wait all existing dio workers, newcomers will block on i_mutex */
4809         ext4_inode_block_unlocked_dio(inode);
4810         inode_dio_wait(inode);
4811
4812         /* Preallocate the range including the unaligned edges */
4813         if (partial_begin || partial_end) {
4814                 ret = ext4_alloc_file_blocks(file,
4815                                 round_down(offset, 1 << blkbits) >> blkbits,
4816                                 (round_up((offset + len), 1 << blkbits) -
4817                                  round_down(offset, 1 << blkbits)) >> blkbits,
4818                                 new_size, flags);
4819                 if (ret)
4820                         goto out_dio;
4821
4822         }
4823
4824         /* Zero range excluding the unaligned edges */
4825         if (max_blocks > 0) {
4826                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4827                           EXT4_EX_NOCACHE);
4828
4829                 /*
4830                  * Prevent page faults from reinstantiating pages we have
4831                  * released from page cache.
4832                  */
4833                 down_write(&EXT4_I(inode)->i_mmap_sem);
4834                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4835                 if (ret) {
4836                         up_write(&EXT4_I(inode)->i_mmap_sem);
4837                         goto out_dio;
4838                 }
4839                 /* Now release the pages and zero block aligned part of pages */
4840                 truncate_pagecache_range(inode, start, end - 1);
4841                 inode->i_mtime = inode->i_ctime = current_time(inode);
4842
4843                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4844                                              flags);
4845                 up_write(&EXT4_I(inode)->i_mmap_sem);
4846                 if (ret)
4847                         goto out_dio;
4848         }
4849         if (!partial_begin && !partial_end)
4850                 goto out_dio;
4851
4852         /*
4853          * In worst case we have to writeout two nonadjacent unwritten
4854          * blocks and update the inode
4855          */
4856         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4857         if (ext4_should_journal_data(inode))
4858                 credits += 2;
4859         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4860         if (IS_ERR(handle)) {
4861                 ret = PTR_ERR(handle);
4862                 ext4_std_error(inode->i_sb, ret);
4863                 goto out_dio;
4864         }
4865
4866         inode->i_mtime = inode->i_ctime = current_time(inode);
4867         if (new_size) {
4868                 ext4_update_inode_size(inode, new_size);
4869         } else {
4870                 /*
4871                 * Mark that we allocate beyond EOF so the subsequent truncate
4872                 * can proceed even if the new size is the same as i_size.
4873                 */
4874                 if ((offset + len) > i_size_read(inode))
4875                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4876         }
4877         ext4_mark_inode_dirty(handle, inode);
4878
4879         /* Zero out partial block at the edges of the range */
4880         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4881         if (ret >= 0)
4882                 ext4_update_inode_fsync_trans(handle, inode, 1);
4883
4884         if (file->f_flags & O_SYNC)
4885                 ext4_handle_sync(handle);
4886
4887         ext4_journal_stop(handle);
4888 out_dio:
4889         ext4_inode_resume_unlocked_dio(inode);
4890 out_mutex:
4891         inode_unlock(inode);
4892         return ret;
4893 }
4894
4895 /*
4896  * preallocate space for a file. This implements ext4's fallocate file
4897  * operation, which gets called from sys_fallocate system call.
4898  * For block-mapped files, posix_fallocate should fall back to the method
4899  * of writing zeroes to the required new blocks (the same behavior which is
4900  * expected for file systems which do not support fallocate() system call).
4901  */
4902 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4903 {
4904         struct inode *inode = file_inode(file);
4905         loff_t new_size = 0;
4906         unsigned int max_blocks;
4907         int ret = 0;
4908         int flags;
4909         ext4_lblk_t lblk;
4910         unsigned int blkbits = inode->i_blkbits;
4911
4912         /*
4913          * Encrypted inodes can't handle collapse range or insert
4914          * range since we would need to re-encrypt blocks with a
4915          * different IV or XTS tweak (which are based on the logical
4916          * block number).
4917          *
4918          * XXX It's not clear why zero range isn't working, but we'll
4919          * leave it disabled for encrypted inodes for now.  This is a
4920          * bug we should fix....
4921          */
4922         if (ext4_encrypted_inode(inode) &&
4923             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4924                      FALLOC_FL_ZERO_RANGE)))
4925                 return -EOPNOTSUPP;
4926
4927         /* Return error if mode is not supported */
4928         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4929                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4930                      FALLOC_FL_INSERT_RANGE))
4931                 return -EOPNOTSUPP;
4932
4933         if (mode & FALLOC_FL_PUNCH_HOLE)
4934                 return ext4_punch_hole(inode, offset, len);
4935
4936         ret = ext4_convert_inline_data(inode);
4937         if (ret)
4938                 return ret;
4939
4940         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4941                 return ext4_collapse_range(inode, offset, len);
4942
4943         if (mode & FALLOC_FL_INSERT_RANGE)
4944                 return ext4_insert_range(inode, offset, len);
4945
4946         if (mode & FALLOC_FL_ZERO_RANGE)
4947                 return ext4_zero_range(file, offset, len, mode);
4948
4949         trace_ext4_fallocate_enter(inode, offset, len, mode);
4950         lblk = offset >> blkbits;
4951
4952         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4953         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4954         if (mode & FALLOC_FL_KEEP_SIZE)
4955                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4956
4957         inode_lock(inode);
4958
4959         /*
4960          * We only support preallocation for extent-based files only
4961          */
4962         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4963                 ret = -EOPNOTSUPP;
4964                 goto out;
4965         }
4966
4967         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4968              offset + len > i_size_read(inode)) {
4969                 new_size = offset + len;
4970                 ret = inode_newsize_ok(inode, new_size);
4971                 if (ret)
4972                         goto out;
4973         }
4974
4975         /* Wait all existing dio workers, newcomers will block on i_mutex */
4976         ext4_inode_block_unlocked_dio(inode);
4977         inode_dio_wait(inode);
4978
4979         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
4980         ext4_inode_resume_unlocked_dio(inode);
4981         if (ret)
4982                 goto out;
4983
4984         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4985                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4986                                                 EXT4_I(inode)->i_sync_tid);
4987         }
4988 out:
4989         inode_unlock(inode);
4990         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4991         return ret;
4992 }
4993
4994 /*
4995  * This function convert a range of blocks to written extents
4996  * The caller of this function will pass the start offset and the size.
4997  * all unwritten extents within this range will be converted to
4998  * written extents.
4999  *
5000  * This function is called from the direct IO end io call back
5001  * function, to convert the fallocated extents after IO is completed.
5002  * Returns 0 on success.
5003  */
5004 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
5005                                    loff_t offset, ssize_t len)
5006 {
5007         unsigned int max_blocks;
5008         int ret = 0;
5009         int ret2 = 0;
5010         struct ext4_map_blocks map;
5011         unsigned int credits, blkbits = inode->i_blkbits;
5012
5013         map.m_lblk = offset >> blkbits;
5014         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
5015
5016         /*
5017          * This is somewhat ugly but the idea is clear: When transaction is
5018          * reserved, everything goes into it. Otherwise we rather start several
5019          * smaller transactions for conversion of each extent separately.
5020          */
5021         if (handle) {
5022                 handle = ext4_journal_start_reserved(handle,
5023                                                      EXT4_HT_EXT_CONVERT);
5024                 if (IS_ERR(handle))
5025                         return PTR_ERR(handle);
5026                 credits = 0;
5027         } else {
5028                 /*
5029                  * credits to insert 1 extent into extent tree
5030                  */
5031                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5032         }
5033         while (ret >= 0 && ret < max_blocks) {
5034                 map.m_lblk += ret;
5035                 map.m_len = (max_blocks -= ret);
5036                 if (credits) {
5037                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5038                                                     credits);
5039                         if (IS_ERR(handle)) {
5040                                 ret = PTR_ERR(handle);
5041                                 break;
5042                         }
5043                 }
5044                 ret = ext4_map_blocks(handle, inode, &map,
5045                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5046                 if (ret <= 0)
5047                         ext4_warning(inode->i_sb,
5048                                      "inode #%lu: block %u: len %u: "
5049                                      "ext4_ext_map_blocks returned %d",
5050                                      inode->i_ino, map.m_lblk,
5051                                      map.m_len, ret);
5052                 ext4_mark_inode_dirty(handle, inode);
5053                 if (credits)
5054                         ret2 = ext4_journal_stop(handle);
5055                 if (ret <= 0 || ret2)
5056                         break;
5057         }
5058         if (!credits)
5059                 ret2 = ext4_journal_stop(handle);
5060         return ret > 0 ? ret2 : ret;
5061 }
5062
5063 /*
5064  * If newes is not existing extent (newes->ec_pblk equals zero) find
5065  * delayed extent at start of newes and update newes accordingly and
5066  * return start of the next delayed extent.
5067  *
5068  * If newes is existing extent (newes->ec_pblk is not equal zero)
5069  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5070  * extent found. Leave newes unmodified.
5071  */
5072 static int ext4_find_delayed_extent(struct inode *inode,
5073                                     struct extent_status *newes)
5074 {
5075         struct extent_status es;
5076         ext4_lblk_t block, next_del;
5077
5078         if (newes->es_pblk == 0) {
5079                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5080                                 newes->es_lblk + newes->es_len - 1, &es);
5081
5082                 /*
5083                  * No extent in extent-tree contains block @newes->es_pblk,
5084                  * then the block may stay in 1)a hole or 2)delayed-extent.
5085                  */
5086                 if (es.es_len == 0)
5087                         /* A hole found. */
5088                         return 0;
5089
5090                 if (es.es_lblk > newes->es_lblk) {
5091                         /* A hole found. */
5092                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5093                                             newes->es_len);
5094                         return 0;
5095                 }
5096
5097                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5098         }
5099
5100         block = newes->es_lblk + newes->es_len;
5101         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5102         if (es.es_len == 0)
5103                 next_del = EXT_MAX_BLOCKS;
5104         else
5105                 next_del = es.es_lblk;
5106
5107         return next_del;
5108 }
5109 /* fiemap flags we can handle specified here */
5110 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5111
5112 static int ext4_xattr_fiemap(struct inode *inode,
5113                                 struct fiemap_extent_info *fieinfo)
5114 {
5115         __u64 physical = 0;
5116         __u64 length;
5117         __u32 flags = FIEMAP_EXTENT_LAST;
5118         int blockbits = inode->i_sb->s_blocksize_bits;
5119         int error = 0;
5120
5121         /* in-inode? */
5122         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5123                 struct ext4_iloc iloc;
5124                 int offset;     /* offset of xattr in inode */
5125
5126                 error = ext4_get_inode_loc(inode, &iloc);
5127                 if (error)
5128                         return error;
5129                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5130                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5131                                 EXT4_I(inode)->i_extra_isize;
5132                 physical += offset;
5133                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5134                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5135                 brelse(iloc.bh);
5136         } else { /* external block */
5137                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5138                 length = inode->i_sb->s_blocksize;
5139         }
5140
5141         if (physical)
5142                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5143                                                 length, flags);
5144         return (error < 0 ? error : 0);
5145 }
5146
5147 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5148                 __u64 start, __u64 len)
5149 {
5150         ext4_lblk_t start_blk;
5151         int error = 0;
5152
5153         if (ext4_has_inline_data(inode)) {
5154                 int has_inline = 1;
5155
5156                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5157                                                 start, len);
5158
5159                 if (has_inline)
5160                         return error;
5161         }
5162
5163         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5164                 error = ext4_ext_precache(inode);
5165                 if (error)
5166                         return error;
5167         }
5168
5169         /* fallback to generic here if not in extents fmt */
5170         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5171                 return generic_block_fiemap(inode, fieinfo, start, len,
5172                         ext4_get_block);
5173
5174         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5175                 return -EBADR;
5176
5177         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5178                 error = ext4_xattr_fiemap(inode, fieinfo);
5179         } else {
5180                 ext4_lblk_t len_blks;
5181                 __u64 last_blk;
5182
5183                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5184                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5185                 if (last_blk >= EXT_MAX_BLOCKS)
5186                         last_blk = EXT_MAX_BLOCKS-1;
5187                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5188
5189                 /*
5190                  * Walk the extent tree gathering extent information
5191                  * and pushing extents back to the user.
5192                  */
5193                 error = ext4_fill_fiemap_extents(inode, start_blk,
5194                                                  len_blks, fieinfo);
5195         }
5196         return error;
5197 }
5198
5199 /*
5200  * ext4_access_path:
5201  * Function to access the path buffer for marking it dirty.
5202  * It also checks if there are sufficient credits left in the journal handle
5203  * to update path.
5204  */
5205 static int
5206 ext4_access_path(handle_t *handle, struct inode *inode,
5207                 struct ext4_ext_path *path)
5208 {
5209         int credits, err;
5210
5211         if (!ext4_handle_valid(handle))
5212                 return 0;
5213
5214         /*
5215          * Check if need to extend journal credits
5216          * 3 for leaf, sb, and inode plus 2 (bmap and group
5217          * descriptor) for each block group; assume two block
5218          * groups
5219          */
5220         if (handle->h_buffer_credits < 7) {
5221                 credits = ext4_writepage_trans_blocks(inode);
5222                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5223                 /* EAGAIN is success */
5224                 if (err && err != -EAGAIN)
5225                         return err;
5226         }
5227
5228         err = ext4_ext_get_access(handle, inode, path);
5229         return err;
5230 }
5231
5232 /*
5233  * ext4_ext_shift_path_extents:
5234  * Shift the extents of a path structure lying between path[depth].p_ext
5235  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5236  * if it is right shift or left shift operation.
5237  */
5238 static int
5239 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5240                             struct inode *inode, handle_t *handle,
5241                             enum SHIFT_DIRECTION SHIFT)
5242 {
5243         int depth, err = 0;
5244         struct ext4_extent *ex_start, *ex_last;
5245         bool update = 0;
5246         depth = path->p_depth;
5247
5248         while (depth >= 0) {
5249                 if (depth == path->p_depth) {
5250                         ex_start = path[depth].p_ext;
5251                         if (!ex_start)
5252                                 return -EFSCORRUPTED;
5253
5254                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5255
5256                         err = ext4_access_path(handle, inode, path + depth);
5257                         if (err)
5258                                 goto out;
5259
5260                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5261                                 update = 1;
5262
5263                         while (ex_start <= ex_last) {
5264                                 if (SHIFT == SHIFT_LEFT) {
5265                                         le32_add_cpu(&ex_start->ee_block,
5266                                                 -shift);
5267                                         /* Try to merge to the left. */
5268                                         if ((ex_start >
5269                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5270                                             &&
5271                                             ext4_ext_try_to_merge_right(inode,
5272                                             path, ex_start - 1))
5273                                                 ex_last--;
5274                                         else
5275                                                 ex_start++;
5276                                 } else {
5277                                         le32_add_cpu(&ex_last->ee_block, shift);
5278                                         ext4_ext_try_to_merge_right(inode, path,
5279                                                 ex_last);
5280                                         ex_last--;
5281                                 }
5282                         }
5283                         err = ext4_ext_dirty(handle, inode, path + depth);
5284                         if (err)
5285                                 goto out;
5286
5287                         if (--depth < 0 || !update)
5288                                 break;
5289                 }
5290
5291                 /* Update index too */
5292                 err = ext4_access_path(handle, inode, path + depth);
5293                 if (err)
5294                         goto out;
5295
5296                 if (SHIFT == SHIFT_LEFT)
5297                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5298                 else
5299                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5300                 err = ext4_ext_dirty(handle, inode, path + depth);
5301                 if (err)
5302                         goto out;
5303
5304                 /* we are done if current index is not a starting index */
5305                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5306                         break;
5307
5308                 depth--;
5309         }
5310
5311 out:
5312         return err;
5313 }
5314
5315 /*
5316  * ext4_ext_shift_extents:
5317  * All the extents which lies in the range from @start to the last allocated
5318  * block for the @inode are shifted either towards left or right (depending
5319  * upon @SHIFT) by @shift blocks.
5320  * On success, 0 is returned, error otherwise.
5321  */
5322 static int
5323 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5324                        ext4_lblk_t start, ext4_lblk_t shift,
5325                        enum SHIFT_DIRECTION SHIFT)
5326 {
5327         struct ext4_ext_path *path;
5328         int ret = 0, depth;
5329         struct ext4_extent *extent;
5330         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5331
5332         /* Let path point to the last extent */
5333         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5334                                 EXT4_EX_NOCACHE);
5335         if (IS_ERR(path))
5336                 return PTR_ERR(path);
5337
5338         depth = path->p_depth;
5339         extent = path[depth].p_ext;
5340         if (!extent)
5341                 goto out;
5342
5343         stop = le32_to_cpu(extent->ee_block);
5344
5345        /*
5346          * In case of left shift, Don't start shifting extents until we make
5347          * sure the hole is big enough to accommodate the shift.
5348         */
5349         if (SHIFT == SHIFT_LEFT) {
5350                 path = ext4_find_extent(inode, start - 1, &path,
5351                                         EXT4_EX_NOCACHE);
5352                 if (IS_ERR(path))
5353                         return PTR_ERR(path);
5354                 depth = path->p_depth;
5355                 extent =  path[depth].p_ext;
5356                 if (extent) {
5357                         ex_start = le32_to_cpu(extent->ee_block);
5358                         ex_end = le32_to_cpu(extent->ee_block) +
5359                                 ext4_ext_get_actual_len(extent);
5360                 } else {
5361                         ex_start = 0;
5362                         ex_end = 0;
5363                 }
5364
5365                 if ((start == ex_start && shift > ex_start) ||
5366                     (shift > start - ex_end)) {
5367                         ext4_ext_drop_refs(path);
5368                         kfree(path);
5369                         return -EINVAL;
5370                 }
5371         }
5372
5373         /*
5374          * In case of left shift, iterator points to start and it is increased
5375          * till we reach stop. In case of right shift, iterator points to stop
5376          * and it is decreased till we reach start.
5377          */
5378         if (SHIFT == SHIFT_LEFT)
5379                 iterator = &start;
5380         else
5381                 iterator = &stop;
5382
5383         /*
5384          * Its safe to start updating extents.  Start and stop are unsigned, so
5385          * in case of right shift if extent with 0 block is reached, iterator
5386          * becomes NULL to indicate the end of the loop.
5387          */
5388         while (iterator && start <= stop) {
5389                 path = ext4_find_extent(inode, *iterator, &path,
5390                                         EXT4_EX_NOCACHE);
5391                 if (IS_ERR(path))
5392                         return PTR_ERR(path);
5393                 depth = path->p_depth;
5394                 extent = path[depth].p_ext;
5395                 if (!extent) {
5396                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5397                                          (unsigned long) *iterator);
5398                         return -EFSCORRUPTED;
5399                 }
5400                 if (SHIFT == SHIFT_LEFT && *iterator >
5401                     le32_to_cpu(extent->ee_block)) {
5402                         /* Hole, move to the next extent */
5403                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5404                                 path[depth].p_ext++;
5405                         } else {
5406                                 *iterator = ext4_ext_next_allocated_block(path);
5407                                 continue;
5408                         }
5409                 }
5410
5411                 if (SHIFT == SHIFT_LEFT) {
5412                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5413                         *iterator = le32_to_cpu(extent->ee_block) +
5414                                         ext4_ext_get_actual_len(extent);
5415                 } else {
5416                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5417                         if (le32_to_cpu(extent->ee_block) > 0)
5418                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5419                         else
5420                                 /* Beginning is reached, end of the loop */
5421                                 iterator = NULL;
5422                         /* Update path extent in case we need to stop */
5423                         while (le32_to_cpu(extent->ee_block) < start)
5424                                 extent++;
5425                         path[depth].p_ext = extent;
5426                 }
5427                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5428                                 handle, SHIFT);
5429                 if (ret)
5430                         break;
5431         }
5432 out:
5433         ext4_ext_drop_refs(path);
5434         kfree(path);
5435         return ret;
5436 }
5437
5438 /*
5439  * ext4_collapse_range:
5440  * This implements the fallocate's collapse range functionality for ext4
5441  * Returns: 0 and non-zero on error.
5442  */
5443 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5444 {
5445         struct super_block *sb = inode->i_sb;
5446         ext4_lblk_t punch_start, punch_stop;
5447         handle_t *handle;
5448         unsigned int credits;
5449         loff_t new_size, ioffset;
5450         int ret;
5451
5452         /*
5453          * We need to test this early because xfstests assumes that a
5454          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5455          * system does not support collapse range.
5456          */
5457         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5458                 return -EOPNOTSUPP;
5459
5460         /* Collapse range works only on fs block size aligned offsets. */
5461         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5462             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5463                 return -EINVAL;
5464
5465         if (!S_ISREG(inode->i_mode))
5466                 return -EINVAL;
5467
5468         trace_ext4_collapse_range(inode, offset, len);
5469
5470         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5471         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5472
5473         /* Call ext4_force_commit to flush all data in case of data=journal. */
5474         if (ext4_should_journal_data(inode)) {
5475                 ret = ext4_force_commit(inode->i_sb);
5476                 if (ret)
5477                         return ret;
5478         }
5479
5480         inode_lock(inode);
5481         /*
5482          * There is no need to overlap collapse range with EOF, in which case
5483          * it is effectively a truncate operation
5484          */
5485         if (offset + len >= i_size_read(inode)) {
5486                 ret = -EINVAL;
5487                 goto out_mutex;
5488         }
5489
5490         /* Currently just for extent based files */
5491         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5492                 ret = -EOPNOTSUPP;
5493                 goto out_mutex;
5494         }
5495
5496         /* Wait for existing dio to complete */
5497         ext4_inode_block_unlocked_dio(inode);
5498         inode_dio_wait(inode);
5499
5500         /*
5501          * Prevent page faults from reinstantiating pages we have released from
5502          * page cache.
5503          */
5504         down_write(&EXT4_I(inode)->i_mmap_sem);
5505         /*
5506          * Need to round down offset to be aligned with page size boundary
5507          * for page size > block size.
5508          */
5509         ioffset = round_down(offset, PAGE_SIZE);
5510         /*
5511          * Write tail of the last page before removed range since it will get
5512          * removed from the page cache below.
5513          */
5514         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5515         if (ret)
5516                 goto out_mmap;
5517         /*
5518          * Write data that will be shifted to preserve them when discarding
5519          * page cache below. We are also protected from pages becoming dirty
5520          * by i_mmap_sem.
5521          */
5522         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5523                                            LLONG_MAX);
5524         if (ret)
5525                 goto out_mmap;
5526         truncate_pagecache(inode, ioffset);
5527
5528         credits = ext4_writepage_trans_blocks(inode);
5529         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5530         if (IS_ERR(handle)) {
5531                 ret = PTR_ERR(handle);
5532                 goto out_mmap;
5533         }
5534
5535         down_write(&EXT4_I(inode)->i_data_sem);
5536         ext4_discard_preallocations(inode);
5537
5538         ret = ext4_es_remove_extent(inode, punch_start,
5539                                     EXT_MAX_BLOCKS - punch_start);
5540         if (ret) {
5541                 up_write(&EXT4_I(inode)->i_data_sem);
5542                 goto out_stop;
5543         }
5544
5545         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5546         if (ret) {
5547                 up_write(&EXT4_I(inode)->i_data_sem);
5548                 goto out_stop;
5549         }
5550         ext4_discard_preallocations(inode);
5551
5552         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5553                                      punch_stop - punch_start, SHIFT_LEFT);
5554         if (ret) {
5555                 up_write(&EXT4_I(inode)->i_data_sem);
5556                 goto out_stop;
5557         }
5558
5559         new_size = i_size_read(inode) - len;
5560         i_size_write(inode, new_size);
5561         EXT4_I(inode)->i_disksize = new_size;
5562
5563         up_write(&EXT4_I(inode)->i_data_sem);
5564         if (IS_SYNC(inode))
5565                 ext4_handle_sync(handle);
5566         inode->i_mtime = inode->i_ctime = current_time(inode);
5567         ext4_mark_inode_dirty(handle, inode);
5568         ext4_update_inode_fsync_trans(handle, inode, 1);
5569
5570 out_stop:
5571         ext4_journal_stop(handle);
5572 out_mmap:
5573         up_write(&EXT4_I(inode)->i_mmap_sem);
5574         ext4_inode_resume_unlocked_dio(inode);
5575 out_mutex:
5576         inode_unlock(inode);
5577         return ret;
5578 }
5579
5580 /*
5581  * ext4_insert_range:
5582  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5583  * The data blocks starting from @offset to the EOF are shifted by @len
5584  * towards right to create a hole in the @inode. Inode size is increased
5585  * by len bytes.
5586  * Returns 0 on success, error otherwise.
5587  */
5588 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5589 {
5590         struct super_block *sb = inode->i_sb;
5591         handle_t *handle;
5592         struct ext4_ext_path *path;
5593         struct ext4_extent *extent;
5594         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5595         unsigned int credits, ee_len;
5596         int ret = 0, depth, split_flag = 0;
5597         loff_t ioffset;
5598
5599         /*
5600          * We need to test this early because xfstests assumes that an
5601          * insert range of (0, 1) will return EOPNOTSUPP if the file
5602          * system does not support insert range.
5603          */
5604         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5605                 return -EOPNOTSUPP;
5606
5607         /* Insert range works only on fs block size aligned offsets. */
5608         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5609                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5610                 return -EINVAL;
5611
5612         if (!S_ISREG(inode->i_mode))
5613                 return -EOPNOTSUPP;
5614
5615         trace_ext4_insert_range(inode, offset, len);
5616
5617         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5618         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5619
5620         /* Call ext4_force_commit to flush all data in case of data=journal */
5621         if (ext4_should_journal_data(inode)) {
5622                 ret = ext4_force_commit(inode->i_sb);
5623                 if (ret)
5624                         return ret;
5625         }
5626
5627         inode_lock(inode);
5628         /* Currently just for extent based files */
5629         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5630                 ret = -EOPNOTSUPP;
5631                 goto out_mutex;
5632         }
5633
5634         /* Check for wrap through zero */
5635         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5636                 ret = -EFBIG;
5637                 goto out_mutex;
5638         }
5639
5640         /* Offset should be less than i_size */
5641         if (offset >= i_size_read(inode)) {
5642                 ret = -EINVAL;
5643                 goto out_mutex;
5644         }
5645
5646         /* Wait for existing dio to complete */
5647         ext4_inode_block_unlocked_dio(inode);
5648         inode_dio_wait(inode);
5649
5650         /*
5651          * Prevent page faults from reinstantiating pages we have released from
5652          * page cache.
5653          */
5654         down_write(&EXT4_I(inode)->i_mmap_sem);
5655         /*
5656          * Need to round down to align start offset to page size boundary
5657          * for page size > block size.
5658          */
5659         ioffset = round_down(offset, PAGE_SIZE);
5660         /* Write out all dirty pages */
5661         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5662                         LLONG_MAX);
5663         if (ret)
5664                 goto out_mmap;
5665         truncate_pagecache(inode, ioffset);
5666
5667         credits = ext4_writepage_trans_blocks(inode);
5668         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5669         if (IS_ERR(handle)) {
5670                 ret = PTR_ERR(handle);
5671                 goto out_mmap;
5672         }
5673
5674         /* Expand file to avoid data loss if there is error while shifting */
5675         inode->i_size += len;
5676         EXT4_I(inode)->i_disksize += len;
5677         inode->i_mtime = inode->i_ctime = current_time(inode);
5678         ret = ext4_mark_inode_dirty(handle, inode);
5679         if (ret)
5680                 goto out_stop;
5681
5682         down_write(&EXT4_I(inode)->i_data_sem);
5683         ext4_discard_preallocations(inode);
5684
5685         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5686         if (IS_ERR(path)) {
5687                 up_write(&EXT4_I(inode)->i_data_sem);
5688                 goto out_stop;
5689         }
5690
5691         depth = ext_depth(inode);
5692         extent = path[depth].p_ext;
5693         if (extent) {
5694                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5695                 ee_len = ext4_ext_get_actual_len(extent);
5696
5697                 /*
5698                  * If offset_lblk is not the starting block of extent, split
5699                  * the extent @offset_lblk
5700                  */
5701                 if ((offset_lblk > ee_start_lblk) &&
5702                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5703                         if (ext4_ext_is_unwritten(extent))
5704                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5705                                         EXT4_EXT_MARK_UNWRIT2;
5706                         ret = ext4_split_extent_at(handle, inode, &path,
5707                                         offset_lblk, split_flag,
5708                                         EXT4_EX_NOCACHE |
5709                                         EXT4_GET_BLOCKS_PRE_IO |
5710                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5711                 }
5712
5713                 ext4_ext_drop_refs(path);
5714                 kfree(path);
5715                 if (ret < 0) {
5716                         up_write(&EXT4_I(inode)->i_data_sem);
5717                         goto out_stop;
5718                 }
5719         } else {
5720                 ext4_ext_drop_refs(path);
5721                 kfree(path);
5722         }
5723
5724         ret = ext4_es_remove_extent(inode, offset_lblk,
5725                         EXT_MAX_BLOCKS - offset_lblk);
5726         if (ret) {
5727                 up_write(&EXT4_I(inode)->i_data_sem);
5728                 goto out_stop;
5729         }
5730
5731         /*
5732          * if offset_lblk lies in a hole which is at start of file, use
5733          * ee_start_lblk to shift extents
5734          */
5735         ret = ext4_ext_shift_extents(inode, handle,
5736                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5737                 len_lblk, SHIFT_RIGHT);
5738
5739         up_write(&EXT4_I(inode)->i_data_sem);
5740         if (IS_SYNC(inode))
5741                 ext4_handle_sync(handle);
5742         if (ret >= 0)
5743                 ext4_update_inode_fsync_trans(handle, inode, 1);
5744
5745 out_stop:
5746         ext4_journal_stop(handle);
5747 out_mmap:
5748         up_write(&EXT4_I(inode)->i_mmap_sem);
5749         ext4_inode_resume_unlocked_dio(inode);
5750 out_mutex:
5751         inode_unlock(inode);
5752         return ret;
5753 }
5754
5755 /**
5756  * ext4_swap_extents - Swap extents between two inodes
5757  *
5758  * @inode1:     First inode
5759  * @inode2:     Second inode
5760  * @lblk1:      Start block for first inode
5761  * @lblk2:      Start block for second inode
5762  * @count:      Number of blocks to swap
5763  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5764  * @erp:        Pointer to save error value
5765  *
5766  * This helper routine does exactly what is promise "swap extents". All other
5767  * stuff such as page-cache locking consistency, bh mapping consistency or
5768  * extent's data copying must be performed by caller.
5769  * Locking:
5770  *              i_mutex is held for both inodes
5771  *              i_data_sem is locked for write for both inodes
5772  * Assumptions:
5773  *              All pages from requested range are locked for both inodes
5774  */
5775 int
5776 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5777                      struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5778                   ext4_lblk_t count, int unwritten, int *erp)
5779 {
5780         struct ext4_ext_path *path1 = NULL;
5781         struct ext4_ext_path *path2 = NULL;
5782         int replaced_count = 0;
5783
5784         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5785         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5786         BUG_ON(!inode_is_locked(inode1));
5787         BUG_ON(!inode_is_locked(inode2));
5788
5789         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5790         if (unlikely(*erp))
5791                 return 0;
5792         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5793         if (unlikely(*erp))
5794                 return 0;
5795
5796         while (count) {
5797                 struct ext4_extent *ex1, *ex2, tmp_ex;
5798                 ext4_lblk_t e1_blk, e2_blk;
5799                 int e1_len, e2_len, len;
5800                 int split = 0;
5801
5802                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5803                 if (IS_ERR(path1)) {
5804                         *erp = PTR_ERR(path1);
5805                         path1 = NULL;
5806                 finish:
5807                         count = 0;
5808                         goto repeat;
5809                 }
5810                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5811                 if (IS_ERR(path2)) {
5812                         *erp = PTR_ERR(path2);
5813                         path2 = NULL;
5814                         goto finish;
5815                 }
5816                 ex1 = path1[path1->p_depth].p_ext;
5817                 ex2 = path2[path2->p_depth].p_ext;
5818                 /* Do we have somthing to swap ? */
5819                 if (unlikely(!ex2 || !ex1))
5820                         goto finish;
5821
5822                 e1_blk = le32_to_cpu(ex1->ee_block);
5823                 e2_blk = le32_to_cpu(ex2->ee_block);
5824                 e1_len = ext4_ext_get_actual_len(ex1);
5825                 e2_len = ext4_ext_get_actual_len(ex2);
5826
5827                 /* Hole handling */
5828                 if (!in_range(lblk1, e1_blk, e1_len) ||
5829                     !in_range(lblk2, e2_blk, e2_len)) {
5830                         ext4_lblk_t next1, next2;
5831
5832                         /* if hole after extent, then go to next extent */
5833                         next1 = ext4_ext_next_allocated_block(path1);
5834                         next2 = ext4_ext_next_allocated_block(path2);
5835                         /* If hole before extent, then shift to that extent */
5836                         if (e1_blk > lblk1)
5837                                 next1 = e1_blk;
5838                         if (e2_blk > lblk2)
5839                                 next2 = e2_blk;
5840                         /* Do we have something to swap */
5841                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5842                                 goto finish;
5843                         /* Move to the rightest boundary */
5844                         len = next1 - lblk1;
5845                         if (len < next2 - lblk2)
5846                                 len = next2 - lblk2;
5847                         if (len > count)
5848                                 len = count;
5849                         lblk1 += len;
5850                         lblk2 += len;
5851                         count -= len;
5852                         goto repeat;
5853                 }
5854
5855                 /* Prepare left boundary */
5856                 if (e1_blk < lblk1) {
5857                         split = 1;
5858                         *erp = ext4_force_split_extent_at(handle, inode1,
5859                                                 &path1, lblk1, 0);
5860                         if (unlikely(*erp))
5861                                 goto finish;
5862                 }
5863                 if (e2_blk < lblk2) {
5864                         split = 1;
5865                         *erp = ext4_force_split_extent_at(handle, inode2,
5866                                                 &path2,  lblk2, 0);
5867                         if (unlikely(*erp))
5868                                 goto finish;
5869                 }
5870                 /* ext4_split_extent_at() may result in leaf extent split,
5871                  * path must to be revalidated. */
5872                 if (split)
5873                         goto repeat;
5874
5875                 /* Prepare right boundary */
5876                 len = count;
5877                 if (len > e1_blk + e1_len - lblk1)
5878                         len = e1_blk + e1_len - lblk1;
5879                 if (len > e2_blk + e2_len - lblk2)
5880                         len = e2_blk + e2_len - lblk2;
5881
5882                 if (len != e1_len) {
5883                         split = 1;
5884                         *erp = ext4_force_split_extent_at(handle, inode1,
5885                                                 &path1, lblk1 + len, 0);
5886                         if (unlikely(*erp))
5887                                 goto finish;
5888                 }
5889                 if (len != e2_len) {
5890                         split = 1;
5891                         *erp = ext4_force_split_extent_at(handle, inode2,
5892                                                 &path2, lblk2 + len, 0);
5893                         if (*erp)
5894                                 goto finish;
5895                 }
5896                 /* ext4_split_extent_at() may result in leaf extent split,
5897                  * path must to be revalidated. */
5898                 if (split)
5899                         goto repeat;
5900
5901                 BUG_ON(e2_len != e1_len);
5902                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5903                 if (unlikely(*erp))
5904                         goto finish;
5905                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5906                 if (unlikely(*erp))
5907                         goto finish;
5908
5909                 /* Both extents are fully inside boundaries. Swap it now */
5910                 tmp_ex = *ex1;
5911                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5912                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5913                 ex1->ee_len = cpu_to_le16(e2_len);
5914                 ex2->ee_len = cpu_to_le16(e1_len);
5915                 if (unwritten)
5916                         ext4_ext_mark_unwritten(ex2);
5917                 if (ext4_ext_is_unwritten(&tmp_ex))
5918                         ext4_ext_mark_unwritten(ex1);
5919
5920                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5921                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5922                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5923                                       path2->p_depth);
5924                 if (unlikely(*erp))
5925                         goto finish;
5926                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5927                                       path1->p_depth);
5928                 /*
5929                  * Looks scarry ah..? second inode already points to new blocks,
5930                  * and it was successfully dirtied. But luckily error may happen
5931                  * only due to journal error, so full transaction will be
5932                  * aborted anyway.
5933                  */
5934                 if (unlikely(*erp))
5935                         goto finish;
5936                 lblk1 += len;
5937                 lblk2 += len;
5938                 replaced_count += len;
5939                 count -= len;
5940
5941         repeat:
5942                 ext4_ext_drop_refs(path1);
5943                 kfree(path1);
5944                 ext4_ext_drop_refs(path2);
5945                 kfree(path2);
5946                 path1 = path2 = NULL;
5947         }
5948         return replaced_count;
5949 }