fs/ext4/extents.c

   1 /*
   2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3  * Written by Alex Tomas <alex@clusterfs.com>
   4  *
   5  * Architecture independence:
   6  *   Copyright (c) 2005, Bull S.A.
   7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public Licens
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <asm/uaccess.h>
  41 #include <linux/fiemap.h>
  42 #include "ext4_jbd2.h"
  43 #include "ext4_extents.h"
  44 #include "xattr.h"
  45
  46 #include <trace/events/ext4.h>
  47
  48 /*
  49  * used by extent splitting.
  50  */
  51 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  52                                         due to ENOSPC */
  53 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  54 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  55
  56 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  57 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  58
  59 static __le32 ext4_extent_block_csum(struct inode *inode,
  60                                      struct ext4_extent_header *eh)
  61 {
  62         struct ext4_inode_info *ei = EXT4_I(inode);
  63         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  64         __u32 csum;
  65
  66         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  67                            EXT4_EXTENT_TAIL_OFFSET(eh));
  68         return cpu_to_le32(csum);
  69 }
  70
  71 static int ext4_extent_block_csum_verify(struct inode *inode,
  72                                          struct ext4_extent_header *eh)
  73 {
  74         struct ext4_extent_tail *et;
  75
  76         if (!ext4_has_metadata_csum(inode->i_sb))
  77                 return 1;
  78
  79         et = find_ext4_extent_tail(eh);
  80         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  81                 return 0;
  82         return 1;
  83 }
  84
  85 static void ext4_extent_block_csum_set(struct inode *inode,
  86                                        struct ext4_extent_header *eh)
  87 {
  88         struct ext4_extent_tail *et;
  89
  90         if (!ext4_has_metadata_csum(inode->i_sb))
  91                 return;
  92
  93         et = find_ext4_extent_tail(eh);
  94         et->et_checksum = ext4_extent_block_csum(inode, eh);
  95 }
  96
  97 static int ext4_split_extent(handle_t *handle,
  98                                 struct inode *inode,
  99                                 struct ext4_ext_path **ppath,
 100                                 struct ext4_map_blocks *map,
 101                                 int split_flag,
 102                                 int flags);
 103
 104 static int ext4_split_extent_at(handle_t *handle,
 105                              struct inode *inode,
 106                              struct ext4_ext_path **ppath,
 107                              ext4_lblk_t split,
 108                              int split_flag,
 109                              int flags);
 110
 111 static int ext4_find_delayed_extent(struct inode *inode,
 112                                     struct extent_status *newes);
 113
 114 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 115                                             struct inode *inode,
 116                                             int needed)
 117 {
 118         int err;
 119
 120         if (!ext4_handle_valid(handle))
 121                 return 0;
 122         if (handle->h_buffer_credits > needed)
 123                 return 0;
 124         err = ext4_journal_extend(handle, needed);
 125         if (err <= 0)
 126                 return err;
 127         err = ext4_truncate_restart_trans(handle, inode, needed);
 128         if (err == 0)
 129                 err = -EAGAIN;
 130
 131         return err;
 132 }
 133
 134 /*
 135  * could return:
 136  *  - EROFS
 137  *  - ENOMEM
 138  */
 139 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 140                                 struct ext4_ext_path *path)
 141 {
 142         if (path->p_bh) {
 143                 /* path points to block */
 144                 BUFFER_TRACE(path->p_bh, "get_write_access");
 145                 return ext4_journal_get_write_access(handle, path->p_bh);
 146         }
 147         /* path points to leaf/index in inode body */
 148         /* we use in-core data, no need to protect them */
 149         return 0;
 150 }
 151
 152 /*
 153  * could return:
 154  *  - EROFS
 155  *  - ENOMEM
 156  *  - EIO
 157  */
 158 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 159                      struct inode *inode, struct ext4_ext_path *path)
 160 {
 161         int err;
 162
 163         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 164         if (path->p_bh) {
 165                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 166                 /* path points to block */
 167                 err = __ext4_handle_dirty_metadata(where, line, handle,
 168                                                    inode, path->p_bh);
 169         } else {
 170                 /* path points to leaf/index in inode body */
 171                 err = ext4_mark_inode_dirty(handle, inode);
 172         }
 173         return err;
 174 }
 175
 176 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 177                               struct ext4_ext_path *path,
 178                               ext4_lblk_t block)
 179 {
 180         if (path) {
 181                 int depth = path->p_depth;
 182                 struct ext4_extent *ex;
 183
 184                 /*
 185                  * Try to predict block placement assuming that we are
 186                  * filling in a file which will eventually be
 187                  * non-sparse --- i.e., in the case of libbfd writing
 188                  * an ELF object sections out-of-order but in a way
 189                  * the eventually results in a contiguous object or
 190                  * executable file, or some database extending a table
 191                  * space file.  However, this is actually somewhat
 192                  * non-ideal if we are writing a sparse file such as
 193                  * qemu or KVM writing a raw image file that is going
 194                  * to stay fairly sparse, since it will end up
 195                  * fragmenting the file system's free space.  Maybe we
 196                  * should have some hueristics or some way to allow
 197                  * userspace to pass a hint to file system,
 198                  * especially if the latter case turns out to be
 199                  * common.
 200                  */
 201                 ex = path[depth].p_ext;
 202                 if (ex) {
 203                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 204                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 205
 206                         if (block > ext_block)
 207                                 return ext_pblk + (block - ext_block);
 208                         else
 209                                 return ext_pblk - (ext_block - block);
 210                 }
 211
 212                 /* it looks like index is empty;
 213                  * try to find starting block from index itself */
 214                 if (path[depth].p_bh)
 215                         return path[depth].p_bh->b_blocknr;
 216         }
 217
 218         /* OK. use inode's group */
 219         return ext4_inode_to_goal_block(inode);
 220 }
 221
 222 /*
 223  * Allocation for a meta data block
 224  */
 225 static ext4_fsblk_t
 226 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 227                         struct ext4_ext_path *path,
 228                         struct ext4_extent *ex, int *err, unsigned int flags)
 229 {
 230         ext4_fsblk_t goal, newblock;
 231
 232         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 233         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 234                                         NULL, err);
 235         return newblock;
 236 }
 237
 238 static inline int ext4_ext_space_block(struct inode *inode, int check)
 239 {
 240         int size;
 241
 242         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 243                         / sizeof(struct ext4_extent);
 244 #ifdef AGGRESSIVE_TEST
 245         if (!check && size > 6)
 246                 size = 6;
 247 #endif
 248         return size;
 249 }
 250
 251 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 252 {
 253         int size;
 254
 255         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 256                         / sizeof(struct ext4_extent_idx);
 257 #ifdef AGGRESSIVE_TEST
 258         if (!check && size > 5)
 259                 size = 5;
 260 #endif
 261         return size;
 262 }
 263
 264 static inline int ext4_ext_space_root(struct inode *inode, int check)
 265 {
 266         int size;
 267
 268         size = sizeof(EXT4_I(inode)->i_data);
 269         size -= sizeof(struct ext4_extent_header);
 270         size /= sizeof(struct ext4_extent);
 271 #ifdef AGGRESSIVE_TEST
 272         if (!check && size > 3)
 273                 size = 3;
 274 #endif
 275         return size;
 276 }
 277
 278 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 279 {
 280         int size;
 281
 282         size = sizeof(EXT4_I(inode)->i_data);
 283         size -= sizeof(struct ext4_extent_header);
 284         size /= sizeof(struct ext4_extent_idx);
 285 #ifdef AGGRESSIVE_TEST
 286         if (!check && size > 4)
 287                 size = 4;
 288 #endif
 289         return size;
 290 }
 291
 292 static inline int
 293 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 294                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 295                            int nofail)
 296 {
 297         struct ext4_ext_path *path = *ppath;
 298         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 299
 300         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 301                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 302                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 303                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 304 }
 305
 306 /*
 307  * Calculate the number of metadata blocks needed
 308  * to allocate @blocks
 309  * Worse case is one block per extent
 310  */
 311 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 312 {
 313         struct ext4_inode_info *ei = EXT4_I(inode);
 314         int idxs;
 315
 316         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 317                 / sizeof(struct ext4_extent_idx));
 318
 319         /*
 320          * If the new delayed allocation block is contiguous with the
 321          * previous da block, it can share index blocks with the
 322          * previous block, so we only need to allocate a new index
 323          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 324          * an additional index block, and at ldxs**3 blocks, yet
 325          * another index blocks.
 326          */
 327         if (ei->i_da_metadata_calc_len &&
 328             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 329                 int num = 0;
 330
 331                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 332                         num++;
 333                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 334                         num++;
 335                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 336                         num++;
 337                         ei->i_da_metadata_calc_len = 0;
 338                 } else
 339                         ei->i_da_metadata_calc_len++;
 340                 ei->i_da_metadata_calc_last_lblock++;
 341                 return num;
 342         }
 343
 344         /*
 345          * In the worst case we need a new set of index blocks at
 346          * every level of the inode's extent tree.
 347          */
 348         ei->i_da_metadata_calc_len = 1;
 349         ei->i_da_metadata_calc_last_lblock = lblock;
 350         return ext_depth(inode) + 1;
 351 }
 352
 353 static int
 354 ext4_ext_max_entries(struct inode *inode, int depth)
 355 {
 356         int max;
 357
 358         if (depth == ext_depth(inode)) {
 359                 if (depth == 0)
 360                         max = ext4_ext_space_root(inode, 1);
 361                 else
 362                         max = ext4_ext_space_root_idx(inode, 1);
 363         } else {
 364                 if (depth == 0)
 365                         max = ext4_ext_space_block(inode, 1);
 366                 else
 367                         max = ext4_ext_space_block_idx(inode, 1);
 368         }
 369
 370         return max;
 371 }
 372
 373 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 374 {
 375         ext4_fsblk_t block = ext4_ext_pblock(ext);
 376         int len = ext4_ext_get_actual_len(ext);
 377         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 378         ext4_lblk_t last = lblock + len - 1;
 379
 380         if (lblock > last)
 381                 return 0;
 382         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 383 }
 384
 385 static int ext4_valid_extent_idx(struct inode *inode,
 386                                 struct ext4_extent_idx *ext_idx)
 387 {
 388         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 389
 390         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 391 }
 392
 393 static int ext4_valid_extent_entries(struct inode *inode,
 394                                 struct ext4_extent_header *eh,
 395                                 int depth)
 396 {
 397         unsigned short entries;
 398         if (eh->eh_entries == 0)
 399                 return 1;
 400
 401         entries = le16_to_cpu(eh->eh_entries);
 402
 403         if (depth == 0) {
 404                 /* leaf entries */
 405                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 406                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 407                 ext4_fsblk_t pblock = 0;
 408                 ext4_lblk_t lblock = 0;
 409                 ext4_lblk_t prev = 0;
 410                 int len = 0;
 411                 while (entries) {
 412                         if (!ext4_valid_extent(inode, ext))
 413                                 return 0;
 414
 415                         /* Check for overlapping extents */
 416                         lblock = le32_to_cpu(ext->ee_block);
 417                         len = ext4_ext_get_actual_len(ext);
 418                         if ((lblock <= prev) && prev) {
 419                                 pblock = ext4_ext_pblock(ext);
 420                                 es->s_last_error_block = cpu_to_le64(pblock);
 421                                 return 0;
 422                         }
 423                         ext++;
 424                         entries--;
 425                         prev = lblock + len - 1;
 426                 }
 427         } else {
 428                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 429                 while (entries) {
 430                         if (!ext4_valid_extent_idx(inode, ext_idx))
 431                                 return 0;
 432                         ext_idx++;
 433                         entries--;
 434                 }
 435         }
 436         return 1;
 437 }
 438
 439 static int __ext4_ext_check(const char *function, unsigned int line,
 440                             struct inode *inode, struct ext4_extent_header *eh,
 441                             int depth, ext4_fsblk_t pblk)
 442 {
 443         const char *error_msg;
 444         int max = 0;
 445
 446         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 447                 error_msg = "invalid magic";
 448                 goto corrupted;
 449         }
 450         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 451                 error_msg = "unexpected eh_depth";
 452                 goto corrupted;
 453         }
 454         if (unlikely(eh->eh_max == 0)) {
 455                 error_msg = "invalid eh_max";
 456                 goto corrupted;
 457         }
 458         max = ext4_ext_max_entries(inode, depth);
 459         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 460                 error_msg = "too large eh_max";
 461                 goto corrupted;
 462         }
 463         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 464                 error_msg = "invalid eh_entries";
 465                 goto corrupted;
 466         }
 467         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 468                 error_msg = "invalid extent entries";
 469                 goto corrupted;
 470         }
 471         /* Verify checksum on non-root extent tree nodes */
 472         if (ext_depth(inode) != depth &&
 473             !ext4_extent_block_csum_verify(inode, eh)) {
 474                 error_msg = "extent tree corrupted";
 475                 goto corrupted;
 476         }
 477         return 0;
 478
 479 corrupted:
 480         ext4_error_inode(inode, function, line, 0,
 481                          "pblk %llu bad header/extent: %s - magic %x, "
 482                          "entries %u, max %u(%u), depth %u(%u)",
 483                          (unsigned long long) pblk, error_msg,
 484                          le16_to_cpu(eh->eh_magic),
 485                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 486                          max, le16_to_cpu(eh->eh_depth), depth);
 487         return -EIO;
 488 }
 489
 490 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 491         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 492
 493 int ext4_ext_check_inode(struct inode *inode)
 494 {
 495         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 496 }
 497
 498 static struct buffer_head *
 499 __read_extent_tree_block(const char *function, unsigned int line,
 500                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 501                          int flags)
 502 {
 503         struct buffer_head              *bh;
 504         int                             err;
 505
 506         bh = sb_getblk(inode->i_sb, pblk);
 507         if (unlikely(!bh))
 508                 return ERR_PTR(-ENOMEM);
 509
 510         if (!bh_uptodate_or_lock(bh)) {
 511                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 512                 err = bh_submit_read(bh);
 513                 if (err < 0)
 514                         goto errout;
 515         }
 516         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 517                 return bh;
 518         err = __ext4_ext_check(function, line, inode,
 519                                ext_block_hdr(bh), depth, pblk);
 520         if (err)
 521                 goto errout;
 522         set_buffer_verified(bh);
 523         /*
 524          * If this is a leaf block, cache all of its entries
 525          */
 526         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 527                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 528                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 529                 ext4_lblk_t prev = 0;
 530                 int i;
 531
 532                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 533                         unsigned int status = EXTENT_STATUS_WRITTEN;
 534                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 535                         int len = ext4_ext_get_actual_len(ex);
 536
 537                         if (prev && (prev != lblk))
 538                                 ext4_es_cache_extent(inode, prev,
 539                                                      lblk - prev, ~0,
 540                                                      EXTENT_STATUS_HOLE);
 541
 542                         if (ext4_ext_is_unwritten(ex))
 543                                 status = EXTENT_STATUS_UNWRITTEN;
 544                         ext4_es_cache_extent(inode, lblk, len,
 545                                              ext4_ext_pblock(ex), status);
 546                         prev = lblk + len;
 547                 }
 548         }
 549         return bh;
 550 errout:
 551         put_bh(bh);
 552         return ERR_PTR(err);
 553
 554 }
 555
 556 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 557         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 558                                  (depth), (flags))
 559
 560 /*
 561  * This function is called to cache a file's extent information in the
 562  * extent status tree
 563  */
 564 int ext4_ext_precache(struct inode *inode)
 565 {
 566         struct ext4_inode_info *ei = EXT4_I(inode);
 567         struct ext4_ext_path *path = NULL;
 568         struct buffer_head *bh;
 569         int i = 0, depth, ret = 0;
 570
 571         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 572                 return 0;       /* not an extent-mapped inode */
 573
 574         down_read(&ei->i_data_sem);
 575         depth = ext_depth(inode);
 576
 577         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 578                        GFP_NOFS);
 579         if (path == NULL) {
 580                 up_read(&ei->i_data_sem);
 581                 return -ENOMEM;
 582         }
 583
 584         /* Don't cache anything if there are no external extent blocks */
 585         if (depth == 0)
 586                 goto out;
 587         path[0].p_hdr = ext_inode_hdr(inode);
 588         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 589         if (ret)
 590                 goto out;
 591         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 592         while (i >= 0) {
 593                 /*
 594                  * If this is a leaf block or we've reached the end of
 595                  * the index block, go up
 596                  */
 597                 if ((i == depth) ||
 598                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 599                         brelse(path[i].p_bh);
 600                         path[i].p_bh = NULL;
 601                         i--;
 602                         continue;
 603                 }
 604                 bh = read_extent_tree_block(inode,
 605                                             ext4_idx_pblock(path[i].p_idx++),
 606                                             depth - i - 1,
 607                                             EXT4_EX_FORCE_CACHE);
 608                 if (IS_ERR(bh)) {
 609                         ret = PTR_ERR(bh);
 610                         break;
 611                 }
 612                 i++;
 613                 path[i].p_bh = bh;
 614                 path[i].p_hdr = ext_block_hdr(bh);
 615                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 616         }
 617         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 618 out:
 619         up_read(&ei->i_data_sem);
 620         ext4_ext_drop_refs(path);
 621         kfree(path);
 622         return ret;
 623 }
 624
 625 #ifdef EXT_DEBUG
 626 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 627 {
 628         int k, l = path->p_depth;
 629
 630         ext_debug("path:");
 631         for (k = 0; k <= l; k++, path++) {
 632                 if (path->p_idx) {
 633                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 634                             ext4_idx_pblock(path->p_idx));
 635                 } else if (path->p_ext) {
 636                         ext_debug("  %d:[%d]%d:%llu ",
 637                                   le32_to_cpu(path->p_ext->ee_block),
 638                                   ext4_ext_is_unwritten(path->p_ext),
 639                                   ext4_ext_get_actual_len(path->p_ext),
 640                                   ext4_ext_pblock(path->p_ext));
 641                 } else
 642                         ext_debug("  []");
 643         }
 644         ext_debug("\n");
 645 }
 646
 647 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 648 {
 649         int depth = ext_depth(inode);
 650         struct ext4_extent_header *eh;
 651         struct ext4_extent *ex;
 652         int i;
 653
 654         if (!path)
 655                 return;
 656
 657         eh = path[depth].p_hdr;
 658         ex = EXT_FIRST_EXTENT(eh);
 659
 660         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 661
 662         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 663                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 664                           ext4_ext_is_unwritten(ex),
 665                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 666         }
 667         ext_debug("\n");
 668 }
 669
 670 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 671                         ext4_fsblk_t newblock, int level)
 672 {
 673         int depth = ext_depth(inode);
 674         struct ext4_extent *ex;
 675
 676         if (depth != level) {
 677                 struct ext4_extent_idx *idx;
 678                 idx = path[level].p_idx;
 679                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 680                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 681                                         le32_to_cpu(idx->ei_block),
 682                                         ext4_idx_pblock(idx),
 683                                         newblock);
 684                         idx++;
 685                 }
 686
 687                 return;
 688         }
 689
 690         ex = path[depth].p_ext;
 691         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 692                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 693                                 le32_to_cpu(ex->ee_block),
 694                                 ext4_ext_pblock(ex),
 695                                 ext4_ext_is_unwritten(ex),
 696                                 ext4_ext_get_actual_len(ex),
 697                                 newblock);
 698                 ex++;
 699         }
 700 }
 701
 702 #else
 703 #define ext4_ext_show_path(inode, path)
 704 #define ext4_ext_show_leaf(inode, path)
 705 #define ext4_ext_show_move(inode, path, newblock, level)
 706 #endif
 707
 708 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 709 {
 710         int depth, i;
 711
 712         if (!path)
 713                 return;
 714         depth = path->p_depth;
 715         for (i = 0; i <= depth; i++, path++)
 716                 if (path->p_bh) {
 717                         brelse(path->p_bh);
 718                         path->p_bh = NULL;
 719                 }
 720 }
 721
 722 /*
 723  * ext4_ext_binsearch_idx:
 724  * binary search for the closest index of the given block
 725  * the header must be checked before calling this
 726  */
 727 static void
 728 ext4_ext_binsearch_idx(struct inode *inode,
 729                         struct ext4_ext_path *path, ext4_lblk_t block)
 730 {
 731         struct ext4_extent_header *eh = path->p_hdr;
 732         struct ext4_extent_idx *r, *l, *m;
 733
 734
 735         ext_debug("binsearch for %u(idx):  ", block);
 736
 737         l = EXT_FIRST_INDEX(eh) + 1;
 738         r = EXT_LAST_INDEX(eh);
 739         while (l <= r) {
 740                 m = l + (r - l) / 2;
 741                 if (block < le32_to_cpu(m->ei_block))
 742                         r = m - 1;
 743                 else
 744                         l = m + 1;
 745                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 746                                 m, le32_to_cpu(m->ei_block),
 747                                 r, le32_to_cpu(r->ei_block));
 748         }
 749
 750         path->p_idx = l - 1;
 751         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 752                   ext4_idx_pblock(path->p_idx));
 753
 754 #ifdef CHECK_BINSEARCH
 755         {
 756                 struct ext4_extent_idx *chix, *ix;
 757                 int k;
 758
 759                 chix = ix = EXT_FIRST_INDEX(eh);
 760                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 761                   if (k != 0 &&
 762                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 763                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 764                                        "first=0x%p\n", k,
 765                                        ix, EXT_FIRST_INDEX(eh));
 766                                 printk(KERN_DEBUG "%u <= %u\n",
 767                                        le32_to_cpu(ix->ei_block),
 768                                        le32_to_cpu(ix[-1].ei_block));
 769                         }
 770                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 771                                            <= le32_to_cpu(ix[-1].ei_block));
 772                         if (block < le32_to_cpu(ix->ei_block))
 773                                 break;
 774                         chix = ix;
 775                 }
 776                 BUG_ON(chix != path->p_idx);
 777         }
 778 #endif
 779
 780 }
 781
 782 /*
 783  * ext4_ext_binsearch:
 784  * binary search for closest extent of the given block
 785  * the header must be checked before calling this
 786  */
 787 static void
 788 ext4_ext_binsearch(struct inode *inode,
 789                 struct ext4_ext_path *path, ext4_lblk_t block)
 790 {
 791         struct ext4_extent_header *eh = path->p_hdr;
 792         struct ext4_extent *r, *l, *m;
 793
 794         if (eh->eh_entries == 0) {
 795                 /*
 796                  * this leaf is empty:
 797                  * we get such a leaf in split/add case
 798                  */
 799                 return;
 800         }
 801
 802         ext_debug("binsearch for %u:  ", block);
 803
 804         l = EXT_FIRST_EXTENT(eh) + 1;
 805         r = EXT_LAST_EXTENT(eh);
 806
 807         while (l <= r) {
 808                 m = l + (r - l) / 2;
 809                 if (block < le32_to_cpu(m->ee_block))
 810                         r = m - 1;
 811                 else
 812                         l = m + 1;
 813                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 814                                 m, le32_to_cpu(m->ee_block),
 815                                 r, le32_to_cpu(r->ee_block));
 816         }
 817
 818         path->p_ext = l - 1;
 819         ext_debug("  -> %d:%llu:[%d]%d ",
 820                         le32_to_cpu(path->p_ext->ee_block),
 821                         ext4_ext_pblock(path->p_ext),
 822                         ext4_ext_is_unwritten(path->p_ext),
 823                         ext4_ext_get_actual_len(path->p_ext));
 824
 825 #ifdef CHECK_BINSEARCH
 826         {
 827                 struct ext4_extent *chex, *ex;
 828                 int k;
 829
 830                 chex = ex = EXT_FIRST_EXTENT(eh);
 831                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 832                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 833                                           <= le32_to_cpu(ex[-1].ee_block));
 834                         if (block < le32_to_cpu(ex->ee_block))
 835                                 break;
 836                         chex = ex;
 837                 }
 838                 BUG_ON(chex != path->p_ext);
 839         }
 840 #endif
 841
 842 }
 843
 844 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 845 {
 846         struct ext4_extent_header *eh;
 847
 848         eh = ext_inode_hdr(inode);
 849         eh->eh_depth = 0;
 850         eh->eh_entries = 0;
 851         eh->eh_magic = EXT4_EXT_MAGIC;
 852         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 853         ext4_mark_inode_dirty(handle, inode);
 854         return 0;
 855 }
 856
 857 struct ext4_ext_path *
 858 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 859                  struct ext4_ext_path **orig_path, int flags)
 860 {
 861         struct ext4_extent_header *eh;
 862         struct buffer_head *bh;
 863         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 864         short int depth, i, ppos = 0;
 865         int ret;
 866
 867         eh = ext_inode_hdr(inode);
 868         depth = ext_depth(inode);
 869
 870         if (path) {
 871                 ext4_ext_drop_refs(path);
 872                 if (depth > path[0].p_maxdepth) {
 873                         kfree(path);
 874                         *orig_path = path = NULL;
 875                 }
 876         }
 877         if (!path) {
 878                 /* account possible depth increase */
 879                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 880                                 GFP_NOFS);
 881                 if (unlikely(!path))
 882                         return ERR_PTR(-ENOMEM);
 883                 path[0].p_maxdepth = depth + 1;
 884         }
 885         path[0].p_hdr = eh;
 886         path[0].p_bh = NULL;
 887
 888         i = depth;
 889         /* walk through the tree */
 890         while (i) {
 891                 ext_debug("depth %d: num %d, max %d\n",
 892                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 893
 894                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 895                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 896                 path[ppos].p_depth = i;
 897                 path[ppos].p_ext = NULL;
 898
 899                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 900                                             flags);
 901                 if (unlikely(IS_ERR(bh))) {
 902                         ret = PTR_ERR(bh);
 903                         goto err;
 904                 }
 905
 906                 eh = ext_block_hdr(bh);
 907                 ppos++;
 908                 if (unlikely(ppos > depth)) {
 909                         put_bh(bh);
 910                         EXT4_ERROR_INODE(inode,
 911                                          "ppos %d > depth %d", ppos, depth);
 912                         ret = -EIO;
 913                         goto err;
 914                 }
 915                 path[ppos].p_bh = bh;
 916                 path[ppos].p_hdr = eh;
 917         }
 918
 919         path[ppos].p_depth = i;
 920         path[ppos].p_ext = NULL;
 921         path[ppos].p_idx = NULL;
 922
 923         /* find extent */
 924         ext4_ext_binsearch(inode, path + ppos, block);
 925         /* if not an empty leaf */
 926         if (path[ppos].p_ext)
 927                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 928
 929         ext4_ext_show_path(inode, path);
 930
 931         return path;
 932
 933 err:
 934         ext4_ext_drop_refs(path);
 935         kfree(path);
 936         if (orig_path)
 937                 *orig_path = NULL;
 938         return ERR_PTR(ret);
 939 }
 940
 941 /*
 942  * ext4_ext_insert_index:
 943  * insert new index [@logical;@ptr] into the block at @curp;
 944  * check where to insert: before @curp or after @curp
 945  */
 946 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 947                                  struct ext4_ext_path *curp,
 948                                  int logical, ext4_fsblk_t ptr)
 949 {
 950         struct ext4_extent_idx *ix;
 951         int len, err;
 952
 953         err = ext4_ext_get_access(handle, inode, curp);
 954         if (err)
 955                 return err;
 956
 957         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 958                 EXT4_ERROR_INODE(inode,
 959                                  "logical %d == ei_block %d!",
 960                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 961                 return -EIO;
 962         }
 963
 964         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 965                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 966                 EXT4_ERROR_INODE(inode,
 967                                  "eh_entries %d >= eh_max %d!",
 968                                  le16_to_cpu(curp->p_hdr->eh_entries),
 969                                  le16_to_cpu(curp->p_hdr->eh_max));
 970                 return -EIO;
 971         }
 972
 973         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 974                 /* insert after */
 975                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 976                 ix = curp->p_idx + 1;
 977         } else {
 978                 /* insert before */
 979                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 980                 ix = curp->p_idx;
 981         }
 982
 983         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 984         BUG_ON(len < 0);
 985         if (len > 0) {
 986                 ext_debug("insert new index %d: "
 987                                 "move %d indices from 0x%p to 0x%p\n",
 988                                 logical, len, ix, ix + 1);
 989                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 990         }
 991
 992         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 993                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 994                 return -EIO;
 995         }
 996
 997         ix->ei_block = cpu_to_le32(logical);
 998         ext4_idx_store_pblock(ix, ptr);
 999         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1000
1001         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1002                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1003                 return -EIO;
1004         }
1005
1006         err = ext4_ext_dirty(handle, inode, curp);
1007         ext4_std_error(inode->i_sb, err);
1008
1009         return err;
1010 }
1011
1012 /*
1013  * ext4_ext_split:
1014  * inserts new subtree into the path, using free index entry
1015  * at depth @at:
1016  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1017  * - makes decision where to split
1018  * - moves remaining extents and index entries (right to the split point)
1019  *   into the newly allocated blocks
1020  * - initializes subtree
1021  */
1022 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1023                           unsigned int flags,
1024                           struct ext4_ext_path *path,
1025                           struct ext4_extent *newext, int at)
1026 {
1027         struct buffer_head *bh = NULL;
1028         int depth = ext_depth(inode);
1029         struct ext4_extent_header *neh;
1030         struct ext4_extent_idx *fidx;
1031         int i = at, k, m, a;
1032         ext4_fsblk_t newblock, oldblock;
1033         __le32 border;
1034         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1035         int err = 0;
1036
1037         /* make decision: where to split? */
1038         /* FIXME: now decision is simplest: at current extent */
1039
1040         /* if current leaf will be split, then we should use
1041          * border from split point */
1042         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1043                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1044                 return -EIO;
1045         }
1046         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1047                 border = path[depth].p_ext[1].ee_block;
1048                 ext_debug("leaf will be split."
1049                                 " next leaf starts at %d\n",
1050                                   le32_to_cpu(border));
1051         } else {
1052                 border = newext->ee_block;
1053                 ext_debug("leaf will be added."
1054                                 " next leaf starts at %d\n",
1055                                 le32_to_cpu(border));
1056         }
1057
1058         /*
1059          * If error occurs, then we break processing
1060          * and mark filesystem read-only. index won't
1061          * be inserted and tree will be in consistent
1062          * state. Next mount will repair buffers too.
1063          */
1064
1065         /*
1066          * Get array to track all allocated blocks.
1067          * We need this to handle errors and free blocks
1068          * upon them.
1069          */
1070         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1071         if (!ablocks)
1072                 return -ENOMEM;
1073
1074         /* allocate all needed blocks */
1075         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1076         for (a = 0; a < depth - at; a++) {
1077                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1078                                                    newext, &err, flags);
1079                 if (newblock == 0)
1080                         goto cleanup;
1081                 ablocks[a] = newblock;
1082         }
1083
1084         /* initialize new leaf */
1085         newblock = ablocks[--a];
1086         if (unlikely(newblock == 0)) {
1087                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1088                 err = -EIO;
1089                 goto cleanup;
1090         }
1091         bh = sb_getblk(inode->i_sb, newblock);
1092         if (unlikely(!bh)) {
1093                 err = -ENOMEM;
1094                 goto cleanup;
1095         }
1096         lock_buffer(bh);
1097
1098         err = ext4_journal_get_create_access(handle, bh);
1099         if (err)
1100                 goto cleanup;
1101
1102         neh = ext_block_hdr(bh);
1103         neh->eh_entries = 0;
1104         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1105         neh->eh_magic = EXT4_EXT_MAGIC;
1106         neh->eh_depth = 0;
1107
1108         /* move remainder of path[depth] to the new leaf */
1109         if (unlikely(path[depth].p_hdr->eh_entries !=
1110                      path[depth].p_hdr->eh_max)) {
1111                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1112                                  path[depth].p_hdr->eh_entries,
1113                                  path[depth].p_hdr->eh_max);
1114                 err = -EIO;
1115                 goto cleanup;
1116         }
1117         /* start copy from next extent */
1118         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1119         ext4_ext_show_move(inode, path, newblock, depth);
1120         if (m) {
1121                 struct ext4_extent *ex;
1122                 ex = EXT_FIRST_EXTENT(neh);
1123                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1124                 le16_add_cpu(&neh->eh_entries, m);
1125         }
1126
1127         ext4_extent_block_csum_set(inode, neh);
1128         set_buffer_uptodate(bh);
1129         unlock_buffer(bh);
1130
1131         err = ext4_handle_dirty_metadata(handle, inode, bh);
1132         if (err)
1133                 goto cleanup;
1134         brelse(bh);
1135         bh = NULL;
1136
1137         /* correct old leaf */
1138         if (m) {
1139                 err = ext4_ext_get_access(handle, inode, path + depth);
1140                 if (err)
1141                         goto cleanup;
1142                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1143                 err = ext4_ext_dirty(handle, inode, path + depth);
1144                 if (err)
1145                         goto cleanup;
1146
1147         }
1148
1149         /* create intermediate indexes */
1150         k = depth - at - 1;
1151         if (unlikely(k < 0)) {
1152                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1153                 err = -EIO;
1154                 goto cleanup;
1155         }
1156         if (k)
1157                 ext_debug("create %d intermediate indices\n", k);
1158         /* insert new index into current index block */
1159         /* current depth stored in i var */
1160         i = depth - 1;
1161         while (k--) {
1162                 oldblock = newblock;
1163                 newblock = ablocks[--a];
1164                 bh = sb_getblk(inode->i_sb, newblock);
1165                 if (unlikely(!bh)) {
1166                         err = -ENOMEM;
1167                         goto cleanup;
1168                 }
1169                 lock_buffer(bh);
1170
1171                 err = ext4_journal_get_create_access(handle, bh);
1172                 if (err)
1173                         goto cleanup;
1174
1175                 neh = ext_block_hdr(bh);
1176                 neh->eh_entries = cpu_to_le16(1);
1177                 neh->eh_magic = EXT4_EXT_MAGIC;
1178                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1179                 neh->eh_depth = cpu_to_le16(depth - i);
1180                 fidx = EXT_FIRST_INDEX(neh);
1181                 fidx->ei_block = border;
1182                 ext4_idx_store_pblock(fidx, oldblock);
1183
1184                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1185                                 i, newblock, le32_to_cpu(border), oldblock);
1186
1187                 /* move remainder of path[i] to the new index block */
1188                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1189                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1190                         EXT4_ERROR_INODE(inode,
1191                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1192                                          le32_to_cpu(path[i].p_ext->ee_block));
1193                         err = -EIO;
1194                         goto cleanup;
1195                 }
1196                 /* start copy indexes */
1197                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1198                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1199                                 EXT_MAX_INDEX(path[i].p_hdr));
1200                 ext4_ext_show_move(inode, path, newblock, i);
1201                 if (m) {
1202                         memmove(++fidx, path[i].p_idx,
1203                                 sizeof(struct ext4_extent_idx) * m);
1204                         le16_add_cpu(&neh->eh_entries, m);
1205                 }
1206                 ext4_extent_block_csum_set(inode, neh);
1207                 set_buffer_uptodate(bh);
1208                 unlock_buffer(bh);
1209
1210                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1211                 if (err)
1212                         goto cleanup;
1213                 brelse(bh);
1214                 bh = NULL;
1215
1216                 /* correct old index */
1217                 if (m) {
1218                         err = ext4_ext_get_access(handle, inode, path + i);
1219                         if (err)
1220                                 goto cleanup;
1221                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1222                         err = ext4_ext_dirty(handle, inode, path + i);
1223                         if (err)
1224                                 goto cleanup;
1225                 }
1226
1227                 i--;
1228         }
1229
1230         /* insert new index */
1231         err = ext4_ext_insert_index(handle, inode, path + at,
1232                                     le32_to_cpu(border), newblock);
1233
1234 cleanup:
1235         if (bh) {
1236                 if (buffer_locked(bh))
1237                         unlock_buffer(bh);
1238                 brelse(bh);
1239         }
1240
1241         if (err) {
1242                 /* free all allocated blocks in error case */
1243                 for (i = 0; i < depth; i++) {
1244                         if (!ablocks[i])
1245                                 continue;
1246                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1247                                          EXT4_FREE_BLOCKS_METADATA);
1248                 }
1249         }
1250         kfree(ablocks);
1251
1252         return err;
1253 }
1254
1255 /*
1256  * ext4_ext_grow_indepth:
1257  * implements tree growing procedure:
1258  * - allocates new block
1259  * - moves top-level data (index block or leaf) into the new block
1260  * - initializes new top-level, creating index that points to the
1261  *   just created block
1262  */
1263 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1264                                  unsigned int flags)
1265 {
1266         struct ext4_extent_header *neh;
1267         struct buffer_head *bh;
1268         ext4_fsblk_t newblock, goal = 0;
1269         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1270         int err = 0;
1271
1272         /* Try to prepend new index to old one */
1273         if (ext_depth(inode))
1274                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1275         if (goal > le32_to_cpu(es->s_first_data_block)) {
1276                 flags |= EXT4_MB_HINT_TRY_GOAL;
1277                 goal--;
1278         } else
1279                 goal = ext4_inode_to_goal_block(inode);
1280         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1281                                         NULL, &err);
1282         if (newblock == 0)
1283                 return err;
1284
1285         bh = sb_getblk(inode->i_sb, newblock);
1286         if (unlikely(!bh))
1287                 return -ENOMEM;
1288         lock_buffer(bh);
1289
1290         err = ext4_journal_get_create_access(handle, bh);
1291         if (err) {
1292                 unlock_buffer(bh);
1293                 goto out;
1294         }
1295
1296         /* move top-level index/leaf into new block */
1297         memmove(bh->b_data, EXT4_I(inode)->i_data,
1298                 sizeof(EXT4_I(inode)->i_data));
1299
1300         /* set size of new block */
1301         neh = ext_block_hdr(bh);
1302         /* old root could have indexes or leaves
1303          * so calculate e_max right way */
1304         if (ext_depth(inode))
1305                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1306         else
1307                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1308         neh->eh_magic = EXT4_EXT_MAGIC;
1309         ext4_extent_block_csum_set(inode, neh);
1310         set_buffer_uptodate(bh);
1311         unlock_buffer(bh);
1312
1313         err = ext4_handle_dirty_metadata(handle, inode, bh);
1314         if (err)
1315                 goto out;
1316
1317         /* Update top-level index: num,max,pointer */
1318         neh = ext_inode_hdr(inode);
1319         neh->eh_entries = cpu_to_le16(1);
1320         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1321         if (neh->eh_depth == 0) {
1322                 /* Root extent block becomes index block */
1323                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1324                 EXT_FIRST_INDEX(neh)->ei_block =
1325                         EXT_FIRST_EXTENT(neh)->ee_block;
1326         }
1327         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1328                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1329                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1330                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1331
1332         le16_add_cpu(&neh->eh_depth, 1);
1333         ext4_mark_inode_dirty(handle, inode);
1334 out:
1335         brelse(bh);
1336
1337         return err;
1338 }
1339
1340 /*
1341  * ext4_ext_create_new_leaf:
1342  * finds empty index and adds new leaf.
1343  * if no free index is found, then it requests in-depth growing.
1344  */
1345 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1346                                     unsigned int mb_flags,
1347                                     unsigned int gb_flags,
1348                                     struct ext4_ext_path **ppath,
1349                                     struct ext4_extent *newext)
1350 {
1351         struct ext4_ext_path *path = *ppath;
1352         struct ext4_ext_path *curp;
1353         int depth, i, err = 0;
1354
1355 repeat:
1356         i = depth = ext_depth(inode);
1357
1358         /* walk up to the tree and look for free index entry */
1359         curp = path + depth;
1360         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1361                 i--;
1362                 curp--;
1363         }
1364
1365         /* we use already allocated block for index block,
1366          * so subsequent data blocks should be contiguous */
1367         if (EXT_HAS_FREE_INDEX(curp)) {
1368                 /* if we found index with free entry, then use that
1369                  * entry: create all needed subtree and add new leaf */
1370                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1371                 if (err)
1372                         goto out;
1373
1374                 /* refill path */
1375                 path = ext4_find_extent(inode,
1376                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1377                                     ppath, gb_flags);
1378                 if (IS_ERR(path))
1379                         err = PTR_ERR(path);
1380         } else {
1381                 /* tree is full, time to grow in depth */
1382                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1383                 if (err)
1384                         goto out;
1385
1386                 /* refill path */
1387                 path = ext4_find_extent(inode,
1388                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1389                                     ppath, gb_flags);
1390                 if (IS_ERR(path)) {
1391                         err = PTR_ERR(path);
1392                         goto out;
1393                 }
1394
1395                 /*
1396                  * only first (depth 0 -> 1) produces free space;
1397                  * in all other cases we have to split the grown tree
1398                  */
1399                 depth = ext_depth(inode);
1400                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1401                         /* now we need to split */
1402                         goto repeat;
1403                 }
1404         }
1405
1406 out:
1407         return err;
1408 }
1409
1410 /*
1411  * search the closest allocated block to the left for *logical
1412  * and returns it at @logical + it's physical address at @phys
1413  * if *logical is the smallest allocated block, the function
1414  * returns 0 at @phys
1415  * return value contains 0 (success) or error code
1416  */
1417 static int ext4_ext_search_left(struct inode *inode,
1418                                 struct ext4_ext_path *path,
1419                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1420 {
1421         struct ext4_extent_idx *ix;
1422         struct ext4_extent *ex;
1423         int depth, ee_len;
1424
1425         if (unlikely(path == NULL)) {
1426                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1427                 return -EIO;
1428         }
1429         depth = path->p_depth;
1430         *phys = 0;
1431
1432         if (depth == 0 && path->p_ext == NULL)
1433                 return 0;
1434
1435         /* usually extent in the path covers blocks smaller
1436          * then *logical, but it can be that extent is the
1437          * first one in the file */
1438
1439         ex = path[depth].p_ext;
1440         ee_len = ext4_ext_get_actual_len(ex);
1441         if (*logical < le32_to_cpu(ex->ee_block)) {
1442                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1443                         EXT4_ERROR_INODE(inode,
1444                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1445                                          *logical, le32_to_cpu(ex->ee_block));
1446                         return -EIO;
1447                 }
1448                 while (--depth >= 0) {
1449                         ix = path[depth].p_idx;
1450                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1451                                 EXT4_ERROR_INODE(inode,
1452                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1453                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1454                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1455                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1456                                   depth);
1457                                 return -EIO;
1458                         }
1459                 }
1460                 return 0;
1461         }
1462
1463         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1464                 EXT4_ERROR_INODE(inode,
1465                                  "logical %d < ee_block %d + ee_len %d!",
1466                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1467                 return -EIO;
1468         }
1469
1470         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1471         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1472         return 0;
1473 }
1474
1475 /*
1476  * search the closest allocated block to the right for *logical
1477  * and returns it at @logical + it's physical address at @phys
1478  * if *logical is the largest allocated block, the function
1479  * returns 0 at @phys
1480  * return value contains 0 (success) or error code
1481  */
1482 static int ext4_ext_search_right(struct inode *inode,
1483                                  struct ext4_ext_path *path,
1484                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1485                                  struct ext4_extent **ret_ex)
1486 {
1487         struct buffer_head *bh = NULL;
1488         struct ext4_extent_header *eh;
1489         struct ext4_extent_idx *ix;
1490         struct ext4_extent *ex;
1491         ext4_fsblk_t block;
1492         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1493         int ee_len;
1494
1495         if (unlikely(path == NULL)) {
1496                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1497                 return -EIO;
1498         }
1499         depth = path->p_depth;
1500         *phys = 0;
1501
1502         if (depth == 0 && path->p_ext == NULL)
1503                 return 0;
1504
1505         /* usually extent in the path covers blocks smaller
1506          * then *logical, but it can be that extent is the
1507          * first one in the file */
1508
1509         ex = path[depth].p_ext;
1510         ee_len = ext4_ext_get_actual_len(ex);
1511         if (*logical < le32_to_cpu(ex->ee_block)) {
1512                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1513                         EXT4_ERROR_INODE(inode,
1514                                          "first_extent(path[%d].p_hdr) != ex",
1515                                          depth);
1516                         return -EIO;
1517                 }
1518                 while (--depth >= 0) {
1519                         ix = path[depth].p_idx;
1520                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1521                                 EXT4_ERROR_INODE(inode,
1522                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1523                                                  *logical);
1524                                 return -EIO;
1525                         }
1526                 }
1527                 goto found_extent;
1528         }
1529
1530         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1531                 EXT4_ERROR_INODE(inode,
1532                                  "logical %d < ee_block %d + ee_len %d!",
1533                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1534                 return -EIO;
1535         }
1536
1537         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1538                 /* next allocated block in this leaf */
1539                 ex++;
1540                 goto found_extent;
1541         }
1542
1543         /* go up and search for index to the right */
1544         while (--depth >= 0) {
1545                 ix = path[depth].p_idx;
1546                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1547                         goto got_index;
1548         }
1549
1550         /* we've gone up to the root and found no index to the right */
1551         return 0;
1552
1553 got_index:
1554         /* we've found index to the right, let's
1555          * follow it and find the closest allocated
1556          * block to the right */
1557         ix++;
1558         block = ext4_idx_pblock(ix);
1559         while (++depth < path->p_depth) {
1560                 /* subtract from p_depth to get proper eh_depth */
1561                 bh = read_extent_tree_block(inode, block,
1562                                             path->p_depth - depth, 0);
1563                 if (IS_ERR(bh))
1564                         return PTR_ERR(bh);
1565                 eh = ext_block_hdr(bh);
1566                 ix = EXT_FIRST_INDEX(eh);
1567                 block = ext4_idx_pblock(ix);
1568                 put_bh(bh);
1569         }
1570
1571         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1572         if (IS_ERR(bh))
1573                 return PTR_ERR(bh);
1574         eh = ext_block_hdr(bh);
1575         ex = EXT_FIRST_EXTENT(eh);
1576 found_extent:
1577         *logical = le32_to_cpu(ex->ee_block);
1578         *phys = ext4_ext_pblock(ex);
1579         *ret_ex = ex;
1580         if (bh)
1581                 put_bh(bh);
1582         return 0;
1583 }
1584
1585 /*
1586  * ext4_ext_next_allocated_block:
1587  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1588  * NOTE: it considers block number from index entry as
1589  * allocated block. Thus, index entries have to be consistent
1590  * with leaves.
1591  */
1592 ext4_lblk_t
1593 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1594 {
1595         int depth;
1596
1597         BUG_ON(path == NULL);
1598         depth = path->p_depth;
1599
1600         if (depth == 0 && path->p_ext == NULL)
1601                 return EXT_MAX_BLOCKS;
1602
1603         while (depth >= 0) {
1604                 if (depth == path->p_depth) {
1605                         /* leaf */
1606                         if (path[depth].p_ext &&
1607                                 path[depth].p_ext !=
1608                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1609                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1610                 } else {
1611                         /* index */
1612                         if (path[depth].p_idx !=
1613                                         EXT_LAST_INDEX(path[depth].p_hdr))
1614                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1615                 }
1616                 depth--;
1617         }
1618
1619         return EXT_MAX_BLOCKS;
1620 }
1621
1622 /*
1623  * ext4_ext_next_leaf_block:
1624  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1625  */
1626 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1627 {
1628         int depth;
1629
1630         BUG_ON(path == NULL);
1631         depth = path->p_depth;
1632
1633         /* zero-tree has no leaf blocks at all */
1634         if (depth == 0)
1635                 return EXT_MAX_BLOCKS;
1636
1637         /* go to index block */
1638         depth--;
1639
1640         while (depth >= 0) {
1641                 if (path[depth].p_idx !=
1642                                 EXT_LAST_INDEX(path[depth].p_hdr))
1643                         return (ext4_lblk_t)
1644                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1645                 depth--;
1646         }
1647
1648         return EXT_MAX_BLOCKS;
1649 }
1650
1651 /*
1652  * ext4_ext_correct_indexes:
1653  * if leaf gets modified and modified extent is first in the leaf,
1654  * then we have to correct all indexes above.
1655  * TODO: do we need to correct tree in all cases?
1656  */
1657 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1658                                 struct ext4_ext_path *path)
1659 {
1660         struct ext4_extent_header *eh;
1661         int depth = ext_depth(inode);
1662         struct ext4_extent *ex;
1663         __le32 border;
1664         int k, err = 0;
1665
1666         eh = path[depth].p_hdr;
1667         ex = path[depth].p_ext;
1668
1669         if (unlikely(ex == NULL || eh == NULL)) {
1670                 EXT4_ERROR_INODE(inode,
1671                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1672                 return -EIO;
1673         }
1674
1675         if (depth == 0) {
1676                 /* there is no tree at all */
1677                 return 0;
1678         }
1679
1680         if (ex != EXT_FIRST_EXTENT(eh)) {
1681                 /* we correct tree if first leaf got modified only */
1682                 return 0;
1683         }
1684
1685         /*
1686          * TODO: we need correction if border is smaller than current one
1687          */
1688         k = depth - 1;
1689         border = path[depth].p_ext->ee_block;
1690         err = ext4_ext_get_access(handle, inode, path + k);
1691         if (err)
1692                 return err;
1693         path[k].p_idx->ei_block = border;
1694         err = ext4_ext_dirty(handle, inode, path + k);
1695         if (err)
1696                 return err;
1697
1698         while (k--) {
1699                 /* change all left-side indexes */
1700                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1701                         break;
1702                 err = ext4_ext_get_access(handle, inode, path + k);
1703                 if (err)
1704                         break;
1705                 path[k].p_idx->ei_block = border;
1706                 err = ext4_ext_dirty(handle, inode, path + k);
1707                 if (err)
1708                         break;
1709         }
1710
1711         return err;
1712 }
1713
1714 int
1715 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1716                                 struct ext4_extent *ex2)
1717 {
1718         unsigned short ext1_ee_len, ext2_ee_len;
1719
1720         /*
1721          * Make sure that both extents are initialized. We don't merge
1722          * unwritten extents so that we can be sure that end_io code has
1723          * the extent that was written properly split out and conversion to
1724          * initialized is trivial.
1725          */
1726         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1727                 return 0;
1728
1729         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1730         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1731
1732         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1733                         le32_to_cpu(ex2->ee_block))
1734                 return 0;
1735
1736         /*
1737          * To allow future support for preallocated extents to be added
1738          * as an RO_COMPAT feature, refuse to merge to extents if
1739          * this can result in the top bit of ee_len being set.
1740          */
1741         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1742                 return 0;
1743         if (ext4_ext_is_unwritten(ex1) &&
1744             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1745              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1746              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1747                 return 0;
1748 #ifdef AGGRESSIVE_TEST
1749         if (ext1_ee_len >= 4)
1750                 return 0;
1751 #endif
1752
1753         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1754                 return 1;
1755         return 0;
1756 }
1757
1758 /*
1759  * This function tries to merge the "ex" extent to the next extent in the tree.
1760  * It always tries to merge towards right. If you want to merge towards
1761  * left, pass "ex - 1" as argument instead of "ex".
1762  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1763  * 1 if they got merged.
1764  */
1765 static int ext4_ext_try_to_merge_right(struct inode *inode,
1766                                  struct ext4_ext_path *path,
1767                                  struct ext4_extent *ex)
1768 {
1769         struct ext4_extent_header *eh;
1770         unsigned int depth, len;
1771         int merge_done = 0, unwritten;
1772
1773         depth = ext_depth(inode);
1774         BUG_ON(path[depth].p_hdr == NULL);
1775         eh = path[depth].p_hdr;
1776
1777         while (ex < EXT_LAST_EXTENT(eh)) {
1778                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1779                         break;
1780                 /* merge with next extent! */
1781                 unwritten = ext4_ext_is_unwritten(ex);
1782                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1783                                 + ext4_ext_get_actual_len(ex + 1));
1784                 if (unwritten)
1785                         ext4_ext_mark_unwritten(ex);
1786
1787                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1788                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1789                                 * sizeof(struct ext4_extent);
1790                         memmove(ex + 1, ex + 2, len);
1791                 }
1792                 le16_add_cpu(&eh->eh_entries, -1);
1793                 merge_done = 1;
1794                 WARN_ON(eh->eh_entries == 0);
1795                 if (!eh->eh_entries)
1796                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1797         }
1798
1799         return merge_done;
1800 }
1801
1802 /*
1803  * This function does a very simple check to see if we can collapse
1804  * an extent tree with a single extent tree leaf block into the inode.
1805  */
1806 static void ext4_ext_try_to_merge_up(handle_t *handle,
1807                                      struct inode *inode,
1808                                      struct ext4_ext_path *path)
1809 {
1810         size_t s;
1811         unsigned max_root = ext4_ext_space_root(inode, 0);
1812         ext4_fsblk_t blk;
1813
1814         if ((path[0].p_depth != 1) ||
1815             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1816             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1817                 return;
1818
1819         /*
1820          * We need to modify the block allocation bitmap and the block
1821          * group descriptor to release the extent tree block.  If we
1822          * can't get the journal credits, give up.
1823          */
1824         if (ext4_journal_extend(handle, 2))
1825                 return;
1826
1827         /*
1828          * Copy the extent data up to the inode
1829          */
1830         blk = ext4_idx_pblock(path[0].p_idx);
1831         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1832                 sizeof(struct ext4_extent_idx);
1833         s += sizeof(struct ext4_extent_header);
1834
1835         path[1].p_maxdepth = path[0].p_maxdepth;
1836         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1837         path[0].p_depth = 0;
1838         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1839                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1840         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1841
1842         brelse(path[1].p_bh);
1843         ext4_free_blocks(handle, inode, NULL, blk, 1,
1844                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1845 }
1846
1847 /*
1848  * This function tries to merge the @ex extent to neighbours in the tree.
1849  * return 1 if merge left else 0.
1850  */
1851 static void ext4_ext_try_to_merge(handle_t *handle,
1852                                   struct inode *inode,
1853                                   struct ext4_ext_path *path,
1854                                   struct ext4_extent *ex) {
1855         struct ext4_extent_header *eh;
1856         unsigned int depth;
1857         int merge_done = 0;
1858
1859         depth = ext_depth(inode);
1860         BUG_ON(path[depth].p_hdr == NULL);
1861         eh = path[depth].p_hdr;
1862
1863         if (ex > EXT_FIRST_EXTENT(eh))
1864                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1865
1866         if (!merge_done)
1867                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1868
1869         ext4_ext_try_to_merge_up(handle, inode, path);
1870 }
1871
1872 /*
1873  * check if a portion of the "newext" extent overlaps with an
1874  * existing extent.
1875  *
1876  * If there is an overlap discovered, it updates the length of the newext
1877  * such that there will be no overlap, and then returns 1.
1878  * If there is no overlap found, it returns 0.
1879  */
1880 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1881                                            struct inode *inode,
1882                                            struct ext4_extent *newext,
1883                                            struct ext4_ext_path *path)
1884 {
1885         ext4_lblk_t b1, b2;
1886         unsigned int depth, len1;
1887         unsigned int ret = 0;
1888
1889         b1 = le32_to_cpu(newext->ee_block);
1890         len1 = ext4_ext_get_actual_len(newext);
1891         depth = ext_depth(inode);
1892         if (!path[depth].p_ext)
1893                 goto out;
1894         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1895
1896         /*
1897          * get the next allocated block if the extent in the path
1898          * is before the requested block(s)
1899          */
1900         if (b2 < b1) {
1901                 b2 = ext4_ext_next_allocated_block(path);
1902                 if (b2 == EXT_MAX_BLOCKS)
1903                         goto out;
1904                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1905         }
1906
1907         /* check for wrap through zero on extent logical start block*/
1908         if (b1 + len1 < b1) {
1909                 len1 = EXT_MAX_BLOCKS - b1;
1910                 newext->ee_len = cpu_to_le16(len1);
1911                 ret = 1;
1912         }
1913
1914         /* check for overlap */
1915         if (b1 + len1 > b2) {
1916                 newext->ee_len = cpu_to_le16(b2 - b1);
1917                 ret = 1;
1918         }
1919 out:
1920         return ret;
1921 }
1922
1923 /*
1924  * ext4_ext_insert_extent:
1925  * tries to merge requsted extent into the existing extent or
1926  * inserts requested extent as new one into the tree,
1927  * creating new leaf in the no-space case.
1928  */
1929 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1930                                 struct ext4_ext_path **ppath,
1931                                 struct ext4_extent *newext, int gb_flags)
1932 {
1933         struct ext4_ext_path *path = *ppath;
1934         struct ext4_extent_header *eh;
1935         struct ext4_extent *ex, *fex;
1936         struct ext4_extent *nearex; /* nearest extent */
1937         struct ext4_ext_path *npath = NULL;
1938         int depth, len, err;
1939         ext4_lblk_t next;
1940         int mb_flags = 0, unwritten;
1941
1942         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1943                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1944         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1945                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1946                 return -EIO;
1947         }
1948         depth = ext_depth(inode);
1949         ex = path[depth].p_ext;
1950         eh = path[depth].p_hdr;
1951         if (unlikely(path[depth].p_hdr == NULL)) {
1952                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1953                 return -EIO;
1954         }
1955
1956         /* try to insert block into found extent and return */
1957         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1958
1959                 /*
1960                  * Try to see whether we should rather test the extent on
1961                  * right from ex, or from the left of ex. This is because
1962                  * ext4_find_extent() can return either extent on the
1963                  * left, or on the right from the searched position. This
1964                  * will make merging more effective.
1965                  */
1966                 if (ex < EXT_LAST_EXTENT(eh) &&
1967                     (le32_to_cpu(ex->ee_block) +
1968                     ext4_ext_get_actual_len(ex) <
1969                     le32_to_cpu(newext->ee_block))) {
1970                         ex += 1;
1971                         goto prepend;
1972                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1973                            (le32_to_cpu(newext->ee_block) +
1974                            ext4_ext_get_actual_len(newext) <
1975                            le32_to_cpu(ex->ee_block)))
1976                         ex -= 1;
1977
1978                 /* Try to append newex to the ex */
1979                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1980                         ext_debug("append [%d]%d block to %u:[%d]%d"
1981                                   "(from %llu)\n",
1982                                   ext4_ext_is_unwritten(newext),
1983                                   ext4_ext_get_actual_len(newext),
1984                                   le32_to_cpu(ex->ee_block),
1985                                   ext4_ext_is_unwritten(ex),
1986                                   ext4_ext_get_actual_len(ex),
1987                                   ext4_ext_pblock(ex));
1988                         err = ext4_ext_get_access(handle, inode,
1989                                                   path + depth);
1990                         if (err)
1991                                 return err;
1992                         unwritten = ext4_ext_is_unwritten(ex);
1993                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1994                                         + ext4_ext_get_actual_len(newext));
1995                         if (unwritten)
1996                                 ext4_ext_mark_unwritten(ex);
1997                         eh = path[depth].p_hdr;
1998                         nearex = ex;
1999                         goto merge;
2000                 }
2001
2002 prepend:
2003                 /* Try to prepend newex to the ex */
2004                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2005                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2006                                   "(from %llu)\n",
2007                                   le32_to_cpu(newext->ee_block),
2008                                   ext4_ext_is_unwritten(newext),
2009                                   ext4_ext_get_actual_len(newext),
2010                                   le32_to_cpu(ex->ee_block),
2011                                   ext4_ext_is_unwritten(ex),
2012                                   ext4_ext_get_actual_len(ex),
2013                                   ext4_ext_pblock(ex));
2014                         err = ext4_ext_get_access(handle, inode,
2015                                                   path + depth);
2016                         if (err)
2017                                 return err;
2018
2019                         unwritten = ext4_ext_is_unwritten(ex);
2020                         ex->ee_block = newext->ee_block;
2021                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2022                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2023                                         + ext4_ext_get_actual_len(newext));
2024                         if (unwritten)
2025                                 ext4_ext_mark_unwritten(ex);
2026                         eh = path[depth].p_hdr;
2027                         nearex = ex;
2028                         goto merge;
2029                 }
2030         }
2031
2032         depth = ext_depth(inode);
2033         eh = path[depth].p_hdr;
2034         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2035                 goto has_space;
2036
2037         /* probably next leaf has space for us? */
2038         fex = EXT_LAST_EXTENT(eh);
2039         next = EXT_MAX_BLOCKS;
2040         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2041                 next = ext4_ext_next_leaf_block(path);
2042         if (next != EXT_MAX_BLOCKS) {
2043                 ext_debug("next leaf block - %u\n", next);
2044                 BUG_ON(npath != NULL);
2045                 npath = ext4_find_extent(inode, next, NULL, 0);
2046                 if (IS_ERR(npath))
2047                         return PTR_ERR(npath);
2048                 BUG_ON(npath->p_depth != path->p_depth);
2049                 eh = npath[depth].p_hdr;
2050                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2051                         ext_debug("next leaf isn't full(%d)\n",
2052                                   le16_to_cpu(eh->eh_entries));
2053                         path = npath;
2054                         goto has_space;
2055                 }
2056                 ext_debug("next leaf has no free space(%d,%d)\n",
2057                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2058         }
2059
2060         /*
2061          * There is no free space in the found leaf.
2062          * We're gonna add a new leaf in the tree.
2063          */
2064         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2065                 mb_flags |= EXT4_MB_USE_RESERVED;
2066         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2067                                        ppath, newext);
2068         if (err)
2069                 goto cleanup;
2070         depth = ext_depth(inode);
2071         eh = path[depth].p_hdr;
2072
2073 has_space:
2074         nearex = path[depth].p_ext;
2075
2076         err = ext4_ext_get_access(handle, inode, path + depth);
2077         if (err)
2078                 goto cleanup;
2079
2080         if (!nearex) {
2081                 /* there is no extent in this leaf, create first one */
2082                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2083                                 le32_to_cpu(newext->ee_block),
2084                                 ext4_ext_pblock(newext),
2085                                 ext4_ext_is_unwritten(newext),
2086                                 ext4_ext_get_actual_len(newext));
2087                 nearex = EXT_FIRST_EXTENT(eh);
2088         } else {
2089                 if (le32_to_cpu(newext->ee_block)
2090                            > le32_to_cpu(nearex->ee_block)) {
2091                         /* Insert after */
2092                         ext_debug("insert %u:%llu:[%d]%d before: "
2093                                         "nearest %p\n",
2094                                         le32_to_cpu(newext->ee_block),
2095                                         ext4_ext_pblock(newext),
2096                                         ext4_ext_is_unwritten(newext),
2097                                         ext4_ext_get_actual_len(newext),
2098                                         nearex);
2099                         nearex++;
2100                 } else {
2101                         /* Insert before */
2102                         BUG_ON(newext->ee_block == nearex->ee_block);
2103                         ext_debug("insert %u:%llu:[%d]%d after: "
2104                                         "nearest %p\n",
2105                                         le32_to_cpu(newext->ee_block),
2106                                         ext4_ext_pblock(newext),
2107                                         ext4_ext_is_unwritten(newext),
2108                                         ext4_ext_get_actual_len(newext),
2109                                         nearex);
2110                 }
2111                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2112                 if (len > 0) {
2113                         ext_debug("insert %u:%llu:[%d]%d: "
2114                                         "move %d extents from 0x%p to 0x%p\n",
2115                                         le32_to_cpu(newext->ee_block),
2116                                         ext4_ext_pblock(newext),
2117                                         ext4_ext_is_unwritten(newext),
2118                                         ext4_ext_get_actual_len(newext),
2119                                         len, nearex, nearex + 1);
2120                         memmove(nearex + 1, nearex,
2121                                 len * sizeof(struct ext4_extent));
2122                 }
2123         }
2124
2125         le16_add_cpu(&eh->eh_entries, 1);
2126         path[depth].p_ext = nearex;
2127         nearex->ee_block = newext->ee_block;
2128         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2129         nearex->ee_len = newext->ee_len;
2130
2131 merge:
2132         /* try to merge extents */
2133         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2134                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2135
2136
2137         /* time to correct all indexes above */
2138         err = ext4_ext_correct_indexes(handle, inode, path);
2139         if (err)
2140                 goto cleanup;
2141
2142         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2143
2144 cleanup:
2145         ext4_ext_drop_refs(npath);
2146         kfree(npath);
2147         return err;
2148 }
2149
2150 static int ext4_fill_fiemap_extents(struct inode *inode,
2151                                     ext4_lblk_t block, ext4_lblk_t num,
2152                                     struct fiemap_extent_info *fieinfo)
2153 {
2154         struct ext4_ext_path *path = NULL;
2155         struct ext4_extent *ex;
2156         struct extent_status es;
2157         ext4_lblk_t next, next_del, start = 0, end = 0;
2158         ext4_lblk_t last = block + num;
2159         int exists, depth = 0, err = 0;
2160         unsigned int flags = 0;
2161         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2162
2163         while (block < last && block != EXT_MAX_BLOCKS) {
2164                 num = last - block;
2165                 /* find extent for this block */
2166                 down_read(&EXT4_I(inode)->i_data_sem);
2167
2168                 path = ext4_find_extent(inode, block, &path, 0);
2169                 if (IS_ERR(path)) {
2170                         up_read(&EXT4_I(inode)->i_data_sem);
2171                         err = PTR_ERR(path);
2172                         path = NULL;
2173                         break;
2174                 }
2175
2176                 depth = ext_depth(inode);
2177                 if (unlikely(path[depth].p_hdr == NULL)) {
2178                         up_read(&EXT4_I(inode)->i_data_sem);
2179                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2180                         err = -EIO;
2181                         break;
2182                 }
2183                 ex = path[depth].p_ext;
2184                 next = ext4_ext_next_allocated_block(path);
2185
2186                 flags = 0;
2187                 exists = 0;
2188                 if (!ex) {
2189                         /* there is no extent yet, so try to allocate
2190                          * all requested space */
2191                         start = block;
2192                         end = block + num;
2193                 } else if (le32_to_cpu(ex->ee_block) > block) {
2194                         /* need to allocate space before found extent */
2195                         start = block;
2196                         end = le32_to_cpu(ex->ee_block);
2197                         if (block + num < end)
2198                                 end = block + num;
2199                 } else if (block >= le32_to_cpu(ex->ee_block)
2200                                         + ext4_ext_get_actual_len(ex)) {
2201                         /* need to allocate space after found extent */
2202                         start = block;
2203                         end = block + num;
2204                         if (end >= next)
2205                                 end = next;
2206                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2207                         /*
2208                          * some part of requested space is covered
2209                          * by found extent
2210                          */
2211                         start = block;
2212                         end = le32_to_cpu(ex->ee_block)
2213                                 + ext4_ext_get_actual_len(ex);
2214                         if (block + num < end)
2215                                 end = block + num;
2216                         exists = 1;
2217                 } else {
2218                         BUG();
2219                 }
2220                 BUG_ON(end <= start);
2221
2222                 if (!exists) {
2223                         es.es_lblk = start;
2224                         es.es_len = end - start;
2225                         es.es_pblk = 0;
2226                 } else {
2227                         es.es_lblk = le32_to_cpu(ex->ee_block);
2228                         es.es_len = ext4_ext_get_actual_len(ex);
2229                         es.es_pblk = ext4_ext_pblock(ex);
2230                         if (ext4_ext_is_unwritten(ex))
2231                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2232                 }
2233
2234                 /*
2235                  * Find delayed extent and update es accordingly. We call
2236                  * it even in !exists case to find out whether es is the
2237                  * last existing extent or not.
2238                  */
2239                 next_del = ext4_find_delayed_extent(inode, &es);
2240                 if (!exists && next_del) {
2241                         exists = 1;
2242                         flags |= (FIEMAP_EXTENT_DELALLOC |
2243                                   FIEMAP_EXTENT_UNKNOWN);
2244                 }
2245                 up_read(&EXT4_I(inode)->i_data_sem);
2246
2247                 if (unlikely(es.es_len == 0)) {
2248                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2249                         err = -EIO;
2250                         break;
2251                 }
2252
2253                 /*
2254                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2255                  * we need to check next == EXT_MAX_BLOCKS because it is
2256                  * possible that an extent is with unwritten and delayed
2257                  * status due to when an extent is delayed allocated and
2258                  * is allocated by fallocate status tree will track both of
2259                  * them in a extent.
2260                  *
2261                  * So we could return a unwritten and delayed extent, and
2262                  * its block is equal to 'next'.
2263                  */
2264                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2265                         flags |= FIEMAP_EXTENT_LAST;
2266                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2267                                      next != EXT_MAX_BLOCKS)) {
2268                                 EXT4_ERROR_INODE(inode,
2269                                                  "next extent == %u, next "
2270                                                  "delalloc extent = %u",
2271                                                  next, next_del);
2272                                 err = -EIO;
2273                                 break;
2274                         }
2275                 }
2276
2277                 if (exists) {
2278                         err = fiemap_fill_next_extent(fieinfo,
2279                                 (__u64)es.es_lblk << blksize_bits,
2280                                 (__u64)es.es_pblk << blksize_bits,
2281                                 (__u64)es.es_len << blksize_bits,
2282                                 flags);
2283                         if (err < 0)
2284                                 break;
2285                         if (err == 1) {
2286                                 err = 0;
2287                                 break;
2288                         }
2289                 }
2290
2291                 block = es.es_lblk + es.es_len;
2292         }
2293
2294         ext4_ext_drop_refs(path);
2295         kfree(path);
2296         return err;
2297 }
2298
2299 /*
2300  * ext4_ext_put_gap_in_cache:
2301  * calculate boundaries of the gap that the requested block fits into
2302  * and cache this gap
2303  */
2304 static void
2305 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2306                                 ext4_lblk_t block)
2307 {
2308         int depth = ext_depth(inode);
2309         ext4_lblk_t len;
2310         ext4_lblk_t lblock;
2311         struct ext4_extent *ex;
2312         struct extent_status es;
2313
2314         ex = path[depth].p_ext;
2315         if (ex == NULL) {
2316                 /* there is no extent yet, so gap is [0;-] */
2317                 lblock = 0;
2318                 len = EXT_MAX_BLOCKS;
2319                 ext_debug("cache gap(whole file):");
2320         } else if (block < le32_to_cpu(ex->ee_block)) {
2321                 lblock = block;
2322                 len = le32_to_cpu(ex->ee_block) - block;
2323                 ext_debug("cache gap(before): %u [%u:%u]",
2324                                 block,
2325                                 le32_to_cpu(ex->ee_block),
2326                                  ext4_ext_get_actual_len(ex));
2327         } else if (block >= le32_to_cpu(ex->ee_block)
2328                         + ext4_ext_get_actual_len(ex)) {
2329                 ext4_lblk_t next;
2330                 lblock = le32_to_cpu(ex->ee_block)
2331                         + ext4_ext_get_actual_len(ex);
2332
2333                 next = ext4_ext_next_allocated_block(path);
2334                 ext_debug("cache gap(after): [%u:%u] %u",
2335                                 le32_to_cpu(ex->ee_block),
2336                                 ext4_ext_get_actual_len(ex),
2337                                 block);
2338                 BUG_ON(next == lblock);
2339                 len = next - lblock;
2340         } else {
2341                 BUG();
2342         }
2343
2344         ext4_es_find_delayed_extent_range(inode, lblock, lblock + len - 1, &es);
2345         if (es.es_len) {
2346                 /* There's delayed extent containing lblock? */
2347                 if (es.es_lblk <= lblock)
2348                         return;
2349                 len = min(es.es_lblk - lblock, len);
2350         }
2351         ext_debug(" -> %u:%u\n", lblock, len);
2352         ext4_es_insert_extent(inode, lblock, len, ~0, EXTENT_STATUS_HOLE);
2353 }
2354
2355 /*
2356  * ext4_ext_rm_idx:
2357  * removes index from the index block.
2358  */
2359 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2360                         struct ext4_ext_path *path, int depth)
2361 {
2362         int err;
2363         ext4_fsblk_t leaf;
2364
2365         /* free index block */
2366         depth--;
2367         path = path + depth;
2368         leaf = ext4_idx_pblock(path->p_idx);
2369         if (unlikely(path->p_hdr->eh_entries == 0)) {
2370                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2371                 return -EIO;
2372         }
2373         err = ext4_ext_get_access(handle, inode, path);
2374         if (err)
2375                 return err;
2376
2377         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2378                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2379                 len *= sizeof(struct ext4_extent_idx);
2380                 memmove(path->p_idx, path->p_idx + 1, len);
2381         }
2382
2383         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2384         err = ext4_ext_dirty(handle, inode, path);
2385         if (err)
2386                 return err;
2387         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2388         trace_ext4_ext_rm_idx(inode, leaf);
2389
2390         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2391                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2392
2393         while (--depth >= 0) {
2394                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2395                         break;
2396                 path--;
2397                 err = ext4_ext_get_access(handle, inode, path);
2398                 if (err)
2399                         break;
2400                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2401                 err = ext4_ext_dirty(handle, inode, path);
2402                 if (err)
2403                         break;
2404         }
2405         return err;
2406 }
2407
2408 /*
2409  * ext4_ext_calc_credits_for_single_extent:
2410  * This routine returns max. credits that needed to insert an extent
2411  * to the extent tree.
2412  * When pass the actual path, the caller should calculate credits
2413  * under i_data_sem.
2414  */
2415 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2416                                                 struct ext4_ext_path *path)
2417 {
2418         if (path) {
2419                 int depth = ext_depth(inode);
2420                 int ret = 0;
2421
2422                 /* probably there is space in leaf? */
2423                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2424                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2425
2426                         /*
2427                          *  There are some space in the leaf tree, no
2428                          *  need to account for leaf block credit
2429                          *
2430                          *  bitmaps and block group descriptor blocks
2431                          *  and other metadata blocks still need to be
2432                          *  accounted.
2433                          */
2434                         /* 1 bitmap, 1 block group descriptor */
2435                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2436                         return ret;
2437                 }
2438         }
2439
2440         return ext4_chunk_trans_blocks(inode, nrblocks);
2441 }
2442
2443 /*
2444  * How many index/leaf blocks need to change/allocate to add @extents extents?
2445  *
2446  * If we add a single extent, then in the worse case, each tree level
2447  * index/leaf need to be changed in case of the tree split.
2448  *
2449  * If more extents are inserted, they could cause the whole tree split more
2450  * than once, but this is really rare.
2451  */
2452 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2453 {
2454         int index;
2455         int depth;
2456
2457         /* If we are converting the inline data, only one is needed here. */
2458         if (ext4_has_inline_data(inode))
2459                 return 1;
2460
2461         depth = ext_depth(inode);
2462
2463         if (extents <= 1)
2464                 index = depth * 2;
2465         else
2466                 index = depth * 3;
2467
2468         return index;
2469 }
2470
2471 static inline int get_default_free_blocks_flags(struct inode *inode)
2472 {
2473         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2474                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2475         else if (ext4_should_journal_data(inode))
2476                 return EXT4_FREE_BLOCKS_FORGET;
2477         return 0;
2478 }
2479
2480 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2481                               struct ext4_extent *ex,
2482                               long long *partial_cluster,
2483                               ext4_lblk_t from, ext4_lblk_t to)
2484 {
2485         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2486         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2487         ext4_fsblk_t pblk;
2488         int flags = get_default_free_blocks_flags(inode);
2489
2490         /*
2491          * For bigalloc file systems, we never free a partial cluster
2492          * at the beginning of the extent.  Instead, we make a note
2493          * that we tried freeing the cluster, and check to see if we
2494          * need to free it on a subsequent call to ext4_remove_blocks,
2495          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2496          */
2497         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2498
2499         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2500         /*
2501          * If we have a partial cluster, and it's different from the
2502          * cluster of the last block, we need to explicitly free the
2503          * partial cluster here.
2504          */
2505         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2506         if (*partial_cluster > 0 &&
2507             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2508                 ext4_free_blocks(handle, inode, NULL,
2509                                  EXT4_C2B(sbi, *partial_cluster),
2510                                  sbi->s_cluster_ratio, flags);
2511                 *partial_cluster = 0;
2512         }
2513
2514 #ifdef EXTENTS_STATS
2515         {
2516                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2517                 spin_lock(&sbi->s_ext_stats_lock);
2518                 sbi->s_ext_blocks += ee_len;
2519                 sbi->s_ext_extents++;
2520                 if (ee_len < sbi->s_ext_min)
2521                         sbi->s_ext_min = ee_len;
2522                 if (ee_len > sbi->s_ext_max)
2523                         sbi->s_ext_max = ee_len;
2524                 if (ext_depth(inode) > sbi->s_depth_max)
2525                         sbi->s_depth_max = ext_depth(inode);
2526                 spin_unlock(&sbi->s_ext_stats_lock);
2527         }
2528 #endif
2529         if (from >= le32_to_cpu(ex->ee_block)
2530             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2531                 /* tail removal */
2532                 ext4_lblk_t num;
2533                 long long first_cluster;
2534
2535                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2536                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2537                 /*
2538                  * Usually we want to free partial cluster at the end of the
2539                  * extent, except for the situation when the cluster is still
2540                  * used by any other extent (partial_cluster is negative).
2541                  */
2542                 if (*partial_cluster < 0 &&
2543                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2544                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2545
2546                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2547                           num, pblk, *partial_cluster);
2548                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2549                 /*
2550                  * If the block range to be freed didn't start at the
2551                  * beginning of a cluster, and we removed the entire
2552                  * extent and the cluster is not used by any other extent,
2553                  * save the partial cluster here, since we might need to
2554                  * delete if we determine that the truncate or punch hole
2555                  * operation has removed all of the blocks in the cluster.
2556                  * If that cluster is used by another extent, preserve its
2557                  * negative value so it isn't freed later on.
2558                  *
2559                  * If the whole extent wasn't freed, we've reached the
2560                  * start of the truncated/punched region and have finished
2561                  * removing blocks.  If there's a partial cluster here it's
2562                  * shared with the remainder of the extent and is no longer
2563                  * a candidate for removal.
2564                  */
2565                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2566                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2567                         if (first_cluster != -*partial_cluster)
2568                                 *partial_cluster = first_cluster;
2569                 } else {
2570                         *partial_cluster = 0;
2571                 }
2572         } else
2573                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2574                            "%u-%u from %u:%u\n",
2575                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2576         return 0;
2577 }
2578
2579
2580 /*
2581  * ext4_ext_rm_leaf() Removes the extents associated with the
2582  * blocks appearing between "start" and "end".  Both "start"
2583  * and "end" must appear in the same extent or EIO is returned.
2584  *
2585  * @handle: The journal handle
2586  * @inode:  The files inode
2587  * @path:   The path to the leaf
2588  * @partial_cluster: The cluster which we'll have to free if all extents
2589  *                   has been released from it.  However, if this value is
2590  *                   negative, it's a cluster just to the right of the
2591  *                   punched region and it must not be freed.
2592  * @start:  The first block to remove
2593  * @end:   The last block to remove
2594  */
2595 static int
2596 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2597                  struct ext4_ext_path *path,
2598                  long long *partial_cluster,
2599                  ext4_lblk_t start, ext4_lblk_t end)
2600 {
2601         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2602         int err = 0, correct_index = 0;
2603         int depth = ext_depth(inode), credits;
2604         struct ext4_extent_header *eh;
2605         ext4_lblk_t a, b;
2606         unsigned num;
2607         ext4_lblk_t ex_ee_block;
2608         unsigned short ex_ee_len;
2609         unsigned unwritten = 0;
2610         struct ext4_extent *ex;
2611         ext4_fsblk_t pblk;
2612
2613         /* the header must be checked already in ext4_ext_remove_space() */
2614         ext_debug("truncate since %u in leaf to %u\n", start, end);
2615         if (!path[depth].p_hdr)
2616                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2617         eh = path[depth].p_hdr;
2618         if (unlikely(path[depth].p_hdr == NULL)) {
2619                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2620                 return -EIO;
2621         }
2622         /* find where to start removing */
2623         ex = path[depth].p_ext;
2624         if (!ex)
2625                 ex = EXT_LAST_EXTENT(eh);
2626
2627         ex_ee_block = le32_to_cpu(ex->ee_block);
2628         ex_ee_len = ext4_ext_get_actual_len(ex);
2629
2630         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2631
2632         while (ex >= EXT_FIRST_EXTENT(eh) &&
2633                         ex_ee_block + ex_ee_len > start) {
2634
2635                 if (ext4_ext_is_unwritten(ex))
2636                         unwritten = 1;
2637                 else
2638                         unwritten = 0;
2639
2640                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2641                           unwritten, ex_ee_len);
2642                 path[depth].p_ext = ex;
2643
2644                 a = ex_ee_block > start ? ex_ee_block : start;
2645                 b = ex_ee_block+ex_ee_len - 1 < end ?
2646                         ex_ee_block+ex_ee_len - 1 : end;
2647
2648                 ext_debug("  border %u:%u\n", a, b);
2649
2650                 /* If this extent is beyond the end of the hole, skip it */
2651                 if (end < ex_ee_block) {
2652                         /*
2653                          * We're going to skip this extent and move to another,
2654                          * so note that its first cluster is in use to avoid
2655                          * freeing it when removing blocks.  Eventually, the
2656                          * right edge of the truncated/punched region will
2657                          * be just to the left.
2658                          */
2659                         if (sbi->s_cluster_ratio > 1) {
2660                                 pblk = ext4_ext_pblock(ex);
2661                                 *partial_cluster =
2662                                         -(long long) EXT4_B2C(sbi, pblk);
2663                         }
2664                         ex--;
2665                         ex_ee_block = le32_to_cpu(ex->ee_block);
2666                         ex_ee_len = ext4_ext_get_actual_len(ex);
2667                         continue;
2668                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2669                         EXT4_ERROR_INODE(inode,
2670                                          "can not handle truncate %u:%u "
2671                                          "on extent %u:%u",
2672                                          start, end, ex_ee_block,
2673                                          ex_ee_block + ex_ee_len - 1);
2674                         err = -EIO;
2675                         goto out;
2676                 } else if (a != ex_ee_block) {
2677                         /* remove tail of the extent */
2678                         num = a - ex_ee_block;
2679                 } else {
2680                         /* remove whole extent: excellent! */
2681                         num = 0;
2682                 }
2683                 /*
2684                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2685                  * descriptor) for each block group; assume two block
2686                  * groups plus ex_ee_len/blocks_per_block_group for
2687                  * the worst case
2688                  */
2689                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2690                 if (ex == EXT_FIRST_EXTENT(eh)) {
2691                         correct_index = 1;
2692                         credits += (ext_depth(inode)) + 1;
2693                 }
2694                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2695
2696                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2697                 if (err)
2698                         goto out;
2699
2700                 err = ext4_ext_get_access(handle, inode, path + depth);
2701                 if (err)
2702                         goto out;
2703
2704                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2705                                          a, b);
2706                 if (err)
2707                         goto out;
2708
2709                 if (num == 0)
2710                         /* this extent is removed; mark slot entirely unused */
2711                         ext4_ext_store_pblock(ex, 0);
2712
2713                 ex->ee_len = cpu_to_le16(num);
2714                 /*
2715                  * Do not mark unwritten if all the blocks in the
2716                  * extent have been removed.
2717                  */
2718                 if (unwritten && num)
2719                         ext4_ext_mark_unwritten(ex);
2720                 /*
2721                  * If the extent was completely released,
2722                  * we need to remove it from the leaf
2723                  */
2724                 if (num == 0) {
2725                         if (end != EXT_MAX_BLOCKS - 1) {
2726                                 /*
2727                                  * For hole punching, we need to scoot all the
2728                                  * extents up when an extent is removed so that
2729                                  * we dont have blank extents in the middle
2730                                  */
2731                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2732                                         sizeof(struct ext4_extent));
2733
2734                                 /* Now get rid of the one at the end */
2735                                 memset(EXT_LAST_EXTENT(eh), 0,
2736                                         sizeof(struct ext4_extent));
2737                         }
2738                         le16_add_cpu(&eh->eh_entries, -1);
2739                 }
2740
2741                 err = ext4_ext_dirty(handle, inode, path + depth);
2742                 if (err)
2743                         goto out;
2744
2745                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2746                                 ext4_ext_pblock(ex));
2747                 ex--;
2748                 ex_ee_block = le32_to_cpu(ex->ee_block);
2749                 ex_ee_len = ext4_ext_get_actual_len(ex);
2750         }
2751
2752         if (correct_index && eh->eh_entries)
2753                 err = ext4_ext_correct_indexes(handle, inode, path);
2754
2755         /*
2756          * If there's a partial cluster and at least one extent remains in
2757          * the leaf, free the partial cluster if it isn't shared with the
2758          * current extent.  If it is shared with the current extent
2759          * we zero partial_cluster because we've reached the start of the
2760          * truncated/punched region and we're done removing blocks.
2761          */
2762         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2763                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2764                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2765                         ext4_free_blocks(handle, inode, NULL,
2766                                          EXT4_C2B(sbi, *partial_cluster),
2767                                          sbi->s_cluster_ratio,
2768                                          get_default_free_blocks_flags(inode));
2769                 }
2770                 *partial_cluster = 0;
2771         }
2772
2773         /* if this leaf is free, then we should
2774          * remove it from index block above */
2775         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2776                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2777
2778 out:
2779         return err;
2780 }
2781
2782 /*
2783  * ext4_ext_more_to_rm:
2784  * returns 1 if current index has to be freed (even partial)
2785  */
2786 static int
2787 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2788 {
2789         BUG_ON(path->p_idx == NULL);
2790
2791         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2792                 return 0;
2793
2794         /*
2795          * if truncate on deeper level happened, it wasn't partial,
2796          * so we have to consider current index for truncation
2797          */
2798         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2799                 return 0;
2800         return 1;
2801 }
2802
2803 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2804                           ext4_lblk_t end)
2805 {
2806         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2807         int depth = ext_depth(inode);
2808         struct ext4_ext_path *path = NULL;
2809         long long partial_cluster = 0;
2810         handle_t *handle;
2811         int i = 0, err = 0;
2812
2813         ext_debug("truncate since %u to %u\n", start, end);
2814
2815         /* probably first extent we're gonna free will be last in block */
2816         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2817         if (IS_ERR(handle))
2818                 return PTR_ERR(handle);
2819
2820 again:
2821         trace_ext4_ext_remove_space(inode, start, end, depth);
2822
2823         /*
2824          * Check if we are removing extents inside the extent tree. If that
2825          * is the case, we are going to punch a hole inside the extent tree
2826          * so we have to check whether we need to split the extent covering
2827          * the last block to remove so we can easily remove the part of it
2828          * in ext4_ext_rm_leaf().
2829          */
2830         if (end < EXT_MAX_BLOCKS - 1) {
2831                 struct ext4_extent *ex;
2832                 ext4_lblk_t ee_block, ex_end, lblk;
2833                 ext4_fsblk_t pblk;
2834
2835                 /* find extent for or closest extent to this block */
2836                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2837                 if (IS_ERR(path)) {
2838                         ext4_journal_stop(handle);
2839                         return PTR_ERR(path);
2840                 }
2841                 depth = ext_depth(inode);
2842                 /* Leaf not may not exist only if inode has no blocks at all */
2843                 ex = path[depth].p_ext;
2844                 if (!ex) {
2845                         if (depth) {
2846                                 EXT4_ERROR_INODE(inode,
2847                                                  "path[%d].p_hdr == NULL",
2848                                                  depth);
2849                                 err = -EIO;
2850                         }
2851                         goto out;
2852                 }
2853
2854                 ee_block = le32_to_cpu(ex->ee_block);
2855                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2856
2857                 /*
2858                  * See if the last block is inside the extent, if so split
2859                  * the extent at 'end' block so we can easily remove the
2860                  * tail of the first part of the split extent in
2861                  * ext4_ext_rm_leaf().
2862                  */
2863                 if (end >= ee_block && end < ex_end) {
2864
2865                         /*
2866                          * If we're going to split the extent, note that
2867                          * the cluster containing the block after 'end' is
2868                          * in use to avoid freeing it when removing blocks.
2869                          */
2870                         if (sbi->s_cluster_ratio > 1) {
2871                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2872                                 partial_cluster =
2873                                         -(long long) EXT4_B2C(sbi, pblk);
2874                         }
2875
2876                         /*
2877                          * Split the extent in two so that 'end' is the last
2878                          * block in the first new extent. Also we should not
2879                          * fail removing space due to ENOSPC so try to use
2880                          * reserved block if that happens.
2881                          */
2882                         err = ext4_force_split_extent_at(handle, inode, &path,
2883                                                          end + 1, 1);
2884                         if (err < 0)
2885                                 goto out;
2886
2887                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2888                         /*
2889                          * If there's an extent to the right its first cluster
2890                          * contains the immediate right boundary of the
2891                          * truncated/punched region.  Set partial_cluster to
2892                          * its negative value so it won't be freed if shared
2893                          * with the current extent.  The end < ee_block case
2894                          * is handled in ext4_ext_rm_leaf().
2895                          */
2896                         lblk = ex_end + 1;
2897                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2898                                                     &ex);
2899                         if (err)
2900                                 goto out;
2901                         if (pblk)
2902                                 partial_cluster =
2903                                         -(long long) EXT4_B2C(sbi, pblk);
2904                 }
2905         }
2906         /*
2907          * We start scanning from right side, freeing all the blocks
2908          * after i_size and walking into the tree depth-wise.
2909          */
2910         depth = ext_depth(inode);
2911         if (path) {
2912                 int k = i = depth;
2913                 while (--k > 0)
2914                         path[k].p_block =
2915                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2916         } else {
2917                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2918                                GFP_NOFS);
2919                 if (path == NULL) {
2920                         ext4_journal_stop(handle);
2921                         return -ENOMEM;
2922                 }
2923                 path[0].p_maxdepth = path[0].p_depth = depth;
2924                 path[0].p_hdr = ext_inode_hdr(inode);
2925                 i = 0;
2926
2927                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2928                         err = -EIO;
2929                         goto out;
2930                 }
2931         }
2932         err = 0;
2933
2934         while (i >= 0 && err == 0) {
2935                 if (i == depth) {
2936                         /* this is leaf block */
2937                         err = ext4_ext_rm_leaf(handle, inode, path,
2938                                                &partial_cluster, start,
2939                                                end);
2940                         /* root level has p_bh == NULL, brelse() eats this */
2941                         brelse(path[i].p_bh);
2942                         path[i].p_bh = NULL;
2943                         i--;
2944                         continue;
2945                 }
2946
2947                 /* this is index block */
2948                 if (!path[i].p_hdr) {
2949                         ext_debug("initialize header\n");
2950                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2951                 }
2952
2953                 if (!path[i].p_idx) {
2954                         /* this level hasn't been touched yet */
2955                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2956                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2957                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2958                                   path[i].p_hdr,
2959                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2960                 } else {
2961                         /* we were already here, see at next index */
2962                         path[i].p_idx--;
2963                 }
2964
2965                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2966                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2967                                 path[i].p_idx);
2968                 if (ext4_ext_more_to_rm(path + i)) {
2969                         struct buffer_head *bh;
2970                         /* go to the next level */
2971                         ext_debug("move to level %d (block %llu)\n",
2972                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2973                         memset(path + i + 1, 0, sizeof(*path));
2974                         bh = read_extent_tree_block(inode,
2975                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2976                                 EXT4_EX_NOCACHE);
2977                         if (IS_ERR(bh)) {
2978                                 /* should we reset i_size? */
2979                                 err = PTR_ERR(bh);
2980                                 break;
2981                         }
2982                         /* Yield here to deal with large extent trees.
2983                          * Should be a no-op if we did IO above. */
2984                         cond_resched();
2985                         if (WARN_ON(i + 1 > depth)) {
2986                                 err = -EIO;
2987                                 break;
2988                         }
2989                         path[i + 1].p_bh = bh;
2990
2991                         /* save actual number of indexes since this
2992                          * number is changed at the next iteration */
2993                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2994                         i++;
2995                 } else {
2996                         /* we finished processing this index, go up */
2997                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2998                                 /* index is empty, remove it;
2999                                  * handle must be already prepared by the
3000                                  * truncatei_leaf() */
3001                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3002                         }
3003                         /* root level has p_bh == NULL, brelse() eats this */
3004                         brelse(path[i].p_bh);
3005                         path[i].p_bh = NULL;
3006                         i--;
3007                         ext_debug("return to level %d\n", i);
3008                 }
3009         }
3010
3011         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3012                         partial_cluster, path->p_hdr->eh_entries);
3013
3014         /*
3015          * If we still have something in the partial cluster and we have removed
3016          * even the first extent, then we should free the blocks in the partial
3017          * cluster as well.  (This code will only run when there are no leaves
3018          * to the immediate left of the truncated/punched region.)
3019          */
3020         if (partial_cluster > 0 && err == 0) {
3021                 /* don't zero partial_cluster since it's not used afterwards */
3022                 ext4_free_blocks(handle, inode, NULL,
3023                                  EXT4_C2B(sbi, partial_cluster),
3024                                  sbi->s_cluster_ratio,
3025                                  get_default_free_blocks_flags(inode));
3026         }
3027
3028         /* TODO: flexible tree reduction should be here */
3029         if (path->p_hdr->eh_entries == 0) {
3030                 /*
3031                  * truncate to zero freed all the tree,
3032                  * so we need to correct eh_depth
3033                  */
3034                 err = ext4_ext_get_access(handle, inode, path);
3035                 if (err == 0) {
3036                         ext_inode_hdr(inode)->eh_depth = 0;
3037                         ext_inode_hdr(inode)->eh_max =
3038                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3039                         err = ext4_ext_dirty(handle, inode, path);
3040                 }
3041         }
3042 out:
3043         ext4_ext_drop_refs(path);
3044         kfree(path);
3045         path = NULL;
3046         if (err == -EAGAIN)
3047                 goto again;
3048         ext4_journal_stop(handle);
3049
3050         return err;
3051 }
3052
3053 /*
3054  * called at mount time
3055  */
3056 void ext4_ext_init(struct super_block *sb)
3057 {
3058         /*
3059          * possible initialization would be here
3060          */
3061
3062         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
3063 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3064                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3065 #ifdef AGGRESSIVE_TEST
3066                        ", aggressive tests"
3067 #endif
3068 #ifdef CHECK_BINSEARCH
3069                        ", check binsearch"
3070 #endif
3071 #ifdef EXTENTS_STATS
3072                        ", stats"
3073 #endif
3074                        "\n");
3075 #endif
3076 #ifdef EXTENTS_STATS
3077                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3078                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3079                 EXT4_SB(sb)->s_ext_max = 0;
3080 #endif
3081         }
3082 }
3083
3084 /*
3085  * called at umount time
3086  */
3087 void ext4_ext_release(struct super_block *sb)
3088 {
3089         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
3090                 return;
3091
3092 #ifdef EXTENTS_STATS
3093         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3094                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3095                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3096                         sbi->s_ext_blocks, sbi->s_ext_extents,
3097                         sbi->s_ext_blocks / sbi->s_ext_extents);
3098                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3099                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3100         }
3101 #endif
3102 }
3103
3104 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3105 {
3106         ext4_lblk_t  ee_block;
3107         ext4_fsblk_t ee_pblock;
3108         unsigned int ee_len;
3109
3110         ee_block  = le32_to_cpu(ex->ee_block);
3111         ee_len    = ext4_ext_get_actual_len(ex);
3112         ee_pblock = ext4_ext_pblock(ex);
3113
3114         if (ee_len == 0)
3115                 return 0;
3116
3117         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3118                                      EXTENT_STATUS_WRITTEN);
3119 }
3120
3121 /* FIXME!! we need to try to merge to left or right after zero-out  */
3122 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3123 {
3124         ext4_fsblk_t ee_pblock;
3125         unsigned int ee_len;
3126         int ret;
3127
3128         ee_len    = ext4_ext_get_actual_len(ex);
3129         ee_pblock = ext4_ext_pblock(ex);
3130
3131         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
3132         if (ret > 0)
3133                 ret = 0;
3134
3135         return ret;
3136 }
3137
3138 /*
3139  * ext4_split_extent_at() splits an extent at given block.
3140  *
3141  * @handle: the journal handle
3142  * @inode: the file inode
3143  * @path: the path to the extent
3144  * @split: the logical block where the extent is splitted.
3145  * @split_flags: indicates if the extent could be zeroout if split fails, and
3146  *               the states(init or unwritten) of new extents.
3147  * @flags: flags used to insert new extent to extent tree.
3148  *
3149  *
3150  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3151  * of which are deterimined by split_flag.
3152  *
3153  * There are two cases:
3154  *  a> the extent are splitted into two extent.
3155  *  b> split is not needed, and just mark the extent.
3156  *
3157  * return 0 on success.
3158  */
3159 static int ext4_split_extent_at(handle_t *handle,
3160                              struct inode *inode,
3161                              struct ext4_ext_path **ppath,
3162                              ext4_lblk_t split,
3163                              int split_flag,
3164                              int flags)
3165 {
3166         struct ext4_ext_path *path = *ppath;
3167         ext4_fsblk_t newblock;
3168         ext4_lblk_t ee_block;
3169         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3170         struct ext4_extent *ex2 = NULL;
3171         unsigned int ee_len, depth;
3172         int err = 0;
3173
3174         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3175                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3176
3177         ext_debug("ext4_split_extents_at: inode %lu, logical"
3178                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3179
3180         ext4_ext_show_leaf(inode, path);
3181
3182         depth = ext_depth(inode);
3183         ex = path[depth].p_ext;
3184         ee_block = le32_to_cpu(ex->ee_block);
3185         ee_len = ext4_ext_get_actual_len(ex);
3186         newblock = split - ee_block + ext4_ext_pblock(ex);
3187
3188         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3189         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3190                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3191                              EXT4_EXT_MARK_UNWRIT1 |
3192                              EXT4_EXT_MARK_UNWRIT2));
3193
3194         err = ext4_ext_get_access(handle, inode, path + depth);
3195         if (err)
3196                 goto out;
3197
3198         if (split == ee_block) {
3199                 /*
3200                  * case b: block @split is the block that the extent begins with
3201                  * then we just change the state of the extent, and splitting
3202                  * is not needed.
3203                  */
3204                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3205                         ext4_ext_mark_unwritten(ex);
3206                 else
3207                         ext4_ext_mark_initialized(ex);
3208
3209                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3210                         ext4_ext_try_to_merge(handle, inode, path, ex);
3211
3212                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3213                 goto out;
3214         }
3215
3216         /* case a */
3217         memcpy(&orig_ex, ex, sizeof(orig_ex));
3218         ex->ee_len = cpu_to_le16(split - ee_block);
3219         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3220                 ext4_ext_mark_unwritten(ex);
3221
3222         /*
3223          * path may lead to new leaf, not to original leaf any more
3224          * after ext4_ext_insert_extent() returns,
3225          */
3226         err = ext4_ext_dirty(handle, inode, path + depth);
3227         if (err)
3228                 goto fix_extent_len;
3229
3230         ex2 = &newex;
3231         ex2->ee_block = cpu_to_le32(split);
3232         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3233         ext4_ext_store_pblock(ex2, newblock);
3234         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3235                 ext4_ext_mark_unwritten(ex2);
3236
3237         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3238         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3239                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3240                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3241                                 err = ext4_ext_zeroout(inode, ex2);
3242                                 zero_ex.ee_block = ex2->ee_block;
3243                                 zero_ex.ee_len = cpu_to_le16(
3244                                                 ext4_ext_get_actual_len(ex2));
3245                                 ext4_ext_store_pblock(&zero_ex,
3246                                                       ext4_ext_pblock(ex2));
3247                         } else {
3248                                 err = ext4_ext_zeroout(inode, ex);
3249                                 zero_ex.ee_block = ex->ee_block;
3250                                 zero_ex.ee_len = cpu_to_le16(
3251                                                 ext4_ext_get_actual_len(ex));
3252                                 ext4_ext_store_pblock(&zero_ex,
3253                                                       ext4_ext_pblock(ex));
3254                         }
3255                 } else {
3256                         err = ext4_ext_zeroout(inode, &orig_ex);
3257                         zero_ex.ee_block = orig_ex.ee_block;
3258                         zero_ex.ee_len = cpu_to_le16(
3259                                                 ext4_ext_get_actual_len(&orig_ex));
3260                         ext4_ext_store_pblock(&zero_ex,
3261                                               ext4_ext_pblock(&orig_ex));
3262                 }
3263
3264                 if (err)
3265                         goto fix_extent_len;
3266                 /* update the extent length and mark as initialized */
3267                 ex->ee_len = cpu_to_le16(ee_len);
3268                 ext4_ext_try_to_merge(handle, inode, path, ex);
3269                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3270                 if (err)
3271                         goto fix_extent_len;
3272
3273                 /* update extent status tree */
3274                 err = ext4_zeroout_es(inode, &zero_ex);
3275
3276                 goto out;
3277         } else if (err)
3278                 goto fix_extent_len;
3279
3280 out:
3281         ext4_ext_show_leaf(inode, path);
3282         return err;
3283
3284 fix_extent_len:
3285         ex->ee_len = orig_ex.ee_len;
3286         ext4_ext_dirty(handle, inode, path + path->p_depth);
3287         return err;
3288 }
3289
3290 /*
3291  * ext4_split_extents() splits an extent and mark extent which is covered
3292  * by @map as split_flags indicates
3293  *
3294  * It may result in splitting the extent into multiple extents (up to three)
3295  * There are three possibilities:
3296  *   a> There is no split required
3297  *   b> Splits in two extents: Split is happening at either end of the extent
3298  *   c> Splits in three extents: Somone is splitting in middle of the extent
3299  *
3300  */
3301 static int ext4_split_extent(handle_t *handle,
3302                               struct inode *inode,
3303                               struct ext4_ext_path **ppath,
3304                               struct ext4_map_blocks *map,
3305                               int split_flag,
3306                               int flags)
3307 {
3308         struct ext4_ext_path *path = *ppath;
3309         ext4_lblk_t ee_block;
3310         struct ext4_extent *ex;
3311         unsigned int ee_len, depth;
3312         int err = 0;
3313         int unwritten;
3314         int split_flag1, flags1;
3315         int allocated = map->m_len;
3316
3317         depth = ext_depth(inode);
3318         ex = path[depth].p_ext;
3319         ee_block = le32_to_cpu(ex->ee_block);
3320         ee_len = ext4_ext_get_actual_len(ex);
3321         unwritten = ext4_ext_is_unwritten(ex);
3322
3323         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3324                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3325                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3326                 if (unwritten)
3327                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3328                                        EXT4_EXT_MARK_UNWRIT2;
3329                 if (split_flag & EXT4_EXT_DATA_VALID2)
3330                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3331                 err = ext4_split_extent_at(handle, inode, ppath,
3332                                 map->m_lblk + map->m_len, split_flag1, flags1);
3333                 if (err)
3334                         goto out;
3335         } else {
3336                 allocated = ee_len - (map->m_lblk - ee_block);
3337         }
3338         /*
3339          * Update path is required because previous ext4_split_extent_at() may
3340          * result in split of original leaf or extent zeroout.
3341          */
3342         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3343         if (IS_ERR(path))
3344                 return PTR_ERR(path);
3345         depth = ext_depth(inode);
3346         ex = path[depth].p_ext;
3347         if (!ex) {
3348                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3349                                  (unsigned long) map->m_lblk);
3350                 return -EIO;
3351         }
3352         unwritten = ext4_ext_is_unwritten(ex);
3353         split_flag1 = 0;
3354
3355         if (map->m_lblk >= ee_block) {
3356                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3357                 if (unwritten) {
3358                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3359                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3360                                                      EXT4_EXT_MARK_UNWRIT2);
3361                 }
3362                 err = ext4_split_extent_at(handle, inode, ppath,
3363                                 map->m_lblk, split_flag1, flags);
3364                 if (err)
3365                         goto out;
3366         }
3367
3368         ext4_ext_show_leaf(inode, path);
3369 out:
3370         return err ? err : allocated;
3371 }
3372
3373 /*
3374  * This function is called by ext4_ext_map_blocks() if someone tries to write
3375  * to an unwritten extent. It may result in splitting the unwritten
3376  * extent into multiple extents (up to three - one initialized and two
3377  * unwritten).
3378  * There are three possibilities:
3379  *   a> There is no split required: Entire extent should be initialized
3380  *   b> Splits in two extents: Write is happening at either end of the extent
3381  *   c> Splits in three extents: Somone is writing in middle of the extent
3382  *
3383  * Pre-conditions:
3384  *  - The extent pointed to by 'path' is unwritten.
3385  *  - The extent pointed to by 'path' contains a superset
3386  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3387  *
3388  * Post-conditions on success:
3389  *  - the returned value is the number of blocks beyond map->l_lblk
3390  *    that are allocated and initialized.
3391  *    It is guaranteed to be >= map->m_len.
3392  */
3393 static int ext4_ext_convert_to_initialized(handle_t *handle,
3394                                            struct inode *inode,
3395                                            struct ext4_map_blocks *map,
3396                                            struct ext4_ext_path **ppath,
3397                                            int flags)
3398 {
3399         struct ext4_ext_path *path = *ppath;
3400         struct ext4_sb_info *sbi;
3401         struct ext4_extent_header *eh;
3402         struct ext4_map_blocks split_map;
3403         struct ext4_extent zero_ex;
3404         struct ext4_extent *ex, *abut_ex;
3405         ext4_lblk_t ee_block, eof_block;
3406         unsigned int ee_len, depth, map_len = map->m_len;
3407         int allocated = 0, max_zeroout = 0;
3408         int err = 0;
3409         int split_flag = 0;
3410
3411         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3412                 "block %llu, max_blocks %u\n", inode->i_ino,
3413                 (unsigned long long)map->m_lblk, map_len);
3414
3415         sbi = EXT4_SB(inode->i_sb);
3416         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3417                 inode->i_sb->s_blocksize_bits;
3418         if (eof_block < map->m_lblk + map_len)
3419                 eof_block = map->m_lblk + map_len;
3420
3421         depth = ext_depth(inode);
3422         eh = path[depth].p_hdr;
3423         ex = path[depth].p_ext;
3424         ee_block = le32_to_cpu(ex->ee_block);
3425         ee_len = ext4_ext_get_actual_len(ex);
3426         zero_ex.ee_len = 0;
3427
3428         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3429
3430         /* Pre-conditions */
3431         BUG_ON(!ext4_ext_is_unwritten(ex));
3432         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3433
3434         /*
3435          * Attempt to transfer newly initialized blocks from the currently
3436          * unwritten extent to its neighbor. This is much cheaper
3437          * than an insertion followed by a merge as those involve costly
3438          * memmove() calls. Transferring to the left is the common case in
3439          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3440          * followed by append writes.
3441          *
3442          * Limitations of the current logic:
3443          *  - L1: we do not deal with writes covering the whole extent.
3444          *    This would require removing the extent if the transfer
3445          *    is possible.
3446          *  - L2: we only attempt to merge with an extent stored in the
3447          *    same extent tree node.
3448          */
3449         if ((map->m_lblk == ee_block) &&
3450                 /* See if we can merge left */
3451                 (map_len < ee_len) &&           /*L1*/
3452                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3453                 ext4_lblk_t prev_lblk;
3454                 ext4_fsblk_t prev_pblk, ee_pblk;
3455                 unsigned int prev_len;
3456
3457                 abut_ex = ex - 1;
3458                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3459                 prev_len = ext4_ext_get_actual_len(abut_ex);
3460                 prev_pblk = ext4_ext_pblock(abut_ex);
3461                 ee_pblk = ext4_ext_pblock(ex);
3462
3463                 /*
3464                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3465                  * upon those conditions:
3466                  * - C1: abut_ex is initialized,
3467                  * - C2: abut_ex is logically abutting ex,
3468                  * - C3: abut_ex is physically abutting ex,
3469                  * - C4: abut_ex can receive the additional blocks without
3470                  *   overflowing the (initialized) length limit.
3471                  */
3472                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3473                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3474                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3475                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3476                         err = ext4_ext_get_access(handle, inode, path + depth);
3477                         if (err)
3478                                 goto out;
3479
3480                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3481                                 map, ex, abut_ex);
3482
3483                         /* Shift the start of ex by 'map_len' blocks */
3484                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3485                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3486                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3487                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3488
3489                         /* Extend abut_ex by 'map_len' blocks */
3490                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3491
3492                         /* Result: number of initialized blocks past m_lblk */
3493                         allocated = map_len;
3494                 }
3495         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3496                    (map_len < ee_len) &&        /*L1*/
3497                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3498                 /* See if we can merge right */
3499                 ext4_lblk_t next_lblk;
3500                 ext4_fsblk_t next_pblk, ee_pblk;
3501                 unsigned int next_len;
3502
3503                 abut_ex = ex + 1;
3504                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3505                 next_len = ext4_ext_get_actual_len(abut_ex);
3506                 next_pblk = ext4_ext_pblock(abut_ex);
3507                 ee_pblk = ext4_ext_pblock(ex);
3508
3509                 /*
3510                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3511                  * upon those conditions:
3512                  * - C1: abut_ex is initialized,
3513                  * - C2: abut_ex is logically abutting ex,
3514                  * - C3: abut_ex is physically abutting ex,
3515                  * - C4: abut_ex can receive the additional blocks without
3516                  *   overflowing the (initialized) length limit.
3517                  */
3518                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3519                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3520                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3521                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3522                         err = ext4_ext_get_access(handle, inode, path + depth);
3523                         if (err)
3524                                 goto out;
3525
3526                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3527                                 map, ex, abut_ex);
3528
3529                         /* Shift the start of abut_ex by 'map_len' blocks */
3530                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3531                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3532                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3533                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3534
3535                         /* Extend abut_ex by 'map_len' blocks */
3536                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3537
3538                         /* Result: number of initialized blocks past m_lblk */
3539                         allocated = map_len;
3540                 }
3541         }
3542         if (allocated) {
3543                 /* Mark the block containing both extents as dirty */
3544                 ext4_ext_dirty(handle, inode, path + depth);
3545
3546                 /* Update path to point to the right extent */
3547                 path[depth].p_ext = abut_ex;
3548                 goto out;
3549         } else
3550                 allocated = ee_len - (map->m_lblk - ee_block);
3551
3552         WARN_ON(map->m_lblk < ee_block);
3553         /*
3554          * It is safe to convert extent to initialized via explicit
3555          * zeroout only if extent is fully inside i_size or new_size.
3556          */
3557         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3558
3559         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3560                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3561                         (inode->i_sb->s_blocksize_bits - 10);
3562
3563         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3564         if (max_zeroout && (ee_len <= max_zeroout)) {
3565                 err = ext4_ext_zeroout(inode, ex);
3566                 if (err)
3567                         goto out;
3568                 zero_ex.ee_block = ex->ee_block;
3569                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3570                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3571
3572                 err = ext4_ext_get_access(handle, inode, path + depth);
3573                 if (err)
3574                         goto out;
3575                 ext4_ext_mark_initialized(ex);
3576                 ext4_ext_try_to_merge(handle, inode, path, ex);
3577                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3578                 goto out;
3579         }
3580
3581         /*
3582          * four cases:
3583          * 1. split the extent into three extents.
3584          * 2. split the extent into two extents, zeroout the first half.
3585          * 3. split the extent into two extents, zeroout the second half.
3586          * 4. split the extent into two extents with out zeroout.
3587          */
3588         split_map.m_lblk = map->m_lblk;
3589         split_map.m_len = map->m_len;
3590
3591         if (max_zeroout && (allocated > map->m_len)) {
3592                 if (allocated <= max_zeroout) {
3593                         /* case 3 */
3594                         zero_ex.ee_block =
3595                                          cpu_to_le32(map->m_lblk);
3596                         zero_ex.ee_len = cpu_to_le16(allocated);
3597                         ext4_ext_store_pblock(&zero_ex,
3598                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3599                         err = ext4_ext_zeroout(inode, &zero_ex);
3600                         if (err)
3601                                 goto out;
3602                         split_map.m_lblk = map->m_lblk;
3603                         split_map.m_len = allocated;
3604                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3605                         /* case 2 */
3606                         if (map->m_lblk != ee_block) {
3607                                 zero_ex.ee_block = ex->ee_block;
3608                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3609                                                         ee_block);
3610                                 ext4_ext_store_pblock(&zero_ex,
3611                                                       ext4_ext_pblock(ex));
3612                                 err = ext4_ext_zeroout(inode, &zero_ex);
3613                                 if (err)
3614                                         goto out;
3615                         }
3616
3617                         split_map.m_lblk = ee_block;
3618                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3619                         allocated = map->m_len;
3620                 }
3621         }
3622
3623         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3624                                 flags);
3625         if (err > 0)
3626                 err = 0;
3627 out:
3628         /* If we have gotten a failure, don't zero out status tree */
3629         if (!err)
3630                 err = ext4_zeroout_es(inode, &zero_ex);
3631         return err ? err : allocated;
3632 }
3633
3634 /*
3635  * This function is called by ext4_ext_map_blocks() from
3636  * ext4_get_blocks_dio_write() when DIO to write
3637  * to an unwritten extent.
3638  *
3639  * Writing to an unwritten extent may result in splitting the unwritten
3640  * extent into multiple initialized/unwritten extents (up to three)
3641  * There are three possibilities:
3642  *   a> There is no split required: Entire extent should be unwritten
3643  *   b> Splits in two extents: Write is happening at either end of the extent
3644  *   c> Splits in three extents: Somone is writing in middle of the extent
3645  *
3646  * This works the same way in the case of initialized -> unwritten conversion.
3647  *
3648  * One of more index blocks maybe needed if the extent tree grow after
3649  * the unwritten extent split. To prevent ENOSPC occur at the IO
3650  * complete, we need to split the unwritten extent before DIO submit
3651  * the IO. The unwritten extent called at this time will be split
3652  * into three unwritten extent(at most). After IO complete, the part
3653  * being filled will be convert to initialized by the end_io callback function
3654  * via ext4_convert_unwritten_extents().
3655  *
3656  * Returns the size of unwritten extent to be written on success.
3657  */
3658 static int ext4_split_convert_extents(handle_t *handle,
3659                                         struct inode *inode,
3660                                         struct ext4_map_blocks *map,
3661                                         struct ext4_ext_path **ppath,
3662                                         int flags)
3663 {
3664         struct ext4_ext_path *path = *ppath;
3665         ext4_lblk_t eof_block;
3666         ext4_lblk_t ee_block;
3667         struct ext4_extent *ex;
3668         unsigned int ee_len;
3669         int split_flag = 0, depth;
3670
3671         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3672                   __func__, inode->i_ino,
3673                   (unsigned long long)map->m_lblk, map->m_len);
3674
3675         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3676                 inode->i_sb->s_blocksize_bits;
3677         if (eof_block < map->m_lblk + map->m_len)
3678                 eof_block = map->m_lblk + map->m_len;
3679         /*
3680          * It is safe to convert extent to initialized via explicit
3681          * zeroout only if extent is fully insde i_size or new_size.
3682          */
3683         depth = ext_depth(inode);
3684         ex = path[depth].p_ext;
3685         ee_block = le32_to_cpu(ex->ee_block);
3686         ee_len = ext4_ext_get_actual_len(ex);
3687
3688         /* Convert to unwritten */
3689         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3690                 split_flag |= EXT4_EXT_DATA_VALID1;
3691         /* Convert to initialized */
3692         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3693                 split_flag |= ee_block + ee_len <= eof_block ?
3694                               EXT4_EXT_MAY_ZEROOUT : 0;
3695                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3696         }
3697         flags |= EXT4_GET_BLOCKS_PRE_IO;
3698         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3699 }
3700
3701 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3702                                                 struct inode *inode,
3703                                                 struct ext4_map_blocks *map,
3704                                                 struct ext4_ext_path **ppath)
3705 {
3706         struct ext4_ext_path *path = *ppath;
3707         struct ext4_extent *ex;
3708         ext4_lblk_t ee_block;
3709         unsigned int ee_len;
3710         int depth;
3711         int err = 0;
3712
3713         depth = ext_depth(inode);
3714         ex = path[depth].p_ext;
3715         ee_block = le32_to_cpu(ex->ee_block);
3716         ee_len = ext4_ext_get_actual_len(ex);
3717
3718         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3719                 "block %llu, max_blocks %u\n", inode->i_ino,
3720                   (unsigned long long)ee_block, ee_len);
3721
3722         /* If extent is larger than requested it is a clear sign that we still
3723          * have some extent state machine issues left. So extent_split is still
3724          * required.
3725          * TODO: Once all related issues will be fixed this situation should be
3726          * illegal.
3727          */
3728         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3729 #ifdef EXT4_DEBUG
3730                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3731                              " len %u; IO logical block %llu, len %u\n",
3732                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3733                              (unsigned long long)map->m_lblk, map->m_len);
3734 #endif
3735                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3736                                                  EXT4_GET_BLOCKS_CONVERT);
3737                 if (err < 0)
3738                         return err;
3739                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3740                 if (IS_ERR(path))
3741                         return PTR_ERR(path);
3742                 depth = ext_depth(inode);
3743                 ex = path[depth].p_ext;
3744         }
3745
3746         err = ext4_ext_get_access(handle, inode, path + depth);
3747         if (err)
3748                 goto out;
3749         /* first mark the extent as initialized */
3750         ext4_ext_mark_initialized(ex);
3751
3752         /* note: ext4_ext_correct_indexes() isn't needed here because
3753          * borders are not changed
3754          */
3755         ext4_ext_try_to_merge(handle, inode, path, ex);
3756
3757         /* Mark modified extent as dirty */
3758         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3759 out:
3760         ext4_ext_show_leaf(inode, path);
3761         return err;
3762 }
3763
3764 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3765                         sector_t block, int count)
3766 {
3767         int i;
3768         for (i = 0; i < count; i++)
3769                 unmap_underlying_metadata(bdev, block + i);
3770 }
3771
3772 /*
3773  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3774  */
3775 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3776                               ext4_lblk_t lblk,
3777                               struct ext4_ext_path *path,
3778                               unsigned int len)
3779 {
3780         int i, depth;
3781         struct ext4_extent_header *eh;
3782         struct ext4_extent *last_ex;
3783
3784         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3785                 return 0;
3786
3787         depth = ext_depth(inode);
3788         eh = path[depth].p_hdr;
3789
3790         /*
3791          * We're going to remove EOFBLOCKS_FL entirely in future so we
3792          * do not care for this case anymore. Simply remove the flag
3793          * if there are no extents.
3794          */
3795         if (unlikely(!eh->eh_entries))
3796                 goto out;
3797         last_ex = EXT_LAST_EXTENT(eh);
3798         /*
3799          * We should clear the EOFBLOCKS_FL flag if we are writing the
3800          * last block in the last extent in the file.  We test this by
3801          * first checking to see if the caller to
3802          * ext4_ext_get_blocks() was interested in the last block (or
3803          * a block beyond the last block) in the current extent.  If
3804          * this turns out to be false, we can bail out from this
3805          * function immediately.
3806          */
3807         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3808             ext4_ext_get_actual_len(last_ex))
3809                 return 0;
3810         /*
3811          * If the caller does appear to be planning to write at or
3812          * beyond the end of the current extent, we then test to see
3813          * if the current extent is the last extent in the file, by
3814          * checking to make sure it was reached via the rightmost node
3815          * at each level of the tree.
3816          */
3817         for (i = depth-1; i >= 0; i--)
3818                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3819                         return 0;
3820 out:
3821         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3822         return ext4_mark_inode_dirty(handle, inode);
3823 }
3824
3825 /**
3826  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3827  *
3828  * Return 1 if there is a delalloc block in the range, otherwise 0.
3829  */
3830 int ext4_find_delalloc_range(struct inode *inode,
3831                              ext4_lblk_t lblk_start,
3832                              ext4_lblk_t lblk_end)
3833 {
3834         struct extent_status es;
3835
3836         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3837         if (es.es_len == 0)
3838                 return 0; /* there is no delay extent in this tree */
3839         else if (es.es_lblk <= lblk_start &&
3840                  lblk_start < es.es_lblk + es.es_len)
3841                 return 1;
3842         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3843                 return 1;
3844         else
3845                 return 0;
3846 }
3847
3848 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3849 {
3850         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3851         ext4_lblk_t lblk_start, lblk_end;
3852         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3853         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3854
3855         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3856 }
3857
3858 /**
3859  * Determines how many complete clusters (out of those specified by the 'map')
3860  * are under delalloc and were reserved quota for.
3861  * This function is called when we are writing out the blocks that were
3862  * originally written with their allocation delayed, but then the space was
3863  * allocated using fallocate() before the delayed allocation could be resolved.
3864  * The cases to look for are:
3865  * ('=' indicated delayed allocated blocks
3866  *  '-' indicates non-delayed allocated blocks)
3867  * (a) partial clusters towards beginning and/or end outside of allocated range
3868  *     are not delalloc'ed.
3869  *      Ex:
3870  *      |----c---=|====c====|====c====|===-c----|
3871  *               |++++++ allocated ++++++|
3872  *      ==> 4 complete clusters in above example
3873  *
3874  * (b) partial cluster (outside of allocated range) towards either end is
3875  *     marked for delayed allocation. In this case, we will exclude that
3876  *     cluster.
3877  *      Ex:
3878  *      |----====c========|========c========|
3879  *           |++++++ allocated ++++++|
3880  *      ==> 1 complete clusters in above example
3881  *
3882  *      Ex:
3883  *      |================c================|
3884  *            |++++++ allocated ++++++|
3885  *      ==> 0 complete clusters in above example
3886  *
3887  * The ext4_da_update_reserve_space will be called only if we
3888  * determine here that there were some "entire" clusters that span
3889  * this 'allocated' range.
3890  * In the non-bigalloc case, this function will just end up returning num_blks
3891  * without ever calling ext4_find_delalloc_range.
3892  */
3893 static unsigned int
3894 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3895                            unsigned int num_blks)
3896 {
3897         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3898         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3899         ext4_lblk_t lblk_from, lblk_to, c_offset;
3900         unsigned int allocated_clusters = 0;
3901
3902         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3903         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3904
3905         /* max possible clusters for this allocation */
3906         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3907
3908         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3909
3910         /* Check towards left side */
3911         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3912         if (c_offset) {
3913                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3914                 lblk_to = lblk_from + c_offset - 1;
3915
3916                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3917                         allocated_clusters--;
3918         }
3919
3920         /* Now check towards right. */
3921         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3922         if (allocated_clusters && c_offset) {
3923                 lblk_from = lblk_start + num_blks;
3924                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3925
3926                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3927                         allocated_clusters--;
3928         }
3929
3930         return allocated_clusters;
3931 }
3932
3933 static int
3934 convert_initialized_extent(handle_t *handle, struct inode *inode,
3935                            struct ext4_map_blocks *map,
3936                            struct ext4_ext_path **ppath, int flags,
3937                            unsigned int allocated, ext4_fsblk_t newblock)
3938 {
3939         struct ext4_ext_path *path = *ppath;
3940         struct ext4_extent *ex;
3941         ext4_lblk_t ee_block;
3942         unsigned int ee_len;
3943         int depth;
3944         int err = 0;
3945
3946         /*
3947          * Make sure that the extent is no bigger than we support with
3948          * unwritten extent
3949          */
3950         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3951                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3952
3953         depth = ext_depth(inode);
3954         ex = path[depth].p_ext;
3955         ee_block = le32_to_cpu(ex->ee_block);
3956         ee_len = ext4_ext_get_actual_len(ex);
3957
3958         ext_debug("%s: inode %lu, logical"
3959                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3960                   (unsigned long long)ee_block, ee_len);
3961
3962         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3963                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3964                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3965                 if (err < 0)
3966                         return err;
3967                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3968                 if (IS_ERR(path))
3969                         return PTR_ERR(path);
3970                 depth = ext_depth(inode);
3971                 ex = path[depth].p_ext;
3972                 if (!ex) {
3973                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3974                                          (unsigned long) map->m_lblk);
3975                         return -EIO;
3976                 }
3977         }
3978
3979         err = ext4_ext_get_access(handle, inode, path + depth);
3980         if (err)
3981                 return err;
3982         /* first mark the extent as unwritten */
3983         ext4_ext_mark_unwritten(ex);
3984
3985         /* note: ext4_ext_correct_indexes() isn't needed here because
3986          * borders are not changed
3987          */
3988         ext4_ext_try_to_merge(handle, inode, path, ex);
3989
3990         /* Mark modified extent as dirty */
3991         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3992         if (err)
3993                 return err;
3994         ext4_ext_show_leaf(inode, path);
3995
3996         ext4_update_inode_fsync_trans(handle, inode, 1);
3997         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3998         if (err)
3999                 return err;
4000         map->m_flags |= EXT4_MAP_UNWRITTEN;
4001         if (allocated > map->m_len)
4002                 allocated = map->m_len;
4003         map->m_len = allocated;
4004         return allocated;
4005 }
4006
4007 static int
4008 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4009                         struct ext4_map_blocks *map,
4010                         struct ext4_ext_path **ppath, int flags,
4011                         unsigned int allocated, ext4_fsblk_t newblock)
4012 {
4013         struct ext4_ext_path *path = *ppath;
4014         int ret = 0;
4015         int err = 0;
4016         ext4_io_end_t *io = ext4_inode_aio(inode);
4017
4018         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4019                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4020                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4021                   flags, allocated);
4022         ext4_ext_show_leaf(inode, path);
4023
4024         /*
4025          * When writing into unwritten space, we should not fail to
4026          * allocate metadata blocks for the new extent block if needed.
4027          */
4028         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4029
4030         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4031                                                     allocated, newblock);
4032
4033         /* get_block() before submit the IO, split the extent */
4034         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4035                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4036                                          flags | EXT4_GET_BLOCKS_CONVERT);
4037                 if (ret <= 0)
4038                         goto out;
4039                 /*
4040                  * Flag the inode(non aio case) or end_io struct (aio case)
4041                  * that this IO needs to conversion to written when IO is
4042                  * completed
4043                  */
4044                 if (io)
4045                         ext4_set_io_unwritten_flag(inode, io);
4046                 else
4047                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4048                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4049                 goto out;
4050         }
4051         /* IO end_io complete, convert the filled extent to written */
4052         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4053                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4054                                                            ppath);
4055                 if (ret >= 0) {
4056                         ext4_update_inode_fsync_trans(handle, inode, 1);
4057                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4058                                                  path, map->m_len);
4059                 } else
4060                         err = ret;
4061                 map->m_flags |= EXT4_MAP_MAPPED;
4062                 map->m_pblk = newblock;
4063                 if (allocated > map->m_len)
4064                         allocated = map->m_len;
4065                 map->m_len = allocated;
4066                 goto out2;
4067         }
4068         /* buffered IO case */
4069         /*
4070          * repeat fallocate creation request
4071          * we already have an unwritten extent
4072          */
4073         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4074                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4075                 goto map_out;
4076         }
4077
4078         /* buffered READ or buffered write_begin() lookup */
4079         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4080                 /*
4081                  * We have blocks reserved already.  We
4082                  * return allocated blocks so that delalloc
4083                  * won't do block reservation for us.  But
4084                  * the buffer head will be unmapped so that
4085                  * a read from the block returns 0s.
4086                  */
4087                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4088                 goto out1;
4089         }
4090
4091         /* buffered write, writepage time, convert*/
4092         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4093         if (ret >= 0)
4094                 ext4_update_inode_fsync_trans(handle, inode, 1);
4095 out:
4096         if (ret <= 0) {
4097                 err = ret;
4098                 goto out2;
4099         } else
4100                 allocated = ret;
4101         map->m_flags |= EXT4_MAP_NEW;
4102         /*
4103          * if we allocated more blocks than requested
4104          * we need to make sure we unmap the extra block
4105          * allocated. The actual needed block will get
4106          * unmapped later when we find the buffer_head marked
4107          * new.
4108          */
4109         if (allocated > map->m_len) {
4110                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
4111                                         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_io_end_t *io = ext4_inode_aio(inode);
4285         ext4_lblk_t cluster_offset;
4286         int set_unwritten = 0;
4287         bool map_from_cluster = false;
4288
4289         ext_debug("blocks %u/%u requested for inode %lu\n",
4290                   map->m_lblk, map->m_len, inode->i_ino);
4291         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4292
4293         /* find extent for this block */
4294         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4295         if (IS_ERR(path)) {
4296                 err = PTR_ERR(path);
4297                 path = NULL;
4298                 goto out2;
4299         }
4300
4301         depth = ext_depth(inode);
4302
4303         /*
4304          * consistent leaf must not be empty;
4305          * this situation is possible, though, _during_ tree modification;
4306          * this is why assert can't be put in ext4_find_extent()
4307          */
4308         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4309                 EXT4_ERROR_INODE(inode, "bad extent address "
4310                                  "lblock: %lu, depth: %d pblock %lld",
4311                                  (unsigned long) map->m_lblk, depth,
4312                                  path[depth].p_block);
4313                 err = -EIO;
4314                 goto out2;
4315         }
4316
4317         ex = path[depth].p_ext;
4318         if (ex) {
4319                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4320                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4321                 unsigned short ee_len;
4322
4323
4324                 /*
4325                  * unwritten extents are treated as holes, except that
4326                  * we split out initialized portions during a write.
4327                  */
4328                 ee_len = ext4_ext_get_actual_len(ex);
4329
4330                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4331
4332                 /* if found extent covers block, simply return it */
4333                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4334                         newblock = map->m_lblk - ee_block + ee_start;
4335                         /* number of remaining blocks in the extent */
4336                         allocated = ee_len - (map->m_lblk - ee_block);
4337                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4338                                   ee_block, ee_len, newblock);
4339
4340                         /*
4341                          * If the extent is initialized check whether the
4342                          * caller wants to convert it to unwritten.
4343                          */
4344                         if ((!ext4_ext_is_unwritten(ex)) &&
4345                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4346                                 allocated = convert_initialized_extent(
4347                                                 handle, inode, map, &path,
4348                                                 flags, allocated, newblock);
4349                                 goto out2;
4350                         } else if (!ext4_ext_is_unwritten(ex))
4351                                 goto out;
4352
4353                         ret = ext4_ext_handle_unwritten_extents(
4354                                 handle, inode, map, &path, flags,
4355                                 allocated, newblock);
4356                         if (ret < 0)
4357                                 err = ret;
4358                         else
4359                                 allocated = ret;
4360                         goto out2;
4361                 }
4362         }
4363
4364         /*
4365          * requested block isn't allocated yet;
4366          * we couldn't try to create block if create flag is zero
4367          */
4368         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4369                 /*
4370                  * put just found gap into cache to speed up
4371                  * subsequent requests
4372                  */
4373                 ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4374                 goto out2;
4375         }
4376
4377         /*
4378          * Okay, we need to do block allocation.
4379          */
4380         newex.ee_block = cpu_to_le32(map->m_lblk);
4381         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4382
4383         /*
4384          * If we are doing bigalloc, check to see if the extent returned
4385          * by ext4_find_extent() implies a cluster we can use.
4386          */
4387         if (cluster_offset && ex &&
4388             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4389                 ar.len = allocated = map->m_len;
4390                 newblock = map->m_pblk;
4391                 map_from_cluster = true;
4392                 goto got_allocated_blocks;
4393         }
4394
4395         /* find neighbour allocated blocks */
4396         ar.lleft = map->m_lblk;
4397         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4398         if (err)
4399                 goto out2;
4400         ar.lright = map->m_lblk;
4401         ex2 = NULL;
4402         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4403         if (err)
4404                 goto out2;
4405
4406         /* Check if the extent after searching to the right implies a
4407          * cluster we can use. */
4408         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4409             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4410                 ar.len = allocated = map->m_len;
4411                 newblock = map->m_pblk;
4412                 map_from_cluster = true;
4413                 goto got_allocated_blocks;
4414         }
4415
4416         /*
4417          * See if request is beyond maximum number of blocks we can have in
4418          * a single extent. For an initialized extent this limit is
4419          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4420          * EXT_UNWRITTEN_MAX_LEN.
4421          */
4422         if (map->m_len > EXT_INIT_MAX_LEN &&
4423             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4424                 map->m_len = EXT_INIT_MAX_LEN;
4425         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4426                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4427                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4428
4429         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4430         newex.ee_len = cpu_to_le16(map->m_len);
4431         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4432         if (err)
4433                 allocated = ext4_ext_get_actual_len(&newex);
4434         else
4435                 allocated = map->m_len;
4436
4437         /* allocate new block */
4438         ar.inode = inode;
4439         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4440         ar.logical = map->m_lblk;
4441         /*
4442          * We calculate the offset from the beginning of the cluster
4443          * for the logical block number, since when we allocate a
4444          * physical cluster, the physical block should start at the
4445          * same offset from the beginning of the cluster.  This is
4446          * needed so that future calls to get_implied_cluster_alloc()
4447          * work correctly.
4448          */
4449         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4450         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4451         ar.goal -= offset;
4452         ar.logical -= offset;
4453         if (S_ISREG(inode->i_mode))
4454                 ar.flags = EXT4_MB_HINT_DATA;
4455         else
4456                 /* disable in-core preallocation for non-regular files */
4457                 ar.flags = 0;
4458         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4459                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4460         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4461                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4462         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4463         if (!newblock)
4464                 goto out2;
4465         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4466                   ar.goal, newblock, allocated);
4467         free_on_err = 1;
4468         allocated_clusters = ar.len;
4469         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4470         if (ar.len > allocated)
4471                 ar.len = allocated;
4472
4473 got_allocated_blocks:
4474         /* try to insert new extent into found leaf and return */
4475         ext4_ext_store_pblock(&newex, newblock + offset);
4476         newex.ee_len = cpu_to_le16(ar.len);
4477         /* Mark unwritten */
4478         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4479                 ext4_ext_mark_unwritten(&newex);
4480                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4481                 /*
4482                  * io_end structure was created for every IO write to an
4483                  * unwritten extent. To avoid unnecessary conversion,
4484                  * here we flag the IO that really needs the conversion.
4485                  * For non asycn direct IO case, flag the inode state
4486                  * that we need to perform conversion when IO is done.
4487                  */
4488                 if (flags & EXT4_GET_BLOCKS_PRE_IO)
4489                         set_unwritten = 1;
4490         }
4491
4492         err = 0;
4493         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4494                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4495                                          path, ar.len);
4496         if (!err)
4497                 err = ext4_ext_insert_extent(handle, inode, &path,
4498                                              &newex, flags);
4499
4500         if (!err && set_unwritten) {
4501                 if (io)
4502                         ext4_set_io_unwritten_flag(inode, io);
4503                 else
4504                         ext4_set_inode_state(inode,
4505                                              EXT4_STATE_DIO_UNWRITTEN);
4506         }
4507
4508         if (err && free_on_err) {
4509                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4510                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4511                 /* free data blocks we just allocated */
4512                 /* not a good idea to call discard here directly,
4513                  * but otherwise we'd need to call it every free() */
4514                 ext4_discard_preallocations(inode);
4515                 ext4_free_blocks(handle, inode, NULL, newblock,
4516                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4517                 goto out2;
4518         }
4519
4520         /* previous routine could use block we allocated */
4521         newblock = ext4_ext_pblock(&newex);
4522         allocated = ext4_ext_get_actual_len(&newex);
4523         if (allocated > map->m_len)
4524                 allocated = map->m_len;
4525         map->m_flags |= EXT4_MAP_NEW;
4526
4527         /*
4528          * Update reserved blocks/metadata blocks after successful
4529          * block allocation which had been deferred till now.
4530          */
4531         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4532                 unsigned int reserved_clusters;
4533                 /*
4534                  * Check how many clusters we had reserved this allocated range
4535                  */
4536                 reserved_clusters = get_reserved_cluster_alloc(inode,
4537                                                 map->m_lblk, allocated);
4538                 if (map_from_cluster) {
4539                         if (reserved_clusters) {
4540                                 /*
4541                                  * We have clusters reserved for this range.
4542                                  * But since we are not doing actual allocation
4543                                  * and are simply using blocks from previously
4544                                  * allocated cluster, we should release the
4545                                  * reservation and not claim quota.
4546                                  */
4547                                 ext4_da_update_reserve_space(inode,
4548                                                 reserved_clusters, 0);
4549                         }
4550                 } else {
4551                         BUG_ON(allocated_clusters < reserved_clusters);
4552                         if (reserved_clusters < allocated_clusters) {
4553                                 struct ext4_inode_info *ei = EXT4_I(inode);
4554                                 int reservation = allocated_clusters -
4555                                                   reserved_clusters;
4556                                 /*
4557                                  * It seems we claimed few clusters outside of
4558                                  * the range of this allocation. We should give
4559                                  * it back to the reservation pool. This can
4560                                  * happen in the following case:
4561                                  *
4562                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4563                                  *   cluster has 4 blocks. Thus, the clusters
4564                                  *   are [0-3],[4-7],[8-11]...
4565                                  * * First comes delayed allocation write for
4566                                  *   logical blocks 10 & 11. Since there were no
4567                                  *   previous delayed allocated blocks in the
4568                                  *   range [8-11], we would reserve 1 cluster
4569                                  *   for this write.
4570                                  * * Next comes write for logical blocks 3 to 8.
4571                                  *   In this case, we will reserve 2 clusters
4572                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4573                                  *   that range has a delayed allocated blocks.
4574                                  *   Thus total reserved clusters now becomes 3.
4575                                  * * Now, during the delayed allocation writeout
4576                                  *   time, we will first write blocks [3-8] and
4577                                  *   allocate 3 clusters for writing these
4578                                  *   blocks. Also, we would claim all these
4579                                  *   three clusters above.
4580                                  * * Now when we come here to writeout the
4581                                  *   blocks [10-11], we would expect to claim
4582                                  *   the reservation of 1 cluster we had made
4583                                  *   (and we would claim it since there are no
4584                                  *   more delayed allocated blocks in the range
4585                                  *   [8-11]. But our reserved cluster count had
4586                                  *   already gone to 0.
4587                                  *
4588                                  *   Thus, at the step 4 above when we determine
4589                                  *   that there are still some unwritten delayed
4590                                  *   allocated blocks outside of our current
4591                                  *   block range, we should increment the
4592                                  *   reserved clusters count so that when the
4593                                  *   remaining blocks finally gets written, we
4594                                  *   could claim them.
4595                                  */
4596                                 dquot_reserve_block(inode,
4597                                                 EXT4_C2B(sbi, reservation));
4598                                 spin_lock(&ei->i_block_reservation_lock);
4599                                 ei->i_reserved_data_blocks += reservation;
4600                                 spin_unlock(&ei->i_block_reservation_lock);
4601                         }
4602                         /*
4603                          * We will claim quota for all newly allocated blocks.
4604                          * We're updating the reserved space *after* the
4605                          * correction above so we do not accidentally free
4606                          * all the metadata reservation because we might
4607                          * actually need it later on.
4608                          */
4609                         ext4_da_update_reserve_space(inode, allocated_clusters,
4610                                                         1);
4611                 }
4612         }
4613
4614         /*
4615          * Cache the extent and update transaction to commit on fdatasync only
4616          * when it is _not_ an unwritten extent.
4617          */
4618         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4619                 ext4_update_inode_fsync_trans(handle, inode, 1);
4620         else
4621                 ext4_update_inode_fsync_trans(handle, inode, 0);
4622 out:
4623         if (allocated > map->m_len)
4624                 allocated = map->m_len;
4625         ext4_ext_show_leaf(inode, path);
4626         map->m_flags |= EXT4_MAP_MAPPED;
4627         map->m_pblk = newblock;
4628         map->m_len = allocated;
4629 out2:
4630         ext4_ext_drop_refs(path);
4631         kfree(path);
4632
4633         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4634                                        err ? err : allocated);
4635         return err ? err : allocated;
4636 }
4637
4638 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4639 {
4640         struct super_block *sb = inode->i_sb;
4641         ext4_lblk_t last_block;
4642         int err = 0;
4643
4644         /*
4645          * TODO: optimization is possible here.
4646          * Probably we need not scan at all,
4647          * because page truncation is enough.
4648          */
4649
4650         /* we have to know where to truncate from in crash case */
4651         EXT4_I(inode)->i_disksize = inode->i_size;
4652         ext4_mark_inode_dirty(handle, inode);
4653
4654         last_block = (inode->i_size + sb->s_blocksize - 1)
4655                         >> EXT4_BLOCK_SIZE_BITS(sb);
4656 retry:
4657         err = ext4_es_remove_extent(inode, last_block,
4658                                     EXT_MAX_BLOCKS - last_block);
4659         if (err == -ENOMEM) {
4660                 cond_resched();
4661                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4662                 goto retry;
4663         }
4664         if (err) {
4665                 ext4_std_error(inode->i_sb, err);
4666                 return;
4667         }
4668         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4669         ext4_std_error(inode->i_sb, err);
4670 }
4671
4672 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4673                                   ext4_lblk_t len, loff_t new_size,
4674                                   int flags, int mode)
4675 {
4676         struct inode *inode = file_inode(file);
4677         handle_t *handle;
4678         int ret = 0;
4679         int ret2 = 0;
4680         int retries = 0;
4681         struct ext4_map_blocks map;
4682         unsigned int credits;
4683         loff_t epos;
4684
4685         map.m_lblk = offset;
4686         map.m_len = len;
4687         /*
4688          * Don't normalize the request if it can fit in one extent so
4689          * that it doesn't get unnecessarily split into multiple
4690          * extents.
4691          */
4692         if (len <= EXT_UNWRITTEN_MAX_LEN)
4693                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4694
4695         /*
4696          * credits to insert 1 extent into extent tree
4697          */
4698         credits = ext4_chunk_trans_blocks(inode, len);
4699
4700 retry:
4701         while (ret >= 0 && len) {
4702                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4703                                             credits);
4704                 if (IS_ERR(handle)) {
4705                         ret = PTR_ERR(handle);
4706                         break;
4707                 }
4708                 ret = ext4_map_blocks(handle, inode, &map, flags);
4709                 if (ret <= 0) {
4710                         ext4_debug("inode #%lu: block %u: len %u: "
4711                                    "ext4_ext_map_blocks returned %d",
4712                                    inode->i_ino, map.m_lblk,
4713                                    map.m_len, ret);
4714                         ext4_mark_inode_dirty(handle, inode);
4715                         ret2 = ext4_journal_stop(handle);
4716                         break;
4717                 }
4718                 map.m_lblk += ret;
4719                 map.m_len = len = len - ret;
4720                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4721                 inode->i_ctime = ext4_current_time(inode);
4722                 if (new_size) {
4723                         if (epos > new_size)
4724                                 epos = new_size;
4725                         if (ext4_update_inode_size(inode, epos) & 0x1)
4726                                 inode->i_mtime = inode->i_ctime;
4727                 } else {
4728                         if (epos > inode->i_size)
4729                                 ext4_set_inode_flag(inode,
4730                                                     EXT4_INODE_EOFBLOCKS);
4731                 }
4732                 ext4_mark_inode_dirty(handle, inode);
4733                 ret2 = ext4_journal_stop(handle);
4734                 if (ret2)
4735                         break;
4736         }
4737         if (ret == -ENOSPC &&
4738                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4739                 ret = 0;
4740                 goto retry;
4741         }
4742
4743         return ret > 0 ? ret2 : ret;
4744 }
4745
4746 static long ext4_zero_range(struct file *file, loff_t offset,
4747                             loff_t len, int mode)
4748 {
4749         struct inode *inode = file_inode(file);
4750         handle_t *handle = NULL;
4751         unsigned int max_blocks;
4752         loff_t new_size = 0;
4753         int ret = 0;
4754         int flags;
4755         int credits;
4756         int partial_begin, partial_end;
4757         loff_t start, end;
4758         ext4_lblk_t lblk;
4759         struct address_space *mapping = inode->i_mapping;
4760         unsigned int blkbits = inode->i_blkbits;
4761
4762         trace_ext4_zero_range(inode, offset, len, mode);
4763
4764         if (!S_ISREG(inode->i_mode))
4765                 return -EINVAL;
4766
4767         /* Call ext4_force_commit to flush all data in case of data=journal. */
4768         if (ext4_should_journal_data(inode)) {
4769                 ret = ext4_force_commit(inode->i_sb);
4770                 if (ret)
4771                         return ret;
4772         }
4773
4774         /*
4775          * Write out all dirty pages to avoid race conditions
4776          * Then release them.
4777          */
4778         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4779                 ret = filemap_write_and_wait_range(mapping, offset,
4780                                                    offset + len - 1);
4781                 if (ret)
4782                         return ret;
4783         }
4784
4785         /*
4786          * Round up offset. This is not fallocate, we neet to zero out
4787          * blocks, so convert interior block aligned part of the range to
4788          * unwritten and possibly manually zero out unaligned parts of the
4789          * range.
4790          */
4791         start = round_up(offset, 1 << blkbits);
4792         end = round_down((offset + len), 1 << blkbits);
4793
4794         if (start < offset || end > offset + len)
4795                 return -EINVAL;
4796         partial_begin = offset & ((1 << blkbits) - 1);
4797         partial_end = (offset + len) & ((1 << blkbits) - 1);
4798
4799         lblk = start >> blkbits;
4800         max_blocks = (end >> blkbits);
4801         if (max_blocks < lblk)
4802                 max_blocks = 0;
4803         else
4804                 max_blocks -= lblk;
4805
4806         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT |
4807                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4808                 EXT4_EX_NOCACHE;
4809         if (mode & FALLOC_FL_KEEP_SIZE)
4810                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4811
4812         mutex_lock(&inode->i_mutex);
4813
4814         /*
4815          * Indirect files do not support unwritten extnets
4816          */
4817         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4818                 ret = -EOPNOTSUPP;
4819                 goto out_mutex;
4820         }
4821
4822         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4823              offset + len > i_size_read(inode)) {
4824                 new_size = offset + len;
4825                 ret = inode_newsize_ok(inode, new_size);
4826                 if (ret)
4827                         goto out_mutex;
4828                 /*
4829                  * If we have a partial block after EOF we have to allocate
4830                  * the entire block.
4831                  */
4832                 if (partial_end)
4833                         max_blocks += 1;
4834         }
4835
4836         if (max_blocks > 0) {
4837
4838                 /* Now release the pages and zero block aligned part of pages*/
4839                 truncate_pagecache_range(inode, start, end - 1);
4840                 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4841
4842                 /* Wait all existing dio workers, newcomers will block on i_mutex */
4843                 ext4_inode_block_unlocked_dio(inode);
4844                 inode_dio_wait(inode);
4845
4846                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4847                                              flags, mode);
4848                 if (ret)
4849                         goto out_dio;
4850                 /*
4851                  * Remove entire range from the extent status tree.
4852                  *
4853                  * ext4_es_remove_extent(inode, lblk, max_blocks) is
4854                  * NOT sufficient.  I'm not sure why this is the case,
4855                  * but let's be conservative and remove the extent
4856                  * status tree for the entire inode.  There should be
4857                  * no outstanding delalloc extents thanks to the
4858                  * filemap_write_and_wait_range() call above.
4859                  */
4860                 ret = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
4861                 if (ret)
4862                         goto out_dio;
4863         }
4864         if (!partial_begin && !partial_end)
4865                 goto out_dio;
4866
4867         /*
4868          * In worst case we have to writeout two nonadjacent unwritten
4869          * blocks and update the inode
4870          */
4871         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4872         if (ext4_should_journal_data(inode))
4873                 credits += 2;
4874         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4875         if (IS_ERR(handle)) {
4876                 ret = PTR_ERR(handle);
4877                 ext4_std_error(inode->i_sb, ret);
4878                 goto out_dio;
4879         }
4880
4881         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4882         if (new_size) {
4883                 ext4_update_inode_size(inode, new_size);
4884         } else {
4885                 /*
4886                 * Mark that we allocate beyond EOF so the subsequent truncate
4887                 * can proceed even if the new size is the same as i_size.
4888                 */
4889                 if ((offset + len) > i_size_read(inode))
4890                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4891         }
4892         ext4_mark_inode_dirty(handle, inode);
4893
4894         /* Zero out partial block at the edges of the range */
4895         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4896
4897         if (file->f_flags & O_SYNC)
4898                 ext4_handle_sync(handle);
4899
4900         ext4_journal_stop(handle);
4901 out_dio:
4902         ext4_inode_resume_unlocked_dio(inode);
4903 out_mutex:
4904         mutex_unlock(&inode->i_mutex);
4905         return ret;
4906 }
4907
4908 /*
4909  * preallocate space for a file. This implements ext4's fallocate file
4910  * operation, which gets called from sys_fallocate system call.
4911  * For block-mapped files, posix_fallocate should fall back to the method
4912  * of writing zeroes to the required new blocks (the same behavior which is
4913  * expected for file systems which do not support fallocate() system call).
4914  */
4915 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4916 {
4917         struct inode *inode = file_inode(file);
4918         loff_t new_size = 0;
4919         unsigned int max_blocks;
4920         int ret = 0;
4921         int flags;
4922         ext4_lblk_t lblk;
4923         unsigned int blkbits = inode->i_blkbits;
4924
4925         /* Return error if mode is not supported */
4926         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4927                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
4928                 return -EOPNOTSUPP;
4929
4930         if (mode & FALLOC_FL_PUNCH_HOLE)
4931                 return ext4_punch_hole(inode, offset, len);
4932
4933         ret = ext4_convert_inline_data(inode);
4934         if (ret)
4935                 return ret;
4936
4937         /*
4938          * currently supporting (pre)allocate mode for extent-based
4939          * files _only_
4940          */
4941         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4942                 return -EOPNOTSUPP;
4943
4944         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4945                 return ext4_collapse_range(inode, offset, len);
4946
4947         if (mode & FALLOC_FL_ZERO_RANGE)
4948                 return ext4_zero_range(file, offset, len, mode);
4949
4950         trace_ext4_fallocate_enter(inode, offset, len, mode);
4951         lblk = offset >> blkbits;
4952         /*
4953          * We can't just convert len to max_blocks because
4954          * If blocksize = 4096 offset = 3072 and len = 2048
4955          */
4956         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4957                 - lblk;
4958
4959         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4960         if (mode & FALLOC_FL_KEEP_SIZE)
4961                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4962
4963         mutex_lock(&inode->i_mutex);
4964
4965         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4966              offset + len > i_size_read(inode)) {
4967                 new_size = offset + len;
4968                 ret = inode_newsize_ok(inode, new_size);
4969                 if (ret)
4970                         goto out;
4971         }
4972
4973         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4974                                      flags, mode);
4975         if (ret)
4976                 goto out;
4977
4978         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4979                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4980                                                 EXT4_I(inode)->i_sync_tid);
4981         }
4982 out:
4983         mutex_unlock(&inode->i_mutex);
4984         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4985         return ret;
4986 }
4987
4988 /*
4989  * This function convert a range of blocks to written extents
4990  * The caller of this function will pass the start offset and the size.
4991  * all unwritten extents within this range will be converted to
4992  * written extents.
4993  *
4994  * This function is called from the direct IO end io call back
4995  * function, to convert the fallocated extents after IO is completed.
4996  * Returns 0 on success.
4997  */
4998 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4999                                    loff_t offset, ssize_t len)
5000 {
5001         unsigned int max_blocks;
5002         int ret = 0;
5003         int ret2 = 0;
5004         struct ext4_map_blocks map;
5005         unsigned int credits, blkbits = inode->i_blkbits;
5006
5007         map.m_lblk = offset >> blkbits;
5008         /*
5009          * We can't just convert len to max_blocks because
5010          * If blocksize = 4096 offset = 3072 and len = 2048
5011          */
5012         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
5013                       map.m_lblk);
5014         /*
5015          * This is somewhat ugly but the idea is clear: When transaction is
5016          * reserved, everything goes into it. Otherwise we rather start several
5017          * smaller transactions for conversion of each extent separately.
5018          */
5019         if (handle) {
5020                 handle = ext4_journal_start_reserved(handle,
5021                                                      EXT4_HT_EXT_CONVERT);
5022                 if (IS_ERR(handle))
5023                         return PTR_ERR(handle);
5024                 credits = 0;
5025         } else {
5026                 /*
5027                  * credits to insert 1 extent into extent tree
5028                  */
5029                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5030         }
5031         while (ret >= 0 && ret < max_blocks) {
5032                 map.m_lblk += ret;
5033                 map.m_len = (max_blocks -= ret);
5034                 if (credits) {
5035                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5036                                                     credits);
5037                         if (IS_ERR(handle)) {
5038                                 ret = PTR_ERR(handle);
5039                                 break;
5040                         }
5041                 }
5042                 ret = ext4_map_blocks(handle, inode, &map,
5043                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5044                 if (ret <= 0)
5045                         ext4_warning(inode->i_sb,
5046                                      "inode #%lu: block %u: len %u: "
5047                                      "ext4_ext_map_blocks returned %d",
5048                                      inode->i_ino, map.m_lblk,
5049                                      map.m_len, ret);
5050                 ext4_mark_inode_dirty(handle, inode);
5051                 if (credits)
5052                         ret2 = ext4_journal_stop(handle);
5053                 if (ret <= 0 || ret2)
5054                         break;
5055         }
5056         if (!credits)
5057                 ret2 = ext4_journal_stop(handle);
5058         return ret > 0 ? ret2 : ret;
5059 }
5060
5061 /*
5062  * If newes is not existing extent (newes->ec_pblk equals zero) find
5063  * delayed extent at start of newes and update newes accordingly and
5064  * return start of the next delayed extent.
5065  *
5066  * If newes is existing extent (newes->ec_pblk is not equal zero)
5067  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5068  * extent found. Leave newes unmodified.
5069  */
5070 static int ext4_find_delayed_extent(struct inode *inode,
5071                                     struct extent_status *newes)
5072 {
5073         struct extent_status es;
5074         ext4_lblk_t block, next_del;
5075
5076         if (newes->es_pblk == 0) {
5077                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5078                                 newes->es_lblk + newes->es_len - 1, &es);
5079
5080                 /*
5081                  * No extent in extent-tree contains block @newes->es_pblk,
5082                  * then the block may stay in 1)a hole or 2)delayed-extent.
5083                  */
5084                 if (es.es_len == 0)
5085                         /* A hole found. */
5086                         return 0;
5087
5088                 if (es.es_lblk > newes->es_lblk) {
5089                         /* A hole found. */
5090                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5091                                             newes->es_len);
5092                         return 0;
5093                 }
5094
5095                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5096         }
5097
5098         block = newes->es_lblk + newes->es_len;
5099         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5100         if (es.es_len == 0)
5101                 next_del = EXT_MAX_BLOCKS;
5102         else
5103                 next_del = es.es_lblk;
5104
5105         return next_del;
5106 }
5107 /* fiemap flags we can handle specified here */
5108 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5109
5110 static int ext4_xattr_fiemap(struct inode *inode,
5111                                 struct fiemap_extent_info *fieinfo)
5112 {
5113         __u64 physical = 0;
5114         __u64 length;
5115         __u32 flags = FIEMAP_EXTENT_LAST;
5116         int blockbits = inode->i_sb->s_blocksize_bits;
5117         int error = 0;
5118
5119         /* in-inode? */
5120         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5121                 struct ext4_iloc iloc;
5122                 int offset;     /* offset of xattr in inode */
5123
5124                 error = ext4_get_inode_loc(inode, &iloc);
5125                 if (error)
5126                         return error;
5127                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5128                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5129                                 EXT4_I(inode)->i_extra_isize;
5130                 physical += offset;
5131                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5132                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5133                 brelse(iloc.bh);
5134         } else { /* external block */
5135                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5136                 length = inode->i_sb->s_blocksize;
5137         }
5138
5139         if (physical)
5140                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5141                                                 length, flags);
5142         return (error < 0 ? error : 0);
5143 }
5144
5145 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5146                 __u64 start, __u64 len)
5147 {
5148         ext4_lblk_t start_blk;
5149         int error = 0;
5150
5151         if (ext4_has_inline_data(inode)) {
5152                 int has_inline = 1;
5153
5154                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5155                                                 start, len);
5156
5157                 if (has_inline)
5158                         return error;
5159         }
5160
5161         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5162                 error = ext4_ext_precache(inode);
5163                 if (error)
5164                         return error;
5165         }
5166
5167         /* fallback to generic here if not in extents fmt */
5168         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5169                 return __generic_block_fiemap(inode, fieinfo, start, len,
5170                                               ext4_get_block);
5171
5172         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5173                 return -EBADR;
5174
5175         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5176                 error = ext4_xattr_fiemap(inode, fieinfo);
5177         } else {
5178                 ext4_lblk_t len_blks;
5179                 __u64 last_blk;
5180
5181                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5182                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5183                 if (last_blk >= EXT_MAX_BLOCKS)
5184                         last_blk = EXT_MAX_BLOCKS-1;
5185                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5186
5187                 /*
5188                  * Walk the extent tree gathering extent information
5189                  * and pushing extents back to the user.
5190                  */
5191                 error = ext4_fill_fiemap_extents(inode, start_blk,
5192                                                  len_blks, fieinfo);
5193         }
5194         return error;
5195 }
5196
5197 /*
5198  * ext4_access_path:
5199  * Function to access the path buffer for marking it dirty.
5200  * It also checks if there are sufficient credits left in the journal handle
5201  * to update path.
5202  */
5203 static int
5204 ext4_access_path(handle_t *handle, struct inode *inode,
5205                 struct ext4_ext_path *path)
5206 {
5207         int credits, err;
5208
5209         if (!ext4_handle_valid(handle))
5210                 return 0;
5211
5212         /*
5213          * Check if need to extend journal credits
5214          * 3 for leaf, sb, and inode plus 2 (bmap and group
5215          * descriptor) for each block group; assume two block
5216          * groups
5217          */
5218         if (handle->h_buffer_credits < 7) {
5219                 credits = ext4_writepage_trans_blocks(inode);
5220                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5221                 /* EAGAIN is success */
5222                 if (err && err != -EAGAIN)
5223                         return err;
5224         }
5225
5226         err = ext4_ext_get_access(handle, inode, path);
5227         return err;
5228 }
5229
5230 /*
5231  * ext4_ext_shift_path_extents:
5232  * Shift the extents of a path structure lying between path[depth].p_ext
5233  * and EXT_LAST_EXTENT(path[depth].p_hdr) downwards, by subtracting shift
5234  * from starting block for each extent.
5235  */
5236 static int
5237 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5238                             struct inode *inode, handle_t *handle,
5239                             ext4_lblk_t *start)
5240 {
5241         int depth, err = 0;
5242         struct ext4_extent *ex_start, *ex_last;
5243         bool update = 0;
5244         depth = path->p_depth;
5245
5246         while (depth >= 0) {
5247                 if (depth == path->p_depth) {
5248                         ex_start = path[depth].p_ext;
5249                         if (!ex_start)
5250                                 return -EIO;
5251
5252                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5253
5254                         err = ext4_access_path(handle, inode, path + depth);
5255                         if (err)
5256                                 goto out;
5257
5258                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5259                                 update = 1;
5260
5261                         *start = le32_to_cpu(ex_last->ee_block) +
5262                                 ext4_ext_get_actual_len(ex_last);
5263
5264                         while (ex_start <= ex_last) {
5265                                 le32_add_cpu(&ex_start->ee_block, -shift);
5266                                 /* Try to merge to the left. */
5267                                 if ((ex_start >
5268                                      EXT_FIRST_EXTENT(path[depth].p_hdr)) &&
5269                                     ext4_ext_try_to_merge_right(inode,
5270                                                         path, ex_start - 1))
5271                                         ex_last--;
5272                                 else
5273                                         ex_start++;
5274                         }
5275                         err = ext4_ext_dirty(handle, inode, path + depth);
5276                         if (err)
5277                                 goto out;
5278
5279                         if (--depth < 0 || !update)
5280                                 break;
5281                 }
5282
5283                 /* Update index too */
5284                 err = ext4_access_path(handle, inode, path + depth);
5285                 if (err)
5286                         goto out;
5287
5288                 le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5289                 err = ext4_ext_dirty(handle, inode, path + depth);
5290                 if (err)
5291                         goto out;
5292
5293                 /* we are done if current index is not a starting index */
5294                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5295                         break;
5296
5297                 depth--;
5298         }
5299
5300 out:
5301         return err;
5302 }
5303
5304 /*
5305  * ext4_ext_shift_extents:
5306  * All the extents which lies in the range from start to the last allocated
5307  * block for the file are shifted downwards by shift blocks.
5308  * On success, 0 is returned, error otherwise.
5309  */
5310 static int
5311 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5312                        ext4_lblk_t start, ext4_lblk_t shift)
5313 {
5314         struct ext4_ext_path *path;
5315         int ret = 0, depth;
5316         struct ext4_extent *extent;
5317         ext4_lblk_t stop_block;
5318         ext4_lblk_t ex_start, ex_end;
5319
5320         /* Let path point to the last extent */
5321         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 0);
5322         if (IS_ERR(path))
5323                 return PTR_ERR(path);
5324
5325         depth = path->p_depth;
5326         extent = path[depth].p_ext;
5327         if (!extent)
5328                 goto out;
5329
5330         stop_block = le32_to_cpu(extent->ee_block) +
5331                         ext4_ext_get_actual_len(extent);
5332
5333         /* Nothing to shift, if hole is at the end of file */
5334         if (start >= stop_block)
5335                 goto out;
5336
5337         /*
5338          * Don't start shifting extents until we make sure the hole is big
5339          * enough to accomodate the shift.
5340          */
5341         path = ext4_find_extent(inode, start - 1, &path, 0);
5342         if (IS_ERR(path))
5343                 return PTR_ERR(path);
5344         depth = path->p_depth;
5345         extent =  path[depth].p_ext;
5346         if (extent) {
5347                 ex_start = le32_to_cpu(extent->ee_block);
5348                 ex_end = le32_to_cpu(extent->ee_block) +
5349                         ext4_ext_get_actual_len(extent);
5350         } else {
5351                 ex_start = 0;
5352                 ex_end = 0;
5353         }
5354
5355         if ((start == ex_start && shift > ex_start) ||
5356             (shift > start - ex_end))
5357                 return -EINVAL;
5358
5359         /* Its safe to start updating extents */
5360         while (start < stop_block) {
5361                 path = ext4_find_extent(inode, start, &path, 0);
5362                 if (IS_ERR(path))
5363                         return PTR_ERR(path);
5364                 depth = path->p_depth;
5365                 extent = path[depth].p_ext;
5366                 if (!extent) {
5367                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5368                                          (unsigned long) start);
5369                         return -EIO;
5370                 }
5371                 if (start > le32_to_cpu(extent->ee_block)) {
5372                         /* Hole, move to the next extent */
5373                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5374                                 path[depth].p_ext++;
5375                         } else {
5376                                 start = ext4_ext_next_allocated_block(path);
5377                                 continue;
5378                         }
5379                 }
5380                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5381                                 handle, &start);
5382                 if (ret)
5383                         break;
5384         }
5385 out:
5386         ext4_ext_drop_refs(path);
5387         kfree(path);
5388         return ret;
5389 }
5390
5391 /*
5392  * ext4_collapse_range:
5393  * This implements the fallocate's collapse range functionality for ext4
5394  * Returns: 0 and non-zero on error.
5395  */
5396 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5397 {
5398         struct super_block *sb = inode->i_sb;
5399         ext4_lblk_t punch_start, punch_stop;
5400         handle_t *handle;
5401         unsigned int credits;
5402         loff_t new_size, ioffset;
5403         int ret;
5404
5405         /* Collapse range works only on fs block size aligned offsets. */
5406         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5407             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5408                 return -EINVAL;
5409
5410         if (!S_ISREG(inode->i_mode))
5411                 return -EINVAL;
5412
5413         trace_ext4_collapse_range(inode, offset, len);
5414
5415         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5416         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5417
5418         /* Call ext4_force_commit to flush all data in case of data=journal. */
5419         if (ext4_should_journal_data(inode)) {
5420                 ret = ext4_force_commit(inode->i_sb);
5421                 if (ret)
5422                         return ret;
5423         }
5424
5425         /*
5426          * Need to round down offset to be aligned with page size boundary
5427          * for page size > block size.
5428          */
5429         ioffset = round_down(offset, PAGE_SIZE);
5430
5431         /* Write out all dirty pages */
5432         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5433                                            LLONG_MAX);
5434         if (ret)
5435                 return ret;
5436
5437         /* Take mutex lock */
5438         mutex_lock(&inode->i_mutex);
5439
5440         /*
5441          * There is no need to overlap collapse range with EOF, in which case
5442          * it is effectively a truncate operation
5443          */
5444         if (offset + len >= i_size_read(inode)) {
5445                 ret = -EINVAL;
5446                 goto out_mutex;
5447         }
5448
5449         /* Currently just for extent based files */
5450         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5451                 ret = -EOPNOTSUPP;
5452                 goto out_mutex;
5453         }
5454
5455         truncate_pagecache(inode, ioffset);
5456
5457         /* Wait for existing dio to complete */
5458         ext4_inode_block_unlocked_dio(inode);
5459         inode_dio_wait(inode);
5460
5461         credits = ext4_writepage_trans_blocks(inode);
5462         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5463         if (IS_ERR(handle)) {
5464                 ret = PTR_ERR(handle);
5465                 goto out_dio;
5466         }
5467
5468         down_write(&EXT4_I(inode)->i_data_sem);
5469         ext4_discard_preallocations(inode);
5470
5471         ret = ext4_es_remove_extent(inode, punch_start,
5472                                     EXT_MAX_BLOCKS - punch_start);
5473         if (ret) {
5474                 up_write(&EXT4_I(inode)->i_data_sem);
5475                 goto out_stop;
5476         }
5477
5478         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5479         if (ret) {
5480                 up_write(&EXT4_I(inode)->i_data_sem);
5481                 goto out_stop;
5482         }
5483         ext4_discard_preallocations(inode);
5484
5485         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5486                                      punch_stop - punch_start);
5487         if (ret) {
5488                 up_write(&EXT4_I(inode)->i_data_sem);
5489                 goto out_stop;
5490         }
5491
5492         new_size = i_size_read(inode) - len;
5493         i_size_write(inode, new_size);
5494         EXT4_I(inode)->i_disksize = new_size;
5495
5496         up_write(&EXT4_I(inode)->i_data_sem);
5497         if (IS_SYNC(inode))
5498                 ext4_handle_sync(handle);
5499         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5500         ext4_mark_inode_dirty(handle, inode);
5501
5502 out_stop:
5503         ext4_journal_stop(handle);
5504 out_dio:
5505         ext4_inode_resume_unlocked_dio(inode);
5506 out_mutex:
5507         mutex_unlock(&inode->i_mutex);
5508         return ret;
5509 }
5510
5511 /**
5512  * ext4_swap_extents - Swap extents between two inodes
5513  *
5514  * @inode1:     First inode
5515  * @inode2:     Second inode
5516  * @lblk1:      Start block for first inode
5517  * @lblk2:      Start block for second inode
5518  * @count:      Number of blocks to swap
5519  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5520  * @erp:        Pointer to save error value
5521  *
5522  * This helper routine does exactly what is promise "swap extents". All other
5523  * stuff such as page-cache locking consistency, bh mapping consistency or
5524  * extent's data copying must be performed by caller.
5525  * Locking:
5526  *              i_mutex is held for both inodes
5527  *              i_data_sem is locked for write for both inodes
5528  * Assumptions:
5529  *              All pages from requested range are locked for both inodes
5530  */
5531 int
5532 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5533                      struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5534                   ext4_lblk_t count, int unwritten, int *erp)
5535 {
5536         struct ext4_ext_path *path1 = NULL;
5537         struct ext4_ext_path *path2 = NULL;
5538         int replaced_count = 0;
5539
5540         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5541         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5542         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5543         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5544
5545         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5546         if (unlikely(*erp))
5547                 return 0;
5548         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5549         if (unlikely(*erp))
5550                 return 0;
5551
5552         while (count) {
5553                 struct ext4_extent *ex1, *ex2, tmp_ex;
5554                 ext4_lblk_t e1_blk, e2_blk;
5555                 int e1_len, e2_len, len;
5556                 int split = 0;
5557
5558                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5559                 if (unlikely(IS_ERR(path1))) {
5560                         *erp = PTR_ERR(path1);
5561                         path1 = NULL;
5562                 finish:
5563                         count = 0;
5564                         goto repeat;
5565                 }
5566                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5567                 if (unlikely(IS_ERR(path2))) {
5568                         *erp = PTR_ERR(path2);
5569                         path2 = NULL;
5570                         goto finish;
5571                 }
5572                 ex1 = path1[path1->p_depth].p_ext;
5573                 ex2 = path2[path2->p_depth].p_ext;
5574                 /* Do we have somthing to swap ? */
5575                 if (unlikely(!ex2 || !ex1))
5576                         goto finish;
5577
5578                 e1_blk = le32_to_cpu(ex1->ee_block);
5579                 e2_blk = le32_to_cpu(ex2->ee_block);
5580                 e1_len = ext4_ext_get_actual_len(ex1);
5581                 e2_len = ext4_ext_get_actual_len(ex2);
5582
5583                 /* Hole handling */
5584                 if (!in_range(lblk1, e1_blk, e1_len) ||
5585                     !in_range(lblk2, e2_blk, e2_len)) {
5586                         ext4_lblk_t next1, next2;
5587
5588                         /* if hole after extent, then go to next extent */
5589                         next1 = ext4_ext_next_allocated_block(path1);
5590                         next2 = ext4_ext_next_allocated_block(path2);
5591                         /* If hole before extent, then shift to that extent */
5592                         if (e1_blk > lblk1)
5593                                 next1 = e1_blk;
5594                         if (e2_blk > lblk2)
5595                                 next2 = e1_blk;
5596                         /* Do we have something to swap */
5597                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5598                                 goto finish;
5599                         /* Move to the rightest boundary */
5600                         len = next1 - lblk1;
5601                         if (len < next2 - lblk2)
5602                                 len = next2 - lblk2;
5603                         if (len > count)
5604                                 len = count;
5605                         lblk1 += len;
5606                         lblk2 += len;
5607                         count -= len;
5608                         goto repeat;
5609                 }
5610
5611                 /* Prepare left boundary */
5612                 if (e1_blk < lblk1) {
5613                         split = 1;
5614                         *erp = ext4_force_split_extent_at(handle, inode1,
5615                                                 &path1, lblk1, 0);
5616                         if (unlikely(*erp))
5617                                 goto finish;
5618                 }
5619                 if (e2_blk < lblk2) {
5620                         split = 1;
5621                         *erp = ext4_force_split_extent_at(handle, inode2,
5622                                                 &path2,  lblk2, 0);
5623                         if (unlikely(*erp))
5624                                 goto finish;
5625                 }
5626                 /* ext4_split_extent_at() may result in leaf extent split,
5627                  * path must to be revalidated. */
5628                 if (split)
5629                         goto repeat;
5630
5631                 /* Prepare right boundary */
5632                 len = count;
5633                 if (len > e1_blk + e1_len - lblk1)
5634                         len = e1_blk + e1_len - lblk1;
5635                 if (len > e2_blk + e2_len - lblk2)
5636                         len = e2_blk + e2_len - lblk2;
5637
5638                 if (len != e1_len) {
5639                         split = 1;
5640                         *erp = ext4_force_split_extent_at(handle, inode1,
5641                                                 &path1, lblk1 + len, 0);
5642                         if (unlikely(*erp))
5643                                 goto finish;
5644                 }
5645                 if (len != e2_len) {
5646                         split = 1;
5647                         *erp = ext4_force_split_extent_at(handle, inode2,
5648                                                 &path2, lblk2 + len, 0);
5649                         if (*erp)
5650                                 goto finish;
5651                 }
5652                 /* ext4_split_extent_at() may result in leaf extent split,
5653                  * path must to be revalidated. */
5654                 if (split)
5655                         goto repeat;
5656
5657                 BUG_ON(e2_len != e1_len);
5658                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5659                 if (unlikely(*erp))
5660                         goto finish;
5661                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5662                 if (unlikely(*erp))
5663                         goto finish;
5664
5665                 /* Both extents are fully inside boundaries. Swap it now */
5666                 tmp_ex = *ex1;
5667                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5668                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5669                 ex1->ee_len = cpu_to_le16(e2_len);
5670                 ex2->ee_len = cpu_to_le16(e1_len);
5671                 if (unwritten)
5672                         ext4_ext_mark_unwritten(ex2);
5673                 if (ext4_ext_is_unwritten(&tmp_ex))
5674                         ext4_ext_mark_unwritten(ex1);
5675
5676                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5677                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5678                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5679                                       path2->p_depth);
5680                 if (unlikely(*erp))
5681                         goto finish;
5682                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5683                                       path1->p_depth);
5684                 /*
5685                  * Looks scarry ah..? second inode already points to new blocks,
5686                  * and it was successfully dirtied. But luckily error may happen
5687                  * only due to journal error, so full transaction will be
5688                  * aborted anyway.
5689                  */
5690                 if (unlikely(*erp))
5691                         goto finish;
5692                 lblk1 += len;
5693                 lblk2 += len;
5694                 replaced_count += len;
5695                 count -= len;
5696
5697         repeat:
5698                 ext4_ext_drop_refs(path1);
5699                 kfree(path1);
5700                 ext4_ext_drop_refs(path2);
5701                 kfree(path2);
5702                 path1 = path2 = NULL;
5703         }
5704         return replaced_count;
5705 }