fs/ext4/extents.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2003-2006, Cluster File Systems, Inc, [email protected]
   4  * Written by Alex Tomas <[email protected]>
   5  *
   6  * Architecture independence:
   7  *   Copyright (c) 2005, Bull S.A.
   8  *   Written by Pierre Peiffer <[email protected]>
   9  */
  10
  11 /*
  12  * Extents support for EXT4
  13  *
  14  * TODO:
  15  *   - ext4*_error() should be used in some situations
  16  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  17  *   - smart tree reduction
  18  */
  19
  20 #include <linux/fs.h>
  21 #include <linux/time.h>
  22 #include <linux/jbd2.h>
  23 #include <linux/highuid.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/quotaops.h>
  26 #include <linux/string.h>
  27 #include <linux/slab.h>
  28 #include <linux/uaccess.h>
  29 #include <linux/fiemap.h>
  30 #include <linux/backing-dev.h>
  31 #include "ext4_jbd2.h"
  32 #include "ext4_extents.h"
  33 #include "xattr.h"
  34
  35 #include <trace/events/ext4.h>
  36
  37 /*
  38  * used by extent splitting.
  39  */
  40 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  41                                         due to ENOSPC */
  42 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  43 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  44
  45 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  46 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  47
  48 static __le32 ext4_extent_block_csum(struct inode *inode,
  49                                      struct ext4_extent_header *eh)
  50 {
  51         struct ext4_inode_info *ei = EXT4_I(inode);
  52         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  53         __u32 csum;
  54
  55         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  56                            EXT4_EXTENT_TAIL_OFFSET(eh));
  57         return cpu_to_le32(csum);
  58 }
  59
  60 static int ext4_extent_block_csum_verify(struct inode *inode,
  61                                          struct ext4_extent_header *eh)
  62 {
  63         struct ext4_extent_tail *et;
  64
  65         if (!ext4_has_metadata_csum(inode->i_sb))
  66                 return 1;
  67
  68         et = find_ext4_extent_tail(eh);
  69         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  70                 return 0;
  71         return 1;
  72 }
  73
  74 static void ext4_extent_block_csum_set(struct inode *inode,
  75                                        struct ext4_extent_header *eh)
  76 {
  77         struct ext4_extent_tail *et;
  78
  79         if (!ext4_has_metadata_csum(inode->i_sb))
  80                 return;
  81
  82         et = find_ext4_extent_tail(eh);
  83         et->et_checksum = ext4_extent_block_csum(inode, eh);
  84 }
  85
  86 static int ext4_split_extent(handle_t *handle,
  87                                 struct inode *inode,
  88                                 struct ext4_ext_path **ppath,
  89                                 struct ext4_map_blocks *map,
  90                                 int split_flag,
  91                                 int flags);
  92
  93 static int ext4_split_extent_at(handle_t *handle,
  94                              struct inode *inode,
  95                              struct ext4_ext_path **ppath,
  96                              ext4_lblk_t split,
  97                              int split_flag,
  98                              int flags);
  99
 100 static int ext4_find_delayed_extent(struct inode *inode,
 101                                     struct extent_status *newes);
 102
 103 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 104                                             struct inode *inode,
 105                                             int needed)
 106 {
 107         int err;
 108
 109         if (!ext4_handle_valid(handle))
 110                 return 0;
 111         if (handle->h_buffer_credits >= needed)
 112                 return 0;
 113         /*
 114          * If we need to extend the journal get a few extra blocks
 115          * while we're at it for efficiency's sake.
 116          */
 117         needed += 3;
 118         err = ext4_journal_extend(handle, needed - handle->h_buffer_credits);
 119         if (err <= 0)
 120                 return err;
 121         err = ext4_truncate_restart_trans(handle, inode, needed);
 122         if (err == 0)
 123                 err = -EAGAIN;
 124
 125         return err;
 126 }
 127
 128 /*
 129  * could return:
 130  *  - EROFS
 131  *  - ENOMEM
 132  */
 133 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 134                                 struct ext4_ext_path *path)
 135 {
 136         if (path->p_bh) {
 137                 /* path points to block */
 138                 BUFFER_TRACE(path->p_bh, "get_write_access");
 139                 return ext4_journal_get_write_access(handle, path->p_bh);
 140         }
 141         /* path points to leaf/index in inode body */
 142         /* we use in-core data, no need to protect them */
 143         return 0;
 144 }
 145
 146 /*
 147  * could return:
 148  *  - EROFS
 149  *  - ENOMEM
 150  *  - EIO
 151  */
 152 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 153                      struct inode *inode, struct ext4_ext_path *path)
 154 {
 155         int err;
 156
 157         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 158         if (path->p_bh) {
 159                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 160                 /* path points to block */
 161                 err = __ext4_handle_dirty_metadata(where, line, handle,
 162                                                    inode, path->p_bh);
 163         } else {
 164                 /* path points to leaf/index in inode body */
 165                 err = ext4_mark_inode_dirty(handle, inode);
 166         }
 167         return err;
 168 }
 169
 170 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 171                               struct ext4_ext_path *path,
 172                               ext4_lblk_t block)
 173 {
 174         if (path) {
 175                 int depth = path->p_depth;
 176                 struct ext4_extent *ex;
 177
 178                 /*
 179                  * Try to predict block placement assuming that we are
 180                  * filling in a file which will eventually be
 181                  * non-sparse --- i.e., in the case of libbfd writing
 182                  * an ELF object sections out-of-order but in a way
 183                  * the eventually results in a contiguous object or
 184                  * executable file, or some database extending a table
 185                  * space file.  However, this is actually somewhat
 186                  * non-ideal if we are writing a sparse file such as
 187                  * qemu or KVM writing a raw image file that is going
 188                  * to stay fairly sparse, since it will end up
 189                  * fragmenting the file system's free space.  Maybe we
 190                  * should have some hueristics or some way to allow
 191                  * userspace to pass a hint to file system,
 192                  * especially if the latter case turns out to be
 193                  * common.
 194                  */
 195                 ex = path[depth].p_ext;
 196                 if (ex) {
 197                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 198                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 199
 200                         if (block > ext_block)
 201                                 return ext_pblk + (block - ext_block);
 202                         else
 203                                 return ext_pblk - (ext_block - block);
 204                 }
 205
 206                 /* it looks like index is empty;
 207                  * try to find starting block from index itself */
 208                 if (path[depth].p_bh)
 209                         return path[depth].p_bh->b_blocknr;
 210         }
 211
 212         /* OK. use inode's group */
 213         return ext4_inode_to_goal_block(inode);
 214 }
 215
 216 /*
 217  * Allocation for a meta data block
 218  */
 219 static ext4_fsblk_t
 220 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 221                         struct ext4_ext_path *path,
 222                         struct ext4_extent *ex, int *err, unsigned int flags)
 223 {
 224         ext4_fsblk_t goal, newblock;
 225
 226         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 227         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 228                                         NULL, err);
 229         return newblock;
 230 }
 231
 232 static inline int ext4_ext_space_block(struct inode *inode, int check)
 233 {
 234         int size;
 235
 236         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 237                         / sizeof(struct ext4_extent);
 238 #ifdef AGGRESSIVE_TEST
 239         if (!check && size > 6)
 240                 size = 6;
 241 #endif
 242         return size;
 243 }
 244
 245 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 246 {
 247         int size;
 248
 249         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 250                         / sizeof(struct ext4_extent_idx);
 251 #ifdef AGGRESSIVE_TEST
 252         if (!check && size > 5)
 253                 size = 5;
 254 #endif
 255         return size;
 256 }
 257
 258 static inline int ext4_ext_space_root(struct inode *inode, int check)
 259 {
 260         int size;
 261
 262         size = sizeof(EXT4_I(inode)->i_data);
 263         size -= sizeof(struct ext4_extent_header);
 264         size /= sizeof(struct ext4_extent);
 265 #ifdef AGGRESSIVE_TEST
 266         if (!check && size > 3)
 267                 size = 3;
 268 #endif
 269         return size;
 270 }
 271
 272 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 273 {
 274         int size;
 275
 276         size = sizeof(EXT4_I(inode)->i_data);
 277         size -= sizeof(struct ext4_extent_header);
 278         size /= sizeof(struct ext4_extent_idx);
 279 #ifdef AGGRESSIVE_TEST
 280         if (!check && size > 4)
 281                 size = 4;
 282 #endif
 283         return size;
 284 }
 285
 286 static inline int
 287 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 288                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 289                            int nofail)
 290 {
 291         struct ext4_ext_path *path = *ppath;
 292         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 293
 294         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 295                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 296                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 297                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 298 }
 299
 300 /*
 301  * Calculate the number of metadata blocks needed
 302  * to allocate @blocks
 303  * Worse case is one block per extent
 304  */
 305 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 306 {
 307         struct ext4_inode_info *ei = EXT4_I(inode);
 308         int idxs;
 309
 310         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 311                 / sizeof(struct ext4_extent_idx));
 312
 313         /*
 314          * If the new delayed allocation block is contiguous with the
 315          * previous da block, it can share index blocks with the
 316          * previous block, so we only need to allocate a new index
 317          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 318          * an additional index block, and at ldxs**3 blocks, yet
 319          * another index blocks.
 320          */
 321         if (ei->i_da_metadata_calc_len &&
 322             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 323                 int num = 0;
 324
 325                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 326                         num++;
 327                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 328                         num++;
 329                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 330                         num++;
 331                         ei->i_da_metadata_calc_len = 0;
 332                 } else
 333                         ei->i_da_metadata_calc_len++;
 334                 ei->i_da_metadata_calc_last_lblock++;
 335                 return num;
 336         }
 337
 338         /*
 339          * In the worst case we need a new set of index blocks at
 340          * every level of the inode's extent tree.
 341          */
 342         ei->i_da_metadata_calc_len = 1;
 343         ei->i_da_metadata_calc_last_lblock = lblock;
 344         return ext_depth(inode) + 1;
 345 }
 346
 347 static int
 348 ext4_ext_max_entries(struct inode *inode, int depth)
 349 {
 350         int max;
 351
 352         if (depth == ext_depth(inode)) {
 353                 if (depth == 0)
 354                         max = ext4_ext_space_root(inode, 1);
 355                 else
 356                         max = ext4_ext_space_root_idx(inode, 1);
 357         } else {
 358                 if (depth == 0)
 359                         max = ext4_ext_space_block(inode, 1);
 360                 else
 361                         max = ext4_ext_space_block_idx(inode, 1);
 362         }
 363
 364         return max;
 365 }
 366
 367 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 368 {
 369         ext4_fsblk_t block = ext4_ext_pblock(ext);
 370         int len = ext4_ext_get_actual_len(ext);
 371         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 372
 373         /*
 374          * We allow neither:
 375          *  - zero length
 376          *  - overflow/wrap-around
 377          */
 378         if (lblock + len <= lblock)
 379                 return 0;
 380         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 381 }
 382
 383 static int ext4_valid_extent_idx(struct inode *inode,
 384                                 struct ext4_extent_idx *ext_idx)
 385 {
 386         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 387
 388         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 389 }
 390
 391 static int ext4_valid_extent_entries(struct inode *inode,
 392                                 struct ext4_extent_header *eh,
 393                                 int depth)
 394 {
 395         unsigned short entries;
 396         if (eh->eh_entries == 0)
 397                 return 1;
 398
 399         entries = le16_to_cpu(eh->eh_entries);
 400
 401         if (depth == 0) {
 402                 /* leaf entries */
 403                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 404                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 405                 ext4_fsblk_t pblock = 0;
 406                 ext4_lblk_t lblock = 0;
 407                 ext4_lblk_t prev = 0;
 408                 int len = 0;
 409                 while (entries) {
 410                         if (!ext4_valid_extent(inode, ext))
 411                                 return 0;
 412
 413                         /* Check for overlapping extents */
 414                         lblock = le32_to_cpu(ext->ee_block);
 415                         len = ext4_ext_get_actual_len(ext);
 416                         if ((lblock <= prev) && prev) {
 417                                 pblock = ext4_ext_pblock(ext);
 418                                 es->s_last_error_block = cpu_to_le64(pblock);
 419                                 return 0;
 420                         }
 421                         ext++;
 422                         entries--;
 423                         prev = lblock + len - 1;
 424                 }
 425         } else {
 426                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 427                 while (entries) {
 428                         if (!ext4_valid_extent_idx(inode, ext_idx))
 429                                 return 0;
 430                         ext_idx++;
 431                         entries--;
 432                 }
 433         }
 434         return 1;
 435 }
 436
 437 static int __ext4_ext_check(const char *function, unsigned int line,
 438                             struct inode *inode, struct ext4_extent_header *eh,
 439                             int depth, ext4_fsblk_t pblk)
 440 {
 441         const char *error_msg;
 442         int max = 0, err = -EFSCORRUPTED;
 443
 444         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 445                 error_msg = "invalid magic";
 446                 goto corrupted;
 447         }
 448         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 449                 error_msg = "unexpected eh_depth";
 450                 goto corrupted;
 451         }
 452         if (unlikely(eh->eh_max == 0)) {
 453                 error_msg = "invalid eh_max";
 454                 goto corrupted;
 455         }
 456         max = ext4_ext_max_entries(inode, depth);
 457         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 458                 error_msg = "too large eh_max";
 459                 goto corrupted;
 460         }
 461         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 462                 error_msg = "invalid eh_entries";
 463                 goto corrupted;
 464         }
 465         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 466                 error_msg = "invalid extent entries";
 467                 goto corrupted;
 468         }
 469         if (unlikely(depth > 32)) {
 470                 error_msg = "too large eh_depth";
 471                 goto corrupted;
 472         }
 473         /* Verify checksum on non-root extent tree nodes */
 474         if (ext_depth(inode) != depth &&
 475             !ext4_extent_block_csum_verify(inode, eh)) {
 476                 error_msg = "extent tree corrupted";
 477                 err = -EFSBADCRC;
 478                 goto corrupted;
 479         }
 480         return 0;
 481
 482 corrupted:
 483         ext4_error_inode(inode, function, line, 0,
 484                          "pblk %llu bad header/extent: %s - magic %x, "
 485                          "entries %u, max %u(%u), depth %u(%u)",
 486                          (unsigned long long) pblk, error_msg,
 487                          le16_to_cpu(eh->eh_magic),
 488                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 489                          max, le16_to_cpu(eh->eh_depth), depth);
 490         return err;
 491 }
 492
 493 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 494         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 495
 496 int ext4_ext_check_inode(struct inode *inode)
 497 {
 498         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 499 }
 500
 501 static struct buffer_head *
 502 __read_extent_tree_block(const char *function, unsigned int line,
 503                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 504                          int flags)
 505 {
 506         struct buffer_head              *bh;
 507         int                             err;
 508
 509         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
 510         if (unlikely(!bh))
 511                 return ERR_PTR(-ENOMEM);
 512
 513         if (!bh_uptodate_or_lock(bh)) {
 514                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 515                 err = bh_submit_read(bh);
 516                 if (err < 0)
 517                         goto errout;
 518         }
 519         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 520                 return bh;
 521         err = __ext4_ext_check(function, line, inode,
 522                                ext_block_hdr(bh), depth, pblk);
 523         if (err)
 524                 goto errout;
 525         set_buffer_verified(bh);
 526         /*
 527          * If this is a leaf block, cache all of its entries
 528          */
 529         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 530                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 531                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 532                 ext4_lblk_t prev = 0;
 533                 int i;
 534
 535                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 536                         unsigned int status = EXTENT_STATUS_WRITTEN;
 537                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 538                         int len = ext4_ext_get_actual_len(ex);
 539
 540                         if (prev && (prev != lblk))
 541                                 ext4_es_cache_extent(inode, prev,
 542                                                      lblk - prev, ~0,
 543                                                      EXTENT_STATUS_HOLE);
 544
 545                         if (ext4_ext_is_unwritten(ex))
 546                                 status = EXTENT_STATUS_UNWRITTEN;
 547                         ext4_es_cache_extent(inode, lblk, len,
 548                                              ext4_ext_pblock(ex), status);
 549                         prev = lblk + len;
 550                 }
 551         }
 552         return bh;
 553 errout:
 554         put_bh(bh);
 555         return ERR_PTR(err);
 556
 557 }
 558
 559 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 560         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 561                                  (depth), (flags))
 562
 563 /*
 564  * This function is called to cache a file's extent information in the
 565  * extent status tree
 566  */
 567 int ext4_ext_precache(struct inode *inode)
 568 {
 569         struct ext4_inode_info *ei = EXT4_I(inode);
 570         struct ext4_ext_path *path = NULL;
 571         struct buffer_head *bh;
 572         int i = 0, depth, ret = 0;
 573
 574         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 575                 return 0;       /* not an extent-mapped inode */
 576
 577         down_read(&ei->i_data_sem);
 578         depth = ext_depth(inode);
 579
 580         path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
 581                        GFP_NOFS);
 582         if (path == NULL) {
 583                 up_read(&ei->i_data_sem);
 584                 return -ENOMEM;
 585         }
 586
 587         /* Don't cache anything if there are no external extent blocks */
 588         if (depth == 0)
 589                 goto out;
 590         path[0].p_hdr = ext_inode_hdr(inode);
 591         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 592         if (ret)
 593                 goto out;
 594         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 595         while (i >= 0) {
 596                 /*
 597                  * If this is a leaf block or we've reached the end of
 598                  * the index block, go up
 599                  */
 600                 if ((i == depth) ||
 601                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 602                         brelse(path[i].p_bh);
 603                         path[i].p_bh = NULL;
 604                         i--;
 605                         continue;
 606                 }
 607                 bh = read_extent_tree_block(inode,
 608                                             ext4_idx_pblock(path[i].p_idx++),
 609                                             depth - i - 1,
 610                                             EXT4_EX_FORCE_CACHE);
 611                 if (IS_ERR(bh)) {
 612                         ret = PTR_ERR(bh);
 613                         break;
 614                 }
 615                 i++;
 616                 path[i].p_bh = bh;
 617                 path[i].p_hdr = ext_block_hdr(bh);
 618                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 619         }
 620         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 621 out:
 622         up_read(&ei->i_data_sem);
 623         ext4_ext_drop_refs(path);
 624         kfree(path);
 625         return ret;
 626 }
 627
 628 #ifdef EXT_DEBUG
 629 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 630 {
 631         int k, l = path->p_depth;
 632
 633         ext_debug("path:");
 634         for (k = 0; k <= l; k++, path++) {
 635                 if (path->p_idx) {
 636                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 637                             ext4_idx_pblock(path->p_idx));
 638                 } else if (path->p_ext) {
 639                         ext_debug("  %d:[%d]%d:%llu ",
 640                                   le32_to_cpu(path->p_ext->ee_block),
 641                                   ext4_ext_is_unwritten(path->p_ext),
 642                                   ext4_ext_get_actual_len(path->p_ext),
 643                                   ext4_ext_pblock(path->p_ext));
 644                 } else
 645                         ext_debug("  []");
 646         }
 647         ext_debug("\n");
 648 }
 649
 650 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 651 {
 652         int depth = ext_depth(inode);
 653         struct ext4_extent_header *eh;
 654         struct ext4_extent *ex;
 655         int i;
 656
 657         if (!path)
 658                 return;
 659
 660         eh = path[depth].p_hdr;
 661         ex = EXT_FIRST_EXTENT(eh);
 662
 663         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 664
 665         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 666                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 667                           ext4_ext_is_unwritten(ex),
 668                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 669         }
 670         ext_debug("\n");
 671 }
 672
 673 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 674                         ext4_fsblk_t newblock, int level)
 675 {
 676         int depth = ext_depth(inode);
 677         struct ext4_extent *ex;
 678
 679         if (depth != level) {
 680                 struct ext4_extent_idx *idx;
 681                 idx = path[level].p_idx;
 682                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 683                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 684                                         le32_to_cpu(idx->ei_block),
 685                                         ext4_idx_pblock(idx),
 686                                         newblock);
 687                         idx++;
 688                 }
 689
 690                 return;
 691         }
 692
 693         ex = path[depth].p_ext;
 694         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 695                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 696                                 le32_to_cpu(ex->ee_block),
 697                                 ext4_ext_pblock(ex),
 698                                 ext4_ext_is_unwritten(ex),
 699                                 ext4_ext_get_actual_len(ex),
 700                                 newblock);
 701                 ex++;
 702         }
 703 }
 704
 705 #else
 706 #define ext4_ext_show_path(inode, path)
 707 #define ext4_ext_show_leaf(inode, path)
 708 #define ext4_ext_show_move(inode, path, newblock, level)
 709 #endif
 710
 711 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 712 {
 713         int depth, i;
 714
 715         if (!path)
 716                 return;
 717         depth = path->p_depth;
 718         for (i = 0; i <= depth; i++, path++)
 719                 if (path->p_bh) {
 720                         brelse(path->p_bh);
 721                         path->p_bh = NULL;
 722                 }
 723 }
 724
 725 /*
 726  * ext4_ext_binsearch_idx:
 727  * binary search for the closest index of the given block
 728  * the header must be checked before calling this
 729  */
 730 static void
 731 ext4_ext_binsearch_idx(struct inode *inode,
 732                         struct ext4_ext_path *path, ext4_lblk_t block)
 733 {
 734         struct ext4_extent_header *eh = path->p_hdr;
 735         struct ext4_extent_idx *r, *l, *m;
 736
 737
 738         ext_debug("binsearch for %u(idx):  ", block);
 739
 740         l = EXT_FIRST_INDEX(eh) + 1;
 741         r = EXT_LAST_INDEX(eh);
 742         while (l <= r) {
 743                 m = l + (r - l) / 2;
 744                 if (block < le32_to_cpu(m->ei_block))
 745                         r = m - 1;
 746                 else
 747                         l = m + 1;
 748                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 749                                 m, le32_to_cpu(m->ei_block),
 750                                 r, le32_to_cpu(r->ei_block));
 751         }
 752
 753         path->p_idx = l - 1;
 754         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 755                   ext4_idx_pblock(path->p_idx));
 756
 757 #ifdef CHECK_BINSEARCH
 758         {
 759                 struct ext4_extent_idx *chix, *ix;
 760                 int k;
 761
 762                 chix = ix = EXT_FIRST_INDEX(eh);
 763                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 764                   if (k != 0 &&
 765                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 766                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 767                                        "first=0x%p\n", k,
 768                                        ix, EXT_FIRST_INDEX(eh));
 769                                 printk(KERN_DEBUG "%u <= %u\n",
 770                                        le32_to_cpu(ix->ei_block),
 771                                        le32_to_cpu(ix[-1].ei_block));
 772                         }
 773                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 774                                            <= le32_to_cpu(ix[-1].ei_block));
 775                         if (block < le32_to_cpu(ix->ei_block))
 776                                 break;
 777                         chix = ix;
 778                 }
 779                 BUG_ON(chix != path->p_idx);
 780         }
 781 #endif
 782
 783 }
 784
 785 /*
 786  * ext4_ext_binsearch:
 787  * binary search for closest extent of the given block
 788  * the header must be checked before calling this
 789  */
 790 static void
 791 ext4_ext_binsearch(struct inode *inode,
 792                 struct ext4_ext_path *path, ext4_lblk_t block)
 793 {
 794         struct ext4_extent_header *eh = path->p_hdr;
 795         struct ext4_extent *r, *l, *m;
 796
 797         if (eh->eh_entries == 0) {
 798                 /*
 799                  * this leaf is empty:
 800                  * we get such a leaf in split/add case
 801                  */
 802                 return;
 803         }
 804
 805         ext_debug("binsearch for %u:  ", block);
 806
 807         l = EXT_FIRST_EXTENT(eh) + 1;
 808         r = EXT_LAST_EXTENT(eh);
 809
 810         while (l <= r) {
 811                 m = l + (r - l) / 2;
 812                 if (block < le32_to_cpu(m->ee_block))
 813                         r = m - 1;
 814                 else
 815                         l = m + 1;
 816                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 817                                 m, le32_to_cpu(m->ee_block),
 818                                 r, le32_to_cpu(r->ee_block));
 819         }
 820
 821         path->p_ext = l - 1;
 822         ext_debug("  -> %d:%llu:[%d]%d ",
 823                         le32_to_cpu(path->p_ext->ee_block),
 824                         ext4_ext_pblock(path->p_ext),
 825                         ext4_ext_is_unwritten(path->p_ext),
 826                         ext4_ext_get_actual_len(path->p_ext));
 827
 828 #ifdef CHECK_BINSEARCH
 829         {
 830                 struct ext4_extent *chex, *ex;
 831                 int k;
 832
 833                 chex = ex = EXT_FIRST_EXTENT(eh);
 834                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 835                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 836                                           <= le32_to_cpu(ex[-1].ee_block));
 837                         if (block < le32_to_cpu(ex->ee_block))
 838                                 break;
 839                         chex = ex;
 840                 }
 841                 BUG_ON(chex != path->p_ext);
 842         }
 843 #endif
 844
 845 }
 846
 847 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 848 {
 849         struct ext4_extent_header *eh;
 850
 851         eh = ext_inode_hdr(inode);
 852         eh->eh_depth = 0;
 853         eh->eh_entries = 0;
 854         eh->eh_magic = EXT4_EXT_MAGIC;
 855         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 856         ext4_mark_inode_dirty(handle, inode);
 857         return 0;
 858 }
 859
 860 struct ext4_ext_path *
 861 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 862                  struct ext4_ext_path **orig_path, int flags)
 863 {
 864         struct ext4_extent_header *eh;
 865         struct buffer_head *bh;
 866         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 867         short int depth, i, ppos = 0;
 868         int ret;
 869
 870         eh = ext_inode_hdr(inode);
 871         depth = ext_depth(inode);
 872         if (depth < 0 || depth > EXT4_MAX_EXTENT_DEPTH) {
 873                 EXT4_ERROR_INODE(inode, "inode has invalid extent depth: %d",
 874                                  depth);
 875                 ret = -EFSCORRUPTED;
 876                 goto err;
 877         }
 878
 879         if (path) {
 880                 ext4_ext_drop_refs(path);
 881                 if (depth > path[0].p_maxdepth) {
 882                         kfree(path);
 883                         *orig_path = path = NULL;
 884                 }
 885         }
 886         if (!path) {
 887                 /* account possible depth increase */
 888                 path = kcalloc(depth + 2, sizeof(struct ext4_ext_path),
 889                                 GFP_NOFS);
 890                 if (unlikely(!path))
 891                         return ERR_PTR(-ENOMEM);
 892                 path[0].p_maxdepth = depth + 1;
 893         }
 894         path[0].p_hdr = eh;
 895         path[0].p_bh = NULL;
 896
 897         i = depth;
 898         /* walk through the tree */
 899         while (i) {
 900                 ext_debug("depth %d: num %d, max %d\n",
 901                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 902
 903                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 904                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 905                 path[ppos].p_depth = i;
 906                 path[ppos].p_ext = NULL;
 907
 908                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 909                                             flags);
 910                 if (IS_ERR(bh)) {
 911                         ret = PTR_ERR(bh);
 912                         goto err;
 913                 }
 914
 915                 eh = ext_block_hdr(bh);
 916                 ppos++;
 917                 path[ppos].p_bh = bh;
 918                 path[ppos].p_hdr = eh;
 919         }
 920
 921         path[ppos].p_depth = i;
 922         path[ppos].p_ext = NULL;
 923         path[ppos].p_idx = NULL;
 924
 925         /* find extent */
 926         ext4_ext_binsearch(inode, path + ppos, block);
 927         /* if not an empty leaf */
 928         if (path[ppos].p_ext)
 929                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 930
 931         ext4_ext_show_path(inode, path);
 932
 933         return path;
 934
 935 err:
 936         ext4_ext_drop_refs(path);
 937         kfree(path);
 938         if (orig_path)
 939                 *orig_path = NULL;
 940         return ERR_PTR(ret);
 941 }
 942
 943 /*
 944  * ext4_ext_insert_index:
 945  * insert new index [@logical;@ptr] into the block at @curp;
 946  * check where to insert: before @curp or after @curp
 947  */
 948 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 949                                  struct ext4_ext_path *curp,
 950                                  int logical, ext4_fsblk_t ptr)
 951 {
 952         struct ext4_extent_idx *ix;
 953         int len, err;
 954
 955         err = ext4_ext_get_access(handle, inode, curp);
 956         if (err)
 957                 return err;
 958
 959         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 960                 EXT4_ERROR_INODE(inode,
 961                                  "logical %d == ei_block %d!",
 962                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 963                 return -EFSCORRUPTED;
 964         }
 965
 966         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 967                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 968                 EXT4_ERROR_INODE(inode,
 969                                  "eh_entries %d >= eh_max %d!",
 970                                  le16_to_cpu(curp->p_hdr->eh_entries),
 971                                  le16_to_cpu(curp->p_hdr->eh_max));
 972                 return -EFSCORRUPTED;
 973         }
 974
 975         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 976                 /* insert after */
 977                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 978                 ix = curp->p_idx + 1;
 979         } else {
 980                 /* insert before */
 981                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 982                 ix = curp->p_idx;
 983         }
 984
 985         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 986         BUG_ON(len < 0);
 987         if (len > 0) {
 988                 ext_debug("insert new index %d: "
 989                                 "move %d indices from 0x%p to 0x%p\n",
 990                                 logical, len, ix, ix + 1);
 991                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 992         }
 993
 994         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 995                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 996                 return -EFSCORRUPTED;
 997         }
 998
 999         ix->ei_block = cpu_to_le32(logical);
1000         ext4_idx_store_pblock(ix, ptr);
1001         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1002
1003         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1004                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1005                 return -EFSCORRUPTED;
1006         }
1007
1008         err = ext4_ext_dirty(handle, inode, curp);
1009         ext4_std_error(inode->i_sb, err);
1010
1011         return err;
1012 }
1013
1014 /*
1015  * ext4_ext_split:
1016  * inserts new subtree into the path, using free index entry
1017  * at depth @at:
1018  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1019  * - makes decision where to split
1020  * - moves remaining extents and index entries (right to the split point)
1021  *   into the newly allocated blocks
1022  * - initializes subtree
1023  */
1024 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1025                           unsigned int flags,
1026                           struct ext4_ext_path *path,
1027                           struct ext4_extent *newext, int at)
1028 {
1029         struct buffer_head *bh = NULL;
1030         int depth = ext_depth(inode);
1031         struct ext4_extent_header *neh;
1032         struct ext4_extent_idx *fidx;
1033         int i = at, k, m, a;
1034         ext4_fsblk_t newblock, oldblock;
1035         __le32 border;
1036         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1037         int err = 0;
1038
1039         /* make decision: where to split? */
1040         /* FIXME: now decision is simplest: at current extent */
1041
1042         /* if current leaf will be split, then we should use
1043          * border from split point */
1044         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1045                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1046                 return -EFSCORRUPTED;
1047         }
1048         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1049                 border = path[depth].p_ext[1].ee_block;
1050                 ext_debug("leaf will be split."
1051                                 " next leaf starts at %d\n",
1052                                   le32_to_cpu(border));
1053         } else {
1054                 border = newext->ee_block;
1055                 ext_debug("leaf will be added."
1056                                 " next leaf starts at %d\n",
1057                                 le32_to_cpu(border));
1058         }
1059
1060         /*
1061          * If error occurs, then we break processing
1062          * and mark filesystem read-only. index won't
1063          * be inserted and tree will be in consistent
1064          * state. Next mount will repair buffers too.
1065          */
1066
1067         /*
1068          * Get array to track all allocated blocks.
1069          * We need this to handle errors and free blocks
1070          * upon them.
1071          */
1072         ablocks = kcalloc(depth, sizeof(ext4_fsblk_t), GFP_NOFS);
1073         if (!ablocks)
1074                 return -ENOMEM;
1075
1076         /* allocate all needed blocks */
1077         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1078         for (a = 0; a < depth - at; a++) {
1079                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1080                                                    newext, &err, flags);
1081                 if (newblock == 0)
1082                         goto cleanup;
1083                 ablocks[a] = newblock;
1084         }
1085
1086         /* initialize new leaf */
1087         newblock = ablocks[--a];
1088         if (unlikely(newblock == 0)) {
1089                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1090                 err = -EFSCORRUPTED;
1091                 goto cleanup;
1092         }
1093         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1094         if (unlikely(!bh)) {
1095                 err = -ENOMEM;
1096                 goto cleanup;
1097         }
1098         lock_buffer(bh);
1099
1100         err = ext4_journal_get_create_access(handle, bh);
1101         if (err)
1102                 goto cleanup;
1103
1104         neh = ext_block_hdr(bh);
1105         neh->eh_entries = 0;
1106         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1107         neh->eh_magic = EXT4_EXT_MAGIC;
1108         neh->eh_depth = 0;
1109
1110         /* move remainder of path[depth] to the new leaf */
1111         if (unlikely(path[depth].p_hdr->eh_entries !=
1112                      path[depth].p_hdr->eh_max)) {
1113                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1114                                  path[depth].p_hdr->eh_entries,
1115                                  path[depth].p_hdr->eh_max);
1116                 err = -EFSCORRUPTED;
1117                 goto cleanup;
1118         }
1119         /* start copy from next extent */
1120         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1121         ext4_ext_show_move(inode, path, newblock, depth);
1122         if (m) {
1123                 struct ext4_extent *ex;
1124                 ex = EXT_FIRST_EXTENT(neh);
1125                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1126                 le16_add_cpu(&neh->eh_entries, m);
1127         }
1128
1129         ext4_extent_block_csum_set(inode, neh);
1130         set_buffer_uptodate(bh);
1131         unlock_buffer(bh);
1132
1133         err = ext4_handle_dirty_metadata(handle, inode, bh);
1134         if (err)
1135                 goto cleanup;
1136         brelse(bh);
1137         bh = NULL;
1138
1139         /* correct old leaf */
1140         if (m) {
1141                 err = ext4_ext_get_access(handle, inode, path + depth);
1142                 if (err)
1143                         goto cleanup;
1144                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1145                 err = ext4_ext_dirty(handle, inode, path + depth);
1146                 if (err)
1147                         goto cleanup;
1148
1149         }
1150
1151         /* create intermediate indexes */
1152         k = depth - at - 1;
1153         if (unlikely(k < 0)) {
1154                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1155                 err = -EFSCORRUPTED;
1156                 goto cleanup;
1157         }
1158         if (k)
1159                 ext_debug("create %d intermediate indices\n", k);
1160         /* insert new index into current index block */
1161         /* current depth stored in i var */
1162         i = depth - 1;
1163         while (k--) {
1164                 oldblock = newblock;
1165                 newblock = ablocks[--a];
1166                 bh = sb_getblk(inode->i_sb, newblock);
1167                 if (unlikely(!bh)) {
1168                         err = -ENOMEM;
1169                         goto cleanup;
1170                 }
1171                 lock_buffer(bh);
1172
1173                 err = ext4_journal_get_create_access(handle, bh);
1174                 if (err)
1175                         goto cleanup;
1176
1177                 neh = ext_block_hdr(bh);
1178                 neh->eh_entries = cpu_to_le16(1);
1179                 neh->eh_magic = EXT4_EXT_MAGIC;
1180                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1181                 neh->eh_depth = cpu_to_le16(depth - i);
1182                 fidx = EXT_FIRST_INDEX(neh);
1183                 fidx->ei_block = border;
1184                 ext4_idx_store_pblock(fidx, oldblock);
1185
1186                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1187                                 i, newblock, le32_to_cpu(border), oldblock);
1188
1189                 /* move remainder of path[i] to the new index block */
1190                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1191                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1192                         EXT4_ERROR_INODE(inode,
1193                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1194                                          le32_to_cpu(path[i].p_ext->ee_block));
1195                         err = -EFSCORRUPTED;
1196                         goto cleanup;
1197                 }
1198                 /* start copy indexes */
1199                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1200                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1201                                 EXT_MAX_INDEX(path[i].p_hdr));
1202                 ext4_ext_show_move(inode, path, newblock, i);
1203                 if (m) {
1204                         memmove(++fidx, path[i].p_idx,
1205                                 sizeof(struct ext4_extent_idx) * m);
1206                         le16_add_cpu(&neh->eh_entries, m);
1207                 }
1208                 ext4_extent_block_csum_set(inode, neh);
1209                 set_buffer_uptodate(bh);
1210                 unlock_buffer(bh);
1211
1212                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1213                 if (err)
1214                         goto cleanup;
1215                 brelse(bh);
1216                 bh = NULL;
1217
1218                 /* correct old index */
1219                 if (m) {
1220                         err = ext4_ext_get_access(handle, inode, path + i);
1221                         if (err)
1222                                 goto cleanup;
1223                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1224                         err = ext4_ext_dirty(handle, inode, path + i);
1225                         if (err)
1226                                 goto cleanup;
1227                 }
1228
1229                 i--;
1230         }
1231
1232         /* insert new index */
1233         err = ext4_ext_insert_index(handle, inode, path + at,
1234                                     le32_to_cpu(border), newblock);
1235
1236 cleanup:
1237         if (bh) {
1238                 if (buffer_locked(bh))
1239                         unlock_buffer(bh);
1240                 brelse(bh);
1241         }
1242
1243         if (err) {
1244                 /* free all allocated blocks in error case */
1245                 for (i = 0; i < depth; i++) {
1246                         if (!ablocks[i])
1247                                 continue;
1248                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1249                                          EXT4_FREE_BLOCKS_METADATA);
1250                 }
1251         }
1252         kfree(ablocks);
1253
1254         return err;
1255 }
1256
1257 /*
1258  * ext4_ext_grow_indepth:
1259  * implements tree growing procedure:
1260  * - allocates new block
1261  * - moves top-level data (index block or leaf) into the new block
1262  * - initializes new top-level, creating index that points to the
1263  *   just created block
1264  */
1265 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1266                                  unsigned int flags)
1267 {
1268         struct ext4_extent_header *neh;
1269         struct buffer_head *bh;
1270         ext4_fsblk_t newblock, goal = 0;
1271         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1272         int err = 0;
1273
1274         /* Try to prepend new index to old one */
1275         if (ext_depth(inode))
1276                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1277         if (goal > le32_to_cpu(es->s_first_data_block)) {
1278                 flags |= EXT4_MB_HINT_TRY_GOAL;
1279                 goal--;
1280         } else
1281                 goal = ext4_inode_to_goal_block(inode);
1282         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1283                                         NULL, &err);
1284         if (newblock == 0)
1285                 return err;
1286
1287         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1288         if (unlikely(!bh))
1289                 return -ENOMEM;
1290         lock_buffer(bh);
1291
1292         err = ext4_journal_get_create_access(handle, bh);
1293         if (err) {
1294                 unlock_buffer(bh);
1295                 goto out;
1296         }
1297
1298         /* move top-level index/leaf into new block */
1299         memmove(bh->b_data, EXT4_I(inode)->i_data,
1300                 sizeof(EXT4_I(inode)->i_data));
1301
1302         /* set size of new block */
1303         neh = ext_block_hdr(bh);
1304         /* old root could have indexes or leaves
1305          * so calculate e_max right way */
1306         if (ext_depth(inode))
1307                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1308         else
1309                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1310         neh->eh_magic = EXT4_EXT_MAGIC;
1311         ext4_extent_block_csum_set(inode, neh);
1312         set_buffer_uptodate(bh);
1313         unlock_buffer(bh);
1314
1315         err = ext4_handle_dirty_metadata(handle, inode, bh);
1316         if (err)
1317                 goto out;
1318
1319         /* Update top-level index: num,max,pointer */
1320         neh = ext_inode_hdr(inode);
1321         neh->eh_entries = cpu_to_le16(1);
1322         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1323         if (neh->eh_depth == 0) {
1324                 /* Root extent block becomes index block */
1325                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1326                 EXT_FIRST_INDEX(neh)->ei_block =
1327                         EXT_FIRST_EXTENT(neh)->ee_block;
1328         }
1329         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1330                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1331                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1332                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1333
1334         le16_add_cpu(&neh->eh_depth, 1);
1335         ext4_mark_inode_dirty(handle, inode);
1336 out:
1337         brelse(bh);
1338
1339         return err;
1340 }
1341
1342 /*
1343  * ext4_ext_create_new_leaf:
1344  * finds empty index and adds new leaf.
1345  * if no free index is found, then it requests in-depth growing.
1346  */
1347 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1348                                     unsigned int mb_flags,
1349                                     unsigned int gb_flags,
1350                                     struct ext4_ext_path **ppath,
1351                                     struct ext4_extent *newext)
1352 {
1353         struct ext4_ext_path *path = *ppath;
1354         struct ext4_ext_path *curp;
1355         int depth, i, err = 0;
1356
1357 repeat:
1358         i = depth = ext_depth(inode);
1359
1360         /* walk up to the tree and look for free index entry */
1361         curp = path + depth;
1362         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1363                 i--;
1364                 curp--;
1365         }
1366
1367         /* we use already allocated block for index block,
1368          * so subsequent data blocks should be contiguous */
1369         if (EXT_HAS_FREE_INDEX(curp)) {
1370                 /* if we found index with free entry, then use that
1371                  * entry: create all needed subtree and add new leaf */
1372                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1373                 if (err)
1374                         goto out;
1375
1376                 /* refill path */
1377                 path = ext4_find_extent(inode,
1378                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1379                                     ppath, gb_flags);
1380                 if (IS_ERR(path))
1381                         err = PTR_ERR(path);
1382         } else {
1383                 /* tree is full, time to grow in depth */
1384                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1385                 if (err)
1386                         goto out;
1387
1388                 /* refill path */
1389                 path = ext4_find_extent(inode,
1390                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1391                                     ppath, gb_flags);
1392                 if (IS_ERR(path)) {
1393                         err = PTR_ERR(path);
1394                         goto out;
1395                 }
1396
1397                 /*
1398                  * only first (depth 0 -> 1) produces free space;
1399                  * in all other cases we have to split the grown tree
1400                  */
1401                 depth = ext_depth(inode);
1402                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1403                         /* now we need to split */
1404                         goto repeat;
1405                 }
1406         }
1407
1408 out:
1409         return err;
1410 }
1411
1412 /*
1413  * search the closest allocated block to the left for *logical
1414  * and returns it at @logical + it's physical address at @phys
1415  * if *logical is the smallest allocated block, the function
1416  * returns 0 at @phys
1417  * return value contains 0 (success) or error code
1418  */
1419 static int ext4_ext_search_left(struct inode *inode,
1420                                 struct ext4_ext_path *path,
1421                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1422 {
1423         struct ext4_extent_idx *ix;
1424         struct ext4_extent *ex;
1425         int depth, ee_len;
1426
1427         if (unlikely(path == NULL)) {
1428                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1429                 return -EFSCORRUPTED;
1430         }
1431         depth = path->p_depth;
1432         *phys = 0;
1433
1434         if (depth == 0 && path->p_ext == NULL)
1435                 return 0;
1436
1437         /* usually extent in the path covers blocks smaller
1438          * then *logical, but it can be that extent is the
1439          * first one in the file */
1440
1441         ex = path[depth].p_ext;
1442         ee_len = ext4_ext_get_actual_len(ex);
1443         if (*logical < le32_to_cpu(ex->ee_block)) {
1444                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1445                         EXT4_ERROR_INODE(inode,
1446                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1447                                          *logical, le32_to_cpu(ex->ee_block));
1448                         return -EFSCORRUPTED;
1449                 }
1450                 while (--depth >= 0) {
1451                         ix = path[depth].p_idx;
1452                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1453                                 EXT4_ERROR_INODE(inode,
1454                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1455                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1456                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1457                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1458                                   depth);
1459                                 return -EFSCORRUPTED;
1460                         }
1461                 }
1462                 return 0;
1463         }
1464
1465         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1466                 EXT4_ERROR_INODE(inode,
1467                                  "logical %d < ee_block %d + ee_len %d!",
1468                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1469                 return -EFSCORRUPTED;
1470         }
1471
1472         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1473         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1474         return 0;
1475 }
1476
1477 /*
1478  * search the closest allocated block to the right for *logical
1479  * and returns it at @logical + it's physical address at @phys
1480  * if *logical is the largest allocated block, the function
1481  * returns 0 at @phys
1482  * return value contains 0 (success) or error code
1483  */
1484 static int ext4_ext_search_right(struct inode *inode,
1485                                  struct ext4_ext_path *path,
1486                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1487                                  struct ext4_extent **ret_ex)
1488 {
1489         struct buffer_head *bh = NULL;
1490         struct ext4_extent_header *eh;
1491         struct ext4_extent_idx *ix;
1492         struct ext4_extent *ex;
1493         ext4_fsblk_t block;
1494         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1495         int ee_len;
1496
1497         if (unlikely(path == NULL)) {
1498                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1499                 return -EFSCORRUPTED;
1500         }
1501         depth = path->p_depth;
1502         *phys = 0;
1503
1504         if (depth == 0 && path->p_ext == NULL)
1505                 return 0;
1506
1507         /* usually extent in the path covers blocks smaller
1508          * then *logical, but it can be that extent is the
1509          * first one in the file */
1510
1511         ex = path[depth].p_ext;
1512         ee_len = ext4_ext_get_actual_len(ex);
1513         if (*logical < le32_to_cpu(ex->ee_block)) {
1514                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1515                         EXT4_ERROR_INODE(inode,
1516                                          "first_extent(path[%d].p_hdr) != ex",
1517                                          depth);
1518                         return -EFSCORRUPTED;
1519                 }
1520                 while (--depth >= 0) {
1521                         ix = path[depth].p_idx;
1522                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1523                                 EXT4_ERROR_INODE(inode,
1524                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1525                                                  *logical);
1526                                 return -EFSCORRUPTED;
1527                         }
1528                 }
1529                 goto found_extent;
1530         }
1531
1532         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1533                 EXT4_ERROR_INODE(inode,
1534                                  "logical %d < ee_block %d + ee_len %d!",
1535                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1536                 return -EFSCORRUPTED;
1537         }
1538
1539         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1540                 /* next allocated block in this leaf */
1541                 ex++;
1542                 goto found_extent;
1543         }
1544
1545         /* go up and search for index to the right */
1546         while (--depth >= 0) {
1547                 ix = path[depth].p_idx;
1548                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1549                         goto got_index;
1550         }
1551
1552         /* we've gone up to the root and found no index to the right */
1553         return 0;
1554
1555 got_index:
1556         /* we've found index to the right, let's
1557          * follow it and find the closest allocated
1558          * block to the right */
1559         ix++;
1560         block = ext4_idx_pblock(ix);
1561         while (++depth < path->p_depth) {
1562                 /* subtract from p_depth to get proper eh_depth */
1563                 bh = read_extent_tree_block(inode, block,
1564                                             path->p_depth - depth, 0);
1565                 if (IS_ERR(bh))
1566                         return PTR_ERR(bh);
1567                 eh = ext_block_hdr(bh);
1568                 ix = EXT_FIRST_INDEX(eh);
1569                 block = ext4_idx_pblock(ix);
1570                 put_bh(bh);
1571         }
1572
1573         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1574         if (IS_ERR(bh))
1575                 return PTR_ERR(bh);
1576         eh = ext_block_hdr(bh);
1577         ex = EXT_FIRST_EXTENT(eh);
1578 found_extent:
1579         *logical = le32_to_cpu(ex->ee_block);
1580         *phys = ext4_ext_pblock(ex);
1581         *ret_ex = ex;
1582         if (bh)
1583                 put_bh(bh);
1584         return 0;
1585 }
1586
1587 /*
1588  * ext4_ext_next_allocated_block:
1589  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1590  * NOTE: it considers block number from index entry as
1591  * allocated block. Thus, index entries have to be consistent
1592  * with leaves.
1593  */
1594 ext4_lblk_t
1595 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1596 {
1597         int depth;
1598
1599         BUG_ON(path == NULL);
1600         depth = path->p_depth;
1601
1602         if (depth == 0 && path->p_ext == NULL)
1603                 return EXT_MAX_BLOCKS;
1604
1605         while (depth >= 0) {
1606                 if (depth == path->p_depth) {
1607                         /* leaf */
1608                         if (path[depth].p_ext &&
1609                                 path[depth].p_ext !=
1610                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1611                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1612                 } else {
1613                         /* index */
1614                         if (path[depth].p_idx !=
1615                                         EXT_LAST_INDEX(path[depth].p_hdr))
1616                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1617                 }
1618                 depth--;
1619         }
1620
1621         return EXT_MAX_BLOCKS;
1622 }
1623
1624 /*
1625  * ext4_ext_next_leaf_block:
1626  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1627  */
1628 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1629 {
1630         int depth;
1631
1632         BUG_ON(path == NULL);
1633         depth = path->p_depth;
1634
1635         /* zero-tree has no leaf blocks at all */
1636         if (depth == 0)
1637                 return EXT_MAX_BLOCKS;
1638
1639         /* go to index block */
1640         depth--;
1641
1642         while (depth >= 0) {
1643                 if (path[depth].p_idx !=
1644                                 EXT_LAST_INDEX(path[depth].p_hdr))
1645                         return (ext4_lblk_t)
1646                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1647                 depth--;
1648         }
1649
1650         return EXT_MAX_BLOCKS;
1651 }
1652
1653 /*
1654  * ext4_ext_correct_indexes:
1655  * if leaf gets modified and modified extent is first in the leaf,
1656  * then we have to correct all indexes above.
1657  * TODO: do we need to correct tree in all cases?
1658  */
1659 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1660                                 struct ext4_ext_path *path)
1661 {
1662         struct ext4_extent_header *eh;
1663         int depth = ext_depth(inode);
1664         struct ext4_extent *ex;
1665         __le32 border;
1666         int k, err = 0;
1667
1668         eh = path[depth].p_hdr;
1669         ex = path[depth].p_ext;
1670
1671         if (unlikely(ex == NULL || eh == NULL)) {
1672                 EXT4_ERROR_INODE(inode,
1673                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1674                 return -EFSCORRUPTED;
1675         }
1676
1677         if (depth == 0) {
1678                 /* there is no tree at all */
1679                 return 0;
1680         }
1681
1682         if (ex != EXT_FIRST_EXTENT(eh)) {
1683                 /* we correct tree if first leaf got modified only */
1684                 return 0;
1685         }
1686
1687         /*
1688          * TODO: we need correction if border is smaller than current one
1689          */
1690         k = depth - 1;
1691         border = path[depth].p_ext->ee_block;
1692         err = ext4_ext_get_access(handle, inode, path + k);
1693         if (err)
1694                 return err;
1695         path[k].p_idx->ei_block = border;
1696         err = ext4_ext_dirty(handle, inode, path + k);
1697         if (err)
1698                 return err;
1699
1700         while (k--) {
1701                 /* change all left-side indexes */
1702                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1703                         break;
1704                 err = ext4_ext_get_access(handle, inode, path + k);
1705                 if (err)
1706                         break;
1707                 path[k].p_idx->ei_block = border;
1708                 err = ext4_ext_dirty(handle, inode, path + k);
1709                 if (err)
1710                         break;
1711         }
1712
1713         return err;
1714 }
1715
1716 int
1717 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1718                                 struct ext4_extent *ex2)
1719 {
1720         unsigned short ext1_ee_len, ext2_ee_len;
1721
1722         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1723                 return 0;
1724
1725         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1726         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1727
1728         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1729                         le32_to_cpu(ex2->ee_block))
1730                 return 0;
1731
1732         /*
1733          * To allow future support for preallocated extents to be added
1734          * as an RO_COMPAT feature, refuse to merge to extents if
1735          * this can result in the top bit of ee_len being set.
1736          */
1737         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1738                 return 0;
1739         /*
1740          * The check for IO to unwritten extent is somewhat racy as we
1741          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1742          * dropping i_data_sem. But reserved blocks should save us in that
1743          * case.
1744          */
1745         if (ext4_ext_is_unwritten(ex1) &&
1746             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1747              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1748              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1749                 return 0;
1750 #ifdef AGGRESSIVE_TEST
1751         if (ext1_ee_len >= 4)
1752                 return 0;
1753 #endif
1754
1755         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1756                 return 1;
1757         return 0;
1758 }
1759
1760 /*
1761  * This function tries to merge the "ex" extent to the next extent in the tree.
1762  * It always tries to merge towards right. If you want to merge towards
1763  * left, pass "ex - 1" as argument instead of "ex".
1764  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1765  * 1 if they got merged.
1766  */
1767 static int ext4_ext_try_to_merge_right(struct inode *inode,
1768                                  struct ext4_ext_path *path,
1769                                  struct ext4_extent *ex)
1770 {
1771         struct ext4_extent_header *eh;
1772         unsigned int depth, len;
1773         int merge_done = 0, unwritten;
1774
1775         depth = ext_depth(inode);
1776         BUG_ON(path[depth].p_hdr == NULL);
1777         eh = path[depth].p_hdr;
1778
1779         while (ex < EXT_LAST_EXTENT(eh)) {
1780                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1781                         break;
1782                 /* merge with next extent! */
1783                 unwritten = ext4_ext_is_unwritten(ex);
1784                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1785                                 + ext4_ext_get_actual_len(ex + 1));
1786                 if (unwritten)
1787                         ext4_ext_mark_unwritten(ex);
1788
1789                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1790                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1791                                 * sizeof(struct ext4_extent);
1792                         memmove(ex + 1, ex + 2, len);
1793                 }
1794                 le16_add_cpu(&eh->eh_entries, -1);
1795                 merge_done = 1;
1796                 WARN_ON(eh->eh_entries == 0);
1797                 if (!eh->eh_entries)
1798                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1799         }
1800
1801         return merge_done;
1802 }
1803
1804 /*
1805  * This function does a very simple check to see if we can collapse
1806  * an extent tree with a single extent tree leaf block into the inode.
1807  */
1808 static void ext4_ext_try_to_merge_up(handle_t *handle,
1809                                      struct inode *inode,
1810                                      struct ext4_ext_path *path)
1811 {
1812         size_t s;
1813         unsigned max_root = ext4_ext_space_root(inode, 0);
1814         ext4_fsblk_t blk;
1815
1816         if ((path[0].p_depth != 1) ||
1817             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1818             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1819                 return;
1820
1821         /*
1822          * We need to modify the block allocation bitmap and the block
1823          * group descriptor to release the extent tree block.  If we
1824          * can't get the journal credits, give up.
1825          */
1826         if (ext4_journal_extend(handle, 2))
1827                 return;
1828
1829         /*
1830          * Copy the extent data up to the inode
1831          */
1832         blk = ext4_idx_pblock(path[0].p_idx);
1833         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1834                 sizeof(struct ext4_extent_idx);
1835         s += sizeof(struct ext4_extent_header);
1836
1837         path[1].p_maxdepth = path[0].p_maxdepth;
1838         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1839         path[0].p_depth = 0;
1840         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1841                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1842         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1843
1844         brelse(path[1].p_bh);
1845         ext4_free_blocks(handle, inode, NULL, blk, 1,
1846                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1847 }
1848
1849 /*
1850  * This function tries to merge the @ex extent to neighbours in the tree.
1851  * return 1 if merge left else 0.
1852  */
1853 static void ext4_ext_try_to_merge(handle_t *handle,
1854                                   struct inode *inode,
1855                                   struct ext4_ext_path *path,
1856                                   struct ext4_extent *ex) {
1857         struct ext4_extent_header *eh;
1858         unsigned int depth;
1859         int merge_done = 0;
1860
1861         depth = ext_depth(inode);
1862         BUG_ON(path[depth].p_hdr == NULL);
1863         eh = path[depth].p_hdr;
1864
1865         if (ex > EXT_FIRST_EXTENT(eh))
1866                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1867
1868         if (!merge_done)
1869                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1870
1871         ext4_ext_try_to_merge_up(handle, inode, path);
1872 }
1873
1874 /*
1875  * check if a portion of the "newext" extent overlaps with an
1876  * existing extent.
1877  *
1878  * If there is an overlap discovered, it updates the length of the newext
1879  * such that there will be no overlap, and then returns 1.
1880  * If there is no overlap found, it returns 0.
1881  */
1882 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1883                                            struct inode *inode,
1884                                            struct ext4_extent *newext,
1885                                            struct ext4_ext_path *path)
1886 {
1887         ext4_lblk_t b1, b2;
1888         unsigned int depth, len1;
1889         unsigned int ret = 0;
1890
1891         b1 = le32_to_cpu(newext->ee_block);
1892         len1 = ext4_ext_get_actual_len(newext);
1893         depth = ext_depth(inode);
1894         if (!path[depth].p_ext)
1895                 goto out;
1896         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1897
1898         /*
1899          * get the next allocated block if the extent in the path
1900          * is before the requested block(s)
1901          */
1902         if (b2 < b1) {
1903                 b2 = ext4_ext_next_allocated_block(path);
1904                 if (b2 == EXT_MAX_BLOCKS)
1905                         goto out;
1906                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1907         }
1908
1909         /* check for wrap through zero on extent logical start block*/
1910         if (b1 + len1 < b1) {
1911                 len1 = EXT_MAX_BLOCKS - b1;
1912                 newext->ee_len = cpu_to_le16(len1);
1913                 ret = 1;
1914         }
1915
1916         /* check for overlap */
1917         if (b1 + len1 > b2) {
1918                 newext->ee_len = cpu_to_le16(b2 - b1);
1919                 ret = 1;
1920         }
1921 out:
1922         return ret;
1923 }
1924
1925 /*
1926  * ext4_ext_insert_extent:
1927  * tries to merge requsted extent into the existing extent or
1928  * inserts requested extent as new one into the tree,
1929  * creating new leaf in the no-space case.
1930  */
1931 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1932                                 struct ext4_ext_path **ppath,
1933                                 struct ext4_extent *newext, int gb_flags)
1934 {
1935         struct ext4_ext_path *path = *ppath;
1936         struct ext4_extent_header *eh;
1937         struct ext4_extent *ex, *fex;
1938         struct ext4_extent *nearex; /* nearest extent */
1939         struct ext4_ext_path *npath = NULL;
1940         int depth, len, err;
1941         ext4_lblk_t next;
1942         int mb_flags = 0, unwritten;
1943
1944         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1945                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1946         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1947                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1948                 return -EFSCORRUPTED;
1949         }
1950         depth = ext_depth(inode);
1951         ex = path[depth].p_ext;
1952         eh = path[depth].p_hdr;
1953         if (unlikely(path[depth].p_hdr == NULL)) {
1954                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1955                 return -EFSCORRUPTED;
1956         }
1957
1958         /* try to insert block into found extent and return */
1959         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1960
1961                 /*
1962                  * Try to see whether we should rather test the extent on
1963                  * right from ex, or from the left of ex. This is because
1964                  * ext4_find_extent() can return either extent on the
1965                  * left, or on the right from the searched position. This
1966                  * will make merging more effective.
1967                  */
1968                 if (ex < EXT_LAST_EXTENT(eh) &&
1969                     (le32_to_cpu(ex->ee_block) +
1970                     ext4_ext_get_actual_len(ex) <
1971                     le32_to_cpu(newext->ee_block))) {
1972                         ex += 1;
1973                         goto prepend;
1974                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1975                            (le32_to_cpu(newext->ee_block) +
1976                            ext4_ext_get_actual_len(newext) <
1977                            le32_to_cpu(ex->ee_block)))
1978                         ex -= 1;
1979
1980                 /* Try to append newex to the ex */
1981                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1982                         ext_debug("append [%d]%d block to %u:[%d]%d"
1983                                   "(from %llu)\n",
1984                                   ext4_ext_is_unwritten(newext),
1985                                   ext4_ext_get_actual_len(newext),
1986                                   le32_to_cpu(ex->ee_block),
1987                                   ext4_ext_is_unwritten(ex),
1988                                   ext4_ext_get_actual_len(ex),
1989                                   ext4_ext_pblock(ex));
1990                         err = ext4_ext_get_access(handle, inode,
1991                                                   path + depth);
1992                         if (err)
1993                                 return err;
1994                         unwritten = ext4_ext_is_unwritten(ex);
1995                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1996                                         + ext4_ext_get_actual_len(newext));
1997                         if (unwritten)
1998                                 ext4_ext_mark_unwritten(ex);
1999                         eh = path[depth].p_hdr;
2000                         nearex = ex;
2001                         goto merge;
2002                 }
2003
2004 prepend:
2005                 /* Try to prepend newex to the ex */
2006                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2007                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2008                                   "(from %llu)\n",
2009                                   le32_to_cpu(newext->ee_block),
2010                                   ext4_ext_is_unwritten(newext),
2011                                   ext4_ext_get_actual_len(newext),
2012                                   le32_to_cpu(ex->ee_block),
2013                                   ext4_ext_is_unwritten(ex),
2014                                   ext4_ext_get_actual_len(ex),
2015                                   ext4_ext_pblock(ex));
2016                         err = ext4_ext_get_access(handle, inode,
2017                                                   path + depth);
2018                         if (err)
2019                                 return err;
2020
2021                         unwritten = ext4_ext_is_unwritten(ex);
2022                         ex->ee_block = newext->ee_block;
2023                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2024                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2025                                         + ext4_ext_get_actual_len(newext));
2026                         if (unwritten)
2027                                 ext4_ext_mark_unwritten(ex);
2028                         eh = path[depth].p_hdr;
2029                         nearex = ex;
2030                         goto merge;
2031                 }
2032         }
2033
2034         depth = ext_depth(inode);
2035         eh = path[depth].p_hdr;
2036         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2037                 goto has_space;
2038
2039         /* probably next leaf has space for us? */
2040         fex = EXT_LAST_EXTENT(eh);
2041         next = EXT_MAX_BLOCKS;
2042         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2043                 next = ext4_ext_next_leaf_block(path);
2044         if (next != EXT_MAX_BLOCKS) {
2045                 ext_debug("next leaf block - %u\n", next);
2046                 BUG_ON(npath != NULL);
2047                 npath = ext4_find_extent(inode, next, NULL, 0);
2048                 if (IS_ERR(npath))
2049                         return PTR_ERR(npath);
2050                 BUG_ON(npath->p_depth != path->p_depth);
2051                 eh = npath[depth].p_hdr;
2052                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2053                         ext_debug("next leaf isn't full(%d)\n",
2054                                   le16_to_cpu(eh->eh_entries));
2055                         path = npath;
2056                         goto has_space;
2057                 }
2058                 ext_debug("next leaf has no free space(%d,%d)\n",
2059                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2060         }
2061
2062         /*
2063          * There is no free space in the found leaf.
2064          * We're gonna add a new leaf in the tree.
2065          */
2066         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2067                 mb_flags |= EXT4_MB_USE_RESERVED;
2068         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2069                                        ppath, newext);
2070         if (err)
2071                 goto cleanup;
2072         depth = ext_depth(inode);
2073         eh = path[depth].p_hdr;
2074
2075 has_space:
2076         nearex = path[depth].p_ext;
2077
2078         err = ext4_ext_get_access(handle, inode, path + depth);
2079         if (err)
2080                 goto cleanup;
2081
2082         if (!nearex) {
2083                 /* there is no extent in this leaf, create first one */
2084                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2085                                 le32_to_cpu(newext->ee_block),
2086                                 ext4_ext_pblock(newext),
2087                                 ext4_ext_is_unwritten(newext),
2088                                 ext4_ext_get_actual_len(newext));
2089                 nearex = EXT_FIRST_EXTENT(eh);
2090         } else {
2091                 if (le32_to_cpu(newext->ee_block)
2092                            > le32_to_cpu(nearex->ee_block)) {
2093                         /* Insert after */
2094                         ext_debug("insert %u:%llu:[%d]%d before: "
2095                                         "nearest %p\n",
2096                                         le32_to_cpu(newext->ee_block),
2097                                         ext4_ext_pblock(newext),
2098                                         ext4_ext_is_unwritten(newext),
2099                                         ext4_ext_get_actual_len(newext),
2100                                         nearex);
2101                         nearex++;
2102                 } else {
2103                         /* Insert before */
2104                         BUG_ON(newext->ee_block == nearex->ee_block);
2105                         ext_debug("insert %u:%llu:[%d]%d after: "
2106                                         "nearest %p\n",
2107                                         le32_to_cpu(newext->ee_block),
2108                                         ext4_ext_pblock(newext),
2109                                         ext4_ext_is_unwritten(newext),
2110                                         ext4_ext_get_actual_len(newext),
2111                                         nearex);
2112                 }
2113                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2114                 if (len > 0) {
2115                         ext_debug("insert %u:%llu:[%d]%d: "
2116                                         "move %d extents from 0x%p to 0x%p\n",
2117                                         le32_to_cpu(newext->ee_block),
2118                                         ext4_ext_pblock(newext),
2119                                         ext4_ext_is_unwritten(newext),
2120                                         ext4_ext_get_actual_len(newext),
2121                                         len, nearex, nearex + 1);
2122                         memmove(nearex + 1, nearex,
2123                                 len * sizeof(struct ext4_extent));
2124                 }
2125         }
2126
2127         le16_add_cpu(&eh->eh_entries, 1);
2128         path[depth].p_ext = nearex;
2129         nearex->ee_block = newext->ee_block;
2130         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2131         nearex->ee_len = newext->ee_len;
2132
2133 merge:
2134         /* try to merge extents */
2135         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2136                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2137
2138
2139         /* time to correct all indexes above */
2140         err = ext4_ext_correct_indexes(handle, inode, path);
2141         if (err)
2142                 goto cleanup;
2143
2144         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2145
2146 cleanup:
2147         ext4_ext_drop_refs(npath);
2148         kfree(npath);
2149         return err;
2150 }
2151
2152 static int ext4_fill_fiemap_extents(struct inode *inode,
2153                                     ext4_lblk_t block, ext4_lblk_t num,
2154                                     struct fiemap_extent_info *fieinfo)
2155 {
2156         struct ext4_ext_path *path = NULL;
2157         struct ext4_extent *ex;
2158         struct extent_status es;
2159         ext4_lblk_t next, next_del, start = 0, end = 0;
2160         ext4_lblk_t last = block + num;
2161         int exists, depth = 0, err = 0;
2162         unsigned int flags = 0;
2163         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2164
2165         while (block < last && block != EXT_MAX_BLOCKS) {
2166                 num = last - block;
2167                 /* find extent for this block */
2168                 down_read(&EXT4_I(inode)->i_data_sem);
2169
2170                 path = ext4_find_extent(inode, block, &path, 0);
2171                 if (IS_ERR(path)) {
2172                         up_read(&EXT4_I(inode)->i_data_sem);
2173                         err = PTR_ERR(path);
2174                         path = NULL;
2175                         break;
2176                 }
2177
2178                 depth = ext_depth(inode);
2179                 if (unlikely(path[depth].p_hdr == NULL)) {
2180                         up_read(&EXT4_I(inode)->i_data_sem);
2181                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2182                         err = -EFSCORRUPTED;
2183                         break;
2184                 }
2185                 ex = path[depth].p_ext;
2186                 next = ext4_ext_next_allocated_block(path);
2187
2188                 flags = 0;
2189                 exists = 0;
2190                 if (!ex) {
2191                         /* there is no extent yet, so try to allocate
2192                          * all requested space */
2193                         start = block;
2194                         end = block + num;
2195                 } else if (le32_to_cpu(ex->ee_block) > block) {
2196                         /* need to allocate space before found extent */
2197                         start = block;
2198                         end = le32_to_cpu(ex->ee_block);
2199                         if (block + num < end)
2200                                 end = block + num;
2201                 } else if (block >= le32_to_cpu(ex->ee_block)
2202                                         + ext4_ext_get_actual_len(ex)) {
2203                         /* need to allocate space after found extent */
2204                         start = block;
2205                         end = block + num;
2206                         if (end >= next)
2207                                 end = next;
2208                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2209                         /*
2210                          * some part of requested space is covered
2211                          * by found extent
2212                          */
2213                         start = block;
2214                         end = le32_to_cpu(ex->ee_block)
2215                                 + ext4_ext_get_actual_len(ex);
2216                         if (block + num < end)
2217                                 end = block + num;
2218                         exists = 1;
2219                 } else {
2220                         BUG();
2221                 }
2222                 BUG_ON(end <= start);
2223
2224                 if (!exists) {
2225                         es.es_lblk = start;
2226                         es.es_len = end - start;
2227                         es.es_pblk = 0;
2228                 } else {
2229                         es.es_lblk = le32_to_cpu(ex->ee_block);
2230                         es.es_len = ext4_ext_get_actual_len(ex);
2231                         es.es_pblk = ext4_ext_pblock(ex);
2232                         if (ext4_ext_is_unwritten(ex))
2233                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2234                 }
2235
2236                 /*
2237                  * Find delayed extent and update es accordingly. We call
2238                  * it even in !exists case to find out whether es is the
2239                  * last existing extent or not.
2240                  */
2241                 next_del = ext4_find_delayed_extent(inode, &es);
2242                 if (!exists && next_del) {
2243                         exists = 1;
2244                         flags |= (FIEMAP_EXTENT_DELALLOC |
2245                                   FIEMAP_EXTENT_UNKNOWN);
2246                 }
2247                 up_read(&EXT4_I(inode)->i_data_sem);
2248
2249                 if (unlikely(es.es_len == 0)) {
2250                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2251                         err = -EFSCORRUPTED;
2252                         break;
2253                 }
2254
2255                 /*
2256                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2257                  * we need to check next == EXT_MAX_BLOCKS because it is
2258                  * possible that an extent is with unwritten and delayed
2259                  * status due to when an extent is delayed allocated and
2260                  * is allocated by fallocate status tree will track both of
2261                  * them in a extent.
2262                  *
2263                  * So we could return a unwritten and delayed extent, and
2264                  * its block is equal to 'next'.
2265                  */
2266                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2267                         flags |= FIEMAP_EXTENT_LAST;
2268                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2269                                      next != EXT_MAX_BLOCKS)) {
2270                                 EXT4_ERROR_INODE(inode,
2271                                                  "next extent == %u, next "
2272                                                  "delalloc extent = %u",
2273                                                  next, next_del);
2274                                 err = -EFSCORRUPTED;
2275                                 break;
2276                         }
2277                 }
2278
2279                 if (exists) {
2280                         err = fiemap_fill_next_extent(fieinfo,
2281                                 (__u64)es.es_lblk << blksize_bits,
2282                                 (__u64)es.es_pblk << blksize_bits,
2283                                 (__u64)es.es_len << blksize_bits,
2284                                 flags);
2285                         if (err < 0)
2286                                 break;
2287                         if (err == 1) {
2288                                 err = 0;
2289                                 break;
2290                         }
2291                 }
2292
2293                 block = es.es_lblk + es.es_len;
2294         }
2295
2296         ext4_ext_drop_refs(path);
2297         kfree(path);
2298         return err;
2299 }
2300
2301 /*
2302  * ext4_ext_determine_hole - determine hole around given block
2303  * @inode:      inode we lookup in
2304  * @path:       path in extent tree to @lblk
2305  * @lblk:       pointer to logical block around which we want to determine hole
2306  *
2307  * Determine hole length (and start if easily possible) around given logical
2308  * block. We don't try too hard to find the beginning of the hole but @path
2309  * actually points to extent before @lblk, we provide it.
2310  *
2311  * The function returns the length of a hole starting at @lblk. We update @lblk
2312  * to the beginning of the hole if we managed to find it.
2313  */
2314 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2315                                            struct ext4_ext_path *path,
2316                                            ext4_lblk_t *lblk)
2317 {
2318         int depth = ext_depth(inode);
2319         struct ext4_extent *ex;
2320         ext4_lblk_t len;
2321
2322         ex = path[depth].p_ext;
2323         if (ex == NULL) {
2324                 /* there is no extent yet, so gap is [0;-] */
2325                 *lblk = 0;
2326                 len = EXT_MAX_BLOCKS;
2327         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2328                 len = le32_to_cpu(ex->ee_block) - *lblk;
2329         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2330                         + ext4_ext_get_actual_len(ex)) {
2331                 ext4_lblk_t next;
2332
2333                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2334                 next = ext4_ext_next_allocated_block(path);
2335                 BUG_ON(next == *lblk);
2336                 len = next - *lblk;
2337         } else {
2338                 BUG();
2339         }
2340         return len;
2341 }
2342
2343 /*
2344  * ext4_ext_put_gap_in_cache:
2345  * calculate boundaries of the gap that the requested block fits into
2346  * and cache this gap
2347  */
2348 static void
2349 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2350                           ext4_lblk_t hole_len)
2351 {
2352         struct extent_status es;
2353
2354         ext4_es_find_delayed_extent_range(inode, hole_start,
2355                                           hole_start + hole_len - 1, &es);
2356         if (es.es_len) {
2357                 /* There's delayed extent containing lblock? */
2358                 if (es.es_lblk <= hole_start)
2359                         return;
2360                 hole_len = min(es.es_lblk - hole_start, hole_len);
2361         }
2362         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2363         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2364                               EXTENT_STATUS_HOLE);
2365 }
2366
2367 /*
2368  * ext4_ext_rm_idx:
2369  * removes index from the index block.
2370  */
2371 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2372                         struct ext4_ext_path *path, int depth)
2373 {
2374         int err;
2375         ext4_fsblk_t leaf;
2376
2377         /* free index block */
2378         depth--;
2379         path = path + depth;
2380         leaf = ext4_idx_pblock(path->p_idx);
2381         if (unlikely(path->p_hdr->eh_entries == 0)) {
2382                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2383                 return -EFSCORRUPTED;
2384         }
2385         err = ext4_ext_get_access(handle, inode, path);
2386         if (err)
2387                 return err;
2388
2389         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2390                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2391                 len *= sizeof(struct ext4_extent_idx);
2392                 memmove(path->p_idx, path->p_idx + 1, len);
2393         }
2394
2395         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2396         err = ext4_ext_dirty(handle, inode, path);
2397         if (err)
2398                 return err;
2399         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2400         trace_ext4_ext_rm_idx(inode, leaf);
2401
2402         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2403                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2404
2405         while (--depth >= 0) {
2406                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2407                         break;
2408                 path--;
2409                 err = ext4_ext_get_access(handle, inode, path);
2410                 if (err)
2411                         break;
2412                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2413                 err = ext4_ext_dirty(handle, inode, path);
2414                 if (err)
2415                         break;
2416         }
2417         return err;
2418 }
2419
2420 /*
2421  * ext4_ext_calc_credits_for_single_extent:
2422  * This routine returns max. credits that needed to insert an extent
2423  * to the extent tree.
2424  * When pass the actual path, the caller should calculate credits
2425  * under i_data_sem.
2426  */
2427 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2428                                                 struct ext4_ext_path *path)
2429 {
2430         if (path) {
2431                 int depth = ext_depth(inode);
2432                 int ret = 0;
2433
2434                 /* probably there is space in leaf? */
2435                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2436                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2437
2438                         /*
2439                          *  There are some space in the leaf tree, no
2440                          *  need to account for leaf block credit
2441                          *
2442                          *  bitmaps and block group descriptor blocks
2443                          *  and other metadata blocks still need to be
2444                          *  accounted.
2445                          */
2446                         /* 1 bitmap, 1 block group descriptor */
2447                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2448                         return ret;
2449                 }
2450         }
2451
2452         return ext4_chunk_trans_blocks(inode, nrblocks);
2453 }
2454
2455 /*
2456  * How many index/leaf blocks need to change/allocate to add @extents extents?
2457  *
2458  * If we add a single extent, then in the worse case, each tree level
2459  * index/leaf need to be changed in case of the tree split.
2460  *
2461  * If more extents are inserted, they could cause the whole tree split more
2462  * than once, but this is really rare.
2463  */
2464 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2465 {
2466         int index;
2467         int depth;
2468
2469         /* If we are converting the inline data, only one is needed here. */
2470         if (ext4_has_inline_data(inode))
2471                 return 1;
2472
2473         depth = ext_depth(inode);
2474
2475         if (extents <= 1)
2476                 index = depth * 2;
2477         else
2478                 index = depth * 3;
2479
2480         return index;
2481 }
2482
2483 static inline int get_default_free_blocks_flags(struct inode *inode)
2484 {
2485         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2486             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2487                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2488         else if (ext4_should_journal_data(inode))
2489                 return EXT4_FREE_BLOCKS_FORGET;
2490         return 0;
2491 }
2492
2493 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2494                               struct ext4_extent *ex,
2495                               long long *partial_cluster,
2496                               ext4_lblk_t from, ext4_lblk_t to)
2497 {
2498         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2499         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2500         ext4_fsblk_t pblk;
2501         int flags = get_default_free_blocks_flags(inode);
2502
2503         /*
2504          * For bigalloc file systems, we never free a partial cluster
2505          * at the beginning of the extent.  Instead, we make a note
2506          * that we tried freeing the cluster, and check to see if we
2507          * need to free it on a subsequent call to ext4_remove_blocks,
2508          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2509          */
2510         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2511
2512         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2513         /*
2514          * If we have a partial cluster, and it's different from the
2515          * cluster of the last block, we need to explicitly free the
2516          * partial cluster here.
2517          */
2518         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2519         if (*partial_cluster > 0 &&
2520             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2521                 ext4_free_blocks(handle, inode, NULL,
2522                                  EXT4_C2B(sbi, *partial_cluster),
2523                                  sbi->s_cluster_ratio, flags);
2524                 *partial_cluster = 0;
2525         }
2526
2527 #ifdef EXTENTS_STATS
2528         {
2529                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2530                 spin_lock(&sbi->s_ext_stats_lock);
2531                 sbi->s_ext_blocks += ee_len;
2532                 sbi->s_ext_extents++;
2533                 if (ee_len < sbi->s_ext_min)
2534                         sbi->s_ext_min = ee_len;
2535                 if (ee_len > sbi->s_ext_max)
2536                         sbi->s_ext_max = ee_len;
2537                 if (ext_depth(inode) > sbi->s_depth_max)
2538                         sbi->s_depth_max = ext_depth(inode);
2539                 spin_unlock(&sbi->s_ext_stats_lock);
2540         }
2541 #endif
2542         if (from >= le32_to_cpu(ex->ee_block)
2543             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2544                 /* tail removal */
2545                 ext4_lblk_t num;
2546                 long long first_cluster;
2547
2548                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2549                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2550                 /*
2551                  * Usually we want to free partial cluster at the end of the
2552                  * extent, except for the situation when the cluster is still
2553                  * used by any other extent (partial_cluster is negative).
2554                  */
2555                 if (*partial_cluster < 0 &&
2556                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2557                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2558
2559                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2560                           num, pblk, *partial_cluster);
2561                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2562                 /*
2563                  * If the block range to be freed didn't start at the
2564                  * beginning of a cluster, and we removed the entire
2565                  * extent and the cluster is not used by any other extent,
2566                  * save the partial cluster here, since we might need to
2567                  * delete if we determine that the truncate or punch hole
2568                  * operation has removed all of the blocks in the cluster.
2569                  * If that cluster is used by another extent, preserve its
2570                  * negative value so it isn't freed later on.
2571                  *
2572                  * If the whole extent wasn't freed, we've reached the
2573                  * start of the truncated/punched region and have finished
2574                  * removing blocks.  If there's a partial cluster here it's
2575                  * shared with the remainder of the extent and is no longer
2576                  * a candidate for removal.
2577                  */
2578                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2579                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2580                         if (first_cluster != -*partial_cluster)
2581                                 *partial_cluster = first_cluster;
2582                 } else {
2583                         *partial_cluster = 0;
2584                 }
2585         } else
2586                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2587                            "%u-%u from %u:%u",
2588                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2589         return 0;
2590 }
2591
2592
2593 /*
2594  * ext4_ext_rm_leaf() Removes the extents associated with the
2595  * blocks appearing between "start" and "end".  Both "start"
2596  * and "end" must appear in the same extent or EIO is returned.
2597  *
2598  * @handle: The journal handle
2599  * @inode:  The files inode
2600  * @path:   The path to the leaf
2601  * @partial_cluster: The cluster which we'll have to free if all extents
2602  *                   has been released from it.  However, if this value is
2603  *                   negative, it's a cluster just to the right of the
2604  *                   punched region and it must not be freed.
2605  * @start:  The first block to remove
2606  * @end:   The last block to remove
2607  */
2608 static int
2609 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2610                  struct ext4_ext_path *path,
2611                  long long *partial_cluster,
2612                  ext4_lblk_t start, ext4_lblk_t end)
2613 {
2614         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2615         int err = 0, correct_index = 0;
2616         int depth = ext_depth(inode), credits;
2617         struct ext4_extent_header *eh;
2618         ext4_lblk_t a, b;
2619         unsigned num;
2620         ext4_lblk_t ex_ee_block;
2621         unsigned short ex_ee_len;
2622         unsigned unwritten = 0;
2623         struct ext4_extent *ex;
2624         ext4_fsblk_t pblk;
2625
2626         /* the header must be checked already in ext4_ext_remove_space() */
2627         ext_debug("truncate since %u in leaf to %u\n", start, end);
2628         if (!path[depth].p_hdr)
2629                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2630         eh = path[depth].p_hdr;
2631         if (unlikely(path[depth].p_hdr == NULL)) {
2632                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2633                 return -EFSCORRUPTED;
2634         }
2635         /* find where to start removing */
2636         ex = path[depth].p_ext;
2637         if (!ex)
2638                 ex = EXT_LAST_EXTENT(eh);
2639
2640         ex_ee_block = le32_to_cpu(ex->ee_block);
2641         ex_ee_len = ext4_ext_get_actual_len(ex);
2642
2643         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2644
2645         while (ex >= EXT_FIRST_EXTENT(eh) &&
2646                         ex_ee_block + ex_ee_len > start) {
2647
2648                 if (ext4_ext_is_unwritten(ex))
2649                         unwritten = 1;
2650                 else
2651                         unwritten = 0;
2652
2653                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2654                           unwritten, ex_ee_len);
2655                 path[depth].p_ext = ex;
2656
2657                 a = ex_ee_block > start ? ex_ee_block : start;
2658                 b = ex_ee_block+ex_ee_len - 1 < end ?
2659                         ex_ee_block+ex_ee_len - 1 : end;
2660
2661                 ext_debug("  border %u:%u\n", a, b);
2662
2663                 /* If this extent is beyond the end of the hole, skip it */
2664                 if (end < ex_ee_block) {
2665                         /*
2666                          * We're going to skip this extent and move to another,
2667                          * so note that its first cluster is in use to avoid
2668                          * freeing it when removing blocks.  Eventually, the
2669                          * right edge of the truncated/punched region will
2670                          * be just to the left.
2671                          */
2672                         if (sbi->s_cluster_ratio > 1) {
2673                                 pblk = ext4_ext_pblock(ex);
2674                                 *partial_cluster =
2675                                         -(long long) EXT4_B2C(sbi, pblk);
2676                         }
2677                         ex--;
2678                         ex_ee_block = le32_to_cpu(ex->ee_block);
2679                         ex_ee_len = ext4_ext_get_actual_len(ex);
2680                         continue;
2681                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2682                         EXT4_ERROR_INODE(inode,
2683                                          "can not handle truncate %u:%u "
2684                                          "on extent %u:%u",
2685                                          start, end, ex_ee_block,
2686                                          ex_ee_block + ex_ee_len - 1);
2687                         err = -EFSCORRUPTED;
2688                         goto out;
2689                 } else if (a != ex_ee_block) {
2690                         /* remove tail of the extent */
2691                         num = a - ex_ee_block;
2692                 } else {
2693                         /* remove whole extent: excellent! */
2694                         num = 0;
2695                 }
2696                 /*
2697                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2698                  * descriptor) for each block group; assume two block
2699                  * groups plus ex_ee_len/blocks_per_block_group for
2700                  * the worst case
2701                  */
2702                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2703                 if (ex == EXT_FIRST_EXTENT(eh)) {
2704                         correct_index = 1;
2705                         credits += (ext_depth(inode)) + 1;
2706                 }
2707                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2708
2709                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2710                 if (err)
2711                         goto out;
2712
2713                 err = ext4_ext_get_access(handle, inode, path + depth);
2714                 if (err)
2715                         goto out;
2716
2717                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2718                                          a, b);
2719                 if (err)
2720                         goto out;
2721
2722                 if (num == 0)
2723                         /* this extent is removed; mark slot entirely unused */
2724                         ext4_ext_store_pblock(ex, 0);
2725
2726                 ex->ee_len = cpu_to_le16(num);
2727                 /*
2728                  * Do not mark unwritten if all the blocks in the
2729                  * extent have been removed.
2730                  */
2731                 if (unwritten && num)
2732                         ext4_ext_mark_unwritten(ex);
2733                 /*
2734                  * If the extent was completely released,
2735                  * we need to remove it from the leaf
2736                  */
2737                 if (num == 0) {
2738                         if (end != EXT_MAX_BLOCKS - 1) {
2739                                 /*
2740                                  * For hole punching, we need to scoot all the
2741                                  * extents up when an extent is removed so that
2742                                  * we dont have blank extents in the middle
2743                                  */
2744                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2745                                         sizeof(struct ext4_extent));
2746
2747                                 /* Now get rid of the one at the end */
2748                                 memset(EXT_LAST_EXTENT(eh), 0,
2749                                         sizeof(struct ext4_extent));
2750                         }
2751                         le16_add_cpu(&eh->eh_entries, -1);
2752                 }
2753
2754                 err = ext4_ext_dirty(handle, inode, path + depth);
2755                 if (err)
2756                         goto out;
2757
2758                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2759                                 ext4_ext_pblock(ex));
2760                 ex--;
2761                 ex_ee_block = le32_to_cpu(ex->ee_block);
2762                 ex_ee_len = ext4_ext_get_actual_len(ex);
2763         }
2764
2765         if (correct_index && eh->eh_entries)
2766                 err = ext4_ext_correct_indexes(handle, inode, path);
2767
2768         /*
2769          * If there's a partial cluster and at least one extent remains in
2770          * the leaf, free the partial cluster if it isn't shared with the
2771          * current extent.  If it is shared with the current extent
2772          * we zero partial_cluster because we've reached the start of the
2773          * truncated/punched region and we're done removing blocks.
2774          */
2775         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2776                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2777                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2778                         ext4_free_blocks(handle, inode, NULL,
2779                                          EXT4_C2B(sbi, *partial_cluster),
2780                                          sbi->s_cluster_ratio,
2781                                          get_default_free_blocks_flags(inode));
2782                 }
2783                 *partial_cluster = 0;
2784         }
2785
2786         /* if this leaf is free, then we should
2787          * remove it from index block above */
2788         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2789                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2790
2791 out:
2792         return err;
2793 }
2794
2795 /*
2796  * ext4_ext_more_to_rm:
2797  * returns 1 if current index has to be freed (even partial)
2798  */
2799 static int
2800 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2801 {
2802         BUG_ON(path->p_idx == NULL);
2803
2804         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2805                 return 0;
2806
2807         /*
2808          * if truncate on deeper level happened, it wasn't partial,
2809          * so we have to consider current index for truncation
2810          */
2811         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2812                 return 0;
2813         return 1;
2814 }
2815
2816 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2817                           ext4_lblk_t end)
2818 {
2819         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2820         int depth = ext_depth(inode);
2821         struct ext4_ext_path *path = NULL;
2822         long long partial_cluster = 0;
2823         handle_t *handle;
2824         int i = 0, err = 0;
2825
2826         ext_debug("truncate since %u to %u\n", start, end);
2827
2828         /* probably first extent we're gonna free will be last in block */
2829         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2830         if (IS_ERR(handle))
2831                 return PTR_ERR(handle);
2832
2833 again:
2834         trace_ext4_ext_remove_space(inode, start, end, depth);
2835
2836         /*
2837          * Check if we are removing extents inside the extent tree. If that
2838          * is the case, we are going to punch a hole inside the extent tree
2839          * so we have to check whether we need to split the extent covering
2840          * the last block to remove so we can easily remove the part of it
2841          * in ext4_ext_rm_leaf().
2842          */
2843         if (end < EXT_MAX_BLOCKS - 1) {
2844                 struct ext4_extent *ex;
2845                 ext4_lblk_t ee_block, ex_end, lblk;
2846                 ext4_fsblk_t pblk;
2847
2848                 /* find extent for or closest extent to this block */
2849                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2850                 if (IS_ERR(path)) {
2851                         ext4_journal_stop(handle);
2852                         return PTR_ERR(path);
2853                 }
2854                 depth = ext_depth(inode);
2855                 /* Leaf not may not exist only if inode has no blocks at all */
2856                 ex = path[depth].p_ext;
2857                 if (!ex) {
2858                         if (depth) {
2859                                 EXT4_ERROR_INODE(inode,
2860                                                  "path[%d].p_hdr == NULL",
2861                                                  depth);
2862                                 err = -EFSCORRUPTED;
2863                         }
2864                         goto out;
2865                 }
2866
2867                 ee_block = le32_to_cpu(ex->ee_block);
2868                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2869
2870                 /*
2871                  * See if the last block is inside the extent, if so split
2872                  * the extent at 'end' block so we can easily remove the
2873                  * tail of the first part of the split extent in
2874                  * ext4_ext_rm_leaf().
2875                  */
2876                 if (end >= ee_block && end < ex_end) {
2877
2878                         /*
2879                          * If we're going to split the extent, note that
2880                          * the cluster containing the block after 'end' is
2881                          * in use to avoid freeing it when removing blocks.
2882                          */
2883                         if (sbi->s_cluster_ratio > 1) {
2884                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2885                                 partial_cluster =
2886                                         -(long long) EXT4_B2C(sbi, pblk);
2887                         }
2888
2889                         /*
2890                          * Split the extent in two so that 'end' is the last
2891                          * block in the first new extent. Also we should not
2892                          * fail removing space due to ENOSPC so try to use
2893                          * reserved block if that happens.
2894                          */
2895                         err = ext4_force_split_extent_at(handle, inode, &path,
2896                                                          end + 1, 1);
2897                         if (err < 0)
2898                                 goto out;
2899
2900                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2901                         /*
2902                          * If there's an extent to the right its first cluster
2903                          * contains the immediate right boundary of the
2904                          * truncated/punched region.  Set partial_cluster to
2905                          * its negative value so it won't be freed if shared
2906                          * with the current extent.  The end < ee_block case
2907                          * is handled in ext4_ext_rm_leaf().
2908                          */
2909                         lblk = ex_end + 1;
2910                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2911                                                     &ex);
2912                         if (err)
2913                                 goto out;
2914                         if (pblk)
2915                                 partial_cluster =
2916                                         -(long long) EXT4_B2C(sbi, pblk);
2917                 }
2918         }
2919         /*
2920          * We start scanning from right side, freeing all the blocks
2921          * after i_size and walking into the tree depth-wise.
2922          */
2923         depth = ext_depth(inode);
2924         if (path) {
2925                 int k = i = depth;
2926                 while (--k > 0)
2927                         path[k].p_block =
2928                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2929         } else {
2930                 path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
2931                                GFP_NOFS);
2932                 if (path == NULL) {
2933                         ext4_journal_stop(handle);
2934                         return -ENOMEM;
2935                 }
2936                 path[0].p_maxdepth = path[0].p_depth = depth;
2937                 path[0].p_hdr = ext_inode_hdr(inode);
2938                 i = 0;
2939
2940                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2941                         err = -EFSCORRUPTED;
2942                         goto out;
2943                 }
2944         }
2945         err = 0;
2946
2947         while (i >= 0 && err == 0) {
2948                 if (i == depth) {
2949                         /* this is leaf block */
2950                         err = ext4_ext_rm_leaf(handle, inode, path,
2951                                                &partial_cluster, start,
2952                                                end);
2953                         /* root level has p_bh == NULL, brelse() eats this */
2954                         brelse(path[i].p_bh);
2955                         path[i].p_bh = NULL;
2956                         i--;
2957                         continue;
2958                 }
2959
2960                 /* this is index block */
2961                 if (!path[i].p_hdr) {
2962                         ext_debug("initialize header\n");
2963                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2964                 }
2965
2966                 if (!path[i].p_idx) {
2967                         /* this level hasn't been touched yet */
2968                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2969                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2970                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2971                                   path[i].p_hdr,
2972                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2973                 } else {
2974                         /* we were already here, see at next index */
2975                         path[i].p_idx--;
2976                 }
2977
2978                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2979                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2980                                 path[i].p_idx);
2981                 if (ext4_ext_more_to_rm(path + i)) {
2982                         struct buffer_head *bh;
2983                         /* go to the next level */
2984                         ext_debug("move to level %d (block %llu)\n",
2985                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2986                         memset(path + i + 1, 0, sizeof(*path));
2987                         bh = read_extent_tree_block(inode,
2988                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2989                                 EXT4_EX_NOCACHE);
2990                         if (IS_ERR(bh)) {
2991                                 /* should we reset i_size? */
2992                                 err = PTR_ERR(bh);
2993                                 break;
2994                         }
2995                         /* Yield here to deal with large extent trees.
2996                          * Should be a no-op if we did IO above. */
2997                         cond_resched();
2998                         if (WARN_ON(i + 1 > depth)) {
2999                                 err = -EFSCORRUPTED;
3000                                 break;
3001                         }
3002                         path[i + 1].p_bh = bh;
3003
3004                         /* save actual number of indexes since this
3005                          * number is changed at the next iteration */
3006                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3007                         i++;
3008                 } else {
3009                         /* we finished processing this index, go up */
3010                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3011                                 /* index is empty, remove it;
3012                                  * handle must be already prepared by the
3013                                  * truncatei_leaf() */
3014                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3015                         }
3016                         /* root level has p_bh == NULL, brelse() eats this */
3017                         brelse(path[i].p_bh);
3018                         path[i].p_bh = NULL;
3019                         i--;
3020                         ext_debug("return to level %d\n", i);
3021                 }
3022         }
3023
3024         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3025                         partial_cluster, path->p_hdr->eh_entries);
3026
3027         /*
3028          * If we still have something in the partial cluster and we have removed
3029          * even the first extent, then we should free the blocks in the partial
3030          * cluster as well.  (This code will only run when there are no leaves
3031          * to the immediate left of the truncated/punched region.)
3032          */
3033         if (partial_cluster > 0 && err == 0) {
3034                 /* don't zero partial_cluster since it's not used afterwards */
3035                 ext4_free_blocks(handle, inode, NULL,
3036                                  EXT4_C2B(sbi, partial_cluster),
3037                                  sbi->s_cluster_ratio,
3038                                  get_default_free_blocks_flags(inode));
3039         }
3040
3041         /* TODO: flexible tree reduction should be here */
3042         if (path->p_hdr->eh_entries == 0) {
3043                 /*
3044                  * truncate to zero freed all the tree,
3045                  * so we need to correct eh_depth
3046                  */
3047                 err = ext4_ext_get_access(handle, inode, path);
3048                 if (err == 0) {
3049                         ext_inode_hdr(inode)->eh_depth = 0;
3050                         ext_inode_hdr(inode)->eh_max =
3051                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3052                         err = ext4_ext_dirty(handle, inode, path);
3053                 }
3054         }
3055 out:
3056         ext4_ext_drop_refs(path);
3057         kfree(path);
3058         path = NULL;
3059         if (err == -EAGAIN)
3060                 goto again;
3061         ext4_journal_stop(handle);
3062
3063         return err;
3064 }
3065
3066 /*
3067  * called at mount time
3068  */
3069 void ext4_ext_init(struct super_block *sb)
3070 {
3071         /*
3072          * possible initialization would be here
3073          */
3074
3075         if (ext4_has_feature_extents(sb)) {
3076 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3077                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3078 #ifdef AGGRESSIVE_TEST
3079                        ", aggressive tests"
3080 #endif
3081 #ifdef CHECK_BINSEARCH
3082                        ", check binsearch"
3083 #endif
3084 #ifdef EXTENTS_STATS
3085                        ", stats"
3086 #endif
3087                        "\n");
3088 #endif
3089 #ifdef EXTENTS_STATS
3090                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3091                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3092                 EXT4_SB(sb)->s_ext_max = 0;
3093 #endif
3094         }
3095 }
3096
3097 /*
3098  * called at umount time
3099  */
3100 void ext4_ext_release(struct super_block *sb)
3101 {
3102         if (!ext4_has_feature_extents(sb))
3103                 return;
3104
3105 #ifdef EXTENTS_STATS
3106         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3107                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3108                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3109                         sbi->s_ext_blocks, sbi->s_ext_extents,
3110                         sbi->s_ext_blocks / sbi->s_ext_extents);
3111                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3112                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3113         }
3114 #endif
3115 }
3116
3117 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3118 {
3119         ext4_lblk_t  ee_block;
3120         ext4_fsblk_t ee_pblock;
3121         unsigned int ee_len;
3122
3123         ee_block  = le32_to_cpu(ex->ee_block);
3124         ee_len    = ext4_ext_get_actual_len(ex);
3125         ee_pblock = ext4_ext_pblock(ex);
3126
3127         if (ee_len == 0)
3128                 return 0;
3129
3130         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3131                                      EXTENT_STATUS_WRITTEN);
3132 }
3133
3134 /* FIXME!! we need to try to merge to left or right after zero-out  */
3135 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3136 {
3137         ext4_fsblk_t ee_pblock;
3138         unsigned int ee_len;
3139
3140         ee_len    = ext4_ext_get_actual_len(ex);
3141         ee_pblock = ext4_ext_pblock(ex);
3142         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3143                                   ee_len);
3144 }
3145
3146 /*
3147  * ext4_split_extent_at() splits an extent at given block.
3148  *
3149  * @handle: the journal handle
3150  * @inode: the file inode
3151  * @path: the path to the extent
3152  * @split: the logical block where the extent is splitted.
3153  * @split_flags: indicates if the extent could be zeroout if split fails, and
3154  *               the states(init or unwritten) of new extents.
3155  * @flags: flags used to insert new extent to extent tree.
3156  *
3157  *
3158  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3159  * of which are deterimined by split_flag.
3160  *
3161  * There are two cases:
3162  *  a> the extent are splitted into two extent.
3163  *  b> split is not needed, and just mark the extent.
3164  *
3165  * return 0 on success.
3166  */
3167 static int ext4_split_extent_at(handle_t *handle,
3168                              struct inode *inode,
3169                              struct ext4_ext_path **ppath,
3170                              ext4_lblk_t split,
3171                              int split_flag,
3172                              int flags)
3173 {
3174         struct ext4_ext_path *path = *ppath;
3175         ext4_fsblk_t newblock;
3176         ext4_lblk_t ee_block;
3177         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3178         struct ext4_extent *ex2 = NULL;
3179         unsigned int ee_len, depth;
3180         int err = 0;
3181
3182         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3183                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3184
3185         ext_debug("ext4_split_extents_at: inode %lu, logical"
3186                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3187
3188         ext4_ext_show_leaf(inode, path);
3189
3190         depth = ext_depth(inode);
3191         ex = path[depth].p_ext;
3192         ee_block = le32_to_cpu(ex->ee_block);
3193         ee_len = ext4_ext_get_actual_len(ex);
3194         newblock = split - ee_block + ext4_ext_pblock(ex);
3195
3196         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3197         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3198                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3199                              EXT4_EXT_MARK_UNWRIT1 |
3200                              EXT4_EXT_MARK_UNWRIT2));
3201
3202         err = ext4_ext_get_access(handle, inode, path + depth);
3203         if (err)
3204                 goto out;
3205
3206         if (split == ee_block) {
3207                 /*
3208                  * case b: block @split is the block that the extent begins with
3209                  * then we just change the state of the extent, and splitting
3210                  * is not needed.
3211                  */
3212                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3213                         ext4_ext_mark_unwritten(ex);
3214                 else
3215                         ext4_ext_mark_initialized(ex);
3216
3217                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3218                         ext4_ext_try_to_merge(handle, inode, path, ex);
3219
3220                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3221                 goto out;
3222         }
3223
3224         /* case a */
3225         memcpy(&orig_ex, ex, sizeof(orig_ex));
3226         ex->ee_len = cpu_to_le16(split - ee_block);
3227         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3228                 ext4_ext_mark_unwritten(ex);
3229
3230         /*
3231          * path may lead to new leaf, not to original leaf any more
3232          * after ext4_ext_insert_extent() returns,
3233          */
3234         err = ext4_ext_dirty(handle, inode, path + depth);
3235         if (err)
3236                 goto fix_extent_len;
3237
3238         ex2 = &newex;
3239         ex2->ee_block = cpu_to_le32(split);
3240         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3241         ext4_ext_store_pblock(ex2, newblock);
3242         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3243                 ext4_ext_mark_unwritten(ex2);
3244
3245         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3246         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3247                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3248                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3249                                 err = ext4_ext_zeroout(inode, ex2);
3250                                 zero_ex.ee_block = ex2->ee_block;
3251                                 zero_ex.ee_len = cpu_to_le16(
3252                                                 ext4_ext_get_actual_len(ex2));
3253                                 ext4_ext_store_pblock(&zero_ex,
3254                                                       ext4_ext_pblock(ex2));
3255                         } else {
3256                                 err = ext4_ext_zeroout(inode, ex);
3257                                 zero_ex.ee_block = ex->ee_block;
3258                                 zero_ex.ee_len = cpu_to_le16(
3259                                                 ext4_ext_get_actual_len(ex));
3260                                 ext4_ext_store_pblock(&zero_ex,
3261                                                       ext4_ext_pblock(ex));
3262                         }
3263                 } else {
3264                         err = ext4_ext_zeroout(inode, &orig_ex);
3265                         zero_ex.ee_block = orig_ex.ee_block;
3266                         zero_ex.ee_len = cpu_to_le16(
3267                                                 ext4_ext_get_actual_len(&orig_ex));
3268                         ext4_ext_store_pblock(&zero_ex,
3269                                               ext4_ext_pblock(&orig_ex));
3270                 }
3271
3272                 if (err)
3273                         goto fix_extent_len;
3274                 /* update the extent length and mark as initialized */
3275                 ex->ee_len = cpu_to_le16(ee_len);
3276                 ext4_ext_try_to_merge(handle, inode, path, ex);
3277                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3278                 if (err)
3279                         goto fix_extent_len;
3280
3281                 /* update extent status tree */
3282                 err = ext4_zeroout_es(inode, &zero_ex);
3283
3284                 goto out;
3285         } else if (err)
3286                 goto fix_extent_len;
3287
3288 out:
3289         ext4_ext_show_leaf(inode, path);
3290         return err;
3291
3292 fix_extent_len:
3293         ex->ee_len = orig_ex.ee_len;
3294         ext4_ext_dirty(handle, inode, path + path->p_depth);
3295         return err;
3296 }
3297
3298 /*
3299  * ext4_split_extents() splits an extent and mark extent which is covered
3300  * by @map as split_flags indicates
3301  *
3302  * It may result in splitting the extent into multiple extents (up to three)
3303  * There are three possibilities:
3304  *   a> There is no split required
3305  *   b> Splits in two extents: Split is happening at either end of the extent
3306  *   c> Splits in three extents: Somone is splitting in middle of the extent
3307  *
3308  */
3309 static int ext4_split_extent(handle_t *handle,
3310                               struct inode *inode,
3311                               struct ext4_ext_path **ppath,
3312                               struct ext4_map_blocks *map,
3313                               int split_flag,
3314                               int flags)
3315 {
3316         struct ext4_ext_path *path = *ppath;
3317         ext4_lblk_t ee_block;
3318         struct ext4_extent *ex;
3319         unsigned int ee_len, depth;
3320         int err = 0;
3321         int unwritten;
3322         int split_flag1, flags1;
3323         int allocated = map->m_len;
3324
3325         depth = ext_depth(inode);
3326         ex = path[depth].p_ext;
3327         ee_block = le32_to_cpu(ex->ee_block);
3328         ee_len = ext4_ext_get_actual_len(ex);
3329         unwritten = ext4_ext_is_unwritten(ex);
3330
3331         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3332                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3333                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3334                 if (unwritten)
3335                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3336                                        EXT4_EXT_MARK_UNWRIT2;
3337                 if (split_flag & EXT4_EXT_DATA_VALID2)
3338                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3339                 err = ext4_split_extent_at(handle, inode, ppath,
3340                                 map->m_lblk + map->m_len, split_flag1, flags1);
3341                 if (err)
3342                         goto out;
3343         } else {
3344                 allocated = ee_len - (map->m_lblk - ee_block);
3345         }
3346         /*
3347          * Update path is required because previous ext4_split_extent_at() may
3348          * result in split of original leaf or extent zeroout.
3349          */
3350         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3351         if (IS_ERR(path))
3352                 return PTR_ERR(path);
3353         depth = ext_depth(inode);
3354         ex = path[depth].p_ext;
3355         if (!ex) {
3356                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3357                                  (unsigned long) map->m_lblk);
3358                 return -EFSCORRUPTED;
3359         }
3360         unwritten = ext4_ext_is_unwritten(ex);
3361         split_flag1 = 0;
3362
3363         if (map->m_lblk >= ee_block) {
3364                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3365                 if (unwritten) {
3366                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3367                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3368                                                      EXT4_EXT_MARK_UNWRIT2);
3369                 }
3370                 err = ext4_split_extent_at(handle, inode, ppath,
3371                                 map->m_lblk, split_flag1, flags);
3372                 if (err)
3373                         goto out;
3374         }
3375
3376         ext4_ext_show_leaf(inode, path);
3377 out:
3378         return err ? err : allocated;
3379 }
3380
3381 /*
3382  * This function is called by ext4_ext_map_blocks() if someone tries to write
3383  * to an unwritten extent. It may result in splitting the unwritten
3384  * extent into multiple extents (up to three - one initialized and two
3385  * unwritten).
3386  * There are three possibilities:
3387  *   a> There is no split required: Entire extent should be initialized
3388  *   b> Splits in two extents: Write is happening at either end of the extent
3389  *   c> Splits in three extents: Somone is writing in middle of the extent
3390  *
3391  * Pre-conditions:
3392  *  - The extent pointed to by 'path' is unwritten.
3393  *  - The extent pointed to by 'path' contains a superset
3394  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3395  *
3396  * Post-conditions on success:
3397  *  - the returned value is the number of blocks beyond map->l_lblk
3398  *    that are allocated and initialized.
3399  *    It is guaranteed to be >= map->m_len.
3400  */
3401 static int ext4_ext_convert_to_initialized(handle_t *handle,
3402                                            struct inode *inode,
3403                                            struct ext4_map_blocks *map,
3404                                            struct ext4_ext_path **ppath,
3405                                            int flags)
3406 {
3407         struct ext4_ext_path *path = *ppath;
3408         struct ext4_sb_info *sbi;
3409         struct ext4_extent_header *eh;
3410         struct ext4_map_blocks split_map;
3411         struct ext4_extent zero_ex1, zero_ex2;
3412         struct ext4_extent *ex, *abut_ex;
3413         ext4_lblk_t ee_block, eof_block;
3414         unsigned int ee_len, depth, map_len = map->m_len;
3415         int allocated = 0, max_zeroout = 0;
3416         int err = 0;
3417         int split_flag = EXT4_EXT_DATA_VALID2;
3418
3419         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3420                 "block %llu, max_blocks %u\n", inode->i_ino,
3421                 (unsigned long long)map->m_lblk, map_len);
3422
3423         sbi = EXT4_SB(inode->i_sb);
3424         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3425                 inode->i_sb->s_blocksize_bits;
3426         if (eof_block < map->m_lblk + map_len)
3427                 eof_block = map->m_lblk + map_len;
3428
3429         depth = ext_depth(inode);
3430         eh = path[depth].p_hdr;
3431         ex = path[depth].p_ext;
3432         ee_block = le32_to_cpu(ex->ee_block);
3433         ee_len = ext4_ext_get_actual_len(ex);
3434         zero_ex1.ee_len = 0;
3435         zero_ex2.ee_len = 0;
3436
3437         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3438
3439         /* Pre-conditions */
3440         BUG_ON(!ext4_ext_is_unwritten(ex));
3441         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3442
3443         /*
3444          * Attempt to transfer newly initialized blocks from the currently
3445          * unwritten extent to its neighbor. This is much cheaper
3446          * than an insertion followed by a merge as those involve costly
3447          * memmove() calls. Transferring to the left is the common case in
3448          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3449          * followed by append writes.
3450          *
3451          * Limitations of the current logic:
3452          *  - L1: we do not deal with writes covering the whole extent.
3453          *    This would require removing the extent if the transfer
3454          *    is possible.
3455          *  - L2: we only attempt to merge with an extent stored in the
3456          *    same extent tree node.
3457          */
3458         if ((map->m_lblk == ee_block) &&
3459                 /* See if we can merge left */
3460                 (map_len < ee_len) &&           /*L1*/
3461                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3462                 ext4_lblk_t prev_lblk;
3463                 ext4_fsblk_t prev_pblk, ee_pblk;
3464                 unsigned int prev_len;
3465
3466                 abut_ex = ex - 1;
3467                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3468                 prev_len = ext4_ext_get_actual_len(abut_ex);
3469                 prev_pblk = ext4_ext_pblock(abut_ex);
3470                 ee_pblk = ext4_ext_pblock(ex);
3471
3472                 /*
3473                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3474                  * upon those conditions:
3475                  * - C1: abut_ex is initialized,
3476                  * - C2: abut_ex is logically abutting ex,
3477                  * - C3: abut_ex is physically abutting ex,
3478                  * - C4: abut_ex can receive the additional blocks without
3479                  *   overflowing the (initialized) length limit.
3480                  */
3481                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3482                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3483                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3484                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3485                         err = ext4_ext_get_access(handle, inode, path + depth);
3486                         if (err)
3487                                 goto out;
3488
3489                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3490                                 map, ex, abut_ex);
3491
3492                         /* Shift the start of ex by 'map_len' blocks */
3493                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3494                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3495                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3496                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3497
3498                         /* Extend abut_ex by 'map_len' blocks */
3499                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3500
3501                         /* Result: number of initialized blocks past m_lblk */
3502                         allocated = map_len;
3503                 }
3504         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3505                    (map_len < ee_len) &&        /*L1*/
3506                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3507                 /* See if we can merge right */
3508                 ext4_lblk_t next_lblk;
3509                 ext4_fsblk_t next_pblk, ee_pblk;
3510                 unsigned int next_len;
3511
3512                 abut_ex = ex + 1;
3513                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3514                 next_len = ext4_ext_get_actual_len(abut_ex);
3515                 next_pblk = ext4_ext_pblock(abut_ex);
3516                 ee_pblk = ext4_ext_pblock(ex);
3517
3518                 /*
3519                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3520                  * upon those conditions:
3521                  * - C1: abut_ex is initialized,
3522                  * - C2: abut_ex is logically abutting ex,
3523                  * - C3: abut_ex is physically abutting ex,
3524                  * - C4: abut_ex can receive the additional blocks without
3525                  *   overflowing the (initialized) length limit.
3526                  */
3527                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3528                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3529                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3530                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3531                         err = ext4_ext_get_access(handle, inode, path + depth);
3532                         if (err)
3533                                 goto out;
3534
3535                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3536                                 map, ex, abut_ex);
3537
3538                         /* Shift the start of abut_ex by 'map_len' blocks */
3539                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3540                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3541                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3542                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3543
3544                         /* Extend abut_ex by 'map_len' blocks */
3545                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3546
3547                         /* Result: number of initialized blocks past m_lblk */
3548                         allocated = map_len;
3549                 }
3550         }
3551         if (allocated) {
3552                 /* Mark the block containing both extents as dirty */
3553                 ext4_ext_dirty(handle, inode, path + depth);
3554
3555                 /* Update path to point to the right extent */
3556                 path[depth].p_ext = abut_ex;
3557                 goto out;
3558         } else
3559                 allocated = ee_len - (map->m_lblk - ee_block);
3560
3561         WARN_ON(map->m_lblk < ee_block);
3562         /*
3563          * It is safe to convert extent to initialized via explicit
3564          * zeroout only if extent is fully inside i_size or new_size.
3565          */
3566         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3567
3568         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3569                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3570                         (inode->i_sb->s_blocksize_bits - 10);
3571
3572         if (ext4_encrypted_inode(inode))
3573                 max_zeroout = 0;
3574
3575         /*
3576          * five cases:
3577          * 1. split the extent into three extents.
3578          * 2. split the extent into two extents, zeroout the head of the first
3579          *    extent.
3580          * 3. split the extent into two extents, zeroout the tail of the second
3581          *    extent.
3582          * 4. split the extent into two extents with out zeroout.
3583          * 5. no splitting needed, just possibly zeroout the head and / or the
3584          *    tail of the extent.
3585          */
3586         split_map.m_lblk = map->m_lblk;
3587         split_map.m_len = map->m_len;
3588
3589         if (max_zeroout && (allocated > split_map.m_len)) {
3590                 if (allocated <= max_zeroout) {
3591                         /* case 3 or 5 */
3592                         zero_ex1.ee_block =
3593                                  cpu_to_le32(split_map.m_lblk +
3594                                              split_map.m_len);
3595                         zero_ex1.ee_len =
3596                                 cpu_to_le16(allocated - split_map.m_len);
3597                         ext4_ext_store_pblock(&zero_ex1,
3598                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3599                                 split_map.m_len - ee_block);
3600                         err = ext4_ext_zeroout(inode, &zero_ex1);
3601                         if (err)
3602                                 goto out;
3603                         split_map.m_len = allocated;
3604                 }
3605                 if (split_map.m_lblk - ee_block + split_map.m_len <
3606                                                                 max_zeroout) {
3607                         /* case 2 or 5 */
3608                         if (split_map.m_lblk != ee_block) {
3609                                 zero_ex2.ee_block = ex->ee_block;
3610                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3611                                                         ee_block);
3612                                 ext4_ext_store_pblock(&zero_ex2,
3613                                                       ext4_ext_pblock(ex));
3614                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3615                                 if (err)
3616                                         goto out;
3617                         }
3618
3619                         split_map.m_len += split_map.m_lblk - ee_block;
3620                         split_map.m_lblk = ee_block;
3621                         allocated = map->m_len;
3622                 }
3623         }
3624
3625         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3626                                 flags);
3627         if (err > 0)
3628                 err = 0;
3629 out:
3630         /* If we have gotten a failure, don't zero out status tree */
3631         if (!err) {
3632                 err = ext4_zeroout_es(inode, &zero_ex1);
3633                 if (!err)
3634                         err = ext4_zeroout_es(inode, &zero_ex2);
3635         }
3636         return err ? err : allocated;
3637 }
3638
3639 /*
3640  * This function is called by ext4_ext_map_blocks() from
3641  * ext4_get_blocks_dio_write() when DIO to write
3642  * to an unwritten extent.
3643  *
3644  * Writing to an unwritten extent may result in splitting the unwritten
3645  * extent into multiple initialized/unwritten extents (up to three)
3646  * There are three possibilities:
3647  *   a> There is no split required: Entire extent should be unwritten
3648  *   b> Splits in two extents: Write is happening at either end of the extent
3649  *   c> Splits in three extents: Somone is writing in middle of the extent
3650  *
3651  * This works the same way in the case of initialized -> unwritten conversion.
3652  *
3653  * One of more index blocks maybe needed if the extent tree grow after
3654  * the unwritten extent split. To prevent ENOSPC occur at the IO
3655  * complete, we need to split the unwritten extent before DIO submit
3656  * the IO. The unwritten extent called at this time will be split
3657  * into three unwritten extent(at most). After IO complete, the part
3658  * being filled will be convert to initialized by the end_io callback function
3659  * via ext4_convert_unwritten_extents().
3660  *
3661  * Returns the size of unwritten extent to be written on success.
3662  */
3663 static int ext4_split_convert_extents(handle_t *handle,
3664                                         struct inode *inode,
3665                                         struct ext4_map_blocks *map,
3666                                         struct ext4_ext_path **ppath,
3667                                         int flags)
3668 {
3669         struct ext4_ext_path *path = *ppath;
3670         ext4_lblk_t eof_block;
3671         ext4_lblk_t ee_block;
3672         struct ext4_extent *ex;
3673         unsigned int ee_len;
3674         int split_flag = 0, depth;
3675
3676         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3677                   __func__, inode->i_ino,
3678                   (unsigned long long)map->m_lblk, map->m_len);
3679
3680         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3681                 inode->i_sb->s_blocksize_bits;
3682         if (eof_block < map->m_lblk + map->m_len)
3683                 eof_block = map->m_lblk + map->m_len;
3684         /*
3685          * It is safe to convert extent to initialized via explicit
3686          * zeroout only if extent is fully insde i_size or new_size.
3687          */
3688         depth = ext_depth(inode);
3689         ex = path[depth].p_ext;
3690         ee_block = le32_to_cpu(ex->ee_block);
3691         ee_len = ext4_ext_get_actual_len(ex);
3692
3693         /* Convert to unwritten */
3694         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3695                 split_flag |= EXT4_EXT_DATA_VALID1;
3696         /* Convert to initialized */
3697         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3698                 split_flag |= ee_block + ee_len <= eof_block ?
3699                               EXT4_EXT_MAY_ZEROOUT : 0;
3700                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3701         }
3702         flags |= EXT4_GET_BLOCKS_PRE_IO;
3703         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3704 }
3705
3706 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3707                                                 struct inode *inode,
3708                                                 struct ext4_map_blocks *map,
3709                                                 struct ext4_ext_path **ppath)
3710 {
3711         struct ext4_ext_path *path = *ppath;
3712         struct ext4_extent *ex;
3713         ext4_lblk_t ee_block;
3714         unsigned int ee_len;
3715         int depth;
3716         int err = 0;
3717
3718         depth = ext_depth(inode);
3719         ex = path[depth].p_ext;
3720         ee_block = le32_to_cpu(ex->ee_block);
3721         ee_len = ext4_ext_get_actual_len(ex);
3722
3723         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3724                 "block %llu, max_blocks %u\n", inode->i_ino,
3725                   (unsigned long long)ee_block, ee_len);
3726
3727         /* If extent is larger than requested it is a clear sign that we still
3728          * have some extent state machine issues left. So extent_split is still
3729          * required.
3730          * TODO: Once all related issues will be fixed this situation should be
3731          * illegal.
3732          */
3733         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3734 #ifdef EXT4_DEBUG
3735                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3736                              " len %u; IO logical block %llu, len %u",
3737                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3738                              (unsigned long long)map->m_lblk, map->m_len);
3739 #endif
3740                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3741                                                  EXT4_GET_BLOCKS_CONVERT);
3742                 if (err < 0)
3743                         return err;
3744                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3745                 if (IS_ERR(path))
3746                         return PTR_ERR(path);
3747                 depth = ext_depth(inode);
3748                 ex = path[depth].p_ext;
3749         }
3750
3751         err = ext4_ext_get_access(handle, inode, path + depth);
3752         if (err)
3753                 goto out;
3754         /* first mark the extent as initialized */
3755         ext4_ext_mark_initialized(ex);
3756
3757         /* note: ext4_ext_correct_indexes() isn't needed here because
3758          * borders are not changed
3759          */
3760         ext4_ext_try_to_merge(handle, inode, path, ex);
3761
3762         /* Mark modified extent as dirty */
3763         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3764 out:
3765         ext4_ext_show_leaf(inode, path);
3766         return err;
3767 }
3768
3769 /*
3770  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3771  */
3772 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3773                               ext4_lblk_t lblk,
3774                               struct ext4_ext_path *path,
3775                               unsigned int len)
3776 {
3777         int i, depth;
3778         struct ext4_extent_header *eh;
3779         struct ext4_extent *last_ex;
3780
3781         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3782                 return 0;
3783
3784         depth = ext_depth(inode);
3785         eh = path[depth].p_hdr;
3786
3787         /*
3788          * We're going to remove EOFBLOCKS_FL entirely in future so we
3789          * do not care for this case anymore. Simply remove the flag
3790          * if there are no extents.
3791          */
3792         if (unlikely(!eh->eh_entries))
3793                 goto out;
3794         last_ex = EXT_LAST_EXTENT(eh);
3795         /*
3796          * We should clear the EOFBLOCKS_FL flag if we are writing the
3797          * last block in the last extent in the file.  We test this by
3798          * first checking to see if the caller to
3799          * ext4_ext_get_blocks() was interested in the last block (or
3800          * a block beyond the last block) in the current extent.  If
3801          * this turns out to be false, we can bail out from this
3802          * function immediately.
3803          */
3804         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3805             ext4_ext_get_actual_len(last_ex))
3806                 return 0;
3807         /*
3808          * If the caller does appear to be planning to write at or
3809          * beyond the end of the current extent, we then test to see
3810          * if the current extent is the last extent in the file, by
3811          * checking to make sure it was reached via the rightmost node
3812          * at each level of the tree.
3813          */
3814         for (i = depth-1; i >= 0; i--)
3815                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3816                         return 0;
3817 out:
3818         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3819         return ext4_mark_inode_dirty(handle, inode);
3820 }
3821
3822 /**
3823  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3824  *
3825  * Return 1 if there is a delalloc block in the range, otherwise 0.
3826  */
3827 int ext4_find_delalloc_range(struct inode *inode,
3828                              ext4_lblk_t lblk_start,
3829                              ext4_lblk_t lblk_end)
3830 {
3831         struct extent_status es;
3832
3833         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3834         if (es.es_len == 0)
3835                 return 0; /* there is no delay extent in this tree */
3836         else if (es.es_lblk <= lblk_start &&
3837                  lblk_start < es.es_lblk + es.es_len)
3838                 return 1;
3839         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3840                 return 1;
3841         else
3842                 return 0;
3843 }
3844
3845 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3846 {
3847         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3848         ext4_lblk_t lblk_start, lblk_end;
3849         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3850         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3851
3852         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3853 }
3854
3855 /**
3856  * Determines how many complete clusters (out of those specified by the 'map')
3857  * are under delalloc and were reserved quota for.
3858  * This function is called when we are writing out the blocks that were
3859  * originally written with their allocation delayed, but then the space was
3860  * allocated using fallocate() before the delayed allocation could be resolved.
3861  * The cases to look for are:
3862  * ('=' indicated delayed allocated blocks
3863  *  '-' indicates non-delayed allocated blocks)
3864  * (a) partial clusters towards beginning and/or end outside of allocated range
3865  *     are not delalloc'ed.
3866  *      Ex:
3867  *      |----c---=|====c====|====c====|===-c----|
3868  *               |++++++ allocated ++++++|
3869  *      ==> 4 complete clusters in above example
3870  *
3871  * (b) partial cluster (outside of allocated range) towards either end is
3872  *     marked for delayed allocation. In this case, we will exclude that
3873  *     cluster.
3874  *      Ex:
3875  *      |----====c========|========c========|
3876  *           |++++++ allocated ++++++|
3877  *      ==> 1 complete clusters in above example
3878  *
3879  *      Ex:
3880  *      |================c================|
3881  *            |++++++ allocated ++++++|
3882  *      ==> 0 complete clusters in above example
3883  *
3884  * The ext4_da_update_reserve_space will be called only if we
3885  * determine here that there were some "entire" clusters that span
3886  * this 'allocated' range.
3887  * In the non-bigalloc case, this function will just end up returning num_blks
3888  * without ever calling ext4_find_delalloc_range.
3889  */
3890 static unsigned int
3891 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3892                            unsigned int num_blks)
3893 {
3894         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3895         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3896         ext4_lblk_t lblk_from, lblk_to, c_offset;
3897         unsigned int allocated_clusters = 0;
3898
3899         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3900         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3901
3902         /* max possible clusters for this allocation */
3903         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3904
3905         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3906
3907         /* Check towards left side */
3908         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3909         if (c_offset) {
3910                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3911                 lblk_to = lblk_from + c_offset - 1;
3912
3913                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3914                         allocated_clusters--;
3915         }
3916
3917         /* Now check towards right. */
3918         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3919         if (allocated_clusters && c_offset) {
3920                 lblk_from = lblk_start + num_blks;
3921                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3922
3923                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3924                         allocated_clusters--;
3925         }
3926
3927         return allocated_clusters;
3928 }
3929
3930 static int
3931 convert_initialized_extent(handle_t *handle, struct inode *inode,
3932                            struct ext4_map_blocks *map,
3933                            struct ext4_ext_path **ppath,
3934                            unsigned int allocated)
3935 {
3936         struct ext4_ext_path *path = *ppath;
3937         struct ext4_extent *ex;
3938         ext4_lblk_t ee_block;
3939         unsigned int ee_len;
3940         int depth;
3941         int err = 0;
3942
3943         /*
3944          * Make sure that the extent is no bigger than we support with
3945          * unwritten extent
3946          */
3947         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3948                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3949
3950         depth = ext_depth(inode);
3951         ex = path[depth].p_ext;
3952         ee_block = le32_to_cpu(ex->ee_block);
3953         ee_len = ext4_ext_get_actual_len(ex);
3954
3955         ext_debug("%s: inode %lu, logical"
3956                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3957                   (unsigned long long)ee_block, ee_len);
3958
3959         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3960                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3961                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3962                 if (err < 0)
3963                         return err;
3964                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3965                 if (IS_ERR(path))
3966                         return PTR_ERR(path);
3967                 depth = ext_depth(inode);
3968                 ex = path[depth].p_ext;
3969                 if (!ex) {
3970                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3971                                          (unsigned long) map->m_lblk);
3972                         return -EFSCORRUPTED;
3973                 }
3974         }
3975
3976         err = ext4_ext_get_access(handle, inode, path + depth);
3977         if (err)
3978                 return err;
3979         /* first mark the extent as unwritten */
3980         ext4_ext_mark_unwritten(ex);
3981
3982         /* note: ext4_ext_correct_indexes() isn't needed here because
3983          * borders are not changed
3984          */
3985         ext4_ext_try_to_merge(handle, inode, path, ex);
3986
3987         /* Mark modified extent as dirty */
3988         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3989         if (err)
3990                 return err;
3991         ext4_ext_show_leaf(inode, path);
3992
3993         ext4_update_inode_fsync_trans(handle, inode, 1);
3994         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3995         if (err)
3996                 return err;
3997         map->m_flags |= EXT4_MAP_UNWRITTEN;
3998         if (allocated > map->m_len)
3999                 allocated = map->m_len;
4000         map->m_len = allocated;
4001         return allocated;
4002 }
4003
4004 static int
4005 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4006                         struct ext4_map_blocks *map,
4007                         struct ext4_ext_path **ppath, int flags,
4008                         unsigned int allocated, ext4_fsblk_t newblock)
4009 {
4010         struct ext4_ext_path *path = *ppath;
4011         int ret = 0;
4012         int err = 0;
4013
4014         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4015                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4016                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4017                   flags, allocated);
4018         ext4_ext_show_leaf(inode, path);
4019
4020         /*
4021          * When writing into unwritten space, we should not fail to
4022          * allocate metadata blocks for the new extent block if needed.
4023          */
4024         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4025
4026         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4027                                                     allocated, newblock);
4028
4029         /* get_block() before submit the IO, split the extent */
4030         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4031                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4032                                          flags | EXT4_GET_BLOCKS_CONVERT);
4033                 if (ret <= 0)
4034                         goto out;
4035                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4036                 goto out;
4037         }
4038         /* IO end_io complete, convert the filled extent to written */
4039         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4040                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4041                         if (allocated > map->m_len)
4042                                 allocated = map->m_len;
4043                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4044                                                  allocated);
4045                         if (err < 0)
4046                                 goto out2;
4047                 }
4048                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4049                                                            ppath);
4050                 if (ret >= 0) {
4051                         ext4_update_inode_fsync_trans(handle, inode, 1);
4052                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4053                                                  path, map->m_len);
4054                 } else
4055                         err = ret;
4056                 map->m_flags |= EXT4_MAP_MAPPED;
4057                 map->m_pblk = newblock;
4058                 if (allocated > map->m_len)
4059                         allocated = map->m_len;
4060                 map->m_len = allocated;
4061                 goto out2;
4062         }
4063         /* buffered IO case */
4064         /*
4065          * repeat fallocate creation request
4066          * we already have an unwritten extent
4067          */
4068         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4069                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4070                 goto map_out;
4071         }
4072
4073         /* buffered READ or buffered write_begin() lookup */
4074         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4075                 /*
4076                  * We have blocks reserved already.  We
4077                  * return allocated blocks so that delalloc
4078                  * won't do block reservation for us.  But
4079                  * the buffer head will be unmapped so that
4080                  * a read from the block returns 0s.
4081                  */
4082                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4083                 goto out1;
4084         }
4085
4086         /* buffered write, writepage time, convert*/
4087         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4088         if (ret >= 0)
4089                 ext4_update_inode_fsync_trans(handle, inode, 1);
4090 out:
4091         if (ret <= 0) {
4092                 err = ret;
4093                 goto out2;
4094         } else
4095                 allocated = ret;
4096         map->m_flags |= EXT4_MAP_NEW;
4097         /*
4098          * if we allocated more blocks than requested
4099          * we need to make sure we unmap the extra block
4100          * allocated. The actual needed block will get
4101          * unmapped later when we find the buffer_head marked
4102          * new.
4103          */
4104         if (allocated > map->m_len) {
4105                 clean_bdev_aliases(inode->i_sb->s_bdev, newblock + map->m_len,
4106                                    allocated - map->m_len);
4107                 allocated = map->m_len;
4108         }
4109         map->m_len = allocated;
4110
4111         /*
4112          * If we have done fallocate with the offset that is already
4113          * delayed allocated, we would have block reservation
4114          * and quota reservation done in the delayed write path.
4115          * But fallocate would have already updated quota and block
4116          * count for this offset. So cancel these reservation
4117          */
4118         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4119                 unsigned int reserved_clusters;
4120                 reserved_clusters = get_reserved_cluster_alloc(inode,
4121                                 map->m_lblk, map->m_len);
4122                 if (reserved_clusters)
4123                         ext4_da_update_reserve_space(inode,
4124                                                      reserved_clusters,
4125                                                      0);
4126         }
4127
4128 map_out:
4129         map->m_flags |= EXT4_MAP_MAPPED;
4130         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4131                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4132                                          map->m_len);
4133                 if (err < 0)
4134                         goto out2;
4135         }
4136 out1:
4137         if (allocated > map->m_len)
4138                 allocated = map->m_len;
4139         ext4_ext_show_leaf(inode, path);
4140         map->m_pblk = newblock;
4141         map->m_len = allocated;
4142 out2:
4143         return err ? err : allocated;
4144 }
4145
4146 /*
4147  * get_implied_cluster_alloc - check to see if the requested
4148  * allocation (in the map structure) overlaps with a cluster already
4149  * allocated in an extent.
4150  *      @sb     The filesystem superblock structure
4151  *      @map    The requested lblk->pblk mapping
4152  *      @ex     The extent structure which might contain an implied
4153  *                      cluster allocation
4154  *
4155  * This function is called by ext4_ext_map_blocks() after we failed to
4156  * find blocks that were already in the inode's extent tree.  Hence,
4157  * we know that the beginning of the requested region cannot overlap
4158  * the extent from the inode's extent tree.  There are three cases we
4159  * want to catch.  The first is this case:
4160  *
4161  *               |--- cluster # N--|
4162  *    |--- extent ---|  |---- requested region ---|
4163  *                      |==========|
4164  *
4165  * The second case that we need to test for is this one:
4166  *
4167  *   |--------- cluster # N ----------------|
4168  *         |--- requested region --|   |------- extent ----|
4169  *         |=======================|
4170  *
4171  * The third case is when the requested region lies between two extents
4172  * within the same cluster:
4173  *          |------------- cluster # N-------------|
4174  * |----- ex -----|                  |---- ex_right ----|
4175  *                  |------ requested region ------|
4176  *                  |================|
4177  *
4178  * In each of the above cases, we need to set the map->m_pblk and
4179  * map->m_len so it corresponds to the return the extent labelled as
4180  * "|====|" from cluster #N, since it is already in use for data in
4181  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4182  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4183  * as a new "allocated" block region.  Otherwise, we will return 0 and
4184  * ext4_ext_map_blocks() will then allocate one or more new clusters
4185  * by calling ext4_mb_new_blocks().
4186  */
4187 static int get_implied_cluster_alloc(struct super_block *sb,
4188                                      struct ext4_map_blocks *map,
4189                                      struct ext4_extent *ex,
4190                                      struct ext4_ext_path *path)
4191 {
4192         struct ext4_sb_info *sbi = EXT4_SB(sb);
4193         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4194         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4195         ext4_lblk_t rr_cluster_start;
4196         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4197         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4198         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4199
4200         /* The extent passed in that we are trying to match */
4201         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4202         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4203
4204         /* The requested region passed into ext4_map_blocks() */
4205         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4206
4207         if ((rr_cluster_start == ex_cluster_end) ||
4208             (rr_cluster_start == ex_cluster_start)) {
4209                 if (rr_cluster_start == ex_cluster_end)
4210                         ee_start += ee_len - 1;
4211                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4212                 map->m_len = min(map->m_len,
4213                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4214                 /*
4215                  * Check for and handle this case:
4216                  *
4217                  *   |--------- cluster # N-------------|
4218                  *                     |------- extent ----|
4219                  *         |--- requested region ---|
4220                  *         |===========|
4221                  */
4222
4223                 if (map->m_lblk < ee_block)
4224                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4225
4226                 /*
4227                  * Check for the case where there is already another allocated
4228                  * block to the right of 'ex' but before the end of the cluster.
4229                  *
4230                  *          |------------- cluster # N-------------|
4231                  * |----- ex -----|                  |---- ex_right ----|
4232                  *                  |------ requested region ------|
4233                  *                  |================|
4234                  */
4235                 if (map->m_lblk > ee_block) {
4236                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4237                         map->m_len = min(map->m_len, next - map->m_lblk);
4238                 }
4239
4240                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4241                 return 1;
4242         }
4243
4244         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4245         return 0;
4246 }
4247
4248
4249 /*
4250  * Block allocation/map/preallocation routine for extents based files
4251  *
4252  *
4253  * Need to be called with
4254  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4255  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4256  *
4257  * return > 0, number of of blocks already mapped/allocated
4258  *          if create == 0 and these are pre-allocated blocks
4259  *              buffer head is unmapped
4260  *          otherwise blocks are mapped
4261  *
4262  * return = 0, if plain look up failed (blocks have not been allocated)
4263  *          buffer head is unmapped
4264  *
4265  * return < 0, error case.
4266  */
4267 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4268                         struct ext4_map_blocks *map, int flags)
4269 {
4270         struct ext4_ext_path *path = NULL;
4271         struct ext4_extent newex, *ex, *ex2;
4272         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4273         ext4_fsblk_t newblock = 0;
4274         int free_on_err = 0, err = 0, depth, ret;
4275         unsigned int allocated = 0, offset = 0;
4276         unsigned int allocated_clusters = 0;
4277         struct ext4_allocation_request ar;
4278         ext4_lblk_t cluster_offset;
4279         bool map_from_cluster = false;
4280
4281         ext_debug("blocks %u/%u requested for inode %lu\n",
4282                   map->m_lblk, map->m_len, inode->i_ino);
4283         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4284
4285         /* find extent for this block */
4286         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4287         if (IS_ERR(path)) {
4288                 err = PTR_ERR(path);
4289                 path = NULL;
4290                 goto out2;
4291         }
4292
4293         depth = ext_depth(inode);
4294
4295         /*
4296          * consistent leaf must not be empty;
4297          * this situation is possible, though, _during_ tree modification;
4298          * this is why assert can't be put in ext4_find_extent()
4299          */
4300         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4301                 EXT4_ERROR_INODE(inode, "bad extent address "
4302                                  "lblock: %lu, depth: %d pblock %lld",
4303                                  (unsigned long) map->m_lblk, depth,
4304                                  path[depth].p_block);
4305                 err = -EFSCORRUPTED;
4306                 goto out2;
4307         }
4308
4309         ex = path[depth].p_ext;
4310         if (ex) {
4311                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4312                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4313                 unsigned short ee_len;
4314
4315
4316                 /*
4317                  * unwritten extents are treated as holes, except that
4318                  * we split out initialized portions during a write.
4319                  */
4320                 ee_len = ext4_ext_get_actual_len(ex);
4321
4322                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4323
4324                 /* if found extent covers block, simply return it */
4325                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4326                         newblock = map->m_lblk - ee_block + ee_start;
4327                         /* number of remaining blocks in the extent */
4328                         allocated = ee_len - (map->m_lblk - ee_block);
4329                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4330                                   ee_block, ee_len, newblock);
4331
4332                         /*
4333                          * If the extent is initialized check whether the
4334                          * caller wants to convert it to unwritten.
4335                          */
4336                         if ((!ext4_ext_is_unwritten(ex)) &&
4337                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4338                                 allocated = convert_initialized_extent(
4339                                                 handle, inode, map, &path,
4340                                                 allocated);
4341                                 goto out2;
4342                         } else if (!ext4_ext_is_unwritten(ex))
4343                                 goto out;
4344
4345                         ret = ext4_ext_handle_unwritten_extents(
4346                                 handle, inode, map, &path, flags,
4347                                 allocated, newblock);
4348                         if (ret < 0)
4349                                 err = ret;
4350                         else
4351                                 allocated = ret;
4352                         goto out2;
4353                 }
4354         }
4355
4356         /*
4357          * requested block isn't allocated yet;
4358          * we couldn't try to create block if create flag is zero
4359          */
4360         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4361                 ext4_lblk_t hole_start, hole_len;
4362
4363                 hole_start = map->m_lblk;
4364                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4365                 /*
4366                  * put just found gap into cache to speed up
4367                  * subsequent requests
4368                  */
4369                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4370
4371                 /* Update hole_len to reflect hole size after map->m_lblk */
4372                 if (hole_start != map->m_lblk)
4373                         hole_len -= map->m_lblk - hole_start;
4374                 map->m_pblk = 0;
4375                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4376
4377                 goto out2;
4378         }
4379
4380         /*
4381          * Okay, we need to do block allocation.
4382          */
4383         newex.ee_block = cpu_to_le32(map->m_lblk);
4384         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4385
4386         /*
4387          * If we are doing bigalloc, check to see if the extent returned
4388          * by ext4_find_extent() implies a cluster we can use.
4389          */
4390         if (cluster_offset && ex &&
4391             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4392                 ar.len = allocated = map->m_len;
4393                 newblock = map->m_pblk;
4394                 map_from_cluster = true;
4395                 goto got_allocated_blocks;
4396         }
4397
4398         /* find neighbour allocated blocks */
4399         ar.lleft = map->m_lblk;
4400         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4401         if (err)
4402                 goto out2;
4403         ar.lright = map->m_lblk;
4404         ex2 = NULL;
4405         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4406         if (err)
4407                 goto out2;
4408
4409         /* Check if the extent after searching to the right implies a
4410          * cluster we can use. */
4411         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4412             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4413                 ar.len = allocated = map->m_len;
4414                 newblock = map->m_pblk;
4415                 map_from_cluster = true;
4416                 goto got_allocated_blocks;
4417         }
4418
4419         /*
4420          * See if request is beyond maximum number of blocks we can have in
4421          * a single extent. For an initialized extent this limit is
4422          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4423          * EXT_UNWRITTEN_MAX_LEN.
4424          */
4425         if (map->m_len > EXT_INIT_MAX_LEN &&
4426             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4427                 map->m_len = EXT_INIT_MAX_LEN;
4428         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4429                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4430                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4431
4432         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4433         newex.ee_len = cpu_to_le16(map->m_len);
4434         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4435         if (err)
4436                 allocated = ext4_ext_get_actual_len(&newex);
4437         else
4438                 allocated = map->m_len;
4439
4440         /* allocate new block */
4441         ar.inode = inode;
4442         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4443         ar.logical = map->m_lblk;
4444         /*
4445          * We calculate the offset from the beginning of the cluster
4446          * for the logical block number, since when we allocate a
4447          * physical cluster, the physical block should start at the
4448          * same offset from the beginning of the cluster.  This is
4449          * needed so that future calls to get_implied_cluster_alloc()
4450          * work correctly.
4451          */
4452         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4453         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4454         ar.goal -= offset;
4455         ar.logical -= offset;
4456         if (S_ISREG(inode->i_mode))
4457                 ar.flags = EXT4_MB_HINT_DATA;
4458         else
4459                 /* disable in-core preallocation for non-regular files */
4460                 ar.flags = 0;
4461         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4462                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4463         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4464                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4465         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4466                 ar.flags |= EXT4_MB_USE_RESERVED;
4467         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4468         if (!newblock)
4469                 goto out2;
4470         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4471                   ar.goal, newblock, allocated);
4472         free_on_err = 1;
4473         allocated_clusters = ar.len;
4474         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4475         if (ar.len > allocated)
4476                 ar.len = allocated;
4477
4478 got_allocated_blocks:
4479         /* try to insert new extent into found leaf and return */
4480         ext4_ext_store_pblock(&newex, newblock + offset);
4481         newex.ee_len = cpu_to_le16(ar.len);
4482         /* Mark unwritten */
4483         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4484                 ext4_ext_mark_unwritten(&newex);
4485                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4486         }
4487
4488         err = 0;
4489         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4490                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4491                                          path, ar.len);
4492         if (!err)
4493                 err = ext4_ext_insert_extent(handle, inode, &path,
4494                                              &newex, flags);
4495
4496         if (err && free_on_err) {
4497                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4498                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4499                 /* free data blocks we just allocated */
4500                 /* not a good idea to call discard here directly,
4501                  * but otherwise we'd need to call it every free() */
4502                 ext4_discard_preallocations(inode);
4503                 ext4_free_blocks(handle, inode, NULL, newblock,
4504                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4505                 goto out2;
4506         }
4507
4508         /* previous routine could use block we allocated */
4509         newblock = ext4_ext_pblock(&newex);
4510         allocated = ext4_ext_get_actual_len(&newex);
4511         if (allocated > map->m_len)
4512                 allocated = map->m_len;
4513         map->m_flags |= EXT4_MAP_NEW;
4514
4515         /*
4516          * Update reserved blocks/metadata blocks after successful
4517          * block allocation which had been deferred till now.
4518          */
4519         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4520                 unsigned int reserved_clusters;
4521                 /*
4522                  * Check how many clusters we had reserved this allocated range
4523                  */
4524                 reserved_clusters = get_reserved_cluster_alloc(inode,
4525                                                 map->m_lblk, allocated);
4526                 if (!map_from_cluster) {
4527                         BUG_ON(allocated_clusters < reserved_clusters);
4528                         if (reserved_clusters < allocated_clusters) {
4529                                 struct ext4_inode_info *ei = EXT4_I(inode);
4530                                 int reservation = allocated_clusters -
4531                                                   reserved_clusters;
4532                                 /*
4533                                  * It seems we claimed few clusters outside of
4534                                  * the range of this allocation. We should give
4535                                  * it back to the reservation pool. This can
4536                                  * happen in the following case:
4537                                  *
4538                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4539                                  *   cluster has 4 blocks. Thus, the clusters
4540                                  *   are [0-3],[4-7],[8-11]...
4541                                  * * First comes delayed allocation write for
4542                                  *   logical blocks 10 & 11. Since there were no
4543                                  *   previous delayed allocated blocks in the
4544                                  *   range [8-11], we would reserve 1 cluster
4545                                  *   for this write.
4546                                  * * Next comes write for logical blocks 3 to 8.
4547                                  *   In this case, we will reserve 2 clusters
4548                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4549                                  *   that range has a delayed allocated blocks.
4550                                  *   Thus total reserved clusters now becomes 3.
4551                                  * * Now, during the delayed allocation writeout
4552                                  *   time, we will first write blocks [3-8] and
4553                                  *   allocate 3 clusters for writing these
4554                                  *   blocks. Also, we would claim all these
4555                                  *   three clusters above.
4556                                  * * Now when we come here to writeout the
4557                                  *   blocks [10-11], we would expect to claim
4558                                  *   the reservation of 1 cluster we had made
4559                                  *   (and we would claim it since there are no
4560                                  *   more delayed allocated blocks in the range
4561                                  *   [8-11]. But our reserved cluster count had
4562                                  *   already gone to 0.
4563                                  *
4564                                  *   Thus, at the step 4 above when we determine
4565                                  *   that there are still some unwritten delayed
4566                                  *   allocated blocks outside of our current
4567                                  *   block range, we should increment the
4568                                  *   reserved clusters count so that when the
4569                                  *   remaining blocks finally gets written, we
4570                                  *   could claim them.
4571                                  */
4572                                 dquot_reserve_block(inode,
4573                                                 EXT4_C2B(sbi, reservation));
4574                                 spin_lock(&ei->i_block_reservation_lock);
4575                                 ei->i_reserved_data_blocks += reservation;
4576                                 spin_unlock(&ei->i_block_reservation_lock);
4577                         }
4578                         /*
4579                          * We will claim quota for all newly allocated blocks.
4580                          * We're updating the reserved space *after* the
4581                          * correction above so we do not accidentally free
4582                          * all the metadata reservation because we might
4583                          * actually need it later on.
4584                          */
4585                         ext4_da_update_reserve_space(inode, allocated_clusters,
4586                                                         1);
4587                 }
4588         }
4589
4590         /*
4591          * Cache the extent and update transaction to commit on fdatasync only
4592          * when it is _not_ an unwritten extent.
4593          */
4594         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4595                 ext4_update_inode_fsync_trans(handle, inode, 1);
4596         else
4597                 ext4_update_inode_fsync_trans(handle, inode, 0);
4598 out:
4599         if (allocated > map->m_len)
4600                 allocated = map->m_len;
4601         ext4_ext_show_leaf(inode, path);
4602         map->m_flags |= EXT4_MAP_MAPPED;
4603         map->m_pblk = newblock;
4604         map->m_len = allocated;
4605 out2:
4606         ext4_ext_drop_refs(path);
4607         kfree(path);
4608
4609         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4610                                        err ? err : allocated);
4611         return err ? err : allocated;
4612 }
4613
4614 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4615 {
4616         struct super_block *sb = inode->i_sb;
4617         ext4_lblk_t last_block;
4618         int err = 0;
4619
4620         /*
4621          * TODO: optimization is possible here.
4622          * Probably we need not scan at all,
4623          * because page truncation is enough.
4624          */
4625
4626         /* we have to know where to truncate from in crash case */
4627         EXT4_I(inode)->i_disksize = inode->i_size;
4628         err = ext4_mark_inode_dirty(handle, inode);
4629         if (err)
4630                 return err;
4631
4632         last_block = (inode->i_size + sb->s_blocksize - 1)
4633                         >> EXT4_BLOCK_SIZE_BITS(sb);
4634 retry:
4635         err = ext4_es_remove_extent(inode, last_block,
4636                                     EXT_MAX_BLOCKS - last_block);
4637         if (err == -ENOMEM) {
4638                 cond_resched();
4639                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4640                 goto retry;
4641         }
4642         if (err)
4643                 return err;
4644         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4645 }
4646
4647 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4648                                   ext4_lblk_t len, loff_t new_size,
4649                                   int flags)
4650 {
4651         struct inode *inode = file_inode(file);
4652         handle_t *handle;
4653         int ret = 0;
4654         int ret2 = 0;
4655         int retries = 0;
4656         int depth = 0;
4657         struct ext4_map_blocks map;
4658         unsigned int credits;
4659         loff_t epos;
4660
4661         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4662         map.m_lblk = offset;
4663         map.m_len = len;
4664         /*
4665          * Don't normalize the request if it can fit in one extent so
4666          * that it doesn't get unnecessarily split into multiple
4667          * extents.
4668          */
4669         if (len <= EXT_UNWRITTEN_MAX_LEN)
4670                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4671
4672         /*
4673          * credits to insert 1 extent into extent tree
4674          */
4675         credits = ext4_chunk_trans_blocks(inode, len);
4676         depth = ext_depth(inode);
4677
4678 retry:
4679         while (ret >= 0 && len) {
4680                 /*
4681                  * Recalculate credits when extent tree depth changes.
4682                  */
4683                 if (depth != ext_depth(inode)) {
4684                         credits = ext4_chunk_trans_blocks(inode, len);
4685                         depth = ext_depth(inode);
4686                 }
4687
4688                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4689                                             credits);
4690                 if (IS_ERR(handle)) {
4691                         ret = PTR_ERR(handle);
4692                         break;
4693                 }
4694                 ret = ext4_map_blocks(handle, inode, &map, flags);
4695                 if (ret <= 0) {
4696                         ext4_debug("inode #%lu: block %u: len %u: "
4697                                    "ext4_ext_map_blocks returned %d",
4698                                    inode->i_ino, map.m_lblk,
4699                                    map.m_len, ret);
4700                         ext4_mark_inode_dirty(handle, inode);
4701                         ret2 = ext4_journal_stop(handle);
4702                         break;
4703                 }
4704                 map.m_lblk += ret;
4705                 map.m_len = len = len - ret;
4706                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4707                 inode->i_ctime = current_time(inode);
4708                 if (new_size) {
4709                         if (epos > new_size)
4710                                 epos = new_size;
4711                         if (ext4_update_inode_size(inode, epos) & 0x1)
4712                                 inode->i_mtime = inode->i_ctime;
4713                 } else {
4714                         if (epos > inode->i_size)
4715                                 ext4_set_inode_flag(inode,
4716                                                     EXT4_INODE_EOFBLOCKS);
4717                 }
4718                 ext4_mark_inode_dirty(handle, inode);
4719                 ext4_update_inode_fsync_trans(handle, inode, 1);
4720                 ret2 = ext4_journal_stop(handle);
4721                 if (ret2)
4722                         break;
4723         }
4724         if (ret == -ENOSPC &&
4725                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4726                 ret = 0;
4727                 goto retry;
4728         }
4729
4730         return ret > 0 ? ret2 : ret;
4731 }
4732
4733 static long ext4_zero_range(struct file *file, loff_t offset,
4734                             loff_t len, int mode)
4735 {
4736         struct inode *inode = file_inode(file);
4737         handle_t *handle = NULL;
4738         unsigned int max_blocks;
4739         loff_t new_size = 0;
4740         int ret = 0;
4741         int flags;
4742         int credits;
4743         int partial_begin, partial_end;
4744         loff_t start, end;
4745         ext4_lblk_t lblk;
4746         unsigned int blkbits = inode->i_blkbits;
4747
4748         trace_ext4_zero_range(inode, offset, len, mode);
4749
4750         if (!S_ISREG(inode->i_mode))
4751                 return -EINVAL;
4752
4753         /* Call ext4_force_commit to flush all data in case of data=journal. */
4754         if (ext4_should_journal_data(inode)) {
4755                 ret = ext4_force_commit(inode->i_sb);
4756                 if (ret)
4757                         return ret;
4758         }
4759
4760         /*
4761          * Round up offset. This is not fallocate, we neet to zero out
4762          * blocks, so convert interior block aligned part of the range to
4763          * unwritten and possibly manually zero out unaligned parts of the
4764          * range.
4765          */
4766         start = round_up(offset, 1 << blkbits);
4767         end = round_down((offset + len), 1 << blkbits);
4768
4769         if (start < offset || end > offset + len)
4770                 return -EINVAL;
4771         partial_begin = offset & ((1 << blkbits) - 1);
4772         partial_end = (offset + len) & ((1 << blkbits) - 1);
4773
4774         lblk = start >> blkbits;
4775         max_blocks = (end >> blkbits);
4776         if (max_blocks < lblk)
4777                 max_blocks = 0;
4778         else
4779                 max_blocks -= lblk;
4780
4781         inode_lock(inode);
4782
4783         /*
4784          * Indirect files do not support unwritten extnets
4785          */
4786         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4787                 ret = -EOPNOTSUPP;
4788                 goto out_mutex;
4789         }
4790
4791         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4792             (offset + len > i_size_read(inode) ||
4793              offset + len > EXT4_I(inode)->i_disksize)) {
4794                 new_size = offset + len;
4795                 ret = inode_newsize_ok(inode, new_size);
4796                 if (ret)
4797                         goto out_mutex;
4798         }
4799
4800         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4801         if (mode & FALLOC_FL_KEEP_SIZE)
4802                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4803
4804         /* Wait all existing dio workers, newcomers will block on i_mutex */
4805         inode_dio_wait(inode);
4806
4807         /* Preallocate the range including the unaligned edges */
4808         if (partial_begin || partial_end) {
4809                 ret = ext4_alloc_file_blocks(file,
4810                                 round_down(offset, 1 << blkbits) >> blkbits,
4811                                 (round_up((offset + len), 1 << blkbits) -
4812                                  round_down(offset, 1 << blkbits)) >> blkbits,
4813                                 new_size, flags);
4814                 if (ret)
4815                         goto out_mutex;
4816
4817         }
4818
4819         /* Zero range excluding the unaligned edges */
4820         if (max_blocks > 0) {
4821                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4822                           EXT4_EX_NOCACHE);
4823
4824                 /*
4825                  * Prevent page faults from reinstantiating pages we have
4826                  * released from page cache.
4827                  */
4828                 down_write(&EXT4_I(inode)->i_mmap_sem);
4829
4830                 ret = ext4_break_layouts(inode);
4831                 if (ret) {
4832                         up_write(&EXT4_I(inode)->i_mmap_sem);
4833                         goto out_mutex;
4834                 }
4835
4836                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4837                 if (ret) {
4838                         up_write(&EXT4_I(inode)->i_mmap_sem);
4839                         goto out_mutex;
4840                 }
4841                 /* Now release the pages and zero block aligned part of pages */
4842                 truncate_pagecache_range(inode, start, end - 1);
4843                 inode->i_mtime = inode->i_ctime = current_time(inode);
4844
4845                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4846                                              flags);
4847                 up_write(&EXT4_I(inode)->i_mmap_sem);
4848                 if (ret)
4849                         goto out_mutex;
4850         }
4851         if (!partial_begin && !partial_end)
4852                 goto out_mutex;
4853
4854         /*
4855          * In worst case we have to writeout two nonadjacent unwritten
4856          * blocks and update the inode
4857          */
4858         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4859         if (ext4_should_journal_data(inode))
4860                 credits += 2;
4861         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4862         if (IS_ERR(handle)) {
4863                 ret = PTR_ERR(handle);
4864                 ext4_std_error(inode->i_sb, ret);
4865                 goto out_mutex;
4866         }
4867
4868         inode->i_mtime = inode->i_ctime = current_time(inode);
4869         if (new_size) {
4870                 ext4_update_inode_size(inode, new_size);
4871         } else {
4872                 /*
4873                 * Mark that we allocate beyond EOF so the subsequent truncate
4874                 * can proceed even if the new size is the same as i_size.
4875                 */
4876                 if ((offset + len) > i_size_read(inode))
4877                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4878         }
4879         ext4_mark_inode_dirty(handle, inode);
4880
4881         /* Zero out partial block at the edges of the range */
4882         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4883         if (ret >= 0)
4884                 ext4_update_inode_fsync_trans(handle, inode, 1);
4885
4886         if (file->f_flags & O_SYNC)
4887                 ext4_handle_sync(handle);
4888
4889         ext4_journal_stop(handle);
4890 out_mutex:
4891         inode_unlock(inode);
4892         return ret;
4893 }
4894
4895 /*
4896  * preallocate space for a file. This implements ext4's fallocate file
4897  * operation, which gets called from sys_fallocate system call.
4898  * For block-mapped files, posix_fallocate should fall back to the method
4899  * of writing zeroes to the required new blocks (the same behavior which is
4900  * expected for file systems which do not support fallocate() system call).
4901  */
4902 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4903 {
4904         struct inode *inode = file_inode(file);
4905         loff_t new_size = 0;
4906         unsigned int max_blocks;
4907         int ret = 0;
4908         int flags;
4909         ext4_lblk_t lblk;
4910         unsigned int blkbits = inode->i_blkbits;
4911
4912         /*
4913          * Encrypted inodes can't handle collapse range or insert
4914          * range since we would need to re-encrypt blocks with a
4915          * different IV or XTS tweak (which are based on the logical
4916          * block number).
4917          *
4918          * XXX It's not clear why zero range isn't working, but we'll
4919          * leave it disabled for encrypted inodes for now.  This is a
4920          * bug we should fix....
4921          */
4922         if (ext4_encrypted_inode(inode) &&
4923             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4924                      FALLOC_FL_ZERO_RANGE)))
4925                 return -EOPNOTSUPP;
4926
4927         /* Return error if mode is not supported */
4928         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4929                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4930                      FALLOC_FL_INSERT_RANGE))
4931                 return -EOPNOTSUPP;
4932
4933         if (mode & FALLOC_FL_PUNCH_HOLE)
4934                 return ext4_punch_hole(inode, offset, len);
4935
4936         ret = ext4_convert_inline_data(inode);
4937         if (ret)
4938                 return ret;
4939
4940         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4941                 return ext4_collapse_range(inode, offset, len);
4942
4943         if (mode & FALLOC_FL_INSERT_RANGE)
4944                 return ext4_insert_range(inode, offset, len);
4945
4946         if (mode & FALLOC_FL_ZERO_RANGE)
4947                 return ext4_zero_range(file, offset, len, mode);
4948
4949         trace_ext4_fallocate_enter(inode, offset, len, mode);
4950         lblk = offset >> blkbits;
4951
4952         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4953         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4954         if (mode & FALLOC_FL_KEEP_SIZE)
4955                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4956
4957         inode_lock(inode);
4958
4959         /*
4960          * We only support preallocation for extent-based files only
4961          */
4962         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4963                 ret = -EOPNOTSUPP;
4964                 goto out;
4965         }
4966
4967         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4968             (offset + len > i_size_read(inode) ||
4969              offset + len > EXT4_I(inode)->i_disksize)) {
4970                 new_size = offset + len;
4971                 ret = inode_newsize_ok(inode, new_size);
4972                 if (ret)
4973                         goto out;
4974         }
4975
4976         /* Wait all existing dio workers, newcomers will block on i_mutex */
4977         inode_dio_wait(inode);
4978
4979         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
4980         if (ret)
4981                 goto out;
4982
4983         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4984                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4985                                                 EXT4_I(inode)->i_sync_tid);
4986         }
4987 out:
4988         inode_unlock(inode);
4989         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4990         return ret;
4991 }
4992
4993 /*
4994  * This function convert a range of blocks to written extents
4995  * The caller of this function will pass the start offset and the size.
4996  * all unwritten extents within this range will be converted to
4997  * written extents.
4998  *
4999  * This function is called from the direct IO end io call back
5000  * function, to convert the fallocated extents after IO is completed.
5001  * Returns 0 on success.
5002  */
5003 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
5004                                    loff_t offset, ssize_t len)
5005 {
5006         unsigned int max_blocks;
5007         int ret = 0;
5008         int ret2 = 0;
5009         struct ext4_map_blocks map;
5010         unsigned int credits, blkbits = inode->i_blkbits;
5011
5012         map.m_lblk = offset >> blkbits;
5013         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
5014
5015         /*
5016          * This is somewhat ugly but the idea is clear: When transaction is
5017          * reserved, everything goes into it. Otherwise we rather start several
5018          * smaller transactions for conversion of each extent separately.
5019          */
5020         if (handle) {
5021                 handle = ext4_journal_start_reserved(handle,
5022                                                      EXT4_HT_EXT_CONVERT);
5023                 if (IS_ERR(handle))
5024                         return PTR_ERR(handle);
5025                 credits = 0;
5026         } else {
5027                 /*
5028                  * credits to insert 1 extent into extent tree
5029                  */
5030                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5031         }
5032         while (ret >= 0 && ret < max_blocks) {
5033                 map.m_lblk += ret;
5034                 map.m_len = (max_blocks -= ret);
5035                 if (credits) {
5036                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5037                                                     credits);
5038                         if (IS_ERR(handle)) {
5039                                 ret = PTR_ERR(handle);
5040                                 break;
5041                         }
5042                 }
5043                 ret = ext4_map_blocks(handle, inode, &map,
5044                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5045                 if (ret <= 0)
5046                         ext4_warning(inode->i_sb,
5047                                      "inode #%lu: block %u: len %u: "
5048                                      "ext4_ext_map_blocks returned %d",
5049                                      inode->i_ino, map.m_lblk,
5050                                      map.m_len, ret);
5051                 ext4_mark_inode_dirty(handle, inode);
5052                 if (credits)
5053                         ret2 = ext4_journal_stop(handle);
5054                 if (ret <= 0 || ret2)
5055                         break;
5056         }
5057         if (!credits)
5058                 ret2 = ext4_journal_stop(handle);
5059         return ret > 0 ? ret2 : ret;
5060 }
5061
5062 /*
5063  * If newes is not existing extent (newes->ec_pblk equals zero) find
5064  * delayed extent at start of newes and update newes accordingly and
5065  * return start of the next delayed extent.
5066  *
5067  * If newes is existing extent (newes->ec_pblk is not equal zero)
5068  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5069  * extent found. Leave newes unmodified.
5070  */
5071 static int ext4_find_delayed_extent(struct inode *inode,
5072                                     struct extent_status *newes)
5073 {
5074         struct extent_status es;
5075         ext4_lblk_t block, next_del;
5076
5077         if (newes->es_pblk == 0) {
5078                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5079                                 newes->es_lblk + newes->es_len - 1, &es);
5080
5081                 /*
5082                  * No extent in extent-tree contains block @newes->es_pblk,
5083                  * then the block may stay in 1)a hole or 2)delayed-extent.
5084                  */
5085                 if (es.es_len == 0)
5086                         /* A hole found. */
5087                         return 0;
5088
5089                 if (es.es_lblk > newes->es_lblk) {
5090                         /* A hole found. */
5091                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5092                                             newes->es_len);
5093                         return 0;
5094                 }
5095
5096                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5097         }
5098
5099         block = newes->es_lblk + newes->es_len;
5100         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5101         if (es.es_len == 0)
5102                 next_del = EXT_MAX_BLOCKS;
5103         else
5104                 next_del = es.es_lblk;
5105
5106         return next_del;
5107 }
5108 /* fiemap flags we can handle specified here */
5109 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5110
5111 static int ext4_xattr_fiemap(struct inode *inode,
5112                                 struct fiemap_extent_info *fieinfo)
5113 {
5114         __u64 physical = 0;
5115         __u64 length;
5116         __u32 flags = FIEMAP_EXTENT_LAST;
5117         int blockbits = inode->i_sb->s_blocksize_bits;
5118         int error = 0;
5119
5120         /* in-inode? */
5121         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5122                 struct ext4_iloc iloc;
5123                 int offset;     /* offset of xattr in inode */
5124
5125                 error = ext4_get_inode_loc(inode, &iloc);
5126                 if (error)
5127                         return error;
5128                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5129                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5130                                 EXT4_I(inode)->i_extra_isize;
5131                 physical += offset;
5132                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5133                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5134                 brelse(iloc.bh);
5135         } else { /* external block */
5136                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5137                 length = inode->i_sb->s_blocksize;
5138         }
5139
5140         if (physical)
5141                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5142                                                 length, flags);
5143         return (error < 0 ? error : 0);
5144 }
5145
5146 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5147                 __u64 start, __u64 len)
5148 {
5149         ext4_lblk_t start_blk;
5150         int error = 0;
5151
5152         if (ext4_has_inline_data(inode)) {
5153                 int has_inline = 1;
5154
5155                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5156                                                 start, len);
5157
5158                 if (has_inline)
5159                         return error;
5160         }
5161
5162         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5163                 error = ext4_ext_precache(inode);
5164                 if (error)
5165                         return error;
5166         }
5167
5168         /* fallback to generic here if not in extents fmt */
5169         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5170                 return generic_block_fiemap(inode, fieinfo, start, len,
5171                         ext4_get_block);
5172
5173         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5174                 return -EBADR;
5175
5176         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5177                 error = ext4_xattr_fiemap(inode, fieinfo);
5178         } else {
5179                 ext4_lblk_t len_blks;
5180                 __u64 last_blk;
5181
5182                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5183                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5184                 if (last_blk >= EXT_MAX_BLOCKS)
5185                         last_blk = EXT_MAX_BLOCKS-1;
5186                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5187
5188                 /*
5189                  * Walk the extent tree gathering extent information
5190                  * and pushing extents back to the user.
5191                  */
5192                 error = ext4_fill_fiemap_extents(inode, start_blk,
5193                                                  len_blks, fieinfo);
5194         }
5195         return error;
5196 }
5197
5198 /*
5199  * ext4_access_path:
5200  * Function to access the path buffer for marking it dirty.
5201  * It also checks if there are sufficient credits left in the journal handle
5202  * to update path.
5203  */
5204 static int
5205 ext4_access_path(handle_t *handle, struct inode *inode,
5206                 struct ext4_ext_path *path)
5207 {
5208         int credits, err;
5209
5210         if (!ext4_handle_valid(handle))
5211                 return 0;
5212
5213         /*
5214          * Check if need to extend journal credits
5215          * 3 for leaf, sb, and inode plus 2 (bmap and group
5216          * descriptor) for each block group; assume two block
5217          * groups
5218          */
5219         if (handle->h_buffer_credits < 7) {
5220                 credits = ext4_writepage_trans_blocks(inode);
5221                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5222                 /* EAGAIN is success */
5223                 if (err && err != -EAGAIN)
5224                         return err;
5225         }
5226
5227         err = ext4_ext_get_access(handle, inode, path);
5228         return err;
5229 }
5230
5231 /*
5232  * ext4_ext_shift_path_extents:
5233  * Shift the extents of a path structure lying between path[depth].p_ext
5234  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5235  * if it is right shift or left shift operation.
5236  */
5237 static int
5238 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5239                             struct inode *inode, handle_t *handle,
5240                             enum SHIFT_DIRECTION SHIFT)
5241 {
5242         int depth, err = 0;
5243         struct ext4_extent *ex_start, *ex_last;
5244         bool update = 0;
5245         depth = path->p_depth;
5246
5247         while (depth >= 0) {
5248                 if (depth == path->p_depth) {
5249                         ex_start = path[depth].p_ext;
5250                         if (!ex_start)
5251                                 return -EFSCORRUPTED;
5252
5253                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5254
5255                         err = ext4_access_path(handle, inode, path + depth);
5256                         if (err)
5257                                 goto out;
5258
5259                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5260                                 update = 1;
5261
5262                         while (ex_start <= ex_last) {
5263                                 if (SHIFT == SHIFT_LEFT) {
5264                                         le32_add_cpu(&ex_start->ee_block,
5265                                                 -shift);
5266                                         /* Try to merge to the left. */
5267                                         if ((ex_start >
5268                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5269                                             &&
5270                                             ext4_ext_try_to_merge_right(inode,
5271                                             path, ex_start - 1))
5272                                                 ex_last--;
5273                                         else
5274                                                 ex_start++;
5275                                 } else {
5276                                         le32_add_cpu(&ex_last->ee_block, shift);
5277                                         ext4_ext_try_to_merge_right(inode, path,
5278                                                 ex_last);
5279                                         ex_last--;
5280                                 }
5281                         }
5282                         err = ext4_ext_dirty(handle, inode, path + depth);
5283                         if (err)
5284                                 goto out;
5285
5286                         if (--depth < 0 || !update)
5287                                 break;
5288                 }
5289
5290                 /* Update index too */
5291                 err = ext4_access_path(handle, inode, path + depth);
5292                 if (err)
5293                         goto out;
5294
5295                 if (SHIFT == SHIFT_LEFT)
5296                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5297                 else
5298                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5299                 err = ext4_ext_dirty(handle, inode, path + depth);
5300                 if (err)
5301                         goto out;
5302
5303                 /* we are done if current index is not a starting index */
5304                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5305                         break;
5306
5307                 depth--;
5308         }
5309
5310 out:
5311         return err;
5312 }
5313
5314 /*
5315  * ext4_ext_shift_extents:
5316  * All the extents which lies in the range from @start to the last allocated
5317  * block for the @inode are shifted either towards left or right (depending
5318  * upon @SHIFT) by @shift blocks.
5319  * On success, 0 is returned, error otherwise.
5320  */
5321 static int
5322 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5323                        ext4_lblk_t start, ext4_lblk_t shift,
5324                        enum SHIFT_DIRECTION SHIFT)
5325 {
5326         struct ext4_ext_path *path;
5327         int ret = 0, depth;
5328         struct ext4_extent *extent;
5329         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5330
5331         /* Let path point to the last extent */
5332         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5333                                 EXT4_EX_NOCACHE);
5334         if (IS_ERR(path))
5335                 return PTR_ERR(path);
5336
5337         depth = path->p_depth;
5338         extent = path[depth].p_ext;
5339         if (!extent)
5340                 goto out;
5341
5342         stop = le32_to_cpu(extent->ee_block);
5343
5344        /*
5345         * For left shifts, make sure the hole on the left is big enough to
5346         * accommodate the shift.  For right shifts, make sure the last extent
5347         * won't be shifted beyond EXT_MAX_BLOCKS.
5348         */
5349         if (SHIFT == SHIFT_LEFT) {
5350                 path = ext4_find_extent(inode, start - 1, &path,
5351                                         EXT4_EX_NOCACHE);
5352                 if (IS_ERR(path))
5353                         return PTR_ERR(path);
5354                 depth = path->p_depth;
5355                 extent =  path[depth].p_ext;
5356                 if (extent) {
5357                         ex_start = le32_to_cpu(extent->ee_block);
5358                         ex_end = le32_to_cpu(extent->ee_block) +
5359                                 ext4_ext_get_actual_len(extent);
5360                 } else {
5361                         ex_start = 0;
5362                         ex_end = 0;
5363                 }
5364
5365                 if ((start == ex_start && shift > ex_start) ||
5366                     (shift > start - ex_end)) {
5367                         ret = -EINVAL;
5368                         goto out;
5369                 }
5370         } else {
5371                 if (shift > EXT_MAX_BLOCKS -
5372                     (stop + ext4_ext_get_actual_len(extent))) {
5373                         ret = -EINVAL;
5374                         goto out;
5375                 }
5376         }
5377
5378         /*
5379          * In case of left shift, iterator points to start and it is increased
5380          * till we reach stop. In case of right shift, iterator points to stop
5381          * and it is decreased till we reach start.
5382          */
5383         if (SHIFT == SHIFT_LEFT)
5384                 iterator = &start;
5385         else
5386                 iterator = &stop;
5387
5388         /*
5389          * Its safe to start updating extents.  Start and stop are unsigned, so
5390          * in case of right shift if extent with 0 block is reached, iterator
5391          * becomes NULL to indicate the end of the loop.
5392          */
5393         while (iterator && start <= stop) {
5394                 path = ext4_find_extent(inode, *iterator, &path,
5395                                         EXT4_EX_NOCACHE);
5396                 if (IS_ERR(path))
5397                         return PTR_ERR(path);
5398                 depth = path->p_depth;
5399                 extent = path[depth].p_ext;
5400                 if (!extent) {
5401                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5402                                          (unsigned long) *iterator);
5403                         return -EFSCORRUPTED;
5404                 }
5405                 if (SHIFT == SHIFT_LEFT && *iterator >
5406                     le32_to_cpu(extent->ee_block)) {
5407                         /* Hole, move to the next extent */
5408                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5409                                 path[depth].p_ext++;
5410                         } else {
5411                                 *iterator = ext4_ext_next_allocated_block(path);
5412                                 continue;
5413                         }
5414                 }
5415
5416                 if (SHIFT == SHIFT_LEFT) {
5417                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5418                         *iterator = le32_to_cpu(extent->ee_block) +
5419                                         ext4_ext_get_actual_len(extent);
5420                 } else {
5421                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5422                         if (le32_to_cpu(extent->ee_block) > 0)
5423                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5424                         else
5425                                 /* Beginning is reached, end of the loop */
5426                                 iterator = NULL;
5427                         /* Update path extent in case we need to stop */
5428                         while (le32_to_cpu(extent->ee_block) < start)
5429                                 extent++;
5430                         path[depth].p_ext = extent;
5431                 }
5432                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5433                                 handle, SHIFT);
5434                 if (ret)
5435                         break;
5436         }
5437 out:
5438         ext4_ext_drop_refs(path);
5439         kfree(path);
5440         return ret;
5441 }
5442
5443 /*
5444  * ext4_collapse_range:
5445  * This implements the fallocate's collapse range functionality for ext4
5446  * Returns: 0 and non-zero on error.
5447  */
5448 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5449 {
5450         struct super_block *sb = inode->i_sb;
5451         ext4_lblk_t punch_start, punch_stop;
5452         handle_t *handle;
5453         unsigned int credits;
5454         loff_t new_size, ioffset;
5455         int ret;
5456
5457         /*
5458          * We need to test this early because xfstests assumes that a
5459          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5460          * system does not support collapse range.
5461          */
5462         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5463                 return -EOPNOTSUPP;
5464
5465         /* Collapse range works only on fs block size aligned offsets. */
5466         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5467             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5468                 return -EINVAL;
5469
5470         if (!S_ISREG(inode->i_mode))
5471                 return -EINVAL;
5472
5473         trace_ext4_collapse_range(inode, offset, len);
5474
5475         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5476         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5477
5478         /* Call ext4_force_commit to flush all data in case of data=journal. */
5479         if (ext4_should_journal_data(inode)) {
5480                 ret = ext4_force_commit(inode->i_sb);
5481                 if (ret)
5482                         return ret;
5483         }
5484
5485         inode_lock(inode);
5486         /*
5487          * There is no need to overlap collapse range with EOF, in which case
5488          * it is effectively a truncate operation
5489          */
5490         if (offset + len >= i_size_read(inode)) {
5491                 ret = -EINVAL;
5492                 goto out_mutex;
5493         }
5494
5495         /* Currently just for extent based files */
5496         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5497                 ret = -EOPNOTSUPP;
5498                 goto out_mutex;
5499         }
5500
5501         /* Wait for existing dio to complete */
5502         inode_dio_wait(inode);
5503
5504         /*
5505          * Prevent page faults from reinstantiating pages we have released from
5506          * page cache.
5507          */
5508         down_write(&EXT4_I(inode)->i_mmap_sem);
5509
5510         ret = ext4_break_layouts(inode);
5511         if (ret)
5512                 goto out_mmap;
5513
5514         /*
5515          * Need to round down offset to be aligned with page size boundary
5516          * for page size > block size.
5517          */
5518         ioffset = round_down(offset, PAGE_SIZE);
5519         /*
5520          * Write tail of the last page before removed range since it will get
5521          * removed from the page cache below.
5522          */
5523         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5524         if (ret)
5525                 goto out_mmap;
5526         /*
5527          * Write data that will be shifted to preserve them when discarding
5528          * page cache below. We are also protected from pages becoming dirty
5529          * by i_mmap_sem.
5530          */
5531         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5532                                            LLONG_MAX);
5533         if (ret)
5534                 goto out_mmap;
5535         truncate_pagecache(inode, ioffset);
5536
5537         credits = ext4_writepage_trans_blocks(inode);
5538         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5539         if (IS_ERR(handle)) {
5540                 ret = PTR_ERR(handle);
5541                 goto out_mmap;
5542         }
5543
5544         down_write(&EXT4_I(inode)->i_data_sem);
5545         ext4_discard_preallocations(inode);
5546
5547         ret = ext4_es_remove_extent(inode, punch_start,
5548                                     EXT_MAX_BLOCKS - punch_start);
5549         if (ret) {
5550                 up_write(&EXT4_I(inode)->i_data_sem);
5551                 goto out_stop;
5552         }
5553
5554         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5555         if (ret) {
5556                 up_write(&EXT4_I(inode)->i_data_sem);
5557                 goto out_stop;
5558         }
5559         ext4_discard_preallocations(inode);
5560
5561         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5562                                      punch_stop - punch_start, SHIFT_LEFT);
5563         if (ret) {
5564                 up_write(&EXT4_I(inode)->i_data_sem);
5565                 goto out_stop;
5566         }
5567
5568         new_size = i_size_read(inode) - len;
5569         i_size_write(inode, new_size);
5570         EXT4_I(inode)->i_disksize = new_size;
5571
5572         up_write(&EXT4_I(inode)->i_data_sem);
5573         if (IS_SYNC(inode))
5574                 ext4_handle_sync(handle);
5575         inode->i_mtime = inode->i_ctime = current_time(inode);
5576         ext4_mark_inode_dirty(handle, inode);
5577         ext4_update_inode_fsync_trans(handle, inode, 1);
5578
5579 out_stop:
5580         ext4_journal_stop(handle);
5581 out_mmap:
5582         up_write(&EXT4_I(inode)->i_mmap_sem);
5583 out_mutex:
5584         inode_unlock(inode);
5585         return ret;
5586 }
5587
5588 /*
5589  * ext4_insert_range:
5590  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5591  * The data blocks starting from @offset to the EOF are shifted by @len
5592  * towards right to create a hole in the @inode. Inode size is increased
5593  * by len bytes.
5594  * Returns 0 on success, error otherwise.
5595  */
5596 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5597 {
5598         struct super_block *sb = inode->i_sb;
5599         handle_t *handle;
5600         struct ext4_ext_path *path;
5601         struct ext4_extent *extent;
5602         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5603         unsigned int credits, ee_len;
5604         int ret = 0, depth, split_flag = 0;
5605         loff_t ioffset;
5606
5607         /*
5608          * We need to test this early because xfstests assumes that an
5609          * insert range of (0, 1) will return EOPNOTSUPP if the file
5610          * system does not support insert range.
5611          */
5612         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5613                 return -EOPNOTSUPP;
5614
5615         /* Insert range works only on fs block size aligned offsets. */
5616         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5617                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5618                 return -EINVAL;
5619
5620         if (!S_ISREG(inode->i_mode))
5621                 return -EOPNOTSUPP;
5622
5623         trace_ext4_insert_range(inode, offset, len);
5624
5625         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5626         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5627
5628         /* Call ext4_force_commit to flush all data in case of data=journal */
5629         if (ext4_should_journal_data(inode)) {
5630                 ret = ext4_force_commit(inode->i_sb);
5631                 if (ret)
5632                         return ret;
5633         }
5634
5635         inode_lock(inode);
5636         /* Currently just for extent based files */
5637         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5638                 ret = -EOPNOTSUPP;
5639                 goto out_mutex;
5640         }
5641
5642         /* Check for wrap through zero */
5643         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5644                 ret = -EFBIG;
5645                 goto out_mutex;
5646         }
5647
5648         /* Offset should be less than i_size */
5649         if (offset >= i_size_read(inode)) {
5650                 ret = -EINVAL;
5651                 goto out_mutex;
5652         }
5653
5654         /* Wait for existing dio to complete */
5655         inode_dio_wait(inode);
5656
5657         /*
5658          * Prevent page faults from reinstantiating pages we have released from
5659          * page cache.
5660          */
5661         down_write(&EXT4_I(inode)->i_mmap_sem);
5662
5663         ret = ext4_break_layouts(inode);
5664         if (ret)
5665                 goto out_mmap;
5666
5667         /*
5668          * Need to round down to align start offset to page size boundary
5669          * for page size > block size.
5670          */
5671         ioffset = round_down(offset, PAGE_SIZE);
5672         /* Write out all dirty pages */
5673         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5674                         LLONG_MAX);
5675         if (ret)
5676                 goto out_mmap;
5677         truncate_pagecache(inode, ioffset);
5678
5679         credits = ext4_writepage_trans_blocks(inode);
5680         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5681         if (IS_ERR(handle)) {
5682                 ret = PTR_ERR(handle);
5683                 goto out_mmap;
5684         }
5685
5686         /* Expand file to avoid data loss if there is error while shifting */
5687         inode->i_size += len;
5688         EXT4_I(inode)->i_disksize += len;
5689         inode->i_mtime = inode->i_ctime = current_time(inode);
5690         ret = ext4_mark_inode_dirty(handle, inode);
5691         if (ret)
5692                 goto out_stop;
5693
5694         down_write(&EXT4_I(inode)->i_data_sem);
5695         ext4_discard_preallocations(inode);
5696
5697         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5698         if (IS_ERR(path)) {
5699                 up_write(&EXT4_I(inode)->i_data_sem);
5700                 goto out_stop;
5701         }
5702
5703         depth = ext_depth(inode);
5704         extent = path[depth].p_ext;
5705         if (extent) {
5706                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5707                 ee_len = ext4_ext_get_actual_len(extent);
5708
5709                 /*
5710                  * If offset_lblk is not the starting block of extent, split
5711                  * the extent @offset_lblk
5712                  */
5713                 if ((offset_lblk > ee_start_lblk) &&
5714                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5715                         if (ext4_ext_is_unwritten(extent))
5716                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5717                                         EXT4_EXT_MARK_UNWRIT2;
5718                         ret = ext4_split_extent_at(handle, inode, &path,
5719                                         offset_lblk, split_flag,
5720                                         EXT4_EX_NOCACHE |
5721                                         EXT4_GET_BLOCKS_PRE_IO |
5722                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5723                 }
5724
5725                 ext4_ext_drop_refs(path);
5726                 kfree(path);
5727                 if (ret < 0) {
5728                         up_write(&EXT4_I(inode)->i_data_sem);
5729                         goto out_stop;
5730                 }
5731         } else {
5732                 ext4_ext_drop_refs(path);
5733                 kfree(path);
5734         }
5735
5736         ret = ext4_es_remove_extent(inode, offset_lblk,
5737                         EXT_MAX_BLOCKS - offset_lblk);
5738         if (ret) {
5739                 up_write(&EXT4_I(inode)->i_data_sem);
5740                 goto out_stop;
5741         }
5742
5743         /*
5744          * if offset_lblk lies in a hole which is at start of file, use
5745          * ee_start_lblk to shift extents
5746          */
5747         ret = ext4_ext_shift_extents(inode, handle,
5748                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5749                 len_lblk, SHIFT_RIGHT);
5750
5751         up_write(&EXT4_I(inode)->i_data_sem);
5752         if (IS_SYNC(inode))
5753                 ext4_handle_sync(handle);
5754         if (ret >= 0)
5755                 ext4_update_inode_fsync_trans(handle, inode, 1);
5756
5757 out_stop:
5758         ext4_journal_stop(handle);
5759 out_mmap:
5760         up_write(&EXT4_I(inode)->i_mmap_sem);
5761 out_mutex:
5762         inode_unlock(inode);
5763         return ret;
5764 }
5765
5766 /**
5767  * ext4_swap_extents - Swap extents between two inodes
5768  *
5769  * @inode1:     First inode
5770  * @inode2:     Second inode
5771  * @lblk1:      Start block for first inode
5772  * @lblk2:      Start block for second inode
5773  * @count:      Number of blocks to swap
5774  * @unwritten: Mark second inode's extents as unwritten after swap
5775  * @erp:        Pointer to save error value
5776  *
5777  * This helper routine does exactly what is promise "swap extents". All other
5778  * stuff such as page-cache locking consistency, bh mapping consistency or
5779  * extent's data copying must be performed by caller.
5780  * Locking:
5781  *              i_mutex is held for both inodes
5782  *              i_data_sem is locked for write for both inodes
5783  * Assumptions:
5784  *              All pages from requested range are locked for both inodes
5785  */
5786 int
5787 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5788                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5789                   ext4_lblk_t count, int unwritten, int *erp)
5790 {
5791         struct ext4_ext_path *path1 = NULL;
5792         struct ext4_ext_path *path2 = NULL;
5793         int replaced_count = 0;
5794
5795         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5796         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5797         BUG_ON(!inode_is_locked(inode1));
5798         BUG_ON(!inode_is_locked(inode2));
5799
5800         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5801         if (unlikely(*erp))
5802                 return 0;
5803         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5804         if (unlikely(*erp))
5805                 return 0;
5806
5807         while (count) {
5808                 struct ext4_extent *ex1, *ex2, tmp_ex;
5809                 ext4_lblk_t e1_blk, e2_blk;
5810                 int e1_len, e2_len, len;
5811                 int split = 0;
5812
5813                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5814                 if (IS_ERR(path1)) {
5815                         *erp = PTR_ERR(path1);
5816                         path1 = NULL;
5817                 finish:
5818                         count = 0;
5819                         goto repeat;
5820                 }
5821                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5822                 if (IS_ERR(path2)) {
5823                         *erp = PTR_ERR(path2);
5824                         path2 = NULL;
5825                         goto finish;
5826                 }
5827                 ex1 = path1[path1->p_depth].p_ext;
5828                 ex2 = path2[path2->p_depth].p_ext;
5829                 /* Do we have somthing to swap ? */
5830                 if (unlikely(!ex2 || !ex1))
5831                         goto finish;
5832
5833                 e1_blk = le32_to_cpu(ex1->ee_block);
5834                 e2_blk = le32_to_cpu(ex2->ee_block);
5835                 e1_len = ext4_ext_get_actual_len(ex1);
5836                 e2_len = ext4_ext_get_actual_len(ex2);
5837
5838                 /* Hole handling */
5839                 if (!in_range(lblk1, e1_blk, e1_len) ||
5840                     !in_range(lblk2, e2_blk, e2_len)) {
5841                         ext4_lblk_t next1, next2;
5842
5843                         /* if hole after extent, then go to next extent */
5844                         next1 = ext4_ext_next_allocated_block(path1);
5845                         next2 = ext4_ext_next_allocated_block(path2);
5846                         /* If hole before extent, then shift to that extent */
5847                         if (e1_blk > lblk1)
5848                                 next1 = e1_blk;
5849                         if (e2_blk > lblk2)
5850                                 next2 = e2_blk;
5851                         /* Do we have something to swap */
5852                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5853                                 goto finish;
5854                         /* Move to the rightest boundary */
5855                         len = next1 - lblk1;
5856                         if (len < next2 - lblk2)
5857                                 len = next2 - lblk2;
5858                         if (len > count)
5859                                 len = count;
5860                         lblk1 += len;
5861                         lblk2 += len;
5862                         count -= len;
5863                         goto repeat;
5864                 }
5865
5866                 /* Prepare left boundary */
5867                 if (e1_blk < lblk1) {
5868                         split = 1;
5869                         *erp = ext4_force_split_extent_at(handle, inode1,
5870                                                 &path1, lblk1, 0);
5871                         if (unlikely(*erp))
5872                                 goto finish;
5873                 }
5874                 if (e2_blk < lblk2) {
5875                         split = 1;
5876                         *erp = ext4_force_split_extent_at(handle, inode2,
5877                                                 &path2,  lblk2, 0);
5878                         if (unlikely(*erp))
5879                                 goto finish;
5880                 }
5881                 /* ext4_split_extent_at() may result in leaf extent split,
5882                  * path must to be revalidated. */
5883                 if (split)
5884                         goto repeat;
5885
5886                 /* Prepare right boundary */
5887                 len = count;
5888                 if (len > e1_blk + e1_len - lblk1)
5889                         len = e1_blk + e1_len - lblk1;
5890                 if (len > e2_blk + e2_len - lblk2)
5891                         len = e2_blk + e2_len - lblk2;
5892
5893                 if (len != e1_len) {
5894                         split = 1;
5895                         *erp = ext4_force_split_extent_at(handle, inode1,
5896                                                 &path1, lblk1 + len, 0);
5897                         if (unlikely(*erp))
5898                                 goto finish;
5899                 }
5900                 if (len != e2_len) {
5901                         split = 1;
5902                         *erp = ext4_force_split_extent_at(handle, inode2,
5903                                                 &path2, lblk2 + len, 0);
5904                         if (*erp)
5905                                 goto finish;
5906                 }
5907                 /* ext4_split_extent_at() may result in leaf extent split,
5908                  * path must to be revalidated. */
5909                 if (split)
5910                         goto repeat;
5911
5912                 BUG_ON(e2_len != e1_len);
5913                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5914                 if (unlikely(*erp))
5915                         goto finish;
5916                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5917                 if (unlikely(*erp))
5918                         goto finish;
5919
5920                 /* Both extents are fully inside boundaries. Swap it now */
5921                 tmp_ex = *ex1;
5922                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5923                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5924                 ex1->ee_len = cpu_to_le16(e2_len);
5925                 ex2->ee_len = cpu_to_le16(e1_len);
5926                 if (unwritten)
5927                         ext4_ext_mark_unwritten(ex2);
5928                 if (ext4_ext_is_unwritten(&tmp_ex))
5929                         ext4_ext_mark_unwritten(ex1);
5930
5931                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5932                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5933                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5934                                       path2->p_depth);
5935                 if (unlikely(*erp))
5936                         goto finish;
5937                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5938                                       path1->p_depth);
5939                 /*
5940                  * Looks scarry ah..? second inode already points to new blocks,
5941                  * and it was successfully dirtied. But luckily error may happen
5942                  * only due to journal error, so full transaction will be
5943                  * aborted anyway.
5944                  */
5945                 if (unlikely(*erp))
5946                         goto finish;
5947                 lblk1 += len;
5948                 lblk2 += len;
5949                 replaced_count += len;
5950                 count -= len;
5951
5952         repeat:
5953                 ext4_ext_drop_refs(path1);
5954                 kfree(path1);
5955                 ext4_ext_drop_refs(path2);
5956                 kfree(path2);
5957                 path1 = path2 = NULL;
5958         }
5959         return replaced_count;
5960 }