fs/ext4/extents.c

   1 /*
   2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3  * Written by Alex Tomas <alex@clusterfs.com>
   4  *
   5  * Architecture independence:
   6  *   Copyright (c) 2005, Bull S.A.
   7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public Licens
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <asm/uaccess.h>
  41 #include <linux/fiemap.h>
  42 #include <linux/backing-dev.h>
  43 #include "ext4_jbd2.h"
  44 #include "ext4_extents.h"
  45 #include "xattr.h"
  46
  47 #include <trace/events/ext4.h>
  48
  49 /*
  50  * used by extent splitting.
  51  */
  52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  53                                         due to ENOSPC */
  54 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
  55 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
  56
  57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  59
  60 static __le32 ext4_extent_block_csum(struct inode *inode,
  61                                      struct ext4_extent_header *eh)
  62 {
  63         struct ext4_inode_info *ei = EXT4_I(inode);
  64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  65         __u32 csum;
  66
  67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  68                            EXT4_EXTENT_TAIL_OFFSET(eh));
  69         return cpu_to_le32(csum);
  70 }
  71
  72 static int ext4_extent_block_csum_verify(struct inode *inode,
  73                                          struct ext4_extent_header *eh)
  74 {
  75         struct ext4_extent_tail *et;
  76
  77         if (!ext4_has_metadata_csum(inode->i_sb))
  78                 return 1;
  79
  80         et = find_ext4_extent_tail(eh);
  81         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  82                 return 0;
  83         return 1;
  84 }
  85
  86 static void ext4_extent_block_csum_set(struct inode *inode,
  87                                        struct ext4_extent_header *eh)
  88 {
  89         struct ext4_extent_tail *et;
  90
  91         if (!ext4_has_metadata_csum(inode->i_sb))
  92                 return;
  93
  94         et = find_ext4_extent_tail(eh);
  95         et->et_checksum = ext4_extent_block_csum(inode, eh);
  96 }
  97
  98 static int ext4_split_extent(handle_t *handle,
  99                                 struct inode *inode,
 100                                 struct ext4_ext_path **ppath,
 101                                 struct ext4_map_blocks *map,
 102                                 int split_flag,
 103                                 int flags);
 104
 105 static int ext4_split_extent_at(handle_t *handle,
 106                              struct inode *inode,
 107                              struct ext4_ext_path **ppath,
 108                              ext4_lblk_t split,
 109                              int split_flag,
 110                              int flags);
 111
 112 static int ext4_find_delayed_extent(struct inode *inode,
 113                                     struct extent_status *newes);
 114
 115 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 116                                             struct inode *inode,
 117                                             int needed)
 118 {
 119         int err;
 120
 121         if (!ext4_handle_valid(handle))
 122                 return 0;
 123         if (handle->h_buffer_credits > needed)
 124                 return 0;
 125         err = ext4_journal_extend(handle, needed);
 126         if (err <= 0)
 127                 return err;
 128         err = ext4_truncate_restart_trans(handle, inode, needed);
 129         if (err == 0)
 130                 err = -EAGAIN;
 131
 132         return err;
 133 }
 134
 135 /*
 136  * could return:
 137  *  - EROFS
 138  *  - ENOMEM
 139  */
 140 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 141                                 struct ext4_ext_path *path)
 142 {
 143         if (path->p_bh) {
 144                 /* path points to block */
 145                 BUFFER_TRACE(path->p_bh, "get_write_access");
 146                 return ext4_journal_get_write_access(handle, path->p_bh);
 147         }
 148         /* path points to leaf/index in inode body */
 149         /* we use in-core data, no need to protect them */
 150         return 0;
 151 }
 152
 153 /*
 154  * could return:
 155  *  - EROFS
 156  *  - ENOMEM
 157  *  - EIO
 158  */
 159 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 160                      struct inode *inode, struct ext4_ext_path *path)
 161 {
 162         int err;
 163
 164         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 165         if (path->p_bh) {
 166                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 167                 /* path points to block */
 168                 err = __ext4_handle_dirty_metadata(where, line, handle,
 169                                                    inode, path->p_bh);
 170         } else {
 171                 /* path points to leaf/index in inode body */
 172                 err = ext4_mark_inode_dirty(handle, inode);
 173         }
 174         return err;
 175 }
 176
 177 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 178                               struct ext4_ext_path *path,
 179                               ext4_lblk_t block)
 180 {
 181         if (path) {
 182                 int depth = path->p_depth;
 183                 struct ext4_extent *ex;
 184
 185                 /*
 186                  * Try to predict block placement assuming that we are
 187                  * filling in a file which will eventually be
 188                  * non-sparse --- i.e., in the case of libbfd writing
 189                  * an ELF object sections out-of-order but in a way
 190                  * the eventually results in a contiguous object or
 191                  * executable file, or some database extending a table
 192                  * space file.  However, this is actually somewhat
 193                  * non-ideal if we are writing a sparse file such as
 194                  * qemu or KVM writing a raw image file that is going
 195                  * to stay fairly sparse, since it will end up
 196                  * fragmenting the file system's free space.  Maybe we
 197                  * should have some hueristics or some way to allow
 198                  * userspace to pass a hint to file system,
 199                  * especially if the latter case turns out to be
 200                  * common.
 201                  */
 202                 ex = path[depth].p_ext;
 203                 if (ex) {
 204                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 205                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 206
 207                         if (block > ext_block)
 208                                 return ext_pblk + (block - ext_block);
 209                         else
 210                                 return ext_pblk - (ext_block - block);
 211                 }
 212
 213                 /* it looks like index is empty;
 214                  * try to find starting block from index itself */
 215                 if (path[depth].p_bh)
 216                         return path[depth].p_bh->b_blocknr;
 217         }
 218
 219         /* OK. use inode's group */
 220         return ext4_inode_to_goal_block(inode);
 221 }
 222
 223 /*
 224  * Allocation for a meta data block
 225  */
 226 static ext4_fsblk_t
 227 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 228                         struct ext4_ext_path *path,
 229                         struct ext4_extent *ex, int *err, unsigned int flags)
 230 {
 231         ext4_fsblk_t goal, newblock;
 232
 233         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 234         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 235                                         NULL, err);
 236         return newblock;
 237 }
 238
 239 static inline int ext4_ext_space_block(struct inode *inode, int check)
 240 {
 241         int size;
 242
 243         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 244                         / sizeof(struct ext4_extent);
 245 #ifdef AGGRESSIVE_TEST
 246         if (!check && size > 6)
 247                 size = 6;
 248 #endif
 249         return size;
 250 }
 251
 252 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 253 {
 254         int size;
 255
 256         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 257                         / sizeof(struct ext4_extent_idx);
 258 #ifdef AGGRESSIVE_TEST
 259         if (!check && size > 5)
 260                 size = 5;
 261 #endif
 262         return size;
 263 }
 264
 265 static inline int ext4_ext_space_root(struct inode *inode, int check)
 266 {
 267         int size;
 268
 269         size = sizeof(EXT4_I(inode)->i_data);
 270         size -= sizeof(struct ext4_extent_header);
 271         size /= sizeof(struct ext4_extent);
 272 #ifdef AGGRESSIVE_TEST
 273         if (!check && size > 3)
 274                 size = 3;
 275 #endif
 276         return size;
 277 }
 278
 279 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 280 {
 281         int size;
 282
 283         size = sizeof(EXT4_I(inode)->i_data);
 284         size -= sizeof(struct ext4_extent_header);
 285         size /= sizeof(struct ext4_extent_idx);
 286 #ifdef AGGRESSIVE_TEST
 287         if (!check && size > 4)
 288                 size = 4;
 289 #endif
 290         return size;
 291 }
 292
 293 static inline int
 294 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
 295                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
 296                            int nofail)
 297 {
 298         struct ext4_ext_path *path = *ppath;
 299         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
 300
 301         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
 302                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
 303                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
 304                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
 305 }
 306
 307 /*
 308  * Calculate the number of metadata blocks needed
 309  * to allocate @blocks
 310  * Worse case is one block per extent
 311  */
 312 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 313 {
 314         struct ext4_inode_info *ei = EXT4_I(inode);
 315         int idxs;
 316
 317         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 318                 / sizeof(struct ext4_extent_idx));
 319
 320         /*
 321          * If the new delayed allocation block is contiguous with the
 322          * previous da block, it can share index blocks with the
 323          * previous block, so we only need to allocate a new index
 324          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 325          * an additional index block, and at ldxs**3 blocks, yet
 326          * another index blocks.
 327          */
 328         if (ei->i_da_metadata_calc_len &&
 329             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 330                 int num = 0;
 331
 332                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 333                         num++;
 334                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 335                         num++;
 336                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 337                         num++;
 338                         ei->i_da_metadata_calc_len = 0;
 339                 } else
 340                         ei->i_da_metadata_calc_len++;
 341                 ei->i_da_metadata_calc_last_lblock++;
 342                 return num;
 343         }
 344
 345         /*
 346          * In the worst case we need a new set of index blocks at
 347          * every level of the inode's extent tree.
 348          */
 349         ei->i_da_metadata_calc_len = 1;
 350         ei->i_da_metadata_calc_last_lblock = lblock;
 351         return ext_depth(inode) + 1;
 352 }
 353
 354 static int
 355 ext4_ext_max_entries(struct inode *inode, int depth)
 356 {
 357         int max;
 358
 359         if (depth == ext_depth(inode)) {
 360                 if (depth == 0)
 361                         max = ext4_ext_space_root(inode, 1);
 362                 else
 363                         max = ext4_ext_space_root_idx(inode, 1);
 364         } else {
 365                 if (depth == 0)
 366                         max = ext4_ext_space_block(inode, 1);
 367                 else
 368                         max = ext4_ext_space_block_idx(inode, 1);
 369         }
 370
 371         return max;
 372 }
 373
 374 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 375 {
 376         ext4_fsblk_t block = ext4_ext_pblock(ext);
 377         int len = ext4_ext_get_actual_len(ext);
 378         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 379         ext4_lblk_t last = lblock + len - 1;
 380
 381         if (len == 0 || lblock > last)
 382                 return 0;
 383         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 384 }
 385
 386 static int ext4_valid_extent_idx(struct inode *inode,
 387                                 struct ext4_extent_idx *ext_idx)
 388 {
 389         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 390
 391         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 392 }
 393
 394 static int ext4_valid_extent_entries(struct inode *inode,
 395                                 struct ext4_extent_header *eh,
 396                                 int depth)
 397 {
 398         unsigned short entries;
 399         if (eh->eh_entries == 0)
 400                 return 1;
 401
 402         entries = le16_to_cpu(eh->eh_entries);
 403
 404         if (depth == 0) {
 405                 /* leaf entries */
 406                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 407                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 408                 ext4_fsblk_t pblock = 0;
 409                 ext4_lblk_t lblock = 0;
 410                 ext4_lblk_t prev = 0;
 411                 int len = 0;
 412                 while (entries) {
 413                         if (!ext4_valid_extent(inode, ext))
 414                                 return 0;
 415
 416                         /* Check for overlapping extents */
 417                         lblock = le32_to_cpu(ext->ee_block);
 418                         len = ext4_ext_get_actual_len(ext);
 419                         if ((lblock <= prev) && prev) {
 420                                 pblock = ext4_ext_pblock(ext);
 421                                 es->s_last_error_block = cpu_to_le64(pblock);
 422                                 return 0;
 423                         }
 424                         ext++;
 425                         entries--;
 426                         prev = lblock + len - 1;
 427                 }
 428         } else {
 429                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 430                 while (entries) {
 431                         if (!ext4_valid_extent_idx(inode, ext_idx))
 432                                 return 0;
 433                         ext_idx++;
 434                         entries--;
 435                 }
 436         }
 437         return 1;
 438 }
 439
 440 static int __ext4_ext_check(const char *function, unsigned int line,
 441                             struct inode *inode, struct ext4_extent_header *eh,
 442                             int depth, ext4_fsblk_t pblk)
 443 {
 444         const char *error_msg;
 445         int max = 0;
 446
 447         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 448                 error_msg = "invalid magic";
 449                 goto corrupted;
 450         }
 451         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 452                 error_msg = "unexpected eh_depth";
 453                 goto corrupted;
 454         }
 455         if (unlikely(eh->eh_max == 0)) {
 456                 error_msg = "invalid eh_max";
 457                 goto corrupted;
 458         }
 459         max = ext4_ext_max_entries(inode, depth);
 460         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 461                 error_msg = "too large eh_max";
 462                 goto corrupted;
 463         }
 464         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 465                 error_msg = "invalid eh_entries";
 466                 goto corrupted;
 467         }
 468         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 469                 error_msg = "invalid extent entries";
 470                 goto corrupted;
 471         }
 472         /* Verify checksum on non-root extent tree nodes */
 473         if (ext_depth(inode) != depth &&
 474             !ext4_extent_block_csum_verify(inode, eh)) {
 475                 error_msg = "extent tree corrupted";
 476                 goto corrupted;
 477         }
 478         return 0;
 479
 480 corrupted:
 481         ext4_error_inode(inode, function, line, 0,
 482                          "pblk %llu bad header/extent: %s - magic %x, "
 483                          "entries %u, max %u(%u), depth %u(%u)",
 484                          (unsigned long long) pblk, error_msg,
 485                          le16_to_cpu(eh->eh_magic),
 486                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 487                          max, le16_to_cpu(eh->eh_depth), depth);
 488         return -EIO;
 489 }
 490
 491 #define ext4_ext_check(inode, eh, depth, pblk)                  \
 492         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
 493
 494 int ext4_ext_check_inode(struct inode *inode)
 495 {
 496         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
 497 }
 498
 499 static struct buffer_head *
 500 __read_extent_tree_block(const char *function, unsigned int line,
 501                          struct inode *inode, ext4_fsblk_t pblk, int depth,
 502                          int flags)
 503 {
 504         struct buffer_head              *bh;
 505         int                             err;
 506
 507         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
 508         if (unlikely(!bh))
 509                 return ERR_PTR(-ENOMEM);
 510
 511         if (!bh_uptodate_or_lock(bh)) {
 512                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
 513                 err = bh_submit_read(bh);
 514                 if (err < 0)
 515                         goto errout;
 516         }
 517         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 518                 return bh;
 519         err = __ext4_ext_check(function, line, inode,
 520                                ext_block_hdr(bh), depth, pblk);
 521         if (err)
 522                 goto errout;
 523         set_buffer_verified(bh);
 524         /*
 525          * If this is a leaf block, cache all of its entries
 526          */
 527         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
 528                 struct ext4_extent_header *eh = ext_block_hdr(bh);
 529                 struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
 530                 ext4_lblk_t prev = 0;
 531                 int i;
 532
 533                 for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
 534                         unsigned int status = EXTENT_STATUS_WRITTEN;
 535                         ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
 536                         int len = ext4_ext_get_actual_len(ex);
 537
 538                         if (prev && (prev != lblk))
 539                                 ext4_es_cache_extent(inode, prev,
 540                                                      lblk - prev, ~0,
 541                                                      EXTENT_STATUS_HOLE);
 542
 543                         if (ext4_ext_is_unwritten(ex))
 544                                 status = EXTENT_STATUS_UNWRITTEN;
 545                         ext4_es_cache_extent(inode, lblk, len,
 546                                              ext4_ext_pblock(ex), status);
 547                         prev = lblk + len;
 548                 }
 549         }
 550         return bh;
 551 errout:
 552         put_bh(bh);
 553         return ERR_PTR(err);
 554
 555 }
 556
 557 #define read_extent_tree_block(inode, pblk, depth, flags)               \
 558         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 559                                  (depth), (flags))
 560
 561 /*
 562  * This function is called to cache a file's extent information in the
 563  * extent status tree
 564  */
 565 int ext4_ext_precache(struct inode *inode)
 566 {
 567         struct ext4_inode_info *ei = EXT4_I(inode);
 568         struct ext4_ext_path *path = NULL;
 569         struct buffer_head *bh;
 570         int i = 0, depth, ret = 0;
 571
 572         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 573                 return 0;       /* not an extent-mapped inode */
 574
 575         down_read(&ei->i_data_sem);
 576         depth = ext_depth(inode);
 577
 578         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
 579                        GFP_NOFS);
 580         if (path == NULL) {
 581                 up_read(&ei->i_data_sem);
 582                 return -ENOMEM;
 583         }
 584
 585         /* Don't cache anything if there are no external extent blocks */
 586         if (depth == 0)
 587                 goto out;
 588         path[0].p_hdr = ext_inode_hdr(inode);
 589         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
 590         if (ret)
 591                 goto out;
 592         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
 593         while (i >= 0) {
 594                 /*
 595                  * If this is a leaf block or we've reached the end of
 596                  * the index block, go up
 597                  */
 598                 if ((i == depth) ||
 599                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
 600                         brelse(path[i].p_bh);
 601                         path[i].p_bh = NULL;
 602                         i--;
 603                         continue;
 604                 }
 605                 bh = read_extent_tree_block(inode,
 606                                             ext4_idx_pblock(path[i].p_idx++),
 607                                             depth - i - 1,
 608                                             EXT4_EX_FORCE_CACHE);
 609                 if (IS_ERR(bh)) {
 610                         ret = PTR_ERR(bh);
 611                         break;
 612                 }
 613                 i++;
 614                 path[i].p_bh = bh;
 615                 path[i].p_hdr = ext_block_hdr(bh);
 616                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
 617         }
 618         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
 619 out:
 620         up_read(&ei->i_data_sem);
 621         ext4_ext_drop_refs(path);
 622         kfree(path);
 623         return ret;
 624 }
 625
 626 #ifdef EXT_DEBUG
 627 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 628 {
 629         int k, l = path->p_depth;
 630
 631         ext_debug("path:");
 632         for (k = 0; k <= l; k++, path++) {
 633                 if (path->p_idx) {
 634                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 635                             ext4_idx_pblock(path->p_idx));
 636                 } else if (path->p_ext) {
 637                         ext_debug("  %d:[%d]%d:%llu ",
 638                                   le32_to_cpu(path->p_ext->ee_block),
 639                                   ext4_ext_is_unwritten(path->p_ext),
 640                                   ext4_ext_get_actual_len(path->p_ext),
 641                                   ext4_ext_pblock(path->p_ext));
 642                 } else
 643                         ext_debug("  []");
 644         }
 645         ext_debug("\n");
 646 }
 647
 648 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 649 {
 650         int depth = ext_depth(inode);
 651         struct ext4_extent_header *eh;
 652         struct ext4_extent *ex;
 653         int i;
 654
 655         if (!path)
 656                 return;
 657
 658         eh = path[depth].p_hdr;
 659         ex = EXT_FIRST_EXTENT(eh);
 660
 661         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 662
 663         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 664                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 665                           ext4_ext_is_unwritten(ex),
 666                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 667         }
 668         ext_debug("\n");
 669 }
 670
 671 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 672                         ext4_fsblk_t newblock, int level)
 673 {
 674         int depth = ext_depth(inode);
 675         struct ext4_extent *ex;
 676
 677         if (depth != level) {
 678                 struct ext4_extent_idx *idx;
 679                 idx = path[level].p_idx;
 680                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 681                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 682                                         le32_to_cpu(idx->ei_block),
 683                                         ext4_idx_pblock(idx),
 684                                         newblock);
 685                         idx++;
 686                 }
 687
 688                 return;
 689         }
 690
 691         ex = path[depth].p_ext;
 692         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 693                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 694                                 le32_to_cpu(ex->ee_block),
 695                                 ext4_ext_pblock(ex),
 696                                 ext4_ext_is_unwritten(ex),
 697                                 ext4_ext_get_actual_len(ex),
 698                                 newblock);
 699                 ex++;
 700         }
 701 }
 702
 703 #else
 704 #define ext4_ext_show_path(inode, path)
 705 #define ext4_ext_show_leaf(inode, path)
 706 #define ext4_ext_show_move(inode, path, newblock, level)
 707 #endif
 708
 709 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 710 {
 711         int depth, i;
 712
 713         if (!path)
 714                 return;
 715         depth = path->p_depth;
 716         for (i = 0; i <= depth; i++, path++)
 717                 if (path->p_bh) {
 718                         brelse(path->p_bh);
 719                         path->p_bh = NULL;
 720                 }
 721 }
 722
 723 /*
 724  * ext4_ext_binsearch_idx:
 725  * binary search for the closest index of the given block
 726  * the header must be checked before calling this
 727  */
 728 static void
 729 ext4_ext_binsearch_idx(struct inode *inode,
 730                         struct ext4_ext_path *path, ext4_lblk_t block)
 731 {
 732         struct ext4_extent_header *eh = path->p_hdr;
 733         struct ext4_extent_idx *r, *l, *m;
 734
 735
 736         ext_debug("binsearch for %u(idx):  ", block);
 737
 738         l = EXT_FIRST_INDEX(eh) + 1;
 739         r = EXT_LAST_INDEX(eh);
 740         while (l <= r) {
 741                 m = l + (r - l) / 2;
 742                 if (block < le32_to_cpu(m->ei_block))
 743                         r = m - 1;
 744                 else
 745                         l = m + 1;
 746                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 747                                 m, le32_to_cpu(m->ei_block),
 748                                 r, le32_to_cpu(r->ei_block));
 749         }
 750
 751         path->p_idx = l - 1;
 752         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 753                   ext4_idx_pblock(path->p_idx));
 754
 755 #ifdef CHECK_BINSEARCH
 756         {
 757                 struct ext4_extent_idx *chix, *ix;
 758                 int k;
 759
 760                 chix = ix = EXT_FIRST_INDEX(eh);
 761                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 762                   if (k != 0 &&
 763                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 764                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 765                                        "first=0x%p\n", k,
 766                                        ix, EXT_FIRST_INDEX(eh));
 767                                 printk(KERN_DEBUG "%u <= %u\n",
 768                                        le32_to_cpu(ix->ei_block),
 769                                        le32_to_cpu(ix[-1].ei_block));
 770                         }
 771                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 772                                            <= le32_to_cpu(ix[-1].ei_block));
 773                         if (block < le32_to_cpu(ix->ei_block))
 774                                 break;
 775                         chix = ix;
 776                 }
 777                 BUG_ON(chix != path->p_idx);
 778         }
 779 #endif
 780
 781 }
 782
 783 /*
 784  * ext4_ext_binsearch:
 785  * binary search for closest extent of the given block
 786  * the header must be checked before calling this
 787  */
 788 static void
 789 ext4_ext_binsearch(struct inode *inode,
 790                 struct ext4_ext_path *path, ext4_lblk_t block)
 791 {
 792         struct ext4_extent_header *eh = path->p_hdr;
 793         struct ext4_extent *r, *l, *m;
 794
 795         if (eh->eh_entries == 0) {
 796                 /*
 797                  * this leaf is empty:
 798                  * we get such a leaf in split/add case
 799                  */
 800                 return;
 801         }
 802
 803         ext_debug("binsearch for %u:  ", block);
 804
 805         l = EXT_FIRST_EXTENT(eh) + 1;
 806         r = EXT_LAST_EXTENT(eh);
 807
 808         while (l <= r) {
 809                 m = l + (r - l) / 2;
 810                 if (block < le32_to_cpu(m->ee_block))
 811                         r = m - 1;
 812                 else
 813                         l = m + 1;
 814                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 815                                 m, le32_to_cpu(m->ee_block),
 816                                 r, le32_to_cpu(r->ee_block));
 817         }
 818
 819         path->p_ext = l - 1;
 820         ext_debug("  -> %d:%llu:[%d]%d ",
 821                         le32_to_cpu(path->p_ext->ee_block),
 822                         ext4_ext_pblock(path->p_ext),
 823                         ext4_ext_is_unwritten(path->p_ext),
 824                         ext4_ext_get_actual_len(path->p_ext));
 825
 826 #ifdef CHECK_BINSEARCH
 827         {
 828                 struct ext4_extent *chex, *ex;
 829                 int k;
 830
 831                 chex = ex = EXT_FIRST_EXTENT(eh);
 832                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 833                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 834                                           <= le32_to_cpu(ex[-1].ee_block));
 835                         if (block < le32_to_cpu(ex->ee_block))
 836                                 break;
 837                         chex = ex;
 838                 }
 839                 BUG_ON(chex != path->p_ext);
 840         }
 841 #endif
 842
 843 }
 844
 845 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 846 {
 847         struct ext4_extent_header *eh;
 848
 849         eh = ext_inode_hdr(inode);
 850         eh->eh_depth = 0;
 851         eh->eh_entries = 0;
 852         eh->eh_magic = EXT4_EXT_MAGIC;
 853         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 854         ext4_mark_inode_dirty(handle, inode);
 855         return 0;
 856 }
 857
 858 struct ext4_ext_path *
 859 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
 860                  struct ext4_ext_path **orig_path, int flags)
 861 {
 862         struct ext4_extent_header *eh;
 863         struct buffer_head *bh;
 864         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
 865         short int depth, i, ppos = 0;
 866         int ret;
 867
 868         eh = ext_inode_hdr(inode);
 869         depth = ext_depth(inode);
 870
 871         if (path) {
 872                 ext4_ext_drop_refs(path);
 873                 if (depth > path[0].p_maxdepth) {
 874                         kfree(path);
 875                         *orig_path = path = NULL;
 876                 }
 877         }
 878         if (!path) {
 879                 /* account possible depth increase */
 880                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 881                                 GFP_NOFS);
 882                 if (unlikely(!path))
 883                         return ERR_PTR(-ENOMEM);
 884                 path[0].p_maxdepth = depth + 1;
 885         }
 886         path[0].p_hdr = eh;
 887         path[0].p_bh = NULL;
 888
 889         i = depth;
 890         /* walk through the tree */
 891         while (i) {
 892                 ext_debug("depth %d: num %d, max %d\n",
 893                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 894
 895                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 896                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 897                 path[ppos].p_depth = i;
 898                 path[ppos].p_ext = NULL;
 899
 900                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
 901                                             flags);
 902                 if (unlikely(IS_ERR(bh))) {
 903                         ret = PTR_ERR(bh);
 904                         goto err;
 905                 }
 906
 907                 eh = ext_block_hdr(bh);
 908                 ppos++;
 909                 if (unlikely(ppos > depth)) {
 910                         put_bh(bh);
 911                         EXT4_ERROR_INODE(inode,
 912                                          "ppos %d > depth %d", ppos, depth);
 913                         ret = -EIO;
 914                         goto err;
 915                 }
 916                 path[ppos].p_bh = bh;
 917                 path[ppos].p_hdr = eh;
 918         }
 919
 920         path[ppos].p_depth = i;
 921         path[ppos].p_ext = NULL;
 922         path[ppos].p_idx = NULL;
 923
 924         /* find extent */
 925         ext4_ext_binsearch(inode, path + ppos, block);
 926         /* if not an empty leaf */
 927         if (path[ppos].p_ext)
 928                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 929
 930         ext4_ext_show_path(inode, path);
 931
 932         return path;
 933
 934 err:
 935         ext4_ext_drop_refs(path);
 936         kfree(path);
 937         if (orig_path)
 938                 *orig_path = NULL;
 939         return ERR_PTR(ret);
 940 }
 941
 942 /*
 943  * ext4_ext_insert_index:
 944  * insert new index [@logical;@ptr] into the block at @curp;
 945  * check where to insert: before @curp or after @curp
 946  */
 947 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 948                                  struct ext4_ext_path *curp,
 949                                  int logical, ext4_fsblk_t ptr)
 950 {
 951         struct ext4_extent_idx *ix;
 952         int len, err;
 953
 954         err = ext4_ext_get_access(handle, inode, curp);
 955         if (err)
 956                 return err;
 957
 958         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 959                 EXT4_ERROR_INODE(inode,
 960                                  "logical %d == ei_block %d!",
 961                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 962                 return -EIO;
 963         }
 964
 965         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 966                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 967                 EXT4_ERROR_INODE(inode,
 968                                  "eh_entries %d >= eh_max %d!",
 969                                  le16_to_cpu(curp->p_hdr->eh_entries),
 970                                  le16_to_cpu(curp->p_hdr->eh_max));
 971                 return -EIO;
 972         }
 973
 974         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 975                 /* insert after */
 976                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 977                 ix = curp->p_idx + 1;
 978         } else {
 979                 /* insert before */
 980                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 981                 ix = curp->p_idx;
 982         }
 983
 984         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 985         BUG_ON(len < 0);
 986         if (len > 0) {
 987                 ext_debug("insert new index %d: "
 988                                 "move %d indices from 0x%p to 0x%p\n",
 989                                 logical, len, ix, ix + 1);
 990                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 991         }
 992
 993         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 994                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 995                 return -EIO;
 996         }
 997
 998         ix->ei_block = cpu_to_le32(logical);
 999         ext4_idx_store_pblock(ix, ptr);
1000         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1001
1002         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1003                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1004                 return -EIO;
1005         }
1006
1007         err = ext4_ext_dirty(handle, inode, curp);
1008         ext4_std_error(inode->i_sb, err);
1009
1010         return err;
1011 }
1012
1013 /*
1014  * ext4_ext_split:
1015  * inserts new subtree into the path, using free index entry
1016  * at depth @at:
1017  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1018  * - makes decision where to split
1019  * - moves remaining extents and index entries (right to the split point)
1020  *   into the newly allocated blocks
1021  * - initializes subtree
1022  */
1023 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1024                           unsigned int flags,
1025                           struct ext4_ext_path *path,
1026                           struct ext4_extent *newext, int at)
1027 {
1028         struct buffer_head *bh = NULL;
1029         int depth = ext_depth(inode);
1030         struct ext4_extent_header *neh;
1031         struct ext4_extent_idx *fidx;
1032         int i = at, k, m, a;
1033         ext4_fsblk_t newblock, oldblock;
1034         __le32 border;
1035         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1036         int err = 0;
1037
1038         /* make decision: where to split? */
1039         /* FIXME: now decision is simplest: at current extent */
1040
1041         /* if current leaf will be split, then we should use
1042          * border from split point */
1043         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1044                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1045                 return -EIO;
1046         }
1047         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1048                 border = path[depth].p_ext[1].ee_block;
1049                 ext_debug("leaf will be split."
1050                                 " next leaf starts at %d\n",
1051                                   le32_to_cpu(border));
1052         } else {
1053                 border = newext->ee_block;
1054                 ext_debug("leaf will be added."
1055                                 " next leaf starts at %d\n",
1056                                 le32_to_cpu(border));
1057         }
1058
1059         /*
1060          * If error occurs, then we break processing
1061          * and mark filesystem read-only. index won't
1062          * be inserted and tree will be in consistent
1063          * state. Next mount will repair buffers too.
1064          */
1065
1066         /*
1067          * Get array to track all allocated blocks.
1068          * We need this to handle errors and free blocks
1069          * upon them.
1070          */
1071         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1072         if (!ablocks)
1073                 return -ENOMEM;
1074
1075         /* allocate all needed blocks */
1076         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1077         for (a = 0; a < depth - at; a++) {
1078                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1079                                                    newext, &err, flags);
1080                 if (newblock == 0)
1081                         goto cleanup;
1082                 ablocks[a] = newblock;
1083         }
1084
1085         /* initialize new leaf */
1086         newblock = ablocks[--a];
1087         if (unlikely(newblock == 0)) {
1088                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1089                 err = -EIO;
1090                 goto cleanup;
1091         }
1092         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1093         if (unlikely(!bh)) {
1094                 err = -ENOMEM;
1095                 goto cleanup;
1096         }
1097         lock_buffer(bh);
1098
1099         err = ext4_journal_get_create_access(handle, bh);
1100         if (err)
1101                 goto cleanup;
1102
1103         neh = ext_block_hdr(bh);
1104         neh->eh_entries = 0;
1105         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1106         neh->eh_magic = EXT4_EXT_MAGIC;
1107         neh->eh_depth = 0;
1108
1109         /* move remainder of path[depth] to the new leaf */
1110         if (unlikely(path[depth].p_hdr->eh_entries !=
1111                      path[depth].p_hdr->eh_max)) {
1112                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1113                                  path[depth].p_hdr->eh_entries,
1114                                  path[depth].p_hdr->eh_max);
1115                 err = -EIO;
1116                 goto cleanup;
1117         }
1118         /* start copy from next extent */
1119         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1120         ext4_ext_show_move(inode, path, newblock, depth);
1121         if (m) {
1122                 struct ext4_extent *ex;
1123                 ex = EXT_FIRST_EXTENT(neh);
1124                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1125                 le16_add_cpu(&neh->eh_entries, m);
1126         }
1127
1128         ext4_extent_block_csum_set(inode, neh);
1129         set_buffer_uptodate(bh);
1130         unlock_buffer(bh);
1131
1132         err = ext4_handle_dirty_metadata(handle, inode, bh);
1133         if (err)
1134                 goto cleanup;
1135         brelse(bh);
1136         bh = NULL;
1137
1138         /* correct old leaf */
1139         if (m) {
1140                 err = ext4_ext_get_access(handle, inode, path + depth);
1141                 if (err)
1142                         goto cleanup;
1143                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1144                 err = ext4_ext_dirty(handle, inode, path + depth);
1145                 if (err)
1146                         goto cleanup;
1147
1148         }
1149
1150         /* create intermediate indexes */
1151         k = depth - at - 1;
1152         if (unlikely(k < 0)) {
1153                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1154                 err = -EIO;
1155                 goto cleanup;
1156         }
1157         if (k)
1158                 ext_debug("create %d intermediate indices\n", k);
1159         /* insert new index into current index block */
1160         /* current depth stored in i var */
1161         i = depth - 1;
1162         while (k--) {
1163                 oldblock = newblock;
1164                 newblock = ablocks[--a];
1165                 bh = sb_getblk(inode->i_sb, newblock);
1166                 if (unlikely(!bh)) {
1167                         err = -ENOMEM;
1168                         goto cleanup;
1169                 }
1170                 lock_buffer(bh);
1171
1172                 err = ext4_journal_get_create_access(handle, bh);
1173                 if (err)
1174                         goto cleanup;
1175
1176                 neh = ext_block_hdr(bh);
1177                 neh->eh_entries = cpu_to_le16(1);
1178                 neh->eh_magic = EXT4_EXT_MAGIC;
1179                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1180                 neh->eh_depth = cpu_to_le16(depth - i);
1181                 fidx = EXT_FIRST_INDEX(neh);
1182                 fidx->ei_block = border;
1183                 ext4_idx_store_pblock(fidx, oldblock);
1184
1185                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1186                                 i, newblock, le32_to_cpu(border), oldblock);
1187
1188                 /* move remainder of path[i] to the new index block */
1189                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1190                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1191                         EXT4_ERROR_INODE(inode,
1192                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1193                                          le32_to_cpu(path[i].p_ext->ee_block));
1194                         err = -EIO;
1195                         goto cleanup;
1196                 }
1197                 /* start copy indexes */
1198                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1199                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1200                                 EXT_MAX_INDEX(path[i].p_hdr));
1201                 ext4_ext_show_move(inode, path, newblock, i);
1202                 if (m) {
1203                         memmove(++fidx, path[i].p_idx,
1204                                 sizeof(struct ext4_extent_idx) * m);
1205                         le16_add_cpu(&neh->eh_entries, m);
1206                 }
1207                 ext4_extent_block_csum_set(inode, neh);
1208                 set_buffer_uptodate(bh);
1209                 unlock_buffer(bh);
1210
1211                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1212                 if (err)
1213                         goto cleanup;
1214                 brelse(bh);
1215                 bh = NULL;
1216
1217                 /* correct old index */
1218                 if (m) {
1219                         err = ext4_ext_get_access(handle, inode, path + i);
1220                         if (err)
1221                                 goto cleanup;
1222                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1223                         err = ext4_ext_dirty(handle, inode, path + i);
1224                         if (err)
1225                                 goto cleanup;
1226                 }
1227
1228                 i--;
1229         }
1230
1231         /* insert new index */
1232         err = ext4_ext_insert_index(handle, inode, path + at,
1233                                     le32_to_cpu(border), newblock);
1234
1235 cleanup:
1236         if (bh) {
1237                 if (buffer_locked(bh))
1238                         unlock_buffer(bh);
1239                 brelse(bh);
1240         }
1241
1242         if (err) {
1243                 /* free all allocated blocks in error case */
1244                 for (i = 0; i < depth; i++) {
1245                         if (!ablocks[i])
1246                                 continue;
1247                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1248                                          EXT4_FREE_BLOCKS_METADATA);
1249                 }
1250         }
1251         kfree(ablocks);
1252
1253         return err;
1254 }
1255
1256 /*
1257  * ext4_ext_grow_indepth:
1258  * implements tree growing procedure:
1259  * - allocates new block
1260  * - moves top-level data (index block or leaf) into the new block
1261  * - initializes new top-level, creating index that points to the
1262  *   just created block
1263  */
1264 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1265                                  unsigned int flags)
1266 {
1267         struct ext4_extent_header *neh;
1268         struct buffer_head *bh;
1269         ext4_fsblk_t newblock, goal = 0;
1270         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1271         int err = 0;
1272
1273         /* Try to prepend new index to old one */
1274         if (ext_depth(inode))
1275                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1276         if (goal > le32_to_cpu(es->s_first_data_block)) {
1277                 flags |= EXT4_MB_HINT_TRY_GOAL;
1278                 goal--;
1279         } else
1280                 goal = ext4_inode_to_goal_block(inode);
1281         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1282                                         NULL, &err);
1283         if (newblock == 0)
1284                 return err;
1285
1286         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1287         if (unlikely(!bh))
1288                 return -ENOMEM;
1289         lock_buffer(bh);
1290
1291         err = ext4_journal_get_create_access(handle, bh);
1292         if (err) {
1293                 unlock_buffer(bh);
1294                 goto out;
1295         }
1296
1297         /* move top-level index/leaf into new block */
1298         memmove(bh->b_data, EXT4_I(inode)->i_data,
1299                 sizeof(EXT4_I(inode)->i_data));
1300
1301         /* set size of new block */
1302         neh = ext_block_hdr(bh);
1303         /* old root could have indexes or leaves
1304          * so calculate e_max right way */
1305         if (ext_depth(inode))
1306                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1307         else
1308                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1309         neh->eh_magic = EXT4_EXT_MAGIC;
1310         ext4_extent_block_csum_set(inode, neh);
1311         set_buffer_uptodate(bh);
1312         unlock_buffer(bh);
1313
1314         err = ext4_handle_dirty_metadata(handle, inode, bh);
1315         if (err)
1316                 goto out;
1317
1318         /* Update top-level index: num,max,pointer */
1319         neh = ext_inode_hdr(inode);
1320         neh->eh_entries = cpu_to_le16(1);
1321         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1322         if (neh->eh_depth == 0) {
1323                 /* Root extent block becomes index block */
1324                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1325                 EXT_FIRST_INDEX(neh)->ei_block =
1326                         EXT_FIRST_EXTENT(neh)->ee_block;
1327         }
1328         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1329                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1330                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1331                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1332
1333         le16_add_cpu(&neh->eh_depth, 1);
1334         ext4_mark_inode_dirty(handle, inode);
1335 out:
1336         brelse(bh);
1337
1338         return err;
1339 }
1340
1341 /*
1342  * ext4_ext_create_new_leaf:
1343  * finds empty index and adds new leaf.
1344  * if no free index is found, then it requests in-depth growing.
1345  */
1346 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1347                                     unsigned int mb_flags,
1348                                     unsigned int gb_flags,
1349                                     struct ext4_ext_path **ppath,
1350                                     struct ext4_extent *newext)
1351 {
1352         struct ext4_ext_path *path = *ppath;
1353         struct ext4_ext_path *curp;
1354         int depth, i, err = 0;
1355
1356 repeat:
1357         i = depth = ext_depth(inode);
1358
1359         /* walk up to the tree and look for free index entry */
1360         curp = path + depth;
1361         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1362                 i--;
1363                 curp--;
1364         }
1365
1366         /* we use already allocated block for index block,
1367          * so subsequent data blocks should be contiguous */
1368         if (EXT_HAS_FREE_INDEX(curp)) {
1369                 /* if we found index with free entry, then use that
1370                  * entry: create all needed subtree and add new leaf */
1371                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1372                 if (err)
1373                         goto out;
1374
1375                 /* refill path */
1376                 path = ext4_find_extent(inode,
1377                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1378                                     ppath, gb_flags);
1379                 if (IS_ERR(path))
1380                         err = PTR_ERR(path);
1381         } else {
1382                 /* tree is full, time to grow in depth */
1383                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1384                 if (err)
1385                         goto out;
1386
1387                 /* refill path */
1388                 path = ext4_find_extent(inode,
1389                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1390                                     ppath, gb_flags);
1391                 if (IS_ERR(path)) {
1392                         err = PTR_ERR(path);
1393                         goto out;
1394                 }
1395
1396                 /*
1397                  * only first (depth 0 -> 1) produces free space;
1398                  * in all other cases we have to split the grown tree
1399                  */
1400                 depth = ext_depth(inode);
1401                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1402                         /* now we need to split */
1403                         goto repeat;
1404                 }
1405         }
1406
1407 out:
1408         return err;
1409 }
1410
1411 /*
1412  * search the closest allocated block to the left for *logical
1413  * and returns it at @logical + it's physical address at @phys
1414  * if *logical is the smallest allocated block, the function
1415  * returns 0 at @phys
1416  * return value contains 0 (success) or error code
1417  */
1418 static int ext4_ext_search_left(struct inode *inode,
1419                                 struct ext4_ext_path *path,
1420                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1421 {
1422         struct ext4_extent_idx *ix;
1423         struct ext4_extent *ex;
1424         int depth, ee_len;
1425
1426         if (unlikely(path == NULL)) {
1427                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1428                 return -EIO;
1429         }
1430         depth = path->p_depth;
1431         *phys = 0;
1432
1433         if (depth == 0 && path->p_ext == NULL)
1434                 return 0;
1435
1436         /* usually extent in the path covers blocks smaller
1437          * then *logical, but it can be that extent is the
1438          * first one in the file */
1439
1440         ex = path[depth].p_ext;
1441         ee_len = ext4_ext_get_actual_len(ex);
1442         if (*logical < le32_to_cpu(ex->ee_block)) {
1443                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1444                         EXT4_ERROR_INODE(inode,
1445                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1446                                          *logical, le32_to_cpu(ex->ee_block));
1447                         return -EIO;
1448                 }
1449                 while (--depth >= 0) {
1450                         ix = path[depth].p_idx;
1451                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1452                                 EXT4_ERROR_INODE(inode,
1453                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1454                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1455                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1456                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1457                                   depth);
1458                                 return -EIO;
1459                         }
1460                 }
1461                 return 0;
1462         }
1463
1464         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1465                 EXT4_ERROR_INODE(inode,
1466                                  "logical %d < ee_block %d + ee_len %d!",
1467                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1468                 return -EIO;
1469         }
1470
1471         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1472         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1473         return 0;
1474 }
1475
1476 /*
1477  * search the closest allocated block to the right for *logical
1478  * and returns it at @logical + it's physical address at @phys
1479  * if *logical is the largest allocated block, the function
1480  * returns 0 at @phys
1481  * return value contains 0 (success) or error code
1482  */
1483 static int ext4_ext_search_right(struct inode *inode,
1484                                  struct ext4_ext_path *path,
1485                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1486                                  struct ext4_extent **ret_ex)
1487 {
1488         struct buffer_head *bh = NULL;
1489         struct ext4_extent_header *eh;
1490         struct ext4_extent_idx *ix;
1491         struct ext4_extent *ex;
1492         ext4_fsblk_t block;
1493         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1494         int ee_len;
1495
1496         if (unlikely(path == NULL)) {
1497                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1498                 return -EIO;
1499         }
1500         depth = path->p_depth;
1501         *phys = 0;
1502
1503         if (depth == 0 && path->p_ext == NULL)
1504                 return 0;
1505
1506         /* usually extent in the path covers blocks smaller
1507          * then *logical, but it can be that extent is the
1508          * first one in the file */
1509
1510         ex = path[depth].p_ext;
1511         ee_len = ext4_ext_get_actual_len(ex);
1512         if (*logical < le32_to_cpu(ex->ee_block)) {
1513                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1514                         EXT4_ERROR_INODE(inode,
1515                                          "first_extent(path[%d].p_hdr) != ex",
1516                                          depth);
1517                         return -EIO;
1518                 }
1519                 while (--depth >= 0) {
1520                         ix = path[depth].p_idx;
1521                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1522                                 EXT4_ERROR_INODE(inode,
1523                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1524                                                  *logical);
1525                                 return -EIO;
1526                         }
1527                 }
1528                 goto found_extent;
1529         }
1530
1531         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1532                 EXT4_ERROR_INODE(inode,
1533                                  "logical %d < ee_block %d + ee_len %d!",
1534                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1535                 return -EIO;
1536         }
1537
1538         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1539                 /* next allocated block in this leaf */
1540                 ex++;
1541                 goto found_extent;
1542         }
1543
1544         /* go up and search for index to the right */
1545         while (--depth >= 0) {
1546                 ix = path[depth].p_idx;
1547                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1548                         goto got_index;
1549         }
1550
1551         /* we've gone up to the root and found no index to the right */
1552         return 0;
1553
1554 got_index:
1555         /* we've found index to the right, let's
1556          * follow it and find the closest allocated
1557          * block to the right */
1558         ix++;
1559         block = ext4_idx_pblock(ix);
1560         while (++depth < path->p_depth) {
1561                 /* subtract from p_depth to get proper eh_depth */
1562                 bh = read_extent_tree_block(inode, block,
1563                                             path->p_depth - depth, 0);
1564                 if (IS_ERR(bh))
1565                         return PTR_ERR(bh);
1566                 eh = ext_block_hdr(bh);
1567                 ix = EXT_FIRST_INDEX(eh);
1568                 block = ext4_idx_pblock(ix);
1569                 put_bh(bh);
1570         }
1571
1572         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1573         if (IS_ERR(bh))
1574                 return PTR_ERR(bh);
1575         eh = ext_block_hdr(bh);
1576         ex = EXT_FIRST_EXTENT(eh);
1577 found_extent:
1578         *logical = le32_to_cpu(ex->ee_block);
1579         *phys = ext4_ext_pblock(ex);
1580         *ret_ex = ex;
1581         if (bh)
1582                 put_bh(bh);
1583         return 0;
1584 }
1585
1586 /*
1587  * ext4_ext_next_allocated_block:
1588  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1589  * NOTE: it considers block number from index entry as
1590  * allocated block. Thus, index entries have to be consistent
1591  * with leaves.
1592  */
1593 ext4_lblk_t
1594 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1595 {
1596         int depth;
1597
1598         BUG_ON(path == NULL);
1599         depth = path->p_depth;
1600
1601         if (depth == 0 && path->p_ext == NULL)
1602                 return EXT_MAX_BLOCKS;
1603
1604         while (depth >= 0) {
1605                 if (depth == path->p_depth) {
1606                         /* leaf */
1607                         if (path[depth].p_ext &&
1608                                 path[depth].p_ext !=
1609                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1610                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1611                 } else {
1612                         /* index */
1613                         if (path[depth].p_idx !=
1614                                         EXT_LAST_INDEX(path[depth].p_hdr))
1615                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1616                 }
1617                 depth--;
1618         }
1619
1620         return EXT_MAX_BLOCKS;
1621 }
1622
1623 /*
1624  * ext4_ext_next_leaf_block:
1625  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1626  */
1627 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1628 {
1629         int depth;
1630
1631         BUG_ON(path == NULL);
1632         depth = path->p_depth;
1633
1634         /* zero-tree has no leaf blocks at all */
1635         if (depth == 0)
1636                 return EXT_MAX_BLOCKS;
1637
1638         /* go to index block */
1639         depth--;
1640
1641         while (depth >= 0) {
1642                 if (path[depth].p_idx !=
1643                                 EXT_LAST_INDEX(path[depth].p_hdr))
1644                         return (ext4_lblk_t)
1645                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1646                 depth--;
1647         }
1648
1649         return EXT_MAX_BLOCKS;
1650 }
1651
1652 /*
1653  * ext4_ext_correct_indexes:
1654  * if leaf gets modified and modified extent is first in the leaf,
1655  * then we have to correct all indexes above.
1656  * TODO: do we need to correct tree in all cases?
1657  */
1658 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1659                                 struct ext4_ext_path *path)
1660 {
1661         struct ext4_extent_header *eh;
1662         int depth = ext_depth(inode);
1663         struct ext4_extent *ex;
1664         __le32 border;
1665         int k, err = 0;
1666
1667         eh = path[depth].p_hdr;
1668         ex = path[depth].p_ext;
1669
1670         if (unlikely(ex == NULL || eh == NULL)) {
1671                 EXT4_ERROR_INODE(inode,
1672                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1673                 return -EIO;
1674         }
1675
1676         if (depth == 0) {
1677                 /* there is no tree at all */
1678                 return 0;
1679         }
1680
1681         if (ex != EXT_FIRST_EXTENT(eh)) {
1682                 /* we correct tree if first leaf got modified only */
1683                 return 0;
1684         }
1685
1686         /*
1687          * TODO: we need correction if border is smaller than current one
1688          */
1689         k = depth - 1;
1690         border = path[depth].p_ext->ee_block;
1691         err = ext4_ext_get_access(handle, inode, path + k);
1692         if (err)
1693                 return err;
1694         path[k].p_idx->ei_block = border;
1695         err = ext4_ext_dirty(handle, inode, path + k);
1696         if (err)
1697                 return err;
1698
1699         while (k--) {
1700                 /* change all left-side indexes */
1701                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1702                         break;
1703                 err = ext4_ext_get_access(handle, inode, path + k);
1704                 if (err)
1705                         break;
1706                 path[k].p_idx->ei_block = border;
1707                 err = ext4_ext_dirty(handle, inode, path + k);
1708                 if (err)
1709                         break;
1710         }
1711
1712         return err;
1713 }
1714
1715 int
1716 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1717                                 struct ext4_extent *ex2)
1718 {
1719         unsigned short ext1_ee_len, ext2_ee_len;
1720
1721         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1722                 return 0;
1723
1724         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1725         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1726
1727         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1728                         le32_to_cpu(ex2->ee_block))
1729                 return 0;
1730
1731         /*
1732          * To allow future support for preallocated extents to be added
1733          * as an RO_COMPAT feature, refuse to merge to extents if
1734          * this can result in the top bit of ee_len being set.
1735          */
1736         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1737                 return 0;
1738         if (ext4_ext_is_unwritten(ex1) &&
1739             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1740              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1741              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1742                 return 0;
1743 #ifdef AGGRESSIVE_TEST
1744         if (ext1_ee_len >= 4)
1745                 return 0;
1746 #endif
1747
1748         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1749                 return 1;
1750         return 0;
1751 }
1752
1753 /*
1754  * This function tries to merge the "ex" extent to the next extent in the tree.
1755  * It always tries to merge towards right. If you want to merge towards
1756  * left, pass "ex - 1" as argument instead of "ex".
1757  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1758  * 1 if they got merged.
1759  */
1760 static int ext4_ext_try_to_merge_right(struct inode *inode,
1761                                  struct ext4_ext_path *path,
1762                                  struct ext4_extent *ex)
1763 {
1764         struct ext4_extent_header *eh;
1765         unsigned int depth, len;
1766         int merge_done = 0, unwritten;
1767
1768         depth = ext_depth(inode);
1769         BUG_ON(path[depth].p_hdr == NULL);
1770         eh = path[depth].p_hdr;
1771
1772         while (ex < EXT_LAST_EXTENT(eh)) {
1773                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1774                         break;
1775                 /* merge with next extent! */
1776                 unwritten = ext4_ext_is_unwritten(ex);
1777                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1778                                 + ext4_ext_get_actual_len(ex + 1));
1779                 if (unwritten)
1780                         ext4_ext_mark_unwritten(ex);
1781
1782                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1783                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1784                                 * sizeof(struct ext4_extent);
1785                         memmove(ex + 1, ex + 2, len);
1786                 }
1787                 le16_add_cpu(&eh->eh_entries, -1);
1788                 merge_done = 1;
1789                 WARN_ON(eh->eh_entries == 0);
1790                 if (!eh->eh_entries)
1791                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1792         }
1793
1794         return merge_done;
1795 }
1796
1797 /*
1798  * This function does a very simple check to see if we can collapse
1799  * an extent tree with a single extent tree leaf block into the inode.
1800  */
1801 static void ext4_ext_try_to_merge_up(handle_t *handle,
1802                                      struct inode *inode,
1803                                      struct ext4_ext_path *path)
1804 {
1805         size_t s;
1806         unsigned max_root = ext4_ext_space_root(inode, 0);
1807         ext4_fsblk_t blk;
1808
1809         if ((path[0].p_depth != 1) ||
1810             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1811             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1812                 return;
1813
1814         /*
1815          * We need to modify the block allocation bitmap and the block
1816          * group descriptor to release the extent tree block.  If we
1817          * can't get the journal credits, give up.
1818          */
1819         if (ext4_journal_extend(handle, 2))
1820                 return;
1821
1822         /*
1823          * Copy the extent data up to the inode
1824          */
1825         blk = ext4_idx_pblock(path[0].p_idx);
1826         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1827                 sizeof(struct ext4_extent_idx);
1828         s += sizeof(struct ext4_extent_header);
1829
1830         path[1].p_maxdepth = path[0].p_maxdepth;
1831         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1832         path[0].p_depth = 0;
1833         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1834                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1835         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1836
1837         brelse(path[1].p_bh);
1838         ext4_free_blocks(handle, inode, NULL, blk, 1,
1839                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1840 }
1841
1842 /*
1843  * This function tries to merge the @ex extent to neighbours in the tree.
1844  * return 1 if merge left else 0.
1845  */
1846 static void ext4_ext_try_to_merge(handle_t *handle,
1847                                   struct inode *inode,
1848                                   struct ext4_ext_path *path,
1849                                   struct ext4_extent *ex) {
1850         struct ext4_extent_header *eh;
1851         unsigned int depth;
1852         int merge_done = 0;
1853
1854         depth = ext_depth(inode);
1855         BUG_ON(path[depth].p_hdr == NULL);
1856         eh = path[depth].p_hdr;
1857
1858         if (ex > EXT_FIRST_EXTENT(eh))
1859                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1860
1861         if (!merge_done)
1862                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1863
1864         ext4_ext_try_to_merge_up(handle, inode, path);
1865 }
1866
1867 /*
1868  * check if a portion of the "newext" extent overlaps with an
1869  * existing extent.
1870  *
1871  * If there is an overlap discovered, it updates the length of the newext
1872  * such that there will be no overlap, and then returns 1.
1873  * If there is no overlap found, it returns 0.
1874  */
1875 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1876                                            struct inode *inode,
1877                                            struct ext4_extent *newext,
1878                                            struct ext4_ext_path *path)
1879 {
1880         ext4_lblk_t b1, b2;
1881         unsigned int depth, len1;
1882         unsigned int ret = 0;
1883
1884         b1 = le32_to_cpu(newext->ee_block);
1885         len1 = ext4_ext_get_actual_len(newext);
1886         depth = ext_depth(inode);
1887         if (!path[depth].p_ext)
1888                 goto out;
1889         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1890
1891         /*
1892          * get the next allocated block if the extent in the path
1893          * is before the requested block(s)
1894          */
1895         if (b2 < b1) {
1896                 b2 = ext4_ext_next_allocated_block(path);
1897                 if (b2 == EXT_MAX_BLOCKS)
1898                         goto out;
1899                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1900         }
1901
1902         /* check for wrap through zero on extent logical start block*/
1903         if (b1 + len1 < b1) {
1904                 len1 = EXT_MAX_BLOCKS - b1;
1905                 newext->ee_len = cpu_to_le16(len1);
1906                 ret = 1;
1907         }
1908
1909         /* check for overlap */
1910         if (b1 + len1 > b2) {
1911                 newext->ee_len = cpu_to_le16(b2 - b1);
1912                 ret = 1;
1913         }
1914 out:
1915         return ret;
1916 }
1917
1918 /*
1919  * ext4_ext_insert_extent:
1920  * tries to merge requsted extent into the existing extent or
1921  * inserts requested extent as new one into the tree,
1922  * creating new leaf in the no-space case.
1923  */
1924 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1925                                 struct ext4_ext_path **ppath,
1926                                 struct ext4_extent *newext, int gb_flags)
1927 {
1928         struct ext4_ext_path *path = *ppath;
1929         struct ext4_extent_header *eh;
1930         struct ext4_extent *ex, *fex;
1931         struct ext4_extent *nearex; /* nearest extent */
1932         struct ext4_ext_path *npath = NULL;
1933         int depth, len, err;
1934         ext4_lblk_t next;
1935         int mb_flags = 0, unwritten;
1936
1937         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1938                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1939         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1940                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1941                 return -EIO;
1942         }
1943         depth = ext_depth(inode);
1944         ex = path[depth].p_ext;
1945         eh = path[depth].p_hdr;
1946         if (unlikely(path[depth].p_hdr == NULL)) {
1947                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1948                 return -EIO;
1949         }
1950
1951         /* try to insert block into found extent and return */
1952         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1953
1954                 /*
1955                  * Try to see whether we should rather test the extent on
1956                  * right from ex, or from the left of ex. This is because
1957                  * ext4_find_extent() can return either extent on the
1958                  * left, or on the right from the searched position. This
1959                  * will make merging more effective.
1960                  */
1961                 if (ex < EXT_LAST_EXTENT(eh) &&
1962                     (le32_to_cpu(ex->ee_block) +
1963                     ext4_ext_get_actual_len(ex) <
1964                     le32_to_cpu(newext->ee_block))) {
1965                         ex += 1;
1966                         goto prepend;
1967                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1968                            (le32_to_cpu(newext->ee_block) +
1969                            ext4_ext_get_actual_len(newext) <
1970                            le32_to_cpu(ex->ee_block)))
1971                         ex -= 1;
1972
1973                 /* Try to append newex to the ex */
1974                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1975                         ext_debug("append [%d]%d block to %u:[%d]%d"
1976                                   "(from %llu)\n",
1977                                   ext4_ext_is_unwritten(newext),
1978                                   ext4_ext_get_actual_len(newext),
1979                                   le32_to_cpu(ex->ee_block),
1980                                   ext4_ext_is_unwritten(ex),
1981                                   ext4_ext_get_actual_len(ex),
1982                                   ext4_ext_pblock(ex));
1983                         err = ext4_ext_get_access(handle, inode,
1984                                                   path + depth);
1985                         if (err)
1986                                 return err;
1987                         unwritten = ext4_ext_is_unwritten(ex);
1988                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1989                                         + ext4_ext_get_actual_len(newext));
1990                         if (unwritten)
1991                                 ext4_ext_mark_unwritten(ex);
1992                         eh = path[depth].p_hdr;
1993                         nearex = ex;
1994                         goto merge;
1995                 }
1996
1997 prepend:
1998                 /* Try to prepend newex to the ex */
1999                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2000                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2001                                   "(from %llu)\n",
2002                                   le32_to_cpu(newext->ee_block),
2003                                   ext4_ext_is_unwritten(newext),
2004                                   ext4_ext_get_actual_len(newext),
2005                                   le32_to_cpu(ex->ee_block),
2006                                   ext4_ext_is_unwritten(ex),
2007                                   ext4_ext_get_actual_len(ex),
2008                                   ext4_ext_pblock(ex));
2009                         err = ext4_ext_get_access(handle, inode,
2010                                                   path + depth);
2011                         if (err)
2012                                 return err;
2013
2014                         unwritten = ext4_ext_is_unwritten(ex);
2015                         ex->ee_block = newext->ee_block;
2016                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2017                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2018                                         + ext4_ext_get_actual_len(newext));
2019                         if (unwritten)
2020                                 ext4_ext_mark_unwritten(ex);
2021                         eh = path[depth].p_hdr;
2022                         nearex = ex;
2023                         goto merge;
2024                 }
2025         }
2026
2027         depth = ext_depth(inode);
2028         eh = path[depth].p_hdr;
2029         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2030                 goto has_space;
2031
2032         /* probably next leaf has space for us? */
2033         fex = EXT_LAST_EXTENT(eh);
2034         next = EXT_MAX_BLOCKS;
2035         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2036                 next = ext4_ext_next_leaf_block(path);
2037         if (next != EXT_MAX_BLOCKS) {
2038                 ext_debug("next leaf block - %u\n", next);
2039                 BUG_ON(npath != NULL);
2040                 npath = ext4_find_extent(inode, next, NULL, 0);
2041                 if (IS_ERR(npath))
2042                         return PTR_ERR(npath);
2043                 BUG_ON(npath->p_depth != path->p_depth);
2044                 eh = npath[depth].p_hdr;
2045                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2046                         ext_debug("next leaf isn't full(%d)\n",
2047                                   le16_to_cpu(eh->eh_entries));
2048                         path = npath;
2049                         goto has_space;
2050                 }
2051                 ext_debug("next leaf has no free space(%d,%d)\n",
2052                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2053         }
2054
2055         /*
2056          * There is no free space in the found leaf.
2057          * We're gonna add a new leaf in the tree.
2058          */
2059         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2060                 mb_flags |= EXT4_MB_USE_RESERVED;
2061         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2062                                        ppath, newext);
2063         if (err)
2064                 goto cleanup;
2065         depth = ext_depth(inode);
2066         eh = path[depth].p_hdr;
2067
2068 has_space:
2069         nearex = path[depth].p_ext;
2070
2071         err = ext4_ext_get_access(handle, inode, path + depth);
2072         if (err)
2073                 goto cleanup;
2074
2075         if (!nearex) {
2076                 /* there is no extent in this leaf, create first one */
2077                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2078                                 le32_to_cpu(newext->ee_block),
2079                                 ext4_ext_pblock(newext),
2080                                 ext4_ext_is_unwritten(newext),
2081                                 ext4_ext_get_actual_len(newext));
2082                 nearex = EXT_FIRST_EXTENT(eh);
2083         } else {
2084                 if (le32_to_cpu(newext->ee_block)
2085                            > le32_to_cpu(nearex->ee_block)) {
2086                         /* Insert after */
2087                         ext_debug("insert %u:%llu:[%d]%d before: "
2088                                         "nearest %p\n",
2089                                         le32_to_cpu(newext->ee_block),
2090                                         ext4_ext_pblock(newext),
2091                                         ext4_ext_is_unwritten(newext),
2092                                         ext4_ext_get_actual_len(newext),
2093                                         nearex);
2094                         nearex++;
2095                 } else {
2096                         /* Insert before */
2097                         BUG_ON(newext->ee_block == nearex->ee_block);
2098                         ext_debug("insert %u:%llu:[%d]%d after: "
2099                                         "nearest %p\n",
2100                                         le32_to_cpu(newext->ee_block),
2101                                         ext4_ext_pblock(newext),
2102                                         ext4_ext_is_unwritten(newext),
2103                                         ext4_ext_get_actual_len(newext),
2104                                         nearex);
2105                 }
2106                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2107                 if (len > 0) {
2108                         ext_debug("insert %u:%llu:[%d]%d: "
2109                                         "move %d extents from 0x%p to 0x%p\n",
2110                                         le32_to_cpu(newext->ee_block),
2111                                         ext4_ext_pblock(newext),
2112                                         ext4_ext_is_unwritten(newext),
2113                                         ext4_ext_get_actual_len(newext),
2114                                         len, nearex, nearex + 1);
2115                         memmove(nearex + 1, nearex,
2116                                 len * sizeof(struct ext4_extent));
2117                 }
2118         }
2119
2120         le16_add_cpu(&eh->eh_entries, 1);
2121         path[depth].p_ext = nearex;
2122         nearex->ee_block = newext->ee_block;
2123         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2124         nearex->ee_len = newext->ee_len;
2125
2126 merge:
2127         /* try to merge extents */
2128         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2129                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2130
2131
2132         /* time to correct all indexes above */
2133         err = ext4_ext_correct_indexes(handle, inode, path);
2134         if (err)
2135                 goto cleanup;
2136
2137         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2138
2139 cleanup:
2140         ext4_ext_drop_refs(npath);
2141         kfree(npath);
2142         return err;
2143 }
2144
2145 static int ext4_fill_fiemap_extents(struct inode *inode,
2146                                     ext4_lblk_t block, ext4_lblk_t num,
2147                                     struct fiemap_extent_info *fieinfo)
2148 {
2149         struct ext4_ext_path *path = NULL;
2150         struct ext4_extent *ex;
2151         struct extent_status es;
2152         ext4_lblk_t next, next_del, start = 0, end = 0;
2153         ext4_lblk_t last = block + num;
2154         int exists, depth = 0, err = 0;
2155         unsigned int flags = 0;
2156         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2157
2158         while (block < last && block != EXT_MAX_BLOCKS) {
2159                 num = last - block;
2160                 /* find extent for this block */
2161                 down_read(&EXT4_I(inode)->i_data_sem);
2162
2163                 path = ext4_find_extent(inode, block, &path, 0);
2164                 if (IS_ERR(path)) {
2165                         up_read(&EXT4_I(inode)->i_data_sem);
2166                         err = PTR_ERR(path);
2167                         path = NULL;
2168                         break;
2169                 }
2170
2171                 depth = ext_depth(inode);
2172                 if (unlikely(path[depth].p_hdr == NULL)) {
2173                         up_read(&EXT4_I(inode)->i_data_sem);
2174                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2175                         err = -EIO;
2176                         break;
2177                 }
2178                 ex = path[depth].p_ext;
2179                 next = ext4_ext_next_allocated_block(path);
2180
2181                 flags = 0;
2182                 exists = 0;
2183                 if (!ex) {
2184                         /* there is no extent yet, so try to allocate
2185                          * all requested space */
2186                         start = block;
2187                         end = block + num;
2188                 } else if (le32_to_cpu(ex->ee_block) > block) {
2189                         /* need to allocate space before found extent */
2190                         start = block;
2191                         end = le32_to_cpu(ex->ee_block);
2192                         if (block + num < end)
2193                                 end = block + num;
2194                 } else if (block >= le32_to_cpu(ex->ee_block)
2195                                         + ext4_ext_get_actual_len(ex)) {
2196                         /* need to allocate space after found extent */
2197                         start = block;
2198                         end = block + num;
2199                         if (end >= next)
2200                                 end = next;
2201                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2202                         /*
2203                          * some part of requested space is covered
2204                          * by found extent
2205                          */
2206                         start = block;
2207                         end = le32_to_cpu(ex->ee_block)
2208                                 + ext4_ext_get_actual_len(ex);
2209                         if (block + num < end)
2210                                 end = block + num;
2211                         exists = 1;
2212                 } else {
2213                         BUG();
2214                 }
2215                 BUG_ON(end <= start);
2216
2217                 if (!exists) {
2218                         es.es_lblk = start;
2219                         es.es_len = end - start;
2220                         es.es_pblk = 0;
2221                 } else {
2222                         es.es_lblk = le32_to_cpu(ex->ee_block);
2223                         es.es_len = ext4_ext_get_actual_len(ex);
2224                         es.es_pblk = ext4_ext_pblock(ex);
2225                         if (ext4_ext_is_unwritten(ex))
2226                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2227                 }
2228
2229                 /*
2230                  * Find delayed extent and update es accordingly. We call
2231                  * it even in !exists case to find out whether es is the
2232                  * last existing extent or not.
2233                  */
2234                 next_del = ext4_find_delayed_extent(inode, &es);
2235                 if (!exists && next_del) {
2236                         exists = 1;
2237                         flags |= (FIEMAP_EXTENT_DELALLOC |
2238                                   FIEMAP_EXTENT_UNKNOWN);
2239                 }
2240                 up_read(&EXT4_I(inode)->i_data_sem);
2241
2242                 if (unlikely(es.es_len == 0)) {
2243                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2244                         err = -EIO;
2245                         break;
2246                 }
2247
2248                 /*
2249                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2250                  * we need to check next == EXT_MAX_BLOCKS because it is
2251                  * possible that an extent is with unwritten and delayed
2252                  * status due to when an extent is delayed allocated and
2253                  * is allocated by fallocate status tree will track both of
2254                  * them in a extent.
2255                  *
2256                  * So we could return a unwritten and delayed extent, and
2257                  * its block is equal to 'next'.
2258                  */
2259                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2260                         flags |= FIEMAP_EXTENT_LAST;
2261                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2262                                      next != EXT_MAX_BLOCKS)) {
2263                                 EXT4_ERROR_INODE(inode,
2264                                                  "next extent == %u, next "
2265                                                  "delalloc extent = %u",
2266                                                  next, next_del);
2267                                 err = -EIO;
2268                                 break;
2269                         }
2270                 }
2271
2272                 if (exists) {
2273                         err = fiemap_fill_next_extent(fieinfo,
2274                                 (__u64)es.es_lblk << blksize_bits,
2275                                 (__u64)es.es_pblk << blksize_bits,
2276                                 (__u64)es.es_len << blksize_bits,
2277                                 flags);
2278                         if (err < 0)
2279                                 break;
2280                         if (err == 1) {
2281                                 err = 0;
2282                                 break;
2283                         }
2284                 }
2285
2286                 block = es.es_lblk + es.es_len;
2287         }
2288
2289         ext4_ext_drop_refs(path);
2290         kfree(path);
2291         return err;
2292 }
2293
2294 /*
2295  * ext4_ext_put_gap_in_cache:
2296  * calculate boundaries of the gap that the requested block fits into
2297  * and cache this gap
2298  */
2299 static void
2300 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2301                                 ext4_lblk_t block)
2302 {
2303         int depth = ext_depth(inode);
2304         ext4_lblk_t len;
2305         ext4_lblk_t lblock;
2306         struct ext4_extent *ex;
2307         struct extent_status es;
2308
2309         ex = path[depth].p_ext;
2310         if (ex == NULL) {
2311                 /* there is no extent yet, so gap is [0;-] */
2312                 lblock = 0;
2313                 len = EXT_MAX_BLOCKS;
2314                 ext_debug("cache gap(whole file):");
2315         } else if (block < le32_to_cpu(ex->ee_block)) {
2316                 lblock = block;
2317                 len = le32_to_cpu(ex->ee_block) - block;
2318                 ext_debug("cache gap(before): %u [%u:%u]",
2319                                 block,
2320                                 le32_to_cpu(ex->ee_block),
2321                                  ext4_ext_get_actual_len(ex));
2322         } else if (block >= le32_to_cpu(ex->ee_block)
2323                         + ext4_ext_get_actual_len(ex)) {
2324                 ext4_lblk_t next;
2325                 lblock = le32_to_cpu(ex->ee_block)
2326                         + ext4_ext_get_actual_len(ex);
2327
2328                 next = ext4_ext_next_allocated_block(path);
2329                 ext_debug("cache gap(after): [%u:%u] %u",
2330                                 le32_to_cpu(ex->ee_block),
2331                                 ext4_ext_get_actual_len(ex),
2332                                 block);
2333                 BUG_ON(next == lblock);
2334                 len = next - lblock;
2335         } else {
2336                 BUG();
2337         }
2338
2339         ext4_es_find_delayed_extent_range(inode, lblock, lblock + len - 1, &es);
2340         if (es.es_len) {
2341                 /* There's delayed extent containing lblock? */
2342                 if (es.es_lblk <= lblock)
2343                         return;
2344                 len = min(es.es_lblk - lblock, len);
2345         }
2346         ext_debug(" -> %u:%u\n", lblock, len);
2347         ext4_es_insert_extent(inode, lblock, len, ~0, EXTENT_STATUS_HOLE);
2348 }
2349
2350 /*
2351  * ext4_ext_rm_idx:
2352  * removes index from the index block.
2353  */
2354 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2355                         struct ext4_ext_path *path, int depth)
2356 {
2357         int err;
2358         ext4_fsblk_t leaf;
2359
2360         /* free index block */
2361         depth--;
2362         path = path + depth;
2363         leaf = ext4_idx_pblock(path->p_idx);
2364         if (unlikely(path->p_hdr->eh_entries == 0)) {
2365                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2366                 return -EIO;
2367         }
2368         err = ext4_ext_get_access(handle, inode, path);
2369         if (err)
2370                 return err;
2371
2372         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2373                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2374                 len *= sizeof(struct ext4_extent_idx);
2375                 memmove(path->p_idx, path->p_idx + 1, len);
2376         }
2377
2378         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2379         err = ext4_ext_dirty(handle, inode, path);
2380         if (err)
2381                 return err;
2382         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2383         trace_ext4_ext_rm_idx(inode, leaf);
2384
2385         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2386                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2387
2388         while (--depth >= 0) {
2389                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2390                         break;
2391                 path--;
2392                 err = ext4_ext_get_access(handle, inode, path);
2393                 if (err)
2394                         break;
2395                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2396                 err = ext4_ext_dirty(handle, inode, path);
2397                 if (err)
2398                         break;
2399         }
2400         return err;
2401 }
2402
2403 /*
2404  * ext4_ext_calc_credits_for_single_extent:
2405  * This routine returns max. credits that needed to insert an extent
2406  * to the extent tree.
2407  * When pass the actual path, the caller should calculate credits
2408  * under i_data_sem.
2409  */
2410 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2411                                                 struct ext4_ext_path *path)
2412 {
2413         if (path) {
2414                 int depth = ext_depth(inode);
2415                 int ret = 0;
2416
2417                 /* probably there is space in leaf? */
2418                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2419                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2420
2421                         /*
2422                          *  There are some space in the leaf tree, no
2423                          *  need to account for leaf block credit
2424                          *
2425                          *  bitmaps and block group descriptor blocks
2426                          *  and other metadata blocks still need to be
2427                          *  accounted.
2428                          */
2429                         /* 1 bitmap, 1 block group descriptor */
2430                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2431                         return ret;
2432                 }
2433         }
2434
2435         return ext4_chunk_trans_blocks(inode, nrblocks);
2436 }
2437
2438 /*
2439  * How many index/leaf blocks need to change/allocate to add @extents extents?
2440  *
2441  * If we add a single extent, then in the worse case, each tree level
2442  * index/leaf need to be changed in case of the tree split.
2443  *
2444  * If more extents are inserted, they could cause the whole tree split more
2445  * than once, but this is really rare.
2446  */
2447 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2448 {
2449         int index;
2450         int depth;
2451
2452         /* If we are converting the inline data, only one is needed here. */
2453         if (ext4_has_inline_data(inode))
2454                 return 1;
2455
2456         depth = ext_depth(inode);
2457
2458         if (extents <= 1)
2459                 index = depth * 2;
2460         else
2461                 index = depth * 3;
2462
2463         return index;
2464 }
2465
2466 static inline int get_default_free_blocks_flags(struct inode *inode)
2467 {
2468         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2469                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2470         else if (ext4_should_journal_data(inode))
2471                 return EXT4_FREE_BLOCKS_FORGET;
2472         return 0;
2473 }
2474
2475 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2476                               struct ext4_extent *ex,
2477                               long long *partial_cluster,
2478                               ext4_lblk_t from, ext4_lblk_t to)
2479 {
2480         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2481         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2482         ext4_fsblk_t pblk;
2483         int flags = get_default_free_blocks_flags(inode);
2484
2485         /*
2486          * For bigalloc file systems, we never free a partial cluster
2487          * at the beginning of the extent.  Instead, we make a note
2488          * that we tried freeing the cluster, and check to see if we
2489          * need to free it on a subsequent call to ext4_remove_blocks,
2490          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2491          */
2492         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2493
2494         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2495         /*
2496          * If we have a partial cluster, and it's different from the
2497          * cluster of the last block, we need to explicitly free the
2498          * partial cluster here.
2499          */
2500         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2501         if (*partial_cluster > 0 &&
2502             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2503                 ext4_free_blocks(handle, inode, NULL,
2504                                  EXT4_C2B(sbi, *partial_cluster),
2505                                  sbi->s_cluster_ratio, flags);
2506                 *partial_cluster = 0;
2507         }
2508
2509 #ifdef EXTENTS_STATS
2510         {
2511                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2512                 spin_lock(&sbi->s_ext_stats_lock);
2513                 sbi->s_ext_blocks += ee_len;
2514                 sbi->s_ext_extents++;
2515                 if (ee_len < sbi->s_ext_min)
2516                         sbi->s_ext_min = ee_len;
2517                 if (ee_len > sbi->s_ext_max)
2518                         sbi->s_ext_max = ee_len;
2519                 if (ext_depth(inode) > sbi->s_depth_max)
2520                         sbi->s_depth_max = ext_depth(inode);
2521                 spin_unlock(&sbi->s_ext_stats_lock);
2522         }
2523 #endif
2524         if (from >= le32_to_cpu(ex->ee_block)
2525             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2526                 /* tail removal */
2527                 ext4_lblk_t num;
2528                 long long first_cluster;
2529
2530                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2531                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2532                 /*
2533                  * Usually we want to free partial cluster at the end of the
2534                  * extent, except for the situation when the cluster is still
2535                  * used by any other extent (partial_cluster is negative).
2536                  */
2537                 if (*partial_cluster < 0 &&
2538                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2539                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2540
2541                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2542                           num, pblk, *partial_cluster);
2543                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2544                 /*
2545                  * If the block range to be freed didn't start at the
2546                  * beginning of a cluster, and we removed the entire
2547                  * extent and the cluster is not used by any other extent,
2548                  * save the partial cluster here, since we might need to
2549                  * delete if we determine that the truncate or punch hole
2550                  * operation has removed all of the blocks in the cluster.
2551                  * If that cluster is used by another extent, preserve its
2552                  * negative value so it isn't freed later on.
2553                  *
2554                  * If the whole extent wasn't freed, we've reached the
2555                  * start of the truncated/punched region and have finished
2556                  * removing blocks.  If there's a partial cluster here it's
2557                  * shared with the remainder of the extent and is no longer
2558                  * a candidate for removal.
2559                  */
2560                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2561                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2562                         if (first_cluster != -*partial_cluster)
2563                                 *partial_cluster = first_cluster;
2564                 } else {
2565                         *partial_cluster = 0;
2566                 }
2567         } else
2568                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2569                            "%u-%u from %u:%u\n",
2570                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2571         return 0;
2572 }
2573
2574
2575 /*
2576  * ext4_ext_rm_leaf() Removes the extents associated with the
2577  * blocks appearing between "start" and "end".  Both "start"
2578  * and "end" must appear in the same extent or EIO is returned.
2579  *
2580  * @handle: The journal handle
2581  * @inode:  The files inode
2582  * @path:   The path to the leaf
2583  * @partial_cluster: The cluster which we'll have to free if all extents
2584  *                   has been released from it.  However, if this value is
2585  *                   negative, it's a cluster just to the right of the
2586  *                   punched region and it must not be freed.
2587  * @start:  The first block to remove
2588  * @end:   The last block to remove
2589  */
2590 static int
2591 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2592                  struct ext4_ext_path *path,
2593                  long long *partial_cluster,
2594                  ext4_lblk_t start, ext4_lblk_t end)
2595 {
2596         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2597         int err = 0, correct_index = 0;
2598         int depth = ext_depth(inode), credits;
2599         struct ext4_extent_header *eh;
2600         ext4_lblk_t a, b;
2601         unsigned num;
2602         ext4_lblk_t ex_ee_block;
2603         unsigned short ex_ee_len;
2604         unsigned unwritten = 0;
2605         struct ext4_extent *ex;
2606         ext4_fsblk_t pblk;
2607
2608         /* the header must be checked already in ext4_ext_remove_space() */
2609         ext_debug("truncate since %u in leaf to %u\n", start, end);
2610         if (!path[depth].p_hdr)
2611                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2612         eh = path[depth].p_hdr;
2613         if (unlikely(path[depth].p_hdr == NULL)) {
2614                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2615                 return -EIO;
2616         }
2617         /* find where to start removing */
2618         ex = path[depth].p_ext;
2619         if (!ex)
2620                 ex = EXT_LAST_EXTENT(eh);
2621
2622         ex_ee_block = le32_to_cpu(ex->ee_block);
2623         ex_ee_len = ext4_ext_get_actual_len(ex);
2624
2625         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2626
2627         while (ex >= EXT_FIRST_EXTENT(eh) &&
2628                         ex_ee_block + ex_ee_len > start) {
2629
2630                 if (ext4_ext_is_unwritten(ex))
2631                         unwritten = 1;
2632                 else
2633                         unwritten = 0;
2634
2635                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2636                           unwritten, ex_ee_len);
2637                 path[depth].p_ext = ex;
2638
2639                 a = ex_ee_block > start ? ex_ee_block : start;
2640                 b = ex_ee_block+ex_ee_len - 1 < end ?
2641                         ex_ee_block+ex_ee_len - 1 : end;
2642
2643                 ext_debug("  border %u:%u\n", a, b);
2644
2645                 /* If this extent is beyond the end of the hole, skip it */
2646                 if (end < ex_ee_block) {
2647                         /*
2648                          * We're going to skip this extent and move to another,
2649                          * so note that its first cluster is in use to avoid
2650                          * freeing it when removing blocks.  Eventually, the
2651                          * right edge of the truncated/punched region will
2652                          * be just to the left.
2653                          */
2654                         if (sbi->s_cluster_ratio > 1) {
2655                                 pblk = ext4_ext_pblock(ex);
2656                                 *partial_cluster =
2657                                         -(long long) EXT4_B2C(sbi, pblk);
2658                         }
2659                         ex--;
2660                         ex_ee_block = le32_to_cpu(ex->ee_block);
2661                         ex_ee_len = ext4_ext_get_actual_len(ex);
2662                         continue;
2663                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2664                         EXT4_ERROR_INODE(inode,
2665                                          "can not handle truncate %u:%u "
2666                                          "on extent %u:%u",
2667                                          start, end, ex_ee_block,
2668                                          ex_ee_block + ex_ee_len - 1);
2669                         err = -EIO;
2670                         goto out;
2671                 } else if (a != ex_ee_block) {
2672                         /* remove tail of the extent */
2673                         num = a - ex_ee_block;
2674                 } else {
2675                         /* remove whole extent: excellent! */
2676                         num = 0;
2677                 }
2678                 /*
2679                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2680                  * descriptor) for each block group; assume two block
2681                  * groups plus ex_ee_len/blocks_per_block_group for
2682                  * the worst case
2683                  */
2684                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2685                 if (ex == EXT_FIRST_EXTENT(eh)) {
2686                         correct_index = 1;
2687                         credits += (ext_depth(inode)) + 1;
2688                 }
2689                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2690
2691                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2692                 if (err)
2693                         goto out;
2694
2695                 err = ext4_ext_get_access(handle, inode, path + depth);
2696                 if (err)
2697                         goto out;
2698
2699                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2700                                          a, b);
2701                 if (err)
2702                         goto out;
2703
2704                 if (num == 0)
2705                         /* this extent is removed; mark slot entirely unused */
2706                         ext4_ext_store_pblock(ex, 0);
2707
2708                 ex->ee_len = cpu_to_le16(num);
2709                 /*
2710                  * Do not mark unwritten if all the blocks in the
2711                  * extent have been removed.
2712                  */
2713                 if (unwritten && num)
2714                         ext4_ext_mark_unwritten(ex);
2715                 /*
2716                  * If the extent was completely released,
2717                  * we need to remove it from the leaf
2718                  */
2719                 if (num == 0) {
2720                         if (end != EXT_MAX_BLOCKS - 1) {
2721                                 /*
2722                                  * For hole punching, we need to scoot all the
2723                                  * extents up when an extent is removed so that
2724                                  * we dont have blank extents in the middle
2725                                  */
2726                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2727                                         sizeof(struct ext4_extent));
2728
2729                                 /* Now get rid of the one at the end */
2730                                 memset(EXT_LAST_EXTENT(eh), 0,
2731                                         sizeof(struct ext4_extent));
2732                         }
2733                         le16_add_cpu(&eh->eh_entries, -1);
2734                 }
2735
2736                 err = ext4_ext_dirty(handle, inode, path + depth);
2737                 if (err)
2738                         goto out;
2739
2740                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2741                                 ext4_ext_pblock(ex));
2742                 ex--;
2743                 ex_ee_block = le32_to_cpu(ex->ee_block);
2744                 ex_ee_len = ext4_ext_get_actual_len(ex);
2745         }
2746
2747         if (correct_index && eh->eh_entries)
2748                 err = ext4_ext_correct_indexes(handle, inode, path);
2749
2750         /*
2751          * If there's a partial cluster and at least one extent remains in
2752          * the leaf, free the partial cluster if it isn't shared with the
2753          * current extent.  If it is shared with the current extent
2754          * we zero partial_cluster because we've reached the start of the
2755          * truncated/punched region and we're done removing blocks.
2756          */
2757         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2758                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2759                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2760                         ext4_free_blocks(handle, inode, NULL,
2761                                          EXT4_C2B(sbi, *partial_cluster),
2762                                          sbi->s_cluster_ratio,
2763                                          get_default_free_blocks_flags(inode));
2764                 }
2765                 *partial_cluster = 0;
2766         }
2767
2768         /* if this leaf is free, then we should
2769          * remove it from index block above */
2770         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2771                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2772
2773 out:
2774         return err;
2775 }
2776
2777 /*
2778  * ext4_ext_more_to_rm:
2779  * returns 1 if current index has to be freed (even partial)
2780  */
2781 static int
2782 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2783 {
2784         BUG_ON(path->p_idx == NULL);
2785
2786         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2787                 return 0;
2788
2789         /*
2790          * if truncate on deeper level happened, it wasn't partial,
2791          * so we have to consider current index for truncation
2792          */
2793         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2794                 return 0;
2795         return 1;
2796 }
2797
2798 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2799                           ext4_lblk_t end)
2800 {
2801         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2802         int depth = ext_depth(inode);
2803         struct ext4_ext_path *path = NULL;
2804         long long partial_cluster = 0;
2805         handle_t *handle;
2806         int i = 0, err = 0;
2807
2808         ext_debug("truncate since %u to %u\n", start, end);
2809
2810         /* probably first extent we're gonna free will be last in block */
2811         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2812         if (IS_ERR(handle))
2813                 return PTR_ERR(handle);
2814
2815 again:
2816         trace_ext4_ext_remove_space(inode, start, end, depth);
2817
2818         /*
2819          * Check if we are removing extents inside the extent tree. If that
2820          * is the case, we are going to punch a hole inside the extent tree
2821          * so we have to check whether we need to split the extent covering
2822          * the last block to remove so we can easily remove the part of it
2823          * in ext4_ext_rm_leaf().
2824          */
2825         if (end < EXT_MAX_BLOCKS - 1) {
2826                 struct ext4_extent *ex;
2827                 ext4_lblk_t ee_block, ex_end, lblk;
2828                 ext4_fsblk_t pblk;
2829
2830                 /* find extent for or closest extent to this block */
2831                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2832                 if (IS_ERR(path)) {
2833                         ext4_journal_stop(handle);
2834                         return PTR_ERR(path);
2835                 }
2836                 depth = ext_depth(inode);
2837                 /* Leaf not may not exist only if inode has no blocks at all */
2838                 ex = path[depth].p_ext;
2839                 if (!ex) {
2840                         if (depth) {
2841                                 EXT4_ERROR_INODE(inode,
2842                                                  "path[%d].p_hdr == NULL",
2843                                                  depth);
2844                                 err = -EIO;
2845                         }
2846                         goto out;
2847                 }
2848
2849                 ee_block = le32_to_cpu(ex->ee_block);
2850                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2851
2852                 /*
2853                  * See if the last block is inside the extent, if so split
2854                  * the extent at 'end' block so we can easily remove the
2855                  * tail of the first part of the split extent in
2856                  * ext4_ext_rm_leaf().
2857                  */
2858                 if (end >= ee_block && end < ex_end) {
2859
2860                         /*
2861                          * If we're going to split the extent, note that
2862                          * the cluster containing the block after 'end' is
2863                          * in use to avoid freeing it when removing blocks.
2864                          */
2865                         if (sbi->s_cluster_ratio > 1) {
2866                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
2867                                 partial_cluster =
2868                                         -(long long) EXT4_B2C(sbi, pblk);
2869                         }
2870
2871                         /*
2872                          * Split the extent in two so that 'end' is the last
2873                          * block in the first new extent. Also we should not
2874                          * fail removing space due to ENOSPC so try to use
2875                          * reserved block if that happens.
2876                          */
2877                         err = ext4_force_split_extent_at(handle, inode, &path,
2878                                                          end + 1, 1);
2879                         if (err < 0)
2880                                 goto out;
2881
2882                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2883                         /*
2884                          * If there's an extent to the right its first cluster
2885                          * contains the immediate right boundary of the
2886                          * truncated/punched region.  Set partial_cluster to
2887                          * its negative value so it won't be freed if shared
2888                          * with the current extent.  The end < ee_block case
2889                          * is handled in ext4_ext_rm_leaf().
2890                          */
2891                         lblk = ex_end + 1;
2892                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2893                                                     &ex);
2894                         if (err)
2895                                 goto out;
2896                         if (pblk)
2897                                 partial_cluster =
2898                                         -(long long) EXT4_B2C(sbi, pblk);
2899                 }
2900         }
2901         /*
2902          * We start scanning from right side, freeing all the blocks
2903          * after i_size and walking into the tree depth-wise.
2904          */
2905         depth = ext_depth(inode);
2906         if (path) {
2907                 int k = i = depth;
2908                 while (--k > 0)
2909                         path[k].p_block =
2910                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2911         } else {
2912                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2913                                GFP_NOFS);
2914                 if (path == NULL) {
2915                         ext4_journal_stop(handle);
2916                         return -ENOMEM;
2917                 }
2918                 path[0].p_maxdepth = path[0].p_depth = depth;
2919                 path[0].p_hdr = ext_inode_hdr(inode);
2920                 i = 0;
2921
2922                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2923                         err = -EIO;
2924                         goto out;
2925                 }
2926         }
2927         err = 0;
2928
2929         while (i >= 0 && err == 0) {
2930                 if (i == depth) {
2931                         /* this is leaf block */
2932                         err = ext4_ext_rm_leaf(handle, inode, path,
2933                                                &partial_cluster, start,
2934                                                end);
2935                         /* root level has p_bh == NULL, brelse() eats this */
2936                         brelse(path[i].p_bh);
2937                         path[i].p_bh = NULL;
2938                         i--;
2939                         continue;
2940                 }
2941
2942                 /* this is index block */
2943                 if (!path[i].p_hdr) {
2944                         ext_debug("initialize header\n");
2945                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2946                 }
2947
2948                 if (!path[i].p_idx) {
2949                         /* this level hasn't been touched yet */
2950                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2951                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2952                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2953                                   path[i].p_hdr,
2954                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2955                 } else {
2956                         /* we were already here, see at next index */
2957                         path[i].p_idx--;
2958                 }
2959
2960                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2961                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2962                                 path[i].p_idx);
2963                 if (ext4_ext_more_to_rm(path + i)) {
2964                         struct buffer_head *bh;
2965                         /* go to the next level */
2966                         ext_debug("move to level %d (block %llu)\n",
2967                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2968                         memset(path + i + 1, 0, sizeof(*path));
2969                         bh = read_extent_tree_block(inode,
2970                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2971                                 EXT4_EX_NOCACHE);
2972                         if (IS_ERR(bh)) {
2973                                 /* should we reset i_size? */
2974                                 err = PTR_ERR(bh);
2975                                 break;
2976                         }
2977                         /* Yield here to deal with large extent trees.
2978                          * Should be a no-op if we did IO above. */
2979                         cond_resched();
2980                         if (WARN_ON(i + 1 > depth)) {
2981                                 err = -EIO;
2982                                 break;
2983                         }
2984                         path[i + 1].p_bh = bh;
2985
2986                         /* save actual number of indexes since this
2987                          * number is changed at the next iteration */
2988                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2989                         i++;
2990                 } else {
2991                         /* we finished processing this index, go up */
2992                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2993                                 /* index is empty, remove it;
2994                                  * handle must be already prepared by the
2995                                  * truncatei_leaf() */
2996                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2997                         }
2998                         /* root level has p_bh == NULL, brelse() eats this */
2999                         brelse(path[i].p_bh);
3000                         path[i].p_bh = NULL;
3001                         i--;
3002                         ext_debug("return to level %d\n", i);
3003                 }
3004         }
3005
3006         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3007                         partial_cluster, path->p_hdr->eh_entries);
3008
3009         /*
3010          * If we still have something in the partial cluster and we have removed
3011          * even the first extent, then we should free the blocks in the partial
3012          * cluster as well.  (This code will only run when there are no leaves
3013          * to the immediate left of the truncated/punched region.)
3014          */
3015         if (partial_cluster > 0 && err == 0) {
3016                 /* don't zero partial_cluster since it's not used afterwards */
3017                 ext4_free_blocks(handle, inode, NULL,
3018                                  EXT4_C2B(sbi, partial_cluster),
3019                                  sbi->s_cluster_ratio,
3020                                  get_default_free_blocks_flags(inode));
3021         }
3022
3023         /* TODO: flexible tree reduction should be here */
3024         if (path->p_hdr->eh_entries == 0) {
3025                 /*
3026                  * truncate to zero freed all the tree,
3027                  * so we need to correct eh_depth
3028                  */
3029                 err = ext4_ext_get_access(handle, inode, path);
3030                 if (err == 0) {
3031                         ext_inode_hdr(inode)->eh_depth = 0;
3032                         ext_inode_hdr(inode)->eh_max =
3033                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3034                         err = ext4_ext_dirty(handle, inode, path);
3035                 }
3036         }
3037 out:
3038         ext4_ext_drop_refs(path);
3039         kfree(path);
3040         path = NULL;
3041         if (err == -EAGAIN)
3042                 goto again;
3043         ext4_journal_stop(handle);
3044
3045         return err;
3046 }
3047
3048 /*
3049  * called at mount time
3050  */
3051 void ext4_ext_init(struct super_block *sb)
3052 {
3053         /*
3054          * possible initialization would be here
3055          */
3056
3057         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
3058 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3059                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3060 #ifdef AGGRESSIVE_TEST
3061                        ", aggressive tests"
3062 #endif
3063 #ifdef CHECK_BINSEARCH
3064                        ", check binsearch"
3065 #endif
3066 #ifdef EXTENTS_STATS
3067                        ", stats"
3068 #endif
3069                        "\n");
3070 #endif
3071 #ifdef EXTENTS_STATS
3072                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3073                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3074                 EXT4_SB(sb)->s_ext_max = 0;
3075 #endif
3076         }
3077 }
3078
3079 /*
3080  * called at umount time
3081  */
3082 void ext4_ext_release(struct super_block *sb)
3083 {
3084         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
3085                 return;
3086
3087 #ifdef EXTENTS_STATS
3088         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3089                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3090                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3091                         sbi->s_ext_blocks, sbi->s_ext_extents,
3092                         sbi->s_ext_blocks / sbi->s_ext_extents);
3093                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3094                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3095         }
3096 #endif
3097 }
3098
3099 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3100 {
3101         ext4_lblk_t  ee_block;
3102         ext4_fsblk_t ee_pblock;
3103         unsigned int ee_len;
3104
3105         ee_block  = le32_to_cpu(ex->ee_block);
3106         ee_len    = ext4_ext_get_actual_len(ex);
3107         ee_pblock = ext4_ext_pblock(ex);
3108
3109         if (ee_len == 0)
3110                 return 0;
3111
3112         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3113                                      EXTENT_STATUS_WRITTEN);
3114 }
3115
3116 /* FIXME!! we need to try to merge to left or right after zero-out  */
3117 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3118 {
3119         ext4_fsblk_t ee_pblock;
3120         unsigned int ee_len;
3121         int ret;
3122
3123         ee_len    = ext4_ext_get_actual_len(ex);
3124         ee_pblock = ext4_ext_pblock(ex);
3125
3126         if (ext4_encrypted_inode(inode))
3127                 return ext4_encrypted_zeroout(inode, ex);
3128
3129         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
3130         if (ret > 0)
3131                 ret = 0;
3132
3133         return ret;
3134 }
3135
3136 /*
3137  * ext4_split_extent_at() splits an extent at given block.
3138  *
3139  * @handle: the journal handle
3140  * @inode: the file inode
3141  * @path: the path to the extent
3142  * @split: the logical block where the extent is splitted.
3143  * @split_flags: indicates if the extent could be zeroout if split fails, and
3144  *               the states(init or unwritten) of new extents.
3145  * @flags: flags used to insert new extent to extent tree.
3146  *
3147  *
3148  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3149  * of which are deterimined by split_flag.
3150  *
3151  * There are two cases:
3152  *  a> the extent are splitted into two extent.
3153  *  b> split is not needed, and just mark the extent.
3154  *
3155  * return 0 on success.
3156  */
3157 static int ext4_split_extent_at(handle_t *handle,
3158                              struct inode *inode,
3159                              struct ext4_ext_path **ppath,
3160                              ext4_lblk_t split,
3161                              int split_flag,
3162                              int flags)
3163 {
3164         struct ext4_ext_path *path = *ppath;
3165         ext4_fsblk_t newblock;
3166         ext4_lblk_t ee_block;
3167         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3168         struct ext4_extent *ex2 = NULL;
3169         unsigned int ee_len, depth;
3170         int err = 0;
3171
3172         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3173                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3174
3175         ext_debug("ext4_split_extents_at: inode %lu, logical"
3176                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3177
3178         ext4_ext_show_leaf(inode, path);
3179
3180         depth = ext_depth(inode);
3181         ex = path[depth].p_ext;
3182         ee_block = le32_to_cpu(ex->ee_block);
3183         ee_len = ext4_ext_get_actual_len(ex);
3184         newblock = split - ee_block + ext4_ext_pblock(ex);
3185
3186         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3187         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3188                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3189                              EXT4_EXT_MARK_UNWRIT1 |
3190                              EXT4_EXT_MARK_UNWRIT2));
3191
3192         err = ext4_ext_get_access(handle, inode, path + depth);
3193         if (err)
3194                 goto out;
3195
3196         if (split == ee_block) {
3197                 /*
3198                  * case b: block @split is the block that the extent begins with
3199                  * then we just change the state of the extent, and splitting
3200                  * is not needed.
3201                  */
3202                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3203                         ext4_ext_mark_unwritten(ex);
3204                 else
3205                         ext4_ext_mark_initialized(ex);
3206
3207                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3208                         ext4_ext_try_to_merge(handle, inode, path, ex);
3209
3210                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3211                 goto out;
3212         }
3213
3214         /* case a */
3215         memcpy(&orig_ex, ex, sizeof(orig_ex));
3216         ex->ee_len = cpu_to_le16(split - ee_block);
3217         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3218                 ext4_ext_mark_unwritten(ex);
3219
3220         /*
3221          * path may lead to new leaf, not to original leaf any more
3222          * after ext4_ext_insert_extent() returns,
3223          */
3224         err = ext4_ext_dirty(handle, inode, path + depth);
3225         if (err)
3226                 goto fix_extent_len;
3227
3228         ex2 = &newex;
3229         ex2->ee_block = cpu_to_le32(split);
3230         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3231         ext4_ext_store_pblock(ex2, newblock);
3232         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3233                 ext4_ext_mark_unwritten(ex2);
3234
3235         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3236         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3237                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3238                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3239                                 err = ext4_ext_zeroout(inode, ex2);
3240                                 zero_ex.ee_block = ex2->ee_block;
3241                                 zero_ex.ee_len = cpu_to_le16(
3242                                                 ext4_ext_get_actual_len(ex2));
3243                                 ext4_ext_store_pblock(&zero_ex,
3244                                                       ext4_ext_pblock(ex2));
3245                         } else {
3246                                 err = ext4_ext_zeroout(inode, ex);
3247                                 zero_ex.ee_block = ex->ee_block;
3248                                 zero_ex.ee_len = cpu_to_le16(
3249                                                 ext4_ext_get_actual_len(ex));
3250                                 ext4_ext_store_pblock(&zero_ex,
3251                                                       ext4_ext_pblock(ex));
3252                         }
3253                 } else {
3254                         err = ext4_ext_zeroout(inode, &orig_ex);
3255                         zero_ex.ee_block = orig_ex.ee_block;
3256                         zero_ex.ee_len = cpu_to_le16(
3257                                                 ext4_ext_get_actual_len(&orig_ex));
3258                         ext4_ext_store_pblock(&zero_ex,
3259                                               ext4_ext_pblock(&orig_ex));
3260                 }
3261
3262                 if (err)
3263                         goto fix_extent_len;
3264                 /* update the extent length and mark as initialized */
3265                 ex->ee_len = cpu_to_le16(ee_len);
3266                 ext4_ext_try_to_merge(handle, inode, path, ex);
3267                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3268                 if (err)
3269                         goto fix_extent_len;
3270
3271                 /* update extent status tree */
3272                 err = ext4_zeroout_es(inode, &zero_ex);
3273
3274                 goto out;
3275         } else if (err)
3276                 goto fix_extent_len;
3277
3278 out:
3279         ext4_ext_show_leaf(inode, path);
3280         return err;
3281
3282 fix_extent_len:
3283         ex->ee_len = orig_ex.ee_len;
3284         ext4_ext_dirty(handle, inode, path + path->p_depth);
3285         return err;
3286 }
3287
3288 /*
3289  * ext4_split_extents() splits an extent and mark extent which is covered
3290  * by @map as split_flags indicates
3291  *
3292  * It may result in splitting the extent into multiple extents (up to three)
3293  * There are three possibilities:
3294  *   a> There is no split required
3295  *   b> Splits in two extents: Split is happening at either end of the extent
3296  *   c> Splits in three extents: Somone is splitting in middle of the extent
3297  *
3298  */
3299 static int ext4_split_extent(handle_t *handle,
3300                               struct inode *inode,
3301                               struct ext4_ext_path **ppath,
3302                               struct ext4_map_blocks *map,
3303                               int split_flag,
3304                               int flags)
3305 {
3306         struct ext4_ext_path *path = *ppath;
3307         ext4_lblk_t ee_block;
3308         struct ext4_extent *ex;
3309         unsigned int ee_len, depth;
3310         int err = 0;
3311         int unwritten;
3312         int split_flag1, flags1;
3313         int allocated = map->m_len;
3314
3315         depth = ext_depth(inode);
3316         ex = path[depth].p_ext;
3317         ee_block = le32_to_cpu(ex->ee_block);
3318         ee_len = ext4_ext_get_actual_len(ex);
3319         unwritten = ext4_ext_is_unwritten(ex);
3320
3321         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3322                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3323                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3324                 if (unwritten)
3325                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3326                                        EXT4_EXT_MARK_UNWRIT2;
3327                 if (split_flag & EXT4_EXT_DATA_VALID2)
3328                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3329                 err = ext4_split_extent_at(handle, inode, ppath,
3330                                 map->m_lblk + map->m_len, split_flag1, flags1);
3331                 if (err)
3332                         goto out;
3333         } else {
3334                 allocated = ee_len - (map->m_lblk - ee_block);
3335         }
3336         /*
3337          * Update path is required because previous ext4_split_extent_at() may
3338          * result in split of original leaf or extent zeroout.
3339          */
3340         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3341         if (IS_ERR(path))
3342                 return PTR_ERR(path);
3343         depth = ext_depth(inode);
3344         ex = path[depth].p_ext;
3345         if (!ex) {
3346                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3347                                  (unsigned long) map->m_lblk);
3348                 return -EIO;
3349         }
3350         unwritten = ext4_ext_is_unwritten(ex);
3351         split_flag1 = 0;
3352
3353         if (map->m_lblk >= ee_block) {
3354                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3355                 if (unwritten) {
3356                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3357                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3358                                                      EXT4_EXT_MARK_UNWRIT2);
3359                 }
3360                 err = ext4_split_extent_at(handle, inode, ppath,
3361                                 map->m_lblk, split_flag1, flags);
3362                 if (err)
3363                         goto out;
3364         }
3365
3366         ext4_ext_show_leaf(inode, path);
3367 out:
3368         return err ? err : allocated;
3369 }
3370
3371 /*
3372  * This function is called by ext4_ext_map_blocks() if someone tries to write
3373  * to an unwritten extent. It may result in splitting the unwritten
3374  * extent into multiple extents (up to three - one initialized and two
3375  * unwritten).
3376  * There are three possibilities:
3377  *   a> There is no split required: Entire extent should be initialized
3378  *   b> Splits in two extents: Write is happening at either end of the extent
3379  *   c> Splits in three extents: Somone is writing in middle of the extent
3380  *
3381  * Pre-conditions:
3382  *  - The extent pointed to by 'path' is unwritten.
3383  *  - The extent pointed to by 'path' contains a superset
3384  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3385  *
3386  * Post-conditions on success:
3387  *  - the returned value is the number of blocks beyond map->l_lblk
3388  *    that are allocated and initialized.
3389  *    It is guaranteed to be >= map->m_len.
3390  */
3391 static int ext4_ext_convert_to_initialized(handle_t *handle,
3392                                            struct inode *inode,
3393                                            struct ext4_map_blocks *map,
3394                                            struct ext4_ext_path **ppath,
3395                                            int flags)
3396 {
3397         struct ext4_ext_path *path = *ppath;
3398         struct ext4_sb_info *sbi;
3399         struct ext4_extent_header *eh;
3400         struct ext4_map_blocks split_map;
3401         struct ext4_extent zero_ex;
3402         struct ext4_extent *ex, *abut_ex;
3403         ext4_lblk_t ee_block, eof_block;
3404         unsigned int ee_len, depth, map_len = map->m_len;
3405         int allocated = 0, max_zeroout = 0;
3406         int err = 0;
3407         int split_flag = 0;
3408
3409         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3410                 "block %llu, max_blocks %u\n", inode->i_ino,
3411                 (unsigned long long)map->m_lblk, map_len);
3412
3413         sbi = EXT4_SB(inode->i_sb);
3414         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3415                 inode->i_sb->s_blocksize_bits;
3416         if (eof_block < map->m_lblk + map_len)
3417                 eof_block = map->m_lblk + map_len;
3418
3419         depth = ext_depth(inode);
3420         eh = path[depth].p_hdr;
3421         ex = path[depth].p_ext;
3422         ee_block = le32_to_cpu(ex->ee_block);
3423         ee_len = ext4_ext_get_actual_len(ex);
3424         zero_ex.ee_len = 0;
3425
3426         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3427
3428         /* Pre-conditions */
3429         BUG_ON(!ext4_ext_is_unwritten(ex));
3430         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3431
3432         /*
3433          * Attempt to transfer newly initialized blocks from the currently
3434          * unwritten extent to its neighbor. This is much cheaper
3435          * than an insertion followed by a merge as those involve costly
3436          * memmove() calls. Transferring to the left is the common case in
3437          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3438          * followed by append writes.
3439          *
3440          * Limitations of the current logic:
3441          *  - L1: we do not deal with writes covering the whole extent.
3442          *    This would require removing the extent if the transfer
3443          *    is possible.
3444          *  - L2: we only attempt to merge with an extent stored in the
3445          *    same extent tree node.
3446          */
3447         if ((map->m_lblk == ee_block) &&
3448                 /* See if we can merge left */
3449                 (map_len < ee_len) &&           /*L1*/
3450                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3451                 ext4_lblk_t prev_lblk;
3452                 ext4_fsblk_t prev_pblk, ee_pblk;
3453                 unsigned int prev_len;
3454
3455                 abut_ex = ex - 1;
3456                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3457                 prev_len = ext4_ext_get_actual_len(abut_ex);
3458                 prev_pblk = ext4_ext_pblock(abut_ex);
3459                 ee_pblk = ext4_ext_pblock(ex);
3460
3461                 /*
3462                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3463                  * upon those conditions:
3464                  * - C1: abut_ex is initialized,
3465                  * - C2: abut_ex is logically abutting ex,
3466                  * - C3: abut_ex is physically abutting ex,
3467                  * - C4: abut_ex can receive the additional blocks without
3468                  *   overflowing the (initialized) length limit.
3469                  */
3470                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3471                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3472                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3473                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3474                         err = ext4_ext_get_access(handle, inode, path + depth);
3475                         if (err)
3476                                 goto out;
3477
3478                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3479                                 map, ex, abut_ex);
3480
3481                         /* Shift the start of ex by 'map_len' blocks */
3482                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3483                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3484                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3485                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3486
3487                         /* Extend abut_ex by 'map_len' blocks */
3488                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3489
3490                         /* Result: number of initialized blocks past m_lblk */
3491                         allocated = map_len;
3492                 }
3493         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3494                    (map_len < ee_len) &&        /*L1*/
3495                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3496                 /* See if we can merge right */
3497                 ext4_lblk_t next_lblk;
3498                 ext4_fsblk_t next_pblk, ee_pblk;
3499                 unsigned int next_len;
3500
3501                 abut_ex = ex + 1;
3502                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3503                 next_len = ext4_ext_get_actual_len(abut_ex);
3504                 next_pblk = ext4_ext_pblock(abut_ex);
3505                 ee_pblk = ext4_ext_pblock(ex);
3506
3507                 /*
3508                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3509                  * upon those conditions:
3510                  * - C1: abut_ex is initialized,
3511                  * - C2: abut_ex is logically abutting ex,
3512                  * - C3: abut_ex is physically abutting ex,
3513                  * - C4: abut_ex can receive the additional blocks without
3514                  *   overflowing the (initialized) length limit.
3515                  */
3516                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3517                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3518                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3519                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3520                         err = ext4_ext_get_access(handle, inode, path + depth);
3521                         if (err)
3522                                 goto out;
3523
3524                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3525                                 map, ex, abut_ex);
3526
3527                         /* Shift the start of abut_ex by 'map_len' blocks */
3528                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3529                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3530                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3531                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3532
3533                         /* Extend abut_ex by 'map_len' blocks */
3534                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3535
3536                         /* Result: number of initialized blocks past m_lblk */
3537                         allocated = map_len;
3538                 }
3539         }
3540         if (allocated) {
3541                 /* Mark the block containing both extents as dirty */
3542                 ext4_ext_dirty(handle, inode, path + depth);
3543
3544                 /* Update path to point to the right extent */
3545                 path[depth].p_ext = abut_ex;
3546                 goto out;
3547         } else
3548                 allocated = ee_len - (map->m_lblk - ee_block);
3549
3550         WARN_ON(map->m_lblk < ee_block);
3551         /*
3552          * It is safe to convert extent to initialized via explicit
3553          * zeroout only if extent is fully inside i_size or new_size.
3554          */
3555         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3556
3557         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3558                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3559                         (inode->i_sb->s_blocksize_bits - 10);
3560
3561         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3562         if (max_zeroout && (ee_len <= max_zeroout)) {
3563                 err = ext4_ext_zeroout(inode, ex);
3564                 if (err)
3565                         goto out;
3566                 zero_ex.ee_block = ex->ee_block;
3567                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3568                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3569
3570                 err = ext4_ext_get_access(handle, inode, path + depth);
3571                 if (err)
3572                         goto out;
3573                 ext4_ext_mark_initialized(ex);
3574                 ext4_ext_try_to_merge(handle, inode, path, ex);
3575                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3576                 goto out;
3577         }
3578
3579         /*
3580          * four cases:
3581          * 1. split the extent into three extents.
3582          * 2. split the extent into two extents, zeroout the first half.
3583          * 3. split the extent into two extents, zeroout the second half.
3584          * 4. split the extent into two extents with out zeroout.
3585          */
3586         split_map.m_lblk = map->m_lblk;
3587         split_map.m_len = map->m_len;
3588
3589         if (max_zeroout && (allocated > map->m_len)) {
3590                 if (allocated <= max_zeroout) {
3591                         /* case 3 */
3592                         zero_ex.ee_block =
3593                                          cpu_to_le32(map->m_lblk);
3594                         zero_ex.ee_len = cpu_to_le16(allocated);
3595                         ext4_ext_store_pblock(&zero_ex,
3596                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3597                         err = ext4_ext_zeroout(inode, &zero_ex);
3598                         if (err)
3599                                 goto out;
3600                         split_map.m_lblk = map->m_lblk;
3601                         split_map.m_len = allocated;
3602                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3603                         /* case 2 */
3604                         if (map->m_lblk != ee_block) {
3605                                 zero_ex.ee_block = ex->ee_block;
3606                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3607                                                         ee_block);
3608                                 ext4_ext_store_pblock(&zero_ex,
3609                                                       ext4_ext_pblock(ex));
3610                                 err = ext4_ext_zeroout(inode, &zero_ex);
3611                                 if (err)
3612                                         goto out;
3613                         }
3614
3615                         split_map.m_lblk = ee_block;
3616                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3617                         allocated = map->m_len;
3618                 }
3619         }
3620
3621         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3622                                 flags);
3623         if (err > 0)
3624                 err = 0;
3625 out:
3626         /* If we have gotten a failure, don't zero out status tree */
3627         if (!err)
3628                 err = ext4_zeroout_es(inode, &zero_ex);
3629         return err ? err : allocated;
3630 }
3631
3632 /*
3633  * This function is called by ext4_ext_map_blocks() from
3634  * ext4_get_blocks_dio_write() when DIO to write
3635  * to an unwritten extent.
3636  *
3637  * Writing to an unwritten extent may result in splitting the unwritten
3638  * extent into multiple initialized/unwritten extents (up to three)
3639  * There are three possibilities:
3640  *   a> There is no split required: Entire extent should be unwritten
3641  *   b> Splits in two extents: Write is happening at either end of the extent
3642  *   c> Splits in three extents: Somone is writing in middle of the extent
3643  *
3644  * This works the same way in the case of initialized -> unwritten conversion.
3645  *
3646  * One of more index blocks maybe needed if the extent tree grow after
3647  * the unwritten extent split. To prevent ENOSPC occur at the IO
3648  * complete, we need to split the unwritten extent before DIO submit
3649  * the IO. The unwritten extent called at this time will be split
3650  * into three unwritten extent(at most). After IO complete, the part
3651  * being filled will be convert to initialized by the end_io callback function
3652  * via ext4_convert_unwritten_extents().
3653  *
3654  * Returns the size of unwritten extent to be written on success.
3655  */
3656 static int ext4_split_convert_extents(handle_t *handle,
3657                                         struct inode *inode,
3658                                         struct ext4_map_blocks *map,
3659                                         struct ext4_ext_path **ppath,
3660                                         int flags)
3661 {
3662         struct ext4_ext_path *path = *ppath;
3663         ext4_lblk_t eof_block;
3664         ext4_lblk_t ee_block;
3665         struct ext4_extent *ex;
3666         unsigned int ee_len;
3667         int split_flag = 0, depth;
3668
3669         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3670                   __func__, inode->i_ino,
3671                   (unsigned long long)map->m_lblk, map->m_len);
3672
3673         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3674                 inode->i_sb->s_blocksize_bits;
3675         if (eof_block < map->m_lblk + map->m_len)
3676                 eof_block = map->m_lblk + map->m_len;
3677         /*
3678          * It is safe to convert extent to initialized via explicit
3679          * zeroout only if extent is fully insde i_size or new_size.
3680          */
3681         depth = ext_depth(inode);
3682         ex = path[depth].p_ext;
3683         ee_block = le32_to_cpu(ex->ee_block);
3684         ee_len = ext4_ext_get_actual_len(ex);
3685
3686         /* Convert to unwritten */
3687         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3688                 split_flag |= EXT4_EXT_DATA_VALID1;
3689         /* Convert to initialized */
3690         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3691                 split_flag |= ee_block + ee_len <= eof_block ?
3692                               EXT4_EXT_MAY_ZEROOUT : 0;
3693                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3694         }
3695         flags |= EXT4_GET_BLOCKS_PRE_IO;
3696         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3697 }
3698
3699 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3700                                                 struct inode *inode,
3701                                                 struct ext4_map_blocks *map,
3702                                                 struct ext4_ext_path **ppath)
3703 {
3704         struct ext4_ext_path *path = *ppath;
3705         struct ext4_extent *ex;
3706         ext4_lblk_t ee_block;
3707         unsigned int ee_len;
3708         int depth;
3709         int err = 0;
3710
3711         depth = ext_depth(inode);
3712         ex = path[depth].p_ext;
3713         ee_block = le32_to_cpu(ex->ee_block);
3714         ee_len = ext4_ext_get_actual_len(ex);
3715
3716         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3717                 "block %llu, max_blocks %u\n", inode->i_ino,
3718                   (unsigned long long)ee_block, ee_len);
3719
3720         /* If extent is larger than requested it is a clear sign that we still
3721          * have some extent state machine issues left. So extent_split is still
3722          * required.
3723          * TODO: Once all related issues will be fixed this situation should be
3724          * illegal.
3725          */
3726         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3727 #ifdef EXT4_DEBUG
3728                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3729                              " len %u; IO logical block %llu, len %u\n",
3730                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3731                              (unsigned long long)map->m_lblk, map->m_len);
3732 #endif
3733                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3734                                                  EXT4_GET_BLOCKS_CONVERT);
3735                 if (err < 0)
3736                         return err;
3737                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3738                 if (IS_ERR(path))
3739                         return PTR_ERR(path);
3740                 depth = ext_depth(inode);
3741                 ex = path[depth].p_ext;
3742         }
3743
3744         err = ext4_ext_get_access(handle, inode, path + depth);
3745         if (err)
3746                 goto out;
3747         /* first mark the extent as initialized */
3748         ext4_ext_mark_initialized(ex);
3749
3750         /* note: ext4_ext_correct_indexes() isn't needed here because
3751          * borders are not changed
3752          */
3753         ext4_ext_try_to_merge(handle, inode, path, ex);
3754
3755         /* Mark modified extent as dirty */
3756         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3757 out:
3758         ext4_ext_show_leaf(inode, path);
3759         return err;
3760 }
3761
3762 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3763                         sector_t block, int count)
3764 {
3765         int i;
3766         for (i = 0; i < count; i++)
3767                 unmap_underlying_metadata(bdev, block + i);
3768 }
3769
3770 /*
3771  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3772  */
3773 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3774                               ext4_lblk_t lblk,
3775                               struct ext4_ext_path *path,
3776                               unsigned int len)
3777 {
3778         int i, depth;
3779         struct ext4_extent_header *eh;
3780         struct ext4_extent *last_ex;
3781
3782         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3783                 return 0;
3784
3785         depth = ext_depth(inode);
3786         eh = path[depth].p_hdr;
3787
3788         /*
3789          * We're going to remove EOFBLOCKS_FL entirely in future so we
3790          * do not care for this case anymore. Simply remove the flag
3791          * if there are no extents.
3792          */
3793         if (unlikely(!eh->eh_entries))
3794                 goto out;
3795         last_ex = EXT_LAST_EXTENT(eh);
3796         /*
3797          * We should clear the EOFBLOCKS_FL flag if we are writing the
3798          * last block in the last extent in the file.  We test this by
3799          * first checking to see if the caller to
3800          * ext4_ext_get_blocks() was interested in the last block (or
3801          * a block beyond the last block) in the current extent.  If
3802          * this turns out to be false, we can bail out from this
3803          * function immediately.
3804          */
3805         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3806             ext4_ext_get_actual_len(last_ex))
3807                 return 0;
3808         /*
3809          * If the caller does appear to be planning to write at or
3810          * beyond the end of the current extent, we then test to see
3811          * if the current extent is the last extent in the file, by
3812          * checking to make sure it was reached via the rightmost node
3813          * at each level of the tree.
3814          */
3815         for (i = depth-1; i >= 0; i--)
3816                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3817                         return 0;
3818 out:
3819         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3820         return ext4_mark_inode_dirty(handle, inode);
3821 }
3822
3823 /**
3824  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3825  *
3826  * Return 1 if there is a delalloc block in the range, otherwise 0.
3827  */
3828 int ext4_find_delalloc_range(struct inode *inode,
3829                              ext4_lblk_t lblk_start,
3830                              ext4_lblk_t lblk_end)
3831 {
3832         struct extent_status es;
3833
3834         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3835         if (es.es_len == 0)
3836                 return 0; /* there is no delay extent in this tree */
3837         else if (es.es_lblk <= lblk_start &&
3838                  lblk_start < es.es_lblk + es.es_len)
3839                 return 1;
3840         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3841                 return 1;
3842         else
3843                 return 0;
3844 }
3845
3846 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3847 {
3848         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3849         ext4_lblk_t lblk_start, lblk_end;
3850         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3851         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3852
3853         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3854 }
3855
3856 /**
3857  * Determines how many complete clusters (out of those specified by the 'map')
3858  * are under delalloc and were reserved quota for.
3859  * This function is called when we are writing out the blocks that were
3860  * originally written with their allocation delayed, but then the space was
3861  * allocated using fallocate() before the delayed allocation could be resolved.
3862  * The cases to look for are:
3863  * ('=' indicated delayed allocated blocks
3864  *  '-' indicates non-delayed allocated blocks)
3865  * (a) partial clusters towards beginning and/or end outside of allocated range
3866  *     are not delalloc'ed.
3867  *      Ex:
3868  *      |----c---=|====c====|====c====|===-c----|
3869  *               |++++++ allocated ++++++|
3870  *      ==> 4 complete clusters in above example
3871  *
3872  * (b) partial cluster (outside of allocated range) towards either end is
3873  *     marked for delayed allocation. In this case, we will exclude that
3874  *     cluster.
3875  *      Ex:
3876  *      |----====c========|========c========|
3877  *           |++++++ allocated ++++++|
3878  *      ==> 1 complete clusters in above example
3879  *
3880  *      Ex:
3881  *      |================c================|
3882  *            |++++++ allocated ++++++|
3883  *      ==> 0 complete clusters in above example
3884  *
3885  * The ext4_da_update_reserve_space will be called only if we
3886  * determine here that there were some "entire" clusters that span
3887  * this 'allocated' range.
3888  * In the non-bigalloc case, this function will just end up returning num_blks
3889  * without ever calling ext4_find_delalloc_range.
3890  */
3891 static unsigned int
3892 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3893                            unsigned int num_blks)
3894 {
3895         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3896         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3897         ext4_lblk_t lblk_from, lblk_to, c_offset;
3898         unsigned int allocated_clusters = 0;
3899
3900         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3901         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3902
3903         /* max possible clusters for this allocation */
3904         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3905
3906         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3907
3908         /* Check towards left side */
3909         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3910         if (c_offset) {
3911                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3912                 lblk_to = lblk_from + c_offset - 1;
3913
3914                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3915                         allocated_clusters--;
3916         }
3917
3918         /* Now check towards right. */
3919         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3920         if (allocated_clusters && c_offset) {
3921                 lblk_from = lblk_start + num_blks;
3922                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3923
3924                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3925                         allocated_clusters--;
3926         }
3927
3928         return allocated_clusters;
3929 }
3930
3931 static int
3932 convert_initialized_extent(handle_t *handle, struct inode *inode,
3933                            struct ext4_map_blocks *map,
3934                            struct ext4_ext_path **ppath, int flags,
3935                            unsigned int allocated, ext4_fsblk_t newblock)
3936 {
3937         struct ext4_ext_path *path = *ppath;
3938         struct ext4_extent *ex;
3939         ext4_lblk_t ee_block;
3940         unsigned int ee_len;
3941         int depth;
3942         int err = 0;
3943
3944         /*
3945          * Make sure that the extent is no bigger than we support with
3946          * unwritten extent
3947          */
3948         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3949                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3950
3951         depth = ext_depth(inode);
3952         ex = path[depth].p_ext;
3953         ee_block = le32_to_cpu(ex->ee_block);
3954         ee_len = ext4_ext_get_actual_len(ex);
3955
3956         ext_debug("%s: inode %lu, logical"
3957                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3958                   (unsigned long long)ee_block, ee_len);
3959
3960         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3961                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3962                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3963                 if (err < 0)
3964                         return err;
3965                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3966                 if (IS_ERR(path))
3967                         return PTR_ERR(path);
3968                 depth = ext_depth(inode);
3969                 ex = path[depth].p_ext;
3970                 if (!ex) {
3971                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3972                                          (unsigned long) map->m_lblk);
3973                         return -EIO;
3974                 }
3975         }
3976
3977         err = ext4_ext_get_access(handle, inode, path + depth);
3978         if (err)
3979                 return err;
3980         /* first mark the extent as unwritten */
3981         ext4_ext_mark_unwritten(ex);
3982
3983         /* note: ext4_ext_correct_indexes() isn't needed here because
3984          * borders are not changed
3985          */
3986         ext4_ext_try_to_merge(handle, inode, path, ex);
3987
3988         /* Mark modified extent as dirty */
3989         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3990         if (err)
3991                 return err;
3992         ext4_ext_show_leaf(inode, path);
3993
3994         ext4_update_inode_fsync_trans(handle, inode, 1);
3995         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3996         if (err)
3997                 return err;
3998         map->m_flags |= EXT4_MAP_UNWRITTEN;
3999         if (allocated > map->m_len)
4000                 allocated = map->m_len;
4001         map->m_len = allocated;
4002         return allocated;
4003 }
4004
4005 static int
4006 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4007                         struct ext4_map_blocks *map,
4008                         struct ext4_ext_path **ppath, int flags,
4009                         unsigned int allocated, ext4_fsblk_t newblock)
4010 {
4011         struct ext4_ext_path *path = *ppath;
4012         int ret = 0;
4013         int err = 0;
4014         ext4_io_end_t *io = ext4_inode_aio(inode);
4015
4016         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4017                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4018                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4019                   flags, allocated);
4020         ext4_ext_show_leaf(inode, path);
4021
4022         /*
4023          * When writing into unwritten space, we should not fail to
4024          * allocate metadata blocks for the new extent block if needed.
4025          */
4026         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4027
4028         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4029                                                     allocated, newblock);
4030
4031         /* get_block() before submit the IO, split the extent */
4032         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4033                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4034                                          flags | EXT4_GET_BLOCKS_CONVERT);
4035                 if (ret <= 0)
4036                         goto out;
4037                 /*
4038                  * Flag the inode(non aio case) or end_io struct (aio case)
4039                  * that this IO needs to conversion to written when IO is
4040                  * completed
4041                  */
4042                 if (io)
4043                         ext4_set_io_unwritten_flag(inode, io);
4044                 else
4045                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4046                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4047                 goto out;
4048         }
4049         /* IO end_io complete, convert the filled extent to written */
4050         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4051                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4052                                                            ppath);
4053                 if (ret >= 0) {
4054                         ext4_update_inode_fsync_trans(handle, inode, 1);
4055                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4056                                                  path, map->m_len);
4057                 } else
4058                         err = ret;
4059                 map->m_flags |= EXT4_MAP_MAPPED;
4060                 map->m_pblk = newblock;
4061                 if (allocated > map->m_len)
4062                         allocated = map->m_len;
4063                 map->m_len = allocated;
4064                 goto out2;
4065         }
4066         /* buffered IO case */
4067         /*
4068          * repeat fallocate creation request
4069          * we already have an unwritten extent
4070          */
4071         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4072                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4073                 goto map_out;
4074         }
4075
4076         /* buffered READ or buffered write_begin() lookup */
4077         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4078                 /*
4079                  * We have blocks reserved already.  We
4080                  * return allocated blocks so that delalloc
4081                  * won't do block reservation for us.  But
4082                  * the buffer head will be unmapped so that
4083                  * a read from the block returns 0s.
4084                  */
4085                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4086                 goto out1;
4087         }
4088
4089         /* buffered write, writepage time, convert*/
4090         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4091         if (ret >= 0)
4092                 ext4_update_inode_fsync_trans(handle, inode, 1);
4093 out:
4094         if (ret <= 0) {
4095                 err = ret;
4096                 goto out2;
4097         } else
4098                 allocated = ret;
4099         map->m_flags |= EXT4_MAP_NEW;
4100         /*
4101          * if we allocated more blocks than requested
4102          * we need to make sure we unmap the extra block
4103          * allocated. The actual needed block will get
4104          * unmapped later when we find the buffer_head marked
4105          * new.
4106          */
4107         if (allocated > map->m_len) {
4108                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
4109                                         newblock + map->m_len,
4110                                         allocated - map->m_len);
4111                 allocated = map->m_len;
4112         }
4113         map->m_len = allocated;
4114
4115         /*
4116          * If we have done fallocate with the offset that is already
4117          * delayed allocated, we would have block reservation
4118          * and quota reservation done in the delayed write path.
4119          * But fallocate would have already updated quota and block
4120          * count for this offset. So cancel these reservation
4121          */
4122         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4123                 unsigned int reserved_clusters;
4124                 reserved_clusters = get_reserved_cluster_alloc(inode,
4125                                 map->m_lblk, map->m_len);
4126                 if (reserved_clusters)
4127                         ext4_da_update_reserve_space(inode,
4128                                                      reserved_clusters,
4129                                                      0);
4130         }
4131
4132 map_out:
4133         map->m_flags |= EXT4_MAP_MAPPED;
4134         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4135                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4136                                          map->m_len);
4137                 if (err < 0)
4138                         goto out2;
4139         }
4140 out1:
4141         if (allocated > map->m_len)
4142                 allocated = map->m_len;
4143         ext4_ext_show_leaf(inode, path);
4144         map->m_pblk = newblock;
4145         map->m_len = allocated;
4146 out2:
4147         return err ? err : allocated;
4148 }
4149
4150 /*
4151  * get_implied_cluster_alloc - check to see if the requested
4152  * allocation (in the map structure) overlaps with a cluster already
4153  * allocated in an extent.
4154  *      @sb     The filesystem superblock structure
4155  *      @map    The requested lblk->pblk mapping
4156  *      @ex     The extent structure which might contain an implied
4157  *                      cluster allocation
4158  *
4159  * This function is called by ext4_ext_map_blocks() after we failed to
4160  * find blocks that were already in the inode's extent tree.  Hence,
4161  * we know that the beginning of the requested region cannot overlap
4162  * the extent from the inode's extent tree.  There are three cases we
4163  * want to catch.  The first is this case:
4164  *
4165  *               |--- cluster # N--|
4166  *    |--- extent ---|  |---- requested region ---|
4167  *                      |==========|
4168  *
4169  * The second case that we need to test for is this one:
4170  *
4171  *   |--------- cluster # N ----------------|
4172  *         |--- requested region --|   |------- extent ----|
4173  *         |=======================|
4174  *
4175  * The third case is when the requested region lies between two extents
4176  * within the same cluster:
4177  *          |------------- cluster # N-------------|
4178  * |----- ex -----|                  |---- ex_right ----|
4179  *                  |------ requested region ------|
4180  *                  |================|
4181  *
4182  * In each of the above cases, we need to set the map->m_pblk and
4183  * map->m_len so it corresponds to the return the extent labelled as
4184  * "|====|" from cluster #N, since it is already in use for data in
4185  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4186  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4187  * as a new "allocated" block region.  Otherwise, we will return 0 and
4188  * ext4_ext_map_blocks() will then allocate one or more new clusters
4189  * by calling ext4_mb_new_blocks().
4190  */
4191 static int get_implied_cluster_alloc(struct super_block *sb,
4192                                      struct ext4_map_blocks *map,
4193                                      struct ext4_extent *ex,
4194                                      struct ext4_ext_path *path)
4195 {
4196         struct ext4_sb_info *sbi = EXT4_SB(sb);
4197         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4198         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4199         ext4_lblk_t rr_cluster_start;
4200         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4201         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4202         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4203
4204         /* The extent passed in that we are trying to match */
4205         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4206         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4207
4208         /* The requested region passed into ext4_map_blocks() */
4209         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4210
4211         if ((rr_cluster_start == ex_cluster_end) ||
4212             (rr_cluster_start == ex_cluster_start)) {
4213                 if (rr_cluster_start == ex_cluster_end)
4214                         ee_start += ee_len - 1;
4215                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4216                 map->m_len = min(map->m_len,
4217                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4218                 /*
4219                  * Check for and handle this case:
4220                  *
4221                  *   |--------- cluster # N-------------|
4222                  *                     |------- extent ----|
4223                  *         |--- requested region ---|
4224                  *         |===========|
4225                  */
4226
4227                 if (map->m_lblk < ee_block)
4228                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4229
4230                 /*
4231                  * Check for the case where there is already another allocated
4232                  * block to the right of 'ex' but before the end of the cluster.
4233                  *
4234                  *          |------------- cluster # N-------------|
4235                  * |----- ex -----|                  |---- ex_right ----|
4236                  *                  |------ requested region ------|
4237                  *                  |================|
4238                  */
4239                 if (map->m_lblk > ee_block) {
4240                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4241                         map->m_len = min(map->m_len, next - map->m_lblk);
4242                 }
4243
4244                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4245                 return 1;
4246         }
4247
4248         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4249         return 0;
4250 }
4251
4252
4253 /*
4254  * Block allocation/map/preallocation routine for extents based files
4255  *
4256  *
4257  * Need to be called with
4258  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4259  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4260  *
4261  * return > 0, number of of blocks already mapped/allocated
4262  *          if create == 0 and these are pre-allocated blocks
4263  *              buffer head is unmapped
4264  *          otherwise blocks are mapped
4265  *
4266  * return = 0, if plain look up failed (blocks have not been allocated)
4267  *          buffer head is unmapped
4268  *
4269  * return < 0, error case.
4270  */
4271 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4272                         struct ext4_map_blocks *map, int flags)
4273 {
4274         struct ext4_ext_path *path = NULL;
4275         struct ext4_extent newex, *ex, *ex2;
4276         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4277         ext4_fsblk_t newblock = 0;
4278         int free_on_err = 0, err = 0, depth, ret;
4279         unsigned int allocated = 0, offset = 0;
4280         unsigned int allocated_clusters = 0;
4281         struct ext4_allocation_request ar;
4282         ext4_io_end_t *io = ext4_inode_aio(inode);
4283         ext4_lblk_t cluster_offset;
4284         int set_unwritten = 0;
4285         bool map_from_cluster = false;
4286
4287         ext_debug("blocks %u/%u requested for inode %lu\n",
4288                   map->m_lblk, map->m_len, inode->i_ino);
4289         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4290
4291         /* find extent for this block */
4292         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4293         if (IS_ERR(path)) {
4294                 err = PTR_ERR(path);
4295                 path = NULL;
4296                 goto out2;
4297         }
4298
4299         depth = ext_depth(inode);
4300
4301         /*
4302          * consistent leaf must not be empty;
4303          * this situation is possible, though, _during_ tree modification;
4304          * this is why assert can't be put in ext4_find_extent()
4305          */
4306         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4307                 EXT4_ERROR_INODE(inode, "bad extent address "
4308                                  "lblock: %lu, depth: %d pblock %lld",
4309                                  (unsigned long) map->m_lblk, depth,
4310                                  path[depth].p_block);
4311                 err = -EIO;
4312                 goto out2;
4313         }
4314
4315         ex = path[depth].p_ext;
4316         if (ex) {
4317                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4318                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4319                 unsigned short ee_len;
4320
4321
4322                 /*
4323                  * unwritten extents are treated as holes, except that
4324                  * we split out initialized portions during a write.
4325                  */
4326                 ee_len = ext4_ext_get_actual_len(ex);
4327
4328                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4329
4330                 /* if found extent covers block, simply return it */
4331                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4332                         newblock = map->m_lblk - ee_block + ee_start;
4333                         /* number of remaining blocks in the extent */
4334                         allocated = ee_len - (map->m_lblk - ee_block);
4335                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4336                                   ee_block, ee_len, newblock);
4337
4338                         /*
4339                          * If the extent is initialized check whether the
4340                          * caller wants to convert it to unwritten.
4341                          */
4342                         if ((!ext4_ext_is_unwritten(ex)) &&
4343                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4344                                 allocated = convert_initialized_extent(
4345                                                 handle, inode, map, &path,
4346                                                 flags, allocated, newblock);
4347                                 goto out2;
4348                         } else if (!ext4_ext_is_unwritten(ex))
4349                                 goto out;
4350
4351                         ret = ext4_ext_handle_unwritten_extents(
4352                                 handle, inode, map, &path, flags,
4353                                 allocated, newblock);
4354                         if (ret < 0)
4355                                 err = ret;
4356                         else
4357                                 allocated = ret;
4358                         goto out2;
4359                 }
4360         }
4361
4362         /*
4363          * requested block isn't allocated yet;
4364          * we couldn't try to create block if create flag is zero
4365          */
4366         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4367                 /*
4368                  * put just found gap into cache to speed up
4369                  * subsequent requests
4370                  */
4371                 ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4372                 goto out2;
4373         }
4374
4375         /*
4376          * Okay, we need to do block allocation.
4377          */
4378         newex.ee_block = cpu_to_le32(map->m_lblk);
4379         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4380
4381         /*
4382          * If we are doing bigalloc, check to see if the extent returned
4383          * by ext4_find_extent() implies a cluster we can use.
4384          */
4385         if (cluster_offset && ex &&
4386             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4387                 ar.len = allocated = map->m_len;
4388                 newblock = map->m_pblk;
4389                 map_from_cluster = true;
4390                 goto got_allocated_blocks;
4391         }
4392
4393         /* find neighbour allocated blocks */
4394         ar.lleft = map->m_lblk;
4395         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4396         if (err)
4397                 goto out2;
4398         ar.lright = map->m_lblk;
4399         ex2 = NULL;
4400         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4401         if (err)
4402                 goto out2;
4403
4404         /* Check if the extent after searching to the right implies a
4405          * cluster we can use. */
4406         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4407             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4408                 ar.len = allocated = map->m_len;
4409                 newblock = map->m_pblk;
4410                 map_from_cluster = true;
4411                 goto got_allocated_blocks;
4412         }
4413
4414         /*
4415          * See if request is beyond maximum number of blocks we can have in
4416          * a single extent. For an initialized extent this limit is
4417          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4418          * EXT_UNWRITTEN_MAX_LEN.
4419          */
4420         if (map->m_len > EXT_INIT_MAX_LEN &&
4421             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4422                 map->m_len = EXT_INIT_MAX_LEN;
4423         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4424                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4425                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4426
4427         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4428         newex.ee_len = cpu_to_le16(map->m_len);
4429         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4430         if (err)
4431                 allocated = ext4_ext_get_actual_len(&newex);
4432         else
4433                 allocated = map->m_len;
4434
4435         /* allocate new block */
4436         ar.inode = inode;
4437         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4438         ar.logical = map->m_lblk;
4439         /*
4440          * We calculate the offset from the beginning of the cluster
4441          * for the logical block number, since when we allocate a
4442          * physical cluster, the physical block should start at the
4443          * same offset from the beginning of the cluster.  This is
4444          * needed so that future calls to get_implied_cluster_alloc()
4445          * work correctly.
4446          */
4447         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4448         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4449         ar.goal -= offset;
4450         ar.logical -= offset;
4451         if (S_ISREG(inode->i_mode))
4452                 ar.flags = EXT4_MB_HINT_DATA;
4453         else
4454                 /* disable in-core preallocation for non-regular files */
4455                 ar.flags = 0;
4456         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4457                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4458         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4459                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4460         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4461                 ar.flags |= EXT4_MB_USE_RESERVED;
4462         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4463         if (!newblock)
4464                 goto out2;
4465         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4466                   ar.goal, newblock, allocated);
4467         free_on_err = 1;
4468         allocated_clusters = ar.len;
4469         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4470         if (ar.len > allocated)
4471                 ar.len = allocated;
4472
4473 got_allocated_blocks:
4474         /* try to insert new extent into found leaf and return */
4475         ext4_ext_store_pblock(&newex, newblock + offset);
4476         newex.ee_len = cpu_to_le16(ar.len);
4477         /* Mark unwritten */
4478         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4479                 ext4_ext_mark_unwritten(&newex);
4480                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4481                 /*
4482                  * io_end structure was created for every IO write to an
4483                  * unwritten extent. To avoid unnecessary conversion,
4484                  * here we flag the IO that really needs the conversion.
4485                  * For non asycn direct IO case, flag the inode state
4486                  * that we need to perform conversion when IO is done.
4487                  */
4488                 if (flags & EXT4_GET_BLOCKS_PRE_IO)
4489                         set_unwritten = 1;
4490         }
4491
4492         err = 0;
4493         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4494                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4495                                          path, ar.len);
4496         if (!err)
4497                 err = ext4_ext_insert_extent(handle, inode, &path,
4498                                              &newex, flags);
4499
4500         if (!err && set_unwritten) {
4501                 if (io)
4502                         ext4_set_io_unwritten_flag(inode, io);
4503                 else
4504                         ext4_set_inode_state(inode,
4505                                              EXT4_STATE_DIO_UNWRITTEN);
4506         }
4507
4508         if (err && free_on_err) {
4509                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4510                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4511                 /* free data blocks we just allocated */
4512                 /* not a good idea to call discard here directly,
4513                  * but otherwise we'd need to call it every free() */
4514                 ext4_discard_preallocations(inode);
4515                 ext4_free_blocks(handle, inode, NULL, newblock,
4516                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4517                 goto out2;
4518         }
4519
4520         /* previous routine could use block we allocated */
4521         newblock = ext4_ext_pblock(&newex);
4522         allocated = ext4_ext_get_actual_len(&newex);
4523         if (allocated > map->m_len)
4524                 allocated = map->m_len;
4525         map->m_flags |= EXT4_MAP_NEW;
4526
4527         /*
4528          * Update reserved blocks/metadata blocks after successful
4529          * block allocation which had been deferred till now.
4530          */
4531         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4532                 unsigned int reserved_clusters;
4533                 /*
4534                  * Check how many clusters we had reserved this allocated range
4535                  */
4536                 reserved_clusters = get_reserved_cluster_alloc(inode,
4537                                                 map->m_lblk, allocated);
4538                 if (!map_from_cluster) {
4539                         BUG_ON(allocated_clusters < reserved_clusters);
4540                         if (reserved_clusters < allocated_clusters) {
4541                                 struct ext4_inode_info *ei = EXT4_I(inode);
4542                                 int reservation = allocated_clusters -
4543                                                   reserved_clusters;
4544                                 /*
4545                                  * It seems we claimed few clusters outside of
4546                                  * the range of this allocation. We should give
4547                                  * it back to the reservation pool. This can
4548                                  * happen in the following case:
4549                                  *
4550                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4551                                  *   cluster has 4 blocks. Thus, the clusters
4552                                  *   are [0-3],[4-7],[8-11]...
4553                                  * * First comes delayed allocation write for
4554                                  *   logical blocks 10 & 11. Since there were no
4555                                  *   previous delayed allocated blocks in the
4556                                  *   range [8-11], we would reserve 1 cluster
4557                                  *   for this write.
4558                                  * * Next comes write for logical blocks 3 to 8.
4559                                  *   In this case, we will reserve 2 clusters
4560                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4561                                  *   that range has a delayed allocated blocks.
4562                                  *   Thus total reserved clusters now becomes 3.
4563                                  * * Now, during the delayed allocation writeout
4564                                  *   time, we will first write blocks [3-8] and
4565                                  *   allocate 3 clusters for writing these
4566                                  *   blocks. Also, we would claim all these
4567                                  *   three clusters above.
4568                                  * * Now when we come here to writeout the
4569                                  *   blocks [10-11], we would expect to claim
4570                                  *   the reservation of 1 cluster we had made
4571                                  *   (and we would claim it since there are no
4572                                  *   more delayed allocated blocks in the range
4573                                  *   [8-11]. But our reserved cluster count had
4574                                  *   already gone to 0.
4575                                  *
4576                                  *   Thus, at the step 4 above when we determine
4577                                  *   that there are still some unwritten delayed
4578                                  *   allocated blocks outside of our current
4579                                  *   block range, we should increment the
4580                                  *   reserved clusters count so that when the
4581                                  *   remaining blocks finally gets written, we
4582                                  *   could claim them.
4583                                  */
4584                                 dquot_reserve_block(inode,
4585                                                 EXT4_C2B(sbi, reservation));
4586                                 spin_lock(&ei->i_block_reservation_lock);
4587                                 ei->i_reserved_data_blocks += reservation;
4588                                 spin_unlock(&ei->i_block_reservation_lock);
4589                         }
4590                         /*
4591                          * We will claim quota for all newly allocated blocks.
4592                          * We're updating the reserved space *after* the
4593                          * correction above so we do not accidentally free
4594                          * all the metadata reservation because we might
4595                          * actually need it later on.
4596                          */
4597                         ext4_da_update_reserve_space(inode, allocated_clusters,
4598                                                         1);
4599                 }
4600         }
4601
4602         /*
4603          * Cache the extent and update transaction to commit on fdatasync only
4604          * when it is _not_ an unwritten extent.
4605          */
4606         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4607                 ext4_update_inode_fsync_trans(handle, inode, 1);
4608         else
4609                 ext4_update_inode_fsync_trans(handle, inode, 0);
4610 out:
4611         if (allocated > map->m_len)
4612                 allocated = map->m_len;
4613         ext4_ext_show_leaf(inode, path);
4614         map->m_flags |= EXT4_MAP_MAPPED;
4615         map->m_pblk = newblock;
4616         map->m_len = allocated;
4617 out2:
4618         ext4_ext_drop_refs(path);
4619         kfree(path);
4620
4621         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4622                                        err ? err : allocated);
4623         return err ? err : allocated;
4624 }
4625
4626 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4627 {
4628         struct super_block *sb = inode->i_sb;
4629         ext4_lblk_t last_block;
4630         int err = 0;
4631
4632         /*
4633          * TODO: optimization is possible here.
4634          * Probably we need not scan at all,
4635          * because page truncation is enough.
4636          */
4637
4638         /* we have to know where to truncate from in crash case */
4639         EXT4_I(inode)->i_disksize = inode->i_size;
4640         ext4_mark_inode_dirty(handle, inode);
4641
4642         last_block = (inode->i_size + sb->s_blocksize - 1)
4643                         >> EXT4_BLOCK_SIZE_BITS(sb);
4644 retry:
4645         err = ext4_es_remove_extent(inode, last_block,
4646                                     EXT_MAX_BLOCKS - last_block);
4647         if (err == -ENOMEM) {
4648                 cond_resched();
4649                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4650                 goto retry;
4651         }
4652         if (err) {
4653                 ext4_std_error(inode->i_sb, err);
4654                 return;
4655         }
4656         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4657         ext4_std_error(inode->i_sb, err);
4658 }
4659
4660 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4661                                   ext4_lblk_t len, loff_t new_size,
4662                                   int flags, int mode)
4663 {
4664         struct inode *inode = file_inode(file);
4665         handle_t *handle;
4666         int ret = 0;
4667         int ret2 = 0;
4668         int retries = 0;
4669         int depth = 0;
4670         struct ext4_map_blocks map;
4671         unsigned int credits;
4672         loff_t epos;
4673
4674         map.m_lblk = offset;
4675         map.m_len = len;
4676         /*
4677          * Don't normalize the request if it can fit in one extent so
4678          * that it doesn't get unnecessarily split into multiple
4679          * extents.
4680          */
4681         if (len <= EXT_UNWRITTEN_MAX_LEN)
4682                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4683
4684         /* Wait all existing dio workers, newcomers will block on i_mutex */
4685         ext4_inode_block_unlocked_dio(inode);
4686         inode_dio_wait(inode);
4687
4688         /*
4689          * credits to insert 1 extent into extent tree
4690          */
4691         credits = ext4_chunk_trans_blocks(inode, len);
4692         /*
4693          * We can only call ext_depth() on extent based inodes
4694          */
4695         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4696                 depth = ext_depth(inode);
4697         else
4698                 depth = -1;
4699
4700 retry:
4701         while (ret >= 0 && len) {
4702                 /*
4703                  * Recalculate credits when extent tree depth changes.
4704                  */
4705                 if (depth >= 0 && depth != ext_depth(inode)) {
4706                         credits = ext4_chunk_trans_blocks(inode, len);
4707                         depth = ext_depth(inode);
4708                 }
4709
4710                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4711                                             credits);
4712                 if (IS_ERR(handle)) {
4713                         ret = PTR_ERR(handle);
4714                         break;
4715                 }
4716                 ret = ext4_map_blocks(handle, inode, &map, flags);
4717                 if (ret <= 0) {
4718                         ext4_debug("inode #%lu: block %u: len %u: "
4719                                    "ext4_ext_map_blocks returned %d",
4720                                    inode->i_ino, map.m_lblk,
4721                                    map.m_len, ret);
4722                         ext4_mark_inode_dirty(handle, inode);
4723                         ret2 = ext4_journal_stop(handle);
4724                         break;
4725                 }
4726                 map.m_lblk += ret;
4727                 map.m_len = len = len - ret;
4728                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4729                 inode->i_ctime = ext4_current_time(inode);
4730                 if (new_size) {
4731                         if (epos > new_size)
4732                                 epos = new_size;
4733                         if (ext4_update_inode_size(inode, epos) & 0x1)
4734                                 inode->i_mtime = inode->i_ctime;
4735                 } else {
4736                         if (epos > inode->i_size)
4737                                 ext4_set_inode_flag(inode,
4738                                                     EXT4_INODE_EOFBLOCKS);
4739                 }
4740                 ext4_mark_inode_dirty(handle, inode);
4741                 ret2 = ext4_journal_stop(handle);
4742                 if (ret2)
4743                         break;
4744         }
4745         if (ret == -ENOSPC &&
4746                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4747                 ret = 0;
4748                 goto retry;
4749         }
4750
4751         ext4_inode_resume_unlocked_dio(inode);
4752
4753         return ret > 0 ? ret2 : ret;
4754 }
4755
4756 static long ext4_zero_range(struct file *file, loff_t offset,
4757                             loff_t len, int mode)
4758 {
4759         struct inode *inode = file_inode(file);
4760         handle_t *handle = NULL;
4761         unsigned int max_blocks;
4762         loff_t new_size = 0;
4763         int ret = 0;
4764         int flags;
4765         int credits;
4766         int partial_begin, partial_end;
4767         loff_t start, end;
4768         ext4_lblk_t lblk;
4769         struct address_space *mapping = inode->i_mapping;
4770         unsigned int blkbits = inode->i_blkbits;
4771
4772         trace_ext4_zero_range(inode, offset, len, mode);
4773
4774         if (!S_ISREG(inode->i_mode))
4775                 return -EINVAL;
4776
4777         /* Call ext4_force_commit to flush all data in case of data=journal. */
4778         if (ext4_should_journal_data(inode)) {
4779                 ret = ext4_force_commit(inode->i_sb);
4780                 if (ret)
4781                         return ret;
4782         }
4783
4784         /*
4785          * Write out all dirty pages to avoid race conditions
4786          * Then release them.
4787          */
4788         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4789                 ret = filemap_write_and_wait_range(mapping, offset,
4790                                                    offset + len - 1);
4791                 if (ret)
4792                         return ret;
4793         }
4794
4795         /*
4796          * Round up offset. This is not fallocate, we neet to zero out
4797          * blocks, so convert interior block aligned part of the range to
4798          * unwritten and possibly manually zero out unaligned parts of the
4799          * range.
4800          */
4801         start = round_up(offset, 1 << blkbits);
4802         end = round_down((offset + len), 1 << blkbits);
4803
4804         if (start < offset || end > offset + len)
4805                 return -EINVAL;
4806         partial_begin = offset & ((1 << blkbits) - 1);
4807         partial_end = (offset + len) & ((1 << blkbits) - 1);
4808
4809         lblk = start >> blkbits;
4810         max_blocks = (end >> blkbits);
4811         if (max_blocks < lblk)
4812                 max_blocks = 0;
4813         else
4814                 max_blocks -= lblk;
4815
4816         mutex_lock(&inode->i_mutex);
4817
4818         /*
4819          * Indirect files do not support unwritten extnets
4820          */
4821         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4822                 ret = -EOPNOTSUPP;
4823                 goto out_mutex;
4824         }
4825
4826         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4827              offset + len > i_size_read(inode)) {
4828                 new_size = offset + len;
4829                 ret = inode_newsize_ok(inode, new_size);
4830                 if (ret)
4831                         goto out_mutex;
4832         }
4833
4834         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4835         if (mode & FALLOC_FL_KEEP_SIZE)
4836                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4837
4838         /* Preallocate the range including the unaligned edges */
4839         if (partial_begin || partial_end) {
4840                 ret = ext4_alloc_file_blocks(file,
4841                                 round_down(offset, 1 << blkbits) >> blkbits,
4842                                 (round_up((offset + len), 1 << blkbits) -
4843                                  round_down(offset, 1 << blkbits)) >> blkbits,
4844                                 new_size, flags, mode);
4845                 if (ret)
4846                         goto out_mutex;
4847
4848         }
4849
4850         /* Zero range excluding the unaligned edges */
4851         if (max_blocks > 0) {
4852                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4853                           EXT4_EX_NOCACHE);
4854
4855                 /* Now release the pages and zero block aligned part of pages*/
4856                 truncate_pagecache_range(inode, start, end - 1);
4857                 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4858
4859                 /* Wait all existing dio workers, newcomers will block on i_mutex */
4860                 ext4_inode_block_unlocked_dio(inode);
4861                 inode_dio_wait(inode);
4862
4863                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4864                                              flags, mode);
4865                 if (ret)
4866                         goto out_dio;
4867         }
4868         if (!partial_begin && !partial_end)
4869                 goto out_dio;
4870
4871         /*
4872          * In worst case we have to writeout two nonadjacent unwritten
4873          * blocks and update the inode
4874          */
4875         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4876         if (ext4_should_journal_data(inode))
4877                 credits += 2;
4878         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4879         if (IS_ERR(handle)) {
4880                 ret = PTR_ERR(handle);
4881                 ext4_std_error(inode->i_sb, ret);
4882                 goto out_dio;
4883         }
4884
4885         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4886         if (new_size) {
4887                 ext4_update_inode_size(inode, new_size);
4888         } else {
4889                 /*
4890                 * Mark that we allocate beyond EOF so the subsequent truncate
4891                 * can proceed even if the new size is the same as i_size.
4892                 */
4893                 if ((offset + len) > i_size_read(inode))
4894                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4895         }
4896         ext4_mark_inode_dirty(handle, inode);
4897
4898         /* Zero out partial block at the edges of the range */
4899         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4900
4901         if (file->f_flags & O_SYNC)
4902                 ext4_handle_sync(handle);
4903
4904         ext4_journal_stop(handle);
4905 out_dio:
4906         ext4_inode_resume_unlocked_dio(inode);
4907 out_mutex:
4908         mutex_unlock(&inode->i_mutex);
4909         return ret;
4910 }
4911
4912 /*
4913  * preallocate space for a file. This implements ext4's fallocate file
4914  * operation, which gets called from sys_fallocate system call.
4915  * For block-mapped files, posix_fallocate should fall back to the method
4916  * of writing zeroes to the required new blocks (the same behavior which is
4917  * expected for file systems which do not support fallocate() system call).
4918  */
4919 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4920 {
4921         struct inode *inode = file_inode(file);
4922         loff_t new_size = 0;
4923         unsigned int max_blocks;
4924         int ret = 0;
4925         int flags;
4926         ext4_lblk_t lblk;
4927         unsigned int blkbits = inode->i_blkbits;
4928
4929         /*
4930          * Encrypted inodes can't handle collapse range or insert
4931          * range since we would need to re-encrypt blocks with a
4932          * different IV or XTS tweak (which are based on the logical
4933          * block number).
4934          *
4935          * XXX It's not clear why zero range isn't working, but we'll
4936          * leave it disabled for encrypted inodes for now.  This is a
4937          * bug we should fix....
4938          */
4939         if (ext4_encrypted_inode(inode) &&
4940             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4941                      FALLOC_FL_ZERO_RANGE)))
4942                 return -EOPNOTSUPP;
4943
4944         /* Return error if mode is not supported */
4945         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4946                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4947                      FALLOC_FL_INSERT_RANGE))
4948                 return -EOPNOTSUPP;
4949
4950         if (mode & FALLOC_FL_PUNCH_HOLE)
4951                 return ext4_punch_hole(inode, offset, len);
4952
4953         ret = ext4_convert_inline_data(inode);
4954         if (ret)
4955                 return ret;
4956
4957         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4958                 return ext4_collapse_range(inode, offset, len);
4959
4960         if (mode & FALLOC_FL_INSERT_RANGE)
4961                 return ext4_insert_range(inode, offset, len);
4962
4963         if (mode & FALLOC_FL_ZERO_RANGE)
4964                 return ext4_zero_range(file, offset, len, mode);
4965
4966         trace_ext4_fallocate_enter(inode, offset, len, mode);
4967         lblk = offset >> blkbits;
4968         /*
4969          * We can't just convert len to max_blocks because
4970          * If blocksize = 4096 offset = 3072 and len = 2048
4971          */
4972         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4973                 - lblk;
4974
4975         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4976         if (mode & FALLOC_FL_KEEP_SIZE)
4977                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4978
4979         mutex_lock(&inode->i_mutex);
4980
4981         /*
4982          * We only support preallocation for extent-based files only
4983          */
4984         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4985                 ret = -EOPNOTSUPP;
4986                 goto out;
4987         }
4988
4989         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4990              offset + len > i_size_read(inode)) {
4991                 new_size = offset + len;
4992                 ret = inode_newsize_ok(inode, new_size);
4993                 if (ret)
4994                         goto out;
4995         }
4996
4997         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4998                                      flags, mode);
4999         if (ret)
5000                 goto out;
5001
5002         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
5003                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
5004                                                 EXT4_I(inode)->i_sync_tid);
5005         }
5006 out:
5007         mutex_unlock(&inode->i_mutex);
5008         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
5009         return ret;
5010 }
5011
5012 /*
5013  * This function convert a range of blocks to written extents
5014  * The caller of this function will pass the start offset and the size.
5015  * all unwritten extents within this range will be converted to
5016  * written extents.
5017  *
5018  * This function is called from the direct IO end io call back
5019  * function, to convert the fallocated extents after IO is completed.
5020  * Returns 0 on success.
5021  */
5022 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
5023                                    loff_t offset, ssize_t len)
5024 {
5025         unsigned int max_blocks;
5026         int ret = 0;
5027         int ret2 = 0;
5028         struct ext4_map_blocks map;
5029         unsigned int credits, blkbits = inode->i_blkbits;
5030
5031         map.m_lblk = offset >> blkbits;
5032         /*
5033          * We can't just convert len to max_blocks because
5034          * If blocksize = 4096 offset = 3072 and len = 2048
5035          */
5036         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
5037                       map.m_lblk);
5038         /*
5039          * This is somewhat ugly but the idea is clear: When transaction is
5040          * reserved, everything goes into it. Otherwise we rather start several
5041          * smaller transactions for conversion of each extent separately.
5042          */
5043         if (handle) {
5044                 handle = ext4_journal_start_reserved(handle,
5045                                                      EXT4_HT_EXT_CONVERT);
5046                 if (IS_ERR(handle))
5047                         return PTR_ERR(handle);
5048                 credits = 0;
5049         } else {
5050                 /*
5051                  * credits to insert 1 extent into extent tree
5052                  */
5053                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5054         }
5055         while (ret >= 0 && ret < max_blocks) {
5056                 map.m_lblk += ret;
5057                 map.m_len = (max_blocks -= ret);
5058                 if (credits) {
5059                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5060                                                     credits);
5061                         if (IS_ERR(handle)) {
5062                                 ret = PTR_ERR(handle);
5063                                 break;
5064                         }
5065                 }
5066                 ret = ext4_map_blocks(handle, inode, &map,
5067                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5068                 if (ret <= 0)
5069                         ext4_warning(inode->i_sb,
5070                                      "inode #%lu: block %u: len %u: "
5071                                      "ext4_ext_map_blocks returned %d",
5072                                      inode->i_ino, map.m_lblk,
5073                                      map.m_len, ret);
5074                 ext4_mark_inode_dirty(handle, inode);
5075                 if (credits)
5076                         ret2 = ext4_journal_stop(handle);
5077                 if (ret <= 0 || ret2)
5078                         break;
5079         }
5080         if (!credits)
5081                 ret2 = ext4_journal_stop(handle);
5082         return ret > 0 ? ret2 : ret;
5083 }
5084
5085 /*
5086  * If newes is not existing extent (newes->ec_pblk equals zero) find
5087  * delayed extent at start of newes and update newes accordingly and
5088  * return start of the next delayed extent.
5089  *
5090  * If newes is existing extent (newes->ec_pblk is not equal zero)
5091  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5092  * extent found. Leave newes unmodified.
5093  */
5094 static int ext4_find_delayed_extent(struct inode *inode,
5095                                     struct extent_status *newes)
5096 {
5097         struct extent_status es;
5098         ext4_lblk_t block, next_del;
5099
5100         if (newes->es_pblk == 0) {
5101                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5102                                 newes->es_lblk + newes->es_len - 1, &es);
5103
5104                 /*
5105                  * No extent in extent-tree contains block @newes->es_pblk,
5106                  * then the block may stay in 1)a hole or 2)delayed-extent.
5107                  */
5108                 if (es.es_len == 0)
5109                         /* A hole found. */
5110                         return 0;
5111
5112                 if (es.es_lblk > newes->es_lblk) {
5113                         /* A hole found. */
5114                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5115                                             newes->es_len);
5116                         return 0;
5117                 }
5118
5119                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5120         }
5121
5122         block = newes->es_lblk + newes->es_len;
5123         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5124         if (es.es_len == 0)
5125                 next_del = EXT_MAX_BLOCKS;
5126         else
5127                 next_del = es.es_lblk;
5128
5129         return next_del;
5130 }
5131 /* fiemap flags we can handle specified here */
5132 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5133
5134 static int ext4_xattr_fiemap(struct inode *inode,
5135                                 struct fiemap_extent_info *fieinfo)
5136 {
5137         __u64 physical = 0;
5138         __u64 length;
5139         __u32 flags = FIEMAP_EXTENT_LAST;
5140         int blockbits = inode->i_sb->s_blocksize_bits;
5141         int error = 0;
5142
5143         /* in-inode? */
5144         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5145                 struct ext4_iloc iloc;
5146                 int offset;     /* offset of xattr in inode */
5147
5148                 error = ext4_get_inode_loc(inode, &iloc);
5149                 if (error)
5150                         return error;
5151                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5152                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5153                                 EXT4_I(inode)->i_extra_isize;
5154                 physical += offset;
5155                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5156                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5157                 brelse(iloc.bh);
5158         } else { /* external block */
5159                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5160                 length = inode->i_sb->s_blocksize;
5161         }
5162
5163         if (physical)
5164                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5165                                                 length, flags);
5166         return (error < 0 ? error : 0);
5167 }
5168
5169 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5170                 __u64 start, __u64 len)
5171 {
5172         ext4_lblk_t start_blk;
5173         int error = 0;
5174
5175         if (ext4_has_inline_data(inode)) {
5176                 int has_inline = 1;
5177
5178                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5179                                                 start, len);
5180
5181                 if (has_inline)
5182                         return error;
5183         }
5184
5185         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5186                 error = ext4_ext_precache(inode);
5187                 if (error)
5188                         return error;
5189         }
5190
5191         /* fallback to generic here if not in extents fmt */
5192         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5193                 return generic_block_fiemap(inode, fieinfo, start, len,
5194                         ext4_get_block);
5195
5196         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5197                 return -EBADR;
5198
5199         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5200                 error = ext4_xattr_fiemap(inode, fieinfo);
5201         } else {
5202                 ext4_lblk_t len_blks;
5203                 __u64 last_blk;
5204
5205                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5206                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5207                 if (last_blk >= EXT_MAX_BLOCKS)
5208                         last_blk = EXT_MAX_BLOCKS-1;
5209                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5210
5211                 /*
5212                  * Walk the extent tree gathering extent information
5213                  * and pushing extents back to the user.
5214                  */
5215                 error = ext4_fill_fiemap_extents(inode, start_blk,
5216                                                  len_blks, fieinfo);
5217         }
5218         return error;
5219 }
5220
5221 /*
5222  * ext4_access_path:
5223  * Function to access the path buffer for marking it dirty.
5224  * It also checks if there are sufficient credits left in the journal handle
5225  * to update path.
5226  */
5227 static int
5228 ext4_access_path(handle_t *handle, struct inode *inode,
5229                 struct ext4_ext_path *path)
5230 {
5231         int credits, err;
5232
5233         if (!ext4_handle_valid(handle))
5234                 return 0;
5235
5236         /*
5237          * Check if need to extend journal credits
5238          * 3 for leaf, sb, and inode plus 2 (bmap and group
5239          * descriptor) for each block group; assume two block
5240          * groups
5241          */
5242         if (handle->h_buffer_credits < 7) {
5243                 credits = ext4_writepage_trans_blocks(inode);
5244                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5245                 /* EAGAIN is success */
5246                 if (err && err != -EAGAIN)
5247                         return err;
5248         }
5249
5250         err = ext4_ext_get_access(handle, inode, path);
5251         return err;
5252 }
5253
5254 /*
5255  * ext4_ext_shift_path_extents:
5256  * Shift the extents of a path structure lying between path[depth].p_ext
5257  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5258  * if it is right shift or left shift operation.
5259  */
5260 static int
5261 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5262                             struct inode *inode, handle_t *handle,
5263                             enum SHIFT_DIRECTION SHIFT)
5264 {
5265         int depth, err = 0;
5266         struct ext4_extent *ex_start, *ex_last;
5267         bool update = 0;
5268         depth = path->p_depth;
5269
5270         while (depth >= 0) {
5271                 if (depth == path->p_depth) {
5272                         ex_start = path[depth].p_ext;
5273                         if (!ex_start)
5274                                 return -EIO;
5275
5276                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5277
5278                         err = ext4_access_path(handle, inode, path + depth);
5279                         if (err)
5280                                 goto out;
5281
5282                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5283                                 update = 1;
5284
5285                         while (ex_start <= ex_last) {
5286                                 if (SHIFT == SHIFT_LEFT) {
5287                                         le32_add_cpu(&ex_start->ee_block,
5288                                                 -shift);
5289                                         /* Try to merge to the left. */
5290                                         if ((ex_start >
5291                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5292                                             &&
5293                                             ext4_ext_try_to_merge_right(inode,
5294                                             path, ex_start - 1))
5295                                                 ex_last--;
5296                                         else
5297                                                 ex_start++;
5298                                 } else {
5299                                         le32_add_cpu(&ex_last->ee_block, shift);
5300                                         ext4_ext_try_to_merge_right(inode, path,
5301                                                 ex_last);
5302                                         ex_last--;
5303                                 }
5304                         }
5305                         err = ext4_ext_dirty(handle, inode, path + depth);
5306                         if (err)
5307                                 goto out;
5308
5309                         if (--depth < 0 || !update)
5310                                 break;
5311                 }
5312
5313                 /* Update index too */
5314                 err = ext4_access_path(handle, inode, path + depth);
5315                 if (err)
5316                         goto out;
5317
5318                 if (SHIFT == SHIFT_LEFT)
5319                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5320                 else
5321                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5322                 err = ext4_ext_dirty(handle, inode, path + depth);
5323                 if (err)
5324                         goto out;
5325
5326                 /* we are done if current index is not a starting index */
5327                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5328                         break;
5329
5330                 depth--;
5331         }
5332
5333 out:
5334         return err;
5335 }
5336
5337 /*
5338  * ext4_ext_shift_extents:
5339  * All the extents which lies in the range from @start to the last allocated
5340  * block for the @inode are shifted either towards left or right (depending
5341  * upon @SHIFT) by @shift blocks.
5342  * On success, 0 is returned, error otherwise.
5343  */
5344 static int
5345 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5346                        ext4_lblk_t start, ext4_lblk_t shift,
5347                        enum SHIFT_DIRECTION SHIFT)
5348 {
5349         struct ext4_ext_path *path;
5350         int ret = 0, depth;
5351         struct ext4_extent *extent;
5352         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5353
5354         /* Let path point to the last extent */
5355         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 0);
5356         if (IS_ERR(path))
5357                 return PTR_ERR(path);
5358
5359         depth = path->p_depth;
5360         extent = path[depth].p_ext;
5361         if (!extent)
5362                 goto out;
5363
5364         stop = le32_to_cpu(extent->ee_block) +
5365                         ext4_ext_get_actual_len(extent);
5366
5367        /*
5368          * In case of left shift, Don't start shifting extents until we make
5369          * sure the hole is big enough to accommodate the shift.
5370         */
5371         if (SHIFT == SHIFT_LEFT) {
5372                 path = ext4_find_extent(inode, start - 1, &path, 0);
5373                 if (IS_ERR(path))
5374                         return PTR_ERR(path);
5375                 depth = path->p_depth;
5376                 extent =  path[depth].p_ext;
5377                 if (extent) {
5378                         ex_start = le32_to_cpu(extent->ee_block);
5379                         ex_end = le32_to_cpu(extent->ee_block) +
5380                                 ext4_ext_get_actual_len(extent);
5381                 } else {
5382                         ex_start = 0;
5383                         ex_end = 0;
5384                 }
5385
5386                 if ((start == ex_start && shift > ex_start) ||
5387                     (shift > start - ex_end)) {
5388                         ext4_ext_drop_refs(path);
5389                         kfree(path);
5390                         return -EINVAL;
5391                 }
5392         }
5393
5394         /*
5395          * In case of left shift, iterator points to start and it is increased
5396          * till we reach stop. In case of right shift, iterator points to stop
5397          * and it is decreased till we reach start.
5398          */
5399         if (SHIFT == SHIFT_LEFT)
5400                 iterator = &start;
5401         else
5402                 iterator = &stop;
5403
5404         /* Its safe to start updating extents */
5405         while (start < stop) {
5406                 path = ext4_find_extent(inode, *iterator, &path, 0);
5407                 if (IS_ERR(path))
5408                         return PTR_ERR(path);
5409                 depth = path->p_depth;
5410                 extent = path[depth].p_ext;
5411                 if (!extent) {
5412                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5413                                          (unsigned long) *iterator);
5414                         return -EIO;
5415                 }
5416                 if (SHIFT == SHIFT_LEFT && *iterator >
5417                     le32_to_cpu(extent->ee_block)) {
5418                         /* Hole, move to the next extent */
5419                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5420                                 path[depth].p_ext++;
5421                         } else {
5422                                 *iterator = ext4_ext_next_allocated_block(path);
5423                                 continue;
5424                         }
5425                 }
5426
5427                 if (SHIFT == SHIFT_LEFT) {
5428                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5429                         *iterator = le32_to_cpu(extent->ee_block) +
5430                                         ext4_ext_get_actual_len(extent);
5431                 } else {
5432                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5433                         *iterator =  le32_to_cpu(extent->ee_block) > 0 ?
5434                                 le32_to_cpu(extent->ee_block) - 1 : 0;
5435                         /* Update path extent in case we need to stop */
5436                         while (le32_to_cpu(extent->ee_block) < start)
5437                                 extent++;
5438                         path[depth].p_ext = extent;
5439                 }
5440                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5441                                 handle, SHIFT);
5442                 if (ret)
5443                         break;
5444         }
5445 out:
5446         ext4_ext_drop_refs(path);
5447         kfree(path);
5448         return ret;
5449 }
5450
5451 /*
5452  * ext4_collapse_range:
5453  * This implements the fallocate's collapse range functionality for ext4
5454  * Returns: 0 and non-zero on error.
5455  */
5456 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5457 {
5458         struct super_block *sb = inode->i_sb;
5459         ext4_lblk_t punch_start, punch_stop;
5460         handle_t *handle;
5461         unsigned int credits;
5462         loff_t new_size, ioffset;
5463         int ret;
5464
5465         /*
5466          * We need to test this early because xfstests assumes that a
5467          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5468          * system does not support collapse range.
5469          */
5470         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5471                 return -EOPNOTSUPP;
5472
5473         /* Collapse range works only on fs block size aligned offsets. */
5474         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5475             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5476                 return -EINVAL;
5477
5478         if (!S_ISREG(inode->i_mode))
5479                 return -EINVAL;
5480
5481         trace_ext4_collapse_range(inode, offset, len);
5482
5483         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5484         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5485
5486         /* Call ext4_force_commit to flush all data in case of data=journal. */
5487         if (ext4_should_journal_data(inode)) {
5488                 ret = ext4_force_commit(inode->i_sb);
5489                 if (ret)
5490                         return ret;
5491         }
5492
5493         /*
5494          * Need to round down offset to be aligned with page size boundary
5495          * for page size > block size.
5496          */
5497         ioffset = round_down(offset, PAGE_SIZE);
5498
5499         /* Write out all dirty pages */
5500         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5501                                            LLONG_MAX);
5502         if (ret)
5503                 return ret;
5504
5505         /* Take mutex lock */
5506         mutex_lock(&inode->i_mutex);
5507
5508         /*
5509          * There is no need to overlap collapse range with EOF, in which case
5510          * it is effectively a truncate operation
5511          */
5512         if (offset + len >= i_size_read(inode)) {
5513                 ret = -EINVAL;
5514                 goto out_mutex;
5515         }
5516
5517         /* Currently just for extent based files */
5518         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5519                 ret = -EOPNOTSUPP;
5520                 goto out_mutex;
5521         }
5522
5523         truncate_pagecache(inode, ioffset);
5524
5525         /* Wait for existing dio to complete */
5526         ext4_inode_block_unlocked_dio(inode);
5527         inode_dio_wait(inode);
5528
5529         credits = ext4_writepage_trans_blocks(inode);
5530         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5531         if (IS_ERR(handle)) {
5532                 ret = PTR_ERR(handle);
5533                 goto out_dio;
5534         }
5535
5536         down_write(&EXT4_I(inode)->i_data_sem);
5537         ext4_discard_preallocations(inode);
5538
5539         ret = ext4_es_remove_extent(inode, punch_start,
5540                                     EXT_MAX_BLOCKS - punch_start);
5541         if (ret) {
5542                 up_write(&EXT4_I(inode)->i_data_sem);
5543                 goto out_stop;
5544         }
5545
5546         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5547         if (ret) {
5548                 up_write(&EXT4_I(inode)->i_data_sem);
5549                 goto out_stop;
5550         }
5551         ext4_discard_preallocations(inode);
5552
5553         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5554                                      punch_stop - punch_start, SHIFT_LEFT);
5555         if (ret) {
5556                 up_write(&EXT4_I(inode)->i_data_sem);
5557                 goto out_stop;
5558         }
5559
5560         new_size = i_size_read(inode) - len;
5561         i_size_write(inode, new_size);
5562         EXT4_I(inode)->i_disksize = new_size;
5563
5564         up_write(&EXT4_I(inode)->i_data_sem);
5565         if (IS_SYNC(inode))
5566                 ext4_handle_sync(handle);
5567         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5568         ext4_mark_inode_dirty(handle, inode);
5569
5570 out_stop:
5571         ext4_journal_stop(handle);
5572 out_dio:
5573         ext4_inode_resume_unlocked_dio(inode);
5574 out_mutex:
5575         mutex_unlock(&inode->i_mutex);
5576         return ret;
5577 }
5578
5579 /*
5580  * ext4_insert_range:
5581  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5582  * The data blocks starting from @offset to the EOF are shifted by @len
5583  * towards right to create a hole in the @inode. Inode size is increased
5584  * by len bytes.
5585  * Returns 0 on success, error otherwise.
5586  */
5587 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5588 {
5589         struct super_block *sb = inode->i_sb;
5590         handle_t *handle;
5591         struct ext4_ext_path *path;
5592         struct ext4_extent *extent;
5593         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5594         unsigned int credits, ee_len;
5595         int ret = 0, depth, split_flag = 0;
5596         loff_t ioffset;
5597
5598         /*
5599          * We need to test this early because xfstests assumes that an
5600          * insert range of (0, 1) will return EOPNOTSUPP if the file
5601          * system does not support insert range.
5602          */
5603         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5604                 return -EOPNOTSUPP;
5605
5606         /* Insert range works only on fs block size aligned offsets. */
5607         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5608                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5609                 return -EINVAL;
5610
5611         if (!S_ISREG(inode->i_mode))
5612                 return -EOPNOTSUPP;
5613
5614         trace_ext4_insert_range(inode, offset, len);
5615
5616         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5617         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5618
5619         /* Call ext4_force_commit to flush all data in case of data=journal */
5620         if (ext4_should_journal_data(inode)) {
5621                 ret = ext4_force_commit(inode->i_sb);
5622                 if (ret)
5623                         return ret;
5624         }
5625
5626         /*
5627          * Need to round down to align start offset to page size boundary
5628          * for page size > block size.
5629          */
5630         ioffset = round_down(offset, PAGE_SIZE);
5631
5632         /* Write out all dirty pages */
5633         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5634                         LLONG_MAX);
5635         if (ret)
5636                 return ret;
5637
5638         /* Take mutex lock */
5639         mutex_lock(&inode->i_mutex);
5640
5641         /* Currently just for extent based files */
5642         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5643                 ret = -EOPNOTSUPP;
5644                 goto out_mutex;
5645         }
5646
5647         /* Check for wrap through zero */
5648         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5649                 ret = -EFBIG;
5650                 goto out_mutex;
5651         }
5652
5653         /* Offset should be less than i_size */
5654         if (offset >= i_size_read(inode)) {
5655                 ret = -EINVAL;
5656                 goto out_mutex;
5657         }
5658
5659         truncate_pagecache(inode, ioffset);
5660
5661         /* Wait for existing dio to complete */
5662         ext4_inode_block_unlocked_dio(inode);
5663         inode_dio_wait(inode);
5664
5665         credits = ext4_writepage_trans_blocks(inode);
5666         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5667         if (IS_ERR(handle)) {
5668                 ret = PTR_ERR(handle);
5669                 goto out_dio;
5670         }
5671
5672         /* Expand file to avoid data loss if there is error while shifting */
5673         inode->i_size += len;
5674         EXT4_I(inode)->i_disksize += len;
5675         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5676         ret = ext4_mark_inode_dirty(handle, inode);
5677         if (ret)
5678                 goto out_stop;
5679
5680         down_write(&EXT4_I(inode)->i_data_sem);
5681         ext4_discard_preallocations(inode);
5682
5683         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5684         if (IS_ERR(path)) {
5685                 up_write(&EXT4_I(inode)->i_data_sem);
5686                 goto out_stop;
5687         }
5688
5689         depth = ext_depth(inode);
5690         extent = path[depth].p_ext;
5691         if (extent) {
5692                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5693                 ee_len = ext4_ext_get_actual_len(extent);
5694
5695                 /*
5696                  * If offset_lblk is not the starting block of extent, split
5697                  * the extent @offset_lblk
5698                  */
5699                 if ((offset_lblk > ee_start_lblk) &&
5700                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5701                         if (ext4_ext_is_unwritten(extent))
5702                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5703                                         EXT4_EXT_MARK_UNWRIT2;
5704                         ret = ext4_split_extent_at(handle, inode, &path,
5705                                         offset_lblk, split_flag,
5706                                         EXT4_EX_NOCACHE |
5707                                         EXT4_GET_BLOCKS_PRE_IO |
5708                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5709                 }
5710
5711                 ext4_ext_drop_refs(path);
5712                 kfree(path);
5713                 if (ret < 0) {
5714                         up_write(&EXT4_I(inode)->i_data_sem);
5715                         goto out_stop;
5716                 }
5717         }
5718
5719         ret = ext4_es_remove_extent(inode, offset_lblk,
5720                         EXT_MAX_BLOCKS - offset_lblk);
5721         if (ret) {
5722                 up_write(&EXT4_I(inode)->i_data_sem);
5723                 goto out_stop;
5724         }
5725
5726         /*
5727          * if offset_lblk lies in a hole which is at start of file, use
5728          * ee_start_lblk to shift extents
5729          */
5730         ret = ext4_ext_shift_extents(inode, handle,
5731                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5732                 len_lblk, SHIFT_RIGHT);
5733
5734         up_write(&EXT4_I(inode)->i_data_sem);
5735         if (IS_SYNC(inode))
5736                 ext4_handle_sync(handle);
5737
5738 out_stop:
5739         ext4_journal_stop(handle);
5740 out_dio:
5741         ext4_inode_resume_unlocked_dio(inode);
5742 out_mutex:
5743         mutex_unlock(&inode->i_mutex);
5744         return ret;
5745 }
5746
5747 /**
5748  * ext4_swap_extents - Swap extents between two inodes
5749  *
5750  * @inode1:     First inode
5751  * @inode2:     Second inode
5752  * @lblk1:      Start block for first inode
5753  * @lblk2:      Start block for second inode
5754  * @count:      Number of blocks to swap
5755  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5756  * @erp:        Pointer to save error value
5757  *
5758  * This helper routine does exactly what is promise "swap extents". All other
5759  * stuff such as page-cache locking consistency, bh mapping consistency or
5760  * extent's data copying must be performed by caller.
5761  * Locking:
5762  *              i_mutex is held for both inodes
5763  *              i_data_sem is locked for write for both inodes
5764  * Assumptions:
5765  *              All pages from requested range are locked for both inodes
5766  */
5767 int
5768 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5769                      struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5770                   ext4_lblk_t count, int unwritten, int *erp)
5771 {
5772         struct ext4_ext_path *path1 = NULL;
5773         struct ext4_ext_path *path2 = NULL;
5774         int replaced_count = 0;
5775
5776         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5777         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5778         BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5779         BUG_ON(!mutex_is_locked(&inode2->i_mutex));
5780
5781         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5782         if (unlikely(*erp))
5783                 return 0;
5784         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5785         if (unlikely(*erp))
5786                 return 0;
5787
5788         while (count) {
5789                 struct ext4_extent *ex1, *ex2, tmp_ex;
5790                 ext4_lblk_t e1_blk, e2_blk;
5791                 int e1_len, e2_len, len;
5792                 int split = 0;
5793
5794                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5795                 if (unlikely(IS_ERR(path1))) {
5796                         *erp = PTR_ERR(path1);
5797                         path1 = NULL;
5798                 finish:
5799                         count = 0;
5800                         goto repeat;
5801                 }
5802                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5803                 if (unlikely(IS_ERR(path2))) {
5804                         *erp = PTR_ERR(path2);
5805                         path2 = NULL;
5806                         goto finish;
5807                 }
5808                 ex1 = path1[path1->p_depth].p_ext;
5809                 ex2 = path2[path2->p_depth].p_ext;
5810                 /* Do we have somthing to swap ? */
5811                 if (unlikely(!ex2 || !ex1))
5812                         goto finish;
5813
5814                 e1_blk = le32_to_cpu(ex1->ee_block);
5815                 e2_blk = le32_to_cpu(ex2->ee_block);
5816                 e1_len = ext4_ext_get_actual_len(ex1);
5817                 e2_len = ext4_ext_get_actual_len(ex2);
5818
5819                 /* Hole handling */
5820                 if (!in_range(lblk1, e1_blk, e1_len) ||
5821                     !in_range(lblk2, e2_blk, e2_len)) {
5822                         ext4_lblk_t next1, next2;
5823
5824                         /* if hole after extent, then go to next extent */
5825                         next1 = ext4_ext_next_allocated_block(path1);
5826                         next2 = ext4_ext_next_allocated_block(path2);
5827                         /* If hole before extent, then shift to that extent */
5828                         if (e1_blk > lblk1)
5829                                 next1 = e1_blk;
5830                         if (e2_blk > lblk2)
5831                                 next2 = e1_blk;
5832                         /* Do we have something to swap */
5833                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5834                                 goto finish;
5835                         /* Move to the rightest boundary */
5836                         len = next1 - lblk1;
5837                         if (len < next2 - lblk2)
5838                                 len = next2 - lblk2;
5839                         if (len > count)
5840                                 len = count;
5841                         lblk1 += len;
5842                         lblk2 += len;
5843                         count -= len;
5844                         goto repeat;
5845                 }
5846
5847                 /* Prepare left boundary */
5848                 if (e1_blk < lblk1) {
5849                         split = 1;
5850                         *erp = ext4_force_split_extent_at(handle, inode1,
5851                                                 &path1, lblk1, 0);
5852                         if (unlikely(*erp))
5853                                 goto finish;
5854                 }
5855                 if (e2_blk < lblk2) {
5856                         split = 1;
5857                         *erp = ext4_force_split_extent_at(handle, inode2,
5858                                                 &path2,  lblk2, 0);
5859                         if (unlikely(*erp))
5860                                 goto finish;
5861                 }
5862                 /* ext4_split_extent_at() may result in leaf extent split,
5863                  * path must to be revalidated. */
5864                 if (split)
5865                         goto repeat;
5866
5867                 /* Prepare right boundary */
5868                 len = count;
5869                 if (len > e1_blk + e1_len - lblk1)
5870                         len = e1_blk + e1_len - lblk1;
5871                 if (len > e2_blk + e2_len - lblk2)
5872                         len = e2_blk + e2_len - lblk2;
5873
5874                 if (len != e1_len) {
5875                         split = 1;
5876                         *erp = ext4_force_split_extent_at(handle, inode1,
5877                                                 &path1, lblk1 + len, 0);
5878                         if (unlikely(*erp))
5879                                 goto finish;
5880                 }
5881                 if (len != e2_len) {
5882                         split = 1;
5883                         *erp = ext4_force_split_extent_at(handle, inode2,
5884                                                 &path2, lblk2 + len, 0);
5885                         if (*erp)
5886                                 goto finish;
5887                 }
5888                 /* ext4_split_extent_at() may result in leaf extent split,
5889                  * path must to be revalidated. */
5890                 if (split)
5891                         goto repeat;
5892
5893                 BUG_ON(e2_len != e1_len);
5894                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5895                 if (unlikely(*erp))
5896                         goto finish;
5897                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5898                 if (unlikely(*erp))
5899                         goto finish;
5900
5901                 /* Both extents are fully inside boundaries. Swap it now */
5902                 tmp_ex = *ex1;
5903                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5904                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5905                 ex1->ee_len = cpu_to_le16(e2_len);
5906                 ex2->ee_len = cpu_to_le16(e1_len);
5907                 if (unwritten)
5908                         ext4_ext_mark_unwritten(ex2);
5909                 if (ext4_ext_is_unwritten(&tmp_ex))
5910                         ext4_ext_mark_unwritten(ex1);
5911
5912                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5913                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5914                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5915                                       path2->p_depth);
5916                 if (unlikely(*erp))
5917                         goto finish;
5918                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5919                                       path1->p_depth);
5920                 /*
5921                  * Looks scarry ah..? second inode already points to new blocks,
5922                  * and it was successfully dirtied. But luckily error may happen
5923                  * only due to journal error, so full transaction will be
5924                  * aborted anyway.
5925                  */
5926                 if (unlikely(*erp))
5927                         goto finish;
5928                 lblk1 += len;
5929                 lblk2 += len;
5930                 replaced_count += len;
5931                 count -= len;
5932
5933         repeat:
5934                 ext4_ext_drop_refs(path1);
5935                 kfree(path1);
5936                 ext4_ext_drop_refs(path2);
5937                 kfree(path2);
5938                 path1 = path2 = NULL;
5939         }
5940         return replaced_count;
5941 }