fs/ext4/dir.c

   1 /*
   2  *  linux/fs/ext4/dir.c
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  from
  10  *
  11  *  linux/fs/minix/dir.c
  12  *
  13  *  Copyright (C) 1991, 1992  Linus Torvalds
  14  *
  15  *  ext4 directory handling functions
  16  *
  17  *  Big-endian to little-endian byte-swapping/bitmaps by
  18  *        David S. Miller (davem@caip.rutgers.edu), 1995
  19  *
  20  * Hash Tree Directory indexing (c) 2001  Daniel Phillips
  21  *
  22  */
  23
  24 #include <linux/fs.h>
  25 #include <linux/jbd2.h>
  26 #include <linux/buffer_head.h>
  27 #include <linux/slab.h>
  28 #include <linux/rbtree.h>
  29 #include "ext4.h"
  30 #include "xattr.h"
  31
  32 static int ext4_dx_readdir(struct file *filp,
  33                            void *dirent, filldir_t filldir);
  34
  35 /**
  36  * Check if the given dir-inode refers to an htree-indexed directory
  37  * (or a directory which chould potentially get coverted to use htree
  38  * indexing).
  39  *
  40  * Return 1 if it is a dx dir, 0 if not
  41  */
  42 static int is_dx_dir(struct inode *inode)
  43 {
  44         struct super_block *sb = inode->i_sb;
  45
  46         if (EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
  47                      EXT4_FEATURE_COMPAT_DIR_INDEX) &&
  48             ((ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) ||
  49              ((inode->i_size >> sb->s_blocksize_bits) == 1)))
  50                 return 1;
  51
  52         return 0;
  53 }
  54
  55 /*
  56  * Return 0 if the directory entry is OK, and 1 if there is a problem
  57  *
  58  * Note: this is the opposite of what ext2 and ext3 historically returned...
  59  *
  60  * bh passed here can be an inode block or a dir data block, depending
  61  * on the inode inline data flag.
  62  */
  63 int __ext4_check_dir_entry(const char *function, unsigned int line,
  64                            struct inode *dir, struct file *filp,
  65                            struct ext4_dir_entry_2 *de,
  66                            struct buffer_head *bh, char *buf, int size,
  67                            unsigned int offset)
  68 {
  69         const char *error_msg = NULL;
  70         const int rlen = ext4_rec_len_from_disk(de->rec_len,
  71                                                 dir->i_sb->s_blocksize);
  72
  73         if (unlikely(rlen < EXT4_DIR_REC_LEN(1)))
  74                 error_msg = "rec_len is smaller than minimal";
  75         else if (unlikely(rlen % 4 != 0))
  76                 error_msg = "rec_len % 4 != 0";
  77         else if (unlikely(rlen < EXT4_DIR_REC_LEN(de->name_len)))
  78                 error_msg = "rec_len is too small for name_len";
  79         else if (unlikely(((char *) de - buf) + rlen > size))
  80                 error_msg = "directory entry across range";
  81         else if (unlikely(le32_to_cpu(de->inode) >
  82                         le32_to_cpu(EXT4_SB(dir->i_sb)->s_es->s_inodes_count)))
  83                 error_msg = "inode out of bounds";
  84         else
  85                 return 0;
  86
  87         if (filp)
  88                 ext4_error_file(filp, function, line, bh->b_blocknr,
  89                                 "bad entry in directory: %s - offset=%u(%u), "
  90                                 "inode=%u, rec_len=%d, name_len=%d",
  91                                 error_msg, (unsigned) (offset % size),
  92                                 offset, le32_to_cpu(de->inode),
  93                                 rlen, de->name_len);
  94         else
  95                 ext4_error_inode(dir, function, line, bh->b_blocknr,
  96                                 "bad entry in directory: %s - offset=%u(%u), "
  97                                 "inode=%u, rec_len=%d, name_len=%d",
  98                                 error_msg, (unsigned) (offset % size),
  99                                 offset, le32_to_cpu(de->inode),
 100                                 rlen, de->name_len);
 101
 102         return 1;
 103 }
 104
 105 static int ext4_readdir(struct file *filp,
 106                          void *dirent, filldir_t filldir)
 107 {
 108         int error = 0;
 109         unsigned int offset;
 110         int i, stored;
 111         struct ext4_dir_entry_2 *de;
 112         int err;
 113         struct inode *inode = filp->f_path.dentry->d_inode;
 114         struct super_block *sb = inode->i_sb;
 115         int ret = 0;
 116         int dir_has_error = 0;
 117
 118         if (ext4_has_inline_data(inode)) {
 119                 int has_inline_data = 1;
 120                 ret = ext4_read_inline_dir(filp, dirent, filldir,
 121                                            &has_inline_data);
 122                 if (has_inline_data)
 123                         return ret;
 124         }
 125
 126         if (is_dx_dir(inode)) {
 127                 err = ext4_dx_readdir(filp, dirent, filldir);
 128                 if (err != ERR_BAD_DX_DIR) {
 129                         ret = err;
 130                         goto out;
 131                 }
 132                 /*
 133                  * We don't set the inode dirty flag since it's not
 134                  * critical that it get flushed back to the disk.
 135                  */
 136                 ext4_clear_inode_flag(filp->f_path.dentry->d_inode,
 137                                       EXT4_INODE_INDEX);
 138         }
 139         stored = 0;
 140         offset = filp->f_pos & (sb->s_blocksize - 1);
 141
 142         while (!error && !stored && filp->f_pos < inode->i_size) {
 143                 struct ext4_map_blocks map;
 144                 struct buffer_head *bh = NULL;
 145
 146                 map.m_lblk = filp->f_pos >> EXT4_BLOCK_SIZE_BITS(sb);
 147                 map.m_len = 1;
 148                 err = ext4_map_blocks(NULL, inode, &map, 0);
 149                 if (err > 0) {
 150                         pgoff_t index = map.m_pblk >>
 151                                         (PAGE_CACHE_SHIFT - inode->i_blkbits);
 152                         if (!ra_has_index(&filp->f_ra, index))
 153                                 page_cache_sync_readahead(
 154                                         sb->s_bdev->bd_inode->i_mapping,
 155                                         &filp->f_ra, filp,
 156                                         index, 1);
 157                         filp->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 158                         bh = ext4_bread(NULL, inode, map.m_lblk, 0, &err);
 159                 }
 160
 161                 /*
 162                  * We ignore I/O errors on directories so users have a chance
 163                  * of recovering data when there's a bad sector
 164                  */
 165                 if (!bh) {
 166                         if (!dir_has_error) {
 167                                 EXT4_ERROR_FILE(filp, 0,
 168                                                 "directory contains a "
 169                                                 "hole at offset %llu",
 170                                            (unsigned long long) filp->f_pos);
 171                                 dir_has_error = 1;
 172                         }
 173                         /* corrupt size?  Maybe no more blocks to read */
 174                         if (filp->f_pos > inode->i_blocks << 9)
 175                                 break;
 176                         filp->f_pos += sb->s_blocksize - offset;
 177                         continue;
 178                 }
 179
 180                 /* Check the checksum */
 181                 if (!buffer_verified(bh) &&
 182                     !ext4_dirent_csum_verify(inode,
 183                                 (struct ext4_dir_entry *)bh->b_data)) {
 184                         EXT4_ERROR_FILE(filp, 0, "directory fails checksum "
 185                                         "at offset %llu",
 186                                         (unsigned long long)filp->f_pos);
 187                         filp->f_pos += sb->s_blocksize - offset;
 188                         continue;
 189                 }
 190                 set_buffer_verified(bh);
 191
 192 revalidate:
 193                 /* If the dir block has changed since the last call to
 194                  * readdir(2), then we might be pointing to an invalid
 195                  * dirent right now.  Scan from the start of the block
 196                  * to make sure. */
 197                 if (filp->f_version != inode->i_version) {
 198                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
 199                                 de = (struct ext4_dir_entry_2 *)
 200                                         (bh->b_data + i);
 201                                 /* It's too expensive to do a full
 202                                  * dirent test each time round this
 203                                  * loop, but we do have to test at
 204                                  * least that it is non-zero.  A
 205                                  * failure will be detected in the
 206                                  * dirent test below. */
 207                                 if (ext4_rec_len_from_disk(de->rec_len,
 208                                         sb->s_blocksize) < EXT4_DIR_REC_LEN(1))
 209                                         break;
 210                                 i += ext4_rec_len_from_disk(de->rec_len,
 211                                                             sb->s_blocksize);
 212                         }
 213                         offset = i;
 214                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
 215                                 | offset;
 216                         filp->f_version = inode->i_version;
 217                 }
 218
 219                 while (!error && filp->f_pos < inode->i_size
 220                        && offset < sb->s_blocksize) {
 221                         de = (struct ext4_dir_entry_2 *) (bh->b_data + offset);
 222                         if (ext4_check_dir_entry(inode, filp, de, bh,
 223                                                  bh->b_data, bh->b_size,
 224                                                  offset)) {
 225                                 /*
 226                                  * On error, skip the f_pos to the next block
 227                                  */
 228                                 filp->f_pos = (filp->f_pos |
 229                                                 (sb->s_blocksize - 1)) + 1;
 230                                 brelse(bh);
 231                                 ret = stored;
 232                                 goto out;
 233                         }
 234                         offset += ext4_rec_len_from_disk(de->rec_len,
 235                                         sb->s_blocksize);
 236                         if (le32_to_cpu(de->inode)) {
 237                                 /* We might block in the next section
 238                                  * if the data destination is
 239                                  * currently swapped out.  So, use a
 240                                  * version stamp to detect whether or
 241                                  * not the directory has been modified
 242                                  * during the copy operation.
 243                                  */
 244                                 u64 version = filp->f_version;
 245
 246                                 error = filldir(dirent, de->name,
 247                                                 de->name_len,
 248                                                 filp->f_pos,
 249                                                 le32_to_cpu(de->inode),
 250                                                 get_dtype(sb, de->file_type));
 251                                 if (error)
 252                                         break;
 253                                 if (version != filp->f_version)
 254                                         goto revalidate;
 255                                 stored++;
 256                         }
 257                         filp->f_pos += ext4_rec_len_from_disk(de->rec_len,
 258                                                 sb->s_blocksize);
 259                 }
 260                 offset = 0;
 261                 brelse(bh);
 262         }
 263 out:
 264         return ret;
 265 }
 266
 267 static inline int is_32bit_api(void)
 268 {
 269 #ifdef CONFIG_COMPAT
 270         return is_compat_task();
 271 #else
 272         return (BITS_PER_LONG == 32);
 273 #endif
 274 }
 275
 276 /*
 277  * These functions convert from the major/minor hash to an f_pos
 278  * value for dx directories
 279  *
 280  * Upper layer (for example NFS) should specify FMODE_32BITHASH or
 281  * FMODE_64BITHASH explicitly. On the other hand, we allow ext4 to be mounted
 282  * directly on both 32-bit and 64-bit nodes, under such case, neither
 283  * FMODE_32BITHASH nor FMODE_64BITHASH is specified.
 284  */
 285 static inline loff_t hash2pos(struct file *filp, __u32 major, __u32 minor)
 286 {
 287         if ((filp->f_mode & FMODE_32BITHASH) ||
 288             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
 289                 return major >> 1;
 290         else
 291                 return ((__u64)(major >> 1) << 32) | (__u64)minor;
 292 }
 293
 294 static inline __u32 pos2maj_hash(struct file *filp, loff_t pos)
 295 {
 296         if ((filp->f_mode & FMODE_32BITHASH) ||
 297             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
 298                 return (pos << 1) & 0xffffffff;
 299         else
 300                 return ((pos >> 32) << 1) & 0xffffffff;
 301 }
 302
 303 static inline __u32 pos2min_hash(struct file *filp, loff_t pos)
 304 {
 305         if ((filp->f_mode & FMODE_32BITHASH) ||
 306             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
 307                 return 0;
 308         else
 309                 return pos & 0xffffffff;
 310 }
 311
 312 /*
 313  * Return 32- or 64-bit end-of-file for dx directories
 314  */
 315 static inline loff_t ext4_get_htree_eof(struct file *filp)
 316 {
 317         if ((filp->f_mode & FMODE_32BITHASH) ||
 318             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
 319                 return EXT4_HTREE_EOF_32BIT;
 320         else
 321                 return EXT4_HTREE_EOF_64BIT;
 322 }
 323
 324
 325 /*
 326  * ext4_dir_llseek() calls generic_file_llseek_size to handle htree
 327  * directories, where the "offset" is in terms of the filename hash
 328  * value instead of the byte offset.
 329  *
 330  * Because we may return a 64-bit hash that is well beyond offset limits,
 331  * we need to pass the max hash as the maximum allowable offset in
 332  * the htree directory case.
 333  *
 334  * For non-htree, ext4_llseek already chooses the proper max offset.
 335  */
 336 loff_t ext4_dir_llseek(struct file *file, loff_t offset, int whence)
 337 {
 338         struct inode *inode = file->f_mapping->host;
 339         int dx_dir = is_dx_dir(inode);
 340         loff_t htree_max = ext4_get_htree_eof(file);
 341
 342         if (likely(dx_dir))
 343                 return generic_file_llseek_size(file, offset, whence,
 344                                                     htree_max, htree_max);
 345         else
 346                 return ext4_llseek(file, offset, whence);
 347 }
 348
 349 /*
 350  * This structure holds the nodes of the red-black tree used to store
 351  * the directory entry in hash order.
 352  */
 353 struct fname {
 354         __u32           hash;
 355         __u32           minor_hash;
 356         struct rb_node  rb_hash;
 357         struct fname    *next;
 358         __u32           inode;
 359         __u8            name_len;
 360         __u8            file_type;
 361         char            name[0];
 362 };
 363
 364 /*
 365  * This functoin implements a non-recursive way of freeing all of the
 366  * nodes in the red-black tree.
 367  */
 368 static void free_rb_tree_fname(struct rb_root *root)
 369 {
 370         struct rb_node  *n = root->rb_node;
 371         struct rb_node  *parent;
 372         struct fname    *fname;
 373
 374         while (n) {
 375                 /* Do the node's children first */
 376                 if (n->rb_left) {
 377                         n = n->rb_left;
 378                         continue;
 379                 }
 380                 if (n->rb_right) {
 381                         n = n->rb_right;
 382                         continue;
 383                 }
 384                 /*
 385                  * The node has no children; free it, and then zero
 386                  * out parent's link to it.  Finally go to the
 387                  * beginning of the loop and try to free the parent
 388                  * node.
 389                  */
 390                 parent = rb_parent(n);
 391                 fname = rb_entry(n, struct fname, rb_hash);
 392                 while (fname) {
 393                         struct fname *old = fname;
 394                         fname = fname->next;
 395                         kfree(old);
 396                 }
 397                 if (!parent)
 398                         *root = RB_ROOT;
 399                 else if (parent->rb_left == n)
 400                         parent->rb_left = NULL;
 401                 else if (parent->rb_right == n)
 402                         parent->rb_right = NULL;
 403                 n = parent;
 404         }
 405 }
 406
 407
 408 static struct dir_private_info *ext4_htree_create_dir_info(struct file *filp,
 409                                                            loff_t pos)
 410 {
 411         struct dir_private_info *p;
 412
 413         p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
 414         if (!p)
 415                 return NULL;
 416         p->curr_hash = pos2maj_hash(filp, pos);
 417         p->curr_minor_hash = pos2min_hash(filp, pos);
 418         return p;
 419 }
 420
 421 void ext4_htree_free_dir_info(struct dir_private_info *p)
 422 {
 423         free_rb_tree_fname(&p->root);
 424         kfree(p);
 425 }
 426
 427 /*
 428  * Given a directory entry, enter it into the fname rb tree.
 429  */
 430 int ext4_htree_store_dirent(struct file *dir_file, __u32 hash,
 431                              __u32 minor_hash,
 432                              struct ext4_dir_entry_2 *dirent)
 433 {
 434         struct rb_node **p, *parent = NULL;
 435         struct fname *fname, *new_fn;
 436         struct dir_private_info *info;
 437         int len;
 438
 439         info = dir_file->private_data;
 440         p = &info->root.rb_node;
 441
 442         /* Create and allocate the fname structure */
 443         len = sizeof(struct fname) + dirent->name_len + 1;
 444         new_fn = kzalloc(len, GFP_KERNEL);
 445         if (!new_fn)
 446                 return -ENOMEM;
 447         new_fn->hash = hash;
 448         new_fn->minor_hash = minor_hash;
 449         new_fn->inode = le32_to_cpu(dirent->inode);
 450         new_fn->name_len = dirent->name_len;
 451         new_fn->file_type = dirent->file_type;
 452         memcpy(new_fn->name, dirent->name, dirent->name_len);
 453         new_fn->name[dirent->name_len] = 0;
 454
 455         while (*p) {
 456                 parent = *p;
 457                 fname = rb_entry(parent, struct fname, rb_hash);
 458
 459                 /*
 460                  * If the hash and minor hash match up, then we put
 461                  * them on a linked list.  This rarely happens...
 462                  */
 463                 if ((new_fn->hash == fname->hash) &&
 464                     (new_fn->minor_hash == fname->minor_hash)) {
 465                         new_fn->next = fname->next;
 466                         fname->next = new_fn;
 467                         return 0;
 468                 }
 469
 470                 if (new_fn->hash < fname->hash)
 471                         p = &(*p)->rb_left;
 472                 else if (new_fn->hash > fname->hash)
 473                         p = &(*p)->rb_right;
 474                 else if (new_fn->minor_hash < fname->minor_hash)
 475                         p = &(*p)->rb_left;
 476                 else /* if (new_fn->minor_hash > fname->minor_hash) */
 477                         p = &(*p)->rb_right;
 478         }
 479
 480         rb_link_node(&new_fn->rb_hash, parent, p);
 481         rb_insert_color(&new_fn->rb_hash, &info->root);
 482         return 0;
 483 }
 484
 485
 486
 487 /*
 488  * This is a helper function for ext4_dx_readdir.  It calls filldir
 489  * for all entres on the fname linked list.  (Normally there is only
 490  * one entry on the linked list, unless there are 62 bit hash collisions.)
 491  */
 492 static int call_filldir(struct file *filp, void *dirent,
 493                         filldir_t filldir, struct fname *fname)
 494 {
 495         struct dir_private_info *info = filp->private_data;
 496         loff_t  curr_pos;
 497         struct inode *inode = filp->f_path.dentry->d_inode;
 498         struct super_block *sb;
 499         int error;
 500
 501         sb = inode->i_sb;
 502
 503         if (!fname) {
 504                 ext4_msg(sb, KERN_ERR, "%s:%d: inode #%lu: comm %s: "
 505                          "called with null fname?!?", __func__, __LINE__,
 506                          inode->i_ino, current->comm);
 507                 return 0;
 508         }
 509         curr_pos = hash2pos(filp, fname->hash, fname->minor_hash);
 510         while (fname) {
 511                 error = filldir(dirent, fname->name,
 512                                 fname->name_len, curr_pos,
 513                                 fname->inode,
 514                                 get_dtype(sb, fname->file_type));
 515                 if (error) {
 516                         filp->f_pos = curr_pos;
 517                         info->extra_fname = fname;
 518                         return error;
 519                 }
 520                 fname = fname->next;
 521         }
 522         return 0;
 523 }
 524
 525 static int ext4_dx_readdir(struct file *filp,
 526                          void *dirent, filldir_t filldir)
 527 {
 528         struct dir_private_info *info = filp->private_data;
 529         struct inode *inode = filp->f_path.dentry->d_inode;
 530         struct fname *fname;
 531         int     ret;
 532
 533         if (!info) {
 534                 info = ext4_htree_create_dir_info(filp, filp->f_pos);
 535                 if (!info)
 536                         return -ENOMEM;
 537                 filp->private_data = info;
 538         }
 539
 540         if (filp->f_pos == ext4_get_htree_eof(filp))
 541                 return 0;       /* EOF */
 542
 543         /* Some one has messed with f_pos; reset the world */
 544         if (info->last_pos != filp->f_pos) {
 545                 free_rb_tree_fname(&info->root);
 546                 info->curr_node = NULL;
 547                 info->extra_fname = NULL;
 548                 info->curr_hash = pos2maj_hash(filp, filp->f_pos);
 549                 info->curr_minor_hash = pos2min_hash(filp, filp->f_pos);
 550         }
 551
 552         /*
 553          * If there are any leftover names on the hash collision
 554          * chain, return them first.
 555          */
 556         if (info->extra_fname) {
 557                 if (call_filldir(filp, dirent, filldir, info->extra_fname))
 558                         goto finished;
 559                 info->extra_fname = NULL;
 560                 goto next_node;
 561         } else if (!info->curr_node)
 562                 info->curr_node = rb_first(&info->root);
 563
 564         while (1) {
 565                 /*
 566                  * Fill the rbtree if we have no more entries,
 567                  * or the inode has changed since we last read in the
 568                  * cached entries.
 569                  */
 570                 if ((!info->curr_node) ||
 571                     (filp->f_version != inode->i_version)) {
 572                         info->curr_node = NULL;
 573                         free_rb_tree_fname(&info->root);
 574                         filp->f_version = inode->i_version;
 575                         ret = ext4_htree_fill_tree(filp, info->curr_hash,
 576                                                    info->curr_minor_hash,
 577                                                    &info->next_hash);
 578                         if (ret < 0)
 579                                 return ret;
 580                         if (ret == 0) {
 581                                 filp->f_pos = ext4_get_htree_eof(filp);
 582                                 break;
 583                         }
 584                         info->curr_node = rb_first(&info->root);
 585                 }
 586
 587                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
 588                 info->curr_hash = fname->hash;
 589                 info->curr_minor_hash = fname->minor_hash;
 590                 if (call_filldir(filp, dirent, filldir, fname))
 591                         break;
 592         next_node:
 593                 info->curr_node = rb_next(info->curr_node);
 594                 if (info->curr_node) {
 595                         fname = rb_entry(info->curr_node, struct fname,
 596                                          rb_hash);
 597                         info->curr_hash = fname->hash;
 598                         info->curr_minor_hash = fname->minor_hash;
 599                 } else {
 600                         if (info->next_hash == ~0) {
 601                                 filp->f_pos = ext4_get_htree_eof(filp);
 602                                 break;
 603                         }
 604                         info->curr_hash = info->next_hash;
 605                         info->curr_minor_hash = 0;
 606                 }
 607         }
 608 finished:
 609         info->last_pos = filp->f_pos;
 610         return 0;
 611 }
 612
 613 static int ext4_release_dir(struct inode *inode, struct file *filp)
 614 {
 615         if (filp->private_data)
 616                 ext4_htree_free_dir_info(filp->private_data);
 617
 618         return 0;
 619 }
 620
 621 const struct file_operations ext4_dir_operations = {
 622         .llseek         = ext4_dir_llseek,
 623         .read           = generic_read_dir,
 624         .readdir        = ext4_readdir,
 625         .unlocked_ioctl = ext4_ioctl,
 626 #ifdef CONFIG_COMPAT
 627         .compat_ioctl   = ext4_compat_ioctl,
 628 #endif
 629         .fsync          = ext4_sync_file,
 630         .release        = ext4_release_dir,
 631 };