fs/ext3/dir.c

   1 /*
   2  *  linux/fs/ext3/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  *  ext3 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/jbd.h>
  26 #include <linux/ext3_fs.h>
  27 #include <linux/buffer_head.h>
  28 #include <linux/slab.h>
  29 #include <linux/rbtree.h>
  30
  31 static unsigned char ext3_filetype_table[] = {
  32         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
  33 };
  34
  35 static int ext3_readdir(struct file *, void *, filldir_t);
  36 static int ext3_dx_readdir(struct file * filp,
  37                            void * dirent, filldir_t filldir);
  38 static int ext3_release_dir (struct inode * inode,
  39                                 struct file * filp);
  40
  41 const struct file_operations ext3_dir_operations = {
  42         .llseek         = generic_file_llseek,
  43         .read           = generic_read_dir,
  44         .readdir        = ext3_readdir,         /* we take BKL. needed?*/
  45         .unlocked_ioctl = ext3_ioctl,
  46 #ifdef CONFIG_COMPAT
  47         .compat_ioctl   = ext3_compat_ioctl,
  48 #endif
  49         .fsync          = ext3_sync_file,       /* BKL held */
  50         .release        = ext3_release_dir,
  51 };
  52
  53
  54 static unsigned char get_dtype(struct super_block *sb, int filetype)
  55 {
  56         if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_FILETYPE) ||
  57             (filetype >= EXT3_FT_MAX))
  58                 return DT_UNKNOWN;
  59
  60         return (ext3_filetype_table[filetype]);
  61 }
  62
  63
  64 int ext3_check_dir_entry (const char * function, struct inode * dir,
  65                           struct ext3_dir_entry_2 * de,
  66                           struct buffer_head * bh,
  67                           unsigned long offset)
  68 {
  69         const char * error_msg = NULL;
  70         const int rlen = ext3_rec_len_from_disk(de->rec_len);
  71
  72         if (unlikely(rlen < EXT3_DIR_REC_LEN(1)))
  73                 error_msg = "rec_len is smaller than minimal";
  74         else if (unlikely(rlen % 4 != 0))
  75                 error_msg = "rec_len % 4 != 0";
  76         else if (unlikely(rlen < EXT3_DIR_REC_LEN(de->name_len)))
  77                 error_msg = "rec_len is too small for name_len";
  78         else if (unlikely((((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize)))
  79                 error_msg = "directory entry across blocks";
  80         else if (unlikely(le32_to_cpu(de->inode) >
  81                         le32_to_cpu(EXT3_SB(dir->i_sb)->s_es->s_inodes_count)))
  82                 error_msg = "inode out of bounds";
  83
  84         if (unlikely(error_msg != NULL))
  85                 ext3_error (dir->i_sb, function,
  86                         "bad entry in directory #%lu: %s - "
  87                         "offset=%lu, inode=%lu, rec_len=%d, name_len=%d",
  88                         dir->i_ino, error_msg, offset,
  89                         (unsigned long) le32_to_cpu(de->inode),
  90                         rlen, de->name_len);
  91
  92         return error_msg == NULL ? 1 : 0;
  93 }
  94
  95 static int ext3_readdir(struct file * filp,
  96                          void * dirent, filldir_t filldir)
  97 {
  98         int error = 0;
  99         unsigned long offset;
 100         int i, stored;
 101         struct ext3_dir_entry_2 *de;
 102         struct super_block *sb;
 103         int err;
 104         struct inode *inode = filp->f_path.dentry->d_inode;
 105         int ret = 0;
 106         int dir_has_error = 0;
 107
 108         sb = inode->i_sb;
 109
 110         if (EXT3_HAS_COMPAT_FEATURE(inode->i_sb,
 111                                     EXT3_FEATURE_COMPAT_DIR_INDEX) &&
 112             ((EXT3_I(inode)->i_flags & EXT3_INDEX_FL) ||
 113              ((inode->i_size >> sb->s_blocksize_bits) == 1))) {
 114                 err = ext3_dx_readdir(filp, dirent, filldir);
 115                 if (err != ERR_BAD_DX_DIR) {
 116                         ret = err;
 117                         goto out;
 118                 }
 119                 /*
 120                  * We don't set the inode dirty flag since it's not
 121                  * critical that it get flushed back to the disk.
 122                  */
 123                 EXT3_I(filp->f_path.dentry->d_inode)->i_flags &= ~EXT3_INDEX_FL;
 124         }
 125         stored = 0;
 126         offset = filp->f_pos & (sb->s_blocksize - 1);
 127
 128         while (!error && !stored && filp->f_pos < inode->i_size) {
 129                 unsigned long blk = filp->f_pos >> EXT3_BLOCK_SIZE_BITS(sb);
 130                 struct buffer_head map_bh;
 131                 struct buffer_head *bh = NULL;
 132
 133                 map_bh.b_state = 0;
 134                 err = ext3_get_blocks_handle(NULL, inode, blk, 1, &map_bh, 0);
 135                 if (err > 0) {
 136                         pgoff_t index = map_bh.b_blocknr >>
 137                                         (PAGE_CACHE_SHIFT - inode->i_blkbits);
 138                         if (!ra_has_index(&filp->f_ra, index))
 139                                 page_cache_sync_readahead(
 140                                         sb->s_bdev->bd_inode->i_mapping,
 141                                         &filp->f_ra, filp,
 142                                         index, 1);
 143                         filp->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 144                         bh = ext3_bread(NULL, inode, blk, 0, &err);
 145                 }
 146
 147                 /*
 148                  * We ignore I/O errors on directories so users have a chance
 149                  * of recovering data when there's a bad sector
 150                  */
 151                 if (!bh) {
 152                         if (!dir_has_error) {
 153                                 ext3_error(sb, __func__, "directory #%lu "
 154                                         "contains a hole at offset %lld",
 155                                         inode->i_ino, filp->f_pos);
 156                                 dir_has_error = 1;
 157                         }
 158                         /* corrupt size?  Maybe no more blocks to read */
 159                         if (filp->f_pos > inode->i_blocks << 9)
 160                                 break;
 161                         filp->f_pos += sb->s_blocksize - offset;
 162                         continue;
 163                 }
 164
 165 revalidate:
 166                 /* If the dir block has changed since the last call to
 167                  * readdir(2), then we might be pointing to an invalid
 168                  * dirent right now.  Scan from the start of the block
 169                  * to make sure. */
 170                 if (filp->f_version != inode->i_version) {
 171                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
 172                                 de = (struct ext3_dir_entry_2 *)
 173                                         (bh->b_data + i);
 174                                 /* It's too expensive to do a full
 175                                  * dirent test each time round this
 176                                  * loop, but we do have to test at
 177                                  * least that it is non-zero.  A
 178                                  * failure will be detected in the
 179                                  * dirent test below. */
 180                                 if (ext3_rec_len_from_disk(de->rec_len) <
 181                                                 EXT3_DIR_REC_LEN(1))
 182                                         break;
 183                                 i += ext3_rec_len_from_disk(de->rec_len);
 184                         }
 185                         offset = i;
 186                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
 187                                 | offset;
 188                         filp->f_version = inode->i_version;
 189                 }
 190
 191                 while (!error && filp->f_pos < inode->i_size
 192                        && offset < sb->s_blocksize) {
 193                         de = (struct ext3_dir_entry_2 *) (bh->b_data + offset);
 194                         if (!ext3_check_dir_entry ("ext3_readdir", inode, de,
 195                                                    bh, offset)) {
 196                                 /* On error, skip the f_pos to the
 197                                    next block. */
 198                                 filp->f_pos = (filp->f_pos |
 199                                                 (sb->s_blocksize - 1)) + 1;
 200                                 brelse (bh);
 201                                 ret = stored;
 202                                 goto out;
 203                         }
 204                         offset += ext3_rec_len_from_disk(de->rec_len);
 205                         if (le32_to_cpu(de->inode)) {
 206                                 /* We might block in the next section
 207                                  * if the data destination is
 208                                  * currently swapped out.  So, use a
 209                                  * version stamp to detect whether or
 210                                  * not the directory has been modified
 211                                  * during the copy operation.
 212                                  */
 213                                 u64 version = filp->f_version;
 214
 215                                 error = filldir(dirent, de->name,
 216                                                 de->name_len,
 217                                                 filp->f_pos,
 218                                                 le32_to_cpu(de->inode),
 219                                                 get_dtype(sb, de->file_type));
 220                                 if (error)
 221                                         break;
 222                                 if (version != filp->f_version)
 223                                         goto revalidate;
 224                                 stored ++;
 225                         }
 226                         filp->f_pos += ext3_rec_len_from_disk(de->rec_len);
 227                 }
 228                 offset = 0;
 229                 brelse (bh);
 230         }
 231 out:
 232         return ret;
 233 }
 234
 235 /*
 236  * These functions convert from the major/minor hash to an f_pos
 237  * value.
 238  *
 239  * Currently we only use major hash numer.  This is unfortunate, but
 240  * on 32-bit machines, the same VFS interface is used for lseek and
 241  * llseek, so if we use the 64 bit offset, then the 32-bit versions of
 242  * lseek/telldir/seekdir will blow out spectacularly, and from within
 243  * the ext2 low-level routine, we don't know if we're being called by
 244  * a 64-bit version of the system call or the 32-bit version of the
 245  * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
 246  * cookie.  Sigh.
 247  */
 248 #define hash2pos(major, minor)  (major >> 1)
 249 #define pos2maj_hash(pos)       ((pos << 1) & 0xffffffff)
 250 #define pos2min_hash(pos)       (0)
 251
 252 /*
 253  * This structure holds the nodes of the red-black tree used to store
 254  * the directory entry in hash order.
 255  */
 256 struct fname {
 257         __u32           hash;
 258         __u32           minor_hash;
 259         struct rb_node  rb_hash;
 260         struct fname    *next;
 261         __u32           inode;
 262         __u8            name_len;
 263         __u8            file_type;
 264         char            name[0];
 265 };
 266
 267 /*
 268  * This functoin implements a non-recursive way of freeing all of the
 269  * nodes in the red-black tree.
 270  */
 271 static void free_rb_tree_fname(struct rb_root *root)
 272 {
 273         struct rb_node  *n = root->rb_node;
 274         struct rb_node  *parent;
 275         struct fname    *fname;
 276
 277         while (n) {
 278                 /* Do the node's children first */
 279                 if (n->rb_left) {
 280                         n = n->rb_left;
 281                         continue;
 282                 }
 283                 if (n->rb_right) {
 284                         n = n->rb_right;
 285                         continue;
 286                 }
 287                 /*
 288                  * The node has no children; free it, and then zero
 289                  * out parent's link to it.  Finally go to the
 290                  * beginning of the loop and try to free the parent
 291                  * node.
 292                  */
 293                 parent = rb_parent(n);
 294                 fname = rb_entry(n, struct fname, rb_hash);
 295                 while (fname) {
 296                         struct fname * old = fname;
 297                         fname = fname->next;
 298                         kfree (old);
 299                 }
 300                 if (!parent)
 301                         *root = RB_ROOT;
 302                 else if (parent->rb_left == n)
 303                         parent->rb_left = NULL;
 304                 else if (parent->rb_right == n)
 305                         parent->rb_right = NULL;
 306                 n = parent;
 307         }
 308 }
 309
 310
 311 static struct dir_private_info *ext3_htree_create_dir_info(loff_t pos)
 312 {
 313         struct dir_private_info *p;
 314
 315         p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
 316         if (!p)
 317                 return NULL;
 318         p->curr_hash = pos2maj_hash(pos);
 319         p->curr_minor_hash = pos2min_hash(pos);
 320         return p;
 321 }
 322
 323 void ext3_htree_free_dir_info(struct dir_private_info *p)
 324 {
 325         free_rb_tree_fname(&p->root);
 326         kfree(p);
 327 }
 328
 329 /*
 330  * Given a directory entry, enter it into the fname rb tree.
 331  */
 332 int ext3_htree_store_dirent(struct file *dir_file, __u32 hash,
 333                              __u32 minor_hash,
 334                              struct ext3_dir_entry_2 *dirent)
 335 {
 336         struct rb_node **p, *parent = NULL;
 337         struct fname * fname, *new_fn;
 338         struct dir_private_info *info;
 339         int len;
 340
 341         info = (struct dir_private_info *) dir_file->private_data;
 342         p = &info->root.rb_node;
 343
 344         /* Create and allocate the fname structure */
 345         len = sizeof(struct fname) + dirent->name_len + 1;
 346         new_fn = kzalloc(len, GFP_KERNEL);
 347         if (!new_fn)
 348                 return -ENOMEM;
 349         new_fn->hash = hash;
 350         new_fn->minor_hash = minor_hash;
 351         new_fn->inode = le32_to_cpu(dirent->inode);
 352         new_fn->name_len = dirent->name_len;
 353         new_fn->file_type = dirent->file_type;
 354         memcpy(new_fn->name, dirent->name, dirent->name_len);
 355         new_fn->name[dirent->name_len] = 0;
 356
 357         while (*p) {
 358                 parent = *p;
 359                 fname = rb_entry(parent, struct fname, rb_hash);
 360
 361                 /*
 362                  * If the hash and minor hash match up, then we put
 363                  * them on a linked list.  This rarely happens...
 364                  */
 365                 if ((new_fn->hash == fname->hash) &&
 366                     (new_fn->minor_hash == fname->minor_hash)) {
 367                         new_fn->next = fname->next;
 368                         fname->next = new_fn;
 369                         return 0;
 370                 }
 371
 372                 if (new_fn->hash < fname->hash)
 373                         p = &(*p)->rb_left;
 374                 else if (new_fn->hash > fname->hash)
 375                         p = &(*p)->rb_right;
 376                 else if (new_fn->minor_hash < fname->minor_hash)
 377                         p = &(*p)->rb_left;
 378                 else /* if (new_fn->minor_hash > fname->minor_hash) */
 379                         p = &(*p)->rb_right;
 380         }
 381
 382         rb_link_node(&new_fn->rb_hash, parent, p);
 383         rb_insert_color(&new_fn->rb_hash, &info->root);
 384         return 0;
 385 }
 386
 387
 388
 389 /*
 390  * This is a helper function for ext3_dx_readdir.  It calls filldir
 391  * for all entres on the fname linked list.  (Normally there is only
 392  * one entry on the linked list, unless there are 62 bit hash collisions.)
 393  */
 394 static int call_filldir(struct file * filp, void * dirent,
 395                         filldir_t filldir, struct fname *fname)
 396 {
 397         struct dir_private_info *info = filp->private_data;
 398         loff_t  curr_pos;
 399         struct inode *inode = filp->f_path.dentry->d_inode;
 400         struct super_block * sb;
 401         int error;
 402
 403         sb = inode->i_sb;
 404
 405         if (!fname) {
 406                 printk("call_filldir: called with null fname?!?\n");
 407                 return 0;
 408         }
 409         curr_pos = hash2pos(fname->hash, fname->minor_hash);
 410         while (fname) {
 411                 error = filldir(dirent, fname->name,
 412                                 fname->name_len, curr_pos,
 413                                 fname->inode,
 414                                 get_dtype(sb, fname->file_type));
 415                 if (error) {
 416                         filp->f_pos = curr_pos;
 417                         info->extra_fname = fname;
 418                         return error;
 419                 }
 420                 fname = fname->next;
 421         }
 422         return 0;
 423 }
 424
 425 static int ext3_dx_readdir(struct file * filp,
 426                          void * dirent, filldir_t filldir)
 427 {
 428         struct dir_private_info *info = filp->private_data;
 429         struct inode *inode = filp->f_path.dentry->d_inode;
 430         struct fname *fname;
 431         int     ret;
 432
 433         if (!info) {
 434                 info = ext3_htree_create_dir_info(filp->f_pos);
 435                 if (!info)
 436                         return -ENOMEM;
 437                 filp->private_data = info;
 438         }
 439
 440         if (filp->f_pos == EXT3_HTREE_EOF)
 441                 return 0;       /* EOF */
 442
 443         /* Some one has messed with f_pos; reset the world */
 444         if (info->last_pos != filp->f_pos) {
 445                 free_rb_tree_fname(&info->root);
 446                 info->curr_node = NULL;
 447                 info->extra_fname = NULL;
 448                 info->curr_hash = pos2maj_hash(filp->f_pos);
 449                 info->curr_minor_hash = pos2min_hash(filp->f_pos);
 450         }
 451
 452         /*
 453          * If there are any leftover names on the hash collision
 454          * chain, return them first.
 455          */
 456         if (info->extra_fname) {
 457                 if (call_filldir(filp, dirent, filldir, info->extra_fname))
 458                         goto finished;
 459                 info->extra_fname = NULL;
 460                 goto next_node;
 461         } else if (!info->curr_node)
 462                 info->curr_node = rb_first(&info->root);
 463
 464         while (1) {
 465                 /*
 466                  * Fill the rbtree if we have no more entries,
 467                  * or the inode has changed since we last read in the
 468                  * cached entries.
 469                  */
 470                 if ((!info->curr_node) ||
 471                     (filp->f_version != inode->i_version)) {
 472                         info->curr_node = NULL;
 473                         free_rb_tree_fname(&info->root);
 474                         filp->f_version = inode->i_version;
 475                         ret = ext3_htree_fill_tree(filp, info->curr_hash,
 476                                                    info->curr_minor_hash,
 477                                                    &info->next_hash);
 478                         if (ret < 0)
 479                                 return ret;
 480                         if (ret == 0) {
 481                                 filp->f_pos = EXT3_HTREE_EOF;
 482                                 break;
 483                         }
 484                         info->curr_node = rb_first(&info->root);
 485                 }
 486
 487                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
 488                 info->curr_hash = fname->hash;
 489                 info->curr_minor_hash = fname->minor_hash;
 490                 if (call_filldir(filp, dirent, filldir, fname))
 491                         break;
 492         next_node:
 493                 info->curr_node = rb_next(info->curr_node);
 494                 if (info->curr_node) {
 495                         fname = rb_entry(info->curr_node, struct fname,
 496                                          rb_hash);
 497                         info->curr_hash = fname->hash;
 498                         info->curr_minor_hash = fname->minor_hash;
 499                 } else {
 500                         if (info->next_hash == ~0) {
 501                                 filp->f_pos = EXT3_HTREE_EOF;
 502                                 break;
 503                         }
 504                         info->curr_hash = info->next_hash;
 505                         info->curr_minor_hash = 0;
 506                 }
 507         }
 508 finished:
 509         info->last_pos = filp->f_pos;
 510         return 0;
 511 }
 512
 513 static int ext3_release_dir (struct inode * inode, struct file * filp)
 514 {
 515        if (filp->private_data)
 516                 ext3_htree_free_dir_info(filp->private_data);
 517
 518         return 0;
 519 }