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