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/ext4_fs.h>
  27 #include <linux/buffer_head.h>
  28 #include <linux/slab.h>
  29 #include <linux/rbtree.h>
  30
  31 static unsigned char ext4_filetype_table[] = {
  32         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
  33 };
  34
  35 static int ext4_readdir(struct file *, void *, filldir_t);
  36 static int ext4_dx_readdir(struct file * filp,
  37                            void * dirent, filldir_t filldir);
  38 static int ext4_release_dir (struct inode * inode,
  39                                 struct file * filp);
  40
  41 const struct file_operations ext4_dir_operations = {
  42         .llseek         = generic_file_llseek,
  43         .read           = generic_read_dir,
  44         .readdir        = ext4_readdir,         /* we take BKL. needed?*/
  45         .ioctl          = ext4_ioctl,           /* BKL held */
  46 #ifdef CONFIG_COMPAT
  47         .compat_ioctl   = ext4_compat_ioctl,
  48 #endif
  49         .fsync          = ext4_sync_file,       /* BKL held */
  50 #ifdef CONFIG_EXT4_INDEX
  51         .release        = ext4_release_dir,
  52 #endif
  53 };
  54
  55
  56 static unsigned char get_dtype(struct super_block *sb, int filetype)
  57 {
  58         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_FILETYPE) ||
  59             (filetype >= EXT4_FT_MAX))
  60                 return DT_UNKNOWN;
  61
  62         return (ext4_filetype_table[filetype]);
  63 }
  64
  65
  66 int ext4_check_dir_entry (const char * function, struct inode * dir,
  67                           struct ext4_dir_entry_2 * de,
  68                           struct buffer_head * bh,
  69                           unsigned long offset)
  70 {
  71         const char * error_msg = NULL;
  72         const int rlen = le16_to_cpu(de->rec_len);
  73
  74         if (rlen < EXT4_DIR_REC_LEN(1))
  75                 error_msg = "rec_len is smaller than minimal";
  76         else if (rlen % 4 != 0)
  77                 error_msg = "rec_len % 4 != 0";
  78         else if (rlen < EXT4_DIR_REC_LEN(de->name_len))
  79                 error_msg = "rec_len is too small for name_len";
  80         else if (((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize)
  81                 error_msg = "directory entry across blocks";
  82         else if (le32_to_cpu(de->inode) >
  83                         le32_to_cpu(EXT4_SB(dir->i_sb)->s_es->s_inodes_count))
  84                 error_msg = "inode out of bounds";
  85
  86         if (error_msg != NULL)
  87                 ext4_error (dir->i_sb, function,
  88                         "bad entry in directory #%lu: %s - "
  89                         "offset=%lu, inode=%lu, rec_len=%d, name_len=%d",
  90                         dir->i_ino, error_msg, offset,
  91                         (unsigned long) le32_to_cpu(de->inode),
  92                         rlen, de->name_len);
  93         return error_msg == NULL ? 1 : 0;
  94 }
  95
  96 static int ext4_readdir(struct file * filp,
  97                          void * dirent, filldir_t filldir)
  98 {
  99         int error = 0;
 100         unsigned long offset;
 101         int i, stored;
 102         struct ext4_dir_entry_2 *de;
 103         struct super_block *sb;
 104         int err;
 105         struct inode *inode = filp->f_path.dentry->d_inode;
 106         int ret = 0;
 107
 108         sb = inode->i_sb;
 109
 110 #ifdef CONFIG_EXT4_INDEX
 111         if (EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
 112                                     EXT4_FEATURE_COMPAT_DIR_INDEX) &&
 113             ((EXT4_I(inode)->i_flags & EXT4_INDEX_FL) ||
 114              ((inode->i_size >> sb->s_blocksize_bits) == 1))) {
 115                 err = ext4_dx_readdir(filp, dirent, filldir);
 116                 if (err != ERR_BAD_DX_DIR) {
 117                         ret = err;
 118                         goto out;
 119                 }
 120                 /*
 121                  * We don't set the inode dirty flag since it's not
 122                  * critical that it get flushed back to the disk.
 123                  */
 124                 EXT4_I(filp->f_path.dentry->d_inode)->i_flags &= ~EXT4_INDEX_FL;
 125         }
 126 #endif
 127         stored = 0;
 128         offset = filp->f_pos & (sb->s_blocksize - 1);
 129
 130         while (!error && !stored && filp->f_pos < inode->i_size) {
 131                 unsigned long blk = filp->f_pos >> EXT4_BLOCK_SIZE_BITS(sb);
 132                 struct buffer_head map_bh;
 133                 struct buffer_head *bh = NULL;
 134
 135                 map_bh.b_state = 0;
 136                 err = ext4_get_blocks_wrap(NULL, inode, blk, 1, &map_bh, 0, 0);
 137                 if (err > 0) {
 138                         pgoff_t index = map_bh.b_blocknr >>
 139                                         (PAGE_CACHE_SHIFT - inode->i_blkbits);
 140                         if (!ra_has_index(&filp->f_ra, index))
 141                                 page_cache_sync_readahead(
 142                                         sb->s_bdev->bd_inode->i_mapping,
 143                                         &filp->f_ra, filp,
 144                                         index, 1);
 145                         filp->f_ra.prev_index = index;
 146                         bh = ext4_bread(NULL, inode, blk, 0, &err);
 147                 }
 148
 149                 /*
 150                  * We ignore I/O errors on directories so users have a chance
 151                  * of recovering data when there's a bad sector
 152                  */
 153                 if (!bh) {
 154                         ext4_error (sb, "ext4_readdir",
 155                                 "directory #%lu contains a hole at offset %lu",
 156                                 inode->i_ino, (unsigned long)filp->f_pos);
 157                         /* corrupt size?  Maybe no more blocks to read */
 158                         if (filp->f_pos > inode->i_blocks << 9)
 159                                 break;
 160                         filp->f_pos += sb->s_blocksize - offset;
 161                         continue;
 162                 }
 163
 164 revalidate:
 165                 /* If the dir block has changed since the last call to
 166                  * readdir(2), then we might be pointing to an invalid
 167                  * dirent right now.  Scan from the start of the block
 168                  * to make sure. */
 169                 if (filp->f_version != inode->i_version) {
 170                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
 171                                 de = (struct ext4_dir_entry_2 *)
 172                                         (bh->b_data + i);
 173                                 /* It's too expensive to do a full
 174                                  * dirent test each time round this
 175                                  * loop, but we do have to test at
 176                                  * least that it is non-zero.  A
 177                                  * failure will be detected in the
 178                                  * dirent test below. */
 179                                 if (le16_to_cpu(de->rec_len) <
 180                                                 EXT4_DIR_REC_LEN(1))
 181                                         break;
 182                                 i += le16_to_cpu(de->rec_len);
 183                         }
 184                         offset = i;
 185                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
 186                                 | offset;
 187                         filp->f_version = inode->i_version;
 188                 }
 189
 190                 while (!error && filp->f_pos < inode->i_size
 191                        && offset < sb->s_blocksize) {
 192                         de = (struct ext4_dir_entry_2 *) (bh->b_data + offset);
 193                         if (!ext4_check_dir_entry ("ext4_readdir", inode, de,
 194                                                    bh, offset)) {
 195                                 /*
 196                                  * On error, skip the f_pos to the next block
 197                                  */
 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 += le16_to_cpu(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                                 unsigned long 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 += le16_to_cpu(de->rec_len);
 227                 }
 228                 offset = 0;
 229                 brelse (bh);
 230         }
 231 out:
 232         return ret;
 233 }
 234
 235 #ifdef CONFIG_EXT4_INDEX
 236 /*
 237  * These functions convert from the major/minor hash to an f_pos
 238  * value.
 239  *
 240  * Currently we only use major hash numer.  This is unfortunate, but
 241  * on 32-bit machines, the same VFS interface is used for lseek and
 242  * llseek, so if we use the 64 bit offset, then the 32-bit versions of
 243  * lseek/telldir/seekdir will blow out spectacularly, and from within
 244  * the ext2 low-level routine, we don't know if we're being called by
 245  * a 64-bit version of the system call or the 32-bit version of the
 246  * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
 247  * cookie.  Sigh.
 248  */
 249 #define hash2pos(major, minor)  (major >> 1)
 250 #define pos2maj_hash(pos)       ((pos << 1) & 0xffffffff)
 251 #define pos2min_hash(pos)       (0)
 252
 253 /*
 254  * This structure holds the nodes of the red-black tree used to store
 255  * the directory entry in hash order.
 256  */
 257 struct fname {
 258         __u32           hash;
 259         __u32           minor_hash;
 260         struct rb_node  rb_hash;
 261         struct fname    *next;
 262         __u32           inode;
 263         __u8            name_len;
 264         __u8            file_type;
 265         char            name[0];
 266 };
 267
 268 /*
 269  * This functoin implements a non-recursive way of freeing all of the
 270  * nodes in the red-black tree.
 271  */
 272 static void free_rb_tree_fname(struct rb_root *root)
 273 {
 274         struct rb_node  *n = root->rb_node;
 275         struct rb_node  *parent;
 276         struct fname    *fname;
 277
 278         while (n) {
 279                 /* Do the node's children first */
 280                 if ((n)->rb_left) {
 281                         n = n->rb_left;
 282                         continue;
 283                 }
 284                 if (n->rb_right) {
 285                         n = n->rb_right;
 286                         continue;
 287                 }
 288                 /*
 289                  * The node has no children; free it, and then zero
 290                  * out parent's link to it.  Finally go to the
 291                  * beginning of the loop and try to free the parent
 292                  * node.
 293                  */
 294                 parent = rb_parent(n);
 295                 fname = rb_entry(n, struct fname, rb_hash);
 296                 while (fname) {
 297                         struct fname * old = fname;
 298                         fname = fname->next;
 299                         kfree (old);
 300                 }
 301                 if (!parent)
 302                         root->rb_node = NULL;
 303                 else if (parent->rb_left == n)
 304                         parent->rb_left = NULL;
 305                 else if (parent->rb_right == n)
 306                         parent->rb_right = NULL;
 307                 n = parent;
 308         }
 309         root->rb_node = NULL;
 310 }
 311
 312
 313 static struct dir_private_info *create_dir_info(loff_t pos)
 314 {
 315         struct dir_private_info *p;
 316
 317         p = kmalloc(sizeof(struct dir_private_info), GFP_KERNEL);
 318         if (!p)
 319                 return NULL;
 320         p->root.rb_node = NULL;
 321         p->curr_node = NULL;
 322         p->extra_fname = NULL;
 323         p->last_pos = 0;
 324         p->curr_hash = pos2maj_hash(pos);
 325         p->curr_minor_hash = pos2min_hash(pos);
 326         p->next_hash = 0;
 327         return p;
 328 }
 329
 330 void ext4_htree_free_dir_info(struct dir_private_info *p)
 331 {
 332         free_rb_tree_fname(&p->root);
 333         kfree(p);
 334 }
 335
 336 /*
 337  * Given a directory entry, enter it into the fname rb tree.
 338  */
 339 int ext4_htree_store_dirent(struct file *dir_file, __u32 hash,
 340                              __u32 minor_hash,
 341                              struct ext4_dir_entry_2 *dirent)
 342 {
 343         struct rb_node **p, *parent = NULL;
 344         struct fname * fname, *new_fn;
 345         struct dir_private_info *info;
 346         int len;
 347
 348         info = (struct dir_private_info *) dir_file->private_data;
 349         p = &info->root.rb_node;
 350
 351         /* Create and allocate the fname structure */
 352         len = sizeof(struct fname) + dirent->name_len + 1;
 353         new_fn = kzalloc(len, GFP_KERNEL);
 354         if (!new_fn)
 355                 return -ENOMEM;
 356         new_fn->hash = hash;
 357         new_fn->minor_hash = minor_hash;
 358         new_fn->inode = le32_to_cpu(dirent->inode);
 359         new_fn->name_len = dirent->name_len;
 360         new_fn->file_type = dirent->file_type;
 361         memcpy(new_fn->name, dirent->name, dirent->name_len);
 362         new_fn->name[dirent->name_len] = 0;
 363
 364         while (*p) {
 365                 parent = *p;
 366                 fname = rb_entry(parent, struct fname, rb_hash);
 367
 368                 /*
 369                  * If the hash and minor hash match up, then we put
 370                  * them on a linked list.  This rarely happens...
 371                  */
 372                 if ((new_fn->hash == fname->hash) &&
 373                     (new_fn->minor_hash == fname->minor_hash)) {
 374                         new_fn->next = fname->next;
 375                         fname->next = new_fn;
 376                         return 0;
 377                 }
 378
 379                 if (new_fn->hash < fname->hash)
 380                         p = &(*p)->rb_left;
 381                 else if (new_fn->hash > fname->hash)
 382                         p = &(*p)->rb_right;
 383                 else if (new_fn->minor_hash < fname->minor_hash)
 384                         p = &(*p)->rb_left;
 385                 else /* if (new_fn->minor_hash > fname->minor_hash) */
 386                         p = &(*p)->rb_right;
 387         }
 388
 389         rb_link_node(&new_fn->rb_hash, parent, p);
 390         rb_insert_color(&new_fn->rb_hash, &info->root);
 391         return 0;
 392 }
 393
 394
 395
 396 /*
 397  * This is a helper function for ext4_dx_readdir.  It calls filldir
 398  * for all entres on the fname linked list.  (Normally there is only
 399  * one entry on the linked list, unless there are 62 bit hash collisions.)
 400  */
 401 static int call_filldir(struct file * filp, void * dirent,
 402                         filldir_t filldir, struct fname *fname)
 403 {
 404         struct dir_private_info *info = filp->private_data;
 405         loff_t  curr_pos;
 406         struct inode *inode = filp->f_path.dentry->d_inode;
 407         struct super_block * sb;
 408         int error;
 409
 410         sb = inode->i_sb;
 411
 412         if (!fname) {
 413                 printk("call_filldir: called with null fname?!?\n");
 414                 return 0;
 415         }
 416         curr_pos = hash2pos(fname->hash, fname->minor_hash);
 417         while (fname) {
 418                 error = filldir(dirent, fname->name,
 419                                 fname->name_len, curr_pos,
 420                                 fname->inode,
 421                                 get_dtype(sb, fname->file_type));
 422                 if (error) {
 423                         filp->f_pos = curr_pos;
 424                         info->extra_fname = fname->next;
 425                         return error;
 426                 }
 427                 fname = fname->next;
 428         }
 429         return 0;
 430 }
 431
 432 static int ext4_dx_readdir(struct file * filp,
 433                          void * dirent, filldir_t filldir)
 434 {
 435         struct dir_private_info *info = filp->private_data;
 436         struct inode *inode = filp->f_path.dentry->d_inode;
 437         struct fname *fname;
 438         int     ret;
 439
 440         if (!info) {
 441                 info = create_dir_info(filp->f_pos);
 442                 if (!info)
 443                         return -ENOMEM;
 444                 filp->private_data = info;
 445         }
 446
 447         if (filp->f_pos == EXT4_HTREE_EOF)
 448                 return 0;       /* EOF */
 449
 450         /* Some one has messed with f_pos; reset the world */
 451         if (info->last_pos != filp->f_pos) {
 452                 free_rb_tree_fname(&info->root);
 453                 info->curr_node = NULL;
 454                 info->extra_fname = NULL;
 455                 info->curr_hash = pos2maj_hash(filp->f_pos);
 456                 info->curr_minor_hash = pos2min_hash(filp->f_pos);
 457         }
 458
 459         /*
 460          * If there are any leftover names on the hash collision
 461          * chain, return them first.
 462          */
 463         if (info->extra_fname &&
 464             call_filldir(filp, dirent, filldir, info->extra_fname))
 465                 goto finished;
 466
 467         if (!info->curr_node)
 468                 info->curr_node = rb_first(&info->root);
 469
 470         while (1) {
 471                 /*
 472                  * Fill the rbtree if we have no more entries,
 473                  * or the inode has changed since we last read in the
 474                  * cached entries.
 475                  */
 476                 if ((!info->curr_node) ||
 477                     (filp->f_version != inode->i_version)) {
 478                         info->curr_node = NULL;
 479                         free_rb_tree_fname(&info->root);
 480                         filp->f_version = inode->i_version;
 481                         ret = ext4_htree_fill_tree(filp, info->curr_hash,
 482                                                    info->curr_minor_hash,
 483                                                    &info->next_hash);
 484                         if (ret < 0)
 485                                 return ret;
 486                         if (ret == 0) {
 487                                 filp->f_pos = EXT4_HTREE_EOF;
 488                                 break;
 489                         }
 490                         info->curr_node = rb_first(&info->root);
 491                 }
 492
 493                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
 494                 info->curr_hash = fname->hash;
 495                 info->curr_minor_hash = fname->minor_hash;
 496                 if (call_filldir(filp, dirent, filldir, fname))
 497                         break;
 498
 499                 info->curr_node = rb_next(info->curr_node);
 500                 if (!info->curr_node) {
 501                         if (info->next_hash == ~0) {
 502                                 filp->f_pos = EXT4_HTREE_EOF;
 503                                 break;
 504                         }
 505                         info->curr_hash = info->next_hash;
 506                         info->curr_minor_hash = 0;
 507                 }
 508         }
 509 finished:
 510         info->last_pos = filp->f_pos;
 511         return 0;
 512 }
 513
 514 static int ext4_release_dir (struct inode * inode, struct file * filp)
 515 {
 516         if (filp->private_data)
 517                 ext4_htree_free_dir_info(filp->private_data);
 518
 519         return 0;
 520 }
 521
 522 #endif