29857ddd9e2602e6cdaa59ba7976422f54815c8e
[linux-2.6.git] / 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
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         .unlocked_ioctl = ext4_ioctl,
46 #ifdef CONFIG_COMPAT
47         .compat_ioctl   = ext4_compat_ioctl,
48 #endif
49         .fsync          = ext4_sync_file,
50         .release        = ext4_release_dir,
51 };
52
53
54 static unsigned char get_dtype(struct super_block *sb, int filetype)
55 {
56         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_FILETYPE) ||
57             (filetype >= EXT4_FT_MAX))
58                 return DT_UNKNOWN;
59
60         return (ext4_filetype_table[filetype]);
61 }
62
63
64 int ext4_check_dir_entry(const char *function, struct inode *dir,
65                          struct ext4_dir_entry_2 *de,
66                          struct buffer_head *bh,
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 (rlen < EXT4_DIR_REC_LEN(1))
74                 error_msg = "rec_len is smaller than minimal";
75         else if (rlen % 4 != 0)
76                 error_msg = "rec_len % 4 != 0";
77         else if (rlen < EXT4_DIR_REC_LEN(de->name_len))
78                 error_msg = "rec_len is too small for name_len";
79         else if (((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize)
80                 error_msg = "directory entry across blocks";
81         else if (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
85         if (error_msg != NULL)
86                 __ext4_error(dir->i_sb, function,
87                         "bad entry in directory #%lu: %s - block=%llu"
88                         "offset=%u(%u), inode=%u, rec_len=%d, name_len=%d",
89                         dir->i_ino, error_msg, 
90                         (unsigned long long) bh->b_blocknr,     
91                         (unsigned) (offset%bh->b_size), offset,
92                         le32_to_cpu(de->inode),
93                         rlen, de->name_len);
94         return error_msg == NULL ? 1 : 0;
95 }
96
97 static int ext4_readdir(struct file *filp,
98                          void *dirent, filldir_t filldir)
99 {
100         int error = 0;
101         unsigned int offset;
102         int i, stored;
103         struct ext4_dir_entry_2 *de;
104         struct super_block *sb;
105         int err;
106         struct inode *inode = filp->f_path.dentry->d_inode;
107         int ret = 0;
108         int dir_has_error = 0;
109
110         sb = inode->i_sb;
111
112         if (EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
113                                     EXT4_FEATURE_COMPAT_DIR_INDEX) &&
114             ((EXT4_I(inode)->i_flags & EXT4_INDEX_FL) ||
115              ((inode->i_size >> sb->s_blocksize_bits) == 1))) {
116                 err = ext4_dx_readdir(filp, dirent, filldir);
117                 if (err != ERR_BAD_DX_DIR) {
118                         ret = err;
119                         goto out;
120                 }
121                 /*
122                  * We don't set the inode dirty flag since it's not
123                  * critical that it get flushed back to the disk.
124                  */
125                 EXT4_I(filp->f_path.dentry->d_inode)->i_flags &= ~EXT4_INDEX_FL;
126         }
127         stored = 0;
128         offset = filp->f_pos & (sb->s_blocksize - 1);
129
130         while (!error && !stored && filp->f_pos < inode->i_size) {
131                 ext4_lblk_t 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(NULL, inode, blk, 1, &map_bh, 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_pos = (loff_t)index << PAGE_CACHE_SHIFT;
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                         if (!dir_has_error) {
155                                 ext4_error(sb, "directory #%lu "
156                                            "contains a hole at offset %Lu",
157                                            inode->i_ino,
158                                            (unsigned long long) filp->f_pos);
159                                 dir_has_error = 1;
160                         }
161                         /* corrupt size?  Maybe no more blocks to read */
162                         if (filp->f_pos > inode->i_blocks << 9)
163                                 break;
164                         filp->f_pos += sb->s_blocksize - offset;
165                         continue;
166                 }
167
168 revalidate:
169                 /* If the dir block has changed since the last call to
170                  * readdir(2), then we might be pointing to an invalid
171                  * dirent right now.  Scan from the start of the block
172                  * to make sure. */
173                 if (filp->f_version != inode->i_version) {
174                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
175                                 de = (struct ext4_dir_entry_2 *)
176                                         (bh->b_data + i);
177                                 /* It's too expensive to do a full
178                                  * dirent test each time round this
179                                  * loop, but we do have to test at
180                                  * least that it is non-zero.  A
181                                  * failure will be detected in the
182                                  * dirent test below. */
183                                 if (ext4_rec_len_from_disk(de->rec_len,
184                                         sb->s_blocksize) < EXT4_DIR_REC_LEN(1))
185                                         break;
186                                 i += ext4_rec_len_from_disk(de->rec_len,
187                                                             sb->s_blocksize);
188                         }
189                         offset = i;
190                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
191                                 | offset;
192                         filp->f_version = inode->i_version;
193                 }
194
195                 while (!error && filp->f_pos < inode->i_size
196                        && offset < sb->s_blocksize) {
197                         de = (struct ext4_dir_entry_2 *) (bh->b_data + offset);
198                         if (!ext4_check_dir_entry("ext4_readdir", inode, de,
199                                                   bh, offset)) {
200                                 /*
201                                  * On error, skip the f_pos to the next block
202                                  */
203                                 filp->f_pos = (filp->f_pos |
204                                                 (sb->s_blocksize - 1)) + 1;
205                                 brelse(bh);
206                                 ret = stored;
207                                 goto out;
208                         }
209                         offset += ext4_rec_len_from_disk(de->rec_len,
210                                         sb->s_blocksize);
211                         if (le32_to_cpu(de->inode)) {
212                                 /* We might block in the next section
213                                  * if the data destination is
214                                  * currently swapped out.  So, use a
215                                  * version stamp to detect whether or
216                                  * not the directory has been modified
217                                  * during the copy operation.
218                                  */
219                                 u64 version = filp->f_version;
220
221                                 error = filldir(dirent, de->name,
222                                                 de->name_len,
223                                                 filp->f_pos,
224                                                 le32_to_cpu(de->inode),
225                                                 get_dtype(sb, de->file_type));
226                                 if (error)
227                                         break;
228                                 if (version != filp->f_version)
229                                         goto revalidate;
230                                 stored++;
231                         }
232                         filp->f_pos += ext4_rec_len_from_disk(de->rec_len,
233                                                 sb->s_blocksize);
234                 }
235                 offset = 0;
236                 brelse(bh);
237         }
238 out:
239         return ret;
240 }
241
242 /*
243  * These functions convert from the major/minor hash to an f_pos
244  * value.
245  *
246  * Currently we only use major hash numer.  This is unfortunate, but
247  * on 32-bit machines, the same VFS interface is used for lseek and
248  * llseek, so if we use the 64 bit offset, then the 32-bit versions of
249  * lseek/telldir/seekdir will blow out spectacularly, and from within
250  * the ext2 low-level routine, we don't know if we're being called by
251  * a 64-bit version of the system call or the 32-bit version of the
252  * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
253  * cookie.  Sigh.
254  */
255 #define hash2pos(major, minor)  (major >> 1)
256 #define pos2maj_hash(pos)       ((pos << 1) & 0xffffffff)
257 #define pos2min_hash(pos)       (0)
258
259 /*
260  * This structure holds the nodes of the red-black tree used to store
261  * the directory entry in hash order.
262  */
263 struct fname {
264         __u32           hash;
265         __u32           minor_hash;
266         struct rb_node  rb_hash;
267         struct fname    *next;
268         __u32           inode;
269         __u8            name_len;
270         __u8            file_type;
271         char            name[0];
272 };
273
274 /*
275  * This functoin implements a non-recursive way of freeing all of the
276  * nodes in the red-black tree.
277  */
278 static void free_rb_tree_fname(struct rb_root *root)
279 {
280         struct rb_node  *n = root->rb_node;
281         struct rb_node  *parent;
282         struct fname    *fname;
283
284         while (n) {
285                 /* Do the node's children first */
286                 if (n->rb_left) {
287                         n = n->rb_left;
288                         continue;
289                 }
290                 if (n->rb_right) {
291                         n = n->rb_right;
292                         continue;
293                 }
294                 /*
295                  * The node has no children; free it, and then zero
296                  * out parent's link to it.  Finally go to the
297                  * beginning of the loop and try to free the parent
298                  * node.
299                  */
300                 parent = rb_parent(n);
301                 fname = rb_entry(n, struct fname, rb_hash);
302                 while (fname) {
303                         struct fname *old = fname;
304                         fname = fname->next;
305                         kfree(old);
306                 }
307                 if (!parent)
308                         root->rb_node = NULL;
309                 else if (parent->rb_left == n)
310                         parent->rb_left = NULL;
311                 else if (parent->rb_right == n)
312                         parent->rb_right = NULL;
313                 n = parent;
314         }
315 }
316
317
318 static struct dir_private_info *ext4_htree_create_dir_info(loff_t pos)
319 {
320         struct dir_private_info *p;
321
322         p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
323         if (!p)
324                 return NULL;
325         p->curr_hash = pos2maj_hash(pos);
326         p->curr_minor_hash = pos2min_hash(pos);
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(KERN_ERR "EXT4-fs: call_filldir: called with "
414                        "null fname?!?\n");
415                 return 0;
416         }
417         curr_pos = hash2pos(fname->hash, fname->minor_hash);
418         while (fname) {
419                 error = filldir(dirent, fname->name,
420                                 fname->name_len, curr_pos,
421                                 fname->inode,
422                                 get_dtype(sb, fname->file_type));
423                 if (error) {
424                         filp->f_pos = curr_pos;
425                         info->extra_fname = fname;
426                         return error;
427                 }
428                 fname = fname->next;
429         }
430         return 0;
431 }
432
433 static int ext4_dx_readdir(struct file *filp,
434                          void *dirent, filldir_t filldir)
435 {
436         struct dir_private_info *info = filp->private_data;
437         struct inode *inode = filp->f_path.dentry->d_inode;
438         struct fname *fname;
439         int     ret;
440
441         if (!info) {
442                 info = ext4_htree_create_dir_info(filp->f_pos);
443                 if (!info)
444                         return -ENOMEM;
445                 filp->private_data = info;
446         }
447
448         if (filp->f_pos == EXT4_HTREE_EOF)
449                 return 0;       /* EOF */
450
451         /* Some one has messed with f_pos; reset the world */
452         if (info->last_pos != filp->f_pos) {
453                 free_rb_tree_fname(&info->root);
454                 info->curr_node = NULL;
455                 info->extra_fname = NULL;
456                 info->curr_hash = pos2maj_hash(filp->f_pos);
457                 info->curr_minor_hash = pos2min_hash(filp->f_pos);
458         }
459
460         /*
461          * If there are any leftover names on the hash collision
462          * chain, return them first.
463          */
464         if (info->extra_fname) {
465                 if (call_filldir(filp, dirent, filldir, info->extra_fname))
466                         goto finished;
467                 info->extra_fname = NULL;
468                 goto next_node;
469         } else if (!info->curr_node)
470                 info->curr_node = rb_first(&info->root);
471
472         while (1) {
473                 /*
474                  * Fill the rbtree if we have no more entries,
475                  * or the inode has changed since we last read in the
476                  * cached entries.
477                  */
478                 if ((!info->curr_node) ||
479                     (filp->f_version != inode->i_version)) {
480                         info->curr_node = NULL;
481                         free_rb_tree_fname(&info->root);
482                         filp->f_version = inode->i_version;
483                         ret = ext4_htree_fill_tree(filp, info->curr_hash,
484                                                    info->curr_minor_hash,
485                                                    &info->next_hash);
486                         if (ret < 0)
487                                 return ret;
488                         if (ret == 0) {
489                                 filp->f_pos = EXT4_HTREE_EOF;
490                                 break;
491                         }
492                         info->curr_node = rb_first(&info->root);
493                 }
494
495                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
496                 info->curr_hash = fname->hash;
497                 info->curr_minor_hash = fname->minor_hash;
498                 if (call_filldir(filp, dirent, filldir, fname))
499                         break;
500         next_node:
501                 info->curr_node = rb_next(info->curr_node);
502                 if (info->curr_node) {
503                         fname = rb_entry(info->curr_node, struct fname,
504                                          rb_hash);
505                         info->curr_hash = fname->hash;
506                         info->curr_minor_hash = fname->minor_hash;
507                 } else {
508                         if (info->next_hash == ~0) {
509                                 filp->f_pos = EXT4_HTREE_EOF;
510                                 break;
511                         }
512                         info->curr_hash = info->next_hash;
513                         info->curr_minor_hash = 0;
514                 }
515         }
516 finished:
517         info->last_pos = filp->f_pos;
518         return 0;
519 }
520
521 static int ext4_release_dir(struct inode *inode, struct file *filp)
522 {
523         if (filp->private_data)
524                 ext4_htree_free_dir_info(filp->private_data);
525
526         return 0;
527 }