Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[linux-2.6] / 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 - "
88                         "offset=%u, inode=%u, rec_len=%d, name_len=%d",
89                         dir->i_ino, error_msg, offset,
90                         le32_to_cpu(de->inode),
91                         rlen, de->name_len);
92         return error_msg == NULL ? 1 : 0;
93 }
94
95 static int ext4_readdir(struct file *filp,
96                          void *dirent, filldir_t filldir)
97 {
98         int error = 0;
99         unsigned int offset;
100         int i, stored;
101         struct ext4_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 (EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
111                                     EXT4_FEATURE_COMPAT_DIR_INDEX) &&
112             ((EXT4_I(inode)->i_flags & EXT4_INDEX_FL) ||
113              ((inode->i_size >> sb->s_blocksize_bits) == 1))) {
114                 err = ext4_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                 EXT4_I(filp->f_path.dentry->d_inode)->i_flags &= ~EXT4_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                 ext4_lblk_t blk = filp->f_pos >> EXT4_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 = ext4_get_blocks_wrap(NULL, inode, blk, 1, &map_bh,
135                                                 0, 0, 0);
136                 if (err > 0) {
137                         pgoff_t index = map_bh.b_blocknr >>
138                                         (PAGE_CACHE_SHIFT - inode->i_blkbits);
139                         if (!ra_has_index(&filp->f_ra, index))
140                                 page_cache_sync_readahead(
141                                         sb->s_bdev->bd_inode->i_mapping,
142                                         &filp->f_ra, filp,
143                                         index, 1);
144                         filp->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
145                         bh = ext4_bread(NULL, inode, blk, 0, &err);
146                 }
147
148                 /*
149                  * We ignore I/O errors on directories so users have a chance
150                  * of recovering data when there's a bad sector
151                  */
152                 if (!bh) {
153                         if (!dir_has_error) {
154                                 ext4_error(sb, __func__, "directory #%lu "
155                                            "contains a hole at offset %Lu",
156                                            inode->i_ino,
157                                            (unsigned long long) filp->f_pos);
158                                 dir_has_error = 1;
159                         }
160                         /* corrupt size?  Maybe no more blocks to read */
161                         if (filp->f_pos > inode->i_blocks << 9)
162                                 break;
163                         filp->f_pos += sb->s_blocksize - offset;
164                         continue;
165                 }
166
167 revalidate:
168                 /* If the dir block has changed since the last call to
169                  * readdir(2), then we might be pointing to an invalid
170                  * dirent right now.  Scan from the start of the block
171                  * to make sure. */
172                 if (filp->f_version != inode->i_version) {
173                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
174                                 de = (struct ext4_dir_entry_2 *)
175                                         (bh->b_data + i);
176                                 /* It's too expensive to do a full
177                                  * dirent test each time round this
178                                  * loop, but we do have to test at
179                                  * least that it is non-zero.  A
180                                  * failure will be detected in the
181                                  * dirent test below. */
182                                 if (ext4_rec_len_from_disk(de->rec_len,
183                                         sb->s_blocksize) < EXT4_DIR_REC_LEN(1))
184                                         break;
185                                 i += ext4_rec_len_from_disk(de->rec_len,
186                                                             sb->s_blocksize);
187                         }
188                         offset = i;
189                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
190                                 | offset;
191                         filp->f_version = inode->i_version;
192                 }
193
194                 while (!error && filp->f_pos < inode->i_size
195                        && offset < sb->s_blocksize) {
196                         de = (struct ext4_dir_entry_2 *) (bh->b_data + offset);
197                         if (!ext4_check_dir_entry("ext4_readdir", inode, de,
198                                                   bh, offset)) {
199                                 /*
200                                  * On error, skip the f_pos to the next block
201                                  */
202                                 filp->f_pos = (filp->f_pos |
203                                                 (sb->s_blocksize - 1)) + 1;
204                                 brelse(bh);
205                                 ret = stored;
206                                 goto out;
207                         }
208                         offset += ext4_rec_len_from_disk(de->rec_len,
209                                         sb->s_blocksize);
210                         if (le32_to_cpu(de->inode)) {
211                                 /* We might block in the next section
212                                  * if the data destination is
213                                  * currently swapped out.  So, use a
214                                  * version stamp to detect whether or
215                                  * not the directory has been modified
216                                  * during the copy operation.
217                                  */
218                                 u64 version = filp->f_version;
219
220                                 error = filldir(dirent, de->name,
221                                                 de->name_len,
222                                                 filp->f_pos,
223                                                 le32_to_cpu(de->inode),
224                                                 get_dtype(sb, de->file_type));
225                                 if (error)
226                                         break;
227                                 if (version != filp->f_version)
228                                         goto revalidate;
229                                 stored++;
230                         }
231                         filp->f_pos += ext4_rec_len_from_disk(de->rec_len,
232                                                 sb->s_blocksize);
233                 }
234                 offset = 0;
235                 brelse(bh);
236         }
237 out:
238         return ret;
239 }
240
241 /*
242  * These functions convert from the major/minor hash to an f_pos
243  * value.
244  *
245  * Currently we only use major hash numer.  This is unfortunate, but
246  * on 32-bit machines, the same VFS interface is used for lseek and
247  * llseek, so if we use the 64 bit offset, then the 32-bit versions of
248  * lseek/telldir/seekdir will blow out spectacularly, and from within
249  * the ext2 low-level routine, we don't know if we're being called by
250  * a 64-bit version of the system call or the 32-bit version of the
251  * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
252  * cookie.  Sigh.
253  */
254 #define hash2pos(major, minor)  (major >> 1)
255 #define pos2maj_hash(pos)       ((pos << 1) & 0xffffffff)
256 #define pos2min_hash(pos)       (0)
257
258 /*
259  * This structure holds the nodes of the red-black tree used to store
260  * the directory entry in hash order.
261  */
262 struct fname {
263         __u32           hash;
264         __u32           minor_hash;
265         struct rb_node  rb_hash;
266         struct fname    *next;
267         __u32           inode;
268         __u8            name_len;
269         __u8            file_type;
270         char            name[0];
271 };
272
273 /*
274  * This functoin implements a non-recursive way of freeing all of the
275  * nodes in the red-black tree.
276  */
277 static void free_rb_tree_fname(struct rb_root *root)
278 {
279         struct rb_node  *n = root->rb_node;
280         struct rb_node  *parent;
281         struct fname    *fname;
282
283         while (n) {
284                 /* Do the node's children first */
285                 if (n->rb_left) {
286                         n = n->rb_left;
287                         continue;
288                 }
289                 if (n->rb_right) {
290                         n = n->rb_right;
291                         continue;
292                 }
293                 /*
294                  * The node has no children; free it, and then zero
295                  * out parent's link to it.  Finally go to the
296                  * beginning of the loop and try to free the parent
297                  * node.
298                  */
299                 parent = rb_parent(n);
300                 fname = rb_entry(n, struct fname, rb_hash);
301                 while (fname) {
302                         struct fname *old = fname;
303                         fname = fname->next;
304                         kfree(old);
305                 }
306                 if (!parent)
307                         root->rb_node = NULL;
308                 else if (parent->rb_left == n)
309                         parent->rb_left = NULL;
310                 else if (parent->rb_right == n)
311                         parent->rb_right = NULL;
312                 n = parent;
313         }
314 }
315
316
317 static struct dir_private_info *ext4_htree_create_dir_info(loff_t pos)
318 {
319         struct dir_private_info *p;
320
321         p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
322         if (!p)
323                 return NULL;
324         p->curr_hash = pos2maj_hash(pos);
325         p->curr_minor_hash = pos2min_hash(pos);
326         return p;
327 }
328
329 void ext4_htree_free_dir_info(struct dir_private_info *p)
330 {
331         free_rb_tree_fname(&p->root);
332         kfree(p);
333 }
334
335 /*
336  * Given a directory entry, enter it into the fname rb tree.
337  */
338 int ext4_htree_store_dirent(struct file *dir_file, __u32 hash,
339                              __u32 minor_hash,
340                              struct ext4_dir_entry_2 *dirent)
341 {
342         struct rb_node **p, *parent = NULL;
343         struct fname *fname, *new_fn;
344         struct dir_private_info *info;
345         int len;
346
347         info = (struct dir_private_info *) dir_file->private_data;
348         p = &info->root.rb_node;
349
350         /* Create and allocate the fname structure */
351         len = sizeof(struct fname) + dirent->name_len + 1;
352         new_fn = kzalloc(len, GFP_KERNEL);
353         if (!new_fn)
354                 return -ENOMEM;
355         new_fn->hash = hash;
356         new_fn->minor_hash = minor_hash;
357         new_fn->inode = le32_to_cpu(dirent->inode);
358         new_fn->name_len = dirent->name_len;
359         new_fn->file_type = dirent->file_type;
360         memcpy(new_fn->name, dirent->name, dirent->name_len);
361         new_fn->name[dirent->name_len] = 0;
362
363         while (*p) {
364                 parent = *p;
365                 fname = rb_entry(parent, struct fname, rb_hash);
366
367                 /*
368                  * If the hash and minor hash match up, then we put
369                  * them on a linked list.  This rarely happens...
370                  */
371                 if ((new_fn->hash == fname->hash) &&
372                     (new_fn->minor_hash == fname->minor_hash)) {
373                         new_fn->next = fname->next;
374                         fname->next = new_fn;
375                         return 0;
376                 }
377
378                 if (new_fn->hash < fname->hash)
379                         p = &(*p)->rb_left;
380                 else if (new_fn->hash > fname->hash)
381                         p = &(*p)->rb_right;
382                 else if (new_fn->minor_hash < fname->minor_hash)
383                         p = &(*p)->rb_left;
384                 else /* if (new_fn->minor_hash > fname->minor_hash) */
385                         p = &(*p)->rb_right;
386         }
387
388         rb_link_node(&new_fn->rb_hash, parent, p);
389         rb_insert_color(&new_fn->rb_hash, &info->root);
390         return 0;
391 }
392
393
394
395 /*
396  * This is a helper function for ext4_dx_readdir.  It calls filldir
397  * for all entres on the fname linked list.  (Normally there is only
398  * one entry on the linked list, unless there are 62 bit hash collisions.)
399  */
400 static int call_filldir(struct file *filp, void *dirent,
401                         filldir_t filldir, struct fname *fname)
402 {
403         struct dir_private_info *info = filp->private_data;
404         loff_t  curr_pos;
405         struct inode *inode = filp->f_path.dentry->d_inode;
406         struct super_block *sb;
407         int error;
408
409         sb = inode->i_sb;
410
411         if (!fname) {
412                 printk(KERN_ERR "EXT4-fs: call_filldir: called with "
413                        "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;
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 = ext4_htree_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                 if (call_filldir(filp, dirent, filldir, info->extra_fname))
465                         goto finished;
466                 info->extra_fname = NULL;
467                 goto next_node;
468         } else if (!info->curr_node)
469                 info->curr_node = rb_first(&info->root);
470
471         while (1) {
472                 /*
473                  * Fill the rbtree if we have no more entries,
474                  * or the inode has changed since we last read in the
475                  * cached entries.
476                  */
477                 if ((!info->curr_node) ||
478                     (filp->f_version != inode->i_version)) {
479                         info->curr_node = NULL;
480                         free_rb_tree_fname(&info->root);
481                         filp->f_version = inode->i_version;
482                         ret = ext4_htree_fill_tree(filp, info->curr_hash,
483                                                    info->curr_minor_hash,
484                                                    &info->next_hash);
485                         if (ret < 0)
486                                 return ret;
487                         if (ret == 0) {
488                                 filp->f_pos = EXT4_HTREE_EOF;
489                                 break;
490                         }
491                         info->curr_node = rb_first(&info->root);
492                 }
493
494                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
495                 info->curr_hash = fname->hash;
496                 info->curr_minor_hash = fname->minor_hash;
497                 if (call_filldir(filp, dirent, filldir, fname))
498                         break;
499         next_node:
500                 info->curr_node = rb_next(info->curr_node);
501                 if (info->curr_node) {
502                         fname = rb_entry(info->curr_node, struct fname,
503                                          rb_hash);
504                         info->curr_hash = fname->hash;
505                         info->curr_minor_hash = fname->minor_hash;
506                 } else {
507                         if (info->next_hash == ~0) {
508                                 filp->f_pos = EXT4_HTREE_EOF;
509                                 break;
510                         }
511                         info->curr_hash = info->next_hash;
512                         info->curr_minor_hash = 0;
513                 }
514         }
515 finished:
516         info->last_pos = filp->f_pos;
517         return 0;
518 }
519
520 static int ext4_release_dir(struct inode *inode, struct file *filp)
521 {
522         if (filp->private_data)
523                 ext4_htree_free_dir_info(filp->private_data);
524
525         return 0;
526 }