nfsd: only register cld pipe notifier when CONFIG_NFSD_V4 is enabled
[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_dx_readdir(struct file *filp,
36                            void *dirent, filldir_t filldir);
37
38 static unsigned char get_dtype(struct super_block *sb, int filetype)
39 {
40         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_FILETYPE) ||
41             (filetype >= EXT4_FT_MAX))
42                 return DT_UNKNOWN;
43
44         return (ext4_filetype_table[filetype]);
45 }
46
47 /**
48  * Check if the given dir-inode refers to an htree-indexed directory
49  * (or a directory which chould potentially get coverted to use htree
50  * indexing).
51  *
52  * Return 1 if it is a dx dir, 0 if not
53  */
54 static int is_dx_dir(struct inode *inode)
55 {
56         struct super_block *sb = inode->i_sb;
57
58         if (EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
59                      EXT4_FEATURE_COMPAT_DIR_INDEX) &&
60             ((ext4_test_inode_flag(inode, EXT4_INODE_INDEX)) ||
61              ((inode->i_size >> sb->s_blocksize_bits) == 1)))
62                 return 1;
63
64         return 0;
65 }
66
67 /*
68  * Return 0 if the directory entry is OK, and 1 if there is a problem
69  *
70  * Note: this is the opposite of what ext2 and ext3 historically returned...
71  */
72 int __ext4_check_dir_entry(const char *function, unsigned int line,
73                            struct inode *dir, struct file *filp,
74                            struct ext4_dir_entry_2 *de,
75                            struct buffer_head *bh,
76                            unsigned int offset)
77 {
78         const char *error_msg = NULL;
79         const int rlen = ext4_rec_len_from_disk(de->rec_len,
80                                                 dir->i_sb->s_blocksize);
81
82         if (unlikely(rlen < EXT4_DIR_REC_LEN(1)))
83                 error_msg = "rec_len is smaller than minimal";
84         else if (unlikely(rlen % 4 != 0))
85                 error_msg = "rec_len % 4 != 0";
86         else if (unlikely(rlen < EXT4_DIR_REC_LEN(de->name_len)))
87                 error_msg = "rec_len is too small for name_len";
88         else if (unlikely(((char *) de - bh->b_data) + rlen >
89                           dir->i_sb->s_blocksize))
90                 error_msg = "directory entry across blocks";
91         else if (unlikely(le32_to_cpu(de->inode) >
92                         le32_to_cpu(EXT4_SB(dir->i_sb)->s_es->s_inodes_count)))
93                 error_msg = "inode out of bounds";
94         else
95                 return 0;
96
97         if (filp)
98                 ext4_error_file(filp, function, line, bh ? bh->b_blocknr : 0,
99                                 "bad entry in directory: %s - offset=%u(%u), "
100                                 "inode=%u, rec_len=%d, name_len=%d",
101                                 error_msg, (unsigned) (offset%bh->b_size),
102                                 offset, le32_to_cpu(de->inode),
103                                 rlen, de->name_len);
104         else
105                 ext4_error_inode(dir, function, line, bh ? bh->b_blocknr : 0,
106                                 "bad entry in directory: %s - offset=%u(%u), "
107                                 "inode=%u, rec_len=%d, name_len=%d",
108                                 error_msg, (unsigned) (offset%bh->b_size),
109                                 offset, le32_to_cpu(de->inode),
110                                 rlen, de->name_len);
111
112         return 1;
113 }
114
115 static int ext4_readdir(struct file *filp,
116                          void *dirent, filldir_t filldir)
117 {
118         int error = 0;
119         unsigned int offset;
120         int i, stored;
121         struct ext4_dir_entry_2 *de;
122         int err;
123         struct inode *inode = filp->f_path.dentry->d_inode;
124         struct super_block *sb = inode->i_sb;
125         int ret = 0;
126         int dir_has_error = 0;
127
128         if (is_dx_dir(inode)) {
129                 err = ext4_dx_readdir(filp, dirent, filldir);
130                 if (err != ERR_BAD_DX_DIR) {
131                         ret = err;
132                         goto out;
133                 }
134                 /*
135                  * We don't set the inode dirty flag since it's not
136                  * critical that it get flushed back to the disk.
137                  */
138                 ext4_clear_inode_flag(filp->f_path.dentry->d_inode,
139                                       EXT4_INODE_INDEX);
140         }
141         stored = 0;
142         offset = filp->f_pos & (sb->s_blocksize - 1);
143
144         while (!error && !stored && filp->f_pos < inode->i_size) {
145                 struct ext4_map_blocks map;
146                 struct buffer_head *bh = NULL;
147
148                 map.m_lblk = filp->f_pos >> EXT4_BLOCK_SIZE_BITS(sb);
149                 map.m_len = 1;
150                 err = ext4_map_blocks(NULL, inode, &map, 0);
151                 if (err > 0) {
152                         pgoff_t index = map.m_pblk >>
153                                         (PAGE_CACHE_SHIFT - inode->i_blkbits);
154                         if (!ra_has_index(&filp->f_ra, index))
155                                 page_cache_sync_readahead(
156                                         sb->s_bdev->bd_inode->i_mapping,
157                                         &filp->f_ra, filp,
158                                         index, 1);
159                         filp->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
160                         bh = ext4_bread(NULL, inode, map.m_lblk, 0, &err);
161                 }
162
163                 /*
164                  * We ignore I/O errors on directories so users have a chance
165                  * of recovering data when there's a bad sector
166                  */
167                 if (!bh) {
168                         if (!dir_has_error) {
169                                 EXT4_ERROR_FILE(filp, 0,
170                                                 "directory contains a "
171                                                 "hole at offset %llu",
172                                            (unsigned long long) filp->f_pos);
173                                 dir_has_error = 1;
174                         }
175                         /* corrupt size?  Maybe no more blocks to read */
176                         if (filp->f_pos > inode->i_blocks << 9)
177                                 break;
178                         filp->f_pos += sb->s_blocksize - offset;
179                         continue;
180                 }
181
182 revalidate:
183                 /* If the dir block has changed since the last call to
184                  * readdir(2), then we might be pointing to an invalid
185                  * dirent right now.  Scan from the start of the block
186                  * to make sure. */
187                 if (filp->f_version != inode->i_version) {
188                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
189                                 de = (struct ext4_dir_entry_2 *)
190                                         (bh->b_data + i);
191                                 /* It's too expensive to do a full
192                                  * dirent test each time round this
193                                  * loop, but we do have to test at
194                                  * least that it is non-zero.  A
195                                  * failure will be detected in the
196                                  * dirent test below. */
197                                 if (ext4_rec_len_from_disk(de->rec_len,
198                                         sb->s_blocksize) < EXT4_DIR_REC_LEN(1))
199                                         break;
200                                 i += ext4_rec_len_from_disk(de->rec_len,
201                                                             sb->s_blocksize);
202                         }
203                         offset = i;
204                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
205                                 | offset;
206                         filp->f_version = inode->i_version;
207                 }
208
209                 while (!error && filp->f_pos < inode->i_size
210                        && offset < sb->s_blocksize) {
211                         de = (struct ext4_dir_entry_2 *) (bh->b_data + offset);
212                         if (ext4_check_dir_entry(inode, filp, de,
213                                                  bh, offset)) {
214                                 /*
215                                  * On error, skip the f_pos to the next block
216                                  */
217                                 filp->f_pos = (filp->f_pos |
218                                                 (sb->s_blocksize - 1)) + 1;
219                                 brelse(bh);
220                                 ret = stored;
221                                 goto out;
222                         }
223                         offset += ext4_rec_len_from_disk(de->rec_len,
224                                         sb->s_blocksize);
225                         if (le32_to_cpu(de->inode)) {
226                                 /* We might block in the next section
227                                  * if the data destination is
228                                  * currently swapped out.  So, use a
229                                  * version stamp to detect whether or
230                                  * not the directory has been modified
231                                  * during the copy operation.
232                                  */
233                                 u64 version = filp->f_version;
234
235                                 error = filldir(dirent, de->name,
236                                                 de->name_len,
237                                                 filp->f_pos,
238                                                 le32_to_cpu(de->inode),
239                                                 get_dtype(sb, de->file_type));
240                                 if (error)
241                                         break;
242                                 if (version != filp->f_version)
243                                         goto revalidate;
244                                 stored++;
245                         }
246                         filp->f_pos += ext4_rec_len_from_disk(de->rec_len,
247                                                 sb->s_blocksize);
248                 }
249                 offset = 0;
250                 brelse(bh);
251         }
252 out:
253         return ret;
254 }
255
256 static inline int is_32bit_api(void)
257 {
258 #ifdef CONFIG_COMPAT
259         return is_compat_task();
260 #else
261         return (BITS_PER_LONG == 32);
262 #endif
263 }
264
265 /*
266  * These functions convert from the major/minor hash to an f_pos
267  * value for dx directories
268  *
269  * Upper layer (for example NFS) should specify FMODE_32BITHASH or
270  * FMODE_64BITHASH explicitly. On the other hand, we allow ext4 to be mounted
271  * directly on both 32-bit and 64-bit nodes, under such case, neither
272  * FMODE_32BITHASH nor FMODE_64BITHASH is specified.
273  */
274 static inline loff_t hash2pos(struct file *filp, __u32 major, __u32 minor)
275 {
276         if ((filp->f_mode & FMODE_32BITHASH) ||
277             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
278                 return major >> 1;
279         else
280                 return ((__u64)(major >> 1) << 32) | (__u64)minor;
281 }
282
283 static inline __u32 pos2maj_hash(struct file *filp, loff_t pos)
284 {
285         if ((filp->f_mode & FMODE_32BITHASH) ||
286             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
287                 return (pos << 1) & 0xffffffff;
288         else
289                 return ((pos >> 32) << 1) & 0xffffffff;
290 }
291
292 static inline __u32 pos2min_hash(struct file *filp, loff_t pos)
293 {
294         if ((filp->f_mode & FMODE_32BITHASH) ||
295             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
296                 return 0;
297         else
298                 return pos & 0xffffffff;
299 }
300
301 /*
302  * Return 32- or 64-bit end-of-file for dx directories
303  */
304 static inline loff_t ext4_get_htree_eof(struct file *filp)
305 {
306         if ((filp->f_mode & FMODE_32BITHASH) ||
307             (!(filp->f_mode & FMODE_64BITHASH) && is_32bit_api()))
308                 return EXT4_HTREE_EOF_32BIT;
309         else
310                 return EXT4_HTREE_EOF_64BIT;
311 }
312
313
314 /*
315  * ext4_dir_llseek() based on generic_file_llseek() to handle both
316  * non-htree and htree directories, where the "offset" is in terms
317  * of the filename hash value instead of the byte offset.
318  *
319  * NOTE: offsets obtained *before* ext4_set_inode_flag(dir, EXT4_INODE_INDEX)
320  *       will be invalid once the directory was converted into a dx directory
321  */
322 loff_t ext4_dir_llseek(struct file *file, loff_t offset, int origin)
323 {
324         struct inode *inode = file->f_mapping->host;
325         loff_t ret = -EINVAL;
326         int dx_dir = is_dx_dir(inode);
327
328         mutex_lock(&inode->i_mutex);
329
330         /* NOTE: relative offsets with dx directories might not work
331          *       as expected, as it is difficult to figure out the
332          *       correct offset between dx hashes */
333
334         switch (origin) {
335         case SEEK_END:
336                 if (unlikely(offset > 0))
337                         goto out_err; /* not supported for directories */
338
339                 /* so only negative offsets are left, does that have a
340                  * meaning for directories at all? */
341                 if (dx_dir)
342                         offset += ext4_get_htree_eof(file);
343                 else
344                         offset += inode->i_size;
345                 break;
346         case SEEK_CUR:
347                 /*
348                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
349                  * position-querying operation.  Avoid rewriting the "same"
350                  * f_pos value back to the file because a concurrent read(),
351                  * write() or lseek() might have altered it
352                  */
353                 if (offset == 0) {
354                         offset = file->f_pos;
355                         goto out_ok;
356                 }
357
358                 offset += file->f_pos;
359                 break;
360         }
361
362         if (unlikely(offset < 0))
363                 goto out_err;
364
365         if (!dx_dir) {
366                 if (offset > inode->i_sb->s_maxbytes)
367                         goto out_err;
368         } else if (offset > ext4_get_htree_eof(file))
369                 goto out_err;
370
371         /* Special lock needed here? */
372         if (offset != file->f_pos) {
373                 file->f_pos = offset;
374                 file->f_version = 0;
375         }
376
377 out_ok:
378         ret = offset;
379 out_err:
380         mutex_unlock(&inode->i_mutex);
381
382         return ret;
383 }
384
385 /*
386  * This structure holds the nodes of the red-black tree used to store
387  * the directory entry in hash order.
388  */
389 struct fname {
390         __u32           hash;
391         __u32           minor_hash;
392         struct rb_node  rb_hash;
393         struct fname    *next;
394         __u32           inode;
395         __u8            name_len;
396         __u8            file_type;
397         char            name[0];
398 };
399
400 /*
401  * This functoin implements a non-recursive way of freeing all of the
402  * nodes in the red-black tree.
403  */
404 static void free_rb_tree_fname(struct rb_root *root)
405 {
406         struct rb_node  *n = root->rb_node;
407         struct rb_node  *parent;
408         struct fname    *fname;
409
410         while (n) {
411                 /* Do the node's children first */
412                 if (n->rb_left) {
413                         n = n->rb_left;
414                         continue;
415                 }
416                 if (n->rb_right) {
417                         n = n->rb_right;
418                         continue;
419                 }
420                 /*
421                  * The node has no children; free it, and then zero
422                  * out parent's link to it.  Finally go to the
423                  * beginning of the loop and try to free the parent
424                  * node.
425                  */
426                 parent = rb_parent(n);
427                 fname = rb_entry(n, struct fname, rb_hash);
428                 while (fname) {
429                         struct fname *old = fname;
430                         fname = fname->next;
431                         kfree(old);
432                 }
433                 if (!parent)
434                         *root = RB_ROOT;
435                 else if (parent->rb_left == n)
436                         parent->rb_left = NULL;
437                 else if (parent->rb_right == n)
438                         parent->rb_right = NULL;
439                 n = parent;
440         }
441 }
442
443
444 static struct dir_private_info *ext4_htree_create_dir_info(struct file *filp,
445                                                            loff_t pos)
446 {
447         struct dir_private_info *p;
448
449         p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
450         if (!p)
451                 return NULL;
452         p->curr_hash = pos2maj_hash(filp, pos);
453         p->curr_minor_hash = pos2min_hash(filp, pos);
454         return p;
455 }
456
457 void ext4_htree_free_dir_info(struct dir_private_info *p)
458 {
459         free_rb_tree_fname(&p->root);
460         kfree(p);
461 }
462
463 /*
464  * Given a directory entry, enter it into the fname rb tree.
465  */
466 int ext4_htree_store_dirent(struct file *dir_file, __u32 hash,
467                              __u32 minor_hash,
468                              struct ext4_dir_entry_2 *dirent)
469 {
470         struct rb_node **p, *parent = NULL;
471         struct fname *fname, *new_fn;
472         struct dir_private_info *info;
473         int len;
474
475         info = dir_file->private_data;
476         p = &info->root.rb_node;
477
478         /* Create and allocate the fname structure */
479         len = sizeof(struct fname) + dirent->name_len + 1;
480         new_fn = kzalloc(len, GFP_KERNEL);
481         if (!new_fn)
482                 return -ENOMEM;
483         new_fn->hash = hash;
484         new_fn->minor_hash = minor_hash;
485         new_fn->inode = le32_to_cpu(dirent->inode);
486         new_fn->name_len = dirent->name_len;
487         new_fn->file_type = dirent->file_type;
488         memcpy(new_fn->name, dirent->name, dirent->name_len);
489         new_fn->name[dirent->name_len] = 0;
490
491         while (*p) {
492                 parent = *p;
493                 fname = rb_entry(parent, struct fname, rb_hash);
494
495                 /*
496                  * If the hash and minor hash match up, then we put
497                  * them on a linked list.  This rarely happens...
498                  */
499                 if ((new_fn->hash == fname->hash) &&
500                     (new_fn->minor_hash == fname->minor_hash)) {
501                         new_fn->next = fname->next;
502                         fname->next = new_fn;
503                         return 0;
504                 }
505
506                 if (new_fn->hash < fname->hash)
507                         p = &(*p)->rb_left;
508                 else if (new_fn->hash > fname->hash)
509                         p = &(*p)->rb_right;
510                 else if (new_fn->minor_hash < fname->minor_hash)
511                         p = &(*p)->rb_left;
512                 else /* if (new_fn->minor_hash > fname->minor_hash) */
513                         p = &(*p)->rb_right;
514         }
515
516         rb_link_node(&new_fn->rb_hash, parent, p);
517         rb_insert_color(&new_fn->rb_hash, &info->root);
518         return 0;
519 }
520
521
522
523 /*
524  * This is a helper function for ext4_dx_readdir.  It calls filldir
525  * for all entres on the fname linked list.  (Normally there is only
526  * one entry on the linked list, unless there are 62 bit hash collisions.)
527  */
528 static int call_filldir(struct file *filp, void *dirent,
529                         filldir_t filldir, struct fname *fname)
530 {
531         struct dir_private_info *info = filp->private_data;
532         loff_t  curr_pos;
533         struct inode *inode = filp->f_path.dentry->d_inode;
534         struct super_block *sb;
535         int error;
536
537         sb = inode->i_sb;
538
539         if (!fname) {
540                 printk(KERN_ERR "EXT4-fs: call_filldir: called with "
541                        "null fname?!?\n");
542                 return 0;
543         }
544         curr_pos = hash2pos(filp, fname->hash, fname->minor_hash);
545         while (fname) {
546                 error = filldir(dirent, fname->name,
547                                 fname->name_len, curr_pos,
548                                 fname->inode,
549                                 get_dtype(sb, fname->file_type));
550                 if (error) {
551                         filp->f_pos = curr_pos;
552                         info->extra_fname = fname;
553                         return error;
554                 }
555                 fname = fname->next;
556         }
557         return 0;
558 }
559
560 static int ext4_dx_readdir(struct file *filp,
561                          void *dirent, filldir_t filldir)
562 {
563         struct dir_private_info *info = filp->private_data;
564         struct inode *inode = filp->f_path.dentry->d_inode;
565         struct fname *fname;
566         int     ret;
567
568         if (!info) {
569                 info = ext4_htree_create_dir_info(filp, filp->f_pos);
570                 if (!info)
571                         return -ENOMEM;
572                 filp->private_data = info;
573         }
574
575         if (filp->f_pos == ext4_get_htree_eof(filp))
576                 return 0;       /* EOF */
577
578         /* Some one has messed with f_pos; reset the world */
579         if (info->last_pos != filp->f_pos) {
580                 free_rb_tree_fname(&info->root);
581                 info->curr_node = NULL;
582                 info->extra_fname = NULL;
583                 info->curr_hash = pos2maj_hash(filp, filp->f_pos);
584                 info->curr_minor_hash = pos2min_hash(filp, filp->f_pos);
585         }
586
587         /*
588          * If there are any leftover names on the hash collision
589          * chain, return them first.
590          */
591         if (info->extra_fname) {
592                 if (call_filldir(filp, dirent, filldir, info->extra_fname))
593                         goto finished;
594                 info->extra_fname = NULL;
595                 goto next_node;
596         } else if (!info->curr_node)
597                 info->curr_node = rb_first(&info->root);
598
599         while (1) {
600                 /*
601                  * Fill the rbtree if we have no more entries,
602                  * or the inode has changed since we last read in the
603                  * cached entries.
604                  */
605                 if ((!info->curr_node) ||
606                     (filp->f_version != inode->i_version)) {
607                         info->curr_node = NULL;
608                         free_rb_tree_fname(&info->root);
609                         filp->f_version = inode->i_version;
610                         ret = ext4_htree_fill_tree(filp, info->curr_hash,
611                                                    info->curr_minor_hash,
612                                                    &info->next_hash);
613                         if (ret < 0)
614                                 return ret;
615                         if (ret == 0) {
616                                 filp->f_pos = ext4_get_htree_eof(filp);
617                                 break;
618                         }
619                         info->curr_node = rb_first(&info->root);
620                 }
621
622                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
623                 info->curr_hash = fname->hash;
624                 info->curr_minor_hash = fname->minor_hash;
625                 if (call_filldir(filp, dirent, filldir, fname))
626                         break;
627         next_node:
628                 info->curr_node = rb_next(info->curr_node);
629                 if (info->curr_node) {
630                         fname = rb_entry(info->curr_node, struct fname,
631                                          rb_hash);
632                         info->curr_hash = fname->hash;
633                         info->curr_minor_hash = fname->minor_hash;
634                 } else {
635                         if (info->next_hash == ~0) {
636                                 filp->f_pos = ext4_get_htree_eof(filp);
637                                 break;
638                         }
639                         info->curr_hash = info->next_hash;
640                         info->curr_minor_hash = 0;
641                 }
642         }
643 finished:
644         info->last_pos = filp->f_pos;
645         return 0;
646 }
647
648 static int ext4_release_dir(struct inode *inode, struct file *filp)
649 {
650         if (filp->private_data)
651                 ext4_htree_free_dir_info(filp->private_data);
652
653         return 0;
654 }
655
656 const struct file_operations ext4_dir_operations = {
657         .llseek         = ext4_dir_llseek,
658         .read           = generic_read_dir,
659         .readdir        = ext4_readdir,
660         .unlocked_ioctl = ext4_ioctl,
661 #ifdef CONFIG_COMPAT
662         .compat_ioctl   = ext4_compat_ioctl,
663 #endif
664         .fsync          = ext4_sync_file,
665         .release        = ext4_release_dir,
666 };