[GFS2] Fix ref count bug that used to bite us on umount
[linux-2.6] / fs / ext3 / dir.c
1 /*
2  *  linux/fs/ext3/dir.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/dir.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  ext3 directory handling functions
16  *
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *
20  * Hash Tree Directory indexing (c) 2001  Daniel Phillips
21  *
22  */
23
24 #include <linux/fs.h>
25 #include <linux/jbd.h>
26 #include <linux/ext3_fs.h>
27 #include <linux/buffer_head.h>
28 #include <linux/smp_lock.h>
29 #include <linux/slab.h>
30 #include <linux/rbtree.h>
31
32 static unsigned char ext3_filetype_table[] = {
33         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
34 };
35
36 static int ext3_readdir(struct file *, void *, filldir_t);
37 static int ext3_dx_readdir(struct file * filp,
38                            void * dirent, filldir_t filldir);
39 static int ext3_release_dir (struct inode * inode,
40                                 struct file * filp);
41
42 const struct file_operations ext3_dir_operations = {
43         .llseek         = generic_file_llseek,
44         .read           = generic_read_dir,
45         .readdir        = ext3_readdir,         /* we take BKL. needed?*/
46         .ioctl          = ext3_ioctl,           /* BKL held */
47         .fsync          = ext3_sync_file,       /* BKL held */
48 #ifdef CONFIG_EXT3_INDEX
49         .release        = ext3_release_dir,
50 #endif
51 };
52
53
54 static unsigned char get_dtype(struct super_block *sb, int filetype)
55 {
56         if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_FILETYPE) ||
57             (filetype >= EXT3_FT_MAX))
58                 return DT_UNKNOWN;
59
60         return (ext3_filetype_table[filetype]);
61 }
62                                
63
64 int ext3_check_dir_entry (const char * function, struct inode * dir,
65                           struct ext3_dir_entry_2 * de,
66                           struct buffer_head * bh,
67                           unsigned long offset)
68 {
69         const char * error_msg = NULL;
70         const int rlen = le16_to_cpu(de->rec_len);
71
72         if (rlen < EXT3_DIR_REC_LEN(1))
73                 error_msg = "rec_len is smaller than minimal";
74         else if (rlen % 4 != 0)
75                 error_msg = "rec_len % 4 != 0";
76         else if (rlen < EXT3_DIR_REC_LEN(de->name_len))
77                 error_msg = "rec_len is too small for name_len";
78         else if (((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize)
79                 error_msg = "directory entry across blocks";
80         else if (le32_to_cpu(de->inode) >
81                         le32_to_cpu(EXT3_SB(dir->i_sb)->s_es->s_inodes_count))
82                 error_msg = "inode out of bounds";
83
84         if (error_msg != NULL)
85                 ext3_error (dir->i_sb, function,
86                         "bad entry in directory #%lu: %s - "
87                         "offset=%lu, inode=%lu, rec_len=%d, name_len=%d",
88                         dir->i_ino, error_msg, offset,
89                         (unsigned long) le32_to_cpu(de->inode),
90                         rlen, de->name_len);
91         return error_msg == NULL ? 1 : 0;
92 }
93
94 static int ext3_readdir(struct file * filp,
95                          void * dirent, filldir_t filldir)
96 {
97         int error = 0;
98         unsigned long offset;
99         int i, stored;
100         struct ext3_dir_entry_2 *de;
101         struct super_block *sb;
102         int err;
103         struct inode *inode = filp->f_dentry->d_inode;
104         int ret = 0;
105
106         sb = inode->i_sb;
107
108 #ifdef CONFIG_EXT3_INDEX
109         if (EXT3_HAS_COMPAT_FEATURE(inode->i_sb,
110                                     EXT3_FEATURE_COMPAT_DIR_INDEX) &&
111             ((EXT3_I(inode)->i_flags & EXT3_INDEX_FL) ||
112              ((inode->i_size >> sb->s_blocksize_bits) == 1))) {
113                 err = ext3_dx_readdir(filp, dirent, filldir);
114                 if (err != ERR_BAD_DX_DIR) {
115                         ret = err;
116                         goto out;
117                 }
118                 /*
119                  * We don't set the inode dirty flag since it's not
120                  * critical that it get flushed back to the disk.
121                  */
122                 EXT3_I(filp->f_dentry->d_inode)->i_flags &= ~EXT3_INDEX_FL;
123         }
124 #endif
125         stored = 0;
126         offset = filp->f_pos & (sb->s_blocksize - 1);
127
128         while (!error && !stored && filp->f_pos < inode->i_size) {
129                 unsigned long blk = filp->f_pos >> EXT3_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 = ext3_get_blocks_handle(NULL, inode, blk, 1,
135                                                 &map_bh, 0, 0);
136                 if (err > 0) {
137                         page_cache_readahead(sb->s_bdev->bd_inode->i_mapping,
138                                 &filp->f_ra,
139                                 filp,
140                                 map_bh.b_blocknr >>
141                                         (PAGE_CACHE_SHIFT - inode->i_blkbits),
142                                 1);
143                         bh = ext3_bread(NULL, inode, blk, 0, &err);
144                 }
145
146                 /*
147                  * We ignore I/O errors on directories so users have a chance
148                  * of recovering data when there's a bad sector
149                  */
150                 if (!bh) {
151                         ext3_error (sb, "ext3_readdir",
152                                 "directory #%lu contains a hole at offset %lu",
153                                 inode->i_ino, (unsigned long)filp->f_pos);
154                         filp->f_pos += sb->s_blocksize - offset;
155                         continue;
156                 }
157
158 revalidate:
159                 /* If the dir block has changed since the last call to
160                  * readdir(2), then we might be pointing to an invalid
161                  * dirent right now.  Scan from the start of the block
162                  * to make sure. */
163                 if (filp->f_version != inode->i_version) {
164                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
165                                 de = (struct ext3_dir_entry_2 *) 
166                                         (bh->b_data + i);
167                                 /* It's too expensive to do a full
168                                  * dirent test each time round this
169                                  * loop, but we do have to test at
170                                  * least that it is non-zero.  A
171                                  * failure will be detected in the
172                                  * dirent test below. */
173                                 if (le16_to_cpu(de->rec_len) <
174                                                 EXT3_DIR_REC_LEN(1))
175                                         break;
176                                 i += le16_to_cpu(de->rec_len);
177                         }
178                         offset = i;
179                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
180                                 | offset;
181                         filp->f_version = inode->i_version;
182                 }
183
184                 while (!error && filp->f_pos < inode->i_size 
185                        && offset < sb->s_blocksize) {
186                         de = (struct ext3_dir_entry_2 *) (bh->b_data + offset);
187                         if (!ext3_check_dir_entry ("ext3_readdir", inode, de,
188                                                    bh, offset)) {
189                                 /* On error, skip the f_pos to the
190                                    next block. */
191                                 filp->f_pos = (filp->f_pos |
192                                                 (sb->s_blocksize - 1)) + 1;
193                                 brelse (bh);
194                                 ret = stored;
195                                 goto out;
196                         }
197                         offset += le16_to_cpu(de->rec_len);
198                         if (le32_to_cpu(de->inode)) {
199                                 /* We might block in the next section
200                                  * if the data destination is
201                                  * currently swapped out.  So, use a
202                                  * version stamp to detect whether or
203                                  * not the directory has been modified
204                                  * during the copy operation.
205                                  */
206                                 unsigned long version = filp->f_version;
207
208                                 error = filldir(dirent, de->name,
209                                                 de->name_len,
210                                                 filp->f_pos,
211                                                 le32_to_cpu(de->inode),
212                                                 get_dtype(sb, de->file_type));
213                                 if (error)
214                                         break;
215                                 if (version != filp->f_version)
216                                         goto revalidate;
217                                 stored ++;
218                         }
219                         filp->f_pos += le16_to_cpu(de->rec_len);
220                 }
221                 offset = 0;
222                 brelse (bh);
223         }
224 out:
225         return ret;
226 }
227
228 #ifdef CONFIG_EXT3_INDEX
229 /*
230  * These functions convert from the major/minor hash to an f_pos
231  * value.
232  * 
233  * Currently we only use major hash numer.  This is unfortunate, but
234  * on 32-bit machines, the same VFS interface is used for lseek and
235  * llseek, so if we use the 64 bit offset, then the 32-bit versions of
236  * lseek/telldir/seekdir will blow out spectacularly, and from within
237  * the ext2 low-level routine, we don't know if we're being called by
238  * a 64-bit version of the system call or the 32-bit version of the
239  * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
240  * cookie.  Sigh.
241  */
242 #define hash2pos(major, minor)  (major >> 1)
243 #define pos2maj_hash(pos)       ((pos << 1) & 0xffffffff)
244 #define pos2min_hash(pos)       (0)
245
246 /*
247  * This structure holds the nodes of the red-black tree used to store
248  * the directory entry in hash order.
249  */
250 struct fname {
251         __u32           hash;
252         __u32           minor_hash;
253         struct rb_node  rb_hash; 
254         struct fname    *next;
255         __u32           inode;
256         __u8            name_len;
257         __u8            file_type;
258         char            name[0];
259 };
260
261 /*
262  * This functoin implements a non-recursive way of freeing all of the
263  * nodes in the red-black tree.
264  */
265 static void free_rb_tree_fname(struct rb_root *root)
266 {
267         struct rb_node  *n = root->rb_node;
268         struct rb_node  *parent;
269         struct fname    *fname;
270
271         while (n) {
272                 /* Do the node's children first */
273                 if ((n)->rb_left) {
274                         n = n->rb_left;
275                         continue;
276                 }
277                 if (n->rb_right) {
278                         n = n->rb_right;
279                         continue;
280                 }
281                 /*
282                  * The node has no children; free it, and then zero
283                  * out parent's link to it.  Finally go to the
284                  * beginning of the loop and try to free the parent
285                  * node.
286                  */
287                 parent = n->rb_parent;
288                 fname = rb_entry(n, struct fname, rb_hash);
289                 while (fname) {
290                         struct fname * old = fname;
291                         fname = fname->next;
292                         kfree (old);
293                 }
294                 if (!parent)
295                         root->rb_node = NULL;
296                 else if (parent->rb_left == n)
297                         parent->rb_left = NULL;
298                 else if (parent->rb_right == n)
299                         parent->rb_right = NULL;
300                 n = parent;
301         }
302         root->rb_node = NULL;
303 }
304
305
306 static struct dir_private_info *create_dir_info(loff_t pos)
307 {
308         struct dir_private_info *p;
309
310         p = kmalloc(sizeof(struct dir_private_info), GFP_KERNEL);
311         if (!p)
312                 return NULL;
313         p->root.rb_node = NULL;
314         p->curr_node = NULL;
315         p->extra_fname = NULL;
316         p->last_pos = 0;
317         p->curr_hash = pos2maj_hash(pos);
318         p->curr_minor_hash = pos2min_hash(pos);
319         p->next_hash = 0;
320         return p;
321 }
322
323 void ext3_htree_free_dir_info(struct dir_private_info *p)
324 {
325         free_rb_tree_fname(&p->root);
326         kfree(p);
327 }
328
329 /*
330  * Given a directory entry, enter it into the fname rb tree.
331  */
332 int ext3_htree_store_dirent(struct file *dir_file, __u32 hash,
333                              __u32 minor_hash,
334                              struct ext3_dir_entry_2 *dirent)
335 {
336         struct rb_node **p, *parent = NULL;
337         struct fname * fname, *new_fn;
338         struct dir_private_info *info;
339         int len;
340
341         info = (struct dir_private_info *) dir_file->private_data;
342         p = &info->root.rb_node;
343
344         /* Create and allocate the fname structure */
345         len = sizeof(struct fname) + dirent->name_len + 1;
346         new_fn = kmalloc(len, GFP_KERNEL);
347         if (!new_fn)
348                 return -ENOMEM;
349         memset(new_fn, 0, len);
350         new_fn->hash = hash;
351         new_fn->minor_hash = minor_hash;
352         new_fn->inode = le32_to_cpu(dirent->inode);
353         new_fn->name_len = dirent->name_len;
354         new_fn->file_type = dirent->file_type;
355         memcpy(new_fn->name, dirent->name, dirent->name_len);
356         new_fn->name[dirent->name_len] = 0;
357
358         while (*p) {
359                 parent = *p;
360                 fname = rb_entry(parent, struct fname, rb_hash);
361
362                 /*
363                  * If the hash and minor hash match up, then we put
364                  * them on a linked list.  This rarely happens...
365                  */
366                 if ((new_fn->hash == fname->hash) &&
367                     (new_fn->minor_hash == fname->minor_hash)) {
368                         new_fn->next = fname->next;
369                         fname->next = new_fn;
370                         return 0;
371                 }
372
373                 if (new_fn->hash < fname->hash)
374                         p = &(*p)->rb_left;
375                 else if (new_fn->hash > fname->hash)
376                         p = &(*p)->rb_right;
377                 else if (new_fn->minor_hash < fname->minor_hash)
378                         p = &(*p)->rb_left;
379                 else /* if (new_fn->minor_hash > fname->minor_hash) */
380                         p = &(*p)->rb_right;
381         }
382
383         rb_link_node(&new_fn->rb_hash, parent, p);
384         rb_insert_color(&new_fn->rb_hash, &info->root);
385         return 0;
386 }
387
388
389
390 /*
391  * This is a helper function for ext3_dx_readdir.  It calls filldir
392  * for all entres on the fname linked list.  (Normally there is only
393  * one entry on the linked list, unless there are 62 bit hash collisions.)
394  */
395 static int call_filldir(struct file * filp, void * dirent,
396                         filldir_t filldir, struct fname *fname)
397 {
398         struct dir_private_info *info = filp->private_data;
399         loff_t  curr_pos;
400         struct inode *inode = filp->f_dentry->d_inode;
401         struct super_block * sb;
402         int error;
403
404         sb = inode->i_sb;
405
406         if (!fname) {
407                 printk("call_filldir: called with null fname?!?\n");
408                 return 0;
409         }
410         curr_pos = hash2pos(fname->hash, fname->minor_hash);
411         while (fname) {
412                 error = filldir(dirent, fname->name,
413                                 fname->name_len, curr_pos, 
414                                 fname->inode,
415                                 get_dtype(sb, fname->file_type));
416                 if (error) {
417                         filp->f_pos = curr_pos;
418                         info->extra_fname = fname->next;
419                         return error;
420                 }
421                 fname = fname->next;
422         }
423         return 0;
424 }
425
426 static int ext3_dx_readdir(struct file * filp,
427                          void * dirent, filldir_t filldir)
428 {
429         struct dir_private_info *info = filp->private_data;
430         struct inode *inode = filp->f_dentry->d_inode;
431         struct fname *fname;
432         int     ret;
433
434         if (!info) {
435                 info = create_dir_info(filp->f_pos);
436                 if (!info)
437                         return -ENOMEM;
438                 filp->private_data = info;
439         }
440
441         if (filp->f_pos == EXT3_HTREE_EOF)
442                 return 0;       /* EOF */
443
444         /* Some one has messed with f_pos; reset the world */
445         if (info->last_pos != filp->f_pos) {
446                 free_rb_tree_fname(&info->root);
447                 info->curr_node = NULL;
448                 info->extra_fname = NULL;
449                 info->curr_hash = pos2maj_hash(filp->f_pos);
450                 info->curr_minor_hash = pos2min_hash(filp->f_pos);
451         }
452
453         /*
454          * If there are any leftover names on the hash collision
455          * chain, return them first.
456          */
457         if (info->extra_fname &&
458             call_filldir(filp, dirent, filldir, info->extra_fname))
459                 goto finished;
460
461         if (!info->curr_node)
462                 info->curr_node = rb_first(&info->root);
463
464         while (1) {
465                 /*
466                  * Fill the rbtree if we have no more entries,
467                  * or the inode has changed since we last read in the
468                  * cached entries. 
469                  */
470                 if ((!info->curr_node) ||
471                     (filp->f_version != inode->i_version)) {
472                         info->curr_node = NULL;
473                         free_rb_tree_fname(&info->root);
474                         filp->f_version = inode->i_version;
475                         ret = ext3_htree_fill_tree(filp, info->curr_hash,
476                                                    info->curr_minor_hash,
477                                                    &info->next_hash);
478                         if (ret < 0)
479                                 return ret;
480                         if (ret == 0) {
481                                 filp->f_pos = EXT3_HTREE_EOF;
482                                 break;
483                         }
484                         info->curr_node = rb_first(&info->root);
485                 }
486
487                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
488                 info->curr_hash = fname->hash;
489                 info->curr_minor_hash = fname->minor_hash;
490                 if (call_filldir(filp, dirent, filldir, fname))
491                         break;
492
493                 info->curr_node = rb_next(info->curr_node);
494                 if (!info->curr_node) {
495                         if (info->next_hash == ~0) {
496                                 filp->f_pos = EXT3_HTREE_EOF;
497                                 break;
498                         }
499                         info->curr_hash = info->next_hash;
500                         info->curr_minor_hash = 0;
501                 }
502         }
503 finished:
504         info->last_pos = filp->f_pos;
505         return 0;
506 }
507
508 static int ext3_release_dir (struct inode * inode, struct file * filp)
509 {
510        if (filp->private_data)
511                 ext3_htree_free_dir_info(filp->private_data);
512
513         return 0;
514 }
515
516 #endif