Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[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 long offset)
68 {
69         const char * error_msg = NULL;
70         const int rlen = ext4_rec_len_from_disk(de->rec_len);
71
72         if (rlen < EXT4_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 < EXT4_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(EXT4_SB(dir->i_sb)->s_es->s_inodes_count))
82                 error_msg = "inode out of bounds";
83
84         if (error_msg != NULL)
85                 ext4_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 ext4_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 ext4_dir_entry_2 *de;
101         struct super_block *sb;
102         int err;
103         struct inode *inode = filp->f_path.dentry->d_inode;
104         int ret = 0;
105
106         sb = inode->i_sb;
107
108         if (EXT4_HAS_COMPAT_FEATURE(inode->i_sb,
109                                     EXT4_FEATURE_COMPAT_DIR_INDEX) &&
110             ((EXT4_I(inode)->i_flags & EXT4_INDEX_FL) ||
111              ((inode->i_size >> sb->s_blocksize_bits) == 1))) {
112                 err = ext4_dx_readdir(filp, dirent, filldir);
113                 if (err != ERR_BAD_DX_DIR) {
114                         ret = err;
115                         goto out;
116                 }
117                 /*
118                  * We don't set the inode dirty flag since it's not
119                  * critical that it get flushed back to the disk.
120                  */
121                 EXT4_I(filp->f_path.dentry->d_inode)->i_flags &= ~EXT4_INDEX_FL;
122         }
123         stored = 0;
124         offset = filp->f_pos & (sb->s_blocksize - 1);
125
126         while (!error && !stored && filp->f_pos < inode->i_size) {
127                 ext4_lblk_t blk = filp->f_pos >> EXT4_BLOCK_SIZE_BITS(sb);
128                 struct buffer_head map_bh;
129                 struct buffer_head *bh = NULL;
130
131                 map_bh.b_state = 0;
132                 err = ext4_get_blocks_wrap(NULL, inode, blk, 1, &map_bh,
133                                                 0, 0, 0);
134                 if (err > 0) {
135                         pgoff_t index = map_bh.b_blocknr >>
136                                         (PAGE_CACHE_SHIFT - inode->i_blkbits);
137                         if (!ra_has_index(&filp->f_ra, index))
138                                 page_cache_sync_readahead(
139                                         sb->s_bdev->bd_inode->i_mapping,
140                                         &filp->f_ra, filp,
141                                         index, 1);
142                         filp->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
143                         bh = ext4_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                         ext4_error (sb, "ext4_readdir",
152                                 "directory #%lu contains a hole at offset %lu",
153                                 inode->i_ino, (unsigned long)filp->f_pos);
154                         /* corrupt size?  Maybe no more blocks to read */
155                         if (filp->f_pos > inode->i_blocks << 9)
156                                 break;
157                         filp->f_pos += sb->s_blocksize - offset;
158                         continue;
159                 }
160
161 revalidate:
162                 /* If the dir block has changed since the last call to
163                  * readdir(2), then we might be pointing to an invalid
164                  * dirent right now.  Scan from the start of the block
165                  * to make sure. */
166                 if (filp->f_version != inode->i_version) {
167                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
168                                 de = (struct ext4_dir_entry_2 *)
169                                         (bh->b_data + i);
170                                 /* It's too expensive to do a full
171                                  * dirent test each time round this
172                                  * loop, but we do have to test at
173                                  * least that it is non-zero.  A
174                                  * failure will be detected in the
175                                  * dirent test below. */
176                                 if (ext4_rec_len_from_disk(de->rec_len)
177                                                 < EXT4_DIR_REC_LEN(1))
178                                         break;
179                                 i += ext4_rec_len_from_disk(de->rec_len);
180                         }
181                         offset = i;
182                         filp->f_pos = (filp->f_pos & ~(sb->s_blocksize - 1))
183                                 | offset;
184                         filp->f_version = inode->i_version;
185                 }
186
187                 while (!error && filp->f_pos < inode->i_size
188                        && offset < sb->s_blocksize) {
189                         de = (struct ext4_dir_entry_2 *) (bh->b_data + offset);
190                         if (!ext4_check_dir_entry ("ext4_readdir", inode, de,
191                                                    bh, offset)) {
192                                 /*
193                                  * On error, skip the f_pos to the next block
194                                  */
195                                 filp->f_pos = (filp->f_pos |
196                                                 (sb->s_blocksize - 1)) + 1;
197                                 brelse (bh);
198                                 ret = stored;
199                                 goto out;
200                         }
201                         offset += ext4_rec_len_from_disk(de->rec_len);
202                         if (le32_to_cpu(de->inode)) {
203                                 /* We might block in the next section
204                                  * if the data destination is
205                                  * currently swapped out.  So, use a
206                                  * version stamp to detect whether or
207                                  * not the directory has been modified
208                                  * during the copy operation.
209                                  */
210                                 u64 version = filp->f_version;
211
212                                 error = filldir(dirent, de->name,
213                                                 de->name_len,
214                                                 filp->f_pos,
215                                                 le32_to_cpu(de->inode),
216                                                 get_dtype(sb, de->file_type));
217                                 if (error)
218                                         break;
219                                 if (version != filp->f_version)
220                                         goto revalidate;
221                                 stored ++;
222                         }
223                         filp->f_pos += ext4_rec_len_from_disk(de->rec_len);
224                 }
225                 offset = 0;
226                 brelse (bh);
227         }
228 out:
229         return ret;
230 }
231
232 /*
233  * These functions convert from the major/minor hash to an f_pos
234  * value.
235  *
236  * Currently we only use major hash numer.  This is unfortunate, but
237  * on 32-bit machines, the same VFS interface is used for lseek and
238  * llseek, so if we use the 64 bit offset, then the 32-bit versions of
239  * lseek/telldir/seekdir will blow out spectacularly, and from within
240  * the ext2 low-level routine, we don't know if we're being called by
241  * a 64-bit version of the system call or the 32-bit version of the
242  * system call.  Worse yet, NFSv2 only allows for a 32-bit readdir
243  * cookie.  Sigh.
244  */
245 #define hash2pos(major, minor)  (major >> 1)
246 #define pos2maj_hash(pos)       ((pos << 1) & 0xffffffff)
247 #define pos2min_hash(pos)       (0)
248
249 /*
250  * This structure holds the nodes of the red-black tree used to store
251  * the directory entry in hash order.
252  */
253 struct fname {
254         __u32           hash;
255         __u32           minor_hash;
256         struct rb_node  rb_hash;
257         struct fname    *next;
258         __u32           inode;
259         __u8            name_len;
260         __u8            file_type;
261         char            name[0];
262 };
263
264 /*
265  * This functoin implements a non-recursive way of freeing all of the
266  * nodes in the red-black tree.
267  */
268 static void free_rb_tree_fname(struct rb_root *root)
269 {
270         struct rb_node  *n = root->rb_node;
271         struct rb_node  *parent;
272         struct fname    *fname;
273
274         while (n) {
275                 /* Do the node's children first */
276                 if (n->rb_left) {
277                         n = n->rb_left;
278                         continue;
279                 }
280                 if (n->rb_right) {
281                         n = n->rb_right;
282                         continue;
283                 }
284                 /*
285                  * The node has no children; free it, and then zero
286                  * out parent's link to it.  Finally go to the
287                  * beginning of the loop and try to free the parent
288                  * node.
289                  */
290                 parent = rb_parent(n);
291                 fname = rb_entry(n, struct fname, rb_hash);
292                 while (fname) {
293                         struct fname * old = fname;
294                         fname = fname->next;
295                         kfree (old);
296                 }
297                 if (!parent)
298                         root->rb_node = NULL;
299                 else if (parent->rb_left == n)
300                         parent->rb_left = NULL;
301                 else if (parent->rb_right == n)
302                         parent->rb_right = NULL;
303                 n = parent;
304         }
305 }
306
307
308 static struct dir_private_info *ext4_htree_create_dir_info(loff_t pos)
309 {
310         struct dir_private_info *p;
311
312         p = kzalloc(sizeof(struct dir_private_info), GFP_KERNEL);
313         if (!p)
314                 return NULL;
315         p->curr_hash = pos2maj_hash(pos);
316         p->curr_minor_hash = pos2min_hash(pos);
317         return p;
318 }
319
320 void ext4_htree_free_dir_info(struct dir_private_info *p)
321 {
322         free_rb_tree_fname(&p->root);
323         kfree(p);
324 }
325
326 /*
327  * Given a directory entry, enter it into the fname rb tree.
328  */
329 int ext4_htree_store_dirent(struct file *dir_file, __u32 hash,
330                              __u32 minor_hash,
331                              struct ext4_dir_entry_2 *dirent)
332 {
333         struct rb_node **p, *parent = NULL;
334         struct fname * fname, *new_fn;
335         struct dir_private_info *info;
336         int len;
337
338         info = (struct dir_private_info *) dir_file->private_data;
339         p = &info->root.rb_node;
340
341         /* Create and allocate the fname structure */
342         len = sizeof(struct fname) + dirent->name_len + 1;
343         new_fn = kzalloc(len, GFP_KERNEL);
344         if (!new_fn)
345                 return -ENOMEM;
346         new_fn->hash = hash;
347         new_fn->minor_hash = minor_hash;
348         new_fn->inode = le32_to_cpu(dirent->inode);
349         new_fn->name_len = dirent->name_len;
350         new_fn->file_type = dirent->file_type;
351         memcpy(new_fn->name, dirent->name, dirent->name_len);
352         new_fn->name[dirent->name_len] = 0;
353
354         while (*p) {
355                 parent = *p;
356                 fname = rb_entry(parent, struct fname, rb_hash);
357
358                 /*
359                  * If the hash and minor hash match up, then we put
360                  * them on a linked list.  This rarely happens...
361                  */
362                 if ((new_fn->hash == fname->hash) &&
363                     (new_fn->minor_hash == fname->minor_hash)) {
364                         new_fn->next = fname->next;
365                         fname->next = new_fn;
366                         return 0;
367                 }
368
369                 if (new_fn->hash < fname->hash)
370                         p = &(*p)->rb_left;
371                 else if (new_fn->hash > fname->hash)
372                         p = &(*p)->rb_right;
373                 else if (new_fn->minor_hash < fname->minor_hash)
374                         p = &(*p)->rb_left;
375                 else /* if (new_fn->minor_hash > fname->minor_hash) */
376                         p = &(*p)->rb_right;
377         }
378
379         rb_link_node(&new_fn->rb_hash, parent, p);
380         rb_insert_color(&new_fn->rb_hash, &info->root);
381         return 0;
382 }
383
384
385
386 /*
387  * This is a helper function for ext4_dx_readdir.  It calls filldir
388  * for all entres on the fname linked list.  (Normally there is only
389  * one entry on the linked list, unless there are 62 bit hash collisions.)
390  */
391 static int call_filldir(struct file * filp, void * dirent,
392                         filldir_t filldir, struct fname *fname)
393 {
394         struct dir_private_info *info = filp->private_data;
395         loff_t  curr_pos;
396         struct inode *inode = filp->f_path.dentry->d_inode;
397         struct super_block * sb;
398         int error;
399
400         sb = inode->i_sb;
401
402         if (!fname) {
403                 printk("call_filldir: called with null fname?!?\n");
404                 return 0;
405         }
406         curr_pos = hash2pos(fname->hash, fname->minor_hash);
407         while (fname) {
408                 error = filldir(dirent, fname->name,
409                                 fname->name_len, curr_pos,
410                                 fname->inode,
411                                 get_dtype(sb, fname->file_type));
412                 if (error) {
413                         filp->f_pos = curr_pos;
414                         info->extra_fname = fname;
415                         return error;
416                 }
417                 fname = fname->next;
418         }
419         return 0;
420 }
421
422 static int ext4_dx_readdir(struct file * filp,
423                          void * dirent, filldir_t filldir)
424 {
425         struct dir_private_info *info = filp->private_data;
426         struct inode *inode = filp->f_path.dentry->d_inode;
427         struct fname *fname;
428         int     ret;
429
430         if (!info) {
431                 info = ext4_htree_create_dir_info(filp->f_pos);
432                 if (!info)
433                         return -ENOMEM;
434                 filp->private_data = info;
435         }
436
437         if (filp->f_pos == EXT4_HTREE_EOF)
438                 return 0;       /* EOF */
439
440         /* Some one has messed with f_pos; reset the world */
441         if (info->last_pos != filp->f_pos) {
442                 free_rb_tree_fname(&info->root);
443                 info->curr_node = NULL;
444                 info->extra_fname = NULL;
445                 info->curr_hash = pos2maj_hash(filp->f_pos);
446                 info->curr_minor_hash = pos2min_hash(filp->f_pos);
447         }
448
449         /*
450          * If there are any leftover names on the hash collision
451          * chain, return them first.
452          */
453         if (info->extra_fname) {
454                 if (call_filldir(filp, dirent, filldir, info->extra_fname))
455                         goto finished;
456
457                 info->extra_fname = NULL;
458                 info->curr_node = rb_next(info->curr_node);
459                 if (!info->curr_node) {
460                         if (info->next_hash == ~0) {
461                                 filp->f_pos = EXT4_HTREE_EOF;
462                                 goto finished;
463                         }
464                         info->curr_hash = info->next_hash;
465                         info->curr_minor_hash = 0;
466                 }
467         } else if (!info->curr_node)
468                 info->curr_node = rb_first(&info->root);
469
470         while (1) {
471                 /*
472                  * Fill the rbtree if we have no more entries,
473                  * or the inode has changed since we last read in the
474                  * cached entries.
475                  */
476                 if ((!info->curr_node) ||
477                     (filp->f_version != inode->i_version)) {
478                         info->curr_node = NULL;
479                         free_rb_tree_fname(&info->root);
480                         filp->f_version = inode->i_version;
481                         ret = ext4_htree_fill_tree(filp, info->curr_hash,
482                                                    info->curr_minor_hash,
483                                                    &info->next_hash);
484                         if (ret < 0)
485                                 return ret;
486                         if (ret == 0) {
487                                 filp->f_pos = EXT4_HTREE_EOF;
488                                 break;
489                         }
490                         info->curr_node = rb_first(&info->root);
491                 }
492
493                 fname = rb_entry(info->curr_node, struct fname, rb_hash);
494                 info->curr_hash = fname->hash;
495                 info->curr_minor_hash = fname->minor_hash;
496                 if (call_filldir(filp, dirent, filldir, fname))
497                         break;
498
499                 info->curr_node = rb_next(info->curr_node);
500                 if (!info->curr_node) {
501                         if (info->next_hash == ~0) {
502                                 filp->f_pos = EXT4_HTREE_EOF;
503                                 break;
504                         }
505                         info->curr_hash = info->next_hash;
506                         info->curr_minor_hash = 0;
507                 }
508         }
509 finished:
510         info->last_pos = filp->f_pos;
511         return 0;
512 }
513
514 static int ext4_release_dir (struct inode * inode, struct file * filp)
515 {
516         if (filp->private_data)
517                 ext4_htree_free_dir_info(filp->private_data);
518
519         return 0;
520 }