[PATCH] ext3: cleanups and WARN_ON()
[linux-2.6.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.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/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/smp_lock.h>
31 #include <linux/highuid.h>
32 #include <linux/pagemap.h>
33 #include <linux/quotaops.h>
34 #include <linux/string.h>
35 #include <linux/buffer_head.h>
36 #include <linux/writeback.h>
37 #include <linux/mpage.h>
38 #include <linux/uio.h>
39 #include "xattr.h"
40 #include "acl.h"
41
42 static int ext3_writepage_trans_blocks(struct inode *inode);
43
44 /*
45  * Test whether an inode is a fast symlink.
46  */
47 static int ext3_inode_is_fast_symlink(struct inode *inode)
48 {
49         int ea_blocks = EXT3_I(inode)->i_file_acl ?
50                 (inode->i_sb->s_blocksize >> 9) : 0;
51
52         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
53 }
54
55 /*
56  * The ext3 forget function must perform a revoke if we are freeing data
57  * which has been journaled.  Metadata (eg. indirect blocks) must be
58  * revoked in all cases. 
59  *
60  * "bh" may be NULL: a metadata block may have been freed from memory
61  * but there may still be a record of it in the journal, and that record
62  * still needs to be revoked.
63  */
64 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
65                         struct buffer_head *bh, int blocknr)
66 {
67         int err;
68
69         might_sleep();
70
71         BUFFER_TRACE(bh, "enter");
72
73         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
74                   "data mode %lx\n",
75                   bh, is_metadata, inode->i_mode,
76                   test_opt(inode->i_sb, DATA_FLAGS));
77
78         /* Never use the revoke function if we are doing full data
79          * journaling: there is no need to, and a V1 superblock won't
80          * support it.  Otherwise, only skip the revoke on un-journaled
81          * data blocks. */
82
83         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
84             (!is_metadata && !ext3_should_journal_data(inode))) {
85                 if (bh) {
86                         BUFFER_TRACE(bh, "call journal_forget");
87                         return ext3_journal_forget(handle, bh);
88                 }
89                 return 0;
90         }
91
92         /*
93          * data!=journal && (is_metadata || should_journal_data(inode))
94          */
95         BUFFER_TRACE(bh, "call ext3_journal_revoke");
96         err = ext3_journal_revoke(handle, blocknr, bh);
97         if (err)
98                 ext3_abort(inode->i_sb, __FUNCTION__,
99                            "error %d when attempting revoke", err);
100         BUFFER_TRACE(bh, "exit");
101         return err;
102 }
103
104 /*
105  * Work out how many blocks we need to proceed with the next chunk of a
106  * truncate transaction.
107  */
108 static unsigned long blocks_for_truncate(struct inode *inode) 
109 {
110         unsigned long needed;
111
112         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
113
114         /* Give ourselves just enough room to cope with inodes in which
115          * i_blocks is corrupt: we've seen disk corruptions in the past
116          * which resulted in random data in an inode which looked enough
117          * like a regular file for ext3 to try to delete it.  Things
118          * will go a bit crazy if that happens, but at least we should
119          * try not to panic the whole kernel. */
120         if (needed < 2)
121                 needed = 2;
122
123         /* But we need to bound the transaction so we don't overflow the
124          * journal. */
125         if (needed > EXT3_MAX_TRANS_DATA) 
126                 needed = EXT3_MAX_TRANS_DATA;
127
128         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
129 }
130
131 /* 
132  * Truncate transactions can be complex and absolutely huge.  So we need to
133  * be able to restart the transaction at a conventient checkpoint to make
134  * sure we don't overflow the journal.
135  *
136  * start_transaction gets us a new handle for a truncate transaction,
137  * and extend_transaction tries to extend the existing one a bit.  If
138  * extend fails, we need to propagate the failure up and restart the
139  * transaction in the top-level truncate loop. --sct 
140  */
141 static handle_t *start_transaction(struct inode *inode) 
142 {
143         handle_t *result;
144
145         result = ext3_journal_start(inode, blocks_for_truncate(inode));
146         if (!IS_ERR(result))
147                 return result;
148
149         ext3_std_error(inode->i_sb, PTR_ERR(result));
150         return result;
151 }
152
153 /*
154  * Try to extend this transaction for the purposes of truncation.
155  *
156  * Returns 0 if we managed to create more room.  If we can't create more
157  * room, and the transaction must be restarted we return 1.
158  */
159 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
160 {
161         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
162                 return 0;
163         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
164                 return 0;
165         return 1;
166 }
167
168 /*
169  * Restart the transaction associated with *handle.  This does a commit,
170  * so before we call here everything must be consistently dirtied against
171  * this transaction.
172  */
173 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
174 {
175         jbd_debug(2, "restarting handle %p\n", handle);
176         return ext3_journal_restart(handle, blocks_for_truncate(inode));
177 }
178
179 /*
180  * Called at the last iput() if i_nlink is zero.
181  */
182 void ext3_delete_inode (struct inode * inode)
183 {
184         handle_t *handle;
185
186         truncate_inode_pages(&inode->i_data, 0);
187
188         if (is_bad_inode(inode))
189                 goto no_delete;
190
191         handle = start_transaction(inode);
192         if (IS_ERR(handle)) {
193                 /*
194                  * If we're going to skip the normal cleanup, we still need to
195                  * make sure that the in-core orphan linked list is properly
196                  * cleaned up.
197                  */
198                 ext3_orphan_del(NULL, inode);
199                 goto no_delete;
200         }
201
202         if (IS_SYNC(inode))
203                 handle->h_sync = 1;
204         inode->i_size = 0;
205         if (inode->i_blocks)
206                 ext3_truncate(inode);
207         /*
208          * Kill off the orphan record which ext3_truncate created.
209          * AKPM: I think this can be inside the above `if'.
210          * Note that ext3_orphan_del() has to be able to cope with the
211          * deletion of a non-existent orphan - this is because we don't
212          * know if ext3_truncate() actually created an orphan record.
213          * (Well, we could do this if we need to, but heck - it works)
214          */
215         ext3_orphan_del(handle, inode);
216         EXT3_I(inode)->i_dtime  = get_seconds();
217
218         /* 
219          * One subtle ordering requirement: if anything has gone wrong
220          * (transaction abort, IO errors, whatever), then we can still
221          * do these next steps (the fs will already have been marked as
222          * having errors), but we can't free the inode if the mark_dirty
223          * fails.  
224          */
225         if (ext3_mark_inode_dirty(handle, inode))
226                 /* If that failed, just do the required in-core inode clear. */
227                 clear_inode(inode);
228         else
229                 ext3_free_inode(handle, inode);
230         ext3_journal_stop(handle);
231         return;
232 no_delete:
233         clear_inode(inode);     /* We must guarantee clearing of inode... */
234 }
235
236 typedef struct {
237         __le32  *p;
238         __le32  key;
239         struct buffer_head *bh;
240 } Indirect;
241
242 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
243 {
244         p->key = *(p->p = v);
245         p->bh = bh;
246 }
247
248 static int verify_chain(Indirect *from, Indirect *to)
249 {
250         while (from <= to && from->key == *from->p)
251                 from++;
252         return (from > to);
253 }
254
255 /**
256  *      ext3_block_to_path - parse the block number into array of offsets
257  *      @inode: inode in question (we are only interested in its superblock)
258  *      @i_block: block number to be parsed
259  *      @offsets: array to store the offsets in
260  *      @boundary: set this non-zero if the referred-to block is likely to be
261  *             followed (on disk) by an indirect block.
262  *
263  *      To store the locations of file's data ext3 uses a data structure common
264  *      for UNIX filesystems - tree of pointers anchored in the inode, with
265  *      data blocks at leaves and indirect blocks in intermediate nodes.
266  *      This function translates the block number into path in that tree -
267  *      return value is the path length and @offsets[n] is the offset of
268  *      pointer to (n+1)th node in the nth one. If @block is out of range
269  *      (negative or too large) warning is printed and zero returned.
270  *
271  *      Note: function doesn't find node addresses, so no IO is needed. All
272  *      we need to know is the capacity of indirect blocks (taken from the
273  *      inode->i_sb).
274  */
275
276 /*
277  * Portability note: the last comparison (check that we fit into triple
278  * indirect block) is spelled differently, because otherwise on an
279  * architecture with 32-bit longs and 8Kb pages we might get into trouble
280  * if our filesystem had 8Kb blocks. We might use long long, but that would
281  * kill us on x86. Oh, well, at least the sign propagation does not matter -
282  * i_block would have to be negative in the very beginning, so we would not
283  * get there at all.
284  */
285
286 static int ext3_block_to_path(struct inode *inode,
287                         long i_block, int offsets[4], int *boundary)
288 {
289         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
290         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
291         const long direct_blocks = EXT3_NDIR_BLOCKS,
292                 indirect_blocks = ptrs,
293                 double_blocks = (1 << (ptrs_bits * 2));
294         int n = 0;
295         int final = 0;
296
297         if (i_block < 0) {
298                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
299         } else if (i_block < direct_blocks) {
300                 offsets[n++] = i_block;
301                 final = direct_blocks;
302         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
303                 offsets[n++] = EXT3_IND_BLOCK;
304                 offsets[n++] = i_block;
305                 final = ptrs;
306         } else if ((i_block -= indirect_blocks) < double_blocks) {
307                 offsets[n++] = EXT3_DIND_BLOCK;
308                 offsets[n++] = i_block >> ptrs_bits;
309                 offsets[n++] = i_block & (ptrs - 1);
310                 final = ptrs;
311         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
312                 offsets[n++] = EXT3_TIND_BLOCK;
313                 offsets[n++] = i_block >> (ptrs_bits * 2);
314                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
315                 offsets[n++] = i_block & (ptrs - 1);
316                 final = ptrs;
317         } else {
318                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
319         }
320         if (boundary)
321                 *boundary = final - 1 - (i_block & (ptrs - 1));
322         return n;
323 }
324
325 /**
326  *      ext3_get_branch - read the chain of indirect blocks leading to data
327  *      @inode: inode in question
328  *      @depth: depth of the chain (1 - direct pointer, etc.)
329  *      @offsets: offsets of pointers in inode/indirect blocks
330  *      @chain: place to store the result
331  *      @err: here we store the error value
332  *
333  *      Function fills the array of triples <key, p, bh> and returns %NULL
334  *      if everything went OK or the pointer to the last filled triple
335  *      (incomplete one) otherwise. Upon the return chain[i].key contains
336  *      the number of (i+1)-th block in the chain (as it is stored in memory,
337  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
338  *      number (it points into struct inode for i==0 and into the bh->b_data
339  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
340  *      block for i>0 and NULL for i==0. In other words, it holds the block
341  *      numbers of the chain, addresses they were taken from (and where we can
342  *      verify that chain did not change) and buffer_heads hosting these
343  *      numbers.
344  *
345  *      Function stops when it stumbles upon zero pointer (absent block)
346  *              (pointer to last triple returned, *@err == 0)
347  *      or when it gets an IO error reading an indirect block
348  *              (ditto, *@err == -EIO)
349  *      or when it notices that chain had been changed while it was reading
350  *              (ditto, *@err == -EAGAIN)
351  *      or when it reads all @depth-1 indirect blocks successfully and finds
352  *      the whole chain, all way to the data (returns %NULL, *err == 0).
353  */
354 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
355                                  Indirect chain[4], int *err)
356 {
357         struct super_block *sb = inode->i_sb;
358         Indirect *p = chain;
359         struct buffer_head *bh;
360
361         *err = 0;
362         /* i_data is not going away, no lock needed */
363         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
364         if (!p->key)
365                 goto no_block;
366         while (--depth) {
367                 bh = sb_bread(sb, le32_to_cpu(p->key));
368                 if (!bh)
369                         goto failure;
370                 /* Reader: pointers */
371                 if (!verify_chain(chain, p))
372                         goto changed;
373                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
374                 /* Reader: end */
375                 if (!p->key)
376                         goto no_block;
377         }
378         return NULL;
379
380 changed:
381         brelse(bh);
382         *err = -EAGAIN;
383         goto no_block;
384 failure:
385         *err = -EIO;
386 no_block:
387         return p;
388 }
389
390 /**
391  *      ext3_find_near - find a place for allocation with sufficient locality
392  *      @inode: owner
393  *      @ind: descriptor of indirect block.
394  *
395  *      This function returns the prefered place for block allocation.
396  *      It is used when heuristic for sequential allocation fails.
397  *      Rules are:
398  *        + if there is a block to the left of our position - allocate near it.
399  *        + if pointer will live in indirect block - allocate near that block.
400  *        + if pointer will live in inode - allocate in the same
401  *          cylinder group. 
402  *
403  * In the latter case we colour the starting block by the callers PID to
404  * prevent it from clashing with concurrent allocations for a different inode
405  * in the same block group.   The PID is used here so that functionally related
406  * files will be close-by on-disk.
407  *
408  *      Caller must make sure that @ind is valid and will stay that way.
409  */
410 static unsigned long ext3_find_near(struct inode *inode, Indirect *ind)
411 {
412         struct ext3_inode_info *ei = EXT3_I(inode);
413         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
414         __le32 *p;
415         unsigned long bg_start;
416         unsigned long colour;
417
418         /* Try to find previous block */
419         for (p = ind->p - 1; p >= start; p--) {
420                 if (*p)
421                         return le32_to_cpu(*p);
422         }
423
424         /* No such thing, so let's try location of indirect block */
425         if (ind->bh)
426                 return ind->bh->b_blocknr;
427
428         /*
429          * It is going to be referred to from the inode itself? OK, just put it
430          * into the same cylinder group then.
431          */
432         bg_start = (ei->i_block_group * EXT3_BLOCKS_PER_GROUP(inode->i_sb)) +
433                 le32_to_cpu(EXT3_SB(inode->i_sb)->s_es->s_first_data_block);
434         colour = (current->pid % 16) *
435                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
436         return bg_start + colour;
437 }
438
439 /**
440  *      ext3_find_goal - find a prefered place for allocation.
441  *      @inode: owner
442  *      @block:  block we want
443  *      @chain:  chain of indirect blocks
444  *      @partial: pointer to the last triple within a chain
445  *      @goal:  place to store the result.
446  *
447  *      Normally this function find the prefered place for block allocation,
448  *      stores it in *@goal and returns zero.
449  */
450
451 static unsigned long ext3_find_goal(struct inode *inode, long block,
452                 Indirect chain[4], Indirect *partial)
453 {
454         struct ext3_block_alloc_info *block_i;
455
456         block_i =  EXT3_I(inode)->i_block_alloc_info;
457
458         /*
459          * try the heuristic for sequential allocation,
460          * failing that at least try to get decent locality.
461          */
462         if (block_i && (block == block_i->last_alloc_logical_block + 1)
463                 && (block_i->last_alloc_physical_block != 0)) {
464                 return block_i->last_alloc_physical_block + 1;
465         }
466
467         return ext3_find_near(inode, partial);
468 }
469
470 /**
471  *      ext3_blks_to_allocate: Look up the block map and count the number
472  *      of direct blocks need to be allocated for the given branch.
473  *
474  *      @branch: chain of indirect blocks
475  *      @k: number of blocks need for indirect blocks
476  *      @blks: number of data blocks to be mapped.
477  *      @blocks_to_boundary:  the offset in the indirect block
478  *
479  *      return the total number of blocks to be allocate, including the
480  *      direct and indirect blocks.
481  */
482 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
483                 int blocks_to_boundary)
484 {
485         unsigned long count = 0;
486
487         /*
488          * Simple case, [t,d]Indirect block(s) has not allocated yet
489          * then it's clear blocks on that path have not allocated
490          */
491         if (k > 0) {
492                 /* right now we don't handle cross boundary allocation */
493                 if (blks < blocks_to_boundary + 1)
494                         count += blks;
495                 else
496                         count += blocks_to_boundary + 1;
497                 return count;
498         }
499
500         count++;
501         while (count < blks && count <= blocks_to_boundary &&
502                 le32_to_cpu(*(branch[0].p + count)) == 0) {
503                 count++;
504         }
505         return count;
506 }
507
508 /**
509  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
510  *      @indirect_blks: the number of blocks need to allocate for indirect
511  *                      blocks
512  *
513  *      @new_blocks: on return it will store the new block numbers for
514  *      the indirect blocks(if needed) and the first direct block,
515  *      @blks:  on return it will store the total number of allocated
516  *              direct blocks
517  */
518 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
519                         unsigned long goal, int indirect_blks, int blks,
520                         unsigned long long new_blocks[4], int *err)
521 {
522         int target, i;
523         unsigned long count = 0;
524         int index = 0;
525         unsigned long current_block = 0;
526         int ret = 0;
527
528         /*
529          * Here we try to allocate the requested multiple blocks at once,
530          * on a best-effort basis.
531          * To build a branch, we should allocate blocks for
532          * the indirect blocks(if not allocated yet), and at least
533          * the first direct block of this branch.  That's the
534          * minimum number of blocks need to allocate(required)
535          */
536         target = blks + indirect_blks;
537
538         while (1) {
539                 count = target;
540                 /* allocating blocks for indirect blocks and direct blocks */
541                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
542                 if (*err)
543                         goto failed_out;
544
545                 target -= count;
546                 /* allocate blocks for indirect blocks */
547                 while (index < indirect_blks && count) {
548                         new_blocks[index++] = current_block++;
549                         count--;
550                 }
551
552                 if (count > 0)
553                         break;
554         }
555
556         /* save the new block number for the first direct block */
557         new_blocks[index] = current_block;
558
559         /* total number of blocks allocated for direct blocks */
560         ret = count;
561         *err = 0;
562         return ret;
563 failed_out:
564         for (i = 0; i <index; i++)
565                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
566         return ret;
567 }
568
569 /**
570  *      ext3_alloc_branch - allocate and set up a chain of blocks.
571  *      @inode: owner
572  *      @indirect_blks: number of allocated indirect blocks
573  *      @blks: number of allocated direct blocks
574  *      @offsets: offsets (in the blocks) to store the pointers to next.
575  *      @branch: place to store the chain in.
576  *
577  *      This function allocates blocks, zeroes out all but the last one,
578  *      links them into chain and (if we are synchronous) writes them to disk.
579  *      In other words, it prepares a branch that can be spliced onto the
580  *      inode. It stores the information about that chain in the branch[], in
581  *      the same format as ext3_get_branch() would do. We are calling it after
582  *      we had read the existing part of chain and partial points to the last
583  *      triple of that (one with zero ->key). Upon the exit we have the same
584  *      picture as after the successful ext3_get_block(), except that in one
585  *      place chain is disconnected - *branch->p is still zero (we did not
586  *      set the last link), but branch->key contains the number that should
587  *      be placed into *branch->p to fill that gap.
588  *
589  *      If allocation fails we free all blocks we've allocated (and forget
590  *      their buffer_heads) and return the error value the from failed
591  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
592  *      as described above and return 0.
593  */
594 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
595                         int indirect_blks, int *blks, unsigned long goal,
596                         int *offsets, Indirect *branch)
597 {
598         int blocksize = inode->i_sb->s_blocksize;
599         int i, n = 0;
600         int err = 0;
601         struct buffer_head *bh;
602         int num;
603         unsigned long long new_blocks[4];
604         unsigned long long current_block;
605
606         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
607                                 *blks, new_blocks, &err);
608         if (err)
609                 return err;
610
611         branch[0].key = cpu_to_le32(new_blocks[0]);
612         /*
613          * metadata blocks and data blocks are allocated.
614          */
615         for (n = 1; n <= indirect_blks;  n++) {
616                 /*
617                  * Get buffer_head for parent block, zero it out
618                  * and set the pointer to new one, then send
619                  * parent to disk.
620                  */
621                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
622                 branch[n].bh = bh;
623                 lock_buffer(bh);
624                 BUFFER_TRACE(bh, "call get_create_access");
625                 err = ext3_journal_get_create_access(handle, bh);
626                 if (err) {
627                         unlock_buffer(bh);
628                         brelse(bh);
629                         goto failed;
630                 }
631
632                 memset(bh->b_data, 0, blocksize);
633                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
634                 branch[n].key = cpu_to_le32(new_blocks[n]);
635                 *branch[n].p = branch[n].key;
636                 if ( n == indirect_blks) {
637                         current_block = new_blocks[n];
638                         /*
639                          * End of chain, update the last new metablock of
640                          * the chain to point to the new allocated
641                          * data blocks numbers
642                          */
643                         for (i=1; i < num; i++)
644                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
645                 }
646                 BUFFER_TRACE(bh, "marking uptodate");
647                 set_buffer_uptodate(bh);
648                 unlock_buffer(bh);
649
650                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
651                 err = ext3_journal_dirty_metadata(handle, bh);
652                 if (err)
653                         goto failed;
654         }
655         *blks = num;
656         return err;
657 failed:
658         /* Allocation failed, free what we already allocated */
659         for (i = 1; i <= n ; i++) {
660                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
661                 ext3_journal_forget(handle, branch[i].bh);
662         }
663         for (i = 0; i <indirect_blks; i++)
664                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
665
666         ext3_free_blocks(handle, inode, new_blocks[i], num);
667
668         return err;
669 }
670
671 /**
672  * ext3_splice_branch - splice the allocated branch onto inode.
673  * @inode: owner
674  * @block: (logical) number of block we are adding
675  * @chain: chain of indirect blocks (with a missing link - see
676  *      ext3_alloc_branch)
677  * @where: location of missing link
678  * @num:   number of indirect blocks we are adding
679  * @blks:  number of direct blocks we are adding
680  *
681  * This function fills the missing link and does all housekeeping needed in
682  * inode (->i_blocks, etc.). In case of success we end up with the full
683  * chain to new block and return 0.
684  */
685 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
686                         long block, Indirect *where, int num, int blks)
687 {
688         int i;
689         int err = 0;
690         struct ext3_block_alloc_info *block_i;
691         unsigned long current_block;
692
693         block_i = EXT3_I(inode)->i_block_alloc_info;
694         /*
695          * If we're splicing into a [td]indirect block (as opposed to the
696          * inode) then we need to get write access to the [td]indirect block
697          * before the splice.
698          */
699         if (where->bh) {
700                 BUFFER_TRACE(where->bh, "get_write_access");
701                 err = ext3_journal_get_write_access(handle, where->bh);
702                 if (err)
703                         goto err_out;
704         }
705         /* That's it */
706
707         *where->p = where->key;
708
709         /*
710          * Update the host buffer_head or inode to point to more just allocated
711          * direct blocks blocks
712          */
713         if (num == 0 && blks > 1) {
714                 current_block = le32_to_cpu(where->key + 1);
715                 for (i = 1; i < blks; i++)
716                         *(where->p + i ) = cpu_to_le32(current_block++);
717         }
718
719         /*
720          * update the most recently allocated logical & physical block
721          * in i_block_alloc_info, to assist find the proper goal block for next
722          * allocation
723          */
724         if (block_i) {
725                 block_i->last_alloc_logical_block = block + blks - 1;
726                 block_i->last_alloc_physical_block =
727                                 le32_to_cpu(where[num].key + blks - 1);
728         }
729
730         /* We are done with atomic stuff, now do the rest of housekeeping */
731
732         inode->i_ctime = CURRENT_TIME_SEC;
733         ext3_mark_inode_dirty(handle, inode);
734
735         /* had we spliced it onto indirect block? */
736         if (where->bh) {
737                 /*
738                  * If we spliced it onto an indirect block, we haven't
739                  * altered the inode.  Note however that if it is being spliced
740                  * onto an indirect block at the very end of the file (the
741                  * file is growing) then we *will* alter the inode to reflect
742                  * the new i_size.  But that is not done here - it is done in
743                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
744                  */
745                 jbd_debug(5, "splicing indirect only\n");
746                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
747                 err = ext3_journal_dirty_metadata(handle, where->bh);
748                 if (err) 
749                         goto err_out;
750         } else {
751                 /*
752                  * OK, we spliced it into the inode itself on a direct block.
753                  * Inode was dirtied above.
754                  */
755                 jbd_debug(5, "splicing direct\n");
756         }
757         return err;
758
759 err_out:
760         for (i = 1; i <= num; i++) {
761                 BUFFER_TRACE(where[i].bh, "call journal_forget");
762                 ext3_journal_forget(handle, where[i].bh);
763                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
764         }
765         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
766
767         return err;
768 }
769
770 /*
771  * Allocation strategy is simple: if we have to allocate something, we will
772  * have to go the whole way to leaf. So let's do it before attaching anything
773  * to tree, set linkage between the newborn blocks, write them if sync is
774  * required, recheck the path, free and repeat if check fails, otherwise
775  * set the last missing link (that will protect us from any truncate-generated
776  * removals - all blocks on the path are immune now) and possibly force the
777  * write on the parent block.
778  * That has a nice additional property: no special recovery from the failed
779  * allocations is needed - we simply release blocks and do not touch anything
780  * reachable from inode.
781  *
782  * `handle' can be NULL if create == 0.
783  *
784  * The BKL may not be held on entry here.  Be sure to take it early.
785  * return > 0, # of blocks mapped or allocated.
786  * return = 0, if plain lookup failed.
787  * return < 0, error case.
788  */
789 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
790                 sector_t iblock, unsigned long maxblocks,
791                 struct buffer_head *bh_result,
792                 int create, int extend_disksize)
793 {
794         int err = -EIO;
795         int offsets[4];
796         Indirect chain[4];
797         Indirect *partial;
798         unsigned long goal;
799         int indirect_blks;
800         int blocks_to_boundary = 0;
801         int depth;
802         struct ext3_inode_info *ei = EXT3_I(inode);
803         int count = 0;
804         unsigned long first_block = 0;
805
806
807         J_ASSERT(handle != NULL || create == 0);
808         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
809
810         if (depth == 0)
811                 goto out;
812
813         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
814
815         /* Simplest case - block found, no allocation needed */
816         if (!partial) {
817                 first_block = chain[depth - 1].key;
818                 clear_buffer_new(bh_result);
819                 count++;
820                 /*map more blocks*/
821                 while (count < maxblocks && count <= blocks_to_boundary) {
822                         if (!verify_chain(chain, partial)) {
823                                 /*
824                                  * Indirect block might be removed by
825                                  * truncate while we were reading it.
826                                  * Handling of that case: forget what we've
827                                  * got now. Flag the err as EAGAIN, so it
828                                  * will reread.
829                                  */
830                                 err = -EAGAIN;
831                                 count = 0;
832                                 break;
833                         }
834                         if (le32_to_cpu(*(chain[depth-1].p+count) ==
835                                         (first_block + count)))
836                                 count++;
837                         else
838                                 break;
839                 }
840                 if (err != -EAGAIN)
841                         goto got_it;
842         }
843
844         /* Next simple case - plain lookup or failed read of indirect block */
845         if (!create || err == -EIO)
846                 goto cleanup;
847
848         mutex_lock(&ei->truncate_mutex);
849
850         /*
851          * If the indirect block is missing while we are reading
852          * the chain(ext3_get_branch() returns -EAGAIN err), or
853          * if the chain has been changed after we grab the semaphore,
854          * (either because another process truncated this branch, or
855          * another get_block allocated this branch) re-grab the chain to see if
856          * the request block has been allocated or not.
857          *
858          * Since we already block the truncate/other get_block
859          * at this point, we will have the current copy of the chain when we
860          * splice the branch into the tree.
861          */
862         if (err == -EAGAIN || !verify_chain(chain, partial)) {
863                 while (partial > chain) {
864                         brelse(partial->bh);
865                         partial--;
866                 }
867                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
868                 if (!partial) {
869                         count++;
870                         mutex_unlock(&ei->truncate_mutex);
871                         if (err)
872                                 goto cleanup;
873                         clear_buffer_new(bh_result);
874                         goto got_it;
875                 }
876         }
877
878         /*
879          * Okay, we need to do block allocation.  Lazily initialize the block
880          * allocation info here if necessary
881         */
882         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
883                 ext3_init_block_alloc_info(inode);
884
885         goal = ext3_find_goal(inode, iblock, chain, partial);
886
887         /* the number of blocks need to allocate for [d,t]indirect blocks */
888         indirect_blks = (chain + depth) - partial - 1;
889
890         /*
891          * Next look up the indirect map to count the totoal number of
892          * direct blocks to allocate for this branch.
893          */
894         count = ext3_blks_to_allocate(partial, indirect_blks,
895                                         maxblocks, blocks_to_boundary);
896         /*
897          * Block out ext3_truncate while we alter the tree
898          */
899         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
900                                 offsets + (partial - chain), partial);
901
902         /*
903          * The ext3_splice_branch call will free and forget any buffers
904          * on the new chain if there is a failure, but that risks using
905          * up transaction credits, especially for bitmaps where the
906          * credits cannot be returned.  Can we handle this somehow?  We
907          * may need to return -EAGAIN upwards in the worst case.  --sct
908          */
909         if (!err)
910                 err = ext3_splice_branch(handle, inode, iblock,
911                                         partial, indirect_blks, count);
912         /*
913          * i_disksize growing is protected by truncate_mutex.  Don't forget to
914          * protect it if you're about to implement concurrent
915          * ext3_get_block() -bzzz
916         */
917         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
918                 ei->i_disksize = inode->i_size;
919         mutex_unlock(&ei->truncate_mutex);
920         if (err)
921                 goto cleanup;
922
923         set_buffer_new(bh_result);
924 got_it:
925         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
926         if (blocks_to_boundary == 0)
927                 set_buffer_boundary(bh_result);
928         err = count;
929         /* Clean up and exit */
930         partial = chain + depth - 1;    /* the whole chain */
931 cleanup:
932         while (partial > chain) {
933                 BUFFER_TRACE(partial->bh, "call brelse");
934                 brelse(partial->bh);
935                 partial--;
936         }
937         BUFFER_TRACE(bh_result, "returned");
938 out:
939         return err;
940 }
941
942 #define DIO_CREDITS (EXT3_RESERVE_TRANS_BLOCKS + 32)
943
944 static int
945 ext3_direct_io_get_blocks(struct inode *inode, sector_t iblock,
946                 struct buffer_head *bh_result, int create)
947 {
948         handle_t *handle = journal_current_handle();
949         int ret = 0;
950         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
951
952         if (!create)
953                 goto get_block;         /* A read */
954
955         if (max_blocks == 1)
956                 goto get_block;         /* A single block get */
957
958         if (handle->h_transaction->t_state == T_LOCKED) {
959                 /*
960                  * Huge direct-io writes can hold off commits for long
961                  * periods of time.  Let this commit run.
962                  */
963                 ext3_journal_stop(handle);
964                 handle = ext3_journal_start(inode, DIO_CREDITS);
965                 if (IS_ERR(handle))
966                         ret = PTR_ERR(handle);
967                 goto get_block;
968         }
969
970         if (handle->h_buffer_credits <= EXT3_RESERVE_TRANS_BLOCKS) {
971                 /*
972                  * Getting low on buffer credits...
973                  */
974                 ret = ext3_journal_extend(handle, DIO_CREDITS);
975                 if (ret > 0) {
976                         /*
977                          * Couldn't extend the transaction.  Start a new one.
978                          */
979                         ret = ext3_journal_restart(handle, DIO_CREDITS);
980                 }
981         }
982
983 get_block:
984         if (ret == 0) {
985                 ret = ext3_get_blocks_handle(handle, inode, iblock,
986                                         max_blocks, bh_result, create, 0);
987                 if (ret > 0) {
988                         bh_result->b_size = (ret << inode->i_blkbits);
989                         ret = 0;
990                 }
991         }
992         return ret;
993 }
994
995 static int ext3_get_block(struct inode *inode, sector_t iblock,
996                         struct buffer_head *bh_result, int create)
997 {
998         return ext3_direct_io_get_blocks(inode, iblock, bh_result, create);
999 }
1000
1001 /*
1002  * `handle' can be NULL if create is zero
1003  */
1004 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1005                                 long block, int create, int *errp)
1006 {
1007         struct buffer_head dummy;
1008         int fatal = 0, err;
1009
1010         J_ASSERT(handle != NULL || create == 0);
1011
1012         dummy.b_state = 0;
1013         dummy.b_blocknr = -1000;
1014         buffer_trace_init(&dummy.b_history);
1015         err = ext3_get_blocks_handle(handle, inode, block, 1,
1016                                         &dummy, create, 1);
1017         if (err == 1) {
1018                 err = 0;
1019         } else if (err >= 0) {
1020                 WARN_ON(1);
1021                 err = -EIO;
1022         }
1023         *errp = err;
1024         if (!err && buffer_mapped(&dummy)) {
1025                 struct buffer_head *bh;
1026                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1027                 if (!bh) {
1028                         *errp = -EIO;
1029                         goto err;
1030                 }
1031                 if (buffer_new(&dummy)) {
1032                         J_ASSERT(create != 0);
1033                         J_ASSERT(handle != 0);
1034
1035                         /*
1036                          * Now that we do not always journal data, we should
1037                          * keep in mind whether this should always journal the
1038                          * new buffer as metadata.  For now, regular file
1039                          * writes use ext3_get_block instead, so it's not a
1040                          * problem.
1041                          */
1042                         lock_buffer(bh);
1043                         BUFFER_TRACE(bh, "call get_create_access");
1044                         fatal = ext3_journal_get_create_access(handle, bh);
1045                         if (!fatal && !buffer_uptodate(bh)) {
1046                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1047                                 set_buffer_uptodate(bh);
1048                         }
1049                         unlock_buffer(bh);
1050                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1051                         err = ext3_journal_dirty_metadata(handle, bh);
1052                         if (!fatal)
1053                                 fatal = err;
1054                 } else {
1055                         BUFFER_TRACE(bh, "not a new buffer");
1056                 }
1057                 if (fatal) {
1058                         *errp = fatal;
1059                         brelse(bh);
1060                         bh = NULL;
1061                 }
1062                 return bh;
1063         }
1064 err:
1065         return NULL;
1066 }
1067
1068 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1069                                int block, int create, int *err)
1070 {
1071         struct buffer_head * bh;
1072
1073         bh = ext3_getblk(handle, inode, block, create, err);
1074         if (!bh)
1075                 return bh;
1076         if (buffer_uptodate(bh))
1077                 return bh;
1078         ll_rw_block(READ, 1, &bh);
1079         wait_on_buffer(bh);
1080         if (buffer_uptodate(bh))
1081                 return bh;
1082         put_bh(bh);
1083         *err = -EIO;
1084         return NULL;
1085 }
1086
1087 static int walk_page_buffers(   handle_t *handle,
1088                                 struct buffer_head *head,
1089                                 unsigned from,
1090                                 unsigned to,
1091                                 int *partial,
1092                                 int (*fn)(      handle_t *handle,
1093                                                 struct buffer_head *bh))
1094 {
1095         struct buffer_head *bh;
1096         unsigned block_start, block_end;
1097         unsigned blocksize = head->b_size;
1098         int err, ret = 0;
1099         struct buffer_head *next;
1100
1101         for (   bh = head, block_start = 0;
1102                 ret == 0 && (bh != head || !block_start);
1103                 block_start = block_end, bh = next)
1104         {
1105                 next = bh->b_this_page;
1106                 block_end = block_start + blocksize;
1107                 if (block_end <= from || block_start >= to) {
1108                         if (partial && !buffer_uptodate(bh))
1109                                 *partial = 1;
1110                         continue;
1111                 }
1112                 err = (*fn)(handle, bh);
1113                 if (!ret)
1114                         ret = err;
1115         }
1116         return ret;
1117 }
1118
1119 /*
1120  * To preserve ordering, it is essential that the hole instantiation and
1121  * the data write be encapsulated in a single transaction.  We cannot
1122  * close off a transaction and start a new one between the ext3_get_block()
1123  * and the commit_write().  So doing the journal_start at the start of
1124  * prepare_write() is the right place.
1125  *
1126  * Also, this function can nest inside ext3_writepage() ->
1127  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1128  * has generated enough buffer credits to do the whole page.  So we won't
1129  * block on the journal in that case, which is good, because the caller may
1130  * be PF_MEMALLOC.
1131  *
1132  * By accident, ext3 can be reentered when a transaction is open via
1133  * quota file writes.  If we were to commit the transaction while thus
1134  * reentered, there can be a deadlock - we would be holding a quota
1135  * lock, and the commit would never complete if another thread had a
1136  * transaction open and was blocking on the quota lock - a ranking
1137  * violation.
1138  *
1139  * So what we do is to rely on the fact that journal_stop/journal_start
1140  * will _not_ run commit under these circumstances because handle->h_ref
1141  * is elevated.  We'll still have enough credits for the tiny quotafile
1142  * write.  
1143  */
1144 static int do_journal_get_write_access(handle_t *handle,
1145                                         struct buffer_head *bh)
1146 {
1147         if (!buffer_mapped(bh) || buffer_freed(bh))
1148                 return 0;
1149         return ext3_journal_get_write_access(handle, bh);
1150 }
1151
1152 static int ext3_prepare_write(struct file *file, struct page *page,
1153                               unsigned from, unsigned to)
1154 {
1155         struct inode *inode = page->mapping->host;
1156         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1157         handle_t *handle;
1158         int retries = 0;
1159
1160 retry:
1161         handle = ext3_journal_start(inode, needed_blocks);
1162         if (IS_ERR(handle)) {
1163                 ret = PTR_ERR(handle);
1164                 goto out;
1165         }
1166         if (test_opt(inode->i_sb, NOBH))
1167                 ret = nobh_prepare_write(page, from, to, ext3_get_block);
1168         else
1169                 ret = block_prepare_write(page, from, to, ext3_get_block);
1170         if (ret)
1171                 goto prepare_write_failed;
1172
1173         if (ext3_should_journal_data(inode)) {
1174                 ret = walk_page_buffers(handle, page_buffers(page),
1175                                 from, to, NULL, do_journal_get_write_access);
1176         }
1177 prepare_write_failed:
1178         if (ret)
1179                 ext3_journal_stop(handle);
1180         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1181                 goto retry;
1182 out:
1183         return ret;
1184 }
1185
1186 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1187 {
1188         int err = journal_dirty_data(handle, bh);
1189         if (err)
1190                 ext3_journal_abort_handle(__FUNCTION__, __FUNCTION__,
1191                                                 bh, handle,err);
1192         return err;
1193 }
1194
1195 /* For commit_write() in data=journal mode */
1196 static int commit_write_fn(handle_t *handle, struct buffer_head *bh)
1197 {
1198         if (!buffer_mapped(bh) || buffer_freed(bh))
1199                 return 0;
1200         set_buffer_uptodate(bh);
1201         return ext3_journal_dirty_metadata(handle, bh);
1202 }
1203
1204 /*
1205  * We need to pick up the new inode size which generic_commit_write gave us
1206  * `file' can be NULL - eg, when called from page_symlink().
1207  *
1208  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1209  * buffers are managed internally.
1210  */
1211 static int ext3_ordered_commit_write(struct file *file, struct page *page,
1212                              unsigned from, unsigned to)
1213 {
1214         handle_t *handle = ext3_journal_current_handle();
1215         struct inode *inode = page->mapping->host;
1216         int ret = 0, ret2;
1217
1218         ret = walk_page_buffers(handle, page_buffers(page),
1219                 from, to, NULL, ext3_journal_dirty_data);
1220
1221         if (ret == 0) {
1222                 /*
1223                  * generic_commit_write() will run mark_inode_dirty() if i_size
1224                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1225                  * into that.
1226                  */
1227                 loff_t new_i_size;
1228
1229                 new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1230                 if (new_i_size > EXT3_I(inode)->i_disksize)
1231                         EXT3_I(inode)->i_disksize = new_i_size;
1232                 ret = generic_commit_write(file, page, from, to);
1233         }
1234         ret2 = ext3_journal_stop(handle);
1235         if (!ret)
1236                 ret = ret2;
1237         return ret;
1238 }
1239
1240 static int ext3_writeback_commit_write(struct file *file, struct page *page,
1241                              unsigned from, unsigned to)
1242 {
1243         handle_t *handle = ext3_journal_current_handle();
1244         struct inode *inode = page->mapping->host;
1245         int ret = 0, ret2;
1246         loff_t new_i_size;
1247
1248         new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1249         if (new_i_size > EXT3_I(inode)->i_disksize)
1250                 EXT3_I(inode)->i_disksize = new_i_size;
1251
1252         if (test_opt(inode->i_sb, NOBH))
1253                 ret = nobh_commit_write(file, page, from, to);
1254         else
1255                 ret = generic_commit_write(file, page, from, to);
1256
1257         ret2 = ext3_journal_stop(handle);
1258         if (!ret)
1259                 ret = ret2;
1260         return ret;
1261 }
1262
1263 static int ext3_journalled_commit_write(struct file *file,
1264                         struct page *page, unsigned from, unsigned to)
1265 {
1266         handle_t *handle = ext3_journal_current_handle();
1267         struct inode *inode = page->mapping->host;
1268         int ret = 0, ret2;
1269         int partial = 0;
1270         loff_t pos;
1271
1272         /*
1273          * Here we duplicate the generic_commit_write() functionality
1274          */
1275         pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1276
1277         ret = walk_page_buffers(handle, page_buffers(page), from,
1278                                 to, &partial, commit_write_fn);
1279         if (!partial)
1280                 SetPageUptodate(page);
1281         if (pos > inode->i_size)
1282                 i_size_write(inode, pos);
1283         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1284         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1285                 EXT3_I(inode)->i_disksize = inode->i_size;
1286                 ret2 = ext3_mark_inode_dirty(handle, inode);
1287                 if (!ret) 
1288                         ret = ret2;
1289         }
1290         ret2 = ext3_journal_stop(handle);
1291         if (!ret)
1292                 ret = ret2;
1293         return ret;
1294 }
1295
1296 /* 
1297  * bmap() is special.  It gets used by applications such as lilo and by
1298  * the swapper to find the on-disk block of a specific piece of data.
1299  *
1300  * Naturally, this is dangerous if the block concerned is still in the
1301  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1302  * filesystem and enables swap, then they may get a nasty shock when the
1303  * data getting swapped to that swapfile suddenly gets overwritten by
1304  * the original zero's written out previously to the journal and
1305  * awaiting writeback in the kernel's buffer cache. 
1306  *
1307  * So, if we see any bmap calls here on a modified, data-journaled file,
1308  * take extra steps to flush any blocks which might be in the cache. 
1309  */
1310 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1311 {
1312         struct inode *inode = mapping->host;
1313         journal_t *journal;
1314         int err;
1315
1316         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1317                 /* 
1318                  * This is a REALLY heavyweight approach, but the use of
1319                  * bmap on dirty files is expected to be extremely rare:
1320                  * only if we run lilo or swapon on a freshly made file
1321                  * do we expect this to happen. 
1322                  *
1323                  * (bmap requires CAP_SYS_RAWIO so this does not
1324                  * represent an unprivileged user DOS attack --- we'd be
1325                  * in trouble if mortal users could trigger this path at
1326                  * will.) 
1327                  *
1328                  * NB. EXT3_STATE_JDATA is not set on files other than
1329                  * regular files.  If somebody wants to bmap a directory
1330                  * or symlink and gets confused because the buffer
1331                  * hasn't yet been flushed to disk, they deserve
1332                  * everything they get.
1333                  */
1334
1335                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1336                 journal = EXT3_JOURNAL(inode);
1337                 journal_lock_updates(journal);
1338                 err = journal_flush(journal);
1339                 journal_unlock_updates(journal);
1340
1341                 if (err)
1342                         return 0;
1343         }
1344
1345         return generic_block_bmap(mapping,block,ext3_get_block);
1346 }
1347
1348 static int bget_one(handle_t *handle, struct buffer_head *bh)
1349 {
1350         get_bh(bh);
1351         return 0;
1352 }
1353
1354 static int bput_one(handle_t *handle, struct buffer_head *bh)
1355 {
1356         put_bh(bh);
1357         return 0;
1358 }
1359
1360 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1361 {
1362         if (buffer_mapped(bh))
1363                 return ext3_journal_dirty_data(handle, bh);
1364         return 0;
1365 }
1366
1367 /*
1368  * Note that we always start a transaction even if we're not journalling
1369  * data.  This is to preserve ordering: any hole instantiation within
1370  * __block_write_full_page -> ext3_get_block() should be journalled
1371  * along with the data so we don't crash and then get metadata which
1372  * refers to old data.
1373  *
1374  * In all journalling modes block_write_full_page() will start the I/O.
1375  *
1376  * Problem:
1377  *
1378  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1379  *              ext3_writepage()
1380  *
1381  * Similar for:
1382  *
1383  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1384  *
1385  * Same applies to ext3_get_block().  We will deadlock on various things like
1386  * lock_journal and i_truncate_mutex.
1387  *
1388  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1389  * allocations fail.
1390  *
1391  * 16May01: If we're reentered then journal_current_handle() will be
1392  *          non-zero. We simply *return*.
1393  *
1394  * 1 July 2001: @@@ FIXME:
1395  *   In journalled data mode, a data buffer may be metadata against the
1396  *   current transaction.  But the same file is part of a shared mapping
1397  *   and someone does a writepage() on it.
1398  *
1399  *   We will move the buffer onto the async_data list, but *after* it has
1400  *   been dirtied. So there's a small window where we have dirty data on
1401  *   BJ_Metadata.
1402  *
1403  *   Note that this only applies to the last partial page in the file.  The
1404  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1405  *   broken code anyway: it's wrong for msync()).
1406  *
1407  *   It's a rare case: affects the final partial page, for journalled data
1408  *   where the file is subject to bith write() and writepage() in the same
1409  *   transction.  To fix it we'll need a custom block_write_full_page().
1410  *   We'll probably need that anyway for journalling writepage() output.
1411  *
1412  * We don't honour synchronous mounts for writepage().  That would be
1413  * disastrous.  Any write() or metadata operation will sync the fs for
1414  * us.
1415  *
1416  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1417  * we don't need to open a transaction here.
1418  */
1419 static int ext3_ordered_writepage(struct page *page,
1420                                 struct writeback_control *wbc)
1421 {
1422         struct inode *inode = page->mapping->host;
1423         struct buffer_head *page_bufs;
1424         handle_t *handle = NULL;
1425         int ret = 0;
1426         int err;
1427
1428         J_ASSERT(PageLocked(page));
1429
1430         /*
1431          * We give up here if we're reentered, because it might be for a
1432          * different filesystem.
1433          */
1434         if (ext3_journal_current_handle())
1435                 goto out_fail;
1436
1437         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1438
1439         if (IS_ERR(handle)) {
1440                 ret = PTR_ERR(handle);
1441                 goto out_fail;
1442         }
1443
1444         if (!page_has_buffers(page)) {
1445                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1446                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1447         }
1448         page_bufs = page_buffers(page);
1449         walk_page_buffers(handle, page_bufs, 0,
1450                         PAGE_CACHE_SIZE, NULL, bget_one);
1451
1452         ret = block_write_full_page(page, ext3_get_block, wbc);
1453
1454         /*
1455          * The page can become unlocked at any point now, and
1456          * truncate can then come in and change things.  So we
1457          * can't touch *page from now on.  But *page_bufs is
1458          * safe due to elevated refcount.
1459          */
1460
1461         /*
1462          * And attach them to the current transaction.  But only if 
1463          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1464          * and generally junk.
1465          */
1466         if (ret == 0) {
1467                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1468                                         NULL, journal_dirty_data_fn);
1469                 if (!ret)
1470                         ret = err;
1471         }
1472         walk_page_buffers(handle, page_bufs, 0,
1473                         PAGE_CACHE_SIZE, NULL, bput_one);
1474         err = ext3_journal_stop(handle);
1475         if (!ret)
1476                 ret = err;
1477         return ret;
1478
1479 out_fail:
1480         redirty_page_for_writepage(wbc, page);
1481         unlock_page(page);
1482         return ret;
1483 }
1484
1485 static int ext3_writeback_writepage(struct page *page,
1486                                 struct writeback_control *wbc)
1487 {
1488         struct inode *inode = page->mapping->host;
1489         handle_t *handle = NULL;
1490         int ret = 0;
1491         int err;
1492
1493         if (ext3_journal_current_handle())
1494                 goto out_fail;
1495
1496         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1497         if (IS_ERR(handle)) {
1498                 ret = PTR_ERR(handle);
1499                 goto out_fail;
1500         }
1501
1502         if (test_opt(inode->i_sb, NOBH))
1503                 ret = nobh_writepage(page, ext3_get_block, wbc);
1504         else
1505                 ret = block_write_full_page(page, ext3_get_block, wbc);
1506
1507         err = ext3_journal_stop(handle);
1508         if (!ret)
1509                 ret = err;
1510         return ret;
1511
1512 out_fail:
1513         redirty_page_for_writepage(wbc, page);
1514         unlock_page(page);
1515         return ret;
1516 }
1517
1518 static int ext3_journalled_writepage(struct page *page,
1519                                 struct writeback_control *wbc)
1520 {
1521         struct inode *inode = page->mapping->host;
1522         handle_t *handle = NULL;
1523         int ret = 0;
1524         int err;
1525
1526         if (ext3_journal_current_handle())
1527                 goto no_write;
1528
1529         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1530         if (IS_ERR(handle)) {
1531                 ret = PTR_ERR(handle);
1532                 goto no_write;
1533         }
1534
1535         if (!page_has_buffers(page) || PageChecked(page)) {
1536                 /*
1537                  * It's mmapped pagecache.  Add buffers and journal it.  There
1538                  * doesn't seem much point in redirtying the page here.
1539                  */
1540                 ClearPageChecked(page);
1541                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1542                                         ext3_get_block);
1543                 if (ret != 0) {
1544                         ext3_journal_stop(handle);
1545                         goto out_unlock;
1546                 }
1547                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1548                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1549
1550                 err = walk_page_buffers(handle, page_buffers(page), 0,
1551                                 PAGE_CACHE_SIZE, NULL, commit_write_fn);
1552                 if (ret == 0)
1553                         ret = err;
1554                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1555                 unlock_page(page);
1556         } else {
1557                 /*
1558                  * It may be a page full of checkpoint-mode buffers.  We don't
1559                  * really know unless we go poke around in the buffer_heads.
1560                  * But block_write_full_page will do the right thing.
1561                  */
1562                 ret = block_write_full_page(page, ext3_get_block, wbc);
1563         }
1564         err = ext3_journal_stop(handle);
1565         if (!ret)
1566                 ret = err;
1567 out:
1568         return ret;
1569
1570 no_write:
1571         redirty_page_for_writepage(wbc, page);
1572 out_unlock:
1573         unlock_page(page);
1574         goto out;
1575 }
1576
1577 static int ext3_readpage(struct file *file, struct page *page)
1578 {
1579         return mpage_readpage(page, ext3_get_block);
1580 }
1581
1582 static int
1583 ext3_readpages(struct file *file, struct address_space *mapping,
1584                 struct list_head *pages, unsigned nr_pages)
1585 {
1586         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1587 }
1588
1589 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1590 {
1591         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1592
1593         /*
1594          * If it's a full truncate we just forget about the pending dirtying
1595          */
1596         if (offset == 0)
1597                 ClearPageChecked(page);
1598
1599         journal_invalidatepage(journal, page, offset);
1600 }
1601
1602 static int ext3_releasepage(struct page *page, gfp_t wait)
1603 {
1604         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1605
1606         WARN_ON(PageChecked(page));
1607         if (!page_has_buffers(page))
1608                 return 0;
1609         return journal_try_to_free_buffers(journal, page, wait);
1610 }
1611
1612 /*
1613  * If the O_DIRECT write will extend the file then add this inode to the
1614  * orphan list.  So recovery will truncate it back to the original size
1615  * if the machine crashes during the write.
1616  *
1617  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1618  * crashes then stale disk data _may_ be exposed inside the file.
1619  */
1620 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1621                         const struct iovec *iov, loff_t offset,
1622                         unsigned long nr_segs)
1623 {
1624         struct file *file = iocb->ki_filp;
1625         struct inode *inode = file->f_mapping->host;
1626         struct ext3_inode_info *ei = EXT3_I(inode);
1627         handle_t *handle = NULL;
1628         ssize_t ret;
1629         int orphan = 0;
1630         size_t count = iov_length(iov, nr_segs);
1631
1632         if (rw == WRITE) {
1633                 loff_t final_size = offset + count;
1634
1635                 handle = ext3_journal_start(inode, DIO_CREDITS);
1636                 if (IS_ERR(handle)) {
1637                         ret = PTR_ERR(handle);
1638                         goto out;
1639                 }
1640                 if (final_size > inode->i_size) {
1641                         ret = ext3_orphan_add(handle, inode);
1642                         if (ret)
1643                                 goto out_stop;
1644                         orphan = 1;
1645                         ei->i_disksize = inode->i_size;
1646                 }
1647         }
1648
1649         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov, 
1650                                  offset, nr_segs,
1651                                  ext3_direct_io_get_blocks, NULL);
1652
1653         /*
1654          * Reacquire the handle: ext3_direct_io_get_block() can restart the
1655          * transaction
1656          */
1657         handle = journal_current_handle();
1658
1659 out_stop:
1660         if (handle) {
1661                 int err;
1662
1663                 if (orphan && inode->i_nlink)
1664                         ext3_orphan_del(handle, inode);
1665                 if (orphan && ret > 0) {
1666                         loff_t end = offset + ret;
1667                         if (end > inode->i_size) {
1668                                 ei->i_disksize = end;
1669                                 i_size_write(inode, end);
1670                                 /*
1671                                  * We're going to return a positive `ret'
1672                                  * here due to non-zero-length I/O, so there's
1673                                  * no way of reporting error returns from
1674                                  * ext3_mark_inode_dirty() to userspace.  So
1675                                  * ignore it.
1676                                  */
1677                                 ext3_mark_inode_dirty(handle, inode);
1678                         }
1679                 }
1680                 err = ext3_journal_stop(handle);
1681                 if (ret == 0)
1682                         ret = err;
1683         }
1684 out:
1685         return ret;
1686 }
1687
1688 /*
1689  * Pages can be marked dirty completely asynchronously from ext3's journalling
1690  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1691  * much here because ->set_page_dirty is called under VFS locks.  The page is
1692  * not necessarily locked.
1693  *
1694  * We cannot just dirty the page and leave attached buffers clean, because the
1695  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1696  * or jbddirty because all the journalling code will explode.
1697  *
1698  * So what we do is to mark the page "pending dirty" and next time writepage
1699  * is called, propagate that into the buffers appropriately.
1700  */
1701 static int ext3_journalled_set_page_dirty(struct page *page)
1702 {
1703         SetPageChecked(page);
1704         return __set_page_dirty_nobuffers(page);
1705 }
1706
1707 static struct address_space_operations ext3_ordered_aops = {
1708         .readpage       = ext3_readpage,
1709         .readpages      = ext3_readpages,
1710         .writepage      = ext3_ordered_writepage,
1711         .sync_page      = block_sync_page,
1712         .prepare_write  = ext3_prepare_write,
1713         .commit_write   = ext3_ordered_commit_write,
1714         .bmap           = ext3_bmap,
1715         .invalidatepage = ext3_invalidatepage,
1716         .releasepage    = ext3_releasepage,
1717         .direct_IO      = ext3_direct_IO,
1718         .migratepage    = buffer_migrate_page,
1719 };
1720
1721 static struct address_space_operations ext3_writeback_aops = {
1722         .readpage       = ext3_readpage,
1723         .readpages      = ext3_readpages,
1724         .writepage      = ext3_writeback_writepage,
1725         .sync_page      = block_sync_page,
1726         .prepare_write  = ext3_prepare_write,
1727         .commit_write   = ext3_writeback_commit_write,
1728         .bmap           = ext3_bmap,
1729         .invalidatepage = ext3_invalidatepage,
1730         .releasepage    = ext3_releasepage,
1731         .direct_IO      = ext3_direct_IO,
1732         .migratepage    = buffer_migrate_page,
1733 };
1734
1735 static struct address_space_operations ext3_journalled_aops = {
1736         .readpage       = ext3_readpage,
1737         .readpages      = ext3_readpages,
1738         .writepage      = ext3_journalled_writepage,
1739         .sync_page      = block_sync_page,
1740         .prepare_write  = ext3_prepare_write,
1741         .commit_write   = ext3_journalled_commit_write,
1742         .set_page_dirty = ext3_journalled_set_page_dirty,
1743         .bmap           = ext3_bmap,
1744         .invalidatepage = ext3_invalidatepage,
1745         .releasepage    = ext3_releasepage,
1746 };
1747
1748 void ext3_set_aops(struct inode *inode)
1749 {
1750         if (ext3_should_order_data(inode))
1751                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1752         else if (ext3_should_writeback_data(inode))
1753                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1754         else
1755                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1756 }
1757
1758 /*
1759  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1760  * up to the end of the block which corresponds to `from'.
1761  * This required during truncate. We need to physically zero the tail end
1762  * of that block so it doesn't yield old data if the file is later grown.
1763  */
1764 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1765                 struct address_space *mapping, loff_t from)
1766 {
1767         unsigned long index = from >> PAGE_CACHE_SHIFT;
1768         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1769         unsigned blocksize, iblock, length, pos;
1770         struct inode *inode = mapping->host;
1771         struct buffer_head *bh;
1772         int err = 0;
1773         void *kaddr;
1774
1775         blocksize = inode->i_sb->s_blocksize;
1776         length = blocksize - (offset & (blocksize - 1));
1777         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1778
1779         /*
1780          * For "nobh" option,  we can only work if we don't need to
1781          * read-in the page - otherwise we create buffers to do the IO.
1782          */
1783         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1784              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1785                 kaddr = kmap_atomic(page, KM_USER0);
1786                 memset(kaddr + offset, 0, length);
1787                 flush_dcache_page(page);
1788                 kunmap_atomic(kaddr, KM_USER0);
1789                 set_page_dirty(page);
1790                 goto unlock;
1791         }
1792
1793         if (!page_has_buffers(page))
1794                 create_empty_buffers(page, blocksize, 0);
1795
1796         /* Find the buffer that contains "offset" */
1797         bh = page_buffers(page);
1798         pos = blocksize;
1799         while (offset >= pos) {
1800                 bh = bh->b_this_page;
1801                 iblock++;
1802                 pos += blocksize;
1803         }
1804
1805         err = 0;
1806         if (buffer_freed(bh)) {
1807                 BUFFER_TRACE(bh, "freed: skip");
1808                 goto unlock;
1809         }
1810
1811         if (!buffer_mapped(bh)) {
1812                 BUFFER_TRACE(bh, "unmapped");
1813                 ext3_get_block(inode, iblock, bh, 0);
1814                 /* unmapped? It's a hole - nothing to do */
1815                 if (!buffer_mapped(bh)) {
1816                         BUFFER_TRACE(bh, "still unmapped");
1817                         goto unlock;
1818                 }
1819         }
1820
1821         /* Ok, it's mapped. Make sure it's up-to-date */
1822         if (PageUptodate(page))
1823                 set_buffer_uptodate(bh);
1824
1825         if (!buffer_uptodate(bh)) {
1826                 err = -EIO;
1827                 ll_rw_block(READ, 1, &bh);
1828                 wait_on_buffer(bh);
1829                 /* Uhhuh. Read error. Complain and punt. */
1830                 if (!buffer_uptodate(bh))
1831                         goto unlock;
1832         }
1833
1834         if (ext3_should_journal_data(inode)) {
1835                 BUFFER_TRACE(bh, "get write access");
1836                 err = ext3_journal_get_write_access(handle, bh);
1837                 if (err)
1838                         goto unlock;
1839         }
1840
1841         kaddr = kmap_atomic(page, KM_USER0);
1842         memset(kaddr + offset, 0, length);
1843         flush_dcache_page(page);
1844         kunmap_atomic(kaddr, KM_USER0);
1845
1846         BUFFER_TRACE(bh, "zeroed end of block");
1847
1848         err = 0;
1849         if (ext3_should_journal_data(inode)) {
1850                 err = ext3_journal_dirty_metadata(handle, bh);
1851         } else {
1852                 if (ext3_should_order_data(inode))
1853                         err = ext3_journal_dirty_data(handle, bh);
1854                 mark_buffer_dirty(bh);
1855         }
1856
1857 unlock:
1858         unlock_page(page);
1859         page_cache_release(page);
1860         return err;
1861 }
1862
1863 /*
1864  * Probably it should be a library function... search for first non-zero word
1865  * or memcmp with zero_page, whatever is better for particular architecture.
1866  * Linus?
1867  */
1868 static inline int all_zeroes(__le32 *p, __le32 *q)
1869 {
1870         while (p < q)
1871                 if (*p++)
1872                         return 0;
1873         return 1;
1874 }
1875
1876 /**
1877  *      ext3_find_shared - find the indirect blocks for partial truncation.
1878  *      @inode:   inode in question
1879  *      @depth:   depth of the affected branch
1880  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1881  *      @chain:   place to store the pointers to partial indirect blocks
1882  *      @top:     place to the (detached) top of branch
1883  *
1884  *      This is a helper function used by ext3_truncate().
1885  *
1886  *      When we do truncate() we may have to clean the ends of several
1887  *      indirect blocks but leave the blocks themselves alive. Block is
1888  *      partially truncated if some data below the new i_size is refered
1889  *      from it (and it is on the path to the first completely truncated
1890  *      data block, indeed).  We have to free the top of that path along
1891  *      with everything to the right of the path. Since no allocation
1892  *      past the truncation point is possible until ext3_truncate()
1893  *      finishes, we may safely do the latter, but top of branch may
1894  *      require special attention - pageout below the truncation point
1895  *      might try to populate it.
1896  *
1897  *      We atomically detach the top of branch from the tree, store the
1898  *      block number of its root in *@top, pointers to buffer_heads of
1899  *      partially truncated blocks - in @chain[].bh and pointers to
1900  *      their last elements that should not be removed - in
1901  *      @chain[].p. Return value is the pointer to last filled element
1902  *      of @chain.
1903  *
1904  *      The work left to caller to do the actual freeing of subtrees:
1905  *              a) free the subtree starting from *@top
1906  *              b) free the subtrees whose roots are stored in
1907  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1908  *              c) free the subtrees growing from the inode past the @chain[0].
1909  *                      (no partially truncated stuff there).  */
1910
1911 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1912                         int offsets[4], Indirect chain[4], __le32 *top)
1913 {
1914         Indirect *partial, *p;
1915         int k, err;
1916
1917         *top = 0;
1918         /* Make k index the deepest non-null offest + 1 */
1919         for (k = depth; k > 1 && !offsets[k-1]; k--)
1920                 ;
1921         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1922         /* Writer: pointers */
1923         if (!partial)
1924                 partial = chain + k-1;
1925         /*
1926          * If the branch acquired continuation since we've looked at it -
1927          * fine, it should all survive and (new) top doesn't belong to us.
1928          */
1929         if (!partial->key && *partial->p)
1930                 /* Writer: end */
1931                 goto no_top;
1932         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1933                 ;
1934         /*
1935          * OK, we've found the last block that must survive. The rest of our
1936          * branch should be detached before unlocking. However, if that rest
1937          * of branch is all ours and does not grow immediately from the inode
1938          * it's easier to cheat and just decrement partial->p.
1939          */
1940         if (p == chain + k - 1 && p > chain) {
1941                 p->p--;
1942         } else {
1943                 *top = *p->p;
1944                 /* Nope, don't do this in ext3.  Must leave the tree intact */
1945 #if 0
1946                 *p->p = 0;
1947 #endif
1948         }
1949         /* Writer: end */
1950
1951         while(partial > p) {
1952                 brelse(partial->bh);
1953                 partial--;
1954         }
1955 no_top:
1956         return partial;
1957 }
1958
1959 /*
1960  * Zero a number of block pointers in either an inode or an indirect block.
1961  * If we restart the transaction we must again get write access to the
1962  * indirect block for further modification.
1963  *
1964  * We release `count' blocks on disk, but (last - first) may be greater
1965  * than `count' because there can be holes in there.
1966  */
1967 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
1968                 struct buffer_head *bh, unsigned long block_to_free,
1969                 unsigned long count, __le32 *first, __le32 *last)
1970 {
1971         __le32 *p;
1972         if (try_to_extend_transaction(handle, inode)) {
1973                 if (bh) {
1974                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1975                         ext3_journal_dirty_metadata(handle, bh);
1976                 }
1977                 ext3_mark_inode_dirty(handle, inode);
1978                 ext3_journal_test_restart(handle, inode);
1979                 if (bh) {
1980                         BUFFER_TRACE(bh, "retaking write access");
1981                         ext3_journal_get_write_access(handle, bh);
1982                 }
1983         }
1984
1985         /*
1986          * Any buffers which are on the journal will be in memory. We find
1987          * them on the hash table so journal_revoke() will run journal_forget()
1988          * on them.  We've already detached each block from the file, so
1989          * bforget() in journal_forget() should be safe.
1990          *
1991          * AKPM: turn on bforget in journal_forget()!!!
1992          */
1993         for (p = first; p < last; p++) {
1994                 u32 nr = le32_to_cpu(*p);
1995                 if (nr) {
1996                         struct buffer_head *bh;
1997
1998                         *p = 0;
1999                         bh = sb_find_get_block(inode->i_sb, nr);
2000                         ext3_forget(handle, 0, inode, bh, nr);
2001                 }
2002         }
2003
2004         ext3_free_blocks(handle, inode, block_to_free, count);
2005 }
2006
2007 /**
2008  * ext3_free_data - free a list of data blocks
2009  * @handle:     handle for this transaction
2010  * @inode:      inode we are dealing with
2011  * @this_bh:    indirect buffer_head which contains *@first and *@last
2012  * @first:      array of block numbers
2013  * @last:       points immediately past the end of array
2014  *
2015  * We are freeing all blocks refered from that array (numbers are stored as
2016  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2017  *
2018  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2019  * blocks are contiguous then releasing them at one time will only affect one
2020  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2021  * actually use a lot of journal space.
2022  *
2023  * @this_bh will be %NULL if @first and @last point into the inode's direct
2024  * block pointers.
2025  */
2026 static void ext3_free_data(handle_t *handle, struct inode *inode,
2027                            struct buffer_head *this_bh,
2028                            __le32 *first, __le32 *last)
2029 {
2030         unsigned long block_to_free = 0;    /* Starting block # of a run */
2031         unsigned long count = 0;            /* Number of blocks in the run */ 
2032         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2033                                                corresponding to
2034                                                block_to_free */
2035         unsigned long nr;                   /* Current block # */
2036         __le32 *p;                          /* Pointer into inode/ind
2037                                                for current block */
2038         int err;
2039
2040         if (this_bh) {                          /* For indirect block */
2041                 BUFFER_TRACE(this_bh, "get_write_access");
2042                 err = ext3_journal_get_write_access(handle, this_bh);
2043                 /* Important: if we can't update the indirect pointers
2044                  * to the blocks, we can't free them. */
2045                 if (err)
2046                         return;
2047         }
2048
2049         for (p = first; p < last; p++) {
2050                 nr = le32_to_cpu(*p);
2051                 if (nr) {
2052                         /* accumulate blocks to free if they're contiguous */
2053                         if (count == 0) {
2054                                 block_to_free = nr;
2055                                 block_to_free_p = p;
2056                                 count = 1;
2057                         } else if (nr == block_to_free + count) {
2058                                 count++;
2059                         } else {
2060                                 ext3_clear_blocks(handle, inode, this_bh, 
2061                                                   block_to_free,
2062                                                   count, block_to_free_p, p);
2063                                 block_to_free = nr;
2064                                 block_to_free_p = p;
2065                                 count = 1;
2066                         }
2067                 }
2068         }
2069
2070         if (count > 0)
2071                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2072                                   count, block_to_free_p, p);
2073
2074         if (this_bh) {
2075                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2076                 ext3_journal_dirty_metadata(handle, this_bh);
2077         }
2078 }
2079
2080 /**
2081  *      ext3_free_branches - free an array of branches
2082  *      @handle: JBD handle for this transaction
2083  *      @inode: inode we are dealing with
2084  *      @parent_bh: the buffer_head which contains *@first and *@last
2085  *      @first: array of block numbers
2086  *      @last:  pointer immediately past the end of array
2087  *      @depth: depth of the branches to free
2088  *
2089  *      We are freeing all blocks refered from these branches (numbers are
2090  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2091  *      appropriately.
2092  */
2093 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2094                                struct buffer_head *parent_bh,
2095                                __le32 *first, __le32 *last, int depth)
2096 {
2097         unsigned long nr;
2098         __le32 *p;
2099
2100         if (is_handle_aborted(handle))
2101                 return;
2102
2103         if (depth--) {
2104                 struct buffer_head *bh;
2105                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2106                 p = last;
2107                 while (--p >= first) {
2108                         nr = le32_to_cpu(*p);
2109                         if (!nr)
2110                                 continue;               /* A hole */
2111
2112                         /* Go read the buffer for the next level down */
2113                         bh = sb_bread(inode->i_sb, nr);
2114
2115                         /*
2116                          * A read failure? Report error and clear slot
2117                          * (should be rare).
2118                          */
2119                         if (!bh) {
2120                                 ext3_error(inode->i_sb, "ext3_free_branches",
2121                                            "Read failure, inode=%ld, block=%ld",
2122                                            inode->i_ino, nr);
2123                                 continue;
2124                         }
2125
2126                         /* This zaps the entire block.  Bottom up. */
2127                         BUFFER_TRACE(bh, "free child branches");
2128                         ext3_free_branches(handle, inode, bh,
2129                                            (__le32*)bh->b_data,
2130                                            (__le32*)bh->b_data + addr_per_block,
2131                                            depth);
2132
2133                         /*
2134                          * We've probably journalled the indirect block several
2135                          * times during the truncate.  But it's no longer
2136                          * needed and we now drop it from the transaction via
2137                          * journal_revoke().
2138                          *
2139                          * That's easy if it's exclusively part of this
2140                          * transaction.  But if it's part of the committing
2141                          * transaction then journal_forget() will simply
2142                          * brelse() it.  That means that if the underlying
2143                          * block is reallocated in ext3_get_block(),
2144                          * unmap_underlying_metadata() will find this block
2145                          * and will try to get rid of it.  damn, damn.
2146                          *
2147                          * If this block has already been committed to the
2148                          * journal, a revoke record will be written.  And
2149                          * revoke records must be emitted *before* clearing
2150                          * this block's bit in the bitmaps.
2151                          */
2152                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2153
2154                         /*
2155                          * Everything below this this pointer has been
2156                          * released.  Now let this top-of-subtree go.
2157                          *
2158                          * We want the freeing of this indirect block to be
2159                          * atomic in the journal with the updating of the
2160                          * bitmap block which owns it.  So make some room in
2161                          * the journal.
2162                          *
2163                          * We zero the parent pointer *after* freeing its
2164                          * pointee in the bitmaps, so if extend_transaction()
2165                          * for some reason fails to put the bitmap changes and
2166                          * the release into the same transaction, recovery
2167                          * will merely complain about releasing a free block,
2168                          * rather than leaking blocks.
2169                          */
2170                         if (is_handle_aborted(handle))
2171                                 return;
2172                         if (try_to_extend_transaction(handle, inode)) {
2173                                 ext3_mark_inode_dirty(handle, inode);
2174                                 ext3_journal_test_restart(handle, inode);
2175                         }
2176
2177                         ext3_free_blocks(handle, inode, nr, 1);
2178
2179                         if (parent_bh) {
2180                                 /*
2181                                  * The block which we have just freed is
2182                                  * pointed to by an indirect block: journal it
2183                                  */
2184                                 BUFFER_TRACE(parent_bh, "get_write_access");
2185                                 if (!ext3_journal_get_write_access(handle,
2186                                                                    parent_bh)){
2187                                         *p = 0;
2188                                         BUFFER_TRACE(parent_bh,
2189                                         "call ext3_journal_dirty_metadata");
2190                                         ext3_journal_dirty_metadata(handle, 
2191                                                                     parent_bh);
2192                                 }
2193                         }
2194                 }
2195         } else {
2196                 /* We have reached the bottom of the tree. */
2197                 BUFFER_TRACE(parent_bh, "free data blocks");
2198                 ext3_free_data(handle, inode, parent_bh, first, last);
2199         }
2200 }
2201
2202 /*
2203  * ext3_truncate()
2204  *
2205  * We block out ext3_get_block() block instantiations across the entire
2206  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2207  * simultaneously on behalf of the same inode.
2208  *
2209  * As we work through the truncate and commmit bits of it to the journal there
2210  * is one core, guiding principle: the file's tree must always be consistent on
2211  * disk.  We must be able to restart the truncate after a crash.
2212  *
2213  * The file's tree may be transiently inconsistent in memory (although it
2214  * probably isn't), but whenever we close off and commit a journal transaction,
2215  * the contents of (the filesystem + the journal) must be consistent and
2216  * restartable.  It's pretty simple, really: bottom up, right to left (although
2217  * left-to-right works OK too).
2218  *
2219  * Note that at recovery time, journal replay occurs *before* the restart of
2220  * truncate against the orphan inode list.
2221  *
2222  * The committed inode has the new, desired i_size (which is the same as
2223  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2224  * that this inode's truncate did not complete and it will again call
2225  * ext3_truncate() to have another go.  So there will be instantiated blocks
2226  * to the right of the truncation point in a crashed ext3 filesystem.  But
2227  * that's fine - as long as they are linked from the inode, the post-crash
2228  * ext3_truncate() run will find them and release them.
2229  */
2230 void ext3_truncate(struct inode *inode)
2231 {
2232         handle_t *handle;
2233         struct ext3_inode_info *ei = EXT3_I(inode);
2234         __le32 *i_data = ei->i_data;
2235         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2236         struct address_space *mapping = inode->i_mapping;
2237         int offsets[4];
2238         Indirect chain[4];
2239         Indirect *partial;
2240         __le32 nr = 0;
2241         int n;
2242         long last_block;
2243         unsigned blocksize = inode->i_sb->s_blocksize;
2244         struct page *page;
2245
2246         if (!(S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
2247             S_ISLNK(inode->i_mode)))
2248                 return;
2249         if (ext3_inode_is_fast_symlink(inode))
2250                 return;
2251         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2252                 return;
2253
2254         /*
2255          * We have to lock the EOF page here, because lock_page() nests
2256          * outside journal_start().
2257          */
2258         if ((inode->i_size & (blocksize - 1)) == 0) {
2259                 /* Block boundary? Nothing to do */
2260                 page = NULL;
2261         } else {
2262                 page = grab_cache_page(mapping,
2263                                 inode->i_size >> PAGE_CACHE_SHIFT);
2264                 if (!page)
2265                         return;
2266         }
2267
2268         handle = start_transaction(inode);
2269         if (IS_ERR(handle)) {
2270                 if (page) {
2271                         clear_highpage(page);
2272                         flush_dcache_page(page);
2273                         unlock_page(page);
2274                         page_cache_release(page);
2275                 }
2276                 return;         /* AKPM: return what? */
2277         }
2278
2279         last_block = (inode->i_size + blocksize-1)
2280                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2281
2282         if (page)
2283                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2284
2285         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2286         if (n == 0)
2287                 goto out_stop;  /* error */
2288
2289         /*
2290          * OK.  This truncate is going to happen.  We add the inode to the
2291          * orphan list, so that if this truncate spans multiple transactions,
2292          * and we crash, we will resume the truncate when the filesystem
2293          * recovers.  It also marks the inode dirty, to catch the new size.
2294          *
2295          * Implication: the file must always be in a sane, consistent
2296          * truncatable state while each transaction commits.
2297          */
2298         if (ext3_orphan_add(handle, inode))
2299                 goto out_stop;
2300
2301         /*
2302          * The orphan list entry will now protect us from any crash which
2303          * occurs before the truncate completes, so it is now safe to propagate
2304          * the new, shorter inode size (held for now in i_size) into the
2305          * on-disk inode. We do this via i_disksize, which is the value which
2306          * ext3 *really* writes onto the disk inode.
2307          */
2308         ei->i_disksize = inode->i_size;
2309
2310         /*
2311          * From here we block out all ext3_get_block() callers who want to
2312          * modify the block allocation tree.
2313          */
2314         mutex_lock(&ei->truncate_mutex);
2315
2316         if (n == 1) {           /* direct blocks */
2317                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2318                                i_data + EXT3_NDIR_BLOCKS);
2319                 goto do_indirects;
2320         }
2321
2322         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2323         /* Kill the top of shared branch (not detached) */
2324         if (nr) {
2325                 if (partial == chain) {
2326                         /* Shared branch grows from the inode */
2327                         ext3_free_branches(handle, inode, NULL,
2328                                            &nr, &nr+1, (chain+n-1) - partial);
2329                         *partial->p = 0;
2330                         /*
2331                          * We mark the inode dirty prior to restart,
2332                          * and prior to stop.  No need for it here.
2333                          */
2334                 } else {
2335                         /* Shared branch grows from an indirect block */
2336                         BUFFER_TRACE(partial->bh, "get_write_access");
2337                         ext3_free_branches(handle, inode, partial->bh,
2338                                         partial->p,
2339                                         partial->p+1, (chain+n-1) - partial);
2340                 }
2341         }
2342         /* Clear the ends of indirect blocks on the shared branch */
2343         while (partial > chain) {
2344                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2345                                    (__le32*)partial->bh->b_data+addr_per_block,
2346                                    (chain+n-1) - partial);
2347                 BUFFER_TRACE(partial->bh, "call brelse");
2348                 brelse (partial->bh);
2349                 partial--;
2350         }
2351 do_indirects:
2352         /* Kill the remaining (whole) subtrees */
2353         switch (offsets[0]) {
2354         default:
2355                 nr = i_data[EXT3_IND_BLOCK];
2356                 if (nr) {
2357                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2358                         i_data[EXT3_IND_BLOCK] = 0;
2359                 }
2360         case EXT3_IND_BLOCK:
2361                 nr = i_data[EXT3_DIND_BLOCK];
2362                 if (nr) {
2363                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2364                         i_data[EXT3_DIND_BLOCK] = 0;
2365                 }
2366         case EXT3_DIND_BLOCK:
2367                 nr = i_data[EXT3_TIND_BLOCK];
2368                 if (nr) {
2369                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2370                         i_data[EXT3_TIND_BLOCK] = 0;
2371                 }
2372         case EXT3_TIND_BLOCK:
2373                 ;
2374         }
2375
2376         ext3_discard_reservation(inode);
2377
2378         mutex_unlock(&ei->truncate_mutex);
2379         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2380         ext3_mark_inode_dirty(handle, inode);
2381
2382         /*
2383          * In a multi-transaction truncate, we only make the final transaction
2384          * synchronous
2385          */
2386         if (IS_SYNC(inode))
2387                 handle->h_sync = 1;
2388 out_stop:
2389         /*
2390          * If this was a simple ftruncate(), and the file will remain alive
2391          * then we need to clear up the orphan record which we created above.
2392          * However, if this was a real unlink then we were called by
2393          * ext3_delete_inode(), and we allow that function to clean up the
2394          * orphan info for us.
2395          */
2396         if (inode->i_nlink)
2397                 ext3_orphan_del(handle, inode);
2398
2399         ext3_journal_stop(handle);
2400 }
2401
2402 static unsigned long ext3_get_inode_block(struct super_block *sb,
2403                 unsigned long ino, struct ext3_iloc *iloc)
2404 {
2405         unsigned long desc, group_desc, block_group;
2406         unsigned long offset, block;
2407         struct buffer_head *bh;
2408         struct ext3_group_desc * gdp;
2409
2410
2411         if ((ino != EXT3_ROOT_INO && ino != EXT3_JOURNAL_INO &&
2412                 ino != EXT3_RESIZE_INO && ino < EXT3_FIRST_INO(sb)) ||
2413                 ino > le32_to_cpu(EXT3_SB(sb)->s_es->s_inodes_count)) {
2414                 ext3_error(sb, "ext3_get_inode_block",
2415                             "bad inode number: %lu", ino);
2416                 return 0;
2417         }
2418         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2419         if (block_group >= EXT3_SB(sb)->s_groups_count) {
2420                 ext3_error(sb,"ext3_get_inode_block","group >= groups count");
2421                 return 0;
2422         }
2423         smp_rmb();
2424         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
2425         desc = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
2426         bh = EXT3_SB(sb)->s_group_desc[group_desc];
2427         if (!bh) {
2428                 ext3_error (sb, "ext3_get_inode_block",
2429                             "Descriptor not loaded");
2430                 return 0;
2431         }
2432
2433         gdp = (struct ext3_group_desc *)bh->b_data;
2434         /*
2435          * Figure out the offset within the block group inode table
2436          */
2437         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2438                 EXT3_INODE_SIZE(sb);
2439         block = le32_to_cpu(gdp[desc].bg_inode_table) +
2440                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2441
2442         iloc->block_group = block_group;
2443         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2444         return block;
2445 }
2446
2447 /*
2448  * ext3_get_inode_loc returns with an extra refcount against the inode's
2449  * underlying buffer_head on success. If 'in_mem' is true, we have all
2450  * data in memory that is needed to recreate the on-disk version of this
2451  * inode.
2452  */
2453 static int __ext3_get_inode_loc(struct inode *inode,
2454                                 struct ext3_iloc *iloc, int in_mem)
2455 {
2456         unsigned long block;
2457         struct buffer_head *bh;
2458
2459         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2460         if (!block)
2461                 return -EIO;
2462
2463         bh = sb_getblk(inode->i_sb, block);
2464         if (!bh) {
2465                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2466                                 "unable to read inode block - "
2467                                 "inode=%lu, block=%lu", inode->i_ino, block);
2468                 return -EIO;
2469         }
2470         if (!buffer_uptodate(bh)) {
2471                 lock_buffer(bh);
2472                 if (buffer_uptodate(bh)) {
2473                         /* someone brought it uptodate while we waited */
2474                         unlock_buffer(bh);
2475                         goto has_buffer;
2476                 }
2477
2478                 /*
2479                  * If we have all information of the inode in memory and this
2480                  * is the only valid inode in the block, we need not read the
2481                  * block.
2482                  */
2483                 if (in_mem) {
2484                         struct buffer_head *bitmap_bh;
2485                         struct ext3_group_desc *desc;
2486                         int inodes_per_buffer;
2487                         int inode_offset, i;
2488                         int block_group;
2489                         int start;
2490
2491                         block_group = (inode->i_ino - 1) /
2492                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2493                         inodes_per_buffer = bh->b_size /
2494                                 EXT3_INODE_SIZE(inode->i_sb);
2495                         inode_offset = ((inode->i_ino - 1) %
2496                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2497                         start = inode_offset & ~(inodes_per_buffer - 1);
2498
2499                         /* Is the inode bitmap in cache? */
2500                         desc = ext3_get_group_desc(inode->i_sb,
2501                                                 block_group, NULL);
2502                         if (!desc)
2503                                 goto make_io;
2504
2505                         bitmap_bh = sb_getblk(inode->i_sb,
2506                                         le32_to_cpu(desc->bg_inode_bitmap));
2507                         if (!bitmap_bh)
2508                                 goto make_io;
2509
2510                         /*
2511                          * If the inode bitmap isn't in cache then the
2512                          * optimisation may end up performing two reads instead
2513                          * of one, so skip it.
2514                          */
2515                         if (!buffer_uptodate(bitmap_bh)) {
2516                                 brelse(bitmap_bh);
2517                                 goto make_io;
2518                         }
2519                         for (i = start; i < start + inodes_per_buffer; i++) {
2520                                 if (i == inode_offset)
2521                                         continue;
2522                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2523                                         break;
2524                         }
2525                         brelse(bitmap_bh);
2526                         if (i == start + inodes_per_buffer) {
2527                                 /* all other inodes are free, so skip I/O */
2528                                 memset(bh->b_data, 0, bh->b_size);
2529                                 set_buffer_uptodate(bh);
2530                                 unlock_buffer(bh);
2531                                 goto has_buffer;
2532                         }
2533                 }
2534
2535 make_io:
2536                 /*
2537                  * There are other valid inodes in the buffer, this inode
2538                  * has in-inode xattrs, or we don't have this inode in memory.
2539                  * Read the block from disk.
2540                  */
2541                 get_bh(bh);
2542                 bh->b_end_io = end_buffer_read_sync;
2543                 submit_bh(READ, bh);
2544                 wait_on_buffer(bh);
2545                 if (!buffer_uptodate(bh)) {
2546                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2547                                         "unable to read inode block - "
2548                                         "inode=%lu, block=%lu",
2549                                         inode->i_ino, block);
2550                         brelse(bh);
2551                         return -EIO;
2552                 }
2553         }
2554 has_buffer:
2555         iloc->bh = bh;
2556         return 0;
2557 }
2558
2559 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2560 {
2561         /* We have all inode data except xattrs in memory here. */
2562         return __ext3_get_inode_loc(inode, iloc,
2563                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2564 }
2565
2566 void ext3_set_inode_flags(struct inode *inode)
2567 {
2568         unsigned int flags = EXT3_I(inode)->i_flags;
2569
2570         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2571         if (flags & EXT3_SYNC_FL)
2572                 inode->i_flags |= S_SYNC;
2573         if (flags & EXT3_APPEND_FL)
2574                 inode->i_flags |= S_APPEND;
2575         if (flags & EXT3_IMMUTABLE_FL)
2576                 inode->i_flags |= S_IMMUTABLE;
2577         if (flags & EXT3_NOATIME_FL)
2578                 inode->i_flags |= S_NOATIME;
2579         if (flags & EXT3_DIRSYNC_FL)
2580                 inode->i_flags |= S_DIRSYNC;
2581 }
2582
2583 void ext3_read_inode(struct inode * inode)
2584 {
2585         struct ext3_iloc iloc;
2586         struct ext3_inode *raw_inode;
2587         struct ext3_inode_info *ei = EXT3_I(inode);
2588         struct buffer_head *bh;
2589         int block;
2590
2591 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2592         ei->i_acl = EXT3_ACL_NOT_CACHED;
2593         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2594 #endif
2595         ei->i_block_alloc_info = NULL;
2596
2597         if (__ext3_get_inode_loc(inode, &iloc, 0))
2598                 goto bad_inode;
2599         bh = iloc.bh;
2600         raw_inode = ext3_raw_inode(&iloc);
2601         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2602         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2603         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2604         if(!(test_opt (inode->i_sb, NO_UID32))) {
2605                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2606                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2607         }
2608         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2609         inode->i_size = le32_to_cpu(raw_inode->i_size);
2610         inode->i_atime.tv_sec = le32_to_cpu(raw_inode->i_atime);
2611         inode->i_ctime.tv_sec = le32_to_cpu(raw_inode->i_ctime);
2612         inode->i_mtime.tv_sec = le32_to_cpu(raw_inode->i_mtime);
2613         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2614
2615         ei->i_state = 0;
2616         ei->i_dir_start_lookup = 0;
2617         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2618         /* We now have enough fields to check if the inode was active or not.
2619          * This is needed because nfsd might try to access dead inodes
2620          * the test is that same one that e2fsck uses
2621          * NeilBrown 1999oct15
2622          */
2623         if (inode->i_nlink == 0) {
2624                 if (inode->i_mode == 0 ||
2625                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2626                         /* this inode is deleted */
2627                         brelse (bh);
2628                         goto bad_inode;
2629                 }
2630                 /* The only unlinked inodes we let through here have
2631                  * valid i_mode and are being read by the orphan
2632                  * recovery code: that's fine, we're about to complete
2633                  * the process of deleting those. */
2634         }
2635         inode->i_blksize = PAGE_SIZE;   /* This is the optimal IO size
2636                                          * (for stat), not the fs block
2637                                          * size */  
2638         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2639         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2640 #ifdef EXT3_FRAGMENTS
2641         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2642         ei->i_frag_no = raw_inode->i_frag;
2643         ei->i_frag_size = raw_inode->i_fsize;
2644 #endif
2645         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2646         if (!S_ISREG(inode->i_mode)) {
2647                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2648         } else {
2649                 inode->i_size |=
2650                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2651         }
2652         ei->i_disksize = inode->i_size;
2653         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2654         ei->i_block_group = iloc.block_group;
2655         /*
2656          * NOTE! The in-memory inode i_data array is in little-endian order
2657          * even on big-endian machines: we do NOT byteswap the block numbers!
2658          */
2659         for (block = 0; block < EXT3_N_BLOCKS; block++)
2660                 ei->i_data[block] = raw_inode->i_block[block];
2661         INIT_LIST_HEAD(&ei->i_orphan);
2662
2663         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2664             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2665                 /*
2666                  * When mke2fs creates big inodes it does not zero out
2667                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2668                  * so ignore those first few inodes.
2669                  */
2670                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2671                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2672                     EXT3_INODE_SIZE(inode->i_sb))
2673                         goto bad_inode;
2674                 if (ei->i_extra_isize == 0) {
2675                         /* The extra space is currently unused. Use it. */
2676                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2677                                             EXT3_GOOD_OLD_INODE_SIZE;
2678                 } else {
2679                         __le32 *magic = (void *)raw_inode +
2680                                         EXT3_GOOD_OLD_INODE_SIZE +
2681                                         ei->i_extra_isize;
2682                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2683                                  ei->i_state |= EXT3_STATE_XATTR;
2684                 }
2685         } else
2686                 ei->i_extra_isize = 0;
2687
2688         if (S_ISREG(inode->i_mode)) {
2689                 inode->i_op = &ext3_file_inode_operations;
2690                 inode->i_fop = &ext3_file_operations;
2691                 ext3_set_aops(inode);
2692         } else if (S_ISDIR(inode->i_mode)) {
2693                 inode->i_op = &ext3_dir_inode_operations;
2694                 inode->i_fop = &ext3_dir_operations;
2695         } else if (S_ISLNK(inode->i_mode)) {
2696                 if (ext3_inode_is_fast_symlink(inode))
2697                         inode->i_op = &ext3_fast_symlink_inode_operations;
2698                 else {
2699                         inode->i_op = &ext3_symlink_inode_operations;
2700                         ext3_set_aops(inode);
2701                 }
2702         } else {
2703                 inode->i_op = &ext3_special_inode_operations;
2704                 if (raw_inode->i_block[0])
2705                         init_special_inode(inode, inode->i_mode,
2706                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2707                 else 
2708                         init_special_inode(inode, inode->i_mode,
2709                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2710         }
2711         brelse (iloc.bh);
2712         ext3_set_inode_flags(inode);
2713         return;
2714
2715 bad_inode:
2716         make_bad_inode(inode);
2717         return;
2718 }
2719
2720 /*
2721  * Post the struct inode info into an on-disk inode location in the
2722  * buffer-cache.  This gobbles the caller's reference to the
2723  * buffer_head in the inode location struct.
2724  *
2725  * The caller must have write access to iloc->bh.
2726  */
2727 static int ext3_do_update_inode(handle_t *handle, 
2728                                 struct inode *inode, 
2729                                 struct ext3_iloc *iloc)
2730 {
2731         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2732         struct ext3_inode_info *ei = EXT3_I(inode);
2733         struct buffer_head *bh = iloc->bh;
2734         int err = 0, rc, block;
2735
2736         /* For fields not not tracking in the in-memory inode,
2737          * initialise them to zero for new inodes. */
2738         if (ei->i_state & EXT3_STATE_NEW)
2739                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2740
2741         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2742         if(!(test_opt(inode->i_sb, NO_UID32))) {
2743                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2744                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2745 /*
2746  * Fix up interoperability with old kernels. Otherwise, old inodes get
2747  * re-used with the upper 16 bits of the uid/gid intact
2748  */
2749                 if(!ei->i_dtime) {
2750                         raw_inode->i_uid_high =
2751                                 cpu_to_le16(high_16_bits(inode->i_uid));
2752                         raw_inode->i_gid_high =
2753                                 cpu_to_le16(high_16_bits(inode->i_gid));
2754                 } else {
2755                         raw_inode->i_uid_high = 0;
2756                         raw_inode->i_gid_high = 0;
2757                 }
2758         } else {
2759                 raw_inode->i_uid_low =
2760                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2761                 raw_inode->i_gid_low =
2762                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2763                 raw_inode->i_uid_high = 0;
2764                 raw_inode->i_gid_high = 0;
2765         }
2766         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2767         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2768         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2769         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2770         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2771         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2772         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2773         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2774 #ifdef EXT3_FRAGMENTS
2775         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2776         raw_inode->i_frag = ei->i_frag_no;
2777         raw_inode->i_fsize = ei->i_frag_size;
2778 #endif
2779         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2780         if (!S_ISREG(inode->i_mode)) {
2781                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2782         } else {
2783                 raw_inode->i_size_high =
2784                         cpu_to_le32(ei->i_disksize >> 32);
2785                 if (ei->i_disksize > 0x7fffffffULL) {
2786                         struct super_block *sb = inode->i_sb;
2787                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2788                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2789                             EXT3_SB(sb)->s_es->s_rev_level ==
2790                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2791                                /* If this is the first large file
2792                                 * created, add a flag to the superblock.
2793                                 */
2794                                 err = ext3_journal_get_write_access(handle,
2795                                                 EXT3_SB(sb)->s_sbh);
2796                                 if (err)
2797                                         goto out_brelse;
2798                                 ext3_update_dynamic_rev(sb);
2799                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2800                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2801                                 sb->s_dirt = 1;
2802                                 handle->h_sync = 1;
2803                                 err = ext3_journal_dirty_metadata(handle,
2804                                                 EXT3_SB(sb)->s_sbh);
2805                         }
2806                 }
2807         }
2808         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2809         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2810                 if (old_valid_dev(inode->i_rdev)) {
2811                         raw_inode->i_block[0] =
2812                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2813                         raw_inode->i_block[1] = 0;
2814                 } else {
2815                         raw_inode->i_block[0] = 0;
2816                         raw_inode->i_block[1] =
2817                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2818                         raw_inode->i_block[2] = 0;
2819                 }
2820         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2821                 raw_inode->i_block[block] = ei->i_data[block];
2822
2823         if (ei->i_extra_isize)
2824                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2825
2826         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2827         rc = ext3_journal_dirty_metadata(handle, bh);
2828         if (!err)
2829                 err = rc;
2830         ei->i_state &= ~EXT3_STATE_NEW;
2831
2832 out_brelse:
2833         brelse (bh);
2834         ext3_std_error(inode->i_sb, err);
2835         return err;
2836 }
2837
2838 /*
2839  * ext3_write_inode()
2840  *
2841  * We are called from a few places:
2842  *
2843  * - Within generic_file_write() for O_SYNC files.
2844  *   Here, there will be no transaction running. We wait for any running
2845  *   trasnaction to commit.
2846  *
2847  * - Within sys_sync(), kupdate and such.
2848  *   We wait on commit, if tol to.
2849  *
2850  * - Within prune_icache() (PF_MEMALLOC == true)
2851  *   Here we simply return.  We can't afford to block kswapd on the
2852  *   journal commit.
2853  *
2854  * In all cases it is actually safe for us to return without doing anything,
2855  * because the inode has been copied into a raw inode buffer in
2856  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2857  * knfsd.
2858  *
2859  * Note that we are absolutely dependent upon all inode dirtiers doing the
2860  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2861  * which we are interested.
2862  *
2863  * It would be a bug for them to not do this.  The code:
2864  *
2865  *      mark_inode_dirty(inode)
2866  *      stuff();
2867  *      inode->i_size = expr;
2868  *
2869  * is in error because a kswapd-driven write_inode() could occur while
2870  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2871  * will no longer be on the superblock's dirty inode list.
2872  */
2873 int ext3_write_inode(struct inode *inode, int wait)
2874 {
2875         if (current->flags & PF_MEMALLOC)
2876                 return 0;
2877
2878         if (ext3_journal_current_handle()) {
2879                 jbd_debug(0, "called recursively, non-PF_MEMALLOC!\n");
2880                 dump_stack();
2881                 return -EIO;
2882         }
2883
2884         if (!wait)
2885                 return 0;
2886
2887         return ext3_force_commit(inode->i_sb);
2888 }
2889
2890 /*
2891  * ext3_setattr()
2892  *
2893  * Called from notify_change.
2894  *
2895  * We want to trap VFS attempts to truncate the file as soon as
2896  * possible.  In particular, we want to make sure that when the VFS
2897  * shrinks i_size, we put the inode on the orphan list and modify
2898  * i_disksize immediately, so that during the subsequent flushing of
2899  * dirty pages and freeing of disk blocks, we can guarantee that any
2900  * commit will leave the blocks being flushed in an unused state on
2901  * disk.  (On recovery, the inode will get truncated and the blocks will
2902  * be freed, so we have a strong guarantee that no future commit will
2903  * leave these blocks visible to the user.)  
2904  *
2905  * Called with inode->sem down.
2906  */
2907 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
2908 {
2909         struct inode *inode = dentry->d_inode;
2910         int error, rc = 0;
2911         const unsigned int ia_valid = attr->ia_valid;
2912
2913         error = inode_change_ok(inode, attr);
2914         if (error)
2915                 return error;
2916
2917         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
2918                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
2919                 handle_t *handle;
2920
2921                 /* (user+group)*(old+new) structure, inode write (sb,
2922                  * inode block, ? - but truncate inode update has it) */
2923                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
2924                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
2925                 if (IS_ERR(handle)) {
2926                         error = PTR_ERR(handle);
2927                         goto err_out;
2928                 }
2929                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
2930                 if (error) {
2931                         ext3_journal_stop(handle);
2932                         return error;
2933                 }
2934                 /* Update corresponding info in inode so that everything is in
2935                  * one transaction */
2936                 if (attr->ia_valid & ATTR_UID)
2937                         inode->i_uid = attr->ia_uid;
2938                 if (attr->ia_valid & ATTR_GID)
2939                         inode->i_gid = attr->ia_gid;
2940                 error = ext3_mark_inode_dirty(handle, inode);
2941                 ext3_journal_stop(handle);
2942         }
2943
2944         if (S_ISREG(inode->i_mode) &&
2945             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
2946                 handle_t *handle;
2947
2948                 handle = ext3_journal_start(inode, 3);
2949                 if (IS_ERR(handle)) {
2950                         error = PTR_ERR(handle);
2951                         goto err_out;
2952                 }
2953
2954                 error = ext3_orphan_add(handle, inode);
2955                 EXT3_I(inode)->i_disksize = attr->ia_size;
2956                 rc = ext3_mark_inode_dirty(handle, inode);
2957                 if (!error)
2958                         error = rc;
2959                 ext3_journal_stop(handle);
2960         }
2961
2962         rc = inode_setattr(inode, attr);
2963
2964         /* If inode_setattr's call to ext3_truncate failed to get a
2965          * transaction handle at all, we need to clean up the in-core
2966          * orphan list manually. */
2967         if (inode->i_nlink)
2968                 ext3_orphan_del(NULL, inode);
2969
2970         if (!rc && (ia_valid & ATTR_MODE))
2971                 rc = ext3_acl_chmod(inode);
2972
2973 err_out:
2974         ext3_std_error(inode->i_sb, error);
2975         if (!error)
2976                 error = rc;
2977         return error;
2978 }
2979
2980
2981 /*
2982  * How many blocks doth make a writepage()?
2983  *
2984  * With N blocks per page, it may be:
2985  * N data blocks
2986  * 2 indirect block
2987  * 2 dindirect
2988  * 1 tindirect
2989  * N+5 bitmap blocks (from the above)
2990  * N+5 group descriptor summary blocks
2991  * 1 inode block
2992  * 1 superblock.
2993  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
2994  *
2995  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
2996  *
2997  * With ordered or writeback data it's the same, less the N data blocks.
2998  *
2999  * If the inode's direct blocks can hold an integral number of pages then a
3000  * page cannot straddle two indirect blocks, and we can only touch one indirect
3001  * and dindirect block, and the "5" above becomes "3".
3002  *
3003  * This still overestimates under most circumstances.  If we were to pass the
3004  * start and end offsets in here as well we could do block_to_path() on each
3005  * block and work out the exact number of indirects which are touched.  Pah.
3006  */
3007
3008 static int ext3_writepage_trans_blocks(struct inode *inode)
3009 {
3010         int bpp = ext3_journal_blocks_per_page(inode);
3011         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3012         int ret;
3013
3014         if (ext3_should_journal_data(inode))
3015                 ret = 3 * (bpp + indirects) + 2;
3016         else
3017                 ret = 2 * (bpp + indirects) + 2;
3018
3019 #ifdef CONFIG_QUOTA
3020         /* We know that structure was already allocated during DQUOT_INIT so
3021          * we will be updating only the data blocks + inodes */
3022         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3023 #endif
3024
3025         return ret;
3026 }
3027
3028 /*
3029  * The caller must have previously called ext3_reserve_inode_write().
3030  * Give this, we know that the caller already has write access to iloc->bh.
3031  */
3032 int ext3_mark_iloc_dirty(handle_t *handle,
3033                 struct inode *inode, struct ext3_iloc *iloc)
3034 {
3035         int err = 0;
3036
3037         /* the do_update_inode consumes one bh->b_count */
3038         get_bh(iloc->bh);
3039
3040         /* ext3_do_update_inode() does journal_dirty_metadata */
3041         err = ext3_do_update_inode(handle, inode, iloc);
3042         put_bh(iloc->bh);
3043         return err;
3044 }
3045
3046 /* 
3047  * On success, We end up with an outstanding reference count against
3048  * iloc->bh.  This _must_ be cleaned up later. 
3049  */
3050
3051 int
3052 ext3_reserve_inode_write(handle_t *handle, struct inode *inode, 
3053                          struct ext3_iloc *iloc)
3054 {
3055         int err = 0;
3056         if (handle) {
3057                 err = ext3_get_inode_loc(inode, iloc);
3058                 if (!err) {
3059                         BUFFER_TRACE(iloc->bh, "get_write_access");
3060                         err = ext3_journal_get_write_access(handle, iloc->bh);
3061                         if (err) {
3062                                 brelse(iloc->bh);
3063                                 iloc->bh = NULL;
3064                         }
3065                 }
3066         }
3067         ext3_std_error(inode->i_sb, err);
3068         return err;
3069 }
3070
3071 /*
3072  * What we do here is to mark the in-core inode as clean with respect to inode
3073  * dirtiness (it may still be data-dirty).
3074  * This means that the in-core inode may be reaped by prune_icache
3075  * without having to perform any I/O.  This is a very good thing,
3076  * because *any* task may call prune_icache - even ones which
3077  * have a transaction open against a different journal.
3078  *
3079  * Is this cheating?  Not really.  Sure, we haven't written the
3080  * inode out, but prune_icache isn't a user-visible syncing function.
3081  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3082  * we start and wait on commits.
3083  *
3084  * Is this efficient/effective?  Well, we're being nice to the system
3085  * by cleaning up our inodes proactively so they can be reaped
3086  * without I/O.  But we are potentially leaving up to five seconds'
3087  * worth of inodes floating about which prune_icache wants us to
3088  * write out.  One way to fix that would be to get prune_icache()
3089  * to do a write_super() to free up some memory.  It has the desired
3090  * effect.
3091  */
3092 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3093 {
3094         struct ext3_iloc iloc;
3095         int err;
3096
3097         might_sleep();
3098         err = ext3_reserve_inode_write(handle, inode, &iloc);
3099         if (!err)
3100                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3101         return err;
3102 }
3103
3104 /*
3105  * ext3_dirty_inode() is called from __mark_inode_dirty()
3106  *
3107  * We're really interested in the case where a file is being extended.
3108  * i_size has been changed by generic_commit_write() and we thus need
3109  * to include the updated inode in the current transaction.
3110  *
3111  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3112  * are allocated to the file.
3113  *
3114  * If the inode is marked synchronous, we don't honour that here - doing
3115  * so would cause a commit on atime updates, which we don't bother doing.
3116  * We handle synchronous inodes at the highest possible level.
3117  */
3118 void ext3_dirty_inode(struct inode *inode)
3119 {
3120         handle_t *current_handle = ext3_journal_current_handle();
3121         handle_t *handle;
3122
3123         handle = ext3_journal_start(inode, 2);
3124         if (IS_ERR(handle))
3125                 goto out;
3126         if (current_handle &&
3127                 current_handle->h_transaction != handle->h_transaction) {
3128                 /* This task has a transaction open against a different fs */
3129                 printk(KERN_EMERG "%s: transactions do not match!\n",
3130                        __FUNCTION__);
3131         } else {
3132                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3133                                 current_handle);
3134                 ext3_mark_inode_dirty(handle, inode);
3135         }
3136         ext3_journal_stop(handle);
3137 out:
3138         return;
3139 }
3140
3141 #if 0
3142 /* 
3143  * Bind an inode's backing buffer_head into this transaction, to prevent
3144  * it from being flushed to disk early.  Unlike
3145  * ext3_reserve_inode_write, this leaves behind no bh reference and
3146  * returns no iloc structure, so the caller needs to repeat the iloc
3147  * lookup to mark the inode dirty later.
3148  */
3149 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3150 {
3151         struct ext3_iloc iloc;
3152
3153         int err = 0;
3154         if (handle) {
3155                 err = ext3_get_inode_loc(inode, &iloc);
3156                 if (!err) {
3157                         BUFFER_TRACE(iloc.bh, "get_write_access");
3158                         err = journal_get_write_access(handle, iloc.bh);
3159                         if (!err)
3160                                 err = ext3_journal_dirty_metadata(handle, 
3161                                                                   iloc.bh);
3162                         brelse(iloc.bh);
3163                 }
3164         }
3165         ext3_std_error(inode->i_sb, err);
3166         return err;
3167 }
3168 #endif
3169
3170 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3171 {
3172         journal_t *journal;
3173         handle_t *handle;
3174         int err;
3175
3176         /*
3177          * We have to be very careful here: changing a data block's
3178          * journaling status dynamically is dangerous.  If we write a
3179          * data block to the journal, change the status and then delete
3180          * that block, we risk forgetting to revoke the old log record
3181          * from the journal and so a subsequent replay can corrupt data.
3182          * So, first we make sure that the journal is empty and that
3183          * nobody is changing anything.
3184          */
3185
3186         journal = EXT3_JOURNAL(inode);
3187         if (is_journal_aborted(journal) || IS_RDONLY(inode))
3188                 return -EROFS;
3189
3190         journal_lock_updates(journal);
3191         journal_flush(journal);
3192
3193         /*
3194          * OK, there are no updates running now, and all cached data is
3195          * synced to disk.  We are now in a completely consistent state
3196          * which doesn't have anything in the journal, and we know that
3197          * no filesystem updates are running, so it is safe to modify
3198          * the inode's in-core data-journaling state flag now.
3199          */
3200
3201         if (val)
3202                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3203         else
3204                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3205         ext3_set_aops(inode);
3206
3207         journal_unlock_updates(journal);
3208
3209         /* Finally we can mark the inode as dirty. */
3210
3211         handle = ext3_journal_start(inode, 1);
3212         if (IS_ERR(handle))
3213                 return PTR_ERR(handle);
3214
3215         err = ext3_mark_inode_dirty(handle, inode);
3216         handle->h_sync = 1;
3217         ext3_journal_stop(handle);
3218         ext3_std_error(inode->i_sb, err);
3219
3220         return err;
3221 }