Pull cpu-hotplug into release branch
[linux-2.6] / 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 = le32_to_cpu(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                         unsigned long blk;
823
824                         if (!verify_chain(chain, partial)) {
825                                 /*
826                                  * Indirect block might be removed by
827                                  * truncate while we were reading it.
828                                  * Handling of that case: forget what we've
829                                  * got now. Flag the err as EAGAIN, so it
830                                  * will reread.
831                                  */
832                                 err = -EAGAIN;
833                                 count = 0;
834                                 break;
835                         }
836                         blk = le32_to_cpu(*(chain[depth-1].p + count));
837
838                         if (blk == first_block + count)
839                                 count++;
840                         else
841                                 break;
842                 }
843                 if (err != -EAGAIN)
844                         goto got_it;
845         }
846
847         /* Next simple case - plain lookup or failed read of indirect block */
848         if (!create || err == -EIO)
849                 goto cleanup;
850
851         mutex_lock(&ei->truncate_mutex);
852
853         /*
854          * If the indirect block is missing while we are reading
855          * the chain(ext3_get_branch() returns -EAGAIN err), or
856          * if the chain has been changed after we grab the semaphore,
857          * (either because another process truncated this branch, or
858          * another get_block allocated this branch) re-grab the chain to see if
859          * the request block has been allocated or not.
860          *
861          * Since we already block the truncate/other get_block
862          * at this point, we will have the current copy of the chain when we
863          * splice the branch into the tree.
864          */
865         if (err == -EAGAIN || !verify_chain(chain, partial)) {
866                 while (partial > chain) {
867                         brelse(partial->bh);
868                         partial--;
869                 }
870                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
871                 if (!partial) {
872                         count++;
873                         mutex_unlock(&ei->truncate_mutex);
874                         if (err)
875                                 goto cleanup;
876                         clear_buffer_new(bh_result);
877                         goto got_it;
878                 }
879         }
880
881         /*
882          * Okay, we need to do block allocation.  Lazily initialize the block
883          * allocation info here if necessary
884         */
885         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
886                 ext3_init_block_alloc_info(inode);
887
888         goal = ext3_find_goal(inode, iblock, chain, partial);
889
890         /* the number of blocks need to allocate for [d,t]indirect blocks */
891         indirect_blks = (chain + depth) - partial - 1;
892
893         /*
894          * Next look up the indirect map to count the totoal number of
895          * direct blocks to allocate for this branch.
896          */
897         count = ext3_blks_to_allocate(partial, indirect_blks,
898                                         maxblocks, blocks_to_boundary);
899         /*
900          * Block out ext3_truncate while we alter the tree
901          */
902         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
903                                 offsets + (partial - chain), partial);
904
905         /*
906          * The ext3_splice_branch call will free and forget any buffers
907          * on the new chain if there is a failure, but that risks using
908          * up transaction credits, especially for bitmaps where the
909          * credits cannot be returned.  Can we handle this somehow?  We
910          * may need to return -EAGAIN upwards in the worst case.  --sct
911          */
912         if (!err)
913                 err = ext3_splice_branch(handle, inode, iblock,
914                                         partial, indirect_blks, count);
915         /*
916          * i_disksize growing is protected by truncate_mutex.  Don't forget to
917          * protect it if you're about to implement concurrent
918          * ext3_get_block() -bzzz
919         */
920         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
921                 ei->i_disksize = inode->i_size;
922         mutex_unlock(&ei->truncate_mutex);
923         if (err)
924                 goto cleanup;
925
926         set_buffer_new(bh_result);
927 got_it:
928         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
929         if (blocks_to_boundary == 0)
930                 set_buffer_boundary(bh_result);
931         err = count;
932         /* Clean up and exit */
933         partial = chain + depth - 1;    /* the whole chain */
934 cleanup:
935         while (partial > chain) {
936                 BUFFER_TRACE(partial->bh, "call brelse");
937                 brelse(partial->bh);
938                 partial--;
939         }
940         BUFFER_TRACE(bh_result, "returned");
941 out:
942         return err;
943 }
944
945 #define DIO_CREDITS (EXT3_RESERVE_TRANS_BLOCKS + 32)
946
947 static int ext3_get_block(struct inode *inode, sector_t iblock,
948                         struct buffer_head *bh_result, int create)
949 {
950         handle_t *handle = journal_current_handle();
951         int ret = 0;
952         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
953
954         if (!create)
955                 goto get_block;         /* A read */
956
957         if (max_blocks == 1)
958                 goto get_block;         /* A single block get */
959
960         if (handle->h_transaction->t_state == T_LOCKED) {
961                 /*
962                  * Huge direct-io writes can hold off commits for long
963                  * periods of time.  Let this commit run.
964                  */
965                 ext3_journal_stop(handle);
966                 handle = ext3_journal_start(inode, DIO_CREDITS);
967                 if (IS_ERR(handle))
968                         ret = PTR_ERR(handle);
969                 goto get_block;
970         }
971
972         if (handle->h_buffer_credits <= EXT3_RESERVE_TRANS_BLOCKS) {
973                 /*
974                  * Getting low on buffer credits...
975                  */
976                 ret = ext3_journal_extend(handle, DIO_CREDITS);
977                 if (ret > 0) {
978                         /*
979                          * Couldn't extend the transaction.  Start a new one.
980                          */
981                         ret = ext3_journal_restart(handle, DIO_CREDITS);
982                 }
983         }
984
985 get_block:
986         if (ret == 0) {
987                 ret = ext3_get_blocks_handle(handle, inode, iblock,
988                                         max_blocks, bh_result, create, 0);
989                 if (ret > 0) {
990                         bh_result->b_size = (ret << inode->i_blkbits);
991                         ret = 0;
992                 }
993         }
994         return ret;
995 }
996
997 /*
998  * `handle' can be NULL if create is zero
999  */
1000 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1001                                 long block, int create, int *errp)
1002 {
1003         struct buffer_head dummy;
1004         int fatal = 0, err;
1005
1006         J_ASSERT(handle != NULL || create == 0);
1007
1008         dummy.b_state = 0;
1009         dummy.b_blocknr = -1000;
1010         buffer_trace_init(&dummy.b_history);
1011         err = ext3_get_blocks_handle(handle, inode, block, 1,
1012                                         &dummy, create, 1);
1013         if (err == 1) {
1014                 err = 0;
1015         } else if (err >= 0) {
1016                 WARN_ON(1);
1017                 err = -EIO;
1018         }
1019         *errp = err;
1020         if (!err && buffer_mapped(&dummy)) {
1021                 struct buffer_head *bh;
1022                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1023                 if (!bh) {
1024                         *errp = -EIO;
1025                         goto err;
1026                 }
1027                 if (buffer_new(&dummy)) {
1028                         J_ASSERT(create != 0);
1029                         J_ASSERT(handle != 0);
1030
1031                         /*
1032                          * Now that we do not always journal data, we should
1033                          * keep in mind whether this should always journal the
1034                          * new buffer as metadata.  For now, regular file
1035                          * writes use ext3_get_block instead, so it's not a
1036                          * problem.
1037                          */
1038                         lock_buffer(bh);
1039                         BUFFER_TRACE(bh, "call get_create_access");
1040                         fatal = ext3_journal_get_create_access(handle, bh);
1041                         if (!fatal && !buffer_uptodate(bh)) {
1042                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1043                                 set_buffer_uptodate(bh);
1044                         }
1045                         unlock_buffer(bh);
1046                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1047                         err = ext3_journal_dirty_metadata(handle, bh);
1048                         if (!fatal)
1049                                 fatal = err;
1050                 } else {
1051                         BUFFER_TRACE(bh, "not a new buffer");
1052                 }
1053                 if (fatal) {
1054                         *errp = fatal;
1055                         brelse(bh);
1056                         bh = NULL;
1057                 }
1058                 return bh;
1059         }
1060 err:
1061         return NULL;
1062 }
1063
1064 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1065                                int block, int create, int *err)
1066 {
1067         struct buffer_head * bh;
1068
1069         bh = ext3_getblk(handle, inode, block, create, err);
1070         if (!bh)
1071                 return bh;
1072         if (buffer_uptodate(bh))
1073                 return bh;
1074         ll_rw_block(READ, 1, &bh);
1075         wait_on_buffer(bh);
1076         if (buffer_uptodate(bh))
1077                 return bh;
1078         put_bh(bh);
1079         *err = -EIO;
1080         return NULL;
1081 }
1082
1083 static int walk_page_buffers(   handle_t *handle,
1084                                 struct buffer_head *head,
1085                                 unsigned from,
1086                                 unsigned to,
1087                                 int *partial,
1088                                 int (*fn)(      handle_t *handle,
1089                                                 struct buffer_head *bh))
1090 {
1091         struct buffer_head *bh;
1092         unsigned block_start, block_end;
1093         unsigned blocksize = head->b_size;
1094         int err, ret = 0;
1095         struct buffer_head *next;
1096
1097         for (   bh = head, block_start = 0;
1098                 ret == 0 && (bh != head || !block_start);
1099                 block_start = block_end, bh = next)
1100         {
1101                 next = bh->b_this_page;
1102                 block_end = block_start + blocksize;
1103                 if (block_end <= from || block_start >= to) {
1104                         if (partial && !buffer_uptodate(bh))
1105                                 *partial = 1;
1106                         continue;
1107                 }
1108                 err = (*fn)(handle, bh);
1109                 if (!ret)
1110                         ret = err;
1111         }
1112         return ret;
1113 }
1114
1115 /*
1116  * To preserve ordering, it is essential that the hole instantiation and
1117  * the data write be encapsulated in a single transaction.  We cannot
1118  * close off a transaction and start a new one between the ext3_get_block()
1119  * and the commit_write().  So doing the journal_start at the start of
1120  * prepare_write() is the right place.
1121  *
1122  * Also, this function can nest inside ext3_writepage() ->
1123  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1124  * has generated enough buffer credits to do the whole page.  So we won't
1125  * block on the journal in that case, which is good, because the caller may
1126  * be PF_MEMALLOC.
1127  *
1128  * By accident, ext3 can be reentered when a transaction is open via
1129  * quota file writes.  If we were to commit the transaction while thus
1130  * reentered, there can be a deadlock - we would be holding a quota
1131  * lock, and the commit would never complete if another thread had a
1132  * transaction open and was blocking on the quota lock - a ranking
1133  * violation.
1134  *
1135  * So what we do is to rely on the fact that journal_stop/journal_start
1136  * will _not_ run commit under these circumstances because handle->h_ref
1137  * is elevated.  We'll still have enough credits for the tiny quotafile
1138  * write.  
1139  */
1140 static int do_journal_get_write_access(handle_t *handle,
1141                                         struct buffer_head *bh)
1142 {
1143         if (!buffer_mapped(bh) || buffer_freed(bh))
1144                 return 0;
1145         return ext3_journal_get_write_access(handle, bh);
1146 }
1147
1148 static int ext3_prepare_write(struct file *file, struct page *page,
1149                               unsigned from, unsigned to)
1150 {
1151         struct inode *inode = page->mapping->host;
1152         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1153         handle_t *handle;
1154         int retries = 0;
1155
1156 retry:
1157         handle = ext3_journal_start(inode, needed_blocks);
1158         if (IS_ERR(handle)) {
1159                 ret = PTR_ERR(handle);
1160                 goto out;
1161         }
1162         if (test_opt(inode->i_sb, NOBH))
1163                 ret = nobh_prepare_write(page, from, to, ext3_get_block);
1164         else
1165                 ret = block_prepare_write(page, from, to, ext3_get_block);
1166         if (ret)
1167                 goto prepare_write_failed;
1168
1169         if (ext3_should_journal_data(inode)) {
1170                 ret = walk_page_buffers(handle, page_buffers(page),
1171                                 from, to, NULL, do_journal_get_write_access);
1172         }
1173 prepare_write_failed:
1174         if (ret)
1175                 ext3_journal_stop(handle);
1176         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1177                 goto retry;
1178 out:
1179         return ret;
1180 }
1181
1182 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1183 {
1184         int err = journal_dirty_data(handle, bh);
1185         if (err)
1186                 ext3_journal_abort_handle(__FUNCTION__, __FUNCTION__,
1187                                                 bh, handle,err);
1188         return err;
1189 }
1190
1191 /* For commit_write() in data=journal mode */
1192 static int commit_write_fn(handle_t *handle, struct buffer_head *bh)
1193 {
1194         if (!buffer_mapped(bh) || buffer_freed(bh))
1195                 return 0;
1196         set_buffer_uptodate(bh);
1197         return ext3_journal_dirty_metadata(handle, bh);
1198 }
1199
1200 /*
1201  * We need to pick up the new inode size which generic_commit_write gave us
1202  * `file' can be NULL - eg, when called from page_symlink().
1203  *
1204  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1205  * buffers are managed internally.
1206  */
1207 static int ext3_ordered_commit_write(struct file *file, struct page *page,
1208                              unsigned from, unsigned to)
1209 {
1210         handle_t *handle = ext3_journal_current_handle();
1211         struct inode *inode = page->mapping->host;
1212         int ret = 0, ret2;
1213
1214         ret = walk_page_buffers(handle, page_buffers(page),
1215                 from, to, NULL, ext3_journal_dirty_data);
1216
1217         if (ret == 0) {
1218                 /*
1219                  * generic_commit_write() will run mark_inode_dirty() if i_size
1220                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1221                  * into that.
1222                  */
1223                 loff_t new_i_size;
1224
1225                 new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1226                 if (new_i_size > EXT3_I(inode)->i_disksize)
1227                         EXT3_I(inode)->i_disksize = new_i_size;
1228                 ret = generic_commit_write(file, page, from, to);
1229         }
1230         ret2 = ext3_journal_stop(handle);
1231         if (!ret)
1232                 ret = ret2;
1233         return ret;
1234 }
1235
1236 static int ext3_writeback_commit_write(struct file *file, struct page *page,
1237                              unsigned from, unsigned to)
1238 {
1239         handle_t *handle = ext3_journal_current_handle();
1240         struct inode *inode = page->mapping->host;
1241         int ret = 0, ret2;
1242         loff_t new_i_size;
1243
1244         new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1245         if (new_i_size > EXT3_I(inode)->i_disksize)
1246                 EXT3_I(inode)->i_disksize = new_i_size;
1247
1248         if (test_opt(inode->i_sb, NOBH))
1249                 ret = nobh_commit_write(file, page, from, to);
1250         else
1251                 ret = generic_commit_write(file, page, from, to);
1252
1253         ret2 = ext3_journal_stop(handle);
1254         if (!ret)
1255                 ret = ret2;
1256         return ret;
1257 }
1258
1259 static int ext3_journalled_commit_write(struct file *file,
1260                         struct page *page, unsigned from, unsigned to)
1261 {
1262         handle_t *handle = ext3_journal_current_handle();
1263         struct inode *inode = page->mapping->host;
1264         int ret = 0, ret2;
1265         int partial = 0;
1266         loff_t pos;
1267
1268         /*
1269          * Here we duplicate the generic_commit_write() functionality
1270          */
1271         pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1272
1273         ret = walk_page_buffers(handle, page_buffers(page), from,
1274                                 to, &partial, commit_write_fn);
1275         if (!partial)
1276                 SetPageUptodate(page);
1277         if (pos > inode->i_size)
1278                 i_size_write(inode, pos);
1279         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1280         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1281                 EXT3_I(inode)->i_disksize = inode->i_size;
1282                 ret2 = ext3_mark_inode_dirty(handle, inode);
1283                 if (!ret) 
1284                         ret = ret2;
1285         }
1286         ret2 = ext3_journal_stop(handle);
1287         if (!ret)
1288                 ret = ret2;
1289         return ret;
1290 }
1291
1292 /* 
1293  * bmap() is special.  It gets used by applications such as lilo and by
1294  * the swapper to find the on-disk block of a specific piece of data.
1295  *
1296  * Naturally, this is dangerous if the block concerned is still in the
1297  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1298  * filesystem and enables swap, then they may get a nasty shock when the
1299  * data getting swapped to that swapfile suddenly gets overwritten by
1300  * the original zero's written out previously to the journal and
1301  * awaiting writeback in the kernel's buffer cache. 
1302  *
1303  * So, if we see any bmap calls here on a modified, data-journaled file,
1304  * take extra steps to flush any blocks which might be in the cache. 
1305  */
1306 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1307 {
1308         struct inode *inode = mapping->host;
1309         journal_t *journal;
1310         int err;
1311
1312         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1313                 /* 
1314                  * This is a REALLY heavyweight approach, but the use of
1315                  * bmap on dirty files is expected to be extremely rare:
1316                  * only if we run lilo or swapon on a freshly made file
1317                  * do we expect this to happen. 
1318                  *
1319                  * (bmap requires CAP_SYS_RAWIO so this does not
1320                  * represent an unprivileged user DOS attack --- we'd be
1321                  * in trouble if mortal users could trigger this path at
1322                  * will.) 
1323                  *
1324                  * NB. EXT3_STATE_JDATA is not set on files other than
1325                  * regular files.  If somebody wants to bmap a directory
1326                  * or symlink and gets confused because the buffer
1327                  * hasn't yet been flushed to disk, they deserve
1328                  * everything they get.
1329                  */
1330
1331                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1332                 journal = EXT3_JOURNAL(inode);
1333                 journal_lock_updates(journal);
1334                 err = journal_flush(journal);
1335                 journal_unlock_updates(journal);
1336
1337                 if (err)
1338                         return 0;
1339         }
1340
1341         return generic_block_bmap(mapping,block,ext3_get_block);
1342 }
1343
1344 static int bget_one(handle_t *handle, struct buffer_head *bh)
1345 {
1346         get_bh(bh);
1347         return 0;
1348 }
1349
1350 static int bput_one(handle_t *handle, struct buffer_head *bh)
1351 {
1352         put_bh(bh);
1353         return 0;
1354 }
1355
1356 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1357 {
1358         if (buffer_mapped(bh))
1359                 return ext3_journal_dirty_data(handle, bh);
1360         return 0;
1361 }
1362
1363 /*
1364  * Note that we always start a transaction even if we're not journalling
1365  * data.  This is to preserve ordering: any hole instantiation within
1366  * __block_write_full_page -> ext3_get_block() should be journalled
1367  * along with the data so we don't crash and then get metadata which
1368  * refers to old data.
1369  *
1370  * In all journalling modes block_write_full_page() will start the I/O.
1371  *
1372  * Problem:
1373  *
1374  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1375  *              ext3_writepage()
1376  *
1377  * Similar for:
1378  *
1379  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1380  *
1381  * Same applies to ext3_get_block().  We will deadlock on various things like
1382  * lock_journal and i_truncate_mutex.
1383  *
1384  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1385  * allocations fail.
1386  *
1387  * 16May01: If we're reentered then journal_current_handle() will be
1388  *          non-zero. We simply *return*.
1389  *
1390  * 1 July 2001: @@@ FIXME:
1391  *   In journalled data mode, a data buffer may be metadata against the
1392  *   current transaction.  But the same file is part of a shared mapping
1393  *   and someone does a writepage() on it.
1394  *
1395  *   We will move the buffer onto the async_data list, but *after* it has
1396  *   been dirtied. So there's a small window where we have dirty data on
1397  *   BJ_Metadata.
1398  *
1399  *   Note that this only applies to the last partial page in the file.  The
1400  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1401  *   broken code anyway: it's wrong for msync()).
1402  *
1403  *   It's a rare case: affects the final partial page, for journalled data
1404  *   where the file is subject to bith write() and writepage() in the same
1405  *   transction.  To fix it we'll need a custom block_write_full_page().
1406  *   We'll probably need that anyway for journalling writepage() output.
1407  *
1408  * We don't honour synchronous mounts for writepage().  That would be
1409  * disastrous.  Any write() or metadata operation will sync the fs for
1410  * us.
1411  *
1412  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1413  * we don't need to open a transaction here.
1414  */
1415 static int ext3_ordered_writepage(struct page *page,
1416                                 struct writeback_control *wbc)
1417 {
1418         struct inode *inode = page->mapping->host;
1419         struct buffer_head *page_bufs;
1420         handle_t *handle = NULL;
1421         int ret = 0;
1422         int err;
1423
1424         J_ASSERT(PageLocked(page));
1425
1426         /*
1427          * We give up here if we're reentered, because it might be for a
1428          * different filesystem.
1429          */
1430         if (ext3_journal_current_handle())
1431                 goto out_fail;
1432
1433         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1434
1435         if (IS_ERR(handle)) {
1436                 ret = PTR_ERR(handle);
1437                 goto out_fail;
1438         }
1439
1440         if (!page_has_buffers(page)) {
1441                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1442                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1443         }
1444         page_bufs = page_buffers(page);
1445         walk_page_buffers(handle, page_bufs, 0,
1446                         PAGE_CACHE_SIZE, NULL, bget_one);
1447
1448         ret = block_write_full_page(page, ext3_get_block, wbc);
1449
1450         /*
1451          * The page can become unlocked at any point now, and
1452          * truncate can then come in and change things.  So we
1453          * can't touch *page from now on.  But *page_bufs is
1454          * safe due to elevated refcount.
1455          */
1456
1457         /*
1458          * And attach them to the current transaction.  But only if 
1459          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1460          * and generally junk.
1461          */
1462         if (ret == 0) {
1463                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1464                                         NULL, journal_dirty_data_fn);
1465                 if (!ret)
1466                         ret = err;
1467         }
1468         walk_page_buffers(handle, page_bufs, 0,
1469                         PAGE_CACHE_SIZE, NULL, bput_one);
1470         err = ext3_journal_stop(handle);
1471         if (!ret)
1472                 ret = err;
1473         return ret;
1474
1475 out_fail:
1476         redirty_page_for_writepage(wbc, page);
1477         unlock_page(page);
1478         return ret;
1479 }
1480
1481 static int ext3_writeback_writepage(struct page *page,
1482                                 struct writeback_control *wbc)
1483 {
1484         struct inode *inode = page->mapping->host;
1485         handle_t *handle = NULL;
1486         int ret = 0;
1487         int err;
1488
1489         if (ext3_journal_current_handle())
1490                 goto out_fail;
1491
1492         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1493         if (IS_ERR(handle)) {
1494                 ret = PTR_ERR(handle);
1495                 goto out_fail;
1496         }
1497
1498         if (test_opt(inode->i_sb, NOBH))
1499                 ret = nobh_writepage(page, ext3_get_block, wbc);
1500         else
1501                 ret = block_write_full_page(page, ext3_get_block, wbc);
1502
1503         err = ext3_journal_stop(handle);
1504         if (!ret)
1505                 ret = err;
1506         return ret;
1507
1508 out_fail:
1509         redirty_page_for_writepage(wbc, page);
1510         unlock_page(page);
1511         return ret;
1512 }
1513
1514 static int ext3_journalled_writepage(struct page *page,
1515                                 struct writeback_control *wbc)
1516 {
1517         struct inode *inode = page->mapping->host;
1518         handle_t *handle = NULL;
1519         int ret = 0;
1520         int err;
1521
1522         if (ext3_journal_current_handle())
1523                 goto no_write;
1524
1525         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1526         if (IS_ERR(handle)) {
1527                 ret = PTR_ERR(handle);
1528                 goto no_write;
1529         }
1530
1531         if (!page_has_buffers(page) || PageChecked(page)) {
1532                 /*
1533                  * It's mmapped pagecache.  Add buffers and journal it.  There
1534                  * doesn't seem much point in redirtying the page here.
1535                  */
1536                 ClearPageChecked(page);
1537                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1538                                         ext3_get_block);
1539                 if (ret != 0) {
1540                         ext3_journal_stop(handle);
1541                         goto out_unlock;
1542                 }
1543                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1544                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1545
1546                 err = walk_page_buffers(handle, page_buffers(page), 0,
1547                                 PAGE_CACHE_SIZE, NULL, commit_write_fn);
1548                 if (ret == 0)
1549                         ret = err;
1550                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1551                 unlock_page(page);
1552         } else {
1553                 /*
1554                  * It may be a page full of checkpoint-mode buffers.  We don't
1555                  * really know unless we go poke around in the buffer_heads.
1556                  * But block_write_full_page will do the right thing.
1557                  */
1558                 ret = block_write_full_page(page, ext3_get_block, wbc);
1559         }
1560         err = ext3_journal_stop(handle);
1561         if (!ret)
1562                 ret = err;
1563 out:
1564         return ret;
1565
1566 no_write:
1567         redirty_page_for_writepage(wbc, page);
1568 out_unlock:
1569         unlock_page(page);
1570         goto out;
1571 }
1572
1573 static int ext3_readpage(struct file *file, struct page *page)
1574 {
1575         return mpage_readpage(page, ext3_get_block);
1576 }
1577
1578 static int
1579 ext3_readpages(struct file *file, struct address_space *mapping,
1580                 struct list_head *pages, unsigned nr_pages)
1581 {
1582         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1583 }
1584
1585 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1586 {
1587         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1588
1589         /*
1590          * If it's a full truncate we just forget about the pending dirtying
1591          */
1592         if (offset == 0)
1593                 ClearPageChecked(page);
1594
1595         journal_invalidatepage(journal, page, offset);
1596 }
1597
1598 static int ext3_releasepage(struct page *page, gfp_t wait)
1599 {
1600         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1601
1602         WARN_ON(PageChecked(page));
1603         if (!page_has_buffers(page))
1604                 return 0;
1605         return journal_try_to_free_buffers(journal, page, wait);
1606 }
1607
1608 /*
1609  * If the O_DIRECT write will extend the file then add this inode to the
1610  * orphan list.  So recovery will truncate it back to the original size
1611  * if the machine crashes during the write.
1612  *
1613  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1614  * crashes then stale disk data _may_ be exposed inside the file.
1615  */
1616 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1617                         const struct iovec *iov, loff_t offset,
1618                         unsigned long nr_segs)
1619 {
1620         struct file *file = iocb->ki_filp;
1621         struct inode *inode = file->f_mapping->host;
1622         struct ext3_inode_info *ei = EXT3_I(inode);
1623         handle_t *handle = NULL;
1624         ssize_t ret;
1625         int orphan = 0;
1626         size_t count = iov_length(iov, nr_segs);
1627
1628         if (rw == WRITE) {
1629                 loff_t final_size = offset + count;
1630
1631                 handle = ext3_journal_start(inode, DIO_CREDITS);
1632                 if (IS_ERR(handle)) {
1633                         ret = PTR_ERR(handle);
1634                         goto out;
1635                 }
1636                 if (final_size > inode->i_size) {
1637                         ret = ext3_orphan_add(handle, inode);
1638                         if (ret)
1639                                 goto out_stop;
1640                         orphan = 1;
1641                         ei->i_disksize = inode->i_size;
1642                 }
1643         }
1644
1645         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov, 
1646                                  offset, nr_segs,
1647                                  ext3_get_block, NULL);
1648
1649         /*
1650          * Reacquire the handle: ext3_get_block() can restart the transaction
1651          */
1652         handle = journal_current_handle();
1653
1654 out_stop:
1655         if (handle) {
1656                 int err;
1657
1658                 if (orphan && inode->i_nlink)
1659                         ext3_orphan_del(handle, inode);
1660                 if (orphan && ret > 0) {
1661                         loff_t end = offset + ret;
1662                         if (end > inode->i_size) {
1663                                 ei->i_disksize = end;
1664                                 i_size_write(inode, end);
1665                                 /*
1666                                  * We're going to return a positive `ret'
1667                                  * here due to non-zero-length I/O, so there's
1668                                  * no way of reporting error returns from
1669                                  * ext3_mark_inode_dirty() to userspace.  So
1670                                  * ignore it.
1671                                  */
1672                                 ext3_mark_inode_dirty(handle, inode);
1673                         }
1674                 }
1675                 err = ext3_journal_stop(handle);
1676                 if (ret == 0)
1677                         ret = err;
1678         }
1679 out:
1680         return ret;
1681 }
1682
1683 /*
1684  * Pages can be marked dirty completely asynchronously from ext3's journalling
1685  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1686  * much here because ->set_page_dirty is called under VFS locks.  The page is
1687  * not necessarily locked.
1688  *
1689  * We cannot just dirty the page and leave attached buffers clean, because the
1690  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1691  * or jbddirty because all the journalling code will explode.
1692  *
1693  * So what we do is to mark the page "pending dirty" and next time writepage
1694  * is called, propagate that into the buffers appropriately.
1695  */
1696 static int ext3_journalled_set_page_dirty(struct page *page)
1697 {
1698         SetPageChecked(page);
1699         return __set_page_dirty_nobuffers(page);
1700 }
1701
1702 static struct address_space_operations ext3_ordered_aops = {
1703         .readpage       = ext3_readpage,
1704         .readpages      = ext3_readpages,
1705         .writepage      = ext3_ordered_writepage,
1706         .sync_page      = block_sync_page,
1707         .prepare_write  = ext3_prepare_write,
1708         .commit_write   = ext3_ordered_commit_write,
1709         .bmap           = ext3_bmap,
1710         .invalidatepage = ext3_invalidatepage,
1711         .releasepage    = ext3_releasepage,
1712         .direct_IO      = ext3_direct_IO,
1713         .migratepage    = buffer_migrate_page,
1714 };
1715
1716 static struct address_space_operations ext3_writeback_aops = {
1717         .readpage       = ext3_readpage,
1718         .readpages      = ext3_readpages,
1719         .writepage      = ext3_writeback_writepage,
1720         .sync_page      = block_sync_page,
1721         .prepare_write  = ext3_prepare_write,
1722         .commit_write   = ext3_writeback_commit_write,
1723         .bmap           = ext3_bmap,
1724         .invalidatepage = ext3_invalidatepage,
1725         .releasepage    = ext3_releasepage,
1726         .direct_IO      = ext3_direct_IO,
1727         .migratepage    = buffer_migrate_page,
1728 };
1729
1730 static struct address_space_operations ext3_journalled_aops = {
1731         .readpage       = ext3_readpage,
1732         .readpages      = ext3_readpages,
1733         .writepage      = ext3_journalled_writepage,
1734         .sync_page      = block_sync_page,
1735         .prepare_write  = ext3_prepare_write,
1736         .commit_write   = ext3_journalled_commit_write,
1737         .set_page_dirty = ext3_journalled_set_page_dirty,
1738         .bmap           = ext3_bmap,
1739         .invalidatepage = ext3_invalidatepage,
1740         .releasepage    = ext3_releasepage,
1741 };
1742
1743 void ext3_set_aops(struct inode *inode)
1744 {
1745         if (ext3_should_order_data(inode))
1746                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1747         else if (ext3_should_writeback_data(inode))
1748                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1749         else
1750                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1751 }
1752
1753 /*
1754  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1755  * up to the end of the block which corresponds to `from'.
1756  * This required during truncate. We need to physically zero the tail end
1757  * of that block so it doesn't yield old data if the file is later grown.
1758  */
1759 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1760                 struct address_space *mapping, loff_t from)
1761 {
1762         unsigned long index = from >> PAGE_CACHE_SHIFT;
1763         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1764         unsigned blocksize, iblock, length, pos;
1765         struct inode *inode = mapping->host;
1766         struct buffer_head *bh;
1767         int err = 0;
1768         void *kaddr;
1769
1770         blocksize = inode->i_sb->s_blocksize;
1771         length = blocksize - (offset & (blocksize - 1));
1772         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1773
1774         /*
1775          * For "nobh" option,  we can only work if we don't need to
1776          * read-in the page - otherwise we create buffers to do the IO.
1777          */
1778         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1779              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1780                 kaddr = kmap_atomic(page, KM_USER0);
1781                 memset(kaddr + offset, 0, length);
1782                 flush_dcache_page(page);
1783                 kunmap_atomic(kaddr, KM_USER0);
1784                 set_page_dirty(page);
1785                 goto unlock;
1786         }
1787
1788         if (!page_has_buffers(page))
1789                 create_empty_buffers(page, blocksize, 0);
1790
1791         /* Find the buffer that contains "offset" */
1792         bh = page_buffers(page);
1793         pos = blocksize;
1794         while (offset >= pos) {
1795                 bh = bh->b_this_page;
1796                 iblock++;
1797                 pos += blocksize;
1798         }
1799
1800         err = 0;
1801         if (buffer_freed(bh)) {
1802                 BUFFER_TRACE(bh, "freed: skip");
1803                 goto unlock;
1804         }
1805
1806         if (!buffer_mapped(bh)) {
1807                 BUFFER_TRACE(bh, "unmapped");
1808                 ext3_get_block(inode, iblock, bh, 0);
1809                 /* unmapped? It's a hole - nothing to do */
1810                 if (!buffer_mapped(bh)) {
1811                         BUFFER_TRACE(bh, "still unmapped");
1812                         goto unlock;
1813                 }
1814         }
1815
1816         /* Ok, it's mapped. Make sure it's up-to-date */
1817         if (PageUptodate(page))
1818                 set_buffer_uptodate(bh);
1819
1820         if (!buffer_uptodate(bh)) {
1821                 err = -EIO;
1822                 ll_rw_block(READ, 1, &bh);
1823                 wait_on_buffer(bh);
1824                 /* Uhhuh. Read error. Complain and punt. */
1825                 if (!buffer_uptodate(bh))
1826                         goto unlock;
1827         }
1828
1829         if (ext3_should_journal_data(inode)) {
1830                 BUFFER_TRACE(bh, "get write access");
1831                 err = ext3_journal_get_write_access(handle, bh);
1832                 if (err)
1833                         goto unlock;
1834         }
1835
1836         kaddr = kmap_atomic(page, KM_USER0);
1837         memset(kaddr + offset, 0, length);
1838         flush_dcache_page(page);
1839         kunmap_atomic(kaddr, KM_USER0);
1840
1841         BUFFER_TRACE(bh, "zeroed end of block");
1842
1843         err = 0;
1844         if (ext3_should_journal_data(inode)) {
1845                 err = ext3_journal_dirty_metadata(handle, bh);
1846         } else {
1847                 if (ext3_should_order_data(inode))
1848                         err = ext3_journal_dirty_data(handle, bh);
1849                 mark_buffer_dirty(bh);
1850         }
1851
1852 unlock:
1853         unlock_page(page);
1854         page_cache_release(page);
1855         return err;
1856 }
1857
1858 /*
1859  * Probably it should be a library function... search for first non-zero word
1860  * or memcmp with zero_page, whatever is better for particular architecture.
1861  * Linus?
1862  */
1863 static inline int all_zeroes(__le32 *p, __le32 *q)
1864 {
1865         while (p < q)
1866                 if (*p++)
1867                         return 0;
1868         return 1;
1869 }
1870
1871 /**
1872  *      ext3_find_shared - find the indirect blocks for partial truncation.
1873  *      @inode:   inode in question
1874  *      @depth:   depth of the affected branch
1875  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1876  *      @chain:   place to store the pointers to partial indirect blocks
1877  *      @top:     place to the (detached) top of branch
1878  *
1879  *      This is a helper function used by ext3_truncate().
1880  *
1881  *      When we do truncate() we may have to clean the ends of several
1882  *      indirect blocks but leave the blocks themselves alive. Block is
1883  *      partially truncated if some data below the new i_size is refered
1884  *      from it (and it is on the path to the first completely truncated
1885  *      data block, indeed).  We have to free the top of that path along
1886  *      with everything to the right of the path. Since no allocation
1887  *      past the truncation point is possible until ext3_truncate()
1888  *      finishes, we may safely do the latter, but top of branch may
1889  *      require special attention - pageout below the truncation point
1890  *      might try to populate it.
1891  *
1892  *      We atomically detach the top of branch from the tree, store the
1893  *      block number of its root in *@top, pointers to buffer_heads of
1894  *      partially truncated blocks - in @chain[].bh and pointers to
1895  *      their last elements that should not be removed - in
1896  *      @chain[].p. Return value is the pointer to last filled element
1897  *      of @chain.
1898  *
1899  *      The work left to caller to do the actual freeing of subtrees:
1900  *              a) free the subtree starting from *@top
1901  *              b) free the subtrees whose roots are stored in
1902  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1903  *              c) free the subtrees growing from the inode past the @chain[0].
1904  *                      (no partially truncated stuff there).  */
1905
1906 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1907                         int offsets[4], Indirect chain[4], __le32 *top)
1908 {
1909         Indirect *partial, *p;
1910         int k, err;
1911
1912         *top = 0;
1913         /* Make k index the deepest non-null offest + 1 */
1914         for (k = depth; k > 1 && !offsets[k-1]; k--)
1915                 ;
1916         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1917         /* Writer: pointers */
1918         if (!partial)
1919                 partial = chain + k-1;
1920         /*
1921          * If the branch acquired continuation since we've looked at it -
1922          * fine, it should all survive and (new) top doesn't belong to us.
1923          */
1924         if (!partial->key && *partial->p)
1925                 /* Writer: end */
1926                 goto no_top;
1927         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1928                 ;
1929         /*
1930          * OK, we've found the last block that must survive. The rest of our
1931          * branch should be detached before unlocking. However, if that rest
1932          * of branch is all ours and does not grow immediately from the inode
1933          * it's easier to cheat and just decrement partial->p.
1934          */
1935         if (p == chain + k - 1 && p > chain) {
1936                 p->p--;
1937         } else {
1938                 *top = *p->p;
1939                 /* Nope, don't do this in ext3.  Must leave the tree intact */
1940 #if 0
1941                 *p->p = 0;
1942 #endif
1943         }
1944         /* Writer: end */
1945
1946         while(partial > p) {
1947                 brelse(partial->bh);
1948                 partial--;
1949         }
1950 no_top:
1951         return partial;
1952 }
1953
1954 /*
1955  * Zero a number of block pointers in either an inode or an indirect block.
1956  * If we restart the transaction we must again get write access to the
1957  * indirect block for further modification.
1958  *
1959  * We release `count' blocks on disk, but (last - first) may be greater
1960  * than `count' because there can be holes in there.
1961  */
1962 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
1963                 struct buffer_head *bh, unsigned long block_to_free,
1964                 unsigned long count, __le32 *first, __le32 *last)
1965 {
1966         __le32 *p;
1967         if (try_to_extend_transaction(handle, inode)) {
1968                 if (bh) {
1969                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1970                         ext3_journal_dirty_metadata(handle, bh);
1971                 }
1972                 ext3_mark_inode_dirty(handle, inode);
1973                 ext3_journal_test_restart(handle, inode);
1974                 if (bh) {
1975                         BUFFER_TRACE(bh, "retaking write access");
1976                         ext3_journal_get_write_access(handle, bh);
1977                 }
1978         }
1979
1980         /*
1981          * Any buffers which are on the journal will be in memory. We find
1982          * them on the hash table so journal_revoke() will run journal_forget()
1983          * on them.  We've already detached each block from the file, so
1984          * bforget() in journal_forget() should be safe.
1985          *
1986          * AKPM: turn on bforget in journal_forget()!!!
1987          */
1988         for (p = first; p < last; p++) {
1989                 u32 nr = le32_to_cpu(*p);
1990                 if (nr) {
1991                         struct buffer_head *bh;
1992
1993                         *p = 0;
1994                         bh = sb_find_get_block(inode->i_sb, nr);
1995                         ext3_forget(handle, 0, inode, bh, nr);
1996                 }
1997         }
1998
1999         ext3_free_blocks(handle, inode, block_to_free, count);
2000 }
2001
2002 /**
2003  * ext3_free_data - free a list of data blocks
2004  * @handle:     handle for this transaction
2005  * @inode:      inode we are dealing with
2006  * @this_bh:    indirect buffer_head which contains *@first and *@last
2007  * @first:      array of block numbers
2008  * @last:       points immediately past the end of array
2009  *
2010  * We are freeing all blocks refered from that array (numbers are stored as
2011  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2012  *
2013  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2014  * blocks are contiguous then releasing them at one time will only affect one
2015  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2016  * actually use a lot of journal space.
2017  *
2018  * @this_bh will be %NULL if @first and @last point into the inode's direct
2019  * block pointers.
2020  */
2021 static void ext3_free_data(handle_t *handle, struct inode *inode,
2022                            struct buffer_head *this_bh,
2023                            __le32 *first, __le32 *last)
2024 {
2025         unsigned long block_to_free = 0;    /* Starting block # of a run */
2026         unsigned long count = 0;            /* Number of blocks in the run */ 
2027         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2028                                                corresponding to
2029                                                block_to_free */
2030         unsigned long nr;                   /* Current block # */
2031         __le32 *p;                          /* Pointer into inode/ind
2032                                                for current block */
2033         int err;
2034
2035         if (this_bh) {                          /* For indirect block */
2036                 BUFFER_TRACE(this_bh, "get_write_access");
2037                 err = ext3_journal_get_write_access(handle, this_bh);
2038                 /* Important: if we can't update the indirect pointers
2039                  * to the blocks, we can't free them. */
2040                 if (err)
2041                         return;
2042         }
2043
2044         for (p = first; p < last; p++) {
2045                 nr = le32_to_cpu(*p);
2046                 if (nr) {
2047                         /* accumulate blocks to free if they're contiguous */
2048                         if (count == 0) {
2049                                 block_to_free = nr;
2050                                 block_to_free_p = p;
2051                                 count = 1;
2052                         } else if (nr == block_to_free + count) {
2053                                 count++;
2054                         } else {
2055                                 ext3_clear_blocks(handle, inode, this_bh, 
2056                                                   block_to_free,
2057                                                   count, block_to_free_p, p);
2058                                 block_to_free = nr;
2059                                 block_to_free_p = p;
2060                                 count = 1;
2061                         }
2062                 }
2063         }
2064
2065         if (count > 0)
2066                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2067                                   count, block_to_free_p, p);
2068
2069         if (this_bh) {
2070                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2071                 ext3_journal_dirty_metadata(handle, this_bh);
2072         }
2073 }
2074
2075 /**
2076  *      ext3_free_branches - free an array of branches
2077  *      @handle: JBD handle for this transaction
2078  *      @inode: inode we are dealing with
2079  *      @parent_bh: the buffer_head which contains *@first and *@last
2080  *      @first: array of block numbers
2081  *      @last:  pointer immediately past the end of array
2082  *      @depth: depth of the branches to free
2083  *
2084  *      We are freeing all blocks refered from these branches (numbers are
2085  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2086  *      appropriately.
2087  */
2088 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2089                                struct buffer_head *parent_bh,
2090                                __le32 *first, __le32 *last, int depth)
2091 {
2092         unsigned long nr;
2093         __le32 *p;
2094
2095         if (is_handle_aborted(handle))
2096                 return;
2097
2098         if (depth--) {
2099                 struct buffer_head *bh;
2100                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2101                 p = last;
2102                 while (--p >= first) {
2103                         nr = le32_to_cpu(*p);
2104                         if (!nr)
2105                                 continue;               /* A hole */
2106
2107                         /* Go read the buffer for the next level down */
2108                         bh = sb_bread(inode->i_sb, nr);
2109
2110                         /*
2111                          * A read failure? Report error and clear slot
2112                          * (should be rare).
2113                          */
2114                         if (!bh) {
2115                                 ext3_error(inode->i_sb, "ext3_free_branches",
2116                                            "Read failure, inode=%ld, block=%ld",
2117                                            inode->i_ino, nr);
2118                                 continue;
2119                         }
2120
2121                         /* This zaps the entire block.  Bottom up. */
2122                         BUFFER_TRACE(bh, "free child branches");
2123                         ext3_free_branches(handle, inode, bh,
2124                                            (__le32*)bh->b_data,
2125                                            (__le32*)bh->b_data + addr_per_block,
2126                                            depth);
2127
2128                         /*
2129                          * We've probably journalled the indirect block several
2130                          * times during the truncate.  But it's no longer
2131                          * needed and we now drop it from the transaction via
2132                          * journal_revoke().
2133                          *
2134                          * That's easy if it's exclusively part of this
2135                          * transaction.  But if it's part of the committing
2136                          * transaction then journal_forget() will simply
2137                          * brelse() it.  That means that if the underlying
2138                          * block is reallocated in ext3_get_block(),
2139                          * unmap_underlying_metadata() will find this block
2140                          * and will try to get rid of it.  damn, damn.
2141                          *
2142                          * If this block has already been committed to the
2143                          * journal, a revoke record will be written.  And
2144                          * revoke records must be emitted *before* clearing
2145                          * this block's bit in the bitmaps.
2146                          */
2147                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2148
2149                         /*
2150                          * Everything below this this pointer has been
2151                          * released.  Now let this top-of-subtree go.
2152                          *
2153                          * We want the freeing of this indirect block to be
2154                          * atomic in the journal with the updating of the
2155                          * bitmap block which owns it.  So make some room in
2156                          * the journal.
2157                          *
2158                          * We zero the parent pointer *after* freeing its
2159                          * pointee in the bitmaps, so if extend_transaction()
2160                          * for some reason fails to put the bitmap changes and
2161                          * the release into the same transaction, recovery
2162                          * will merely complain about releasing a free block,
2163                          * rather than leaking blocks.
2164                          */
2165                         if (is_handle_aborted(handle))
2166                                 return;
2167                         if (try_to_extend_transaction(handle, inode)) {
2168                                 ext3_mark_inode_dirty(handle, inode);
2169                                 ext3_journal_test_restart(handle, inode);
2170                         }
2171
2172                         ext3_free_blocks(handle, inode, nr, 1);
2173
2174                         if (parent_bh) {
2175                                 /*
2176                                  * The block which we have just freed is
2177                                  * pointed to by an indirect block: journal it
2178                                  */
2179                                 BUFFER_TRACE(parent_bh, "get_write_access");
2180                                 if (!ext3_journal_get_write_access(handle,
2181                                                                    parent_bh)){
2182                                         *p = 0;
2183                                         BUFFER_TRACE(parent_bh,
2184                                         "call ext3_journal_dirty_metadata");
2185                                         ext3_journal_dirty_metadata(handle, 
2186                                                                     parent_bh);
2187                                 }
2188                         }
2189                 }
2190         } else {
2191                 /* We have reached the bottom of the tree. */
2192                 BUFFER_TRACE(parent_bh, "free data blocks");
2193                 ext3_free_data(handle, inode, parent_bh, first, last);
2194         }
2195 }
2196
2197 /*
2198  * ext3_truncate()
2199  *
2200  * We block out ext3_get_block() block instantiations across the entire
2201  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2202  * simultaneously on behalf of the same inode.
2203  *
2204  * As we work through the truncate and commmit bits of it to the journal there
2205  * is one core, guiding principle: the file's tree must always be consistent on
2206  * disk.  We must be able to restart the truncate after a crash.
2207  *
2208  * The file's tree may be transiently inconsistent in memory (although it
2209  * probably isn't), but whenever we close off and commit a journal transaction,
2210  * the contents of (the filesystem + the journal) must be consistent and
2211  * restartable.  It's pretty simple, really: bottom up, right to left (although
2212  * left-to-right works OK too).
2213  *
2214  * Note that at recovery time, journal replay occurs *before* the restart of
2215  * truncate against the orphan inode list.
2216  *
2217  * The committed inode has the new, desired i_size (which is the same as
2218  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2219  * that this inode's truncate did not complete and it will again call
2220  * ext3_truncate() to have another go.  So there will be instantiated blocks
2221  * to the right of the truncation point in a crashed ext3 filesystem.  But
2222  * that's fine - as long as they are linked from the inode, the post-crash
2223  * ext3_truncate() run will find them and release them.
2224  */
2225 void ext3_truncate(struct inode *inode)
2226 {
2227         handle_t *handle;
2228         struct ext3_inode_info *ei = EXT3_I(inode);
2229         __le32 *i_data = ei->i_data;
2230         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2231         struct address_space *mapping = inode->i_mapping;
2232         int offsets[4];
2233         Indirect chain[4];
2234         Indirect *partial;
2235         __le32 nr = 0;
2236         int n;
2237         long last_block;
2238         unsigned blocksize = inode->i_sb->s_blocksize;
2239         struct page *page;
2240
2241         if (!(S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
2242             S_ISLNK(inode->i_mode)))
2243                 return;
2244         if (ext3_inode_is_fast_symlink(inode))
2245                 return;
2246         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2247                 return;
2248
2249         /*
2250          * We have to lock the EOF page here, because lock_page() nests
2251          * outside journal_start().
2252          */
2253         if ((inode->i_size & (blocksize - 1)) == 0) {
2254                 /* Block boundary? Nothing to do */
2255                 page = NULL;
2256         } else {
2257                 page = grab_cache_page(mapping,
2258                                 inode->i_size >> PAGE_CACHE_SHIFT);
2259                 if (!page)
2260                         return;
2261         }
2262
2263         handle = start_transaction(inode);
2264         if (IS_ERR(handle)) {
2265                 if (page) {
2266                         clear_highpage(page);
2267                         flush_dcache_page(page);
2268                         unlock_page(page);
2269                         page_cache_release(page);
2270                 }
2271                 return;         /* AKPM: return what? */
2272         }
2273
2274         last_block = (inode->i_size + blocksize-1)
2275                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2276
2277         if (page)
2278                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2279
2280         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2281         if (n == 0)
2282                 goto out_stop;  /* error */
2283
2284         /*
2285          * OK.  This truncate is going to happen.  We add the inode to the
2286          * orphan list, so that if this truncate spans multiple transactions,
2287          * and we crash, we will resume the truncate when the filesystem
2288          * recovers.  It also marks the inode dirty, to catch the new size.
2289          *
2290          * Implication: the file must always be in a sane, consistent
2291          * truncatable state while each transaction commits.
2292          */
2293         if (ext3_orphan_add(handle, inode))
2294                 goto out_stop;
2295
2296         /*
2297          * The orphan list entry will now protect us from any crash which
2298          * occurs before the truncate completes, so it is now safe to propagate
2299          * the new, shorter inode size (held for now in i_size) into the
2300          * on-disk inode. We do this via i_disksize, which is the value which
2301          * ext3 *really* writes onto the disk inode.
2302          */
2303         ei->i_disksize = inode->i_size;
2304
2305         /*
2306          * From here we block out all ext3_get_block() callers who want to
2307          * modify the block allocation tree.
2308          */
2309         mutex_lock(&ei->truncate_mutex);
2310
2311         if (n == 1) {           /* direct blocks */
2312                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2313                                i_data + EXT3_NDIR_BLOCKS);
2314                 goto do_indirects;
2315         }
2316
2317         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2318         /* Kill the top of shared branch (not detached) */
2319         if (nr) {
2320                 if (partial == chain) {
2321                         /* Shared branch grows from the inode */
2322                         ext3_free_branches(handle, inode, NULL,
2323                                            &nr, &nr+1, (chain+n-1) - partial);
2324                         *partial->p = 0;
2325                         /*
2326                          * We mark the inode dirty prior to restart,
2327                          * and prior to stop.  No need for it here.
2328                          */
2329                 } else {
2330                         /* Shared branch grows from an indirect block */
2331                         BUFFER_TRACE(partial->bh, "get_write_access");
2332                         ext3_free_branches(handle, inode, partial->bh,
2333                                         partial->p,
2334                                         partial->p+1, (chain+n-1) - partial);
2335                 }
2336         }
2337         /* Clear the ends of indirect blocks on the shared branch */
2338         while (partial > chain) {
2339                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2340                                    (__le32*)partial->bh->b_data+addr_per_block,
2341                                    (chain+n-1) - partial);
2342                 BUFFER_TRACE(partial->bh, "call brelse");
2343                 brelse (partial->bh);
2344                 partial--;
2345         }
2346 do_indirects:
2347         /* Kill the remaining (whole) subtrees */
2348         switch (offsets[0]) {
2349         default:
2350                 nr = i_data[EXT3_IND_BLOCK];
2351                 if (nr) {
2352                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2353                         i_data[EXT3_IND_BLOCK] = 0;
2354                 }
2355         case EXT3_IND_BLOCK:
2356                 nr = i_data[EXT3_DIND_BLOCK];
2357                 if (nr) {
2358                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2359                         i_data[EXT3_DIND_BLOCK] = 0;
2360                 }
2361         case EXT3_DIND_BLOCK:
2362                 nr = i_data[EXT3_TIND_BLOCK];
2363                 if (nr) {
2364                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2365                         i_data[EXT3_TIND_BLOCK] = 0;
2366                 }
2367         case EXT3_TIND_BLOCK:
2368                 ;
2369         }
2370
2371         ext3_discard_reservation(inode);
2372
2373         mutex_unlock(&ei->truncate_mutex);
2374         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2375         ext3_mark_inode_dirty(handle, inode);
2376
2377         /*
2378          * In a multi-transaction truncate, we only make the final transaction
2379          * synchronous
2380          */
2381         if (IS_SYNC(inode))
2382                 handle->h_sync = 1;
2383 out_stop:
2384         /*
2385          * If this was a simple ftruncate(), and the file will remain alive
2386          * then we need to clear up the orphan record which we created above.
2387          * However, if this was a real unlink then we were called by
2388          * ext3_delete_inode(), and we allow that function to clean up the
2389          * orphan info for us.
2390          */
2391         if (inode->i_nlink)
2392                 ext3_orphan_del(handle, inode);
2393
2394         ext3_journal_stop(handle);
2395 }
2396
2397 static unsigned long ext3_get_inode_block(struct super_block *sb,
2398                 unsigned long ino, struct ext3_iloc *iloc)
2399 {
2400         unsigned long desc, group_desc, block_group;
2401         unsigned long offset, block;
2402         struct buffer_head *bh;
2403         struct ext3_group_desc * gdp;
2404
2405
2406         if ((ino != EXT3_ROOT_INO && ino != EXT3_JOURNAL_INO &&
2407                 ino != EXT3_RESIZE_INO && ino < EXT3_FIRST_INO(sb)) ||
2408                 ino > le32_to_cpu(EXT3_SB(sb)->s_es->s_inodes_count)) {
2409                 ext3_error(sb, "ext3_get_inode_block",
2410                             "bad inode number: %lu", ino);
2411                 return 0;
2412         }
2413         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2414         if (block_group >= EXT3_SB(sb)->s_groups_count) {
2415                 ext3_error(sb,"ext3_get_inode_block","group >= groups count");
2416                 return 0;
2417         }
2418         smp_rmb();
2419         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
2420         desc = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
2421         bh = EXT3_SB(sb)->s_group_desc[group_desc];
2422         if (!bh) {
2423                 ext3_error (sb, "ext3_get_inode_block",
2424                             "Descriptor not loaded");
2425                 return 0;
2426         }
2427
2428         gdp = (struct ext3_group_desc *)bh->b_data;
2429         /*
2430          * Figure out the offset within the block group inode table
2431          */
2432         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2433                 EXT3_INODE_SIZE(sb);
2434         block = le32_to_cpu(gdp[desc].bg_inode_table) +
2435                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2436
2437         iloc->block_group = block_group;
2438         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2439         return block;
2440 }
2441
2442 /*
2443  * ext3_get_inode_loc returns with an extra refcount against the inode's
2444  * underlying buffer_head on success. If 'in_mem' is true, we have all
2445  * data in memory that is needed to recreate the on-disk version of this
2446  * inode.
2447  */
2448 static int __ext3_get_inode_loc(struct inode *inode,
2449                                 struct ext3_iloc *iloc, int in_mem)
2450 {
2451         unsigned long block;
2452         struct buffer_head *bh;
2453
2454         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2455         if (!block)
2456                 return -EIO;
2457
2458         bh = sb_getblk(inode->i_sb, block);
2459         if (!bh) {
2460                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2461                                 "unable to read inode block - "
2462                                 "inode=%lu, block=%lu", inode->i_ino, block);
2463                 return -EIO;
2464         }
2465         if (!buffer_uptodate(bh)) {
2466                 lock_buffer(bh);
2467                 if (buffer_uptodate(bh)) {
2468                         /* someone brought it uptodate while we waited */
2469                         unlock_buffer(bh);
2470                         goto has_buffer;
2471                 }
2472
2473                 /*
2474                  * If we have all information of the inode in memory and this
2475                  * is the only valid inode in the block, we need not read the
2476                  * block.
2477                  */
2478                 if (in_mem) {
2479                         struct buffer_head *bitmap_bh;
2480                         struct ext3_group_desc *desc;
2481                         int inodes_per_buffer;
2482                         int inode_offset, i;
2483                         int block_group;
2484                         int start;
2485
2486                         block_group = (inode->i_ino - 1) /
2487                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2488                         inodes_per_buffer = bh->b_size /
2489                                 EXT3_INODE_SIZE(inode->i_sb);
2490                         inode_offset = ((inode->i_ino - 1) %
2491                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2492                         start = inode_offset & ~(inodes_per_buffer - 1);
2493
2494                         /* Is the inode bitmap in cache? */
2495                         desc = ext3_get_group_desc(inode->i_sb,
2496                                                 block_group, NULL);
2497                         if (!desc)
2498                                 goto make_io;
2499
2500                         bitmap_bh = sb_getblk(inode->i_sb,
2501                                         le32_to_cpu(desc->bg_inode_bitmap));
2502                         if (!bitmap_bh)
2503                                 goto make_io;
2504
2505                         /*
2506                          * If the inode bitmap isn't in cache then the
2507                          * optimisation may end up performing two reads instead
2508                          * of one, so skip it.
2509                          */
2510                         if (!buffer_uptodate(bitmap_bh)) {
2511                                 brelse(bitmap_bh);
2512                                 goto make_io;
2513                         }
2514                         for (i = start; i < start + inodes_per_buffer; i++) {
2515                                 if (i == inode_offset)
2516                                         continue;
2517                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2518                                         break;
2519                         }
2520                         brelse(bitmap_bh);
2521                         if (i == start + inodes_per_buffer) {
2522                                 /* all other inodes are free, so skip I/O */
2523                                 memset(bh->b_data, 0, bh->b_size);
2524                                 set_buffer_uptodate(bh);
2525                                 unlock_buffer(bh);
2526                                 goto has_buffer;
2527                         }
2528                 }
2529
2530 make_io:
2531                 /*
2532                  * There are other valid inodes in the buffer, this inode
2533                  * has in-inode xattrs, or we don't have this inode in memory.
2534                  * Read the block from disk.
2535                  */
2536                 get_bh(bh);
2537                 bh->b_end_io = end_buffer_read_sync;
2538                 submit_bh(READ, bh);
2539                 wait_on_buffer(bh);
2540                 if (!buffer_uptodate(bh)) {
2541                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2542                                         "unable to read inode block - "
2543                                         "inode=%lu, block=%lu",
2544                                         inode->i_ino, block);
2545                         brelse(bh);
2546                         return -EIO;
2547                 }
2548         }
2549 has_buffer:
2550         iloc->bh = bh;
2551         return 0;
2552 }
2553
2554 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2555 {
2556         /* We have all inode data except xattrs in memory here. */
2557         return __ext3_get_inode_loc(inode, iloc,
2558                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2559 }
2560
2561 void ext3_set_inode_flags(struct inode *inode)
2562 {
2563         unsigned int flags = EXT3_I(inode)->i_flags;
2564
2565         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2566         if (flags & EXT3_SYNC_FL)
2567                 inode->i_flags |= S_SYNC;
2568         if (flags & EXT3_APPEND_FL)
2569                 inode->i_flags |= S_APPEND;
2570         if (flags & EXT3_IMMUTABLE_FL)
2571                 inode->i_flags |= S_IMMUTABLE;
2572         if (flags & EXT3_NOATIME_FL)
2573                 inode->i_flags |= S_NOATIME;
2574         if (flags & EXT3_DIRSYNC_FL)
2575                 inode->i_flags |= S_DIRSYNC;
2576 }
2577
2578 void ext3_read_inode(struct inode * inode)
2579 {
2580         struct ext3_iloc iloc;
2581         struct ext3_inode *raw_inode;
2582         struct ext3_inode_info *ei = EXT3_I(inode);
2583         struct buffer_head *bh;
2584         int block;
2585
2586 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2587         ei->i_acl = EXT3_ACL_NOT_CACHED;
2588         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2589 #endif
2590         ei->i_block_alloc_info = NULL;
2591
2592         if (__ext3_get_inode_loc(inode, &iloc, 0))
2593                 goto bad_inode;
2594         bh = iloc.bh;
2595         raw_inode = ext3_raw_inode(&iloc);
2596         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2597         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2598         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2599         if(!(test_opt (inode->i_sb, NO_UID32))) {
2600                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2601                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2602         }
2603         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2604         inode->i_size = le32_to_cpu(raw_inode->i_size);
2605         inode->i_atime.tv_sec = le32_to_cpu(raw_inode->i_atime);
2606         inode->i_ctime.tv_sec = le32_to_cpu(raw_inode->i_ctime);
2607         inode->i_mtime.tv_sec = le32_to_cpu(raw_inode->i_mtime);
2608         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2609
2610         ei->i_state = 0;
2611         ei->i_dir_start_lookup = 0;
2612         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2613         /* We now have enough fields to check if the inode was active or not.
2614          * This is needed because nfsd might try to access dead inodes
2615          * the test is that same one that e2fsck uses
2616          * NeilBrown 1999oct15
2617          */
2618         if (inode->i_nlink == 0) {
2619                 if (inode->i_mode == 0 ||
2620                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2621                         /* this inode is deleted */
2622                         brelse (bh);
2623                         goto bad_inode;
2624                 }
2625                 /* The only unlinked inodes we let through here have
2626                  * valid i_mode and are being read by the orphan
2627                  * recovery code: that's fine, we're about to complete
2628                  * the process of deleting those. */
2629         }
2630         inode->i_blksize = PAGE_SIZE;   /* This is the optimal IO size
2631                                          * (for stat), not the fs block
2632                                          * size */  
2633         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2634         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2635 #ifdef EXT3_FRAGMENTS
2636         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2637         ei->i_frag_no = raw_inode->i_frag;
2638         ei->i_frag_size = raw_inode->i_fsize;
2639 #endif
2640         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2641         if (!S_ISREG(inode->i_mode)) {
2642                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2643         } else {
2644                 inode->i_size |=
2645                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2646         }
2647         ei->i_disksize = inode->i_size;
2648         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2649         ei->i_block_group = iloc.block_group;
2650         /*
2651          * NOTE! The in-memory inode i_data array is in little-endian order
2652          * even on big-endian machines: we do NOT byteswap the block numbers!
2653          */
2654         for (block = 0; block < EXT3_N_BLOCKS; block++)
2655                 ei->i_data[block] = raw_inode->i_block[block];
2656         INIT_LIST_HEAD(&ei->i_orphan);
2657
2658         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2659             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2660                 /*
2661                  * When mke2fs creates big inodes it does not zero out
2662                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2663                  * so ignore those first few inodes.
2664                  */
2665                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2666                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2667                     EXT3_INODE_SIZE(inode->i_sb))
2668                         goto bad_inode;
2669                 if (ei->i_extra_isize == 0) {
2670                         /* The extra space is currently unused. Use it. */
2671                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2672                                             EXT3_GOOD_OLD_INODE_SIZE;
2673                 } else {
2674                         __le32 *magic = (void *)raw_inode +
2675                                         EXT3_GOOD_OLD_INODE_SIZE +
2676                                         ei->i_extra_isize;
2677                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2678                                  ei->i_state |= EXT3_STATE_XATTR;
2679                 }
2680         } else
2681                 ei->i_extra_isize = 0;
2682
2683         if (S_ISREG(inode->i_mode)) {
2684                 inode->i_op = &ext3_file_inode_operations;
2685                 inode->i_fop = &ext3_file_operations;
2686                 ext3_set_aops(inode);
2687         } else if (S_ISDIR(inode->i_mode)) {
2688                 inode->i_op = &ext3_dir_inode_operations;
2689                 inode->i_fop = &ext3_dir_operations;
2690         } else if (S_ISLNK(inode->i_mode)) {
2691                 if (ext3_inode_is_fast_symlink(inode))
2692                         inode->i_op = &ext3_fast_symlink_inode_operations;
2693                 else {
2694                         inode->i_op = &ext3_symlink_inode_operations;
2695                         ext3_set_aops(inode);
2696                 }
2697         } else {
2698                 inode->i_op = &ext3_special_inode_operations;
2699                 if (raw_inode->i_block[0])
2700                         init_special_inode(inode, inode->i_mode,
2701                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2702                 else 
2703                         init_special_inode(inode, inode->i_mode,
2704                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2705         }
2706         brelse (iloc.bh);
2707         ext3_set_inode_flags(inode);
2708         return;
2709
2710 bad_inode:
2711         make_bad_inode(inode);
2712         return;
2713 }
2714
2715 /*
2716  * Post the struct inode info into an on-disk inode location in the
2717  * buffer-cache.  This gobbles the caller's reference to the
2718  * buffer_head in the inode location struct.
2719  *
2720  * The caller must have write access to iloc->bh.
2721  */
2722 static int ext3_do_update_inode(handle_t *handle, 
2723                                 struct inode *inode, 
2724                                 struct ext3_iloc *iloc)
2725 {
2726         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2727         struct ext3_inode_info *ei = EXT3_I(inode);
2728         struct buffer_head *bh = iloc->bh;
2729         int err = 0, rc, block;
2730
2731         /* For fields not not tracking in the in-memory inode,
2732          * initialise them to zero for new inodes. */
2733         if (ei->i_state & EXT3_STATE_NEW)
2734                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2735
2736         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2737         if(!(test_opt(inode->i_sb, NO_UID32))) {
2738                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2739                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2740 /*
2741  * Fix up interoperability with old kernels. Otherwise, old inodes get
2742  * re-used with the upper 16 bits of the uid/gid intact
2743  */
2744                 if(!ei->i_dtime) {
2745                         raw_inode->i_uid_high =
2746                                 cpu_to_le16(high_16_bits(inode->i_uid));
2747                         raw_inode->i_gid_high =
2748                                 cpu_to_le16(high_16_bits(inode->i_gid));
2749                 } else {
2750                         raw_inode->i_uid_high = 0;
2751                         raw_inode->i_gid_high = 0;
2752                 }
2753         } else {
2754                 raw_inode->i_uid_low =
2755                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2756                 raw_inode->i_gid_low =
2757                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2758                 raw_inode->i_uid_high = 0;
2759                 raw_inode->i_gid_high = 0;
2760         }
2761         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2762         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2763         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2764         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2765         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2766         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2767         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2768         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2769 #ifdef EXT3_FRAGMENTS
2770         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2771         raw_inode->i_frag = ei->i_frag_no;
2772         raw_inode->i_fsize = ei->i_frag_size;
2773 #endif
2774         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2775         if (!S_ISREG(inode->i_mode)) {
2776                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2777         } else {
2778                 raw_inode->i_size_high =
2779                         cpu_to_le32(ei->i_disksize >> 32);
2780                 if (ei->i_disksize > 0x7fffffffULL) {
2781                         struct super_block *sb = inode->i_sb;
2782                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2783                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2784                             EXT3_SB(sb)->s_es->s_rev_level ==
2785                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2786                                /* If this is the first large file
2787                                 * created, add a flag to the superblock.
2788                                 */
2789                                 err = ext3_journal_get_write_access(handle,
2790                                                 EXT3_SB(sb)->s_sbh);
2791                                 if (err)
2792                                         goto out_brelse;
2793                                 ext3_update_dynamic_rev(sb);
2794                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2795                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2796                                 sb->s_dirt = 1;
2797                                 handle->h_sync = 1;
2798                                 err = ext3_journal_dirty_metadata(handle,
2799                                                 EXT3_SB(sb)->s_sbh);
2800                         }
2801                 }
2802         }
2803         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2804         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2805                 if (old_valid_dev(inode->i_rdev)) {
2806                         raw_inode->i_block[0] =
2807                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2808                         raw_inode->i_block[1] = 0;
2809                 } else {
2810                         raw_inode->i_block[0] = 0;
2811                         raw_inode->i_block[1] =
2812                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2813                         raw_inode->i_block[2] = 0;
2814                 }
2815         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2816                 raw_inode->i_block[block] = ei->i_data[block];
2817
2818         if (ei->i_extra_isize)
2819                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2820
2821         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2822         rc = ext3_journal_dirty_metadata(handle, bh);
2823         if (!err)
2824                 err = rc;
2825         ei->i_state &= ~EXT3_STATE_NEW;
2826
2827 out_brelse:
2828         brelse (bh);
2829         ext3_std_error(inode->i_sb, err);
2830         return err;
2831 }
2832
2833 /*
2834  * ext3_write_inode()
2835  *
2836  * We are called from a few places:
2837  *
2838  * - Within generic_file_write() for O_SYNC files.
2839  *   Here, there will be no transaction running. We wait for any running
2840  *   trasnaction to commit.
2841  *
2842  * - Within sys_sync(), kupdate and such.
2843  *   We wait on commit, if tol to.
2844  *
2845  * - Within prune_icache() (PF_MEMALLOC == true)
2846  *   Here we simply return.  We can't afford to block kswapd on the
2847  *   journal commit.
2848  *
2849  * In all cases it is actually safe for us to return without doing anything,
2850  * because the inode has been copied into a raw inode buffer in
2851  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2852  * knfsd.
2853  *
2854  * Note that we are absolutely dependent upon all inode dirtiers doing the
2855  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2856  * which we are interested.
2857  *
2858  * It would be a bug for them to not do this.  The code:
2859  *
2860  *      mark_inode_dirty(inode)
2861  *      stuff();
2862  *      inode->i_size = expr;
2863  *
2864  * is in error because a kswapd-driven write_inode() could occur while
2865  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2866  * will no longer be on the superblock's dirty inode list.
2867  */
2868 int ext3_write_inode(struct inode *inode, int wait)
2869 {
2870         if (current->flags & PF_MEMALLOC)
2871                 return 0;
2872
2873         if (ext3_journal_current_handle()) {
2874                 jbd_debug(0, "called recursively, non-PF_MEMALLOC!\n");
2875                 dump_stack();
2876                 return -EIO;
2877         }
2878
2879         if (!wait)
2880                 return 0;
2881
2882         return ext3_force_commit(inode->i_sb);
2883 }
2884
2885 /*
2886  * ext3_setattr()
2887  *
2888  * Called from notify_change.
2889  *
2890  * We want to trap VFS attempts to truncate the file as soon as
2891  * possible.  In particular, we want to make sure that when the VFS
2892  * shrinks i_size, we put the inode on the orphan list and modify
2893  * i_disksize immediately, so that during the subsequent flushing of
2894  * dirty pages and freeing of disk blocks, we can guarantee that any
2895  * commit will leave the blocks being flushed in an unused state on
2896  * disk.  (On recovery, the inode will get truncated and the blocks will
2897  * be freed, so we have a strong guarantee that no future commit will
2898  * leave these blocks visible to the user.)  
2899  *
2900  * Called with inode->sem down.
2901  */
2902 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
2903 {
2904         struct inode *inode = dentry->d_inode;
2905         int error, rc = 0;
2906         const unsigned int ia_valid = attr->ia_valid;
2907
2908         error = inode_change_ok(inode, attr);
2909         if (error)
2910                 return error;
2911
2912         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
2913                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
2914                 handle_t *handle;
2915
2916                 /* (user+group)*(old+new) structure, inode write (sb,
2917                  * inode block, ? - but truncate inode update has it) */
2918                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
2919                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
2920                 if (IS_ERR(handle)) {
2921                         error = PTR_ERR(handle);
2922                         goto err_out;
2923                 }
2924                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
2925                 if (error) {
2926                         ext3_journal_stop(handle);
2927                         return error;
2928                 }
2929                 /* Update corresponding info in inode so that everything is in
2930                  * one transaction */
2931                 if (attr->ia_valid & ATTR_UID)
2932                         inode->i_uid = attr->ia_uid;
2933                 if (attr->ia_valid & ATTR_GID)
2934                         inode->i_gid = attr->ia_gid;
2935                 error = ext3_mark_inode_dirty(handle, inode);
2936                 ext3_journal_stop(handle);
2937         }
2938
2939         if (S_ISREG(inode->i_mode) &&
2940             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
2941                 handle_t *handle;
2942
2943                 handle = ext3_journal_start(inode, 3);
2944                 if (IS_ERR(handle)) {
2945                         error = PTR_ERR(handle);
2946                         goto err_out;
2947                 }
2948
2949                 error = ext3_orphan_add(handle, inode);
2950                 EXT3_I(inode)->i_disksize = attr->ia_size;
2951                 rc = ext3_mark_inode_dirty(handle, inode);
2952                 if (!error)
2953                         error = rc;
2954                 ext3_journal_stop(handle);
2955         }
2956
2957         rc = inode_setattr(inode, attr);
2958
2959         /* If inode_setattr's call to ext3_truncate failed to get a
2960          * transaction handle at all, we need to clean up the in-core
2961          * orphan list manually. */
2962         if (inode->i_nlink)
2963                 ext3_orphan_del(NULL, inode);
2964
2965         if (!rc && (ia_valid & ATTR_MODE))
2966                 rc = ext3_acl_chmod(inode);
2967
2968 err_out:
2969         ext3_std_error(inode->i_sb, error);
2970         if (!error)
2971                 error = rc;
2972         return error;
2973 }
2974
2975
2976 /*
2977  * How many blocks doth make a writepage()?
2978  *
2979  * With N blocks per page, it may be:
2980  * N data blocks
2981  * 2 indirect block
2982  * 2 dindirect
2983  * 1 tindirect
2984  * N+5 bitmap blocks (from the above)
2985  * N+5 group descriptor summary blocks
2986  * 1 inode block
2987  * 1 superblock.
2988  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
2989  *
2990  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
2991  *
2992  * With ordered or writeback data it's the same, less the N data blocks.
2993  *
2994  * If the inode's direct blocks can hold an integral number of pages then a
2995  * page cannot straddle two indirect blocks, and we can only touch one indirect
2996  * and dindirect block, and the "5" above becomes "3".
2997  *
2998  * This still overestimates under most circumstances.  If we were to pass the
2999  * start and end offsets in here as well we could do block_to_path() on each
3000  * block and work out the exact number of indirects which are touched.  Pah.
3001  */
3002
3003 static int ext3_writepage_trans_blocks(struct inode *inode)
3004 {
3005         int bpp = ext3_journal_blocks_per_page(inode);
3006         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3007         int ret;
3008
3009         if (ext3_should_journal_data(inode))
3010                 ret = 3 * (bpp + indirects) + 2;
3011         else
3012                 ret = 2 * (bpp + indirects) + 2;
3013
3014 #ifdef CONFIG_QUOTA
3015         /* We know that structure was already allocated during DQUOT_INIT so
3016          * we will be updating only the data blocks + inodes */
3017         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3018 #endif
3019
3020         return ret;
3021 }
3022
3023 /*
3024  * The caller must have previously called ext3_reserve_inode_write().
3025  * Give this, we know that the caller already has write access to iloc->bh.
3026  */
3027 int ext3_mark_iloc_dirty(handle_t *handle,
3028                 struct inode *inode, struct ext3_iloc *iloc)
3029 {
3030         int err = 0;
3031
3032         /* the do_update_inode consumes one bh->b_count */
3033         get_bh(iloc->bh);
3034
3035         /* ext3_do_update_inode() does journal_dirty_metadata */
3036         err = ext3_do_update_inode(handle, inode, iloc);
3037         put_bh(iloc->bh);
3038         return err;
3039 }
3040
3041 /* 
3042  * On success, We end up with an outstanding reference count against
3043  * iloc->bh.  This _must_ be cleaned up later. 
3044  */
3045
3046 int
3047 ext3_reserve_inode_write(handle_t *handle, struct inode *inode, 
3048                          struct ext3_iloc *iloc)
3049 {
3050         int err = 0;
3051         if (handle) {
3052                 err = ext3_get_inode_loc(inode, iloc);
3053                 if (!err) {
3054                         BUFFER_TRACE(iloc->bh, "get_write_access");
3055                         err = ext3_journal_get_write_access(handle, iloc->bh);
3056                         if (err) {
3057                                 brelse(iloc->bh);
3058                                 iloc->bh = NULL;
3059                         }
3060                 }
3061         }
3062         ext3_std_error(inode->i_sb, err);
3063         return err;
3064 }
3065
3066 /*
3067  * What we do here is to mark the in-core inode as clean with respect to inode
3068  * dirtiness (it may still be data-dirty).
3069  * This means that the in-core inode may be reaped by prune_icache
3070  * without having to perform any I/O.  This is a very good thing,
3071  * because *any* task may call prune_icache - even ones which
3072  * have a transaction open against a different journal.
3073  *
3074  * Is this cheating?  Not really.  Sure, we haven't written the
3075  * inode out, but prune_icache isn't a user-visible syncing function.
3076  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3077  * we start and wait on commits.
3078  *
3079  * Is this efficient/effective?  Well, we're being nice to the system
3080  * by cleaning up our inodes proactively so they can be reaped
3081  * without I/O.  But we are potentially leaving up to five seconds'
3082  * worth of inodes floating about which prune_icache wants us to
3083  * write out.  One way to fix that would be to get prune_icache()
3084  * to do a write_super() to free up some memory.  It has the desired
3085  * effect.
3086  */
3087 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3088 {
3089         struct ext3_iloc iloc;
3090         int err;
3091
3092         might_sleep();
3093         err = ext3_reserve_inode_write(handle, inode, &iloc);
3094         if (!err)
3095                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3096         return err;
3097 }
3098
3099 /*
3100  * ext3_dirty_inode() is called from __mark_inode_dirty()
3101  *
3102  * We're really interested in the case where a file is being extended.
3103  * i_size has been changed by generic_commit_write() and we thus need
3104  * to include the updated inode in the current transaction.
3105  *
3106  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3107  * are allocated to the file.
3108  *
3109  * If the inode is marked synchronous, we don't honour that here - doing
3110  * so would cause a commit on atime updates, which we don't bother doing.
3111  * We handle synchronous inodes at the highest possible level.
3112  */
3113 void ext3_dirty_inode(struct inode *inode)
3114 {
3115         handle_t *current_handle = ext3_journal_current_handle();
3116         handle_t *handle;
3117
3118         handle = ext3_journal_start(inode, 2);
3119         if (IS_ERR(handle))
3120                 goto out;
3121         if (current_handle &&
3122                 current_handle->h_transaction != handle->h_transaction) {
3123                 /* This task has a transaction open against a different fs */
3124                 printk(KERN_EMERG "%s: transactions do not match!\n",
3125                        __FUNCTION__);
3126         } else {
3127                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3128                                 current_handle);
3129                 ext3_mark_inode_dirty(handle, inode);
3130         }
3131         ext3_journal_stop(handle);
3132 out:
3133         return;
3134 }
3135
3136 #if 0
3137 /* 
3138  * Bind an inode's backing buffer_head into this transaction, to prevent
3139  * it from being flushed to disk early.  Unlike
3140  * ext3_reserve_inode_write, this leaves behind no bh reference and
3141  * returns no iloc structure, so the caller needs to repeat the iloc
3142  * lookup to mark the inode dirty later.
3143  */
3144 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3145 {
3146         struct ext3_iloc iloc;
3147
3148         int err = 0;
3149         if (handle) {
3150                 err = ext3_get_inode_loc(inode, &iloc);
3151                 if (!err) {
3152                         BUFFER_TRACE(iloc.bh, "get_write_access");
3153                         err = journal_get_write_access(handle, iloc.bh);
3154                         if (!err)
3155                                 err = ext3_journal_dirty_metadata(handle, 
3156                                                                   iloc.bh);
3157                         brelse(iloc.bh);
3158                 }
3159         }
3160         ext3_std_error(inode->i_sb, err);
3161         return err;
3162 }
3163 #endif
3164
3165 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3166 {
3167         journal_t *journal;
3168         handle_t *handle;
3169         int err;
3170
3171         /*
3172          * We have to be very careful here: changing a data block's
3173          * journaling status dynamically is dangerous.  If we write a
3174          * data block to the journal, change the status and then delete
3175          * that block, we risk forgetting to revoke the old log record
3176          * from the journal and so a subsequent replay can corrupt data.
3177          * So, first we make sure that the journal is empty and that
3178          * nobody is changing anything.
3179          */
3180
3181         journal = EXT3_JOURNAL(inode);
3182         if (is_journal_aborted(journal) || IS_RDONLY(inode))
3183                 return -EROFS;
3184
3185         journal_lock_updates(journal);
3186         journal_flush(journal);
3187
3188         /*
3189          * OK, there are no updates running now, and all cached data is
3190          * synced to disk.  We are now in a completely consistent state
3191          * which doesn't have anything in the journal, and we know that
3192          * no filesystem updates are running, so it is safe to modify
3193          * the inode's in-core data-journaling state flag now.
3194          */
3195
3196         if (val)
3197                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3198         else
3199                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3200         ext3_set_aops(inode);
3201
3202         journal_unlock_updates(journal);
3203
3204         /* Finally we can mark the inode as dirty. */
3205
3206         handle = ext3_journal_start(inode, 1);
3207         if (IS_ERR(handle))
3208                 return PTR_ERR(handle);
3209
3210         err = ext3_mark_inode_dirty(handle, inode);
3211         handle->h_sync = 1;
3212         ext3_journal_stop(handle);
3213         ext3_std_error(inode->i_sb, err);
3214
3215         return err;
3216 }