[PATCH] revert "retries in ext3_prepare_write() violate ordering requirements"
[linux-2.6.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/smp_lock.h>
31 #include <linux/highuid.h>
32 #include <linux/pagemap.h>
33 #include <linux/quotaops.h>
34 #include <linux/string.h>
35 #include <linux/buffer_head.h>
36 #include <linux/writeback.h>
37 #include <linux/mpage.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include "xattr.h"
41 #include "acl.h"
42
43 static int ext3_writepage_trans_blocks(struct inode *inode);
44
45 /*
46  * Test whether an inode is a fast symlink.
47  */
48 static int ext3_inode_is_fast_symlink(struct inode *inode)
49 {
50         int ea_blocks = EXT3_I(inode)->i_file_acl ?
51                 (inode->i_sb->s_blocksize >> 9) : 0;
52
53         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
54 }
55
56 /*
57  * The ext3 forget function must perform a revoke if we are freeing data
58  * which has been journaled.  Metadata (eg. indirect blocks) must be
59  * revoked in all cases.
60  *
61  * "bh" may be NULL: a metadata block may have been freed from memory
62  * but there may still be a record of it in the journal, and that record
63  * still needs to be revoked.
64  */
65 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
66                         struct buffer_head *bh, ext3_fsblk_t blocknr)
67 {
68         int err;
69
70         might_sleep();
71
72         BUFFER_TRACE(bh, "enter");
73
74         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
75                   "data mode %lx\n",
76                   bh, is_metadata, inode->i_mode,
77                   test_opt(inode->i_sb, DATA_FLAGS));
78
79         /* Never use the revoke function if we are doing full data
80          * journaling: there is no need to, and a V1 superblock won't
81          * support it.  Otherwise, only skip the revoke on un-journaled
82          * data blocks. */
83
84         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
85             (!is_metadata && !ext3_should_journal_data(inode))) {
86                 if (bh) {
87                         BUFFER_TRACE(bh, "call journal_forget");
88                         return ext3_journal_forget(handle, bh);
89                 }
90                 return 0;
91         }
92
93         /*
94          * data!=journal && (is_metadata || should_journal_data(inode))
95          */
96         BUFFER_TRACE(bh, "call ext3_journal_revoke");
97         err = ext3_journal_revoke(handle, blocknr, bh);
98         if (err)
99                 ext3_abort(inode->i_sb, __FUNCTION__,
100                            "error %d when attempting revoke", err);
101         BUFFER_TRACE(bh, "exit");
102         return err;
103 }
104
105 /*
106  * Work out how many blocks we need to proceed with the next chunk of a
107  * truncate transaction.
108  */
109 static unsigned long blocks_for_truncate(struct inode *inode)
110 {
111         unsigned long needed;
112
113         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
114
115         /* Give ourselves just enough room to cope with inodes in which
116          * i_blocks is corrupt: we've seen disk corruptions in the past
117          * which resulted in random data in an inode which looked enough
118          * like a regular file for ext3 to try to delete it.  Things
119          * will go a bit crazy if that happens, but at least we should
120          * try not to panic the whole kernel. */
121         if (needed < 2)
122                 needed = 2;
123
124         /* But we need to bound the transaction so we don't overflow the
125          * journal. */
126         if (needed > EXT3_MAX_TRANS_DATA)
127                 needed = EXT3_MAX_TRANS_DATA;
128
129         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
130 }
131
132 /*
133  * Truncate transactions can be complex and absolutely huge.  So we need to
134  * be able to restart the transaction at a conventient checkpoint to make
135  * sure we don't overflow the journal.
136  *
137  * start_transaction gets us a new handle for a truncate transaction,
138  * and extend_transaction tries to extend the existing one a bit.  If
139  * extend fails, we need to propagate the failure up and restart the
140  * transaction in the top-level truncate loop. --sct
141  */
142 static handle_t *start_transaction(struct inode *inode)
143 {
144         handle_t *result;
145
146         result = ext3_journal_start(inode, blocks_for_truncate(inode));
147         if (!IS_ERR(result))
148                 return result;
149
150         ext3_std_error(inode->i_sb, PTR_ERR(result));
151         return result;
152 }
153
154 /*
155  * Try to extend this transaction for the purposes of truncation.
156  *
157  * Returns 0 if we managed to create more room.  If we can't create more
158  * room, and the transaction must be restarted we return 1.
159  */
160 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
161 {
162         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
163                 return 0;
164         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
165                 return 0;
166         return 1;
167 }
168
169 /*
170  * Restart the transaction associated with *handle.  This does a commit,
171  * so before we call here everything must be consistently dirtied against
172  * this transaction.
173  */
174 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
175 {
176         jbd_debug(2, "restarting handle %p\n", handle);
177         return ext3_journal_restart(handle, blocks_for_truncate(inode));
178 }
179
180 /*
181  * Called at the last iput() if i_nlink is zero.
182  */
183 void ext3_delete_inode (struct inode * inode)
184 {
185         handle_t *handle;
186
187         truncate_inode_pages(&inode->i_data, 0);
188
189         if (is_bad_inode(inode))
190                 goto no_delete;
191
192         handle = start_transaction(inode);
193         if (IS_ERR(handle)) {
194                 /*
195                  * If we're going to skip the normal cleanup, we still need to
196                  * make sure that the in-core orphan linked list is properly
197                  * cleaned up.
198                  */
199                 ext3_orphan_del(NULL, inode);
200                 goto no_delete;
201         }
202
203         if (IS_SYNC(inode))
204                 handle->h_sync = 1;
205         inode->i_size = 0;
206         if (inode->i_blocks)
207                 ext3_truncate(inode);
208         /*
209          * Kill off the orphan record which ext3_truncate created.
210          * AKPM: I think this can be inside the above `if'.
211          * Note that ext3_orphan_del() has to be able to cope with the
212          * deletion of a non-existent orphan - this is because we don't
213          * know if ext3_truncate() actually created an orphan record.
214          * (Well, we could do this if we need to, but heck - it works)
215          */
216         ext3_orphan_del(handle, inode);
217         EXT3_I(inode)->i_dtime  = get_seconds();
218
219         /*
220          * One subtle ordering requirement: if anything has gone wrong
221          * (transaction abort, IO errors, whatever), then we can still
222          * do these next steps (the fs will already have been marked as
223          * having errors), but we can't free the inode if the mark_dirty
224          * fails.
225          */
226         if (ext3_mark_inode_dirty(handle, inode))
227                 /* If that failed, just do the required in-core inode clear. */
228                 clear_inode(inode);
229         else
230                 ext3_free_inode(handle, inode);
231         ext3_journal_stop(handle);
232         return;
233 no_delete:
234         clear_inode(inode);     /* We must guarantee clearing of inode... */
235 }
236
237 typedef struct {
238         __le32  *p;
239         __le32  key;
240         struct buffer_head *bh;
241 } Indirect;
242
243 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
244 {
245         p->key = *(p->p = v);
246         p->bh = bh;
247 }
248
249 static int verify_chain(Indirect *from, Indirect *to)
250 {
251         while (from <= to && from->key == *from->p)
252                 from++;
253         return (from > to);
254 }
255
256 /**
257  *      ext3_block_to_path - parse the block number into array of offsets
258  *      @inode: inode in question (we are only interested in its superblock)
259  *      @i_block: block number to be parsed
260  *      @offsets: array to store the offsets in
261  *      @boundary: set this non-zero if the referred-to block is likely to be
262  *             followed (on disk) by an indirect block.
263  *
264  *      To store the locations of file's data ext3 uses a data structure common
265  *      for UNIX filesystems - tree of pointers anchored in the inode, with
266  *      data blocks at leaves and indirect blocks in intermediate nodes.
267  *      This function translates the block number into path in that tree -
268  *      return value is the path length and @offsets[n] is the offset of
269  *      pointer to (n+1)th node in the nth one. If @block is out of range
270  *      (negative or too large) warning is printed and zero returned.
271  *
272  *      Note: function doesn't find node addresses, so no IO is needed. All
273  *      we need to know is the capacity of indirect blocks (taken from the
274  *      inode->i_sb).
275  */
276
277 /*
278  * Portability note: the last comparison (check that we fit into triple
279  * indirect block) is spelled differently, because otherwise on an
280  * architecture with 32-bit longs and 8Kb pages we might get into trouble
281  * if our filesystem had 8Kb blocks. We might use long long, but that would
282  * kill us on x86. Oh, well, at least the sign propagation does not matter -
283  * i_block would have to be negative in the very beginning, so we would not
284  * get there at all.
285  */
286
287 static int ext3_block_to_path(struct inode *inode,
288                         long i_block, int offsets[4], int *boundary)
289 {
290         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
291         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
292         const long direct_blocks = EXT3_NDIR_BLOCKS,
293                 indirect_blocks = ptrs,
294                 double_blocks = (1 << (ptrs_bits * 2));
295         int n = 0;
296         int final = 0;
297
298         if (i_block < 0) {
299                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
300         } else if (i_block < direct_blocks) {
301                 offsets[n++] = i_block;
302                 final = direct_blocks;
303         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
304                 offsets[n++] = EXT3_IND_BLOCK;
305                 offsets[n++] = i_block;
306                 final = ptrs;
307         } else if ((i_block -= indirect_blocks) < double_blocks) {
308                 offsets[n++] = EXT3_DIND_BLOCK;
309                 offsets[n++] = i_block >> ptrs_bits;
310                 offsets[n++] = i_block & (ptrs - 1);
311                 final = ptrs;
312         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
313                 offsets[n++] = EXT3_TIND_BLOCK;
314                 offsets[n++] = i_block >> (ptrs_bits * 2);
315                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
316                 offsets[n++] = i_block & (ptrs - 1);
317                 final = ptrs;
318         } else {
319                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
320         }
321         if (boundary)
322                 *boundary = final - 1 - (i_block & (ptrs - 1));
323         return n;
324 }
325
326 /**
327  *      ext3_get_branch - read the chain of indirect blocks leading to data
328  *      @inode: inode in question
329  *      @depth: depth of the chain (1 - direct pointer, etc.)
330  *      @offsets: offsets of pointers in inode/indirect blocks
331  *      @chain: place to store the result
332  *      @err: here we store the error value
333  *
334  *      Function fills the array of triples <key, p, bh> and returns %NULL
335  *      if everything went OK or the pointer to the last filled triple
336  *      (incomplete one) otherwise. Upon the return chain[i].key contains
337  *      the number of (i+1)-th block in the chain (as it is stored in memory,
338  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
339  *      number (it points into struct inode for i==0 and into the bh->b_data
340  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
341  *      block for i>0 and NULL for i==0. In other words, it holds the block
342  *      numbers of the chain, addresses they were taken from (and where we can
343  *      verify that chain did not change) and buffer_heads hosting these
344  *      numbers.
345  *
346  *      Function stops when it stumbles upon zero pointer (absent block)
347  *              (pointer to last triple returned, *@err == 0)
348  *      or when it gets an IO error reading an indirect block
349  *              (ditto, *@err == -EIO)
350  *      or when it notices that chain had been changed while it was reading
351  *              (ditto, *@err == -EAGAIN)
352  *      or when it reads all @depth-1 indirect blocks successfully and finds
353  *      the whole chain, all way to the data (returns %NULL, *err == 0).
354  */
355 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
356                                  Indirect chain[4], int *err)
357 {
358         struct super_block *sb = inode->i_sb;
359         Indirect *p = chain;
360         struct buffer_head *bh;
361
362         *err = 0;
363         /* i_data is not going away, no lock needed */
364         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
365         if (!p->key)
366                 goto no_block;
367         while (--depth) {
368                 bh = sb_bread(sb, le32_to_cpu(p->key));
369                 if (!bh)
370                         goto failure;
371                 /* Reader: pointers */
372                 if (!verify_chain(chain, p))
373                         goto changed;
374                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
375                 /* Reader: end */
376                 if (!p->key)
377                         goto no_block;
378         }
379         return NULL;
380
381 changed:
382         brelse(bh);
383         *err = -EAGAIN;
384         goto no_block;
385 failure:
386         *err = -EIO;
387 no_block:
388         return p;
389 }
390
391 /**
392  *      ext3_find_near - find a place for allocation with sufficient locality
393  *      @inode: owner
394  *      @ind: descriptor of indirect block.
395  *
396  *      This function returns the prefered place for block allocation.
397  *      It is used when heuristic for sequential allocation fails.
398  *      Rules are:
399  *        + if there is a block to the left of our position - allocate near it.
400  *        + if pointer will live in indirect block - allocate near that block.
401  *        + if pointer will live in inode - allocate in the same
402  *          cylinder group.
403  *
404  * In the latter case we colour the starting block by the callers PID to
405  * prevent it from clashing with concurrent allocations for a different inode
406  * in the same block group.   The PID is used here so that functionally related
407  * files will be close-by on-disk.
408  *
409  *      Caller must make sure that @ind is valid and will stay that way.
410  */
411 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
412 {
413         struct ext3_inode_info *ei = EXT3_I(inode);
414         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
415         __le32 *p;
416         ext3_fsblk_t bg_start;
417         ext3_grpblk_t colour;
418
419         /* Try to find previous block */
420         for (p = ind->p - 1; p >= start; p--) {
421                 if (*p)
422                         return le32_to_cpu(*p);
423         }
424
425         /* No such thing, so let's try location of indirect block */
426         if (ind->bh)
427                 return ind->bh->b_blocknr;
428
429         /*
430          * It is going to be referred to from the inode itself? OK, just put it
431          * into the same cylinder group then.
432          */
433         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
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 ext3_fsblk_t 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                         ext3_fsblk_t goal, int indirect_blks, int blks,
520                         ext3_fsblk_t new_blocks[4], int *err)
521 {
522         int target, i;
523         unsigned long count = 0;
524         int index = 0;
525         ext3_fsblk_t 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, ext3_fsblk_t 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         ext3_fsblk_t new_blocks[4];
604         ext3_fsblk_t 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         ext3_fsblk_t 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         ext3_fsblk_t 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         ext3_fsblk_t 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                         ext3_fsblk_t 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 (count > blocks_to_boundary)
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 = ext3_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         /*
1014          * ext3_get_blocks_handle() returns number of blocks
1015          * mapped. 0 in case of a HOLE.
1016          */
1017         if (err > 0) {
1018                 if (err > 1)
1019                         WARN_ON(1);
1020                 err = 0;
1021         }
1022         *errp = err;
1023         if (!err && buffer_mapped(&dummy)) {
1024                 struct buffer_head *bh;
1025                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1026                 if (!bh) {
1027                         *errp = -EIO;
1028                         goto err;
1029                 }
1030                 if (buffer_new(&dummy)) {
1031                         J_ASSERT(create != 0);
1032                         J_ASSERT(handle != 0);
1033
1034                         /*
1035                          * Now that we do not always journal data, we should
1036                          * keep in mind whether this should always journal the
1037                          * new buffer as metadata.  For now, regular file
1038                          * writes use ext3_get_block instead, so it's not a
1039                          * problem.
1040                          */
1041                         lock_buffer(bh);
1042                         BUFFER_TRACE(bh, "call get_create_access");
1043                         fatal = ext3_journal_get_create_access(handle, bh);
1044                         if (!fatal && !buffer_uptodate(bh)) {
1045                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1046                                 set_buffer_uptodate(bh);
1047                         }
1048                         unlock_buffer(bh);
1049                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1050                         err = ext3_journal_dirty_metadata(handle, bh);
1051                         if (!fatal)
1052                                 fatal = err;
1053                 } else {
1054                         BUFFER_TRACE(bh, "not a new buffer");
1055                 }
1056                 if (fatal) {
1057                         *errp = fatal;
1058                         brelse(bh);
1059                         bh = NULL;
1060                 }
1061                 return bh;
1062         }
1063 err:
1064         return NULL;
1065 }
1066
1067 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1068                                int block, int create, int *err)
1069 {
1070         struct buffer_head * bh;
1071
1072         bh = ext3_getblk(handle, inode, block, create, err);
1073         if (!bh)
1074                 return bh;
1075         if (buffer_uptodate(bh))
1076                 return bh;
1077         ll_rw_block(READ_META, 1, &bh);
1078         wait_on_buffer(bh);
1079         if (buffer_uptodate(bh))
1080                 return bh;
1081         put_bh(bh);
1082         *err = -EIO;
1083         return NULL;
1084 }
1085
1086 static int walk_page_buffers(   handle_t *handle,
1087                                 struct buffer_head *head,
1088                                 unsigned from,
1089                                 unsigned to,
1090                                 int *partial,
1091                                 int (*fn)(      handle_t *handle,
1092                                                 struct buffer_head *bh))
1093 {
1094         struct buffer_head *bh;
1095         unsigned block_start, block_end;
1096         unsigned blocksize = head->b_size;
1097         int err, ret = 0;
1098         struct buffer_head *next;
1099
1100         for (   bh = head, block_start = 0;
1101                 ret == 0 && (bh != head || !block_start);
1102                 block_start = block_end, bh = next)
1103         {
1104                 next = bh->b_this_page;
1105                 block_end = block_start + blocksize;
1106                 if (block_end <= from || block_start >= to) {
1107                         if (partial && !buffer_uptodate(bh))
1108                                 *partial = 1;
1109                         continue;
1110                 }
1111                 err = (*fn)(handle, bh);
1112                 if (!ret)
1113                         ret = err;
1114         }
1115         return ret;
1116 }
1117
1118 /*
1119  * To preserve ordering, it is essential that the hole instantiation and
1120  * the data write be encapsulated in a single transaction.  We cannot
1121  * close off a transaction and start a new one between the ext3_get_block()
1122  * and the commit_write().  So doing the journal_start at the start of
1123  * prepare_write() is the right place.
1124  *
1125  * Also, this function can nest inside ext3_writepage() ->
1126  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1127  * has generated enough buffer credits to do the whole page.  So we won't
1128  * block on the journal in that case, which is good, because the caller may
1129  * be PF_MEMALLOC.
1130  *
1131  * By accident, ext3 can be reentered when a transaction is open via
1132  * quota file writes.  If we were to commit the transaction while thus
1133  * reentered, there can be a deadlock - we would be holding a quota
1134  * lock, and the commit would never complete if another thread had a
1135  * transaction open and was blocking on the quota lock - a ranking
1136  * violation.
1137  *
1138  * So what we do is to rely on the fact that journal_stop/journal_start
1139  * will _not_ run commit under these circumstances because handle->h_ref
1140  * is elevated.  We'll still have enough credits for the tiny quotafile
1141  * write.
1142  */
1143 static int do_journal_get_write_access(handle_t *handle,
1144                                         struct buffer_head *bh)
1145 {
1146         if (!buffer_mapped(bh) || buffer_freed(bh))
1147                 return 0;
1148         return ext3_journal_get_write_access(handle, bh);
1149 }
1150
1151 static int ext3_prepare_write(struct file *file, struct page *page,
1152                               unsigned from, unsigned to)
1153 {
1154         struct inode *inode = page->mapping->host;
1155         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1156         handle_t *handle;
1157         int retries = 0;
1158
1159 retry:
1160         handle = ext3_journal_start(inode, needed_blocks);
1161         if (IS_ERR(handle)) {
1162                 ret = PTR_ERR(handle);
1163                 goto out;
1164         }
1165         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1166                 ret = nobh_prepare_write(page, from, to, ext3_get_block);
1167         else
1168                 ret = block_prepare_write(page, from, to, ext3_get_block);
1169         if (ret)
1170                 goto prepare_write_failed;
1171
1172         if (ext3_should_journal_data(inode)) {
1173                 ret = walk_page_buffers(handle, page_buffers(page),
1174                                 from, to, NULL, do_journal_get_write_access);
1175         }
1176 prepare_write_failed:
1177         if (ret)
1178                 ext3_journal_stop(handle);
1179         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1180                 goto retry;
1181 out:
1182         return ret;
1183 }
1184
1185 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1186 {
1187         int err = journal_dirty_data(handle, bh);
1188         if (err)
1189                 ext3_journal_abort_handle(__FUNCTION__, __FUNCTION__,
1190                                                 bh, handle,err);
1191         return err;
1192 }
1193
1194 /* For commit_write() in data=journal mode */
1195 static int commit_write_fn(handle_t *handle, struct buffer_head *bh)
1196 {
1197         if (!buffer_mapped(bh) || buffer_freed(bh))
1198                 return 0;
1199         set_buffer_uptodate(bh);
1200         return ext3_journal_dirty_metadata(handle, bh);
1201 }
1202
1203 /*
1204  * We need to pick up the new inode size which generic_commit_write gave us
1205  * `file' can be NULL - eg, when called from page_symlink().
1206  *
1207  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1208  * buffers are managed internally.
1209  */
1210 static int ext3_ordered_commit_write(struct file *file, struct page *page,
1211                              unsigned from, unsigned to)
1212 {
1213         handle_t *handle = ext3_journal_current_handle();
1214         struct inode *inode = page->mapping->host;
1215         int ret = 0, ret2;
1216
1217         ret = walk_page_buffers(handle, page_buffers(page),
1218                 from, to, NULL, ext3_journal_dirty_data);
1219
1220         if (ret == 0) {
1221                 /*
1222                  * generic_commit_write() will run mark_inode_dirty() if i_size
1223                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1224                  * into that.
1225                  */
1226                 loff_t new_i_size;
1227
1228                 new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1229                 if (new_i_size > EXT3_I(inode)->i_disksize)
1230                         EXT3_I(inode)->i_disksize = new_i_size;
1231                 ret = generic_commit_write(file, page, from, to);
1232         }
1233         ret2 = ext3_journal_stop(handle);
1234         if (!ret)
1235                 ret = ret2;
1236         return ret;
1237 }
1238
1239 static int ext3_writeback_commit_write(struct file *file, struct page *page,
1240                              unsigned from, unsigned to)
1241 {
1242         handle_t *handle = ext3_journal_current_handle();
1243         struct inode *inode = page->mapping->host;
1244         int ret = 0, ret2;
1245         loff_t new_i_size;
1246
1247         new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1248         if (new_i_size > EXT3_I(inode)->i_disksize)
1249                 EXT3_I(inode)->i_disksize = new_i_size;
1250
1251         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1252                 ret = nobh_commit_write(file, page, from, to);
1253         else
1254                 ret = generic_commit_write(file, page, from, to);
1255
1256         ret2 = ext3_journal_stop(handle);
1257         if (!ret)
1258                 ret = ret2;
1259         return ret;
1260 }
1261
1262 static int ext3_journalled_commit_write(struct file *file,
1263                         struct page *page, unsigned from, unsigned to)
1264 {
1265         handle_t *handle = ext3_journal_current_handle();
1266         struct inode *inode = page->mapping->host;
1267         int ret = 0, ret2;
1268         int partial = 0;
1269         loff_t pos;
1270
1271         /*
1272          * Here we duplicate the generic_commit_write() functionality
1273          */
1274         pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1275
1276         ret = walk_page_buffers(handle, page_buffers(page), from,
1277                                 to, &partial, commit_write_fn);
1278         if (!partial)
1279                 SetPageUptodate(page);
1280         if (pos > inode->i_size)
1281                 i_size_write(inode, pos);
1282         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1283         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1284                 EXT3_I(inode)->i_disksize = inode->i_size;
1285                 ret2 = ext3_mark_inode_dirty(handle, inode);
1286                 if (!ret)
1287                         ret = ret2;
1288         }
1289         ret2 = ext3_journal_stop(handle);
1290         if (!ret)
1291                 ret = ret2;
1292         return ret;
1293 }
1294
1295 /*
1296  * bmap() is special.  It gets used by applications such as lilo and by
1297  * the swapper to find the on-disk block of a specific piece of data.
1298  *
1299  * Naturally, this is dangerous if the block concerned is still in the
1300  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1301  * filesystem and enables swap, then they may get a nasty shock when the
1302  * data getting swapped to that swapfile suddenly gets overwritten by
1303  * the original zero's written out previously to the journal and
1304  * awaiting writeback in the kernel's buffer cache.
1305  *
1306  * So, if we see any bmap calls here on a modified, data-journaled file,
1307  * take extra steps to flush any blocks which might be in the cache.
1308  */
1309 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1310 {
1311         struct inode *inode = mapping->host;
1312         journal_t *journal;
1313         int err;
1314
1315         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1316                 /*
1317                  * This is a REALLY heavyweight approach, but the use of
1318                  * bmap on dirty files is expected to be extremely rare:
1319                  * only if we run lilo or swapon on a freshly made file
1320                  * do we expect this to happen.
1321                  *
1322                  * (bmap requires CAP_SYS_RAWIO so this does not
1323                  * represent an unprivileged user DOS attack --- we'd be
1324                  * in trouble if mortal users could trigger this path at
1325                  * will.)
1326                  *
1327                  * NB. EXT3_STATE_JDATA is not set on files other than
1328                  * regular files.  If somebody wants to bmap a directory
1329                  * or symlink and gets confused because the buffer
1330                  * hasn't yet been flushed to disk, they deserve
1331                  * everything they get.
1332                  */
1333
1334                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1335                 journal = EXT3_JOURNAL(inode);
1336                 journal_lock_updates(journal);
1337                 err = journal_flush(journal);
1338                 journal_unlock_updates(journal);
1339
1340                 if (err)
1341                         return 0;
1342         }
1343
1344         return generic_block_bmap(mapping,block,ext3_get_block);
1345 }
1346
1347 static int bget_one(handle_t *handle, struct buffer_head *bh)
1348 {
1349         get_bh(bh);
1350         return 0;
1351 }
1352
1353 static int bput_one(handle_t *handle, struct buffer_head *bh)
1354 {
1355         put_bh(bh);
1356         return 0;
1357 }
1358
1359 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1360 {
1361         if (buffer_mapped(bh))
1362                 return ext3_journal_dirty_data(handle, bh);
1363         return 0;
1364 }
1365
1366 /*
1367  * Note that we always start a transaction even if we're not journalling
1368  * data.  This is to preserve ordering: any hole instantiation within
1369  * __block_write_full_page -> ext3_get_block() should be journalled
1370  * along with the data so we don't crash and then get metadata which
1371  * refers to old data.
1372  *
1373  * In all journalling modes block_write_full_page() will start the I/O.
1374  *
1375  * Problem:
1376  *
1377  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1378  *              ext3_writepage()
1379  *
1380  * Similar for:
1381  *
1382  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1383  *
1384  * Same applies to ext3_get_block().  We will deadlock on various things like
1385  * lock_journal and i_truncate_mutex.
1386  *
1387  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1388  * allocations fail.
1389  *
1390  * 16May01: If we're reentered then journal_current_handle() will be
1391  *          non-zero. We simply *return*.
1392  *
1393  * 1 July 2001: @@@ FIXME:
1394  *   In journalled data mode, a data buffer may be metadata against the
1395  *   current transaction.  But the same file is part of a shared mapping
1396  *   and someone does a writepage() on it.
1397  *
1398  *   We will move the buffer onto the async_data list, but *after* it has
1399  *   been dirtied. So there's a small window where we have dirty data on
1400  *   BJ_Metadata.
1401  *
1402  *   Note that this only applies to the last partial page in the file.  The
1403  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1404  *   broken code anyway: it's wrong for msync()).
1405  *
1406  *   It's a rare case: affects the final partial page, for journalled data
1407  *   where the file is subject to bith write() and writepage() in the same
1408  *   transction.  To fix it we'll need a custom block_write_full_page().
1409  *   We'll probably need that anyway for journalling writepage() output.
1410  *
1411  * We don't honour synchronous mounts for writepage().  That would be
1412  * disastrous.  Any write() or metadata operation will sync the fs for
1413  * us.
1414  *
1415  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1416  * we don't need to open a transaction here.
1417  */
1418 static int ext3_ordered_writepage(struct page *page,
1419                                 struct writeback_control *wbc)
1420 {
1421         struct inode *inode = page->mapping->host;
1422         struct buffer_head *page_bufs;
1423         handle_t *handle = NULL;
1424         int ret = 0;
1425         int err;
1426
1427         J_ASSERT(PageLocked(page));
1428
1429         /*
1430          * We give up here if we're reentered, because it might be for a
1431          * different filesystem.
1432          */
1433         if (ext3_journal_current_handle())
1434                 goto out_fail;
1435
1436         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1437
1438         if (IS_ERR(handle)) {
1439                 ret = PTR_ERR(handle);
1440                 goto out_fail;
1441         }
1442
1443         if (!page_has_buffers(page)) {
1444                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1445                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1446         }
1447         page_bufs = page_buffers(page);
1448         walk_page_buffers(handle, page_bufs, 0,
1449                         PAGE_CACHE_SIZE, NULL, bget_one);
1450
1451         ret = block_write_full_page(page, ext3_get_block, wbc);
1452
1453         /*
1454          * The page can become unlocked at any point now, and
1455          * truncate can then come in and change things.  So we
1456          * can't touch *page from now on.  But *page_bufs is
1457          * safe due to elevated refcount.
1458          */
1459
1460         /*
1461          * And attach them to the current transaction.  But only if
1462          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1463          * and generally junk.
1464          */
1465         if (ret == 0) {
1466                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1467                                         NULL, journal_dirty_data_fn);
1468                 if (!ret)
1469                         ret = err;
1470         }
1471         walk_page_buffers(handle, page_bufs, 0,
1472                         PAGE_CACHE_SIZE, NULL, bput_one);
1473         err = ext3_journal_stop(handle);
1474         if (!ret)
1475                 ret = err;
1476         return ret;
1477
1478 out_fail:
1479         redirty_page_for_writepage(wbc, page);
1480         unlock_page(page);
1481         return ret;
1482 }
1483
1484 static int ext3_writeback_writepage(struct page *page,
1485                                 struct writeback_control *wbc)
1486 {
1487         struct inode *inode = page->mapping->host;
1488         handle_t *handle = NULL;
1489         int ret = 0;
1490         int err;
1491
1492         if (ext3_journal_current_handle())
1493                 goto out_fail;
1494
1495         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1496         if (IS_ERR(handle)) {
1497                 ret = PTR_ERR(handle);
1498                 goto out_fail;
1499         }
1500
1501         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1502                 ret = nobh_writepage(page, ext3_get_block, wbc);
1503         else
1504                 ret = block_write_full_page(page, ext3_get_block, wbc);
1505
1506         err = ext3_journal_stop(handle);
1507         if (!ret)
1508                 ret = err;
1509         return ret;
1510
1511 out_fail:
1512         redirty_page_for_writepage(wbc, page);
1513         unlock_page(page);
1514         return ret;
1515 }
1516
1517 static int ext3_journalled_writepage(struct page *page,
1518                                 struct writeback_control *wbc)
1519 {
1520         struct inode *inode = page->mapping->host;
1521         handle_t *handle = NULL;
1522         int ret = 0;
1523         int err;
1524
1525         if (ext3_journal_current_handle())
1526                 goto no_write;
1527
1528         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1529         if (IS_ERR(handle)) {
1530                 ret = PTR_ERR(handle);
1531                 goto no_write;
1532         }
1533
1534         if (!page_has_buffers(page) || PageChecked(page)) {
1535                 /*
1536                  * It's mmapped pagecache.  Add buffers and journal it.  There
1537                  * doesn't seem much point in redirtying the page here.
1538                  */
1539                 ClearPageChecked(page);
1540                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1541                                         ext3_get_block);
1542                 if (ret != 0) {
1543                         ext3_journal_stop(handle);
1544                         goto out_unlock;
1545                 }
1546                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1547                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1548
1549                 err = walk_page_buffers(handle, page_buffers(page), 0,
1550                                 PAGE_CACHE_SIZE, NULL, commit_write_fn);
1551                 if (ret == 0)
1552                         ret = err;
1553                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1554                 unlock_page(page);
1555         } else {
1556                 /*
1557                  * It may be a page full of checkpoint-mode buffers.  We don't
1558                  * really know unless we go poke around in the buffer_heads.
1559                  * But block_write_full_page will do the right thing.
1560                  */
1561                 ret = block_write_full_page(page, ext3_get_block, wbc);
1562         }
1563         err = ext3_journal_stop(handle);
1564         if (!ret)
1565                 ret = err;
1566 out:
1567         return ret;
1568
1569 no_write:
1570         redirty_page_for_writepage(wbc, page);
1571 out_unlock:
1572         unlock_page(page);
1573         goto out;
1574 }
1575
1576 static int ext3_readpage(struct file *file, struct page *page)
1577 {
1578         return mpage_readpage(page, ext3_get_block);
1579 }
1580
1581 static int
1582 ext3_readpages(struct file *file, struct address_space *mapping,
1583                 struct list_head *pages, unsigned nr_pages)
1584 {
1585         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1586 }
1587
1588 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1589 {
1590         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1591
1592         /*
1593          * If it's a full truncate we just forget about the pending dirtying
1594          */
1595         if (offset == 0)
1596                 ClearPageChecked(page);
1597
1598         journal_invalidatepage(journal, page, offset);
1599 }
1600
1601 static int ext3_releasepage(struct page *page, gfp_t wait)
1602 {
1603         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1604
1605         WARN_ON(PageChecked(page));
1606         if (!page_has_buffers(page))
1607                 return 0;
1608         return journal_try_to_free_buffers(journal, page, wait);
1609 }
1610
1611 /*
1612  * If the O_DIRECT write will extend the file then add this inode to the
1613  * orphan list.  So recovery will truncate it back to the original size
1614  * if the machine crashes during the write.
1615  *
1616  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1617  * crashes then stale disk data _may_ be exposed inside the file.
1618  */
1619 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1620                         const struct iovec *iov, loff_t offset,
1621                         unsigned long nr_segs)
1622 {
1623         struct file *file = iocb->ki_filp;
1624         struct inode *inode = file->f_mapping->host;
1625         struct ext3_inode_info *ei = EXT3_I(inode);
1626         handle_t *handle = NULL;
1627         ssize_t ret;
1628         int orphan = 0;
1629         size_t count = iov_length(iov, nr_segs);
1630
1631         if (rw == WRITE) {
1632                 loff_t final_size = offset + count;
1633
1634                 handle = ext3_journal_start(inode, DIO_CREDITS);
1635                 if (IS_ERR(handle)) {
1636                         ret = PTR_ERR(handle);
1637                         goto out;
1638                 }
1639                 if (final_size > inode->i_size) {
1640                         ret = ext3_orphan_add(handle, inode);
1641                         if (ret)
1642                                 goto out_stop;
1643                         orphan = 1;
1644                         ei->i_disksize = inode->i_size;
1645                 }
1646         }
1647
1648         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1649                                  offset, nr_segs,
1650                                  ext3_get_block, NULL);
1651
1652         /*
1653          * Reacquire the handle: ext3_get_block() can restart the transaction
1654          */
1655         handle = ext3_journal_current_handle();
1656
1657 out_stop:
1658         if (handle) {
1659                 int err;
1660
1661                 if (orphan && inode->i_nlink)
1662                         ext3_orphan_del(handle, inode);
1663                 if (orphan && ret > 0) {
1664                         loff_t end = offset + ret;
1665                         if (end > inode->i_size) {
1666                                 ei->i_disksize = end;
1667                                 i_size_write(inode, end);
1668                                 /*
1669                                  * We're going to return a positive `ret'
1670                                  * here due to non-zero-length I/O, so there's
1671                                  * no way of reporting error returns from
1672                                  * ext3_mark_inode_dirty() to userspace.  So
1673                                  * ignore it.
1674                                  */
1675                                 ext3_mark_inode_dirty(handle, inode);
1676                         }
1677                 }
1678                 err = ext3_journal_stop(handle);
1679                 if (ret == 0)
1680                         ret = err;
1681         }
1682 out:
1683         return ret;
1684 }
1685
1686 /*
1687  * Pages can be marked dirty completely asynchronously from ext3's journalling
1688  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1689  * much here because ->set_page_dirty is called under VFS locks.  The page is
1690  * not necessarily locked.
1691  *
1692  * We cannot just dirty the page and leave attached buffers clean, because the
1693  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1694  * or jbddirty because all the journalling code will explode.
1695  *
1696  * So what we do is to mark the page "pending dirty" and next time writepage
1697  * is called, propagate that into the buffers appropriately.
1698  */
1699 static int ext3_journalled_set_page_dirty(struct page *page)
1700 {
1701         SetPageChecked(page);
1702         return __set_page_dirty_nobuffers(page);
1703 }
1704
1705 static const struct address_space_operations ext3_ordered_aops = {
1706         .readpage       = ext3_readpage,
1707         .readpages      = ext3_readpages,
1708         .writepage      = ext3_ordered_writepage,
1709         .sync_page      = block_sync_page,
1710         .prepare_write  = ext3_prepare_write,
1711         .commit_write   = ext3_ordered_commit_write,
1712         .bmap           = ext3_bmap,
1713         .invalidatepage = ext3_invalidatepage,
1714         .releasepage    = ext3_releasepage,
1715         .direct_IO      = ext3_direct_IO,
1716         .migratepage    = buffer_migrate_page,
1717 };
1718
1719 static const struct address_space_operations ext3_writeback_aops = {
1720         .readpage       = ext3_readpage,
1721         .readpages      = ext3_readpages,
1722         .writepage      = ext3_writeback_writepage,
1723         .sync_page      = block_sync_page,
1724         .prepare_write  = ext3_prepare_write,
1725         .commit_write   = ext3_writeback_commit_write,
1726         .bmap           = ext3_bmap,
1727         .invalidatepage = ext3_invalidatepage,
1728         .releasepage    = ext3_releasepage,
1729         .direct_IO      = ext3_direct_IO,
1730         .migratepage    = buffer_migrate_page,
1731 };
1732
1733 static const struct address_space_operations ext3_journalled_aops = {
1734         .readpage       = ext3_readpage,
1735         .readpages      = ext3_readpages,
1736         .writepage      = ext3_journalled_writepage,
1737         .sync_page      = block_sync_page,
1738         .prepare_write  = ext3_prepare_write,
1739         .commit_write   = ext3_journalled_commit_write,
1740         .set_page_dirty = ext3_journalled_set_page_dirty,
1741         .bmap           = ext3_bmap,
1742         .invalidatepage = ext3_invalidatepage,
1743         .releasepage    = ext3_releasepage,
1744 };
1745
1746 void ext3_set_aops(struct inode *inode)
1747 {
1748         if (ext3_should_order_data(inode))
1749                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1750         else if (ext3_should_writeback_data(inode))
1751                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1752         else
1753                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1754 }
1755
1756 /*
1757  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1758  * up to the end of the block which corresponds to `from'.
1759  * This required during truncate. We need to physically zero the tail end
1760  * of that block so it doesn't yield old data if the file is later grown.
1761  */
1762 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1763                 struct address_space *mapping, loff_t from)
1764 {
1765         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1766         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1767         unsigned blocksize, iblock, length, pos;
1768         struct inode *inode = mapping->host;
1769         struct buffer_head *bh;
1770         int err = 0;
1771         void *kaddr;
1772
1773         blocksize = inode->i_sb->s_blocksize;
1774         length = blocksize - (offset & (blocksize - 1));
1775         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1776
1777         /*
1778          * For "nobh" option,  we can only work if we don't need to
1779          * read-in the page - otherwise we create buffers to do the IO.
1780          */
1781         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1782              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1783                 kaddr = kmap_atomic(page, KM_USER0);
1784                 memset(kaddr + offset, 0, length);
1785                 flush_dcache_page(page);
1786                 kunmap_atomic(kaddr, KM_USER0);
1787                 set_page_dirty(page);
1788                 goto unlock;
1789         }
1790
1791         if (!page_has_buffers(page))
1792                 create_empty_buffers(page, blocksize, 0);
1793
1794         /* Find the buffer that contains "offset" */
1795         bh = page_buffers(page);
1796         pos = blocksize;
1797         while (offset >= pos) {
1798                 bh = bh->b_this_page;
1799                 iblock++;
1800                 pos += blocksize;
1801         }
1802
1803         err = 0;
1804         if (buffer_freed(bh)) {
1805                 BUFFER_TRACE(bh, "freed: skip");
1806                 goto unlock;
1807         }
1808
1809         if (!buffer_mapped(bh)) {
1810                 BUFFER_TRACE(bh, "unmapped");
1811                 ext3_get_block(inode, iblock, bh, 0);
1812                 /* unmapped? It's a hole - nothing to do */
1813                 if (!buffer_mapped(bh)) {
1814                         BUFFER_TRACE(bh, "still unmapped");
1815                         goto unlock;
1816                 }
1817         }
1818
1819         /* Ok, it's mapped. Make sure it's up-to-date */
1820         if (PageUptodate(page))
1821                 set_buffer_uptodate(bh);
1822
1823         if (!buffer_uptodate(bh)) {
1824                 err = -EIO;
1825                 ll_rw_block(READ, 1, &bh);
1826                 wait_on_buffer(bh);
1827                 /* Uhhuh. Read error. Complain and punt. */
1828                 if (!buffer_uptodate(bh))
1829                         goto unlock;
1830         }
1831
1832         if (ext3_should_journal_data(inode)) {
1833                 BUFFER_TRACE(bh, "get write access");
1834                 err = ext3_journal_get_write_access(handle, bh);
1835                 if (err)
1836                         goto unlock;
1837         }
1838
1839         kaddr = kmap_atomic(page, KM_USER0);
1840         memset(kaddr + offset, 0, length);
1841         flush_dcache_page(page);
1842         kunmap_atomic(kaddr, KM_USER0);
1843
1844         BUFFER_TRACE(bh, "zeroed end of block");
1845
1846         err = 0;
1847         if (ext3_should_journal_data(inode)) {
1848                 err = ext3_journal_dirty_metadata(handle, bh);
1849         } else {
1850                 if (ext3_should_order_data(inode))
1851                         err = ext3_journal_dirty_data(handle, bh);
1852                 mark_buffer_dirty(bh);
1853         }
1854
1855 unlock:
1856         unlock_page(page);
1857         page_cache_release(page);
1858         return err;
1859 }
1860
1861 /*
1862  * Probably it should be a library function... search for first non-zero word
1863  * or memcmp with zero_page, whatever is better for particular architecture.
1864  * Linus?
1865  */
1866 static inline int all_zeroes(__le32 *p, __le32 *q)
1867 {
1868         while (p < q)
1869                 if (*p++)
1870                         return 0;
1871         return 1;
1872 }
1873
1874 /**
1875  *      ext3_find_shared - find the indirect blocks for partial truncation.
1876  *      @inode:   inode in question
1877  *      @depth:   depth of the affected branch
1878  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1879  *      @chain:   place to store the pointers to partial indirect blocks
1880  *      @top:     place to the (detached) top of branch
1881  *
1882  *      This is a helper function used by ext3_truncate().
1883  *
1884  *      When we do truncate() we may have to clean the ends of several
1885  *      indirect blocks but leave the blocks themselves alive. Block is
1886  *      partially truncated if some data below the new i_size is refered
1887  *      from it (and it is on the path to the first completely truncated
1888  *      data block, indeed).  We have to free the top of that path along
1889  *      with everything to the right of the path. Since no allocation
1890  *      past the truncation point is possible until ext3_truncate()
1891  *      finishes, we may safely do the latter, but top of branch may
1892  *      require special attention - pageout below the truncation point
1893  *      might try to populate it.
1894  *
1895  *      We atomically detach the top of branch from the tree, store the
1896  *      block number of its root in *@top, pointers to buffer_heads of
1897  *      partially truncated blocks - in @chain[].bh and pointers to
1898  *      their last elements that should not be removed - in
1899  *      @chain[].p. Return value is the pointer to last filled element
1900  *      of @chain.
1901  *
1902  *      The work left to caller to do the actual freeing of subtrees:
1903  *              a) free the subtree starting from *@top
1904  *              b) free the subtrees whose roots are stored in
1905  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1906  *              c) free the subtrees growing from the inode past the @chain[0].
1907  *                      (no partially truncated stuff there).  */
1908
1909 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1910                         int offsets[4], Indirect chain[4], __le32 *top)
1911 {
1912         Indirect *partial, *p;
1913         int k, err;
1914
1915         *top = 0;
1916         /* Make k index the deepest non-null offest + 1 */
1917         for (k = depth; k > 1 && !offsets[k-1]; k--)
1918                 ;
1919         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1920         /* Writer: pointers */
1921         if (!partial)
1922                 partial = chain + k-1;
1923         /*
1924          * If the branch acquired continuation since we've looked at it -
1925          * fine, it should all survive and (new) top doesn't belong to us.
1926          */
1927         if (!partial->key && *partial->p)
1928                 /* Writer: end */
1929                 goto no_top;
1930         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1931                 ;
1932         /*
1933          * OK, we've found the last block that must survive. The rest of our
1934          * branch should be detached before unlocking. However, if that rest
1935          * of branch is all ours and does not grow immediately from the inode
1936          * it's easier to cheat and just decrement partial->p.
1937          */
1938         if (p == chain + k - 1 && p > chain) {
1939                 p->p--;
1940         } else {
1941                 *top = *p->p;
1942                 /* Nope, don't do this in ext3.  Must leave the tree intact */
1943 #if 0
1944                 *p->p = 0;
1945 #endif
1946         }
1947         /* Writer: end */
1948
1949         while(partial > p) {
1950                 brelse(partial->bh);
1951                 partial--;
1952         }
1953 no_top:
1954         return partial;
1955 }
1956
1957 /*
1958  * Zero a number of block pointers in either an inode or an indirect block.
1959  * If we restart the transaction we must again get write access to the
1960  * indirect block for further modification.
1961  *
1962  * We release `count' blocks on disk, but (last - first) may be greater
1963  * than `count' because there can be holes in there.
1964  */
1965 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
1966                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
1967                 unsigned long count, __le32 *first, __le32 *last)
1968 {
1969         __le32 *p;
1970         if (try_to_extend_transaction(handle, inode)) {
1971                 if (bh) {
1972                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1973                         ext3_journal_dirty_metadata(handle, bh);
1974                 }
1975                 ext3_mark_inode_dirty(handle, inode);
1976                 ext3_journal_test_restart(handle, inode);
1977                 if (bh) {
1978                         BUFFER_TRACE(bh, "retaking write access");
1979                         ext3_journal_get_write_access(handle, bh);
1980                 }
1981         }
1982
1983         /*
1984          * Any buffers which are on the journal will be in memory. We find
1985          * them on the hash table so journal_revoke() will run journal_forget()
1986          * on them.  We've already detached each block from the file, so
1987          * bforget() in journal_forget() should be safe.
1988          *
1989          * AKPM: turn on bforget in journal_forget()!!!
1990          */
1991         for (p = first; p < last; p++) {
1992                 u32 nr = le32_to_cpu(*p);
1993                 if (nr) {
1994                         struct buffer_head *bh;
1995
1996                         *p = 0;
1997                         bh = sb_find_get_block(inode->i_sb, nr);
1998                         ext3_forget(handle, 0, inode, bh, nr);
1999                 }
2000         }
2001
2002         ext3_free_blocks(handle, inode, block_to_free, count);
2003 }
2004
2005 /**
2006  * ext3_free_data - free a list of data blocks
2007  * @handle:     handle for this transaction
2008  * @inode:      inode we are dealing with
2009  * @this_bh:    indirect buffer_head which contains *@first and *@last
2010  * @first:      array of block numbers
2011  * @last:       points immediately past the end of array
2012  *
2013  * We are freeing all blocks refered from that array (numbers are stored as
2014  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2015  *
2016  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2017  * blocks are contiguous then releasing them at one time will only affect one
2018  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2019  * actually use a lot of journal space.
2020  *
2021  * @this_bh will be %NULL if @first and @last point into the inode's direct
2022  * block pointers.
2023  */
2024 static void ext3_free_data(handle_t *handle, struct inode *inode,
2025                            struct buffer_head *this_bh,
2026                            __le32 *first, __le32 *last)
2027 {
2028         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2029         unsigned long count = 0;            /* Number of blocks in the run */
2030         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2031                                                corresponding to
2032                                                block_to_free */
2033         ext3_fsblk_t nr;                    /* Current block # */
2034         __le32 *p;                          /* Pointer into inode/ind
2035                                                for current block */
2036         int err;
2037
2038         if (this_bh) {                          /* For indirect block */
2039                 BUFFER_TRACE(this_bh, "get_write_access");
2040                 err = ext3_journal_get_write_access(handle, this_bh);
2041                 /* Important: if we can't update the indirect pointers
2042                  * to the blocks, we can't free them. */
2043                 if (err)
2044                         return;
2045         }
2046
2047         for (p = first; p < last; p++) {
2048                 nr = le32_to_cpu(*p);
2049                 if (nr) {
2050                         /* accumulate blocks to free if they're contiguous */
2051                         if (count == 0) {
2052                                 block_to_free = nr;
2053                                 block_to_free_p = p;
2054                                 count = 1;
2055                         } else if (nr == block_to_free + count) {
2056                                 count++;
2057                         } else {
2058                                 ext3_clear_blocks(handle, inode, this_bh,
2059                                                   block_to_free,
2060                                                   count, block_to_free_p, p);
2061                                 block_to_free = nr;
2062                                 block_to_free_p = p;
2063                                 count = 1;
2064                         }
2065                 }
2066         }
2067
2068         if (count > 0)
2069                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2070                                   count, block_to_free_p, p);
2071
2072         if (this_bh) {
2073                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2074                 ext3_journal_dirty_metadata(handle, this_bh);
2075         }
2076 }
2077
2078 /**
2079  *      ext3_free_branches - free an array of branches
2080  *      @handle: JBD handle for this transaction
2081  *      @inode: inode we are dealing with
2082  *      @parent_bh: the buffer_head which contains *@first and *@last
2083  *      @first: array of block numbers
2084  *      @last:  pointer immediately past the end of array
2085  *      @depth: depth of the branches to free
2086  *
2087  *      We are freeing all blocks refered from these branches (numbers are
2088  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2089  *      appropriately.
2090  */
2091 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2092                                struct buffer_head *parent_bh,
2093                                __le32 *first, __le32 *last, int depth)
2094 {
2095         ext3_fsblk_t nr;
2096         __le32 *p;
2097
2098         if (is_handle_aborted(handle))
2099                 return;
2100
2101         if (depth--) {
2102                 struct buffer_head *bh;
2103                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2104                 p = last;
2105                 while (--p >= first) {
2106                         nr = le32_to_cpu(*p);
2107                         if (!nr)
2108                                 continue;               /* A hole */
2109
2110                         /* Go read the buffer for the next level down */
2111                         bh = sb_bread(inode->i_sb, nr);
2112
2113                         /*
2114                          * A read failure? Report error and clear slot
2115                          * (should be rare).
2116                          */
2117                         if (!bh) {
2118                                 ext3_error(inode->i_sb, "ext3_free_branches",
2119                                            "Read failure, inode=%lu, block="E3FSBLK,
2120                                            inode->i_ino, nr);
2121                                 continue;
2122                         }
2123
2124                         /* This zaps the entire block.  Bottom up. */
2125                         BUFFER_TRACE(bh, "free child branches");
2126                         ext3_free_branches(handle, inode, bh,
2127                                            (__le32*)bh->b_data,
2128                                            (__le32*)bh->b_data + addr_per_block,
2129                                            depth);
2130
2131                         /*
2132                          * We've probably journalled the indirect block several
2133                          * times during the truncate.  But it's no longer
2134                          * needed and we now drop it from the transaction via
2135                          * journal_revoke().
2136                          *
2137                          * That's easy if it's exclusively part of this
2138                          * transaction.  But if it's part of the committing
2139                          * transaction then journal_forget() will simply
2140                          * brelse() it.  That means that if the underlying
2141                          * block is reallocated in ext3_get_block(),
2142                          * unmap_underlying_metadata() will find this block
2143                          * and will try to get rid of it.  damn, damn.
2144                          *
2145                          * If this block has already been committed to the
2146                          * journal, a revoke record will be written.  And
2147                          * revoke records must be emitted *before* clearing
2148                          * this block's bit in the bitmaps.
2149                          */
2150                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2151
2152                         /*
2153                          * Everything below this this pointer has been
2154                          * released.  Now let this top-of-subtree go.
2155                          *
2156                          * We want the freeing of this indirect block to be
2157                          * atomic in the journal with the updating of the
2158                          * bitmap block which owns it.  So make some room in
2159                          * the journal.
2160                          *
2161                          * We zero the parent pointer *after* freeing its
2162                          * pointee in the bitmaps, so if extend_transaction()
2163                          * for some reason fails to put the bitmap changes and
2164                          * the release into the same transaction, recovery
2165                          * will merely complain about releasing a free block,
2166                          * rather than leaking blocks.
2167                          */
2168                         if (is_handle_aborted(handle))
2169                                 return;
2170                         if (try_to_extend_transaction(handle, inode)) {
2171                                 ext3_mark_inode_dirty(handle, inode);
2172                                 ext3_journal_test_restart(handle, inode);
2173                         }
2174
2175                         ext3_free_blocks(handle, inode, nr, 1);
2176
2177                         if (parent_bh) {
2178                                 /*
2179                                  * The block which we have just freed is
2180                                  * pointed to by an indirect block: journal it
2181                                  */
2182                                 BUFFER_TRACE(parent_bh, "get_write_access");
2183                                 if (!ext3_journal_get_write_access(handle,
2184                                                                    parent_bh)){
2185                                         *p = 0;
2186                                         BUFFER_TRACE(parent_bh,
2187                                         "call ext3_journal_dirty_metadata");
2188                                         ext3_journal_dirty_metadata(handle,
2189                                                                     parent_bh);
2190                                 }
2191                         }
2192                 }
2193         } else {
2194                 /* We have reached the bottom of the tree. */
2195                 BUFFER_TRACE(parent_bh, "free data blocks");
2196                 ext3_free_data(handle, inode, parent_bh, first, last);
2197         }
2198 }
2199
2200 /*
2201  * ext3_truncate()
2202  *
2203  * We block out ext3_get_block() block instantiations across the entire
2204  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2205  * simultaneously on behalf of the same inode.
2206  *
2207  * As we work through the truncate and commmit bits of it to the journal there
2208  * is one core, guiding principle: the file's tree must always be consistent on
2209  * disk.  We must be able to restart the truncate after a crash.
2210  *
2211  * The file's tree may be transiently inconsistent in memory (although it
2212  * probably isn't), but whenever we close off and commit a journal transaction,
2213  * the contents of (the filesystem + the journal) must be consistent and
2214  * restartable.  It's pretty simple, really: bottom up, right to left (although
2215  * left-to-right works OK too).
2216  *
2217  * Note that at recovery time, journal replay occurs *before* the restart of
2218  * truncate against the orphan inode list.
2219  *
2220  * The committed inode has the new, desired i_size (which is the same as
2221  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2222  * that this inode's truncate did not complete and it will again call
2223  * ext3_truncate() to have another go.  So there will be instantiated blocks
2224  * to the right of the truncation point in a crashed ext3 filesystem.  But
2225  * that's fine - as long as they are linked from the inode, the post-crash
2226  * ext3_truncate() run will find them and release them.
2227  */
2228 void ext3_truncate(struct inode *inode)
2229 {
2230         handle_t *handle;
2231         struct ext3_inode_info *ei = EXT3_I(inode);
2232         __le32 *i_data = ei->i_data;
2233         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2234         struct address_space *mapping = inode->i_mapping;
2235         int offsets[4];
2236         Indirect chain[4];
2237         Indirect *partial;
2238         __le32 nr = 0;
2239         int n;
2240         long last_block;
2241         unsigned blocksize = inode->i_sb->s_blocksize;
2242         struct page *page;
2243
2244         if (!(S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
2245             S_ISLNK(inode->i_mode)))
2246                 return;
2247         if (ext3_inode_is_fast_symlink(inode))
2248                 return;
2249         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2250                 return;
2251
2252         /*
2253          * We have to lock the EOF page here, because lock_page() nests
2254          * outside journal_start().
2255          */
2256         if ((inode->i_size & (blocksize - 1)) == 0) {
2257                 /* Block boundary? Nothing to do */
2258                 page = NULL;
2259         } else {
2260                 page = grab_cache_page(mapping,
2261                                 inode->i_size >> PAGE_CACHE_SHIFT);
2262                 if (!page)
2263                         return;
2264         }
2265
2266         handle = start_transaction(inode);
2267         if (IS_ERR(handle)) {
2268                 if (page) {
2269                         clear_highpage(page);
2270                         flush_dcache_page(page);
2271                         unlock_page(page);
2272                         page_cache_release(page);
2273                 }
2274                 return;         /* AKPM: return what? */
2275         }
2276
2277         last_block = (inode->i_size + blocksize-1)
2278                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2279
2280         if (page)
2281                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2282
2283         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2284         if (n == 0)
2285                 goto out_stop;  /* error */
2286
2287         /*
2288          * OK.  This truncate is going to happen.  We add the inode to the
2289          * orphan list, so that if this truncate spans multiple transactions,
2290          * and we crash, we will resume the truncate when the filesystem
2291          * recovers.  It also marks the inode dirty, to catch the new size.
2292          *
2293          * Implication: the file must always be in a sane, consistent
2294          * truncatable state while each transaction commits.
2295          */
2296         if (ext3_orphan_add(handle, inode))
2297                 goto out_stop;
2298
2299         /*
2300          * The orphan list entry will now protect us from any crash which
2301          * occurs before the truncate completes, so it is now safe to propagate
2302          * the new, shorter inode size (held for now in i_size) into the
2303          * on-disk inode. We do this via i_disksize, which is the value which
2304          * ext3 *really* writes onto the disk inode.
2305          */
2306         ei->i_disksize = inode->i_size;
2307
2308         /*
2309          * From here we block out all ext3_get_block() callers who want to
2310          * modify the block allocation tree.
2311          */
2312         mutex_lock(&ei->truncate_mutex);
2313
2314         if (n == 1) {           /* direct blocks */
2315                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2316                                i_data + EXT3_NDIR_BLOCKS);
2317                 goto do_indirects;
2318         }
2319
2320         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2321         /* Kill the top of shared branch (not detached) */
2322         if (nr) {
2323                 if (partial == chain) {
2324                         /* Shared branch grows from the inode */
2325                         ext3_free_branches(handle, inode, NULL,
2326                                            &nr, &nr+1, (chain+n-1) - partial);
2327                         *partial->p = 0;
2328                         /*
2329                          * We mark the inode dirty prior to restart,
2330                          * and prior to stop.  No need for it here.
2331                          */
2332                 } else {
2333                         /* Shared branch grows from an indirect block */
2334                         BUFFER_TRACE(partial->bh, "get_write_access");
2335                         ext3_free_branches(handle, inode, partial->bh,
2336                                         partial->p,
2337                                         partial->p+1, (chain+n-1) - partial);
2338                 }
2339         }
2340         /* Clear the ends of indirect blocks on the shared branch */
2341         while (partial > chain) {
2342                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2343                                    (__le32*)partial->bh->b_data+addr_per_block,
2344                                    (chain+n-1) - partial);
2345                 BUFFER_TRACE(partial->bh, "call brelse");
2346                 brelse (partial->bh);
2347                 partial--;
2348         }
2349 do_indirects:
2350         /* Kill the remaining (whole) subtrees */
2351         switch (offsets[0]) {
2352         default:
2353                 nr = i_data[EXT3_IND_BLOCK];
2354                 if (nr) {
2355                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2356                         i_data[EXT3_IND_BLOCK] = 0;
2357                 }
2358         case EXT3_IND_BLOCK:
2359                 nr = i_data[EXT3_DIND_BLOCK];
2360                 if (nr) {
2361                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2362                         i_data[EXT3_DIND_BLOCK] = 0;
2363                 }
2364         case EXT3_DIND_BLOCK:
2365                 nr = i_data[EXT3_TIND_BLOCK];
2366                 if (nr) {
2367                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2368                         i_data[EXT3_TIND_BLOCK] = 0;
2369                 }
2370         case EXT3_TIND_BLOCK:
2371                 ;
2372         }
2373
2374         ext3_discard_reservation(inode);
2375
2376         mutex_unlock(&ei->truncate_mutex);
2377         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2378         ext3_mark_inode_dirty(handle, inode);
2379
2380         /*
2381          * In a multi-transaction truncate, we only make the final transaction
2382          * synchronous
2383          */
2384         if (IS_SYNC(inode))
2385                 handle->h_sync = 1;
2386 out_stop:
2387         /*
2388          * If this was a simple ftruncate(), and the file will remain alive
2389          * then we need to clear up the orphan record which we created above.
2390          * However, if this was a real unlink then we were called by
2391          * ext3_delete_inode(), and we allow that function to clean up the
2392          * orphan info for us.
2393          */
2394         if (inode->i_nlink)
2395                 ext3_orphan_del(handle, inode);
2396
2397         ext3_journal_stop(handle);
2398 }
2399
2400 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2401                 unsigned long ino, struct ext3_iloc *iloc)
2402 {
2403         unsigned long desc, group_desc, block_group;
2404         unsigned long offset;
2405         ext3_fsblk_t block;
2406         struct buffer_head *bh;
2407         struct ext3_group_desc * gdp;
2408
2409         if (!ext3_valid_inum(sb, ino)) {
2410                 /*
2411                  * This error is already checked for in namei.c unless we are
2412                  * looking at an NFS filehandle, in which case no error
2413                  * report is needed
2414                  */
2415                 return 0;
2416         }
2417
2418         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2419         if (block_group >= EXT3_SB(sb)->s_groups_count) {
2420                 ext3_error(sb,"ext3_get_inode_block","group >= groups count");
2421                 return 0;
2422         }
2423         smp_rmb();
2424         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
2425         desc = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
2426         bh = EXT3_SB(sb)->s_group_desc[group_desc];
2427         if (!bh) {
2428                 ext3_error (sb, "ext3_get_inode_block",
2429                             "Descriptor not loaded");
2430                 return 0;
2431         }
2432
2433         gdp = (struct ext3_group_desc *)bh->b_data;
2434         /*
2435          * Figure out the offset within the block group inode table
2436          */
2437         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2438                 EXT3_INODE_SIZE(sb);
2439         block = le32_to_cpu(gdp[desc].bg_inode_table) +
2440                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2441
2442         iloc->block_group = block_group;
2443         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2444         return block;
2445 }
2446
2447 /*
2448  * ext3_get_inode_loc returns with an extra refcount against the inode's
2449  * underlying buffer_head on success. If 'in_mem' is true, we have all
2450  * data in memory that is needed to recreate the on-disk version of this
2451  * inode.
2452  */
2453 static int __ext3_get_inode_loc(struct inode *inode,
2454                                 struct ext3_iloc *iloc, int in_mem)
2455 {
2456         ext3_fsblk_t block;
2457         struct buffer_head *bh;
2458
2459         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2460         if (!block)
2461                 return -EIO;
2462
2463         bh = sb_getblk(inode->i_sb, block);
2464         if (!bh) {
2465                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2466                                 "unable to read inode block - "
2467                                 "inode=%lu, block="E3FSBLK,
2468                                  inode->i_ino, block);
2469                 return -EIO;
2470         }
2471         if (!buffer_uptodate(bh)) {
2472                 lock_buffer(bh);
2473                 if (buffer_uptodate(bh)) {
2474                         /* someone brought it uptodate while we waited */
2475                         unlock_buffer(bh);
2476                         goto has_buffer;
2477                 }
2478
2479                 /*
2480                  * If we have all information of the inode in memory and this
2481                  * is the only valid inode in the block, we need not read the
2482                  * block.
2483                  */
2484                 if (in_mem) {
2485                         struct buffer_head *bitmap_bh;
2486                         struct ext3_group_desc *desc;
2487                         int inodes_per_buffer;
2488                         int inode_offset, i;
2489                         int block_group;
2490                         int start;
2491
2492                         block_group = (inode->i_ino - 1) /
2493                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2494                         inodes_per_buffer = bh->b_size /
2495                                 EXT3_INODE_SIZE(inode->i_sb);
2496                         inode_offset = ((inode->i_ino - 1) %
2497                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2498                         start = inode_offset & ~(inodes_per_buffer - 1);
2499
2500                         /* Is the inode bitmap in cache? */
2501                         desc = ext3_get_group_desc(inode->i_sb,
2502                                                 block_group, NULL);
2503                         if (!desc)
2504                                 goto make_io;
2505
2506                         bitmap_bh = sb_getblk(inode->i_sb,
2507                                         le32_to_cpu(desc->bg_inode_bitmap));
2508                         if (!bitmap_bh)
2509                                 goto make_io;
2510
2511                         /*
2512                          * If the inode bitmap isn't in cache then the
2513                          * optimisation may end up performing two reads instead
2514                          * of one, so skip it.
2515                          */
2516                         if (!buffer_uptodate(bitmap_bh)) {
2517                                 brelse(bitmap_bh);
2518                                 goto make_io;
2519                         }
2520                         for (i = start; i < start + inodes_per_buffer; i++) {
2521                                 if (i == inode_offset)
2522                                         continue;
2523                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2524                                         break;
2525                         }
2526                         brelse(bitmap_bh);
2527                         if (i == start + inodes_per_buffer) {
2528                                 /* all other inodes are free, so skip I/O */
2529                                 memset(bh->b_data, 0, bh->b_size);
2530                                 set_buffer_uptodate(bh);
2531                                 unlock_buffer(bh);
2532                                 goto has_buffer;
2533                         }
2534                 }
2535
2536 make_io:
2537                 /*
2538                  * There are other valid inodes in the buffer, this inode
2539                  * has in-inode xattrs, or we don't have this inode in memory.
2540                  * Read the block from disk.
2541                  */
2542                 get_bh(bh);
2543                 bh->b_end_io = end_buffer_read_sync;
2544                 submit_bh(READ_META, bh);
2545                 wait_on_buffer(bh);
2546                 if (!buffer_uptodate(bh)) {
2547                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2548                                         "unable to read inode block - "
2549                                         "inode=%lu, block="E3FSBLK,
2550                                         inode->i_ino, block);
2551                         brelse(bh);
2552                         return -EIO;
2553                 }
2554         }
2555 has_buffer:
2556         iloc->bh = bh;
2557         return 0;
2558 }
2559
2560 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2561 {
2562         /* We have all inode data except xattrs in memory here. */
2563         return __ext3_get_inode_loc(inode, iloc,
2564                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2565 }
2566
2567 void ext3_set_inode_flags(struct inode *inode)
2568 {
2569         unsigned int flags = EXT3_I(inode)->i_flags;
2570
2571         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2572         if (flags & EXT3_SYNC_FL)
2573                 inode->i_flags |= S_SYNC;
2574         if (flags & EXT3_APPEND_FL)
2575                 inode->i_flags |= S_APPEND;
2576         if (flags & EXT3_IMMUTABLE_FL)
2577                 inode->i_flags |= S_IMMUTABLE;
2578         if (flags & EXT3_NOATIME_FL)
2579                 inode->i_flags |= S_NOATIME;
2580         if (flags & EXT3_DIRSYNC_FL)
2581                 inode->i_flags |= S_DIRSYNC;
2582 }
2583
2584 void ext3_read_inode(struct inode * inode)
2585 {
2586         struct ext3_iloc iloc;
2587         struct ext3_inode *raw_inode;
2588         struct ext3_inode_info *ei = EXT3_I(inode);
2589         struct buffer_head *bh;
2590         int block;
2591
2592 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2593         ei->i_acl = EXT3_ACL_NOT_CACHED;
2594         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2595 #endif
2596         ei->i_block_alloc_info = NULL;
2597
2598         if (__ext3_get_inode_loc(inode, &iloc, 0))
2599                 goto bad_inode;
2600         bh = iloc.bh;
2601         raw_inode = ext3_raw_inode(&iloc);
2602         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2603         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2604         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2605         if(!(test_opt (inode->i_sb, NO_UID32))) {
2606                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2607                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2608         }
2609         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2610         inode->i_size = le32_to_cpu(raw_inode->i_size);
2611         inode->i_atime.tv_sec = le32_to_cpu(raw_inode->i_atime);
2612         inode->i_ctime.tv_sec = le32_to_cpu(raw_inode->i_ctime);
2613         inode->i_mtime.tv_sec = le32_to_cpu(raw_inode->i_mtime);
2614         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2615
2616         ei->i_state = 0;
2617         ei->i_dir_start_lookup = 0;
2618         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2619         /* We now have enough fields to check if the inode was active or not.
2620          * This is needed because nfsd might try to access dead inodes
2621          * the test is that same one that e2fsck uses
2622          * NeilBrown 1999oct15
2623          */
2624         if (inode->i_nlink == 0) {
2625                 if (inode->i_mode == 0 ||
2626                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2627                         /* this inode is deleted */
2628                         brelse (bh);
2629                         goto bad_inode;
2630                 }
2631                 /* The only unlinked inodes we let through here have
2632                  * valid i_mode and are being read by the orphan
2633                  * recovery code: that's fine, we're about to complete
2634                  * the process of deleting those. */
2635         }
2636         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2637         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2638 #ifdef EXT3_FRAGMENTS
2639         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2640         ei->i_frag_no = raw_inode->i_frag;
2641         ei->i_frag_size = raw_inode->i_fsize;
2642 #endif
2643         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2644         if (!S_ISREG(inode->i_mode)) {
2645                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2646         } else {
2647                 inode->i_size |=
2648                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2649         }
2650         ei->i_disksize = inode->i_size;
2651         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2652         ei->i_block_group = iloc.block_group;
2653         /*
2654          * NOTE! The in-memory inode i_data array is in little-endian order
2655          * even on big-endian machines: we do NOT byteswap the block numbers!
2656          */
2657         for (block = 0; block < EXT3_N_BLOCKS; block++)
2658                 ei->i_data[block] = raw_inode->i_block[block];
2659         INIT_LIST_HEAD(&ei->i_orphan);
2660
2661         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2662             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2663                 /*
2664                  * When mke2fs creates big inodes it does not zero out
2665                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2666                  * so ignore those first few inodes.
2667                  */
2668                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2669                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2670                     EXT3_INODE_SIZE(inode->i_sb))
2671                         goto bad_inode;
2672                 if (ei->i_extra_isize == 0) {
2673                         /* The extra space is currently unused. Use it. */
2674                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2675                                             EXT3_GOOD_OLD_INODE_SIZE;
2676                 } else {
2677                         __le32 *magic = (void *)raw_inode +
2678                                         EXT3_GOOD_OLD_INODE_SIZE +
2679                                         ei->i_extra_isize;
2680                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2681                                  ei->i_state |= EXT3_STATE_XATTR;
2682                 }
2683         } else
2684                 ei->i_extra_isize = 0;
2685
2686         if (S_ISREG(inode->i_mode)) {
2687                 inode->i_op = &ext3_file_inode_operations;
2688                 inode->i_fop = &ext3_file_operations;
2689                 ext3_set_aops(inode);
2690         } else if (S_ISDIR(inode->i_mode)) {
2691                 inode->i_op = &ext3_dir_inode_operations;
2692                 inode->i_fop = &ext3_dir_operations;
2693         } else if (S_ISLNK(inode->i_mode)) {
2694                 if (ext3_inode_is_fast_symlink(inode))
2695                         inode->i_op = &ext3_fast_symlink_inode_operations;
2696                 else {
2697                         inode->i_op = &ext3_symlink_inode_operations;
2698                         ext3_set_aops(inode);
2699                 }
2700         } else {
2701                 inode->i_op = &ext3_special_inode_operations;
2702                 if (raw_inode->i_block[0])
2703                         init_special_inode(inode, inode->i_mode,
2704                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2705                 else
2706                         init_special_inode(inode, inode->i_mode,
2707                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2708         }
2709         brelse (iloc.bh);
2710         ext3_set_inode_flags(inode);
2711         return;
2712
2713 bad_inode:
2714         make_bad_inode(inode);
2715         return;
2716 }
2717
2718 /*
2719  * Post the struct inode info into an on-disk inode location in the
2720  * buffer-cache.  This gobbles the caller's reference to the
2721  * buffer_head in the inode location struct.
2722  *
2723  * The caller must have write access to iloc->bh.
2724  */
2725 static int ext3_do_update_inode(handle_t *handle,
2726                                 struct inode *inode,
2727                                 struct ext3_iloc *iloc)
2728 {
2729         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2730         struct ext3_inode_info *ei = EXT3_I(inode);
2731         struct buffer_head *bh = iloc->bh;
2732         int err = 0, rc, block;
2733
2734         /* For fields not not tracking in the in-memory inode,
2735          * initialise them to zero for new inodes. */
2736         if (ei->i_state & EXT3_STATE_NEW)
2737                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2738
2739         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2740         if(!(test_opt(inode->i_sb, NO_UID32))) {
2741                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2742                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2743 /*
2744  * Fix up interoperability with old kernels. Otherwise, old inodes get
2745  * re-used with the upper 16 bits of the uid/gid intact
2746  */
2747                 if(!ei->i_dtime) {
2748                         raw_inode->i_uid_high =
2749                                 cpu_to_le16(high_16_bits(inode->i_uid));
2750                         raw_inode->i_gid_high =
2751                                 cpu_to_le16(high_16_bits(inode->i_gid));
2752                 } else {
2753                         raw_inode->i_uid_high = 0;
2754                         raw_inode->i_gid_high = 0;
2755                 }
2756         } else {
2757                 raw_inode->i_uid_low =
2758                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2759                 raw_inode->i_gid_low =
2760                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2761                 raw_inode->i_uid_high = 0;
2762                 raw_inode->i_gid_high = 0;
2763         }
2764         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2765         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2766         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2767         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2768         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2769         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2770         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2771         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2772 #ifdef EXT3_FRAGMENTS
2773         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2774         raw_inode->i_frag = ei->i_frag_no;
2775         raw_inode->i_fsize = ei->i_frag_size;
2776 #endif
2777         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2778         if (!S_ISREG(inode->i_mode)) {
2779                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2780         } else {
2781                 raw_inode->i_size_high =
2782                         cpu_to_le32(ei->i_disksize >> 32);
2783                 if (ei->i_disksize > 0x7fffffffULL) {
2784                         struct super_block *sb = inode->i_sb;
2785                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2786                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2787                             EXT3_SB(sb)->s_es->s_rev_level ==
2788                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2789                                /* If this is the first large file
2790                                 * created, add a flag to the superblock.
2791                                 */
2792                                 err = ext3_journal_get_write_access(handle,
2793                                                 EXT3_SB(sb)->s_sbh);
2794                                 if (err)
2795                                         goto out_brelse;
2796                                 ext3_update_dynamic_rev(sb);
2797                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2798                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2799                                 sb->s_dirt = 1;
2800                                 handle->h_sync = 1;
2801                                 err = ext3_journal_dirty_metadata(handle,
2802                                                 EXT3_SB(sb)->s_sbh);
2803                         }
2804                 }
2805         }
2806         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2807         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2808                 if (old_valid_dev(inode->i_rdev)) {
2809                         raw_inode->i_block[0] =
2810                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2811                         raw_inode->i_block[1] = 0;
2812                 } else {
2813                         raw_inode->i_block[0] = 0;
2814                         raw_inode->i_block[1] =
2815                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2816                         raw_inode->i_block[2] = 0;
2817                 }
2818         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2819                 raw_inode->i_block[block] = ei->i_data[block];
2820
2821         if (ei->i_extra_isize)
2822                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2823
2824         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2825         rc = ext3_journal_dirty_metadata(handle, bh);
2826         if (!err)
2827                 err = rc;
2828         ei->i_state &= ~EXT3_STATE_NEW;
2829
2830 out_brelse:
2831         brelse (bh);
2832         ext3_std_error(inode->i_sb, err);
2833         return err;
2834 }
2835
2836 /*
2837  * ext3_write_inode()
2838  *
2839  * We are called from a few places:
2840  *
2841  * - Within generic_file_write() for O_SYNC files.
2842  *   Here, there will be no transaction running. We wait for any running
2843  *   trasnaction to commit.
2844  *
2845  * - Within sys_sync(), kupdate and such.
2846  *   We wait on commit, if tol to.
2847  *
2848  * - Within prune_icache() (PF_MEMALLOC == true)
2849  *   Here we simply return.  We can't afford to block kswapd on the
2850  *   journal commit.
2851  *
2852  * In all cases it is actually safe for us to return without doing anything,
2853  * because the inode has been copied into a raw inode buffer in
2854  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2855  * knfsd.
2856  *
2857  * Note that we are absolutely dependent upon all inode dirtiers doing the
2858  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2859  * which we are interested.
2860  *
2861  * It would be a bug for them to not do this.  The code:
2862  *
2863  *      mark_inode_dirty(inode)
2864  *      stuff();
2865  *      inode->i_size = expr;
2866  *
2867  * is in error because a kswapd-driven write_inode() could occur while
2868  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2869  * will no longer be on the superblock's dirty inode list.
2870  */
2871 int ext3_write_inode(struct inode *inode, int wait)
2872 {
2873         if (current->flags & PF_MEMALLOC)
2874                 return 0;
2875
2876         if (ext3_journal_current_handle()) {
2877                 jbd_debug(0, "called recursively, non-PF_MEMALLOC!\n");
2878                 dump_stack();
2879                 return -EIO;
2880         }
2881
2882         if (!wait)
2883                 return 0;
2884
2885         return ext3_force_commit(inode->i_sb);
2886 }
2887
2888 /*
2889  * ext3_setattr()
2890  *
2891  * Called from notify_change.
2892  *
2893  * We want to trap VFS attempts to truncate the file as soon as
2894  * possible.  In particular, we want to make sure that when the VFS
2895  * shrinks i_size, we put the inode on the orphan list and modify
2896  * i_disksize immediately, so that during the subsequent flushing of
2897  * dirty pages and freeing of disk blocks, we can guarantee that any
2898  * commit will leave the blocks being flushed in an unused state on
2899  * disk.  (On recovery, the inode will get truncated and the blocks will
2900  * be freed, so we have a strong guarantee that no future commit will
2901  * leave these blocks visible to the user.)
2902  *
2903  * Called with inode->sem down.
2904  */
2905 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
2906 {
2907         struct inode *inode = dentry->d_inode;
2908         int error, rc = 0;
2909         const unsigned int ia_valid = attr->ia_valid;
2910
2911         error = inode_change_ok(inode, attr);
2912         if (error)
2913                 return error;
2914
2915         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
2916                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
2917                 handle_t *handle;
2918
2919                 /* (user+group)*(old+new) structure, inode write (sb,
2920                  * inode block, ? - but truncate inode update has it) */
2921                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
2922                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
2923                 if (IS_ERR(handle)) {
2924                         error = PTR_ERR(handle);
2925                         goto err_out;
2926                 }
2927                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
2928                 if (error) {
2929                         ext3_journal_stop(handle);
2930                         return error;
2931                 }
2932                 /* Update corresponding info in inode so that everything is in
2933                  * one transaction */
2934                 if (attr->ia_valid & ATTR_UID)
2935                         inode->i_uid = attr->ia_uid;
2936                 if (attr->ia_valid & ATTR_GID)
2937                         inode->i_gid = attr->ia_gid;
2938                 error = ext3_mark_inode_dirty(handle, inode);
2939                 ext3_journal_stop(handle);
2940         }
2941
2942         if (S_ISREG(inode->i_mode) &&
2943             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
2944                 handle_t *handle;
2945
2946                 handle = ext3_journal_start(inode, 3);
2947                 if (IS_ERR(handle)) {
2948                         error = PTR_ERR(handle);
2949                         goto err_out;
2950                 }
2951
2952                 error = ext3_orphan_add(handle, inode);
2953                 EXT3_I(inode)->i_disksize = attr->ia_size;
2954                 rc = ext3_mark_inode_dirty(handle, inode);
2955                 if (!error)
2956                         error = rc;
2957                 ext3_journal_stop(handle);
2958         }
2959
2960         rc = inode_setattr(inode, attr);
2961
2962         /* If inode_setattr's call to ext3_truncate failed to get a
2963          * transaction handle at all, we need to clean up the in-core
2964          * orphan list manually. */
2965         if (inode->i_nlink)
2966                 ext3_orphan_del(NULL, inode);
2967
2968         if (!rc && (ia_valid & ATTR_MODE))
2969                 rc = ext3_acl_chmod(inode);
2970
2971 err_out:
2972         ext3_std_error(inode->i_sb, error);
2973         if (!error)
2974                 error = rc;
2975         return error;
2976 }
2977
2978
2979 /*
2980  * How many blocks doth make a writepage()?
2981  *
2982  * With N blocks per page, it may be:
2983  * N data blocks
2984  * 2 indirect block
2985  * 2 dindirect
2986  * 1 tindirect
2987  * N+5 bitmap blocks (from the above)
2988  * N+5 group descriptor summary blocks
2989  * 1 inode block
2990  * 1 superblock.
2991  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
2992  *
2993  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
2994  *
2995  * With ordered or writeback data it's the same, less the N data blocks.
2996  *
2997  * If the inode's direct blocks can hold an integral number of pages then a
2998  * page cannot straddle two indirect blocks, and we can only touch one indirect
2999  * and dindirect block, and the "5" above becomes "3".
3000  *
3001  * This still overestimates under most circumstances.  If we were to pass the
3002  * start and end offsets in here as well we could do block_to_path() on each
3003  * block and work out the exact number of indirects which are touched.  Pah.
3004  */
3005
3006 static int ext3_writepage_trans_blocks(struct inode *inode)
3007 {
3008         int bpp = ext3_journal_blocks_per_page(inode);
3009         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3010         int ret;
3011
3012         if (ext3_should_journal_data(inode))
3013                 ret = 3 * (bpp + indirects) + 2;
3014         else
3015                 ret = 2 * (bpp + indirects) + 2;
3016
3017 #ifdef CONFIG_QUOTA
3018         /* We know that structure was already allocated during DQUOT_INIT so
3019          * we will be updating only the data blocks + inodes */
3020         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3021 #endif
3022
3023         return ret;
3024 }
3025
3026 /*
3027  * The caller must have previously called ext3_reserve_inode_write().
3028  * Give this, we know that the caller already has write access to iloc->bh.
3029  */
3030 int ext3_mark_iloc_dirty(handle_t *handle,
3031                 struct inode *inode, struct ext3_iloc *iloc)
3032 {
3033         int err = 0;
3034
3035         /* the do_update_inode consumes one bh->b_count */
3036         get_bh(iloc->bh);
3037
3038         /* ext3_do_update_inode() does journal_dirty_metadata */
3039         err = ext3_do_update_inode(handle, inode, iloc);
3040         put_bh(iloc->bh);
3041         return err;
3042 }
3043
3044 /*
3045  * On success, We end up with an outstanding reference count against
3046  * iloc->bh.  This _must_ be cleaned up later.
3047  */
3048
3049 int
3050 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3051                          struct ext3_iloc *iloc)
3052 {
3053         int err = 0;
3054         if (handle) {
3055                 err = ext3_get_inode_loc(inode, iloc);
3056                 if (!err) {
3057                         BUFFER_TRACE(iloc->bh, "get_write_access");
3058                         err = ext3_journal_get_write_access(handle, iloc->bh);
3059                         if (err) {
3060                                 brelse(iloc->bh);
3061                                 iloc->bh = NULL;
3062                         }
3063                 }
3064         }
3065         ext3_std_error(inode->i_sb, err);
3066         return err;
3067 }
3068
3069 /*
3070  * What we do here is to mark the in-core inode as clean with respect to inode
3071  * dirtiness (it may still be data-dirty).
3072  * This means that the in-core inode may be reaped by prune_icache
3073  * without having to perform any I/O.  This is a very good thing,
3074  * because *any* task may call prune_icache - even ones which
3075  * have a transaction open against a different journal.
3076  *
3077  * Is this cheating?  Not really.  Sure, we haven't written the
3078  * inode out, but prune_icache isn't a user-visible syncing function.
3079  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3080  * we start and wait on commits.
3081  *
3082  * Is this efficient/effective?  Well, we're being nice to the system
3083  * by cleaning up our inodes proactively so they can be reaped
3084  * without I/O.  But we are potentially leaving up to five seconds'
3085  * worth of inodes floating about which prune_icache wants us to
3086  * write out.  One way to fix that would be to get prune_icache()
3087  * to do a write_super() to free up some memory.  It has the desired
3088  * effect.
3089  */
3090 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3091 {
3092         struct ext3_iloc iloc;
3093         int err;
3094
3095         might_sleep();
3096         err = ext3_reserve_inode_write(handle, inode, &iloc);
3097         if (!err)
3098                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3099         return err;
3100 }
3101
3102 /*
3103  * ext3_dirty_inode() is called from __mark_inode_dirty()
3104  *
3105  * We're really interested in the case where a file is being extended.
3106  * i_size has been changed by generic_commit_write() and we thus need
3107  * to include the updated inode in the current transaction.
3108  *
3109  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3110  * are allocated to the file.
3111  *
3112  * If the inode is marked synchronous, we don't honour that here - doing
3113  * so would cause a commit on atime updates, which we don't bother doing.
3114  * We handle synchronous inodes at the highest possible level.
3115  */
3116 void ext3_dirty_inode(struct inode *inode)
3117 {
3118         handle_t *current_handle = ext3_journal_current_handle();
3119         handle_t *handle;
3120
3121         handle = ext3_journal_start(inode, 2);
3122         if (IS_ERR(handle))
3123                 goto out;
3124         if (current_handle &&
3125                 current_handle->h_transaction != handle->h_transaction) {
3126                 /* This task has a transaction open against a different fs */
3127                 printk(KERN_EMERG "%s: transactions do not match!\n",
3128                        __FUNCTION__);
3129         } else {
3130                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3131                                 current_handle);
3132                 ext3_mark_inode_dirty(handle, inode);
3133         }
3134         ext3_journal_stop(handle);
3135 out:
3136         return;
3137 }
3138
3139 #if 0
3140 /*
3141  * Bind an inode's backing buffer_head into this transaction, to prevent
3142  * it from being flushed to disk early.  Unlike
3143  * ext3_reserve_inode_write, this leaves behind no bh reference and
3144  * returns no iloc structure, so the caller needs to repeat the iloc
3145  * lookup to mark the inode dirty later.
3146  */
3147 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3148 {
3149         struct ext3_iloc iloc;
3150
3151         int err = 0;
3152         if (handle) {
3153                 err = ext3_get_inode_loc(inode, &iloc);
3154                 if (!err) {
3155                         BUFFER_TRACE(iloc.bh, "get_write_access");
3156                         err = journal_get_write_access(handle, iloc.bh);
3157                         if (!err)
3158                                 err = ext3_journal_dirty_metadata(handle,
3159                                                                   iloc.bh);
3160                         brelse(iloc.bh);
3161                 }
3162         }
3163         ext3_std_error(inode->i_sb, err);
3164         return err;
3165 }
3166 #endif
3167
3168 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3169 {
3170         journal_t *journal;
3171         handle_t *handle;
3172         int err;
3173
3174         /*
3175          * We have to be very careful here: changing a data block's
3176          * journaling status dynamically is dangerous.  If we write a
3177          * data block to the journal, change the status and then delete
3178          * that block, we risk forgetting to revoke the old log record
3179          * from the journal and so a subsequent replay can corrupt data.
3180          * So, first we make sure that the journal is empty and that
3181          * nobody is changing anything.
3182          */
3183
3184         journal = EXT3_JOURNAL(inode);
3185         if (is_journal_aborted(journal) || IS_RDONLY(inode))
3186                 return -EROFS;
3187
3188         journal_lock_updates(journal);
3189         journal_flush(journal);
3190
3191         /*
3192          * OK, there are no updates running now, and all cached data is
3193          * synced to disk.  We are now in a completely consistent state
3194          * which doesn't have anything in the journal, and we know that
3195          * no filesystem updates are running, so it is safe to modify
3196          * the inode's in-core data-journaling state flag now.
3197          */
3198
3199         if (val)
3200                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3201         else
3202                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3203         ext3_set_aops(inode);
3204
3205         journal_unlock_updates(journal);
3206
3207         /* Finally we can mark the inode as dirty. */
3208
3209         handle = ext3_journal_start(inode, 1);
3210         if (IS_ERR(handle))
3211                 return PTR_ERR(handle);
3212
3213         err = ext3_mark_inode_dirty(handle, inode);
3214         handle->h_sync = 1;
3215         ext3_journal_stop(handle);
3216         ext3_std_error(inode->i_sb, err);
3217
3218         return err;
3219 }