Pull cpuidle into release branch
[linux-2.6] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include "xattr.h"
40 #include "acl.h"
41
42 static int ext3_writepage_trans_blocks(struct inode *inode);
43
44 /*
45  * Test whether an inode is a fast symlink.
46  */
47 static int ext3_inode_is_fast_symlink(struct inode *inode)
48 {
49         int ea_blocks = EXT3_I(inode)->i_file_acl ?
50                 (inode->i_sb->s_blocksize >> 9) : 0;
51
52         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
53 }
54
55 /*
56  * The ext3 forget function must perform a revoke if we are freeing data
57  * which has been journaled.  Metadata (eg. indirect blocks) must be
58  * revoked in all cases.
59  *
60  * "bh" may be NULL: a metadata block may have been freed from memory
61  * but there may still be a record of it in the journal, and that record
62  * still needs to be revoked.
63  */
64 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
65                         struct buffer_head *bh, ext3_fsblk_t blocknr)
66 {
67         int err;
68
69         might_sleep();
70
71         BUFFER_TRACE(bh, "enter");
72
73         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
74                   "data mode %lx\n",
75                   bh, is_metadata, inode->i_mode,
76                   test_opt(inode->i_sb, DATA_FLAGS));
77
78         /* Never use the revoke function if we are doing full data
79          * journaling: there is no need to, and a V1 superblock won't
80          * support it.  Otherwise, only skip the revoke on un-journaled
81          * data blocks. */
82
83         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
84             (!is_metadata && !ext3_should_journal_data(inode))) {
85                 if (bh) {
86                         BUFFER_TRACE(bh, "call journal_forget");
87                         return ext3_journal_forget(handle, bh);
88                 }
89                 return 0;
90         }
91
92         /*
93          * data!=journal && (is_metadata || should_journal_data(inode))
94          */
95         BUFFER_TRACE(bh, "call ext3_journal_revoke");
96         err = ext3_journal_revoke(handle, blocknr, bh);
97         if (err)
98                 ext3_abort(inode->i_sb, __FUNCTION__,
99                            "error %d when attempting revoke", err);
100         BUFFER_TRACE(bh, "exit");
101         return err;
102 }
103
104 /*
105  * Work out how many blocks we need to proceed with the next chunk of a
106  * truncate transaction.
107  */
108 static unsigned long blocks_for_truncate(struct inode *inode)
109 {
110         unsigned long needed;
111
112         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
113
114         /* Give ourselves just enough room to cope with inodes in which
115          * i_blocks is corrupt: we've seen disk corruptions in the past
116          * which resulted in random data in an inode which looked enough
117          * like a regular file for ext3 to try to delete it.  Things
118          * will go a bit crazy if that happens, but at least we should
119          * try not to panic the whole kernel. */
120         if (needed < 2)
121                 needed = 2;
122
123         /* But we need to bound the transaction so we don't overflow the
124          * journal. */
125         if (needed > EXT3_MAX_TRANS_DATA)
126                 needed = EXT3_MAX_TRANS_DATA;
127
128         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
129 }
130
131 /*
132  * Truncate transactions can be complex and absolutely huge.  So we need to
133  * be able to restart the transaction at a conventient checkpoint to make
134  * sure we don't overflow the journal.
135  *
136  * start_transaction gets us a new handle for a truncate transaction,
137  * and extend_transaction tries to extend the existing one a bit.  If
138  * extend fails, we need to propagate the failure up and restart the
139  * transaction in the top-level truncate loop. --sct
140  */
141 static handle_t *start_transaction(struct inode *inode)
142 {
143         handle_t *result;
144
145         result = ext3_journal_start(inode, blocks_for_truncate(inode));
146         if (!IS_ERR(result))
147                 return result;
148
149         ext3_std_error(inode->i_sb, PTR_ERR(result));
150         return result;
151 }
152
153 /*
154  * Try to extend this transaction for the purposes of truncation.
155  *
156  * Returns 0 if we managed to create more room.  If we can't create more
157  * room, and the transaction must be restarted we return 1.
158  */
159 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
160 {
161         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
162                 return 0;
163         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
164                 return 0;
165         return 1;
166 }
167
168 /*
169  * Restart the transaction associated with *handle.  This does a commit,
170  * so before we call here everything must be consistently dirtied against
171  * this transaction.
172  */
173 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
174 {
175         jbd_debug(2, "restarting handle %p\n", handle);
176         return ext3_journal_restart(handle, blocks_for_truncate(inode));
177 }
178
179 /*
180  * Called at the last iput() if i_nlink is zero.
181  */
182 void ext3_delete_inode (struct inode * inode)
183 {
184         handle_t *handle;
185
186         truncate_inode_pages(&inode->i_data, 0);
187
188         if (is_bad_inode(inode))
189                 goto no_delete;
190
191         handle = start_transaction(inode);
192         if (IS_ERR(handle)) {
193                 /*
194                  * If we're going to skip the normal cleanup, we still need to
195                  * make sure that the in-core orphan linked list is properly
196                  * cleaned up.
197                  */
198                 ext3_orphan_del(NULL, inode);
199                 goto no_delete;
200         }
201
202         if (IS_SYNC(inode))
203                 handle->h_sync = 1;
204         inode->i_size = 0;
205         if (inode->i_blocks)
206                 ext3_truncate(inode);
207         /*
208          * Kill off the orphan record which ext3_truncate created.
209          * AKPM: I think this can be inside the above `if'.
210          * Note that ext3_orphan_del() has to be able to cope with the
211          * deletion of a non-existent orphan - this is because we don't
212          * know if ext3_truncate() actually created an orphan record.
213          * (Well, we could do this if we need to, but heck - it works)
214          */
215         ext3_orphan_del(handle, inode);
216         EXT3_I(inode)->i_dtime  = get_seconds();
217
218         /*
219          * One subtle ordering requirement: if anything has gone wrong
220          * (transaction abort, IO errors, whatever), then we can still
221          * do these next steps (the fs will already have been marked as
222          * having errors), but we can't free the inode if the mark_dirty
223          * fails.
224          */
225         if (ext3_mark_inode_dirty(handle, inode))
226                 /* If that failed, just do the required in-core inode clear. */
227                 clear_inode(inode);
228         else
229                 ext3_free_inode(handle, inode);
230         ext3_journal_stop(handle);
231         return;
232 no_delete:
233         clear_inode(inode);     /* We must guarantee clearing of inode... */
234 }
235
236 typedef struct {
237         __le32  *p;
238         __le32  key;
239         struct buffer_head *bh;
240 } Indirect;
241
242 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
243 {
244         p->key = *(p->p = v);
245         p->bh = bh;
246 }
247
248 static int verify_chain(Indirect *from, Indirect *to)
249 {
250         while (from <= to && from->key == *from->p)
251                 from++;
252         return (from > to);
253 }
254
255 /**
256  *      ext3_block_to_path - parse the block number into array of offsets
257  *      @inode: inode in question (we are only interested in its superblock)
258  *      @i_block: block number to be parsed
259  *      @offsets: array to store the offsets in
260  *      @boundary: set this non-zero if the referred-to block is likely to be
261  *             followed (on disk) by an indirect block.
262  *
263  *      To store the locations of file's data ext3 uses a data structure common
264  *      for UNIX filesystems - tree of pointers anchored in the inode, with
265  *      data blocks at leaves and indirect blocks in intermediate nodes.
266  *      This function translates the block number into path in that tree -
267  *      return value is the path length and @offsets[n] is the offset of
268  *      pointer to (n+1)th node in the nth one. If @block is out of range
269  *      (negative or too large) warning is printed and zero returned.
270  *
271  *      Note: function doesn't find node addresses, so no IO is needed. All
272  *      we need to know is the capacity of indirect blocks (taken from the
273  *      inode->i_sb).
274  */
275
276 /*
277  * Portability note: the last comparison (check that we fit into triple
278  * indirect block) is spelled differently, because otherwise on an
279  * architecture with 32-bit longs and 8Kb pages we might get into trouble
280  * if our filesystem had 8Kb blocks. We might use long long, but that would
281  * kill us on x86. Oh, well, at least the sign propagation does not matter -
282  * i_block would have to be negative in the very beginning, so we would not
283  * get there at all.
284  */
285
286 static int ext3_block_to_path(struct inode *inode,
287                         long i_block, int offsets[4], int *boundary)
288 {
289         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
290         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
291         const long direct_blocks = EXT3_NDIR_BLOCKS,
292                 indirect_blocks = ptrs,
293                 double_blocks = (1 << (ptrs_bits * 2));
294         int n = 0;
295         int final = 0;
296
297         if (i_block < 0) {
298                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
299         } else if (i_block < direct_blocks) {
300                 offsets[n++] = i_block;
301                 final = direct_blocks;
302         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
303                 offsets[n++] = EXT3_IND_BLOCK;
304                 offsets[n++] = i_block;
305                 final = ptrs;
306         } else if ((i_block -= indirect_blocks) < double_blocks) {
307                 offsets[n++] = EXT3_DIND_BLOCK;
308                 offsets[n++] = i_block >> ptrs_bits;
309                 offsets[n++] = i_block & (ptrs - 1);
310                 final = ptrs;
311         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
312                 offsets[n++] = EXT3_TIND_BLOCK;
313                 offsets[n++] = i_block >> (ptrs_bits * 2);
314                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
315                 offsets[n++] = i_block & (ptrs - 1);
316                 final = ptrs;
317         } else {
318                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
319         }
320         if (boundary)
321                 *boundary = final - 1 - (i_block & (ptrs - 1));
322         return n;
323 }
324
325 /**
326  *      ext3_get_branch - read the chain of indirect blocks leading to data
327  *      @inode: inode in question
328  *      @depth: depth of the chain (1 - direct pointer, etc.)
329  *      @offsets: offsets of pointers in inode/indirect blocks
330  *      @chain: place to store the result
331  *      @err: here we store the error value
332  *
333  *      Function fills the array of triples <key, p, bh> and returns %NULL
334  *      if everything went OK or the pointer to the last filled triple
335  *      (incomplete one) otherwise. Upon the return chain[i].key contains
336  *      the number of (i+1)-th block in the chain (as it is stored in memory,
337  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
338  *      number (it points into struct inode for i==0 and into the bh->b_data
339  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
340  *      block for i>0 and NULL for i==0. In other words, it holds the block
341  *      numbers of the chain, addresses they were taken from (and where we can
342  *      verify that chain did not change) and buffer_heads hosting these
343  *      numbers.
344  *
345  *      Function stops when it stumbles upon zero pointer (absent block)
346  *              (pointer to last triple returned, *@err == 0)
347  *      or when it gets an IO error reading an indirect block
348  *              (ditto, *@err == -EIO)
349  *      or when it notices that chain had been changed while it was reading
350  *              (ditto, *@err == -EAGAIN)
351  *      or when it reads all @depth-1 indirect blocks successfully and finds
352  *      the whole chain, all way to the data (returns %NULL, *err == 0).
353  */
354 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
355                                  Indirect chain[4], int *err)
356 {
357         struct super_block *sb = inode->i_sb;
358         Indirect *p = chain;
359         struct buffer_head *bh;
360
361         *err = 0;
362         /* i_data is not going away, no lock needed */
363         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
364         if (!p->key)
365                 goto no_block;
366         while (--depth) {
367                 bh = sb_bread(sb, le32_to_cpu(p->key));
368                 if (!bh)
369                         goto failure;
370                 /* Reader: pointers */
371                 if (!verify_chain(chain, p))
372                         goto changed;
373                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
374                 /* Reader: end */
375                 if (!p->key)
376                         goto no_block;
377         }
378         return NULL;
379
380 changed:
381         brelse(bh);
382         *err = -EAGAIN;
383         goto no_block;
384 failure:
385         *err = -EIO;
386 no_block:
387         return p;
388 }
389
390 /**
391  *      ext3_find_near - find a place for allocation with sufficient locality
392  *      @inode: owner
393  *      @ind: descriptor of indirect block.
394  *
395  *      This function returns the prefered place for block allocation.
396  *      It is used when heuristic for sequential allocation fails.
397  *      Rules are:
398  *        + if there is a block to the left of our position - allocate near it.
399  *        + if pointer will live in indirect block - allocate near that block.
400  *        + if pointer will live in inode - allocate in the same
401  *          cylinder group.
402  *
403  * In the latter case we colour the starting block by the callers PID to
404  * prevent it from clashing with concurrent allocations for a different inode
405  * in the same block group.   The PID is used here so that functionally related
406  * files will be close-by on-disk.
407  *
408  *      Caller must make sure that @ind is valid and will stay that way.
409  */
410 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
411 {
412         struct ext3_inode_info *ei = EXT3_I(inode);
413         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
414         __le32 *p;
415         ext3_fsblk_t bg_start;
416         ext3_grpblk_t colour;
417
418         /* Try to find previous block */
419         for (p = ind->p - 1; p >= start; p--) {
420                 if (*p)
421                         return le32_to_cpu(*p);
422         }
423
424         /* No such thing, so let's try location of indirect block */
425         if (ind->bh)
426                 return ind->bh->b_blocknr;
427
428         /*
429          * It is going to be referred to from the inode itself? OK, just put it
430          * into the same cylinder group then.
431          */
432         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
433         colour = (current->pid % 16) *
434                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
435         return bg_start + colour;
436 }
437
438 /**
439  *      ext3_find_goal - find a prefered place for allocation.
440  *      @inode: owner
441  *      @block:  block we want
442  *      @chain:  chain of indirect blocks
443  *      @partial: pointer to the last triple within a chain
444  *      @goal:  place to store the result.
445  *
446  *      Normally this function find the prefered place for block allocation,
447  *      stores it in *@goal and returns zero.
448  */
449
450 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
451                 Indirect chain[4], Indirect *partial)
452 {
453         struct ext3_block_alloc_info *block_i;
454
455         block_i =  EXT3_I(inode)->i_block_alloc_info;
456
457         /*
458          * try the heuristic for sequential allocation,
459          * failing that at least try to get decent locality.
460          */
461         if (block_i && (block == block_i->last_alloc_logical_block + 1)
462                 && (block_i->last_alloc_physical_block != 0)) {
463                 return block_i->last_alloc_physical_block + 1;
464         }
465
466         return ext3_find_near(inode, partial);
467 }
468
469 /**
470  *      ext3_blks_to_allocate: Look up the block map and count the number
471  *      of direct blocks need to be allocated for the given branch.
472  *
473  *      @branch: chain of indirect blocks
474  *      @k: number of blocks need for indirect blocks
475  *      @blks: number of data blocks to be mapped.
476  *      @blocks_to_boundary:  the offset in the indirect block
477  *
478  *      return the total number of blocks to be allocate, including the
479  *      direct and indirect blocks.
480  */
481 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
482                 int blocks_to_boundary)
483 {
484         unsigned long count = 0;
485
486         /*
487          * Simple case, [t,d]Indirect block(s) has not allocated yet
488          * then it's clear blocks on that path have not allocated
489          */
490         if (k > 0) {
491                 /* right now we don't handle cross boundary allocation */
492                 if (blks < blocks_to_boundary + 1)
493                         count += blks;
494                 else
495                         count += blocks_to_boundary + 1;
496                 return count;
497         }
498
499         count++;
500         while (count < blks && count <= blocks_to_boundary &&
501                 le32_to_cpu(*(branch[0].p + count)) == 0) {
502                 count++;
503         }
504         return count;
505 }
506
507 /**
508  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
509  *      @indirect_blks: the number of blocks need to allocate for indirect
510  *                      blocks
511  *
512  *      @new_blocks: on return it will store the new block numbers for
513  *      the indirect blocks(if needed) and the first direct block,
514  *      @blks:  on return it will store the total number of allocated
515  *              direct blocks
516  */
517 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
518                         ext3_fsblk_t goal, int indirect_blks, int blks,
519                         ext3_fsblk_t new_blocks[4], int *err)
520 {
521         int target, i;
522         unsigned long count = 0;
523         int index = 0;
524         ext3_fsblk_t current_block = 0;
525         int ret = 0;
526
527         /*
528          * Here we try to allocate the requested multiple blocks at once,
529          * on a best-effort basis.
530          * To build a branch, we should allocate blocks for
531          * the indirect blocks(if not allocated yet), and at least
532          * the first direct block of this branch.  That's the
533          * minimum number of blocks need to allocate(required)
534          */
535         target = blks + indirect_blks;
536
537         while (1) {
538                 count = target;
539                 /* allocating blocks for indirect blocks and direct blocks */
540                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
541                 if (*err)
542                         goto failed_out;
543
544                 target -= count;
545                 /* allocate blocks for indirect blocks */
546                 while (index < indirect_blks && count) {
547                         new_blocks[index++] = current_block++;
548                         count--;
549                 }
550
551                 if (count > 0)
552                         break;
553         }
554
555         /* save the new block number for the first direct block */
556         new_blocks[index] = current_block;
557
558         /* total number of blocks allocated for direct blocks */
559         ret = count;
560         *err = 0;
561         return ret;
562 failed_out:
563         for (i = 0; i <index; i++)
564                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
565         return ret;
566 }
567
568 /**
569  *      ext3_alloc_branch - allocate and set up a chain of blocks.
570  *      @inode: owner
571  *      @indirect_blks: number of allocated indirect blocks
572  *      @blks: number of allocated direct blocks
573  *      @offsets: offsets (in the blocks) to store the pointers to next.
574  *      @branch: place to store the chain in.
575  *
576  *      This function allocates blocks, zeroes out all but the last one,
577  *      links them into chain and (if we are synchronous) writes them to disk.
578  *      In other words, it prepares a branch that can be spliced onto the
579  *      inode. It stores the information about that chain in the branch[], in
580  *      the same format as ext3_get_branch() would do. We are calling it after
581  *      we had read the existing part of chain and partial points to the last
582  *      triple of that (one with zero ->key). Upon the exit we have the same
583  *      picture as after the successful ext3_get_block(), except that in one
584  *      place chain is disconnected - *branch->p is still zero (we did not
585  *      set the last link), but branch->key contains the number that should
586  *      be placed into *branch->p to fill that gap.
587  *
588  *      If allocation fails we free all blocks we've allocated (and forget
589  *      their buffer_heads) and return the error value the from failed
590  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
591  *      as described above and return 0.
592  */
593 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
594                         int indirect_blks, int *blks, ext3_fsblk_t goal,
595                         int *offsets, Indirect *branch)
596 {
597         int blocksize = inode->i_sb->s_blocksize;
598         int i, n = 0;
599         int err = 0;
600         struct buffer_head *bh;
601         int num;
602         ext3_fsblk_t new_blocks[4];
603         ext3_fsblk_t current_block;
604
605         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
606                                 *blks, new_blocks, &err);
607         if (err)
608                 return err;
609
610         branch[0].key = cpu_to_le32(new_blocks[0]);
611         /*
612          * metadata blocks and data blocks are allocated.
613          */
614         for (n = 1; n <= indirect_blks;  n++) {
615                 /*
616                  * Get buffer_head for parent block, zero it out
617                  * and set the pointer to new one, then send
618                  * parent to disk.
619                  */
620                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
621                 branch[n].bh = bh;
622                 lock_buffer(bh);
623                 BUFFER_TRACE(bh, "call get_create_access");
624                 err = ext3_journal_get_create_access(handle, bh);
625                 if (err) {
626                         unlock_buffer(bh);
627                         brelse(bh);
628                         goto failed;
629                 }
630
631                 memset(bh->b_data, 0, blocksize);
632                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
633                 branch[n].key = cpu_to_le32(new_blocks[n]);
634                 *branch[n].p = branch[n].key;
635                 if ( n == indirect_blks) {
636                         current_block = new_blocks[n];
637                         /*
638                          * End of chain, update the last new metablock of
639                          * the chain to point to the new allocated
640                          * data blocks numbers
641                          */
642                         for (i=1; i < num; i++)
643                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
644                 }
645                 BUFFER_TRACE(bh, "marking uptodate");
646                 set_buffer_uptodate(bh);
647                 unlock_buffer(bh);
648
649                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
650                 err = ext3_journal_dirty_metadata(handle, bh);
651                 if (err)
652                         goto failed;
653         }
654         *blks = num;
655         return err;
656 failed:
657         /* Allocation failed, free what we already allocated */
658         for (i = 1; i <= n ; i++) {
659                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
660                 ext3_journal_forget(handle, branch[i].bh);
661         }
662         for (i = 0; i <indirect_blks; i++)
663                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
664
665         ext3_free_blocks(handle, inode, new_blocks[i], num);
666
667         return err;
668 }
669
670 /**
671  * ext3_splice_branch - splice the allocated branch onto inode.
672  * @inode: owner
673  * @block: (logical) number of block we are adding
674  * @chain: chain of indirect blocks (with a missing link - see
675  *      ext3_alloc_branch)
676  * @where: location of missing link
677  * @num:   number of indirect blocks we are adding
678  * @blks:  number of direct blocks we are adding
679  *
680  * This function fills the missing link and does all housekeeping needed in
681  * inode (->i_blocks, etc.). In case of success we end up with the full
682  * chain to new block and return 0.
683  */
684 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
685                         long block, Indirect *where, int num, int blks)
686 {
687         int i;
688         int err = 0;
689         struct ext3_block_alloc_info *block_i;
690         ext3_fsblk_t current_block;
691
692         block_i = EXT3_I(inode)->i_block_alloc_info;
693         /*
694          * If we're splicing into a [td]indirect block (as opposed to the
695          * inode) then we need to get write access to the [td]indirect block
696          * before the splice.
697          */
698         if (where->bh) {
699                 BUFFER_TRACE(where->bh, "get_write_access");
700                 err = ext3_journal_get_write_access(handle, where->bh);
701                 if (err)
702                         goto err_out;
703         }
704         /* That's it */
705
706         *where->p = where->key;
707
708         /*
709          * Update the host buffer_head or inode to point to more just allocated
710          * direct blocks blocks
711          */
712         if (num == 0 && blks > 1) {
713                 current_block = le32_to_cpu(where->key) + 1;
714                 for (i = 1; i < blks; i++)
715                         *(where->p + i ) = cpu_to_le32(current_block++);
716         }
717
718         /*
719          * update the most recently allocated logical & physical block
720          * in i_block_alloc_info, to assist find the proper goal block for next
721          * allocation
722          */
723         if (block_i) {
724                 block_i->last_alloc_logical_block = block + blks - 1;
725                 block_i->last_alloc_physical_block =
726                                 le32_to_cpu(where[num].key) + blks - 1;
727         }
728
729         /* We are done with atomic stuff, now do the rest of housekeeping */
730
731         inode->i_ctime = CURRENT_TIME_SEC;
732         ext3_mark_inode_dirty(handle, inode);
733
734         /* had we spliced it onto indirect block? */
735         if (where->bh) {
736                 /*
737                  * If we spliced it onto an indirect block, we haven't
738                  * altered the inode.  Note however that if it is being spliced
739                  * onto an indirect block at the very end of the file (the
740                  * file is growing) then we *will* alter the inode to reflect
741                  * the new i_size.  But that is not done here - it is done in
742                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
743                  */
744                 jbd_debug(5, "splicing indirect only\n");
745                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
746                 err = ext3_journal_dirty_metadata(handle, where->bh);
747                 if (err)
748                         goto err_out;
749         } else {
750                 /*
751                  * OK, we spliced it into the inode itself on a direct block.
752                  * Inode was dirtied above.
753                  */
754                 jbd_debug(5, "splicing direct\n");
755         }
756         return err;
757
758 err_out:
759         for (i = 1; i <= num; i++) {
760                 BUFFER_TRACE(where[i].bh, "call journal_forget");
761                 ext3_journal_forget(handle, where[i].bh);
762                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
763         }
764         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
765
766         return err;
767 }
768
769 /*
770  * Allocation strategy is simple: if we have to allocate something, we will
771  * have to go the whole way to leaf. So let's do it before attaching anything
772  * to tree, set linkage between the newborn blocks, write them if sync is
773  * required, recheck the path, free and repeat if check fails, otherwise
774  * set the last missing link (that will protect us from any truncate-generated
775  * removals - all blocks on the path are immune now) and possibly force the
776  * write on the parent block.
777  * That has a nice additional property: no special recovery from the failed
778  * allocations is needed - we simply release blocks and do not touch anything
779  * reachable from inode.
780  *
781  * `handle' can be NULL if create == 0.
782  *
783  * The BKL may not be held on entry here.  Be sure to take it early.
784  * return > 0, # of blocks mapped or allocated.
785  * return = 0, if plain lookup failed.
786  * return < 0, error case.
787  */
788 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
789                 sector_t iblock, unsigned long maxblocks,
790                 struct buffer_head *bh_result,
791                 int create, int extend_disksize)
792 {
793         int err = -EIO;
794         int offsets[4];
795         Indirect chain[4];
796         Indirect *partial;
797         ext3_fsblk_t goal;
798         int indirect_blks;
799         int blocks_to_boundary = 0;
800         int depth;
801         struct ext3_inode_info *ei = EXT3_I(inode);
802         int count = 0;
803         ext3_fsblk_t first_block = 0;
804
805
806         J_ASSERT(handle != NULL || create == 0);
807         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
808
809         if (depth == 0)
810                 goto out;
811
812         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
813
814         /* Simplest case - block found, no allocation needed */
815         if (!partial) {
816                 first_block = le32_to_cpu(chain[depth - 1].key);
817                 clear_buffer_new(bh_result);
818                 count++;
819                 /*map more blocks*/
820                 while (count < maxblocks && count <= blocks_to_boundary) {
821                         ext3_fsblk_t blk;
822
823                         if (!verify_chain(chain, partial)) {
824                                 /*
825                                  * Indirect block might be removed by
826                                  * truncate while we were reading it.
827                                  * Handling of that case: forget what we've
828                                  * got now. Flag the err as EAGAIN, so it
829                                  * will reread.
830                                  */
831                                 err = -EAGAIN;
832                                 count = 0;
833                                 break;
834                         }
835                         blk = le32_to_cpu(*(chain[depth-1].p + count));
836
837                         if (blk == first_block + count)
838                                 count++;
839                         else
840                                 break;
841                 }
842                 if (err != -EAGAIN)
843                         goto got_it;
844         }
845
846         /* Next simple case - plain lookup or failed read of indirect block */
847         if (!create || err == -EIO)
848                 goto cleanup;
849
850         mutex_lock(&ei->truncate_mutex);
851
852         /*
853          * If the indirect block is missing while we are reading
854          * the chain(ext3_get_branch() returns -EAGAIN err), or
855          * if the chain has been changed after we grab the semaphore,
856          * (either because another process truncated this branch, or
857          * another get_block allocated this branch) re-grab the chain to see if
858          * the request block has been allocated or not.
859          *
860          * Since we already block the truncate/other get_block
861          * at this point, we will have the current copy of the chain when we
862          * splice the branch into the tree.
863          */
864         if (err == -EAGAIN || !verify_chain(chain, partial)) {
865                 while (partial > chain) {
866                         brelse(partial->bh);
867                         partial--;
868                 }
869                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
870                 if (!partial) {
871                         count++;
872                         mutex_unlock(&ei->truncate_mutex);
873                         if (err)
874                                 goto cleanup;
875                         clear_buffer_new(bh_result);
876                         goto got_it;
877                 }
878         }
879
880         /*
881          * Okay, we need to do block allocation.  Lazily initialize the block
882          * allocation info here if necessary
883         */
884         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
885                 ext3_init_block_alloc_info(inode);
886
887         goal = ext3_find_goal(inode, iblock, chain, partial);
888
889         /* the number of blocks need to allocate for [d,t]indirect blocks */
890         indirect_blks = (chain + depth) - partial - 1;
891
892         /*
893          * Next look up the indirect map to count the totoal number of
894          * direct blocks to allocate for this branch.
895          */
896         count = ext3_blks_to_allocate(partial, indirect_blks,
897                                         maxblocks, blocks_to_boundary);
898         /*
899          * Block out ext3_truncate while we alter the tree
900          */
901         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
902                                 offsets + (partial - chain), partial);
903
904         /*
905          * The ext3_splice_branch call will free and forget any buffers
906          * on the new chain if there is a failure, but that risks using
907          * up transaction credits, especially for bitmaps where the
908          * credits cannot be returned.  Can we handle this somehow?  We
909          * may need to return -EAGAIN upwards in the worst case.  --sct
910          */
911         if (!err)
912                 err = ext3_splice_branch(handle, inode, iblock,
913                                         partial, indirect_blks, count);
914         /*
915          * i_disksize growing is protected by truncate_mutex.  Don't forget to
916          * protect it if you're about to implement concurrent
917          * ext3_get_block() -bzzz
918         */
919         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
920                 ei->i_disksize = inode->i_size;
921         mutex_unlock(&ei->truncate_mutex);
922         if (err)
923                 goto cleanup;
924
925         set_buffer_new(bh_result);
926 got_it:
927         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
928         if (count > blocks_to_boundary)
929                 set_buffer_boundary(bh_result);
930         err = count;
931         /* Clean up and exit */
932         partial = chain + depth - 1;    /* the whole chain */
933 cleanup:
934         while (partial > chain) {
935                 BUFFER_TRACE(partial->bh, "call brelse");
936                 brelse(partial->bh);
937                 partial--;
938         }
939         BUFFER_TRACE(bh_result, "returned");
940 out:
941         return err;
942 }
943
944 #define DIO_CREDITS (EXT3_RESERVE_TRANS_BLOCKS + 32)
945
946 static int ext3_get_block(struct inode *inode, sector_t iblock,
947                         struct buffer_head *bh_result, int create)
948 {
949         handle_t *handle = ext3_journal_current_handle();
950         int ret = 0;
951         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
952
953         if (!create)
954                 goto get_block;         /* A read */
955
956         if (max_blocks == 1)
957                 goto get_block;         /* A single block get */
958
959         if (handle->h_transaction->t_state == T_LOCKED) {
960                 /*
961                  * Huge direct-io writes can hold off commits for long
962                  * periods of time.  Let this commit run.
963                  */
964                 ext3_journal_stop(handle);
965                 handle = ext3_journal_start(inode, DIO_CREDITS);
966                 if (IS_ERR(handle))
967                         ret = PTR_ERR(handle);
968                 goto get_block;
969         }
970
971         if (handle->h_buffer_credits <= EXT3_RESERVE_TRANS_BLOCKS) {
972                 /*
973                  * Getting low on buffer credits...
974                  */
975                 ret = ext3_journal_extend(handle, DIO_CREDITS);
976                 if (ret > 0) {
977                         /*
978                          * Couldn't extend the transaction.  Start a new one.
979                          */
980                         ret = ext3_journal_restart(handle, DIO_CREDITS);
981                 }
982         }
983
984 get_block:
985         if (ret == 0) {
986                 ret = ext3_get_blocks_handle(handle, inode, iblock,
987                                         max_blocks, bh_result, create, 0);
988                 if (ret > 0) {
989                         bh_result->b_size = (ret << inode->i_blkbits);
990                         ret = 0;
991                 }
992         }
993         return ret;
994 }
995
996 /*
997  * `handle' can be NULL if create is zero
998  */
999 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1000                                 long block, int create, int *errp)
1001 {
1002         struct buffer_head dummy;
1003         int fatal = 0, err;
1004
1005         J_ASSERT(handle != NULL || create == 0);
1006
1007         dummy.b_state = 0;
1008         dummy.b_blocknr = -1000;
1009         buffer_trace_init(&dummy.b_history);
1010         err = ext3_get_blocks_handle(handle, inode, block, 1,
1011                                         &dummy, create, 1);
1012         /*
1013          * ext3_get_blocks_handle() returns number of blocks
1014          * mapped. 0 in case of a HOLE.
1015          */
1016         if (err > 0) {
1017                 if (err > 1)
1018                         WARN_ON(1);
1019                 err = 0;
1020         }
1021         *errp = err;
1022         if (!err && buffer_mapped(&dummy)) {
1023                 struct buffer_head *bh;
1024                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1025                 if (!bh) {
1026                         *errp = -EIO;
1027                         goto err;
1028                 }
1029                 if (buffer_new(&dummy)) {
1030                         J_ASSERT(create != 0);
1031                         J_ASSERT(handle != NULL);
1032
1033                         /*
1034                          * Now that we do not always journal data, we should
1035                          * keep in mind whether this should always journal the
1036                          * new buffer as metadata.  For now, regular file
1037                          * writes use ext3_get_block instead, so it's not a
1038                          * problem.
1039                          */
1040                         lock_buffer(bh);
1041                         BUFFER_TRACE(bh, "call get_create_access");
1042                         fatal = ext3_journal_get_create_access(handle, bh);
1043                         if (!fatal && !buffer_uptodate(bh)) {
1044                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1045                                 set_buffer_uptodate(bh);
1046                         }
1047                         unlock_buffer(bh);
1048                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1049                         err = ext3_journal_dirty_metadata(handle, bh);
1050                         if (!fatal)
1051                                 fatal = err;
1052                 } else {
1053                         BUFFER_TRACE(bh, "not a new buffer");
1054                 }
1055                 if (fatal) {
1056                         *errp = fatal;
1057                         brelse(bh);
1058                         bh = NULL;
1059                 }
1060                 return bh;
1061         }
1062 err:
1063         return NULL;
1064 }
1065
1066 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1067                                int block, int create, int *err)
1068 {
1069         struct buffer_head * bh;
1070
1071         bh = ext3_getblk(handle, inode, block, create, err);
1072         if (!bh)
1073                 return bh;
1074         if (buffer_uptodate(bh))
1075                 return bh;
1076         ll_rw_block(READ_META, 1, &bh);
1077         wait_on_buffer(bh);
1078         if (buffer_uptodate(bh))
1079                 return bh;
1080         put_bh(bh);
1081         *err = -EIO;
1082         return NULL;
1083 }
1084
1085 static int walk_page_buffers(   handle_t *handle,
1086                                 struct buffer_head *head,
1087                                 unsigned from,
1088                                 unsigned to,
1089                                 int *partial,
1090                                 int (*fn)(      handle_t *handle,
1091                                                 struct buffer_head *bh))
1092 {
1093         struct buffer_head *bh;
1094         unsigned block_start, block_end;
1095         unsigned blocksize = head->b_size;
1096         int err, ret = 0;
1097         struct buffer_head *next;
1098
1099         for (   bh = head, block_start = 0;
1100                 ret == 0 && (bh != head || !block_start);
1101                 block_start = block_end, bh = next)
1102         {
1103                 next = bh->b_this_page;
1104                 block_end = block_start + blocksize;
1105                 if (block_end <= from || block_start >= to) {
1106                         if (partial && !buffer_uptodate(bh))
1107                                 *partial = 1;
1108                         continue;
1109                 }
1110                 err = (*fn)(handle, bh);
1111                 if (!ret)
1112                         ret = err;
1113         }
1114         return ret;
1115 }
1116
1117 /*
1118  * To preserve ordering, it is essential that the hole instantiation and
1119  * the data write be encapsulated in a single transaction.  We cannot
1120  * close off a transaction and start a new one between the ext3_get_block()
1121  * and the commit_write().  So doing the journal_start at the start of
1122  * prepare_write() is the right place.
1123  *
1124  * Also, this function can nest inside ext3_writepage() ->
1125  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1126  * has generated enough buffer credits to do the whole page.  So we won't
1127  * block on the journal in that case, which is good, because the caller may
1128  * be PF_MEMALLOC.
1129  *
1130  * By accident, ext3 can be reentered when a transaction is open via
1131  * quota file writes.  If we were to commit the transaction while thus
1132  * reentered, there can be a deadlock - we would be holding a quota
1133  * lock, and the commit would never complete if another thread had a
1134  * transaction open and was blocking on the quota lock - a ranking
1135  * violation.
1136  *
1137  * So what we do is to rely on the fact that journal_stop/journal_start
1138  * will _not_ run commit under these circumstances because handle->h_ref
1139  * is elevated.  We'll still have enough credits for the tiny quotafile
1140  * write.
1141  */
1142 static int do_journal_get_write_access(handle_t *handle,
1143                                         struct buffer_head *bh)
1144 {
1145         if (!buffer_mapped(bh) || buffer_freed(bh))
1146                 return 0;
1147         return ext3_journal_get_write_access(handle, bh);
1148 }
1149
1150 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1151                                 loff_t pos, unsigned len, unsigned flags,
1152                                 struct page **pagep, void **fsdata)
1153 {
1154         struct inode *inode = mapping->host;
1155         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1156         handle_t *handle;
1157         int retries = 0;
1158         struct page *page;
1159         pgoff_t index;
1160         unsigned from, to;
1161
1162         index = pos >> PAGE_CACHE_SHIFT;
1163         from = pos & (PAGE_CACHE_SIZE - 1);
1164         to = from + len;
1165
1166 retry:
1167         page = __grab_cache_page(mapping, index);
1168         if (!page)
1169                 return -ENOMEM;
1170         *pagep = page;
1171
1172         handle = ext3_journal_start(inode, needed_blocks);
1173         if (IS_ERR(handle)) {
1174                 unlock_page(page);
1175                 page_cache_release(page);
1176                 ret = PTR_ERR(handle);
1177                 goto out;
1178         }
1179         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1180                                                         ext3_get_block);
1181         if (ret)
1182                 goto write_begin_failed;
1183
1184         if (ext3_should_journal_data(inode)) {
1185                 ret = walk_page_buffers(handle, page_buffers(page),
1186                                 from, to, NULL, do_journal_get_write_access);
1187         }
1188 write_begin_failed:
1189         if (ret) {
1190                 ext3_journal_stop(handle);
1191                 unlock_page(page);
1192                 page_cache_release(page);
1193         }
1194         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1195                 goto retry;
1196 out:
1197         return ret;
1198 }
1199
1200
1201 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1202 {
1203         int err = journal_dirty_data(handle, bh);
1204         if (err)
1205                 ext3_journal_abort_handle(__FUNCTION__, __FUNCTION__,
1206                                                 bh, handle, err);
1207         return err;
1208 }
1209
1210 /* For write_end() in data=journal mode */
1211 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1212 {
1213         if (!buffer_mapped(bh) || buffer_freed(bh))
1214                 return 0;
1215         set_buffer_uptodate(bh);
1216         return ext3_journal_dirty_metadata(handle, bh);
1217 }
1218
1219 /*
1220  * Generic write_end handler for ordered and writeback ext3 journal modes.
1221  * We can't use generic_write_end, because that unlocks the page and we need to
1222  * unlock the page after ext3_journal_stop, but ext3_journal_stop must run
1223  * after block_write_end.
1224  */
1225 static int ext3_generic_write_end(struct file *file,
1226                                 struct address_space *mapping,
1227                                 loff_t pos, unsigned len, unsigned copied,
1228                                 struct page *page, void *fsdata)
1229 {
1230         struct inode *inode = file->f_mapping->host;
1231
1232         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1233
1234         if (pos+copied > inode->i_size) {
1235                 i_size_write(inode, pos+copied);
1236                 mark_inode_dirty(inode);
1237         }
1238
1239         return copied;
1240 }
1241
1242 /*
1243  * We need to pick up the new inode size which generic_commit_write gave us
1244  * `file' can be NULL - eg, when called from page_symlink().
1245  *
1246  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1247  * buffers are managed internally.
1248  */
1249 static int ext3_ordered_write_end(struct file *file,
1250                                 struct address_space *mapping,
1251                                 loff_t pos, unsigned len, unsigned copied,
1252                                 struct page *page, void *fsdata)
1253 {
1254         handle_t *handle = ext3_journal_current_handle();
1255         struct inode *inode = file->f_mapping->host;
1256         unsigned from, to;
1257         int ret = 0, ret2;
1258
1259         from = pos & (PAGE_CACHE_SIZE - 1);
1260         to = from + len;
1261
1262         ret = walk_page_buffers(handle, page_buffers(page),
1263                 from, to, NULL, ext3_journal_dirty_data);
1264
1265         if (ret == 0) {
1266                 /*
1267                  * generic_write_end() will run mark_inode_dirty() if i_size
1268                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1269                  * into that.
1270                  */
1271                 loff_t new_i_size;
1272
1273                 new_i_size = pos + copied;
1274                 if (new_i_size > EXT3_I(inode)->i_disksize)
1275                         EXT3_I(inode)->i_disksize = new_i_size;
1276                 copied = ext3_generic_write_end(file, mapping, pos, len, copied,
1277                                                         page, fsdata);
1278                 if (copied < 0)
1279                         ret = copied;
1280         }
1281         ret2 = ext3_journal_stop(handle);
1282         if (!ret)
1283                 ret = ret2;
1284         unlock_page(page);
1285         page_cache_release(page);
1286
1287         return ret ? ret : copied;
1288 }
1289
1290 static int ext3_writeback_write_end(struct file *file,
1291                                 struct address_space *mapping,
1292                                 loff_t pos, unsigned len, unsigned copied,
1293                                 struct page *page, void *fsdata)
1294 {
1295         handle_t *handle = ext3_journal_current_handle();
1296         struct inode *inode = file->f_mapping->host;
1297         int ret = 0, ret2;
1298         loff_t new_i_size;
1299
1300         new_i_size = pos + copied;
1301         if (new_i_size > EXT3_I(inode)->i_disksize)
1302                 EXT3_I(inode)->i_disksize = new_i_size;
1303
1304         copied = ext3_generic_write_end(file, mapping, pos, len, copied,
1305                                                         page, fsdata);
1306         if (copied < 0)
1307                 ret = copied;
1308
1309         ret2 = ext3_journal_stop(handle);
1310         if (!ret)
1311                 ret = ret2;
1312         unlock_page(page);
1313         page_cache_release(page);
1314
1315         return ret ? ret : copied;
1316 }
1317
1318 static int ext3_journalled_write_end(struct file *file,
1319                                 struct address_space *mapping,
1320                                 loff_t pos, unsigned len, unsigned copied,
1321                                 struct page *page, void *fsdata)
1322 {
1323         handle_t *handle = ext3_journal_current_handle();
1324         struct inode *inode = mapping->host;
1325         int ret = 0, ret2;
1326         int partial = 0;
1327         unsigned from, to;
1328
1329         from = pos & (PAGE_CACHE_SIZE - 1);
1330         to = from + len;
1331
1332         if (copied < len) {
1333                 if (!PageUptodate(page))
1334                         copied = 0;
1335                 page_zero_new_buffers(page, from+copied, to);
1336         }
1337
1338         ret = walk_page_buffers(handle, page_buffers(page), from,
1339                                 to, &partial, write_end_fn);
1340         if (!partial)
1341                 SetPageUptodate(page);
1342         if (pos+copied > inode->i_size)
1343                 i_size_write(inode, pos+copied);
1344         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1345         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1346                 EXT3_I(inode)->i_disksize = inode->i_size;
1347                 ret2 = ext3_mark_inode_dirty(handle, inode);
1348                 if (!ret)
1349                         ret = ret2;
1350         }
1351
1352         ret2 = ext3_journal_stop(handle);
1353         if (!ret)
1354                 ret = ret2;
1355         unlock_page(page);
1356         page_cache_release(page);
1357
1358         return ret ? ret : copied;
1359 }
1360
1361 /*
1362  * bmap() is special.  It gets used by applications such as lilo and by
1363  * the swapper to find the on-disk block of a specific piece of data.
1364  *
1365  * Naturally, this is dangerous if the block concerned is still in the
1366  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1367  * filesystem and enables swap, then they may get a nasty shock when the
1368  * data getting swapped to that swapfile suddenly gets overwritten by
1369  * the original zero's written out previously to the journal and
1370  * awaiting writeback in the kernel's buffer cache.
1371  *
1372  * So, if we see any bmap calls here on a modified, data-journaled file,
1373  * take extra steps to flush any blocks which might be in the cache.
1374  */
1375 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1376 {
1377         struct inode *inode = mapping->host;
1378         journal_t *journal;
1379         int err;
1380
1381         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1382                 /*
1383                  * This is a REALLY heavyweight approach, but the use of
1384                  * bmap on dirty files is expected to be extremely rare:
1385                  * only if we run lilo or swapon on a freshly made file
1386                  * do we expect this to happen.
1387                  *
1388                  * (bmap requires CAP_SYS_RAWIO so this does not
1389                  * represent an unprivileged user DOS attack --- we'd be
1390                  * in trouble if mortal users could trigger this path at
1391                  * will.)
1392                  *
1393                  * NB. EXT3_STATE_JDATA is not set on files other than
1394                  * regular files.  If somebody wants to bmap a directory
1395                  * or symlink and gets confused because the buffer
1396                  * hasn't yet been flushed to disk, they deserve
1397                  * everything they get.
1398                  */
1399
1400                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1401                 journal = EXT3_JOURNAL(inode);
1402                 journal_lock_updates(journal);
1403                 err = journal_flush(journal);
1404                 journal_unlock_updates(journal);
1405
1406                 if (err)
1407                         return 0;
1408         }
1409
1410         return generic_block_bmap(mapping,block,ext3_get_block);
1411 }
1412
1413 static int bget_one(handle_t *handle, struct buffer_head *bh)
1414 {
1415         get_bh(bh);
1416         return 0;
1417 }
1418
1419 static int bput_one(handle_t *handle, struct buffer_head *bh)
1420 {
1421         put_bh(bh);
1422         return 0;
1423 }
1424
1425 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1426 {
1427         if (buffer_mapped(bh))
1428                 return ext3_journal_dirty_data(handle, bh);
1429         return 0;
1430 }
1431
1432 /*
1433  * Note that we always start a transaction even if we're not journalling
1434  * data.  This is to preserve ordering: any hole instantiation within
1435  * __block_write_full_page -> ext3_get_block() should be journalled
1436  * along with the data so we don't crash and then get metadata which
1437  * refers to old data.
1438  *
1439  * In all journalling modes block_write_full_page() will start the I/O.
1440  *
1441  * Problem:
1442  *
1443  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1444  *              ext3_writepage()
1445  *
1446  * Similar for:
1447  *
1448  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1449  *
1450  * Same applies to ext3_get_block().  We will deadlock on various things like
1451  * lock_journal and i_truncate_mutex.
1452  *
1453  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1454  * allocations fail.
1455  *
1456  * 16May01: If we're reentered then journal_current_handle() will be
1457  *          non-zero. We simply *return*.
1458  *
1459  * 1 July 2001: @@@ FIXME:
1460  *   In journalled data mode, a data buffer may be metadata against the
1461  *   current transaction.  But the same file is part of a shared mapping
1462  *   and someone does a writepage() on it.
1463  *
1464  *   We will move the buffer onto the async_data list, but *after* it has
1465  *   been dirtied. So there's a small window where we have dirty data on
1466  *   BJ_Metadata.
1467  *
1468  *   Note that this only applies to the last partial page in the file.  The
1469  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1470  *   broken code anyway: it's wrong for msync()).
1471  *
1472  *   It's a rare case: affects the final partial page, for journalled data
1473  *   where the file is subject to bith write() and writepage() in the same
1474  *   transction.  To fix it we'll need a custom block_write_full_page().
1475  *   We'll probably need that anyway for journalling writepage() output.
1476  *
1477  * We don't honour synchronous mounts for writepage().  That would be
1478  * disastrous.  Any write() or metadata operation will sync the fs for
1479  * us.
1480  *
1481  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1482  * we don't need to open a transaction here.
1483  */
1484 static int ext3_ordered_writepage(struct page *page,
1485                                 struct writeback_control *wbc)
1486 {
1487         struct inode *inode = page->mapping->host;
1488         struct buffer_head *page_bufs;
1489         handle_t *handle = NULL;
1490         int ret = 0;
1491         int err;
1492
1493         J_ASSERT(PageLocked(page));
1494
1495         /*
1496          * We give up here if we're reentered, because it might be for a
1497          * different filesystem.
1498          */
1499         if (ext3_journal_current_handle())
1500                 goto out_fail;
1501
1502         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1503
1504         if (IS_ERR(handle)) {
1505                 ret = PTR_ERR(handle);
1506                 goto out_fail;
1507         }
1508
1509         if (!page_has_buffers(page)) {
1510                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1511                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1512         }
1513         page_bufs = page_buffers(page);
1514         walk_page_buffers(handle, page_bufs, 0,
1515                         PAGE_CACHE_SIZE, NULL, bget_one);
1516
1517         ret = block_write_full_page(page, ext3_get_block, wbc);
1518
1519         /*
1520          * The page can become unlocked at any point now, and
1521          * truncate can then come in and change things.  So we
1522          * can't touch *page from now on.  But *page_bufs is
1523          * safe due to elevated refcount.
1524          */
1525
1526         /*
1527          * And attach them to the current transaction.  But only if
1528          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1529          * and generally junk.
1530          */
1531         if (ret == 0) {
1532                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1533                                         NULL, journal_dirty_data_fn);
1534                 if (!ret)
1535                         ret = err;
1536         }
1537         walk_page_buffers(handle, page_bufs, 0,
1538                         PAGE_CACHE_SIZE, NULL, bput_one);
1539         err = ext3_journal_stop(handle);
1540         if (!ret)
1541                 ret = err;
1542         return ret;
1543
1544 out_fail:
1545         redirty_page_for_writepage(wbc, page);
1546         unlock_page(page);
1547         return ret;
1548 }
1549
1550 static int ext3_writeback_writepage(struct page *page,
1551                                 struct writeback_control *wbc)
1552 {
1553         struct inode *inode = page->mapping->host;
1554         handle_t *handle = NULL;
1555         int ret = 0;
1556         int err;
1557
1558         if (ext3_journal_current_handle())
1559                 goto out_fail;
1560
1561         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1562         if (IS_ERR(handle)) {
1563                 ret = PTR_ERR(handle);
1564                 goto out_fail;
1565         }
1566
1567         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1568                 ret = nobh_writepage(page, ext3_get_block, wbc);
1569         else
1570                 ret = block_write_full_page(page, ext3_get_block, wbc);
1571
1572         err = ext3_journal_stop(handle);
1573         if (!ret)
1574                 ret = err;
1575         return ret;
1576
1577 out_fail:
1578         redirty_page_for_writepage(wbc, page);
1579         unlock_page(page);
1580         return ret;
1581 }
1582
1583 static int ext3_journalled_writepage(struct page *page,
1584                                 struct writeback_control *wbc)
1585 {
1586         struct inode *inode = page->mapping->host;
1587         handle_t *handle = NULL;
1588         int ret = 0;
1589         int err;
1590
1591         if (ext3_journal_current_handle())
1592                 goto no_write;
1593
1594         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1595         if (IS_ERR(handle)) {
1596                 ret = PTR_ERR(handle);
1597                 goto no_write;
1598         }
1599
1600         if (!page_has_buffers(page) || PageChecked(page)) {
1601                 /*
1602                  * It's mmapped pagecache.  Add buffers and journal it.  There
1603                  * doesn't seem much point in redirtying the page here.
1604                  */
1605                 ClearPageChecked(page);
1606                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1607                                         ext3_get_block);
1608                 if (ret != 0) {
1609                         ext3_journal_stop(handle);
1610                         goto out_unlock;
1611                 }
1612                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1613                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1614
1615                 err = walk_page_buffers(handle, page_buffers(page), 0,
1616                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1617                 if (ret == 0)
1618                         ret = err;
1619                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1620                 unlock_page(page);
1621         } else {
1622                 /*
1623                  * It may be a page full of checkpoint-mode buffers.  We don't
1624                  * really know unless we go poke around in the buffer_heads.
1625                  * But block_write_full_page will do the right thing.
1626                  */
1627                 ret = block_write_full_page(page, ext3_get_block, wbc);
1628         }
1629         err = ext3_journal_stop(handle);
1630         if (!ret)
1631                 ret = err;
1632 out:
1633         return ret;
1634
1635 no_write:
1636         redirty_page_for_writepage(wbc, page);
1637 out_unlock:
1638         unlock_page(page);
1639         goto out;
1640 }
1641
1642 static int ext3_readpage(struct file *file, struct page *page)
1643 {
1644         return mpage_readpage(page, ext3_get_block);
1645 }
1646
1647 static int
1648 ext3_readpages(struct file *file, struct address_space *mapping,
1649                 struct list_head *pages, unsigned nr_pages)
1650 {
1651         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1652 }
1653
1654 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1655 {
1656         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1657
1658         /*
1659          * If it's a full truncate we just forget about the pending dirtying
1660          */
1661         if (offset == 0)
1662                 ClearPageChecked(page);
1663
1664         journal_invalidatepage(journal, page, offset);
1665 }
1666
1667 static int ext3_releasepage(struct page *page, gfp_t wait)
1668 {
1669         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1670
1671         WARN_ON(PageChecked(page));
1672         if (!page_has_buffers(page))
1673                 return 0;
1674         return journal_try_to_free_buffers(journal, page, wait);
1675 }
1676
1677 /*
1678  * If the O_DIRECT write will extend the file then add this inode to the
1679  * orphan list.  So recovery will truncate it back to the original size
1680  * if the machine crashes during the write.
1681  *
1682  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1683  * crashes then stale disk data _may_ be exposed inside the file.
1684  */
1685 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1686                         const struct iovec *iov, loff_t offset,
1687                         unsigned long nr_segs)
1688 {
1689         struct file *file = iocb->ki_filp;
1690         struct inode *inode = file->f_mapping->host;
1691         struct ext3_inode_info *ei = EXT3_I(inode);
1692         handle_t *handle = NULL;
1693         ssize_t ret;
1694         int orphan = 0;
1695         size_t count = iov_length(iov, nr_segs);
1696
1697         if (rw == WRITE) {
1698                 loff_t final_size = offset + count;
1699
1700                 handle = ext3_journal_start(inode, DIO_CREDITS);
1701                 if (IS_ERR(handle)) {
1702                         ret = PTR_ERR(handle);
1703                         goto out;
1704                 }
1705                 if (final_size > inode->i_size) {
1706                         ret = ext3_orphan_add(handle, inode);
1707                         if (ret)
1708                                 goto out_stop;
1709                         orphan = 1;
1710                         ei->i_disksize = inode->i_size;
1711                 }
1712         }
1713
1714         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1715                                  offset, nr_segs,
1716                                  ext3_get_block, NULL);
1717
1718         /*
1719          * Reacquire the handle: ext3_get_block() can restart the transaction
1720          */
1721         handle = ext3_journal_current_handle();
1722
1723 out_stop:
1724         if (handle) {
1725                 int err;
1726
1727                 if (orphan && inode->i_nlink)
1728                         ext3_orphan_del(handle, inode);
1729                 if (orphan && ret > 0) {
1730                         loff_t end = offset + ret;
1731                         if (end > inode->i_size) {
1732                                 ei->i_disksize = end;
1733                                 i_size_write(inode, end);
1734                                 /*
1735                                  * We're going to return a positive `ret'
1736                                  * here due to non-zero-length I/O, so there's
1737                                  * no way of reporting error returns from
1738                                  * ext3_mark_inode_dirty() to userspace.  So
1739                                  * ignore it.
1740                                  */
1741                                 ext3_mark_inode_dirty(handle, inode);
1742                         }
1743                 }
1744                 err = ext3_journal_stop(handle);
1745                 if (ret == 0)
1746                         ret = err;
1747         }
1748 out:
1749         return ret;
1750 }
1751
1752 /*
1753  * Pages can be marked dirty completely asynchronously from ext3's journalling
1754  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1755  * much here because ->set_page_dirty is called under VFS locks.  The page is
1756  * not necessarily locked.
1757  *
1758  * We cannot just dirty the page and leave attached buffers clean, because the
1759  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1760  * or jbddirty because all the journalling code will explode.
1761  *
1762  * So what we do is to mark the page "pending dirty" and next time writepage
1763  * is called, propagate that into the buffers appropriately.
1764  */
1765 static int ext3_journalled_set_page_dirty(struct page *page)
1766 {
1767         SetPageChecked(page);
1768         return __set_page_dirty_nobuffers(page);
1769 }
1770
1771 static const struct address_space_operations ext3_ordered_aops = {
1772         .readpage       = ext3_readpage,
1773         .readpages      = ext3_readpages,
1774         .writepage      = ext3_ordered_writepage,
1775         .sync_page      = block_sync_page,
1776         .write_begin    = ext3_write_begin,
1777         .write_end      = ext3_ordered_write_end,
1778         .bmap           = ext3_bmap,
1779         .invalidatepage = ext3_invalidatepage,
1780         .releasepage    = ext3_releasepage,
1781         .direct_IO      = ext3_direct_IO,
1782         .migratepage    = buffer_migrate_page,
1783 };
1784
1785 static const struct address_space_operations ext3_writeback_aops = {
1786         .readpage       = ext3_readpage,
1787         .readpages      = ext3_readpages,
1788         .writepage      = ext3_writeback_writepage,
1789         .sync_page      = block_sync_page,
1790         .write_begin    = ext3_write_begin,
1791         .write_end      = ext3_writeback_write_end,
1792         .bmap           = ext3_bmap,
1793         .invalidatepage = ext3_invalidatepage,
1794         .releasepage    = ext3_releasepage,
1795         .direct_IO      = ext3_direct_IO,
1796         .migratepage    = buffer_migrate_page,
1797 };
1798
1799 static const struct address_space_operations ext3_journalled_aops = {
1800         .readpage       = ext3_readpage,
1801         .readpages      = ext3_readpages,
1802         .writepage      = ext3_journalled_writepage,
1803         .sync_page      = block_sync_page,
1804         .write_begin    = ext3_write_begin,
1805         .write_end      = ext3_journalled_write_end,
1806         .set_page_dirty = ext3_journalled_set_page_dirty,
1807         .bmap           = ext3_bmap,
1808         .invalidatepage = ext3_invalidatepage,
1809         .releasepage    = ext3_releasepage,
1810 };
1811
1812 void ext3_set_aops(struct inode *inode)
1813 {
1814         if (ext3_should_order_data(inode))
1815                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1816         else if (ext3_should_writeback_data(inode))
1817                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1818         else
1819                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1820 }
1821
1822 /*
1823  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1824  * up to the end of the block which corresponds to `from'.
1825  * This required during truncate. We need to physically zero the tail end
1826  * of that block so it doesn't yield old data if the file is later grown.
1827  */
1828 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1829                 struct address_space *mapping, loff_t from)
1830 {
1831         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1832         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1833         unsigned blocksize, iblock, length, pos;
1834         struct inode *inode = mapping->host;
1835         struct buffer_head *bh;
1836         int err = 0;
1837
1838         blocksize = inode->i_sb->s_blocksize;
1839         length = blocksize - (offset & (blocksize - 1));
1840         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1841
1842         /*
1843          * For "nobh" option,  we can only work if we don't need to
1844          * read-in the page - otherwise we create buffers to do the IO.
1845          */
1846         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1847              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1848                 zero_user_page(page, offset, length, KM_USER0);
1849                 set_page_dirty(page);
1850                 goto unlock;
1851         }
1852
1853         if (!page_has_buffers(page))
1854                 create_empty_buffers(page, blocksize, 0);
1855
1856         /* Find the buffer that contains "offset" */
1857         bh = page_buffers(page);
1858         pos = blocksize;
1859         while (offset >= pos) {
1860                 bh = bh->b_this_page;
1861                 iblock++;
1862                 pos += blocksize;
1863         }
1864
1865         err = 0;
1866         if (buffer_freed(bh)) {
1867                 BUFFER_TRACE(bh, "freed: skip");
1868                 goto unlock;
1869         }
1870
1871         if (!buffer_mapped(bh)) {
1872                 BUFFER_TRACE(bh, "unmapped");
1873                 ext3_get_block(inode, iblock, bh, 0);
1874                 /* unmapped? It's a hole - nothing to do */
1875                 if (!buffer_mapped(bh)) {
1876                         BUFFER_TRACE(bh, "still unmapped");
1877                         goto unlock;
1878                 }
1879         }
1880
1881         /* Ok, it's mapped. Make sure it's up-to-date */
1882         if (PageUptodate(page))
1883                 set_buffer_uptodate(bh);
1884
1885         if (!buffer_uptodate(bh)) {
1886                 err = -EIO;
1887                 ll_rw_block(READ, 1, &bh);
1888                 wait_on_buffer(bh);
1889                 /* Uhhuh. Read error. Complain and punt. */
1890                 if (!buffer_uptodate(bh))
1891                         goto unlock;
1892         }
1893
1894         if (ext3_should_journal_data(inode)) {
1895                 BUFFER_TRACE(bh, "get write access");
1896                 err = ext3_journal_get_write_access(handle, bh);
1897                 if (err)
1898                         goto unlock;
1899         }
1900
1901         zero_user_page(page, offset, length, KM_USER0);
1902         BUFFER_TRACE(bh, "zeroed end of block");
1903
1904         err = 0;
1905         if (ext3_should_journal_data(inode)) {
1906                 err = ext3_journal_dirty_metadata(handle, bh);
1907         } else {
1908                 if (ext3_should_order_data(inode))
1909                         err = ext3_journal_dirty_data(handle, bh);
1910                 mark_buffer_dirty(bh);
1911         }
1912
1913 unlock:
1914         unlock_page(page);
1915         page_cache_release(page);
1916         return err;
1917 }
1918
1919 /*
1920  * Probably it should be a library function... search for first non-zero word
1921  * or memcmp with zero_page, whatever is better for particular architecture.
1922  * Linus?
1923  */
1924 static inline int all_zeroes(__le32 *p, __le32 *q)
1925 {
1926         while (p < q)
1927                 if (*p++)
1928                         return 0;
1929         return 1;
1930 }
1931
1932 /**
1933  *      ext3_find_shared - find the indirect blocks for partial truncation.
1934  *      @inode:   inode in question
1935  *      @depth:   depth of the affected branch
1936  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1937  *      @chain:   place to store the pointers to partial indirect blocks
1938  *      @top:     place to the (detached) top of branch
1939  *
1940  *      This is a helper function used by ext3_truncate().
1941  *
1942  *      When we do truncate() we may have to clean the ends of several
1943  *      indirect blocks but leave the blocks themselves alive. Block is
1944  *      partially truncated if some data below the new i_size is refered
1945  *      from it (and it is on the path to the first completely truncated
1946  *      data block, indeed).  We have to free the top of that path along
1947  *      with everything to the right of the path. Since no allocation
1948  *      past the truncation point is possible until ext3_truncate()
1949  *      finishes, we may safely do the latter, but top of branch may
1950  *      require special attention - pageout below the truncation point
1951  *      might try to populate it.
1952  *
1953  *      We atomically detach the top of branch from the tree, store the
1954  *      block number of its root in *@top, pointers to buffer_heads of
1955  *      partially truncated blocks - in @chain[].bh and pointers to
1956  *      their last elements that should not be removed - in
1957  *      @chain[].p. Return value is the pointer to last filled element
1958  *      of @chain.
1959  *
1960  *      The work left to caller to do the actual freeing of subtrees:
1961  *              a) free the subtree starting from *@top
1962  *              b) free the subtrees whose roots are stored in
1963  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1964  *              c) free the subtrees growing from the inode past the @chain[0].
1965  *                      (no partially truncated stuff there).  */
1966
1967 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1968                         int offsets[4], Indirect chain[4], __le32 *top)
1969 {
1970         Indirect *partial, *p;
1971         int k, err;
1972
1973         *top = 0;
1974         /* Make k index the deepest non-null offest + 1 */
1975         for (k = depth; k > 1 && !offsets[k-1]; k--)
1976                 ;
1977         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1978         /* Writer: pointers */
1979         if (!partial)
1980                 partial = chain + k-1;
1981         /*
1982          * If the branch acquired continuation since we've looked at it -
1983          * fine, it should all survive and (new) top doesn't belong to us.
1984          */
1985         if (!partial->key && *partial->p)
1986                 /* Writer: end */
1987                 goto no_top;
1988         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1989                 ;
1990         /*
1991          * OK, we've found the last block that must survive. The rest of our
1992          * branch should be detached before unlocking. However, if that rest
1993          * of branch is all ours and does not grow immediately from the inode
1994          * it's easier to cheat and just decrement partial->p.
1995          */
1996         if (p == chain + k - 1 && p > chain) {
1997                 p->p--;
1998         } else {
1999                 *top = *p->p;
2000                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2001 #if 0
2002                 *p->p = 0;
2003 #endif
2004         }
2005         /* Writer: end */
2006
2007         while(partial > p) {
2008                 brelse(partial->bh);
2009                 partial--;
2010         }
2011 no_top:
2012         return partial;
2013 }
2014
2015 /*
2016  * Zero a number of block pointers in either an inode or an indirect block.
2017  * If we restart the transaction we must again get write access to the
2018  * indirect block for further modification.
2019  *
2020  * We release `count' blocks on disk, but (last - first) may be greater
2021  * than `count' because there can be holes in there.
2022  */
2023 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2024                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2025                 unsigned long count, __le32 *first, __le32 *last)
2026 {
2027         __le32 *p;
2028         if (try_to_extend_transaction(handle, inode)) {
2029                 if (bh) {
2030                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2031                         ext3_journal_dirty_metadata(handle, bh);
2032                 }
2033                 ext3_mark_inode_dirty(handle, inode);
2034                 ext3_journal_test_restart(handle, inode);
2035                 if (bh) {
2036                         BUFFER_TRACE(bh, "retaking write access");
2037                         ext3_journal_get_write_access(handle, bh);
2038                 }
2039         }
2040
2041         /*
2042          * Any buffers which are on the journal will be in memory. We find
2043          * them on the hash table so journal_revoke() will run journal_forget()
2044          * on them.  We've already detached each block from the file, so
2045          * bforget() in journal_forget() should be safe.
2046          *
2047          * AKPM: turn on bforget in journal_forget()!!!
2048          */
2049         for (p = first; p < last; p++) {
2050                 u32 nr = le32_to_cpu(*p);
2051                 if (nr) {
2052                         struct buffer_head *bh;
2053
2054                         *p = 0;
2055                         bh = sb_find_get_block(inode->i_sb, nr);
2056                         ext3_forget(handle, 0, inode, bh, nr);
2057                 }
2058         }
2059
2060         ext3_free_blocks(handle, inode, block_to_free, count);
2061 }
2062
2063 /**
2064  * ext3_free_data - free a list of data blocks
2065  * @handle:     handle for this transaction
2066  * @inode:      inode we are dealing with
2067  * @this_bh:    indirect buffer_head which contains *@first and *@last
2068  * @first:      array of block numbers
2069  * @last:       points immediately past the end of array
2070  *
2071  * We are freeing all blocks refered from that array (numbers are stored as
2072  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2073  *
2074  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2075  * blocks are contiguous then releasing them at one time will only affect one
2076  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2077  * actually use a lot of journal space.
2078  *
2079  * @this_bh will be %NULL if @first and @last point into the inode's direct
2080  * block pointers.
2081  */
2082 static void ext3_free_data(handle_t *handle, struct inode *inode,
2083                            struct buffer_head *this_bh,
2084                            __le32 *first, __le32 *last)
2085 {
2086         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2087         unsigned long count = 0;            /* Number of blocks in the run */
2088         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2089                                                corresponding to
2090                                                block_to_free */
2091         ext3_fsblk_t nr;                    /* Current block # */
2092         __le32 *p;                          /* Pointer into inode/ind
2093                                                for current block */
2094         int err;
2095
2096         if (this_bh) {                          /* For indirect block */
2097                 BUFFER_TRACE(this_bh, "get_write_access");
2098                 err = ext3_journal_get_write_access(handle, this_bh);
2099                 /* Important: if we can't update the indirect pointers
2100                  * to the blocks, we can't free them. */
2101                 if (err)
2102                         return;
2103         }
2104
2105         for (p = first; p < last; p++) {
2106                 nr = le32_to_cpu(*p);
2107                 if (nr) {
2108                         /* accumulate blocks to free if they're contiguous */
2109                         if (count == 0) {
2110                                 block_to_free = nr;
2111                                 block_to_free_p = p;
2112                                 count = 1;
2113                         } else if (nr == block_to_free + count) {
2114                                 count++;
2115                         } else {
2116                                 ext3_clear_blocks(handle, inode, this_bh,
2117                                                   block_to_free,
2118                                                   count, block_to_free_p, p);
2119                                 block_to_free = nr;
2120                                 block_to_free_p = p;
2121                                 count = 1;
2122                         }
2123                 }
2124         }
2125
2126         if (count > 0)
2127                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2128                                   count, block_to_free_p, p);
2129
2130         if (this_bh) {
2131                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2132                 ext3_journal_dirty_metadata(handle, this_bh);
2133         }
2134 }
2135
2136 /**
2137  *      ext3_free_branches - free an array of branches
2138  *      @handle: JBD handle for this transaction
2139  *      @inode: inode we are dealing with
2140  *      @parent_bh: the buffer_head which contains *@first and *@last
2141  *      @first: array of block numbers
2142  *      @last:  pointer immediately past the end of array
2143  *      @depth: depth of the branches to free
2144  *
2145  *      We are freeing all blocks refered from these branches (numbers are
2146  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2147  *      appropriately.
2148  */
2149 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2150                                struct buffer_head *parent_bh,
2151                                __le32 *first, __le32 *last, int depth)
2152 {
2153         ext3_fsblk_t nr;
2154         __le32 *p;
2155
2156         if (is_handle_aborted(handle))
2157                 return;
2158
2159         if (depth--) {
2160                 struct buffer_head *bh;
2161                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2162                 p = last;
2163                 while (--p >= first) {
2164                         nr = le32_to_cpu(*p);
2165                         if (!nr)
2166                                 continue;               /* A hole */
2167
2168                         /* Go read the buffer for the next level down */
2169                         bh = sb_bread(inode->i_sb, nr);
2170
2171                         /*
2172                          * A read failure? Report error and clear slot
2173                          * (should be rare).
2174                          */
2175                         if (!bh) {
2176                                 ext3_error(inode->i_sb, "ext3_free_branches",
2177                                            "Read failure, inode=%lu, block="E3FSBLK,
2178                                            inode->i_ino, nr);
2179                                 continue;
2180                         }
2181
2182                         /* This zaps the entire block.  Bottom up. */
2183                         BUFFER_TRACE(bh, "free child branches");
2184                         ext3_free_branches(handle, inode, bh,
2185                                            (__le32*)bh->b_data,
2186                                            (__le32*)bh->b_data + addr_per_block,
2187                                            depth);
2188
2189                         /*
2190                          * We've probably journalled the indirect block several
2191                          * times during the truncate.  But it's no longer
2192                          * needed and we now drop it from the transaction via
2193                          * journal_revoke().
2194                          *
2195                          * That's easy if it's exclusively part of this
2196                          * transaction.  But if it's part of the committing
2197                          * transaction then journal_forget() will simply
2198                          * brelse() it.  That means that if the underlying
2199                          * block is reallocated in ext3_get_block(),
2200                          * unmap_underlying_metadata() will find this block
2201                          * and will try to get rid of it.  damn, damn.
2202                          *
2203                          * If this block has already been committed to the
2204                          * journal, a revoke record will be written.  And
2205                          * revoke records must be emitted *before* clearing
2206                          * this block's bit in the bitmaps.
2207                          */
2208                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2209
2210                         /*
2211                          * Everything below this this pointer has been
2212                          * released.  Now let this top-of-subtree go.
2213                          *
2214                          * We want the freeing of this indirect block to be
2215                          * atomic in the journal with the updating of the
2216                          * bitmap block which owns it.  So make some room in
2217                          * the journal.
2218                          *
2219                          * We zero the parent pointer *after* freeing its
2220                          * pointee in the bitmaps, so if extend_transaction()
2221                          * for some reason fails to put the bitmap changes and
2222                          * the release into the same transaction, recovery
2223                          * will merely complain about releasing a free block,
2224                          * rather than leaking blocks.
2225                          */
2226                         if (is_handle_aborted(handle))
2227                                 return;
2228                         if (try_to_extend_transaction(handle, inode)) {
2229                                 ext3_mark_inode_dirty(handle, inode);
2230                                 ext3_journal_test_restart(handle, inode);
2231                         }
2232
2233                         ext3_free_blocks(handle, inode, nr, 1);
2234
2235                         if (parent_bh) {
2236                                 /*
2237                                  * The block which we have just freed is
2238                                  * pointed to by an indirect block: journal it
2239                                  */
2240                                 BUFFER_TRACE(parent_bh, "get_write_access");
2241                                 if (!ext3_journal_get_write_access(handle,
2242                                                                    parent_bh)){
2243                                         *p = 0;
2244                                         BUFFER_TRACE(parent_bh,
2245                                         "call ext3_journal_dirty_metadata");
2246                                         ext3_journal_dirty_metadata(handle,
2247                                                                     parent_bh);
2248                                 }
2249                         }
2250                 }
2251         } else {
2252                 /* We have reached the bottom of the tree. */
2253                 BUFFER_TRACE(parent_bh, "free data blocks");
2254                 ext3_free_data(handle, inode, parent_bh, first, last);
2255         }
2256 }
2257
2258 /*
2259  * ext3_truncate()
2260  *
2261  * We block out ext3_get_block() block instantiations across the entire
2262  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2263  * simultaneously on behalf of the same inode.
2264  *
2265  * As we work through the truncate and commmit bits of it to the journal there
2266  * is one core, guiding principle: the file's tree must always be consistent on
2267  * disk.  We must be able to restart the truncate after a crash.
2268  *
2269  * The file's tree may be transiently inconsistent in memory (although it
2270  * probably isn't), but whenever we close off and commit a journal transaction,
2271  * the contents of (the filesystem + the journal) must be consistent and
2272  * restartable.  It's pretty simple, really: bottom up, right to left (although
2273  * left-to-right works OK too).
2274  *
2275  * Note that at recovery time, journal replay occurs *before* the restart of
2276  * truncate against the orphan inode list.
2277  *
2278  * The committed inode has the new, desired i_size (which is the same as
2279  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2280  * that this inode's truncate did not complete and it will again call
2281  * ext3_truncate() to have another go.  So there will be instantiated blocks
2282  * to the right of the truncation point in a crashed ext3 filesystem.  But
2283  * that's fine - as long as they are linked from the inode, the post-crash
2284  * ext3_truncate() run will find them and release them.
2285  */
2286 void ext3_truncate(struct inode *inode)
2287 {
2288         handle_t *handle;
2289         struct ext3_inode_info *ei = EXT3_I(inode);
2290         __le32 *i_data = ei->i_data;
2291         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2292         struct address_space *mapping = inode->i_mapping;
2293         int offsets[4];
2294         Indirect chain[4];
2295         Indirect *partial;
2296         __le32 nr = 0;
2297         int n;
2298         long last_block;
2299         unsigned blocksize = inode->i_sb->s_blocksize;
2300         struct page *page;
2301
2302         if (!(S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
2303             S_ISLNK(inode->i_mode)))
2304                 return;
2305         if (ext3_inode_is_fast_symlink(inode))
2306                 return;
2307         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2308                 return;
2309
2310         /*
2311          * We have to lock the EOF page here, because lock_page() nests
2312          * outside journal_start().
2313          */
2314         if ((inode->i_size & (blocksize - 1)) == 0) {
2315                 /* Block boundary? Nothing to do */
2316                 page = NULL;
2317         } else {
2318                 page = grab_cache_page(mapping,
2319                                 inode->i_size >> PAGE_CACHE_SHIFT);
2320                 if (!page)
2321                         return;
2322         }
2323
2324         handle = start_transaction(inode);
2325         if (IS_ERR(handle)) {
2326                 if (page) {
2327                         clear_highpage(page);
2328                         flush_dcache_page(page);
2329                         unlock_page(page);
2330                         page_cache_release(page);
2331                 }
2332                 return;         /* AKPM: return what? */
2333         }
2334
2335         last_block = (inode->i_size + blocksize-1)
2336                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2337
2338         if (page)
2339                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2340
2341         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2342         if (n == 0)
2343                 goto out_stop;  /* error */
2344
2345         /*
2346          * OK.  This truncate is going to happen.  We add the inode to the
2347          * orphan list, so that if this truncate spans multiple transactions,
2348          * and we crash, we will resume the truncate when the filesystem
2349          * recovers.  It also marks the inode dirty, to catch the new size.
2350          *
2351          * Implication: the file must always be in a sane, consistent
2352          * truncatable state while each transaction commits.
2353          */
2354         if (ext3_orphan_add(handle, inode))
2355                 goto out_stop;
2356
2357         /*
2358          * The orphan list entry will now protect us from any crash which
2359          * occurs before the truncate completes, so it is now safe to propagate
2360          * the new, shorter inode size (held for now in i_size) into the
2361          * on-disk inode. We do this via i_disksize, which is the value which
2362          * ext3 *really* writes onto the disk inode.
2363          */
2364         ei->i_disksize = inode->i_size;
2365
2366         /*
2367          * From here we block out all ext3_get_block() callers who want to
2368          * modify the block allocation tree.
2369          */
2370         mutex_lock(&ei->truncate_mutex);
2371
2372         if (n == 1) {           /* direct blocks */
2373                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2374                                i_data + EXT3_NDIR_BLOCKS);
2375                 goto do_indirects;
2376         }
2377
2378         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2379         /* Kill the top of shared branch (not detached) */
2380         if (nr) {
2381                 if (partial == chain) {
2382                         /* Shared branch grows from the inode */
2383                         ext3_free_branches(handle, inode, NULL,
2384                                            &nr, &nr+1, (chain+n-1) - partial);
2385                         *partial->p = 0;
2386                         /*
2387                          * We mark the inode dirty prior to restart,
2388                          * and prior to stop.  No need for it here.
2389                          */
2390                 } else {
2391                         /* Shared branch grows from an indirect block */
2392                         BUFFER_TRACE(partial->bh, "get_write_access");
2393                         ext3_free_branches(handle, inode, partial->bh,
2394                                         partial->p,
2395                                         partial->p+1, (chain+n-1) - partial);
2396                 }
2397         }
2398         /* Clear the ends of indirect blocks on the shared branch */
2399         while (partial > chain) {
2400                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2401                                    (__le32*)partial->bh->b_data+addr_per_block,
2402                                    (chain+n-1) - partial);
2403                 BUFFER_TRACE(partial->bh, "call brelse");
2404                 brelse (partial->bh);
2405                 partial--;
2406         }
2407 do_indirects:
2408         /* Kill the remaining (whole) subtrees */
2409         switch (offsets[0]) {
2410         default:
2411                 nr = i_data[EXT3_IND_BLOCK];
2412                 if (nr) {
2413                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2414                         i_data[EXT3_IND_BLOCK] = 0;
2415                 }
2416         case EXT3_IND_BLOCK:
2417                 nr = i_data[EXT3_DIND_BLOCK];
2418                 if (nr) {
2419                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2420                         i_data[EXT3_DIND_BLOCK] = 0;
2421                 }
2422         case EXT3_DIND_BLOCK:
2423                 nr = i_data[EXT3_TIND_BLOCK];
2424                 if (nr) {
2425                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2426                         i_data[EXT3_TIND_BLOCK] = 0;
2427                 }
2428         case EXT3_TIND_BLOCK:
2429                 ;
2430         }
2431
2432         ext3_discard_reservation(inode);
2433
2434         mutex_unlock(&ei->truncate_mutex);
2435         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2436         ext3_mark_inode_dirty(handle, inode);
2437
2438         /*
2439          * In a multi-transaction truncate, we only make the final transaction
2440          * synchronous
2441          */
2442         if (IS_SYNC(inode))
2443                 handle->h_sync = 1;
2444 out_stop:
2445         /*
2446          * If this was a simple ftruncate(), and the file will remain alive
2447          * then we need to clear up the orphan record which we created above.
2448          * However, if this was a real unlink then we were called by
2449          * ext3_delete_inode(), and we allow that function to clean up the
2450          * orphan info for us.
2451          */
2452         if (inode->i_nlink)
2453                 ext3_orphan_del(handle, inode);
2454
2455         ext3_journal_stop(handle);
2456 }
2457
2458 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2459                 unsigned long ino, struct ext3_iloc *iloc)
2460 {
2461         unsigned long desc, group_desc, block_group;
2462         unsigned long offset;
2463         ext3_fsblk_t block;
2464         struct buffer_head *bh;
2465         struct ext3_group_desc * gdp;
2466
2467         if (!ext3_valid_inum(sb, ino)) {
2468                 /*
2469                  * This error is already checked for in namei.c unless we are
2470                  * looking at an NFS filehandle, in which case no error
2471                  * report is needed
2472                  */
2473                 return 0;
2474         }
2475
2476         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2477         if (block_group >= EXT3_SB(sb)->s_groups_count) {
2478                 ext3_error(sb,"ext3_get_inode_block","group >= groups count");
2479                 return 0;
2480         }
2481         smp_rmb();
2482         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
2483         desc = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
2484         bh = EXT3_SB(sb)->s_group_desc[group_desc];
2485         if (!bh) {
2486                 ext3_error (sb, "ext3_get_inode_block",
2487                             "Descriptor not loaded");
2488                 return 0;
2489         }
2490
2491         gdp = (struct ext3_group_desc *)bh->b_data;
2492         /*
2493          * Figure out the offset within the block group inode table
2494          */
2495         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2496                 EXT3_INODE_SIZE(sb);
2497         block = le32_to_cpu(gdp[desc].bg_inode_table) +
2498                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2499
2500         iloc->block_group = block_group;
2501         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2502         return block;
2503 }
2504
2505 /*
2506  * ext3_get_inode_loc returns with an extra refcount against the inode's
2507  * underlying buffer_head on success. If 'in_mem' is true, we have all
2508  * data in memory that is needed to recreate the on-disk version of this
2509  * inode.
2510  */
2511 static int __ext3_get_inode_loc(struct inode *inode,
2512                                 struct ext3_iloc *iloc, int in_mem)
2513 {
2514         ext3_fsblk_t block;
2515         struct buffer_head *bh;
2516
2517         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2518         if (!block)
2519                 return -EIO;
2520
2521         bh = sb_getblk(inode->i_sb, block);
2522         if (!bh) {
2523                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2524                                 "unable to read inode block - "
2525                                 "inode=%lu, block="E3FSBLK,
2526                                  inode->i_ino, block);
2527                 return -EIO;
2528         }
2529         if (!buffer_uptodate(bh)) {
2530                 lock_buffer(bh);
2531                 if (buffer_uptodate(bh)) {
2532                         /* someone brought it uptodate while we waited */
2533                         unlock_buffer(bh);
2534                         goto has_buffer;
2535                 }
2536
2537                 /*
2538                  * If we have all information of the inode in memory and this
2539                  * is the only valid inode in the block, we need not read the
2540                  * block.
2541                  */
2542                 if (in_mem) {
2543                         struct buffer_head *bitmap_bh;
2544                         struct ext3_group_desc *desc;
2545                         int inodes_per_buffer;
2546                         int inode_offset, i;
2547                         int block_group;
2548                         int start;
2549
2550                         block_group = (inode->i_ino - 1) /
2551                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2552                         inodes_per_buffer = bh->b_size /
2553                                 EXT3_INODE_SIZE(inode->i_sb);
2554                         inode_offset = ((inode->i_ino - 1) %
2555                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2556                         start = inode_offset & ~(inodes_per_buffer - 1);
2557
2558                         /* Is the inode bitmap in cache? */
2559                         desc = ext3_get_group_desc(inode->i_sb,
2560                                                 block_group, NULL);
2561                         if (!desc)
2562                                 goto make_io;
2563
2564                         bitmap_bh = sb_getblk(inode->i_sb,
2565                                         le32_to_cpu(desc->bg_inode_bitmap));
2566                         if (!bitmap_bh)
2567                                 goto make_io;
2568
2569                         /*
2570                          * If the inode bitmap isn't in cache then the
2571                          * optimisation may end up performing two reads instead
2572                          * of one, so skip it.
2573                          */
2574                         if (!buffer_uptodate(bitmap_bh)) {
2575                                 brelse(bitmap_bh);
2576                                 goto make_io;
2577                         }
2578                         for (i = start; i < start + inodes_per_buffer; i++) {
2579                                 if (i == inode_offset)
2580                                         continue;
2581                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2582                                         break;
2583                         }
2584                         brelse(bitmap_bh);
2585                         if (i == start + inodes_per_buffer) {
2586                                 /* all other inodes are free, so skip I/O */
2587                                 memset(bh->b_data, 0, bh->b_size);
2588                                 set_buffer_uptodate(bh);
2589                                 unlock_buffer(bh);
2590                                 goto has_buffer;
2591                         }
2592                 }
2593
2594 make_io:
2595                 /*
2596                  * There are other valid inodes in the buffer, this inode
2597                  * has in-inode xattrs, or we don't have this inode in memory.
2598                  * Read the block from disk.
2599                  */
2600                 get_bh(bh);
2601                 bh->b_end_io = end_buffer_read_sync;
2602                 submit_bh(READ_META, bh);
2603                 wait_on_buffer(bh);
2604                 if (!buffer_uptodate(bh)) {
2605                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2606                                         "unable to read inode block - "
2607                                         "inode=%lu, block="E3FSBLK,
2608                                         inode->i_ino, block);
2609                         brelse(bh);
2610                         return -EIO;
2611                 }
2612         }
2613 has_buffer:
2614         iloc->bh = bh;
2615         return 0;
2616 }
2617
2618 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2619 {
2620         /* We have all inode data except xattrs in memory here. */
2621         return __ext3_get_inode_loc(inode, iloc,
2622                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2623 }
2624
2625 void ext3_set_inode_flags(struct inode *inode)
2626 {
2627         unsigned int flags = EXT3_I(inode)->i_flags;
2628
2629         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2630         if (flags & EXT3_SYNC_FL)
2631                 inode->i_flags |= S_SYNC;
2632         if (flags & EXT3_APPEND_FL)
2633                 inode->i_flags |= S_APPEND;
2634         if (flags & EXT3_IMMUTABLE_FL)
2635                 inode->i_flags |= S_IMMUTABLE;
2636         if (flags & EXT3_NOATIME_FL)
2637                 inode->i_flags |= S_NOATIME;
2638         if (flags & EXT3_DIRSYNC_FL)
2639                 inode->i_flags |= S_DIRSYNC;
2640 }
2641
2642 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2643 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2644 {
2645         unsigned int flags = ei->vfs_inode.i_flags;
2646
2647         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2648                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2649         if (flags & S_SYNC)
2650                 ei->i_flags |= EXT3_SYNC_FL;
2651         if (flags & S_APPEND)
2652                 ei->i_flags |= EXT3_APPEND_FL;
2653         if (flags & S_IMMUTABLE)
2654                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2655         if (flags & S_NOATIME)
2656                 ei->i_flags |= EXT3_NOATIME_FL;
2657         if (flags & S_DIRSYNC)
2658                 ei->i_flags |= EXT3_DIRSYNC_FL;
2659 }
2660
2661 void ext3_read_inode(struct inode * inode)
2662 {
2663         struct ext3_iloc iloc;
2664         struct ext3_inode *raw_inode;
2665         struct ext3_inode_info *ei = EXT3_I(inode);
2666         struct buffer_head *bh;
2667         int block;
2668
2669 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2670         ei->i_acl = EXT3_ACL_NOT_CACHED;
2671         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2672 #endif
2673         ei->i_block_alloc_info = NULL;
2674
2675         if (__ext3_get_inode_loc(inode, &iloc, 0))
2676                 goto bad_inode;
2677         bh = iloc.bh;
2678         raw_inode = ext3_raw_inode(&iloc);
2679         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2680         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2681         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2682         if(!(test_opt (inode->i_sb, NO_UID32))) {
2683                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2684                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2685         }
2686         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2687         inode->i_size = le32_to_cpu(raw_inode->i_size);
2688         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2689         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2690         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2691         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2692
2693         ei->i_state = 0;
2694         ei->i_dir_start_lookup = 0;
2695         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2696         /* We now have enough fields to check if the inode was active or not.
2697          * This is needed because nfsd might try to access dead inodes
2698          * the test is that same one that e2fsck uses
2699          * NeilBrown 1999oct15
2700          */
2701         if (inode->i_nlink == 0) {
2702                 if (inode->i_mode == 0 ||
2703                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2704                         /* this inode is deleted */
2705                         brelse (bh);
2706                         goto bad_inode;
2707                 }
2708                 /* The only unlinked inodes we let through here have
2709                  * valid i_mode and are being read by the orphan
2710                  * recovery code: that's fine, we're about to complete
2711                  * the process of deleting those. */
2712         }
2713         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2714         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2715 #ifdef EXT3_FRAGMENTS
2716         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2717         ei->i_frag_no = raw_inode->i_frag;
2718         ei->i_frag_size = raw_inode->i_fsize;
2719 #endif
2720         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2721         if (!S_ISREG(inode->i_mode)) {
2722                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2723         } else {
2724                 inode->i_size |=
2725                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2726         }
2727         ei->i_disksize = inode->i_size;
2728         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2729         ei->i_block_group = iloc.block_group;
2730         /*
2731          * NOTE! The in-memory inode i_data array is in little-endian order
2732          * even on big-endian machines: we do NOT byteswap the block numbers!
2733          */
2734         for (block = 0; block < EXT3_N_BLOCKS; block++)
2735                 ei->i_data[block] = raw_inode->i_block[block];
2736         INIT_LIST_HEAD(&ei->i_orphan);
2737
2738         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2739             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2740                 /*
2741                  * When mke2fs creates big inodes it does not zero out
2742                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2743                  * so ignore those first few inodes.
2744                  */
2745                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2746                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2747                     EXT3_INODE_SIZE(inode->i_sb)) {
2748                         brelse (bh);
2749                         goto bad_inode;
2750                 }
2751                 if (ei->i_extra_isize == 0) {
2752                         /* The extra space is currently unused. Use it. */
2753                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2754                                             EXT3_GOOD_OLD_INODE_SIZE;
2755                 } else {
2756                         __le32 *magic = (void *)raw_inode +
2757                                         EXT3_GOOD_OLD_INODE_SIZE +
2758                                         ei->i_extra_isize;
2759                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2760                                  ei->i_state |= EXT3_STATE_XATTR;
2761                 }
2762         } else
2763                 ei->i_extra_isize = 0;
2764
2765         if (S_ISREG(inode->i_mode)) {
2766                 inode->i_op = &ext3_file_inode_operations;
2767                 inode->i_fop = &ext3_file_operations;
2768                 ext3_set_aops(inode);
2769         } else if (S_ISDIR(inode->i_mode)) {
2770                 inode->i_op = &ext3_dir_inode_operations;
2771                 inode->i_fop = &ext3_dir_operations;
2772         } else if (S_ISLNK(inode->i_mode)) {
2773                 if (ext3_inode_is_fast_symlink(inode))
2774                         inode->i_op = &ext3_fast_symlink_inode_operations;
2775                 else {
2776                         inode->i_op = &ext3_symlink_inode_operations;
2777                         ext3_set_aops(inode);
2778                 }
2779         } else {
2780                 inode->i_op = &ext3_special_inode_operations;
2781                 if (raw_inode->i_block[0])
2782                         init_special_inode(inode, inode->i_mode,
2783                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2784                 else
2785                         init_special_inode(inode, inode->i_mode,
2786                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2787         }
2788         brelse (iloc.bh);
2789         ext3_set_inode_flags(inode);
2790         return;
2791
2792 bad_inode:
2793         make_bad_inode(inode);
2794         return;
2795 }
2796
2797 /*
2798  * Post the struct inode info into an on-disk inode location in the
2799  * buffer-cache.  This gobbles the caller's reference to the
2800  * buffer_head in the inode location struct.
2801  *
2802  * The caller must have write access to iloc->bh.
2803  */
2804 static int ext3_do_update_inode(handle_t *handle,
2805                                 struct inode *inode,
2806                                 struct ext3_iloc *iloc)
2807 {
2808         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2809         struct ext3_inode_info *ei = EXT3_I(inode);
2810         struct buffer_head *bh = iloc->bh;
2811         int err = 0, rc, block;
2812
2813         /* For fields not not tracking in the in-memory inode,
2814          * initialise them to zero for new inodes. */
2815         if (ei->i_state & EXT3_STATE_NEW)
2816                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2817
2818         ext3_get_inode_flags(ei);
2819         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2820         if(!(test_opt(inode->i_sb, NO_UID32))) {
2821                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2822                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2823 /*
2824  * Fix up interoperability with old kernels. Otherwise, old inodes get
2825  * re-used with the upper 16 bits of the uid/gid intact
2826  */
2827                 if(!ei->i_dtime) {
2828                         raw_inode->i_uid_high =
2829                                 cpu_to_le16(high_16_bits(inode->i_uid));
2830                         raw_inode->i_gid_high =
2831                                 cpu_to_le16(high_16_bits(inode->i_gid));
2832                 } else {
2833                         raw_inode->i_uid_high = 0;
2834                         raw_inode->i_gid_high = 0;
2835                 }
2836         } else {
2837                 raw_inode->i_uid_low =
2838                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2839                 raw_inode->i_gid_low =
2840                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2841                 raw_inode->i_uid_high = 0;
2842                 raw_inode->i_gid_high = 0;
2843         }
2844         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2845         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2846         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2847         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2848         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2849         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2850         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2851         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2852 #ifdef EXT3_FRAGMENTS
2853         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2854         raw_inode->i_frag = ei->i_frag_no;
2855         raw_inode->i_fsize = ei->i_frag_size;
2856 #endif
2857         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2858         if (!S_ISREG(inode->i_mode)) {
2859                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2860         } else {
2861                 raw_inode->i_size_high =
2862                         cpu_to_le32(ei->i_disksize >> 32);
2863                 if (ei->i_disksize > 0x7fffffffULL) {
2864                         struct super_block *sb = inode->i_sb;
2865                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2866                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2867                             EXT3_SB(sb)->s_es->s_rev_level ==
2868                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2869                                /* If this is the first large file
2870                                 * created, add a flag to the superblock.
2871                                 */
2872                                 err = ext3_journal_get_write_access(handle,
2873                                                 EXT3_SB(sb)->s_sbh);
2874                                 if (err)
2875                                         goto out_brelse;
2876                                 ext3_update_dynamic_rev(sb);
2877                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2878                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2879                                 sb->s_dirt = 1;
2880                                 handle->h_sync = 1;
2881                                 err = ext3_journal_dirty_metadata(handle,
2882                                                 EXT3_SB(sb)->s_sbh);
2883                         }
2884                 }
2885         }
2886         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2887         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2888                 if (old_valid_dev(inode->i_rdev)) {
2889                         raw_inode->i_block[0] =
2890                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2891                         raw_inode->i_block[1] = 0;
2892                 } else {
2893                         raw_inode->i_block[0] = 0;
2894                         raw_inode->i_block[1] =
2895                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2896                         raw_inode->i_block[2] = 0;
2897                 }
2898         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2899                 raw_inode->i_block[block] = ei->i_data[block];
2900
2901         if (ei->i_extra_isize)
2902                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2903
2904         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2905         rc = ext3_journal_dirty_metadata(handle, bh);
2906         if (!err)
2907                 err = rc;
2908         ei->i_state &= ~EXT3_STATE_NEW;
2909
2910 out_brelse:
2911         brelse (bh);
2912         ext3_std_error(inode->i_sb, err);
2913         return err;
2914 }
2915
2916 /*
2917  * ext3_write_inode()
2918  *
2919  * We are called from a few places:
2920  *
2921  * - Within generic_file_write() for O_SYNC files.
2922  *   Here, there will be no transaction running. We wait for any running
2923  *   trasnaction to commit.
2924  *
2925  * - Within sys_sync(), kupdate and such.
2926  *   We wait on commit, if tol to.
2927  *
2928  * - Within prune_icache() (PF_MEMALLOC == true)
2929  *   Here we simply return.  We can't afford to block kswapd on the
2930  *   journal commit.
2931  *
2932  * In all cases it is actually safe for us to return without doing anything,
2933  * because the inode has been copied into a raw inode buffer in
2934  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2935  * knfsd.
2936  *
2937  * Note that we are absolutely dependent upon all inode dirtiers doing the
2938  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2939  * which we are interested.
2940  *
2941  * It would be a bug for them to not do this.  The code:
2942  *
2943  *      mark_inode_dirty(inode)
2944  *      stuff();
2945  *      inode->i_size = expr;
2946  *
2947  * is in error because a kswapd-driven write_inode() could occur while
2948  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2949  * will no longer be on the superblock's dirty inode list.
2950  */
2951 int ext3_write_inode(struct inode *inode, int wait)
2952 {
2953         if (current->flags & PF_MEMALLOC)
2954                 return 0;
2955
2956         if (ext3_journal_current_handle()) {
2957                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
2958                 dump_stack();
2959                 return -EIO;
2960         }
2961
2962         if (!wait)
2963                 return 0;
2964
2965         return ext3_force_commit(inode->i_sb);
2966 }
2967
2968 /*
2969  * ext3_setattr()
2970  *
2971  * Called from notify_change.
2972  *
2973  * We want to trap VFS attempts to truncate the file as soon as
2974  * possible.  In particular, we want to make sure that when the VFS
2975  * shrinks i_size, we put the inode on the orphan list and modify
2976  * i_disksize immediately, so that during the subsequent flushing of
2977  * dirty pages and freeing of disk blocks, we can guarantee that any
2978  * commit will leave the blocks being flushed in an unused state on
2979  * disk.  (On recovery, the inode will get truncated and the blocks will
2980  * be freed, so we have a strong guarantee that no future commit will
2981  * leave these blocks visible to the user.)
2982  *
2983  * Called with inode->sem down.
2984  */
2985 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
2986 {
2987         struct inode *inode = dentry->d_inode;
2988         int error, rc = 0;
2989         const unsigned int ia_valid = attr->ia_valid;
2990
2991         error = inode_change_ok(inode, attr);
2992         if (error)
2993                 return error;
2994
2995         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
2996                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
2997                 handle_t *handle;
2998
2999                 /* (user+group)*(old+new) structure, inode write (sb,
3000                  * inode block, ? - but truncate inode update has it) */
3001                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3002                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3003                 if (IS_ERR(handle)) {
3004                         error = PTR_ERR(handle);
3005                         goto err_out;
3006                 }
3007                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
3008                 if (error) {
3009                         ext3_journal_stop(handle);
3010                         return error;
3011                 }
3012                 /* Update corresponding info in inode so that everything is in
3013                  * one transaction */
3014                 if (attr->ia_valid & ATTR_UID)
3015                         inode->i_uid = attr->ia_uid;
3016                 if (attr->ia_valid & ATTR_GID)
3017                         inode->i_gid = attr->ia_gid;
3018                 error = ext3_mark_inode_dirty(handle, inode);
3019                 ext3_journal_stop(handle);
3020         }
3021
3022         if (S_ISREG(inode->i_mode) &&
3023             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3024                 handle_t *handle;
3025
3026                 handle = ext3_journal_start(inode, 3);
3027                 if (IS_ERR(handle)) {
3028                         error = PTR_ERR(handle);
3029                         goto err_out;
3030                 }
3031
3032                 error = ext3_orphan_add(handle, inode);
3033                 EXT3_I(inode)->i_disksize = attr->ia_size;
3034                 rc = ext3_mark_inode_dirty(handle, inode);
3035                 if (!error)
3036                         error = rc;
3037                 ext3_journal_stop(handle);
3038         }
3039
3040         rc = inode_setattr(inode, attr);
3041
3042         /* If inode_setattr's call to ext3_truncate failed to get a
3043          * transaction handle at all, we need to clean up the in-core
3044          * orphan list manually. */
3045         if (inode->i_nlink)
3046                 ext3_orphan_del(NULL, inode);
3047
3048         if (!rc && (ia_valid & ATTR_MODE))
3049                 rc = ext3_acl_chmod(inode);
3050
3051 err_out:
3052         ext3_std_error(inode->i_sb, error);
3053         if (!error)
3054                 error = rc;
3055         return error;
3056 }
3057
3058
3059 /*
3060  * How many blocks doth make a writepage()?
3061  *
3062  * With N blocks per page, it may be:
3063  * N data blocks
3064  * 2 indirect block
3065  * 2 dindirect
3066  * 1 tindirect
3067  * N+5 bitmap blocks (from the above)
3068  * N+5 group descriptor summary blocks
3069  * 1 inode block
3070  * 1 superblock.
3071  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3072  *
3073  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3074  *
3075  * With ordered or writeback data it's the same, less the N data blocks.
3076  *
3077  * If the inode's direct blocks can hold an integral number of pages then a
3078  * page cannot straddle two indirect blocks, and we can only touch one indirect
3079  * and dindirect block, and the "5" above becomes "3".
3080  *
3081  * This still overestimates under most circumstances.  If we were to pass the
3082  * start and end offsets in here as well we could do block_to_path() on each
3083  * block and work out the exact number of indirects which are touched.  Pah.
3084  */
3085
3086 static int ext3_writepage_trans_blocks(struct inode *inode)
3087 {
3088         int bpp = ext3_journal_blocks_per_page(inode);
3089         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3090         int ret;
3091
3092         if (ext3_should_journal_data(inode))
3093                 ret = 3 * (bpp + indirects) + 2;
3094         else
3095                 ret = 2 * (bpp + indirects) + 2;
3096
3097 #ifdef CONFIG_QUOTA
3098         /* We know that structure was already allocated during DQUOT_INIT so
3099          * we will be updating only the data blocks + inodes */
3100         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3101 #endif
3102
3103         return ret;
3104 }
3105
3106 /*
3107  * The caller must have previously called ext3_reserve_inode_write().
3108  * Give this, we know that the caller already has write access to iloc->bh.
3109  */
3110 int ext3_mark_iloc_dirty(handle_t *handle,
3111                 struct inode *inode, struct ext3_iloc *iloc)
3112 {
3113         int err = 0;
3114
3115         /* the do_update_inode consumes one bh->b_count */
3116         get_bh(iloc->bh);
3117
3118         /* ext3_do_update_inode() does journal_dirty_metadata */
3119         err = ext3_do_update_inode(handle, inode, iloc);
3120         put_bh(iloc->bh);
3121         return err;
3122 }
3123
3124 /*
3125  * On success, We end up with an outstanding reference count against
3126  * iloc->bh.  This _must_ be cleaned up later.
3127  */
3128
3129 int
3130 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3131                          struct ext3_iloc *iloc)
3132 {
3133         int err = 0;
3134         if (handle) {
3135                 err = ext3_get_inode_loc(inode, iloc);
3136                 if (!err) {
3137                         BUFFER_TRACE(iloc->bh, "get_write_access");
3138                         err = ext3_journal_get_write_access(handle, iloc->bh);
3139                         if (err) {
3140                                 brelse(iloc->bh);
3141                                 iloc->bh = NULL;
3142                         }
3143                 }
3144         }
3145         ext3_std_error(inode->i_sb, err);
3146         return err;
3147 }
3148
3149 /*
3150  * What we do here is to mark the in-core inode as clean with respect to inode
3151  * dirtiness (it may still be data-dirty).
3152  * This means that the in-core inode may be reaped by prune_icache
3153  * without having to perform any I/O.  This is a very good thing,
3154  * because *any* task may call prune_icache - even ones which
3155  * have a transaction open against a different journal.
3156  *
3157  * Is this cheating?  Not really.  Sure, we haven't written the
3158  * inode out, but prune_icache isn't a user-visible syncing function.
3159  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3160  * we start and wait on commits.
3161  *
3162  * Is this efficient/effective?  Well, we're being nice to the system
3163  * by cleaning up our inodes proactively so they can be reaped
3164  * without I/O.  But we are potentially leaving up to five seconds'
3165  * worth of inodes floating about which prune_icache wants us to
3166  * write out.  One way to fix that would be to get prune_icache()
3167  * to do a write_super() to free up some memory.  It has the desired
3168  * effect.
3169  */
3170 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3171 {
3172         struct ext3_iloc iloc;
3173         int err;
3174
3175         might_sleep();
3176         err = ext3_reserve_inode_write(handle, inode, &iloc);
3177         if (!err)
3178                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3179         return err;
3180 }
3181
3182 /*
3183  * ext3_dirty_inode() is called from __mark_inode_dirty()
3184  *
3185  * We're really interested in the case where a file is being extended.
3186  * i_size has been changed by generic_commit_write() and we thus need
3187  * to include the updated inode in the current transaction.
3188  *
3189  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3190  * are allocated to the file.
3191  *
3192  * If the inode is marked synchronous, we don't honour that here - doing
3193  * so would cause a commit on atime updates, which we don't bother doing.
3194  * We handle synchronous inodes at the highest possible level.
3195  */
3196 void ext3_dirty_inode(struct inode *inode)
3197 {
3198         handle_t *current_handle = ext3_journal_current_handle();
3199         handle_t *handle;
3200
3201         handle = ext3_journal_start(inode, 2);
3202         if (IS_ERR(handle))
3203                 goto out;
3204         if (current_handle &&
3205                 current_handle->h_transaction != handle->h_transaction) {
3206                 /* This task has a transaction open against a different fs */
3207                 printk(KERN_EMERG "%s: transactions do not match!\n",
3208                        __FUNCTION__);
3209         } else {
3210                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3211                                 current_handle);
3212                 ext3_mark_inode_dirty(handle, inode);
3213         }
3214         ext3_journal_stop(handle);
3215 out:
3216         return;
3217 }
3218
3219 #if 0
3220 /*
3221  * Bind an inode's backing buffer_head into this transaction, to prevent
3222  * it from being flushed to disk early.  Unlike
3223  * ext3_reserve_inode_write, this leaves behind no bh reference and
3224  * returns no iloc structure, so the caller needs to repeat the iloc
3225  * lookup to mark the inode dirty later.
3226  */
3227 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3228 {
3229         struct ext3_iloc iloc;
3230
3231         int err = 0;
3232         if (handle) {
3233                 err = ext3_get_inode_loc(inode, &iloc);
3234                 if (!err) {
3235                         BUFFER_TRACE(iloc.bh, "get_write_access");
3236                         err = journal_get_write_access(handle, iloc.bh);
3237                         if (!err)
3238                                 err = ext3_journal_dirty_metadata(handle,
3239                                                                   iloc.bh);
3240                         brelse(iloc.bh);
3241                 }
3242         }
3243         ext3_std_error(inode->i_sb, err);
3244         return err;
3245 }
3246 #endif
3247
3248 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3249 {
3250         journal_t *journal;
3251         handle_t *handle;
3252         int err;
3253
3254         /*
3255          * We have to be very careful here: changing a data block's
3256          * journaling status dynamically is dangerous.  If we write a
3257          * data block to the journal, change the status and then delete
3258          * that block, we risk forgetting to revoke the old log record
3259          * from the journal and so a subsequent replay can corrupt data.
3260          * So, first we make sure that the journal is empty and that
3261          * nobody is changing anything.
3262          */
3263
3264         journal = EXT3_JOURNAL(inode);
3265         if (is_journal_aborted(journal))
3266                 return -EROFS;
3267
3268         journal_lock_updates(journal);
3269         journal_flush(journal);
3270
3271         /*
3272          * OK, there are no updates running now, and all cached data is
3273          * synced to disk.  We are now in a completely consistent state
3274          * which doesn't have anything in the journal, and we know that
3275          * no filesystem updates are running, so it is safe to modify
3276          * the inode's in-core data-journaling state flag now.
3277          */
3278
3279         if (val)
3280                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3281         else
3282                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3283         ext3_set_aops(inode);
3284
3285         journal_unlock_updates(journal);
3286
3287         /* Finally we can mark the inode as dirty. */
3288
3289         handle = ext3_journal_start(inode, 1);
3290         if (IS_ERR(handle))
3291                 return PTR_ERR(handle);
3292
3293         err = ext3_mark_inode_dirty(handle, inode);
3294         handle->h_sync = 1;
3295         ext3_journal_stop(handle);
3296         ext3_std_error(inode->i_sb, err);
3297
3298         return err;
3299 }