]> pilppa.org Git - linux-2.6-omap-h63xx.git/blob - fs/ext3/inode.c
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ieee1394...
[linux-2.6-omap-h63xx.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/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 <linux/fiemap.h>
40 #include <linux/namei.h>
41 #include "xattr.h"
42 #include "acl.h"
43
44 static int ext3_writepage_trans_blocks(struct inode *inode);
45
46 /*
47  * Test whether an inode is a fast symlink.
48  */
49 static int ext3_inode_is_fast_symlink(struct inode *inode)
50 {
51         int ea_blocks = EXT3_I(inode)->i_file_acl ?
52                 (inode->i_sb->s_blocksize >> 9) : 0;
53
54         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
55 }
56
57 /*
58  * The ext3 forget function must perform a revoke if we are freeing data
59  * which has been journaled.  Metadata (eg. indirect blocks) must be
60  * revoked in all cases.
61  *
62  * "bh" may be NULL: a metadata block may have been freed from memory
63  * but there may still be a record of it in the journal, and that record
64  * still needs to be revoked.
65  */
66 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
67                         struct buffer_head *bh, ext3_fsblk_t blocknr)
68 {
69         int err;
70
71         might_sleep();
72
73         BUFFER_TRACE(bh, "enter");
74
75         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
76                   "data mode %lx\n",
77                   bh, is_metadata, inode->i_mode,
78                   test_opt(inode->i_sb, DATA_FLAGS));
79
80         /* Never use the revoke function if we are doing full data
81          * journaling: there is no need to, and a V1 superblock won't
82          * support it.  Otherwise, only skip the revoke on un-journaled
83          * data blocks. */
84
85         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
86             (!is_metadata && !ext3_should_journal_data(inode))) {
87                 if (bh) {
88                         BUFFER_TRACE(bh, "call journal_forget");
89                         return ext3_journal_forget(handle, bh);
90                 }
91                 return 0;
92         }
93
94         /*
95          * data!=journal && (is_metadata || should_journal_data(inode))
96          */
97         BUFFER_TRACE(bh, "call ext3_journal_revoke");
98         err = ext3_journal_revoke(handle, blocknr, bh);
99         if (err)
100                 ext3_abort(inode->i_sb, __func__,
101                            "error %d when attempting revoke", err);
102         BUFFER_TRACE(bh, "exit");
103         return err;
104 }
105
106 /*
107  * Work out how many blocks we need to proceed with the next chunk of a
108  * truncate transaction.
109  */
110 static unsigned long blocks_for_truncate(struct inode *inode)
111 {
112         unsigned long needed;
113
114         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
115
116         /* Give ourselves just enough room to cope with inodes in which
117          * i_blocks is corrupt: we've seen disk corruptions in the past
118          * which resulted in random data in an inode which looked enough
119          * like a regular file for ext3 to try to delete it.  Things
120          * will go a bit crazy if that happens, but at least we should
121          * try not to panic the whole kernel. */
122         if (needed < 2)
123                 needed = 2;
124
125         /* But we need to bound the transaction so we don't overflow the
126          * journal. */
127         if (needed > EXT3_MAX_TRANS_DATA)
128                 needed = EXT3_MAX_TRANS_DATA;
129
130         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
131 }
132
133 /*
134  * Truncate transactions can be complex and absolutely huge.  So we need to
135  * be able to restart the transaction at a conventient checkpoint to make
136  * sure we don't overflow the journal.
137  *
138  * start_transaction gets us a new handle for a truncate transaction,
139  * and extend_transaction tries to extend the existing one a bit.  If
140  * extend fails, we need to propagate the failure up and restart the
141  * transaction in the top-level truncate loop. --sct
142  */
143 static handle_t *start_transaction(struct inode *inode)
144 {
145         handle_t *result;
146
147         result = ext3_journal_start(inode, blocks_for_truncate(inode));
148         if (!IS_ERR(result))
149                 return result;
150
151         ext3_std_error(inode->i_sb, PTR_ERR(result));
152         return result;
153 }
154
155 /*
156  * Try to extend this transaction for the purposes of truncation.
157  *
158  * Returns 0 if we managed to create more room.  If we can't create more
159  * room, and the transaction must be restarted we return 1.
160  */
161 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
162 {
163         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
164                 return 0;
165         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
166                 return 0;
167         return 1;
168 }
169
170 /*
171  * Restart the transaction associated with *handle.  This does a commit,
172  * so before we call here everything must be consistently dirtied against
173  * this transaction.
174  */
175 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
176 {
177         jbd_debug(2, "restarting handle %p\n", handle);
178         return ext3_journal_restart(handle, blocks_for_truncate(inode));
179 }
180
181 /*
182  * Called at the last iput() if i_nlink is zero.
183  */
184 void ext3_delete_inode (struct inode * inode)
185 {
186         handle_t *handle;
187
188         truncate_inode_pages(&inode->i_data, 0);
189
190         if (is_bad_inode(inode))
191                 goto no_delete;
192
193         handle = start_transaction(inode);
194         if (IS_ERR(handle)) {
195                 /*
196                  * If we're going to skip the normal cleanup, we still need to
197                  * make sure that the in-core orphan linked list is properly
198                  * cleaned up.
199                  */
200                 ext3_orphan_del(NULL, inode);
201                 goto no_delete;
202         }
203
204         if (IS_SYNC(inode))
205                 handle->h_sync = 1;
206         inode->i_size = 0;
207         if (inode->i_blocks)
208                 ext3_truncate(inode);
209         /*
210          * Kill off the orphan record which ext3_truncate created.
211          * AKPM: I think this can be inside the above `if'.
212          * Note that ext3_orphan_del() has to be able to cope with the
213          * deletion of a non-existent orphan - this is because we don't
214          * know if ext3_truncate() actually created an orphan record.
215          * (Well, we could do this if we need to, but heck - it works)
216          */
217         ext3_orphan_del(handle, inode);
218         EXT3_I(inode)->i_dtime  = get_seconds();
219
220         /*
221          * One subtle ordering requirement: if anything has gone wrong
222          * (transaction abort, IO errors, whatever), then we can still
223          * do these next steps (the fs will already have been marked as
224          * having errors), but we can't free the inode if the mark_dirty
225          * fails.
226          */
227         if (ext3_mark_inode_dirty(handle, inode))
228                 /* If that failed, just do the required in-core inode clear. */
229                 clear_inode(inode);
230         else
231                 ext3_free_inode(handle, inode);
232         ext3_journal_stop(handle);
233         return;
234 no_delete:
235         clear_inode(inode);     /* We must guarantee clearing of inode... */
236 }
237
238 typedef struct {
239         __le32  *p;
240         __le32  key;
241         struct buffer_head *bh;
242 } Indirect;
243
244 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
245 {
246         p->key = *(p->p = v);
247         p->bh = bh;
248 }
249
250 static int verify_chain(Indirect *from, Indirect *to)
251 {
252         while (from <= to && from->key == *from->p)
253                 from++;
254         return (from > to);
255 }
256
257 /**
258  *      ext3_block_to_path - parse the block number into array of offsets
259  *      @inode: inode in question (we are only interested in its superblock)
260  *      @i_block: block number to be parsed
261  *      @offsets: array to store the offsets in
262  *      @boundary: set this non-zero if the referred-to block is likely to be
263  *             followed (on disk) by an indirect block.
264  *
265  *      To store the locations of file's data ext3 uses a data structure common
266  *      for UNIX filesystems - tree of pointers anchored in the inode, with
267  *      data blocks at leaves and indirect blocks in intermediate nodes.
268  *      This function translates the block number into path in that tree -
269  *      return value is the path length and @offsets[n] is the offset of
270  *      pointer to (n+1)th node in the nth one. If @block is out of range
271  *      (negative or too large) warning is printed and zero returned.
272  *
273  *      Note: function doesn't find node addresses, so no IO is needed. All
274  *      we need to know is the capacity of indirect blocks (taken from the
275  *      inode->i_sb).
276  */
277
278 /*
279  * Portability note: the last comparison (check that we fit into triple
280  * indirect block) is spelled differently, because otherwise on an
281  * architecture with 32-bit longs and 8Kb pages we might get into trouble
282  * if our filesystem had 8Kb blocks. We might use long long, but that would
283  * kill us on x86. Oh, well, at least the sign propagation does not matter -
284  * i_block would have to be negative in the very beginning, so we would not
285  * get there at all.
286  */
287
288 static int ext3_block_to_path(struct inode *inode,
289                         long i_block, int offsets[4], int *boundary)
290 {
291         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
292         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
293         const long direct_blocks = EXT3_NDIR_BLOCKS,
294                 indirect_blocks = ptrs,
295                 double_blocks = (1 << (ptrs_bits * 2));
296         int n = 0;
297         int final = 0;
298
299         if (i_block < 0) {
300                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
301         } else if (i_block < direct_blocks) {
302                 offsets[n++] = i_block;
303                 final = direct_blocks;
304         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
305                 offsets[n++] = EXT3_IND_BLOCK;
306                 offsets[n++] = i_block;
307                 final = ptrs;
308         } else if ((i_block -= indirect_blocks) < double_blocks) {
309                 offsets[n++] = EXT3_DIND_BLOCK;
310                 offsets[n++] = i_block >> ptrs_bits;
311                 offsets[n++] = i_block & (ptrs - 1);
312                 final = ptrs;
313         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
314                 offsets[n++] = EXT3_TIND_BLOCK;
315                 offsets[n++] = i_block >> (ptrs_bits * 2);
316                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
317                 offsets[n++] = i_block & (ptrs - 1);
318                 final = ptrs;
319         } else {
320                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
321         }
322         if (boundary)
323                 *boundary = final - 1 - (i_block & (ptrs - 1));
324         return n;
325 }
326
327 /**
328  *      ext3_get_branch - read the chain of indirect blocks leading to data
329  *      @inode: inode in question
330  *      @depth: depth of the chain (1 - direct pointer, etc.)
331  *      @offsets: offsets of pointers in inode/indirect blocks
332  *      @chain: place to store the result
333  *      @err: here we store the error value
334  *
335  *      Function fills the array of triples <key, p, bh> and returns %NULL
336  *      if everything went OK or the pointer to the last filled triple
337  *      (incomplete one) otherwise. Upon the return chain[i].key contains
338  *      the number of (i+1)-th block in the chain (as it is stored in memory,
339  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
340  *      number (it points into struct inode for i==0 and into the bh->b_data
341  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
342  *      block for i>0 and NULL for i==0. In other words, it holds the block
343  *      numbers of the chain, addresses they were taken from (and where we can
344  *      verify that chain did not change) and buffer_heads hosting these
345  *      numbers.
346  *
347  *      Function stops when it stumbles upon zero pointer (absent block)
348  *              (pointer to last triple returned, *@err == 0)
349  *      or when it gets an IO error reading an indirect block
350  *              (ditto, *@err == -EIO)
351  *      or when it notices that chain had been changed while it was reading
352  *              (ditto, *@err == -EAGAIN)
353  *      or when it reads all @depth-1 indirect blocks successfully and finds
354  *      the whole chain, all way to the data (returns %NULL, *err == 0).
355  */
356 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
357                                  Indirect chain[4], int *err)
358 {
359         struct super_block *sb = inode->i_sb;
360         Indirect *p = chain;
361         struct buffer_head *bh;
362
363         *err = 0;
364         /* i_data is not going away, no lock needed */
365         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
366         if (!p->key)
367                 goto no_block;
368         while (--depth) {
369                 bh = sb_bread(sb, le32_to_cpu(p->key));
370                 if (!bh)
371                         goto failure;
372                 /* Reader: pointers */
373                 if (!verify_chain(chain, p))
374                         goto changed;
375                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
376                 /* Reader: end */
377                 if (!p->key)
378                         goto no_block;
379         }
380         return NULL;
381
382 changed:
383         brelse(bh);
384         *err = -EAGAIN;
385         goto no_block;
386 failure:
387         *err = -EIO;
388 no_block:
389         return p;
390 }
391
392 /**
393  *      ext3_find_near - find a place for allocation with sufficient locality
394  *      @inode: owner
395  *      @ind: descriptor of indirect block.
396  *
397  *      This function returns the preferred place for block allocation.
398  *      It is used when heuristic for sequential allocation fails.
399  *      Rules are:
400  *        + if there is a block to the left of our position - allocate near it.
401  *        + if pointer will live in indirect block - allocate near that block.
402  *        + if pointer will live in inode - allocate in the same
403  *          cylinder group.
404  *
405  * In the latter case we colour the starting block by the callers PID to
406  * prevent it from clashing with concurrent allocations for a different inode
407  * in the same block group.   The PID is used here so that functionally related
408  * files will be close-by on-disk.
409  *
410  *      Caller must make sure that @ind is valid and will stay that way.
411  */
412 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
413 {
414         struct ext3_inode_info *ei = EXT3_I(inode);
415         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
416         __le32 *p;
417         ext3_fsblk_t bg_start;
418         ext3_grpblk_t colour;
419
420         /* Try to find previous block */
421         for (p = ind->p - 1; p >= start; p--) {
422                 if (*p)
423                         return le32_to_cpu(*p);
424         }
425
426         /* No such thing, so let's try location of indirect block */
427         if (ind->bh)
428                 return ind->bh->b_blocknr;
429
430         /*
431          * It is going to be referred to from the inode itself? OK, just put it
432          * into the same cylinder group then.
433          */
434         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
435         colour = (current->pid % 16) *
436                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
437         return bg_start + colour;
438 }
439
440 /**
441  *      ext3_find_goal - find a preferred place for allocation.
442  *      @inode: owner
443  *      @block:  block we want
444  *      @partial: pointer to the last triple within a chain
445  *
446  *      Normally this function find the preferred place for block allocation,
447  *      returns it.
448  */
449
450 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
451                                    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, 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 /* Maximum number of blocks we map for direct IO at once. */
945 #define DIO_MAX_BLOCKS 4096
946 /*
947  * Number of credits we need for writing DIO_MAX_BLOCKS:
948  * We need sb + group descriptor + bitmap + inode -> 4
949  * For B blocks with A block pointers per block we need:
950  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
951  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
952  */
953 #define DIO_CREDITS 25
954
955 static int ext3_get_block(struct inode *inode, sector_t iblock,
956                         struct buffer_head *bh_result, int create)
957 {
958         handle_t *handle = ext3_journal_current_handle();
959         int ret = 0, started = 0;
960         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
961
962         if (create && !handle) {        /* Direct IO write... */
963                 if (max_blocks > DIO_MAX_BLOCKS)
964                         max_blocks = DIO_MAX_BLOCKS;
965                 handle = ext3_journal_start(inode, DIO_CREDITS +
966                                 2 * EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb));
967                 if (IS_ERR(handle)) {
968                         ret = PTR_ERR(handle);
969                         goto out;
970                 }
971                 started = 1;
972         }
973
974         ret = ext3_get_blocks_handle(handle, inode, iblock,
975                                         max_blocks, bh_result, create, 0);
976         if (ret > 0) {
977                 bh_result->b_size = (ret << inode->i_blkbits);
978                 ret = 0;
979         }
980         if (started)
981                 ext3_journal_stop(handle);
982 out:
983         return ret;
984 }
985
986 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
987                 u64 start, u64 len)
988 {
989         return generic_block_fiemap(inode, fieinfo, start, len,
990                                     ext3_get_block);
991 }
992
993 /*
994  * `handle' can be NULL if create is zero
995  */
996 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
997                                 long block, int create, int *errp)
998 {
999         struct buffer_head dummy;
1000         int fatal = 0, err;
1001
1002         J_ASSERT(handle != NULL || create == 0);
1003
1004         dummy.b_state = 0;
1005         dummy.b_blocknr = -1000;
1006         buffer_trace_init(&dummy.b_history);
1007         err = ext3_get_blocks_handle(handle, inode, block, 1,
1008                                         &dummy, create, 1);
1009         /*
1010          * ext3_get_blocks_handle() returns number of blocks
1011          * mapped. 0 in case of a HOLE.
1012          */
1013         if (err > 0) {
1014                 if (err > 1)
1015                         WARN_ON(1);
1016                 err = 0;
1017         }
1018         *errp = err;
1019         if (!err && buffer_mapped(&dummy)) {
1020                 struct buffer_head *bh;
1021                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1022                 if (!bh) {
1023                         *errp = -EIO;
1024                         goto err;
1025                 }
1026                 if (buffer_new(&dummy)) {
1027                         J_ASSERT(create != 0);
1028                         J_ASSERT(handle != NULL);
1029
1030                         /*
1031                          * Now that we do not always journal data, we should
1032                          * keep in mind whether this should always journal the
1033                          * new buffer as metadata.  For now, regular file
1034                          * writes use ext3_get_block instead, so it's not a
1035                          * problem.
1036                          */
1037                         lock_buffer(bh);
1038                         BUFFER_TRACE(bh, "call get_create_access");
1039                         fatal = ext3_journal_get_create_access(handle, bh);
1040                         if (!fatal && !buffer_uptodate(bh)) {
1041                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1042                                 set_buffer_uptodate(bh);
1043                         }
1044                         unlock_buffer(bh);
1045                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1046                         err = ext3_journal_dirty_metadata(handle, bh);
1047                         if (!fatal)
1048                                 fatal = err;
1049                 } else {
1050                         BUFFER_TRACE(bh, "not a new buffer");
1051                 }
1052                 if (fatal) {
1053                         *errp = fatal;
1054                         brelse(bh);
1055                         bh = NULL;
1056                 }
1057                 return bh;
1058         }
1059 err:
1060         return NULL;
1061 }
1062
1063 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1064                                int block, int create, int *err)
1065 {
1066         struct buffer_head * bh;
1067
1068         bh = ext3_getblk(handle, inode, block, create, err);
1069         if (!bh)
1070                 return bh;
1071         if (buffer_uptodate(bh))
1072                 return bh;
1073         ll_rw_block(READ_META, 1, &bh);
1074         wait_on_buffer(bh);
1075         if (buffer_uptodate(bh))
1076                 return bh;
1077         put_bh(bh);
1078         *err = -EIO;
1079         return NULL;
1080 }
1081
1082 static int walk_page_buffers(   handle_t *handle,
1083                                 struct buffer_head *head,
1084                                 unsigned from,
1085                                 unsigned to,
1086                                 int *partial,
1087                                 int (*fn)(      handle_t *handle,
1088                                                 struct buffer_head *bh))
1089 {
1090         struct buffer_head *bh;
1091         unsigned block_start, block_end;
1092         unsigned blocksize = head->b_size;
1093         int err, ret = 0;
1094         struct buffer_head *next;
1095
1096         for (   bh = head, block_start = 0;
1097                 ret == 0 && (bh != head || !block_start);
1098                 block_start = block_end, bh = next)
1099         {
1100                 next = bh->b_this_page;
1101                 block_end = block_start + blocksize;
1102                 if (block_end <= from || block_start >= to) {
1103                         if (partial && !buffer_uptodate(bh))
1104                                 *partial = 1;
1105                         continue;
1106                 }
1107                 err = (*fn)(handle, bh);
1108                 if (!ret)
1109                         ret = err;
1110         }
1111         return ret;
1112 }
1113
1114 /*
1115  * To preserve ordering, it is essential that the hole instantiation and
1116  * the data write be encapsulated in a single transaction.  We cannot
1117  * close off a transaction and start a new one between the ext3_get_block()
1118  * and the commit_write().  So doing the journal_start at the start of
1119  * prepare_write() is the right place.
1120  *
1121  * Also, this function can nest inside ext3_writepage() ->
1122  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1123  * has generated enough buffer credits to do the whole page.  So we won't
1124  * block on the journal in that case, which is good, because the caller may
1125  * be PF_MEMALLOC.
1126  *
1127  * By accident, ext3 can be reentered when a transaction is open via
1128  * quota file writes.  If we were to commit the transaction while thus
1129  * reentered, there can be a deadlock - we would be holding a quota
1130  * lock, and the commit would never complete if another thread had a
1131  * transaction open and was blocking on the quota lock - a ranking
1132  * violation.
1133  *
1134  * So what we do is to rely on the fact that journal_stop/journal_start
1135  * will _not_ run commit under these circumstances because handle->h_ref
1136  * is elevated.  We'll still have enough credits for the tiny quotafile
1137  * write.
1138  */
1139 static int do_journal_get_write_access(handle_t *handle,
1140                                         struct buffer_head *bh)
1141 {
1142         if (!buffer_mapped(bh) || buffer_freed(bh))
1143                 return 0;
1144         return ext3_journal_get_write_access(handle, bh);
1145 }
1146
1147 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1148                                 loff_t pos, unsigned len, unsigned flags,
1149                                 struct page **pagep, void **fsdata)
1150 {
1151         struct inode *inode = mapping->host;
1152         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1153         handle_t *handle;
1154         int retries = 0;
1155         struct page *page;
1156         pgoff_t index;
1157         unsigned from, to;
1158
1159         index = pos >> PAGE_CACHE_SHIFT;
1160         from = pos & (PAGE_CACHE_SIZE - 1);
1161         to = from + len;
1162
1163 retry:
1164         page = grab_cache_page_write_begin(mapping, index, flags);
1165         if (!page)
1166                 return -ENOMEM;
1167         *pagep = page;
1168
1169         handle = ext3_journal_start(inode, needed_blocks);
1170         if (IS_ERR(handle)) {
1171                 unlock_page(page);
1172                 page_cache_release(page);
1173                 ret = PTR_ERR(handle);
1174                 goto out;
1175         }
1176         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1177                                                         ext3_get_block);
1178         if (ret)
1179                 goto write_begin_failed;
1180
1181         if (ext3_should_journal_data(inode)) {
1182                 ret = walk_page_buffers(handle, page_buffers(page),
1183                                 from, to, NULL, do_journal_get_write_access);
1184         }
1185 write_begin_failed:
1186         if (ret) {
1187                 ext3_journal_stop(handle);
1188                 unlock_page(page);
1189                 page_cache_release(page);
1190                 /*
1191                  * block_write_begin may have instantiated a few blocks
1192                  * outside i_size.  Trim these off again. Don't need
1193                  * i_size_read because we hold i_mutex.
1194                  */
1195                 if (pos + len > inode->i_size)
1196                         vmtruncate(inode, inode->i_size);
1197         }
1198         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1199                 goto retry;
1200 out:
1201         return ret;
1202 }
1203
1204
1205 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1206 {
1207         int err = journal_dirty_data(handle, bh);
1208         if (err)
1209                 ext3_journal_abort_handle(__func__, __func__,
1210                                                 bh, handle, err);
1211         return err;
1212 }
1213
1214 /* For write_end() in data=journal mode */
1215 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1216 {
1217         if (!buffer_mapped(bh) || buffer_freed(bh))
1218                 return 0;
1219         set_buffer_uptodate(bh);
1220         return ext3_journal_dirty_metadata(handle, bh);
1221 }
1222
1223 /*
1224  * Generic write_end handler for ordered and writeback ext3 journal modes.
1225  * We can't use generic_write_end, because that unlocks the page and we need to
1226  * unlock the page after ext3_journal_stop, but ext3_journal_stop must run
1227  * after block_write_end.
1228  */
1229 static int ext3_generic_write_end(struct file *file,
1230                                 struct address_space *mapping,
1231                                 loff_t pos, unsigned len, unsigned copied,
1232                                 struct page *page, void *fsdata)
1233 {
1234         struct inode *inode = file->f_mapping->host;
1235
1236         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1237
1238         if (pos+copied > inode->i_size) {
1239                 i_size_write(inode, pos+copied);
1240                 mark_inode_dirty(inode);
1241         }
1242
1243         return copied;
1244 }
1245
1246 /*
1247  * We need to pick up the new inode size which generic_commit_write gave us
1248  * `file' can be NULL - eg, when called from page_symlink().
1249  *
1250  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1251  * buffers are managed internally.
1252  */
1253 static int ext3_ordered_write_end(struct file *file,
1254                                 struct address_space *mapping,
1255                                 loff_t pos, unsigned len, unsigned copied,
1256                                 struct page *page, void *fsdata)
1257 {
1258         handle_t *handle = ext3_journal_current_handle();
1259         struct inode *inode = file->f_mapping->host;
1260         unsigned from, to;
1261         int ret = 0, ret2;
1262
1263         from = pos & (PAGE_CACHE_SIZE - 1);
1264         to = from + len;
1265
1266         ret = walk_page_buffers(handle, page_buffers(page),
1267                 from, to, NULL, ext3_journal_dirty_data);
1268
1269         if (ret == 0) {
1270                 /*
1271                  * generic_write_end() will run mark_inode_dirty() if i_size
1272                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1273                  * into that.
1274                  */
1275                 loff_t new_i_size;
1276
1277                 new_i_size = pos + copied;
1278                 if (new_i_size > EXT3_I(inode)->i_disksize)
1279                         EXT3_I(inode)->i_disksize = new_i_size;
1280                 ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1281                                                         page, fsdata);
1282                 copied = ret2;
1283                 if (ret2 < 0)
1284                         ret = ret2;
1285         }
1286         ret2 = ext3_journal_stop(handle);
1287         if (!ret)
1288                 ret = ret2;
1289         unlock_page(page);
1290         page_cache_release(page);
1291
1292         return ret ? ret : copied;
1293 }
1294
1295 static int ext3_writeback_write_end(struct file *file,
1296                                 struct address_space *mapping,
1297                                 loff_t pos, unsigned len, unsigned copied,
1298                                 struct page *page, void *fsdata)
1299 {
1300         handle_t *handle = ext3_journal_current_handle();
1301         struct inode *inode = file->f_mapping->host;
1302         int ret = 0, ret2;
1303         loff_t new_i_size;
1304
1305         new_i_size = pos + copied;
1306         if (new_i_size > EXT3_I(inode)->i_disksize)
1307                 EXT3_I(inode)->i_disksize = new_i_size;
1308
1309         ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1310                                                         page, fsdata);
1311         copied = ret2;
1312         if (ret2 < 0)
1313                 ret = ret2;
1314
1315         ret2 = ext3_journal_stop(handle);
1316         if (!ret)
1317                 ret = ret2;
1318         unlock_page(page);
1319         page_cache_release(page);
1320
1321         return ret ? ret : copied;
1322 }
1323
1324 static int ext3_journalled_write_end(struct file *file,
1325                                 struct address_space *mapping,
1326                                 loff_t pos, unsigned len, unsigned copied,
1327                                 struct page *page, void *fsdata)
1328 {
1329         handle_t *handle = ext3_journal_current_handle();
1330         struct inode *inode = mapping->host;
1331         int ret = 0, ret2;
1332         int partial = 0;
1333         unsigned from, to;
1334
1335         from = pos & (PAGE_CACHE_SIZE - 1);
1336         to = from + len;
1337
1338         if (copied < len) {
1339                 if (!PageUptodate(page))
1340                         copied = 0;
1341                 page_zero_new_buffers(page, from+copied, to);
1342         }
1343
1344         ret = walk_page_buffers(handle, page_buffers(page), from,
1345                                 to, &partial, write_end_fn);
1346         if (!partial)
1347                 SetPageUptodate(page);
1348         if (pos+copied > inode->i_size)
1349                 i_size_write(inode, pos+copied);
1350         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1351         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1352                 EXT3_I(inode)->i_disksize = inode->i_size;
1353                 ret2 = ext3_mark_inode_dirty(handle, inode);
1354                 if (!ret)
1355                         ret = ret2;
1356         }
1357
1358         ret2 = ext3_journal_stop(handle);
1359         if (!ret)
1360                 ret = ret2;
1361         unlock_page(page);
1362         page_cache_release(page);
1363
1364         return ret ? ret : copied;
1365 }
1366
1367 /*
1368  * bmap() is special.  It gets used by applications such as lilo and by
1369  * the swapper to find the on-disk block of a specific piece of data.
1370  *
1371  * Naturally, this is dangerous if the block concerned is still in the
1372  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1373  * filesystem and enables swap, then they may get a nasty shock when the
1374  * data getting swapped to that swapfile suddenly gets overwritten by
1375  * the original zero's written out previously to the journal and
1376  * awaiting writeback in the kernel's buffer cache.
1377  *
1378  * So, if we see any bmap calls here on a modified, data-journaled file,
1379  * take extra steps to flush any blocks which might be in the cache.
1380  */
1381 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1382 {
1383         struct inode *inode = mapping->host;
1384         journal_t *journal;
1385         int err;
1386
1387         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1388                 /*
1389                  * This is a REALLY heavyweight approach, but the use of
1390                  * bmap on dirty files is expected to be extremely rare:
1391                  * only if we run lilo or swapon on a freshly made file
1392                  * do we expect this to happen.
1393                  *
1394                  * (bmap requires CAP_SYS_RAWIO so this does not
1395                  * represent an unprivileged user DOS attack --- we'd be
1396                  * in trouble if mortal users could trigger this path at
1397                  * will.)
1398                  *
1399                  * NB. EXT3_STATE_JDATA is not set on files other than
1400                  * regular files.  If somebody wants to bmap a directory
1401                  * or symlink and gets confused because the buffer
1402                  * hasn't yet been flushed to disk, they deserve
1403                  * everything they get.
1404                  */
1405
1406                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1407                 journal = EXT3_JOURNAL(inode);
1408                 journal_lock_updates(journal);
1409                 err = journal_flush(journal);
1410                 journal_unlock_updates(journal);
1411
1412                 if (err)
1413                         return 0;
1414         }
1415
1416         return generic_block_bmap(mapping,block,ext3_get_block);
1417 }
1418
1419 static int bget_one(handle_t *handle, struct buffer_head *bh)
1420 {
1421         get_bh(bh);
1422         return 0;
1423 }
1424
1425 static int bput_one(handle_t *handle, struct buffer_head *bh)
1426 {
1427         put_bh(bh);
1428         return 0;
1429 }
1430
1431 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1432 {
1433         if (buffer_mapped(bh))
1434                 return ext3_journal_dirty_data(handle, bh);
1435         return 0;
1436 }
1437
1438 /*
1439  * Note that we always start a transaction even if we're not journalling
1440  * data.  This is to preserve ordering: any hole instantiation within
1441  * __block_write_full_page -> ext3_get_block() should be journalled
1442  * along with the data so we don't crash and then get metadata which
1443  * refers to old data.
1444  *
1445  * In all journalling modes block_write_full_page() will start the I/O.
1446  *
1447  * Problem:
1448  *
1449  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1450  *              ext3_writepage()
1451  *
1452  * Similar for:
1453  *
1454  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1455  *
1456  * Same applies to ext3_get_block().  We will deadlock on various things like
1457  * lock_journal and i_truncate_mutex.
1458  *
1459  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1460  * allocations fail.
1461  *
1462  * 16May01: If we're reentered then journal_current_handle() will be
1463  *          non-zero. We simply *return*.
1464  *
1465  * 1 July 2001: @@@ FIXME:
1466  *   In journalled data mode, a data buffer may be metadata against the
1467  *   current transaction.  But the same file is part of a shared mapping
1468  *   and someone does a writepage() on it.
1469  *
1470  *   We will move the buffer onto the async_data list, but *after* it has
1471  *   been dirtied. So there's a small window where we have dirty data on
1472  *   BJ_Metadata.
1473  *
1474  *   Note that this only applies to the last partial page in the file.  The
1475  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1476  *   broken code anyway: it's wrong for msync()).
1477  *
1478  *   It's a rare case: affects the final partial page, for journalled data
1479  *   where the file is subject to bith write() and writepage() in the same
1480  *   transction.  To fix it we'll need a custom block_write_full_page().
1481  *   We'll probably need that anyway for journalling writepage() output.
1482  *
1483  * We don't honour synchronous mounts for writepage().  That would be
1484  * disastrous.  Any write() or metadata operation will sync the fs for
1485  * us.
1486  *
1487  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1488  * we don't need to open a transaction here.
1489  */
1490 static int ext3_ordered_writepage(struct page *page,
1491                                 struct writeback_control *wbc)
1492 {
1493         struct inode *inode = page->mapping->host;
1494         struct buffer_head *page_bufs;
1495         handle_t *handle = NULL;
1496         int ret = 0;
1497         int err;
1498
1499         J_ASSERT(PageLocked(page));
1500
1501         /*
1502          * We give up here if we're reentered, because it might be for a
1503          * different filesystem.
1504          */
1505         if (ext3_journal_current_handle())
1506                 goto out_fail;
1507
1508         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1509
1510         if (IS_ERR(handle)) {
1511                 ret = PTR_ERR(handle);
1512                 goto out_fail;
1513         }
1514
1515         if (!page_has_buffers(page)) {
1516                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1517                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1518         }
1519         page_bufs = page_buffers(page);
1520         walk_page_buffers(handle, page_bufs, 0,
1521                         PAGE_CACHE_SIZE, NULL, bget_one);
1522
1523         ret = block_write_full_page(page, ext3_get_block, wbc);
1524
1525         /*
1526          * The page can become unlocked at any point now, and
1527          * truncate can then come in and change things.  So we
1528          * can't touch *page from now on.  But *page_bufs is
1529          * safe due to elevated refcount.
1530          */
1531
1532         /*
1533          * And attach them to the current transaction.  But only if
1534          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1535          * and generally junk.
1536          */
1537         if (ret == 0) {
1538                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1539                                         NULL, journal_dirty_data_fn);
1540                 if (!ret)
1541                         ret = err;
1542         }
1543         walk_page_buffers(handle, page_bufs, 0,
1544                         PAGE_CACHE_SIZE, NULL, bput_one);
1545         err = ext3_journal_stop(handle);
1546         if (!ret)
1547                 ret = err;
1548         return ret;
1549
1550 out_fail:
1551         redirty_page_for_writepage(wbc, page);
1552         unlock_page(page);
1553         return ret;
1554 }
1555
1556 static int ext3_writeback_writepage(struct page *page,
1557                                 struct writeback_control *wbc)
1558 {
1559         struct inode *inode = page->mapping->host;
1560         handle_t *handle = NULL;
1561         int ret = 0;
1562         int err;
1563
1564         if (ext3_journal_current_handle())
1565                 goto out_fail;
1566
1567         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1568         if (IS_ERR(handle)) {
1569                 ret = PTR_ERR(handle);
1570                 goto out_fail;
1571         }
1572
1573         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1574                 ret = nobh_writepage(page, ext3_get_block, wbc);
1575         else
1576                 ret = block_write_full_page(page, ext3_get_block, wbc);
1577
1578         err = ext3_journal_stop(handle);
1579         if (!ret)
1580                 ret = err;
1581         return ret;
1582
1583 out_fail:
1584         redirty_page_for_writepage(wbc, page);
1585         unlock_page(page);
1586         return ret;
1587 }
1588
1589 static int ext3_journalled_writepage(struct page *page,
1590                                 struct writeback_control *wbc)
1591 {
1592         struct inode *inode = page->mapping->host;
1593         handle_t *handle = NULL;
1594         int ret = 0;
1595         int err;
1596
1597         if (ext3_journal_current_handle())
1598                 goto no_write;
1599
1600         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1601         if (IS_ERR(handle)) {
1602                 ret = PTR_ERR(handle);
1603                 goto no_write;
1604         }
1605
1606         if (!page_has_buffers(page) || PageChecked(page)) {
1607                 /*
1608                  * It's mmapped pagecache.  Add buffers and journal it.  There
1609                  * doesn't seem much point in redirtying the page here.
1610                  */
1611                 ClearPageChecked(page);
1612                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1613                                         ext3_get_block);
1614                 if (ret != 0) {
1615                         ext3_journal_stop(handle);
1616                         goto out_unlock;
1617                 }
1618                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1619                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1620
1621                 err = walk_page_buffers(handle, page_buffers(page), 0,
1622                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1623                 if (ret == 0)
1624                         ret = err;
1625                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1626                 unlock_page(page);
1627         } else {
1628                 /*
1629                  * It may be a page full of checkpoint-mode buffers.  We don't
1630                  * really know unless we go poke around in the buffer_heads.
1631                  * But block_write_full_page will do the right thing.
1632                  */
1633                 ret = block_write_full_page(page, ext3_get_block, wbc);
1634         }
1635         err = ext3_journal_stop(handle);
1636         if (!ret)
1637                 ret = err;
1638 out:
1639         return ret;
1640
1641 no_write:
1642         redirty_page_for_writepage(wbc, page);
1643 out_unlock:
1644         unlock_page(page);
1645         goto out;
1646 }
1647
1648 static int ext3_readpage(struct file *file, struct page *page)
1649 {
1650         return mpage_readpage(page, ext3_get_block);
1651 }
1652
1653 static int
1654 ext3_readpages(struct file *file, struct address_space *mapping,
1655                 struct list_head *pages, unsigned nr_pages)
1656 {
1657         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1658 }
1659
1660 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1661 {
1662         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1663
1664         /*
1665          * If it's a full truncate we just forget about the pending dirtying
1666          */
1667         if (offset == 0)
1668                 ClearPageChecked(page);
1669
1670         journal_invalidatepage(journal, page, offset);
1671 }
1672
1673 static int ext3_releasepage(struct page *page, gfp_t wait)
1674 {
1675         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1676
1677         WARN_ON(PageChecked(page));
1678         if (!page_has_buffers(page))
1679                 return 0;
1680         return journal_try_to_free_buffers(journal, page, wait);
1681 }
1682
1683 /*
1684  * If the O_DIRECT write will extend the file then add this inode to the
1685  * orphan list.  So recovery will truncate it back to the original size
1686  * if the machine crashes during the write.
1687  *
1688  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1689  * crashes then stale disk data _may_ be exposed inside the file. But current
1690  * VFS code falls back into buffered path in that case so we are safe.
1691  */
1692 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1693                         const struct iovec *iov, loff_t offset,
1694                         unsigned long nr_segs)
1695 {
1696         struct file *file = iocb->ki_filp;
1697         struct inode *inode = file->f_mapping->host;
1698         struct ext3_inode_info *ei = EXT3_I(inode);
1699         handle_t *handle;
1700         ssize_t ret;
1701         int orphan = 0;
1702         size_t count = iov_length(iov, nr_segs);
1703
1704         if (rw == WRITE) {
1705                 loff_t final_size = offset + count;
1706
1707                 if (final_size > inode->i_size) {
1708                         /* Credits for sb + inode write */
1709                         handle = ext3_journal_start(inode, 2);
1710                         if (IS_ERR(handle)) {
1711                                 ret = PTR_ERR(handle);
1712                                 goto out;
1713                         }
1714                         ret = ext3_orphan_add(handle, inode);
1715                         if (ret) {
1716                                 ext3_journal_stop(handle);
1717                                 goto out;
1718                         }
1719                         orphan = 1;
1720                         ei->i_disksize = inode->i_size;
1721                         ext3_journal_stop(handle);
1722                 }
1723         }
1724
1725         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1726                                  offset, nr_segs,
1727                                  ext3_get_block, NULL);
1728
1729         if (orphan) {
1730                 int err;
1731
1732                 /* Credits for sb + inode write */
1733                 handle = ext3_journal_start(inode, 2);
1734                 if (IS_ERR(handle)) {
1735                         /* This is really bad luck. We've written the data
1736                          * but cannot extend i_size. Bail out and pretend
1737                          * the write failed... */
1738                         ret = PTR_ERR(handle);
1739                         goto out;
1740                 }
1741                 if (inode->i_nlink)
1742                         ext3_orphan_del(handle, inode);
1743                 if (ret > 0) {
1744                         loff_t end = offset + ret;
1745                         if (end > inode->i_size) {
1746                                 ei->i_disksize = end;
1747                                 i_size_write(inode, end);
1748                                 /*
1749                                  * We're going to return a positive `ret'
1750                                  * here due to non-zero-length I/O, so there's
1751                                  * no way of reporting error returns from
1752                                  * ext3_mark_inode_dirty() to userspace.  So
1753                                  * ignore it.
1754                                  */
1755                                 ext3_mark_inode_dirty(handle, inode);
1756                         }
1757                 }
1758                 err = ext3_journal_stop(handle);
1759                 if (ret == 0)
1760                         ret = err;
1761         }
1762 out:
1763         return ret;
1764 }
1765
1766 /*
1767  * Pages can be marked dirty completely asynchronously from ext3's journalling
1768  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1769  * much here because ->set_page_dirty is called under VFS locks.  The page is
1770  * not necessarily locked.
1771  *
1772  * We cannot just dirty the page and leave attached buffers clean, because the
1773  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1774  * or jbddirty because all the journalling code will explode.
1775  *
1776  * So what we do is to mark the page "pending dirty" and next time writepage
1777  * is called, propagate that into the buffers appropriately.
1778  */
1779 static int ext3_journalled_set_page_dirty(struct page *page)
1780 {
1781         SetPageChecked(page);
1782         return __set_page_dirty_nobuffers(page);
1783 }
1784
1785 static const struct address_space_operations ext3_ordered_aops = {
1786         .readpage               = ext3_readpage,
1787         .readpages              = ext3_readpages,
1788         .writepage              = ext3_ordered_writepage,
1789         .sync_page              = block_sync_page,
1790         .write_begin            = ext3_write_begin,
1791         .write_end              = ext3_ordered_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         .is_partially_uptodate  = block_is_partially_uptodate,
1798 };
1799
1800 static const struct address_space_operations ext3_writeback_aops = {
1801         .readpage               = ext3_readpage,
1802         .readpages              = ext3_readpages,
1803         .writepage              = ext3_writeback_writepage,
1804         .sync_page              = block_sync_page,
1805         .write_begin            = ext3_write_begin,
1806         .write_end              = ext3_writeback_write_end,
1807         .bmap                   = ext3_bmap,
1808         .invalidatepage         = ext3_invalidatepage,
1809         .releasepage            = ext3_releasepage,
1810         .direct_IO              = ext3_direct_IO,
1811         .migratepage            = buffer_migrate_page,
1812         .is_partially_uptodate  = block_is_partially_uptodate,
1813 };
1814
1815 static const struct address_space_operations ext3_journalled_aops = {
1816         .readpage               = ext3_readpage,
1817         .readpages              = ext3_readpages,
1818         .writepage              = ext3_journalled_writepage,
1819         .sync_page              = block_sync_page,
1820         .write_begin            = ext3_write_begin,
1821         .write_end              = ext3_journalled_write_end,
1822         .set_page_dirty         = ext3_journalled_set_page_dirty,
1823         .bmap                   = ext3_bmap,
1824         .invalidatepage         = ext3_invalidatepage,
1825         .releasepage            = ext3_releasepage,
1826         .is_partially_uptodate  = block_is_partially_uptodate,
1827 };
1828
1829 void ext3_set_aops(struct inode *inode)
1830 {
1831         if (ext3_should_order_data(inode))
1832                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1833         else if (ext3_should_writeback_data(inode))
1834                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1835         else
1836                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1837 }
1838
1839 /*
1840  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1841  * up to the end of the block which corresponds to `from'.
1842  * This required during truncate. We need to physically zero the tail end
1843  * of that block so it doesn't yield old data if the file is later grown.
1844  */
1845 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1846                 struct address_space *mapping, loff_t from)
1847 {
1848         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1849         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1850         unsigned blocksize, iblock, length, pos;
1851         struct inode *inode = mapping->host;
1852         struct buffer_head *bh;
1853         int err = 0;
1854
1855         blocksize = inode->i_sb->s_blocksize;
1856         length = blocksize - (offset & (blocksize - 1));
1857         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1858
1859         /*
1860          * For "nobh" option,  we can only work if we don't need to
1861          * read-in the page - otherwise we create buffers to do the IO.
1862          */
1863         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1864              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1865                 zero_user(page, offset, length);
1866                 set_page_dirty(page);
1867                 goto unlock;
1868         }
1869
1870         if (!page_has_buffers(page))
1871                 create_empty_buffers(page, blocksize, 0);
1872
1873         /* Find the buffer that contains "offset" */
1874         bh = page_buffers(page);
1875         pos = blocksize;
1876         while (offset >= pos) {
1877                 bh = bh->b_this_page;
1878                 iblock++;
1879                 pos += blocksize;
1880         }
1881
1882         err = 0;
1883         if (buffer_freed(bh)) {
1884                 BUFFER_TRACE(bh, "freed: skip");
1885                 goto unlock;
1886         }
1887
1888         if (!buffer_mapped(bh)) {
1889                 BUFFER_TRACE(bh, "unmapped");
1890                 ext3_get_block(inode, iblock, bh, 0);
1891                 /* unmapped? It's a hole - nothing to do */
1892                 if (!buffer_mapped(bh)) {
1893                         BUFFER_TRACE(bh, "still unmapped");
1894                         goto unlock;
1895                 }
1896         }
1897
1898         /* Ok, it's mapped. Make sure it's up-to-date */
1899         if (PageUptodate(page))
1900                 set_buffer_uptodate(bh);
1901
1902         if (!buffer_uptodate(bh)) {
1903                 err = -EIO;
1904                 ll_rw_block(READ, 1, &bh);
1905                 wait_on_buffer(bh);
1906                 /* Uhhuh. Read error. Complain and punt. */
1907                 if (!buffer_uptodate(bh))
1908                         goto unlock;
1909         }
1910
1911         if (ext3_should_journal_data(inode)) {
1912                 BUFFER_TRACE(bh, "get write access");
1913                 err = ext3_journal_get_write_access(handle, bh);
1914                 if (err)
1915                         goto unlock;
1916         }
1917
1918         zero_user(page, offset, length);
1919         BUFFER_TRACE(bh, "zeroed end of block");
1920
1921         err = 0;
1922         if (ext3_should_journal_data(inode)) {
1923                 err = ext3_journal_dirty_metadata(handle, bh);
1924         } else {
1925                 if (ext3_should_order_data(inode))
1926                         err = ext3_journal_dirty_data(handle, bh);
1927                 mark_buffer_dirty(bh);
1928         }
1929
1930 unlock:
1931         unlock_page(page);
1932         page_cache_release(page);
1933         return err;
1934 }
1935
1936 /*
1937  * Probably it should be a library function... search for first non-zero word
1938  * or memcmp with zero_page, whatever is better for particular architecture.
1939  * Linus?
1940  */
1941 static inline int all_zeroes(__le32 *p, __le32 *q)
1942 {
1943         while (p < q)
1944                 if (*p++)
1945                         return 0;
1946         return 1;
1947 }
1948
1949 /**
1950  *      ext3_find_shared - find the indirect blocks for partial truncation.
1951  *      @inode:   inode in question
1952  *      @depth:   depth of the affected branch
1953  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1954  *      @chain:   place to store the pointers to partial indirect blocks
1955  *      @top:     place to the (detached) top of branch
1956  *
1957  *      This is a helper function used by ext3_truncate().
1958  *
1959  *      When we do truncate() we may have to clean the ends of several
1960  *      indirect blocks but leave the blocks themselves alive. Block is
1961  *      partially truncated if some data below the new i_size is refered
1962  *      from it (and it is on the path to the first completely truncated
1963  *      data block, indeed).  We have to free the top of that path along
1964  *      with everything to the right of the path. Since no allocation
1965  *      past the truncation point is possible until ext3_truncate()
1966  *      finishes, we may safely do the latter, but top of branch may
1967  *      require special attention - pageout below the truncation point
1968  *      might try to populate it.
1969  *
1970  *      We atomically detach the top of branch from the tree, store the
1971  *      block number of its root in *@top, pointers to buffer_heads of
1972  *      partially truncated blocks - in @chain[].bh and pointers to
1973  *      their last elements that should not be removed - in
1974  *      @chain[].p. Return value is the pointer to last filled element
1975  *      of @chain.
1976  *
1977  *      The work left to caller to do the actual freeing of subtrees:
1978  *              a) free the subtree starting from *@top
1979  *              b) free the subtrees whose roots are stored in
1980  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1981  *              c) free the subtrees growing from the inode past the @chain[0].
1982  *                      (no partially truncated stuff there).  */
1983
1984 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1985                         int offsets[4], Indirect chain[4], __le32 *top)
1986 {
1987         Indirect *partial, *p;
1988         int k, err;
1989
1990         *top = 0;
1991         /* Make k index the deepest non-null offest + 1 */
1992         for (k = depth; k > 1 && !offsets[k-1]; k--)
1993                 ;
1994         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1995         /* Writer: pointers */
1996         if (!partial)
1997                 partial = chain + k-1;
1998         /*
1999          * If the branch acquired continuation since we've looked at it -
2000          * fine, it should all survive and (new) top doesn't belong to us.
2001          */
2002         if (!partial->key && *partial->p)
2003                 /* Writer: end */
2004                 goto no_top;
2005         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2006                 ;
2007         /*
2008          * OK, we've found the last block that must survive. The rest of our
2009          * branch should be detached before unlocking. However, if that rest
2010          * of branch is all ours and does not grow immediately from the inode
2011          * it's easier to cheat and just decrement partial->p.
2012          */
2013         if (p == chain + k - 1 && p > chain) {
2014                 p->p--;
2015         } else {
2016                 *top = *p->p;
2017                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2018 #if 0
2019                 *p->p = 0;
2020 #endif
2021         }
2022         /* Writer: end */
2023
2024         while(partial > p) {
2025                 brelse(partial->bh);
2026                 partial--;
2027         }
2028 no_top:
2029         return partial;
2030 }
2031
2032 /*
2033  * Zero a number of block pointers in either an inode or an indirect block.
2034  * If we restart the transaction we must again get write access to the
2035  * indirect block for further modification.
2036  *
2037  * We release `count' blocks on disk, but (last - first) may be greater
2038  * than `count' because there can be holes in there.
2039  */
2040 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2041                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2042                 unsigned long count, __le32 *first, __le32 *last)
2043 {
2044         __le32 *p;
2045         if (try_to_extend_transaction(handle, inode)) {
2046                 if (bh) {
2047                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2048                         ext3_journal_dirty_metadata(handle, bh);
2049                 }
2050                 ext3_mark_inode_dirty(handle, inode);
2051                 ext3_journal_test_restart(handle, inode);
2052                 if (bh) {
2053                         BUFFER_TRACE(bh, "retaking write access");
2054                         ext3_journal_get_write_access(handle, bh);
2055                 }
2056         }
2057
2058         /*
2059          * Any buffers which are on the journal will be in memory. We find
2060          * them on the hash table so journal_revoke() will run journal_forget()
2061          * on them.  We've already detached each block from the file, so
2062          * bforget() in journal_forget() should be safe.
2063          *
2064          * AKPM: turn on bforget in journal_forget()!!!
2065          */
2066         for (p = first; p < last; p++) {
2067                 u32 nr = le32_to_cpu(*p);
2068                 if (nr) {
2069                         struct buffer_head *bh;
2070
2071                         *p = 0;
2072                         bh = sb_find_get_block(inode->i_sb, nr);
2073                         ext3_forget(handle, 0, inode, bh, nr);
2074                 }
2075         }
2076
2077         ext3_free_blocks(handle, inode, block_to_free, count);
2078 }
2079
2080 /**
2081  * ext3_free_data - free a list of data blocks
2082  * @handle:     handle for this transaction
2083  * @inode:      inode we are dealing with
2084  * @this_bh:    indirect buffer_head which contains *@first and *@last
2085  * @first:      array of block numbers
2086  * @last:       points immediately past the end of array
2087  *
2088  * We are freeing all blocks refered from that array (numbers are stored as
2089  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2090  *
2091  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2092  * blocks are contiguous then releasing them at one time will only affect one
2093  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2094  * actually use a lot of journal space.
2095  *
2096  * @this_bh will be %NULL if @first and @last point into the inode's direct
2097  * block pointers.
2098  */
2099 static void ext3_free_data(handle_t *handle, struct inode *inode,
2100                            struct buffer_head *this_bh,
2101                            __le32 *first, __le32 *last)
2102 {
2103         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2104         unsigned long count = 0;            /* Number of blocks in the run */
2105         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2106                                                corresponding to
2107                                                block_to_free */
2108         ext3_fsblk_t nr;                    /* Current block # */
2109         __le32 *p;                          /* Pointer into inode/ind
2110                                                for current block */
2111         int err;
2112
2113         if (this_bh) {                          /* For indirect block */
2114                 BUFFER_TRACE(this_bh, "get_write_access");
2115                 err = ext3_journal_get_write_access(handle, this_bh);
2116                 /* Important: if we can't update the indirect pointers
2117                  * to the blocks, we can't free them. */
2118                 if (err)
2119                         return;
2120         }
2121
2122         for (p = first; p < last; p++) {
2123                 nr = le32_to_cpu(*p);
2124                 if (nr) {
2125                         /* accumulate blocks to free if they're contiguous */
2126                         if (count == 0) {
2127                                 block_to_free = nr;
2128                                 block_to_free_p = p;
2129                                 count = 1;
2130                         } else if (nr == block_to_free + count) {
2131                                 count++;
2132                         } else {
2133                                 ext3_clear_blocks(handle, inode, this_bh,
2134                                                   block_to_free,
2135                                                   count, block_to_free_p, p);
2136                                 block_to_free = nr;
2137                                 block_to_free_p = p;
2138                                 count = 1;
2139                         }
2140                 }
2141         }
2142
2143         if (count > 0)
2144                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2145                                   count, block_to_free_p, p);
2146
2147         if (this_bh) {
2148                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2149
2150                 /*
2151                  * The buffer head should have an attached journal head at this
2152                  * point. However, if the data is corrupted and an indirect
2153                  * block pointed to itself, it would have been detached when
2154                  * the block was cleared. Check for this instead of OOPSing.
2155                  */
2156                 if (bh2jh(this_bh))
2157                         ext3_journal_dirty_metadata(handle, this_bh);
2158                 else
2159                         ext3_error(inode->i_sb, "ext3_free_data",
2160                                    "circular indirect block detected, "
2161                                    "inode=%lu, block=%llu",
2162                                    inode->i_ino,
2163                                    (unsigned long long)this_bh->b_blocknr);
2164         }
2165 }
2166
2167 /**
2168  *      ext3_free_branches - free an array of branches
2169  *      @handle: JBD handle for this transaction
2170  *      @inode: inode we are dealing with
2171  *      @parent_bh: the buffer_head which contains *@first and *@last
2172  *      @first: array of block numbers
2173  *      @last:  pointer immediately past the end of array
2174  *      @depth: depth of the branches to free
2175  *
2176  *      We are freeing all blocks refered from these branches (numbers are
2177  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2178  *      appropriately.
2179  */
2180 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2181                                struct buffer_head *parent_bh,
2182                                __le32 *first, __le32 *last, int depth)
2183 {
2184         ext3_fsblk_t nr;
2185         __le32 *p;
2186
2187         if (is_handle_aborted(handle))
2188                 return;
2189
2190         if (depth--) {
2191                 struct buffer_head *bh;
2192                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2193                 p = last;
2194                 while (--p >= first) {
2195                         nr = le32_to_cpu(*p);
2196                         if (!nr)
2197                                 continue;               /* A hole */
2198
2199                         /* Go read the buffer for the next level down */
2200                         bh = sb_bread(inode->i_sb, nr);
2201
2202                         /*
2203                          * A read failure? Report error and clear slot
2204                          * (should be rare).
2205                          */
2206                         if (!bh) {
2207                                 ext3_error(inode->i_sb, "ext3_free_branches",
2208                                            "Read failure, inode=%lu, block="E3FSBLK,
2209                                            inode->i_ino, nr);
2210                                 continue;
2211                         }
2212
2213                         /* This zaps the entire block.  Bottom up. */
2214                         BUFFER_TRACE(bh, "free child branches");
2215                         ext3_free_branches(handle, inode, bh,
2216                                            (__le32*)bh->b_data,
2217                                            (__le32*)bh->b_data + addr_per_block,
2218                                            depth);
2219
2220                         /*
2221                          * We've probably journalled the indirect block several
2222                          * times during the truncate.  But it's no longer
2223                          * needed and we now drop it from the transaction via
2224                          * journal_revoke().
2225                          *
2226                          * That's easy if it's exclusively part of this
2227                          * transaction.  But if it's part of the committing
2228                          * transaction then journal_forget() will simply
2229                          * brelse() it.  That means that if the underlying
2230                          * block is reallocated in ext3_get_block(),
2231                          * unmap_underlying_metadata() will find this block
2232                          * and will try to get rid of it.  damn, damn.
2233                          *
2234                          * If this block has already been committed to the
2235                          * journal, a revoke record will be written.  And
2236                          * revoke records must be emitted *before* clearing
2237                          * this block's bit in the bitmaps.
2238                          */
2239                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2240
2241                         /*
2242                          * Everything below this this pointer has been
2243                          * released.  Now let this top-of-subtree go.
2244                          *
2245                          * We want the freeing of this indirect block to be
2246                          * atomic in the journal with the updating of the
2247                          * bitmap block which owns it.  So make some room in
2248                          * the journal.
2249                          *
2250                          * We zero the parent pointer *after* freeing its
2251                          * pointee in the bitmaps, so if extend_transaction()
2252                          * for some reason fails to put the bitmap changes and
2253                          * the release into the same transaction, recovery
2254                          * will merely complain about releasing a free block,
2255                          * rather than leaking blocks.
2256                          */
2257                         if (is_handle_aborted(handle))
2258                                 return;
2259                         if (try_to_extend_transaction(handle, inode)) {
2260                                 ext3_mark_inode_dirty(handle, inode);
2261                                 ext3_journal_test_restart(handle, inode);
2262                         }
2263
2264                         ext3_free_blocks(handle, inode, nr, 1);
2265
2266                         if (parent_bh) {
2267                                 /*
2268                                  * The block which we have just freed is
2269                                  * pointed to by an indirect block: journal it
2270                                  */
2271                                 BUFFER_TRACE(parent_bh, "get_write_access");
2272                                 if (!ext3_journal_get_write_access(handle,
2273                                                                    parent_bh)){
2274                                         *p = 0;
2275                                         BUFFER_TRACE(parent_bh,
2276                                         "call ext3_journal_dirty_metadata");
2277                                         ext3_journal_dirty_metadata(handle,
2278                                                                     parent_bh);
2279                                 }
2280                         }
2281                 }
2282         } else {
2283                 /* We have reached the bottom of the tree. */
2284                 BUFFER_TRACE(parent_bh, "free data blocks");
2285                 ext3_free_data(handle, inode, parent_bh, first, last);
2286         }
2287 }
2288
2289 int ext3_can_truncate(struct inode *inode)
2290 {
2291         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2292                 return 0;
2293         if (S_ISREG(inode->i_mode))
2294                 return 1;
2295         if (S_ISDIR(inode->i_mode))
2296                 return 1;
2297         if (S_ISLNK(inode->i_mode))
2298                 return !ext3_inode_is_fast_symlink(inode);
2299         return 0;
2300 }
2301
2302 /*
2303  * ext3_truncate()
2304  *
2305  * We block out ext3_get_block() block instantiations across the entire
2306  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2307  * simultaneously on behalf of the same inode.
2308  *
2309  * As we work through the truncate and commmit bits of it to the journal there
2310  * is one core, guiding principle: the file's tree must always be consistent on
2311  * disk.  We must be able to restart the truncate after a crash.
2312  *
2313  * The file's tree may be transiently inconsistent in memory (although it
2314  * probably isn't), but whenever we close off and commit a journal transaction,
2315  * the contents of (the filesystem + the journal) must be consistent and
2316  * restartable.  It's pretty simple, really: bottom up, right to left (although
2317  * left-to-right works OK too).
2318  *
2319  * Note that at recovery time, journal replay occurs *before* the restart of
2320  * truncate against the orphan inode list.
2321  *
2322  * The committed inode has the new, desired i_size (which is the same as
2323  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2324  * that this inode's truncate did not complete and it will again call
2325  * ext3_truncate() to have another go.  So there will be instantiated blocks
2326  * to the right of the truncation point in a crashed ext3 filesystem.  But
2327  * that's fine - as long as they are linked from the inode, the post-crash
2328  * ext3_truncate() run will find them and release them.
2329  */
2330 void ext3_truncate(struct inode *inode)
2331 {
2332         handle_t *handle;
2333         struct ext3_inode_info *ei = EXT3_I(inode);
2334         __le32 *i_data = ei->i_data;
2335         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2336         struct address_space *mapping = inode->i_mapping;
2337         int offsets[4];
2338         Indirect chain[4];
2339         Indirect *partial;
2340         __le32 nr = 0;
2341         int n;
2342         long last_block;
2343         unsigned blocksize = inode->i_sb->s_blocksize;
2344         struct page *page;
2345
2346         if (!ext3_can_truncate(inode))
2347                 return;
2348
2349         /*
2350          * We have to lock the EOF page here, because lock_page() nests
2351          * outside journal_start().
2352          */
2353         if ((inode->i_size & (blocksize - 1)) == 0) {
2354                 /* Block boundary? Nothing to do */
2355                 page = NULL;
2356         } else {
2357                 page = grab_cache_page(mapping,
2358                                 inode->i_size >> PAGE_CACHE_SHIFT);
2359                 if (!page)
2360                         return;
2361         }
2362
2363         handle = start_transaction(inode);
2364         if (IS_ERR(handle)) {
2365                 if (page) {
2366                         clear_highpage(page);
2367                         flush_dcache_page(page);
2368                         unlock_page(page);
2369                         page_cache_release(page);
2370                 }
2371                 return;         /* AKPM: return what? */
2372         }
2373
2374         last_block = (inode->i_size + blocksize-1)
2375                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2376
2377         if (page)
2378                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2379
2380         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2381         if (n == 0)
2382                 goto out_stop;  /* error */
2383
2384         /*
2385          * OK.  This truncate is going to happen.  We add the inode to the
2386          * orphan list, so that if this truncate spans multiple transactions,
2387          * and we crash, we will resume the truncate when the filesystem
2388          * recovers.  It also marks the inode dirty, to catch the new size.
2389          *
2390          * Implication: the file must always be in a sane, consistent
2391          * truncatable state while each transaction commits.
2392          */
2393         if (ext3_orphan_add(handle, inode))
2394                 goto out_stop;
2395
2396         /*
2397          * The orphan list entry will now protect us from any crash which
2398          * occurs before the truncate completes, so it is now safe to propagate
2399          * the new, shorter inode size (held for now in i_size) into the
2400          * on-disk inode. We do this via i_disksize, which is the value which
2401          * ext3 *really* writes onto the disk inode.
2402          */
2403         ei->i_disksize = inode->i_size;
2404
2405         /*
2406          * From here we block out all ext3_get_block() callers who want to
2407          * modify the block allocation tree.
2408          */
2409         mutex_lock(&ei->truncate_mutex);
2410
2411         if (n == 1) {           /* direct blocks */
2412                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2413                                i_data + EXT3_NDIR_BLOCKS);
2414                 goto do_indirects;
2415         }
2416
2417         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2418         /* Kill the top of shared branch (not detached) */
2419         if (nr) {
2420                 if (partial == chain) {
2421                         /* Shared branch grows from the inode */
2422                         ext3_free_branches(handle, inode, NULL,
2423                                            &nr, &nr+1, (chain+n-1) - partial);
2424                         *partial->p = 0;
2425                         /*
2426                          * We mark the inode dirty prior to restart,
2427                          * and prior to stop.  No need for it here.
2428                          */
2429                 } else {
2430                         /* Shared branch grows from an indirect block */
2431                         BUFFER_TRACE(partial->bh, "get_write_access");
2432                         ext3_free_branches(handle, inode, partial->bh,
2433                                         partial->p,
2434                                         partial->p+1, (chain+n-1) - partial);
2435                 }
2436         }
2437         /* Clear the ends of indirect blocks on the shared branch */
2438         while (partial > chain) {
2439                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2440                                    (__le32*)partial->bh->b_data+addr_per_block,
2441                                    (chain+n-1) - partial);
2442                 BUFFER_TRACE(partial->bh, "call brelse");
2443                 brelse (partial->bh);
2444                 partial--;
2445         }
2446 do_indirects:
2447         /* Kill the remaining (whole) subtrees */
2448         switch (offsets[0]) {
2449         default:
2450                 nr = i_data[EXT3_IND_BLOCK];
2451                 if (nr) {
2452                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2453                         i_data[EXT3_IND_BLOCK] = 0;
2454                 }
2455         case EXT3_IND_BLOCK:
2456                 nr = i_data[EXT3_DIND_BLOCK];
2457                 if (nr) {
2458                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2459                         i_data[EXT3_DIND_BLOCK] = 0;
2460                 }
2461         case EXT3_DIND_BLOCK:
2462                 nr = i_data[EXT3_TIND_BLOCK];
2463                 if (nr) {
2464                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2465                         i_data[EXT3_TIND_BLOCK] = 0;
2466                 }
2467         case EXT3_TIND_BLOCK:
2468                 ;
2469         }
2470
2471         ext3_discard_reservation(inode);
2472
2473         mutex_unlock(&ei->truncate_mutex);
2474         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2475         ext3_mark_inode_dirty(handle, inode);
2476
2477         /*
2478          * In a multi-transaction truncate, we only make the final transaction
2479          * synchronous
2480          */
2481         if (IS_SYNC(inode))
2482                 handle->h_sync = 1;
2483 out_stop:
2484         /*
2485          * If this was a simple ftruncate(), and the file will remain alive
2486          * then we need to clear up the orphan record which we created above.
2487          * However, if this was a real unlink then we were called by
2488          * ext3_delete_inode(), and we allow that function to clean up the
2489          * orphan info for us.
2490          */
2491         if (inode->i_nlink)
2492                 ext3_orphan_del(handle, inode);
2493
2494         ext3_journal_stop(handle);
2495 }
2496
2497 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2498                 unsigned long ino, struct ext3_iloc *iloc)
2499 {
2500         unsigned long block_group;
2501         unsigned long offset;
2502         ext3_fsblk_t block;
2503         struct ext3_group_desc *gdp;
2504
2505         if (!ext3_valid_inum(sb, ino)) {
2506                 /*
2507                  * This error is already checked for in namei.c unless we are
2508                  * looking at an NFS filehandle, in which case no error
2509                  * report is needed
2510                  */
2511                 return 0;
2512         }
2513
2514         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2515         gdp = ext3_get_group_desc(sb, block_group, NULL);
2516         if (!gdp)
2517                 return 0;
2518         /*
2519          * Figure out the offset within the block group inode table
2520          */
2521         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2522                 EXT3_INODE_SIZE(sb);
2523         block = le32_to_cpu(gdp->bg_inode_table) +
2524                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2525
2526         iloc->block_group = block_group;
2527         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2528         return block;
2529 }
2530
2531 /*
2532  * ext3_get_inode_loc returns with an extra refcount against the inode's
2533  * underlying buffer_head on success. If 'in_mem' is true, we have all
2534  * data in memory that is needed to recreate the on-disk version of this
2535  * inode.
2536  */
2537 static int __ext3_get_inode_loc(struct inode *inode,
2538                                 struct ext3_iloc *iloc, int in_mem)
2539 {
2540         ext3_fsblk_t block;
2541         struct buffer_head *bh;
2542
2543         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2544         if (!block)
2545                 return -EIO;
2546
2547         bh = sb_getblk(inode->i_sb, block);
2548         if (!bh) {
2549                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2550                                 "unable to read inode block - "
2551                                 "inode=%lu, block="E3FSBLK,
2552                                  inode->i_ino, block);
2553                 return -EIO;
2554         }
2555         if (!buffer_uptodate(bh)) {
2556                 lock_buffer(bh);
2557
2558                 /*
2559                  * If the buffer has the write error flag, we have failed
2560                  * to write out another inode in the same block.  In this
2561                  * case, we don't have to read the block because we may
2562                  * read the old inode data successfully.
2563                  */
2564                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2565                         set_buffer_uptodate(bh);
2566
2567                 if (buffer_uptodate(bh)) {
2568                         /* someone brought it uptodate while we waited */
2569                         unlock_buffer(bh);
2570                         goto has_buffer;
2571                 }
2572
2573                 /*
2574                  * If we have all information of the inode in memory and this
2575                  * is the only valid inode in the block, we need not read the
2576                  * block.
2577                  */
2578                 if (in_mem) {
2579                         struct buffer_head *bitmap_bh;
2580                         struct ext3_group_desc *desc;
2581                         int inodes_per_buffer;
2582                         int inode_offset, i;
2583                         int block_group;
2584                         int start;
2585
2586                         block_group = (inode->i_ino - 1) /
2587                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2588                         inodes_per_buffer = bh->b_size /
2589                                 EXT3_INODE_SIZE(inode->i_sb);
2590                         inode_offset = ((inode->i_ino - 1) %
2591                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2592                         start = inode_offset & ~(inodes_per_buffer - 1);
2593
2594                         /* Is the inode bitmap in cache? */
2595                         desc = ext3_get_group_desc(inode->i_sb,
2596                                                 block_group, NULL);
2597                         if (!desc)
2598                                 goto make_io;
2599
2600                         bitmap_bh = sb_getblk(inode->i_sb,
2601                                         le32_to_cpu(desc->bg_inode_bitmap));
2602                         if (!bitmap_bh)
2603                                 goto make_io;
2604
2605                         /*
2606                          * If the inode bitmap isn't in cache then the
2607                          * optimisation may end up performing two reads instead
2608                          * of one, so skip it.
2609                          */
2610                         if (!buffer_uptodate(bitmap_bh)) {
2611                                 brelse(bitmap_bh);
2612                                 goto make_io;
2613                         }
2614                         for (i = start; i < start + inodes_per_buffer; i++) {
2615                                 if (i == inode_offset)
2616                                         continue;
2617                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2618                                         break;
2619                         }
2620                         brelse(bitmap_bh);
2621                         if (i == start + inodes_per_buffer) {
2622                                 /* all other inodes are free, so skip I/O */
2623                                 memset(bh->b_data, 0, bh->b_size);
2624                                 set_buffer_uptodate(bh);
2625                                 unlock_buffer(bh);
2626                                 goto has_buffer;
2627                         }
2628                 }
2629
2630 make_io:
2631                 /*
2632                  * There are other valid inodes in the buffer, this inode
2633                  * has in-inode xattrs, or we don't have this inode in memory.
2634                  * Read the block from disk.
2635                  */
2636                 get_bh(bh);
2637                 bh->b_end_io = end_buffer_read_sync;
2638                 submit_bh(READ_META, bh);
2639                 wait_on_buffer(bh);
2640                 if (!buffer_uptodate(bh)) {
2641                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2642                                         "unable to read inode block - "
2643                                         "inode=%lu, block="E3FSBLK,
2644                                         inode->i_ino, block);
2645                         brelse(bh);
2646                         return -EIO;
2647                 }
2648         }
2649 has_buffer:
2650         iloc->bh = bh;
2651         return 0;
2652 }
2653
2654 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2655 {
2656         /* We have all inode data except xattrs in memory here. */
2657         return __ext3_get_inode_loc(inode, iloc,
2658                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2659 }
2660
2661 void ext3_set_inode_flags(struct inode *inode)
2662 {
2663         unsigned int flags = EXT3_I(inode)->i_flags;
2664
2665         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2666         if (flags & EXT3_SYNC_FL)
2667                 inode->i_flags |= S_SYNC;
2668         if (flags & EXT3_APPEND_FL)
2669                 inode->i_flags |= S_APPEND;
2670         if (flags & EXT3_IMMUTABLE_FL)
2671                 inode->i_flags |= S_IMMUTABLE;
2672         if (flags & EXT3_NOATIME_FL)
2673                 inode->i_flags |= S_NOATIME;
2674         if (flags & EXT3_DIRSYNC_FL)
2675                 inode->i_flags |= S_DIRSYNC;
2676 }
2677
2678 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2679 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2680 {
2681         unsigned int flags = ei->vfs_inode.i_flags;
2682
2683         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2684                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2685         if (flags & S_SYNC)
2686                 ei->i_flags |= EXT3_SYNC_FL;
2687         if (flags & S_APPEND)
2688                 ei->i_flags |= EXT3_APPEND_FL;
2689         if (flags & S_IMMUTABLE)
2690                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2691         if (flags & S_NOATIME)
2692                 ei->i_flags |= EXT3_NOATIME_FL;
2693         if (flags & S_DIRSYNC)
2694                 ei->i_flags |= EXT3_DIRSYNC_FL;
2695 }
2696
2697 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2698 {
2699         struct ext3_iloc iloc;
2700         struct ext3_inode *raw_inode;
2701         struct ext3_inode_info *ei;
2702         struct buffer_head *bh;
2703         struct inode *inode;
2704         long ret;
2705         int block;
2706
2707         inode = iget_locked(sb, ino);
2708         if (!inode)
2709                 return ERR_PTR(-ENOMEM);
2710         if (!(inode->i_state & I_NEW))
2711                 return inode;
2712
2713         ei = EXT3_I(inode);
2714 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2715         ei->i_acl = EXT3_ACL_NOT_CACHED;
2716         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2717 #endif
2718         ei->i_block_alloc_info = NULL;
2719
2720         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2721         if (ret < 0)
2722                 goto bad_inode;
2723         bh = iloc.bh;
2724         raw_inode = ext3_raw_inode(&iloc);
2725         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2726         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2727         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2728         if(!(test_opt (inode->i_sb, NO_UID32))) {
2729                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2730                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2731         }
2732         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2733         inode->i_size = le32_to_cpu(raw_inode->i_size);
2734         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2735         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2736         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2737         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2738
2739         ei->i_state = 0;
2740         ei->i_dir_start_lookup = 0;
2741         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2742         /* We now have enough fields to check if the inode was active or not.
2743          * This is needed because nfsd might try to access dead inodes
2744          * the test is that same one that e2fsck uses
2745          * NeilBrown 1999oct15
2746          */
2747         if (inode->i_nlink == 0) {
2748                 if (inode->i_mode == 0 ||
2749                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2750                         /* this inode is deleted */
2751                         brelse (bh);
2752                         ret = -ESTALE;
2753                         goto bad_inode;
2754                 }
2755                 /* The only unlinked inodes we let through here have
2756                  * valid i_mode and are being read by the orphan
2757                  * recovery code: that's fine, we're about to complete
2758                  * the process of deleting those. */
2759         }
2760         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2761         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2762 #ifdef EXT3_FRAGMENTS
2763         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2764         ei->i_frag_no = raw_inode->i_frag;
2765         ei->i_frag_size = raw_inode->i_fsize;
2766 #endif
2767         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2768         if (!S_ISREG(inode->i_mode)) {
2769                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2770         } else {
2771                 inode->i_size |=
2772                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2773         }
2774         ei->i_disksize = inode->i_size;
2775         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2776         ei->i_block_group = iloc.block_group;
2777         /*
2778          * NOTE! The in-memory inode i_data array is in little-endian order
2779          * even on big-endian machines: we do NOT byteswap the block numbers!
2780          */
2781         for (block = 0; block < EXT3_N_BLOCKS; block++)
2782                 ei->i_data[block] = raw_inode->i_block[block];
2783         INIT_LIST_HEAD(&ei->i_orphan);
2784
2785         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2786             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2787                 /*
2788                  * When mke2fs creates big inodes it does not zero out
2789                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2790                  * so ignore those first few inodes.
2791                  */
2792                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2793                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2794                     EXT3_INODE_SIZE(inode->i_sb)) {
2795                         brelse (bh);
2796                         ret = -EIO;
2797                         goto bad_inode;
2798                 }
2799                 if (ei->i_extra_isize == 0) {
2800                         /* The extra space is currently unused. Use it. */
2801                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2802                                             EXT3_GOOD_OLD_INODE_SIZE;
2803                 } else {
2804                         __le32 *magic = (void *)raw_inode +
2805                                         EXT3_GOOD_OLD_INODE_SIZE +
2806                                         ei->i_extra_isize;
2807                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2808                                  ei->i_state |= EXT3_STATE_XATTR;
2809                 }
2810         } else
2811                 ei->i_extra_isize = 0;
2812
2813         if (S_ISREG(inode->i_mode)) {
2814                 inode->i_op = &ext3_file_inode_operations;
2815                 inode->i_fop = &ext3_file_operations;
2816                 ext3_set_aops(inode);
2817         } else if (S_ISDIR(inode->i_mode)) {
2818                 inode->i_op = &ext3_dir_inode_operations;
2819                 inode->i_fop = &ext3_dir_operations;
2820         } else if (S_ISLNK(inode->i_mode)) {
2821                 if (ext3_inode_is_fast_symlink(inode)) {
2822                         inode->i_op = &ext3_fast_symlink_inode_operations;
2823                         nd_terminate_link(ei->i_data, inode->i_size,
2824                                 sizeof(ei->i_data) - 1);
2825                 } else {
2826                         inode->i_op = &ext3_symlink_inode_operations;
2827                         ext3_set_aops(inode);
2828                 }
2829         } else {
2830                 inode->i_op = &ext3_special_inode_operations;
2831                 if (raw_inode->i_block[0])
2832                         init_special_inode(inode, inode->i_mode,
2833                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2834                 else
2835                         init_special_inode(inode, inode->i_mode,
2836                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2837         }
2838         brelse (iloc.bh);
2839         ext3_set_inode_flags(inode);
2840         unlock_new_inode(inode);
2841         return inode;
2842
2843 bad_inode:
2844         iget_failed(inode);
2845         return ERR_PTR(ret);
2846 }
2847
2848 /*
2849  * Post the struct inode info into an on-disk inode location in the
2850  * buffer-cache.  This gobbles the caller's reference to the
2851  * buffer_head in the inode location struct.
2852  *
2853  * The caller must have write access to iloc->bh.
2854  */
2855 static int ext3_do_update_inode(handle_t *handle,
2856                                 struct inode *inode,
2857                                 struct ext3_iloc *iloc)
2858 {
2859         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2860         struct ext3_inode_info *ei = EXT3_I(inode);
2861         struct buffer_head *bh = iloc->bh;
2862         int err = 0, rc, block;
2863
2864         /* For fields not not tracking in the in-memory inode,
2865          * initialise them to zero for new inodes. */
2866         if (ei->i_state & EXT3_STATE_NEW)
2867                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2868
2869         ext3_get_inode_flags(ei);
2870         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2871         if(!(test_opt(inode->i_sb, NO_UID32))) {
2872                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2873                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2874 /*
2875  * Fix up interoperability with old kernels. Otherwise, old inodes get
2876  * re-used with the upper 16 bits of the uid/gid intact
2877  */
2878                 if(!ei->i_dtime) {
2879                         raw_inode->i_uid_high =
2880                                 cpu_to_le16(high_16_bits(inode->i_uid));
2881                         raw_inode->i_gid_high =
2882                                 cpu_to_le16(high_16_bits(inode->i_gid));
2883                 } else {
2884                         raw_inode->i_uid_high = 0;
2885                         raw_inode->i_gid_high = 0;
2886                 }
2887         } else {
2888                 raw_inode->i_uid_low =
2889                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2890                 raw_inode->i_gid_low =
2891                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2892                 raw_inode->i_uid_high = 0;
2893                 raw_inode->i_gid_high = 0;
2894         }
2895         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2896         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2897         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2898         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2899         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2900         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2901         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2902         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2903 #ifdef EXT3_FRAGMENTS
2904         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2905         raw_inode->i_frag = ei->i_frag_no;
2906         raw_inode->i_fsize = ei->i_frag_size;
2907 #endif
2908         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2909         if (!S_ISREG(inode->i_mode)) {
2910                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2911         } else {
2912                 raw_inode->i_size_high =
2913                         cpu_to_le32(ei->i_disksize >> 32);
2914                 if (ei->i_disksize > 0x7fffffffULL) {
2915                         struct super_block *sb = inode->i_sb;
2916                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2917                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2918                             EXT3_SB(sb)->s_es->s_rev_level ==
2919                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2920                                /* If this is the first large file
2921                                 * created, add a flag to the superblock.
2922                                 */
2923                                 err = ext3_journal_get_write_access(handle,
2924                                                 EXT3_SB(sb)->s_sbh);
2925                                 if (err)
2926                                         goto out_brelse;
2927                                 ext3_update_dynamic_rev(sb);
2928                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2929                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2930                                 sb->s_dirt = 1;
2931                                 handle->h_sync = 1;
2932                                 err = ext3_journal_dirty_metadata(handle,
2933                                                 EXT3_SB(sb)->s_sbh);
2934                         }
2935                 }
2936         }
2937         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2938         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2939                 if (old_valid_dev(inode->i_rdev)) {
2940                         raw_inode->i_block[0] =
2941                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2942                         raw_inode->i_block[1] = 0;
2943                 } else {
2944                         raw_inode->i_block[0] = 0;
2945                         raw_inode->i_block[1] =
2946                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2947                         raw_inode->i_block[2] = 0;
2948                 }
2949         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2950                 raw_inode->i_block[block] = ei->i_data[block];
2951
2952         if (ei->i_extra_isize)
2953                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2954
2955         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2956         rc = ext3_journal_dirty_metadata(handle, bh);
2957         if (!err)
2958                 err = rc;
2959         ei->i_state &= ~EXT3_STATE_NEW;
2960
2961 out_brelse:
2962         brelse (bh);
2963         ext3_std_error(inode->i_sb, err);
2964         return err;
2965 }
2966
2967 /*
2968  * ext3_write_inode()
2969  *
2970  * We are called from a few places:
2971  *
2972  * - Within generic_file_write() for O_SYNC files.
2973  *   Here, there will be no transaction running. We wait for any running
2974  *   trasnaction to commit.
2975  *
2976  * - Within sys_sync(), kupdate and such.
2977  *   We wait on commit, if tol to.
2978  *
2979  * - Within prune_icache() (PF_MEMALLOC == true)
2980  *   Here we simply return.  We can't afford to block kswapd on the
2981  *   journal commit.
2982  *
2983  * In all cases it is actually safe for us to return without doing anything,
2984  * because the inode has been copied into a raw inode buffer in
2985  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2986  * knfsd.
2987  *
2988  * Note that we are absolutely dependent upon all inode dirtiers doing the
2989  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2990  * which we are interested.
2991  *
2992  * It would be a bug for them to not do this.  The code:
2993  *
2994  *      mark_inode_dirty(inode)
2995  *      stuff();
2996  *      inode->i_size = expr;
2997  *
2998  * is in error because a kswapd-driven write_inode() could occur while
2999  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3000  * will no longer be on the superblock's dirty inode list.
3001  */
3002 int ext3_write_inode(struct inode *inode, int wait)
3003 {
3004         if (current->flags & PF_MEMALLOC)
3005                 return 0;
3006
3007         if (ext3_journal_current_handle()) {
3008                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3009                 dump_stack();
3010                 return -EIO;
3011         }
3012
3013         if (!wait)
3014                 return 0;
3015
3016         return ext3_force_commit(inode->i_sb);
3017 }
3018
3019 /*
3020  * ext3_setattr()
3021  *
3022  * Called from notify_change.
3023  *
3024  * We want to trap VFS attempts to truncate the file as soon as
3025  * possible.  In particular, we want to make sure that when the VFS
3026  * shrinks i_size, we put the inode on the orphan list and modify
3027  * i_disksize immediately, so that during the subsequent flushing of
3028  * dirty pages and freeing of disk blocks, we can guarantee that any
3029  * commit will leave the blocks being flushed in an unused state on
3030  * disk.  (On recovery, the inode will get truncated and the blocks will
3031  * be freed, so we have a strong guarantee that no future commit will
3032  * leave these blocks visible to the user.)
3033  *
3034  * Called with inode->sem down.
3035  */
3036 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3037 {
3038         struct inode *inode = dentry->d_inode;
3039         int error, rc = 0;
3040         const unsigned int ia_valid = attr->ia_valid;
3041
3042         error = inode_change_ok(inode, attr);
3043         if (error)
3044                 return error;
3045
3046         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3047                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3048                 handle_t *handle;
3049
3050                 /* (user+group)*(old+new) structure, inode write (sb,
3051                  * inode block, ? - but truncate inode update has it) */
3052                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3053                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3054                 if (IS_ERR(handle)) {
3055                         error = PTR_ERR(handle);
3056                         goto err_out;
3057                 }
3058                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
3059                 if (error) {
3060                         ext3_journal_stop(handle);
3061                         return error;
3062                 }
3063                 /* Update corresponding info in inode so that everything is in
3064                  * one transaction */
3065                 if (attr->ia_valid & ATTR_UID)
3066                         inode->i_uid = attr->ia_uid;
3067                 if (attr->ia_valid & ATTR_GID)
3068                         inode->i_gid = attr->ia_gid;
3069                 error = ext3_mark_inode_dirty(handle, inode);
3070                 ext3_journal_stop(handle);
3071         }
3072
3073         if (S_ISREG(inode->i_mode) &&
3074             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3075                 handle_t *handle;
3076
3077                 handle = ext3_journal_start(inode, 3);
3078                 if (IS_ERR(handle)) {
3079                         error = PTR_ERR(handle);
3080                         goto err_out;
3081                 }
3082
3083                 error = ext3_orphan_add(handle, inode);
3084                 EXT3_I(inode)->i_disksize = attr->ia_size;
3085                 rc = ext3_mark_inode_dirty(handle, inode);
3086                 if (!error)
3087                         error = rc;
3088                 ext3_journal_stop(handle);
3089         }
3090
3091         rc = inode_setattr(inode, attr);
3092
3093         /* If inode_setattr's call to ext3_truncate failed to get a
3094          * transaction handle at all, we need to clean up the in-core
3095          * orphan list manually. */
3096         if (inode->i_nlink)
3097                 ext3_orphan_del(NULL, inode);
3098
3099         if (!rc && (ia_valid & ATTR_MODE))
3100                 rc = ext3_acl_chmod(inode);
3101
3102 err_out:
3103         ext3_std_error(inode->i_sb, error);
3104         if (!error)
3105                 error = rc;
3106         return error;
3107 }
3108
3109
3110 /*
3111  * How many blocks doth make a writepage()?
3112  *
3113  * With N blocks per page, it may be:
3114  * N data blocks
3115  * 2 indirect block
3116  * 2 dindirect
3117  * 1 tindirect
3118  * N+5 bitmap blocks (from the above)
3119  * N+5 group descriptor summary blocks
3120  * 1 inode block
3121  * 1 superblock.
3122  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3123  *
3124  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3125  *
3126  * With ordered or writeback data it's the same, less the N data blocks.
3127  *
3128  * If the inode's direct blocks can hold an integral number of pages then a
3129  * page cannot straddle two indirect blocks, and we can only touch one indirect
3130  * and dindirect block, and the "5" above becomes "3".
3131  *
3132  * This still overestimates under most circumstances.  If we were to pass the
3133  * start and end offsets in here as well we could do block_to_path() on each
3134  * block and work out the exact number of indirects which are touched.  Pah.
3135  */
3136
3137 static int ext3_writepage_trans_blocks(struct inode *inode)
3138 {
3139         int bpp = ext3_journal_blocks_per_page(inode);
3140         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3141         int ret;
3142
3143         if (ext3_should_journal_data(inode))
3144                 ret = 3 * (bpp + indirects) + 2;
3145         else
3146                 ret = 2 * (bpp + indirects) + 2;
3147
3148 #ifdef CONFIG_QUOTA
3149         /* We know that structure was already allocated during DQUOT_INIT so
3150          * we will be updating only the data blocks + inodes */
3151         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3152 #endif
3153
3154         return ret;
3155 }
3156
3157 /*
3158  * The caller must have previously called ext3_reserve_inode_write().
3159  * Give this, we know that the caller already has write access to iloc->bh.
3160  */
3161 int ext3_mark_iloc_dirty(handle_t *handle,
3162                 struct inode *inode, struct ext3_iloc *iloc)
3163 {
3164         int err = 0;
3165
3166         /* the do_update_inode consumes one bh->b_count */
3167         get_bh(iloc->bh);
3168
3169         /* ext3_do_update_inode() does journal_dirty_metadata */
3170         err = ext3_do_update_inode(handle, inode, iloc);
3171         put_bh(iloc->bh);
3172         return err;
3173 }
3174
3175 /*
3176  * On success, We end up with an outstanding reference count against
3177  * iloc->bh.  This _must_ be cleaned up later.
3178  */
3179
3180 int
3181 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3182                          struct ext3_iloc *iloc)
3183 {
3184         int err = 0;
3185         if (handle) {
3186                 err = ext3_get_inode_loc(inode, iloc);
3187                 if (!err) {
3188                         BUFFER_TRACE(iloc->bh, "get_write_access");
3189                         err = ext3_journal_get_write_access(handle, iloc->bh);
3190                         if (err) {
3191                                 brelse(iloc->bh);
3192                                 iloc->bh = NULL;
3193                         }
3194                 }
3195         }
3196         ext3_std_error(inode->i_sb, err);
3197         return err;
3198 }
3199
3200 /*
3201  * What we do here is to mark the in-core inode as clean with respect to inode
3202  * dirtiness (it may still be data-dirty).
3203  * This means that the in-core inode may be reaped by prune_icache
3204  * without having to perform any I/O.  This is a very good thing,
3205  * because *any* task may call prune_icache - even ones which
3206  * have a transaction open against a different journal.
3207  *
3208  * Is this cheating?  Not really.  Sure, we haven't written the
3209  * inode out, but prune_icache isn't a user-visible syncing function.
3210  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3211  * we start and wait on commits.
3212  *
3213  * Is this efficient/effective?  Well, we're being nice to the system
3214  * by cleaning up our inodes proactively so they can be reaped
3215  * without I/O.  But we are potentially leaving up to five seconds'
3216  * worth of inodes floating about which prune_icache wants us to
3217  * write out.  One way to fix that would be to get prune_icache()
3218  * to do a write_super() to free up some memory.  It has the desired
3219  * effect.
3220  */
3221 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3222 {
3223         struct ext3_iloc iloc;
3224         int err;
3225
3226         might_sleep();
3227         err = ext3_reserve_inode_write(handle, inode, &iloc);
3228         if (!err)
3229                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3230         return err;
3231 }
3232
3233 /*
3234  * ext3_dirty_inode() is called from __mark_inode_dirty()
3235  *
3236  * We're really interested in the case where a file is being extended.
3237  * i_size has been changed by generic_commit_write() and we thus need
3238  * to include the updated inode in the current transaction.
3239  *
3240  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3241  * are allocated to the file.
3242  *
3243  * If the inode is marked synchronous, we don't honour that here - doing
3244  * so would cause a commit on atime updates, which we don't bother doing.
3245  * We handle synchronous inodes at the highest possible level.
3246  */
3247 void ext3_dirty_inode(struct inode *inode)
3248 {
3249         handle_t *current_handle = ext3_journal_current_handle();
3250         handle_t *handle;
3251
3252         handle = ext3_journal_start(inode, 2);
3253         if (IS_ERR(handle))
3254                 goto out;
3255         if (current_handle &&
3256                 current_handle->h_transaction != handle->h_transaction) {
3257                 /* This task has a transaction open against a different fs */
3258                 printk(KERN_EMERG "%s: transactions do not match!\n",
3259                        __func__);
3260         } else {
3261                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3262                                 current_handle);
3263                 ext3_mark_inode_dirty(handle, inode);
3264         }
3265         ext3_journal_stop(handle);
3266 out:
3267         return;
3268 }
3269
3270 #if 0
3271 /*
3272  * Bind an inode's backing buffer_head into this transaction, to prevent
3273  * it from being flushed to disk early.  Unlike
3274  * ext3_reserve_inode_write, this leaves behind no bh reference and
3275  * returns no iloc structure, so the caller needs to repeat the iloc
3276  * lookup to mark the inode dirty later.
3277  */
3278 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3279 {
3280         struct ext3_iloc iloc;
3281
3282         int err = 0;
3283         if (handle) {
3284                 err = ext3_get_inode_loc(inode, &iloc);
3285                 if (!err) {
3286                         BUFFER_TRACE(iloc.bh, "get_write_access");
3287                         err = journal_get_write_access(handle, iloc.bh);
3288                         if (!err)
3289                                 err = ext3_journal_dirty_metadata(handle,
3290                                                                   iloc.bh);
3291                         brelse(iloc.bh);
3292                 }
3293         }
3294         ext3_std_error(inode->i_sb, err);
3295         return err;
3296 }
3297 #endif
3298
3299 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3300 {
3301         journal_t *journal;
3302         handle_t *handle;
3303         int err;
3304
3305         /*
3306          * We have to be very careful here: changing a data block's
3307          * journaling status dynamically is dangerous.  If we write a
3308          * data block to the journal, change the status and then delete
3309          * that block, we risk forgetting to revoke the old log record
3310          * from the journal and so a subsequent replay can corrupt data.
3311          * So, first we make sure that the journal is empty and that
3312          * nobody is changing anything.
3313          */
3314
3315         journal = EXT3_JOURNAL(inode);
3316         if (is_journal_aborted(journal))
3317                 return -EROFS;
3318
3319         journal_lock_updates(journal);
3320         journal_flush(journal);
3321
3322         /*
3323          * OK, there are no updates running now, and all cached data is
3324          * synced to disk.  We are now in a completely consistent state
3325          * which doesn't have anything in the journal, and we know that
3326          * no filesystem updates are running, so it is safe to modify
3327          * the inode's in-core data-journaling state flag now.
3328          */
3329
3330         if (val)
3331                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3332         else
3333                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3334         ext3_set_aops(inode);
3335
3336         journal_unlock_updates(journal);
3337
3338         /* Finally we can mark the inode as dirty. */
3339
3340         handle = ext3_journal_start(inode, 1);
3341         if (IS_ERR(handle))
3342                 return PTR_ERR(handle);
3343
3344         err = ext3_mark_inode_dirty(handle, inode);
3345         handle->h_sync = 1;
3346         ext3_journal_stop(handle);
3347         ext3_std_error(inode->i_sb, err);
3348
3349         return err;
3350 }