[PATCH] retries in ext3_prepare_write() violate ordering requirements
[linux-2.6.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/smp_lock.h>
31 #include <linux/highuid.h>
32 #include <linux/pagemap.h>
33 #include <linux/quotaops.h>
34 #include <linux/string.h>
35 #include <linux/buffer_head.h>
36 #include <linux/writeback.h>
37 #include <linux/mpage.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include "xattr.h"
41 #include "acl.h"
42
43 static int ext3_writepage_trans_blocks(struct inode *inode);
44
45 /*
46  * Test whether an inode is a fast symlink.
47  */
48 static int ext3_inode_is_fast_symlink(struct inode *inode)
49 {
50         int ea_blocks = EXT3_I(inode)->i_file_acl ?
51                 (inode->i_sb->s_blocksize >> 9) : 0;
52
53         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
54 }
55
56 /*
57  * The ext3 forget function must perform a revoke if we are freeing data
58  * which has been journaled.  Metadata (eg. indirect blocks) must be
59  * revoked in all cases.
60  *
61  * "bh" may be NULL: a metadata block may have been freed from memory
62  * but there may still be a record of it in the journal, and that record
63  * still needs to be revoked.
64  */
65 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
66                         struct buffer_head *bh, ext3_fsblk_t blocknr)
67 {
68         int err;
69
70         might_sleep();
71
72         BUFFER_TRACE(bh, "enter");
73
74         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
75                   "data mode %lx\n",
76                   bh, is_metadata, inode->i_mode,
77                   test_opt(inode->i_sb, DATA_FLAGS));
78
79         /* Never use the revoke function if we are doing full data
80          * journaling: there is no need to, and a V1 superblock won't
81          * support it.  Otherwise, only skip the revoke on un-journaled
82          * data blocks. */
83
84         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
85             (!is_metadata && !ext3_should_journal_data(inode))) {
86                 if (bh) {
87                         BUFFER_TRACE(bh, "call journal_forget");
88                         return ext3_journal_forget(handle, bh);
89                 }
90                 return 0;
91         }
92
93         /*
94          * data!=journal && (is_metadata || should_journal_data(inode))
95          */
96         BUFFER_TRACE(bh, "call ext3_journal_revoke");
97         err = ext3_journal_revoke(handle, blocknr, bh);
98         if (err)
99                 ext3_abort(inode->i_sb, __FUNCTION__,
100                            "error %d when attempting revoke", err);
101         BUFFER_TRACE(bh, "exit");
102         return err;
103 }
104
105 /*
106  * Work out how many blocks we need to proceed with the next chunk of a
107  * truncate transaction.
108  */
109 static unsigned long blocks_for_truncate(struct inode *inode)
110 {
111         unsigned long needed;
112
113         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
114
115         /* Give ourselves just enough room to cope with inodes in which
116          * i_blocks is corrupt: we've seen disk corruptions in the past
117          * which resulted in random data in an inode which looked enough
118          * like a regular file for ext3 to try to delete it.  Things
119          * will go a bit crazy if that happens, but at least we should
120          * try not to panic the whole kernel. */
121         if (needed < 2)
122                 needed = 2;
123
124         /* But we need to bound the transaction so we don't overflow the
125          * journal. */
126         if (needed > EXT3_MAX_TRANS_DATA)
127                 needed = EXT3_MAX_TRANS_DATA;
128
129         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
130 }
131
132 /*
133  * Truncate transactions can be complex and absolutely huge.  So we need to
134  * be able to restart the transaction at a conventient checkpoint to make
135  * sure we don't overflow the journal.
136  *
137  * start_transaction gets us a new handle for a truncate transaction,
138  * and extend_transaction tries to extend the existing one a bit.  If
139  * extend fails, we need to propagate the failure up and restart the
140  * transaction in the top-level truncate loop. --sct
141  */
142 static handle_t *start_transaction(struct inode *inode)
143 {
144         handle_t *result;
145
146         result = ext3_journal_start(inode, blocks_for_truncate(inode));
147         if (!IS_ERR(result))
148                 return result;
149
150         ext3_std_error(inode->i_sb, PTR_ERR(result));
151         return result;
152 }
153
154 /*
155  * Try to extend this transaction for the purposes of truncation.
156  *
157  * Returns 0 if we managed to create more room.  If we can't create more
158  * room, and the transaction must be restarted we return 1.
159  */
160 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
161 {
162         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
163                 return 0;
164         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
165                 return 0;
166         return 1;
167 }
168
169 /*
170  * Restart the transaction associated with *handle.  This does a commit,
171  * so before we call here everything must be consistently dirtied against
172  * this transaction.
173  */
174 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
175 {
176         jbd_debug(2, "restarting handle %p\n", handle);
177         return ext3_journal_restart(handle, blocks_for_truncate(inode));
178 }
179
180 /*
181  * Called at the last iput() if i_nlink is zero.
182  */
183 void ext3_delete_inode (struct inode * inode)
184 {
185         handle_t *handle;
186
187         truncate_inode_pages(&inode->i_data, 0);
188
189         if (is_bad_inode(inode))
190                 goto no_delete;
191
192         handle = start_transaction(inode);
193         if (IS_ERR(handle)) {
194                 /*
195                  * If we're going to skip the normal cleanup, we still need to
196                  * make sure that the in-core orphan linked list is properly
197                  * cleaned up.
198                  */
199                 ext3_orphan_del(NULL, inode);
200                 goto no_delete;
201         }
202
203         if (IS_SYNC(inode))
204                 handle->h_sync = 1;
205         inode->i_size = 0;
206         if (inode->i_blocks)
207                 ext3_truncate(inode);
208         /*
209          * Kill off the orphan record which ext3_truncate created.
210          * AKPM: I think this can be inside the above `if'.
211          * Note that ext3_orphan_del() has to be able to cope with the
212          * deletion of a non-existent orphan - this is because we don't
213          * know if ext3_truncate() actually created an orphan record.
214          * (Well, we could do this if we need to, but heck - it works)
215          */
216         ext3_orphan_del(handle, inode);
217         EXT3_I(inode)->i_dtime  = get_seconds();
218
219         /*
220          * One subtle ordering requirement: if anything has gone wrong
221          * (transaction abort, IO errors, whatever), then we can still
222          * do these next steps (the fs will already have been marked as
223          * having errors), but we can't free the inode if the mark_dirty
224          * fails.
225          */
226         if (ext3_mark_inode_dirty(handle, inode))
227                 /* If that failed, just do the required in-core inode clear. */
228                 clear_inode(inode);
229         else
230                 ext3_free_inode(handle, inode);
231         ext3_journal_stop(handle);
232         return;
233 no_delete:
234         clear_inode(inode);     /* We must guarantee clearing of inode... */
235 }
236
237 typedef struct {
238         __le32  *p;
239         __le32  key;
240         struct buffer_head *bh;
241 } Indirect;
242
243 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
244 {
245         p->key = *(p->p = v);
246         p->bh = bh;
247 }
248
249 static int verify_chain(Indirect *from, Indirect *to)
250 {
251         while (from <= to && from->key == *from->p)
252                 from++;
253         return (from > to);
254 }
255
256 /**
257  *      ext3_block_to_path - parse the block number into array of offsets
258  *      @inode: inode in question (we are only interested in its superblock)
259  *      @i_block: block number to be parsed
260  *      @offsets: array to store the offsets in
261  *      @boundary: set this non-zero if the referred-to block is likely to be
262  *             followed (on disk) by an indirect block.
263  *
264  *      To store the locations of file's data ext3 uses a data structure common
265  *      for UNIX filesystems - tree of pointers anchored in the inode, with
266  *      data blocks at leaves and indirect blocks in intermediate nodes.
267  *      This function translates the block number into path in that tree -
268  *      return value is the path length and @offsets[n] is the offset of
269  *      pointer to (n+1)th node in the nth one. If @block is out of range
270  *      (negative or too large) warning is printed and zero returned.
271  *
272  *      Note: function doesn't find node addresses, so no IO is needed. All
273  *      we need to know is the capacity of indirect blocks (taken from the
274  *      inode->i_sb).
275  */
276
277 /*
278  * Portability note: the last comparison (check that we fit into triple
279  * indirect block) is spelled differently, because otherwise on an
280  * architecture with 32-bit longs and 8Kb pages we might get into trouble
281  * if our filesystem had 8Kb blocks. We might use long long, but that would
282  * kill us on x86. Oh, well, at least the sign propagation does not matter -
283  * i_block would have to be negative in the very beginning, so we would not
284  * get there at all.
285  */
286
287 static int ext3_block_to_path(struct inode *inode,
288                         long i_block, int offsets[4], int *boundary)
289 {
290         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
291         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
292         const long direct_blocks = EXT3_NDIR_BLOCKS,
293                 indirect_blocks = ptrs,
294                 double_blocks = (1 << (ptrs_bits * 2));
295         int n = 0;
296         int final = 0;
297
298         if (i_block < 0) {
299                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
300         } else if (i_block < direct_blocks) {
301                 offsets[n++] = i_block;
302                 final = direct_blocks;
303         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
304                 offsets[n++] = EXT3_IND_BLOCK;
305                 offsets[n++] = i_block;
306                 final = ptrs;
307         } else if ((i_block -= indirect_blocks) < double_blocks) {
308                 offsets[n++] = EXT3_DIND_BLOCK;
309                 offsets[n++] = i_block >> ptrs_bits;
310                 offsets[n++] = i_block & (ptrs - 1);
311                 final = ptrs;
312         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
313                 offsets[n++] = EXT3_TIND_BLOCK;
314                 offsets[n++] = i_block >> (ptrs_bits * 2);
315                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
316                 offsets[n++] = i_block & (ptrs - 1);
317                 final = ptrs;
318         } else {
319                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
320         }
321         if (boundary)
322                 *boundary = final - 1 - (i_block & (ptrs - 1));
323         return n;
324 }
325
326 /**
327  *      ext3_get_branch - read the chain of indirect blocks leading to data
328  *      @inode: inode in question
329  *      @depth: depth of the chain (1 - direct pointer, etc.)
330  *      @offsets: offsets of pointers in inode/indirect blocks
331  *      @chain: place to store the result
332  *      @err: here we store the error value
333  *
334  *      Function fills the array of triples <key, p, bh> and returns %NULL
335  *      if everything went OK or the pointer to the last filled triple
336  *      (incomplete one) otherwise. Upon the return chain[i].key contains
337  *      the number of (i+1)-th block in the chain (as it is stored in memory,
338  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
339  *      number (it points into struct inode for i==0 and into the bh->b_data
340  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
341  *      block for i>0 and NULL for i==0. In other words, it holds the block
342  *      numbers of the chain, addresses they were taken from (and where we can
343  *      verify that chain did not change) and buffer_heads hosting these
344  *      numbers.
345  *
346  *      Function stops when it stumbles upon zero pointer (absent block)
347  *              (pointer to last triple returned, *@err == 0)
348  *      or when it gets an IO error reading an indirect block
349  *              (ditto, *@err == -EIO)
350  *      or when it notices that chain had been changed while it was reading
351  *              (ditto, *@err == -EAGAIN)
352  *      or when it reads all @depth-1 indirect blocks successfully and finds
353  *      the whole chain, all way to the data (returns %NULL, *err == 0).
354  */
355 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
356                                  Indirect chain[4], int *err)
357 {
358         struct super_block *sb = inode->i_sb;
359         Indirect *p = chain;
360         struct buffer_head *bh;
361
362         *err = 0;
363         /* i_data is not going away, no lock needed */
364         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
365         if (!p->key)
366                 goto no_block;
367         while (--depth) {
368                 bh = sb_bread(sb, le32_to_cpu(p->key));
369                 if (!bh)
370                         goto failure;
371                 /* Reader: pointers */
372                 if (!verify_chain(chain, p))
373                         goto changed;
374                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
375                 /* Reader: end */
376                 if (!p->key)
377                         goto no_block;
378         }
379         return NULL;
380
381 changed:
382         brelse(bh);
383         *err = -EAGAIN;
384         goto no_block;
385 failure:
386         *err = -EIO;
387 no_block:
388         return p;
389 }
390
391 /**
392  *      ext3_find_near - find a place for allocation with sufficient locality
393  *      @inode: owner
394  *      @ind: descriptor of indirect block.
395  *
396  *      This function returns the prefered place for block allocation.
397  *      It is used when heuristic for sequential allocation fails.
398  *      Rules are:
399  *        + if there is a block to the left of our position - allocate near it.
400  *        + if pointer will live in indirect block - allocate near that block.
401  *        + if pointer will live in inode - allocate in the same
402  *          cylinder group.
403  *
404  * In the latter case we colour the starting block by the callers PID to
405  * prevent it from clashing with concurrent allocations for a different inode
406  * in the same block group.   The PID is used here so that functionally related
407  * files will be close-by on-disk.
408  *
409  *      Caller must make sure that @ind is valid and will stay that way.
410  */
411 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
412 {
413         struct ext3_inode_info *ei = EXT3_I(inode);
414         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
415         __le32 *p;
416         ext3_fsblk_t bg_start;
417         ext3_grpblk_t colour;
418
419         /* Try to find previous block */
420         for (p = ind->p - 1; p >= start; p--) {
421                 if (*p)
422                         return le32_to_cpu(*p);
423         }
424
425         /* No such thing, so let's try location of indirect block */
426         if (ind->bh)
427                 return ind->bh->b_blocknr;
428
429         /*
430          * It is going to be referred to from the inode itself? OK, just put it
431          * into the same cylinder group then.
432          */
433         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
434         colour = (current->pid % 16) *
435                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
436         return bg_start + colour;
437 }
438
439 /**
440  *      ext3_find_goal - find a prefered place for allocation.
441  *      @inode: owner
442  *      @block:  block we want
443  *      @chain:  chain of indirect blocks
444  *      @partial: pointer to the last triple within a chain
445  *      @goal:  place to store the result.
446  *
447  *      Normally this function find the prefered place for block allocation,
448  *      stores it in *@goal and returns zero.
449  */
450
451 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
452                 Indirect chain[4], Indirect *partial)
453 {
454         struct ext3_block_alloc_info *block_i;
455
456         block_i =  EXT3_I(inode)->i_block_alloc_info;
457
458         /*
459          * try the heuristic for sequential allocation,
460          * failing that at least try to get decent locality.
461          */
462         if (block_i && (block == block_i->last_alloc_logical_block + 1)
463                 && (block_i->last_alloc_physical_block != 0)) {
464                 return block_i->last_alloc_physical_block + 1;
465         }
466
467         return ext3_find_near(inode, partial);
468 }
469
470 /**
471  *      ext3_blks_to_allocate: Look up the block map and count the number
472  *      of direct blocks need to be allocated for the given branch.
473  *
474  *      @branch: chain of indirect blocks
475  *      @k: number of blocks need for indirect blocks
476  *      @blks: number of data blocks to be mapped.
477  *      @blocks_to_boundary:  the offset in the indirect block
478  *
479  *      return the total number of blocks to be allocate, including the
480  *      direct and indirect blocks.
481  */
482 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
483                 int blocks_to_boundary)
484 {
485         unsigned long count = 0;
486
487         /*
488          * Simple case, [t,d]Indirect block(s) has not allocated yet
489          * then it's clear blocks on that path have not allocated
490          */
491         if (k > 0) {
492                 /* right now we don't handle cross boundary allocation */
493                 if (blks < blocks_to_boundary + 1)
494                         count += blks;
495                 else
496                         count += blocks_to_boundary + 1;
497                 return count;
498         }
499
500         count++;
501         while (count < blks && count <= blocks_to_boundary &&
502                 le32_to_cpu(*(branch[0].p + count)) == 0) {
503                 count++;
504         }
505         return count;
506 }
507
508 /**
509  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
510  *      @indirect_blks: the number of blocks need to allocate for indirect
511  *                      blocks
512  *
513  *      @new_blocks: on return it will store the new block numbers for
514  *      the indirect blocks(if needed) and the first direct block,
515  *      @blks:  on return it will store the total number of allocated
516  *              direct blocks
517  */
518 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
519                         ext3_fsblk_t goal, int indirect_blks, int blks,
520                         ext3_fsblk_t new_blocks[4], int *err)
521 {
522         int target, i;
523         unsigned long count = 0;
524         int index = 0;
525         ext3_fsblk_t current_block = 0;
526         int ret = 0;
527
528         /*
529          * Here we try to allocate the requested multiple blocks at once,
530          * on a best-effort basis.
531          * To build a branch, we should allocate blocks for
532          * the indirect blocks(if not allocated yet), and at least
533          * the first direct block of this branch.  That's the
534          * minimum number of blocks need to allocate(required)
535          */
536         target = blks + indirect_blks;
537
538         while (1) {
539                 count = target;
540                 /* allocating blocks for indirect blocks and direct blocks */
541                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
542                 if (*err)
543                         goto failed_out;
544
545                 target -= count;
546                 /* allocate blocks for indirect blocks */
547                 while (index < indirect_blks && count) {
548                         new_blocks[index++] = current_block++;
549                         count--;
550                 }
551
552                 if (count > 0)
553                         break;
554         }
555
556         /* save the new block number for the first direct block */
557         new_blocks[index] = current_block;
558
559         /* total number of blocks allocated for direct blocks */
560         ret = count;
561         *err = 0;
562         return ret;
563 failed_out:
564         for (i = 0; i <index; i++)
565                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
566         return ret;
567 }
568
569 /**
570  *      ext3_alloc_branch - allocate and set up a chain of blocks.
571  *      @inode: owner
572  *      @indirect_blks: number of allocated indirect blocks
573  *      @blks: number of allocated direct blocks
574  *      @offsets: offsets (in the blocks) to store the pointers to next.
575  *      @branch: place to store the chain in.
576  *
577  *      This function allocates blocks, zeroes out all but the last one,
578  *      links them into chain and (if we are synchronous) writes them to disk.
579  *      In other words, it prepares a branch that can be spliced onto the
580  *      inode. It stores the information about that chain in the branch[], in
581  *      the same format as ext3_get_branch() would do. We are calling it after
582  *      we had read the existing part of chain and partial points to the last
583  *      triple of that (one with zero ->key). Upon the exit we have the same
584  *      picture as after the successful ext3_get_block(), except that in one
585  *      place chain is disconnected - *branch->p is still zero (we did not
586  *      set the last link), but branch->key contains the number that should
587  *      be placed into *branch->p to fill that gap.
588  *
589  *      If allocation fails we free all blocks we've allocated (and forget
590  *      their buffer_heads) and return the error value the from failed
591  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
592  *      as described above and return 0.
593  */
594 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
595                         int indirect_blks, int *blks, ext3_fsblk_t goal,
596                         int *offsets, Indirect *branch)
597 {
598         int blocksize = inode->i_sb->s_blocksize;
599         int i, n = 0;
600         int err = 0;
601         struct buffer_head *bh;
602         int num;
603         ext3_fsblk_t new_blocks[4];
604         ext3_fsblk_t current_block;
605
606         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
607                                 *blks, new_blocks, &err);
608         if (err)
609                 return err;
610
611         branch[0].key = cpu_to_le32(new_blocks[0]);
612         /*
613          * metadata blocks and data blocks are allocated.
614          */
615         for (n = 1; n <= indirect_blks;  n++) {
616                 /*
617                  * Get buffer_head for parent block, zero it out
618                  * and set the pointer to new one, then send
619                  * parent to disk.
620                  */
621                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
622                 branch[n].bh = bh;
623                 lock_buffer(bh);
624                 BUFFER_TRACE(bh, "call get_create_access");
625                 err = ext3_journal_get_create_access(handle, bh);
626                 if (err) {
627                         unlock_buffer(bh);
628                         brelse(bh);
629                         goto failed;
630                 }
631
632                 memset(bh->b_data, 0, blocksize);
633                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
634                 branch[n].key = cpu_to_le32(new_blocks[n]);
635                 *branch[n].p = branch[n].key;
636                 if ( n == indirect_blks) {
637                         current_block = new_blocks[n];
638                         /*
639                          * End of chain, update the last new metablock of
640                          * the chain to point to the new allocated
641                          * data blocks numbers
642                          */
643                         for (i=1; i < num; i++)
644                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
645                 }
646                 BUFFER_TRACE(bh, "marking uptodate");
647                 set_buffer_uptodate(bh);
648                 unlock_buffer(bh);
649
650                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
651                 err = ext3_journal_dirty_metadata(handle, bh);
652                 if (err)
653                         goto failed;
654         }
655         *blks = num;
656         return err;
657 failed:
658         /* Allocation failed, free what we already allocated */
659         for (i = 1; i <= n ; i++) {
660                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
661                 ext3_journal_forget(handle, branch[i].bh);
662         }
663         for (i = 0; i <indirect_blks; i++)
664                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
665
666         ext3_free_blocks(handle, inode, new_blocks[i], num);
667
668         return err;
669 }
670
671 /**
672  * ext3_splice_branch - splice the allocated branch onto inode.
673  * @inode: owner
674  * @block: (logical) number of block we are adding
675  * @chain: chain of indirect blocks (with a missing link - see
676  *      ext3_alloc_branch)
677  * @where: location of missing link
678  * @num:   number of indirect blocks we are adding
679  * @blks:  number of direct blocks we are adding
680  *
681  * This function fills the missing link and does all housekeeping needed in
682  * inode (->i_blocks, etc.). In case of success we end up with the full
683  * chain to new block and return 0.
684  */
685 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
686                         long block, Indirect *where, int num, int blks)
687 {
688         int i;
689         int err = 0;
690         struct ext3_block_alloc_info *block_i;
691         ext3_fsblk_t current_block;
692
693         block_i = EXT3_I(inode)->i_block_alloc_info;
694         /*
695          * If we're splicing into a [td]indirect block (as opposed to the
696          * inode) then we need to get write access to the [td]indirect block
697          * before the splice.
698          */
699         if (where->bh) {
700                 BUFFER_TRACE(where->bh, "get_write_access");
701                 err = ext3_journal_get_write_access(handle, where->bh);
702                 if (err)
703                         goto err_out;
704         }
705         /* That's it */
706
707         *where->p = where->key;
708
709         /*
710          * Update the host buffer_head or inode to point to more just allocated
711          * direct blocks blocks
712          */
713         if (num == 0 && blks > 1) {
714                 current_block = le32_to_cpu(where->key) + 1;
715                 for (i = 1; i < blks; i++)
716                         *(where->p + i ) = cpu_to_le32(current_block++);
717         }
718
719         /*
720          * update the most recently allocated logical & physical block
721          * in i_block_alloc_info, to assist find the proper goal block for next
722          * allocation
723          */
724         if (block_i) {
725                 block_i->last_alloc_logical_block = block + blks - 1;
726                 block_i->last_alloc_physical_block =
727                                 le32_to_cpu(where[num].key) + blks - 1;
728         }
729
730         /* We are done with atomic stuff, now do the rest of housekeeping */
731
732         inode->i_ctime = CURRENT_TIME_SEC;
733         ext3_mark_inode_dirty(handle, inode);
734
735         /* had we spliced it onto indirect block? */
736         if (where->bh) {
737                 /*
738                  * If we spliced it onto an indirect block, we haven't
739                  * altered the inode.  Note however that if it is being spliced
740                  * onto an indirect block at the very end of the file (the
741                  * file is growing) then we *will* alter the inode to reflect
742                  * the new i_size.  But that is not done here - it is done in
743                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
744                  */
745                 jbd_debug(5, "splicing indirect only\n");
746                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
747                 err = ext3_journal_dirty_metadata(handle, where->bh);
748                 if (err)
749                         goto err_out;
750         } else {
751                 /*
752                  * OK, we spliced it into the inode itself on a direct block.
753                  * Inode was dirtied above.
754                  */
755                 jbd_debug(5, "splicing direct\n");
756         }
757         return err;
758
759 err_out:
760         for (i = 1; i <= num; i++) {
761                 BUFFER_TRACE(where[i].bh, "call journal_forget");
762                 ext3_journal_forget(handle, where[i].bh);
763                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
764         }
765         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
766
767         return err;
768 }
769
770 /*
771  * Allocation strategy is simple: if we have to allocate something, we will
772  * have to go the whole way to leaf. So let's do it before attaching anything
773  * to tree, set linkage between the newborn blocks, write them if sync is
774  * required, recheck the path, free and repeat if check fails, otherwise
775  * set the last missing link (that will protect us from any truncate-generated
776  * removals - all blocks on the path are immune now) and possibly force the
777  * write on the parent block.
778  * That has a nice additional property: no special recovery from the failed
779  * allocations is needed - we simply release blocks and do not touch anything
780  * reachable from inode.
781  *
782  * `handle' can be NULL if create == 0.
783  *
784  * The BKL may not be held on entry here.  Be sure to take it early.
785  * return > 0, # of blocks mapped or allocated.
786  * return = 0, if plain lookup failed.
787  * return < 0, error case.
788  */
789 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
790                 sector_t iblock, unsigned long maxblocks,
791                 struct buffer_head *bh_result,
792                 int create, int extend_disksize)
793 {
794         int err = -EIO;
795         int offsets[4];
796         Indirect chain[4];
797         Indirect *partial;
798         ext3_fsblk_t goal;
799         int indirect_blks;
800         int blocks_to_boundary = 0;
801         int depth;
802         struct ext3_inode_info *ei = EXT3_I(inode);
803         int count = 0;
804         ext3_fsblk_t first_block = 0;
805
806
807         J_ASSERT(handle != NULL || create == 0);
808         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
809
810         if (depth == 0)
811                 goto out;
812
813         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
814
815         /* Simplest case - block found, no allocation needed */
816         if (!partial) {
817                 first_block = le32_to_cpu(chain[depth - 1].key);
818                 clear_buffer_new(bh_result);
819                 count++;
820                 /*map more blocks*/
821                 while (count < maxblocks && count <= blocks_to_boundary) {
822                         ext3_fsblk_t blk;
823
824                         if (!verify_chain(chain, partial)) {
825                                 /*
826                                  * Indirect block might be removed by
827                                  * truncate while we were reading it.
828                                  * Handling of that case: forget what we've
829                                  * got now. Flag the err as EAGAIN, so it
830                                  * will reread.
831                                  */
832                                 err = -EAGAIN;
833                                 count = 0;
834                                 break;
835                         }
836                         blk = le32_to_cpu(*(chain[depth-1].p + count));
837
838                         if (blk == first_block + count)
839                                 count++;
840                         else
841                                 break;
842                 }
843                 if (err != -EAGAIN)
844                         goto got_it;
845         }
846
847         /* Next simple case - plain lookup or failed read of indirect block */
848         if (!create || err == -EIO)
849                 goto cleanup;
850
851         mutex_lock(&ei->truncate_mutex);
852
853         /*
854          * If the indirect block is missing while we are reading
855          * the chain(ext3_get_branch() returns -EAGAIN err), or
856          * if the chain has been changed after we grab the semaphore,
857          * (either because another process truncated this branch, or
858          * another get_block allocated this branch) re-grab the chain to see if
859          * the request block has been allocated or not.
860          *
861          * Since we already block the truncate/other get_block
862          * at this point, we will have the current copy of the chain when we
863          * splice the branch into the tree.
864          */
865         if (err == -EAGAIN || !verify_chain(chain, partial)) {
866                 while (partial > chain) {
867                         brelse(partial->bh);
868                         partial--;
869                 }
870                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
871                 if (!partial) {
872                         count++;
873                         mutex_unlock(&ei->truncate_mutex);
874                         if (err)
875                                 goto cleanup;
876                         clear_buffer_new(bh_result);
877                         goto got_it;
878                 }
879         }
880
881         /*
882          * Okay, we need to do block allocation.  Lazily initialize the block
883          * allocation info here if necessary
884         */
885         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
886                 ext3_init_block_alloc_info(inode);
887
888         goal = ext3_find_goal(inode, iblock, chain, partial);
889
890         /* the number of blocks need to allocate for [d,t]indirect blocks */
891         indirect_blks = (chain + depth) - partial - 1;
892
893         /*
894          * Next look up the indirect map to count the totoal number of
895          * direct blocks to allocate for this branch.
896          */
897         count = ext3_blks_to_allocate(partial, indirect_blks,
898                                         maxblocks, blocks_to_boundary);
899         /*
900          * Block out ext3_truncate while we alter the tree
901          */
902         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
903                                 offsets + (partial - chain), partial);
904
905         /*
906          * The ext3_splice_branch call will free and forget any buffers
907          * on the new chain if there is a failure, but that risks using
908          * up transaction credits, especially for bitmaps where the
909          * credits cannot be returned.  Can we handle this somehow?  We
910          * may need to return -EAGAIN upwards in the worst case.  --sct
911          */
912         if (!err)
913                 err = ext3_splice_branch(handle, inode, iblock,
914                                         partial, indirect_blks, count);
915         /*
916          * i_disksize growing is protected by truncate_mutex.  Don't forget to
917          * protect it if you're about to implement concurrent
918          * ext3_get_block() -bzzz
919         */
920         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
921                 ei->i_disksize = inode->i_size;
922         mutex_unlock(&ei->truncate_mutex);
923         if (err)
924                 goto cleanup;
925
926         set_buffer_new(bh_result);
927 got_it:
928         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
929         if (count > blocks_to_boundary)
930                 set_buffer_boundary(bh_result);
931         err = count;
932         /* Clean up and exit */
933         partial = chain + depth - 1;    /* the whole chain */
934 cleanup:
935         while (partial > chain) {
936                 BUFFER_TRACE(partial->bh, "call brelse");
937                 brelse(partial->bh);
938                 partial--;
939         }
940         BUFFER_TRACE(bh_result, "returned");
941 out:
942         return err;
943 }
944
945 #define DIO_CREDITS (EXT3_RESERVE_TRANS_BLOCKS + 32)
946
947 static int ext3_get_block(struct inode *inode, sector_t iblock,
948                         struct buffer_head *bh_result, int create)
949 {
950         handle_t *handle = journal_current_handle();
951         int ret = 0;
952         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
953
954         if (!create)
955                 goto get_block;         /* A read */
956
957         if (max_blocks == 1)
958                 goto get_block;         /* A single block get */
959
960         if (handle->h_transaction->t_state == T_LOCKED) {
961                 /*
962                  * Huge direct-io writes can hold off commits for long
963                  * periods of time.  Let this commit run.
964                  */
965                 ext3_journal_stop(handle);
966                 handle = ext3_journal_start(inode, DIO_CREDITS);
967                 if (IS_ERR(handle))
968                         ret = PTR_ERR(handle);
969                 goto get_block;
970         }
971
972         if (handle->h_buffer_credits <= EXT3_RESERVE_TRANS_BLOCKS) {
973                 /*
974                  * Getting low on buffer credits...
975                  */
976                 ret = ext3_journal_extend(handle, DIO_CREDITS);
977                 if (ret > 0) {
978                         /*
979                          * Couldn't extend the transaction.  Start a new one.
980                          */
981                         ret = ext3_journal_restart(handle, DIO_CREDITS);
982                 }
983         }
984
985 get_block:
986         if (ret == 0) {
987                 ret = ext3_get_blocks_handle(handle, inode, iblock,
988                                         max_blocks, bh_result, create, 0);
989                 if (ret > 0) {
990                         bh_result->b_size = (ret << inode->i_blkbits);
991                         ret = 0;
992                 }
993         }
994         return ret;
995 }
996
997 /*
998  * `handle' can be NULL if create is zero
999  */
1000 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1001                                 long block, int create, int *errp)
1002 {
1003         struct buffer_head dummy;
1004         int fatal = 0, err;
1005
1006         J_ASSERT(handle != NULL || create == 0);
1007
1008         dummy.b_state = 0;
1009         dummy.b_blocknr = -1000;
1010         buffer_trace_init(&dummy.b_history);
1011         err = ext3_get_blocks_handle(handle, inode, block, 1,
1012                                         &dummy, create, 1);
1013         /*
1014          * ext3_get_blocks_handle() returns number of blocks
1015          * mapped. 0 in case of a HOLE.
1016          */
1017         if (err > 0) {
1018                 if (err > 1)
1019                         WARN_ON(1);
1020                 err = 0;
1021         }
1022         *errp = err;
1023         if (!err && buffer_mapped(&dummy)) {
1024                 struct buffer_head *bh;
1025                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1026                 if (!bh) {
1027                         *errp = -EIO;
1028                         goto err;
1029                 }
1030                 if (buffer_new(&dummy)) {
1031                         J_ASSERT(create != 0);
1032                         J_ASSERT(handle != 0);
1033
1034                         /*
1035                          * Now that we do not always journal data, we should
1036                          * keep in mind whether this should always journal the
1037                          * new buffer as metadata.  For now, regular file
1038                          * writes use ext3_get_block instead, so it's not a
1039                          * problem.
1040                          */
1041                         lock_buffer(bh);
1042                         BUFFER_TRACE(bh, "call get_create_access");
1043                         fatal = ext3_journal_get_create_access(handle, bh);
1044                         if (!fatal && !buffer_uptodate(bh)) {
1045                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1046                                 set_buffer_uptodate(bh);
1047                         }
1048                         unlock_buffer(bh);
1049                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1050                         err = ext3_journal_dirty_metadata(handle, bh);
1051                         if (!fatal)
1052                                 fatal = err;
1053                 } else {
1054                         BUFFER_TRACE(bh, "not a new buffer");
1055                 }
1056                 if (fatal) {
1057                         *errp = fatal;
1058                         brelse(bh);
1059                         bh = NULL;
1060                 }
1061                 return bh;
1062         }
1063 err:
1064         return NULL;
1065 }
1066
1067 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1068                                int block, int create, int *err)
1069 {
1070         struct buffer_head * bh;
1071
1072         bh = ext3_getblk(handle, inode, block, create, err);
1073         if (!bh)
1074                 return bh;
1075         if (buffer_uptodate(bh))
1076                 return bh;
1077         ll_rw_block(READ_META, 1, &bh);
1078         wait_on_buffer(bh);
1079         if (buffer_uptodate(bh))
1080                 return bh;
1081         put_bh(bh);
1082         *err = -EIO;
1083         return NULL;
1084 }
1085
1086 static int walk_page_buffers(   handle_t *handle,
1087                                 struct buffer_head *head,
1088                                 unsigned from,
1089                                 unsigned to,
1090                                 int *partial,
1091                                 int (*fn)(      handle_t *handle,
1092                                                 struct buffer_head *bh))
1093 {
1094         struct buffer_head *bh;
1095         unsigned block_start, block_end;
1096         unsigned blocksize = head->b_size;
1097         int err, ret = 0;
1098         struct buffer_head *next;
1099
1100         for (   bh = head, block_start = 0;
1101                 ret == 0 && (bh != head || !block_start);
1102                 block_start = block_end, bh = next)
1103         {
1104                 next = bh->b_this_page;
1105                 block_end = block_start + blocksize;
1106                 if (block_end <= from || block_start >= to) {
1107                         if (partial && !buffer_uptodate(bh))
1108                                 *partial = 1;
1109                         continue;
1110                 }
1111                 err = (*fn)(handle, bh);
1112                 if (!ret)
1113                         ret = err;
1114         }
1115         return ret;
1116 }
1117
1118 /*
1119  * To preserve ordering, it is essential that the hole instantiation and
1120  * the data write be encapsulated in a single transaction.  We cannot
1121  * close off a transaction and start a new one between the ext3_get_block()
1122  * and the commit_write().  So doing the journal_start at the start of
1123  * prepare_write() is the right place.
1124  *
1125  * Also, this function can nest inside ext3_writepage() ->
1126  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1127  * has generated enough buffer credits to do the whole page.  So we won't
1128  * block on the journal in that case, which is good, because the caller may
1129  * be PF_MEMALLOC.
1130  *
1131  * By accident, ext3 can be reentered when a transaction is open via
1132  * quota file writes.  If we were to commit the transaction while thus
1133  * reentered, there can be a deadlock - we would be holding a quota
1134  * lock, and the commit would never complete if another thread had a
1135  * transaction open and was blocking on the quota lock - a ranking
1136  * violation.
1137  *
1138  * So what we do is to rely on the fact that journal_stop/journal_start
1139  * will _not_ run commit under these circumstances because handle->h_ref
1140  * is elevated.  We'll still have enough credits for the tiny quotafile
1141  * write.
1142  */
1143 static int do_journal_get_write_access(handle_t *handle,
1144                                         struct buffer_head *bh)
1145 {
1146         if (!buffer_mapped(bh) || buffer_freed(bh))
1147                 return 0;
1148         return ext3_journal_get_write_access(handle, bh);
1149 }
1150
1151 /*
1152  * The idea of this helper function is following:
1153  * if prepare_write has allocated some blocks, but not all of them, the
1154  * transaction must include the content of the newly allocated blocks.
1155  * This content is expected to be set to zeroes by block_prepare_write().
1156  * 2006/10/14  SAW
1157  */
1158 static int ext3_prepare_failure(struct file *file, struct page *page,
1159                                 unsigned from, unsigned to)
1160 {
1161         struct address_space *mapping;
1162         struct buffer_head *bh, *head, *next;
1163         unsigned block_start, block_end;
1164         unsigned blocksize;
1165         int ret;
1166         handle_t *handle = ext3_journal_current_handle();
1167
1168         mapping = page->mapping;
1169         if (ext3_should_writeback_data(mapping->host)) {
1170                 /* optimization: no constraints about data */
1171 skip:
1172                 return ext3_journal_stop(handle);
1173         }
1174
1175         head = page_buffers(page);
1176         blocksize = head->b_size;
1177         for (   bh = head, block_start = 0;
1178                 bh != head || !block_start;
1179                 block_start = block_end, bh = next)
1180         {
1181                 next = bh->b_this_page;
1182                 block_end = block_start + blocksize;
1183                 if (block_end <= from)
1184                         continue;
1185                 if (block_start >= to) {
1186                         block_start = to;
1187                         break;
1188                 }
1189                 if (!buffer_mapped(bh))
1190                 /* prepare_write failed on this bh */
1191                         break;
1192                 if (ext3_should_journal_data(mapping->host)) {
1193                         ret = do_journal_get_write_access(handle, bh);
1194                         if (ret) {
1195                                 ext3_journal_stop(handle);
1196                                 return ret;
1197                         }
1198                 }
1199         /*
1200          * block_start here becomes the first block where the current iteration
1201          * of prepare_write failed.
1202          */
1203         }
1204         if (block_start <= from)
1205                 goto skip;
1206
1207         /* commit allocated and zeroed buffers */
1208         return mapping->a_ops->commit_write(file, page, from, block_start);
1209 }
1210
1211 static int ext3_prepare_write(struct file *file, struct page *page,
1212                               unsigned from, unsigned to)
1213 {
1214         struct inode *inode = page->mapping->host;
1215         int ret, ret2;
1216         int needed_blocks = ext3_writepage_trans_blocks(inode);
1217         handle_t *handle;
1218         int retries = 0;
1219
1220 retry:
1221         handle = ext3_journal_start(inode, needed_blocks);
1222         if (IS_ERR(handle))
1223                 return PTR_ERR(handle);
1224         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1225                 ret = nobh_prepare_write(page, from, to, ext3_get_block);
1226         else
1227                 ret = block_prepare_write(page, from, to, ext3_get_block);
1228         if (ret)
1229                 goto failure;
1230
1231         if (ext3_should_journal_data(inode)) {
1232                 ret = walk_page_buffers(handle, page_buffers(page),
1233                                 from, to, NULL, do_journal_get_write_access);
1234                 if (ret)
1235                         /* fatal error, just put the handle and return */
1236                         journal_stop(handle);
1237         }
1238         return ret;
1239
1240 failure:
1241         ret2 = ext3_prepare_failure(file, page, from, to);
1242         if (ret2 < 0)
1243                 return ret2;
1244         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1245                 goto retry;
1246         /* retry number exceeded, or other error like -EDQUOT */
1247         return ret;
1248 }
1249
1250 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1251 {
1252         int err = journal_dirty_data(handle, bh);
1253         if (err)
1254                 ext3_journal_abort_handle(__FUNCTION__, __FUNCTION__,
1255                                                 bh, handle,err);
1256         return err;
1257 }
1258
1259 /* For commit_write() in data=journal mode */
1260 static int commit_write_fn(handle_t *handle, struct buffer_head *bh)
1261 {
1262         if (!buffer_mapped(bh) || buffer_freed(bh))
1263                 return 0;
1264         set_buffer_uptodate(bh);
1265         return ext3_journal_dirty_metadata(handle, bh);
1266 }
1267
1268 /*
1269  * We need to pick up the new inode size which generic_commit_write gave us
1270  * `file' can be NULL - eg, when called from page_symlink().
1271  *
1272  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1273  * buffers are managed internally.
1274  */
1275 static int ext3_ordered_commit_write(struct file *file, struct page *page,
1276                              unsigned from, unsigned to)
1277 {
1278         handle_t *handle = ext3_journal_current_handle();
1279         struct inode *inode = page->mapping->host;
1280         int ret = 0, ret2;
1281
1282         ret = walk_page_buffers(handle, page_buffers(page),
1283                 from, to, NULL, ext3_journal_dirty_data);
1284
1285         if (ret == 0) {
1286                 /*
1287                  * generic_commit_write() will run mark_inode_dirty() if i_size
1288                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1289                  * into that.
1290                  */
1291                 loff_t new_i_size;
1292
1293                 new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1294                 if (new_i_size > EXT3_I(inode)->i_disksize)
1295                         EXT3_I(inode)->i_disksize = new_i_size;
1296                 ret = generic_commit_write(file, page, from, to);
1297         }
1298         ret2 = ext3_journal_stop(handle);
1299         if (!ret)
1300                 ret = ret2;
1301         return ret;
1302 }
1303
1304 static int ext3_writeback_commit_write(struct file *file, struct page *page,
1305                              unsigned from, unsigned to)
1306 {
1307         handle_t *handle = ext3_journal_current_handle();
1308         struct inode *inode = page->mapping->host;
1309         int ret = 0, ret2;
1310         loff_t new_i_size;
1311
1312         new_i_size = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1313         if (new_i_size > EXT3_I(inode)->i_disksize)
1314                 EXT3_I(inode)->i_disksize = new_i_size;
1315
1316         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1317                 ret = nobh_commit_write(file, page, from, to);
1318         else
1319                 ret = generic_commit_write(file, page, from, to);
1320
1321         ret2 = ext3_journal_stop(handle);
1322         if (!ret)
1323                 ret = ret2;
1324         return ret;
1325 }
1326
1327 static int ext3_journalled_commit_write(struct file *file,
1328                         struct page *page, unsigned from, unsigned to)
1329 {
1330         handle_t *handle = ext3_journal_current_handle();
1331         struct inode *inode = page->mapping->host;
1332         int ret = 0, ret2;
1333         int partial = 0;
1334         loff_t pos;
1335
1336         /*
1337          * Here we duplicate the generic_commit_write() functionality
1338          */
1339         pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1340
1341         ret = walk_page_buffers(handle, page_buffers(page), from,
1342                                 to, &partial, commit_write_fn);
1343         if (!partial)
1344                 SetPageUptodate(page);
1345         if (pos > inode->i_size)
1346                 i_size_write(inode, pos);
1347         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1348         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1349                 EXT3_I(inode)->i_disksize = inode->i_size;
1350                 ret2 = ext3_mark_inode_dirty(handle, inode);
1351                 if (!ret)
1352                         ret = ret2;
1353         }
1354         ret2 = ext3_journal_stop(handle);
1355         if (!ret)
1356                 ret = ret2;
1357         return ret;
1358 }
1359
1360 /*
1361  * bmap() is special.  It gets used by applications such as lilo and by
1362  * the swapper to find the on-disk block of a specific piece of data.
1363  *
1364  * Naturally, this is dangerous if the block concerned is still in the
1365  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1366  * filesystem and enables swap, then they may get a nasty shock when the
1367  * data getting swapped to that swapfile suddenly gets overwritten by
1368  * the original zero's written out previously to the journal and
1369  * awaiting writeback in the kernel's buffer cache.
1370  *
1371  * So, if we see any bmap calls here on a modified, data-journaled file,
1372  * take extra steps to flush any blocks which might be in the cache.
1373  */
1374 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1375 {
1376         struct inode *inode = mapping->host;
1377         journal_t *journal;
1378         int err;
1379
1380         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1381                 /*
1382                  * This is a REALLY heavyweight approach, but the use of
1383                  * bmap on dirty files is expected to be extremely rare:
1384                  * only if we run lilo or swapon on a freshly made file
1385                  * do we expect this to happen.
1386                  *
1387                  * (bmap requires CAP_SYS_RAWIO so this does not
1388                  * represent an unprivileged user DOS attack --- we'd be
1389                  * in trouble if mortal users could trigger this path at
1390                  * will.)
1391                  *
1392                  * NB. EXT3_STATE_JDATA is not set on files other than
1393                  * regular files.  If somebody wants to bmap a directory
1394                  * or symlink and gets confused because the buffer
1395                  * hasn't yet been flushed to disk, they deserve
1396                  * everything they get.
1397                  */
1398
1399                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1400                 journal = EXT3_JOURNAL(inode);
1401                 journal_lock_updates(journal);
1402                 err = journal_flush(journal);
1403                 journal_unlock_updates(journal);
1404
1405                 if (err)
1406                         return 0;
1407         }
1408
1409         return generic_block_bmap(mapping,block,ext3_get_block);
1410 }
1411
1412 static int bget_one(handle_t *handle, struct buffer_head *bh)
1413 {
1414         get_bh(bh);
1415         return 0;
1416 }
1417
1418 static int bput_one(handle_t *handle, struct buffer_head *bh)
1419 {
1420         put_bh(bh);
1421         return 0;
1422 }
1423
1424 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1425 {
1426         if (buffer_mapped(bh))
1427                 return ext3_journal_dirty_data(handle, bh);
1428         return 0;
1429 }
1430
1431 /*
1432  * Note that we always start a transaction even if we're not journalling
1433  * data.  This is to preserve ordering: any hole instantiation within
1434  * __block_write_full_page -> ext3_get_block() should be journalled
1435  * along with the data so we don't crash and then get metadata which
1436  * refers to old data.
1437  *
1438  * In all journalling modes block_write_full_page() will start the I/O.
1439  *
1440  * Problem:
1441  *
1442  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1443  *              ext3_writepage()
1444  *
1445  * Similar for:
1446  *
1447  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1448  *
1449  * Same applies to ext3_get_block().  We will deadlock on various things like
1450  * lock_journal and i_truncate_mutex.
1451  *
1452  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1453  * allocations fail.
1454  *
1455  * 16May01: If we're reentered then journal_current_handle() will be
1456  *          non-zero. We simply *return*.
1457  *
1458  * 1 July 2001: @@@ FIXME:
1459  *   In journalled data mode, a data buffer may be metadata against the
1460  *   current transaction.  But the same file is part of a shared mapping
1461  *   and someone does a writepage() on it.
1462  *
1463  *   We will move the buffer onto the async_data list, but *after* it has
1464  *   been dirtied. So there's a small window where we have dirty data on
1465  *   BJ_Metadata.
1466  *
1467  *   Note that this only applies to the last partial page in the file.  The
1468  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1469  *   broken code anyway: it's wrong for msync()).
1470  *
1471  *   It's a rare case: affects the final partial page, for journalled data
1472  *   where the file is subject to bith write() and writepage() in the same
1473  *   transction.  To fix it we'll need a custom block_write_full_page().
1474  *   We'll probably need that anyway for journalling writepage() output.
1475  *
1476  * We don't honour synchronous mounts for writepage().  That would be
1477  * disastrous.  Any write() or metadata operation will sync the fs for
1478  * us.
1479  *
1480  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1481  * we don't need to open a transaction here.
1482  */
1483 static int ext3_ordered_writepage(struct page *page,
1484                                 struct writeback_control *wbc)
1485 {
1486         struct inode *inode = page->mapping->host;
1487         struct buffer_head *page_bufs;
1488         handle_t *handle = NULL;
1489         int ret = 0;
1490         int err;
1491
1492         J_ASSERT(PageLocked(page));
1493
1494         /*
1495          * We give up here if we're reentered, because it might be for a
1496          * different filesystem.
1497          */
1498         if (ext3_journal_current_handle())
1499                 goto out_fail;
1500
1501         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1502
1503         if (IS_ERR(handle)) {
1504                 ret = PTR_ERR(handle);
1505                 goto out_fail;
1506         }
1507
1508         if (!page_has_buffers(page)) {
1509                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1510                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1511         }
1512         page_bufs = page_buffers(page);
1513         walk_page_buffers(handle, page_bufs, 0,
1514                         PAGE_CACHE_SIZE, NULL, bget_one);
1515
1516         ret = block_write_full_page(page, ext3_get_block, wbc);
1517
1518         /*
1519          * The page can become unlocked at any point now, and
1520          * truncate can then come in and change things.  So we
1521          * can't touch *page from now on.  But *page_bufs is
1522          * safe due to elevated refcount.
1523          */
1524
1525         /*
1526          * And attach them to the current transaction.  But only if
1527          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1528          * and generally junk.
1529          */
1530         if (ret == 0) {
1531                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1532                                         NULL, journal_dirty_data_fn);
1533                 if (!ret)
1534                         ret = err;
1535         }
1536         walk_page_buffers(handle, page_bufs, 0,
1537                         PAGE_CACHE_SIZE, NULL, bput_one);
1538         err = ext3_journal_stop(handle);
1539         if (!ret)
1540                 ret = err;
1541         return ret;
1542
1543 out_fail:
1544         redirty_page_for_writepage(wbc, page);
1545         unlock_page(page);
1546         return ret;
1547 }
1548
1549 static int ext3_writeback_writepage(struct page *page,
1550                                 struct writeback_control *wbc)
1551 {
1552         struct inode *inode = page->mapping->host;
1553         handle_t *handle = NULL;
1554         int ret = 0;
1555         int err;
1556
1557         if (ext3_journal_current_handle())
1558                 goto out_fail;
1559
1560         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1561         if (IS_ERR(handle)) {
1562                 ret = PTR_ERR(handle);
1563                 goto out_fail;
1564         }
1565
1566         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1567                 ret = nobh_writepage(page, ext3_get_block, wbc);
1568         else
1569                 ret = block_write_full_page(page, ext3_get_block, wbc);
1570
1571         err = ext3_journal_stop(handle);
1572         if (!ret)
1573                 ret = err;
1574         return ret;
1575
1576 out_fail:
1577         redirty_page_for_writepage(wbc, page);
1578         unlock_page(page);
1579         return ret;
1580 }
1581
1582 static int ext3_journalled_writepage(struct page *page,
1583                                 struct writeback_control *wbc)
1584 {
1585         struct inode *inode = page->mapping->host;
1586         handle_t *handle = NULL;
1587         int ret = 0;
1588         int err;
1589
1590         if (ext3_journal_current_handle())
1591                 goto no_write;
1592
1593         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1594         if (IS_ERR(handle)) {
1595                 ret = PTR_ERR(handle);
1596                 goto no_write;
1597         }
1598
1599         if (!page_has_buffers(page) || PageChecked(page)) {
1600                 /*
1601                  * It's mmapped pagecache.  Add buffers and journal it.  There
1602                  * doesn't seem much point in redirtying the page here.
1603                  */
1604                 ClearPageChecked(page);
1605                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1606                                         ext3_get_block);
1607                 if (ret != 0) {
1608                         ext3_journal_stop(handle);
1609                         goto out_unlock;
1610                 }
1611                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1612                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1613
1614                 err = walk_page_buffers(handle, page_buffers(page), 0,
1615                                 PAGE_CACHE_SIZE, NULL, commit_write_fn);
1616                 if (ret == 0)
1617                         ret = err;
1618                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1619                 unlock_page(page);
1620         } else {
1621                 /*
1622                  * It may be a page full of checkpoint-mode buffers.  We don't
1623                  * really know unless we go poke around in the buffer_heads.
1624                  * But block_write_full_page will do the right thing.
1625                  */
1626                 ret = block_write_full_page(page, ext3_get_block, wbc);
1627         }
1628         err = ext3_journal_stop(handle);
1629         if (!ret)
1630                 ret = err;
1631 out:
1632         return ret;
1633
1634 no_write:
1635         redirty_page_for_writepage(wbc, page);
1636 out_unlock:
1637         unlock_page(page);
1638         goto out;
1639 }
1640
1641 static int ext3_readpage(struct file *file, struct page *page)
1642 {
1643         return mpage_readpage(page, ext3_get_block);
1644 }
1645
1646 static int
1647 ext3_readpages(struct file *file, struct address_space *mapping,
1648                 struct list_head *pages, unsigned nr_pages)
1649 {
1650         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1651 }
1652
1653 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1654 {
1655         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1656
1657         /*
1658          * If it's a full truncate we just forget about the pending dirtying
1659          */
1660         if (offset == 0)
1661                 ClearPageChecked(page);
1662
1663         journal_invalidatepage(journal, page, offset);
1664 }
1665
1666 static int ext3_releasepage(struct page *page, gfp_t wait)
1667 {
1668         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1669
1670         WARN_ON(PageChecked(page));
1671         if (!page_has_buffers(page))
1672                 return 0;
1673         return journal_try_to_free_buffers(journal, page, wait);
1674 }
1675
1676 /*
1677  * If the O_DIRECT write will extend the file then add this inode to the
1678  * orphan list.  So recovery will truncate it back to the original size
1679  * if the machine crashes during the write.
1680  *
1681  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1682  * crashes then stale disk data _may_ be exposed inside the file.
1683  */
1684 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1685                         const struct iovec *iov, loff_t offset,
1686                         unsigned long nr_segs)
1687 {
1688         struct file *file = iocb->ki_filp;
1689         struct inode *inode = file->f_mapping->host;
1690         struct ext3_inode_info *ei = EXT3_I(inode);
1691         handle_t *handle = NULL;
1692         ssize_t ret;
1693         int orphan = 0;
1694         size_t count = iov_length(iov, nr_segs);
1695
1696         if (rw == WRITE) {
1697                 loff_t final_size = offset + count;
1698
1699                 handle = ext3_journal_start(inode, DIO_CREDITS);
1700                 if (IS_ERR(handle)) {
1701                         ret = PTR_ERR(handle);
1702                         goto out;
1703                 }
1704                 if (final_size > inode->i_size) {
1705                         ret = ext3_orphan_add(handle, inode);
1706                         if (ret)
1707                                 goto out_stop;
1708                         orphan = 1;
1709                         ei->i_disksize = inode->i_size;
1710                 }
1711         }
1712
1713         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1714                                  offset, nr_segs,
1715                                  ext3_get_block, NULL);
1716
1717         /*
1718          * Reacquire the handle: ext3_get_block() can restart the transaction
1719          */
1720         handle = journal_current_handle();
1721
1722 out_stop:
1723         if (handle) {
1724                 int err;
1725
1726                 if (orphan && inode->i_nlink)
1727                         ext3_orphan_del(handle, inode);
1728                 if (orphan && ret > 0) {
1729                         loff_t end = offset + ret;
1730                         if (end > inode->i_size) {
1731                                 ei->i_disksize = end;
1732                                 i_size_write(inode, end);
1733                                 /*
1734                                  * We're going to return a positive `ret'
1735                                  * here due to non-zero-length I/O, so there's
1736                                  * no way of reporting error returns from
1737                                  * ext3_mark_inode_dirty() to userspace.  So
1738                                  * ignore it.
1739                                  */
1740                                 ext3_mark_inode_dirty(handle, inode);
1741                         }
1742                 }
1743                 err = ext3_journal_stop(handle);
1744                 if (ret == 0)
1745                         ret = err;
1746         }
1747 out:
1748         return ret;
1749 }
1750
1751 /*
1752  * Pages can be marked dirty completely asynchronously from ext3's journalling
1753  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1754  * much here because ->set_page_dirty is called under VFS locks.  The page is
1755  * not necessarily locked.
1756  *
1757  * We cannot just dirty the page and leave attached buffers clean, because the
1758  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1759  * or jbddirty because all the journalling code will explode.
1760  *
1761  * So what we do is to mark the page "pending dirty" and next time writepage
1762  * is called, propagate that into the buffers appropriately.
1763  */
1764 static int ext3_journalled_set_page_dirty(struct page *page)
1765 {
1766         SetPageChecked(page);
1767         return __set_page_dirty_nobuffers(page);
1768 }
1769
1770 static const struct address_space_operations ext3_ordered_aops = {
1771         .readpage       = ext3_readpage,
1772         .readpages      = ext3_readpages,
1773         .writepage      = ext3_ordered_writepage,
1774         .sync_page      = block_sync_page,
1775         .prepare_write  = ext3_prepare_write,
1776         .commit_write   = ext3_ordered_commit_write,
1777         .bmap           = ext3_bmap,
1778         .invalidatepage = ext3_invalidatepage,
1779         .releasepage    = ext3_releasepage,
1780         .direct_IO      = ext3_direct_IO,
1781         .migratepage    = buffer_migrate_page,
1782 };
1783
1784 static const struct address_space_operations ext3_writeback_aops = {
1785         .readpage       = ext3_readpage,
1786         .readpages      = ext3_readpages,
1787         .writepage      = ext3_writeback_writepage,
1788         .sync_page      = block_sync_page,
1789         .prepare_write  = ext3_prepare_write,
1790         .commit_write   = ext3_writeback_commit_write,
1791         .bmap           = ext3_bmap,
1792         .invalidatepage = ext3_invalidatepage,
1793         .releasepage    = ext3_releasepage,
1794         .direct_IO      = ext3_direct_IO,
1795         .migratepage    = buffer_migrate_page,
1796 };
1797
1798 static const struct address_space_operations ext3_journalled_aops = {
1799         .readpage       = ext3_readpage,
1800         .readpages      = ext3_readpages,
1801         .writepage      = ext3_journalled_writepage,
1802         .sync_page      = block_sync_page,
1803         .prepare_write  = ext3_prepare_write,
1804         .commit_write   = ext3_journalled_commit_write,
1805         .set_page_dirty = ext3_journalled_set_page_dirty,
1806         .bmap           = ext3_bmap,
1807         .invalidatepage = ext3_invalidatepage,
1808         .releasepage    = ext3_releasepage,
1809 };
1810
1811 void ext3_set_aops(struct inode *inode)
1812 {
1813         if (ext3_should_order_data(inode))
1814                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1815         else if (ext3_should_writeback_data(inode))
1816                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1817         else
1818                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1819 }
1820
1821 /*
1822  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1823  * up to the end of the block which corresponds to `from'.
1824  * This required during truncate. We need to physically zero the tail end
1825  * of that block so it doesn't yield old data if the file is later grown.
1826  */
1827 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1828                 struct address_space *mapping, loff_t from)
1829 {
1830         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1831         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1832         unsigned blocksize, iblock, length, pos;
1833         struct inode *inode = mapping->host;
1834         struct buffer_head *bh;
1835         int err = 0;
1836         void *kaddr;
1837
1838         blocksize = inode->i_sb->s_blocksize;
1839         length = blocksize - (offset & (blocksize - 1));
1840         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1841
1842         /*
1843          * For "nobh" option,  we can only work if we don't need to
1844          * read-in the page - otherwise we create buffers to do the IO.
1845          */
1846         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1847              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1848                 kaddr = kmap_atomic(page, KM_USER0);
1849                 memset(kaddr + offset, 0, length);
1850                 flush_dcache_page(page);
1851                 kunmap_atomic(kaddr, KM_USER0);
1852                 set_page_dirty(page);
1853                 goto unlock;
1854         }
1855
1856         if (!page_has_buffers(page))
1857                 create_empty_buffers(page, blocksize, 0);
1858
1859         /* Find the buffer that contains "offset" */
1860         bh = page_buffers(page);
1861         pos = blocksize;
1862         while (offset >= pos) {
1863                 bh = bh->b_this_page;
1864                 iblock++;
1865                 pos += blocksize;
1866         }
1867
1868         err = 0;
1869         if (buffer_freed(bh)) {
1870                 BUFFER_TRACE(bh, "freed: skip");
1871                 goto unlock;
1872         }
1873
1874         if (!buffer_mapped(bh)) {
1875                 BUFFER_TRACE(bh, "unmapped");
1876                 ext3_get_block(inode, iblock, bh, 0);
1877                 /* unmapped? It's a hole - nothing to do */
1878                 if (!buffer_mapped(bh)) {
1879                         BUFFER_TRACE(bh, "still unmapped");
1880                         goto unlock;
1881                 }
1882         }
1883
1884         /* Ok, it's mapped. Make sure it's up-to-date */
1885         if (PageUptodate(page))
1886                 set_buffer_uptodate(bh);
1887
1888         if (!buffer_uptodate(bh)) {
1889                 err = -EIO;
1890                 ll_rw_block(READ, 1, &bh);
1891                 wait_on_buffer(bh);
1892                 /* Uhhuh. Read error. Complain and punt. */
1893                 if (!buffer_uptodate(bh))
1894                         goto unlock;
1895         }
1896
1897         if (ext3_should_journal_data(inode)) {
1898                 BUFFER_TRACE(bh, "get write access");
1899                 err = ext3_journal_get_write_access(handle, bh);
1900                 if (err)
1901                         goto unlock;
1902         }
1903
1904         kaddr = kmap_atomic(page, KM_USER0);
1905         memset(kaddr + offset, 0, length);
1906         flush_dcache_page(page);
1907         kunmap_atomic(kaddr, KM_USER0);
1908
1909         BUFFER_TRACE(bh, "zeroed end of block");
1910
1911         err = 0;
1912         if (ext3_should_journal_data(inode)) {
1913                 err = ext3_journal_dirty_metadata(handle, bh);
1914         } else {
1915                 if (ext3_should_order_data(inode))
1916                         err = ext3_journal_dirty_data(handle, bh);
1917                 mark_buffer_dirty(bh);
1918         }
1919
1920 unlock:
1921         unlock_page(page);
1922         page_cache_release(page);
1923         return err;
1924 }
1925
1926 /*
1927  * Probably it should be a library function... search for first non-zero word
1928  * or memcmp with zero_page, whatever is better for particular architecture.
1929  * Linus?
1930  */
1931 static inline int all_zeroes(__le32 *p, __le32 *q)
1932 {
1933         while (p < q)
1934                 if (*p++)
1935                         return 0;
1936         return 1;
1937 }
1938
1939 /**
1940  *      ext3_find_shared - find the indirect blocks for partial truncation.
1941  *      @inode:   inode in question
1942  *      @depth:   depth of the affected branch
1943  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1944  *      @chain:   place to store the pointers to partial indirect blocks
1945  *      @top:     place to the (detached) top of branch
1946  *
1947  *      This is a helper function used by ext3_truncate().
1948  *
1949  *      When we do truncate() we may have to clean the ends of several
1950  *      indirect blocks but leave the blocks themselves alive. Block is
1951  *      partially truncated if some data below the new i_size is refered
1952  *      from it (and it is on the path to the first completely truncated
1953  *      data block, indeed).  We have to free the top of that path along
1954  *      with everything to the right of the path. Since no allocation
1955  *      past the truncation point is possible until ext3_truncate()
1956  *      finishes, we may safely do the latter, but top of branch may
1957  *      require special attention - pageout below the truncation point
1958  *      might try to populate it.
1959  *
1960  *      We atomically detach the top of branch from the tree, store the
1961  *      block number of its root in *@top, pointers to buffer_heads of
1962  *      partially truncated blocks - in @chain[].bh and pointers to
1963  *      their last elements that should not be removed - in
1964  *      @chain[].p. Return value is the pointer to last filled element
1965  *      of @chain.
1966  *
1967  *      The work left to caller to do the actual freeing of subtrees:
1968  *              a) free the subtree starting from *@top
1969  *              b) free the subtrees whose roots are stored in
1970  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1971  *              c) free the subtrees growing from the inode past the @chain[0].
1972  *                      (no partially truncated stuff there).  */
1973
1974 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1975                         int offsets[4], Indirect chain[4], __le32 *top)
1976 {
1977         Indirect *partial, *p;
1978         int k, err;
1979
1980         *top = 0;
1981         /* Make k index the deepest non-null offest + 1 */
1982         for (k = depth; k > 1 && !offsets[k-1]; k--)
1983                 ;
1984         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1985         /* Writer: pointers */
1986         if (!partial)
1987                 partial = chain + k-1;
1988         /*
1989          * If the branch acquired continuation since we've looked at it -
1990          * fine, it should all survive and (new) top doesn't belong to us.
1991          */
1992         if (!partial->key && *partial->p)
1993                 /* Writer: end */
1994                 goto no_top;
1995         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1996                 ;
1997         /*
1998          * OK, we've found the last block that must survive. The rest of our
1999          * branch should be detached before unlocking. However, if that rest
2000          * of branch is all ours and does not grow immediately from the inode
2001          * it's easier to cheat and just decrement partial->p.
2002          */
2003         if (p == chain + k - 1 && p > chain) {
2004                 p->p--;
2005         } else {
2006                 *top = *p->p;
2007                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2008 #if 0
2009                 *p->p = 0;
2010 #endif
2011         }
2012         /* Writer: end */
2013
2014         while(partial > p) {
2015                 brelse(partial->bh);
2016                 partial--;
2017         }
2018 no_top:
2019         return partial;
2020 }
2021
2022 /*
2023  * Zero a number of block pointers in either an inode or an indirect block.
2024  * If we restart the transaction we must again get write access to the
2025  * indirect block for further modification.
2026  *
2027  * We release `count' blocks on disk, but (last - first) may be greater
2028  * than `count' because there can be holes in there.
2029  */
2030 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2031                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2032                 unsigned long count, __le32 *first, __le32 *last)
2033 {
2034         __le32 *p;
2035         if (try_to_extend_transaction(handle, inode)) {
2036                 if (bh) {
2037                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2038                         ext3_journal_dirty_metadata(handle, bh);
2039                 }
2040                 ext3_mark_inode_dirty(handle, inode);
2041                 ext3_journal_test_restart(handle, inode);
2042                 if (bh) {
2043                         BUFFER_TRACE(bh, "retaking write access");
2044                         ext3_journal_get_write_access(handle, bh);
2045                 }
2046         }
2047
2048         /*
2049          * Any buffers which are on the journal will be in memory. We find
2050          * them on the hash table so journal_revoke() will run journal_forget()
2051          * on them.  We've already detached each block from the file, so
2052          * bforget() in journal_forget() should be safe.
2053          *
2054          * AKPM: turn on bforget in journal_forget()!!!
2055          */
2056         for (p = first; p < last; p++) {
2057                 u32 nr = le32_to_cpu(*p);
2058                 if (nr) {
2059                         struct buffer_head *bh;
2060
2061                         *p = 0;
2062                         bh = sb_find_get_block(inode->i_sb, nr);
2063                         ext3_forget(handle, 0, inode, bh, nr);
2064                 }
2065         }
2066
2067         ext3_free_blocks(handle, inode, block_to_free, count);
2068 }
2069
2070 /**
2071  * ext3_free_data - free a list of data blocks
2072  * @handle:     handle for this transaction
2073  * @inode:      inode we are dealing with
2074  * @this_bh:    indirect buffer_head which contains *@first and *@last
2075  * @first:      array of block numbers
2076  * @last:       points immediately past the end of array
2077  *
2078  * We are freeing all blocks refered from that array (numbers are stored as
2079  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2080  *
2081  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2082  * blocks are contiguous then releasing them at one time will only affect one
2083  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2084  * actually use a lot of journal space.
2085  *
2086  * @this_bh will be %NULL if @first and @last point into the inode's direct
2087  * block pointers.
2088  */
2089 static void ext3_free_data(handle_t *handle, struct inode *inode,
2090                            struct buffer_head *this_bh,
2091                            __le32 *first, __le32 *last)
2092 {
2093         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2094         unsigned long count = 0;            /* Number of blocks in the run */
2095         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2096                                                corresponding to
2097                                                block_to_free */
2098         ext3_fsblk_t nr;                    /* Current block # */
2099         __le32 *p;                          /* Pointer into inode/ind
2100                                                for current block */
2101         int err;
2102
2103         if (this_bh) {                          /* For indirect block */
2104                 BUFFER_TRACE(this_bh, "get_write_access");
2105                 err = ext3_journal_get_write_access(handle, this_bh);
2106                 /* Important: if we can't update the indirect pointers
2107                  * to the blocks, we can't free them. */
2108                 if (err)
2109                         return;
2110         }
2111
2112         for (p = first; p < last; p++) {
2113                 nr = le32_to_cpu(*p);
2114                 if (nr) {
2115                         /* accumulate blocks to free if they're contiguous */
2116                         if (count == 0) {
2117                                 block_to_free = nr;
2118                                 block_to_free_p = p;
2119                                 count = 1;
2120                         } else if (nr == block_to_free + count) {
2121                                 count++;
2122                         } else {
2123                                 ext3_clear_blocks(handle, inode, this_bh,
2124                                                   block_to_free,
2125                                                   count, block_to_free_p, p);
2126                                 block_to_free = nr;
2127                                 block_to_free_p = p;
2128                                 count = 1;
2129                         }
2130                 }
2131         }
2132
2133         if (count > 0)
2134                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2135                                   count, block_to_free_p, p);
2136
2137         if (this_bh) {
2138                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2139                 ext3_journal_dirty_metadata(handle, this_bh);
2140         }
2141 }
2142
2143 /**
2144  *      ext3_free_branches - free an array of branches
2145  *      @handle: JBD handle for this transaction
2146  *      @inode: inode we are dealing with
2147  *      @parent_bh: the buffer_head which contains *@first and *@last
2148  *      @first: array of block numbers
2149  *      @last:  pointer immediately past the end of array
2150  *      @depth: depth of the branches to free
2151  *
2152  *      We are freeing all blocks refered from these branches (numbers are
2153  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2154  *      appropriately.
2155  */
2156 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2157                                struct buffer_head *parent_bh,
2158                                __le32 *first, __le32 *last, int depth)
2159 {
2160         ext3_fsblk_t nr;
2161         __le32 *p;
2162
2163         if (is_handle_aborted(handle))
2164                 return;
2165
2166         if (depth--) {
2167                 struct buffer_head *bh;
2168                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2169                 p = last;
2170                 while (--p >= first) {
2171                         nr = le32_to_cpu(*p);
2172                         if (!nr)
2173                                 continue;               /* A hole */
2174
2175                         /* Go read the buffer for the next level down */
2176                         bh = sb_bread(inode->i_sb, nr);
2177
2178                         /*
2179                          * A read failure? Report error and clear slot
2180                          * (should be rare).
2181                          */
2182                         if (!bh) {
2183                                 ext3_error(inode->i_sb, "ext3_free_branches",
2184                                            "Read failure, inode=%lu, block="E3FSBLK,
2185                                            inode->i_ino, nr);
2186                                 continue;
2187                         }
2188
2189                         /* This zaps the entire block.  Bottom up. */
2190                         BUFFER_TRACE(bh, "free child branches");
2191                         ext3_free_branches(handle, inode, bh,
2192                                            (__le32*)bh->b_data,
2193                                            (__le32*)bh->b_data + addr_per_block,
2194                                            depth);
2195
2196                         /*
2197                          * We've probably journalled the indirect block several
2198                          * times during the truncate.  But it's no longer
2199                          * needed and we now drop it from the transaction via
2200                          * journal_revoke().
2201                          *
2202                          * That's easy if it's exclusively part of this
2203                          * transaction.  But if it's part of the committing
2204                          * transaction then journal_forget() will simply
2205                          * brelse() it.  That means that if the underlying
2206                          * block is reallocated in ext3_get_block(),
2207                          * unmap_underlying_metadata() will find this block
2208                          * and will try to get rid of it.  damn, damn.
2209                          *
2210                          * If this block has already been committed to the
2211                          * journal, a revoke record will be written.  And
2212                          * revoke records must be emitted *before* clearing
2213                          * this block's bit in the bitmaps.
2214                          */
2215                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2216
2217                         /*
2218                          * Everything below this this pointer has been
2219                          * released.  Now let this top-of-subtree go.
2220                          *
2221                          * We want the freeing of this indirect block to be
2222                          * atomic in the journal with the updating of the
2223                          * bitmap block which owns it.  So make some room in
2224                          * the journal.
2225                          *
2226                          * We zero the parent pointer *after* freeing its
2227                          * pointee in the bitmaps, so if extend_transaction()
2228                          * for some reason fails to put the bitmap changes and
2229                          * the release into the same transaction, recovery
2230                          * will merely complain about releasing a free block,
2231                          * rather than leaking blocks.
2232                          */
2233                         if (is_handle_aborted(handle))
2234                                 return;
2235                         if (try_to_extend_transaction(handle, inode)) {
2236                                 ext3_mark_inode_dirty(handle, inode);
2237                                 ext3_journal_test_restart(handle, inode);
2238                         }
2239
2240                         ext3_free_blocks(handle, inode, nr, 1);
2241
2242                         if (parent_bh) {
2243                                 /*
2244                                  * The block which we have just freed is
2245                                  * pointed to by an indirect block: journal it
2246                                  */
2247                                 BUFFER_TRACE(parent_bh, "get_write_access");
2248                                 if (!ext3_journal_get_write_access(handle,
2249                                                                    parent_bh)){
2250                                         *p = 0;
2251                                         BUFFER_TRACE(parent_bh,
2252                                         "call ext3_journal_dirty_metadata");
2253                                         ext3_journal_dirty_metadata(handle,
2254                                                                     parent_bh);
2255                                 }
2256                         }
2257                 }
2258         } else {
2259                 /* We have reached the bottom of the tree. */
2260                 BUFFER_TRACE(parent_bh, "free data blocks");
2261                 ext3_free_data(handle, inode, parent_bh, first, last);
2262         }
2263 }
2264
2265 /*
2266  * ext3_truncate()
2267  *
2268  * We block out ext3_get_block() block instantiations across the entire
2269  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2270  * simultaneously on behalf of the same inode.
2271  *
2272  * As we work through the truncate and commmit bits of it to the journal there
2273  * is one core, guiding principle: the file's tree must always be consistent on
2274  * disk.  We must be able to restart the truncate after a crash.
2275  *
2276  * The file's tree may be transiently inconsistent in memory (although it
2277  * probably isn't), but whenever we close off and commit a journal transaction,
2278  * the contents of (the filesystem + the journal) must be consistent and
2279  * restartable.  It's pretty simple, really: bottom up, right to left (although
2280  * left-to-right works OK too).
2281  *
2282  * Note that at recovery time, journal replay occurs *before* the restart of
2283  * truncate against the orphan inode list.
2284  *
2285  * The committed inode has the new, desired i_size (which is the same as
2286  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2287  * that this inode's truncate did not complete and it will again call
2288  * ext3_truncate() to have another go.  So there will be instantiated blocks
2289  * to the right of the truncation point in a crashed ext3 filesystem.  But
2290  * that's fine - as long as they are linked from the inode, the post-crash
2291  * ext3_truncate() run will find them and release them.
2292  */
2293 void ext3_truncate(struct inode *inode)
2294 {
2295         handle_t *handle;
2296         struct ext3_inode_info *ei = EXT3_I(inode);
2297         __le32 *i_data = ei->i_data;
2298         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2299         struct address_space *mapping = inode->i_mapping;
2300         int offsets[4];
2301         Indirect chain[4];
2302         Indirect *partial;
2303         __le32 nr = 0;
2304         int n;
2305         long last_block;
2306         unsigned blocksize = inode->i_sb->s_blocksize;
2307         struct page *page;
2308
2309         if (!(S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
2310             S_ISLNK(inode->i_mode)))
2311                 return;
2312         if (ext3_inode_is_fast_symlink(inode))
2313                 return;
2314         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2315                 return;
2316
2317         /*
2318          * We have to lock the EOF page here, because lock_page() nests
2319          * outside journal_start().
2320          */
2321         if ((inode->i_size & (blocksize - 1)) == 0) {
2322                 /* Block boundary? Nothing to do */
2323                 page = NULL;
2324         } else {
2325                 page = grab_cache_page(mapping,
2326                                 inode->i_size >> PAGE_CACHE_SHIFT);
2327                 if (!page)
2328                         return;
2329         }
2330
2331         handle = start_transaction(inode);
2332         if (IS_ERR(handle)) {
2333                 if (page) {
2334                         clear_highpage(page);
2335                         flush_dcache_page(page);
2336                         unlock_page(page);
2337                         page_cache_release(page);
2338                 }
2339                 return;         /* AKPM: return what? */
2340         }
2341
2342         last_block = (inode->i_size + blocksize-1)
2343                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2344
2345         if (page)
2346                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2347
2348         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2349         if (n == 0)
2350                 goto out_stop;  /* error */
2351
2352         /*
2353          * OK.  This truncate is going to happen.  We add the inode to the
2354          * orphan list, so that if this truncate spans multiple transactions,
2355          * and we crash, we will resume the truncate when the filesystem
2356          * recovers.  It also marks the inode dirty, to catch the new size.
2357          *
2358          * Implication: the file must always be in a sane, consistent
2359          * truncatable state while each transaction commits.
2360          */
2361         if (ext3_orphan_add(handle, inode))
2362                 goto out_stop;
2363
2364         /*
2365          * The orphan list entry will now protect us from any crash which
2366          * occurs before the truncate completes, so it is now safe to propagate
2367          * the new, shorter inode size (held for now in i_size) into the
2368          * on-disk inode. We do this via i_disksize, which is the value which
2369          * ext3 *really* writes onto the disk inode.
2370          */
2371         ei->i_disksize = inode->i_size;
2372
2373         /*
2374          * From here we block out all ext3_get_block() callers who want to
2375          * modify the block allocation tree.
2376          */
2377         mutex_lock(&ei->truncate_mutex);
2378
2379         if (n == 1) {           /* direct blocks */
2380                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2381                                i_data + EXT3_NDIR_BLOCKS);
2382                 goto do_indirects;
2383         }
2384
2385         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2386         /* Kill the top of shared branch (not detached) */
2387         if (nr) {
2388                 if (partial == chain) {
2389                         /* Shared branch grows from the inode */
2390                         ext3_free_branches(handle, inode, NULL,
2391                                            &nr, &nr+1, (chain+n-1) - partial);
2392                         *partial->p = 0;
2393                         /*
2394                          * We mark the inode dirty prior to restart,
2395                          * and prior to stop.  No need for it here.
2396                          */
2397                 } else {
2398                         /* Shared branch grows from an indirect block */
2399                         BUFFER_TRACE(partial->bh, "get_write_access");
2400                         ext3_free_branches(handle, inode, partial->bh,
2401                                         partial->p,
2402                                         partial->p+1, (chain+n-1) - partial);
2403                 }
2404         }
2405         /* Clear the ends of indirect blocks on the shared branch */
2406         while (partial > chain) {
2407                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2408                                    (__le32*)partial->bh->b_data+addr_per_block,
2409                                    (chain+n-1) - partial);
2410                 BUFFER_TRACE(partial->bh, "call brelse");
2411                 brelse (partial->bh);
2412                 partial--;
2413         }
2414 do_indirects:
2415         /* Kill the remaining (whole) subtrees */
2416         switch (offsets[0]) {
2417         default:
2418                 nr = i_data[EXT3_IND_BLOCK];
2419                 if (nr) {
2420                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2421                         i_data[EXT3_IND_BLOCK] = 0;
2422                 }
2423         case EXT3_IND_BLOCK:
2424                 nr = i_data[EXT3_DIND_BLOCK];
2425                 if (nr) {
2426                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2427                         i_data[EXT3_DIND_BLOCK] = 0;
2428                 }
2429         case EXT3_DIND_BLOCK:
2430                 nr = i_data[EXT3_TIND_BLOCK];
2431                 if (nr) {
2432                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2433                         i_data[EXT3_TIND_BLOCK] = 0;
2434                 }
2435         case EXT3_TIND_BLOCK:
2436                 ;
2437         }
2438
2439         ext3_discard_reservation(inode);
2440
2441         mutex_unlock(&ei->truncate_mutex);
2442         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2443         ext3_mark_inode_dirty(handle, inode);
2444
2445         /*
2446          * In a multi-transaction truncate, we only make the final transaction
2447          * synchronous
2448          */
2449         if (IS_SYNC(inode))
2450                 handle->h_sync = 1;
2451 out_stop:
2452         /*
2453          * If this was a simple ftruncate(), and the file will remain alive
2454          * then we need to clear up the orphan record which we created above.
2455          * However, if this was a real unlink then we were called by
2456          * ext3_delete_inode(), and we allow that function to clean up the
2457          * orphan info for us.
2458          */
2459         if (inode->i_nlink)
2460                 ext3_orphan_del(handle, inode);
2461
2462         ext3_journal_stop(handle);
2463 }
2464
2465 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2466                 unsigned long ino, struct ext3_iloc *iloc)
2467 {
2468         unsigned long desc, group_desc, block_group;
2469         unsigned long offset;
2470         ext3_fsblk_t block;
2471         struct buffer_head *bh;
2472         struct ext3_group_desc * gdp;
2473
2474         if (!ext3_valid_inum(sb, ino)) {
2475                 /*
2476                  * This error is already checked for in namei.c unless we are
2477                  * looking at an NFS filehandle, in which case no error
2478                  * report is needed
2479                  */
2480                 return 0;
2481         }
2482
2483         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2484         if (block_group >= EXT3_SB(sb)->s_groups_count) {
2485                 ext3_error(sb,"ext3_get_inode_block","group >= groups count");
2486                 return 0;
2487         }
2488         smp_rmb();
2489         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
2490         desc = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
2491         bh = EXT3_SB(sb)->s_group_desc[group_desc];
2492         if (!bh) {
2493                 ext3_error (sb, "ext3_get_inode_block",
2494                             "Descriptor not loaded");
2495                 return 0;
2496         }
2497
2498         gdp = (struct ext3_group_desc *)bh->b_data;
2499         /*
2500          * Figure out the offset within the block group inode table
2501          */
2502         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2503                 EXT3_INODE_SIZE(sb);
2504         block = le32_to_cpu(gdp[desc].bg_inode_table) +
2505                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2506
2507         iloc->block_group = block_group;
2508         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2509         return block;
2510 }
2511
2512 /*
2513  * ext3_get_inode_loc returns with an extra refcount against the inode's
2514  * underlying buffer_head on success. If 'in_mem' is true, we have all
2515  * data in memory that is needed to recreate the on-disk version of this
2516  * inode.
2517  */
2518 static int __ext3_get_inode_loc(struct inode *inode,
2519                                 struct ext3_iloc *iloc, int in_mem)
2520 {
2521         ext3_fsblk_t block;
2522         struct buffer_head *bh;
2523
2524         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2525         if (!block)
2526                 return -EIO;
2527
2528         bh = sb_getblk(inode->i_sb, block);
2529         if (!bh) {
2530                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2531                                 "unable to read inode block - "
2532                                 "inode=%lu, block="E3FSBLK,
2533                                  inode->i_ino, block);
2534                 return -EIO;
2535         }
2536         if (!buffer_uptodate(bh)) {
2537                 lock_buffer(bh);
2538                 if (buffer_uptodate(bh)) {
2539                         /* someone brought it uptodate while we waited */
2540                         unlock_buffer(bh);
2541                         goto has_buffer;
2542                 }
2543
2544                 /*
2545                  * If we have all information of the inode in memory and this
2546                  * is the only valid inode in the block, we need not read the
2547                  * block.
2548                  */
2549                 if (in_mem) {
2550                         struct buffer_head *bitmap_bh;
2551                         struct ext3_group_desc *desc;
2552                         int inodes_per_buffer;
2553                         int inode_offset, i;
2554                         int block_group;
2555                         int start;
2556
2557                         block_group = (inode->i_ino - 1) /
2558                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2559                         inodes_per_buffer = bh->b_size /
2560                                 EXT3_INODE_SIZE(inode->i_sb);
2561                         inode_offset = ((inode->i_ino - 1) %
2562                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2563                         start = inode_offset & ~(inodes_per_buffer - 1);
2564
2565                         /* Is the inode bitmap in cache? */
2566                         desc = ext3_get_group_desc(inode->i_sb,
2567                                                 block_group, NULL);
2568                         if (!desc)
2569                                 goto make_io;
2570
2571                         bitmap_bh = sb_getblk(inode->i_sb,
2572                                         le32_to_cpu(desc->bg_inode_bitmap));
2573                         if (!bitmap_bh)
2574                                 goto make_io;
2575
2576                         /*
2577                          * If the inode bitmap isn't in cache then the
2578                          * optimisation may end up performing two reads instead
2579                          * of one, so skip it.
2580                          */
2581                         if (!buffer_uptodate(bitmap_bh)) {
2582                                 brelse(bitmap_bh);
2583                                 goto make_io;
2584                         }
2585                         for (i = start; i < start + inodes_per_buffer; i++) {
2586                                 if (i == inode_offset)
2587                                         continue;
2588                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2589                                         break;
2590                         }
2591                         brelse(bitmap_bh);
2592                         if (i == start + inodes_per_buffer) {
2593                                 /* all other inodes are free, so skip I/O */
2594                                 memset(bh->b_data, 0, bh->b_size);
2595                                 set_buffer_uptodate(bh);
2596                                 unlock_buffer(bh);
2597                                 goto has_buffer;
2598                         }
2599                 }
2600
2601 make_io:
2602                 /*
2603                  * There are other valid inodes in the buffer, this inode
2604                  * has in-inode xattrs, or we don't have this inode in memory.
2605                  * Read the block from disk.
2606                  */
2607                 get_bh(bh);
2608                 bh->b_end_io = end_buffer_read_sync;
2609                 submit_bh(READ_META, bh);
2610                 wait_on_buffer(bh);
2611                 if (!buffer_uptodate(bh)) {
2612                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2613                                         "unable to read inode block - "
2614                                         "inode=%lu, block="E3FSBLK,
2615                                         inode->i_ino, block);
2616                         brelse(bh);
2617                         return -EIO;
2618                 }
2619         }
2620 has_buffer:
2621         iloc->bh = bh;
2622         return 0;
2623 }
2624
2625 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2626 {
2627         /* We have all inode data except xattrs in memory here. */
2628         return __ext3_get_inode_loc(inode, iloc,
2629                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2630 }
2631
2632 void ext3_set_inode_flags(struct inode *inode)
2633 {
2634         unsigned int flags = EXT3_I(inode)->i_flags;
2635
2636         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2637         if (flags & EXT3_SYNC_FL)
2638                 inode->i_flags |= S_SYNC;
2639         if (flags & EXT3_APPEND_FL)
2640                 inode->i_flags |= S_APPEND;
2641         if (flags & EXT3_IMMUTABLE_FL)
2642                 inode->i_flags |= S_IMMUTABLE;
2643         if (flags & EXT3_NOATIME_FL)
2644                 inode->i_flags |= S_NOATIME;
2645         if (flags & EXT3_DIRSYNC_FL)
2646                 inode->i_flags |= S_DIRSYNC;
2647 }
2648
2649 void ext3_read_inode(struct inode * inode)
2650 {
2651         struct ext3_iloc iloc;
2652         struct ext3_inode *raw_inode;
2653         struct ext3_inode_info *ei = EXT3_I(inode);
2654         struct buffer_head *bh;
2655         int block;
2656
2657 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2658         ei->i_acl = EXT3_ACL_NOT_CACHED;
2659         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2660 #endif
2661         ei->i_block_alloc_info = NULL;
2662
2663         if (__ext3_get_inode_loc(inode, &iloc, 0))
2664                 goto bad_inode;
2665         bh = iloc.bh;
2666         raw_inode = ext3_raw_inode(&iloc);
2667         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2668         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2669         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2670         if(!(test_opt (inode->i_sb, NO_UID32))) {
2671                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2672                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2673         }
2674         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2675         inode->i_size = le32_to_cpu(raw_inode->i_size);
2676         inode->i_atime.tv_sec = le32_to_cpu(raw_inode->i_atime);
2677         inode->i_ctime.tv_sec = le32_to_cpu(raw_inode->i_ctime);
2678         inode->i_mtime.tv_sec = le32_to_cpu(raw_inode->i_mtime);
2679         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2680
2681         ei->i_state = 0;
2682         ei->i_dir_start_lookup = 0;
2683         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2684         /* We now have enough fields to check if the inode was active or not.
2685          * This is needed because nfsd might try to access dead inodes
2686          * the test is that same one that e2fsck uses
2687          * NeilBrown 1999oct15
2688          */
2689         if (inode->i_nlink == 0) {
2690                 if (inode->i_mode == 0 ||
2691                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2692                         /* this inode is deleted */
2693                         brelse (bh);
2694                         goto bad_inode;
2695                 }
2696                 /* The only unlinked inodes we let through here have
2697                  * valid i_mode and are being read by the orphan
2698                  * recovery code: that's fine, we're about to complete
2699                  * the process of deleting those. */
2700         }
2701         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2702         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2703 #ifdef EXT3_FRAGMENTS
2704         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2705         ei->i_frag_no = raw_inode->i_frag;
2706         ei->i_frag_size = raw_inode->i_fsize;
2707 #endif
2708         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2709         if (!S_ISREG(inode->i_mode)) {
2710                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2711         } else {
2712                 inode->i_size |=
2713                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2714         }
2715         ei->i_disksize = inode->i_size;
2716         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2717         ei->i_block_group = iloc.block_group;
2718         /*
2719          * NOTE! The in-memory inode i_data array is in little-endian order
2720          * even on big-endian machines: we do NOT byteswap the block numbers!
2721          */
2722         for (block = 0; block < EXT3_N_BLOCKS; block++)
2723                 ei->i_data[block] = raw_inode->i_block[block];
2724         INIT_LIST_HEAD(&ei->i_orphan);
2725
2726         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2727             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2728                 /*
2729                  * When mke2fs creates big inodes it does not zero out
2730                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2731                  * so ignore those first few inodes.
2732                  */
2733                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2734                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2735                     EXT3_INODE_SIZE(inode->i_sb))
2736                         goto bad_inode;
2737                 if (ei->i_extra_isize == 0) {
2738                         /* The extra space is currently unused. Use it. */
2739                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2740                                             EXT3_GOOD_OLD_INODE_SIZE;
2741                 } else {
2742                         __le32 *magic = (void *)raw_inode +
2743                                         EXT3_GOOD_OLD_INODE_SIZE +
2744                                         ei->i_extra_isize;
2745                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2746                                  ei->i_state |= EXT3_STATE_XATTR;
2747                 }
2748         } else
2749                 ei->i_extra_isize = 0;
2750
2751         if (S_ISREG(inode->i_mode)) {
2752                 inode->i_op = &ext3_file_inode_operations;
2753                 inode->i_fop = &ext3_file_operations;
2754                 ext3_set_aops(inode);
2755         } else if (S_ISDIR(inode->i_mode)) {
2756                 inode->i_op = &ext3_dir_inode_operations;
2757                 inode->i_fop = &ext3_dir_operations;
2758         } else if (S_ISLNK(inode->i_mode)) {
2759                 if (ext3_inode_is_fast_symlink(inode))
2760                         inode->i_op = &ext3_fast_symlink_inode_operations;
2761                 else {
2762                         inode->i_op = &ext3_symlink_inode_operations;
2763                         ext3_set_aops(inode);
2764                 }
2765         } else {
2766                 inode->i_op = &ext3_special_inode_operations;
2767                 if (raw_inode->i_block[0])
2768                         init_special_inode(inode, inode->i_mode,
2769                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2770                 else
2771                         init_special_inode(inode, inode->i_mode,
2772                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2773         }
2774         brelse (iloc.bh);
2775         ext3_set_inode_flags(inode);
2776         return;
2777
2778 bad_inode:
2779         make_bad_inode(inode);
2780         return;
2781 }
2782
2783 /*
2784  * Post the struct inode info into an on-disk inode location in the
2785  * buffer-cache.  This gobbles the caller's reference to the
2786  * buffer_head in the inode location struct.
2787  *
2788  * The caller must have write access to iloc->bh.
2789  */
2790 static int ext3_do_update_inode(handle_t *handle,
2791                                 struct inode *inode,
2792                                 struct ext3_iloc *iloc)
2793 {
2794         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2795         struct ext3_inode_info *ei = EXT3_I(inode);
2796         struct buffer_head *bh = iloc->bh;
2797         int err = 0, rc, block;
2798
2799         /* For fields not not tracking in the in-memory inode,
2800          * initialise them to zero for new inodes. */
2801         if (ei->i_state & EXT3_STATE_NEW)
2802                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2803
2804         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2805         if(!(test_opt(inode->i_sb, NO_UID32))) {
2806                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2807                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2808 /*
2809  * Fix up interoperability with old kernels. Otherwise, old inodes get
2810  * re-used with the upper 16 bits of the uid/gid intact
2811  */
2812                 if(!ei->i_dtime) {
2813                         raw_inode->i_uid_high =
2814                                 cpu_to_le16(high_16_bits(inode->i_uid));
2815                         raw_inode->i_gid_high =
2816                                 cpu_to_le16(high_16_bits(inode->i_gid));
2817                 } else {
2818                         raw_inode->i_uid_high = 0;
2819                         raw_inode->i_gid_high = 0;
2820                 }
2821         } else {
2822                 raw_inode->i_uid_low =
2823                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2824                 raw_inode->i_gid_low =
2825                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2826                 raw_inode->i_uid_high = 0;
2827                 raw_inode->i_gid_high = 0;
2828         }
2829         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2830         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2831         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2832         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2833         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2834         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2835         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2836         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2837 #ifdef EXT3_FRAGMENTS
2838         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2839         raw_inode->i_frag = ei->i_frag_no;
2840         raw_inode->i_fsize = ei->i_frag_size;
2841 #endif
2842         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2843         if (!S_ISREG(inode->i_mode)) {
2844                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2845         } else {
2846                 raw_inode->i_size_high =
2847                         cpu_to_le32(ei->i_disksize >> 32);
2848                 if (ei->i_disksize > 0x7fffffffULL) {
2849                         struct super_block *sb = inode->i_sb;
2850                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2851                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2852                             EXT3_SB(sb)->s_es->s_rev_level ==
2853                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2854                                /* If this is the first large file
2855                                 * created, add a flag to the superblock.
2856                                 */
2857                                 err = ext3_journal_get_write_access(handle,
2858                                                 EXT3_SB(sb)->s_sbh);
2859                                 if (err)
2860                                         goto out_brelse;
2861                                 ext3_update_dynamic_rev(sb);
2862                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2863                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2864                                 sb->s_dirt = 1;
2865                                 handle->h_sync = 1;
2866                                 err = ext3_journal_dirty_metadata(handle,
2867                                                 EXT3_SB(sb)->s_sbh);
2868                         }
2869                 }
2870         }
2871         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2872         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2873                 if (old_valid_dev(inode->i_rdev)) {
2874                         raw_inode->i_block[0] =
2875                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2876                         raw_inode->i_block[1] = 0;
2877                 } else {
2878                         raw_inode->i_block[0] = 0;
2879                         raw_inode->i_block[1] =
2880                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2881                         raw_inode->i_block[2] = 0;
2882                 }
2883         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2884                 raw_inode->i_block[block] = ei->i_data[block];
2885
2886         if (ei->i_extra_isize)
2887                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2888
2889         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2890         rc = ext3_journal_dirty_metadata(handle, bh);
2891         if (!err)
2892                 err = rc;
2893         ei->i_state &= ~EXT3_STATE_NEW;
2894
2895 out_brelse:
2896         brelse (bh);
2897         ext3_std_error(inode->i_sb, err);
2898         return err;
2899 }
2900
2901 /*
2902  * ext3_write_inode()
2903  *
2904  * We are called from a few places:
2905  *
2906  * - Within generic_file_write() for O_SYNC files.
2907  *   Here, there will be no transaction running. We wait for any running
2908  *   trasnaction to commit.
2909  *
2910  * - Within sys_sync(), kupdate and such.
2911  *   We wait on commit, if tol to.
2912  *
2913  * - Within prune_icache() (PF_MEMALLOC == true)
2914  *   Here we simply return.  We can't afford to block kswapd on the
2915  *   journal commit.
2916  *
2917  * In all cases it is actually safe for us to return without doing anything,
2918  * because the inode has been copied into a raw inode buffer in
2919  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2920  * knfsd.
2921  *
2922  * Note that we are absolutely dependent upon all inode dirtiers doing the
2923  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2924  * which we are interested.
2925  *
2926  * It would be a bug for them to not do this.  The code:
2927  *
2928  *      mark_inode_dirty(inode)
2929  *      stuff();
2930  *      inode->i_size = expr;
2931  *
2932  * is in error because a kswapd-driven write_inode() could occur while
2933  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2934  * will no longer be on the superblock's dirty inode list.
2935  */
2936 int ext3_write_inode(struct inode *inode, int wait)
2937 {
2938         if (current->flags & PF_MEMALLOC)
2939                 return 0;
2940
2941         if (ext3_journal_current_handle()) {
2942                 jbd_debug(0, "called recursively, non-PF_MEMALLOC!\n");
2943                 dump_stack();
2944                 return -EIO;
2945         }
2946
2947         if (!wait)
2948                 return 0;
2949
2950         return ext3_force_commit(inode->i_sb);
2951 }
2952
2953 /*
2954  * ext3_setattr()
2955  *
2956  * Called from notify_change.
2957  *
2958  * We want to trap VFS attempts to truncate the file as soon as
2959  * possible.  In particular, we want to make sure that when the VFS
2960  * shrinks i_size, we put the inode on the orphan list and modify
2961  * i_disksize immediately, so that during the subsequent flushing of
2962  * dirty pages and freeing of disk blocks, we can guarantee that any
2963  * commit will leave the blocks being flushed in an unused state on
2964  * disk.  (On recovery, the inode will get truncated and the blocks will
2965  * be freed, so we have a strong guarantee that no future commit will
2966  * leave these blocks visible to the user.)
2967  *
2968  * Called with inode->sem down.
2969  */
2970 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
2971 {
2972         struct inode *inode = dentry->d_inode;
2973         int error, rc = 0;
2974         const unsigned int ia_valid = attr->ia_valid;
2975
2976         error = inode_change_ok(inode, attr);
2977         if (error)
2978                 return error;
2979
2980         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
2981                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
2982                 handle_t *handle;
2983
2984                 /* (user+group)*(old+new) structure, inode write (sb,
2985                  * inode block, ? - but truncate inode update has it) */
2986                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
2987                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
2988                 if (IS_ERR(handle)) {
2989                         error = PTR_ERR(handle);
2990                         goto err_out;
2991                 }
2992                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
2993                 if (error) {
2994                         ext3_journal_stop(handle);
2995                         return error;
2996                 }
2997                 /* Update corresponding info in inode so that everything is in
2998                  * one transaction */
2999                 if (attr->ia_valid & ATTR_UID)
3000                         inode->i_uid = attr->ia_uid;
3001                 if (attr->ia_valid & ATTR_GID)
3002                         inode->i_gid = attr->ia_gid;
3003                 error = ext3_mark_inode_dirty(handle, inode);
3004                 ext3_journal_stop(handle);
3005         }
3006
3007         if (S_ISREG(inode->i_mode) &&
3008             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3009                 handle_t *handle;
3010
3011                 handle = ext3_journal_start(inode, 3);
3012                 if (IS_ERR(handle)) {
3013                         error = PTR_ERR(handle);
3014                         goto err_out;
3015                 }
3016
3017                 error = ext3_orphan_add(handle, inode);
3018                 EXT3_I(inode)->i_disksize = attr->ia_size;
3019                 rc = ext3_mark_inode_dirty(handle, inode);
3020                 if (!error)
3021                         error = rc;
3022                 ext3_journal_stop(handle);
3023         }
3024
3025         rc = inode_setattr(inode, attr);
3026
3027         /* If inode_setattr's call to ext3_truncate failed to get a
3028          * transaction handle at all, we need to clean up the in-core
3029          * orphan list manually. */
3030         if (inode->i_nlink)
3031                 ext3_orphan_del(NULL, inode);
3032
3033         if (!rc && (ia_valid & ATTR_MODE))
3034                 rc = ext3_acl_chmod(inode);
3035
3036 err_out:
3037         ext3_std_error(inode->i_sb, error);
3038         if (!error)
3039                 error = rc;
3040         return error;
3041 }
3042
3043
3044 /*
3045  * How many blocks doth make a writepage()?
3046  *
3047  * With N blocks per page, it may be:
3048  * N data blocks
3049  * 2 indirect block
3050  * 2 dindirect
3051  * 1 tindirect
3052  * N+5 bitmap blocks (from the above)
3053  * N+5 group descriptor summary blocks
3054  * 1 inode block
3055  * 1 superblock.
3056  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3057  *
3058  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3059  *
3060  * With ordered or writeback data it's the same, less the N data blocks.
3061  *
3062  * If the inode's direct blocks can hold an integral number of pages then a
3063  * page cannot straddle two indirect blocks, and we can only touch one indirect
3064  * and dindirect block, and the "5" above becomes "3".
3065  *
3066  * This still overestimates under most circumstances.  If we were to pass the
3067  * start and end offsets in here as well we could do block_to_path() on each
3068  * block and work out the exact number of indirects which are touched.  Pah.
3069  */
3070
3071 static int ext3_writepage_trans_blocks(struct inode *inode)
3072 {
3073         int bpp = ext3_journal_blocks_per_page(inode);
3074         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3075         int ret;
3076
3077         if (ext3_should_journal_data(inode))
3078                 ret = 3 * (bpp + indirects) + 2;
3079         else
3080                 ret = 2 * (bpp + indirects) + 2;
3081
3082 #ifdef CONFIG_QUOTA
3083         /* We know that structure was already allocated during DQUOT_INIT so
3084          * we will be updating only the data blocks + inodes */
3085         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3086 #endif
3087
3088         return ret;
3089 }
3090
3091 /*
3092  * The caller must have previously called ext3_reserve_inode_write().
3093  * Give this, we know that the caller already has write access to iloc->bh.
3094  */
3095 int ext3_mark_iloc_dirty(handle_t *handle,
3096                 struct inode *inode, struct ext3_iloc *iloc)
3097 {
3098         int err = 0;
3099
3100         /* the do_update_inode consumes one bh->b_count */
3101         get_bh(iloc->bh);
3102
3103         /* ext3_do_update_inode() does journal_dirty_metadata */
3104         err = ext3_do_update_inode(handle, inode, iloc);
3105         put_bh(iloc->bh);
3106         return err;
3107 }
3108
3109 /*
3110  * On success, We end up with an outstanding reference count against
3111  * iloc->bh.  This _must_ be cleaned up later.
3112  */
3113
3114 int
3115 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3116                          struct ext3_iloc *iloc)
3117 {
3118         int err = 0;
3119         if (handle) {
3120                 err = ext3_get_inode_loc(inode, iloc);
3121                 if (!err) {
3122                         BUFFER_TRACE(iloc->bh, "get_write_access");
3123                         err = ext3_journal_get_write_access(handle, iloc->bh);
3124                         if (err) {
3125                                 brelse(iloc->bh);
3126                                 iloc->bh = NULL;
3127                         }
3128                 }
3129         }
3130         ext3_std_error(inode->i_sb, err);
3131         return err;
3132 }
3133
3134 /*
3135  * What we do here is to mark the in-core inode as clean with respect to inode
3136  * dirtiness (it may still be data-dirty).
3137  * This means that the in-core inode may be reaped by prune_icache
3138  * without having to perform any I/O.  This is a very good thing,
3139  * because *any* task may call prune_icache - even ones which
3140  * have a transaction open against a different journal.
3141  *
3142  * Is this cheating?  Not really.  Sure, we haven't written the
3143  * inode out, but prune_icache isn't a user-visible syncing function.
3144  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3145  * we start and wait on commits.
3146  *
3147  * Is this efficient/effective?  Well, we're being nice to the system
3148  * by cleaning up our inodes proactively so they can be reaped
3149  * without I/O.  But we are potentially leaving up to five seconds'
3150  * worth of inodes floating about which prune_icache wants us to
3151  * write out.  One way to fix that would be to get prune_icache()
3152  * to do a write_super() to free up some memory.  It has the desired
3153  * effect.
3154  */
3155 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3156 {
3157         struct ext3_iloc iloc;
3158         int err;
3159
3160         might_sleep();
3161         err = ext3_reserve_inode_write(handle, inode, &iloc);
3162         if (!err)
3163                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3164         return err;
3165 }
3166
3167 /*
3168  * ext3_dirty_inode() is called from __mark_inode_dirty()
3169  *
3170  * We're really interested in the case where a file is being extended.
3171  * i_size has been changed by generic_commit_write() and we thus need
3172  * to include the updated inode in the current transaction.
3173  *
3174  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3175  * are allocated to the file.
3176  *
3177  * If the inode is marked synchronous, we don't honour that here - doing
3178  * so would cause a commit on atime updates, which we don't bother doing.
3179  * We handle synchronous inodes at the highest possible level.
3180  */
3181 void ext3_dirty_inode(struct inode *inode)
3182 {
3183         handle_t *current_handle = ext3_journal_current_handle();
3184         handle_t *handle;
3185
3186         handle = ext3_journal_start(inode, 2);
3187         if (IS_ERR(handle))
3188                 goto out;
3189         if (current_handle &&
3190                 current_handle->h_transaction != handle->h_transaction) {
3191                 /* This task has a transaction open against a different fs */
3192                 printk(KERN_EMERG "%s: transactions do not match!\n",
3193                        __FUNCTION__);
3194         } else {
3195                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3196                                 current_handle);
3197                 ext3_mark_inode_dirty(handle, inode);
3198         }
3199         ext3_journal_stop(handle);
3200 out:
3201         return;
3202 }
3203
3204 #if 0
3205 /*
3206  * Bind an inode's backing buffer_head into this transaction, to prevent
3207  * it from being flushed to disk early.  Unlike
3208  * ext3_reserve_inode_write, this leaves behind no bh reference and
3209  * returns no iloc structure, so the caller needs to repeat the iloc
3210  * lookup to mark the inode dirty later.
3211  */
3212 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3213 {
3214         struct ext3_iloc iloc;
3215
3216         int err = 0;
3217         if (handle) {
3218                 err = ext3_get_inode_loc(inode, &iloc);
3219                 if (!err) {
3220                         BUFFER_TRACE(iloc.bh, "get_write_access");
3221                         err = journal_get_write_access(handle, iloc.bh);
3222                         if (!err)
3223                                 err = ext3_journal_dirty_metadata(handle,
3224                                                                   iloc.bh);
3225                         brelse(iloc.bh);
3226                 }
3227         }
3228         ext3_std_error(inode->i_sb, err);
3229         return err;
3230 }
3231 #endif
3232
3233 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3234 {
3235         journal_t *journal;
3236         handle_t *handle;
3237         int err;
3238
3239         /*
3240          * We have to be very careful here: changing a data block's
3241          * journaling status dynamically is dangerous.  If we write a
3242          * data block to the journal, change the status and then delete
3243          * that block, we risk forgetting to revoke the old log record
3244          * from the journal and so a subsequent replay can corrupt data.
3245          * So, first we make sure that the journal is empty and that
3246          * nobody is changing anything.
3247          */
3248
3249         journal = EXT3_JOURNAL(inode);
3250         if (is_journal_aborted(journal) || IS_RDONLY(inode))
3251                 return -EROFS;
3252
3253         journal_lock_updates(journal);
3254         journal_flush(journal);
3255
3256         /*
3257          * OK, there are no updates running now, and all cached data is
3258          * synced to disk.  We are now in a completely consistent state
3259          * which doesn't have anything in the journal, and we know that
3260          * no filesystem updates are running, so it is safe to modify
3261          * the inode's in-core data-journaling state flag now.
3262          */
3263
3264         if (val)
3265                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3266         else
3267                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3268         ext3_set_aops(inode);
3269
3270         journal_unlock_updates(journal);
3271
3272         /* Finally we can mark the inode as dirty. */
3273
3274         handle = ext3_journal_start(inode, 1);
3275         if (IS_ERR(handle))
3276                 return PTR_ERR(handle);
3277
3278         err = ext3_mark_inode_dirty(handle, inode);
3279         handle->h_sync = 1;
3280         ext3_journal_stop(handle);
3281         ext3_std_error(inode->i_sb, err);
3282
3283         return err;
3284 }