Merge branch 'linus' into x86/signal
[linux-2.6] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include <linux/fiemap.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, __func__,
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 preferred 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 preferred place for allocation.
441  *      @inode: owner
442  *      @block:  block we want
443  *      @partial: pointer to the last triple within a chain
444  *
445  *      Normally this function find the preferred place for block allocation,
446  *      returns it.
447  */
448
449 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
450                                    Indirect *partial)
451 {
452         struct ext3_block_alloc_info *block_i;
453
454         block_i =  EXT3_I(inode)->i_block_alloc_info;
455
456         /*
457          * try the heuristic for sequential allocation,
458          * failing that at least try to get decent locality.
459          */
460         if (block_i && (block == block_i->last_alloc_logical_block + 1)
461                 && (block_i->last_alloc_physical_block != 0)) {
462                 return block_i->last_alloc_physical_block + 1;
463         }
464
465         return ext3_find_near(inode, partial);
466 }
467
468 /**
469  *      ext3_blks_to_allocate: Look up the block map and count the number
470  *      of direct blocks need to be allocated for the given branch.
471  *
472  *      @branch: chain of indirect blocks
473  *      @k: number of blocks need for indirect blocks
474  *      @blks: number of data blocks to be mapped.
475  *      @blocks_to_boundary:  the offset in the indirect block
476  *
477  *      return the total number of blocks to be allocate, including the
478  *      direct and indirect blocks.
479  */
480 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
481                 int blocks_to_boundary)
482 {
483         unsigned long count = 0;
484
485         /*
486          * Simple case, [t,d]Indirect block(s) has not allocated yet
487          * then it's clear blocks on that path have not allocated
488          */
489         if (k > 0) {
490                 /* right now we don't handle cross boundary allocation */
491                 if (blks < blocks_to_boundary + 1)
492                         count += blks;
493                 else
494                         count += blocks_to_boundary + 1;
495                 return count;
496         }
497
498         count++;
499         while (count < blks && count <= blocks_to_boundary &&
500                 le32_to_cpu(*(branch[0].p + count)) == 0) {
501                 count++;
502         }
503         return count;
504 }
505
506 /**
507  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
508  *      @indirect_blks: the number of blocks need to allocate for indirect
509  *                      blocks
510  *
511  *      @new_blocks: on return it will store the new block numbers for
512  *      the indirect blocks(if needed) and the first direct block,
513  *      @blks:  on return it will store the total number of allocated
514  *              direct blocks
515  */
516 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
517                         ext3_fsblk_t goal, int indirect_blks, int blks,
518                         ext3_fsblk_t new_blocks[4], int *err)
519 {
520         int target, i;
521         unsigned long count = 0;
522         int index = 0;
523         ext3_fsblk_t current_block = 0;
524         int ret = 0;
525
526         /*
527          * Here we try to allocate the requested multiple blocks at once,
528          * on a best-effort basis.
529          * To build a branch, we should allocate blocks for
530          * the indirect blocks(if not allocated yet), and at least
531          * the first direct block of this branch.  That's the
532          * minimum number of blocks need to allocate(required)
533          */
534         target = blks + indirect_blks;
535
536         while (1) {
537                 count = target;
538                 /* allocating blocks for indirect blocks and direct blocks */
539                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
540                 if (*err)
541                         goto failed_out;
542
543                 target -= count;
544                 /* allocate blocks for indirect blocks */
545                 while (index < indirect_blks && count) {
546                         new_blocks[index++] = current_block++;
547                         count--;
548                 }
549
550                 if (count > 0)
551                         break;
552         }
553
554         /* save the new block number for the first direct block */
555         new_blocks[index] = current_block;
556
557         /* total number of blocks allocated for direct blocks */
558         ret = count;
559         *err = 0;
560         return ret;
561 failed_out:
562         for (i = 0; i <index; i++)
563                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
564         return ret;
565 }
566
567 /**
568  *      ext3_alloc_branch - allocate and set up a chain of blocks.
569  *      @inode: owner
570  *      @indirect_blks: number of allocated indirect blocks
571  *      @blks: number of allocated direct blocks
572  *      @offsets: offsets (in the blocks) to store the pointers to next.
573  *      @branch: place to store the chain in.
574  *
575  *      This function allocates blocks, zeroes out all but the last one,
576  *      links them into chain and (if we are synchronous) writes them to disk.
577  *      In other words, it prepares a branch that can be spliced onto the
578  *      inode. It stores the information about that chain in the branch[], in
579  *      the same format as ext3_get_branch() would do. We are calling it after
580  *      we had read the existing part of chain and partial points to the last
581  *      triple of that (one with zero ->key). Upon the exit we have the same
582  *      picture as after the successful ext3_get_block(), except that in one
583  *      place chain is disconnected - *branch->p is still zero (we did not
584  *      set the last link), but branch->key contains the number that should
585  *      be placed into *branch->p to fill that gap.
586  *
587  *      If allocation fails we free all blocks we've allocated (and forget
588  *      their buffer_heads) and return the error value the from failed
589  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
590  *      as described above and return 0.
591  */
592 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
593                         int indirect_blks, int *blks, ext3_fsblk_t goal,
594                         int *offsets, Indirect *branch)
595 {
596         int blocksize = inode->i_sb->s_blocksize;
597         int i, n = 0;
598         int err = 0;
599         struct buffer_head *bh;
600         int num;
601         ext3_fsblk_t new_blocks[4];
602         ext3_fsblk_t current_block;
603
604         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
605                                 *blks, new_blocks, &err);
606         if (err)
607                 return err;
608
609         branch[0].key = cpu_to_le32(new_blocks[0]);
610         /*
611          * metadata blocks and data blocks are allocated.
612          */
613         for (n = 1; n <= indirect_blks;  n++) {
614                 /*
615                  * Get buffer_head for parent block, zero it out
616                  * and set the pointer to new one, then send
617                  * parent to disk.
618                  */
619                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
620                 branch[n].bh = bh;
621                 lock_buffer(bh);
622                 BUFFER_TRACE(bh, "call get_create_access");
623                 err = ext3_journal_get_create_access(handle, bh);
624                 if (err) {
625                         unlock_buffer(bh);
626                         brelse(bh);
627                         goto failed;
628                 }
629
630                 memset(bh->b_data, 0, blocksize);
631                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
632                 branch[n].key = cpu_to_le32(new_blocks[n]);
633                 *branch[n].p = branch[n].key;
634                 if ( n == indirect_blks) {
635                         current_block = new_blocks[n];
636                         /*
637                          * End of chain, update the last new metablock of
638                          * the chain to point to the new allocated
639                          * data blocks numbers
640                          */
641                         for (i=1; i < num; i++)
642                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
643                 }
644                 BUFFER_TRACE(bh, "marking uptodate");
645                 set_buffer_uptodate(bh);
646                 unlock_buffer(bh);
647
648                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
649                 err = ext3_journal_dirty_metadata(handle, bh);
650                 if (err)
651                         goto failed;
652         }
653         *blks = num;
654         return err;
655 failed:
656         /* Allocation failed, free what we already allocated */
657         for (i = 1; i <= n ; i++) {
658                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
659                 ext3_journal_forget(handle, branch[i].bh);
660         }
661         for (i = 0; i <indirect_blks; i++)
662                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
663
664         ext3_free_blocks(handle, inode, new_blocks[i], num);
665
666         return err;
667 }
668
669 /**
670  * ext3_splice_branch - splice the allocated branch onto inode.
671  * @inode: owner
672  * @block: (logical) number of block we are adding
673  * @chain: chain of indirect blocks (with a missing link - see
674  *      ext3_alloc_branch)
675  * @where: location of missing link
676  * @num:   number of indirect blocks we are adding
677  * @blks:  number of direct blocks we are adding
678  *
679  * This function fills the missing link and does all housekeeping needed in
680  * inode (->i_blocks, etc.). In case of success we end up with the full
681  * chain to new block and return 0.
682  */
683 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
684                         long block, Indirect *where, int num, int blks)
685 {
686         int i;
687         int err = 0;
688         struct ext3_block_alloc_info *block_i;
689         ext3_fsblk_t current_block;
690
691         block_i = EXT3_I(inode)->i_block_alloc_info;
692         /*
693          * If we're splicing into a [td]indirect block (as opposed to the
694          * inode) then we need to get write access to the [td]indirect block
695          * before the splice.
696          */
697         if (where->bh) {
698                 BUFFER_TRACE(where->bh, "get_write_access");
699                 err = ext3_journal_get_write_access(handle, where->bh);
700                 if (err)
701                         goto err_out;
702         }
703         /* That's it */
704
705         *where->p = where->key;
706
707         /*
708          * Update the host buffer_head or inode to point to more just allocated
709          * direct blocks blocks
710          */
711         if (num == 0 && blks > 1) {
712                 current_block = le32_to_cpu(where->key) + 1;
713                 for (i = 1; i < blks; i++)
714                         *(where->p + i ) = cpu_to_le32(current_block++);
715         }
716
717         /*
718          * update the most recently allocated logical & physical block
719          * in i_block_alloc_info, to assist find the proper goal block for next
720          * allocation
721          */
722         if (block_i) {
723                 block_i->last_alloc_logical_block = block + blks - 1;
724                 block_i->last_alloc_physical_block =
725                                 le32_to_cpu(where[num].key) + blks - 1;
726         }
727
728         /* We are done with atomic stuff, now do the rest of housekeeping */
729
730         inode->i_ctime = CURRENT_TIME_SEC;
731         ext3_mark_inode_dirty(handle, inode);
732
733         /* had we spliced it onto indirect block? */
734         if (where->bh) {
735                 /*
736                  * If we spliced it onto an indirect block, we haven't
737                  * altered the inode.  Note however that if it is being spliced
738                  * onto an indirect block at the very end of the file (the
739                  * file is growing) then we *will* alter the inode to reflect
740                  * the new i_size.  But that is not done here - it is done in
741                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
742                  */
743                 jbd_debug(5, "splicing indirect only\n");
744                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
745                 err = ext3_journal_dirty_metadata(handle, where->bh);
746                 if (err)
747                         goto err_out;
748         } else {
749                 /*
750                  * OK, we spliced it into the inode itself on a direct block.
751                  * Inode was dirtied above.
752                  */
753                 jbd_debug(5, "splicing direct\n");
754         }
755         return err;
756
757 err_out:
758         for (i = 1; i <= num; i++) {
759                 BUFFER_TRACE(where[i].bh, "call journal_forget");
760                 ext3_journal_forget(handle, where[i].bh);
761                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
762         }
763         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
764
765         return err;
766 }
767
768 /*
769  * Allocation strategy is simple: if we have to allocate something, we will
770  * have to go the whole way to leaf. So let's do it before attaching anything
771  * to tree, set linkage between the newborn blocks, write them if sync is
772  * required, recheck the path, free and repeat if check fails, otherwise
773  * set the last missing link (that will protect us from any truncate-generated
774  * removals - all blocks on the path are immune now) and possibly force the
775  * write on the parent block.
776  * That has a nice additional property: no special recovery from the failed
777  * allocations is needed - we simply release blocks and do not touch anything
778  * reachable from inode.
779  *
780  * `handle' can be NULL if create == 0.
781  *
782  * The BKL may not be held on entry here.  Be sure to take it early.
783  * return > 0, # of blocks mapped or allocated.
784  * return = 0, if plain lookup failed.
785  * return < 0, error case.
786  */
787 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
788                 sector_t iblock, unsigned long maxblocks,
789                 struct buffer_head *bh_result,
790                 int create, int extend_disksize)
791 {
792         int err = -EIO;
793         int offsets[4];
794         Indirect chain[4];
795         Indirect *partial;
796         ext3_fsblk_t goal;
797         int indirect_blks;
798         int blocks_to_boundary = 0;
799         int depth;
800         struct ext3_inode_info *ei = EXT3_I(inode);
801         int count = 0;
802         ext3_fsblk_t first_block = 0;
803
804
805         J_ASSERT(handle != NULL || create == 0);
806         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
807
808         if (depth == 0)
809                 goto out;
810
811         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
812
813         /* Simplest case - block found, no allocation needed */
814         if (!partial) {
815                 first_block = le32_to_cpu(chain[depth - 1].key);
816                 clear_buffer_new(bh_result);
817                 count++;
818                 /*map more blocks*/
819                 while (count < maxblocks && count <= blocks_to_boundary) {
820                         ext3_fsblk_t blk;
821
822                         if (!verify_chain(chain, partial)) {
823                                 /*
824                                  * Indirect block might be removed by
825                                  * truncate while we were reading it.
826                                  * Handling of that case: forget what we've
827                                  * got now. Flag the err as EAGAIN, so it
828                                  * will reread.
829                                  */
830                                 err = -EAGAIN;
831                                 count = 0;
832                                 break;
833                         }
834                         blk = le32_to_cpu(*(chain[depth-1].p + count));
835
836                         if (blk == first_block + count)
837                                 count++;
838                         else
839                                 break;
840                 }
841                 if (err != -EAGAIN)
842                         goto got_it;
843         }
844
845         /* Next simple case - plain lookup or failed read of indirect block */
846         if (!create || err == -EIO)
847                 goto cleanup;
848
849         mutex_lock(&ei->truncate_mutex);
850
851         /*
852          * If the indirect block is missing while we are reading
853          * the chain(ext3_get_branch() returns -EAGAIN err), or
854          * if the chain has been changed after we grab the semaphore,
855          * (either because another process truncated this branch, or
856          * another get_block allocated this branch) re-grab the chain to see if
857          * the request block has been allocated or not.
858          *
859          * Since we already block the truncate/other get_block
860          * at this point, we will have the current copy of the chain when we
861          * splice the branch into the tree.
862          */
863         if (err == -EAGAIN || !verify_chain(chain, partial)) {
864                 while (partial > chain) {
865                         brelse(partial->bh);
866                         partial--;
867                 }
868                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
869                 if (!partial) {
870                         count++;
871                         mutex_unlock(&ei->truncate_mutex);
872                         if (err)
873                                 goto cleanup;
874                         clear_buffer_new(bh_result);
875                         goto got_it;
876                 }
877         }
878
879         /*
880          * Okay, we need to do block allocation.  Lazily initialize the block
881          * allocation info here if necessary
882         */
883         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
884                 ext3_init_block_alloc_info(inode);
885
886         goal = ext3_find_goal(inode, iblock, partial);
887
888         /* the number of blocks need to allocate for [d,t]indirect blocks */
889         indirect_blks = (chain + depth) - partial - 1;
890
891         /*
892          * Next look up the indirect map to count the totoal number of
893          * direct blocks to allocate for this branch.
894          */
895         count = ext3_blks_to_allocate(partial, indirect_blks,
896                                         maxblocks, blocks_to_boundary);
897         /*
898          * Block out ext3_truncate while we alter the tree
899          */
900         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
901                                 offsets + (partial - chain), partial);
902
903         /*
904          * The ext3_splice_branch call will free and forget any buffers
905          * on the new chain if there is a failure, but that risks using
906          * up transaction credits, especially for bitmaps where the
907          * credits cannot be returned.  Can we handle this somehow?  We
908          * may need to return -EAGAIN upwards in the worst case.  --sct
909          */
910         if (!err)
911                 err = ext3_splice_branch(handle, inode, iblock,
912                                         partial, indirect_blks, count);
913         /*
914          * i_disksize growing is protected by truncate_mutex.  Don't forget to
915          * protect it if you're about to implement concurrent
916          * ext3_get_block() -bzzz
917         */
918         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
919                 ei->i_disksize = inode->i_size;
920         mutex_unlock(&ei->truncate_mutex);
921         if (err)
922                 goto cleanup;
923
924         set_buffer_new(bh_result);
925 got_it:
926         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
927         if (count > blocks_to_boundary)
928                 set_buffer_boundary(bh_result);
929         err = count;
930         /* Clean up and exit */
931         partial = chain + depth - 1;    /* the whole chain */
932 cleanup:
933         while (partial > chain) {
934                 BUFFER_TRACE(partial->bh, "call brelse");
935                 brelse(partial->bh);
936                 partial--;
937         }
938         BUFFER_TRACE(bh_result, "returned");
939 out:
940         return err;
941 }
942
943 /* Maximum number of blocks we map for direct IO at once. */
944 #define DIO_MAX_BLOCKS 4096
945 /*
946  * Number of credits we need for writing DIO_MAX_BLOCKS:
947  * We need sb + group descriptor + bitmap + inode -> 4
948  * For B blocks with A block pointers per block we need:
949  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
950  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
951  */
952 #define DIO_CREDITS 25
953
954 static int ext3_get_block(struct inode *inode, sector_t iblock,
955                         struct buffer_head *bh_result, int create)
956 {
957         handle_t *handle = ext3_journal_current_handle();
958         int ret = 0, started = 0;
959         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
960
961         if (create && !handle) {        /* Direct IO write... */
962                 if (max_blocks > DIO_MAX_BLOCKS)
963                         max_blocks = DIO_MAX_BLOCKS;
964                 handle = ext3_journal_start(inode, DIO_CREDITS +
965                                 2 * EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb));
966                 if (IS_ERR(handle)) {
967                         ret = PTR_ERR(handle);
968                         goto out;
969                 }
970                 started = 1;
971         }
972
973         ret = ext3_get_blocks_handle(handle, inode, iblock,
974                                         max_blocks, bh_result, create, 0);
975         if (ret > 0) {
976                 bh_result->b_size = (ret << inode->i_blkbits);
977                 ret = 0;
978         }
979         if (started)
980                 ext3_journal_stop(handle);
981 out:
982         return ret;
983 }
984
985 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
986                 u64 start, u64 len)
987 {
988         return generic_block_fiemap(inode, fieinfo, start, len,
989                                     ext3_get_block);
990 }
991
992 /*
993  * `handle' can be NULL if create is zero
994  */
995 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
996                                 long block, int create, int *errp)
997 {
998         struct buffer_head dummy;
999         int fatal = 0, err;
1000
1001         J_ASSERT(handle != NULL || create == 0);
1002
1003         dummy.b_state = 0;
1004         dummy.b_blocknr = -1000;
1005         buffer_trace_init(&dummy.b_history);
1006         err = ext3_get_blocks_handle(handle, inode, block, 1,
1007                                         &dummy, create, 1);
1008         /*
1009          * ext3_get_blocks_handle() returns number of blocks
1010          * mapped. 0 in case of a HOLE.
1011          */
1012         if (err > 0) {
1013                 if (err > 1)
1014                         WARN_ON(1);
1015                 err = 0;
1016         }
1017         *errp = err;
1018         if (!err && buffer_mapped(&dummy)) {
1019                 struct buffer_head *bh;
1020                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1021                 if (!bh) {
1022                         *errp = -EIO;
1023                         goto err;
1024                 }
1025                 if (buffer_new(&dummy)) {
1026                         J_ASSERT(create != 0);
1027                         J_ASSERT(handle != NULL);
1028
1029                         /*
1030                          * Now that we do not always journal data, we should
1031                          * keep in mind whether this should always journal the
1032                          * new buffer as metadata.  For now, regular file
1033                          * writes use ext3_get_block instead, so it's not a
1034                          * problem.
1035                          */
1036                         lock_buffer(bh);
1037                         BUFFER_TRACE(bh, "call get_create_access");
1038                         fatal = ext3_journal_get_create_access(handle, bh);
1039                         if (!fatal && !buffer_uptodate(bh)) {
1040                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1041                                 set_buffer_uptodate(bh);
1042                         }
1043                         unlock_buffer(bh);
1044                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1045                         err = ext3_journal_dirty_metadata(handle, bh);
1046                         if (!fatal)
1047                                 fatal = err;
1048                 } else {
1049                         BUFFER_TRACE(bh, "not a new buffer");
1050                 }
1051                 if (fatal) {
1052                         *errp = fatal;
1053                         brelse(bh);
1054                         bh = NULL;
1055                 }
1056                 return bh;
1057         }
1058 err:
1059         return NULL;
1060 }
1061
1062 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1063                                int block, int create, int *err)
1064 {
1065         struct buffer_head * bh;
1066
1067         bh = ext3_getblk(handle, inode, block, create, err);
1068         if (!bh)
1069                 return bh;
1070         if (buffer_uptodate(bh))
1071                 return bh;
1072         ll_rw_block(READ_META, 1, &bh);
1073         wait_on_buffer(bh);
1074         if (buffer_uptodate(bh))
1075                 return bh;
1076         put_bh(bh);
1077         *err = -EIO;
1078         return NULL;
1079 }
1080
1081 static int walk_page_buffers(   handle_t *handle,
1082                                 struct buffer_head *head,
1083                                 unsigned from,
1084                                 unsigned to,
1085                                 int *partial,
1086                                 int (*fn)(      handle_t *handle,
1087                                                 struct buffer_head *bh))
1088 {
1089         struct buffer_head *bh;
1090         unsigned block_start, block_end;
1091         unsigned blocksize = head->b_size;
1092         int err, ret = 0;
1093         struct buffer_head *next;
1094
1095         for (   bh = head, block_start = 0;
1096                 ret == 0 && (bh != head || !block_start);
1097                 block_start = block_end, bh = next)
1098         {
1099                 next = bh->b_this_page;
1100                 block_end = block_start + blocksize;
1101                 if (block_end <= from || block_start >= to) {
1102                         if (partial && !buffer_uptodate(bh))
1103                                 *partial = 1;
1104                         continue;
1105                 }
1106                 err = (*fn)(handle, bh);
1107                 if (!ret)
1108                         ret = err;
1109         }
1110         return ret;
1111 }
1112
1113 /*
1114  * To preserve ordering, it is essential that the hole instantiation and
1115  * the data write be encapsulated in a single transaction.  We cannot
1116  * close off a transaction and start a new one between the ext3_get_block()
1117  * and the commit_write().  So doing the journal_start at the start of
1118  * prepare_write() is the right place.
1119  *
1120  * Also, this function can nest inside ext3_writepage() ->
1121  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1122  * has generated enough buffer credits to do the whole page.  So we won't
1123  * block on the journal in that case, which is good, because the caller may
1124  * be PF_MEMALLOC.
1125  *
1126  * By accident, ext3 can be reentered when a transaction is open via
1127  * quota file writes.  If we were to commit the transaction while thus
1128  * reentered, there can be a deadlock - we would be holding a quota
1129  * lock, and the commit would never complete if another thread had a
1130  * transaction open and was blocking on the quota lock - a ranking
1131  * violation.
1132  *
1133  * So what we do is to rely on the fact that journal_stop/journal_start
1134  * will _not_ run commit under these circumstances because handle->h_ref
1135  * is elevated.  We'll still have enough credits for the tiny quotafile
1136  * write.
1137  */
1138 static int do_journal_get_write_access(handle_t *handle,
1139                                         struct buffer_head *bh)
1140 {
1141         if (!buffer_mapped(bh) || buffer_freed(bh))
1142                 return 0;
1143         return ext3_journal_get_write_access(handle, bh);
1144 }
1145
1146 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1147                                 loff_t pos, unsigned len, unsigned flags,
1148                                 struct page **pagep, void **fsdata)
1149 {
1150         struct inode *inode = mapping->host;
1151         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1152         handle_t *handle;
1153         int retries = 0;
1154         struct page *page;
1155         pgoff_t index;
1156         unsigned from, to;
1157
1158         index = pos >> PAGE_CACHE_SHIFT;
1159         from = pos & (PAGE_CACHE_SIZE - 1);
1160         to = from + len;
1161
1162 retry:
1163         page = __grab_cache_page(mapping, index);
1164         if (!page)
1165                 return -ENOMEM;
1166         *pagep = page;
1167
1168         handle = ext3_journal_start(inode, needed_blocks);
1169         if (IS_ERR(handle)) {
1170                 unlock_page(page);
1171                 page_cache_release(page);
1172                 ret = PTR_ERR(handle);
1173                 goto out;
1174         }
1175         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1176                                                         ext3_get_block);
1177         if (ret)
1178                 goto write_begin_failed;
1179
1180         if (ext3_should_journal_data(inode)) {
1181                 ret = walk_page_buffers(handle, page_buffers(page),
1182                                 from, to, NULL, do_journal_get_write_access);
1183         }
1184 write_begin_failed:
1185         if (ret) {
1186                 ext3_journal_stop(handle);
1187                 unlock_page(page);
1188                 page_cache_release(page);
1189         }
1190         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1191                 goto retry;
1192 out:
1193         return ret;
1194 }
1195
1196
1197 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1198 {
1199         int err = journal_dirty_data(handle, bh);
1200         if (err)
1201                 ext3_journal_abort_handle(__func__, __func__,
1202                                                 bh, handle, err);
1203         return err;
1204 }
1205
1206 /* For write_end() in data=journal mode */
1207 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1208 {
1209         if (!buffer_mapped(bh) || buffer_freed(bh))
1210                 return 0;
1211         set_buffer_uptodate(bh);
1212         return ext3_journal_dirty_metadata(handle, bh);
1213 }
1214
1215 /*
1216  * Generic write_end handler for ordered and writeback ext3 journal modes.
1217  * We can't use generic_write_end, because that unlocks the page and we need to
1218  * unlock the page after ext3_journal_stop, but ext3_journal_stop must run
1219  * after block_write_end.
1220  */
1221 static int ext3_generic_write_end(struct file *file,
1222                                 struct address_space *mapping,
1223                                 loff_t pos, unsigned len, unsigned copied,
1224                                 struct page *page, void *fsdata)
1225 {
1226         struct inode *inode = file->f_mapping->host;
1227
1228         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1229
1230         if (pos+copied > inode->i_size) {
1231                 i_size_write(inode, pos+copied);
1232                 mark_inode_dirty(inode);
1233         }
1234
1235         return copied;
1236 }
1237
1238 /*
1239  * We need to pick up the new inode size which generic_commit_write gave us
1240  * `file' can be NULL - eg, when called from page_symlink().
1241  *
1242  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1243  * buffers are managed internally.
1244  */
1245 static int ext3_ordered_write_end(struct file *file,
1246                                 struct address_space *mapping,
1247                                 loff_t pos, unsigned len, unsigned copied,
1248                                 struct page *page, void *fsdata)
1249 {
1250         handle_t *handle = ext3_journal_current_handle();
1251         struct inode *inode = file->f_mapping->host;
1252         unsigned from, to;
1253         int ret = 0, ret2;
1254
1255         from = pos & (PAGE_CACHE_SIZE - 1);
1256         to = from + len;
1257
1258         ret = walk_page_buffers(handle, page_buffers(page),
1259                 from, to, NULL, ext3_journal_dirty_data);
1260
1261         if (ret == 0) {
1262                 /*
1263                  * generic_write_end() will run mark_inode_dirty() if i_size
1264                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1265                  * into that.
1266                  */
1267                 loff_t new_i_size;
1268
1269                 new_i_size = pos + copied;
1270                 if (new_i_size > EXT3_I(inode)->i_disksize)
1271                         EXT3_I(inode)->i_disksize = new_i_size;
1272                 ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1273                                                         page, fsdata);
1274                 copied = ret2;
1275                 if (ret2 < 0)
1276                         ret = ret2;
1277         }
1278         ret2 = ext3_journal_stop(handle);
1279         if (!ret)
1280                 ret = ret2;
1281         unlock_page(page);
1282         page_cache_release(page);
1283
1284         return ret ? ret : copied;
1285 }
1286
1287 static int ext3_writeback_write_end(struct file *file,
1288                                 struct address_space *mapping,
1289                                 loff_t pos, unsigned len, unsigned copied,
1290                                 struct page *page, void *fsdata)
1291 {
1292         handle_t *handle = ext3_journal_current_handle();
1293         struct inode *inode = file->f_mapping->host;
1294         int ret = 0, ret2;
1295         loff_t new_i_size;
1296
1297         new_i_size = pos + copied;
1298         if (new_i_size > EXT3_I(inode)->i_disksize)
1299                 EXT3_I(inode)->i_disksize = new_i_size;
1300
1301         ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1302                                                         page, fsdata);
1303         copied = ret2;
1304         if (ret2 < 0)
1305                 ret = ret2;
1306
1307         ret2 = ext3_journal_stop(handle);
1308         if (!ret)
1309                 ret = ret2;
1310         unlock_page(page);
1311         page_cache_release(page);
1312
1313         return ret ? ret : copied;
1314 }
1315
1316 static int ext3_journalled_write_end(struct file *file,
1317                                 struct address_space *mapping,
1318                                 loff_t pos, unsigned len, unsigned copied,
1319                                 struct page *page, void *fsdata)
1320 {
1321         handle_t *handle = ext3_journal_current_handle();
1322         struct inode *inode = mapping->host;
1323         int ret = 0, ret2;
1324         int partial = 0;
1325         unsigned from, to;
1326
1327         from = pos & (PAGE_CACHE_SIZE - 1);
1328         to = from + len;
1329
1330         if (copied < len) {
1331                 if (!PageUptodate(page))
1332                         copied = 0;
1333                 page_zero_new_buffers(page, from+copied, to);
1334         }
1335
1336         ret = walk_page_buffers(handle, page_buffers(page), from,
1337                                 to, &partial, write_end_fn);
1338         if (!partial)
1339                 SetPageUptodate(page);
1340         if (pos+copied > inode->i_size)
1341                 i_size_write(inode, pos+copied);
1342         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1343         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1344                 EXT3_I(inode)->i_disksize = inode->i_size;
1345                 ret2 = ext3_mark_inode_dirty(handle, inode);
1346                 if (!ret)
1347                         ret = ret2;
1348         }
1349
1350         ret2 = ext3_journal_stop(handle);
1351         if (!ret)
1352                 ret = ret2;
1353         unlock_page(page);
1354         page_cache_release(page);
1355
1356         return ret ? ret : copied;
1357 }
1358
1359 /*
1360  * bmap() is special.  It gets used by applications such as lilo and by
1361  * the swapper to find the on-disk block of a specific piece of data.
1362  *
1363  * Naturally, this is dangerous if the block concerned is still in the
1364  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1365  * filesystem and enables swap, then they may get a nasty shock when the
1366  * data getting swapped to that swapfile suddenly gets overwritten by
1367  * the original zero's written out previously to the journal and
1368  * awaiting writeback in the kernel's buffer cache.
1369  *
1370  * So, if we see any bmap calls here on a modified, data-journaled file,
1371  * take extra steps to flush any blocks which might be in the cache.
1372  */
1373 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1374 {
1375         struct inode *inode = mapping->host;
1376         journal_t *journal;
1377         int err;
1378
1379         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1380                 /*
1381                  * This is a REALLY heavyweight approach, but the use of
1382                  * bmap on dirty files is expected to be extremely rare:
1383                  * only if we run lilo or swapon on a freshly made file
1384                  * do we expect this to happen.
1385                  *
1386                  * (bmap requires CAP_SYS_RAWIO so this does not
1387                  * represent an unprivileged user DOS attack --- we'd be
1388                  * in trouble if mortal users could trigger this path at
1389                  * will.)
1390                  *
1391                  * NB. EXT3_STATE_JDATA is not set on files other than
1392                  * regular files.  If somebody wants to bmap a directory
1393                  * or symlink and gets confused because the buffer
1394                  * hasn't yet been flushed to disk, they deserve
1395                  * everything they get.
1396                  */
1397
1398                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1399                 journal = EXT3_JOURNAL(inode);
1400                 journal_lock_updates(journal);
1401                 err = journal_flush(journal);
1402                 journal_unlock_updates(journal);
1403
1404                 if (err)
1405                         return 0;
1406         }
1407
1408         return generic_block_bmap(mapping,block,ext3_get_block);
1409 }
1410
1411 static int bget_one(handle_t *handle, struct buffer_head *bh)
1412 {
1413         get_bh(bh);
1414         return 0;
1415 }
1416
1417 static int bput_one(handle_t *handle, struct buffer_head *bh)
1418 {
1419         put_bh(bh);
1420         return 0;
1421 }
1422
1423 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1424 {
1425         if (buffer_mapped(bh))
1426                 return ext3_journal_dirty_data(handle, bh);
1427         return 0;
1428 }
1429
1430 /*
1431  * Note that we always start a transaction even if we're not journalling
1432  * data.  This is to preserve ordering: any hole instantiation within
1433  * __block_write_full_page -> ext3_get_block() should be journalled
1434  * along with the data so we don't crash and then get metadata which
1435  * refers to old data.
1436  *
1437  * In all journalling modes block_write_full_page() will start the I/O.
1438  *
1439  * Problem:
1440  *
1441  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1442  *              ext3_writepage()
1443  *
1444  * Similar for:
1445  *
1446  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1447  *
1448  * Same applies to ext3_get_block().  We will deadlock on various things like
1449  * lock_journal and i_truncate_mutex.
1450  *
1451  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1452  * allocations fail.
1453  *
1454  * 16May01: If we're reentered then journal_current_handle() will be
1455  *          non-zero. We simply *return*.
1456  *
1457  * 1 July 2001: @@@ FIXME:
1458  *   In journalled data mode, a data buffer may be metadata against the
1459  *   current transaction.  But the same file is part of a shared mapping
1460  *   and someone does a writepage() on it.
1461  *
1462  *   We will move the buffer onto the async_data list, but *after* it has
1463  *   been dirtied. So there's a small window where we have dirty data on
1464  *   BJ_Metadata.
1465  *
1466  *   Note that this only applies to the last partial page in the file.  The
1467  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1468  *   broken code anyway: it's wrong for msync()).
1469  *
1470  *   It's a rare case: affects the final partial page, for journalled data
1471  *   where the file is subject to bith write() and writepage() in the same
1472  *   transction.  To fix it we'll need a custom block_write_full_page().
1473  *   We'll probably need that anyway for journalling writepage() output.
1474  *
1475  * We don't honour synchronous mounts for writepage().  That would be
1476  * disastrous.  Any write() or metadata operation will sync the fs for
1477  * us.
1478  *
1479  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1480  * we don't need to open a transaction here.
1481  */
1482 static int ext3_ordered_writepage(struct page *page,
1483                                 struct writeback_control *wbc)
1484 {
1485         struct inode *inode = page->mapping->host;
1486         struct buffer_head *page_bufs;
1487         handle_t *handle = NULL;
1488         int ret = 0;
1489         int err;
1490
1491         J_ASSERT(PageLocked(page));
1492
1493         /*
1494          * We give up here if we're reentered, because it might be for a
1495          * different filesystem.
1496          */
1497         if (ext3_journal_current_handle())
1498                 goto out_fail;
1499
1500         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1501
1502         if (IS_ERR(handle)) {
1503                 ret = PTR_ERR(handle);
1504                 goto out_fail;
1505         }
1506
1507         if (!page_has_buffers(page)) {
1508                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1509                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1510         }
1511         page_bufs = page_buffers(page);
1512         walk_page_buffers(handle, page_bufs, 0,
1513                         PAGE_CACHE_SIZE, NULL, bget_one);
1514
1515         ret = block_write_full_page(page, ext3_get_block, wbc);
1516
1517         /*
1518          * The page can become unlocked at any point now, and
1519          * truncate can then come in and change things.  So we
1520          * can't touch *page from now on.  But *page_bufs is
1521          * safe due to elevated refcount.
1522          */
1523
1524         /*
1525          * And attach them to the current transaction.  But only if
1526          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1527          * and generally junk.
1528          */
1529         if (ret == 0) {
1530                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1531                                         NULL, journal_dirty_data_fn);
1532                 if (!ret)
1533                         ret = err;
1534         }
1535         walk_page_buffers(handle, page_bufs, 0,
1536                         PAGE_CACHE_SIZE, NULL, bput_one);
1537         err = ext3_journal_stop(handle);
1538         if (!ret)
1539                 ret = err;
1540         return ret;
1541
1542 out_fail:
1543         redirty_page_for_writepage(wbc, page);
1544         unlock_page(page);
1545         return ret;
1546 }
1547
1548 static int ext3_writeback_writepage(struct page *page,
1549                                 struct writeback_control *wbc)
1550 {
1551         struct inode *inode = page->mapping->host;
1552         handle_t *handle = NULL;
1553         int ret = 0;
1554         int err;
1555
1556         if (ext3_journal_current_handle())
1557                 goto out_fail;
1558
1559         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1560         if (IS_ERR(handle)) {
1561                 ret = PTR_ERR(handle);
1562                 goto out_fail;
1563         }
1564
1565         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1566                 ret = nobh_writepage(page, ext3_get_block, wbc);
1567         else
1568                 ret = block_write_full_page(page, ext3_get_block, wbc);
1569
1570         err = ext3_journal_stop(handle);
1571         if (!ret)
1572                 ret = err;
1573         return ret;
1574
1575 out_fail:
1576         redirty_page_for_writepage(wbc, page);
1577         unlock_page(page);
1578         return ret;
1579 }
1580
1581 static int ext3_journalled_writepage(struct page *page,
1582                                 struct writeback_control *wbc)
1583 {
1584         struct inode *inode = page->mapping->host;
1585         handle_t *handle = NULL;
1586         int ret = 0;
1587         int err;
1588
1589         if (ext3_journal_current_handle())
1590                 goto no_write;
1591
1592         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1593         if (IS_ERR(handle)) {
1594                 ret = PTR_ERR(handle);
1595                 goto no_write;
1596         }
1597
1598         if (!page_has_buffers(page) || PageChecked(page)) {
1599                 /*
1600                  * It's mmapped pagecache.  Add buffers and journal it.  There
1601                  * doesn't seem much point in redirtying the page here.
1602                  */
1603                 ClearPageChecked(page);
1604                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1605                                         ext3_get_block);
1606                 if (ret != 0) {
1607                         ext3_journal_stop(handle);
1608                         goto out_unlock;
1609                 }
1610                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1611                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1612
1613                 err = walk_page_buffers(handle, page_buffers(page), 0,
1614                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1615                 if (ret == 0)
1616                         ret = err;
1617                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1618                 unlock_page(page);
1619         } else {
1620                 /*
1621                  * It may be a page full of checkpoint-mode buffers.  We don't
1622                  * really know unless we go poke around in the buffer_heads.
1623                  * But block_write_full_page will do the right thing.
1624                  */
1625                 ret = block_write_full_page(page, ext3_get_block, wbc);
1626         }
1627         err = ext3_journal_stop(handle);
1628         if (!ret)
1629                 ret = err;
1630 out:
1631         return ret;
1632
1633 no_write:
1634         redirty_page_for_writepage(wbc, page);
1635 out_unlock:
1636         unlock_page(page);
1637         goto out;
1638 }
1639
1640 static int ext3_readpage(struct file *file, struct page *page)
1641 {
1642         return mpage_readpage(page, ext3_get_block);
1643 }
1644
1645 static int
1646 ext3_readpages(struct file *file, struct address_space *mapping,
1647                 struct list_head *pages, unsigned nr_pages)
1648 {
1649         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1650 }
1651
1652 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1653 {
1654         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1655
1656         /*
1657          * If it's a full truncate we just forget about the pending dirtying
1658          */
1659         if (offset == 0)
1660                 ClearPageChecked(page);
1661
1662         journal_invalidatepage(journal, page, offset);
1663 }
1664
1665 static int ext3_releasepage(struct page *page, gfp_t wait)
1666 {
1667         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1668
1669         WARN_ON(PageChecked(page));
1670         if (!page_has_buffers(page))
1671                 return 0;
1672         return journal_try_to_free_buffers(journal, page, wait);
1673 }
1674
1675 /*
1676  * If the O_DIRECT write will extend the file then add this inode to the
1677  * orphan list.  So recovery will truncate it back to the original size
1678  * if the machine crashes during the write.
1679  *
1680  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1681  * crashes then stale disk data _may_ be exposed inside the file. But current
1682  * VFS code falls back into buffered path in that case so we are safe.
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;
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                 if (final_size > inode->i_size) {
1700                         /* Credits for sb + inode write */
1701                         handle = ext3_journal_start(inode, 2);
1702                         if (IS_ERR(handle)) {
1703                                 ret = PTR_ERR(handle);
1704                                 goto out;
1705                         }
1706                         ret = ext3_orphan_add(handle, inode);
1707                         if (ret) {
1708                                 ext3_journal_stop(handle);
1709                                 goto out;
1710                         }
1711                         orphan = 1;
1712                         ei->i_disksize = inode->i_size;
1713                         ext3_journal_stop(handle);
1714                 }
1715         }
1716
1717         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1718                                  offset, nr_segs,
1719                                  ext3_get_block, NULL);
1720
1721         if (orphan) {
1722                 int err;
1723
1724                 /* Credits for sb + inode write */
1725                 handle = ext3_journal_start(inode, 2);
1726                 if (IS_ERR(handle)) {
1727                         /* This is really bad luck. We've written the data
1728                          * but cannot extend i_size. Bail out and pretend
1729                          * the write failed... */
1730                         ret = PTR_ERR(handle);
1731                         goto out;
1732                 }
1733                 if (inode->i_nlink)
1734                         ext3_orphan_del(handle, inode);
1735                 if (ret > 0) {
1736                         loff_t end = offset + ret;
1737                         if (end > inode->i_size) {
1738                                 ei->i_disksize = end;
1739                                 i_size_write(inode, end);
1740                                 /*
1741                                  * We're going to return a positive `ret'
1742                                  * here due to non-zero-length I/O, so there's
1743                                  * no way of reporting error returns from
1744                                  * ext3_mark_inode_dirty() to userspace.  So
1745                                  * ignore it.
1746                                  */
1747                                 ext3_mark_inode_dirty(handle, inode);
1748                         }
1749                 }
1750                 err = ext3_journal_stop(handle);
1751                 if (ret == 0)
1752                         ret = err;
1753         }
1754 out:
1755         return ret;
1756 }
1757
1758 /*
1759  * Pages can be marked dirty completely asynchronously from ext3's journalling
1760  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1761  * much here because ->set_page_dirty is called under VFS locks.  The page is
1762  * not necessarily locked.
1763  *
1764  * We cannot just dirty the page and leave attached buffers clean, because the
1765  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1766  * or jbddirty because all the journalling code will explode.
1767  *
1768  * So what we do is to mark the page "pending dirty" and next time writepage
1769  * is called, propagate that into the buffers appropriately.
1770  */
1771 static int ext3_journalled_set_page_dirty(struct page *page)
1772 {
1773         SetPageChecked(page);
1774         return __set_page_dirty_nobuffers(page);
1775 }
1776
1777 static const struct address_space_operations ext3_ordered_aops = {
1778         .readpage               = ext3_readpage,
1779         .readpages              = ext3_readpages,
1780         .writepage              = ext3_ordered_writepage,
1781         .sync_page              = block_sync_page,
1782         .write_begin            = ext3_write_begin,
1783         .write_end              = ext3_ordered_write_end,
1784         .bmap                   = ext3_bmap,
1785         .invalidatepage         = ext3_invalidatepage,
1786         .releasepage            = ext3_releasepage,
1787         .direct_IO              = ext3_direct_IO,
1788         .migratepage            = buffer_migrate_page,
1789         .is_partially_uptodate  = block_is_partially_uptodate,
1790 };
1791
1792 static const struct address_space_operations ext3_writeback_aops = {
1793         .readpage               = ext3_readpage,
1794         .readpages              = ext3_readpages,
1795         .writepage              = ext3_writeback_writepage,
1796         .sync_page              = block_sync_page,
1797         .write_begin            = ext3_write_begin,
1798         .write_end              = ext3_writeback_write_end,
1799         .bmap                   = ext3_bmap,
1800         .invalidatepage         = ext3_invalidatepage,
1801         .releasepage            = ext3_releasepage,
1802         .direct_IO              = ext3_direct_IO,
1803         .migratepage            = buffer_migrate_page,
1804         .is_partially_uptodate  = block_is_partially_uptodate,
1805 };
1806
1807 static const struct address_space_operations ext3_journalled_aops = {
1808         .readpage               = ext3_readpage,
1809         .readpages              = ext3_readpages,
1810         .writepage              = ext3_journalled_writepage,
1811         .sync_page              = block_sync_page,
1812         .write_begin            = ext3_write_begin,
1813         .write_end              = ext3_journalled_write_end,
1814         .set_page_dirty         = ext3_journalled_set_page_dirty,
1815         .bmap                   = ext3_bmap,
1816         .invalidatepage         = ext3_invalidatepage,
1817         .releasepage            = ext3_releasepage,
1818         .is_partially_uptodate  = block_is_partially_uptodate,
1819 };
1820
1821 void ext3_set_aops(struct inode *inode)
1822 {
1823         if (ext3_should_order_data(inode))
1824                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1825         else if (ext3_should_writeback_data(inode))
1826                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1827         else
1828                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1829 }
1830
1831 /*
1832  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1833  * up to the end of the block which corresponds to `from'.
1834  * This required during truncate. We need to physically zero the tail end
1835  * of that block so it doesn't yield old data if the file is later grown.
1836  */
1837 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1838                 struct address_space *mapping, loff_t from)
1839 {
1840         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1841         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1842         unsigned blocksize, iblock, length, pos;
1843         struct inode *inode = mapping->host;
1844         struct buffer_head *bh;
1845         int err = 0;
1846
1847         blocksize = inode->i_sb->s_blocksize;
1848         length = blocksize - (offset & (blocksize - 1));
1849         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1850
1851         /*
1852          * For "nobh" option,  we can only work if we don't need to
1853          * read-in the page - otherwise we create buffers to do the IO.
1854          */
1855         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1856              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1857                 zero_user(page, offset, length);
1858                 set_page_dirty(page);
1859                 goto unlock;
1860         }
1861
1862         if (!page_has_buffers(page))
1863                 create_empty_buffers(page, blocksize, 0);
1864
1865         /* Find the buffer that contains "offset" */
1866         bh = page_buffers(page);
1867         pos = blocksize;
1868         while (offset >= pos) {
1869                 bh = bh->b_this_page;
1870                 iblock++;
1871                 pos += blocksize;
1872         }
1873
1874         err = 0;
1875         if (buffer_freed(bh)) {
1876                 BUFFER_TRACE(bh, "freed: skip");
1877                 goto unlock;
1878         }
1879
1880         if (!buffer_mapped(bh)) {
1881                 BUFFER_TRACE(bh, "unmapped");
1882                 ext3_get_block(inode, iblock, bh, 0);
1883                 /* unmapped? It's a hole - nothing to do */
1884                 if (!buffer_mapped(bh)) {
1885                         BUFFER_TRACE(bh, "still unmapped");
1886                         goto unlock;
1887                 }
1888         }
1889
1890         /* Ok, it's mapped. Make sure it's up-to-date */
1891         if (PageUptodate(page))
1892                 set_buffer_uptodate(bh);
1893
1894         if (!buffer_uptodate(bh)) {
1895                 err = -EIO;
1896                 ll_rw_block(READ, 1, &bh);
1897                 wait_on_buffer(bh);
1898                 /* Uhhuh. Read error. Complain and punt. */
1899                 if (!buffer_uptodate(bh))
1900                         goto unlock;
1901         }
1902
1903         if (ext3_should_journal_data(inode)) {
1904                 BUFFER_TRACE(bh, "get write access");
1905                 err = ext3_journal_get_write_access(handle, bh);
1906                 if (err)
1907                         goto unlock;
1908         }
1909
1910         zero_user(page, offset, length);
1911         BUFFER_TRACE(bh, "zeroed end of block");
1912
1913         err = 0;
1914         if (ext3_should_journal_data(inode)) {
1915                 err = ext3_journal_dirty_metadata(handle, bh);
1916         } else {
1917                 if (ext3_should_order_data(inode))
1918                         err = ext3_journal_dirty_data(handle, bh);
1919                 mark_buffer_dirty(bh);
1920         }
1921
1922 unlock:
1923         unlock_page(page);
1924         page_cache_release(page);
1925         return err;
1926 }
1927
1928 /*
1929  * Probably it should be a library function... search for first non-zero word
1930  * or memcmp with zero_page, whatever is better for particular architecture.
1931  * Linus?
1932  */
1933 static inline int all_zeroes(__le32 *p, __le32 *q)
1934 {
1935         while (p < q)
1936                 if (*p++)
1937                         return 0;
1938         return 1;
1939 }
1940
1941 /**
1942  *      ext3_find_shared - find the indirect blocks for partial truncation.
1943  *      @inode:   inode in question
1944  *      @depth:   depth of the affected branch
1945  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1946  *      @chain:   place to store the pointers to partial indirect blocks
1947  *      @top:     place to the (detached) top of branch
1948  *
1949  *      This is a helper function used by ext3_truncate().
1950  *
1951  *      When we do truncate() we may have to clean the ends of several
1952  *      indirect blocks but leave the blocks themselves alive. Block is
1953  *      partially truncated if some data below the new i_size is refered
1954  *      from it (and it is on the path to the first completely truncated
1955  *      data block, indeed).  We have to free the top of that path along
1956  *      with everything to the right of the path. Since no allocation
1957  *      past the truncation point is possible until ext3_truncate()
1958  *      finishes, we may safely do the latter, but top of branch may
1959  *      require special attention - pageout below the truncation point
1960  *      might try to populate it.
1961  *
1962  *      We atomically detach the top of branch from the tree, store the
1963  *      block number of its root in *@top, pointers to buffer_heads of
1964  *      partially truncated blocks - in @chain[].bh and pointers to
1965  *      their last elements that should not be removed - in
1966  *      @chain[].p. Return value is the pointer to last filled element
1967  *      of @chain.
1968  *
1969  *      The work left to caller to do the actual freeing of subtrees:
1970  *              a) free the subtree starting from *@top
1971  *              b) free the subtrees whose roots are stored in
1972  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1973  *              c) free the subtrees growing from the inode past the @chain[0].
1974  *                      (no partially truncated stuff there).  */
1975
1976 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1977                         int offsets[4], Indirect chain[4], __le32 *top)
1978 {
1979         Indirect *partial, *p;
1980         int k, err;
1981
1982         *top = 0;
1983         /* Make k index the deepest non-null offest + 1 */
1984         for (k = depth; k > 1 && !offsets[k-1]; k--)
1985                 ;
1986         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1987         /* Writer: pointers */
1988         if (!partial)
1989                 partial = chain + k-1;
1990         /*
1991          * If the branch acquired continuation since we've looked at it -
1992          * fine, it should all survive and (new) top doesn't belong to us.
1993          */
1994         if (!partial->key && *partial->p)
1995                 /* Writer: end */
1996                 goto no_top;
1997         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1998                 ;
1999         /*
2000          * OK, we've found the last block that must survive. The rest of our
2001          * branch should be detached before unlocking. However, if that rest
2002          * of branch is all ours and does not grow immediately from the inode
2003          * it's easier to cheat and just decrement partial->p.
2004          */
2005         if (p == chain + k - 1 && p > chain) {
2006                 p->p--;
2007         } else {
2008                 *top = *p->p;
2009                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2010 #if 0
2011                 *p->p = 0;
2012 #endif
2013         }
2014         /* Writer: end */
2015
2016         while(partial > p) {
2017                 brelse(partial->bh);
2018                 partial--;
2019         }
2020 no_top:
2021         return partial;
2022 }
2023
2024 /*
2025  * Zero a number of block pointers in either an inode or an indirect block.
2026  * If we restart the transaction we must again get write access to the
2027  * indirect block for further modification.
2028  *
2029  * We release `count' blocks on disk, but (last - first) may be greater
2030  * than `count' because there can be holes in there.
2031  */
2032 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2033                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2034                 unsigned long count, __le32 *first, __le32 *last)
2035 {
2036         __le32 *p;
2037         if (try_to_extend_transaction(handle, inode)) {
2038                 if (bh) {
2039                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2040                         ext3_journal_dirty_metadata(handle, bh);
2041                 }
2042                 ext3_mark_inode_dirty(handle, inode);
2043                 ext3_journal_test_restart(handle, inode);
2044                 if (bh) {
2045                         BUFFER_TRACE(bh, "retaking write access");
2046                         ext3_journal_get_write_access(handle, bh);
2047                 }
2048         }
2049
2050         /*
2051          * Any buffers which are on the journal will be in memory. We find
2052          * them on the hash table so journal_revoke() will run journal_forget()
2053          * on them.  We've already detached each block from the file, so
2054          * bforget() in journal_forget() should be safe.
2055          *
2056          * AKPM: turn on bforget in journal_forget()!!!
2057          */
2058         for (p = first; p < last; p++) {
2059                 u32 nr = le32_to_cpu(*p);
2060                 if (nr) {
2061                         struct buffer_head *bh;
2062
2063                         *p = 0;
2064                         bh = sb_find_get_block(inode->i_sb, nr);
2065                         ext3_forget(handle, 0, inode, bh, nr);
2066                 }
2067         }
2068
2069         ext3_free_blocks(handle, inode, block_to_free, count);
2070 }
2071
2072 /**
2073  * ext3_free_data - free a list of data blocks
2074  * @handle:     handle for this transaction
2075  * @inode:      inode we are dealing with
2076  * @this_bh:    indirect buffer_head which contains *@first and *@last
2077  * @first:      array of block numbers
2078  * @last:       points immediately past the end of array
2079  *
2080  * We are freeing all blocks refered from that array (numbers are stored as
2081  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2082  *
2083  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2084  * blocks are contiguous then releasing them at one time will only affect one
2085  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2086  * actually use a lot of journal space.
2087  *
2088  * @this_bh will be %NULL if @first and @last point into the inode's direct
2089  * block pointers.
2090  */
2091 static void ext3_free_data(handle_t *handle, struct inode *inode,
2092                            struct buffer_head *this_bh,
2093                            __le32 *first, __le32 *last)
2094 {
2095         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2096         unsigned long count = 0;            /* Number of blocks in the run */
2097         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2098                                                corresponding to
2099                                                block_to_free */
2100         ext3_fsblk_t nr;                    /* Current block # */
2101         __le32 *p;                          /* Pointer into inode/ind
2102                                                for current block */
2103         int err;
2104
2105         if (this_bh) {                          /* For indirect block */
2106                 BUFFER_TRACE(this_bh, "get_write_access");
2107                 err = ext3_journal_get_write_access(handle, this_bh);
2108                 /* Important: if we can't update the indirect pointers
2109                  * to the blocks, we can't free them. */
2110                 if (err)
2111                         return;
2112         }
2113
2114         for (p = first; p < last; p++) {
2115                 nr = le32_to_cpu(*p);
2116                 if (nr) {
2117                         /* accumulate blocks to free if they're contiguous */
2118                         if (count == 0) {
2119                                 block_to_free = nr;
2120                                 block_to_free_p = p;
2121                                 count = 1;
2122                         } else if (nr == block_to_free + count) {
2123                                 count++;
2124                         } else {
2125                                 ext3_clear_blocks(handle, inode, this_bh,
2126                                                   block_to_free,
2127                                                   count, block_to_free_p, p);
2128                                 block_to_free = nr;
2129                                 block_to_free_p = p;
2130                                 count = 1;
2131                         }
2132                 }
2133         }
2134
2135         if (count > 0)
2136                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2137                                   count, block_to_free_p, p);
2138
2139         if (this_bh) {
2140                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2141
2142                 /*
2143                  * The buffer head should have an attached journal head at this
2144                  * point. However, if the data is corrupted and an indirect
2145                  * block pointed to itself, it would have been detached when
2146                  * the block was cleared. Check for this instead of OOPSing.
2147                  */
2148                 if (bh2jh(this_bh))
2149                         ext3_journal_dirty_metadata(handle, this_bh);
2150                 else
2151                         ext3_error(inode->i_sb, "ext3_free_data",
2152                                    "circular indirect block detected, "
2153                                    "inode=%lu, block=%llu",
2154                                    inode->i_ino,
2155                                    (unsigned long long)this_bh->b_blocknr);
2156         }
2157 }
2158
2159 /**
2160  *      ext3_free_branches - free an array of branches
2161  *      @handle: JBD handle for this transaction
2162  *      @inode: inode we are dealing with
2163  *      @parent_bh: the buffer_head which contains *@first and *@last
2164  *      @first: array of block numbers
2165  *      @last:  pointer immediately past the end of array
2166  *      @depth: depth of the branches to free
2167  *
2168  *      We are freeing all blocks refered from these branches (numbers are
2169  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2170  *      appropriately.
2171  */
2172 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2173                                struct buffer_head *parent_bh,
2174                                __le32 *first, __le32 *last, int depth)
2175 {
2176         ext3_fsblk_t nr;
2177         __le32 *p;
2178
2179         if (is_handle_aborted(handle))
2180                 return;
2181
2182         if (depth--) {
2183                 struct buffer_head *bh;
2184                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2185                 p = last;
2186                 while (--p >= first) {
2187                         nr = le32_to_cpu(*p);
2188                         if (!nr)
2189                                 continue;               /* A hole */
2190
2191                         /* Go read the buffer for the next level down */
2192                         bh = sb_bread(inode->i_sb, nr);
2193
2194                         /*
2195                          * A read failure? Report error and clear slot
2196                          * (should be rare).
2197                          */
2198                         if (!bh) {
2199                                 ext3_error(inode->i_sb, "ext3_free_branches",
2200                                            "Read failure, inode=%lu, block="E3FSBLK,
2201                                            inode->i_ino, nr);
2202                                 continue;
2203                         }
2204
2205                         /* This zaps the entire block.  Bottom up. */
2206                         BUFFER_TRACE(bh, "free child branches");
2207                         ext3_free_branches(handle, inode, bh,
2208                                            (__le32*)bh->b_data,
2209                                            (__le32*)bh->b_data + addr_per_block,
2210                                            depth);
2211
2212                         /*
2213                          * We've probably journalled the indirect block several
2214                          * times during the truncate.  But it's no longer
2215                          * needed and we now drop it from the transaction via
2216                          * journal_revoke().
2217                          *
2218                          * That's easy if it's exclusively part of this
2219                          * transaction.  But if it's part of the committing
2220                          * transaction then journal_forget() will simply
2221                          * brelse() it.  That means that if the underlying
2222                          * block is reallocated in ext3_get_block(),
2223                          * unmap_underlying_metadata() will find this block
2224                          * and will try to get rid of it.  damn, damn.
2225                          *
2226                          * If this block has already been committed to the
2227                          * journal, a revoke record will be written.  And
2228                          * revoke records must be emitted *before* clearing
2229                          * this block's bit in the bitmaps.
2230                          */
2231                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2232
2233                         /*
2234                          * Everything below this this pointer has been
2235                          * released.  Now let this top-of-subtree go.
2236                          *
2237                          * We want the freeing of this indirect block to be
2238                          * atomic in the journal with the updating of the
2239                          * bitmap block which owns it.  So make some room in
2240                          * the journal.
2241                          *
2242                          * We zero the parent pointer *after* freeing its
2243                          * pointee in the bitmaps, so if extend_transaction()
2244                          * for some reason fails to put the bitmap changes and
2245                          * the release into the same transaction, recovery
2246                          * will merely complain about releasing a free block,
2247                          * rather than leaking blocks.
2248                          */
2249                         if (is_handle_aborted(handle))
2250                                 return;
2251                         if (try_to_extend_transaction(handle, inode)) {
2252                                 ext3_mark_inode_dirty(handle, inode);
2253                                 ext3_journal_test_restart(handle, inode);
2254                         }
2255
2256                         ext3_free_blocks(handle, inode, nr, 1);
2257
2258                         if (parent_bh) {
2259                                 /*
2260                                  * The block which we have just freed is
2261                                  * pointed to by an indirect block: journal it
2262                                  */
2263                                 BUFFER_TRACE(parent_bh, "get_write_access");
2264                                 if (!ext3_journal_get_write_access(handle,
2265                                                                    parent_bh)){
2266                                         *p = 0;
2267                                         BUFFER_TRACE(parent_bh,
2268                                         "call ext3_journal_dirty_metadata");
2269                                         ext3_journal_dirty_metadata(handle,
2270                                                                     parent_bh);
2271                                 }
2272                         }
2273                 }
2274         } else {
2275                 /* We have reached the bottom of the tree. */
2276                 BUFFER_TRACE(parent_bh, "free data blocks");
2277                 ext3_free_data(handle, inode, parent_bh, first, last);
2278         }
2279 }
2280
2281 int ext3_can_truncate(struct inode *inode)
2282 {
2283         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2284                 return 0;
2285         if (S_ISREG(inode->i_mode))
2286                 return 1;
2287         if (S_ISDIR(inode->i_mode))
2288                 return 1;
2289         if (S_ISLNK(inode->i_mode))
2290                 return !ext3_inode_is_fast_symlink(inode);
2291         return 0;
2292 }
2293
2294 /*
2295  * ext3_truncate()
2296  *
2297  * We block out ext3_get_block() block instantiations across the entire
2298  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2299  * simultaneously on behalf of the same inode.
2300  *
2301  * As we work through the truncate and commmit bits of it to the journal there
2302  * is one core, guiding principle: the file's tree must always be consistent on
2303  * disk.  We must be able to restart the truncate after a crash.
2304  *
2305  * The file's tree may be transiently inconsistent in memory (although it
2306  * probably isn't), but whenever we close off and commit a journal transaction,
2307  * the contents of (the filesystem + the journal) must be consistent and
2308  * restartable.  It's pretty simple, really: bottom up, right to left (although
2309  * left-to-right works OK too).
2310  *
2311  * Note that at recovery time, journal replay occurs *before* the restart of
2312  * truncate against the orphan inode list.
2313  *
2314  * The committed inode has the new, desired i_size (which is the same as
2315  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2316  * that this inode's truncate did not complete and it will again call
2317  * ext3_truncate() to have another go.  So there will be instantiated blocks
2318  * to the right of the truncation point in a crashed ext3 filesystem.  But
2319  * that's fine - as long as they are linked from the inode, the post-crash
2320  * ext3_truncate() run will find them and release them.
2321  */
2322 void ext3_truncate(struct inode *inode)
2323 {
2324         handle_t *handle;
2325         struct ext3_inode_info *ei = EXT3_I(inode);
2326         __le32 *i_data = ei->i_data;
2327         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2328         struct address_space *mapping = inode->i_mapping;
2329         int offsets[4];
2330         Indirect chain[4];
2331         Indirect *partial;
2332         __le32 nr = 0;
2333         int n;
2334         long last_block;
2335         unsigned blocksize = inode->i_sb->s_blocksize;
2336         struct page *page;
2337
2338         if (!ext3_can_truncate(inode))
2339                 return;
2340
2341         /*
2342          * We have to lock the EOF page here, because lock_page() nests
2343          * outside journal_start().
2344          */
2345         if ((inode->i_size & (blocksize - 1)) == 0) {
2346                 /* Block boundary? Nothing to do */
2347                 page = NULL;
2348         } else {
2349                 page = grab_cache_page(mapping,
2350                                 inode->i_size >> PAGE_CACHE_SHIFT);
2351                 if (!page)
2352                         return;
2353         }
2354
2355         handle = start_transaction(inode);
2356         if (IS_ERR(handle)) {
2357                 if (page) {
2358                         clear_highpage(page);
2359                         flush_dcache_page(page);
2360                         unlock_page(page);
2361                         page_cache_release(page);
2362                 }
2363                 return;         /* AKPM: return what? */
2364         }
2365
2366         last_block = (inode->i_size + blocksize-1)
2367                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2368
2369         if (page)
2370                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2371
2372         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2373         if (n == 0)
2374                 goto out_stop;  /* error */
2375
2376         /*
2377          * OK.  This truncate is going to happen.  We add the inode to the
2378          * orphan list, so that if this truncate spans multiple transactions,
2379          * and we crash, we will resume the truncate when the filesystem
2380          * recovers.  It also marks the inode dirty, to catch the new size.
2381          *
2382          * Implication: the file must always be in a sane, consistent
2383          * truncatable state while each transaction commits.
2384          */
2385         if (ext3_orphan_add(handle, inode))
2386                 goto out_stop;
2387
2388         /*
2389          * The orphan list entry will now protect us from any crash which
2390          * occurs before the truncate completes, so it is now safe to propagate
2391          * the new, shorter inode size (held for now in i_size) into the
2392          * on-disk inode. We do this via i_disksize, which is the value which
2393          * ext3 *really* writes onto the disk inode.
2394          */
2395         ei->i_disksize = inode->i_size;
2396
2397         /*
2398          * From here we block out all ext3_get_block() callers who want to
2399          * modify the block allocation tree.
2400          */
2401         mutex_lock(&ei->truncate_mutex);
2402
2403         if (n == 1) {           /* direct blocks */
2404                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2405                                i_data + EXT3_NDIR_BLOCKS);
2406                 goto do_indirects;
2407         }
2408
2409         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2410         /* Kill the top of shared branch (not detached) */
2411         if (nr) {
2412                 if (partial == chain) {
2413                         /* Shared branch grows from the inode */
2414                         ext3_free_branches(handle, inode, NULL,
2415                                            &nr, &nr+1, (chain+n-1) - partial);
2416                         *partial->p = 0;
2417                         /*
2418                          * We mark the inode dirty prior to restart,
2419                          * and prior to stop.  No need for it here.
2420                          */
2421                 } else {
2422                         /* Shared branch grows from an indirect block */
2423                         BUFFER_TRACE(partial->bh, "get_write_access");
2424                         ext3_free_branches(handle, inode, partial->bh,
2425                                         partial->p,
2426                                         partial->p+1, (chain+n-1) - partial);
2427                 }
2428         }
2429         /* Clear the ends of indirect blocks on the shared branch */
2430         while (partial > chain) {
2431                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2432                                    (__le32*)partial->bh->b_data+addr_per_block,
2433                                    (chain+n-1) - partial);
2434                 BUFFER_TRACE(partial->bh, "call brelse");
2435                 brelse (partial->bh);
2436                 partial--;
2437         }
2438 do_indirects:
2439         /* Kill the remaining (whole) subtrees */
2440         switch (offsets[0]) {
2441         default:
2442                 nr = i_data[EXT3_IND_BLOCK];
2443                 if (nr) {
2444                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2445                         i_data[EXT3_IND_BLOCK] = 0;
2446                 }
2447         case EXT3_IND_BLOCK:
2448                 nr = i_data[EXT3_DIND_BLOCK];
2449                 if (nr) {
2450                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2451                         i_data[EXT3_DIND_BLOCK] = 0;
2452                 }
2453         case EXT3_DIND_BLOCK:
2454                 nr = i_data[EXT3_TIND_BLOCK];
2455                 if (nr) {
2456                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2457                         i_data[EXT3_TIND_BLOCK] = 0;
2458                 }
2459         case EXT3_TIND_BLOCK:
2460                 ;
2461         }
2462
2463         ext3_discard_reservation(inode);
2464
2465         mutex_unlock(&ei->truncate_mutex);
2466         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2467         ext3_mark_inode_dirty(handle, inode);
2468
2469         /*
2470          * In a multi-transaction truncate, we only make the final transaction
2471          * synchronous
2472          */
2473         if (IS_SYNC(inode))
2474                 handle->h_sync = 1;
2475 out_stop:
2476         /*
2477          * If this was a simple ftruncate(), and the file will remain alive
2478          * then we need to clear up the orphan record which we created above.
2479          * However, if this was a real unlink then we were called by
2480          * ext3_delete_inode(), and we allow that function to clean up the
2481          * orphan info for us.
2482          */
2483         if (inode->i_nlink)
2484                 ext3_orphan_del(handle, inode);
2485
2486         ext3_journal_stop(handle);
2487 }
2488
2489 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2490                 unsigned long ino, struct ext3_iloc *iloc)
2491 {
2492         unsigned long block_group;
2493         unsigned long offset;
2494         ext3_fsblk_t block;
2495         struct ext3_group_desc *gdp;
2496
2497         if (!ext3_valid_inum(sb, ino)) {
2498                 /*
2499                  * This error is already checked for in namei.c unless we are
2500                  * looking at an NFS filehandle, in which case no error
2501                  * report is needed
2502                  */
2503                 return 0;
2504         }
2505
2506         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2507         gdp = ext3_get_group_desc(sb, block_group, NULL);
2508         if (!gdp)
2509                 return 0;
2510         /*
2511          * Figure out the offset within the block group inode table
2512          */
2513         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2514                 EXT3_INODE_SIZE(sb);
2515         block = le32_to_cpu(gdp->bg_inode_table) +
2516                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2517
2518         iloc->block_group = block_group;
2519         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2520         return block;
2521 }
2522
2523 /*
2524  * ext3_get_inode_loc returns with an extra refcount against the inode's
2525  * underlying buffer_head on success. If 'in_mem' is true, we have all
2526  * data in memory that is needed to recreate the on-disk version of this
2527  * inode.
2528  */
2529 static int __ext3_get_inode_loc(struct inode *inode,
2530                                 struct ext3_iloc *iloc, int in_mem)
2531 {
2532         ext3_fsblk_t block;
2533         struct buffer_head *bh;
2534
2535         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2536         if (!block)
2537                 return -EIO;
2538
2539         bh = sb_getblk(inode->i_sb, block);
2540         if (!bh) {
2541                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2542                                 "unable to read inode block - "
2543                                 "inode=%lu, block="E3FSBLK,
2544                                  inode->i_ino, block);
2545                 return -EIO;
2546         }
2547         if (!buffer_uptodate(bh)) {
2548                 lock_buffer(bh);
2549
2550                 /*
2551                  * If the buffer has the write error flag, we have failed
2552                  * to write out another inode in the same block.  In this
2553                  * case, we don't have to read the block because we may
2554                  * read the old inode data successfully.
2555                  */
2556                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2557                         set_buffer_uptodate(bh);
2558
2559                 if (buffer_uptodate(bh)) {
2560                         /* someone brought it uptodate while we waited */
2561                         unlock_buffer(bh);
2562                         goto has_buffer;
2563                 }
2564
2565                 /*
2566                  * If we have all information of the inode in memory and this
2567                  * is the only valid inode in the block, we need not read the
2568                  * block.
2569                  */
2570                 if (in_mem) {
2571                         struct buffer_head *bitmap_bh;
2572                         struct ext3_group_desc *desc;
2573                         int inodes_per_buffer;
2574                         int inode_offset, i;
2575                         int block_group;
2576                         int start;
2577
2578                         block_group = (inode->i_ino - 1) /
2579                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2580                         inodes_per_buffer = bh->b_size /
2581                                 EXT3_INODE_SIZE(inode->i_sb);
2582                         inode_offset = ((inode->i_ino - 1) %
2583                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2584                         start = inode_offset & ~(inodes_per_buffer - 1);
2585
2586                         /* Is the inode bitmap in cache? */
2587                         desc = ext3_get_group_desc(inode->i_sb,
2588                                                 block_group, NULL);
2589                         if (!desc)
2590                                 goto make_io;
2591
2592                         bitmap_bh = sb_getblk(inode->i_sb,
2593                                         le32_to_cpu(desc->bg_inode_bitmap));
2594                         if (!bitmap_bh)
2595                                 goto make_io;
2596
2597                         /*
2598                          * If the inode bitmap isn't in cache then the
2599                          * optimisation may end up performing two reads instead
2600                          * of one, so skip it.
2601                          */
2602                         if (!buffer_uptodate(bitmap_bh)) {
2603                                 brelse(bitmap_bh);
2604                                 goto make_io;
2605                         }
2606                         for (i = start; i < start + inodes_per_buffer; i++) {
2607                                 if (i == inode_offset)
2608                                         continue;
2609                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2610                                         break;
2611                         }
2612                         brelse(bitmap_bh);
2613                         if (i == start + inodes_per_buffer) {
2614                                 /* all other inodes are free, so skip I/O */
2615                                 memset(bh->b_data, 0, bh->b_size);
2616                                 set_buffer_uptodate(bh);
2617                                 unlock_buffer(bh);
2618                                 goto has_buffer;
2619                         }
2620                 }
2621
2622 make_io:
2623                 /*
2624                  * There are other valid inodes in the buffer, this inode
2625                  * has in-inode xattrs, or we don't have this inode in memory.
2626                  * Read the block from disk.
2627                  */
2628                 get_bh(bh);
2629                 bh->b_end_io = end_buffer_read_sync;
2630                 submit_bh(READ_META, bh);
2631                 wait_on_buffer(bh);
2632                 if (!buffer_uptodate(bh)) {
2633                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2634                                         "unable to read inode block - "
2635                                         "inode=%lu, block="E3FSBLK,
2636                                         inode->i_ino, block);
2637                         brelse(bh);
2638                         return -EIO;
2639                 }
2640         }
2641 has_buffer:
2642         iloc->bh = bh;
2643         return 0;
2644 }
2645
2646 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2647 {
2648         /* We have all inode data except xattrs in memory here. */
2649         return __ext3_get_inode_loc(inode, iloc,
2650                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2651 }
2652
2653 void ext3_set_inode_flags(struct inode *inode)
2654 {
2655         unsigned int flags = EXT3_I(inode)->i_flags;
2656
2657         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2658         if (flags & EXT3_SYNC_FL)
2659                 inode->i_flags |= S_SYNC;
2660         if (flags & EXT3_APPEND_FL)
2661                 inode->i_flags |= S_APPEND;
2662         if (flags & EXT3_IMMUTABLE_FL)
2663                 inode->i_flags |= S_IMMUTABLE;
2664         if (flags & EXT3_NOATIME_FL)
2665                 inode->i_flags |= S_NOATIME;
2666         if (flags & EXT3_DIRSYNC_FL)
2667                 inode->i_flags |= S_DIRSYNC;
2668 }
2669
2670 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2671 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2672 {
2673         unsigned int flags = ei->vfs_inode.i_flags;
2674
2675         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2676                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2677         if (flags & S_SYNC)
2678                 ei->i_flags |= EXT3_SYNC_FL;
2679         if (flags & S_APPEND)
2680                 ei->i_flags |= EXT3_APPEND_FL;
2681         if (flags & S_IMMUTABLE)
2682                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2683         if (flags & S_NOATIME)
2684                 ei->i_flags |= EXT3_NOATIME_FL;
2685         if (flags & S_DIRSYNC)
2686                 ei->i_flags |= EXT3_DIRSYNC_FL;
2687 }
2688
2689 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2690 {
2691         struct ext3_iloc iloc;
2692         struct ext3_inode *raw_inode;
2693         struct ext3_inode_info *ei;
2694         struct buffer_head *bh;
2695         struct inode *inode;
2696         long ret;
2697         int block;
2698
2699         inode = iget_locked(sb, ino);
2700         if (!inode)
2701                 return ERR_PTR(-ENOMEM);
2702         if (!(inode->i_state & I_NEW))
2703                 return inode;
2704
2705         ei = EXT3_I(inode);
2706 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2707         ei->i_acl = EXT3_ACL_NOT_CACHED;
2708         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2709 #endif
2710         ei->i_block_alloc_info = NULL;
2711
2712         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2713         if (ret < 0)
2714                 goto bad_inode;
2715         bh = iloc.bh;
2716         raw_inode = ext3_raw_inode(&iloc);
2717         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2718         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2719         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2720         if(!(test_opt (inode->i_sb, NO_UID32))) {
2721                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2722                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2723         }
2724         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2725         inode->i_size = le32_to_cpu(raw_inode->i_size);
2726         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2727         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2728         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2729         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2730
2731         ei->i_state = 0;
2732         ei->i_dir_start_lookup = 0;
2733         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2734         /* We now have enough fields to check if the inode was active or not.
2735          * This is needed because nfsd might try to access dead inodes
2736          * the test is that same one that e2fsck uses
2737          * NeilBrown 1999oct15
2738          */
2739         if (inode->i_nlink == 0) {
2740                 if (inode->i_mode == 0 ||
2741                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2742                         /* this inode is deleted */
2743                         brelse (bh);
2744                         ret = -ESTALE;
2745                         goto bad_inode;
2746                 }
2747                 /* The only unlinked inodes we let through here have
2748                  * valid i_mode and are being read by the orphan
2749                  * recovery code: that's fine, we're about to complete
2750                  * the process of deleting those. */
2751         }
2752         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2753         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2754 #ifdef EXT3_FRAGMENTS
2755         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2756         ei->i_frag_no = raw_inode->i_frag;
2757         ei->i_frag_size = raw_inode->i_fsize;
2758 #endif
2759         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2760         if (!S_ISREG(inode->i_mode)) {
2761                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2762         } else {
2763                 inode->i_size |=
2764                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2765         }
2766         ei->i_disksize = inode->i_size;
2767         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2768         ei->i_block_group = iloc.block_group;
2769         /*
2770          * NOTE! The in-memory inode i_data array is in little-endian order
2771          * even on big-endian machines: we do NOT byteswap the block numbers!
2772          */
2773         for (block = 0; block < EXT3_N_BLOCKS; block++)
2774                 ei->i_data[block] = raw_inode->i_block[block];
2775         INIT_LIST_HEAD(&ei->i_orphan);
2776
2777         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2778             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2779                 /*
2780                  * When mke2fs creates big inodes it does not zero out
2781                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2782                  * so ignore those first few inodes.
2783                  */
2784                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2785                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2786                     EXT3_INODE_SIZE(inode->i_sb)) {
2787                         brelse (bh);
2788                         ret = -EIO;
2789                         goto bad_inode;
2790                 }
2791                 if (ei->i_extra_isize == 0) {
2792                         /* The extra space is currently unused. Use it. */
2793                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2794                                             EXT3_GOOD_OLD_INODE_SIZE;
2795                 } else {
2796                         __le32 *magic = (void *)raw_inode +
2797                                         EXT3_GOOD_OLD_INODE_SIZE +
2798                                         ei->i_extra_isize;
2799                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2800                                  ei->i_state |= EXT3_STATE_XATTR;
2801                 }
2802         } else
2803                 ei->i_extra_isize = 0;
2804
2805         if (S_ISREG(inode->i_mode)) {
2806                 inode->i_op = &ext3_file_inode_operations;
2807                 inode->i_fop = &ext3_file_operations;
2808                 ext3_set_aops(inode);
2809         } else if (S_ISDIR(inode->i_mode)) {
2810                 inode->i_op = &ext3_dir_inode_operations;
2811                 inode->i_fop = &ext3_dir_operations;
2812         } else if (S_ISLNK(inode->i_mode)) {
2813                 if (ext3_inode_is_fast_symlink(inode))
2814                         inode->i_op = &ext3_fast_symlink_inode_operations;
2815                 else {
2816                         inode->i_op = &ext3_symlink_inode_operations;
2817                         ext3_set_aops(inode);
2818                 }
2819         } else {
2820                 inode->i_op = &ext3_special_inode_operations;
2821                 if (raw_inode->i_block[0])
2822                         init_special_inode(inode, inode->i_mode,
2823                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2824                 else
2825                         init_special_inode(inode, inode->i_mode,
2826                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2827         }
2828         brelse (iloc.bh);
2829         ext3_set_inode_flags(inode);
2830         unlock_new_inode(inode);
2831         return inode;
2832
2833 bad_inode:
2834         iget_failed(inode);
2835         return ERR_PTR(ret);
2836 }
2837
2838 /*
2839  * Post the struct inode info into an on-disk inode location in the
2840  * buffer-cache.  This gobbles the caller's reference to the
2841  * buffer_head in the inode location struct.
2842  *
2843  * The caller must have write access to iloc->bh.
2844  */
2845 static int ext3_do_update_inode(handle_t *handle,
2846                                 struct inode *inode,
2847                                 struct ext3_iloc *iloc)
2848 {
2849         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2850         struct ext3_inode_info *ei = EXT3_I(inode);
2851         struct buffer_head *bh = iloc->bh;
2852         int err = 0, rc, block;
2853
2854         /* For fields not not tracking in the in-memory inode,
2855          * initialise them to zero for new inodes. */
2856         if (ei->i_state & EXT3_STATE_NEW)
2857                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2858
2859         ext3_get_inode_flags(ei);
2860         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2861         if(!(test_opt(inode->i_sb, NO_UID32))) {
2862                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2863                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2864 /*
2865  * Fix up interoperability with old kernels. Otherwise, old inodes get
2866  * re-used with the upper 16 bits of the uid/gid intact
2867  */
2868                 if(!ei->i_dtime) {
2869                         raw_inode->i_uid_high =
2870                                 cpu_to_le16(high_16_bits(inode->i_uid));
2871                         raw_inode->i_gid_high =
2872                                 cpu_to_le16(high_16_bits(inode->i_gid));
2873                 } else {
2874                         raw_inode->i_uid_high = 0;
2875                         raw_inode->i_gid_high = 0;
2876                 }
2877         } else {
2878                 raw_inode->i_uid_low =
2879                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2880                 raw_inode->i_gid_low =
2881                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2882                 raw_inode->i_uid_high = 0;
2883                 raw_inode->i_gid_high = 0;
2884         }
2885         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2886         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2887         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2888         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2889         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2890         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2891         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2892         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2893 #ifdef EXT3_FRAGMENTS
2894         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2895         raw_inode->i_frag = ei->i_frag_no;
2896         raw_inode->i_fsize = ei->i_frag_size;
2897 #endif
2898         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2899         if (!S_ISREG(inode->i_mode)) {
2900                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2901         } else {
2902                 raw_inode->i_size_high =
2903                         cpu_to_le32(ei->i_disksize >> 32);
2904                 if (ei->i_disksize > 0x7fffffffULL) {
2905                         struct super_block *sb = inode->i_sb;
2906                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2907                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2908                             EXT3_SB(sb)->s_es->s_rev_level ==
2909                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2910                                /* If this is the first large file
2911                                 * created, add a flag to the superblock.
2912                                 */
2913                                 err = ext3_journal_get_write_access(handle,
2914                                                 EXT3_SB(sb)->s_sbh);
2915                                 if (err)
2916                                         goto out_brelse;
2917                                 ext3_update_dynamic_rev(sb);
2918                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2919                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2920                                 sb->s_dirt = 1;
2921                                 handle->h_sync = 1;
2922                                 err = ext3_journal_dirty_metadata(handle,
2923                                                 EXT3_SB(sb)->s_sbh);
2924                         }
2925                 }
2926         }
2927         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2928         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2929                 if (old_valid_dev(inode->i_rdev)) {
2930                         raw_inode->i_block[0] =
2931                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2932                         raw_inode->i_block[1] = 0;
2933                 } else {
2934                         raw_inode->i_block[0] = 0;
2935                         raw_inode->i_block[1] =
2936                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2937                         raw_inode->i_block[2] = 0;
2938                 }
2939         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2940                 raw_inode->i_block[block] = ei->i_data[block];
2941
2942         if (ei->i_extra_isize)
2943                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2944
2945         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2946         rc = ext3_journal_dirty_metadata(handle, bh);
2947         if (!err)
2948                 err = rc;
2949         ei->i_state &= ~EXT3_STATE_NEW;
2950
2951 out_brelse:
2952         brelse (bh);
2953         ext3_std_error(inode->i_sb, err);
2954         return err;
2955 }
2956
2957 /*
2958  * ext3_write_inode()
2959  *
2960  * We are called from a few places:
2961  *
2962  * - Within generic_file_write() for O_SYNC files.
2963  *   Here, there will be no transaction running. We wait for any running
2964  *   trasnaction to commit.
2965  *
2966  * - Within sys_sync(), kupdate and such.
2967  *   We wait on commit, if tol to.
2968  *
2969  * - Within prune_icache() (PF_MEMALLOC == true)
2970  *   Here we simply return.  We can't afford to block kswapd on the
2971  *   journal commit.
2972  *
2973  * In all cases it is actually safe for us to return without doing anything,
2974  * because the inode has been copied into a raw inode buffer in
2975  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2976  * knfsd.
2977  *
2978  * Note that we are absolutely dependent upon all inode dirtiers doing the
2979  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2980  * which we are interested.
2981  *
2982  * It would be a bug for them to not do this.  The code:
2983  *
2984  *      mark_inode_dirty(inode)
2985  *      stuff();
2986  *      inode->i_size = expr;
2987  *
2988  * is in error because a kswapd-driven write_inode() could occur while
2989  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2990  * will no longer be on the superblock's dirty inode list.
2991  */
2992 int ext3_write_inode(struct inode *inode, int wait)
2993 {
2994         if (current->flags & PF_MEMALLOC)
2995                 return 0;
2996
2997         if (ext3_journal_current_handle()) {
2998                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
2999                 dump_stack();
3000                 return -EIO;
3001         }
3002
3003         if (!wait)
3004                 return 0;
3005
3006         return ext3_force_commit(inode->i_sb);
3007 }
3008
3009 /*
3010  * ext3_setattr()
3011  *
3012  * Called from notify_change.
3013  *
3014  * We want to trap VFS attempts to truncate the file as soon as
3015  * possible.  In particular, we want to make sure that when the VFS
3016  * shrinks i_size, we put the inode on the orphan list and modify
3017  * i_disksize immediately, so that during the subsequent flushing of
3018  * dirty pages and freeing of disk blocks, we can guarantee that any
3019  * commit will leave the blocks being flushed in an unused state on
3020  * disk.  (On recovery, the inode will get truncated and the blocks will
3021  * be freed, so we have a strong guarantee that no future commit will
3022  * leave these blocks visible to the user.)
3023  *
3024  * Called with inode->sem down.
3025  */
3026 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3027 {
3028         struct inode *inode = dentry->d_inode;
3029         int error, rc = 0;
3030         const unsigned int ia_valid = attr->ia_valid;
3031
3032         error = inode_change_ok(inode, attr);
3033         if (error)
3034                 return error;
3035
3036         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3037                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3038                 handle_t *handle;
3039
3040                 /* (user+group)*(old+new) structure, inode write (sb,
3041                  * inode block, ? - but truncate inode update has it) */
3042                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3043                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3044                 if (IS_ERR(handle)) {
3045                         error = PTR_ERR(handle);
3046                         goto err_out;
3047                 }
3048                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
3049                 if (error) {
3050                         ext3_journal_stop(handle);
3051                         return error;
3052                 }
3053                 /* Update corresponding info in inode so that everything is in
3054                  * one transaction */
3055                 if (attr->ia_valid & ATTR_UID)
3056                         inode->i_uid = attr->ia_uid;
3057                 if (attr->ia_valid & ATTR_GID)
3058                         inode->i_gid = attr->ia_gid;
3059                 error = ext3_mark_inode_dirty(handle, inode);
3060                 ext3_journal_stop(handle);
3061         }
3062
3063         if (S_ISREG(inode->i_mode) &&
3064             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3065                 handle_t *handle;
3066
3067                 handle = ext3_journal_start(inode, 3);
3068                 if (IS_ERR(handle)) {
3069                         error = PTR_ERR(handle);
3070                         goto err_out;
3071                 }
3072
3073                 error = ext3_orphan_add(handle, inode);
3074                 EXT3_I(inode)->i_disksize = attr->ia_size;
3075                 rc = ext3_mark_inode_dirty(handle, inode);
3076                 if (!error)
3077                         error = rc;
3078                 ext3_journal_stop(handle);
3079         }
3080
3081         rc = inode_setattr(inode, attr);
3082
3083         /* If inode_setattr's call to ext3_truncate failed to get a
3084          * transaction handle at all, we need to clean up the in-core
3085          * orphan list manually. */
3086         if (inode->i_nlink)
3087                 ext3_orphan_del(NULL, inode);
3088
3089         if (!rc && (ia_valid & ATTR_MODE))
3090                 rc = ext3_acl_chmod(inode);
3091
3092 err_out:
3093         ext3_std_error(inode->i_sb, error);
3094         if (!error)
3095                 error = rc;
3096         return error;
3097 }
3098
3099
3100 /*
3101  * How many blocks doth make a writepage()?
3102  *
3103  * With N blocks per page, it may be:
3104  * N data blocks
3105  * 2 indirect block
3106  * 2 dindirect
3107  * 1 tindirect
3108  * N+5 bitmap blocks (from the above)
3109  * N+5 group descriptor summary blocks
3110  * 1 inode block
3111  * 1 superblock.
3112  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3113  *
3114  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3115  *
3116  * With ordered or writeback data it's the same, less the N data blocks.
3117  *
3118  * If the inode's direct blocks can hold an integral number of pages then a
3119  * page cannot straddle two indirect blocks, and we can only touch one indirect
3120  * and dindirect block, and the "5" above becomes "3".
3121  *
3122  * This still overestimates under most circumstances.  If we were to pass the
3123  * start and end offsets in here as well we could do block_to_path() on each
3124  * block and work out the exact number of indirects which are touched.  Pah.
3125  */
3126
3127 static int ext3_writepage_trans_blocks(struct inode *inode)
3128 {
3129         int bpp = ext3_journal_blocks_per_page(inode);
3130         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3131         int ret;
3132
3133         if (ext3_should_journal_data(inode))
3134                 ret = 3 * (bpp + indirects) + 2;
3135         else
3136                 ret = 2 * (bpp + indirects) + 2;
3137
3138 #ifdef CONFIG_QUOTA
3139         /* We know that structure was already allocated during DQUOT_INIT so
3140          * we will be updating only the data blocks + inodes */
3141         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3142 #endif
3143
3144         return ret;
3145 }
3146
3147 /*
3148  * The caller must have previously called ext3_reserve_inode_write().
3149  * Give this, we know that the caller already has write access to iloc->bh.
3150  */
3151 int ext3_mark_iloc_dirty(handle_t *handle,
3152                 struct inode *inode, struct ext3_iloc *iloc)
3153 {
3154         int err = 0;
3155
3156         /* the do_update_inode consumes one bh->b_count */
3157         get_bh(iloc->bh);
3158
3159         /* ext3_do_update_inode() does journal_dirty_metadata */
3160         err = ext3_do_update_inode(handle, inode, iloc);
3161         put_bh(iloc->bh);
3162         return err;
3163 }
3164
3165 /*
3166  * On success, We end up with an outstanding reference count against
3167  * iloc->bh.  This _must_ be cleaned up later.
3168  */
3169
3170 int
3171 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3172                          struct ext3_iloc *iloc)
3173 {
3174         int err = 0;
3175         if (handle) {
3176                 err = ext3_get_inode_loc(inode, iloc);
3177                 if (!err) {
3178                         BUFFER_TRACE(iloc->bh, "get_write_access");
3179                         err = ext3_journal_get_write_access(handle, iloc->bh);
3180                         if (err) {
3181                                 brelse(iloc->bh);
3182                                 iloc->bh = NULL;
3183                         }
3184                 }
3185         }
3186         ext3_std_error(inode->i_sb, err);
3187         return err;
3188 }
3189
3190 /*
3191  * What we do here is to mark the in-core inode as clean with respect to inode
3192  * dirtiness (it may still be data-dirty).
3193  * This means that the in-core inode may be reaped by prune_icache
3194  * without having to perform any I/O.  This is a very good thing,
3195  * because *any* task may call prune_icache - even ones which
3196  * have a transaction open against a different journal.
3197  *
3198  * Is this cheating?  Not really.  Sure, we haven't written the
3199  * inode out, but prune_icache isn't a user-visible syncing function.
3200  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3201  * we start and wait on commits.
3202  *
3203  * Is this efficient/effective?  Well, we're being nice to the system
3204  * by cleaning up our inodes proactively so they can be reaped
3205  * without I/O.  But we are potentially leaving up to five seconds'
3206  * worth of inodes floating about which prune_icache wants us to
3207  * write out.  One way to fix that would be to get prune_icache()
3208  * to do a write_super() to free up some memory.  It has the desired
3209  * effect.
3210  */
3211 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3212 {
3213         struct ext3_iloc iloc;
3214         int err;
3215
3216         might_sleep();
3217         err = ext3_reserve_inode_write(handle, inode, &iloc);
3218         if (!err)
3219                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3220         return err;
3221 }
3222
3223 /*
3224  * ext3_dirty_inode() is called from __mark_inode_dirty()
3225  *
3226  * We're really interested in the case where a file is being extended.
3227  * i_size has been changed by generic_commit_write() and we thus need
3228  * to include the updated inode in the current transaction.
3229  *
3230  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3231  * are allocated to the file.
3232  *
3233  * If the inode is marked synchronous, we don't honour that here - doing
3234  * so would cause a commit on atime updates, which we don't bother doing.
3235  * We handle synchronous inodes at the highest possible level.
3236  */
3237 void ext3_dirty_inode(struct inode *inode)
3238 {
3239         handle_t *current_handle = ext3_journal_current_handle();
3240         handle_t *handle;
3241
3242         handle = ext3_journal_start(inode, 2);
3243         if (IS_ERR(handle))
3244                 goto out;
3245         if (current_handle &&
3246                 current_handle->h_transaction != handle->h_transaction) {
3247                 /* This task has a transaction open against a different fs */
3248                 printk(KERN_EMERG "%s: transactions do not match!\n",
3249                        __func__);
3250         } else {
3251                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3252                                 current_handle);
3253                 ext3_mark_inode_dirty(handle, inode);
3254         }
3255         ext3_journal_stop(handle);
3256 out:
3257         return;
3258 }
3259
3260 #if 0
3261 /*
3262  * Bind an inode's backing buffer_head into this transaction, to prevent
3263  * it from being flushed to disk early.  Unlike
3264  * ext3_reserve_inode_write, this leaves behind no bh reference and
3265  * returns no iloc structure, so the caller needs to repeat the iloc
3266  * lookup to mark the inode dirty later.
3267  */
3268 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3269 {
3270         struct ext3_iloc iloc;
3271
3272         int err = 0;
3273         if (handle) {
3274                 err = ext3_get_inode_loc(inode, &iloc);
3275                 if (!err) {
3276                         BUFFER_TRACE(iloc.bh, "get_write_access");
3277                         err = journal_get_write_access(handle, iloc.bh);
3278                         if (!err)
3279                                 err = ext3_journal_dirty_metadata(handle,
3280                                                                   iloc.bh);
3281                         brelse(iloc.bh);
3282                 }
3283         }
3284         ext3_std_error(inode->i_sb, err);
3285         return err;
3286 }
3287 #endif
3288
3289 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3290 {
3291         journal_t *journal;
3292         handle_t *handle;
3293         int err;
3294
3295         /*
3296          * We have to be very careful here: changing a data block's
3297          * journaling status dynamically is dangerous.  If we write a
3298          * data block to the journal, change the status and then delete
3299          * that block, we risk forgetting to revoke the old log record
3300          * from the journal and so a subsequent replay can corrupt data.
3301          * So, first we make sure that the journal is empty and that
3302          * nobody is changing anything.
3303          */
3304
3305         journal = EXT3_JOURNAL(inode);
3306         if (is_journal_aborted(journal))
3307                 return -EROFS;
3308
3309         journal_lock_updates(journal);
3310         journal_flush(journal);
3311
3312         /*
3313          * OK, there are no updates running now, and all cached data is
3314          * synced to disk.  We are now in a completely consistent state
3315          * which doesn't have anything in the journal, and we know that
3316          * no filesystem updates are running, so it is safe to modify
3317          * the inode's in-core data-journaling state flag now.
3318          */
3319
3320         if (val)
3321                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3322         else
3323                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3324         ext3_set_aops(inode);
3325
3326         journal_unlock_updates(journal);
3327
3328         /* Finally we can mark the inode as dirty. */
3329
3330         handle = ext3_journal_start(inode, 1);
3331         if (IS_ERR(handle))
3332                 return PTR_ERR(handle);
3333
3334         err = ext3_mark_inode_dirty(handle, inode);
3335         handle->h_sync = 1;
3336         ext3_journal_stop(handle);
3337         ext3_std_error(inode->i_sb, err);
3338
3339         return err;
3340 }