]> pilppa.org Git - linux-2.6-omap-h63xx.git/blob - fs/ext3/inode.c
Merge commit 'v2.6.28-rc2' into topic/asoc
[linux-2.6-omap-h63xx.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include <linux/fiemap.h>
40 #include "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                  * block_write_begin may have instantiated a few blocks
1191                  * outside i_size.  Trim these off again. Don't need
1192                  * i_size_read because we hold i_mutex.
1193                  */
1194                 if (pos + len > inode->i_size)
1195                         vmtruncate(inode, inode->i_size);
1196         }
1197         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1198                 goto retry;
1199 out:
1200         return ret;
1201 }
1202
1203
1204 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1205 {
1206         int err = journal_dirty_data(handle, bh);
1207         if (err)
1208                 ext3_journal_abort_handle(__func__, __func__,
1209                                                 bh, handle, err);
1210         return err;
1211 }
1212
1213 /* For write_end() in data=journal mode */
1214 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1215 {
1216         if (!buffer_mapped(bh) || buffer_freed(bh))
1217                 return 0;
1218         set_buffer_uptodate(bh);
1219         return ext3_journal_dirty_metadata(handle, bh);
1220 }
1221
1222 /*
1223  * Generic write_end handler for ordered and writeback ext3 journal modes.
1224  * We can't use generic_write_end, because that unlocks the page and we need to
1225  * unlock the page after ext3_journal_stop, but ext3_journal_stop must run
1226  * after block_write_end.
1227  */
1228 static int ext3_generic_write_end(struct file *file,
1229                                 struct address_space *mapping,
1230                                 loff_t pos, unsigned len, unsigned copied,
1231                                 struct page *page, void *fsdata)
1232 {
1233         struct inode *inode = file->f_mapping->host;
1234
1235         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1236
1237         if (pos+copied > inode->i_size) {
1238                 i_size_write(inode, pos+copied);
1239                 mark_inode_dirty(inode);
1240         }
1241
1242         return copied;
1243 }
1244
1245 /*
1246  * We need to pick up the new inode size which generic_commit_write gave us
1247  * `file' can be NULL - eg, when called from page_symlink().
1248  *
1249  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1250  * buffers are managed internally.
1251  */
1252 static int ext3_ordered_write_end(struct file *file,
1253                                 struct address_space *mapping,
1254                                 loff_t pos, unsigned len, unsigned copied,
1255                                 struct page *page, void *fsdata)
1256 {
1257         handle_t *handle = ext3_journal_current_handle();
1258         struct inode *inode = file->f_mapping->host;
1259         unsigned from, to;
1260         int ret = 0, ret2;
1261
1262         from = pos & (PAGE_CACHE_SIZE - 1);
1263         to = from + len;
1264
1265         ret = walk_page_buffers(handle, page_buffers(page),
1266                 from, to, NULL, ext3_journal_dirty_data);
1267
1268         if (ret == 0) {
1269                 /*
1270                  * generic_write_end() will run mark_inode_dirty() if i_size
1271                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1272                  * into that.
1273                  */
1274                 loff_t new_i_size;
1275
1276                 new_i_size = pos + copied;
1277                 if (new_i_size > EXT3_I(inode)->i_disksize)
1278                         EXT3_I(inode)->i_disksize = new_i_size;
1279                 ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1280                                                         page, fsdata);
1281                 copied = ret2;
1282                 if (ret2 < 0)
1283                         ret = ret2;
1284         }
1285         ret2 = ext3_journal_stop(handle);
1286         if (!ret)
1287                 ret = ret2;
1288         unlock_page(page);
1289         page_cache_release(page);
1290
1291         return ret ? ret : copied;
1292 }
1293
1294 static int ext3_writeback_write_end(struct file *file,
1295                                 struct address_space *mapping,
1296                                 loff_t pos, unsigned len, unsigned copied,
1297                                 struct page *page, void *fsdata)
1298 {
1299         handle_t *handle = ext3_journal_current_handle();
1300         struct inode *inode = file->f_mapping->host;
1301         int ret = 0, ret2;
1302         loff_t new_i_size;
1303
1304         new_i_size = pos + copied;
1305         if (new_i_size > EXT3_I(inode)->i_disksize)
1306                 EXT3_I(inode)->i_disksize = new_i_size;
1307
1308         ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1309                                                         page, fsdata);
1310         copied = ret2;
1311         if (ret2 < 0)
1312                 ret = ret2;
1313
1314         ret2 = ext3_journal_stop(handle);
1315         if (!ret)
1316                 ret = ret2;
1317         unlock_page(page);
1318         page_cache_release(page);
1319
1320         return ret ? ret : copied;
1321 }
1322
1323 static int ext3_journalled_write_end(struct file *file,
1324                                 struct address_space *mapping,
1325                                 loff_t pos, unsigned len, unsigned copied,
1326                                 struct page *page, void *fsdata)
1327 {
1328         handle_t *handle = ext3_journal_current_handle();
1329         struct inode *inode = mapping->host;
1330         int ret = 0, ret2;
1331         int partial = 0;
1332         unsigned from, to;
1333
1334         from = pos & (PAGE_CACHE_SIZE - 1);
1335         to = from + len;
1336
1337         if (copied < len) {
1338                 if (!PageUptodate(page))
1339                         copied = 0;
1340                 page_zero_new_buffers(page, from+copied, to);
1341         }
1342
1343         ret = walk_page_buffers(handle, page_buffers(page), from,
1344                                 to, &partial, write_end_fn);
1345         if (!partial)
1346                 SetPageUptodate(page);
1347         if (pos+copied > inode->i_size)
1348                 i_size_write(inode, pos+copied);
1349         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1350         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1351                 EXT3_I(inode)->i_disksize = inode->i_size;
1352                 ret2 = ext3_mark_inode_dirty(handle, inode);
1353                 if (!ret)
1354                         ret = ret2;
1355         }
1356
1357         ret2 = ext3_journal_stop(handle);
1358         if (!ret)
1359                 ret = ret2;
1360         unlock_page(page);
1361         page_cache_release(page);
1362
1363         return ret ? ret : copied;
1364 }
1365
1366 /*
1367  * bmap() is special.  It gets used by applications such as lilo and by
1368  * the swapper to find the on-disk block of a specific piece of data.
1369  *
1370  * Naturally, this is dangerous if the block concerned is still in the
1371  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1372  * filesystem and enables swap, then they may get a nasty shock when the
1373  * data getting swapped to that swapfile suddenly gets overwritten by
1374  * the original zero's written out previously to the journal and
1375  * awaiting writeback in the kernel's buffer cache.
1376  *
1377  * So, if we see any bmap calls here on a modified, data-journaled file,
1378  * take extra steps to flush any blocks which might be in the cache.
1379  */
1380 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1381 {
1382         struct inode *inode = mapping->host;
1383         journal_t *journal;
1384         int err;
1385
1386         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1387                 /*
1388                  * This is a REALLY heavyweight approach, but the use of
1389                  * bmap on dirty files is expected to be extremely rare:
1390                  * only if we run lilo or swapon on a freshly made file
1391                  * do we expect this to happen.
1392                  *
1393                  * (bmap requires CAP_SYS_RAWIO so this does not
1394                  * represent an unprivileged user DOS attack --- we'd be
1395                  * in trouble if mortal users could trigger this path at
1396                  * will.)
1397                  *
1398                  * NB. EXT3_STATE_JDATA is not set on files other than
1399                  * regular files.  If somebody wants to bmap a directory
1400                  * or symlink and gets confused because the buffer
1401                  * hasn't yet been flushed to disk, they deserve
1402                  * everything they get.
1403                  */
1404
1405                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1406                 journal = EXT3_JOURNAL(inode);
1407                 journal_lock_updates(journal);
1408                 err = journal_flush(journal);
1409                 journal_unlock_updates(journal);
1410
1411                 if (err)
1412                         return 0;
1413         }
1414
1415         return generic_block_bmap(mapping,block,ext3_get_block);
1416 }
1417
1418 static int bget_one(handle_t *handle, struct buffer_head *bh)
1419 {
1420         get_bh(bh);
1421         return 0;
1422 }
1423
1424 static int bput_one(handle_t *handle, struct buffer_head *bh)
1425 {
1426         put_bh(bh);
1427         return 0;
1428 }
1429
1430 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1431 {
1432         if (buffer_mapped(bh))
1433                 return ext3_journal_dirty_data(handle, bh);
1434         return 0;
1435 }
1436
1437 /*
1438  * Note that we always start a transaction even if we're not journalling
1439  * data.  This is to preserve ordering: any hole instantiation within
1440  * __block_write_full_page -> ext3_get_block() should be journalled
1441  * along with the data so we don't crash and then get metadata which
1442  * refers to old data.
1443  *
1444  * In all journalling modes block_write_full_page() will start the I/O.
1445  *
1446  * Problem:
1447  *
1448  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1449  *              ext3_writepage()
1450  *
1451  * Similar for:
1452  *
1453  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1454  *
1455  * Same applies to ext3_get_block().  We will deadlock on various things like
1456  * lock_journal and i_truncate_mutex.
1457  *
1458  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1459  * allocations fail.
1460  *
1461  * 16May01: If we're reentered then journal_current_handle() will be
1462  *          non-zero. We simply *return*.
1463  *
1464  * 1 July 2001: @@@ FIXME:
1465  *   In journalled data mode, a data buffer may be metadata against the
1466  *   current transaction.  But the same file is part of a shared mapping
1467  *   and someone does a writepage() on it.
1468  *
1469  *   We will move the buffer onto the async_data list, but *after* it has
1470  *   been dirtied. So there's a small window where we have dirty data on
1471  *   BJ_Metadata.
1472  *
1473  *   Note that this only applies to the last partial page in the file.  The
1474  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1475  *   broken code anyway: it's wrong for msync()).
1476  *
1477  *   It's a rare case: affects the final partial page, for journalled data
1478  *   where the file is subject to bith write() and writepage() in the same
1479  *   transction.  To fix it we'll need a custom block_write_full_page().
1480  *   We'll probably need that anyway for journalling writepage() output.
1481  *
1482  * We don't honour synchronous mounts for writepage().  That would be
1483  * disastrous.  Any write() or metadata operation will sync the fs for
1484  * us.
1485  *
1486  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1487  * we don't need to open a transaction here.
1488  */
1489 static int ext3_ordered_writepage(struct page *page,
1490                                 struct writeback_control *wbc)
1491 {
1492         struct inode *inode = page->mapping->host;
1493         struct buffer_head *page_bufs;
1494         handle_t *handle = NULL;
1495         int ret = 0;
1496         int err;
1497
1498         J_ASSERT(PageLocked(page));
1499
1500         /*
1501          * We give up here if we're reentered, because it might be for a
1502          * different filesystem.
1503          */
1504         if (ext3_journal_current_handle())
1505                 goto out_fail;
1506
1507         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1508
1509         if (IS_ERR(handle)) {
1510                 ret = PTR_ERR(handle);
1511                 goto out_fail;
1512         }
1513
1514         if (!page_has_buffers(page)) {
1515                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1516                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1517         }
1518         page_bufs = page_buffers(page);
1519         walk_page_buffers(handle, page_bufs, 0,
1520                         PAGE_CACHE_SIZE, NULL, bget_one);
1521
1522         ret = block_write_full_page(page, ext3_get_block, wbc);
1523
1524         /*
1525          * The page can become unlocked at any point now, and
1526          * truncate can then come in and change things.  So we
1527          * can't touch *page from now on.  But *page_bufs is
1528          * safe due to elevated refcount.
1529          */
1530
1531         /*
1532          * And attach them to the current transaction.  But only if
1533          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1534          * and generally junk.
1535          */
1536         if (ret == 0) {
1537                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1538                                         NULL, journal_dirty_data_fn);
1539                 if (!ret)
1540                         ret = err;
1541         }
1542         walk_page_buffers(handle, page_bufs, 0,
1543                         PAGE_CACHE_SIZE, NULL, bput_one);
1544         err = ext3_journal_stop(handle);
1545         if (!ret)
1546                 ret = err;
1547         return ret;
1548
1549 out_fail:
1550         redirty_page_for_writepage(wbc, page);
1551         unlock_page(page);
1552         return ret;
1553 }
1554
1555 static int ext3_writeback_writepage(struct page *page,
1556                                 struct writeback_control *wbc)
1557 {
1558         struct inode *inode = page->mapping->host;
1559         handle_t *handle = NULL;
1560         int ret = 0;
1561         int err;
1562
1563         if (ext3_journal_current_handle())
1564                 goto out_fail;
1565
1566         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1567         if (IS_ERR(handle)) {
1568                 ret = PTR_ERR(handle);
1569                 goto out_fail;
1570         }
1571
1572         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1573                 ret = nobh_writepage(page, ext3_get_block, wbc);
1574         else
1575                 ret = block_write_full_page(page, ext3_get_block, wbc);
1576
1577         err = ext3_journal_stop(handle);
1578         if (!ret)
1579                 ret = err;
1580         return ret;
1581
1582 out_fail:
1583         redirty_page_for_writepage(wbc, page);
1584         unlock_page(page);
1585         return ret;
1586 }
1587
1588 static int ext3_journalled_writepage(struct page *page,
1589                                 struct writeback_control *wbc)
1590 {
1591         struct inode *inode = page->mapping->host;
1592         handle_t *handle = NULL;
1593         int ret = 0;
1594         int err;
1595
1596         if (ext3_journal_current_handle())
1597                 goto no_write;
1598
1599         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1600         if (IS_ERR(handle)) {
1601                 ret = PTR_ERR(handle);
1602                 goto no_write;
1603         }
1604
1605         if (!page_has_buffers(page) || PageChecked(page)) {
1606                 /*
1607                  * It's mmapped pagecache.  Add buffers and journal it.  There
1608                  * doesn't seem much point in redirtying the page here.
1609                  */
1610                 ClearPageChecked(page);
1611                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1612                                         ext3_get_block);
1613                 if (ret != 0) {
1614                         ext3_journal_stop(handle);
1615                         goto out_unlock;
1616                 }
1617                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1618                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1619
1620                 err = walk_page_buffers(handle, page_buffers(page), 0,
1621                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1622                 if (ret == 0)
1623                         ret = err;
1624                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1625                 unlock_page(page);
1626         } else {
1627                 /*
1628                  * It may be a page full of checkpoint-mode buffers.  We don't
1629                  * really know unless we go poke around in the buffer_heads.
1630                  * But block_write_full_page will do the right thing.
1631                  */
1632                 ret = block_write_full_page(page, ext3_get_block, wbc);
1633         }
1634         err = ext3_journal_stop(handle);
1635         if (!ret)
1636                 ret = err;
1637 out:
1638         return ret;
1639
1640 no_write:
1641         redirty_page_for_writepage(wbc, page);
1642 out_unlock:
1643         unlock_page(page);
1644         goto out;
1645 }
1646
1647 static int ext3_readpage(struct file *file, struct page *page)
1648 {
1649         return mpage_readpage(page, ext3_get_block);
1650 }
1651
1652 static int
1653 ext3_readpages(struct file *file, struct address_space *mapping,
1654                 struct list_head *pages, unsigned nr_pages)
1655 {
1656         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1657 }
1658
1659 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1660 {
1661         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1662
1663         /*
1664          * If it's a full truncate we just forget about the pending dirtying
1665          */
1666         if (offset == 0)
1667                 ClearPageChecked(page);
1668
1669         journal_invalidatepage(journal, page, offset);
1670 }
1671
1672 static int ext3_releasepage(struct page *page, gfp_t wait)
1673 {
1674         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1675
1676         WARN_ON(PageChecked(page));
1677         if (!page_has_buffers(page))
1678                 return 0;
1679         return journal_try_to_free_buffers(journal, page, wait);
1680 }
1681
1682 /*
1683  * If the O_DIRECT write will extend the file then add this inode to the
1684  * orphan list.  So recovery will truncate it back to the original size
1685  * if the machine crashes during the write.
1686  *
1687  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1688  * crashes then stale disk data _may_ be exposed inside the file. But current
1689  * VFS code falls back into buffered path in that case so we are safe.
1690  */
1691 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1692                         const struct iovec *iov, loff_t offset,
1693                         unsigned long nr_segs)
1694 {
1695         struct file *file = iocb->ki_filp;
1696         struct inode *inode = file->f_mapping->host;
1697         struct ext3_inode_info *ei = EXT3_I(inode);
1698         handle_t *handle;
1699         ssize_t ret;
1700         int orphan = 0;
1701         size_t count = iov_length(iov, nr_segs);
1702
1703         if (rw == WRITE) {
1704                 loff_t final_size = offset + count;
1705
1706                 if (final_size > inode->i_size) {
1707                         /* Credits for sb + inode write */
1708                         handle = ext3_journal_start(inode, 2);
1709                         if (IS_ERR(handle)) {
1710                                 ret = PTR_ERR(handle);
1711                                 goto out;
1712                         }
1713                         ret = ext3_orphan_add(handle, inode);
1714                         if (ret) {
1715                                 ext3_journal_stop(handle);
1716                                 goto out;
1717                         }
1718                         orphan = 1;
1719                         ei->i_disksize = inode->i_size;
1720                         ext3_journal_stop(handle);
1721                 }
1722         }
1723
1724         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1725                                  offset, nr_segs,
1726                                  ext3_get_block, NULL);
1727
1728         if (orphan) {
1729                 int err;
1730
1731                 /* Credits for sb + inode write */
1732                 handle = ext3_journal_start(inode, 2);
1733                 if (IS_ERR(handle)) {
1734                         /* This is really bad luck. We've written the data
1735                          * but cannot extend i_size. Bail out and pretend
1736                          * the write failed... */
1737                         ret = PTR_ERR(handle);
1738                         goto out;
1739                 }
1740                 if (inode->i_nlink)
1741                         ext3_orphan_del(handle, inode);
1742                 if (ret > 0) {
1743                         loff_t end = offset + ret;
1744                         if (end > inode->i_size) {
1745                                 ei->i_disksize = end;
1746                                 i_size_write(inode, end);
1747                                 /*
1748                                  * We're going to return a positive `ret'
1749                                  * here due to non-zero-length I/O, so there's
1750                                  * no way of reporting error returns from
1751                                  * ext3_mark_inode_dirty() to userspace.  So
1752                                  * ignore it.
1753                                  */
1754                                 ext3_mark_inode_dirty(handle, inode);
1755                         }
1756                 }
1757                 err = ext3_journal_stop(handle);
1758                 if (ret == 0)
1759                         ret = err;
1760         }
1761 out:
1762         return ret;
1763 }
1764
1765 /*
1766  * Pages can be marked dirty completely asynchronously from ext3's journalling
1767  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1768  * much here because ->set_page_dirty is called under VFS locks.  The page is
1769  * not necessarily locked.
1770  *
1771  * We cannot just dirty the page and leave attached buffers clean, because the
1772  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1773  * or jbddirty because all the journalling code will explode.
1774  *
1775  * So what we do is to mark the page "pending dirty" and next time writepage
1776  * is called, propagate that into the buffers appropriately.
1777  */
1778 static int ext3_journalled_set_page_dirty(struct page *page)
1779 {
1780         SetPageChecked(page);
1781         return __set_page_dirty_nobuffers(page);
1782 }
1783
1784 static const struct address_space_operations ext3_ordered_aops = {
1785         .readpage               = ext3_readpage,
1786         .readpages              = ext3_readpages,
1787         .writepage              = ext3_ordered_writepage,
1788         .sync_page              = block_sync_page,
1789         .write_begin            = ext3_write_begin,
1790         .write_end              = ext3_ordered_write_end,
1791         .bmap                   = ext3_bmap,
1792         .invalidatepage         = ext3_invalidatepage,
1793         .releasepage            = ext3_releasepage,
1794         .direct_IO              = ext3_direct_IO,
1795         .migratepage            = buffer_migrate_page,
1796         .is_partially_uptodate  = block_is_partially_uptodate,
1797 };
1798
1799 static const struct address_space_operations ext3_writeback_aops = {
1800         .readpage               = ext3_readpage,
1801         .readpages              = ext3_readpages,
1802         .writepage              = ext3_writeback_writepage,
1803         .sync_page              = block_sync_page,
1804         .write_begin            = ext3_write_begin,
1805         .write_end              = ext3_writeback_write_end,
1806         .bmap                   = ext3_bmap,
1807         .invalidatepage         = ext3_invalidatepage,
1808         .releasepage            = ext3_releasepage,
1809         .direct_IO              = ext3_direct_IO,
1810         .migratepage            = buffer_migrate_page,
1811         .is_partially_uptodate  = block_is_partially_uptodate,
1812 };
1813
1814 static const struct address_space_operations ext3_journalled_aops = {
1815         .readpage               = ext3_readpage,
1816         .readpages              = ext3_readpages,
1817         .writepage              = ext3_journalled_writepage,
1818         .sync_page              = block_sync_page,
1819         .write_begin            = ext3_write_begin,
1820         .write_end              = ext3_journalled_write_end,
1821         .set_page_dirty         = ext3_journalled_set_page_dirty,
1822         .bmap                   = ext3_bmap,
1823         .invalidatepage         = ext3_invalidatepage,
1824         .releasepage            = ext3_releasepage,
1825         .is_partially_uptodate  = block_is_partially_uptodate,
1826 };
1827
1828 void ext3_set_aops(struct inode *inode)
1829 {
1830         if (ext3_should_order_data(inode))
1831                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1832         else if (ext3_should_writeback_data(inode))
1833                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1834         else
1835                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1836 }
1837
1838 /*
1839  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1840  * up to the end of the block which corresponds to `from'.
1841  * This required during truncate. We need to physically zero the tail end
1842  * of that block so it doesn't yield old data if the file is later grown.
1843  */
1844 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1845                 struct address_space *mapping, loff_t from)
1846 {
1847         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1848         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1849         unsigned blocksize, iblock, length, pos;
1850         struct inode *inode = mapping->host;
1851         struct buffer_head *bh;
1852         int err = 0;
1853
1854         blocksize = inode->i_sb->s_blocksize;
1855         length = blocksize - (offset & (blocksize - 1));
1856         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1857
1858         /*
1859          * For "nobh" option,  we can only work if we don't need to
1860          * read-in the page - otherwise we create buffers to do the IO.
1861          */
1862         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1863              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1864                 zero_user(page, offset, length);
1865                 set_page_dirty(page);
1866                 goto unlock;
1867         }
1868
1869         if (!page_has_buffers(page))
1870                 create_empty_buffers(page, blocksize, 0);
1871
1872         /* Find the buffer that contains "offset" */
1873         bh = page_buffers(page);
1874         pos = blocksize;
1875         while (offset >= pos) {
1876                 bh = bh->b_this_page;
1877                 iblock++;
1878                 pos += blocksize;
1879         }
1880
1881         err = 0;
1882         if (buffer_freed(bh)) {
1883                 BUFFER_TRACE(bh, "freed: skip");
1884                 goto unlock;
1885         }
1886
1887         if (!buffer_mapped(bh)) {
1888                 BUFFER_TRACE(bh, "unmapped");
1889                 ext3_get_block(inode, iblock, bh, 0);
1890                 /* unmapped? It's a hole - nothing to do */
1891                 if (!buffer_mapped(bh)) {
1892                         BUFFER_TRACE(bh, "still unmapped");
1893                         goto unlock;
1894                 }
1895         }
1896
1897         /* Ok, it's mapped. Make sure it's up-to-date */
1898         if (PageUptodate(page))
1899                 set_buffer_uptodate(bh);
1900
1901         if (!buffer_uptodate(bh)) {
1902                 err = -EIO;
1903                 ll_rw_block(READ, 1, &bh);
1904                 wait_on_buffer(bh);
1905                 /* Uhhuh. Read error. Complain and punt. */
1906                 if (!buffer_uptodate(bh))
1907                         goto unlock;
1908         }
1909
1910         if (ext3_should_journal_data(inode)) {
1911                 BUFFER_TRACE(bh, "get write access");
1912                 err = ext3_journal_get_write_access(handle, bh);
1913                 if (err)
1914                         goto unlock;
1915         }
1916
1917         zero_user(page, offset, length);
1918         BUFFER_TRACE(bh, "zeroed end of block");
1919
1920         err = 0;
1921         if (ext3_should_journal_data(inode)) {
1922                 err = ext3_journal_dirty_metadata(handle, bh);
1923         } else {
1924                 if (ext3_should_order_data(inode))
1925                         err = ext3_journal_dirty_data(handle, bh);
1926                 mark_buffer_dirty(bh);
1927         }
1928
1929 unlock:
1930         unlock_page(page);
1931         page_cache_release(page);
1932         return err;
1933 }
1934
1935 /*
1936  * Probably it should be a library function... search for first non-zero word
1937  * or memcmp with zero_page, whatever is better for particular architecture.
1938  * Linus?
1939  */
1940 static inline int all_zeroes(__le32 *p, __le32 *q)
1941 {
1942         while (p < q)
1943                 if (*p++)
1944                         return 0;
1945         return 1;
1946 }
1947
1948 /**
1949  *      ext3_find_shared - find the indirect blocks for partial truncation.
1950  *      @inode:   inode in question
1951  *      @depth:   depth of the affected branch
1952  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1953  *      @chain:   place to store the pointers to partial indirect blocks
1954  *      @top:     place to the (detached) top of branch
1955  *
1956  *      This is a helper function used by ext3_truncate().
1957  *
1958  *      When we do truncate() we may have to clean the ends of several
1959  *      indirect blocks but leave the blocks themselves alive. Block is
1960  *      partially truncated if some data below the new i_size is refered
1961  *      from it (and it is on the path to the first completely truncated
1962  *      data block, indeed).  We have to free the top of that path along
1963  *      with everything to the right of the path. Since no allocation
1964  *      past the truncation point is possible until ext3_truncate()
1965  *      finishes, we may safely do the latter, but top of branch may
1966  *      require special attention - pageout below the truncation point
1967  *      might try to populate it.
1968  *
1969  *      We atomically detach the top of branch from the tree, store the
1970  *      block number of its root in *@top, pointers to buffer_heads of
1971  *      partially truncated blocks - in @chain[].bh and pointers to
1972  *      their last elements that should not be removed - in
1973  *      @chain[].p. Return value is the pointer to last filled element
1974  *      of @chain.
1975  *
1976  *      The work left to caller to do the actual freeing of subtrees:
1977  *              a) free the subtree starting from *@top
1978  *              b) free the subtrees whose roots are stored in
1979  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1980  *              c) free the subtrees growing from the inode past the @chain[0].
1981  *                      (no partially truncated stuff there).  */
1982
1983 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1984                         int offsets[4], Indirect chain[4], __le32 *top)
1985 {
1986         Indirect *partial, *p;
1987         int k, err;
1988
1989         *top = 0;
1990         /* Make k index the deepest non-null offest + 1 */
1991         for (k = depth; k > 1 && !offsets[k-1]; k--)
1992                 ;
1993         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1994         /* Writer: pointers */
1995         if (!partial)
1996                 partial = chain + k-1;
1997         /*
1998          * If the branch acquired continuation since we've looked at it -
1999          * fine, it should all survive and (new) top doesn't belong to us.
2000          */
2001         if (!partial->key && *partial->p)
2002                 /* Writer: end */
2003                 goto no_top;
2004         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2005                 ;
2006         /*
2007          * OK, we've found the last block that must survive. The rest of our
2008          * branch should be detached before unlocking. However, if that rest
2009          * of branch is all ours and does not grow immediately from the inode
2010          * it's easier to cheat and just decrement partial->p.
2011          */
2012         if (p == chain + k - 1 && p > chain) {
2013                 p->p--;
2014         } else {
2015                 *top = *p->p;
2016                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2017 #if 0
2018                 *p->p = 0;
2019 #endif
2020         }
2021         /* Writer: end */
2022
2023         while(partial > p) {
2024                 brelse(partial->bh);
2025                 partial--;
2026         }
2027 no_top:
2028         return partial;
2029 }
2030
2031 /*
2032  * Zero a number of block pointers in either an inode or an indirect block.
2033  * If we restart the transaction we must again get write access to the
2034  * indirect block for further modification.
2035  *
2036  * We release `count' blocks on disk, but (last - first) may be greater
2037  * than `count' because there can be holes in there.
2038  */
2039 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2040                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2041                 unsigned long count, __le32 *first, __le32 *last)
2042 {
2043         __le32 *p;
2044         if (try_to_extend_transaction(handle, inode)) {
2045                 if (bh) {
2046                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2047                         ext3_journal_dirty_metadata(handle, bh);
2048                 }
2049                 ext3_mark_inode_dirty(handle, inode);
2050                 ext3_journal_test_restart(handle, inode);
2051                 if (bh) {
2052                         BUFFER_TRACE(bh, "retaking write access");
2053                         ext3_journal_get_write_access(handle, bh);
2054                 }
2055         }
2056
2057         /*
2058          * Any buffers which are on the journal will be in memory. We find
2059          * them on the hash table so journal_revoke() will run journal_forget()
2060          * on them.  We've already detached each block from the file, so
2061          * bforget() in journal_forget() should be safe.
2062          *
2063          * AKPM: turn on bforget in journal_forget()!!!
2064          */
2065         for (p = first; p < last; p++) {
2066                 u32 nr = le32_to_cpu(*p);
2067                 if (nr) {
2068                         struct buffer_head *bh;
2069
2070                         *p = 0;
2071                         bh = sb_find_get_block(inode->i_sb, nr);
2072                         ext3_forget(handle, 0, inode, bh, nr);
2073                 }
2074         }
2075
2076         ext3_free_blocks(handle, inode, block_to_free, count);
2077 }
2078
2079 /**
2080  * ext3_free_data - free a list of data blocks
2081  * @handle:     handle for this transaction
2082  * @inode:      inode we are dealing with
2083  * @this_bh:    indirect buffer_head which contains *@first and *@last
2084  * @first:      array of block numbers
2085  * @last:       points immediately past the end of array
2086  *
2087  * We are freeing all blocks refered from that array (numbers are stored as
2088  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2089  *
2090  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2091  * blocks are contiguous then releasing them at one time will only affect one
2092  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2093  * actually use a lot of journal space.
2094  *
2095  * @this_bh will be %NULL if @first and @last point into the inode's direct
2096  * block pointers.
2097  */
2098 static void ext3_free_data(handle_t *handle, struct inode *inode,
2099                            struct buffer_head *this_bh,
2100                            __le32 *first, __le32 *last)
2101 {
2102         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2103         unsigned long count = 0;            /* Number of blocks in the run */
2104         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2105                                                corresponding to
2106                                                block_to_free */
2107         ext3_fsblk_t nr;                    /* Current block # */
2108         __le32 *p;                          /* Pointer into inode/ind
2109                                                for current block */
2110         int err;
2111
2112         if (this_bh) {                          /* For indirect block */
2113                 BUFFER_TRACE(this_bh, "get_write_access");
2114                 err = ext3_journal_get_write_access(handle, this_bh);
2115                 /* Important: if we can't update the indirect pointers
2116                  * to the blocks, we can't free them. */
2117                 if (err)
2118                         return;
2119         }
2120
2121         for (p = first; p < last; p++) {
2122                 nr = le32_to_cpu(*p);
2123                 if (nr) {
2124                         /* accumulate blocks to free if they're contiguous */
2125                         if (count == 0) {
2126                                 block_to_free = nr;
2127                                 block_to_free_p = p;
2128                                 count = 1;
2129                         } else if (nr == block_to_free + count) {
2130                                 count++;
2131                         } else {
2132                                 ext3_clear_blocks(handle, inode, this_bh,
2133                                                   block_to_free,
2134                                                   count, block_to_free_p, p);
2135                                 block_to_free = nr;
2136                                 block_to_free_p = p;
2137                                 count = 1;
2138                         }
2139                 }
2140         }
2141
2142         if (count > 0)
2143                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2144                                   count, block_to_free_p, p);
2145
2146         if (this_bh) {
2147                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2148
2149                 /*
2150                  * The buffer head should have an attached journal head at this
2151                  * point. However, if the data is corrupted and an indirect
2152                  * block pointed to itself, it would have been detached when
2153                  * the block was cleared. Check for this instead of OOPSing.
2154                  */
2155                 if (bh2jh(this_bh))
2156                         ext3_journal_dirty_metadata(handle, this_bh);
2157                 else
2158                         ext3_error(inode->i_sb, "ext3_free_data",
2159                                    "circular indirect block detected, "
2160                                    "inode=%lu, block=%llu",
2161                                    inode->i_ino,
2162                                    (unsigned long long)this_bh->b_blocknr);
2163         }
2164 }
2165
2166 /**
2167  *      ext3_free_branches - free an array of branches
2168  *      @handle: JBD handle for this transaction
2169  *      @inode: inode we are dealing with
2170  *      @parent_bh: the buffer_head which contains *@first and *@last
2171  *      @first: array of block numbers
2172  *      @last:  pointer immediately past the end of array
2173  *      @depth: depth of the branches to free
2174  *
2175  *      We are freeing all blocks refered from these branches (numbers are
2176  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2177  *      appropriately.
2178  */
2179 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2180                                struct buffer_head *parent_bh,
2181                                __le32 *first, __le32 *last, int depth)
2182 {
2183         ext3_fsblk_t nr;
2184         __le32 *p;
2185
2186         if (is_handle_aborted(handle))
2187                 return;
2188
2189         if (depth--) {
2190                 struct buffer_head *bh;
2191                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2192                 p = last;
2193                 while (--p >= first) {
2194                         nr = le32_to_cpu(*p);
2195                         if (!nr)
2196                                 continue;               /* A hole */
2197
2198                         /* Go read the buffer for the next level down */
2199                         bh = sb_bread(inode->i_sb, nr);
2200
2201                         /*
2202                          * A read failure? Report error and clear slot
2203                          * (should be rare).
2204                          */
2205                         if (!bh) {
2206                                 ext3_error(inode->i_sb, "ext3_free_branches",
2207                                            "Read failure, inode=%lu, block="E3FSBLK,
2208                                            inode->i_ino, nr);
2209                                 continue;
2210                         }
2211
2212                         /* This zaps the entire block.  Bottom up. */
2213                         BUFFER_TRACE(bh, "free child branches");
2214                         ext3_free_branches(handle, inode, bh,
2215                                            (__le32*)bh->b_data,
2216                                            (__le32*)bh->b_data + addr_per_block,
2217                                            depth);
2218
2219                         /*
2220                          * We've probably journalled the indirect block several
2221                          * times during the truncate.  But it's no longer
2222                          * needed and we now drop it from the transaction via
2223                          * journal_revoke().
2224                          *
2225                          * That's easy if it's exclusively part of this
2226                          * transaction.  But if it's part of the committing
2227                          * transaction then journal_forget() will simply
2228                          * brelse() it.  That means that if the underlying
2229                          * block is reallocated in ext3_get_block(),
2230                          * unmap_underlying_metadata() will find this block
2231                          * and will try to get rid of it.  damn, damn.
2232                          *
2233                          * If this block has already been committed to the
2234                          * journal, a revoke record will be written.  And
2235                          * revoke records must be emitted *before* clearing
2236                          * this block's bit in the bitmaps.
2237                          */
2238                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2239
2240                         /*
2241                          * Everything below this this pointer has been
2242                          * released.  Now let this top-of-subtree go.
2243                          *
2244                          * We want the freeing of this indirect block to be
2245                          * atomic in the journal with the updating of the
2246                          * bitmap block which owns it.  So make some room in
2247                          * the journal.
2248                          *
2249                          * We zero the parent pointer *after* freeing its
2250                          * pointee in the bitmaps, so if extend_transaction()
2251                          * for some reason fails to put the bitmap changes and
2252                          * the release into the same transaction, recovery
2253                          * will merely complain about releasing a free block,
2254                          * rather than leaking blocks.
2255                          */
2256                         if (is_handle_aborted(handle))
2257                                 return;
2258                         if (try_to_extend_transaction(handle, inode)) {
2259                                 ext3_mark_inode_dirty(handle, inode);
2260                                 ext3_journal_test_restart(handle, inode);
2261                         }
2262
2263                         ext3_free_blocks(handle, inode, nr, 1);
2264
2265                         if (parent_bh) {
2266                                 /*
2267                                  * The block which we have just freed is
2268                                  * pointed to by an indirect block: journal it
2269                                  */
2270                                 BUFFER_TRACE(parent_bh, "get_write_access");
2271                                 if (!ext3_journal_get_write_access(handle,
2272                                                                    parent_bh)){
2273                                         *p = 0;
2274                                         BUFFER_TRACE(parent_bh,
2275                                         "call ext3_journal_dirty_metadata");
2276                                         ext3_journal_dirty_metadata(handle,
2277                                                                     parent_bh);
2278                                 }
2279                         }
2280                 }
2281         } else {
2282                 /* We have reached the bottom of the tree. */
2283                 BUFFER_TRACE(parent_bh, "free data blocks");
2284                 ext3_free_data(handle, inode, parent_bh, first, last);
2285         }
2286 }
2287
2288 int ext3_can_truncate(struct inode *inode)
2289 {
2290         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2291                 return 0;
2292         if (S_ISREG(inode->i_mode))
2293                 return 1;
2294         if (S_ISDIR(inode->i_mode))
2295                 return 1;
2296         if (S_ISLNK(inode->i_mode))
2297                 return !ext3_inode_is_fast_symlink(inode);
2298         return 0;
2299 }
2300
2301 /*
2302  * ext3_truncate()
2303  *
2304  * We block out ext3_get_block() block instantiations across the entire
2305  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2306  * simultaneously on behalf of the same inode.
2307  *
2308  * As we work through the truncate and commmit bits of it to the journal there
2309  * is one core, guiding principle: the file's tree must always be consistent on
2310  * disk.  We must be able to restart the truncate after a crash.
2311  *
2312  * The file's tree may be transiently inconsistent in memory (although it
2313  * probably isn't), but whenever we close off and commit a journal transaction,
2314  * the contents of (the filesystem + the journal) must be consistent and
2315  * restartable.  It's pretty simple, really: bottom up, right to left (although
2316  * left-to-right works OK too).
2317  *
2318  * Note that at recovery time, journal replay occurs *before* the restart of
2319  * truncate against the orphan inode list.
2320  *
2321  * The committed inode has the new, desired i_size (which is the same as
2322  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2323  * that this inode's truncate did not complete and it will again call
2324  * ext3_truncate() to have another go.  So there will be instantiated blocks
2325  * to the right of the truncation point in a crashed ext3 filesystem.  But
2326  * that's fine - as long as they are linked from the inode, the post-crash
2327  * ext3_truncate() run will find them and release them.
2328  */
2329 void ext3_truncate(struct inode *inode)
2330 {
2331         handle_t *handle;
2332         struct ext3_inode_info *ei = EXT3_I(inode);
2333         __le32 *i_data = ei->i_data;
2334         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2335         struct address_space *mapping = inode->i_mapping;
2336         int offsets[4];
2337         Indirect chain[4];
2338         Indirect *partial;
2339         __le32 nr = 0;
2340         int n;
2341         long last_block;
2342         unsigned blocksize = inode->i_sb->s_blocksize;
2343         struct page *page;
2344
2345         if (!ext3_can_truncate(inode))
2346                 return;
2347
2348         /*
2349          * We have to lock the EOF page here, because lock_page() nests
2350          * outside journal_start().
2351          */
2352         if ((inode->i_size & (blocksize - 1)) == 0) {
2353                 /* Block boundary? Nothing to do */
2354                 page = NULL;
2355         } else {
2356                 page = grab_cache_page(mapping,
2357                                 inode->i_size >> PAGE_CACHE_SHIFT);
2358                 if (!page)
2359                         return;
2360         }
2361
2362         handle = start_transaction(inode);
2363         if (IS_ERR(handle)) {
2364                 if (page) {
2365                         clear_highpage(page);
2366                         flush_dcache_page(page);
2367                         unlock_page(page);
2368                         page_cache_release(page);
2369                 }
2370                 return;         /* AKPM: return what? */
2371         }
2372
2373         last_block = (inode->i_size + blocksize-1)
2374                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2375
2376         if (page)
2377                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2378
2379         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2380         if (n == 0)
2381                 goto out_stop;  /* error */
2382
2383         /*
2384          * OK.  This truncate is going to happen.  We add the inode to the
2385          * orphan list, so that if this truncate spans multiple transactions,
2386          * and we crash, we will resume the truncate when the filesystem
2387          * recovers.  It also marks the inode dirty, to catch the new size.
2388          *
2389          * Implication: the file must always be in a sane, consistent
2390          * truncatable state while each transaction commits.
2391          */
2392         if (ext3_orphan_add(handle, inode))
2393                 goto out_stop;
2394
2395         /*
2396          * The orphan list entry will now protect us from any crash which
2397          * occurs before the truncate completes, so it is now safe to propagate
2398          * the new, shorter inode size (held for now in i_size) into the
2399          * on-disk inode. We do this via i_disksize, which is the value which
2400          * ext3 *really* writes onto the disk inode.
2401          */
2402         ei->i_disksize = inode->i_size;
2403
2404         /*
2405          * From here we block out all ext3_get_block() callers who want to
2406          * modify the block allocation tree.
2407          */
2408         mutex_lock(&ei->truncate_mutex);
2409
2410         if (n == 1) {           /* direct blocks */
2411                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2412                                i_data + EXT3_NDIR_BLOCKS);
2413                 goto do_indirects;
2414         }
2415
2416         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2417         /* Kill the top of shared branch (not detached) */
2418         if (nr) {
2419                 if (partial == chain) {
2420                         /* Shared branch grows from the inode */
2421                         ext3_free_branches(handle, inode, NULL,
2422                                            &nr, &nr+1, (chain+n-1) - partial);
2423                         *partial->p = 0;
2424                         /*
2425                          * We mark the inode dirty prior to restart,
2426                          * and prior to stop.  No need for it here.
2427                          */
2428                 } else {
2429                         /* Shared branch grows from an indirect block */
2430                         BUFFER_TRACE(partial->bh, "get_write_access");
2431                         ext3_free_branches(handle, inode, partial->bh,
2432                                         partial->p,
2433                                         partial->p+1, (chain+n-1) - partial);
2434                 }
2435         }
2436         /* Clear the ends of indirect blocks on the shared branch */
2437         while (partial > chain) {
2438                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2439                                    (__le32*)partial->bh->b_data+addr_per_block,
2440                                    (chain+n-1) - partial);
2441                 BUFFER_TRACE(partial->bh, "call brelse");
2442                 brelse (partial->bh);
2443                 partial--;
2444         }
2445 do_indirects:
2446         /* Kill the remaining (whole) subtrees */
2447         switch (offsets[0]) {
2448         default:
2449                 nr = i_data[EXT3_IND_BLOCK];
2450                 if (nr) {
2451                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2452                         i_data[EXT3_IND_BLOCK] = 0;
2453                 }
2454         case EXT3_IND_BLOCK:
2455                 nr = i_data[EXT3_DIND_BLOCK];
2456                 if (nr) {
2457                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2458                         i_data[EXT3_DIND_BLOCK] = 0;
2459                 }
2460         case EXT3_DIND_BLOCK:
2461                 nr = i_data[EXT3_TIND_BLOCK];
2462                 if (nr) {
2463                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2464                         i_data[EXT3_TIND_BLOCK] = 0;
2465                 }
2466         case EXT3_TIND_BLOCK:
2467                 ;
2468         }
2469
2470         ext3_discard_reservation(inode);
2471
2472         mutex_unlock(&ei->truncate_mutex);
2473         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2474         ext3_mark_inode_dirty(handle, inode);
2475
2476         /*
2477          * In a multi-transaction truncate, we only make the final transaction
2478          * synchronous
2479          */
2480         if (IS_SYNC(inode))
2481                 handle->h_sync = 1;
2482 out_stop:
2483         /*
2484          * If this was a simple ftruncate(), and the file will remain alive
2485          * then we need to clear up the orphan record which we created above.
2486          * However, if this was a real unlink then we were called by
2487          * ext3_delete_inode(), and we allow that function to clean up the
2488          * orphan info for us.
2489          */
2490         if (inode->i_nlink)
2491                 ext3_orphan_del(handle, inode);
2492
2493         ext3_journal_stop(handle);
2494 }
2495
2496 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2497                 unsigned long ino, struct ext3_iloc *iloc)
2498 {
2499         unsigned long block_group;
2500         unsigned long offset;
2501         ext3_fsblk_t block;
2502         struct ext3_group_desc *gdp;
2503
2504         if (!ext3_valid_inum(sb, ino)) {
2505                 /*
2506                  * This error is already checked for in namei.c unless we are
2507                  * looking at an NFS filehandle, in which case no error
2508                  * report is needed
2509                  */
2510                 return 0;
2511         }
2512
2513         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2514         gdp = ext3_get_group_desc(sb, block_group, NULL);
2515         if (!gdp)
2516                 return 0;
2517         /*
2518          * Figure out the offset within the block group inode table
2519          */
2520         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2521                 EXT3_INODE_SIZE(sb);
2522         block = le32_to_cpu(gdp->bg_inode_table) +
2523                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2524
2525         iloc->block_group = block_group;
2526         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2527         return block;
2528 }
2529
2530 /*
2531  * ext3_get_inode_loc returns with an extra refcount against the inode's
2532  * underlying buffer_head on success. If 'in_mem' is true, we have all
2533  * data in memory that is needed to recreate the on-disk version of this
2534  * inode.
2535  */
2536 static int __ext3_get_inode_loc(struct inode *inode,
2537                                 struct ext3_iloc *iloc, int in_mem)
2538 {
2539         ext3_fsblk_t block;
2540         struct buffer_head *bh;
2541
2542         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2543         if (!block)
2544                 return -EIO;
2545
2546         bh = sb_getblk(inode->i_sb, block);
2547         if (!bh) {
2548                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2549                                 "unable to read inode block - "
2550                                 "inode=%lu, block="E3FSBLK,
2551                                  inode->i_ino, block);
2552                 return -EIO;
2553         }
2554         if (!buffer_uptodate(bh)) {
2555                 lock_buffer(bh);
2556
2557                 /*
2558                  * If the buffer has the write error flag, we have failed
2559                  * to write out another inode in the same block.  In this
2560                  * case, we don't have to read the block because we may
2561                  * read the old inode data successfully.
2562                  */
2563                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2564                         set_buffer_uptodate(bh);
2565
2566                 if (buffer_uptodate(bh)) {
2567                         /* someone brought it uptodate while we waited */
2568                         unlock_buffer(bh);
2569                         goto has_buffer;
2570                 }
2571
2572                 /*
2573                  * If we have all information of the inode in memory and this
2574                  * is the only valid inode in the block, we need not read the
2575                  * block.
2576                  */
2577                 if (in_mem) {
2578                         struct buffer_head *bitmap_bh;
2579                         struct ext3_group_desc *desc;
2580                         int inodes_per_buffer;
2581                         int inode_offset, i;
2582                         int block_group;
2583                         int start;
2584
2585                         block_group = (inode->i_ino - 1) /
2586                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2587                         inodes_per_buffer = bh->b_size /
2588                                 EXT3_INODE_SIZE(inode->i_sb);
2589                         inode_offset = ((inode->i_ino - 1) %
2590                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2591                         start = inode_offset & ~(inodes_per_buffer - 1);
2592
2593                         /* Is the inode bitmap in cache? */
2594                         desc = ext3_get_group_desc(inode->i_sb,
2595                                                 block_group, NULL);
2596                         if (!desc)
2597                                 goto make_io;
2598
2599                         bitmap_bh = sb_getblk(inode->i_sb,
2600                                         le32_to_cpu(desc->bg_inode_bitmap));
2601                         if (!bitmap_bh)
2602                                 goto make_io;
2603
2604                         /*
2605                          * If the inode bitmap isn't in cache then the
2606                          * optimisation may end up performing two reads instead
2607                          * of one, so skip it.
2608                          */
2609                         if (!buffer_uptodate(bitmap_bh)) {
2610                                 brelse(bitmap_bh);
2611                                 goto make_io;
2612                         }
2613                         for (i = start; i < start + inodes_per_buffer; i++) {
2614                                 if (i == inode_offset)
2615                                         continue;
2616                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2617                                         break;
2618                         }
2619                         brelse(bitmap_bh);
2620                         if (i == start + inodes_per_buffer) {
2621                                 /* all other inodes are free, so skip I/O */
2622                                 memset(bh->b_data, 0, bh->b_size);
2623                                 set_buffer_uptodate(bh);
2624                                 unlock_buffer(bh);
2625                                 goto has_buffer;
2626                         }
2627                 }
2628
2629 make_io:
2630                 /*
2631                  * There are other valid inodes in the buffer, this inode
2632                  * has in-inode xattrs, or we don't have this inode in memory.
2633                  * Read the block from disk.
2634                  */
2635                 get_bh(bh);
2636                 bh->b_end_io = end_buffer_read_sync;
2637                 submit_bh(READ_META, bh);
2638                 wait_on_buffer(bh);
2639                 if (!buffer_uptodate(bh)) {
2640                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2641                                         "unable to read inode block - "
2642                                         "inode=%lu, block="E3FSBLK,
2643                                         inode->i_ino, block);
2644                         brelse(bh);
2645                         return -EIO;
2646                 }
2647         }
2648 has_buffer:
2649         iloc->bh = bh;
2650         return 0;
2651 }
2652
2653 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2654 {
2655         /* We have all inode data except xattrs in memory here. */
2656         return __ext3_get_inode_loc(inode, iloc,
2657                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2658 }
2659
2660 void ext3_set_inode_flags(struct inode *inode)
2661 {
2662         unsigned int flags = EXT3_I(inode)->i_flags;
2663
2664         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2665         if (flags & EXT3_SYNC_FL)
2666                 inode->i_flags |= S_SYNC;
2667         if (flags & EXT3_APPEND_FL)
2668                 inode->i_flags |= S_APPEND;
2669         if (flags & EXT3_IMMUTABLE_FL)
2670                 inode->i_flags |= S_IMMUTABLE;
2671         if (flags & EXT3_NOATIME_FL)
2672                 inode->i_flags |= S_NOATIME;
2673         if (flags & EXT3_DIRSYNC_FL)
2674                 inode->i_flags |= S_DIRSYNC;
2675 }
2676
2677 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2678 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2679 {
2680         unsigned int flags = ei->vfs_inode.i_flags;
2681
2682         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2683                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2684         if (flags & S_SYNC)
2685                 ei->i_flags |= EXT3_SYNC_FL;
2686         if (flags & S_APPEND)
2687                 ei->i_flags |= EXT3_APPEND_FL;
2688         if (flags & S_IMMUTABLE)
2689                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2690         if (flags & S_NOATIME)
2691                 ei->i_flags |= EXT3_NOATIME_FL;
2692         if (flags & S_DIRSYNC)
2693                 ei->i_flags |= EXT3_DIRSYNC_FL;
2694 }
2695
2696 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2697 {
2698         struct ext3_iloc iloc;
2699         struct ext3_inode *raw_inode;
2700         struct ext3_inode_info *ei;
2701         struct buffer_head *bh;
2702         struct inode *inode;
2703         long ret;
2704         int block;
2705
2706         inode = iget_locked(sb, ino);
2707         if (!inode)
2708                 return ERR_PTR(-ENOMEM);
2709         if (!(inode->i_state & I_NEW))
2710                 return inode;
2711
2712         ei = EXT3_I(inode);
2713 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2714         ei->i_acl = EXT3_ACL_NOT_CACHED;
2715         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2716 #endif
2717         ei->i_block_alloc_info = NULL;
2718
2719         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2720         if (ret < 0)
2721                 goto bad_inode;
2722         bh = iloc.bh;
2723         raw_inode = ext3_raw_inode(&iloc);
2724         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2725         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2726         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2727         if(!(test_opt (inode->i_sb, NO_UID32))) {
2728                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2729                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2730         }
2731         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2732         inode->i_size = le32_to_cpu(raw_inode->i_size);
2733         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2734         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2735         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2736         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2737
2738         ei->i_state = 0;
2739         ei->i_dir_start_lookup = 0;
2740         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2741         /* We now have enough fields to check if the inode was active or not.
2742          * This is needed because nfsd might try to access dead inodes
2743          * the test is that same one that e2fsck uses
2744          * NeilBrown 1999oct15
2745          */
2746         if (inode->i_nlink == 0) {
2747                 if (inode->i_mode == 0 ||
2748                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2749                         /* this inode is deleted */
2750                         brelse (bh);
2751                         ret = -ESTALE;
2752                         goto bad_inode;
2753                 }
2754                 /* The only unlinked inodes we let through here have
2755                  * valid i_mode and are being read by the orphan
2756                  * recovery code: that's fine, we're about to complete
2757                  * the process of deleting those. */
2758         }
2759         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2760         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2761 #ifdef EXT3_FRAGMENTS
2762         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2763         ei->i_frag_no = raw_inode->i_frag;
2764         ei->i_frag_size = raw_inode->i_fsize;
2765 #endif
2766         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2767         if (!S_ISREG(inode->i_mode)) {
2768                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2769         } else {
2770                 inode->i_size |=
2771                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2772         }
2773         ei->i_disksize = inode->i_size;
2774         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2775         ei->i_block_group = iloc.block_group;
2776         /*
2777          * NOTE! The in-memory inode i_data array is in little-endian order
2778          * even on big-endian machines: we do NOT byteswap the block numbers!
2779          */
2780         for (block = 0; block < EXT3_N_BLOCKS; block++)
2781                 ei->i_data[block] = raw_inode->i_block[block];
2782         INIT_LIST_HEAD(&ei->i_orphan);
2783
2784         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2785             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2786                 /*
2787                  * When mke2fs creates big inodes it does not zero out
2788                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2789                  * so ignore those first few inodes.
2790                  */
2791                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2792                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2793                     EXT3_INODE_SIZE(inode->i_sb)) {
2794                         brelse (bh);
2795                         ret = -EIO;
2796                         goto bad_inode;
2797                 }
2798                 if (ei->i_extra_isize == 0) {
2799                         /* The extra space is currently unused. Use it. */
2800                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2801                                             EXT3_GOOD_OLD_INODE_SIZE;
2802                 } else {
2803                         __le32 *magic = (void *)raw_inode +
2804                                         EXT3_GOOD_OLD_INODE_SIZE +
2805                                         ei->i_extra_isize;
2806                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2807                                  ei->i_state |= EXT3_STATE_XATTR;
2808                 }
2809         } else
2810                 ei->i_extra_isize = 0;
2811
2812         if (S_ISREG(inode->i_mode)) {
2813                 inode->i_op = &ext3_file_inode_operations;
2814                 inode->i_fop = &ext3_file_operations;
2815                 ext3_set_aops(inode);
2816         } else if (S_ISDIR(inode->i_mode)) {
2817                 inode->i_op = &ext3_dir_inode_operations;
2818                 inode->i_fop = &ext3_dir_operations;
2819         } else if (S_ISLNK(inode->i_mode)) {
2820                 if (ext3_inode_is_fast_symlink(inode))
2821                         inode->i_op = &ext3_fast_symlink_inode_operations;
2822                 else {
2823                         inode->i_op = &ext3_symlink_inode_operations;
2824                         ext3_set_aops(inode);
2825                 }
2826         } else {
2827                 inode->i_op = &ext3_special_inode_operations;
2828                 if (raw_inode->i_block[0])
2829                         init_special_inode(inode, inode->i_mode,
2830                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2831                 else
2832                         init_special_inode(inode, inode->i_mode,
2833                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2834         }
2835         brelse (iloc.bh);
2836         ext3_set_inode_flags(inode);
2837         unlock_new_inode(inode);
2838         return inode;
2839
2840 bad_inode:
2841         iget_failed(inode);
2842         return ERR_PTR(ret);
2843 }
2844
2845 /*
2846  * Post the struct inode info into an on-disk inode location in the
2847  * buffer-cache.  This gobbles the caller's reference to the
2848  * buffer_head in the inode location struct.
2849  *
2850  * The caller must have write access to iloc->bh.
2851  */
2852 static int ext3_do_update_inode(handle_t *handle,
2853                                 struct inode *inode,
2854                                 struct ext3_iloc *iloc)
2855 {
2856         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2857         struct ext3_inode_info *ei = EXT3_I(inode);
2858         struct buffer_head *bh = iloc->bh;
2859         int err = 0, rc, block;
2860
2861         /* For fields not not tracking in the in-memory inode,
2862          * initialise them to zero for new inodes. */
2863         if (ei->i_state & EXT3_STATE_NEW)
2864                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2865
2866         ext3_get_inode_flags(ei);
2867         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2868         if(!(test_opt(inode->i_sb, NO_UID32))) {
2869                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2870                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2871 /*
2872  * Fix up interoperability with old kernels. Otherwise, old inodes get
2873  * re-used with the upper 16 bits of the uid/gid intact
2874  */
2875                 if(!ei->i_dtime) {
2876                         raw_inode->i_uid_high =
2877                                 cpu_to_le16(high_16_bits(inode->i_uid));
2878                         raw_inode->i_gid_high =
2879                                 cpu_to_le16(high_16_bits(inode->i_gid));
2880                 } else {
2881                         raw_inode->i_uid_high = 0;
2882                         raw_inode->i_gid_high = 0;
2883                 }
2884         } else {
2885                 raw_inode->i_uid_low =
2886                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2887                 raw_inode->i_gid_low =
2888                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2889                 raw_inode->i_uid_high = 0;
2890                 raw_inode->i_gid_high = 0;
2891         }
2892         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2893         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2894         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2895         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2896         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2897         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2898         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2899         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2900 #ifdef EXT3_FRAGMENTS
2901         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2902         raw_inode->i_frag = ei->i_frag_no;
2903         raw_inode->i_fsize = ei->i_frag_size;
2904 #endif
2905         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2906         if (!S_ISREG(inode->i_mode)) {
2907                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2908         } else {
2909                 raw_inode->i_size_high =
2910                         cpu_to_le32(ei->i_disksize >> 32);
2911                 if (ei->i_disksize > 0x7fffffffULL) {
2912                         struct super_block *sb = inode->i_sb;
2913                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2914                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2915                             EXT3_SB(sb)->s_es->s_rev_level ==
2916                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2917                                /* If this is the first large file
2918                                 * created, add a flag to the superblock.
2919                                 */
2920                                 err = ext3_journal_get_write_access(handle,
2921                                                 EXT3_SB(sb)->s_sbh);
2922                                 if (err)
2923                                         goto out_brelse;
2924                                 ext3_update_dynamic_rev(sb);
2925                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2926                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2927                                 sb->s_dirt = 1;
2928                                 handle->h_sync = 1;
2929                                 err = ext3_journal_dirty_metadata(handle,
2930                                                 EXT3_SB(sb)->s_sbh);
2931                         }
2932                 }
2933         }
2934         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2935         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2936                 if (old_valid_dev(inode->i_rdev)) {
2937                         raw_inode->i_block[0] =
2938                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2939                         raw_inode->i_block[1] = 0;
2940                 } else {
2941                         raw_inode->i_block[0] = 0;
2942                         raw_inode->i_block[1] =
2943                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2944                         raw_inode->i_block[2] = 0;
2945                 }
2946         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2947                 raw_inode->i_block[block] = ei->i_data[block];
2948
2949         if (ei->i_extra_isize)
2950                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2951
2952         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2953         rc = ext3_journal_dirty_metadata(handle, bh);
2954         if (!err)
2955                 err = rc;
2956         ei->i_state &= ~EXT3_STATE_NEW;
2957
2958 out_brelse:
2959         brelse (bh);
2960         ext3_std_error(inode->i_sb, err);
2961         return err;
2962 }
2963
2964 /*
2965  * ext3_write_inode()
2966  *
2967  * We are called from a few places:
2968  *
2969  * - Within generic_file_write() for O_SYNC files.
2970  *   Here, there will be no transaction running. We wait for any running
2971  *   trasnaction to commit.
2972  *
2973  * - Within sys_sync(), kupdate and such.
2974  *   We wait on commit, if tol to.
2975  *
2976  * - Within prune_icache() (PF_MEMALLOC == true)
2977  *   Here we simply return.  We can't afford to block kswapd on the
2978  *   journal commit.
2979  *
2980  * In all cases it is actually safe for us to return without doing anything,
2981  * because the inode has been copied into a raw inode buffer in
2982  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2983  * knfsd.
2984  *
2985  * Note that we are absolutely dependent upon all inode dirtiers doing the
2986  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2987  * which we are interested.
2988  *
2989  * It would be a bug for them to not do this.  The code:
2990  *
2991  *      mark_inode_dirty(inode)
2992  *      stuff();
2993  *      inode->i_size = expr;
2994  *
2995  * is in error because a kswapd-driven write_inode() could occur while
2996  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2997  * will no longer be on the superblock's dirty inode list.
2998  */
2999 int ext3_write_inode(struct inode *inode, int wait)
3000 {
3001         if (current->flags & PF_MEMALLOC)
3002                 return 0;
3003
3004         if (ext3_journal_current_handle()) {
3005                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3006                 dump_stack();
3007                 return -EIO;
3008         }
3009
3010         if (!wait)
3011                 return 0;
3012
3013         return ext3_force_commit(inode->i_sb);
3014 }
3015
3016 /*
3017  * ext3_setattr()
3018  *
3019  * Called from notify_change.
3020  *
3021  * We want to trap VFS attempts to truncate the file as soon as
3022  * possible.  In particular, we want to make sure that when the VFS
3023  * shrinks i_size, we put the inode on the orphan list and modify
3024  * i_disksize immediately, so that during the subsequent flushing of
3025  * dirty pages and freeing of disk blocks, we can guarantee that any
3026  * commit will leave the blocks being flushed in an unused state on
3027  * disk.  (On recovery, the inode will get truncated and the blocks will
3028  * be freed, so we have a strong guarantee that no future commit will
3029  * leave these blocks visible to the user.)
3030  *
3031  * Called with inode->sem down.
3032  */
3033 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3034 {
3035         struct inode *inode = dentry->d_inode;
3036         int error, rc = 0;
3037         const unsigned int ia_valid = attr->ia_valid;
3038
3039         error = inode_change_ok(inode, attr);
3040         if (error)
3041                 return error;
3042
3043         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3044                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3045                 handle_t *handle;
3046
3047                 /* (user+group)*(old+new) structure, inode write (sb,
3048                  * inode block, ? - but truncate inode update has it) */
3049                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3050                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3051                 if (IS_ERR(handle)) {
3052                         error = PTR_ERR(handle);
3053                         goto err_out;
3054                 }
3055                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
3056                 if (error) {
3057                         ext3_journal_stop(handle);
3058                         return error;
3059                 }
3060                 /* Update corresponding info in inode so that everything is in
3061                  * one transaction */
3062                 if (attr->ia_valid & ATTR_UID)
3063                         inode->i_uid = attr->ia_uid;
3064                 if (attr->ia_valid & ATTR_GID)
3065                         inode->i_gid = attr->ia_gid;
3066                 error = ext3_mark_inode_dirty(handle, inode);
3067                 ext3_journal_stop(handle);
3068         }
3069
3070         if (S_ISREG(inode->i_mode) &&
3071             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3072                 handle_t *handle;
3073
3074                 handle = ext3_journal_start(inode, 3);
3075                 if (IS_ERR(handle)) {
3076                         error = PTR_ERR(handle);
3077                         goto err_out;
3078                 }
3079
3080                 error = ext3_orphan_add(handle, inode);
3081                 EXT3_I(inode)->i_disksize = attr->ia_size;
3082                 rc = ext3_mark_inode_dirty(handle, inode);
3083                 if (!error)
3084                         error = rc;
3085                 ext3_journal_stop(handle);
3086         }
3087
3088         rc = inode_setattr(inode, attr);
3089
3090         /* If inode_setattr's call to ext3_truncate failed to get a
3091          * transaction handle at all, we need to clean up the in-core
3092          * orphan list manually. */
3093         if (inode->i_nlink)
3094                 ext3_orphan_del(NULL, inode);
3095
3096         if (!rc && (ia_valid & ATTR_MODE))
3097                 rc = ext3_acl_chmod(inode);
3098
3099 err_out:
3100         ext3_std_error(inode->i_sb, error);
3101         if (!error)
3102                 error = rc;
3103         return error;
3104 }
3105
3106
3107 /*
3108  * How many blocks doth make a writepage()?
3109  *
3110  * With N blocks per page, it may be:
3111  * N data blocks
3112  * 2 indirect block
3113  * 2 dindirect
3114  * 1 tindirect
3115  * N+5 bitmap blocks (from the above)
3116  * N+5 group descriptor summary blocks
3117  * 1 inode block
3118  * 1 superblock.
3119  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3120  *
3121  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3122  *
3123  * With ordered or writeback data it's the same, less the N data blocks.
3124  *
3125  * If the inode's direct blocks can hold an integral number of pages then a
3126  * page cannot straddle two indirect blocks, and we can only touch one indirect
3127  * and dindirect block, and the "5" above becomes "3".
3128  *
3129  * This still overestimates under most circumstances.  If we were to pass the
3130  * start and end offsets in here as well we could do block_to_path() on each
3131  * block and work out the exact number of indirects which are touched.  Pah.
3132  */
3133
3134 static int ext3_writepage_trans_blocks(struct inode *inode)
3135 {
3136         int bpp = ext3_journal_blocks_per_page(inode);
3137         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3138         int ret;
3139
3140         if (ext3_should_journal_data(inode))
3141                 ret = 3 * (bpp + indirects) + 2;
3142         else
3143                 ret = 2 * (bpp + indirects) + 2;
3144
3145 #ifdef CONFIG_QUOTA
3146         /* We know that structure was already allocated during DQUOT_INIT so
3147          * we will be updating only the data blocks + inodes */
3148         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3149 #endif
3150
3151         return ret;
3152 }
3153
3154 /*
3155  * The caller must have previously called ext3_reserve_inode_write().
3156  * Give this, we know that the caller already has write access to iloc->bh.
3157  */
3158 int ext3_mark_iloc_dirty(handle_t *handle,
3159                 struct inode *inode, struct ext3_iloc *iloc)
3160 {
3161         int err = 0;
3162
3163         /* the do_update_inode consumes one bh->b_count */
3164         get_bh(iloc->bh);
3165
3166         /* ext3_do_update_inode() does journal_dirty_metadata */
3167         err = ext3_do_update_inode(handle, inode, iloc);
3168         put_bh(iloc->bh);
3169         return err;
3170 }
3171
3172 /*
3173  * On success, We end up with an outstanding reference count against
3174  * iloc->bh.  This _must_ be cleaned up later.
3175  */
3176
3177 int
3178 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3179                          struct ext3_iloc *iloc)
3180 {
3181         int err = 0;
3182         if (handle) {
3183                 err = ext3_get_inode_loc(inode, iloc);
3184                 if (!err) {
3185                         BUFFER_TRACE(iloc->bh, "get_write_access");
3186                         err = ext3_journal_get_write_access(handle, iloc->bh);
3187                         if (err) {
3188                                 brelse(iloc->bh);
3189                                 iloc->bh = NULL;
3190                         }
3191                 }
3192         }
3193         ext3_std_error(inode->i_sb, err);
3194         return err;
3195 }
3196
3197 /*
3198  * What we do here is to mark the in-core inode as clean with respect to inode
3199  * dirtiness (it may still be data-dirty).
3200  * This means that the in-core inode may be reaped by prune_icache
3201  * without having to perform any I/O.  This is a very good thing,
3202  * because *any* task may call prune_icache - even ones which
3203  * have a transaction open against a different journal.
3204  *
3205  * Is this cheating?  Not really.  Sure, we haven't written the
3206  * inode out, but prune_icache isn't a user-visible syncing function.
3207  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3208  * we start and wait on commits.
3209  *
3210  * Is this efficient/effective?  Well, we're being nice to the system
3211  * by cleaning up our inodes proactively so they can be reaped
3212  * without I/O.  But we are potentially leaving up to five seconds'
3213  * worth of inodes floating about which prune_icache wants us to
3214  * write out.  One way to fix that would be to get prune_icache()
3215  * to do a write_super() to free up some memory.  It has the desired
3216  * effect.
3217  */
3218 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3219 {
3220         struct ext3_iloc iloc;
3221         int err;
3222
3223         might_sleep();
3224         err = ext3_reserve_inode_write(handle, inode, &iloc);
3225         if (!err)
3226                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3227         return err;
3228 }
3229
3230 /*
3231  * ext3_dirty_inode() is called from __mark_inode_dirty()
3232  *
3233  * We're really interested in the case where a file is being extended.
3234  * i_size has been changed by generic_commit_write() and we thus need
3235  * to include the updated inode in the current transaction.
3236  *
3237  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3238  * are allocated to the file.
3239  *
3240  * If the inode is marked synchronous, we don't honour that here - doing
3241  * so would cause a commit on atime updates, which we don't bother doing.
3242  * We handle synchronous inodes at the highest possible level.
3243  */
3244 void ext3_dirty_inode(struct inode *inode)
3245 {
3246         handle_t *current_handle = ext3_journal_current_handle();
3247         handle_t *handle;
3248
3249         handle = ext3_journal_start(inode, 2);
3250         if (IS_ERR(handle))
3251                 goto out;
3252         if (current_handle &&
3253                 current_handle->h_transaction != handle->h_transaction) {
3254                 /* This task has a transaction open against a different fs */
3255                 printk(KERN_EMERG "%s: transactions do not match!\n",
3256                        __func__);
3257         } else {
3258                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3259                                 current_handle);
3260                 ext3_mark_inode_dirty(handle, inode);
3261         }
3262         ext3_journal_stop(handle);
3263 out:
3264         return;
3265 }
3266
3267 #if 0
3268 /*
3269  * Bind an inode's backing buffer_head into this transaction, to prevent
3270  * it from being flushed to disk early.  Unlike
3271  * ext3_reserve_inode_write, this leaves behind no bh reference and
3272  * returns no iloc structure, so the caller needs to repeat the iloc
3273  * lookup to mark the inode dirty later.
3274  */
3275 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3276 {
3277         struct ext3_iloc iloc;
3278
3279         int err = 0;
3280         if (handle) {
3281                 err = ext3_get_inode_loc(inode, &iloc);
3282                 if (!err) {
3283                         BUFFER_TRACE(iloc.bh, "get_write_access");
3284                         err = journal_get_write_access(handle, iloc.bh);
3285                         if (!err)
3286                                 err = ext3_journal_dirty_metadata(handle,
3287                                                                   iloc.bh);
3288                         brelse(iloc.bh);
3289                 }
3290         }
3291         ext3_std_error(inode->i_sb, err);
3292         return err;
3293 }
3294 #endif
3295
3296 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3297 {
3298         journal_t *journal;
3299         handle_t *handle;
3300         int err;
3301
3302         /*
3303          * We have to be very careful here: changing a data block's
3304          * journaling status dynamically is dangerous.  If we write a
3305          * data block to the journal, change the status and then delete
3306          * that block, we risk forgetting to revoke the old log record
3307          * from the journal and so a subsequent replay can corrupt data.
3308          * So, first we make sure that the journal is empty and that
3309          * nobody is changing anything.
3310          */
3311
3312         journal = EXT3_JOURNAL(inode);
3313         if (is_journal_aborted(journal))
3314                 return -EROFS;
3315
3316         journal_lock_updates(journal);
3317         journal_flush(journal);
3318
3319         /*
3320          * OK, there are no updates running now, and all cached data is
3321          * synced to disk.  We are now in a completely consistent state
3322          * which doesn't have anything in the journal, and we know that
3323          * no filesystem updates are running, so it is safe to modify
3324          * the inode's in-core data-journaling state flag now.
3325          */
3326
3327         if (val)
3328                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3329         else
3330                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3331         ext3_set_aops(inode);
3332
3333         journal_unlock_updates(journal);
3334
3335         /* Finally we can mark the inode as dirty. */
3336
3337         handle = ext3_journal_start(inode, 1);
3338         if (IS_ERR(handle))
3339                 return PTR_ERR(handle);
3340
3341         err = ext3_mark_inode_dirty(handle, inode);
3342         handle->h_sync = 1;
3343         ext3_journal_stop(handle);
3344         ext3_std_error(inode->i_sb, err);
3345
3346         return err;
3347 }