Merge branch 'master' of /home/davem/src/GIT/linux-2.6/
[linux-2.6] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include <linux/fiemap.h>
40 #include <linux/namei.h>
41 #include "xattr.h"
42 #include "acl.h"
43
44 static int ext3_writepage_trans_blocks(struct inode *inode);
45
46 /*
47  * Test whether an inode is a fast symlink.
48  */
49 static int ext3_inode_is_fast_symlink(struct inode *inode)
50 {
51         int ea_blocks = EXT3_I(inode)->i_file_acl ?
52                 (inode->i_sb->s_blocksize >> 9) : 0;
53
54         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
55 }
56
57 /*
58  * The ext3 forget function must perform a revoke if we are freeing data
59  * which has been journaled.  Metadata (eg. indirect blocks) must be
60  * revoked in all cases.
61  *
62  * "bh" may be NULL: a metadata block may have been freed from memory
63  * but there may still be a record of it in the journal, and that record
64  * still needs to be revoked.
65  */
66 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
67                         struct buffer_head *bh, ext3_fsblk_t blocknr)
68 {
69         int err;
70
71         might_sleep();
72
73         BUFFER_TRACE(bh, "enter");
74
75         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
76                   "data mode %lx\n",
77                   bh, is_metadata, inode->i_mode,
78                   test_opt(inode->i_sb, DATA_FLAGS));
79
80         /* Never use the revoke function if we are doing full data
81          * journaling: there is no need to, and a V1 superblock won't
82          * support it.  Otherwise, only skip the revoke on un-journaled
83          * data blocks. */
84
85         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
86             (!is_metadata && !ext3_should_journal_data(inode))) {
87                 if (bh) {
88                         BUFFER_TRACE(bh, "call journal_forget");
89                         return ext3_journal_forget(handle, bh);
90                 }
91                 return 0;
92         }
93
94         /*
95          * data!=journal && (is_metadata || should_journal_data(inode))
96          */
97         BUFFER_TRACE(bh, "call ext3_journal_revoke");
98         err = ext3_journal_revoke(handle, blocknr, bh);
99         if (err)
100                 ext3_abort(inode->i_sb, __func__,
101                            "error %d when attempting revoke", err);
102         BUFFER_TRACE(bh, "exit");
103         return err;
104 }
105
106 /*
107  * Work out how many blocks we need to proceed with the next chunk of a
108  * truncate transaction.
109  */
110 static unsigned long blocks_for_truncate(struct inode *inode)
111 {
112         unsigned long needed;
113
114         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
115
116         /* Give ourselves just enough room to cope with inodes in which
117          * i_blocks is corrupt: we've seen disk corruptions in the past
118          * which resulted in random data in an inode which looked enough
119          * like a regular file for ext3 to try to delete it.  Things
120          * will go a bit crazy if that happens, but at least we should
121          * try not to panic the whole kernel. */
122         if (needed < 2)
123                 needed = 2;
124
125         /* But we need to bound the transaction so we don't overflow the
126          * journal. */
127         if (needed > EXT3_MAX_TRANS_DATA)
128                 needed = EXT3_MAX_TRANS_DATA;
129
130         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
131 }
132
133 /*
134  * Truncate transactions can be complex and absolutely huge.  So we need to
135  * be able to restart the transaction at a conventient checkpoint to make
136  * sure we don't overflow the journal.
137  *
138  * start_transaction gets us a new handle for a truncate transaction,
139  * and extend_transaction tries to extend the existing one a bit.  If
140  * extend fails, we need to propagate the failure up and restart the
141  * transaction in the top-level truncate loop. --sct
142  */
143 static handle_t *start_transaction(struct inode *inode)
144 {
145         handle_t *result;
146
147         result = ext3_journal_start(inode, blocks_for_truncate(inode));
148         if (!IS_ERR(result))
149                 return result;
150
151         ext3_std_error(inode->i_sb, PTR_ERR(result));
152         return result;
153 }
154
155 /*
156  * Try to extend this transaction for the purposes of truncation.
157  *
158  * Returns 0 if we managed to create more room.  If we can't create more
159  * room, and the transaction must be restarted we return 1.
160  */
161 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
162 {
163         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
164                 return 0;
165         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
166                 return 0;
167         return 1;
168 }
169
170 /*
171  * Restart the transaction associated with *handle.  This does a commit,
172  * so before we call here everything must be consistently dirtied against
173  * this transaction.
174  */
175 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
176 {
177         jbd_debug(2, "restarting handle %p\n", handle);
178         return ext3_journal_restart(handle, blocks_for_truncate(inode));
179 }
180
181 /*
182  * Called at the last iput() if i_nlink is zero.
183  */
184 void ext3_delete_inode (struct inode * inode)
185 {
186         handle_t *handle;
187
188         truncate_inode_pages(&inode->i_data, 0);
189
190         if (is_bad_inode(inode))
191                 goto no_delete;
192
193         handle = start_transaction(inode);
194         if (IS_ERR(handle)) {
195                 /*
196                  * If we're going to skip the normal cleanup, we still need to
197                  * make sure that the in-core orphan linked list is properly
198                  * cleaned up.
199                  */
200                 ext3_orphan_del(NULL, inode);
201                 goto no_delete;
202         }
203
204         if (IS_SYNC(inode))
205                 handle->h_sync = 1;
206         inode->i_size = 0;
207         if (inode->i_blocks)
208                 ext3_truncate(inode);
209         /*
210          * Kill off the orphan record which ext3_truncate created.
211          * AKPM: I think this can be inside the above `if'.
212          * Note that ext3_orphan_del() has to be able to cope with the
213          * deletion of a non-existent orphan - this is because we don't
214          * know if ext3_truncate() actually created an orphan record.
215          * (Well, we could do this if we need to, but heck - it works)
216          */
217         ext3_orphan_del(handle, inode);
218         EXT3_I(inode)->i_dtime  = get_seconds();
219
220         /*
221          * One subtle ordering requirement: if anything has gone wrong
222          * (transaction abort, IO errors, whatever), then we can still
223          * do these next steps (the fs will already have been marked as
224          * having errors), but we can't free the inode if the mark_dirty
225          * fails.
226          */
227         if (ext3_mark_inode_dirty(handle, inode))
228                 /* If that failed, just do the required in-core inode clear. */
229                 clear_inode(inode);
230         else
231                 ext3_free_inode(handle, inode);
232         ext3_journal_stop(handle);
233         return;
234 no_delete:
235         clear_inode(inode);     /* We must guarantee clearing of inode... */
236 }
237
238 typedef struct {
239         __le32  *p;
240         __le32  key;
241         struct buffer_head *bh;
242 } Indirect;
243
244 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
245 {
246         p->key = *(p->p = v);
247         p->bh = bh;
248 }
249
250 static int verify_chain(Indirect *from, Indirect *to)
251 {
252         while (from <= to && from->key == *from->p)
253                 from++;
254         return (from > to);
255 }
256
257 /**
258  *      ext3_block_to_path - parse the block number into array of offsets
259  *      @inode: inode in question (we are only interested in its superblock)
260  *      @i_block: block number to be parsed
261  *      @offsets: array to store the offsets in
262  *      @boundary: set this non-zero if the referred-to block is likely to be
263  *             followed (on disk) by an indirect block.
264  *
265  *      To store the locations of file's data ext3 uses a data structure common
266  *      for UNIX filesystems - tree of pointers anchored in the inode, with
267  *      data blocks at leaves and indirect blocks in intermediate nodes.
268  *      This function translates the block number into path in that tree -
269  *      return value is the path length and @offsets[n] is the offset of
270  *      pointer to (n+1)th node in the nth one. If @block is out of range
271  *      (negative or too large) warning is printed and zero returned.
272  *
273  *      Note: function doesn't find node addresses, so no IO is needed. All
274  *      we need to know is the capacity of indirect blocks (taken from the
275  *      inode->i_sb).
276  */
277
278 /*
279  * Portability note: the last comparison (check that we fit into triple
280  * indirect block) is spelled differently, because otherwise on an
281  * architecture with 32-bit longs and 8Kb pages we might get into trouble
282  * if our filesystem had 8Kb blocks. We might use long long, but that would
283  * kill us on x86. Oh, well, at least the sign propagation does not matter -
284  * i_block would have to be negative in the very beginning, so we would not
285  * get there at all.
286  */
287
288 static int ext3_block_to_path(struct inode *inode,
289                         long i_block, int offsets[4], int *boundary)
290 {
291         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
292         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
293         const long direct_blocks = EXT3_NDIR_BLOCKS,
294                 indirect_blocks = ptrs,
295                 double_blocks = (1 << (ptrs_bits * 2));
296         int n = 0;
297         int final = 0;
298
299         if (i_block < 0) {
300                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
301         } else if (i_block < direct_blocks) {
302                 offsets[n++] = i_block;
303                 final = direct_blocks;
304         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
305                 offsets[n++] = EXT3_IND_BLOCK;
306                 offsets[n++] = i_block;
307                 final = ptrs;
308         } else if ((i_block -= indirect_blocks) < double_blocks) {
309                 offsets[n++] = EXT3_DIND_BLOCK;
310                 offsets[n++] = i_block >> ptrs_bits;
311                 offsets[n++] = i_block & (ptrs - 1);
312                 final = ptrs;
313         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
314                 offsets[n++] = EXT3_TIND_BLOCK;
315                 offsets[n++] = i_block >> (ptrs_bits * 2);
316                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
317                 offsets[n++] = i_block & (ptrs - 1);
318                 final = ptrs;
319         } else {
320                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
321         }
322         if (boundary)
323                 *boundary = final - 1 - (i_block & (ptrs - 1));
324         return n;
325 }
326
327 /**
328  *      ext3_get_branch - read the chain of indirect blocks leading to data
329  *      @inode: inode in question
330  *      @depth: depth of the chain (1 - direct pointer, etc.)
331  *      @offsets: offsets of pointers in inode/indirect blocks
332  *      @chain: place to store the result
333  *      @err: here we store the error value
334  *
335  *      Function fills the array of triples <key, p, bh> and returns %NULL
336  *      if everything went OK or the pointer to the last filled triple
337  *      (incomplete one) otherwise. Upon the return chain[i].key contains
338  *      the number of (i+1)-th block in the chain (as it is stored in memory,
339  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
340  *      number (it points into struct inode for i==0 and into the bh->b_data
341  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
342  *      block for i>0 and NULL for i==0. In other words, it holds the block
343  *      numbers of the chain, addresses they were taken from (and where we can
344  *      verify that chain did not change) and buffer_heads hosting these
345  *      numbers.
346  *
347  *      Function stops when it stumbles upon zero pointer (absent block)
348  *              (pointer to last triple returned, *@err == 0)
349  *      or when it gets an IO error reading an indirect block
350  *              (ditto, *@err == -EIO)
351  *      or when it notices that chain had been changed while it was reading
352  *              (ditto, *@err == -EAGAIN)
353  *      or when it reads all @depth-1 indirect blocks successfully and finds
354  *      the whole chain, all way to the data (returns %NULL, *err == 0).
355  */
356 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
357                                  Indirect chain[4], int *err)
358 {
359         struct super_block *sb = inode->i_sb;
360         Indirect *p = chain;
361         struct buffer_head *bh;
362
363         *err = 0;
364         /* i_data is not going away, no lock needed */
365         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
366         if (!p->key)
367                 goto no_block;
368         while (--depth) {
369                 bh = sb_bread(sb, le32_to_cpu(p->key));
370                 if (!bh)
371                         goto failure;
372                 /* Reader: pointers */
373                 if (!verify_chain(chain, p))
374                         goto changed;
375                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
376                 /* Reader: end */
377                 if (!p->key)
378                         goto no_block;
379         }
380         return NULL;
381
382 changed:
383         brelse(bh);
384         *err = -EAGAIN;
385         goto no_block;
386 failure:
387         *err = -EIO;
388 no_block:
389         return p;
390 }
391
392 /**
393  *      ext3_find_near - find a place for allocation with sufficient locality
394  *      @inode: owner
395  *      @ind: descriptor of indirect block.
396  *
397  *      This function returns the preferred place for block allocation.
398  *      It is used when heuristic for sequential allocation fails.
399  *      Rules are:
400  *        + if there is a block to the left of our position - allocate near it.
401  *        + if pointer will live in indirect block - allocate near that block.
402  *        + if pointer will live in inode - allocate in the same
403  *          cylinder group.
404  *
405  * In the latter case we colour the starting block by the callers PID to
406  * prevent it from clashing with concurrent allocations for a different inode
407  * in the same block group.   The PID is used here so that functionally related
408  * files will be close-by on-disk.
409  *
410  *      Caller must make sure that @ind is valid and will stay that way.
411  */
412 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
413 {
414         struct ext3_inode_info *ei = EXT3_I(inode);
415         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
416         __le32 *p;
417         ext3_fsblk_t bg_start;
418         ext3_grpblk_t colour;
419
420         /* Try to find previous block */
421         for (p = ind->p - 1; p >= start; p--) {
422                 if (*p)
423                         return le32_to_cpu(*p);
424         }
425
426         /* No such thing, so let's try location of indirect block */
427         if (ind->bh)
428                 return ind->bh->b_blocknr;
429
430         /*
431          * It is going to be referred to from the inode itself? OK, just put it
432          * into the same cylinder group then.
433          */
434         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
435         colour = (current->pid % 16) *
436                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
437         return bg_start + colour;
438 }
439
440 /**
441  *      ext3_find_goal - find a preferred place for allocation.
442  *      @inode: owner
443  *      @block:  block we want
444  *      @partial: pointer to the last triple within a chain
445  *
446  *      Normally this function find the preferred place for block allocation,
447  *      returns it.
448  */
449
450 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
451                                    Indirect *partial)
452 {
453         struct ext3_block_alloc_info *block_i;
454
455         block_i =  EXT3_I(inode)->i_block_alloc_info;
456
457         /*
458          * try the heuristic for sequential allocation,
459          * failing that at least try to get decent locality.
460          */
461         if (block_i && (block == block_i->last_alloc_logical_block + 1)
462                 && (block_i->last_alloc_physical_block != 0)) {
463                 return block_i->last_alloc_physical_block + 1;
464         }
465
466         return ext3_find_near(inode, partial);
467 }
468
469 /**
470  *      ext3_blks_to_allocate: Look up the block map and count the number
471  *      of direct blocks need to be allocated for the given branch.
472  *
473  *      @branch: chain of indirect blocks
474  *      @k: number of blocks need for indirect blocks
475  *      @blks: number of data blocks to be mapped.
476  *      @blocks_to_boundary:  the offset in the indirect block
477  *
478  *      return the total number of blocks to be allocate, including the
479  *      direct and indirect blocks.
480  */
481 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
482                 int blocks_to_boundary)
483 {
484         unsigned long count = 0;
485
486         /*
487          * Simple case, [t,d]Indirect block(s) has not allocated yet
488          * then it's clear blocks on that path have not allocated
489          */
490         if (k > 0) {
491                 /* right now we don't handle cross boundary allocation */
492                 if (blks < blocks_to_boundary + 1)
493                         count += blks;
494                 else
495                         count += blocks_to_boundary + 1;
496                 return count;
497         }
498
499         count++;
500         while (count < blks && count <= blocks_to_boundary &&
501                 le32_to_cpu(*(branch[0].p + count)) == 0) {
502                 count++;
503         }
504         return count;
505 }
506
507 /**
508  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
509  *      @indirect_blks: the number of blocks need to allocate for indirect
510  *                      blocks
511  *
512  *      @new_blocks: on return it will store the new block numbers for
513  *      the indirect blocks(if needed) and the first direct block,
514  *      @blks:  on return it will store the total number of allocated
515  *              direct blocks
516  */
517 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
518                         ext3_fsblk_t goal, int indirect_blks, int blks,
519                         ext3_fsblk_t new_blocks[4], int *err)
520 {
521         int target, i;
522         unsigned long count = 0;
523         int index = 0;
524         ext3_fsblk_t current_block = 0;
525         int ret = 0;
526
527         /*
528          * Here we try to allocate the requested multiple blocks at once,
529          * on a best-effort basis.
530          * To build a branch, we should allocate blocks for
531          * the indirect blocks(if not allocated yet), and at least
532          * the first direct block of this branch.  That's the
533          * minimum number of blocks need to allocate(required)
534          */
535         target = blks + indirect_blks;
536
537         while (1) {
538                 count = target;
539                 /* allocating blocks for indirect blocks and direct blocks */
540                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
541                 if (*err)
542                         goto failed_out;
543
544                 target -= count;
545                 /* allocate blocks for indirect blocks */
546                 while (index < indirect_blks && count) {
547                         new_blocks[index++] = current_block++;
548                         count--;
549                 }
550
551                 if (count > 0)
552                         break;
553         }
554
555         /* save the new block number for the first direct block */
556         new_blocks[index] = current_block;
557
558         /* total number of blocks allocated for direct blocks */
559         ret = count;
560         *err = 0;
561         return ret;
562 failed_out:
563         for (i = 0; i <index; i++)
564                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
565         return ret;
566 }
567
568 /**
569  *      ext3_alloc_branch - allocate and set up a chain of blocks.
570  *      @inode: owner
571  *      @indirect_blks: number of allocated indirect blocks
572  *      @blks: number of allocated direct blocks
573  *      @offsets: offsets (in the blocks) to store the pointers to next.
574  *      @branch: place to store the chain in.
575  *
576  *      This function allocates blocks, zeroes out all but the last one,
577  *      links them into chain and (if we are synchronous) writes them to disk.
578  *      In other words, it prepares a branch that can be spliced onto the
579  *      inode. It stores the information about that chain in the branch[], in
580  *      the same format as ext3_get_branch() would do. We are calling it after
581  *      we had read the existing part of chain and partial points to the last
582  *      triple of that (one with zero ->key). Upon the exit we have the same
583  *      picture as after the successful ext3_get_block(), except that in one
584  *      place chain is disconnected - *branch->p is still zero (we did not
585  *      set the last link), but branch->key contains the number that should
586  *      be placed into *branch->p to fill that gap.
587  *
588  *      If allocation fails we free all blocks we've allocated (and forget
589  *      their buffer_heads) and return the error value the from failed
590  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
591  *      as described above and return 0.
592  */
593 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
594                         int indirect_blks, int *blks, ext3_fsblk_t goal,
595                         int *offsets, Indirect *branch)
596 {
597         int blocksize = inode->i_sb->s_blocksize;
598         int i, n = 0;
599         int err = 0;
600         struct buffer_head *bh;
601         int num;
602         ext3_fsblk_t new_blocks[4];
603         ext3_fsblk_t current_block;
604
605         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
606                                 *blks, new_blocks, &err);
607         if (err)
608                 return err;
609
610         branch[0].key = cpu_to_le32(new_blocks[0]);
611         /*
612          * metadata blocks and data blocks are allocated.
613          */
614         for (n = 1; n <= indirect_blks;  n++) {
615                 /*
616                  * Get buffer_head for parent block, zero it out
617                  * and set the pointer to new one, then send
618                  * parent to disk.
619                  */
620                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
621                 branch[n].bh = bh;
622                 lock_buffer(bh);
623                 BUFFER_TRACE(bh, "call get_create_access");
624                 err = ext3_journal_get_create_access(handle, bh);
625                 if (err) {
626                         unlock_buffer(bh);
627                         brelse(bh);
628                         goto failed;
629                 }
630
631                 memset(bh->b_data, 0, blocksize);
632                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
633                 branch[n].key = cpu_to_le32(new_blocks[n]);
634                 *branch[n].p = branch[n].key;
635                 if ( n == indirect_blks) {
636                         current_block = new_blocks[n];
637                         /*
638                          * End of chain, update the last new metablock of
639                          * the chain to point to the new allocated
640                          * data blocks numbers
641                          */
642                         for (i=1; i < num; i++)
643                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
644                 }
645                 BUFFER_TRACE(bh, "marking uptodate");
646                 set_buffer_uptodate(bh);
647                 unlock_buffer(bh);
648
649                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
650                 err = ext3_journal_dirty_metadata(handle, bh);
651                 if (err)
652                         goto failed;
653         }
654         *blks = num;
655         return err;
656 failed:
657         /* Allocation failed, free what we already allocated */
658         for (i = 1; i <= n ; i++) {
659                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
660                 ext3_journal_forget(handle, branch[i].bh);
661         }
662         for (i = 0; i <indirect_blks; i++)
663                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
664
665         ext3_free_blocks(handle, inode, new_blocks[i], num);
666
667         return err;
668 }
669
670 /**
671  * ext3_splice_branch - splice the allocated branch onto inode.
672  * @inode: owner
673  * @block: (logical) number of block we are adding
674  * @chain: chain of indirect blocks (with a missing link - see
675  *      ext3_alloc_branch)
676  * @where: location of missing link
677  * @num:   number of indirect blocks we are adding
678  * @blks:  number of direct blocks we are adding
679  *
680  * This function fills the missing link and does all housekeeping needed in
681  * inode (->i_blocks, etc.). In case of success we end up with the full
682  * chain to new block and return 0.
683  */
684 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
685                         long block, Indirect *where, int num, int blks)
686 {
687         int i;
688         int err = 0;
689         struct ext3_block_alloc_info *block_i;
690         ext3_fsblk_t current_block;
691
692         block_i = EXT3_I(inode)->i_block_alloc_info;
693         /*
694          * If we're splicing into a [td]indirect block (as opposed to the
695          * inode) then we need to get write access to the [td]indirect block
696          * before the splice.
697          */
698         if (where->bh) {
699                 BUFFER_TRACE(where->bh, "get_write_access");
700                 err = ext3_journal_get_write_access(handle, where->bh);
701                 if (err)
702                         goto err_out;
703         }
704         /* That's it */
705
706         *where->p = where->key;
707
708         /*
709          * Update the host buffer_head or inode to point to more just allocated
710          * direct blocks blocks
711          */
712         if (num == 0 && blks > 1) {
713                 current_block = le32_to_cpu(where->key) + 1;
714                 for (i = 1; i < blks; i++)
715                         *(where->p + i ) = cpu_to_le32(current_block++);
716         }
717
718         /*
719          * update the most recently allocated logical & physical block
720          * in i_block_alloc_info, to assist find the proper goal block for next
721          * allocation
722          */
723         if (block_i) {
724                 block_i->last_alloc_logical_block = block + blks - 1;
725                 block_i->last_alloc_physical_block =
726                                 le32_to_cpu(where[num].key) + blks - 1;
727         }
728
729         /* We are done with atomic stuff, now do the rest of housekeeping */
730
731         inode->i_ctime = CURRENT_TIME_SEC;
732         ext3_mark_inode_dirty(handle, inode);
733
734         /* had we spliced it onto indirect block? */
735         if (where->bh) {
736                 /*
737                  * If we spliced it onto an indirect block, we haven't
738                  * altered the inode.  Note however that if it is being spliced
739                  * onto an indirect block at the very end of the file (the
740                  * file is growing) then we *will* alter the inode to reflect
741                  * the new i_size.  But that is not done here - it is done in
742                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
743                  */
744                 jbd_debug(5, "splicing indirect only\n");
745                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
746                 err = ext3_journal_dirty_metadata(handle, where->bh);
747                 if (err)
748                         goto err_out;
749         } else {
750                 /*
751                  * OK, we spliced it into the inode itself on a direct block.
752                  * Inode was dirtied above.
753                  */
754                 jbd_debug(5, "splicing direct\n");
755         }
756         return err;
757
758 err_out:
759         for (i = 1; i <= num; i++) {
760                 BUFFER_TRACE(where[i].bh, "call journal_forget");
761                 ext3_journal_forget(handle, where[i].bh);
762                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
763         }
764         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
765
766         return err;
767 }
768
769 /*
770  * Allocation strategy is simple: if we have to allocate something, we will
771  * have to go the whole way to leaf. So let's do it before attaching anything
772  * to tree, set linkage between the newborn blocks, write them if sync is
773  * required, recheck the path, free and repeat if check fails, otherwise
774  * set the last missing link (that will protect us from any truncate-generated
775  * removals - all blocks on the path are immune now) and possibly force the
776  * write on the parent block.
777  * That has a nice additional property: no special recovery from the failed
778  * allocations is needed - we simply release blocks and do not touch anything
779  * reachable from inode.
780  *
781  * `handle' can be NULL if create == 0.
782  *
783  * The BKL may not be held on entry here.  Be sure to take it early.
784  * return > 0, # of blocks mapped or allocated.
785  * return = 0, if plain lookup failed.
786  * return < 0, error case.
787  */
788 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
789                 sector_t iblock, unsigned long maxblocks,
790                 struct buffer_head *bh_result,
791                 int create, int extend_disksize)
792 {
793         int err = -EIO;
794         int offsets[4];
795         Indirect chain[4];
796         Indirect *partial;
797         ext3_fsblk_t goal;
798         int indirect_blks;
799         int blocks_to_boundary = 0;
800         int depth;
801         struct ext3_inode_info *ei = EXT3_I(inode);
802         int count = 0;
803         ext3_fsblk_t first_block = 0;
804
805
806         J_ASSERT(handle != NULL || create == 0);
807         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
808
809         if (depth == 0)
810                 goto out;
811
812         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
813
814         /* Simplest case - block found, no allocation needed */
815         if (!partial) {
816                 first_block = le32_to_cpu(chain[depth - 1].key);
817                 clear_buffer_new(bh_result);
818                 count++;
819                 /*map more blocks*/
820                 while (count < maxblocks && count <= blocks_to_boundary) {
821                         ext3_fsblk_t blk;
822
823                         if (!verify_chain(chain, partial)) {
824                                 /*
825                                  * Indirect block might be removed by
826                                  * truncate while we were reading it.
827                                  * Handling of that case: forget what we've
828                                  * got now. Flag the err as EAGAIN, so it
829                                  * will reread.
830                                  */
831                                 err = -EAGAIN;
832                                 count = 0;
833                                 break;
834                         }
835                         blk = le32_to_cpu(*(chain[depth-1].p + count));
836
837                         if (blk == first_block + count)
838                                 count++;
839                         else
840                                 break;
841                 }
842                 if (err != -EAGAIN)
843                         goto got_it;
844         }
845
846         /* Next simple case - plain lookup or failed read of indirect block */
847         if (!create || err == -EIO)
848                 goto cleanup;
849
850         mutex_lock(&ei->truncate_mutex);
851
852         /*
853          * If the indirect block is missing while we are reading
854          * the chain(ext3_get_branch() returns -EAGAIN err), or
855          * if the chain has been changed after we grab the semaphore,
856          * (either because another process truncated this branch, or
857          * another get_block allocated this branch) re-grab the chain to see if
858          * the request block has been allocated or not.
859          *
860          * Since we already block the truncate/other get_block
861          * at this point, we will have the current copy of the chain when we
862          * splice the branch into the tree.
863          */
864         if (err == -EAGAIN || !verify_chain(chain, partial)) {
865                 while (partial > chain) {
866                         brelse(partial->bh);
867                         partial--;
868                 }
869                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
870                 if (!partial) {
871                         count++;
872                         mutex_unlock(&ei->truncate_mutex);
873                         if (err)
874                                 goto cleanup;
875                         clear_buffer_new(bh_result);
876                         goto got_it;
877                 }
878         }
879
880         /*
881          * Okay, we need to do block allocation.  Lazily initialize the block
882          * allocation info here if necessary
883         */
884         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
885                 ext3_init_block_alloc_info(inode);
886
887         goal = ext3_find_goal(inode, iblock, partial);
888
889         /* the number of blocks need to allocate for [d,t]indirect blocks */
890         indirect_blks = (chain + depth) - partial - 1;
891
892         /*
893          * Next look up the indirect map to count the totoal number of
894          * direct blocks to allocate for this branch.
895          */
896         count = ext3_blks_to_allocate(partial, indirect_blks,
897                                         maxblocks, blocks_to_boundary);
898         /*
899          * Block out ext3_truncate while we alter the tree
900          */
901         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
902                                 offsets + (partial - chain), partial);
903
904         /*
905          * The ext3_splice_branch call will free and forget any buffers
906          * on the new chain if there is a failure, but that risks using
907          * up transaction credits, especially for bitmaps where the
908          * credits cannot be returned.  Can we handle this somehow?  We
909          * may need to return -EAGAIN upwards in the worst case.  --sct
910          */
911         if (!err)
912                 err = ext3_splice_branch(handle, inode, iblock,
913                                         partial, indirect_blks, count);
914         /*
915          * i_disksize growing is protected by truncate_mutex.  Don't forget to
916          * protect it if you're about to implement concurrent
917          * ext3_get_block() -bzzz
918         */
919         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
920                 ei->i_disksize = inode->i_size;
921         mutex_unlock(&ei->truncate_mutex);
922         if (err)
923                 goto cleanup;
924
925         set_buffer_new(bh_result);
926 got_it:
927         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
928         if (count > blocks_to_boundary)
929                 set_buffer_boundary(bh_result);
930         err = count;
931         /* Clean up and exit */
932         partial = chain + depth - 1;    /* the whole chain */
933 cleanup:
934         while (partial > chain) {
935                 BUFFER_TRACE(partial->bh, "call brelse");
936                 brelse(partial->bh);
937                 partial--;
938         }
939         BUFFER_TRACE(bh_result, "returned");
940 out:
941         return err;
942 }
943
944 /* Maximum number of blocks we map for direct IO at once. */
945 #define DIO_MAX_BLOCKS 4096
946 /*
947  * Number of credits we need for writing DIO_MAX_BLOCKS:
948  * We need sb + group descriptor + bitmap + inode -> 4
949  * For B blocks with A block pointers per block we need:
950  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
951  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
952  */
953 #define DIO_CREDITS 25
954
955 static int ext3_get_block(struct inode *inode, sector_t iblock,
956                         struct buffer_head *bh_result, int create)
957 {
958         handle_t *handle = ext3_journal_current_handle();
959         int ret = 0, started = 0;
960         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
961
962         if (create && !handle) {        /* Direct IO write... */
963                 if (max_blocks > DIO_MAX_BLOCKS)
964                         max_blocks = DIO_MAX_BLOCKS;
965                 handle = ext3_journal_start(inode, DIO_CREDITS +
966                                 2 * EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb));
967                 if (IS_ERR(handle)) {
968                         ret = PTR_ERR(handle);
969                         goto out;
970                 }
971                 started = 1;
972         }
973
974         ret = ext3_get_blocks_handle(handle, inode, iblock,
975                                         max_blocks, bh_result, create, 0);
976         if (ret > 0) {
977                 bh_result->b_size = (ret << inode->i_blkbits);
978                 ret = 0;
979         }
980         if (started)
981                 ext3_journal_stop(handle);
982 out:
983         return ret;
984 }
985
986 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
987                 u64 start, u64 len)
988 {
989         return generic_block_fiemap(inode, fieinfo, start, len,
990                                     ext3_get_block);
991 }
992
993 /*
994  * `handle' can be NULL if create is zero
995  */
996 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
997                                 long block, int create, int *errp)
998 {
999         struct buffer_head dummy;
1000         int fatal = 0, err;
1001
1002         J_ASSERT(handle != NULL || create == 0);
1003
1004         dummy.b_state = 0;
1005         dummy.b_blocknr = -1000;
1006         buffer_trace_init(&dummy.b_history);
1007         err = ext3_get_blocks_handle(handle, inode, block, 1,
1008                                         &dummy, create, 1);
1009         /*
1010          * ext3_get_blocks_handle() returns number of blocks
1011          * mapped. 0 in case of a HOLE.
1012          */
1013         if (err > 0) {
1014                 if (err > 1)
1015                         WARN_ON(1);
1016                 err = 0;
1017         }
1018         *errp = err;
1019         if (!err && buffer_mapped(&dummy)) {
1020                 struct buffer_head *bh;
1021                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1022                 if (!bh) {
1023                         *errp = -EIO;
1024                         goto err;
1025                 }
1026                 if (buffer_new(&dummy)) {
1027                         J_ASSERT(create != 0);
1028                         J_ASSERT(handle != NULL);
1029
1030                         /*
1031                          * Now that we do not always journal data, we should
1032                          * keep in mind whether this should always journal the
1033                          * new buffer as metadata.  For now, regular file
1034                          * writes use ext3_get_block instead, so it's not a
1035                          * problem.
1036                          */
1037                         lock_buffer(bh);
1038                         BUFFER_TRACE(bh, "call get_create_access");
1039                         fatal = ext3_journal_get_create_access(handle, bh);
1040                         if (!fatal && !buffer_uptodate(bh)) {
1041                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1042                                 set_buffer_uptodate(bh);
1043                         }
1044                         unlock_buffer(bh);
1045                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1046                         err = ext3_journal_dirty_metadata(handle, bh);
1047                         if (!fatal)
1048                                 fatal = err;
1049                 } else {
1050                         BUFFER_TRACE(bh, "not a new buffer");
1051                 }
1052                 if (fatal) {
1053                         *errp = fatal;
1054                         brelse(bh);
1055                         bh = NULL;
1056                 }
1057                 return bh;
1058         }
1059 err:
1060         return NULL;
1061 }
1062
1063 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1064                                int block, int create, int *err)
1065 {
1066         struct buffer_head * bh;
1067
1068         bh = ext3_getblk(handle, inode, block, create, err);
1069         if (!bh)
1070                 return bh;
1071         if (buffer_uptodate(bh))
1072                 return bh;
1073         ll_rw_block(READ_META, 1, &bh);
1074         wait_on_buffer(bh);
1075         if (buffer_uptodate(bh))
1076                 return bh;
1077         put_bh(bh);
1078         *err = -EIO;
1079         return NULL;
1080 }
1081
1082 static int walk_page_buffers(   handle_t *handle,
1083                                 struct buffer_head *head,
1084                                 unsigned from,
1085                                 unsigned to,
1086                                 int *partial,
1087                                 int (*fn)(      handle_t *handle,
1088                                                 struct buffer_head *bh))
1089 {
1090         struct buffer_head *bh;
1091         unsigned block_start, block_end;
1092         unsigned blocksize = head->b_size;
1093         int err, ret = 0;
1094         struct buffer_head *next;
1095
1096         for (   bh = head, block_start = 0;
1097                 ret == 0 && (bh != head || !block_start);
1098                 block_start = block_end, bh = next)
1099         {
1100                 next = bh->b_this_page;
1101                 block_end = block_start + blocksize;
1102                 if (block_end <= from || block_start >= to) {
1103                         if (partial && !buffer_uptodate(bh))
1104                                 *partial = 1;
1105                         continue;
1106                 }
1107                 err = (*fn)(handle, bh);
1108                 if (!ret)
1109                         ret = err;
1110         }
1111         return ret;
1112 }
1113
1114 /*
1115  * To preserve ordering, it is essential that the hole instantiation and
1116  * the data write be encapsulated in a single transaction.  We cannot
1117  * close off a transaction and start a new one between the ext3_get_block()
1118  * and the commit_write().  So doing the journal_start at the start of
1119  * prepare_write() is the right place.
1120  *
1121  * Also, this function can nest inside ext3_writepage() ->
1122  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1123  * has generated enough buffer credits to do the whole page.  So we won't
1124  * block on the journal in that case, which is good, because the caller may
1125  * be PF_MEMALLOC.
1126  *
1127  * By accident, ext3 can be reentered when a transaction is open via
1128  * quota file writes.  If we were to commit the transaction while thus
1129  * reentered, there can be a deadlock - we would be holding a quota
1130  * lock, and the commit would never complete if another thread had a
1131  * transaction open and was blocking on the quota lock - a ranking
1132  * violation.
1133  *
1134  * So what we do is to rely on the fact that journal_stop/journal_start
1135  * will _not_ run commit under these circumstances because handle->h_ref
1136  * is elevated.  We'll still have enough credits for the tiny quotafile
1137  * write.
1138  */
1139 static int do_journal_get_write_access(handle_t *handle,
1140                                         struct buffer_head *bh)
1141 {
1142         if (!buffer_mapped(bh) || buffer_freed(bh))
1143                 return 0;
1144         return ext3_journal_get_write_access(handle, bh);
1145 }
1146
1147 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1148                                 loff_t pos, unsigned len, unsigned flags,
1149                                 struct page **pagep, void **fsdata)
1150 {
1151         struct inode *inode = mapping->host;
1152         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1153         handle_t *handle;
1154         int retries = 0;
1155         struct page *page;
1156         pgoff_t index;
1157         unsigned from, to;
1158
1159         index = pos >> PAGE_CACHE_SHIFT;
1160         from = pos & (PAGE_CACHE_SIZE - 1);
1161         to = from + len;
1162
1163 retry:
1164         page = grab_cache_page_write_begin(mapping, index, flags);
1165         if (!page)
1166                 return -ENOMEM;
1167         *pagep = page;
1168
1169         handle = ext3_journal_start(inode, needed_blocks);
1170         if (IS_ERR(handle)) {
1171                 unlock_page(page);
1172                 page_cache_release(page);
1173                 ret = PTR_ERR(handle);
1174                 goto out;
1175         }
1176         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1177                                                         ext3_get_block);
1178         if (ret)
1179                 goto write_begin_failed;
1180
1181         if (ext3_should_journal_data(inode)) {
1182                 ret = walk_page_buffers(handle, page_buffers(page),
1183                                 from, to, NULL, do_journal_get_write_access);
1184         }
1185 write_begin_failed:
1186         if (ret) {
1187                 ext3_journal_stop(handle);
1188                 unlock_page(page);
1189                 page_cache_release(page);
1190                 /*
1191                  * block_write_begin may have instantiated a few blocks
1192                  * outside i_size.  Trim these off again. Don't need
1193                  * i_size_read because we hold i_mutex.
1194                  */
1195                 if (pos + len > inode->i_size)
1196                         vmtruncate(inode, inode->i_size);
1197         }
1198         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1199                 goto retry;
1200 out:
1201         return ret;
1202 }
1203
1204
1205 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1206 {
1207         int err = journal_dirty_data(handle, bh);
1208         if (err)
1209                 ext3_journal_abort_handle(__func__, __func__,
1210                                                 bh, handle, err);
1211         return err;
1212 }
1213
1214 /* For write_end() in data=journal mode */
1215 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1216 {
1217         if (!buffer_mapped(bh) || buffer_freed(bh))
1218                 return 0;
1219         set_buffer_uptodate(bh);
1220         return ext3_journal_dirty_metadata(handle, bh);
1221 }
1222
1223 /*
1224  * Generic write_end handler for ordered and writeback ext3 journal modes.
1225  * We can't use generic_write_end, because that unlocks the page and we need to
1226  * unlock the page after ext3_journal_stop, but ext3_journal_stop must run
1227  * after block_write_end.
1228  */
1229 static int ext3_generic_write_end(struct file *file,
1230                                 struct address_space *mapping,
1231                                 loff_t pos, unsigned len, unsigned copied,
1232                                 struct page *page, void *fsdata)
1233 {
1234         struct inode *inode = file->f_mapping->host;
1235
1236         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1237
1238         if (pos+copied > inode->i_size) {
1239                 i_size_write(inode, pos+copied);
1240                 mark_inode_dirty(inode);
1241         }
1242
1243         return copied;
1244 }
1245
1246 /*
1247  * We need to pick up the new inode size which generic_commit_write gave us
1248  * `file' can be NULL - eg, when called from page_symlink().
1249  *
1250  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1251  * buffers are managed internally.
1252  */
1253 static int ext3_ordered_write_end(struct file *file,
1254                                 struct address_space *mapping,
1255                                 loff_t pos, unsigned len, unsigned copied,
1256                                 struct page *page, void *fsdata)
1257 {
1258         handle_t *handle = ext3_journal_current_handle();
1259         struct inode *inode = file->f_mapping->host;
1260         unsigned from, to;
1261         int ret = 0, ret2;
1262
1263         from = pos & (PAGE_CACHE_SIZE - 1);
1264         to = from + len;
1265
1266         ret = walk_page_buffers(handle, page_buffers(page),
1267                 from, to, NULL, ext3_journal_dirty_data);
1268
1269         if (ret == 0) {
1270                 /*
1271                  * generic_write_end() will run mark_inode_dirty() if i_size
1272                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1273                  * into that.
1274                  */
1275                 loff_t new_i_size;
1276
1277                 new_i_size = pos + copied;
1278                 if (new_i_size > EXT3_I(inode)->i_disksize)
1279                         EXT3_I(inode)->i_disksize = new_i_size;
1280                 ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1281                                                         page, fsdata);
1282                 copied = ret2;
1283                 if (ret2 < 0)
1284                         ret = ret2;
1285         }
1286         ret2 = ext3_journal_stop(handle);
1287         if (!ret)
1288                 ret = ret2;
1289         unlock_page(page);
1290         page_cache_release(page);
1291
1292         return ret ? ret : copied;
1293 }
1294
1295 static int ext3_writeback_write_end(struct file *file,
1296                                 struct address_space *mapping,
1297                                 loff_t pos, unsigned len, unsigned copied,
1298                                 struct page *page, void *fsdata)
1299 {
1300         handle_t *handle = ext3_journal_current_handle();
1301         struct inode *inode = file->f_mapping->host;
1302         int ret = 0, ret2;
1303         loff_t new_i_size;
1304
1305         new_i_size = pos + copied;
1306         if (new_i_size > EXT3_I(inode)->i_disksize)
1307                 EXT3_I(inode)->i_disksize = new_i_size;
1308
1309         ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1310                                                         page, fsdata);
1311         copied = ret2;
1312         if (ret2 < 0)
1313                 ret = ret2;
1314
1315         ret2 = ext3_journal_stop(handle);
1316         if (!ret)
1317                 ret = ret2;
1318         unlock_page(page);
1319         page_cache_release(page);
1320
1321         return ret ? ret : copied;
1322 }
1323
1324 static int ext3_journalled_write_end(struct file *file,
1325                                 struct address_space *mapping,
1326                                 loff_t pos, unsigned len, unsigned copied,
1327                                 struct page *page, void *fsdata)
1328 {
1329         handle_t *handle = ext3_journal_current_handle();
1330         struct inode *inode = mapping->host;
1331         int ret = 0, ret2;
1332         int partial = 0;
1333         unsigned from, to;
1334
1335         from = pos & (PAGE_CACHE_SIZE - 1);
1336         to = from + len;
1337
1338         if (copied < len) {
1339                 if (!PageUptodate(page))
1340                         copied = 0;
1341                 page_zero_new_buffers(page, from+copied, to);
1342         }
1343
1344         ret = walk_page_buffers(handle, page_buffers(page), from,
1345                                 to, &partial, write_end_fn);
1346         if (!partial)
1347                 SetPageUptodate(page);
1348         if (pos+copied > inode->i_size)
1349                 i_size_write(inode, pos+copied);
1350         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1351         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1352                 EXT3_I(inode)->i_disksize = inode->i_size;
1353                 ret2 = ext3_mark_inode_dirty(handle, inode);
1354                 if (!ret)
1355                         ret = ret2;
1356         }
1357
1358         ret2 = ext3_journal_stop(handle);
1359         if (!ret)
1360                 ret = ret2;
1361         unlock_page(page);
1362         page_cache_release(page);
1363
1364         return ret ? ret : copied;
1365 }
1366
1367 /*
1368  * bmap() is special.  It gets used by applications such as lilo and by
1369  * the swapper to find the on-disk block of a specific piece of data.
1370  *
1371  * Naturally, this is dangerous if the block concerned is still in the
1372  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1373  * filesystem and enables swap, then they may get a nasty shock when the
1374  * data getting swapped to that swapfile suddenly gets overwritten by
1375  * the original zero's written out previously to the journal and
1376  * awaiting writeback in the kernel's buffer cache.
1377  *
1378  * So, if we see any bmap calls here on a modified, data-journaled file,
1379  * take extra steps to flush any blocks which might be in the cache.
1380  */
1381 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1382 {
1383         struct inode *inode = mapping->host;
1384         journal_t *journal;
1385         int err;
1386
1387         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1388                 /*
1389                  * This is a REALLY heavyweight approach, but the use of
1390                  * bmap on dirty files is expected to be extremely rare:
1391                  * only if we run lilo or swapon on a freshly made file
1392                  * do we expect this to happen.
1393                  *
1394                  * (bmap requires CAP_SYS_RAWIO so this does not
1395                  * represent an unprivileged user DOS attack --- we'd be
1396                  * in trouble if mortal users could trigger this path at
1397                  * will.)
1398                  *
1399                  * NB. EXT3_STATE_JDATA is not set on files other than
1400                  * regular files.  If somebody wants to bmap a directory
1401                  * or symlink and gets confused because the buffer
1402                  * hasn't yet been flushed to disk, they deserve
1403                  * everything they get.
1404                  */
1405
1406                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1407                 journal = EXT3_JOURNAL(inode);
1408                 journal_lock_updates(journal);
1409                 err = journal_flush(journal);
1410                 journal_unlock_updates(journal);
1411
1412                 if (err)
1413                         return 0;
1414         }
1415
1416         return generic_block_bmap(mapping,block,ext3_get_block);
1417 }
1418
1419 static int bget_one(handle_t *handle, struct buffer_head *bh)
1420 {
1421         get_bh(bh);
1422         return 0;
1423 }
1424
1425 static int bput_one(handle_t *handle, struct buffer_head *bh)
1426 {
1427         put_bh(bh);
1428         return 0;
1429 }
1430
1431 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1432 {
1433         if (buffer_mapped(bh))
1434                 return ext3_journal_dirty_data(handle, bh);
1435         return 0;
1436 }
1437
1438 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1439 {
1440         return !buffer_mapped(bh);
1441 }
1442 /*
1443  * Note that we always start a transaction even if we're not journalling
1444  * data.  This is to preserve ordering: any hole instantiation within
1445  * __block_write_full_page -> ext3_get_block() should be journalled
1446  * along with the data so we don't crash and then get metadata which
1447  * refers to old data.
1448  *
1449  * In all journalling modes block_write_full_page() will start the I/O.
1450  *
1451  * Problem:
1452  *
1453  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1454  *              ext3_writepage()
1455  *
1456  * Similar for:
1457  *
1458  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1459  *
1460  * Same applies to ext3_get_block().  We will deadlock on various things like
1461  * lock_journal and i_truncate_mutex.
1462  *
1463  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1464  * allocations fail.
1465  *
1466  * 16May01: If we're reentered then journal_current_handle() will be
1467  *          non-zero. We simply *return*.
1468  *
1469  * 1 July 2001: @@@ FIXME:
1470  *   In journalled data mode, a data buffer may be metadata against the
1471  *   current transaction.  But the same file is part of a shared mapping
1472  *   and someone does a writepage() on it.
1473  *
1474  *   We will move the buffer onto the async_data list, but *after* it has
1475  *   been dirtied. So there's a small window where we have dirty data on
1476  *   BJ_Metadata.
1477  *
1478  *   Note that this only applies to the last partial page in the file.  The
1479  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1480  *   broken code anyway: it's wrong for msync()).
1481  *
1482  *   It's a rare case: affects the final partial page, for journalled data
1483  *   where the file is subject to bith write() and writepage() in the same
1484  *   transction.  To fix it we'll need a custom block_write_full_page().
1485  *   We'll probably need that anyway for journalling writepage() output.
1486  *
1487  * We don't honour synchronous mounts for writepage().  That would be
1488  * disastrous.  Any write() or metadata operation will sync the fs for
1489  * us.
1490  *
1491  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1492  * we don't need to open a transaction here.
1493  */
1494 static int ext3_ordered_writepage(struct page *page,
1495                                 struct writeback_control *wbc)
1496 {
1497         struct inode *inode = page->mapping->host;
1498         struct buffer_head *page_bufs;
1499         handle_t *handle = NULL;
1500         int ret = 0;
1501         int err;
1502
1503         J_ASSERT(PageLocked(page));
1504
1505         /*
1506          * We give up here if we're reentered, because it might be for a
1507          * different filesystem.
1508          */
1509         if (ext3_journal_current_handle())
1510                 goto out_fail;
1511
1512         if (!page_has_buffers(page)) {
1513                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1514                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1515         } else if (!walk_page_buffers(NULL, page_buffers(page), 0, PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1516                 /* Provide NULL instead of get_block so that we catch bugs if buffers weren't really mapped */
1517                 return block_write_full_page(page, NULL, wbc);
1518         }
1519         page_bufs = page_buffers(page);
1520
1521         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1522
1523         if (IS_ERR(handle)) {
1524                 ret = PTR_ERR(handle);
1525                 goto out_fail;
1526         }
1527
1528         walk_page_buffers(handle, page_bufs, 0,
1529                         PAGE_CACHE_SIZE, NULL, bget_one);
1530
1531         ret = block_write_full_page(page, ext3_get_block, wbc);
1532
1533         /*
1534          * The page can become unlocked at any point now, and
1535          * truncate can then come in and change things.  So we
1536          * can't touch *page from now on.  But *page_bufs is
1537          * safe due to elevated refcount.
1538          */
1539
1540         /*
1541          * And attach them to the current transaction.  But only if
1542          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1543          * and generally junk.
1544          */
1545         if (ret == 0) {
1546                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1547                                         NULL, journal_dirty_data_fn);
1548                 if (!ret)
1549                         ret = err;
1550         }
1551         walk_page_buffers(handle, page_bufs, 0,
1552                         PAGE_CACHE_SIZE, NULL, bput_one);
1553         err = ext3_journal_stop(handle);
1554         if (!ret)
1555                 ret = err;
1556         return ret;
1557
1558 out_fail:
1559         redirty_page_for_writepage(wbc, page);
1560         unlock_page(page);
1561         return ret;
1562 }
1563
1564 static int ext3_writeback_writepage(struct page *page,
1565                                 struct writeback_control *wbc)
1566 {
1567         struct inode *inode = page->mapping->host;
1568         handle_t *handle = NULL;
1569         int ret = 0;
1570         int err;
1571
1572         if (ext3_journal_current_handle())
1573                 goto out_fail;
1574
1575         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1576         if (IS_ERR(handle)) {
1577                 ret = PTR_ERR(handle);
1578                 goto out_fail;
1579         }
1580
1581         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1582                 ret = nobh_writepage(page, ext3_get_block, wbc);
1583         else
1584                 ret = block_write_full_page(page, ext3_get_block, wbc);
1585
1586         err = ext3_journal_stop(handle);
1587         if (!ret)
1588                 ret = err;
1589         return ret;
1590
1591 out_fail:
1592         redirty_page_for_writepage(wbc, page);
1593         unlock_page(page);
1594         return ret;
1595 }
1596
1597 static int ext3_journalled_writepage(struct page *page,
1598                                 struct writeback_control *wbc)
1599 {
1600         struct inode *inode = page->mapping->host;
1601         handle_t *handle = NULL;
1602         int ret = 0;
1603         int err;
1604
1605         if (ext3_journal_current_handle())
1606                 goto no_write;
1607
1608         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1609         if (IS_ERR(handle)) {
1610                 ret = PTR_ERR(handle);
1611                 goto no_write;
1612         }
1613
1614         if (!page_has_buffers(page) || PageChecked(page)) {
1615                 /*
1616                  * It's mmapped pagecache.  Add buffers and journal it.  There
1617                  * doesn't seem much point in redirtying the page here.
1618                  */
1619                 ClearPageChecked(page);
1620                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1621                                         ext3_get_block);
1622                 if (ret != 0) {
1623                         ext3_journal_stop(handle);
1624                         goto out_unlock;
1625                 }
1626                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1627                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1628
1629                 err = walk_page_buffers(handle, page_buffers(page), 0,
1630                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1631                 if (ret == 0)
1632                         ret = err;
1633                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1634                 unlock_page(page);
1635         } else {
1636                 /*
1637                  * It may be a page full of checkpoint-mode buffers.  We don't
1638                  * really know unless we go poke around in the buffer_heads.
1639                  * But block_write_full_page will do the right thing.
1640                  */
1641                 ret = block_write_full_page(page, ext3_get_block, wbc);
1642         }
1643         err = ext3_journal_stop(handle);
1644         if (!ret)
1645                 ret = err;
1646 out:
1647         return ret;
1648
1649 no_write:
1650         redirty_page_for_writepage(wbc, page);
1651 out_unlock:
1652         unlock_page(page);
1653         goto out;
1654 }
1655
1656 static int ext3_readpage(struct file *file, struct page *page)
1657 {
1658         return mpage_readpage(page, ext3_get_block);
1659 }
1660
1661 static int
1662 ext3_readpages(struct file *file, struct address_space *mapping,
1663                 struct list_head *pages, unsigned nr_pages)
1664 {
1665         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1666 }
1667
1668 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1669 {
1670         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1671
1672         /*
1673          * If it's a full truncate we just forget about the pending dirtying
1674          */
1675         if (offset == 0)
1676                 ClearPageChecked(page);
1677
1678         journal_invalidatepage(journal, page, offset);
1679 }
1680
1681 static int ext3_releasepage(struct page *page, gfp_t wait)
1682 {
1683         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1684
1685         WARN_ON(PageChecked(page));
1686         if (!page_has_buffers(page))
1687                 return 0;
1688         return journal_try_to_free_buffers(journal, page, wait);
1689 }
1690
1691 /*
1692  * If the O_DIRECT write will extend the file then add this inode to the
1693  * orphan list.  So recovery will truncate it back to the original size
1694  * if the machine crashes during the write.
1695  *
1696  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1697  * crashes then stale disk data _may_ be exposed inside the file. But current
1698  * VFS code falls back into buffered path in that case so we are safe.
1699  */
1700 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1701                         const struct iovec *iov, loff_t offset,
1702                         unsigned long nr_segs)
1703 {
1704         struct file *file = iocb->ki_filp;
1705         struct inode *inode = file->f_mapping->host;
1706         struct ext3_inode_info *ei = EXT3_I(inode);
1707         handle_t *handle;
1708         ssize_t ret;
1709         int orphan = 0;
1710         size_t count = iov_length(iov, nr_segs);
1711
1712         if (rw == WRITE) {
1713                 loff_t final_size = offset + count;
1714
1715                 if (final_size > inode->i_size) {
1716                         /* Credits for sb + inode write */
1717                         handle = ext3_journal_start(inode, 2);
1718                         if (IS_ERR(handle)) {
1719                                 ret = PTR_ERR(handle);
1720                                 goto out;
1721                         }
1722                         ret = ext3_orphan_add(handle, inode);
1723                         if (ret) {
1724                                 ext3_journal_stop(handle);
1725                                 goto out;
1726                         }
1727                         orphan = 1;
1728                         ei->i_disksize = inode->i_size;
1729                         ext3_journal_stop(handle);
1730                 }
1731         }
1732
1733         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1734                                  offset, nr_segs,
1735                                  ext3_get_block, NULL);
1736
1737         if (orphan) {
1738                 int err;
1739
1740                 /* Credits for sb + inode write */
1741                 handle = ext3_journal_start(inode, 2);
1742                 if (IS_ERR(handle)) {
1743                         /* This is really bad luck. We've written the data
1744                          * but cannot extend i_size. Bail out and pretend
1745                          * the write failed... */
1746                         ret = PTR_ERR(handle);
1747                         goto out;
1748                 }
1749                 if (inode->i_nlink)
1750                         ext3_orphan_del(handle, inode);
1751                 if (ret > 0) {
1752                         loff_t end = offset + ret;
1753                         if (end > inode->i_size) {
1754                                 ei->i_disksize = end;
1755                                 i_size_write(inode, end);
1756                                 /*
1757                                  * We're going to return a positive `ret'
1758                                  * here due to non-zero-length I/O, so there's
1759                                  * no way of reporting error returns from
1760                                  * ext3_mark_inode_dirty() to userspace.  So
1761                                  * ignore it.
1762                                  */
1763                                 ext3_mark_inode_dirty(handle, inode);
1764                         }
1765                 }
1766                 err = ext3_journal_stop(handle);
1767                 if (ret == 0)
1768                         ret = err;
1769         }
1770 out:
1771         return ret;
1772 }
1773
1774 /*
1775  * Pages can be marked dirty completely asynchronously from ext3's journalling
1776  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1777  * much here because ->set_page_dirty is called under VFS locks.  The page is
1778  * not necessarily locked.
1779  *
1780  * We cannot just dirty the page and leave attached buffers clean, because the
1781  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1782  * or jbddirty because all the journalling code will explode.
1783  *
1784  * So what we do is to mark the page "pending dirty" and next time writepage
1785  * is called, propagate that into the buffers appropriately.
1786  */
1787 static int ext3_journalled_set_page_dirty(struct page *page)
1788 {
1789         SetPageChecked(page);
1790         return __set_page_dirty_nobuffers(page);
1791 }
1792
1793 static const struct address_space_operations ext3_ordered_aops = {
1794         .readpage               = ext3_readpage,
1795         .readpages              = ext3_readpages,
1796         .writepage              = ext3_ordered_writepage,
1797         .sync_page              = block_sync_page,
1798         .write_begin            = ext3_write_begin,
1799         .write_end              = ext3_ordered_write_end,
1800         .bmap                   = ext3_bmap,
1801         .invalidatepage         = ext3_invalidatepage,
1802         .releasepage            = ext3_releasepage,
1803         .direct_IO              = ext3_direct_IO,
1804         .migratepage            = buffer_migrate_page,
1805         .is_partially_uptodate  = block_is_partially_uptodate,
1806 };
1807
1808 static const struct address_space_operations ext3_writeback_aops = {
1809         .readpage               = ext3_readpage,
1810         .readpages              = ext3_readpages,
1811         .writepage              = ext3_writeback_writepage,
1812         .sync_page              = block_sync_page,
1813         .write_begin            = ext3_write_begin,
1814         .write_end              = ext3_writeback_write_end,
1815         .bmap                   = ext3_bmap,
1816         .invalidatepage         = ext3_invalidatepage,
1817         .releasepage            = ext3_releasepage,
1818         .direct_IO              = ext3_direct_IO,
1819         .migratepage            = buffer_migrate_page,
1820         .is_partially_uptodate  = block_is_partially_uptodate,
1821 };
1822
1823 static const struct address_space_operations ext3_journalled_aops = {
1824         .readpage               = ext3_readpage,
1825         .readpages              = ext3_readpages,
1826         .writepage              = ext3_journalled_writepage,
1827         .sync_page              = block_sync_page,
1828         .write_begin            = ext3_write_begin,
1829         .write_end              = ext3_journalled_write_end,
1830         .set_page_dirty         = ext3_journalled_set_page_dirty,
1831         .bmap                   = ext3_bmap,
1832         .invalidatepage         = ext3_invalidatepage,
1833         .releasepage            = ext3_releasepage,
1834         .is_partially_uptodate  = block_is_partially_uptodate,
1835 };
1836
1837 void ext3_set_aops(struct inode *inode)
1838 {
1839         if (ext3_should_order_data(inode))
1840                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1841         else if (ext3_should_writeback_data(inode))
1842                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1843         else
1844                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1845 }
1846
1847 /*
1848  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1849  * up to the end of the block which corresponds to `from'.
1850  * This required during truncate. We need to physically zero the tail end
1851  * of that block so it doesn't yield old data if the file is later grown.
1852  */
1853 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1854                 struct address_space *mapping, loff_t from)
1855 {
1856         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1857         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1858         unsigned blocksize, iblock, length, pos;
1859         struct inode *inode = mapping->host;
1860         struct buffer_head *bh;
1861         int err = 0;
1862
1863         blocksize = inode->i_sb->s_blocksize;
1864         length = blocksize - (offset & (blocksize - 1));
1865         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1866
1867         /*
1868          * For "nobh" option,  we can only work if we don't need to
1869          * read-in the page - otherwise we create buffers to do the IO.
1870          */
1871         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1872              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1873                 zero_user(page, offset, length);
1874                 set_page_dirty(page);
1875                 goto unlock;
1876         }
1877
1878         if (!page_has_buffers(page))
1879                 create_empty_buffers(page, blocksize, 0);
1880
1881         /* Find the buffer that contains "offset" */
1882         bh = page_buffers(page);
1883         pos = blocksize;
1884         while (offset >= pos) {
1885                 bh = bh->b_this_page;
1886                 iblock++;
1887                 pos += blocksize;
1888         }
1889
1890         err = 0;
1891         if (buffer_freed(bh)) {
1892                 BUFFER_TRACE(bh, "freed: skip");
1893                 goto unlock;
1894         }
1895
1896         if (!buffer_mapped(bh)) {
1897                 BUFFER_TRACE(bh, "unmapped");
1898                 ext3_get_block(inode, iblock, bh, 0);
1899                 /* unmapped? It's a hole - nothing to do */
1900                 if (!buffer_mapped(bh)) {
1901                         BUFFER_TRACE(bh, "still unmapped");
1902                         goto unlock;
1903                 }
1904         }
1905
1906         /* Ok, it's mapped. Make sure it's up-to-date */
1907         if (PageUptodate(page))
1908                 set_buffer_uptodate(bh);
1909
1910         if (!buffer_uptodate(bh)) {
1911                 err = -EIO;
1912                 ll_rw_block(READ, 1, &bh);
1913                 wait_on_buffer(bh);
1914                 /* Uhhuh. Read error. Complain and punt. */
1915                 if (!buffer_uptodate(bh))
1916                         goto unlock;
1917         }
1918
1919         if (ext3_should_journal_data(inode)) {
1920                 BUFFER_TRACE(bh, "get write access");
1921                 err = ext3_journal_get_write_access(handle, bh);
1922                 if (err)
1923                         goto unlock;
1924         }
1925
1926         zero_user(page, offset, length);
1927         BUFFER_TRACE(bh, "zeroed end of block");
1928
1929         err = 0;
1930         if (ext3_should_journal_data(inode)) {
1931                 err = ext3_journal_dirty_metadata(handle, bh);
1932         } else {
1933                 if (ext3_should_order_data(inode))
1934                         err = ext3_journal_dirty_data(handle, bh);
1935                 mark_buffer_dirty(bh);
1936         }
1937
1938 unlock:
1939         unlock_page(page);
1940         page_cache_release(page);
1941         return err;
1942 }
1943
1944 /*
1945  * Probably it should be a library function... search for first non-zero word
1946  * or memcmp with zero_page, whatever is better for particular architecture.
1947  * Linus?
1948  */
1949 static inline int all_zeroes(__le32 *p, __le32 *q)
1950 {
1951         while (p < q)
1952                 if (*p++)
1953                         return 0;
1954         return 1;
1955 }
1956
1957 /**
1958  *      ext3_find_shared - find the indirect blocks for partial truncation.
1959  *      @inode:   inode in question
1960  *      @depth:   depth of the affected branch
1961  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1962  *      @chain:   place to store the pointers to partial indirect blocks
1963  *      @top:     place to the (detached) top of branch
1964  *
1965  *      This is a helper function used by ext3_truncate().
1966  *
1967  *      When we do truncate() we may have to clean the ends of several
1968  *      indirect blocks but leave the blocks themselves alive. Block is
1969  *      partially truncated if some data below the new i_size is refered
1970  *      from it (and it is on the path to the first completely truncated
1971  *      data block, indeed).  We have to free the top of that path along
1972  *      with everything to the right of the path. Since no allocation
1973  *      past the truncation point is possible until ext3_truncate()
1974  *      finishes, we may safely do the latter, but top of branch may
1975  *      require special attention - pageout below the truncation point
1976  *      might try to populate it.
1977  *
1978  *      We atomically detach the top of branch from the tree, store the
1979  *      block number of its root in *@top, pointers to buffer_heads of
1980  *      partially truncated blocks - in @chain[].bh and pointers to
1981  *      their last elements that should not be removed - in
1982  *      @chain[].p. Return value is the pointer to last filled element
1983  *      of @chain.
1984  *
1985  *      The work left to caller to do the actual freeing of subtrees:
1986  *              a) free the subtree starting from *@top
1987  *              b) free the subtrees whose roots are stored in
1988  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1989  *              c) free the subtrees growing from the inode past the @chain[0].
1990  *                      (no partially truncated stuff there).  */
1991
1992 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1993                         int offsets[4], Indirect chain[4], __le32 *top)
1994 {
1995         Indirect *partial, *p;
1996         int k, err;
1997
1998         *top = 0;
1999         /* Make k index the deepest non-null offest + 1 */
2000         for (k = depth; k > 1 && !offsets[k-1]; k--)
2001                 ;
2002         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2003         /* Writer: pointers */
2004         if (!partial)
2005                 partial = chain + k-1;
2006         /*
2007          * If the branch acquired continuation since we've looked at it -
2008          * fine, it should all survive and (new) top doesn't belong to us.
2009          */
2010         if (!partial->key && *partial->p)
2011                 /* Writer: end */
2012                 goto no_top;
2013         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2014                 ;
2015         /*
2016          * OK, we've found the last block that must survive. The rest of our
2017          * branch should be detached before unlocking. However, if that rest
2018          * of branch is all ours and does not grow immediately from the inode
2019          * it's easier to cheat and just decrement partial->p.
2020          */
2021         if (p == chain + k - 1 && p > chain) {
2022                 p->p--;
2023         } else {
2024                 *top = *p->p;
2025                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2026 #if 0
2027                 *p->p = 0;
2028 #endif
2029         }
2030         /* Writer: end */
2031
2032         while(partial > p) {
2033                 brelse(partial->bh);
2034                 partial--;
2035         }
2036 no_top:
2037         return partial;
2038 }
2039
2040 /*
2041  * Zero a number of block pointers in either an inode or an indirect block.
2042  * If we restart the transaction we must again get write access to the
2043  * indirect block for further modification.
2044  *
2045  * We release `count' blocks on disk, but (last - first) may be greater
2046  * than `count' because there can be holes in there.
2047  */
2048 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2049                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2050                 unsigned long count, __le32 *first, __le32 *last)
2051 {
2052         __le32 *p;
2053         if (try_to_extend_transaction(handle, inode)) {
2054                 if (bh) {
2055                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2056                         ext3_journal_dirty_metadata(handle, bh);
2057                 }
2058                 ext3_mark_inode_dirty(handle, inode);
2059                 ext3_journal_test_restart(handle, inode);
2060                 if (bh) {
2061                         BUFFER_TRACE(bh, "retaking write access");
2062                         ext3_journal_get_write_access(handle, bh);
2063                 }
2064         }
2065
2066         /*
2067          * Any buffers which are on the journal will be in memory. We find
2068          * them on the hash table so journal_revoke() will run journal_forget()
2069          * on them.  We've already detached each block from the file, so
2070          * bforget() in journal_forget() should be safe.
2071          *
2072          * AKPM: turn on bforget in journal_forget()!!!
2073          */
2074         for (p = first; p < last; p++) {
2075                 u32 nr = le32_to_cpu(*p);
2076                 if (nr) {
2077                         struct buffer_head *bh;
2078
2079                         *p = 0;
2080                         bh = sb_find_get_block(inode->i_sb, nr);
2081                         ext3_forget(handle, 0, inode, bh, nr);
2082                 }
2083         }
2084
2085         ext3_free_blocks(handle, inode, block_to_free, count);
2086 }
2087
2088 /**
2089  * ext3_free_data - free a list of data blocks
2090  * @handle:     handle for this transaction
2091  * @inode:      inode we are dealing with
2092  * @this_bh:    indirect buffer_head which contains *@first and *@last
2093  * @first:      array of block numbers
2094  * @last:       points immediately past the end of array
2095  *
2096  * We are freeing all blocks refered from that array (numbers are stored as
2097  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2098  *
2099  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2100  * blocks are contiguous then releasing them at one time will only affect one
2101  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2102  * actually use a lot of journal space.
2103  *
2104  * @this_bh will be %NULL if @first and @last point into the inode's direct
2105  * block pointers.
2106  */
2107 static void ext3_free_data(handle_t *handle, struct inode *inode,
2108                            struct buffer_head *this_bh,
2109                            __le32 *first, __le32 *last)
2110 {
2111         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2112         unsigned long count = 0;            /* Number of blocks in the run */
2113         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2114                                                corresponding to
2115                                                block_to_free */
2116         ext3_fsblk_t nr;                    /* Current block # */
2117         __le32 *p;                          /* Pointer into inode/ind
2118                                                for current block */
2119         int err;
2120
2121         if (this_bh) {                          /* For indirect block */
2122                 BUFFER_TRACE(this_bh, "get_write_access");
2123                 err = ext3_journal_get_write_access(handle, this_bh);
2124                 /* Important: if we can't update the indirect pointers
2125                  * to the blocks, we can't free them. */
2126                 if (err)
2127                         return;
2128         }
2129
2130         for (p = first; p < last; p++) {
2131                 nr = le32_to_cpu(*p);
2132                 if (nr) {
2133                         /* accumulate blocks to free if they're contiguous */
2134                         if (count == 0) {
2135                                 block_to_free = nr;
2136                                 block_to_free_p = p;
2137                                 count = 1;
2138                         } else if (nr == block_to_free + count) {
2139                                 count++;
2140                         } else {
2141                                 ext3_clear_blocks(handle, inode, this_bh,
2142                                                   block_to_free,
2143                                                   count, block_to_free_p, p);
2144                                 block_to_free = nr;
2145                                 block_to_free_p = p;
2146                                 count = 1;
2147                         }
2148                 }
2149         }
2150
2151         if (count > 0)
2152                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2153                                   count, block_to_free_p, p);
2154
2155         if (this_bh) {
2156                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2157
2158                 /*
2159                  * The buffer head should have an attached journal head at this
2160                  * point. However, if the data is corrupted and an indirect
2161                  * block pointed to itself, it would have been detached when
2162                  * the block was cleared. Check for this instead of OOPSing.
2163                  */
2164                 if (bh2jh(this_bh))
2165                         ext3_journal_dirty_metadata(handle, this_bh);
2166                 else
2167                         ext3_error(inode->i_sb, "ext3_free_data",
2168                                    "circular indirect block detected, "
2169                                    "inode=%lu, block=%llu",
2170                                    inode->i_ino,
2171                                    (unsigned long long)this_bh->b_blocknr);
2172         }
2173 }
2174
2175 /**
2176  *      ext3_free_branches - free an array of branches
2177  *      @handle: JBD handle for this transaction
2178  *      @inode: inode we are dealing with
2179  *      @parent_bh: the buffer_head which contains *@first and *@last
2180  *      @first: array of block numbers
2181  *      @last:  pointer immediately past the end of array
2182  *      @depth: depth of the branches to free
2183  *
2184  *      We are freeing all blocks refered from these branches (numbers are
2185  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2186  *      appropriately.
2187  */
2188 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2189                                struct buffer_head *parent_bh,
2190                                __le32 *first, __le32 *last, int depth)
2191 {
2192         ext3_fsblk_t nr;
2193         __le32 *p;
2194
2195         if (is_handle_aborted(handle))
2196                 return;
2197
2198         if (depth--) {
2199                 struct buffer_head *bh;
2200                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2201                 p = last;
2202                 while (--p >= first) {
2203                         nr = le32_to_cpu(*p);
2204                         if (!nr)
2205                                 continue;               /* A hole */
2206
2207                         /* Go read the buffer for the next level down */
2208                         bh = sb_bread(inode->i_sb, nr);
2209
2210                         /*
2211                          * A read failure? Report error and clear slot
2212                          * (should be rare).
2213                          */
2214                         if (!bh) {
2215                                 ext3_error(inode->i_sb, "ext3_free_branches",
2216                                            "Read failure, inode=%lu, block="E3FSBLK,
2217                                            inode->i_ino, nr);
2218                                 continue;
2219                         }
2220
2221                         /* This zaps the entire block.  Bottom up. */
2222                         BUFFER_TRACE(bh, "free child branches");
2223                         ext3_free_branches(handle, inode, bh,
2224                                            (__le32*)bh->b_data,
2225                                            (__le32*)bh->b_data + addr_per_block,
2226                                            depth);
2227
2228                         /*
2229                          * We've probably journalled the indirect block several
2230                          * times during the truncate.  But it's no longer
2231                          * needed and we now drop it from the transaction via
2232                          * journal_revoke().
2233                          *
2234                          * That's easy if it's exclusively part of this
2235                          * transaction.  But if it's part of the committing
2236                          * transaction then journal_forget() will simply
2237                          * brelse() it.  That means that if the underlying
2238                          * block is reallocated in ext3_get_block(),
2239                          * unmap_underlying_metadata() will find this block
2240                          * and will try to get rid of it.  damn, damn.
2241                          *
2242                          * If this block has already been committed to the
2243                          * journal, a revoke record will be written.  And
2244                          * revoke records must be emitted *before* clearing
2245                          * this block's bit in the bitmaps.
2246                          */
2247                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2248
2249                         /*
2250                          * Everything below this this pointer has been
2251                          * released.  Now let this top-of-subtree go.
2252                          *
2253                          * We want the freeing of this indirect block to be
2254                          * atomic in the journal with the updating of the
2255                          * bitmap block which owns it.  So make some room in
2256                          * the journal.
2257                          *
2258                          * We zero the parent pointer *after* freeing its
2259                          * pointee in the bitmaps, so if extend_transaction()
2260                          * for some reason fails to put the bitmap changes and
2261                          * the release into the same transaction, recovery
2262                          * will merely complain about releasing a free block,
2263                          * rather than leaking blocks.
2264                          */
2265                         if (is_handle_aborted(handle))
2266                                 return;
2267                         if (try_to_extend_transaction(handle, inode)) {
2268                                 ext3_mark_inode_dirty(handle, inode);
2269                                 ext3_journal_test_restart(handle, inode);
2270                         }
2271
2272                         ext3_free_blocks(handle, inode, nr, 1);
2273
2274                         if (parent_bh) {
2275                                 /*
2276                                  * The block which we have just freed is
2277                                  * pointed to by an indirect block: journal it
2278                                  */
2279                                 BUFFER_TRACE(parent_bh, "get_write_access");
2280                                 if (!ext3_journal_get_write_access(handle,
2281                                                                    parent_bh)){
2282                                         *p = 0;
2283                                         BUFFER_TRACE(parent_bh,
2284                                         "call ext3_journal_dirty_metadata");
2285                                         ext3_journal_dirty_metadata(handle,
2286                                                                     parent_bh);
2287                                 }
2288                         }
2289                 }
2290         } else {
2291                 /* We have reached the bottom of the tree. */
2292                 BUFFER_TRACE(parent_bh, "free data blocks");
2293                 ext3_free_data(handle, inode, parent_bh, first, last);
2294         }
2295 }
2296
2297 int ext3_can_truncate(struct inode *inode)
2298 {
2299         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2300                 return 0;
2301         if (S_ISREG(inode->i_mode))
2302                 return 1;
2303         if (S_ISDIR(inode->i_mode))
2304                 return 1;
2305         if (S_ISLNK(inode->i_mode))
2306                 return !ext3_inode_is_fast_symlink(inode);
2307         return 0;
2308 }
2309
2310 /*
2311  * ext3_truncate()
2312  *
2313  * We block out ext3_get_block() block instantiations across the entire
2314  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2315  * simultaneously on behalf of the same inode.
2316  *
2317  * As we work through the truncate and commmit bits of it to the journal there
2318  * is one core, guiding principle: the file's tree must always be consistent on
2319  * disk.  We must be able to restart the truncate after a crash.
2320  *
2321  * The file's tree may be transiently inconsistent in memory (although it
2322  * probably isn't), but whenever we close off and commit a journal transaction,
2323  * the contents of (the filesystem + the journal) must be consistent and
2324  * restartable.  It's pretty simple, really: bottom up, right to left (although
2325  * left-to-right works OK too).
2326  *
2327  * Note that at recovery time, journal replay occurs *before* the restart of
2328  * truncate against the orphan inode list.
2329  *
2330  * The committed inode has the new, desired i_size (which is the same as
2331  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2332  * that this inode's truncate did not complete and it will again call
2333  * ext3_truncate() to have another go.  So there will be instantiated blocks
2334  * to the right of the truncation point in a crashed ext3 filesystem.  But
2335  * that's fine - as long as they are linked from the inode, the post-crash
2336  * ext3_truncate() run will find them and release them.
2337  */
2338 void ext3_truncate(struct inode *inode)
2339 {
2340         handle_t *handle;
2341         struct ext3_inode_info *ei = EXT3_I(inode);
2342         __le32 *i_data = ei->i_data;
2343         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2344         struct address_space *mapping = inode->i_mapping;
2345         int offsets[4];
2346         Indirect chain[4];
2347         Indirect *partial;
2348         __le32 nr = 0;
2349         int n;
2350         long last_block;
2351         unsigned blocksize = inode->i_sb->s_blocksize;
2352         struct page *page;
2353
2354         if (!ext3_can_truncate(inode))
2355                 return;
2356
2357         /*
2358          * We have to lock the EOF page here, because lock_page() nests
2359          * outside journal_start().
2360          */
2361         if ((inode->i_size & (blocksize - 1)) == 0) {
2362                 /* Block boundary? Nothing to do */
2363                 page = NULL;
2364         } else {
2365                 page = grab_cache_page(mapping,
2366                                 inode->i_size >> PAGE_CACHE_SHIFT);
2367                 if (!page)
2368                         return;
2369         }
2370
2371         handle = start_transaction(inode);
2372         if (IS_ERR(handle)) {
2373                 if (page) {
2374                         clear_highpage(page);
2375                         flush_dcache_page(page);
2376                         unlock_page(page);
2377                         page_cache_release(page);
2378                 }
2379                 return;         /* AKPM: return what? */
2380         }
2381
2382         last_block = (inode->i_size + blocksize-1)
2383                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2384
2385         if (page)
2386                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2387
2388         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2389         if (n == 0)
2390                 goto out_stop;  /* error */
2391
2392         /*
2393          * OK.  This truncate is going to happen.  We add the inode to the
2394          * orphan list, so that if this truncate spans multiple transactions,
2395          * and we crash, we will resume the truncate when the filesystem
2396          * recovers.  It also marks the inode dirty, to catch the new size.
2397          *
2398          * Implication: the file must always be in a sane, consistent
2399          * truncatable state while each transaction commits.
2400          */
2401         if (ext3_orphan_add(handle, inode))
2402                 goto out_stop;
2403
2404         /*
2405          * The orphan list entry will now protect us from any crash which
2406          * occurs before the truncate completes, so it is now safe to propagate
2407          * the new, shorter inode size (held for now in i_size) into the
2408          * on-disk inode. We do this via i_disksize, which is the value which
2409          * ext3 *really* writes onto the disk inode.
2410          */
2411         ei->i_disksize = inode->i_size;
2412
2413         /*
2414          * From here we block out all ext3_get_block() callers who want to
2415          * modify the block allocation tree.
2416          */
2417         mutex_lock(&ei->truncate_mutex);
2418
2419         if (n == 1) {           /* direct blocks */
2420                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2421                                i_data + EXT3_NDIR_BLOCKS);
2422                 goto do_indirects;
2423         }
2424
2425         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2426         /* Kill the top of shared branch (not detached) */
2427         if (nr) {
2428                 if (partial == chain) {
2429                         /* Shared branch grows from the inode */
2430                         ext3_free_branches(handle, inode, NULL,
2431                                            &nr, &nr+1, (chain+n-1) - partial);
2432                         *partial->p = 0;
2433                         /*
2434                          * We mark the inode dirty prior to restart,
2435                          * and prior to stop.  No need for it here.
2436                          */
2437                 } else {
2438                         /* Shared branch grows from an indirect block */
2439                         BUFFER_TRACE(partial->bh, "get_write_access");
2440                         ext3_free_branches(handle, inode, partial->bh,
2441                                         partial->p,
2442                                         partial->p+1, (chain+n-1) - partial);
2443                 }
2444         }
2445         /* Clear the ends of indirect blocks on the shared branch */
2446         while (partial > chain) {
2447                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2448                                    (__le32*)partial->bh->b_data+addr_per_block,
2449                                    (chain+n-1) - partial);
2450                 BUFFER_TRACE(partial->bh, "call brelse");
2451                 brelse (partial->bh);
2452                 partial--;
2453         }
2454 do_indirects:
2455         /* Kill the remaining (whole) subtrees */
2456         switch (offsets[0]) {
2457         default:
2458                 nr = i_data[EXT3_IND_BLOCK];
2459                 if (nr) {
2460                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2461                         i_data[EXT3_IND_BLOCK] = 0;
2462                 }
2463         case EXT3_IND_BLOCK:
2464                 nr = i_data[EXT3_DIND_BLOCK];
2465                 if (nr) {
2466                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2467                         i_data[EXT3_DIND_BLOCK] = 0;
2468                 }
2469         case EXT3_DIND_BLOCK:
2470                 nr = i_data[EXT3_TIND_BLOCK];
2471                 if (nr) {
2472                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2473                         i_data[EXT3_TIND_BLOCK] = 0;
2474                 }
2475         case EXT3_TIND_BLOCK:
2476                 ;
2477         }
2478
2479         ext3_discard_reservation(inode);
2480
2481         mutex_unlock(&ei->truncate_mutex);
2482         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2483         ext3_mark_inode_dirty(handle, inode);
2484
2485         /*
2486          * In a multi-transaction truncate, we only make the final transaction
2487          * synchronous
2488          */
2489         if (IS_SYNC(inode))
2490                 handle->h_sync = 1;
2491 out_stop:
2492         /*
2493          * If this was a simple ftruncate(), and the file will remain alive
2494          * then we need to clear up the orphan record which we created above.
2495          * However, if this was a real unlink then we were called by
2496          * ext3_delete_inode(), and we allow that function to clean up the
2497          * orphan info for us.
2498          */
2499         if (inode->i_nlink)
2500                 ext3_orphan_del(handle, inode);
2501
2502         ext3_journal_stop(handle);
2503 }
2504
2505 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2506                 unsigned long ino, struct ext3_iloc *iloc)
2507 {
2508         unsigned long block_group;
2509         unsigned long offset;
2510         ext3_fsblk_t block;
2511         struct ext3_group_desc *gdp;
2512
2513         if (!ext3_valid_inum(sb, ino)) {
2514                 /*
2515                  * This error is already checked for in namei.c unless we are
2516                  * looking at an NFS filehandle, in which case no error
2517                  * report is needed
2518                  */
2519                 return 0;
2520         }
2521
2522         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2523         gdp = ext3_get_group_desc(sb, block_group, NULL);
2524         if (!gdp)
2525                 return 0;
2526         /*
2527          * Figure out the offset within the block group inode table
2528          */
2529         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2530                 EXT3_INODE_SIZE(sb);
2531         block = le32_to_cpu(gdp->bg_inode_table) +
2532                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2533
2534         iloc->block_group = block_group;
2535         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2536         return block;
2537 }
2538
2539 /*
2540  * ext3_get_inode_loc returns with an extra refcount against the inode's
2541  * underlying buffer_head on success. If 'in_mem' is true, we have all
2542  * data in memory that is needed to recreate the on-disk version of this
2543  * inode.
2544  */
2545 static int __ext3_get_inode_loc(struct inode *inode,
2546                                 struct ext3_iloc *iloc, int in_mem)
2547 {
2548         ext3_fsblk_t block;
2549         struct buffer_head *bh;
2550
2551         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2552         if (!block)
2553                 return -EIO;
2554
2555         bh = sb_getblk(inode->i_sb, block);
2556         if (!bh) {
2557                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2558                                 "unable to read inode block - "
2559                                 "inode=%lu, block="E3FSBLK,
2560                                  inode->i_ino, block);
2561                 return -EIO;
2562         }
2563         if (!buffer_uptodate(bh)) {
2564                 lock_buffer(bh);
2565
2566                 /*
2567                  * If the buffer has the write error flag, we have failed
2568                  * to write out another inode in the same block.  In this
2569                  * case, we don't have to read the block because we may
2570                  * read the old inode data successfully.
2571                  */
2572                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2573                         set_buffer_uptodate(bh);
2574
2575                 if (buffer_uptodate(bh)) {
2576                         /* someone brought it uptodate while we waited */
2577                         unlock_buffer(bh);
2578                         goto has_buffer;
2579                 }
2580
2581                 /*
2582                  * If we have all information of the inode in memory and this
2583                  * is the only valid inode in the block, we need not read the
2584                  * block.
2585                  */
2586                 if (in_mem) {
2587                         struct buffer_head *bitmap_bh;
2588                         struct ext3_group_desc *desc;
2589                         int inodes_per_buffer;
2590                         int inode_offset, i;
2591                         int block_group;
2592                         int start;
2593
2594                         block_group = (inode->i_ino - 1) /
2595                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2596                         inodes_per_buffer = bh->b_size /
2597                                 EXT3_INODE_SIZE(inode->i_sb);
2598                         inode_offset = ((inode->i_ino - 1) %
2599                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2600                         start = inode_offset & ~(inodes_per_buffer - 1);
2601
2602                         /* Is the inode bitmap in cache? */
2603                         desc = ext3_get_group_desc(inode->i_sb,
2604                                                 block_group, NULL);
2605                         if (!desc)
2606                                 goto make_io;
2607
2608                         bitmap_bh = sb_getblk(inode->i_sb,
2609                                         le32_to_cpu(desc->bg_inode_bitmap));
2610                         if (!bitmap_bh)
2611                                 goto make_io;
2612
2613                         /*
2614                          * If the inode bitmap isn't in cache then the
2615                          * optimisation may end up performing two reads instead
2616                          * of one, so skip it.
2617                          */
2618                         if (!buffer_uptodate(bitmap_bh)) {
2619                                 brelse(bitmap_bh);
2620                                 goto make_io;
2621                         }
2622                         for (i = start; i < start + inodes_per_buffer; i++) {
2623                                 if (i == inode_offset)
2624                                         continue;
2625                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2626                                         break;
2627                         }
2628                         brelse(bitmap_bh);
2629                         if (i == start + inodes_per_buffer) {
2630                                 /* all other inodes are free, so skip I/O */
2631                                 memset(bh->b_data, 0, bh->b_size);
2632                                 set_buffer_uptodate(bh);
2633                                 unlock_buffer(bh);
2634                                 goto has_buffer;
2635                         }
2636                 }
2637
2638 make_io:
2639                 /*
2640                  * There are other valid inodes in the buffer, this inode
2641                  * has in-inode xattrs, or we don't have this inode in memory.
2642                  * Read the block from disk.
2643                  */
2644                 get_bh(bh);
2645                 bh->b_end_io = end_buffer_read_sync;
2646                 submit_bh(READ_META, bh);
2647                 wait_on_buffer(bh);
2648                 if (!buffer_uptodate(bh)) {
2649                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2650                                         "unable to read inode block - "
2651                                         "inode=%lu, block="E3FSBLK,
2652                                         inode->i_ino, block);
2653                         brelse(bh);
2654                         return -EIO;
2655                 }
2656         }
2657 has_buffer:
2658         iloc->bh = bh;
2659         return 0;
2660 }
2661
2662 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2663 {
2664         /* We have all inode data except xattrs in memory here. */
2665         return __ext3_get_inode_loc(inode, iloc,
2666                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2667 }
2668
2669 void ext3_set_inode_flags(struct inode *inode)
2670 {
2671         unsigned int flags = EXT3_I(inode)->i_flags;
2672
2673         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2674         if (flags & EXT3_SYNC_FL)
2675                 inode->i_flags |= S_SYNC;
2676         if (flags & EXT3_APPEND_FL)
2677                 inode->i_flags |= S_APPEND;
2678         if (flags & EXT3_IMMUTABLE_FL)
2679                 inode->i_flags |= S_IMMUTABLE;
2680         if (flags & EXT3_NOATIME_FL)
2681                 inode->i_flags |= S_NOATIME;
2682         if (flags & EXT3_DIRSYNC_FL)
2683                 inode->i_flags |= S_DIRSYNC;
2684 }
2685
2686 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2687 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2688 {
2689         unsigned int flags = ei->vfs_inode.i_flags;
2690
2691         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2692                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2693         if (flags & S_SYNC)
2694                 ei->i_flags |= EXT3_SYNC_FL;
2695         if (flags & S_APPEND)
2696                 ei->i_flags |= EXT3_APPEND_FL;
2697         if (flags & S_IMMUTABLE)
2698                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2699         if (flags & S_NOATIME)
2700                 ei->i_flags |= EXT3_NOATIME_FL;
2701         if (flags & S_DIRSYNC)
2702                 ei->i_flags |= EXT3_DIRSYNC_FL;
2703 }
2704
2705 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2706 {
2707         struct ext3_iloc iloc;
2708         struct ext3_inode *raw_inode;
2709         struct ext3_inode_info *ei;
2710         struct buffer_head *bh;
2711         struct inode *inode;
2712         long ret;
2713         int block;
2714
2715         inode = iget_locked(sb, ino);
2716         if (!inode)
2717                 return ERR_PTR(-ENOMEM);
2718         if (!(inode->i_state & I_NEW))
2719                 return inode;
2720
2721         ei = EXT3_I(inode);
2722 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2723         ei->i_acl = EXT3_ACL_NOT_CACHED;
2724         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2725 #endif
2726         ei->i_block_alloc_info = NULL;
2727
2728         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2729         if (ret < 0)
2730                 goto bad_inode;
2731         bh = iloc.bh;
2732         raw_inode = ext3_raw_inode(&iloc);
2733         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2734         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2735         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2736         if(!(test_opt (inode->i_sb, NO_UID32))) {
2737                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2738                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2739         }
2740         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2741         inode->i_size = le32_to_cpu(raw_inode->i_size);
2742         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2743         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2744         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2745         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2746
2747         ei->i_state = 0;
2748         ei->i_dir_start_lookup = 0;
2749         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2750         /* We now have enough fields to check if the inode was active or not.
2751          * This is needed because nfsd might try to access dead inodes
2752          * the test is that same one that e2fsck uses
2753          * NeilBrown 1999oct15
2754          */
2755         if (inode->i_nlink == 0) {
2756                 if (inode->i_mode == 0 ||
2757                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2758                         /* this inode is deleted */
2759                         brelse (bh);
2760                         ret = -ESTALE;
2761                         goto bad_inode;
2762                 }
2763                 /* The only unlinked inodes we let through here have
2764                  * valid i_mode and are being read by the orphan
2765                  * recovery code: that's fine, we're about to complete
2766                  * the process of deleting those. */
2767         }
2768         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2769         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2770 #ifdef EXT3_FRAGMENTS
2771         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2772         ei->i_frag_no = raw_inode->i_frag;
2773         ei->i_frag_size = raw_inode->i_fsize;
2774 #endif
2775         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2776         if (!S_ISREG(inode->i_mode)) {
2777                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2778         } else {
2779                 inode->i_size |=
2780                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2781         }
2782         ei->i_disksize = inode->i_size;
2783         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2784         ei->i_block_group = iloc.block_group;
2785         /*
2786          * NOTE! The in-memory inode i_data array is in little-endian order
2787          * even on big-endian machines: we do NOT byteswap the block numbers!
2788          */
2789         for (block = 0; block < EXT3_N_BLOCKS; block++)
2790                 ei->i_data[block] = raw_inode->i_block[block];
2791         INIT_LIST_HEAD(&ei->i_orphan);
2792
2793         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2794             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2795                 /*
2796                  * When mke2fs creates big inodes it does not zero out
2797                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2798                  * so ignore those first few inodes.
2799                  */
2800                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2801                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2802                     EXT3_INODE_SIZE(inode->i_sb)) {
2803                         brelse (bh);
2804                         ret = -EIO;
2805                         goto bad_inode;
2806                 }
2807                 if (ei->i_extra_isize == 0) {
2808                         /* The extra space is currently unused. Use it. */
2809                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2810                                             EXT3_GOOD_OLD_INODE_SIZE;
2811                 } else {
2812                         __le32 *magic = (void *)raw_inode +
2813                                         EXT3_GOOD_OLD_INODE_SIZE +
2814                                         ei->i_extra_isize;
2815                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2816                                  ei->i_state |= EXT3_STATE_XATTR;
2817                 }
2818         } else
2819                 ei->i_extra_isize = 0;
2820
2821         if (S_ISREG(inode->i_mode)) {
2822                 inode->i_op = &ext3_file_inode_operations;
2823                 inode->i_fop = &ext3_file_operations;
2824                 ext3_set_aops(inode);
2825         } else if (S_ISDIR(inode->i_mode)) {
2826                 inode->i_op = &ext3_dir_inode_operations;
2827                 inode->i_fop = &ext3_dir_operations;
2828         } else if (S_ISLNK(inode->i_mode)) {
2829                 if (ext3_inode_is_fast_symlink(inode)) {
2830                         inode->i_op = &ext3_fast_symlink_inode_operations;
2831                         nd_terminate_link(ei->i_data, inode->i_size,
2832                                 sizeof(ei->i_data) - 1);
2833                 } else {
2834                         inode->i_op = &ext3_symlink_inode_operations;
2835                         ext3_set_aops(inode);
2836                 }
2837         } else {
2838                 inode->i_op = &ext3_special_inode_operations;
2839                 if (raw_inode->i_block[0])
2840                         init_special_inode(inode, inode->i_mode,
2841                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2842                 else
2843                         init_special_inode(inode, inode->i_mode,
2844                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2845         }
2846         brelse (iloc.bh);
2847         ext3_set_inode_flags(inode);
2848         unlock_new_inode(inode);
2849         return inode;
2850
2851 bad_inode:
2852         iget_failed(inode);
2853         return ERR_PTR(ret);
2854 }
2855
2856 /*
2857  * Post the struct inode info into an on-disk inode location in the
2858  * buffer-cache.  This gobbles the caller's reference to the
2859  * buffer_head in the inode location struct.
2860  *
2861  * The caller must have write access to iloc->bh.
2862  */
2863 static int ext3_do_update_inode(handle_t *handle,
2864                                 struct inode *inode,
2865                                 struct ext3_iloc *iloc)
2866 {
2867         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2868         struct ext3_inode_info *ei = EXT3_I(inode);
2869         struct buffer_head *bh = iloc->bh;
2870         int err = 0, rc, block;
2871
2872         /* For fields not not tracking in the in-memory inode,
2873          * initialise them to zero for new inodes. */
2874         if (ei->i_state & EXT3_STATE_NEW)
2875                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2876
2877         ext3_get_inode_flags(ei);
2878         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2879         if(!(test_opt(inode->i_sb, NO_UID32))) {
2880                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2881                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2882 /*
2883  * Fix up interoperability with old kernels. Otherwise, old inodes get
2884  * re-used with the upper 16 bits of the uid/gid intact
2885  */
2886                 if(!ei->i_dtime) {
2887                         raw_inode->i_uid_high =
2888                                 cpu_to_le16(high_16_bits(inode->i_uid));
2889                         raw_inode->i_gid_high =
2890                                 cpu_to_le16(high_16_bits(inode->i_gid));
2891                 } else {
2892                         raw_inode->i_uid_high = 0;
2893                         raw_inode->i_gid_high = 0;
2894                 }
2895         } else {
2896                 raw_inode->i_uid_low =
2897                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2898                 raw_inode->i_gid_low =
2899                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2900                 raw_inode->i_uid_high = 0;
2901                 raw_inode->i_gid_high = 0;
2902         }
2903         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2904         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2905         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2906         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2907         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2908         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2909         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2910         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2911 #ifdef EXT3_FRAGMENTS
2912         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2913         raw_inode->i_frag = ei->i_frag_no;
2914         raw_inode->i_fsize = ei->i_frag_size;
2915 #endif
2916         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2917         if (!S_ISREG(inode->i_mode)) {
2918                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2919         } else {
2920                 raw_inode->i_size_high =
2921                         cpu_to_le32(ei->i_disksize >> 32);
2922                 if (ei->i_disksize > 0x7fffffffULL) {
2923                         struct super_block *sb = inode->i_sb;
2924                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2925                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2926                             EXT3_SB(sb)->s_es->s_rev_level ==
2927                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2928                                /* If this is the first large file
2929                                 * created, add a flag to the superblock.
2930                                 */
2931                                 err = ext3_journal_get_write_access(handle,
2932                                                 EXT3_SB(sb)->s_sbh);
2933                                 if (err)
2934                                         goto out_brelse;
2935                                 ext3_update_dynamic_rev(sb);
2936                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2937                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2938                                 sb->s_dirt = 1;
2939                                 handle->h_sync = 1;
2940                                 err = ext3_journal_dirty_metadata(handle,
2941                                                 EXT3_SB(sb)->s_sbh);
2942                         }
2943                 }
2944         }
2945         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2946         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2947                 if (old_valid_dev(inode->i_rdev)) {
2948                         raw_inode->i_block[0] =
2949                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2950                         raw_inode->i_block[1] = 0;
2951                 } else {
2952                         raw_inode->i_block[0] = 0;
2953                         raw_inode->i_block[1] =
2954                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2955                         raw_inode->i_block[2] = 0;
2956                 }
2957         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2958                 raw_inode->i_block[block] = ei->i_data[block];
2959
2960         if (ei->i_extra_isize)
2961                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2962
2963         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2964         rc = ext3_journal_dirty_metadata(handle, bh);
2965         if (!err)
2966                 err = rc;
2967         ei->i_state &= ~EXT3_STATE_NEW;
2968
2969 out_brelse:
2970         brelse (bh);
2971         ext3_std_error(inode->i_sb, err);
2972         return err;
2973 }
2974
2975 /*
2976  * ext3_write_inode()
2977  *
2978  * We are called from a few places:
2979  *
2980  * - Within generic_file_write() for O_SYNC files.
2981  *   Here, there will be no transaction running. We wait for any running
2982  *   trasnaction to commit.
2983  *
2984  * - Within sys_sync(), kupdate and such.
2985  *   We wait on commit, if tol to.
2986  *
2987  * - Within prune_icache() (PF_MEMALLOC == true)
2988  *   Here we simply return.  We can't afford to block kswapd on the
2989  *   journal commit.
2990  *
2991  * In all cases it is actually safe for us to return without doing anything,
2992  * because the inode has been copied into a raw inode buffer in
2993  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2994  * knfsd.
2995  *
2996  * Note that we are absolutely dependent upon all inode dirtiers doing the
2997  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2998  * which we are interested.
2999  *
3000  * It would be a bug for them to not do this.  The code:
3001  *
3002  *      mark_inode_dirty(inode)
3003  *      stuff();
3004  *      inode->i_size = expr;
3005  *
3006  * is in error because a kswapd-driven write_inode() could occur while
3007  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3008  * will no longer be on the superblock's dirty inode list.
3009  */
3010 int ext3_write_inode(struct inode *inode, int wait)
3011 {
3012         if (current->flags & PF_MEMALLOC)
3013                 return 0;
3014
3015         if (ext3_journal_current_handle()) {
3016                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3017                 dump_stack();
3018                 return -EIO;
3019         }
3020
3021         if (!wait)
3022                 return 0;
3023
3024         return ext3_force_commit(inode->i_sb);
3025 }
3026
3027 /*
3028  * ext3_setattr()
3029  *
3030  * Called from notify_change.
3031  *
3032  * We want to trap VFS attempts to truncate the file as soon as
3033  * possible.  In particular, we want to make sure that when the VFS
3034  * shrinks i_size, we put the inode on the orphan list and modify
3035  * i_disksize immediately, so that during the subsequent flushing of
3036  * dirty pages and freeing of disk blocks, we can guarantee that any
3037  * commit will leave the blocks being flushed in an unused state on
3038  * disk.  (On recovery, the inode will get truncated and the blocks will
3039  * be freed, so we have a strong guarantee that no future commit will
3040  * leave these blocks visible to the user.)
3041  *
3042  * Called with inode->sem down.
3043  */
3044 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3045 {
3046         struct inode *inode = dentry->d_inode;
3047         int error, rc = 0;
3048         const unsigned int ia_valid = attr->ia_valid;
3049
3050         error = inode_change_ok(inode, attr);
3051         if (error)
3052                 return error;
3053
3054         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3055                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3056                 handle_t *handle;
3057
3058                 /* (user+group)*(old+new) structure, inode write (sb,
3059                  * inode block, ? - but truncate inode update has it) */
3060                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3061                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3062                 if (IS_ERR(handle)) {
3063                         error = PTR_ERR(handle);
3064                         goto err_out;
3065                 }
3066                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
3067                 if (error) {
3068                         ext3_journal_stop(handle);
3069                         return error;
3070                 }
3071                 /* Update corresponding info in inode so that everything is in
3072                  * one transaction */
3073                 if (attr->ia_valid & ATTR_UID)
3074                         inode->i_uid = attr->ia_uid;
3075                 if (attr->ia_valid & ATTR_GID)
3076                         inode->i_gid = attr->ia_gid;
3077                 error = ext3_mark_inode_dirty(handle, inode);
3078                 ext3_journal_stop(handle);
3079         }
3080
3081         if (S_ISREG(inode->i_mode) &&
3082             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3083                 handle_t *handle;
3084
3085                 handle = ext3_journal_start(inode, 3);
3086                 if (IS_ERR(handle)) {
3087                         error = PTR_ERR(handle);
3088                         goto err_out;
3089                 }
3090
3091                 error = ext3_orphan_add(handle, inode);
3092                 EXT3_I(inode)->i_disksize = attr->ia_size;
3093                 rc = ext3_mark_inode_dirty(handle, inode);
3094                 if (!error)
3095                         error = rc;
3096                 ext3_journal_stop(handle);
3097         }
3098
3099         rc = inode_setattr(inode, attr);
3100
3101         /* If inode_setattr's call to ext3_truncate failed to get a
3102          * transaction handle at all, we need to clean up the in-core
3103          * orphan list manually. */
3104         if (inode->i_nlink)
3105                 ext3_orphan_del(NULL, inode);
3106
3107         if (!rc && (ia_valid & ATTR_MODE))
3108                 rc = ext3_acl_chmod(inode);
3109
3110 err_out:
3111         ext3_std_error(inode->i_sb, error);
3112         if (!error)
3113                 error = rc;
3114         return error;
3115 }
3116
3117
3118 /*
3119  * How many blocks doth make a writepage()?
3120  *
3121  * With N blocks per page, it may be:
3122  * N data blocks
3123  * 2 indirect block
3124  * 2 dindirect
3125  * 1 tindirect
3126  * N+5 bitmap blocks (from the above)
3127  * N+5 group descriptor summary blocks
3128  * 1 inode block
3129  * 1 superblock.
3130  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3131  *
3132  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3133  *
3134  * With ordered or writeback data it's the same, less the N data blocks.
3135  *
3136  * If the inode's direct blocks can hold an integral number of pages then a
3137  * page cannot straddle two indirect blocks, and we can only touch one indirect
3138  * and dindirect block, and the "5" above becomes "3".
3139  *
3140  * This still overestimates under most circumstances.  If we were to pass the
3141  * start and end offsets in here as well we could do block_to_path() on each
3142  * block and work out the exact number of indirects which are touched.  Pah.
3143  */
3144
3145 static int ext3_writepage_trans_blocks(struct inode *inode)
3146 {
3147         int bpp = ext3_journal_blocks_per_page(inode);
3148         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3149         int ret;
3150
3151         if (ext3_should_journal_data(inode))
3152                 ret = 3 * (bpp + indirects) + 2;
3153         else
3154                 ret = 2 * (bpp + indirects) + 2;
3155
3156 #ifdef CONFIG_QUOTA
3157         /* We know that structure was already allocated during vfs_dq_init so
3158          * we will be updating only the data blocks + inodes */
3159         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3160 #endif
3161
3162         return ret;
3163 }
3164
3165 /*
3166  * The caller must have previously called ext3_reserve_inode_write().
3167  * Give this, we know that the caller already has write access to iloc->bh.
3168  */
3169 int ext3_mark_iloc_dirty(handle_t *handle,
3170                 struct inode *inode, struct ext3_iloc *iloc)
3171 {
3172         int err = 0;
3173
3174         /* the do_update_inode consumes one bh->b_count */
3175         get_bh(iloc->bh);
3176
3177         /* ext3_do_update_inode() does journal_dirty_metadata */
3178         err = ext3_do_update_inode(handle, inode, iloc);
3179         put_bh(iloc->bh);
3180         return err;
3181 }
3182
3183 /*
3184  * On success, We end up with an outstanding reference count against
3185  * iloc->bh.  This _must_ be cleaned up later.
3186  */
3187
3188 int
3189 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3190                          struct ext3_iloc *iloc)
3191 {
3192         int err = 0;
3193         if (handle) {
3194                 err = ext3_get_inode_loc(inode, iloc);
3195                 if (!err) {
3196                         BUFFER_TRACE(iloc->bh, "get_write_access");
3197                         err = ext3_journal_get_write_access(handle, iloc->bh);
3198                         if (err) {
3199                                 brelse(iloc->bh);
3200                                 iloc->bh = NULL;
3201                         }
3202                 }
3203         }
3204         ext3_std_error(inode->i_sb, err);
3205         return err;
3206 }
3207
3208 /*
3209  * What we do here is to mark the in-core inode as clean with respect to inode
3210  * dirtiness (it may still be data-dirty).
3211  * This means that the in-core inode may be reaped by prune_icache
3212  * without having to perform any I/O.  This is a very good thing,
3213  * because *any* task may call prune_icache - even ones which
3214  * have a transaction open against a different journal.
3215  *
3216  * Is this cheating?  Not really.  Sure, we haven't written the
3217  * inode out, but prune_icache isn't a user-visible syncing function.
3218  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3219  * we start and wait on commits.
3220  *
3221  * Is this efficient/effective?  Well, we're being nice to the system
3222  * by cleaning up our inodes proactively so they can be reaped
3223  * without I/O.  But we are potentially leaving up to five seconds'
3224  * worth of inodes floating about which prune_icache wants us to
3225  * write out.  One way to fix that would be to get prune_icache()
3226  * to do a write_super() to free up some memory.  It has the desired
3227  * effect.
3228  */
3229 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3230 {
3231         struct ext3_iloc iloc;
3232         int err;
3233
3234         might_sleep();
3235         err = ext3_reserve_inode_write(handle, inode, &iloc);
3236         if (!err)
3237                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3238         return err;
3239 }
3240
3241 /*
3242  * ext3_dirty_inode() is called from __mark_inode_dirty()
3243  *
3244  * We're really interested in the case where a file is being extended.
3245  * i_size has been changed by generic_commit_write() and we thus need
3246  * to include the updated inode in the current transaction.
3247  *
3248  * Also, vfs_dq_alloc_space() will always dirty the inode when blocks
3249  * are allocated to the file.
3250  *
3251  * If the inode is marked synchronous, we don't honour that here - doing
3252  * so would cause a commit on atime updates, which we don't bother doing.
3253  * We handle synchronous inodes at the highest possible level.
3254  */
3255 void ext3_dirty_inode(struct inode *inode)
3256 {
3257         handle_t *current_handle = ext3_journal_current_handle();
3258         handle_t *handle;
3259
3260         handle = ext3_journal_start(inode, 2);
3261         if (IS_ERR(handle))
3262                 goto out;
3263         if (current_handle &&
3264                 current_handle->h_transaction != handle->h_transaction) {
3265                 /* This task has a transaction open against a different fs */
3266                 printk(KERN_EMERG "%s: transactions do not match!\n",
3267                        __func__);
3268         } else {
3269                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3270                                 current_handle);
3271                 ext3_mark_inode_dirty(handle, inode);
3272         }
3273         ext3_journal_stop(handle);
3274 out:
3275         return;
3276 }
3277
3278 #if 0
3279 /*
3280  * Bind an inode's backing buffer_head into this transaction, to prevent
3281  * it from being flushed to disk early.  Unlike
3282  * ext3_reserve_inode_write, this leaves behind no bh reference and
3283  * returns no iloc structure, so the caller needs to repeat the iloc
3284  * lookup to mark the inode dirty later.
3285  */
3286 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3287 {
3288         struct ext3_iloc iloc;
3289
3290         int err = 0;
3291         if (handle) {
3292                 err = ext3_get_inode_loc(inode, &iloc);
3293                 if (!err) {
3294                         BUFFER_TRACE(iloc.bh, "get_write_access");
3295                         err = journal_get_write_access(handle, iloc.bh);
3296                         if (!err)
3297                                 err = ext3_journal_dirty_metadata(handle,
3298                                                                   iloc.bh);
3299                         brelse(iloc.bh);
3300                 }
3301         }
3302         ext3_std_error(inode->i_sb, err);
3303         return err;
3304 }
3305 #endif
3306
3307 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3308 {
3309         journal_t *journal;
3310         handle_t *handle;
3311         int err;
3312
3313         /*
3314          * We have to be very careful here: changing a data block's
3315          * journaling status dynamically is dangerous.  If we write a
3316          * data block to the journal, change the status and then delete
3317          * that block, we risk forgetting to revoke the old log record
3318          * from the journal and so a subsequent replay can corrupt data.
3319          * So, first we make sure that the journal is empty and that
3320          * nobody is changing anything.
3321          */
3322
3323         journal = EXT3_JOURNAL(inode);
3324         if (is_journal_aborted(journal))
3325                 return -EROFS;
3326
3327         journal_lock_updates(journal);
3328         journal_flush(journal);
3329
3330         /*
3331          * OK, there are no updates running now, and all cached data is
3332          * synced to disk.  We are now in a completely consistent state
3333          * which doesn't have anything in the journal, and we know that
3334          * no filesystem updates are running, so it is safe to modify
3335          * the inode's in-core data-journaling state flag now.
3336          */
3337
3338         if (val)
3339                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3340         else
3341                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3342         ext3_set_aops(inode);
3343
3344         journal_unlock_updates(journal);
3345
3346         /* Finally we can mark the inode as dirty. */
3347
3348         handle = ext3_journal_start(inode, 1);
3349         if (IS_ERR(handle))
3350                 return PTR_ERR(handle);
3351
3352         err = ext3_mark_inode_dirty(handle, inode);
3353         handle->h_sync = 1;
3354         ext3_journal_stop(handle);
3355         ext3_std_error(inode->i_sb, err);
3356
3357         return err;
3358 }