Merge branch 'ext3-latency-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/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;
1153         handle_t *handle;
1154         int retries = 0;
1155         struct page *page;
1156         pgoff_t index;
1157         unsigned from, to;
1158         /* Reserve one block more for addition to orphan list in case
1159          * we allocate blocks but write fails for some reason */
1160         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1161
1162         index = pos >> PAGE_CACHE_SHIFT;
1163         from = pos & (PAGE_CACHE_SIZE - 1);
1164         to = from + len;
1165
1166 retry:
1167         page = grab_cache_page_write_begin(mapping, index, flags);
1168         if (!page)
1169                 return -ENOMEM;
1170         *pagep = page;
1171
1172         handle = ext3_journal_start(inode, needed_blocks);
1173         if (IS_ERR(handle)) {
1174                 unlock_page(page);
1175                 page_cache_release(page);
1176                 ret = PTR_ERR(handle);
1177                 goto out;
1178         }
1179         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1180                                                         ext3_get_block);
1181         if (ret)
1182                 goto write_begin_failed;
1183
1184         if (ext3_should_journal_data(inode)) {
1185                 ret = walk_page_buffers(handle, page_buffers(page),
1186                                 from, to, NULL, do_journal_get_write_access);
1187         }
1188 write_begin_failed:
1189         if (ret) {
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                  * Add inode to orphan list in case we crash before truncate
1196                  * finishes.
1197                  */
1198                 if (pos + len > inode->i_size)
1199                         ext3_orphan_add(handle, inode);
1200                 ext3_journal_stop(handle);
1201                 unlock_page(page);
1202                 page_cache_release(page);
1203                 if (pos + len > inode->i_size)
1204                         vmtruncate(inode, inode->i_size);
1205         }
1206         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1207                 goto retry;
1208 out:
1209         return ret;
1210 }
1211
1212
1213 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1214 {
1215         int err = journal_dirty_data(handle, bh);
1216         if (err)
1217                 ext3_journal_abort_handle(__func__, __func__,
1218                                                 bh, handle, err);
1219         return err;
1220 }
1221
1222 /* For ordered writepage and write_end functions */
1223 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1224 {
1225         /*
1226          * Write could have mapped the buffer but it didn't copy the data in
1227          * yet. So avoid filing such buffer into a transaction.
1228          */
1229         if (buffer_mapped(bh) && buffer_uptodate(bh))
1230                 return ext3_journal_dirty_data(handle, bh);
1231         return 0;
1232 }
1233
1234 /* For write_end() in data=journal mode */
1235 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1236 {
1237         if (!buffer_mapped(bh) || buffer_freed(bh))
1238                 return 0;
1239         set_buffer_uptodate(bh);
1240         return ext3_journal_dirty_metadata(handle, bh);
1241 }
1242
1243 /*
1244  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1245  * for the whole page but later we failed to copy the data in. Update inode
1246  * size according to what we managed to copy. The rest is going to be
1247  * truncated in write_end function.
1248  */
1249 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1250 {
1251         /* What matters to us is i_disksize. We don't write i_size anywhere */
1252         if (pos + copied > inode->i_size)
1253                 i_size_write(inode, pos + copied);
1254         if (pos + copied > EXT3_I(inode)->i_disksize) {
1255                 EXT3_I(inode)->i_disksize = pos + copied;
1256                 mark_inode_dirty(inode);
1257         }
1258 }
1259
1260 /*
1261  * We need to pick up the new inode size which generic_commit_write gave us
1262  * `file' can be NULL - eg, when called from page_symlink().
1263  *
1264  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1265  * buffers are managed internally.
1266  */
1267 static int ext3_ordered_write_end(struct file *file,
1268                                 struct address_space *mapping,
1269                                 loff_t pos, unsigned len, unsigned copied,
1270                                 struct page *page, void *fsdata)
1271 {
1272         handle_t *handle = ext3_journal_current_handle();
1273         struct inode *inode = file->f_mapping->host;
1274         unsigned from, to;
1275         int ret = 0, ret2;
1276
1277         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1278
1279         from = pos & (PAGE_CACHE_SIZE - 1);
1280         to = from + copied;
1281         ret = walk_page_buffers(handle, page_buffers(page),
1282                 from, to, NULL, journal_dirty_data_fn);
1283
1284         if (ret == 0)
1285                 update_file_sizes(inode, pos, copied);
1286         /*
1287          * There may be allocated blocks outside of i_size because
1288          * we failed to copy some data. Prepare for truncate.
1289          */
1290         if (pos + len > inode->i_size)
1291                 ext3_orphan_add(handle, inode);
1292         ret2 = ext3_journal_stop(handle);
1293         if (!ret)
1294                 ret = ret2;
1295         unlock_page(page);
1296         page_cache_release(page);
1297
1298         if (pos + len > inode->i_size)
1299                 vmtruncate(inode, inode->i_size);
1300         return ret ? ret : copied;
1301 }
1302
1303 static int ext3_writeback_write_end(struct file *file,
1304                                 struct address_space *mapping,
1305                                 loff_t pos, unsigned len, unsigned copied,
1306                                 struct page *page, void *fsdata)
1307 {
1308         handle_t *handle = ext3_journal_current_handle();
1309         struct inode *inode = file->f_mapping->host;
1310         int ret;
1311
1312         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1313         update_file_sizes(inode, pos, copied);
1314         /*
1315          * There may be allocated blocks outside of i_size because
1316          * we failed to copy some data. Prepare for truncate.
1317          */
1318         if (pos + len > inode->i_size)
1319                 ext3_orphan_add(handle, inode);
1320         ret = ext3_journal_stop(handle);
1321         unlock_page(page);
1322         page_cache_release(page);
1323
1324         if (pos + len > inode->i_size)
1325                 vmtruncate(inode, inode->i_size);
1326         return ret ? ret : copied;
1327 }
1328
1329 static int ext3_journalled_write_end(struct file *file,
1330                                 struct address_space *mapping,
1331                                 loff_t pos, unsigned len, unsigned copied,
1332                                 struct page *page, void *fsdata)
1333 {
1334         handle_t *handle = ext3_journal_current_handle();
1335         struct inode *inode = mapping->host;
1336         int ret = 0, ret2;
1337         int partial = 0;
1338         unsigned from, to;
1339
1340         from = pos & (PAGE_CACHE_SIZE - 1);
1341         to = from + len;
1342
1343         if (copied < len) {
1344                 if (!PageUptodate(page))
1345                         copied = 0;
1346                 page_zero_new_buffers(page, from + copied, to);
1347                 to = from + copied;
1348         }
1349
1350         ret = walk_page_buffers(handle, page_buffers(page), from,
1351                                 to, &partial, write_end_fn);
1352         if (!partial)
1353                 SetPageUptodate(page);
1354
1355         if (pos + copied > inode->i_size)
1356                 i_size_write(inode, pos + copied);
1357         /*
1358          * There may be allocated blocks outside of i_size because
1359          * we failed to copy some data. Prepare for truncate.
1360          */
1361         if (pos + len > inode->i_size)
1362                 ext3_orphan_add(handle, inode);
1363         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1364         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1365                 EXT3_I(inode)->i_disksize = inode->i_size;
1366                 ret2 = ext3_mark_inode_dirty(handle, inode);
1367                 if (!ret)
1368                         ret = ret2;
1369         }
1370
1371         ret2 = ext3_journal_stop(handle);
1372         if (!ret)
1373                 ret = ret2;
1374         unlock_page(page);
1375         page_cache_release(page);
1376
1377         if (pos + len > inode->i_size)
1378                 vmtruncate(inode, inode->i_size);
1379         return ret ? ret : copied;
1380 }
1381
1382 /*
1383  * bmap() is special.  It gets used by applications such as lilo and by
1384  * the swapper to find the on-disk block of a specific piece of data.
1385  *
1386  * Naturally, this is dangerous if the block concerned is still in the
1387  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1388  * filesystem and enables swap, then they may get a nasty shock when the
1389  * data getting swapped to that swapfile suddenly gets overwritten by
1390  * the original zero's written out previously to the journal and
1391  * awaiting writeback in the kernel's buffer cache.
1392  *
1393  * So, if we see any bmap calls here on a modified, data-journaled file,
1394  * take extra steps to flush any blocks which might be in the cache.
1395  */
1396 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1397 {
1398         struct inode *inode = mapping->host;
1399         journal_t *journal;
1400         int err;
1401
1402         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1403                 /*
1404                  * This is a REALLY heavyweight approach, but the use of
1405                  * bmap on dirty files is expected to be extremely rare:
1406                  * only if we run lilo or swapon on a freshly made file
1407                  * do we expect this to happen.
1408                  *
1409                  * (bmap requires CAP_SYS_RAWIO so this does not
1410                  * represent an unprivileged user DOS attack --- we'd be
1411                  * in trouble if mortal users could trigger this path at
1412                  * will.)
1413                  *
1414                  * NB. EXT3_STATE_JDATA is not set on files other than
1415                  * regular files.  If somebody wants to bmap a directory
1416                  * or symlink and gets confused because the buffer
1417                  * hasn't yet been flushed to disk, they deserve
1418                  * everything they get.
1419                  */
1420
1421                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1422                 journal = EXT3_JOURNAL(inode);
1423                 journal_lock_updates(journal);
1424                 err = journal_flush(journal);
1425                 journal_unlock_updates(journal);
1426
1427                 if (err)
1428                         return 0;
1429         }
1430
1431         return generic_block_bmap(mapping,block,ext3_get_block);
1432 }
1433
1434 static int bget_one(handle_t *handle, struct buffer_head *bh)
1435 {
1436         get_bh(bh);
1437         return 0;
1438 }
1439
1440 static int bput_one(handle_t *handle, struct buffer_head *bh)
1441 {
1442         put_bh(bh);
1443         return 0;
1444 }
1445
1446 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1447 {
1448         return !buffer_mapped(bh);
1449 }
1450
1451 /*
1452  * Note that we always start a transaction even if we're not journalling
1453  * data.  This is to preserve ordering: any hole instantiation within
1454  * __block_write_full_page -> ext3_get_block() should be journalled
1455  * along with the data so we don't crash and then get metadata which
1456  * refers to old data.
1457  *
1458  * In all journalling modes block_write_full_page() will start the I/O.
1459  *
1460  * Problem:
1461  *
1462  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1463  *              ext3_writepage()
1464  *
1465  * Similar for:
1466  *
1467  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1468  *
1469  * Same applies to ext3_get_block().  We will deadlock on various things like
1470  * lock_journal and i_truncate_mutex.
1471  *
1472  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1473  * allocations fail.
1474  *
1475  * 16May01: If we're reentered then journal_current_handle() will be
1476  *          non-zero. We simply *return*.
1477  *
1478  * 1 July 2001: @@@ FIXME:
1479  *   In journalled data mode, a data buffer may be metadata against the
1480  *   current transaction.  But the same file is part of a shared mapping
1481  *   and someone does a writepage() on it.
1482  *
1483  *   We will move the buffer onto the async_data list, but *after* it has
1484  *   been dirtied. So there's a small window where we have dirty data on
1485  *   BJ_Metadata.
1486  *
1487  *   Note that this only applies to the last partial page in the file.  The
1488  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1489  *   broken code anyway: it's wrong for msync()).
1490  *
1491  *   It's a rare case: affects the final partial page, for journalled data
1492  *   where the file is subject to bith write() and writepage() in the same
1493  *   transction.  To fix it we'll need a custom block_write_full_page().
1494  *   We'll probably need that anyway for journalling writepage() output.
1495  *
1496  * We don't honour synchronous mounts for writepage().  That would be
1497  * disastrous.  Any write() or metadata operation will sync the fs for
1498  * us.
1499  *
1500  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1501  * we don't need to open a transaction here.
1502  */
1503 static int ext3_ordered_writepage(struct page *page,
1504                                 struct writeback_control *wbc)
1505 {
1506         struct inode *inode = page->mapping->host;
1507         struct buffer_head *page_bufs;
1508         handle_t *handle = NULL;
1509         int ret = 0;
1510         int err;
1511
1512         J_ASSERT(PageLocked(page));
1513
1514         /*
1515          * We give up here if we're reentered, because it might be for a
1516          * different filesystem.
1517          */
1518         if (ext3_journal_current_handle())
1519                 goto out_fail;
1520
1521         if (!page_has_buffers(page)) {
1522                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1523                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1524         } else if (!walk_page_buffers(NULL, page_buffers(page), 0, PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1525                 /* Provide NULL instead of get_block so that we catch bugs if buffers weren't really mapped */
1526                 return block_write_full_page(page, NULL, wbc);
1527         }
1528         page_bufs = page_buffers(page);
1529
1530         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1531
1532         if (IS_ERR(handle)) {
1533                 ret = PTR_ERR(handle);
1534                 goto out_fail;
1535         }
1536
1537         walk_page_buffers(handle, page_bufs, 0,
1538                         PAGE_CACHE_SIZE, NULL, bget_one);
1539
1540         ret = block_write_full_page(page, ext3_get_block, wbc);
1541
1542         /*
1543          * The page can become unlocked at any point now, and
1544          * truncate can then come in and change things.  So we
1545          * can't touch *page from now on.  But *page_bufs is
1546          * safe due to elevated refcount.
1547          */
1548
1549         /*
1550          * And attach them to the current transaction.  But only if
1551          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1552          * and generally junk.
1553          */
1554         if (ret == 0) {
1555                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1556                                         NULL, journal_dirty_data_fn);
1557                 if (!ret)
1558                         ret = err;
1559         }
1560         walk_page_buffers(handle, page_bufs, 0,
1561                         PAGE_CACHE_SIZE, NULL, bput_one);
1562         err = ext3_journal_stop(handle);
1563         if (!ret)
1564                 ret = err;
1565         return ret;
1566
1567 out_fail:
1568         redirty_page_for_writepage(wbc, page);
1569         unlock_page(page);
1570         return ret;
1571 }
1572
1573 static int ext3_writeback_writepage(struct page *page,
1574                                 struct writeback_control *wbc)
1575 {
1576         struct inode *inode = page->mapping->host;
1577         handle_t *handle = NULL;
1578         int ret = 0;
1579         int err;
1580
1581         if (ext3_journal_current_handle())
1582                 goto out_fail;
1583
1584         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1585         if (IS_ERR(handle)) {
1586                 ret = PTR_ERR(handle);
1587                 goto out_fail;
1588         }
1589
1590         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1591                 ret = nobh_writepage(page, ext3_get_block, wbc);
1592         else
1593                 ret = block_write_full_page(page, ext3_get_block, wbc);
1594
1595         err = ext3_journal_stop(handle);
1596         if (!ret)
1597                 ret = err;
1598         return ret;
1599
1600 out_fail:
1601         redirty_page_for_writepage(wbc, page);
1602         unlock_page(page);
1603         return ret;
1604 }
1605
1606 static int ext3_journalled_writepage(struct page *page,
1607                                 struct writeback_control *wbc)
1608 {
1609         struct inode *inode = page->mapping->host;
1610         handle_t *handle = NULL;
1611         int ret = 0;
1612         int err;
1613
1614         if (ext3_journal_current_handle())
1615                 goto no_write;
1616
1617         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1618         if (IS_ERR(handle)) {
1619                 ret = PTR_ERR(handle);
1620                 goto no_write;
1621         }
1622
1623         if (!page_has_buffers(page) || PageChecked(page)) {
1624                 /*
1625                  * It's mmapped pagecache.  Add buffers and journal it.  There
1626                  * doesn't seem much point in redirtying the page here.
1627                  */
1628                 ClearPageChecked(page);
1629                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1630                                         ext3_get_block);
1631                 if (ret != 0) {
1632                         ext3_journal_stop(handle);
1633                         goto out_unlock;
1634                 }
1635                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1636                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1637
1638                 err = walk_page_buffers(handle, page_buffers(page), 0,
1639                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1640                 if (ret == 0)
1641                         ret = err;
1642                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1643                 unlock_page(page);
1644         } else {
1645                 /*
1646                  * It may be a page full of checkpoint-mode buffers.  We don't
1647                  * really know unless we go poke around in the buffer_heads.
1648                  * But block_write_full_page will do the right thing.
1649                  */
1650                 ret = block_write_full_page(page, ext3_get_block, wbc);
1651         }
1652         err = ext3_journal_stop(handle);
1653         if (!ret)
1654                 ret = err;
1655 out:
1656         return ret;
1657
1658 no_write:
1659         redirty_page_for_writepage(wbc, page);
1660 out_unlock:
1661         unlock_page(page);
1662         goto out;
1663 }
1664
1665 static int ext3_readpage(struct file *file, struct page *page)
1666 {
1667         return mpage_readpage(page, ext3_get_block);
1668 }
1669
1670 static int
1671 ext3_readpages(struct file *file, struct address_space *mapping,
1672                 struct list_head *pages, unsigned nr_pages)
1673 {
1674         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1675 }
1676
1677 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1678 {
1679         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1680
1681         /*
1682          * If it's a full truncate we just forget about the pending dirtying
1683          */
1684         if (offset == 0)
1685                 ClearPageChecked(page);
1686
1687         journal_invalidatepage(journal, page, offset);
1688 }
1689
1690 static int ext3_releasepage(struct page *page, gfp_t wait)
1691 {
1692         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1693
1694         WARN_ON(PageChecked(page));
1695         if (!page_has_buffers(page))
1696                 return 0;
1697         return journal_try_to_free_buffers(journal, page, wait);
1698 }
1699
1700 /*
1701  * If the O_DIRECT write will extend the file then add this inode to the
1702  * orphan list.  So recovery will truncate it back to the original size
1703  * if the machine crashes during the write.
1704  *
1705  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1706  * crashes then stale disk data _may_ be exposed inside the file. But current
1707  * VFS code falls back into buffered path in that case so we are safe.
1708  */
1709 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1710                         const struct iovec *iov, loff_t offset,
1711                         unsigned long nr_segs)
1712 {
1713         struct file *file = iocb->ki_filp;
1714         struct inode *inode = file->f_mapping->host;
1715         struct ext3_inode_info *ei = EXT3_I(inode);
1716         handle_t *handle;
1717         ssize_t ret;
1718         int orphan = 0;
1719         size_t count = iov_length(iov, nr_segs);
1720
1721         if (rw == WRITE) {
1722                 loff_t final_size = offset + count;
1723
1724                 if (final_size > inode->i_size) {
1725                         /* Credits for sb + inode write */
1726                         handle = ext3_journal_start(inode, 2);
1727                         if (IS_ERR(handle)) {
1728                                 ret = PTR_ERR(handle);
1729                                 goto out;
1730                         }
1731                         ret = ext3_orphan_add(handle, inode);
1732                         if (ret) {
1733                                 ext3_journal_stop(handle);
1734                                 goto out;
1735                         }
1736                         orphan = 1;
1737                         ei->i_disksize = inode->i_size;
1738                         ext3_journal_stop(handle);
1739                 }
1740         }
1741
1742         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1743                                  offset, nr_segs,
1744                                  ext3_get_block, NULL);
1745
1746         if (orphan) {
1747                 int err;
1748
1749                 /* Credits for sb + inode write */
1750                 handle = ext3_journal_start(inode, 2);
1751                 if (IS_ERR(handle)) {
1752                         /* This is really bad luck. We've written the data
1753                          * but cannot extend i_size. Bail out and pretend
1754                          * the write failed... */
1755                         ret = PTR_ERR(handle);
1756                         goto out;
1757                 }
1758                 if (inode->i_nlink)
1759                         ext3_orphan_del(handle, inode);
1760                 if (ret > 0) {
1761                         loff_t end = offset + ret;
1762                         if (end > inode->i_size) {
1763                                 ei->i_disksize = end;
1764                                 i_size_write(inode, end);
1765                                 /*
1766                                  * We're going to return a positive `ret'
1767                                  * here due to non-zero-length I/O, so there's
1768                                  * no way of reporting error returns from
1769                                  * ext3_mark_inode_dirty() to userspace.  So
1770                                  * ignore it.
1771                                  */
1772                                 ext3_mark_inode_dirty(handle, inode);
1773                         }
1774                 }
1775                 err = ext3_journal_stop(handle);
1776                 if (ret == 0)
1777                         ret = err;
1778         }
1779 out:
1780         return ret;
1781 }
1782
1783 /*
1784  * Pages can be marked dirty completely asynchronously from ext3's journalling
1785  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1786  * much here because ->set_page_dirty is called under VFS locks.  The page is
1787  * not necessarily locked.
1788  *
1789  * We cannot just dirty the page and leave attached buffers clean, because the
1790  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1791  * or jbddirty because all the journalling code will explode.
1792  *
1793  * So what we do is to mark the page "pending dirty" and next time writepage
1794  * is called, propagate that into the buffers appropriately.
1795  */
1796 static int ext3_journalled_set_page_dirty(struct page *page)
1797 {
1798         SetPageChecked(page);
1799         return __set_page_dirty_nobuffers(page);
1800 }
1801
1802 static const struct address_space_operations ext3_ordered_aops = {
1803         .readpage               = ext3_readpage,
1804         .readpages              = ext3_readpages,
1805         .writepage              = ext3_ordered_writepage,
1806         .sync_page              = block_sync_page,
1807         .write_begin            = ext3_write_begin,
1808         .write_end              = ext3_ordered_write_end,
1809         .bmap                   = ext3_bmap,
1810         .invalidatepage         = ext3_invalidatepage,
1811         .releasepage            = ext3_releasepage,
1812         .direct_IO              = ext3_direct_IO,
1813         .migratepage            = buffer_migrate_page,
1814         .is_partially_uptodate  = block_is_partially_uptodate,
1815 };
1816
1817 static const struct address_space_operations ext3_writeback_aops = {
1818         .readpage               = ext3_readpage,
1819         .readpages              = ext3_readpages,
1820         .writepage              = ext3_writeback_writepage,
1821         .sync_page              = block_sync_page,
1822         .write_begin            = ext3_write_begin,
1823         .write_end              = ext3_writeback_write_end,
1824         .bmap                   = ext3_bmap,
1825         .invalidatepage         = ext3_invalidatepage,
1826         .releasepage            = ext3_releasepage,
1827         .direct_IO              = ext3_direct_IO,
1828         .migratepage            = buffer_migrate_page,
1829         .is_partially_uptodate  = block_is_partially_uptodate,
1830 };
1831
1832 static const struct address_space_operations ext3_journalled_aops = {
1833         .readpage               = ext3_readpage,
1834         .readpages              = ext3_readpages,
1835         .writepage              = ext3_journalled_writepage,
1836         .sync_page              = block_sync_page,
1837         .write_begin            = ext3_write_begin,
1838         .write_end              = ext3_journalled_write_end,
1839         .set_page_dirty         = ext3_journalled_set_page_dirty,
1840         .bmap                   = ext3_bmap,
1841         .invalidatepage         = ext3_invalidatepage,
1842         .releasepage            = ext3_releasepage,
1843         .is_partially_uptodate  = block_is_partially_uptodate,
1844 };
1845
1846 void ext3_set_aops(struct inode *inode)
1847 {
1848         if (ext3_should_order_data(inode))
1849                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1850         else if (ext3_should_writeback_data(inode))
1851                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1852         else
1853                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1854 }
1855
1856 /*
1857  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1858  * up to the end of the block which corresponds to `from'.
1859  * This required during truncate. We need to physically zero the tail end
1860  * of that block so it doesn't yield old data if the file is later grown.
1861  */
1862 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1863                 struct address_space *mapping, loff_t from)
1864 {
1865         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1866         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1867         unsigned blocksize, iblock, length, pos;
1868         struct inode *inode = mapping->host;
1869         struct buffer_head *bh;
1870         int err = 0;
1871
1872         blocksize = inode->i_sb->s_blocksize;
1873         length = blocksize - (offset & (blocksize - 1));
1874         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1875
1876         /*
1877          * For "nobh" option,  we can only work if we don't need to
1878          * read-in the page - otherwise we create buffers to do the IO.
1879          */
1880         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1881              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1882                 zero_user(page, offset, length);
1883                 set_page_dirty(page);
1884                 goto unlock;
1885         }
1886
1887         if (!page_has_buffers(page))
1888                 create_empty_buffers(page, blocksize, 0);
1889
1890         /* Find the buffer that contains "offset" */
1891         bh = page_buffers(page);
1892         pos = blocksize;
1893         while (offset >= pos) {
1894                 bh = bh->b_this_page;
1895                 iblock++;
1896                 pos += blocksize;
1897         }
1898
1899         err = 0;
1900         if (buffer_freed(bh)) {
1901                 BUFFER_TRACE(bh, "freed: skip");
1902                 goto unlock;
1903         }
1904
1905         if (!buffer_mapped(bh)) {
1906                 BUFFER_TRACE(bh, "unmapped");
1907                 ext3_get_block(inode, iblock, bh, 0);
1908                 /* unmapped? It's a hole - nothing to do */
1909                 if (!buffer_mapped(bh)) {
1910                         BUFFER_TRACE(bh, "still unmapped");
1911                         goto unlock;
1912                 }
1913         }
1914
1915         /* Ok, it's mapped. Make sure it's up-to-date */
1916         if (PageUptodate(page))
1917                 set_buffer_uptodate(bh);
1918
1919         if (!buffer_uptodate(bh)) {
1920                 err = -EIO;
1921                 ll_rw_block(READ, 1, &bh);
1922                 wait_on_buffer(bh);
1923                 /* Uhhuh. Read error. Complain and punt. */
1924                 if (!buffer_uptodate(bh))
1925                         goto unlock;
1926         }
1927
1928         if (ext3_should_journal_data(inode)) {
1929                 BUFFER_TRACE(bh, "get write access");
1930                 err = ext3_journal_get_write_access(handle, bh);
1931                 if (err)
1932                         goto unlock;
1933         }
1934
1935         zero_user(page, offset, length);
1936         BUFFER_TRACE(bh, "zeroed end of block");
1937
1938         err = 0;
1939         if (ext3_should_journal_data(inode)) {
1940                 err = ext3_journal_dirty_metadata(handle, bh);
1941         } else {
1942                 if (ext3_should_order_data(inode))
1943                         err = ext3_journal_dirty_data(handle, bh);
1944                 mark_buffer_dirty(bh);
1945         }
1946
1947 unlock:
1948         unlock_page(page);
1949         page_cache_release(page);
1950         return err;
1951 }
1952
1953 /*
1954  * Probably it should be a library function... search for first non-zero word
1955  * or memcmp with zero_page, whatever is better for particular architecture.
1956  * Linus?
1957  */
1958 static inline int all_zeroes(__le32 *p, __le32 *q)
1959 {
1960         while (p < q)
1961                 if (*p++)
1962                         return 0;
1963         return 1;
1964 }
1965
1966 /**
1967  *      ext3_find_shared - find the indirect blocks for partial truncation.
1968  *      @inode:   inode in question
1969  *      @depth:   depth of the affected branch
1970  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1971  *      @chain:   place to store the pointers to partial indirect blocks
1972  *      @top:     place to the (detached) top of branch
1973  *
1974  *      This is a helper function used by ext3_truncate().
1975  *
1976  *      When we do truncate() we may have to clean the ends of several
1977  *      indirect blocks but leave the blocks themselves alive. Block is
1978  *      partially truncated if some data below the new i_size is refered
1979  *      from it (and it is on the path to the first completely truncated
1980  *      data block, indeed).  We have to free the top of that path along
1981  *      with everything to the right of the path. Since no allocation
1982  *      past the truncation point is possible until ext3_truncate()
1983  *      finishes, we may safely do the latter, but top of branch may
1984  *      require special attention - pageout below the truncation point
1985  *      might try to populate it.
1986  *
1987  *      We atomically detach the top of branch from the tree, store the
1988  *      block number of its root in *@top, pointers to buffer_heads of
1989  *      partially truncated blocks - in @chain[].bh and pointers to
1990  *      their last elements that should not be removed - in
1991  *      @chain[].p. Return value is the pointer to last filled element
1992  *      of @chain.
1993  *
1994  *      The work left to caller to do the actual freeing of subtrees:
1995  *              a) free the subtree starting from *@top
1996  *              b) free the subtrees whose roots are stored in
1997  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1998  *              c) free the subtrees growing from the inode past the @chain[0].
1999  *                      (no partially truncated stuff there).  */
2000
2001 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2002                         int offsets[4], Indirect chain[4], __le32 *top)
2003 {
2004         Indirect *partial, *p;
2005         int k, err;
2006
2007         *top = 0;
2008         /* Make k index the deepest non-null offest + 1 */
2009         for (k = depth; k > 1 && !offsets[k-1]; k--)
2010                 ;
2011         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2012         /* Writer: pointers */
2013         if (!partial)
2014                 partial = chain + k-1;
2015         /*
2016          * If the branch acquired continuation since we've looked at it -
2017          * fine, it should all survive and (new) top doesn't belong to us.
2018          */
2019         if (!partial->key && *partial->p)
2020                 /* Writer: end */
2021                 goto no_top;
2022         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2023                 ;
2024         /*
2025          * OK, we've found the last block that must survive. The rest of our
2026          * branch should be detached before unlocking. However, if that rest
2027          * of branch is all ours and does not grow immediately from the inode
2028          * it's easier to cheat and just decrement partial->p.
2029          */
2030         if (p == chain + k - 1 && p > chain) {
2031                 p->p--;
2032         } else {
2033                 *top = *p->p;
2034                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2035 #if 0
2036                 *p->p = 0;
2037 #endif
2038         }
2039         /* Writer: end */
2040
2041         while(partial > p) {
2042                 brelse(partial->bh);
2043                 partial--;
2044         }
2045 no_top:
2046         return partial;
2047 }
2048
2049 /*
2050  * Zero a number of block pointers in either an inode or an indirect block.
2051  * If we restart the transaction we must again get write access to the
2052  * indirect block for further modification.
2053  *
2054  * We release `count' blocks on disk, but (last - first) may be greater
2055  * than `count' because there can be holes in there.
2056  */
2057 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2058                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2059                 unsigned long count, __le32 *first, __le32 *last)
2060 {
2061         __le32 *p;
2062         if (try_to_extend_transaction(handle, inode)) {
2063                 if (bh) {
2064                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2065                         ext3_journal_dirty_metadata(handle, bh);
2066                 }
2067                 ext3_mark_inode_dirty(handle, inode);
2068                 ext3_journal_test_restart(handle, inode);
2069                 if (bh) {
2070                         BUFFER_TRACE(bh, "retaking write access");
2071                         ext3_journal_get_write_access(handle, bh);
2072                 }
2073         }
2074
2075         /*
2076          * Any buffers which are on the journal will be in memory. We find
2077          * them on the hash table so journal_revoke() will run journal_forget()
2078          * on them.  We've already detached each block from the file, so
2079          * bforget() in journal_forget() should be safe.
2080          *
2081          * AKPM: turn on bforget in journal_forget()!!!
2082          */
2083         for (p = first; p < last; p++) {
2084                 u32 nr = le32_to_cpu(*p);
2085                 if (nr) {
2086                         struct buffer_head *bh;
2087
2088                         *p = 0;
2089                         bh = sb_find_get_block(inode->i_sb, nr);
2090                         ext3_forget(handle, 0, inode, bh, nr);
2091                 }
2092         }
2093
2094         ext3_free_blocks(handle, inode, block_to_free, count);
2095 }
2096
2097 /**
2098  * ext3_free_data - free a list of data blocks
2099  * @handle:     handle for this transaction
2100  * @inode:      inode we are dealing with
2101  * @this_bh:    indirect buffer_head which contains *@first and *@last
2102  * @first:      array of block numbers
2103  * @last:       points immediately past the end of array
2104  *
2105  * We are freeing all blocks refered from that array (numbers are stored as
2106  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2107  *
2108  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2109  * blocks are contiguous then releasing them at one time will only affect one
2110  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2111  * actually use a lot of journal space.
2112  *
2113  * @this_bh will be %NULL if @first and @last point into the inode's direct
2114  * block pointers.
2115  */
2116 static void ext3_free_data(handle_t *handle, struct inode *inode,
2117                            struct buffer_head *this_bh,
2118                            __le32 *first, __le32 *last)
2119 {
2120         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2121         unsigned long count = 0;            /* Number of blocks in the run */
2122         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2123                                                corresponding to
2124                                                block_to_free */
2125         ext3_fsblk_t nr;                    /* Current block # */
2126         __le32 *p;                          /* Pointer into inode/ind
2127                                                for current block */
2128         int err;
2129
2130         if (this_bh) {                          /* For indirect block */
2131                 BUFFER_TRACE(this_bh, "get_write_access");
2132                 err = ext3_journal_get_write_access(handle, this_bh);
2133                 /* Important: if we can't update the indirect pointers
2134                  * to the blocks, we can't free them. */
2135                 if (err)
2136                         return;
2137         }
2138
2139         for (p = first; p < last; p++) {
2140                 nr = le32_to_cpu(*p);
2141                 if (nr) {
2142                         /* accumulate blocks to free if they're contiguous */
2143                         if (count == 0) {
2144                                 block_to_free = nr;
2145                                 block_to_free_p = p;
2146                                 count = 1;
2147                         } else if (nr == block_to_free + count) {
2148                                 count++;
2149                         } else {
2150                                 ext3_clear_blocks(handle, inode, this_bh,
2151                                                   block_to_free,
2152                                                   count, block_to_free_p, p);
2153                                 block_to_free = nr;
2154                                 block_to_free_p = p;
2155                                 count = 1;
2156                         }
2157                 }
2158         }
2159
2160         if (count > 0)
2161                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2162                                   count, block_to_free_p, p);
2163
2164         if (this_bh) {
2165                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2166
2167                 /*
2168                  * The buffer head should have an attached journal head at this
2169                  * point. However, if the data is corrupted and an indirect
2170                  * block pointed to itself, it would have been detached when
2171                  * the block was cleared. Check for this instead of OOPSing.
2172                  */
2173                 if (bh2jh(this_bh))
2174                         ext3_journal_dirty_metadata(handle, this_bh);
2175                 else
2176                         ext3_error(inode->i_sb, "ext3_free_data",
2177                                    "circular indirect block detected, "
2178                                    "inode=%lu, block=%llu",
2179                                    inode->i_ino,
2180                                    (unsigned long long)this_bh->b_blocknr);
2181         }
2182 }
2183
2184 /**
2185  *      ext3_free_branches - free an array of branches
2186  *      @handle: JBD handle for this transaction
2187  *      @inode: inode we are dealing with
2188  *      @parent_bh: the buffer_head which contains *@first and *@last
2189  *      @first: array of block numbers
2190  *      @last:  pointer immediately past the end of array
2191  *      @depth: depth of the branches to free
2192  *
2193  *      We are freeing all blocks refered from these branches (numbers are
2194  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2195  *      appropriately.
2196  */
2197 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2198                                struct buffer_head *parent_bh,
2199                                __le32 *first, __le32 *last, int depth)
2200 {
2201         ext3_fsblk_t nr;
2202         __le32 *p;
2203
2204         if (is_handle_aborted(handle))
2205                 return;
2206
2207         if (depth--) {
2208                 struct buffer_head *bh;
2209                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2210                 p = last;
2211                 while (--p >= first) {
2212                         nr = le32_to_cpu(*p);
2213                         if (!nr)
2214                                 continue;               /* A hole */
2215
2216                         /* Go read the buffer for the next level down */
2217                         bh = sb_bread(inode->i_sb, nr);
2218
2219                         /*
2220                          * A read failure? Report error and clear slot
2221                          * (should be rare).
2222                          */
2223                         if (!bh) {
2224                                 ext3_error(inode->i_sb, "ext3_free_branches",
2225                                            "Read failure, inode=%lu, block="E3FSBLK,
2226                                            inode->i_ino, nr);
2227                                 continue;
2228                         }
2229
2230                         /* This zaps the entire block.  Bottom up. */
2231                         BUFFER_TRACE(bh, "free child branches");
2232                         ext3_free_branches(handle, inode, bh,
2233                                            (__le32*)bh->b_data,
2234                                            (__le32*)bh->b_data + addr_per_block,
2235                                            depth);
2236
2237                         /*
2238                          * We've probably journalled the indirect block several
2239                          * times during the truncate.  But it's no longer
2240                          * needed and we now drop it from the transaction via
2241                          * journal_revoke().
2242                          *
2243                          * That's easy if it's exclusively part of this
2244                          * transaction.  But if it's part of the committing
2245                          * transaction then journal_forget() will simply
2246                          * brelse() it.  That means that if the underlying
2247                          * block is reallocated in ext3_get_block(),
2248                          * unmap_underlying_metadata() will find this block
2249                          * and will try to get rid of it.  damn, damn.
2250                          *
2251                          * If this block has already been committed to the
2252                          * journal, a revoke record will be written.  And
2253                          * revoke records must be emitted *before* clearing
2254                          * this block's bit in the bitmaps.
2255                          */
2256                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2257
2258                         /*
2259                          * Everything below this this pointer has been
2260                          * released.  Now let this top-of-subtree go.
2261                          *
2262                          * We want the freeing of this indirect block to be
2263                          * atomic in the journal with the updating of the
2264                          * bitmap block which owns it.  So make some room in
2265                          * the journal.
2266                          *
2267                          * We zero the parent pointer *after* freeing its
2268                          * pointee in the bitmaps, so if extend_transaction()
2269                          * for some reason fails to put the bitmap changes and
2270                          * the release into the same transaction, recovery
2271                          * will merely complain about releasing a free block,
2272                          * rather than leaking blocks.
2273                          */
2274                         if (is_handle_aborted(handle))
2275                                 return;
2276                         if (try_to_extend_transaction(handle, inode)) {
2277                                 ext3_mark_inode_dirty(handle, inode);
2278                                 ext3_journal_test_restart(handle, inode);
2279                         }
2280
2281                         ext3_free_blocks(handle, inode, nr, 1);
2282
2283                         if (parent_bh) {
2284                                 /*
2285                                  * The block which we have just freed is
2286                                  * pointed to by an indirect block: journal it
2287                                  */
2288                                 BUFFER_TRACE(parent_bh, "get_write_access");
2289                                 if (!ext3_journal_get_write_access(handle,
2290                                                                    parent_bh)){
2291                                         *p = 0;
2292                                         BUFFER_TRACE(parent_bh,
2293                                         "call ext3_journal_dirty_metadata");
2294                                         ext3_journal_dirty_metadata(handle,
2295                                                                     parent_bh);
2296                                 }
2297                         }
2298                 }
2299         } else {
2300                 /* We have reached the bottom of the tree. */
2301                 BUFFER_TRACE(parent_bh, "free data blocks");
2302                 ext3_free_data(handle, inode, parent_bh, first, last);
2303         }
2304 }
2305
2306 int ext3_can_truncate(struct inode *inode)
2307 {
2308         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2309                 return 0;
2310         if (S_ISREG(inode->i_mode))
2311                 return 1;
2312         if (S_ISDIR(inode->i_mode))
2313                 return 1;
2314         if (S_ISLNK(inode->i_mode))
2315                 return !ext3_inode_is_fast_symlink(inode);
2316         return 0;
2317 }
2318
2319 /*
2320  * ext3_truncate()
2321  *
2322  * We block out ext3_get_block() block instantiations across the entire
2323  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2324  * simultaneously on behalf of the same inode.
2325  *
2326  * As we work through the truncate and commmit bits of it to the journal there
2327  * is one core, guiding principle: the file's tree must always be consistent on
2328  * disk.  We must be able to restart the truncate after a crash.
2329  *
2330  * The file's tree may be transiently inconsistent in memory (although it
2331  * probably isn't), but whenever we close off and commit a journal transaction,
2332  * the contents of (the filesystem + the journal) must be consistent and
2333  * restartable.  It's pretty simple, really: bottom up, right to left (although
2334  * left-to-right works OK too).
2335  *
2336  * Note that at recovery time, journal replay occurs *before* the restart of
2337  * truncate against the orphan inode list.
2338  *
2339  * The committed inode has the new, desired i_size (which is the same as
2340  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2341  * that this inode's truncate did not complete and it will again call
2342  * ext3_truncate() to have another go.  So there will be instantiated blocks
2343  * to the right of the truncation point in a crashed ext3 filesystem.  But
2344  * that's fine - as long as they are linked from the inode, the post-crash
2345  * ext3_truncate() run will find them and release them.
2346  */
2347 void ext3_truncate(struct inode *inode)
2348 {
2349         handle_t *handle;
2350         struct ext3_inode_info *ei = EXT3_I(inode);
2351         __le32 *i_data = ei->i_data;
2352         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2353         struct address_space *mapping = inode->i_mapping;
2354         int offsets[4];
2355         Indirect chain[4];
2356         Indirect *partial;
2357         __le32 nr = 0;
2358         int n;
2359         long last_block;
2360         unsigned blocksize = inode->i_sb->s_blocksize;
2361         struct page *page;
2362
2363         if (!ext3_can_truncate(inode))
2364                 return;
2365
2366         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2367                 ei->i_state |= EXT3_STATE_FLUSH_ON_CLOSE;
2368
2369         /*
2370          * We have to lock the EOF page here, because lock_page() nests
2371          * outside journal_start().
2372          */
2373         if ((inode->i_size & (blocksize - 1)) == 0) {
2374                 /* Block boundary? Nothing to do */
2375                 page = NULL;
2376         } else {
2377                 page = grab_cache_page(mapping,
2378                                 inode->i_size >> PAGE_CACHE_SHIFT);
2379                 if (!page)
2380                         return;
2381         }
2382
2383         handle = start_transaction(inode);
2384         if (IS_ERR(handle)) {
2385                 if (page) {
2386                         clear_highpage(page);
2387                         flush_dcache_page(page);
2388                         unlock_page(page);
2389                         page_cache_release(page);
2390                 }
2391                 return;         /* AKPM: return what? */
2392         }
2393
2394         last_block = (inode->i_size + blocksize-1)
2395                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2396
2397         if (page)
2398                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2399
2400         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2401         if (n == 0)
2402                 goto out_stop;  /* error */
2403
2404         /*
2405          * OK.  This truncate is going to happen.  We add the inode to the
2406          * orphan list, so that if this truncate spans multiple transactions,
2407          * and we crash, we will resume the truncate when the filesystem
2408          * recovers.  It also marks the inode dirty, to catch the new size.
2409          *
2410          * Implication: the file must always be in a sane, consistent
2411          * truncatable state while each transaction commits.
2412          */
2413         if (ext3_orphan_add(handle, inode))
2414                 goto out_stop;
2415
2416         /*
2417          * The orphan list entry will now protect us from any crash which
2418          * occurs before the truncate completes, so it is now safe to propagate
2419          * the new, shorter inode size (held for now in i_size) into the
2420          * on-disk inode. We do this via i_disksize, which is the value which
2421          * ext3 *really* writes onto the disk inode.
2422          */
2423         ei->i_disksize = inode->i_size;
2424
2425         /*
2426          * From here we block out all ext3_get_block() callers who want to
2427          * modify the block allocation tree.
2428          */
2429         mutex_lock(&ei->truncate_mutex);
2430
2431         if (n == 1) {           /* direct blocks */
2432                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2433                                i_data + EXT3_NDIR_BLOCKS);
2434                 goto do_indirects;
2435         }
2436
2437         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2438         /* Kill the top of shared branch (not detached) */
2439         if (nr) {
2440                 if (partial == chain) {
2441                         /* Shared branch grows from the inode */
2442                         ext3_free_branches(handle, inode, NULL,
2443                                            &nr, &nr+1, (chain+n-1) - partial);
2444                         *partial->p = 0;
2445                         /*
2446                          * We mark the inode dirty prior to restart,
2447                          * and prior to stop.  No need for it here.
2448                          */
2449                 } else {
2450                         /* Shared branch grows from an indirect block */
2451                         BUFFER_TRACE(partial->bh, "get_write_access");
2452                         ext3_free_branches(handle, inode, partial->bh,
2453                                         partial->p,
2454                                         partial->p+1, (chain+n-1) - partial);
2455                 }
2456         }
2457         /* Clear the ends of indirect blocks on the shared branch */
2458         while (partial > chain) {
2459                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2460                                    (__le32*)partial->bh->b_data+addr_per_block,
2461                                    (chain+n-1) - partial);
2462                 BUFFER_TRACE(partial->bh, "call brelse");
2463                 brelse (partial->bh);
2464                 partial--;
2465         }
2466 do_indirects:
2467         /* Kill the remaining (whole) subtrees */
2468         switch (offsets[0]) {
2469         default:
2470                 nr = i_data[EXT3_IND_BLOCK];
2471                 if (nr) {
2472                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2473                         i_data[EXT3_IND_BLOCK] = 0;
2474                 }
2475         case EXT3_IND_BLOCK:
2476                 nr = i_data[EXT3_DIND_BLOCK];
2477                 if (nr) {
2478                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2479                         i_data[EXT3_DIND_BLOCK] = 0;
2480                 }
2481         case EXT3_DIND_BLOCK:
2482                 nr = i_data[EXT3_TIND_BLOCK];
2483                 if (nr) {
2484                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2485                         i_data[EXT3_TIND_BLOCK] = 0;
2486                 }
2487         case EXT3_TIND_BLOCK:
2488                 ;
2489         }
2490
2491         ext3_discard_reservation(inode);
2492
2493         mutex_unlock(&ei->truncate_mutex);
2494         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2495         ext3_mark_inode_dirty(handle, inode);
2496
2497         /*
2498          * In a multi-transaction truncate, we only make the final transaction
2499          * synchronous
2500          */
2501         if (IS_SYNC(inode))
2502                 handle->h_sync = 1;
2503 out_stop:
2504         /*
2505          * If this was a simple ftruncate(), and the file will remain alive
2506          * then we need to clear up the orphan record which we created above.
2507          * However, if this was a real unlink then we were called by
2508          * ext3_delete_inode(), and we allow that function to clean up the
2509          * orphan info for us.
2510          */
2511         if (inode->i_nlink)
2512                 ext3_orphan_del(handle, inode);
2513
2514         ext3_journal_stop(handle);
2515 }
2516
2517 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2518                 unsigned long ino, struct ext3_iloc *iloc)
2519 {
2520         unsigned long block_group;
2521         unsigned long offset;
2522         ext3_fsblk_t block;
2523         struct ext3_group_desc *gdp;
2524
2525         if (!ext3_valid_inum(sb, ino)) {
2526                 /*
2527                  * This error is already checked for in namei.c unless we are
2528                  * looking at an NFS filehandle, in which case no error
2529                  * report is needed
2530                  */
2531                 return 0;
2532         }
2533
2534         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2535         gdp = ext3_get_group_desc(sb, block_group, NULL);
2536         if (!gdp)
2537                 return 0;
2538         /*
2539          * Figure out the offset within the block group inode table
2540          */
2541         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2542                 EXT3_INODE_SIZE(sb);
2543         block = le32_to_cpu(gdp->bg_inode_table) +
2544                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2545
2546         iloc->block_group = block_group;
2547         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2548         return block;
2549 }
2550
2551 /*
2552  * ext3_get_inode_loc returns with an extra refcount against the inode's
2553  * underlying buffer_head on success. If 'in_mem' is true, we have all
2554  * data in memory that is needed to recreate the on-disk version of this
2555  * inode.
2556  */
2557 static int __ext3_get_inode_loc(struct inode *inode,
2558                                 struct ext3_iloc *iloc, int in_mem)
2559 {
2560         ext3_fsblk_t block;
2561         struct buffer_head *bh;
2562
2563         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2564         if (!block)
2565                 return -EIO;
2566
2567         bh = sb_getblk(inode->i_sb, block);
2568         if (!bh) {
2569                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2570                                 "unable to read inode block - "
2571                                 "inode=%lu, block="E3FSBLK,
2572                                  inode->i_ino, block);
2573                 return -EIO;
2574         }
2575         if (!buffer_uptodate(bh)) {
2576                 lock_buffer(bh);
2577
2578                 /*
2579                  * If the buffer has the write error flag, we have failed
2580                  * to write out another inode in the same block.  In this
2581                  * case, we don't have to read the block because we may
2582                  * read the old inode data successfully.
2583                  */
2584                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2585                         set_buffer_uptodate(bh);
2586
2587                 if (buffer_uptodate(bh)) {
2588                         /* someone brought it uptodate while we waited */
2589                         unlock_buffer(bh);
2590                         goto has_buffer;
2591                 }
2592
2593                 /*
2594                  * If we have all information of the inode in memory and this
2595                  * is the only valid inode in the block, we need not read the
2596                  * block.
2597                  */
2598                 if (in_mem) {
2599                         struct buffer_head *bitmap_bh;
2600                         struct ext3_group_desc *desc;
2601                         int inodes_per_buffer;
2602                         int inode_offset, i;
2603                         int block_group;
2604                         int start;
2605
2606                         block_group = (inode->i_ino - 1) /
2607                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2608                         inodes_per_buffer = bh->b_size /
2609                                 EXT3_INODE_SIZE(inode->i_sb);
2610                         inode_offset = ((inode->i_ino - 1) %
2611                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2612                         start = inode_offset & ~(inodes_per_buffer - 1);
2613
2614                         /* Is the inode bitmap in cache? */
2615                         desc = ext3_get_group_desc(inode->i_sb,
2616                                                 block_group, NULL);
2617                         if (!desc)
2618                                 goto make_io;
2619
2620                         bitmap_bh = sb_getblk(inode->i_sb,
2621                                         le32_to_cpu(desc->bg_inode_bitmap));
2622                         if (!bitmap_bh)
2623                                 goto make_io;
2624
2625                         /*
2626                          * If the inode bitmap isn't in cache then the
2627                          * optimisation may end up performing two reads instead
2628                          * of one, so skip it.
2629                          */
2630                         if (!buffer_uptodate(bitmap_bh)) {
2631                                 brelse(bitmap_bh);
2632                                 goto make_io;
2633                         }
2634                         for (i = start; i < start + inodes_per_buffer; i++) {
2635                                 if (i == inode_offset)
2636                                         continue;
2637                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2638                                         break;
2639                         }
2640                         brelse(bitmap_bh);
2641                         if (i == start + inodes_per_buffer) {
2642                                 /* all other inodes are free, so skip I/O */
2643                                 memset(bh->b_data, 0, bh->b_size);
2644                                 set_buffer_uptodate(bh);
2645                                 unlock_buffer(bh);
2646                                 goto has_buffer;
2647                         }
2648                 }
2649
2650 make_io:
2651                 /*
2652                  * There are other valid inodes in the buffer, this inode
2653                  * has in-inode xattrs, or we don't have this inode in memory.
2654                  * Read the block from disk.
2655                  */
2656                 get_bh(bh);
2657                 bh->b_end_io = end_buffer_read_sync;
2658                 submit_bh(READ_META, bh);
2659                 wait_on_buffer(bh);
2660                 if (!buffer_uptodate(bh)) {
2661                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2662                                         "unable to read inode block - "
2663                                         "inode=%lu, block="E3FSBLK,
2664                                         inode->i_ino, block);
2665                         brelse(bh);
2666                         return -EIO;
2667                 }
2668         }
2669 has_buffer:
2670         iloc->bh = bh;
2671         return 0;
2672 }
2673
2674 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2675 {
2676         /* We have all inode data except xattrs in memory here. */
2677         return __ext3_get_inode_loc(inode, iloc,
2678                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2679 }
2680
2681 void ext3_set_inode_flags(struct inode *inode)
2682 {
2683         unsigned int flags = EXT3_I(inode)->i_flags;
2684
2685         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2686         if (flags & EXT3_SYNC_FL)
2687                 inode->i_flags |= S_SYNC;
2688         if (flags & EXT3_APPEND_FL)
2689                 inode->i_flags |= S_APPEND;
2690         if (flags & EXT3_IMMUTABLE_FL)
2691                 inode->i_flags |= S_IMMUTABLE;
2692         if (flags & EXT3_NOATIME_FL)
2693                 inode->i_flags |= S_NOATIME;
2694         if (flags & EXT3_DIRSYNC_FL)
2695                 inode->i_flags |= S_DIRSYNC;
2696 }
2697
2698 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2699 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2700 {
2701         unsigned int flags = ei->vfs_inode.i_flags;
2702
2703         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2704                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2705         if (flags & S_SYNC)
2706                 ei->i_flags |= EXT3_SYNC_FL;
2707         if (flags & S_APPEND)
2708                 ei->i_flags |= EXT3_APPEND_FL;
2709         if (flags & S_IMMUTABLE)
2710                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2711         if (flags & S_NOATIME)
2712                 ei->i_flags |= EXT3_NOATIME_FL;
2713         if (flags & S_DIRSYNC)
2714                 ei->i_flags |= EXT3_DIRSYNC_FL;
2715 }
2716
2717 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2718 {
2719         struct ext3_iloc iloc;
2720         struct ext3_inode *raw_inode;
2721         struct ext3_inode_info *ei;
2722         struct buffer_head *bh;
2723         struct inode *inode;
2724         long ret;
2725         int block;
2726
2727         inode = iget_locked(sb, ino);
2728         if (!inode)
2729                 return ERR_PTR(-ENOMEM);
2730         if (!(inode->i_state & I_NEW))
2731                 return inode;
2732
2733         ei = EXT3_I(inode);
2734 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2735         ei->i_acl = EXT3_ACL_NOT_CACHED;
2736         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2737 #endif
2738         ei->i_block_alloc_info = NULL;
2739
2740         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2741         if (ret < 0)
2742                 goto bad_inode;
2743         bh = iloc.bh;
2744         raw_inode = ext3_raw_inode(&iloc);
2745         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2746         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2747         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2748         if(!(test_opt (inode->i_sb, NO_UID32))) {
2749                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2750                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2751         }
2752         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2753         inode->i_size = le32_to_cpu(raw_inode->i_size);
2754         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2755         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2756         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2757         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2758
2759         ei->i_state = 0;
2760         ei->i_dir_start_lookup = 0;
2761         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2762         /* We now have enough fields to check if the inode was active or not.
2763          * This is needed because nfsd might try to access dead inodes
2764          * the test is that same one that e2fsck uses
2765          * NeilBrown 1999oct15
2766          */
2767         if (inode->i_nlink == 0) {
2768                 if (inode->i_mode == 0 ||
2769                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2770                         /* this inode is deleted */
2771                         brelse (bh);
2772                         ret = -ESTALE;
2773                         goto bad_inode;
2774                 }
2775                 /* The only unlinked inodes we let through here have
2776                  * valid i_mode and are being read by the orphan
2777                  * recovery code: that's fine, we're about to complete
2778                  * the process of deleting those. */
2779         }
2780         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2781         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2782 #ifdef EXT3_FRAGMENTS
2783         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2784         ei->i_frag_no = raw_inode->i_frag;
2785         ei->i_frag_size = raw_inode->i_fsize;
2786 #endif
2787         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2788         if (!S_ISREG(inode->i_mode)) {
2789                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2790         } else {
2791                 inode->i_size |=
2792                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2793         }
2794         ei->i_disksize = inode->i_size;
2795         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2796         ei->i_block_group = iloc.block_group;
2797         /*
2798          * NOTE! The in-memory inode i_data array is in little-endian order
2799          * even on big-endian machines: we do NOT byteswap the block numbers!
2800          */
2801         for (block = 0; block < EXT3_N_BLOCKS; block++)
2802                 ei->i_data[block] = raw_inode->i_block[block];
2803         INIT_LIST_HEAD(&ei->i_orphan);
2804
2805         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2806             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2807                 /*
2808                  * When mke2fs creates big inodes it does not zero out
2809                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2810                  * so ignore those first few inodes.
2811                  */
2812                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2813                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2814                     EXT3_INODE_SIZE(inode->i_sb)) {
2815                         brelse (bh);
2816                         ret = -EIO;
2817                         goto bad_inode;
2818                 }
2819                 if (ei->i_extra_isize == 0) {
2820                         /* The extra space is currently unused. Use it. */
2821                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2822                                             EXT3_GOOD_OLD_INODE_SIZE;
2823                 } else {
2824                         __le32 *magic = (void *)raw_inode +
2825                                         EXT3_GOOD_OLD_INODE_SIZE +
2826                                         ei->i_extra_isize;
2827                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2828                                  ei->i_state |= EXT3_STATE_XATTR;
2829                 }
2830         } else
2831                 ei->i_extra_isize = 0;
2832
2833         if (S_ISREG(inode->i_mode)) {
2834                 inode->i_op = &ext3_file_inode_operations;
2835                 inode->i_fop = &ext3_file_operations;
2836                 ext3_set_aops(inode);
2837         } else if (S_ISDIR(inode->i_mode)) {
2838                 inode->i_op = &ext3_dir_inode_operations;
2839                 inode->i_fop = &ext3_dir_operations;
2840         } else if (S_ISLNK(inode->i_mode)) {
2841                 if (ext3_inode_is_fast_symlink(inode)) {
2842                         inode->i_op = &ext3_fast_symlink_inode_operations;
2843                         nd_terminate_link(ei->i_data, inode->i_size,
2844                                 sizeof(ei->i_data) - 1);
2845                 } else {
2846                         inode->i_op = &ext3_symlink_inode_operations;
2847                         ext3_set_aops(inode);
2848                 }
2849         } else {
2850                 inode->i_op = &ext3_special_inode_operations;
2851                 if (raw_inode->i_block[0])
2852                         init_special_inode(inode, inode->i_mode,
2853                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2854                 else
2855                         init_special_inode(inode, inode->i_mode,
2856                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2857         }
2858         brelse (iloc.bh);
2859         ext3_set_inode_flags(inode);
2860         unlock_new_inode(inode);
2861         return inode;
2862
2863 bad_inode:
2864         iget_failed(inode);
2865         return ERR_PTR(ret);
2866 }
2867
2868 /*
2869  * Post the struct inode info into an on-disk inode location in the
2870  * buffer-cache.  This gobbles the caller's reference to the
2871  * buffer_head in the inode location struct.
2872  *
2873  * The caller must have write access to iloc->bh.
2874  */
2875 static int ext3_do_update_inode(handle_t *handle,
2876                                 struct inode *inode,
2877                                 struct ext3_iloc *iloc)
2878 {
2879         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2880         struct ext3_inode_info *ei = EXT3_I(inode);
2881         struct buffer_head *bh = iloc->bh;
2882         int err = 0, rc, block;
2883
2884         /* For fields not not tracking in the in-memory inode,
2885          * initialise them to zero for new inodes. */
2886         if (ei->i_state & EXT3_STATE_NEW)
2887                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2888
2889         ext3_get_inode_flags(ei);
2890         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2891         if(!(test_opt(inode->i_sb, NO_UID32))) {
2892                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2893                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2894 /*
2895  * Fix up interoperability with old kernels. Otherwise, old inodes get
2896  * re-used with the upper 16 bits of the uid/gid intact
2897  */
2898                 if(!ei->i_dtime) {
2899                         raw_inode->i_uid_high =
2900                                 cpu_to_le16(high_16_bits(inode->i_uid));
2901                         raw_inode->i_gid_high =
2902                                 cpu_to_le16(high_16_bits(inode->i_gid));
2903                 } else {
2904                         raw_inode->i_uid_high = 0;
2905                         raw_inode->i_gid_high = 0;
2906                 }
2907         } else {
2908                 raw_inode->i_uid_low =
2909                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2910                 raw_inode->i_gid_low =
2911                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2912                 raw_inode->i_uid_high = 0;
2913                 raw_inode->i_gid_high = 0;
2914         }
2915         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2916         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2917         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2918         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2919         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2920         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2921         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2922         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2923 #ifdef EXT3_FRAGMENTS
2924         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2925         raw_inode->i_frag = ei->i_frag_no;
2926         raw_inode->i_fsize = ei->i_frag_size;
2927 #endif
2928         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2929         if (!S_ISREG(inode->i_mode)) {
2930                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2931         } else {
2932                 raw_inode->i_size_high =
2933                         cpu_to_le32(ei->i_disksize >> 32);
2934                 if (ei->i_disksize > 0x7fffffffULL) {
2935                         struct super_block *sb = inode->i_sb;
2936                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2937                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2938                             EXT3_SB(sb)->s_es->s_rev_level ==
2939                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2940                                /* If this is the first large file
2941                                 * created, add a flag to the superblock.
2942                                 */
2943                                 err = ext3_journal_get_write_access(handle,
2944                                                 EXT3_SB(sb)->s_sbh);
2945                                 if (err)
2946                                         goto out_brelse;
2947                                 ext3_update_dynamic_rev(sb);
2948                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2949                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2950                                 sb->s_dirt = 1;
2951                                 handle->h_sync = 1;
2952                                 err = ext3_journal_dirty_metadata(handle,
2953                                                 EXT3_SB(sb)->s_sbh);
2954                         }
2955                 }
2956         }
2957         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2958         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2959                 if (old_valid_dev(inode->i_rdev)) {
2960                         raw_inode->i_block[0] =
2961                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2962                         raw_inode->i_block[1] = 0;
2963                 } else {
2964                         raw_inode->i_block[0] = 0;
2965                         raw_inode->i_block[1] =
2966                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2967                         raw_inode->i_block[2] = 0;
2968                 }
2969         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2970                 raw_inode->i_block[block] = ei->i_data[block];
2971
2972         if (ei->i_extra_isize)
2973                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2974
2975         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2976         rc = ext3_journal_dirty_metadata(handle, bh);
2977         if (!err)
2978                 err = rc;
2979         ei->i_state &= ~EXT3_STATE_NEW;
2980
2981 out_brelse:
2982         brelse (bh);
2983         ext3_std_error(inode->i_sb, err);
2984         return err;
2985 }
2986
2987 /*
2988  * ext3_write_inode()
2989  *
2990  * We are called from a few places:
2991  *
2992  * - Within generic_file_write() for O_SYNC files.
2993  *   Here, there will be no transaction running. We wait for any running
2994  *   trasnaction to commit.
2995  *
2996  * - Within sys_sync(), kupdate and such.
2997  *   We wait on commit, if tol to.
2998  *
2999  * - Within prune_icache() (PF_MEMALLOC == true)
3000  *   Here we simply return.  We can't afford to block kswapd on the
3001  *   journal commit.
3002  *
3003  * In all cases it is actually safe for us to return without doing anything,
3004  * because the inode has been copied into a raw inode buffer in
3005  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3006  * knfsd.
3007  *
3008  * Note that we are absolutely dependent upon all inode dirtiers doing the
3009  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3010  * which we are interested.
3011  *
3012  * It would be a bug for them to not do this.  The code:
3013  *
3014  *      mark_inode_dirty(inode)
3015  *      stuff();
3016  *      inode->i_size = expr;
3017  *
3018  * is in error because a kswapd-driven write_inode() could occur while
3019  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3020  * will no longer be on the superblock's dirty inode list.
3021  */
3022 int ext3_write_inode(struct inode *inode, int wait)
3023 {
3024         if (current->flags & PF_MEMALLOC)
3025                 return 0;
3026
3027         if (ext3_journal_current_handle()) {
3028                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3029                 dump_stack();
3030                 return -EIO;
3031         }
3032
3033         if (!wait)
3034                 return 0;
3035
3036         return ext3_force_commit(inode->i_sb);
3037 }
3038
3039 /*
3040  * ext3_setattr()
3041  *
3042  * Called from notify_change.
3043  *
3044  * We want to trap VFS attempts to truncate the file as soon as
3045  * possible.  In particular, we want to make sure that when the VFS
3046  * shrinks i_size, we put the inode on the orphan list and modify
3047  * i_disksize immediately, so that during the subsequent flushing of
3048  * dirty pages and freeing of disk blocks, we can guarantee that any
3049  * commit will leave the blocks being flushed in an unused state on
3050  * disk.  (On recovery, the inode will get truncated and the blocks will
3051  * be freed, so we have a strong guarantee that no future commit will
3052  * leave these blocks visible to the user.)
3053  *
3054  * Called with inode->sem down.
3055  */
3056 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3057 {
3058         struct inode *inode = dentry->d_inode;
3059         int error, rc = 0;
3060         const unsigned int ia_valid = attr->ia_valid;
3061
3062         error = inode_change_ok(inode, attr);
3063         if (error)
3064                 return error;
3065
3066         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3067                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3068                 handle_t *handle;
3069
3070                 /* (user+group)*(old+new) structure, inode write (sb,
3071                  * inode block, ? - but truncate inode update has it) */
3072                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3073                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3074                 if (IS_ERR(handle)) {
3075                         error = PTR_ERR(handle);
3076                         goto err_out;
3077                 }
3078                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
3079                 if (error) {
3080                         ext3_journal_stop(handle);
3081                         return error;
3082                 }
3083                 /* Update corresponding info in inode so that everything is in
3084                  * one transaction */
3085                 if (attr->ia_valid & ATTR_UID)
3086                         inode->i_uid = attr->ia_uid;
3087                 if (attr->ia_valid & ATTR_GID)
3088                         inode->i_gid = attr->ia_gid;
3089                 error = ext3_mark_inode_dirty(handle, inode);
3090                 ext3_journal_stop(handle);
3091         }
3092
3093         if (S_ISREG(inode->i_mode) &&
3094             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3095                 handle_t *handle;
3096
3097                 handle = ext3_journal_start(inode, 3);
3098                 if (IS_ERR(handle)) {
3099                         error = PTR_ERR(handle);
3100                         goto err_out;
3101                 }
3102
3103                 error = ext3_orphan_add(handle, inode);
3104                 EXT3_I(inode)->i_disksize = attr->ia_size;
3105                 rc = ext3_mark_inode_dirty(handle, inode);
3106                 if (!error)
3107                         error = rc;
3108                 ext3_journal_stop(handle);
3109         }
3110
3111         rc = inode_setattr(inode, attr);
3112
3113         /* If inode_setattr's call to ext3_truncate failed to get a
3114          * transaction handle at all, we need to clean up the in-core
3115          * orphan list manually. */
3116         if (inode->i_nlink)
3117                 ext3_orphan_del(NULL, inode);
3118
3119         if (!rc && (ia_valid & ATTR_MODE))
3120                 rc = ext3_acl_chmod(inode);
3121
3122 err_out:
3123         ext3_std_error(inode->i_sb, error);
3124         if (!error)
3125                 error = rc;
3126         return error;
3127 }
3128
3129
3130 /*
3131  * How many blocks doth make a writepage()?
3132  *
3133  * With N blocks per page, it may be:
3134  * N data blocks
3135  * 2 indirect block
3136  * 2 dindirect
3137  * 1 tindirect
3138  * N+5 bitmap blocks (from the above)
3139  * N+5 group descriptor summary blocks
3140  * 1 inode block
3141  * 1 superblock.
3142  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3143  *
3144  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3145  *
3146  * With ordered or writeback data it's the same, less the N data blocks.
3147  *
3148  * If the inode's direct blocks can hold an integral number of pages then a
3149  * page cannot straddle two indirect blocks, and we can only touch one indirect
3150  * and dindirect block, and the "5" above becomes "3".
3151  *
3152  * This still overestimates under most circumstances.  If we were to pass the
3153  * start and end offsets in here as well we could do block_to_path() on each
3154  * block and work out the exact number of indirects which are touched.  Pah.
3155  */
3156
3157 static int ext3_writepage_trans_blocks(struct inode *inode)
3158 {
3159         int bpp = ext3_journal_blocks_per_page(inode);
3160         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3161         int ret;
3162
3163         if (ext3_should_journal_data(inode))
3164                 ret = 3 * (bpp + indirects) + 2;
3165         else
3166                 ret = 2 * (bpp + indirects) + 2;
3167
3168 #ifdef CONFIG_QUOTA
3169         /* We know that structure was already allocated during vfs_dq_init so
3170          * we will be updating only the data blocks + inodes */
3171         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3172 #endif
3173
3174         return ret;
3175 }
3176
3177 /*
3178  * The caller must have previously called ext3_reserve_inode_write().
3179  * Give this, we know that the caller already has write access to iloc->bh.
3180  */
3181 int ext3_mark_iloc_dirty(handle_t *handle,
3182                 struct inode *inode, struct ext3_iloc *iloc)
3183 {
3184         int err = 0;
3185
3186         /* the do_update_inode consumes one bh->b_count */
3187         get_bh(iloc->bh);
3188
3189         /* ext3_do_update_inode() does journal_dirty_metadata */
3190         err = ext3_do_update_inode(handle, inode, iloc);
3191         put_bh(iloc->bh);
3192         return err;
3193 }
3194
3195 /*
3196  * On success, We end up with an outstanding reference count against
3197  * iloc->bh.  This _must_ be cleaned up later.
3198  */
3199
3200 int
3201 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3202                          struct ext3_iloc *iloc)
3203 {
3204         int err = 0;
3205         if (handle) {
3206                 err = ext3_get_inode_loc(inode, iloc);
3207                 if (!err) {
3208                         BUFFER_TRACE(iloc->bh, "get_write_access");
3209                         err = ext3_journal_get_write_access(handle, iloc->bh);
3210                         if (err) {
3211                                 brelse(iloc->bh);
3212                                 iloc->bh = NULL;
3213                         }
3214                 }
3215         }
3216         ext3_std_error(inode->i_sb, err);
3217         return err;
3218 }
3219
3220 /*
3221  * What we do here is to mark the in-core inode as clean with respect to inode
3222  * dirtiness (it may still be data-dirty).
3223  * This means that the in-core inode may be reaped by prune_icache
3224  * without having to perform any I/O.  This is a very good thing,
3225  * because *any* task may call prune_icache - even ones which
3226  * have a transaction open against a different journal.
3227  *
3228  * Is this cheating?  Not really.  Sure, we haven't written the
3229  * inode out, but prune_icache isn't a user-visible syncing function.
3230  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3231  * we start and wait on commits.
3232  *
3233  * Is this efficient/effective?  Well, we're being nice to the system
3234  * by cleaning up our inodes proactively so they can be reaped
3235  * without I/O.  But we are potentially leaving up to five seconds'
3236  * worth of inodes floating about which prune_icache wants us to
3237  * write out.  One way to fix that would be to get prune_icache()
3238  * to do a write_super() to free up some memory.  It has the desired
3239  * effect.
3240  */
3241 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3242 {
3243         struct ext3_iloc iloc;
3244         int err;
3245
3246         might_sleep();
3247         err = ext3_reserve_inode_write(handle, inode, &iloc);
3248         if (!err)
3249                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3250         return err;
3251 }
3252
3253 /*
3254  * ext3_dirty_inode() is called from __mark_inode_dirty()
3255  *
3256  * We're really interested in the case where a file is being extended.
3257  * i_size has been changed by generic_commit_write() and we thus need
3258  * to include the updated inode in the current transaction.
3259  *
3260  * Also, vfs_dq_alloc_space() will always dirty the inode when blocks
3261  * are allocated to the file.
3262  *
3263  * If the inode is marked synchronous, we don't honour that here - doing
3264  * so would cause a commit on atime updates, which we don't bother doing.
3265  * We handle synchronous inodes at the highest possible level.
3266  */
3267 void ext3_dirty_inode(struct inode *inode)
3268 {
3269         handle_t *current_handle = ext3_journal_current_handle();
3270         handle_t *handle;
3271
3272         handle = ext3_journal_start(inode, 2);
3273         if (IS_ERR(handle))
3274                 goto out;
3275         if (current_handle &&
3276                 current_handle->h_transaction != handle->h_transaction) {
3277                 /* This task has a transaction open against a different fs */
3278                 printk(KERN_EMERG "%s: transactions do not match!\n",
3279                        __func__);
3280         } else {
3281                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3282                                 current_handle);
3283                 ext3_mark_inode_dirty(handle, inode);
3284         }
3285         ext3_journal_stop(handle);
3286 out:
3287         return;
3288 }
3289
3290 #if 0
3291 /*
3292  * Bind an inode's backing buffer_head into this transaction, to prevent
3293  * it from being flushed to disk early.  Unlike
3294  * ext3_reserve_inode_write, this leaves behind no bh reference and
3295  * returns no iloc structure, so the caller needs to repeat the iloc
3296  * lookup to mark the inode dirty later.
3297  */
3298 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3299 {
3300         struct ext3_iloc iloc;
3301
3302         int err = 0;
3303         if (handle) {
3304                 err = ext3_get_inode_loc(inode, &iloc);
3305                 if (!err) {
3306                         BUFFER_TRACE(iloc.bh, "get_write_access");
3307                         err = journal_get_write_access(handle, iloc.bh);
3308                         if (!err)
3309                                 err = ext3_journal_dirty_metadata(handle,
3310                                                                   iloc.bh);
3311                         brelse(iloc.bh);
3312                 }
3313         }
3314         ext3_std_error(inode->i_sb, err);
3315         return err;
3316 }
3317 #endif
3318
3319 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3320 {
3321         journal_t *journal;
3322         handle_t *handle;
3323         int err;
3324
3325         /*
3326          * We have to be very careful here: changing a data block's
3327          * journaling status dynamically is dangerous.  If we write a
3328          * data block to the journal, change the status and then delete
3329          * that block, we risk forgetting to revoke the old log record
3330          * from the journal and so a subsequent replay can corrupt data.
3331          * So, first we make sure that the journal is empty and that
3332          * nobody is changing anything.
3333          */
3334
3335         journal = EXT3_JOURNAL(inode);
3336         if (is_journal_aborted(journal))
3337                 return -EROFS;
3338
3339         journal_lock_updates(journal);
3340         journal_flush(journal);
3341
3342         /*
3343          * OK, there are no updates running now, and all cached data is
3344          * synced to disk.  We are now in a completely consistent state
3345          * which doesn't have anything in the journal, and we know that
3346          * no filesystem updates are running, so it is safe to modify
3347          * the inode's in-core data-journaling state flag now.
3348          */
3349
3350         if (val)
3351                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3352         else
3353                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3354         ext3_set_aops(inode);
3355
3356         journal_unlock_updates(journal);
3357
3358         /* Finally we can mark the inode as dirty. */
3359
3360         handle = ext3_journal_start(inode, 1);
3361         if (IS_ERR(handle))
3362                 return PTR_ERR(handle);
3363
3364         err = ext3_mark_inode_dirty(handle, inode);
3365         handle->h_sync = 1;
3366         ext3_journal_stop(handle);
3367         ext3_std_error(inode->i_sb, err);
3368
3369         return err;
3370 }