sort out blockdev_direct_IO variants
[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 truncate_restart_transaction(handle_t *handle, struct inode *inode)
176 {
177         int ret;
178
179         jbd_debug(2, "restarting handle %p\n", handle);
180         /*
181          * Drop truncate_mutex to avoid deadlock with ext3_get_blocks_handle
182          * At this moment, get_block can be called only for blocks inside
183          * i_size since page cache has been already dropped and writes are
184          * blocked by i_mutex. So we can safely drop the truncate_mutex.
185          */
186         mutex_unlock(&EXT3_I(inode)->truncate_mutex);
187         ret = ext3_journal_restart(handle, blocks_for_truncate(inode));
188         mutex_lock(&EXT3_I(inode)->truncate_mutex);
189         return ret;
190 }
191
192 /*
193  * Called at the last iput() if i_nlink is zero.
194  */
195 void ext3_delete_inode (struct inode * inode)
196 {
197         handle_t *handle;
198
199         if (!is_bad_inode(inode))
200                 dquot_initialize(inode);
201
202         truncate_inode_pages(&inode->i_data, 0);
203
204         if (is_bad_inode(inode))
205                 goto no_delete;
206
207         handle = start_transaction(inode);
208         if (IS_ERR(handle)) {
209                 /*
210                  * If we're going to skip the normal cleanup, we still need to
211                  * make sure that the in-core orphan linked list is properly
212                  * cleaned up.
213                  */
214                 ext3_orphan_del(NULL, inode);
215                 goto no_delete;
216         }
217
218         if (IS_SYNC(inode))
219                 handle->h_sync = 1;
220         inode->i_size = 0;
221         if (inode->i_blocks)
222                 ext3_truncate(inode);
223         /*
224          * Kill off the orphan record which ext3_truncate created.
225          * AKPM: I think this can be inside the above `if'.
226          * Note that ext3_orphan_del() has to be able to cope with the
227          * deletion of a non-existent orphan - this is because we don't
228          * know if ext3_truncate() actually created an orphan record.
229          * (Well, we could do this if we need to, but heck - it works)
230          */
231         ext3_orphan_del(handle, inode);
232         EXT3_I(inode)->i_dtime  = get_seconds();
233
234         /*
235          * One subtle ordering requirement: if anything has gone wrong
236          * (transaction abort, IO errors, whatever), then we can still
237          * do these next steps (the fs will already have been marked as
238          * having errors), but we can't free the inode if the mark_dirty
239          * fails.
240          */
241         if (ext3_mark_inode_dirty(handle, inode))
242                 /* If that failed, just do the required in-core inode clear. */
243                 clear_inode(inode);
244         else
245                 ext3_free_inode(handle, inode);
246         ext3_journal_stop(handle);
247         return;
248 no_delete:
249         clear_inode(inode);     /* We must guarantee clearing of inode... */
250 }
251
252 typedef struct {
253         __le32  *p;
254         __le32  key;
255         struct buffer_head *bh;
256 } Indirect;
257
258 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
259 {
260         p->key = *(p->p = v);
261         p->bh = bh;
262 }
263
264 static int verify_chain(Indirect *from, Indirect *to)
265 {
266         while (from <= to && from->key == *from->p)
267                 from++;
268         return (from > to);
269 }
270
271 /**
272  *      ext3_block_to_path - parse the block number into array of offsets
273  *      @inode: inode in question (we are only interested in its superblock)
274  *      @i_block: block number to be parsed
275  *      @offsets: array to store the offsets in
276  *      @boundary: set this non-zero if the referred-to block is likely to be
277  *             followed (on disk) by an indirect block.
278  *
279  *      To store the locations of file's data ext3 uses a data structure common
280  *      for UNIX filesystems - tree of pointers anchored in the inode, with
281  *      data blocks at leaves and indirect blocks in intermediate nodes.
282  *      This function translates the block number into path in that tree -
283  *      return value is the path length and @offsets[n] is the offset of
284  *      pointer to (n+1)th node in the nth one. If @block is out of range
285  *      (negative or too large) warning is printed and zero returned.
286  *
287  *      Note: function doesn't find node addresses, so no IO is needed. All
288  *      we need to know is the capacity of indirect blocks (taken from the
289  *      inode->i_sb).
290  */
291
292 /*
293  * Portability note: the last comparison (check that we fit into triple
294  * indirect block) is spelled differently, because otherwise on an
295  * architecture with 32-bit longs and 8Kb pages we might get into trouble
296  * if our filesystem had 8Kb blocks. We might use long long, but that would
297  * kill us on x86. Oh, well, at least the sign propagation does not matter -
298  * i_block would have to be negative in the very beginning, so we would not
299  * get there at all.
300  */
301
302 static int ext3_block_to_path(struct inode *inode,
303                         long i_block, int offsets[4], int *boundary)
304 {
305         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
306         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
307         const long direct_blocks = EXT3_NDIR_BLOCKS,
308                 indirect_blocks = ptrs,
309                 double_blocks = (1 << (ptrs_bits * 2));
310         int n = 0;
311         int final = 0;
312
313         if (i_block < 0) {
314                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
315         } else if (i_block < direct_blocks) {
316                 offsets[n++] = i_block;
317                 final = direct_blocks;
318         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
319                 offsets[n++] = EXT3_IND_BLOCK;
320                 offsets[n++] = i_block;
321                 final = ptrs;
322         } else if ((i_block -= indirect_blocks) < double_blocks) {
323                 offsets[n++] = EXT3_DIND_BLOCK;
324                 offsets[n++] = i_block >> ptrs_bits;
325                 offsets[n++] = i_block & (ptrs - 1);
326                 final = ptrs;
327         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
328                 offsets[n++] = EXT3_TIND_BLOCK;
329                 offsets[n++] = i_block >> (ptrs_bits * 2);
330                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
331                 offsets[n++] = i_block & (ptrs - 1);
332                 final = ptrs;
333         } else {
334                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
335         }
336         if (boundary)
337                 *boundary = final - 1 - (i_block & (ptrs - 1));
338         return n;
339 }
340
341 /**
342  *      ext3_get_branch - read the chain of indirect blocks leading to data
343  *      @inode: inode in question
344  *      @depth: depth of the chain (1 - direct pointer, etc.)
345  *      @offsets: offsets of pointers in inode/indirect blocks
346  *      @chain: place to store the result
347  *      @err: here we store the error value
348  *
349  *      Function fills the array of triples <key, p, bh> and returns %NULL
350  *      if everything went OK or the pointer to the last filled triple
351  *      (incomplete one) otherwise. Upon the return chain[i].key contains
352  *      the number of (i+1)-th block in the chain (as it is stored in memory,
353  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
354  *      number (it points into struct inode for i==0 and into the bh->b_data
355  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
356  *      block for i>0 and NULL for i==0. In other words, it holds the block
357  *      numbers of the chain, addresses they were taken from (and where we can
358  *      verify that chain did not change) and buffer_heads hosting these
359  *      numbers.
360  *
361  *      Function stops when it stumbles upon zero pointer (absent block)
362  *              (pointer to last triple returned, *@err == 0)
363  *      or when it gets an IO error reading an indirect block
364  *              (ditto, *@err == -EIO)
365  *      or when it notices that chain had been changed while it was reading
366  *              (ditto, *@err == -EAGAIN)
367  *      or when it reads all @depth-1 indirect blocks successfully and finds
368  *      the whole chain, all way to the data (returns %NULL, *err == 0).
369  */
370 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
371                                  Indirect chain[4], int *err)
372 {
373         struct super_block *sb = inode->i_sb;
374         Indirect *p = chain;
375         struct buffer_head *bh;
376
377         *err = 0;
378         /* i_data is not going away, no lock needed */
379         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
380         if (!p->key)
381                 goto no_block;
382         while (--depth) {
383                 bh = sb_bread(sb, le32_to_cpu(p->key));
384                 if (!bh)
385                         goto failure;
386                 /* Reader: pointers */
387                 if (!verify_chain(chain, p))
388                         goto changed;
389                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
390                 /* Reader: end */
391                 if (!p->key)
392                         goto no_block;
393         }
394         return NULL;
395
396 changed:
397         brelse(bh);
398         *err = -EAGAIN;
399         goto no_block;
400 failure:
401         *err = -EIO;
402 no_block:
403         return p;
404 }
405
406 /**
407  *      ext3_find_near - find a place for allocation with sufficient locality
408  *      @inode: owner
409  *      @ind: descriptor of indirect block.
410  *
411  *      This function returns the preferred place for block allocation.
412  *      It is used when heuristic for sequential allocation fails.
413  *      Rules are:
414  *        + if there is a block to the left of our position - allocate near it.
415  *        + if pointer will live in indirect block - allocate near that block.
416  *        + if pointer will live in inode - allocate in the same
417  *          cylinder group.
418  *
419  * In the latter case we colour the starting block by the callers PID to
420  * prevent it from clashing with concurrent allocations for a different inode
421  * in the same block group.   The PID is used here so that functionally related
422  * files will be close-by on-disk.
423  *
424  *      Caller must make sure that @ind is valid and will stay that way.
425  */
426 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
427 {
428         struct ext3_inode_info *ei = EXT3_I(inode);
429         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
430         __le32 *p;
431         ext3_fsblk_t bg_start;
432         ext3_grpblk_t colour;
433
434         /* Try to find previous block */
435         for (p = ind->p - 1; p >= start; p--) {
436                 if (*p)
437                         return le32_to_cpu(*p);
438         }
439
440         /* No such thing, so let's try location of indirect block */
441         if (ind->bh)
442                 return ind->bh->b_blocknr;
443
444         /*
445          * It is going to be referred to from the inode itself? OK, just put it
446          * into the same cylinder group then.
447          */
448         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
449         colour = (current->pid % 16) *
450                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
451         return bg_start + colour;
452 }
453
454 /**
455  *      ext3_find_goal - find a preferred place for allocation.
456  *      @inode: owner
457  *      @block:  block we want
458  *      @partial: pointer to the last triple within a chain
459  *
460  *      Normally this function find the preferred place for block allocation,
461  *      returns it.
462  */
463
464 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
465                                    Indirect *partial)
466 {
467         struct ext3_block_alloc_info *block_i;
468
469         block_i =  EXT3_I(inode)->i_block_alloc_info;
470
471         /*
472          * try the heuristic for sequential allocation,
473          * failing that at least try to get decent locality.
474          */
475         if (block_i && (block == block_i->last_alloc_logical_block + 1)
476                 && (block_i->last_alloc_physical_block != 0)) {
477                 return block_i->last_alloc_physical_block + 1;
478         }
479
480         return ext3_find_near(inode, partial);
481 }
482
483 /**
484  *      ext3_blks_to_allocate: Look up the block map and count the number
485  *      of direct blocks need to be allocated for the given branch.
486  *
487  *      @branch: chain of indirect blocks
488  *      @k: number of blocks need for indirect blocks
489  *      @blks: number of data blocks to be mapped.
490  *      @blocks_to_boundary:  the offset in the indirect block
491  *
492  *      return the total number of blocks to be allocate, including the
493  *      direct and indirect blocks.
494  */
495 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
496                 int blocks_to_boundary)
497 {
498         unsigned long count = 0;
499
500         /*
501          * Simple case, [t,d]Indirect block(s) has not allocated yet
502          * then it's clear blocks on that path have not allocated
503          */
504         if (k > 0) {
505                 /* right now we don't handle cross boundary allocation */
506                 if (blks < blocks_to_boundary + 1)
507                         count += blks;
508                 else
509                         count += blocks_to_boundary + 1;
510                 return count;
511         }
512
513         count++;
514         while (count < blks && count <= blocks_to_boundary &&
515                 le32_to_cpu(*(branch[0].p + count)) == 0) {
516                 count++;
517         }
518         return count;
519 }
520
521 /**
522  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
523  *      @indirect_blks: the number of blocks need to allocate for indirect
524  *                      blocks
525  *
526  *      @new_blocks: on return it will store the new block numbers for
527  *      the indirect blocks(if needed) and the first direct block,
528  *      @blks:  on return it will store the total number of allocated
529  *              direct blocks
530  */
531 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
532                         ext3_fsblk_t goal, int indirect_blks, int blks,
533                         ext3_fsblk_t new_blocks[4], int *err)
534 {
535         int target, i;
536         unsigned long count = 0;
537         int index = 0;
538         ext3_fsblk_t current_block = 0;
539         int ret = 0;
540
541         /*
542          * Here we try to allocate the requested multiple blocks at once,
543          * on a best-effort basis.
544          * To build a branch, we should allocate blocks for
545          * the indirect blocks(if not allocated yet), and at least
546          * the first direct block of this branch.  That's the
547          * minimum number of blocks need to allocate(required)
548          */
549         target = blks + indirect_blks;
550
551         while (1) {
552                 count = target;
553                 /* allocating blocks for indirect blocks and direct blocks */
554                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
555                 if (*err)
556                         goto failed_out;
557
558                 target -= count;
559                 /* allocate blocks for indirect blocks */
560                 while (index < indirect_blks && count) {
561                         new_blocks[index++] = current_block++;
562                         count--;
563                 }
564
565                 if (count > 0)
566                         break;
567         }
568
569         /* save the new block number for the first direct block */
570         new_blocks[index] = current_block;
571
572         /* total number of blocks allocated for direct blocks */
573         ret = count;
574         *err = 0;
575         return ret;
576 failed_out:
577         for (i = 0; i <index; i++)
578                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
579         return ret;
580 }
581
582 /**
583  *      ext3_alloc_branch - allocate and set up a chain of blocks.
584  *      @inode: owner
585  *      @indirect_blks: number of allocated indirect blocks
586  *      @blks: number of allocated direct blocks
587  *      @offsets: offsets (in the blocks) to store the pointers to next.
588  *      @branch: place to store the chain in.
589  *
590  *      This function allocates blocks, zeroes out all but the last one,
591  *      links them into chain and (if we are synchronous) writes them to disk.
592  *      In other words, it prepares a branch that can be spliced onto the
593  *      inode. It stores the information about that chain in the branch[], in
594  *      the same format as ext3_get_branch() would do. We are calling it after
595  *      we had read the existing part of chain and partial points to the last
596  *      triple of that (one with zero ->key). Upon the exit we have the same
597  *      picture as after the successful ext3_get_block(), except that in one
598  *      place chain is disconnected - *branch->p is still zero (we did not
599  *      set the last link), but branch->key contains the number that should
600  *      be placed into *branch->p to fill that gap.
601  *
602  *      If allocation fails we free all blocks we've allocated (and forget
603  *      their buffer_heads) and return the error value the from failed
604  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
605  *      as described above and return 0.
606  */
607 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
608                         int indirect_blks, int *blks, ext3_fsblk_t goal,
609                         int *offsets, Indirect *branch)
610 {
611         int blocksize = inode->i_sb->s_blocksize;
612         int i, n = 0;
613         int err = 0;
614         struct buffer_head *bh;
615         int num;
616         ext3_fsblk_t new_blocks[4];
617         ext3_fsblk_t current_block;
618
619         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
620                                 *blks, new_blocks, &err);
621         if (err)
622                 return err;
623
624         branch[0].key = cpu_to_le32(new_blocks[0]);
625         /*
626          * metadata blocks and data blocks are allocated.
627          */
628         for (n = 1; n <= indirect_blks;  n++) {
629                 /*
630                  * Get buffer_head for parent block, zero it out
631                  * and set the pointer to new one, then send
632                  * parent to disk.
633                  */
634                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
635                 branch[n].bh = bh;
636                 lock_buffer(bh);
637                 BUFFER_TRACE(bh, "call get_create_access");
638                 err = ext3_journal_get_create_access(handle, bh);
639                 if (err) {
640                         unlock_buffer(bh);
641                         brelse(bh);
642                         goto failed;
643                 }
644
645                 memset(bh->b_data, 0, blocksize);
646                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
647                 branch[n].key = cpu_to_le32(new_blocks[n]);
648                 *branch[n].p = branch[n].key;
649                 if ( n == indirect_blks) {
650                         current_block = new_blocks[n];
651                         /*
652                          * End of chain, update the last new metablock of
653                          * the chain to point to the new allocated
654                          * data blocks numbers
655                          */
656                         for (i=1; i < num; i++)
657                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
658                 }
659                 BUFFER_TRACE(bh, "marking uptodate");
660                 set_buffer_uptodate(bh);
661                 unlock_buffer(bh);
662
663                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
664                 err = ext3_journal_dirty_metadata(handle, bh);
665                 if (err)
666                         goto failed;
667         }
668         *blks = num;
669         return err;
670 failed:
671         /* Allocation failed, free what we already allocated */
672         for (i = 1; i <= n ; i++) {
673                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
674                 ext3_journal_forget(handle, branch[i].bh);
675         }
676         for (i = 0; i <indirect_blks; i++)
677                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
678
679         ext3_free_blocks(handle, inode, new_blocks[i], num);
680
681         return err;
682 }
683
684 /**
685  * ext3_splice_branch - splice the allocated branch onto inode.
686  * @inode: owner
687  * @block: (logical) number of block we are adding
688  * @chain: chain of indirect blocks (with a missing link - see
689  *      ext3_alloc_branch)
690  * @where: location of missing link
691  * @num:   number of indirect blocks we are adding
692  * @blks:  number of direct blocks we are adding
693  *
694  * This function fills the missing link and does all housekeeping needed in
695  * inode (->i_blocks, etc.). In case of success we end up with the full
696  * chain to new block and return 0.
697  */
698 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
699                         long block, Indirect *where, int num, int blks)
700 {
701         int i;
702         int err = 0;
703         struct ext3_block_alloc_info *block_i;
704         ext3_fsblk_t current_block;
705         struct ext3_inode_info *ei = EXT3_I(inode);
706
707         block_i = ei->i_block_alloc_info;
708         /*
709          * If we're splicing into a [td]indirect block (as opposed to the
710          * inode) then we need to get write access to the [td]indirect block
711          * before the splice.
712          */
713         if (where->bh) {
714                 BUFFER_TRACE(where->bh, "get_write_access");
715                 err = ext3_journal_get_write_access(handle, where->bh);
716                 if (err)
717                         goto err_out;
718         }
719         /* That's it */
720
721         *where->p = where->key;
722
723         /*
724          * Update the host buffer_head or inode to point to more just allocated
725          * direct blocks blocks
726          */
727         if (num == 0 && blks > 1) {
728                 current_block = le32_to_cpu(where->key) + 1;
729                 for (i = 1; i < blks; i++)
730                         *(where->p + i ) = cpu_to_le32(current_block++);
731         }
732
733         /*
734          * update the most recently allocated logical & physical block
735          * in i_block_alloc_info, to assist find the proper goal block for next
736          * allocation
737          */
738         if (block_i) {
739                 block_i->last_alloc_logical_block = block + blks - 1;
740                 block_i->last_alloc_physical_block =
741                                 le32_to_cpu(where[num].key) + blks - 1;
742         }
743
744         /* We are done with atomic stuff, now do the rest of housekeeping */
745
746         inode->i_ctime = CURRENT_TIME_SEC;
747         ext3_mark_inode_dirty(handle, inode);
748         /* ext3_mark_inode_dirty already updated i_sync_tid */
749         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
750
751         /* had we spliced it onto indirect block? */
752         if (where->bh) {
753                 /*
754                  * If we spliced it onto an indirect block, we haven't
755                  * altered the inode.  Note however that if it is being spliced
756                  * onto an indirect block at the very end of the file (the
757                  * file is growing) then we *will* alter the inode to reflect
758                  * the new i_size.  But that is not done here - it is done in
759                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
760                  */
761                 jbd_debug(5, "splicing indirect only\n");
762                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
763                 err = ext3_journal_dirty_metadata(handle, where->bh);
764                 if (err)
765                         goto err_out;
766         } else {
767                 /*
768                  * OK, we spliced it into the inode itself on a direct block.
769                  * Inode was dirtied above.
770                  */
771                 jbd_debug(5, "splicing direct\n");
772         }
773         return err;
774
775 err_out:
776         for (i = 1; i <= num; i++) {
777                 BUFFER_TRACE(where[i].bh, "call journal_forget");
778                 ext3_journal_forget(handle, where[i].bh);
779                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
780         }
781         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
782
783         return err;
784 }
785
786 /*
787  * Allocation strategy is simple: if we have to allocate something, we will
788  * have to go the whole way to leaf. So let's do it before attaching anything
789  * to tree, set linkage between the newborn blocks, write them if sync is
790  * required, recheck the path, free and repeat if check fails, otherwise
791  * set the last missing link (that will protect us from any truncate-generated
792  * removals - all blocks on the path are immune now) and possibly force the
793  * write on the parent block.
794  * That has a nice additional property: no special recovery from the failed
795  * allocations is needed - we simply release blocks and do not touch anything
796  * reachable from inode.
797  *
798  * `handle' can be NULL if create == 0.
799  *
800  * The BKL may not be held on entry here.  Be sure to take it early.
801  * return > 0, # of blocks mapped or allocated.
802  * return = 0, if plain lookup failed.
803  * return < 0, error case.
804  */
805 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
806                 sector_t iblock, unsigned long maxblocks,
807                 struct buffer_head *bh_result,
808                 int create)
809 {
810         int err = -EIO;
811         int offsets[4];
812         Indirect chain[4];
813         Indirect *partial;
814         ext3_fsblk_t goal;
815         int indirect_blks;
816         int blocks_to_boundary = 0;
817         int depth;
818         struct ext3_inode_info *ei = EXT3_I(inode);
819         int count = 0;
820         ext3_fsblk_t first_block = 0;
821
822
823         J_ASSERT(handle != NULL || create == 0);
824         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
825
826         if (depth == 0)
827                 goto out;
828
829         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
830
831         /* Simplest case - block found, no allocation needed */
832         if (!partial) {
833                 first_block = le32_to_cpu(chain[depth - 1].key);
834                 clear_buffer_new(bh_result);
835                 count++;
836                 /*map more blocks*/
837                 while (count < maxblocks && count <= blocks_to_boundary) {
838                         ext3_fsblk_t blk;
839
840                         if (!verify_chain(chain, chain + depth - 1)) {
841                                 /*
842                                  * Indirect block might be removed by
843                                  * truncate while we were reading it.
844                                  * Handling of that case: forget what we've
845                                  * got now. Flag the err as EAGAIN, so it
846                                  * will reread.
847                                  */
848                                 err = -EAGAIN;
849                                 count = 0;
850                                 break;
851                         }
852                         blk = le32_to_cpu(*(chain[depth-1].p + count));
853
854                         if (blk == first_block + count)
855                                 count++;
856                         else
857                                 break;
858                 }
859                 if (err != -EAGAIN)
860                         goto got_it;
861         }
862
863         /* Next simple case - plain lookup or failed read of indirect block */
864         if (!create || err == -EIO)
865                 goto cleanup;
866
867         mutex_lock(&ei->truncate_mutex);
868
869         /*
870          * If the indirect block is missing while we are reading
871          * the chain(ext3_get_branch() returns -EAGAIN err), or
872          * if the chain has been changed after we grab the semaphore,
873          * (either because another process truncated this branch, or
874          * another get_block allocated this branch) re-grab the chain to see if
875          * the request block has been allocated or not.
876          *
877          * Since we already block the truncate/other get_block
878          * at this point, we will have the current copy of the chain when we
879          * splice the branch into the tree.
880          */
881         if (err == -EAGAIN || !verify_chain(chain, partial)) {
882                 while (partial > chain) {
883                         brelse(partial->bh);
884                         partial--;
885                 }
886                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
887                 if (!partial) {
888                         count++;
889                         mutex_unlock(&ei->truncate_mutex);
890                         if (err)
891                                 goto cleanup;
892                         clear_buffer_new(bh_result);
893                         goto got_it;
894                 }
895         }
896
897         /*
898          * Okay, we need to do block allocation.  Lazily initialize the block
899          * allocation info here if necessary
900         */
901         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
902                 ext3_init_block_alloc_info(inode);
903
904         goal = ext3_find_goal(inode, iblock, partial);
905
906         /* the number of blocks need to allocate for [d,t]indirect blocks */
907         indirect_blks = (chain + depth) - partial - 1;
908
909         /*
910          * Next look up the indirect map to count the totoal number of
911          * direct blocks to allocate for this branch.
912          */
913         count = ext3_blks_to_allocate(partial, indirect_blks,
914                                         maxblocks, blocks_to_boundary);
915         /*
916          * Block out ext3_truncate while we alter the tree
917          */
918         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
919                                 offsets + (partial - chain), partial);
920
921         /*
922          * The ext3_splice_branch call will free and forget any buffers
923          * on the new chain if there is a failure, but that risks using
924          * up transaction credits, especially for bitmaps where the
925          * credits cannot be returned.  Can we handle this somehow?  We
926          * may need to return -EAGAIN upwards in the worst case.  --sct
927          */
928         if (!err)
929                 err = ext3_splice_branch(handle, inode, iblock,
930                                         partial, indirect_blks, count);
931         mutex_unlock(&ei->truncate_mutex);
932         if (err)
933                 goto cleanup;
934
935         set_buffer_new(bh_result);
936 got_it:
937         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
938         if (count > blocks_to_boundary)
939                 set_buffer_boundary(bh_result);
940         err = count;
941         /* Clean up and exit */
942         partial = chain + depth - 1;    /* the whole chain */
943 cleanup:
944         while (partial > chain) {
945                 BUFFER_TRACE(partial->bh, "call brelse");
946                 brelse(partial->bh);
947                 partial--;
948         }
949         BUFFER_TRACE(bh_result, "returned");
950 out:
951         return err;
952 }
953
954 /* Maximum number of blocks we map for direct IO at once. */
955 #define DIO_MAX_BLOCKS 4096
956 /*
957  * Number of credits we need for writing DIO_MAX_BLOCKS:
958  * We need sb + group descriptor + bitmap + inode -> 4
959  * For B blocks with A block pointers per block we need:
960  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
961  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
962  */
963 #define DIO_CREDITS 25
964
965 static int ext3_get_block(struct inode *inode, sector_t iblock,
966                         struct buffer_head *bh_result, int create)
967 {
968         handle_t *handle = ext3_journal_current_handle();
969         int ret = 0, started = 0;
970         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
971
972         if (create && !handle) {        /* Direct IO write... */
973                 if (max_blocks > DIO_MAX_BLOCKS)
974                         max_blocks = DIO_MAX_BLOCKS;
975                 handle = ext3_journal_start(inode, DIO_CREDITS +
976                                 EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb));
977                 if (IS_ERR(handle)) {
978                         ret = PTR_ERR(handle);
979                         goto out;
980                 }
981                 started = 1;
982         }
983
984         ret = ext3_get_blocks_handle(handle, inode, iblock,
985                                         max_blocks, bh_result, create);
986         if (ret > 0) {
987                 bh_result->b_size = (ret << inode->i_blkbits);
988                 ret = 0;
989         }
990         if (started)
991                 ext3_journal_stop(handle);
992 out:
993         return ret;
994 }
995
996 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
997                 u64 start, u64 len)
998 {
999         return generic_block_fiemap(inode, fieinfo, start, len,
1000                                     ext3_get_block);
1001 }
1002
1003 /*
1004  * `handle' can be NULL if create is zero
1005  */
1006 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1007                                 long block, int create, int *errp)
1008 {
1009         struct buffer_head dummy;
1010         int fatal = 0, err;
1011
1012         J_ASSERT(handle != NULL || create == 0);
1013
1014         dummy.b_state = 0;
1015         dummy.b_blocknr = -1000;
1016         buffer_trace_init(&dummy.b_history);
1017         err = ext3_get_blocks_handle(handle, inode, block, 1,
1018                                         &dummy, create);
1019         /*
1020          * ext3_get_blocks_handle() returns number of blocks
1021          * mapped. 0 in case of a HOLE.
1022          */
1023         if (err > 0) {
1024                 if (err > 1)
1025                         WARN_ON(1);
1026                 err = 0;
1027         }
1028         *errp = err;
1029         if (!err && buffer_mapped(&dummy)) {
1030                 struct buffer_head *bh;
1031                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1032                 if (!bh) {
1033                         *errp = -EIO;
1034                         goto err;
1035                 }
1036                 if (buffer_new(&dummy)) {
1037                         J_ASSERT(create != 0);
1038                         J_ASSERT(handle != NULL);
1039
1040                         /*
1041                          * Now that we do not always journal data, we should
1042                          * keep in mind whether this should always journal the
1043                          * new buffer as metadata.  For now, regular file
1044                          * writes use ext3_get_block instead, so it's not a
1045                          * problem.
1046                          */
1047                         lock_buffer(bh);
1048                         BUFFER_TRACE(bh, "call get_create_access");
1049                         fatal = ext3_journal_get_create_access(handle, bh);
1050                         if (!fatal && !buffer_uptodate(bh)) {
1051                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1052                                 set_buffer_uptodate(bh);
1053                         }
1054                         unlock_buffer(bh);
1055                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1056                         err = ext3_journal_dirty_metadata(handle, bh);
1057                         if (!fatal)
1058                                 fatal = err;
1059                 } else {
1060                         BUFFER_TRACE(bh, "not a new buffer");
1061                 }
1062                 if (fatal) {
1063                         *errp = fatal;
1064                         brelse(bh);
1065                         bh = NULL;
1066                 }
1067                 return bh;
1068         }
1069 err:
1070         return NULL;
1071 }
1072
1073 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1074                                int block, int create, int *err)
1075 {
1076         struct buffer_head * bh;
1077
1078         bh = ext3_getblk(handle, inode, block, create, err);
1079         if (!bh)
1080                 return bh;
1081         if (buffer_uptodate(bh))
1082                 return bh;
1083         ll_rw_block(READ_META, 1, &bh);
1084         wait_on_buffer(bh);
1085         if (buffer_uptodate(bh))
1086                 return bh;
1087         put_bh(bh);
1088         *err = -EIO;
1089         return NULL;
1090 }
1091
1092 static int walk_page_buffers(   handle_t *handle,
1093                                 struct buffer_head *head,
1094                                 unsigned from,
1095                                 unsigned to,
1096                                 int *partial,
1097                                 int (*fn)(      handle_t *handle,
1098                                                 struct buffer_head *bh))
1099 {
1100         struct buffer_head *bh;
1101         unsigned block_start, block_end;
1102         unsigned blocksize = head->b_size;
1103         int err, ret = 0;
1104         struct buffer_head *next;
1105
1106         for (   bh = head, block_start = 0;
1107                 ret == 0 && (bh != head || !block_start);
1108                 block_start = block_end, bh = next)
1109         {
1110                 next = bh->b_this_page;
1111                 block_end = block_start + blocksize;
1112                 if (block_end <= from || block_start >= to) {
1113                         if (partial && !buffer_uptodate(bh))
1114                                 *partial = 1;
1115                         continue;
1116                 }
1117                 err = (*fn)(handle, bh);
1118                 if (!ret)
1119                         ret = err;
1120         }
1121         return ret;
1122 }
1123
1124 /*
1125  * To preserve ordering, it is essential that the hole instantiation and
1126  * the data write be encapsulated in a single transaction.  We cannot
1127  * close off a transaction and start a new one between the ext3_get_block()
1128  * and the commit_write().  So doing the journal_start at the start of
1129  * prepare_write() is the right place.
1130  *
1131  * Also, this function can nest inside ext3_writepage() ->
1132  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1133  * has generated enough buffer credits to do the whole page.  So we won't
1134  * block on the journal in that case, which is good, because the caller may
1135  * be PF_MEMALLOC.
1136  *
1137  * By accident, ext3 can be reentered when a transaction is open via
1138  * quota file writes.  If we were to commit the transaction while thus
1139  * reentered, there can be a deadlock - we would be holding a quota
1140  * lock, and the commit would never complete if another thread had a
1141  * transaction open and was blocking on the quota lock - a ranking
1142  * violation.
1143  *
1144  * So what we do is to rely on the fact that journal_stop/journal_start
1145  * will _not_ run commit under these circumstances because handle->h_ref
1146  * is elevated.  We'll still have enough credits for the tiny quotafile
1147  * write.
1148  */
1149 static int do_journal_get_write_access(handle_t *handle,
1150                                         struct buffer_head *bh)
1151 {
1152         if (!buffer_mapped(bh) || buffer_freed(bh))
1153                 return 0;
1154         return ext3_journal_get_write_access(handle, bh);
1155 }
1156
1157 /*
1158  * Truncate blocks that were not used by write. We have to truncate the
1159  * pagecache as well so that corresponding buffers get properly unmapped.
1160  */
1161 static void ext3_truncate_failed_write(struct inode *inode)
1162 {
1163         truncate_inode_pages(inode->i_mapping, inode->i_size);
1164         ext3_truncate(inode);
1165 }
1166
1167 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1168                                 loff_t pos, unsigned len, unsigned flags,
1169                                 struct page **pagep, void **fsdata)
1170 {
1171         struct inode *inode = mapping->host;
1172         int ret;
1173         handle_t *handle;
1174         int retries = 0;
1175         struct page *page;
1176         pgoff_t index;
1177         unsigned from, to;
1178         /* Reserve one block more for addition to orphan list in case
1179          * we allocate blocks but write fails for some reason */
1180         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1181
1182         index = pos >> PAGE_CACHE_SHIFT;
1183         from = pos & (PAGE_CACHE_SIZE - 1);
1184         to = from + len;
1185
1186 retry:
1187         page = grab_cache_page_write_begin(mapping, index, flags);
1188         if (!page)
1189                 return -ENOMEM;
1190         *pagep = page;
1191
1192         handle = ext3_journal_start(inode, needed_blocks);
1193         if (IS_ERR(handle)) {
1194                 unlock_page(page);
1195                 page_cache_release(page);
1196                 ret = PTR_ERR(handle);
1197                 goto out;
1198         }
1199         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1200                                                         ext3_get_block);
1201         if (ret)
1202                 goto write_begin_failed;
1203
1204         if (ext3_should_journal_data(inode)) {
1205                 ret = walk_page_buffers(handle, page_buffers(page),
1206                                 from, to, NULL, do_journal_get_write_access);
1207         }
1208 write_begin_failed:
1209         if (ret) {
1210                 /*
1211                  * block_write_begin may have instantiated a few blocks
1212                  * outside i_size.  Trim these off again. Don't need
1213                  * i_size_read because we hold i_mutex.
1214                  *
1215                  * Add inode to orphan list in case we crash before truncate
1216                  * finishes. Do this only if ext3_can_truncate() agrees so
1217                  * that orphan processing code is happy.
1218                  */
1219                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1220                         ext3_orphan_add(handle, inode);
1221                 ext3_journal_stop(handle);
1222                 unlock_page(page);
1223                 page_cache_release(page);
1224                 if (pos + len > inode->i_size)
1225                         ext3_truncate_failed_write(inode);
1226         }
1227         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1228                 goto retry;
1229 out:
1230         return ret;
1231 }
1232
1233
1234 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1235 {
1236         int err = journal_dirty_data(handle, bh);
1237         if (err)
1238                 ext3_journal_abort_handle(__func__, __func__,
1239                                                 bh, handle, err);
1240         return err;
1241 }
1242
1243 /* For ordered writepage and write_end functions */
1244 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1245 {
1246         /*
1247          * Write could have mapped the buffer but it didn't copy the data in
1248          * yet. So avoid filing such buffer into a transaction.
1249          */
1250         if (buffer_mapped(bh) && buffer_uptodate(bh))
1251                 return ext3_journal_dirty_data(handle, bh);
1252         return 0;
1253 }
1254
1255 /* For write_end() in data=journal mode */
1256 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1257 {
1258         if (!buffer_mapped(bh) || buffer_freed(bh))
1259                 return 0;
1260         set_buffer_uptodate(bh);
1261         return ext3_journal_dirty_metadata(handle, bh);
1262 }
1263
1264 /*
1265  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1266  * for the whole page but later we failed to copy the data in. Update inode
1267  * size according to what we managed to copy. The rest is going to be
1268  * truncated in write_end function.
1269  */
1270 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1271 {
1272         /* What matters to us is i_disksize. We don't write i_size anywhere */
1273         if (pos + copied > inode->i_size)
1274                 i_size_write(inode, pos + copied);
1275         if (pos + copied > EXT3_I(inode)->i_disksize) {
1276                 EXT3_I(inode)->i_disksize = pos + copied;
1277                 mark_inode_dirty(inode);
1278         }
1279 }
1280
1281 /*
1282  * We need to pick up the new inode size which generic_commit_write gave us
1283  * `file' can be NULL - eg, when called from page_symlink().
1284  *
1285  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1286  * buffers are managed internally.
1287  */
1288 static int ext3_ordered_write_end(struct file *file,
1289                                 struct address_space *mapping,
1290                                 loff_t pos, unsigned len, unsigned copied,
1291                                 struct page *page, void *fsdata)
1292 {
1293         handle_t *handle = ext3_journal_current_handle();
1294         struct inode *inode = file->f_mapping->host;
1295         unsigned from, to;
1296         int ret = 0, ret2;
1297
1298         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1299
1300         from = pos & (PAGE_CACHE_SIZE - 1);
1301         to = from + copied;
1302         ret = walk_page_buffers(handle, page_buffers(page),
1303                 from, to, NULL, journal_dirty_data_fn);
1304
1305         if (ret == 0)
1306                 update_file_sizes(inode, pos, copied);
1307         /*
1308          * There may be allocated blocks outside of i_size because
1309          * we failed to copy some data. Prepare for truncate.
1310          */
1311         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1312                 ext3_orphan_add(handle, inode);
1313         ret2 = ext3_journal_stop(handle);
1314         if (!ret)
1315                 ret = ret2;
1316         unlock_page(page);
1317         page_cache_release(page);
1318
1319         if (pos + len > inode->i_size)
1320                 ext3_truncate_failed_write(inode);
1321         return ret ? ret : copied;
1322 }
1323
1324 static int ext3_writeback_write_end(struct file *file,
1325                                 struct address_space *mapping,
1326                                 loff_t pos, unsigned len, unsigned copied,
1327                                 struct page *page, void *fsdata)
1328 {
1329         handle_t *handle = ext3_journal_current_handle();
1330         struct inode *inode = file->f_mapping->host;
1331         int ret;
1332
1333         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1334         update_file_sizes(inode, pos, copied);
1335         /*
1336          * There may be allocated blocks outside of i_size because
1337          * we failed to copy some data. Prepare for truncate.
1338          */
1339         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1340                 ext3_orphan_add(handle, inode);
1341         ret = ext3_journal_stop(handle);
1342         unlock_page(page);
1343         page_cache_release(page);
1344
1345         if (pos + len > inode->i_size)
1346                 ext3_truncate_failed_write(inode);
1347         return ret ? ret : copied;
1348 }
1349
1350 static int ext3_journalled_write_end(struct file *file,
1351                                 struct address_space *mapping,
1352                                 loff_t pos, unsigned len, unsigned copied,
1353                                 struct page *page, void *fsdata)
1354 {
1355         handle_t *handle = ext3_journal_current_handle();
1356         struct inode *inode = mapping->host;
1357         int ret = 0, ret2;
1358         int partial = 0;
1359         unsigned from, to;
1360
1361         from = pos & (PAGE_CACHE_SIZE - 1);
1362         to = from + len;
1363
1364         if (copied < len) {
1365                 if (!PageUptodate(page))
1366                         copied = 0;
1367                 page_zero_new_buffers(page, from + copied, to);
1368                 to = from + copied;
1369         }
1370
1371         ret = walk_page_buffers(handle, page_buffers(page), from,
1372                                 to, &partial, write_end_fn);
1373         if (!partial)
1374                 SetPageUptodate(page);
1375
1376         if (pos + copied > inode->i_size)
1377                 i_size_write(inode, pos + copied);
1378         /*
1379          * There may be allocated blocks outside of i_size because
1380          * we failed to copy some data. Prepare for truncate.
1381          */
1382         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1383                 ext3_orphan_add(handle, inode);
1384         ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1385         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1386                 EXT3_I(inode)->i_disksize = inode->i_size;
1387                 ret2 = ext3_mark_inode_dirty(handle, inode);
1388                 if (!ret)
1389                         ret = ret2;
1390         }
1391
1392         ret2 = ext3_journal_stop(handle);
1393         if (!ret)
1394                 ret = ret2;
1395         unlock_page(page);
1396         page_cache_release(page);
1397
1398         if (pos + len > inode->i_size)
1399                 ext3_truncate_failed_write(inode);
1400         return ret ? ret : copied;
1401 }
1402
1403 /*
1404  * bmap() is special.  It gets used by applications such as lilo and by
1405  * the swapper to find the on-disk block of a specific piece of data.
1406  *
1407  * Naturally, this is dangerous if the block concerned is still in the
1408  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1409  * filesystem and enables swap, then they may get a nasty shock when the
1410  * data getting swapped to that swapfile suddenly gets overwritten by
1411  * the original zero's written out previously to the journal and
1412  * awaiting writeback in the kernel's buffer cache.
1413  *
1414  * So, if we see any bmap calls here on a modified, data-journaled file,
1415  * take extra steps to flush any blocks which might be in the cache.
1416  */
1417 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1418 {
1419         struct inode *inode = mapping->host;
1420         journal_t *journal;
1421         int err;
1422
1423         if (ext3_test_inode_state(inode, EXT3_STATE_JDATA)) {
1424                 /*
1425                  * This is a REALLY heavyweight approach, but the use of
1426                  * bmap on dirty files is expected to be extremely rare:
1427                  * only if we run lilo or swapon on a freshly made file
1428                  * do we expect this to happen.
1429                  *
1430                  * (bmap requires CAP_SYS_RAWIO so this does not
1431                  * represent an unprivileged user DOS attack --- we'd be
1432                  * in trouble if mortal users could trigger this path at
1433                  * will.)
1434                  *
1435                  * NB. EXT3_STATE_JDATA is not set on files other than
1436                  * regular files.  If somebody wants to bmap a directory
1437                  * or symlink and gets confused because the buffer
1438                  * hasn't yet been flushed to disk, they deserve
1439                  * everything they get.
1440                  */
1441
1442                 ext3_clear_inode_state(inode, EXT3_STATE_JDATA);
1443                 journal = EXT3_JOURNAL(inode);
1444                 journal_lock_updates(journal);
1445                 err = journal_flush(journal);
1446                 journal_unlock_updates(journal);
1447
1448                 if (err)
1449                         return 0;
1450         }
1451
1452         return generic_block_bmap(mapping,block,ext3_get_block);
1453 }
1454
1455 static int bget_one(handle_t *handle, struct buffer_head *bh)
1456 {
1457         get_bh(bh);
1458         return 0;
1459 }
1460
1461 static int bput_one(handle_t *handle, struct buffer_head *bh)
1462 {
1463         put_bh(bh);
1464         return 0;
1465 }
1466
1467 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1468 {
1469         return !buffer_mapped(bh);
1470 }
1471
1472 /*
1473  * Note that we always start a transaction even if we're not journalling
1474  * data.  This is to preserve ordering: any hole instantiation within
1475  * __block_write_full_page -> ext3_get_block() should be journalled
1476  * along with the data so we don't crash and then get metadata which
1477  * refers to old data.
1478  *
1479  * In all journalling modes block_write_full_page() will start the I/O.
1480  *
1481  * Problem:
1482  *
1483  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1484  *              ext3_writepage()
1485  *
1486  * Similar for:
1487  *
1488  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1489  *
1490  * Same applies to ext3_get_block().  We will deadlock on various things like
1491  * lock_journal and i_truncate_mutex.
1492  *
1493  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1494  * allocations fail.
1495  *
1496  * 16May01: If we're reentered then journal_current_handle() will be
1497  *          non-zero. We simply *return*.
1498  *
1499  * 1 July 2001: @@@ FIXME:
1500  *   In journalled data mode, a data buffer may be metadata against the
1501  *   current transaction.  But the same file is part of a shared mapping
1502  *   and someone does a writepage() on it.
1503  *
1504  *   We will move the buffer onto the async_data list, but *after* it has
1505  *   been dirtied. So there's a small window where we have dirty data on
1506  *   BJ_Metadata.
1507  *
1508  *   Note that this only applies to the last partial page in the file.  The
1509  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1510  *   broken code anyway: it's wrong for msync()).
1511  *
1512  *   It's a rare case: affects the final partial page, for journalled data
1513  *   where the file is subject to bith write() and writepage() in the same
1514  *   transction.  To fix it we'll need a custom block_write_full_page().
1515  *   We'll probably need that anyway for journalling writepage() output.
1516  *
1517  * We don't honour synchronous mounts for writepage().  That would be
1518  * disastrous.  Any write() or metadata operation will sync the fs for
1519  * us.
1520  *
1521  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1522  * we don't need to open a transaction here.
1523  */
1524 static int ext3_ordered_writepage(struct page *page,
1525                                 struct writeback_control *wbc)
1526 {
1527         struct inode *inode = page->mapping->host;
1528         struct buffer_head *page_bufs;
1529         handle_t *handle = NULL;
1530         int ret = 0;
1531         int err;
1532
1533         J_ASSERT(PageLocked(page));
1534         WARN_ON_ONCE(IS_RDONLY(inode));
1535
1536         /*
1537          * We give up here if we're reentered, because it might be for a
1538          * different filesystem.
1539          */
1540         if (ext3_journal_current_handle())
1541                 goto out_fail;
1542
1543         if (!page_has_buffers(page)) {
1544                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1545                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1546                 page_bufs = page_buffers(page);
1547         } else {
1548                 page_bufs = page_buffers(page);
1549                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1550                                        NULL, buffer_unmapped)) {
1551                         /* Provide NULL get_block() to catch bugs if buffers
1552                          * weren't really mapped */
1553                         return block_write_full_page(page, NULL, wbc);
1554                 }
1555         }
1556         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1557
1558         if (IS_ERR(handle)) {
1559                 ret = PTR_ERR(handle);
1560                 goto out_fail;
1561         }
1562
1563         walk_page_buffers(handle, page_bufs, 0,
1564                         PAGE_CACHE_SIZE, NULL, bget_one);
1565
1566         ret = block_write_full_page(page, ext3_get_block, wbc);
1567
1568         /*
1569          * The page can become unlocked at any point now, and
1570          * truncate can then come in and change things.  So we
1571          * can't touch *page from now on.  But *page_bufs is
1572          * safe due to elevated refcount.
1573          */
1574
1575         /*
1576          * And attach them to the current transaction.  But only if
1577          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1578          * and generally junk.
1579          */
1580         if (ret == 0) {
1581                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1582                                         NULL, journal_dirty_data_fn);
1583                 if (!ret)
1584                         ret = err;
1585         }
1586         walk_page_buffers(handle, page_bufs, 0,
1587                         PAGE_CACHE_SIZE, NULL, bput_one);
1588         err = ext3_journal_stop(handle);
1589         if (!ret)
1590                 ret = err;
1591         return ret;
1592
1593 out_fail:
1594         redirty_page_for_writepage(wbc, page);
1595         unlock_page(page);
1596         return ret;
1597 }
1598
1599 static int ext3_writeback_writepage(struct page *page,
1600                                 struct writeback_control *wbc)
1601 {
1602         struct inode *inode = page->mapping->host;
1603         handle_t *handle = NULL;
1604         int ret = 0;
1605         int err;
1606
1607         J_ASSERT(PageLocked(page));
1608         WARN_ON_ONCE(IS_RDONLY(inode));
1609
1610         if (ext3_journal_current_handle())
1611                 goto out_fail;
1612
1613         if (page_has_buffers(page)) {
1614                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1615                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1616                         /* Provide NULL get_block() to catch bugs if buffers
1617                          * weren't really mapped */
1618                         return block_write_full_page(page, NULL, wbc);
1619                 }
1620         }
1621
1622         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1623         if (IS_ERR(handle)) {
1624                 ret = PTR_ERR(handle);
1625                 goto out_fail;
1626         }
1627
1628         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1629                 ret = nobh_writepage(page, ext3_get_block, wbc);
1630         else
1631                 ret = block_write_full_page(page, ext3_get_block, wbc);
1632
1633         err = ext3_journal_stop(handle);
1634         if (!ret)
1635                 ret = err;
1636         return ret;
1637
1638 out_fail:
1639         redirty_page_for_writepage(wbc, page);
1640         unlock_page(page);
1641         return ret;
1642 }
1643
1644 static int ext3_journalled_writepage(struct page *page,
1645                                 struct writeback_control *wbc)
1646 {
1647         struct inode *inode = page->mapping->host;
1648         handle_t *handle = NULL;
1649         int ret = 0;
1650         int err;
1651
1652         J_ASSERT(PageLocked(page));
1653         WARN_ON_ONCE(IS_RDONLY(inode));
1654
1655         if (ext3_journal_current_handle())
1656                 goto no_write;
1657
1658         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1659         if (IS_ERR(handle)) {
1660                 ret = PTR_ERR(handle);
1661                 goto no_write;
1662         }
1663
1664         if (!page_has_buffers(page) || PageChecked(page)) {
1665                 /*
1666                  * It's mmapped pagecache.  Add buffers and journal it.  There
1667                  * doesn't seem much point in redirtying the page here.
1668                  */
1669                 ClearPageChecked(page);
1670                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1671                                         ext3_get_block);
1672                 if (ret != 0) {
1673                         ext3_journal_stop(handle);
1674                         goto out_unlock;
1675                 }
1676                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1677                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1678
1679                 err = walk_page_buffers(handle, page_buffers(page), 0,
1680                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1681                 if (ret == 0)
1682                         ret = err;
1683                 ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1684                 unlock_page(page);
1685         } else {
1686                 /*
1687                  * It may be a page full of checkpoint-mode buffers.  We don't
1688                  * really know unless we go poke around in the buffer_heads.
1689                  * But block_write_full_page will do the right thing.
1690                  */
1691                 ret = block_write_full_page(page, ext3_get_block, wbc);
1692         }
1693         err = ext3_journal_stop(handle);
1694         if (!ret)
1695                 ret = err;
1696 out:
1697         return ret;
1698
1699 no_write:
1700         redirty_page_for_writepage(wbc, page);
1701 out_unlock:
1702         unlock_page(page);
1703         goto out;
1704 }
1705
1706 static int ext3_readpage(struct file *file, struct page *page)
1707 {
1708         return mpage_readpage(page, ext3_get_block);
1709 }
1710
1711 static int
1712 ext3_readpages(struct file *file, struct address_space *mapping,
1713                 struct list_head *pages, unsigned nr_pages)
1714 {
1715         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1716 }
1717
1718 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1719 {
1720         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1721
1722         /*
1723          * If it's a full truncate we just forget about the pending dirtying
1724          */
1725         if (offset == 0)
1726                 ClearPageChecked(page);
1727
1728         journal_invalidatepage(journal, page, offset);
1729 }
1730
1731 static int ext3_releasepage(struct page *page, gfp_t wait)
1732 {
1733         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1734
1735         WARN_ON(PageChecked(page));
1736         if (!page_has_buffers(page))
1737                 return 0;
1738         return journal_try_to_free_buffers(journal, page, wait);
1739 }
1740
1741 /*
1742  * If the O_DIRECT write will extend the file then add this inode to the
1743  * orphan list.  So recovery will truncate it back to the original size
1744  * if the machine crashes during the write.
1745  *
1746  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1747  * crashes then stale disk data _may_ be exposed inside the file. But current
1748  * VFS code falls back into buffered path in that case so we are safe.
1749  */
1750 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1751                         const struct iovec *iov, loff_t offset,
1752                         unsigned long nr_segs)
1753 {
1754         struct file *file = iocb->ki_filp;
1755         struct inode *inode = file->f_mapping->host;
1756         struct ext3_inode_info *ei = EXT3_I(inode);
1757         handle_t *handle;
1758         ssize_t ret;
1759         int orphan = 0;
1760         size_t count = iov_length(iov, nr_segs);
1761         int retries = 0;
1762
1763         if (rw == WRITE) {
1764                 loff_t final_size = offset + count;
1765
1766                 if (final_size > inode->i_size) {
1767                         /* Credits for sb + inode write */
1768                         handle = ext3_journal_start(inode, 2);
1769                         if (IS_ERR(handle)) {
1770                                 ret = PTR_ERR(handle);
1771                                 goto out;
1772                         }
1773                         ret = ext3_orphan_add(handle, inode);
1774                         if (ret) {
1775                                 ext3_journal_stop(handle);
1776                                 goto out;
1777                         }
1778                         orphan = 1;
1779                         ei->i_disksize = inode->i_size;
1780                         ext3_journal_stop(handle);
1781                 }
1782         }
1783
1784 retry:
1785         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1786                                  offset, nr_segs,
1787                                  ext3_get_block, NULL);
1788         /*
1789          * In case of error extending write may have instantiated a few
1790          * blocks outside i_size. Trim these off again.
1791          */
1792         if (unlikely((rw & WRITE) && ret < 0)) {
1793                 loff_t isize = i_size_read(inode);
1794                 loff_t end = offset + iov_length(iov, nr_segs);
1795
1796                 if (end > isize)
1797                         vmtruncate(inode, isize);
1798         }
1799         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1800                 goto retry;
1801
1802         if (orphan) {
1803                 int err;
1804
1805                 /* Credits for sb + inode write */
1806                 handle = ext3_journal_start(inode, 2);
1807                 if (IS_ERR(handle)) {
1808                         /* This is really bad luck. We've written the data
1809                          * but cannot extend i_size. Truncate allocated blocks
1810                          * and pretend the write failed... */
1811                         ext3_truncate(inode);
1812                         ret = PTR_ERR(handle);
1813                         goto out;
1814                 }
1815                 if (inode->i_nlink)
1816                         ext3_orphan_del(handle, inode);
1817                 if (ret > 0) {
1818                         loff_t end = offset + ret;
1819                         if (end > inode->i_size) {
1820                                 ei->i_disksize = end;
1821                                 i_size_write(inode, end);
1822                                 /*
1823                                  * We're going to return a positive `ret'
1824                                  * here due to non-zero-length I/O, so there's
1825                                  * no way of reporting error returns from
1826                                  * ext3_mark_inode_dirty() to userspace.  So
1827                                  * ignore it.
1828                                  */
1829                                 ext3_mark_inode_dirty(handle, inode);
1830                         }
1831                 }
1832                 err = ext3_journal_stop(handle);
1833                 if (ret == 0)
1834                         ret = err;
1835         }
1836 out:
1837         return ret;
1838 }
1839
1840 /*
1841  * Pages can be marked dirty completely asynchronously from ext3's journalling
1842  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1843  * much here because ->set_page_dirty is called under VFS locks.  The page is
1844  * not necessarily locked.
1845  *
1846  * We cannot just dirty the page and leave attached buffers clean, because the
1847  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1848  * or jbddirty because all the journalling code will explode.
1849  *
1850  * So what we do is to mark the page "pending dirty" and next time writepage
1851  * is called, propagate that into the buffers appropriately.
1852  */
1853 static int ext3_journalled_set_page_dirty(struct page *page)
1854 {
1855         SetPageChecked(page);
1856         return __set_page_dirty_nobuffers(page);
1857 }
1858
1859 static const struct address_space_operations ext3_ordered_aops = {
1860         .readpage               = ext3_readpage,
1861         .readpages              = ext3_readpages,
1862         .writepage              = ext3_ordered_writepage,
1863         .sync_page              = block_sync_page,
1864         .write_begin            = ext3_write_begin,
1865         .write_end              = ext3_ordered_write_end,
1866         .bmap                   = ext3_bmap,
1867         .invalidatepage         = ext3_invalidatepage,
1868         .releasepage            = ext3_releasepage,
1869         .direct_IO              = ext3_direct_IO,
1870         .migratepage            = buffer_migrate_page,
1871         .is_partially_uptodate  = block_is_partially_uptodate,
1872         .error_remove_page      = generic_error_remove_page,
1873 };
1874
1875 static const struct address_space_operations ext3_writeback_aops = {
1876         .readpage               = ext3_readpage,
1877         .readpages              = ext3_readpages,
1878         .writepage              = ext3_writeback_writepage,
1879         .sync_page              = block_sync_page,
1880         .write_begin            = ext3_write_begin,
1881         .write_end              = ext3_writeback_write_end,
1882         .bmap                   = ext3_bmap,
1883         .invalidatepage         = ext3_invalidatepage,
1884         .releasepage            = ext3_releasepage,
1885         .direct_IO              = ext3_direct_IO,
1886         .migratepage            = buffer_migrate_page,
1887         .is_partially_uptodate  = block_is_partially_uptodate,
1888         .error_remove_page      = generic_error_remove_page,
1889 };
1890
1891 static const struct address_space_operations ext3_journalled_aops = {
1892         .readpage               = ext3_readpage,
1893         .readpages              = ext3_readpages,
1894         .writepage              = ext3_journalled_writepage,
1895         .sync_page              = block_sync_page,
1896         .write_begin            = ext3_write_begin,
1897         .write_end              = ext3_journalled_write_end,
1898         .set_page_dirty         = ext3_journalled_set_page_dirty,
1899         .bmap                   = ext3_bmap,
1900         .invalidatepage         = ext3_invalidatepage,
1901         .releasepage            = ext3_releasepage,
1902         .is_partially_uptodate  = block_is_partially_uptodate,
1903         .error_remove_page      = generic_error_remove_page,
1904 };
1905
1906 void ext3_set_aops(struct inode *inode)
1907 {
1908         if (ext3_should_order_data(inode))
1909                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1910         else if (ext3_should_writeback_data(inode))
1911                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1912         else
1913                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1914 }
1915
1916 /*
1917  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1918  * up to the end of the block which corresponds to `from'.
1919  * This required during truncate. We need to physically zero the tail end
1920  * of that block so it doesn't yield old data if the file is later grown.
1921  */
1922 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1923                 struct address_space *mapping, loff_t from)
1924 {
1925         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1926         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1927         unsigned blocksize, iblock, length, pos;
1928         struct inode *inode = mapping->host;
1929         struct buffer_head *bh;
1930         int err = 0;
1931
1932         blocksize = inode->i_sb->s_blocksize;
1933         length = blocksize - (offset & (blocksize - 1));
1934         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1935
1936         /*
1937          * For "nobh" option,  we can only work if we don't need to
1938          * read-in the page - otherwise we create buffers to do the IO.
1939          */
1940         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1941              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1942                 zero_user(page, offset, length);
1943                 set_page_dirty(page);
1944                 goto unlock;
1945         }
1946
1947         if (!page_has_buffers(page))
1948                 create_empty_buffers(page, blocksize, 0);
1949
1950         /* Find the buffer that contains "offset" */
1951         bh = page_buffers(page);
1952         pos = blocksize;
1953         while (offset >= pos) {
1954                 bh = bh->b_this_page;
1955                 iblock++;
1956                 pos += blocksize;
1957         }
1958
1959         err = 0;
1960         if (buffer_freed(bh)) {
1961                 BUFFER_TRACE(bh, "freed: skip");
1962                 goto unlock;
1963         }
1964
1965         if (!buffer_mapped(bh)) {
1966                 BUFFER_TRACE(bh, "unmapped");
1967                 ext3_get_block(inode, iblock, bh, 0);
1968                 /* unmapped? It's a hole - nothing to do */
1969                 if (!buffer_mapped(bh)) {
1970                         BUFFER_TRACE(bh, "still unmapped");
1971                         goto unlock;
1972                 }
1973         }
1974
1975         /* Ok, it's mapped. Make sure it's up-to-date */
1976         if (PageUptodate(page))
1977                 set_buffer_uptodate(bh);
1978
1979         if (!buffer_uptodate(bh)) {
1980                 err = -EIO;
1981                 ll_rw_block(READ, 1, &bh);
1982                 wait_on_buffer(bh);
1983                 /* Uhhuh. Read error. Complain and punt. */
1984                 if (!buffer_uptodate(bh))
1985                         goto unlock;
1986         }
1987
1988         if (ext3_should_journal_data(inode)) {
1989                 BUFFER_TRACE(bh, "get write access");
1990                 err = ext3_journal_get_write_access(handle, bh);
1991                 if (err)
1992                         goto unlock;
1993         }
1994
1995         zero_user(page, offset, length);
1996         BUFFER_TRACE(bh, "zeroed end of block");
1997
1998         err = 0;
1999         if (ext3_should_journal_data(inode)) {
2000                 err = ext3_journal_dirty_metadata(handle, bh);
2001         } else {
2002                 if (ext3_should_order_data(inode))
2003                         err = ext3_journal_dirty_data(handle, bh);
2004                 mark_buffer_dirty(bh);
2005         }
2006
2007 unlock:
2008         unlock_page(page);
2009         page_cache_release(page);
2010         return err;
2011 }
2012
2013 /*
2014  * Probably it should be a library function... search for first non-zero word
2015  * or memcmp with zero_page, whatever is better for particular architecture.
2016  * Linus?
2017  */
2018 static inline int all_zeroes(__le32 *p, __le32 *q)
2019 {
2020         while (p < q)
2021                 if (*p++)
2022                         return 0;
2023         return 1;
2024 }
2025
2026 /**
2027  *      ext3_find_shared - find the indirect blocks for partial truncation.
2028  *      @inode:   inode in question
2029  *      @depth:   depth of the affected branch
2030  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
2031  *      @chain:   place to store the pointers to partial indirect blocks
2032  *      @top:     place to the (detached) top of branch
2033  *
2034  *      This is a helper function used by ext3_truncate().
2035  *
2036  *      When we do truncate() we may have to clean the ends of several
2037  *      indirect blocks but leave the blocks themselves alive. Block is
2038  *      partially truncated if some data below the new i_size is refered
2039  *      from it (and it is on the path to the first completely truncated
2040  *      data block, indeed).  We have to free the top of that path along
2041  *      with everything to the right of the path. Since no allocation
2042  *      past the truncation point is possible until ext3_truncate()
2043  *      finishes, we may safely do the latter, but top of branch may
2044  *      require special attention - pageout below the truncation point
2045  *      might try to populate it.
2046  *
2047  *      We atomically detach the top of branch from the tree, store the
2048  *      block number of its root in *@top, pointers to buffer_heads of
2049  *      partially truncated blocks - in @chain[].bh and pointers to
2050  *      their last elements that should not be removed - in
2051  *      @chain[].p. Return value is the pointer to last filled element
2052  *      of @chain.
2053  *
2054  *      The work left to caller to do the actual freeing of subtrees:
2055  *              a) free the subtree starting from *@top
2056  *              b) free the subtrees whose roots are stored in
2057  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2058  *              c) free the subtrees growing from the inode past the @chain[0].
2059  *                      (no partially truncated stuff there).  */
2060
2061 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2062                         int offsets[4], Indirect chain[4], __le32 *top)
2063 {
2064         Indirect *partial, *p;
2065         int k, err;
2066
2067         *top = 0;
2068         /* Make k index the deepest non-null offset + 1 */
2069         for (k = depth; k > 1 && !offsets[k-1]; k--)
2070                 ;
2071         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2072         /* Writer: pointers */
2073         if (!partial)
2074                 partial = chain + k-1;
2075         /*
2076          * If the branch acquired continuation since we've looked at it -
2077          * fine, it should all survive and (new) top doesn't belong to us.
2078          */
2079         if (!partial->key && *partial->p)
2080                 /* Writer: end */
2081                 goto no_top;
2082         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2083                 ;
2084         /*
2085          * OK, we've found the last block that must survive. The rest of our
2086          * branch should be detached before unlocking. However, if that rest
2087          * of branch is all ours and does not grow immediately from the inode
2088          * it's easier to cheat and just decrement partial->p.
2089          */
2090         if (p == chain + k - 1 && p > chain) {
2091                 p->p--;
2092         } else {
2093                 *top = *p->p;
2094                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2095 #if 0
2096                 *p->p = 0;
2097 #endif
2098         }
2099         /* Writer: end */
2100
2101         while(partial > p) {
2102                 brelse(partial->bh);
2103                 partial--;
2104         }
2105 no_top:
2106         return partial;
2107 }
2108
2109 /*
2110  * Zero a number of block pointers in either an inode or an indirect block.
2111  * If we restart the transaction we must again get write access to the
2112  * indirect block for further modification.
2113  *
2114  * We release `count' blocks on disk, but (last - first) may be greater
2115  * than `count' because there can be holes in there.
2116  */
2117 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2118                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2119                 unsigned long count, __le32 *first, __le32 *last)
2120 {
2121         __le32 *p;
2122         if (try_to_extend_transaction(handle, inode)) {
2123                 if (bh) {
2124                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2125                         ext3_journal_dirty_metadata(handle, bh);
2126                 }
2127                 ext3_mark_inode_dirty(handle, inode);
2128                 truncate_restart_transaction(handle, inode);
2129                 if (bh) {
2130                         BUFFER_TRACE(bh, "retaking write access");
2131                         ext3_journal_get_write_access(handle, bh);
2132                 }
2133         }
2134
2135         /*
2136          * Any buffers which are on the journal will be in memory. We find
2137          * them on the hash table so journal_revoke() will run journal_forget()
2138          * on them.  We've already detached each block from the file, so
2139          * bforget() in journal_forget() should be safe.
2140          *
2141          * AKPM: turn on bforget in journal_forget()!!!
2142          */
2143         for (p = first; p < last; p++) {
2144                 u32 nr = le32_to_cpu(*p);
2145                 if (nr) {
2146                         struct buffer_head *bh;
2147
2148                         *p = 0;
2149                         bh = sb_find_get_block(inode->i_sb, nr);
2150                         ext3_forget(handle, 0, inode, bh, nr);
2151                 }
2152         }
2153
2154         ext3_free_blocks(handle, inode, block_to_free, count);
2155 }
2156
2157 /**
2158  * ext3_free_data - free a list of data blocks
2159  * @handle:     handle for this transaction
2160  * @inode:      inode we are dealing with
2161  * @this_bh:    indirect buffer_head which contains *@first and *@last
2162  * @first:      array of block numbers
2163  * @last:       points immediately past the end of array
2164  *
2165  * We are freeing all blocks refered from that array (numbers are stored as
2166  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2167  *
2168  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2169  * blocks are contiguous then releasing them at one time will only affect one
2170  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2171  * actually use a lot of journal space.
2172  *
2173  * @this_bh will be %NULL if @first and @last point into the inode's direct
2174  * block pointers.
2175  */
2176 static void ext3_free_data(handle_t *handle, struct inode *inode,
2177                            struct buffer_head *this_bh,
2178                            __le32 *first, __le32 *last)
2179 {
2180         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2181         unsigned long count = 0;            /* Number of blocks in the run */
2182         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2183                                                corresponding to
2184                                                block_to_free */
2185         ext3_fsblk_t nr;                    /* Current block # */
2186         __le32 *p;                          /* Pointer into inode/ind
2187                                                for current block */
2188         int err;
2189
2190         if (this_bh) {                          /* For indirect block */
2191                 BUFFER_TRACE(this_bh, "get_write_access");
2192                 err = ext3_journal_get_write_access(handle, this_bh);
2193                 /* Important: if we can't update the indirect pointers
2194                  * to the blocks, we can't free them. */
2195                 if (err)
2196                         return;
2197         }
2198
2199         for (p = first; p < last; p++) {
2200                 nr = le32_to_cpu(*p);
2201                 if (nr) {
2202                         /* accumulate blocks to free if they're contiguous */
2203                         if (count == 0) {
2204                                 block_to_free = nr;
2205                                 block_to_free_p = p;
2206                                 count = 1;
2207                         } else if (nr == block_to_free + count) {
2208                                 count++;
2209                         } else {
2210                                 ext3_clear_blocks(handle, inode, this_bh,
2211                                                   block_to_free,
2212                                                   count, block_to_free_p, p);
2213                                 block_to_free = nr;
2214                                 block_to_free_p = p;
2215                                 count = 1;
2216                         }
2217                 }
2218         }
2219
2220         if (count > 0)
2221                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2222                                   count, block_to_free_p, p);
2223
2224         if (this_bh) {
2225                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2226
2227                 /*
2228                  * The buffer head should have an attached journal head at this
2229                  * point. However, if the data is corrupted and an indirect
2230                  * block pointed to itself, it would have been detached when
2231                  * the block was cleared. Check for this instead of OOPSing.
2232                  */
2233                 if (bh2jh(this_bh))
2234                         ext3_journal_dirty_metadata(handle, this_bh);
2235                 else
2236                         ext3_error(inode->i_sb, "ext3_free_data",
2237                                    "circular indirect block detected, "
2238                                    "inode=%lu, block=%llu",
2239                                    inode->i_ino,
2240                                    (unsigned long long)this_bh->b_blocknr);
2241         }
2242 }
2243
2244 /**
2245  *      ext3_free_branches - free an array of branches
2246  *      @handle: JBD handle for this transaction
2247  *      @inode: inode we are dealing with
2248  *      @parent_bh: the buffer_head which contains *@first and *@last
2249  *      @first: array of block numbers
2250  *      @last:  pointer immediately past the end of array
2251  *      @depth: depth of the branches to free
2252  *
2253  *      We are freeing all blocks refered from these branches (numbers are
2254  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2255  *      appropriately.
2256  */
2257 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2258                                struct buffer_head *parent_bh,
2259                                __le32 *first, __le32 *last, int depth)
2260 {
2261         ext3_fsblk_t nr;
2262         __le32 *p;
2263
2264         if (is_handle_aborted(handle))
2265                 return;
2266
2267         if (depth--) {
2268                 struct buffer_head *bh;
2269                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2270                 p = last;
2271                 while (--p >= first) {
2272                         nr = le32_to_cpu(*p);
2273                         if (!nr)
2274                                 continue;               /* A hole */
2275
2276                         /* Go read the buffer for the next level down */
2277                         bh = sb_bread(inode->i_sb, nr);
2278
2279                         /*
2280                          * A read failure? Report error and clear slot
2281                          * (should be rare).
2282                          */
2283                         if (!bh) {
2284                                 ext3_error(inode->i_sb, "ext3_free_branches",
2285                                            "Read failure, inode=%lu, block="E3FSBLK,
2286                                            inode->i_ino, nr);
2287                                 continue;
2288                         }
2289
2290                         /* This zaps the entire block.  Bottom up. */
2291                         BUFFER_TRACE(bh, "free child branches");
2292                         ext3_free_branches(handle, inode, bh,
2293                                            (__le32*)bh->b_data,
2294                                            (__le32*)bh->b_data + addr_per_block,
2295                                            depth);
2296
2297                         /*
2298                          * We've probably journalled the indirect block several
2299                          * times during the truncate.  But it's no longer
2300                          * needed and we now drop it from the transaction via
2301                          * journal_revoke().
2302                          *
2303                          * That's easy if it's exclusively part of this
2304                          * transaction.  But if it's part of the committing
2305                          * transaction then journal_forget() will simply
2306                          * brelse() it.  That means that if the underlying
2307                          * block is reallocated in ext3_get_block(),
2308                          * unmap_underlying_metadata() will find this block
2309                          * and will try to get rid of it.  damn, damn.
2310                          *
2311                          * If this block has already been committed to the
2312                          * journal, a revoke record will be written.  And
2313                          * revoke records must be emitted *before* clearing
2314                          * this block's bit in the bitmaps.
2315                          */
2316                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2317
2318                         /*
2319                          * Everything below this this pointer has been
2320                          * released.  Now let this top-of-subtree go.
2321                          *
2322                          * We want the freeing of this indirect block to be
2323                          * atomic in the journal with the updating of the
2324                          * bitmap block which owns it.  So make some room in
2325                          * the journal.
2326                          *
2327                          * We zero the parent pointer *after* freeing its
2328                          * pointee in the bitmaps, so if extend_transaction()
2329                          * for some reason fails to put the bitmap changes and
2330                          * the release into the same transaction, recovery
2331                          * will merely complain about releasing a free block,
2332                          * rather than leaking blocks.
2333                          */
2334                         if (is_handle_aborted(handle))
2335                                 return;
2336                         if (try_to_extend_transaction(handle, inode)) {
2337                                 ext3_mark_inode_dirty(handle, inode);
2338                                 truncate_restart_transaction(handle, inode);
2339                         }
2340
2341                         ext3_free_blocks(handle, inode, nr, 1);
2342
2343                         if (parent_bh) {
2344                                 /*
2345                                  * The block which we have just freed is
2346                                  * pointed to by an indirect block: journal it
2347                                  */
2348                                 BUFFER_TRACE(parent_bh, "get_write_access");
2349                                 if (!ext3_journal_get_write_access(handle,
2350                                                                    parent_bh)){
2351                                         *p = 0;
2352                                         BUFFER_TRACE(parent_bh,
2353                                         "call ext3_journal_dirty_metadata");
2354                                         ext3_journal_dirty_metadata(handle,
2355                                                                     parent_bh);
2356                                 }
2357                         }
2358                 }
2359         } else {
2360                 /* We have reached the bottom of the tree. */
2361                 BUFFER_TRACE(parent_bh, "free data blocks");
2362                 ext3_free_data(handle, inode, parent_bh, first, last);
2363         }
2364 }
2365
2366 int ext3_can_truncate(struct inode *inode)
2367 {
2368         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2369                 return 0;
2370         if (S_ISREG(inode->i_mode))
2371                 return 1;
2372         if (S_ISDIR(inode->i_mode))
2373                 return 1;
2374         if (S_ISLNK(inode->i_mode))
2375                 return !ext3_inode_is_fast_symlink(inode);
2376         return 0;
2377 }
2378
2379 /*
2380  * ext3_truncate()
2381  *
2382  * We block out ext3_get_block() block instantiations across the entire
2383  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2384  * simultaneously on behalf of the same inode.
2385  *
2386  * As we work through the truncate and commmit bits of it to the journal there
2387  * is one core, guiding principle: the file's tree must always be consistent on
2388  * disk.  We must be able to restart the truncate after a crash.
2389  *
2390  * The file's tree may be transiently inconsistent in memory (although it
2391  * probably isn't), but whenever we close off and commit a journal transaction,
2392  * the contents of (the filesystem + the journal) must be consistent and
2393  * restartable.  It's pretty simple, really: bottom up, right to left (although
2394  * left-to-right works OK too).
2395  *
2396  * Note that at recovery time, journal replay occurs *before* the restart of
2397  * truncate against the orphan inode list.
2398  *
2399  * The committed inode has the new, desired i_size (which is the same as
2400  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2401  * that this inode's truncate did not complete and it will again call
2402  * ext3_truncate() to have another go.  So there will be instantiated blocks
2403  * to the right of the truncation point in a crashed ext3 filesystem.  But
2404  * that's fine - as long as they are linked from the inode, the post-crash
2405  * ext3_truncate() run will find them and release them.
2406  */
2407 void ext3_truncate(struct inode *inode)
2408 {
2409         handle_t *handle;
2410         struct ext3_inode_info *ei = EXT3_I(inode);
2411         __le32 *i_data = ei->i_data;
2412         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2413         struct address_space *mapping = inode->i_mapping;
2414         int offsets[4];
2415         Indirect chain[4];
2416         Indirect *partial;
2417         __le32 nr = 0;
2418         int n;
2419         long last_block;
2420         unsigned blocksize = inode->i_sb->s_blocksize;
2421         struct page *page;
2422
2423         if (!ext3_can_truncate(inode))
2424                 goto out_notrans;
2425
2426         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2427                 ext3_set_inode_state(inode, EXT3_STATE_FLUSH_ON_CLOSE);
2428
2429         /*
2430          * We have to lock the EOF page here, because lock_page() nests
2431          * outside journal_start().
2432          */
2433         if ((inode->i_size & (blocksize - 1)) == 0) {
2434                 /* Block boundary? Nothing to do */
2435                 page = NULL;
2436         } else {
2437                 page = grab_cache_page(mapping,
2438                                 inode->i_size >> PAGE_CACHE_SHIFT);
2439                 if (!page)
2440                         goto out_notrans;
2441         }
2442
2443         handle = start_transaction(inode);
2444         if (IS_ERR(handle)) {
2445                 if (page) {
2446                         clear_highpage(page);
2447                         flush_dcache_page(page);
2448                         unlock_page(page);
2449                         page_cache_release(page);
2450                 }
2451                 goto out_notrans;
2452         }
2453
2454         last_block = (inode->i_size + blocksize-1)
2455                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2456
2457         if (page)
2458                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2459
2460         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2461         if (n == 0)
2462                 goto out_stop;  /* error */
2463
2464         /*
2465          * OK.  This truncate is going to happen.  We add the inode to the
2466          * orphan list, so that if this truncate spans multiple transactions,
2467          * and we crash, we will resume the truncate when the filesystem
2468          * recovers.  It also marks the inode dirty, to catch the new size.
2469          *
2470          * Implication: the file must always be in a sane, consistent
2471          * truncatable state while each transaction commits.
2472          */
2473         if (ext3_orphan_add(handle, inode))
2474                 goto out_stop;
2475
2476         /*
2477          * The orphan list entry will now protect us from any crash which
2478          * occurs before the truncate completes, so it is now safe to propagate
2479          * the new, shorter inode size (held for now in i_size) into the
2480          * on-disk inode. We do this via i_disksize, which is the value which
2481          * ext3 *really* writes onto the disk inode.
2482          */
2483         ei->i_disksize = inode->i_size;
2484
2485         /*
2486          * From here we block out all ext3_get_block() callers who want to
2487          * modify the block allocation tree.
2488          */
2489         mutex_lock(&ei->truncate_mutex);
2490
2491         if (n == 1) {           /* direct blocks */
2492                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2493                                i_data + EXT3_NDIR_BLOCKS);
2494                 goto do_indirects;
2495         }
2496
2497         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2498         /* Kill the top of shared branch (not detached) */
2499         if (nr) {
2500                 if (partial == chain) {
2501                         /* Shared branch grows from the inode */
2502                         ext3_free_branches(handle, inode, NULL,
2503                                            &nr, &nr+1, (chain+n-1) - partial);
2504                         *partial->p = 0;
2505                         /*
2506                          * We mark the inode dirty prior to restart,
2507                          * and prior to stop.  No need for it here.
2508                          */
2509                 } else {
2510                         /* Shared branch grows from an indirect block */
2511                         BUFFER_TRACE(partial->bh, "get_write_access");
2512                         ext3_free_branches(handle, inode, partial->bh,
2513                                         partial->p,
2514                                         partial->p+1, (chain+n-1) - partial);
2515                 }
2516         }
2517         /* Clear the ends of indirect blocks on the shared branch */
2518         while (partial > chain) {
2519                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2520                                    (__le32*)partial->bh->b_data+addr_per_block,
2521                                    (chain+n-1) - partial);
2522                 BUFFER_TRACE(partial->bh, "call brelse");
2523                 brelse (partial->bh);
2524                 partial--;
2525         }
2526 do_indirects:
2527         /* Kill the remaining (whole) subtrees */
2528         switch (offsets[0]) {
2529         default:
2530                 nr = i_data[EXT3_IND_BLOCK];
2531                 if (nr) {
2532                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2533                         i_data[EXT3_IND_BLOCK] = 0;
2534                 }
2535         case EXT3_IND_BLOCK:
2536                 nr = i_data[EXT3_DIND_BLOCK];
2537                 if (nr) {
2538                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2539                         i_data[EXT3_DIND_BLOCK] = 0;
2540                 }
2541         case EXT3_DIND_BLOCK:
2542                 nr = i_data[EXT3_TIND_BLOCK];
2543                 if (nr) {
2544                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2545                         i_data[EXT3_TIND_BLOCK] = 0;
2546                 }
2547         case EXT3_TIND_BLOCK:
2548                 ;
2549         }
2550
2551         ext3_discard_reservation(inode);
2552
2553         mutex_unlock(&ei->truncate_mutex);
2554         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2555         ext3_mark_inode_dirty(handle, inode);
2556
2557         /*
2558          * In a multi-transaction truncate, we only make the final transaction
2559          * synchronous
2560          */
2561         if (IS_SYNC(inode))
2562                 handle->h_sync = 1;
2563 out_stop:
2564         /*
2565          * If this was a simple ftruncate(), and the file will remain alive
2566          * then we need to clear up the orphan record which we created above.
2567          * However, if this was a real unlink then we were called by
2568          * ext3_delete_inode(), and we allow that function to clean up the
2569          * orphan info for us.
2570          */
2571         if (inode->i_nlink)
2572                 ext3_orphan_del(handle, inode);
2573
2574         ext3_journal_stop(handle);
2575         return;
2576 out_notrans:
2577         /*
2578          * Delete the inode from orphan list so that it doesn't stay there
2579          * forever and trigger assertion on umount.
2580          */
2581         if (inode->i_nlink)
2582                 ext3_orphan_del(NULL, inode);
2583 }
2584
2585 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2586                 unsigned long ino, struct ext3_iloc *iloc)
2587 {
2588         unsigned long block_group;
2589         unsigned long offset;
2590         ext3_fsblk_t block;
2591         struct ext3_group_desc *gdp;
2592
2593         if (!ext3_valid_inum(sb, ino)) {
2594                 /*
2595                  * This error is already checked for in namei.c unless we are
2596                  * looking at an NFS filehandle, in which case no error
2597                  * report is needed
2598                  */
2599                 return 0;
2600         }
2601
2602         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2603         gdp = ext3_get_group_desc(sb, block_group, NULL);
2604         if (!gdp)
2605                 return 0;
2606         /*
2607          * Figure out the offset within the block group inode table
2608          */
2609         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2610                 EXT3_INODE_SIZE(sb);
2611         block = le32_to_cpu(gdp->bg_inode_table) +
2612                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2613
2614         iloc->block_group = block_group;
2615         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2616         return block;
2617 }
2618
2619 /*
2620  * ext3_get_inode_loc returns with an extra refcount against the inode's
2621  * underlying buffer_head on success. If 'in_mem' is true, we have all
2622  * data in memory that is needed to recreate the on-disk version of this
2623  * inode.
2624  */
2625 static int __ext3_get_inode_loc(struct inode *inode,
2626                                 struct ext3_iloc *iloc, int in_mem)
2627 {
2628         ext3_fsblk_t block;
2629         struct buffer_head *bh;
2630
2631         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2632         if (!block)
2633                 return -EIO;
2634
2635         bh = sb_getblk(inode->i_sb, block);
2636         if (!bh) {
2637                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2638                                 "unable to read inode block - "
2639                                 "inode=%lu, block="E3FSBLK,
2640                                  inode->i_ino, block);
2641                 return -EIO;
2642         }
2643         if (!buffer_uptodate(bh)) {
2644                 lock_buffer(bh);
2645
2646                 /*
2647                  * If the buffer has the write error flag, we have failed
2648                  * to write out another inode in the same block.  In this
2649                  * case, we don't have to read the block because we may
2650                  * read the old inode data successfully.
2651                  */
2652                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2653                         set_buffer_uptodate(bh);
2654
2655                 if (buffer_uptodate(bh)) {
2656                         /* someone brought it uptodate while we waited */
2657                         unlock_buffer(bh);
2658                         goto has_buffer;
2659                 }
2660
2661                 /*
2662                  * If we have all information of the inode in memory and this
2663                  * is the only valid inode in the block, we need not read the
2664                  * block.
2665                  */
2666                 if (in_mem) {
2667                         struct buffer_head *bitmap_bh;
2668                         struct ext3_group_desc *desc;
2669                         int inodes_per_buffer;
2670                         int inode_offset, i;
2671                         int block_group;
2672                         int start;
2673
2674                         block_group = (inode->i_ino - 1) /
2675                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2676                         inodes_per_buffer = bh->b_size /
2677                                 EXT3_INODE_SIZE(inode->i_sb);
2678                         inode_offset = ((inode->i_ino - 1) %
2679                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2680                         start = inode_offset & ~(inodes_per_buffer - 1);
2681
2682                         /* Is the inode bitmap in cache? */
2683                         desc = ext3_get_group_desc(inode->i_sb,
2684                                                 block_group, NULL);
2685                         if (!desc)
2686                                 goto make_io;
2687
2688                         bitmap_bh = sb_getblk(inode->i_sb,
2689                                         le32_to_cpu(desc->bg_inode_bitmap));
2690                         if (!bitmap_bh)
2691                                 goto make_io;
2692
2693                         /*
2694                          * If the inode bitmap isn't in cache then the
2695                          * optimisation may end up performing two reads instead
2696                          * of one, so skip it.
2697                          */
2698                         if (!buffer_uptodate(bitmap_bh)) {
2699                                 brelse(bitmap_bh);
2700                                 goto make_io;
2701                         }
2702                         for (i = start; i < start + inodes_per_buffer; i++) {
2703                                 if (i == inode_offset)
2704                                         continue;
2705                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2706                                         break;
2707                         }
2708                         brelse(bitmap_bh);
2709                         if (i == start + inodes_per_buffer) {
2710                                 /* all other inodes are free, so skip I/O */
2711                                 memset(bh->b_data, 0, bh->b_size);
2712                                 set_buffer_uptodate(bh);
2713                                 unlock_buffer(bh);
2714                                 goto has_buffer;
2715                         }
2716                 }
2717
2718 make_io:
2719                 /*
2720                  * There are other valid inodes in the buffer, this inode
2721                  * has in-inode xattrs, or we don't have this inode in memory.
2722                  * Read the block from disk.
2723                  */
2724                 get_bh(bh);
2725                 bh->b_end_io = end_buffer_read_sync;
2726                 submit_bh(READ_META, bh);
2727                 wait_on_buffer(bh);
2728                 if (!buffer_uptodate(bh)) {
2729                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2730                                         "unable to read inode block - "
2731                                         "inode=%lu, block="E3FSBLK,
2732                                         inode->i_ino, block);
2733                         brelse(bh);
2734                         return -EIO;
2735                 }
2736         }
2737 has_buffer:
2738         iloc->bh = bh;
2739         return 0;
2740 }
2741
2742 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2743 {
2744         /* We have all inode data except xattrs in memory here. */
2745         return __ext3_get_inode_loc(inode, iloc,
2746                 !ext3_test_inode_state(inode, EXT3_STATE_XATTR));
2747 }
2748
2749 void ext3_set_inode_flags(struct inode *inode)
2750 {
2751         unsigned int flags = EXT3_I(inode)->i_flags;
2752
2753         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2754         if (flags & EXT3_SYNC_FL)
2755                 inode->i_flags |= S_SYNC;
2756         if (flags & EXT3_APPEND_FL)
2757                 inode->i_flags |= S_APPEND;
2758         if (flags & EXT3_IMMUTABLE_FL)
2759                 inode->i_flags |= S_IMMUTABLE;
2760         if (flags & EXT3_NOATIME_FL)
2761                 inode->i_flags |= S_NOATIME;
2762         if (flags & EXT3_DIRSYNC_FL)
2763                 inode->i_flags |= S_DIRSYNC;
2764 }
2765
2766 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2767 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2768 {
2769         unsigned int flags = ei->vfs_inode.i_flags;
2770
2771         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2772                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2773         if (flags & S_SYNC)
2774                 ei->i_flags |= EXT3_SYNC_FL;
2775         if (flags & S_APPEND)
2776                 ei->i_flags |= EXT3_APPEND_FL;
2777         if (flags & S_IMMUTABLE)
2778                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2779         if (flags & S_NOATIME)
2780                 ei->i_flags |= EXT3_NOATIME_FL;
2781         if (flags & S_DIRSYNC)
2782                 ei->i_flags |= EXT3_DIRSYNC_FL;
2783 }
2784
2785 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2786 {
2787         struct ext3_iloc iloc;
2788         struct ext3_inode *raw_inode;
2789         struct ext3_inode_info *ei;
2790         struct buffer_head *bh;
2791         struct inode *inode;
2792         journal_t *journal = EXT3_SB(sb)->s_journal;
2793         transaction_t *transaction;
2794         long ret;
2795         int block;
2796
2797         inode = iget_locked(sb, ino);
2798         if (!inode)
2799                 return ERR_PTR(-ENOMEM);
2800         if (!(inode->i_state & I_NEW))
2801                 return inode;
2802
2803         ei = EXT3_I(inode);
2804         ei->i_block_alloc_info = NULL;
2805
2806         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2807         if (ret < 0)
2808                 goto bad_inode;
2809         bh = iloc.bh;
2810         raw_inode = ext3_raw_inode(&iloc);
2811         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2812         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2813         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2814         if(!(test_opt (inode->i_sb, NO_UID32))) {
2815                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2816                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2817         }
2818         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2819         inode->i_size = le32_to_cpu(raw_inode->i_size);
2820         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2821         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2822         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2823         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2824
2825         ei->i_state_flags = 0;
2826         ei->i_dir_start_lookup = 0;
2827         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2828         /* We now have enough fields to check if the inode was active or not.
2829          * This is needed because nfsd might try to access dead inodes
2830          * the test is that same one that e2fsck uses
2831          * NeilBrown 1999oct15
2832          */
2833         if (inode->i_nlink == 0) {
2834                 if (inode->i_mode == 0 ||
2835                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2836                         /* this inode is deleted */
2837                         brelse (bh);
2838                         ret = -ESTALE;
2839                         goto bad_inode;
2840                 }
2841                 /* The only unlinked inodes we let through here have
2842                  * valid i_mode and are being read by the orphan
2843                  * recovery code: that's fine, we're about to complete
2844                  * the process of deleting those. */
2845         }
2846         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2847         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2848 #ifdef EXT3_FRAGMENTS
2849         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2850         ei->i_frag_no = raw_inode->i_frag;
2851         ei->i_frag_size = raw_inode->i_fsize;
2852 #endif
2853         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2854         if (!S_ISREG(inode->i_mode)) {
2855                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2856         } else {
2857                 inode->i_size |=
2858                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2859         }
2860         ei->i_disksize = inode->i_size;
2861         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2862         ei->i_block_group = iloc.block_group;
2863         /*
2864          * NOTE! The in-memory inode i_data array is in little-endian order
2865          * even on big-endian machines: we do NOT byteswap the block numbers!
2866          */
2867         for (block = 0; block < EXT3_N_BLOCKS; block++)
2868                 ei->i_data[block] = raw_inode->i_block[block];
2869         INIT_LIST_HEAD(&ei->i_orphan);
2870
2871         /*
2872          * Set transaction id's of transactions that have to be committed
2873          * to finish f[data]sync. We set them to currently running transaction
2874          * as we cannot be sure that the inode or some of its metadata isn't
2875          * part of the transaction - the inode could have been reclaimed and
2876          * now it is reread from disk.
2877          */
2878         if (journal) {
2879                 tid_t tid;
2880
2881                 spin_lock(&journal->j_state_lock);
2882                 if (journal->j_running_transaction)
2883                         transaction = journal->j_running_transaction;
2884                 else
2885                         transaction = journal->j_committing_transaction;
2886                 if (transaction)
2887                         tid = transaction->t_tid;
2888                 else
2889                         tid = journal->j_commit_sequence;
2890                 spin_unlock(&journal->j_state_lock);
2891                 atomic_set(&ei->i_sync_tid, tid);
2892                 atomic_set(&ei->i_datasync_tid, tid);
2893         }
2894
2895         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2896             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2897                 /*
2898                  * When mke2fs creates big inodes it does not zero out
2899                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2900                  * so ignore those first few inodes.
2901                  */
2902                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2903                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2904                     EXT3_INODE_SIZE(inode->i_sb)) {
2905                         brelse (bh);
2906                         ret = -EIO;
2907                         goto bad_inode;
2908                 }
2909                 if (ei->i_extra_isize == 0) {
2910                         /* The extra space is currently unused. Use it. */
2911                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2912                                             EXT3_GOOD_OLD_INODE_SIZE;
2913                 } else {
2914                         __le32 *magic = (void *)raw_inode +
2915                                         EXT3_GOOD_OLD_INODE_SIZE +
2916                                         ei->i_extra_isize;
2917                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2918                                  ext3_set_inode_state(inode, EXT3_STATE_XATTR);
2919                 }
2920         } else
2921                 ei->i_extra_isize = 0;
2922
2923         if (S_ISREG(inode->i_mode)) {
2924                 inode->i_op = &ext3_file_inode_operations;
2925                 inode->i_fop = &ext3_file_operations;
2926                 ext3_set_aops(inode);
2927         } else if (S_ISDIR(inode->i_mode)) {
2928                 inode->i_op = &ext3_dir_inode_operations;
2929                 inode->i_fop = &ext3_dir_operations;
2930         } else if (S_ISLNK(inode->i_mode)) {
2931                 if (ext3_inode_is_fast_symlink(inode)) {
2932                         inode->i_op = &ext3_fast_symlink_inode_operations;
2933                         nd_terminate_link(ei->i_data, inode->i_size,
2934                                 sizeof(ei->i_data) - 1);
2935                 } else {
2936                         inode->i_op = &ext3_symlink_inode_operations;
2937                         ext3_set_aops(inode);
2938                 }
2939         } else {
2940                 inode->i_op = &ext3_special_inode_operations;
2941                 if (raw_inode->i_block[0])
2942                         init_special_inode(inode, inode->i_mode,
2943                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2944                 else
2945                         init_special_inode(inode, inode->i_mode,
2946                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2947         }
2948         brelse (iloc.bh);
2949         ext3_set_inode_flags(inode);
2950         unlock_new_inode(inode);
2951         return inode;
2952
2953 bad_inode:
2954         iget_failed(inode);
2955         return ERR_PTR(ret);
2956 }
2957
2958 /*
2959  * Post the struct inode info into an on-disk inode location in the
2960  * buffer-cache.  This gobbles the caller's reference to the
2961  * buffer_head in the inode location struct.
2962  *
2963  * The caller must have write access to iloc->bh.
2964  */
2965 static int ext3_do_update_inode(handle_t *handle,
2966                                 struct inode *inode,
2967                                 struct ext3_iloc *iloc)
2968 {
2969         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2970         struct ext3_inode_info *ei = EXT3_I(inode);
2971         struct buffer_head *bh = iloc->bh;
2972         int err = 0, rc, block;
2973
2974 again:
2975         /* we can't allow multiple procs in here at once, its a bit racey */
2976         lock_buffer(bh);
2977
2978         /* For fields not not tracking in the in-memory inode,
2979          * initialise them to zero for new inodes. */
2980         if (ext3_test_inode_state(inode, EXT3_STATE_NEW))
2981                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2982
2983         ext3_get_inode_flags(ei);
2984         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2985         if(!(test_opt(inode->i_sb, NO_UID32))) {
2986                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2987                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2988 /*
2989  * Fix up interoperability with old kernels. Otherwise, old inodes get
2990  * re-used with the upper 16 bits of the uid/gid intact
2991  */
2992                 if(!ei->i_dtime) {
2993                         raw_inode->i_uid_high =
2994                                 cpu_to_le16(high_16_bits(inode->i_uid));
2995                         raw_inode->i_gid_high =
2996                                 cpu_to_le16(high_16_bits(inode->i_gid));
2997                 } else {
2998                         raw_inode->i_uid_high = 0;
2999                         raw_inode->i_gid_high = 0;
3000                 }
3001         } else {
3002                 raw_inode->i_uid_low =
3003                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3004                 raw_inode->i_gid_low =
3005                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3006                 raw_inode->i_uid_high = 0;
3007                 raw_inode->i_gid_high = 0;
3008         }
3009         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3010         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
3011         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
3012         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
3013         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
3014         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
3015         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3016         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
3017 #ifdef EXT3_FRAGMENTS
3018         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
3019         raw_inode->i_frag = ei->i_frag_no;
3020         raw_inode->i_fsize = ei->i_frag_size;
3021 #endif
3022         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
3023         if (!S_ISREG(inode->i_mode)) {
3024                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
3025         } else {
3026                 raw_inode->i_size_high =
3027                         cpu_to_le32(ei->i_disksize >> 32);
3028                 if (ei->i_disksize > 0x7fffffffULL) {
3029                         struct super_block *sb = inode->i_sb;
3030                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
3031                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
3032                             EXT3_SB(sb)->s_es->s_rev_level ==
3033                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
3034                                /* If this is the first large file
3035                                 * created, add a flag to the superblock.
3036                                 */
3037                                 unlock_buffer(bh);
3038                                 err = ext3_journal_get_write_access(handle,
3039                                                 EXT3_SB(sb)->s_sbh);
3040                                 if (err)
3041                                         goto out_brelse;
3042
3043                                 ext3_update_dynamic_rev(sb);
3044                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
3045                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
3046                                 handle->h_sync = 1;
3047                                 err = ext3_journal_dirty_metadata(handle,
3048                                                 EXT3_SB(sb)->s_sbh);
3049                                 /* get our lock and start over */
3050                                 goto again;
3051                         }
3052                 }
3053         }
3054         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3055         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3056                 if (old_valid_dev(inode->i_rdev)) {
3057                         raw_inode->i_block[0] =
3058                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3059                         raw_inode->i_block[1] = 0;
3060                 } else {
3061                         raw_inode->i_block[0] = 0;
3062                         raw_inode->i_block[1] =
3063                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3064                         raw_inode->i_block[2] = 0;
3065                 }
3066         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
3067                 raw_inode->i_block[block] = ei->i_data[block];
3068
3069         if (ei->i_extra_isize)
3070                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3071
3072         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
3073         unlock_buffer(bh);
3074         rc = ext3_journal_dirty_metadata(handle, bh);
3075         if (!err)
3076                 err = rc;
3077         ext3_clear_inode_state(inode, EXT3_STATE_NEW);
3078
3079         atomic_set(&ei->i_sync_tid, handle->h_transaction->t_tid);
3080 out_brelse:
3081         brelse (bh);
3082         ext3_std_error(inode->i_sb, err);
3083         return err;
3084 }
3085
3086 /*
3087  * ext3_write_inode()
3088  *
3089  * We are called from a few places:
3090  *
3091  * - Within generic_file_write() for O_SYNC files.
3092  *   Here, there will be no transaction running. We wait for any running
3093  *   trasnaction to commit.
3094  *
3095  * - Within sys_sync(), kupdate and such.
3096  *   We wait on commit, if tol to.
3097  *
3098  * - Within prune_icache() (PF_MEMALLOC == true)
3099  *   Here we simply return.  We can't afford to block kswapd on the
3100  *   journal commit.
3101  *
3102  * In all cases it is actually safe for us to return without doing anything,
3103  * because the inode has been copied into a raw inode buffer in
3104  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3105  * knfsd.
3106  *
3107  * Note that we are absolutely dependent upon all inode dirtiers doing the
3108  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3109  * which we are interested.
3110  *
3111  * It would be a bug for them to not do this.  The code:
3112  *
3113  *      mark_inode_dirty(inode)
3114  *      stuff();
3115  *      inode->i_size = expr;
3116  *
3117  * is in error because a kswapd-driven write_inode() could occur while
3118  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3119  * will no longer be on the superblock's dirty inode list.
3120  */
3121 int ext3_write_inode(struct inode *inode, struct writeback_control *wbc)
3122 {
3123         if (current->flags & PF_MEMALLOC)
3124                 return 0;
3125
3126         if (ext3_journal_current_handle()) {
3127                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3128                 dump_stack();
3129                 return -EIO;
3130         }
3131
3132         if (wbc->sync_mode != WB_SYNC_ALL)
3133                 return 0;
3134
3135         return ext3_force_commit(inode->i_sb);
3136 }
3137
3138 /*
3139  * ext3_setattr()
3140  *
3141  * Called from notify_change.
3142  *
3143  * We want to trap VFS attempts to truncate the file as soon as
3144  * possible.  In particular, we want to make sure that when the VFS
3145  * shrinks i_size, we put the inode on the orphan list and modify
3146  * i_disksize immediately, so that during the subsequent flushing of
3147  * dirty pages and freeing of disk blocks, we can guarantee that any
3148  * commit will leave the blocks being flushed in an unused state on
3149  * disk.  (On recovery, the inode will get truncated and the blocks will
3150  * be freed, so we have a strong guarantee that no future commit will
3151  * leave these blocks visible to the user.)
3152  *
3153  * Called with inode->sem down.
3154  */
3155 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3156 {
3157         struct inode *inode = dentry->d_inode;
3158         int error, rc = 0;
3159         const unsigned int ia_valid = attr->ia_valid;
3160
3161         error = inode_change_ok(inode, attr);
3162         if (error)
3163                 return error;
3164
3165         if (is_quota_modification(inode, attr))
3166                 dquot_initialize(inode);
3167         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3168                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3169                 handle_t *handle;
3170
3171                 /* (user+group)*(old+new) structure, inode write (sb,
3172                  * inode block, ? - but truncate inode update has it) */
3173                 handle = ext3_journal_start(inode, EXT3_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3174                                         EXT3_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)+3);
3175                 if (IS_ERR(handle)) {
3176                         error = PTR_ERR(handle);
3177                         goto err_out;
3178                 }
3179                 error = dquot_transfer(inode, attr);
3180                 if (error) {
3181                         ext3_journal_stop(handle);
3182                         return error;
3183                 }
3184                 /* Update corresponding info in inode so that everything is in
3185                  * one transaction */
3186                 if (attr->ia_valid & ATTR_UID)
3187                         inode->i_uid = attr->ia_uid;
3188                 if (attr->ia_valid & ATTR_GID)
3189                         inode->i_gid = attr->ia_gid;
3190                 error = ext3_mark_inode_dirty(handle, inode);
3191                 ext3_journal_stop(handle);
3192         }
3193
3194         if (S_ISREG(inode->i_mode) &&
3195             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3196                 handle_t *handle;
3197
3198                 handle = ext3_journal_start(inode, 3);
3199                 if (IS_ERR(handle)) {
3200                         error = PTR_ERR(handle);
3201                         goto err_out;
3202                 }
3203
3204                 error = ext3_orphan_add(handle, inode);
3205                 EXT3_I(inode)->i_disksize = attr->ia_size;
3206                 rc = ext3_mark_inode_dirty(handle, inode);
3207                 if (!error)
3208                         error = rc;
3209                 ext3_journal_stop(handle);
3210         }
3211
3212         rc = inode_setattr(inode, attr);
3213
3214         if (!rc && (ia_valid & ATTR_MODE))
3215                 rc = ext3_acl_chmod(inode);
3216
3217 err_out:
3218         ext3_std_error(inode->i_sb, error);
3219         if (!error)
3220                 error = rc;
3221         return error;
3222 }
3223
3224
3225 /*
3226  * How many blocks doth make a writepage()?
3227  *
3228  * With N blocks per page, it may be:
3229  * N data blocks
3230  * 2 indirect block
3231  * 2 dindirect
3232  * 1 tindirect
3233  * N+5 bitmap blocks (from the above)
3234  * N+5 group descriptor summary blocks
3235  * 1 inode block
3236  * 1 superblock.
3237  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3238  *
3239  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3240  *
3241  * With ordered or writeback data it's the same, less the N data blocks.
3242  *
3243  * If the inode's direct blocks can hold an integral number of pages then a
3244  * page cannot straddle two indirect blocks, and we can only touch one indirect
3245  * and dindirect block, and the "5" above becomes "3".
3246  *
3247  * This still overestimates under most circumstances.  If we were to pass the
3248  * start and end offsets in here as well we could do block_to_path() on each
3249  * block and work out the exact number of indirects which are touched.  Pah.
3250  */
3251
3252 static int ext3_writepage_trans_blocks(struct inode *inode)
3253 {
3254         int bpp = ext3_journal_blocks_per_page(inode);
3255         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3256         int ret;
3257
3258         if (ext3_should_journal_data(inode))
3259                 ret = 3 * (bpp + indirects) + 2;
3260         else
3261                 ret = 2 * (bpp + indirects) + 2;
3262
3263 #ifdef CONFIG_QUOTA
3264         /* We know that structure was already allocated during dquot_initialize so
3265          * we will be updating only the data blocks + inodes */
3266         ret += EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
3267 #endif
3268
3269         return ret;
3270 }
3271
3272 /*
3273  * The caller must have previously called ext3_reserve_inode_write().
3274  * Give this, we know that the caller already has write access to iloc->bh.
3275  */
3276 int ext3_mark_iloc_dirty(handle_t *handle,
3277                 struct inode *inode, struct ext3_iloc *iloc)
3278 {
3279         int err = 0;
3280
3281         /* the do_update_inode consumes one bh->b_count */
3282         get_bh(iloc->bh);
3283
3284         /* ext3_do_update_inode() does journal_dirty_metadata */
3285         err = ext3_do_update_inode(handle, inode, iloc);
3286         put_bh(iloc->bh);
3287         return err;
3288 }
3289
3290 /*
3291  * On success, We end up with an outstanding reference count against
3292  * iloc->bh.  This _must_ be cleaned up later.
3293  */
3294
3295 int
3296 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3297                          struct ext3_iloc *iloc)
3298 {
3299         int err = 0;
3300         if (handle) {
3301                 err = ext3_get_inode_loc(inode, iloc);
3302                 if (!err) {
3303                         BUFFER_TRACE(iloc->bh, "get_write_access");
3304                         err = ext3_journal_get_write_access(handle, iloc->bh);
3305                         if (err) {
3306                                 brelse(iloc->bh);
3307                                 iloc->bh = NULL;
3308                         }
3309                 }
3310         }
3311         ext3_std_error(inode->i_sb, err);
3312         return err;
3313 }
3314
3315 /*
3316  * What we do here is to mark the in-core inode as clean with respect to inode
3317  * dirtiness (it may still be data-dirty).
3318  * This means that the in-core inode may be reaped by prune_icache
3319  * without having to perform any I/O.  This is a very good thing,
3320  * because *any* task may call prune_icache - even ones which
3321  * have a transaction open against a different journal.
3322  *
3323  * Is this cheating?  Not really.  Sure, we haven't written the
3324  * inode out, but prune_icache isn't a user-visible syncing function.
3325  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3326  * we start and wait on commits.
3327  *
3328  * Is this efficient/effective?  Well, we're being nice to the system
3329  * by cleaning up our inodes proactively so they can be reaped
3330  * without I/O.  But we are potentially leaving up to five seconds'
3331  * worth of inodes floating about which prune_icache wants us to
3332  * write out.  One way to fix that would be to get prune_icache()
3333  * to do a write_super() to free up some memory.  It has the desired
3334  * effect.
3335  */
3336 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3337 {
3338         struct ext3_iloc iloc;
3339         int err;
3340
3341         might_sleep();
3342         err = ext3_reserve_inode_write(handle, inode, &iloc);
3343         if (!err)
3344                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3345         return err;
3346 }
3347
3348 /*
3349  * ext3_dirty_inode() is called from __mark_inode_dirty()
3350  *
3351  * We're really interested in the case where a file is being extended.
3352  * i_size has been changed by generic_commit_write() and we thus need
3353  * to include the updated inode in the current transaction.
3354  *
3355  * Also, dquot_alloc_space() will always dirty the inode when blocks
3356  * are allocated to the file.
3357  *
3358  * If the inode is marked synchronous, we don't honour that here - doing
3359  * so would cause a commit on atime updates, which we don't bother doing.
3360  * We handle synchronous inodes at the highest possible level.
3361  */
3362 void ext3_dirty_inode(struct inode *inode)
3363 {
3364         handle_t *current_handle = ext3_journal_current_handle();
3365         handle_t *handle;
3366
3367         handle = ext3_journal_start(inode, 2);
3368         if (IS_ERR(handle))
3369                 goto out;
3370         if (current_handle &&
3371                 current_handle->h_transaction != handle->h_transaction) {
3372                 /* This task has a transaction open against a different fs */
3373                 printk(KERN_EMERG "%s: transactions do not match!\n",
3374                        __func__);
3375         } else {
3376                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3377                                 current_handle);
3378                 ext3_mark_inode_dirty(handle, inode);
3379         }
3380         ext3_journal_stop(handle);
3381 out:
3382         return;
3383 }
3384
3385 #if 0
3386 /*
3387  * Bind an inode's backing buffer_head into this transaction, to prevent
3388  * it from being flushed to disk early.  Unlike
3389  * ext3_reserve_inode_write, this leaves behind no bh reference and
3390  * returns no iloc structure, so the caller needs to repeat the iloc
3391  * lookup to mark the inode dirty later.
3392  */
3393 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3394 {
3395         struct ext3_iloc iloc;
3396
3397         int err = 0;
3398         if (handle) {
3399                 err = ext3_get_inode_loc(inode, &iloc);
3400                 if (!err) {
3401                         BUFFER_TRACE(iloc.bh, "get_write_access");
3402                         err = journal_get_write_access(handle, iloc.bh);
3403                         if (!err)
3404                                 err = ext3_journal_dirty_metadata(handle,
3405                                                                   iloc.bh);
3406                         brelse(iloc.bh);
3407                 }
3408         }
3409         ext3_std_error(inode->i_sb, err);
3410         return err;
3411 }
3412 #endif
3413
3414 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3415 {
3416         journal_t *journal;
3417         handle_t *handle;
3418         int err;
3419
3420         /*
3421          * We have to be very careful here: changing a data block's
3422          * journaling status dynamically is dangerous.  If we write a
3423          * data block to the journal, change the status and then delete
3424          * that block, we risk forgetting to revoke the old log record
3425          * from the journal and so a subsequent replay can corrupt data.
3426          * So, first we make sure that the journal is empty and that
3427          * nobody is changing anything.
3428          */
3429
3430         journal = EXT3_JOURNAL(inode);
3431         if (is_journal_aborted(journal))
3432                 return -EROFS;
3433
3434         journal_lock_updates(journal);
3435         journal_flush(journal);
3436
3437         /*
3438          * OK, there are no updates running now, and all cached data is
3439          * synced to disk.  We are now in a completely consistent state
3440          * which doesn't have anything in the journal, and we know that
3441          * no filesystem updates are running, so it is safe to modify
3442          * the inode's in-core data-journaling state flag now.
3443          */
3444
3445         if (val)
3446                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3447         else
3448                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3449         ext3_set_aops(inode);
3450
3451         journal_unlock_updates(journal);
3452
3453         /* Finally we can mark the inode as dirty. */
3454
3455         handle = ext3_journal_start(inode, 1);
3456         if (IS_ERR(handle))
3457                 return PTR_ERR(handle);
3458
3459         err = ext3_mark_inode_dirty(handle, inode);
3460         handle->h_sync = 1;
3461         ext3_journal_stop(handle);
3462         ext3_std_error(inode->i_sb, err);
3463
3464         return err;
3465 }