convert ext3 to ->evict_inode()
[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 inode eviction from icache
194  */
195 void ext3_evict_inode (struct inode *inode)
196 {
197         struct ext3_block_alloc_info *rsv;
198         handle_t *handle;
199         int want_delete = 0;
200
201         if (!inode->i_nlink && !is_bad_inode(inode)) {
202                 dquot_initialize(inode);
203                 want_delete = 1;
204         }
205
206         truncate_inode_pages(&inode->i_data, 0);
207
208         ext3_discard_reservation(inode);
209         rsv = EXT3_I(inode)->i_block_alloc_info;
210         EXT3_I(inode)->i_block_alloc_info = NULL;
211         if (unlikely(rsv))
212                 kfree(rsv);
213
214         if (!want_delete)
215                 goto no_delete;
216
217         handle = start_transaction(inode);
218         if (IS_ERR(handle)) {
219                 /*
220                  * If we're going to skip the normal cleanup, we still need to
221                  * make sure that the in-core orphan linked list is properly
222                  * cleaned up.
223                  */
224                 ext3_orphan_del(NULL, inode);
225                 goto no_delete;
226         }
227
228         if (IS_SYNC(inode))
229                 handle->h_sync = 1;
230         inode->i_size = 0;
231         if (inode->i_blocks)
232                 ext3_truncate(inode);
233         /*
234          * Kill off the orphan record which ext3_truncate created.
235          * AKPM: I think this can be inside the above `if'.
236          * Note that ext3_orphan_del() has to be able to cope with the
237          * deletion of a non-existent orphan - this is because we don't
238          * know if ext3_truncate() actually created an orphan record.
239          * (Well, we could do this if we need to, but heck - it works)
240          */
241         ext3_orphan_del(handle, inode);
242         EXT3_I(inode)->i_dtime  = get_seconds();
243
244         /*
245          * One subtle ordering requirement: if anything has gone wrong
246          * (transaction abort, IO errors, whatever), then we can still
247          * do these next steps (the fs will already have been marked as
248          * having errors), but we can't free the inode if the mark_dirty
249          * fails.
250          */
251         if (ext3_mark_inode_dirty(handle, inode)) {
252                 /* If that failed, just dquot_drop() and be done with that */
253                 dquot_drop(inode);
254                 end_writeback(inode);
255         } else {
256                 ext3_xattr_delete_inode(handle, inode);
257                 dquot_free_inode(inode);
258                 dquot_drop(inode);
259                 end_writeback(inode);
260                 ext3_free_inode(handle, inode);
261         }
262         ext3_journal_stop(handle);
263         return;
264 no_delete:
265         end_writeback(inode);
266         dquot_drop(inode);
267 }
268
269 typedef struct {
270         __le32  *p;
271         __le32  key;
272         struct buffer_head *bh;
273 } Indirect;
274
275 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
276 {
277         p->key = *(p->p = v);
278         p->bh = bh;
279 }
280
281 static int verify_chain(Indirect *from, Indirect *to)
282 {
283         while (from <= to && from->key == *from->p)
284                 from++;
285         return (from > to);
286 }
287
288 /**
289  *      ext3_block_to_path - parse the block number into array of offsets
290  *      @inode: inode in question (we are only interested in its superblock)
291  *      @i_block: block number to be parsed
292  *      @offsets: array to store the offsets in
293  *      @boundary: set this non-zero if the referred-to block is likely to be
294  *             followed (on disk) by an indirect block.
295  *
296  *      To store the locations of file's data ext3 uses a data structure common
297  *      for UNIX filesystems - tree of pointers anchored in the inode, with
298  *      data blocks at leaves and indirect blocks in intermediate nodes.
299  *      This function translates the block number into path in that tree -
300  *      return value is the path length and @offsets[n] is the offset of
301  *      pointer to (n+1)th node in the nth one. If @block is out of range
302  *      (negative or too large) warning is printed and zero returned.
303  *
304  *      Note: function doesn't find node addresses, so no IO is needed. All
305  *      we need to know is the capacity of indirect blocks (taken from the
306  *      inode->i_sb).
307  */
308
309 /*
310  * Portability note: the last comparison (check that we fit into triple
311  * indirect block) is spelled differently, because otherwise on an
312  * architecture with 32-bit longs and 8Kb pages we might get into trouble
313  * if our filesystem had 8Kb blocks. We might use long long, but that would
314  * kill us on x86. Oh, well, at least the sign propagation does not matter -
315  * i_block would have to be negative in the very beginning, so we would not
316  * get there at all.
317  */
318
319 static int ext3_block_to_path(struct inode *inode,
320                         long i_block, int offsets[4], int *boundary)
321 {
322         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
323         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
324         const long direct_blocks = EXT3_NDIR_BLOCKS,
325                 indirect_blocks = ptrs,
326                 double_blocks = (1 << (ptrs_bits * 2));
327         int n = 0;
328         int final = 0;
329
330         if (i_block < 0) {
331                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
332         } else if (i_block < direct_blocks) {
333                 offsets[n++] = i_block;
334                 final = direct_blocks;
335         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
336                 offsets[n++] = EXT3_IND_BLOCK;
337                 offsets[n++] = i_block;
338                 final = ptrs;
339         } else if ((i_block -= indirect_blocks) < double_blocks) {
340                 offsets[n++] = EXT3_DIND_BLOCK;
341                 offsets[n++] = i_block >> ptrs_bits;
342                 offsets[n++] = i_block & (ptrs - 1);
343                 final = ptrs;
344         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
345                 offsets[n++] = EXT3_TIND_BLOCK;
346                 offsets[n++] = i_block >> (ptrs_bits * 2);
347                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
348                 offsets[n++] = i_block & (ptrs - 1);
349                 final = ptrs;
350         } else {
351                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
352         }
353         if (boundary)
354                 *boundary = final - 1 - (i_block & (ptrs - 1));
355         return n;
356 }
357
358 /**
359  *      ext3_get_branch - read the chain of indirect blocks leading to data
360  *      @inode: inode in question
361  *      @depth: depth of the chain (1 - direct pointer, etc.)
362  *      @offsets: offsets of pointers in inode/indirect blocks
363  *      @chain: place to store the result
364  *      @err: here we store the error value
365  *
366  *      Function fills the array of triples <key, p, bh> and returns %NULL
367  *      if everything went OK or the pointer to the last filled triple
368  *      (incomplete one) otherwise. Upon the return chain[i].key contains
369  *      the number of (i+1)-th block in the chain (as it is stored in memory,
370  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
371  *      number (it points into struct inode for i==0 and into the bh->b_data
372  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
373  *      block for i>0 and NULL for i==0. In other words, it holds the block
374  *      numbers of the chain, addresses they were taken from (and where we can
375  *      verify that chain did not change) and buffer_heads hosting these
376  *      numbers.
377  *
378  *      Function stops when it stumbles upon zero pointer (absent block)
379  *              (pointer to last triple returned, *@err == 0)
380  *      or when it gets an IO error reading an indirect block
381  *              (ditto, *@err == -EIO)
382  *      or when it notices that chain had been changed while it was reading
383  *              (ditto, *@err == -EAGAIN)
384  *      or when it reads all @depth-1 indirect blocks successfully and finds
385  *      the whole chain, all way to the data (returns %NULL, *err == 0).
386  */
387 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
388                                  Indirect chain[4], int *err)
389 {
390         struct super_block *sb = inode->i_sb;
391         Indirect *p = chain;
392         struct buffer_head *bh;
393
394         *err = 0;
395         /* i_data is not going away, no lock needed */
396         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
397         if (!p->key)
398                 goto no_block;
399         while (--depth) {
400                 bh = sb_bread(sb, le32_to_cpu(p->key));
401                 if (!bh)
402                         goto failure;
403                 /* Reader: pointers */
404                 if (!verify_chain(chain, p))
405                         goto changed;
406                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
407                 /* Reader: end */
408                 if (!p->key)
409                         goto no_block;
410         }
411         return NULL;
412
413 changed:
414         brelse(bh);
415         *err = -EAGAIN;
416         goto no_block;
417 failure:
418         *err = -EIO;
419 no_block:
420         return p;
421 }
422
423 /**
424  *      ext3_find_near - find a place for allocation with sufficient locality
425  *      @inode: owner
426  *      @ind: descriptor of indirect block.
427  *
428  *      This function returns the preferred place for block allocation.
429  *      It is used when heuristic for sequential allocation fails.
430  *      Rules are:
431  *        + if there is a block to the left of our position - allocate near it.
432  *        + if pointer will live in indirect block - allocate near that block.
433  *        + if pointer will live in inode - allocate in the same
434  *          cylinder group.
435  *
436  * In the latter case we colour the starting block by the callers PID to
437  * prevent it from clashing with concurrent allocations for a different inode
438  * in the same block group.   The PID is used here so that functionally related
439  * files will be close-by on-disk.
440  *
441  *      Caller must make sure that @ind is valid and will stay that way.
442  */
443 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
444 {
445         struct ext3_inode_info *ei = EXT3_I(inode);
446         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
447         __le32 *p;
448         ext3_fsblk_t bg_start;
449         ext3_grpblk_t colour;
450
451         /* Try to find previous block */
452         for (p = ind->p - 1; p >= start; p--) {
453                 if (*p)
454                         return le32_to_cpu(*p);
455         }
456
457         /* No such thing, so let's try location of indirect block */
458         if (ind->bh)
459                 return ind->bh->b_blocknr;
460
461         /*
462          * It is going to be referred to from the inode itself? OK, just put it
463          * into the same cylinder group then.
464          */
465         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
466         colour = (current->pid % 16) *
467                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
468         return bg_start + colour;
469 }
470
471 /**
472  *      ext3_find_goal - find a preferred place for allocation.
473  *      @inode: owner
474  *      @block:  block we want
475  *      @partial: pointer to the last triple within a chain
476  *
477  *      Normally this function find the preferred place for block allocation,
478  *      returns it.
479  */
480
481 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
482                                    Indirect *partial)
483 {
484         struct ext3_block_alloc_info *block_i;
485
486         block_i =  EXT3_I(inode)->i_block_alloc_info;
487
488         /*
489          * try the heuristic for sequential allocation,
490          * failing that at least try to get decent locality.
491          */
492         if (block_i && (block == block_i->last_alloc_logical_block + 1)
493                 && (block_i->last_alloc_physical_block != 0)) {
494                 return block_i->last_alloc_physical_block + 1;
495         }
496
497         return ext3_find_near(inode, partial);
498 }
499
500 /**
501  *      ext3_blks_to_allocate: Look up the block map and count the number
502  *      of direct blocks need to be allocated for the given branch.
503  *
504  *      @branch: chain of indirect blocks
505  *      @k: number of blocks need for indirect blocks
506  *      @blks: number of data blocks to be mapped.
507  *      @blocks_to_boundary:  the offset in the indirect block
508  *
509  *      return the total number of blocks to be allocate, including the
510  *      direct and indirect blocks.
511  */
512 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
513                 int blocks_to_boundary)
514 {
515         unsigned long count = 0;
516
517         /*
518          * Simple case, [t,d]Indirect block(s) has not allocated yet
519          * then it's clear blocks on that path have not allocated
520          */
521         if (k > 0) {
522                 /* right now we don't handle cross boundary allocation */
523                 if (blks < blocks_to_boundary + 1)
524                         count += blks;
525                 else
526                         count += blocks_to_boundary + 1;
527                 return count;
528         }
529
530         count++;
531         while (count < blks && count <= blocks_to_boundary &&
532                 le32_to_cpu(*(branch[0].p + count)) == 0) {
533                 count++;
534         }
535         return count;
536 }
537
538 /**
539  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
540  *      @indirect_blks: the number of blocks need to allocate for indirect
541  *                      blocks
542  *
543  *      @new_blocks: on return it will store the new block numbers for
544  *      the indirect blocks(if needed) and the first direct block,
545  *      @blks:  on return it will store the total number of allocated
546  *              direct blocks
547  */
548 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
549                         ext3_fsblk_t goal, int indirect_blks, int blks,
550                         ext3_fsblk_t new_blocks[4], int *err)
551 {
552         int target, i;
553         unsigned long count = 0;
554         int index = 0;
555         ext3_fsblk_t current_block = 0;
556         int ret = 0;
557
558         /*
559          * Here we try to allocate the requested multiple blocks at once,
560          * on a best-effort basis.
561          * To build a branch, we should allocate blocks for
562          * the indirect blocks(if not allocated yet), and at least
563          * the first direct block of this branch.  That's the
564          * minimum number of blocks need to allocate(required)
565          */
566         target = blks + indirect_blks;
567
568         while (1) {
569                 count = target;
570                 /* allocating blocks for indirect blocks and direct blocks */
571                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
572                 if (*err)
573                         goto failed_out;
574
575                 target -= count;
576                 /* allocate blocks for indirect blocks */
577                 while (index < indirect_blks && count) {
578                         new_blocks[index++] = current_block++;
579                         count--;
580                 }
581
582                 if (count > 0)
583                         break;
584         }
585
586         /* save the new block number for the first direct block */
587         new_blocks[index] = current_block;
588
589         /* total number of blocks allocated for direct blocks */
590         ret = count;
591         *err = 0;
592         return ret;
593 failed_out:
594         for (i = 0; i <index; i++)
595                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
596         return ret;
597 }
598
599 /**
600  *      ext3_alloc_branch - allocate and set up a chain of blocks.
601  *      @inode: owner
602  *      @indirect_blks: number of allocated indirect blocks
603  *      @blks: number of allocated direct blocks
604  *      @offsets: offsets (in the blocks) to store the pointers to next.
605  *      @branch: place to store the chain in.
606  *
607  *      This function allocates blocks, zeroes out all but the last one,
608  *      links them into chain and (if we are synchronous) writes them to disk.
609  *      In other words, it prepares a branch that can be spliced onto the
610  *      inode. It stores the information about that chain in the branch[], in
611  *      the same format as ext3_get_branch() would do. We are calling it after
612  *      we had read the existing part of chain and partial points to the last
613  *      triple of that (one with zero ->key). Upon the exit we have the same
614  *      picture as after the successful ext3_get_block(), except that in one
615  *      place chain is disconnected - *branch->p is still zero (we did not
616  *      set the last link), but branch->key contains the number that should
617  *      be placed into *branch->p to fill that gap.
618  *
619  *      If allocation fails we free all blocks we've allocated (and forget
620  *      their buffer_heads) and return the error value the from failed
621  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
622  *      as described above and return 0.
623  */
624 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
625                         int indirect_blks, int *blks, ext3_fsblk_t goal,
626                         int *offsets, Indirect *branch)
627 {
628         int blocksize = inode->i_sb->s_blocksize;
629         int i, n = 0;
630         int err = 0;
631         struct buffer_head *bh;
632         int num;
633         ext3_fsblk_t new_blocks[4];
634         ext3_fsblk_t current_block;
635
636         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
637                                 *blks, new_blocks, &err);
638         if (err)
639                 return err;
640
641         branch[0].key = cpu_to_le32(new_blocks[0]);
642         /*
643          * metadata blocks and data blocks are allocated.
644          */
645         for (n = 1; n <= indirect_blks;  n++) {
646                 /*
647                  * Get buffer_head for parent block, zero it out
648                  * and set the pointer to new one, then send
649                  * parent to disk.
650                  */
651                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
652                 branch[n].bh = bh;
653                 lock_buffer(bh);
654                 BUFFER_TRACE(bh, "call get_create_access");
655                 err = ext3_journal_get_create_access(handle, bh);
656                 if (err) {
657                         unlock_buffer(bh);
658                         brelse(bh);
659                         goto failed;
660                 }
661
662                 memset(bh->b_data, 0, blocksize);
663                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
664                 branch[n].key = cpu_to_le32(new_blocks[n]);
665                 *branch[n].p = branch[n].key;
666                 if ( n == indirect_blks) {
667                         current_block = new_blocks[n];
668                         /*
669                          * End of chain, update the last new metablock of
670                          * the chain to point to the new allocated
671                          * data blocks numbers
672                          */
673                         for (i=1; i < num; i++)
674                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
675                 }
676                 BUFFER_TRACE(bh, "marking uptodate");
677                 set_buffer_uptodate(bh);
678                 unlock_buffer(bh);
679
680                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
681                 err = ext3_journal_dirty_metadata(handle, bh);
682                 if (err)
683                         goto failed;
684         }
685         *blks = num;
686         return err;
687 failed:
688         /* Allocation failed, free what we already allocated */
689         for (i = 1; i <= n ; i++) {
690                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
691                 ext3_journal_forget(handle, branch[i].bh);
692         }
693         for (i = 0; i <indirect_blks; i++)
694                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
695
696         ext3_free_blocks(handle, inode, new_blocks[i], num);
697
698         return err;
699 }
700
701 /**
702  * ext3_splice_branch - splice the allocated branch onto inode.
703  * @inode: owner
704  * @block: (logical) number of block we are adding
705  * @chain: chain of indirect blocks (with a missing link - see
706  *      ext3_alloc_branch)
707  * @where: location of missing link
708  * @num:   number of indirect blocks we are adding
709  * @blks:  number of direct blocks we are adding
710  *
711  * This function fills the missing link and does all housekeeping needed in
712  * inode (->i_blocks, etc.). In case of success we end up with the full
713  * chain to new block and return 0.
714  */
715 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
716                         long block, Indirect *where, int num, int blks)
717 {
718         int i;
719         int err = 0;
720         struct ext3_block_alloc_info *block_i;
721         ext3_fsblk_t current_block;
722         struct ext3_inode_info *ei = EXT3_I(inode);
723
724         block_i = ei->i_block_alloc_info;
725         /*
726          * If we're splicing into a [td]indirect block (as opposed to the
727          * inode) then we need to get write access to the [td]indirect block
728          * before the splice.
729          */
730         if (where->bh) {
731                 BUFFER_TRACE(where->bh, "get_write_access");
732                 err = ext3_journal_get_write_access(handle, where->bh);
733                 if (err)
734                         goto err_out;
735         }
736         /* That's it */
737
738         *where->p = where->key;
739
740         /*
741          * Update the host buffer_head or inode to point to more just allocated
742          * direct blocks blocks
743          */
744         if (num == 0 && blks > 1) {
745                 current_block = le32_to_cpu(where->key) + 1;
746                 for (i = 1; i < blks; i++)
747                         *(where->p + i ) = cpu_to_le32(current_block++);
748         }
749
750         /*
751          * update the most recently allocated logical & physical block
752          * in i_block_alloc_info, to assist find the proper goal block for next
753          * allocation
754          */
755         if (block_i) {
756                 block_i->last_alloc_logical_block = block + blks - 1;
757                 block_i->last_alloc_physical_block =
758                                 le32_to_cpu(where[num].key) + blks - 1;
759         }
760
761         /* We are done with atomic stuff, now do the rest of housekeeping */
762
763         inode->i_ctime = CURRENT_TIME_SEC;
764         ext3_mark_inode_dirty(handle, inode);
765         /* ext3_mark_inode_dirty already updated i_sync_tid */
766         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
767
768         /* had we spliced it onto indirect block? */
769         if (where->bh) {
770                 /*
771                  * If we spliced it onto an indirect block, we haven't
772                  * altered the inode.  Note however that if it is being spliced
773                  * onto an indirect block at the very end of the file (the
774                  * file is growing) then we *will* alter the inode to reflect
775                  * the new i_size.  But that is not done here - it is done in
776                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
777                  */
778                 jbd_debug(5, "splicing indirect only\n");
779                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
780                 err = ext3_journal_dirty_metadata(handle, where->bh);
781                 if (err)
782                         goto err_out;
783         } else {
784                 /*
785                  * OK, we spliced it into the inode itself on a direct block.
786                  * Inode was dirtied above.
787                  */
788                 jbd_debug(5, "splicing direct\n");
789         }
790         return err;
791
792 err_out:
793         for (i = 1; i <= num; i++) {
794                 BUFFER_TRACE(where[i].bh, "call journal_forget");
795                 ext3_journal_forget(handle, where[i].bh);
796                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
797         }
798         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
799
800         return err;
801 }
802
803 /*
804  * Allocation strategy is simple: if we have to allocate something, we will
805  * have to go the whole way to leaf. So let's do it before attaching anything
806  * to tree, set linkage between the newborn blocks, write them if sync is
807  * required, recheck the path, free and repeat if check fails, otherwise
808  * set the last missing link (that will protect us from any truncate-generated
809  * removals - all blocks on the path are immune now) and possibly force the
810  * write on the parent block.
811  * That has a nice additional property: no special recovery from the failed
812  * allocations is needed - we simply release blocks and do not touch anything
813  * reachable from inode.
814  *
815  * `handle' can be NULL if create == 0.
816  *
817  * The BKL may not be held on entry here.  Be sure to take it early.
818  * return > 0, # of blocks mapped or allocated.
819  * return = 0, if plain lookup failed.
820  * return < 0, error case.
821  */
822 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
823                 sector_t iblock, unsigned long maxblocks,
824                 struct buffer_head *bh_result,
825                 int create)
826 {
827         int err = -EIO;
828         int offsets[4];
829         Indirect chain[4];
830         Indirect *partial;
831         ext3_fsblk_t goal;
832         int indirect_blks;
833         int blocks_to_boundary = 0;
834         int depth;
835         struct ext3_inode_info *ei = EXT3_I(inode);
836         int count = 0;
837         ext3_fsblk_t first_block = 0;
838
839
840         J_ASSERT(handle != NULL || create == 0);
841         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
842
843         if (depth == 0)
844                 goto out;
845
846         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
847
848         /* Simplest case - block found, no allocation needed */
849         if (!partial) {
850                 first_block = le32_to_cpu(chain[depth - 1].key);
851                 clear_buffer_new(bh_result);
852                 count++;
853                 /*map more blocks*/
854                 while (count < maxblocks && count <= blocks_to_boundary) {
855                         ext3_fsblk_t blk;
856
857                         if (!verify_chain(chain, chain + depth - 1)) {
858                                 /*
859                                  * Indirect block might be removed by
860                                  * truncate while we were reading it.
861                                  * Handling of that case: forget what we've
862                                  * got now. Flag the err as EAGAIN, so it
863                                  * will reread.
864                                  */
865                                 err = -EAGAIN;
866                                 count = 0;
867                                 break;
868                         }
869                         blk = le32_to_cpu(*(chain[depth-1].p + count));
870
871                         if (blk == first_block + count)
872                                 count++;
873                         else
874                                 break;
875                 }
876                 if (err != -EAGAIN)
877                         goto got_it;
878         }
879
880         /* Next simple case - plain lookup or failed read of indirect block */
881         if (!create || err == -EIO)
882                 goto cleanup;
883
884         mutex_lock(&ei->truncate_mutex);
885
886         /*
887          * If the indirect block is missing while we are reading
888          * the chain(ext3_get_branch() returns -EAGAIN err), or
889          * if the chain has been changed after we grab the semaphore,
890          * (either because another process truncated this branch, or
891          * another get_block allocated this branch) re-grab the chain to see if
892          * the request block has been allocated or not.
893          *
894          * Since we already block the truncate/other get_block
895          * at this point, we will have the current copy of the chain when we
896          * splice the branch into the tree.
897          */
898         if (err == -EAGAIN || !verify_chain(chain, partial)) {
899                 while (partial > chain) {
900                         brelse(partial->bh);
901                         partial--;
902                 }
903                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
904                 if (!partial) {
905                         count++;
906                         mutex_unlock(&ei->truncate_mutex);
907                         if (err)
908                                 goto cleanup;
909                         clear_buffer_new(bh_result);
910                         goto got_it;
911                 }
912         }
913
914         /*
915          * Okay, we need to do block allocation.  Lazily initialize the block
916          * allocation info here if necessary
917         */
918         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
919                 ext3_init_block_alloc_info(inode);
920
921         goal = ext3_find_goal(inode, iblock, partial);
922
923         /* the number of blocks need to allocate for [d,t]indirect blocks */
924         indirect_blks = (chain + depth) - partial - 1;
925
926         /*
927          * Next look up the indirect map to count the totoal number of
928          * direct blocks to allocate for this branch.
929          */
930         count = ext3_blks_to_allocate(partial, indirect_blks,
931                                         maxblocks, blocks_to_boundary);
932         /*
933          * Block out ext3_truncate while we alter the tree
934          */
935         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
936                                 offsets + (partial - chain), partial);
937
938         /*
939          * The ext3_splice_branch call will free and forget any buffers
940          * on the new chain if there is a failure, but that risks using
941          * up transaction credits, especially for bitmaps where the
942          * credits cannot be returned.  Can we handle this somehow?  We
943          * may need to return -EAGAIN upwards in the worst case.  --sct
944          */
945         if (!err)
946                 err = ext3_splice_branch(handle, inode, iblock,
947                                         partial, indirect_blks, count);
948         mutex_unlock(&ei->truncate_mutex);
949         if (err)
950                 goto cleanup;
951
952         set_buffer_new(bh_result);
953 got_it:
954         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
955         if (count > blocks_to_boundary)
956                 set_buffer_boundary(bh_result);
957         err = count;
958         /* Clean up and exit */
959         partial = chain + depth - 1;    /* the whole chain */
960 cleanup:
961         while (partial > chain) {
962                 BUFFER_TRACE(partial->bh, "call brelse");
963                 brelse(partial->bh);
964                 partial--;
965         }
966         BUFFER_TRACE(bh_result, "returned");
967 out:
968         return err;
969 }
970
971 /* Maximum number of blocks we map for direct IO at once. */
972 #define DIO_MAX_BLOCKS 4096
973 /*
974  * Number of credits we need for writing DIO_MAX_BLOCKS:
975  * We need sb + group descriptor + bitmap + inode -> 4
976  * For B blocks with A block pointers per block we need:
977  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
978  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
979  */
980 #define DIO_CREDITS 25
981
982 static int ext3_get_block(struct inode *inode, sector_t iblock,
983                         struct buffer_head *bh_result, int create)
984 {
985         handle_t *handle = ext3_journal_current_handle();
986         int ret = 0, started = 0;
987         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
988
989         if (create && !handle) {        /* Direct IO write... */
990                 if (max_blocks > DIO_MAX_BLOCKS)
991                         max_blocks = DIO_MAX_BLOCKS;
992                 handle = ext3_journal_start(inode, DIO_CREDITS +
993                                 EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb));
994                 if (IS_ERR(handle)) {
995                         ret = PTR_ERR(handle);
996                         goto out;
997                 }
998                 started = 1;
999         }
1000
1001         ret = ext3_get_blocks_handle(handle, inode, iblock,
1002                                         max_blocks, bh_result, create);
1003         if (ret > 0) {
1004                 bh_result->b_size = (ret << inode->i_blkbits);
1005                 ret = 0;
1006         }
1007         if (started)
1008                 ext3_journal_stop(handle);
1009 out:
1010         return ret;
1011 }
1012
1013 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
1014                 u64 start, u64 len)
1015 {
1016         return generic_block_fiemap(inode, fieinfo, start, len,
1017                                     ext3_get_block);
1018 }
1019
1020 /*
1021  * `handle' can be NULL if create is zero
1022  */
1023 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1024                                 long block, int create, int *errp)
1025 {
1026         struct buffer_head dummy;
1027         int fatal = 0, err;
1028
1029         J_ASSERT(handle != NULL || create == 0);
1030
1031         dummy.b_state = 0;
1032         dummy.b_blocknr = -1000;
1033         buffer_trace_init(&dummy.b_history);
1034         err = ext3_get_blocks_handle(handle, inode, block, 1,
1035                                         &dummy, create);
1036         /*
1037          * ext3_get_blocks_handle() returns number of blocks
1038          * mapped. 0 in case of a HOLE.
1039          */
1040         if (err > 0) {
1041                 if (err > 1)
1042                         WARN_ON(1);
1043                 err = 0;
1044         }
1045         *errp = err;
1046         if (!err && buffer_mapped(&dummy)) {
1047                 struct buffer_head *bh;
1048                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1049                 if (!bh) {
1050                         *errp = -EIO;
1051                         goto err;
1052                 }
1053                 if (buffer_new(&dummy)) {
1054                         J_ASSERT(create != 0);
1055                         J_ASSERT(handle != NULL);
1056
1057                         /*
1058                          * Now that we do not always journal data, we should
1059                          * keep in mind whether this should always journal the
1060                          * new buffer as metadata.  For now, regular file
1061                          * writes use ext3_get_block instead, so it's not a
1062                          * problem.
1063                          */
1064                         lock_buffer(bh);
1065                         BUFFER_TRACE(bh, "call get_create_access");
1066                         fatal = ext3_journal_get_create_access(handle, bh);
1067                         if (!fatal && !buffer_uptodate(bh)) {
1068                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1069                                 set_buffer_uptodate(bh);
1070                         }
1071                         unlock_buffer(bh);
1072                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1073                         err = ext3_journal_dirty_metadata(handle, bh);
1074                         if (!fatal)
1075                                 fatal = err;
1076                 } else {
1077                         BUFFER_TRACE(bh, "not a new buffer");
1078                 }
1079                 if (fatal) {
1080                         *errp = fatal;
1081                         brelse(bh);
1082                         bh = NULL;
1083                 }
1084                 return bh;
1085         }
1086 err:
1087         return NULL;
1088 }
1089
1090 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1091                                int block, int create, int *err)
1092 {
1093         struct buffer_head * bh;
1094
1095         bh = ext3_getblk(handle, inode, block, create, err);
1096         if (!bh)
1097                 return bh;
1098         if (buffer_uptodate(bh))
1099                 return bh;
1100         ll_rw_block(READ_META, 1, &bh);
1101         wait_on_buffer(bh);
1102         if (buffer_uptodate(bh))
1103                 return bh;
1104         put_bh(bh);
1105         *err = -EIO;
1106         return NULL;
1107 }
1108
1109 static int walk_page_buffers(   handle_t *handle,
1110                                 struct buffer_head *head,
1111                                 unsigned from,
1112                                 unsigned to,
1113                                 int *partial,
1114                                 int (*fn)(      handle_t *handle,
1115                                                 struct buffer_head *bh))
1116 {
1117         struct buffer_head *bh;
1118         unsigned block_start, block_end;
1119         unsigned blocksize = head->b_size;
1120         int err, ret = 0;
1121         struct buffer_head *next;
1122
1123         for (   bh = head, block_start = 0;
1124                 ret == 0 && (bh != head || !block_start);
1125                 block_start = block_end, bh = next)
1126         {
1127                 next = bh->b_this_page;
1128                 block_end = block_start + blocksize;
1129                 if (block_end <= from || block_start >= to) {
1130                         if (partial && !buffer_uptodate(bh))
1131                                 *partial = 1;
1132                         continue;
1133                 }
1134                 err = (*fn)(handle, bh);
1135                 if (!ret)
1136                         ret = err;
1137         }
1138         return ret;
1139 }
1140
1141 /*
1142  * To preserve ordering, it is essential that the hole instantiation and
1143  * the data write be encapsulated in a single transaction.  We cannot
1144  * close off a transaction and start a new one between the ext3_get_block()
1145  * and the commit_write().  So doing the journal_start at the start of
1146  * prepare_write() is the right place.
1147  *
1148  * Also, this function can nest inside ext3_writepage() ->
1149  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1150  * has generated enough buffer credits to do the whole page.  So we won't
1151  * block on the journal in that case, which is good, because the caller may
1152  * be PF_MEMALLOC.
1153  *
1154  * By accident, ext3 can be reentered when a transaction is open via
1155  * quota file writes.  If we were to commit the transaction while thus
1156  * reentered, there can be a deadlock - we would be holding a quota
1157  * lock, and the commit would never complete if another thread had a
1158  * transaction open and was blocking on the quota lock - a ranking
1159  * violation.
1160  *
1161  * So what we do is to rely on the fact that journal_stop/journal_start
1162  * will _not_ run commit under these circumstances because handle->h_ref
1163  * is elevated.  We'll still have enough credits for the tiny quotafile
1164  * write.
1165  */
1166 static int do_journal_get_write_access(handle_t *handle,
1167                                         struct buffer_head *bh)
1168 {
1169         if (!buffer_mapped(bh) || buffer_freed(bh))
1170                 return 0;
1171         return ext3_journal_get_write_access(handle, bh);
1172 }
1173
1174 /*
1175  * Truncate blocks that were not used by write. We have to truncate the
1176  * pagecache as well so that corresponding buffers get properly unmapped.
1177  */
1178 static void ext3_truncate_failed_write(struct inode *inode)
1179 {
1180         truncate_inode_pages(inode->i_mapping, inode->i_size);
1181         ext3_truncate(inode);
1182 }
1183
1184 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1185                                 loff_t pos, unsigned len, unsigned flags,
1186                                 struct page **pagep, void **fsdata)
1187 {
1188         struct inode *inode = mapping->host;
1189         int ret;
1190         handle_t *handle;
1191         int retries = 0;
1192         struct page *page;
1193         pgoff_t index;
1194         unsigned from, to;
1195         /* Reserve one block more for addition to orphan list in case
1196          * we allocate blocks but write fails for some reason */
1197         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1198
1199         index = pos >> PAGE_CACHE_SHIFT;
1200         from = pos & (PAGE_CACHE_SIZE - 1);
1201         to = from + len;
1202
1203 retry:
1204         page = grab_cache_page_write_begin(mapping, index, flags);
1205         if (!page)
1206                 return -ENOMEM;
1207         *pagep = page;
1208
1209         handle = ext3_journal_start(inode, needed_blocks);
1210         if (IS_ERR(handle)) {
1211                 unlock_page(page);
1212                 page_cache_release(page);
1213                 ret = PTR_ERR(handle);
1214                 goto out;
1215         }
1216         ret = __block_write_begin(page, pos, len, ext3_get_block);
1217         if (ret)
1218                 goto write_begin_failed;
1219
1220         if (ext3_should_journal_data(inode)) {
1221                 ret = walk_page_buffers(handle, page_buffers(page),
1222                                 from, to, NULL, do_journal_get_write_access);
1223         }
1224 write_begin_failed:
1225         if (ret) {
1226                 /*
1227                  * block_write_begin may have instantiated a few blocks
1228                  * outside i_size.  Trim these off again. Don't need
1229                  * i_size_read because we hold i_mutex.
1230                  *
1231                  * Add inode to orphan list in case we crash before truncate
1232                  * finishes. Do this only if ext3_can_truncate() agrees so
1233                  * that orphan processing code is happy.
1234                  */
1235                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1236                         ext3_orphan_add(handle, inode);
1237                 ext3_journal_stop(handle);
1238                 unlock_page(page);
1239                 page_cache_release(page);
1240                 if (pos + len > inode->i_size)
1241                         ext3_truncate_failed_write(inode);
1242         }
1243         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1244                 goto retry;
1245 out:
1246         return ret;
1247 }
1248
1249
1250 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1251 {
1252         int err = journal_dirty_data(handle, bh);
1253         if (err)
1254                 ext3_journal_abort_handle(__func__, __func__,
1255                                                 bh, handle, err);
1256         return err;
1257 }
1258
1259 /* For ordered writepage and write_end functions */
1260 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1261 {
1262         /*
1263          * Write could have mapped the buffer but it didn't copy the data in
1264          * yet. So avoid filing such buffer into a transaction.
1265          */
1266         if (buffer_mapped(bh) && buffer_uptodate(bh))
1267                 return ext3_journal_dirty_data(handle, bh);
1268         return 0;
1269 }
1270
1271 /* For write_end() in data=journal mode */
1272 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1273 {
1274         if (!buffer_mapped(bh) || buffer_freed(bh))
1275                 return 0;
1276         set_buffer_uptodate(bh);
1277         return ext3_journal_dirty_metadata(handle, bh);
1278 }
1279
1280 /*
1281  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1282  * for the whole page but later we failed to copy the data in. Update inode
1283  * size according to what we managed to copy. The rest is going to be
1284  * truncated in write_end function.
1285  */
1286 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1287 {
1288         /* What matters to us is i_disksize. We don't write i_size anywhere */
1289         if (pos + copied > inode->i_size)
1290                 i_size_write(inode, pos + copied);
1291         if (pos + copied > EXT3_I(inode)->i_disksize) {
1292                 EXT3_I(inode)->i_disksize = pos + copied;
1293                 mark_inode_dirty(inode);
1294         }
1295 }
1296
1297 /*
1298  * We need to pick up the new inode size which generic_commit_write gave us
1299  * `file' can be NULL - eg, when called from page_symlink().
1300  *
1301  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1302  * buffers are managed internally.
1303  */
1304 static int ext3_ordered_write_end(struct file *file,
1305                                 struct address_space *mapping,
1306                                 loff_t pos, unsigned len, unsigned copied,
1307                                 struct page *page, void *fsdata)
1308 {
1309         handle_t *handle = ext3_journal_current_handle();
1310         struct inode *inode = file->f_mapping->host;
1311         unsigned from, to;
1312         int ret = 0, ret2;
1313
1314         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1315
1316         from = pos & (PAGE_CACHE_SIZE - 1);
1317         to = from + copied;
1318         ret = walk_page_buffers(handle, page_buffers(page),
1319                 from, to, NULL, journal_dirty_data_fn);
1320
1321         if (ret == 0)
1322                 update_file_sizes(inode, pos, copied);
1323         /*
1324          * There may be allocated blocks outside of i_size because
1325          * we failed to copy some data. Prepare for truncate.
1326          */
1327         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1328                 ext3_orphan_add(handle, inode);
1329         ret2 = ext3_journal_stop(handle);
1330         if (!ret)
1331                 ret = ret2;
1332         unlock_page(page);
1333         page_cache_release(page);
1334
1335         if (pos + len > inode->i_size)
1336                 ext3_truncate_failed_write(inode);
1337         return ret ? ret : copied;
1338 }
1339
1340 static int ext3_writeback_write_end(struct file *file,
1341                                 struct address_space *mapping,
1342                                 loff_t pos, unsigned len, unsigned copied,
1343                                 struct page *page, void *fsdata)
1344 {
1345         handle_t *handle = ext3_journal_current_handle();
1346         struct inode *inode = file->f_mapping->host;
1347         int ret;
1348
1349         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1350         update_file_sizes(inode, pos, copied);
1351         /*
1352          * There may be allocated blocks outside of i_size because
1353          * we failed to copy some data. Prepare for truncate.
1354          */
1355         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1356                 ext3_orphan_add(handle, inode);
1357         ret = ext3_journal_stop(handle);
1358         unlock_page(page);
1359         page_cache_release(page);
1360
1361         if (pos + len > inode->i_size)
1362                 ext3_truncate_failed_write(inode);
1363         return ret ? ret : copied;
1364 }
1365
1366 static int ext3_journalled_write_end(struct file *file,
1367                                 struct address_space *mapping,
1368                                 loff_t pos, unsigned len, unsigned copied,
1369                                 struct page *page, void *fsdata)
1370 {
1371         handle_t *handle = ext3_journal_current_handle();
1372         struct inode *inode = mapping->host;
1373         int ret = 0, ret2;
1374         int partial = 0;
1375         unsigned from, to;
1376
1377         from = pos & (PAGE_CACHE_SIZE - 1);
1378         to = from + len;
1379
1380         if (copied < len) {
1381                 if (!PageUptodate(page))
1382                         copied = 0;
1383                 page_zero_new_buffers(page, from + copied, to);
1384                 to = from + copied;
1385         }
1386
1387         ret = walk_page_buffers(handle, page_buffers(page), from,
1388                                 to, &partial, write_end_fn);
1389         if (!partial)
1390                 SetPageUptodate(page);
1391
1392         if (pos + copied > inode->i_size)
1393                 i_size_write(inode, pos + copied);
1394         /*
1395          * There may be allocated blocks outside of i_size because
1396          * we failed to copy some data. Prepare for truncate.
1397          */
1398         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1399                 ext3_orphan_add(handle, inode);
1400         ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1401         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1402                 EXT3_I(inode)->i_disksize = inode->i_size;
1403                 ret2 = ext3_mark_inode_dirty(handle, inode);
1404                 if (!ret)
1405                         ret = ret2;
1406         }
1407
1408         ret2 = ext3_journal_stop(handle);
1409         if (!ret)
1410                 ret = ret2;
1411         unlock_page(page);
1412         page_cache_release(page);
1413
1414         if (pos + len > inode->i_size)
1415                 ext3_truncate_failed_write(inode);
1416         return ret ? ret : copied;
1417 }
1418
1419 /*
1420  * bmap() is special.  It gets used by applications such as lilo and by
1421  * the swapper to find the on-disk block of a specific piece of data.
1422  *
1423  * Naturally, this is dangerous if the block concerned is still in the
1424  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1425  * filesystem and enables swap, then they may get a nasty shock when the
1426  * data getting swapped to that swapfile suddenly gets overwritten by
1427  * the original zero's written out previously to the journal and
1428  * awaiting writeback in the kernel's buffer cache.
1429  *
1430  * So, if we see any bmap calls here on a modified, data-journaled file,
1431  * take extra steps to flush any blocks which might be in the cache.
1432  */
1433 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1434 {
1435         struct inode *inode = mapping->host;
1436         journal_t *journal;
1437         int err;
1438
1439         if (ext3_test_inode_state(inode, EXT3_STATE_JDATA)) {
1440                 /*
1441                  * This is a REALLY heavyweight approach, but the use of
1442                  * bmap on dirty files is expected to be extremely rare:
1443                  * only if we run lilo or swapon on a freshly made file
1444                  * do we expect this to happen.
1445                  *
1446                  * (bmap requires CAP_SYS_RAWIO so this does not
1447                  * represent an unprivileged user DOS attack --- we'd be
1448                  * in trouble if mortal users could trigger this path at
1449                  * will.)
1450                  *
1451                  * NB. EXT3_STATE_JDATA is not set on files other than
1452                  * regular files.  If somebody wants to bmap a directory
1453                  * or symlink and gets confused because the buffer
1454                  * hasn't yet been flushed to disk, they deserve
1455                  * everything they get.
1456                  */
1457
1458                 ext3_clear_inode_state(inode, EXT3_STATE_JDATA);
1459                 journal = EXT3_JOURNAL(inode);
1460                 journal_lock_updates(journal);
1461                 err = journal_flush(journal);
1462                 journal_unlock_updates(journal);
1463
1464                 if (err)
1465                         return 0;
1466         }
1467
1468         return generic_block_bmap(mapping,block,ext3_get_block);
1469 }
1470
1471 static int bget_one(handle_t *handle, struct buffer_head *bh)
1472 {
1473         get_bh(bh);
1474         return 0;
1475 }
1476
1477 static int bput_one(handle_t *handle, struct buffer_head *bh)
1478 {
1479         put_bh(bh);
1480         return 0;
1481 }
1482
1483 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1484 {
1485         return !buffer_mapped(bh);
1486 }
1487
1488 /*
1489  * Note that we always start a transaction even if we're not journalling
1490  * data.  This is to preserve ordering: any hole instantiation within
1491  * __block_write_full_page -> ext3_get_block() should be journalled
1492  * along with the data so we don't crash and then get metadata which
1493  * refers to old data.
1494  *
1495  * In all journalling modes block_write_full_page() will start the I/O.
1496  *
1497  * Problem:
1498  *
1499  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1500  *              ext3_writepage()
1501  *
1502  * Similar for:
1503  *
1504  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1505  *
1506  * Same applies to ext3_get_block().  We will deadlock on various things like
1507  * lock_journal and i_truncate_mutex.
1508  *
1509  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1510  * allocations fail.
1511  *
1512  * 16May01: If we're reentered then journal_current_handle() will be
1513  *          non-zero. We simply *return*.
1514  *
1515  * 1 July 2001: @@@ FIXME:
1516  *   In journalled data mode, a data buffer may be metadata against the
1517  *   current transaction.  But the same file is part of a shared mapping
1518  *   and someone does a writepage() on it.
1519  *
1520  *   We will move the buffer onto the async_data list, but *after* it has
1521  *   been dirtied. So there's a small window where we have dirty data on
1522  *   BJ_Metadata.
1523  *
1524  *   Note that this only applies to the last partial page in the file.  The
1525  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1526  *   broken code anyway: it's wrong for msync()).
1527  *
1528  *   It's a rare case: affects the final partial page, for journalled data
1529  *   where the file is subject to bith write() and writepage() in the same
1530  *   transction.  To fix it we'll need a custom block_write_full_page().
1531  *   We'll probably need that anyway for journalling writepage() output.
1532  *
1533  * We don't honour synchronous mounts for writepage().  That would be
1534  * disastrous.  Any write() or metadata operation will sync the fs for
1535  * us.
1536  *
1537  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1538  * we don't need to open a transaction here.
1539  */
1540 static int ext3_ordered_writepage(struct page *page,
1541                                 struct writeback_control *wbc)
1542 {
1543         struct inode *inode = page->mapping->host;
1544         struct buffer_head *page_bufs;
1545         handle_t *handle = NULL;
1546         int ret = 0;
1547         int err;
1548
1549         J_ASSERT(PageLocked(page));
1550         WARN_ON_ONCE(IS_RDONLY(inode));
1551
1552         /*
1553          * We give up here if we're reentered, because it might be for a
1554          * different filesystem.
1555          */
1556         if (ext3_journal_current_handle())
1557                 goto out_fail;
1558
1559         if (!page_has_buffers(page)) {
1560                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1561                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1562                 page_bufs = page_buffers(page);
1563         } else {
1564                 page_bufs = page_buffers(page);
1565                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1566                                        NULL, buffer_unmapped)) {
1567                         /* Provide NULL get_block() to catch bugs if buffers
1568                          * weren't really mapped */
1569                         return block_write_full_page(page, NULL, wbc);
1570                 }
1571         }
1572         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1573
1574         if (IS_ERR(handle)) {
1575                 ret = PTR_ERR(handle);
1576                 goto out_fail;
1577         }
1578
1579         walk_page_buffers(handle, page_bufs, 0,
1580                         PAGE_CACHE_SIZE, NULL, bget_one);
1581
1582         ret = block_write_full_page(page, ext3_get_block, wbc);
1583
1584         /*
1585          * The page can become unlocked at any point now, and
1586          * truncate can then come in and change things.  So we
1587          * can't touch *page from now on.  But *page_bufs is
1588          * safe due to elevated refcount.
1589          */
1590
1591         /*
1592          * And attach them to the current transaction.  But only if
1593          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1594          * and generally junk.
1595          */
1596         if (ret == 0) {
1597                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1598                                         NULL, journal_dirty_data_fn);
1599                 if (!ret)
1600                         ret = err;
1601         }
1602         walk_page_buffers(handle, page_bufs, 0,
1603                         PAGE_CACHE_SIZE, NULL, bput_one);
1604         err = ext3_journal_stop(handle);
1605         if (!ret)
1606                 ret = err;
1607         return ret;
1608
1609 out_fail:
1610         redirty_page_for_writepage(wbc, page);
1611         unlock_page(page);
1612         return ret;
1613 }
1614
1615 static int ext3_writeback_writepage(struct page *page,
1616                                 struct writeback_control *wbc)
1617 {
1618         struct inode *inode = page->mapping->host;
1619         handle_t *handle = NULL;
1620         int ret = 0;
1621         int err;
1622
1623         J_ASSERT(PageLocked(page));
1624         WARN_ON_ONCE(IS_RDONLY(inode));
1625
1626         if (ext3_journal_current_handle())
1627                 goto out_fail;
1628
1629         if (page_has_buffers(page)) {
1630                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1631                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1632                         /* Provide NULL get_block() to catch bugs if buffers
1633                          * weren't really mapped */
1634                         return block_write_full_page(page, NULL, wbc);
1635                 }
1636         }
1637
1638         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1639         if (IS_ERR(handle)) {
1640                 ret = PTR_ERR(handle);
1641                 goto out_fail;
1642         }
1643
1644         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1645                 ret = nobh_writepage(page, ext3_get_block, wbc);
1646         else
1647                 ret = block_write_full_page(page, ext3_get_block, wbc);
1648
1649         err = ext3_journal_stop(handle);
1650         if (!ret)
1651                 ret = err;
1652         return ret;
1653
1654 out_fail:
1655         redirty_page_for_writepage(wbc, page);
1656         unlock_page(page);
1657         return ret;
1658 }
1659
1660 static int ext3_journalled_writepage(struct page *page,
1661                                 struct writeback_control *wbc)
1662 {
1663         struct inode *inode = page->mapping->host;
1664         handle_t *handle = NULL;
1665         int ret = 0;
1666         int err;
1667
1668         J_ASSERT(PageLocked(page));
1669         WARN_ON_ONCE(IS_RDONLY(inode));
1670
1671         if (ext3_journal_current_handle())
1672                 goto no_write;
1673
1674         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1675         if (IS_ERR(handle)) {
1676                 ret = PTR_ERR(handle);
1677                 goto no_write;
1678         }
1679
1680         if (!page_has_buffers(page) || PageChecked(page)) {
1681                 /*
1682                  * It's mmapped pagecache.  Add buffers and journal it.  There
1683                  * doesn't seem much point in redirtying the page here.
1684                  */
1685                 ClearPageChecked(page);
1686                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1687                                         ext3_get_block);
1688                 if (ret != 0) {
1689                         ext3_journal_stop(handle);
1690                         goto out_unlock;
1691                 }
1692                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1693                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1694
1695                 err = walk_page_buffers(handle, page_buffers(page), 0,
1696                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1697                 if (ret == 0)
1698                         ret = err;
1699                 ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1700                 unlock_page(page);
1701         } else {
1702                 /*
1703                  * It may be a page full of checkpoint-mode buffers.  We don't
1704                  * really know unless we go poke around in the buffer_heads.
1705                  * But block_write_full_page will do the right thing.
1706                  */
1707                 ret = block_write_full_page(page, ext3_get_block, wbc);
1708         }
1709         err = ext3_journal_stop(handle);
1710         if (!ret)
1711                 ret = err;
1712 out:
1713         return ret;
1714
1715 no_write:
1716         redirty_page_for_writepage(wbc, page);
1717 out_unlock:
1718         unlock_page(page);
1719         goto out;
1720 }
1721
1722 static int ext3_readpage(struct file *file, struct page *page)
1723 {
1724         return mpage_readpage(page, ext3_get_block);
1725 }
1726
1727 static int
1728 ext3_readpages(struct file *file, struct address_space *mapping,
1729                 struct list_head *pages, unsigned nr_pages)
1730 {
1731         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1732 }
1733
1734 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1735 {
1736         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1737
1738         /*
1739          * If it's a full truncate we just forget about the pending dirtying
1740          */
1741         if (offset == 0)
1742                 ClearPageChecked(page);
1743
1744         journal_invalidatepage(journal, page, offset);
1745 }
1746
1747 static int ext3_releasepage(struct page *page, gfp_t wait)
1748 {
1749         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1750
1751         WARN_ON(PageChecked(page));
1752         if (!page_has_buffers(page))
1753                 return 0;
1754         return journal_try_to_free_buffers(journal, page, wait);
1755 }
1756
1757 /*
1758  * If the O_DIRECT write will extend the file then add this inode to the
1759  * orphan list.  So recovery will truncate it back to the original size
1760  * if the machine crashes during the write.
1761  *
1762  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1763  * crashes then stale disk data _may_ be exposed inside the file. But current
1764  * VFS code falls back into buffered path in that case so we are safe.
1765  */
1766 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1767                         const struct iovec *iov, loff_t offset,
1768                         unsigned long nr_segs)
1769 {
1770         struct file *file = iocb->ki_filp;
1771         struct inode *inode = file->f_mapping->host;
1772         struct ext3_inode_info *ei = EXT3_I(inode);
1773         handle_t *handle;
1774         ssize_t ret;
1775         int orphan = 0;
1776         size_t count = iov_length(iov, nr_segs);
1777         int retries = 0;
1778
1779         if (rw == WRITE) {
1780                 loff_t final_size = offset + count;
1781
1782                 if (final_size > inode->i_size) {
1783                         /* Credits for sb + inode write */
1784                         handle = ext3_journal_start(inode, 2);
1785                         if (IS_ERR(handle)) {
1786                                 ret = PTR_ERR(handle);
1787                                 goto out;
1788                         }
1789                         ret = ext3_orphan_add(handle, inode);
1790                         if (ret) {
1791                                 ext3_journal_stop(handle);
1792                                 goto out;
1793                         }
1794                         orphan = 1;
1795                         ei->i_disksize = inode->i_size;
1796                         ext3_journal_stop(handle);
1797                 }
1798         }
1799
1800 retry:
1801         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1802                                  offset, nr_segs,
1803                                  ext3_get_block, NULL);
1804         /*
1805          * In case of error extending write may have instantiated a few
1806          * blocks outside i_size. Trim these off again.
1807          */
1808         if (unlikely((rw & WRITE) && ret < 0)) {
1809                 loff_t isize = i_size_read(inode);
1810                 loff_t end = offset + iov_length(iov, nr_segs);
1811
1812                 if (end > isize)
1813                         vmtruncate(inode, isize);
1814         }
1815         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1816                 goto retry;
1817
1818         if (orphan) {
1819                 int err;
1820
1821                 /* Credits for sb + inode write */
1822                 handle = ext3_journal_start(inode, 2);
1823                 if (IS_ERR(handle)) {
1824                         /* This is really bad luck. We've written the data
1825                          * but cannot extend i_size. Truncate allocated blocks
1826                          * and pretend the write failed... */
1827                         ext3_truncate(inode);
1828                         ret = PTR_ERR(handle);
1829                         goto out;
1830                 }
1831                 if (inode->i_nlink)
1832                         ext3_orphan_del(handle, inode);
1833                 if (ret > 0) {
1834                         loff_t end = offset + ret;
1835                         if (end > inode->i_size) {
1836                                 ei->i_disksize = end;
1837                                 i_size_write(inode, end);
1838                                 /*
1839                                  * We're going to return a positive `ret'
1840                                  * here due to non-zero-length I/O, so there's
1841                                  * no way of reporting error returns from
1842                                  * ext3_mark_inode_dirty() to userspace.  So
1843                                  * ignore it.
1844                                  */
1845                                 ext3_mark_inode_dirty(handle, inode);
1846                         }
1847                 }
1848                 err = ext3_journal_stop(handle);
1849                 if (ret == 0)
1850                         ret = err;
1851         }
1852 out:
1853         return ret;
1854 }
1855
1856 /*
1857  * Pages can be marked dirty completely asynchronously from ext3's journalling
1858  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1859  * much here because ->set_page_dirty is called under VFS locks.  The page is
1860  * not necessarily locked.
1861  *
1862  * We cannot just dirty the page and leave attached buffers clean, because the
1863  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1864  * or jbddirty because all the journalling code will explode.
1865  *
1866  * So what we do is to mark the page "pending dirty" and next time writepage
1867  * is called, propagate that into the buffers appropriately.
1868  */
1869 static int ext3_journalled_set_page_dirty(struct page *page)
1870 {
1871         SetPageChecked(page);
1872         return __set_page_dirty_nobuffers(page);
1873 }
1874
1875 static const struct address_space_operations ext3_ordered_aops = {
1876         .readpage               = ext3_readpage,
1877         .readpages              = ext3_readpages,
1878         .writepage              = ext3_ordered_writepage,
1879         .sync_page              = block_sync_page,
1880         .write_begin            = ext3_write_begin,
1881         .write_end              = ext3_ordered_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_writeback_aops = {
1892         .readpage               = ext3_readpage,
1893         .readpages              = ext3_readpages,
1894         .writepage              = ext3_writeback_writepage,
1895         .sync_page              = block_sync_page,
1896         .write_begin            = ext3_write_begin,
1897         .write_end              = ext3_writeback_write_end,
1898         .bmap                   = ext3_bmap,
1899         .invalidatepage         = ext3_invalidatepage,
1900         .releasepage            = ext3_releasepage,
1901         .direct_IO              = ext3_direct_IO,
1902         .migratepage            = buffer_migrate_page,
1903         .is_partially_uptodate  = block_is_partially_uptodate,
1904         .error_remove_page      = generic_error_remove_page,
1905 };
1906
1907 static const struct address_space_operations ext3_journalled_aops = {
1908         .readpage               = ext3_readpage,
1909         .readpages              = ext3_readpages,
1910         .writepage              = ext3_journalled_writepage,
1911         .sync_page              = block_sync_page,
1912         .write_begin            = ext3_write_begin,
1913         .write_end              = ext3_journalled_write_end,
1914         .set_page_dirty         = ext3_journalled_set_page_dirty,
1915         .bmap                   = ext3_bmap,
1916         .invalidatepage         = ext3_invalidatepage,
1917         .releasepage            = ext3_releasepage,
1918         .is_partially_uptodate  = block_is_partially_uptodate,
1919         .error_remove_page      = generic_error_remove_page,
1920 };
1921
1922 void ext3_set_aops(struct inode *inode)
1923 {
1924         if (ext3_should_order_data(inode))
1925                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1926         else if (ext3_should_writeback_data(inode))
1927                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1928         else
1929                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1930 }
1931
1932 /*
1933  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1934  * up to the end of the block which corresponds to `from'.
1935  * This required during truncate. We need to physically zero the tail end
1936  * of that block so it doesn't yield old data if the file is later grown.
1937  */
1938 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1939                 struct address_space *mapping, loff_t from)
1940 {
1941         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1942         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1943         unsigned blocksize, iblock, length, pos;
1944         struct inode *inode = mapping->host;
1945         struct buffer_head *bh;
1946         int err = 0;
1947
1948         blocksize = inode->i_sb->s_blocksize;
1949         length = blocksize - (offset & (blocksize - 1));
1950         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1951
1952         /*
1953          * For "nobh" option,  we can only work if we don't need to
1954          * read-in the page - otherwise we create buffers to do the IO.
1955          */
1956         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1957              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1958                 zero_user(page, offset, length);
1959                 set_page_dirty(page);
1960                 goto unlock;
1961         }
1962
1963         if (!page_has_buffers(page))
1964                 create_empty_buffers(page, blocksize, 0);
1965
1966         /* Find the buffer that contains "offset" */
1967         bh = page_buffers(page);
1968         pos = blocksize;
1969         while (offset >= pos) {
1970                 bh = bh->b_this_page;
1971                 iblock++;
1972                 pos += blocksize;
1973         }
1974
1975         err = 0;
1976         if (buffer_freed(bh)) {
1977                 BUFFER_TRACE(bh, "freed: skip");
1978                 goto unlock;
1979         }
1980
1981         if (!buffer_mapped(bh)) {
1982                 BUFFER_TRACE(bh, "unmapped");
1983                 ext3_get_block(inode, iblock, bh, 0);
1984                 /* unmapped? It's a hole - nothing to do */
1985                 if (!buffer_mapped(bh)) {
1986                         BUFFER_TRACE(bh, "still unmapped");
1987                         goto unlock;
1988                 }
1989         }
1990
1991         /* Ok, it's mapped. Make sure it's up-to-date */
1992         if (PageUptodate(page))
1993                 set_buffer_uptodate(bh);
1994
1995         if (!buffer_uptodate(bh)) {
1996                 err = -EIO;
1997                 ll_rw_block(READ, 1, &bh);
1998                 wait_on_buffer(bh);
1999                 /* Uhhuh. Read error. Complain and punt. */
2000                 if (!buffer_uptodate(bh))
2001                         goto unlock;
2002         }
2003
2004         if (ext3_should_journal_data(inode)) {
2005                 BUFFER_TRACE(bh, "get write access");
2006                 err = ext3_journal_get_write_access(handle, bh);
2007                 if (err)
2008                         goto unlock;
2009         }
2010
2011         zero_user(page, offset, length);
2012         BUFFER_TRACE(bh, "zeroed end of block");
2013
2014         err = 0;
2015         if (ext3_should_journal_data(inode)) {
2016                 err = ext3_journal_dirty_metadata(handle, bh);
2017         } else {
2018                 if (ext3_should_order_data(inode))
2019                         err = ext3_journal_dirty_data(handle, bh);
2020                 mark_buffer_dirty(bh);
2021         }
2022
2023 unlock:
2024         unlock_page(page);
2025         page_cache_release(page);
2026         return err;
2027 }
2028
2029 /*
2030  * Probably it should be a library function... search for first non-zero word
2031  * or memcmp with zero_page, whatever is better for particular architecture.
2032  * Linus?
2033  */
2034 static inline int all_zeroes(__le32 *p, __le32 *q)
2035 {
2036         while (p < q)
2037                 if (*p++)
2038                         return 0;
2039         return 1;
2040 }
2041
2042 /**
2043  *      ext3_find_shared - find the indirect blocks for partial truncation.
2044  *      @inode:   inode in question
2045  *      @depth:   depth of the affected branch
2046  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
2047  *      @chain:   place to store the pointers to partial indirect blocks
2048  *      @top:     place to the (detached) top of branch
2049  *
2050  *      This is a helper function used by ext3_truncate().
2051  *
2052  *      When we do truncate() we may have to clean the ends of several
2053  *      indirect blocks but leave the blocks themselves alive. Block is
2054  *      partially truncated if some data below the new i_size is refered
2055  *      from it (and it is on the path to the first completely truncated
2056  *      data block, indeed).  We have to free the top of that path along
2057  *      with everything to the right of the path. Since no allocation
2058  *      past the truncation point is possible until ext3_truncate()
2059  *      finishes, we may safely do the latter, but top of branch may
2060  *      require special attention - pageout below the truncation point
2061  *      might try to populate it.
2062  *
2063  *      We atomically detach the top of branch from the tree, store the
2064  *      block number of its root in *@top, pointers to buffer_heads of
2065  *      partially truncated blocks - in @chain[].bh and pointers to
2066  *      their last elements that should not be removed - in
2067  *      @chain[].p. Return value is the pointer to last filled element
2068  *      of @chain.
2069  *
2070  *      The work left to caller to do the actual freeing of subtrees:
2071  *              a) free the subtree starting from *@top
2072  *              b) free the subtrees whose roots are stored in
2073  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2074  *              c) free the subtrees growing from the inode past the @chain[0].
2075  *                      (no partially truncated stuff there).  */
2076
2077 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2078                         int offsets[4], Indirect chain[4], __le32 *top)
2079 {
2080         Indirect *partial, *p;
2081         int k, err;
2082
2083         *top = 0;
2084         /* Make k index the deepest non-null offset + 1 */
2085         for (k = depth; k > 1 && !offsets[k-1]; k--)
2086                 ;
2087         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2088         /* Writer: pointers */
2089         if (!partial)
2090                 partial = chain + k-1;
2091         /*
2092          * If the branch acquired continuation since we've looked at it -
2093          * fine, it should all survive and (new) top doesn't belong to us.
2094          */
2095         if (!partial->key && *partial->p)
2096                 /* Writer: end */
2097                 goto no_top;
2098         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2099                 ;
2100         /*
2101          * OK, we've found the last block that must survive. The rest of our
2102          * branch should be detached before unlocking. However, if that rest
2103          * of branch is all ours and does not grow immediately from the inode
2104          * it's easier to cheat and just decrement partial->p.
2105          */
2106         if (p == chain + k - 1 && p > chain) {
2107                 p->p--;
2108         } else {
2109                 *top = *p->p;
2110                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2111 #if 0
2112                 *p->p = 0;
2113 #endif
2114         }
2115         /* Writer: end */
2116
2117         while(partial > p) {
2118                 brelse(partial->bh);
2119                 partial--;
2120         }
2121 no_top:
2122         return partial;
2123 }
2124
2125 /*
2126  * Zero a number of block pointers in either an inode or an indirect block.
2127  * If we restart the transaction we must again get write access to the
2128  * indirect block for further modification.
2129  *
2130  * We release `count' blocks on disk, but (last - first) may be greater
2131  * than `count' because there can be holes in there.
2132  */
2133 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2134                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2135                 unsigned long count, __le32 *first, __le32 *last)
2136 {
2137         __le32 *p;
2138         if (try_to_extend_transaction(handle, inode)) {
2139                 if (bh) {
2140                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2141                         ext3_journal_dirty_metadata(handle, bh);
2142                 }
2143                 ext3_mark_inode_dirty(handle, inode);
2144                 truncate_restart_transaction(handle, inode);
2145                 if (bh) {
2146                         BUFFER_TRACE(bh, "retaking write access");
2147                         ext3_journal_get_write_access(handle, bh);
2148                 }
2149         }
2150
2151         /*
2152          * Any buffers which are on the journal will be in memory. We find
2153          * them on the hash table so journal_revoke() will run journal_forget()
2154          * on them.  We've already detached each block from the file, so
2155          * bforget() in journal_forget() should be safe.
2156          *
2157          * AKPM: turn on bforget in journal_forget()!!!
2158          */
2159         for (p = first; p < last; p++) {
2160                 u32 nr = le32_to_cpu(*p);
2161                 if (nr) {
2162                         struct buffer_head *bh;
2163
2164                         *p = 0;
2165                         bh = sb_find_get_block(inode->i_sb, nr);
2166                         ext3_forget(handle, 0, inode, bh, nr);
2167                 }
2168         }
2169
2170         ext3_free_blocks(handle, inode, block_to_free, count);
2171 }
2172
2173 /**
2174  * ext3_free_data - free a list of data blocks
2175  * @handle:     handle for this transaction
2176  * @inode:      inode we are dealing with
2177  * @this_bh:    indirect buffer_head which contains *@first and *@last
2178  * @first:      array of block numbers
2179  * @last:       points immediately past the end of array
2180  *
2181  * We are freeing all blocks refered from that array (numbers are stored as
2182  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2183  *
2184  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2185  * blocks are contiguous then releasing them at one time will only affect one
2186  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2187  * actually use a lot of journal space.
2188  *
2189  * @this_bh will be %NULL if @first and @last point into the inode's direct
2190  * block pointers.
2191  */
2192 static void ext3_free_data(handle_t *handle, struct inode *inode,
2193                            struct buffer_head *this_bh,
2194                            __le32 *first, __le32 *last)
2195 {
2196         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2197         unsigned long count = 0;            /* Number of blocks in the run */
2198         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2199                                                corresponding to
2200                                                block_to_free */
2201         ext3_fsblk_t nr;                    /* Current block # */
2202         __le32 *p;                          /* Pointer into inode/ind
2203                                                for current block */
2204         int err;
2205
2206         if (this_bh) {                          /* For indirect block */
2207                 BUFFER_TRACE(this_bh, "get_write_access");
2208                 err = ext3_journal_get_write_access(handle, this_bh);
2209                 /* Important: if we can't update the indirect pointers
2210                  * to the blocks, we can't free them. */
2211                 if (err)
2212                         return;
2213         }
2214
2215         for (p = first; p < last; p++) {
2216                 nr = le32_to_cpu(*p);
2217                 if (nr) {
2218                         /* accumulate blocks to free if they're contiguous */
2219                         if (count == 0) {
2220                                 block_to_free = nr;
2221                                 block_to_free_p = p;
2222                                 count = 1;
2223                         } else if (nr == block_to_free + count) {
2224                                 count++;
2225                         } else {
2226                                 ext3_clear_blocks(handle, inode, this_bh,
2227                                                   block_to_free,
2228                                                   count, block_to_free_p, p);
2229                                 block_to_free = nr;
2230                                 block_to_free_p = p;
2231                                 count = 1;
2232                         }
2233                 }
2234         }
2235
2236         if (count > 0)
2237                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2238                                   count, block_to_free_p, p);
2239
2240         if (this_bh) {
2241                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2242
2243                 /*
2244                  * The buffer head should have an attached journal head at this
2245                  * point. However, if the data is corrupted and an indirect
2246                  * block pointed to itself, it would have been detached when
2247                  * the block was cleared. Check for this instead of OOPSing.
2248                  */
2249                 if (bh2jh(this_bh))
2250                         ext3_journal_dirty_metadata(handle, this_bh);
2251                 else
2252                         ext3_error(inode->i_sb, "ext3_free_data",
2253                                    "circular indirect block detected, "
2254                                    "inode=%lu, block=%llu",
2255                                    inode->i_ino,
2256                                    (unsigned long long)this_bh->b_blocknr);
2257         }
2258 }
2259
2260 /**
2261  *      ext3_free_branches - free an array of branches
2262  *      @handle: JBD handle for this transaction
2263  *      @inode: inode we are dealing with
2264  *      @parent_bh: the buffer_head which contains *@first and *@last
2265  *      @first: array of block numbers
2266  *      @last:  pointer immediately past the end of array
2267  *      @depth: depth of the branches to free
2268  *
2269  *      We are freeing all blocks refered from these branches (numbers are
2270  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2271  *      appropriately.
2272  */
2273 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2274                                struct buffer_head *parent_bh,
2275                                __le32 *first, __le32 *last, int depth)
2276 {
2277         ext3_fsblk_t nr;
2278         __le32 *p;
2279
2280         if (is_handle_aborted(handle))
2281                 return;
2282
2283         if (depth--) {
2284                 struct buffer_head *bh;
2285                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2286                 p = last;
2287                 while (--p >= first) {
2288                         nr = le32_to_cpu(*p);
2289                         if (!nr)
2290                                 continue;               /* A hole */
2291
2292                         /* Go read the buffer for the next level down */
2293                         bh = sb_bread(inode->i_sb, nr);
2294
2295                         /*
2296                          * A read failure? Report error and clear slot
2297                          * (should be rare).
2298                          */
2299                         if (!bh) {
2300                                 ext3_error(inode->i_sb, "ext3_free_branches",
2301                                            "Read failure, inode=%lu, block="E3FSBLK,
2302                                            inode->i_ino, nr);
2303                                 continue;
2304                         }
2305
2306                         /* This zaps the entire block.  Bottom up. */
2307                         BUFFER_TRACE(bh, "free child branches");
2308                         ext3_free_branches(handle, inode, bh,
2309                                            (__le32*)bh->b_data,
2310                                            (__le32*)bh->b_data + addr_per_block,
2311                                            depth);
2312
2313                         /*
2314                          * We've probably journalled the indirect block several
2315                          * times during the truncate.  But it's no longer
2316                          * needed and we now drop it from the transaction via
2317                          * journal_revoke().
2318                          *
2319                          * That's easy if it's exclusively part of this
2320                          * transaction.  But if it's part of the committing
2321                          * transaction then journal_forget() will simply
2322                          * brelse() it.  That means that if the underlying
2323                          * block is reallocated in ext3_get_block(),
2324                          * unmap_underlying_metadata() will find this block
2325                          * and will try to get rid of it.  damn, damn.
2326                          *
2327                          * If this block has already been committed to the
2328                          * journal, a revoke record will be written.  And
2329                          * revoke records must be emitted *before* clearing
2330                          * this block's bit in the bitmaps.
2331                          */
2332                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2333
2334                         /*
2335                          * Everything below this this pointer has been
2336                          * released.  Now let this top-of-subtree go.
2337                          *
2338                          * We want the freeing of this indirect block to be
2339                          * atomic in the journal with the updating of the
2340                          * bitmap block which owns it.  So make some room in
2341                          * the journal.
2342                          *
2343                          * We zero the parent pointer *after* freeing its
2344                          * pointee in the bitmaps, so if extend_transaction()
2345                          * for some reason fails to put the bitmap changes and
2346                          * the release into the same transaction, recovery
2347                          * will merely complain about releasing a free block,
2348                          * rather than leaking blocks.
2349                          */
2350                         if (is_handle_aborted(handle))
2351                                 return;
2352                         if (try_to_extend_transaction(handle, inode)) {
2353                                 ext3_mark_inode_dirty(handle, inode);
2354                                 truncate_restart_transaction(handle, inode);
2355                         }
2356
2357                         ext3_free_blocks(handle, inode, nr, 1);
2358
2359                         if (parent_bh) {
2360                                 /*
2361                                  * The block which we have just freed is
2362                                  * pointed to by an indirect block: journal it
2363                                  */
2364                                 BUFFER_TRACE(parent_bh, "get_write_access");
2365                                 if (!ext3_journal_get_write_access(handle,
2366                                                                    parent_bh)){
2367                                         *p = 0;
2368                                         BUFFER_TRACE(parent_bh,
2369                                         "call ext3_journal_dirty_metadata");
2370                                         ext3_journal_dirty_metadata(handle,
2371                                                                     parent_bh);
2372                                 }
2373                         }
2374                 }
2375         } else {
2376                 /* We have reached the bottom of the tree. */
2377                 BUFFER_TRACE(parent_bh, "free data blocks");
2378                 ext3_free_data(handle, inode, parent_bh, first, last);
2379         }
2380 }
2381
2382 int ext3_can_truncate(struct inode *inode)
2383 {
2384         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2385                 return 0;
2386         if (S_ISREG(inode->i_mode))
2387                 return 1;
2388         if (S_ISDIR(inode->i_mode))
2389                 return 1;
2390         if (S_ISLNK(inode->i_mode))
2391                 return !ext3_inode_is_fast_symlink(inode);
2392         return 0;
2393 }
2394
2395 /*
2396  * ext3_truncate()
2397  *
2398  * We block out ext3_get_block() block instantiations across the entire
2399  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2400  * simultaneously on behalf of the same inode.
2401  *
2402  * As we work through the truncate and commmit bits of it to the journal there
2403  * is one core, guiding principle: the file's tree must always be consistent on
2404  * disk.  We must be able to restart the truncate after a crash.
2405  *
2406  * The file's tree may be transiently inconsistent in memory (although it
2407  * probably isn't), but whenever we close off and commit a journal transaction,
2408  * the contents of (the filesystem + the journal) must be consistent and
2409  * restartable.  It's pretty simple, really: bottom up, right to left (although
2410  * left-to-right works OK too).
2411  *
2412  * Note that at recovery time, journal replay occurs *before* the restart of
2413  * truncate against the orphan inode list.
2414  *
2415  * The committed inode has the new, desired i_size (which is the same as
2416  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2417  * that this inode's truncate did not complete and it will again call
2418  * ext3_truncate() to have another go.  So there will be instantiated blocks
2419  * to the right of the truncation point in a crashed ext3 filesystem.  But
2420  * that's fine - as long as they are linked from the inode, the post-crash
2421  * ext3_truncate() run will find them and release them.
2422  */
2423 void ext3_truncate(struct inode *inode)
2424 {
2425         handle_t *handle;
2426         struct ext3_inode_info *ei = EXT3_I(inode);
2427         __le32 *i_data = ei->i_data;
2428         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2429         struct address_space *mapping = inode->i_mapping;
2430         int offsets[4];
2431         Indirect chain[4];
2432         Indirect *partial;
2433         __le32 nr = 0;
2434         int n;
2435         long last_block;
2436         unsigned blocksize = inode->i_sb->s_blocksize;
2437         struct page *page;
2438
2439         if (!ext3_can_truncate(inode))
2440                 goto out_notrans;
2441
2442         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2443                 ext3_set_inode_state(inode, EXT3_STATE_FLUSH_ON_CLOSE);
2444
2445         /*
2446          * We have to lock the EOF page here, because lock_page() nests
2447          * outside journal_start().
2448          */
2449         if ((inode->i_size & (blocksize - 1)) == 0) {
2450                 /* Block boundary? Nothing to do */
2451                 page = NULL;
2452         } else {
2453                 page = grab_cache_page(mapping,
2454                                 inode->i_size >> PAGE_CACHE_SHIFT);
2455                 if (!page)
2456                         goto out_notrans;
2457         }
2458
2459         handle = start_transaction(inode);
2460         if (IS_ERR(handle)) {
2461                 if (page) {
2462                         clear_highpage(page);
2463                         flush_dcache_page(page);
2464                         unlock_page(page);
2465                         page_cache_release(page);
2466                 }
2467                 goto out_notrans;
2468         }
2469
2470         last_block = (inode->i_size + blocksize-1)
2471                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2472
2473         if (page)
2474                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2475
2476         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2477         if (n == 0)
2478                 goto out_stop;  /* error */
2479
2480         /*
2481          * OK.  This truncate is going to happen.  We add the inode to the
2482          * orphan list, so that if this truncate spans multiple transactions,
2483          * and we crash, we will resume the truncate when the filesystem
2484          * recovers.  It also marks the inode dirty, to catch the new size.
2485          *
2486          * Implication: the file must always be in a sane, consistent
2487          * truncatable state while each transaction commits.
2488          */
2489         if (ext3_orphan_add(handle, inode))
2490                 goto out_stop;
2491
2492         /*
2493          * The orphan list entry will now protect us from any crash which
2494          * occurs before the truncate completes, so it is now safe to propagate
2495          * the new, shorter inode size (held for now in i_size) into the
2496          * on-disk inode. We do this via i_disksize, which is the value which
2497          * ext3 *really* writes onto the disk inode.
2498          */
2499         ei->i_disksize = inode->i_size;
2500
2501         /*
2502          * From here we block out all ext3_get_block() callers who want to
2503          * modify the block allocation tree.
2504          */
2505         mutex_lock(&ei->truncate_mutex);
2506
2507         if (n == 1) {           /* direct blocks */
2508                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2509                                i_data + EXT3_NDIR_BLOCKS);
2510                 goto do_indirects;
2511         }
2512
2513         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2514         /* Kill the top of shared branch (not detached) */
2515         if (nr) {
2516                 if (partial == chain) {
2517                         /* Shared branch grows from the inode */
2518                         ext3_free_branches(handle, inode, NULL,
2519                                            &nr, &nr+1, (chain+n-1) - partial);
2520                         *partial->p = 0;
2521                         /*
2522                          * We mark the inode dirty prior to restart,
2523                          * and prior to stop.  No need for it here.
2524                          */
2525                 } else {
2526                         /* Shared branch grows from an indirect block */
2527                         BUFFER_TRACE(partial->bh, "get_write_access");
2528                         ext3_free_branches(handle, inode, partial->bh,
2529                                         partial->p,
2530                                         partial->p+1, (chain+n-1) - partial);
2531                 }
2532         }
2533         /* Clear the ends of indirect blocks on the shared branch */
2534         while (partial > chain) {
2535                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2536                                    (__le32*)partial->bh->b_data+addr_per_block,
2537                                    (chain+n-1) - partial);
2538                 BUFFER_TRACE(partial->bh, "call brelse");
2539                 brelse (partial->bh);
2540                 partial--;
2541         }
2542 do_indirects:
2543         /* Kill the remaining (whole) subtrees */
2544         switch (offsets[0]) {
2545         default:
2546                 nr = i_data[EXT3_IND_BLOCK];
2547                 if (nr) {
2548                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2549                         i_data[EXT3_IND_BLOCK] = 0;
2550                 }
2551         case EXT3_IND_BLOCK:
2552                 nr = i_data[EXT3_DIND_BLOCK];
2553                 if (nr) {
2554                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2555                         i_data[EXT3_DIND_BLOCK] = 0;
2556                 }
2557         case EXT3_DIND_BLOCK:
2558                 nr = i_data[EXT3_TIND_BLOCK];
2559                 if (nr) {
2560                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2561                         i_data[EXT3_TIND_BLOCK] = 0;
2562                 }
2563         case EXT3_TIND_BLOCK:
2564                 ;
2565         }
2566
2567         ext3_discard_reservation(inode);
2568
2569         mutex_unlock(&ei->truncate_mutex);
2570         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2571         ext3_mark_inode_dirty(handle, inode);
2572
2573         /*
2574          * In a multi-transaction truncate, we only make the final transaction
2575          * synchronous
2576          */
2577         if (IS_SYNC(inode))
2578                 handle->h_sync = 1;
2579 out_stop:
2580         /*
2581          * If this was a simple ftruncate(), and the file will remain alive
2582          * then we need to clear up the orphan record which we created above.
2583          * However, if this was a real unlink then we were called by
2584          * ext3_evict_inode(), and we allow that function to clean up the
2585          * orphan info for us.
2586          */
2587         if (inode->i_nlink)
2588                 ext3_orphan_del(handle, inode);
2589
2590         ext3_journal_stop(handle);
2591         return;
2592 out_notrans:
2593         /*
2594          * Delete the inode from orphan list so that it doesn't stay there
2595          * forever and trigger assertion on umount.
2596          */
2597         if (inode->i_nlink)
2598                 ext3_orphan_del(NULL, inode);
2599 }
2600
2601 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2602                 unsigned long ino, struct ext3_iloc *iloc)
2603 {
2604         unsigned long block_group;
2605         unsigned long offset;
2606         ext3_fsblk_t block;
2607         struct ext3_group_desc *gdp;
2608
2609         if (!ext3_valid_inum(sb, ino)) {
2610                 /*
2611                  * This error is already checked for in namei.c unless we are
2612                  * looking at an NFS filehandle, in which case no error
2613                  * report is needed
2614                  */
2615                 return 0;
2616         }
2617
2618         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2619         gdp = ext3_get_group_desc(sb, block_group, NULL);
2620         if (!gdp)
2621                 return 0;
2622         /*
2623          * Figure out the offset within the block group inode table
2624          */
2625         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2626                 EXT3_INODE_SIZE(sb);
2627         block = le32_to_cpu(gdp->bg_inode_table) +
2628                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2629
2630         iloc->block_group = block_group;
2631         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2632         return block;
2633 }
2634
2635 /*
2636  * ext3_get_inode_loc returns with an extra refcount against the inode's
2637  * underlying buffer_head on success. If 'in_mem' is true, we have all
2638  * data in memory that is needed to recreate the on-disk version of this
2639  * inode.
2640  */
2641 static int __ext3_get_inode_loc(struct inode *inode,
2642                                 struct ext3_iloc *iloc, int in_mem)
2643 {
2644         ext3_fsblk_t block;
2645         struct buffer_head *bh;
2646
2647         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2648         if (!block)
2649                 return -EIO;
2650
2651         bh = sb_getblk(inode->i_sb, block);
2652         if (!bh) {
2653                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2654                                 "unable to read inode block - "
2655                                 "inode=%lu, block="E3FSBLK,
2656                                  inode->i_ino, block);
2657                 return -EIO;
2658         }
2659         if (!buffer_uptodate(bh)) {
2660                 lock_buffer(bh);
2661
2662                 /*
2663                  * If the buffer has the write error flag, we have failed
2664                  * to write out another inode in the same block.  In this
2665                  * case, we don't have to read the block because we may
2666                  * read the old inode data successfully.
2667                  */
2668                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2669                         set_buffer_uptodate(bh);
2670
2671                 if (buffer_uptodate(bh)) {
2672                         /* someone brought it uptodate while we waited */
2673                         unlock_buffer(bh);
2674                         goto has_buffer;
2675                 }
2676
2677                 /*
2678                  * If we have all information of the inode in memory and this
2679                  * is the only valid inode in the block, we need not read the
2680                  * block.
2681                  */
2682                 if (in_mem) {
2683                         struct buffer_head *bitmap_bh;
2684                         struct ext3_group_desc *desc;
2685                         int inodes_per_buffer;
2686                         int inode_offset, i;
2687                         int block_group;
2688                         int start;
2689
2690                         block_group = (inode->i_ino - 1) /
2691                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2692                         inodes_per_buffer = bh->b_size /
2693                                 EXT3_INODE_SIZE(inode->i_sb);
2694                         inode_offset = ((inode->i_ino - 1) %
2695                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2696                         start = inode_offset & ~(inodes_per_buffer - 1);
2697
2698                         /* Is the inode bitmap in cache? */
2699                         desc = ext3_get_group_desc(inode->i_sb,
2700                                                 block_group, NULL);
2701                         if (!desc)
2702                                 goto make_io;
2703
2704                         bitmap_bh = sb_getblk(inode->i_sb,
2705                                         le32_to_cpu(desc->bg_inode_bitmap));
2706                         if (!bitmap_bh)
2707                                 goto make_io;
2708
2709                         /*
2710                          * If the inode bitmap isn't in cache then the
2711                          * optimisation may end up performing two reads instead
2712                          * of one, so skip it.
2713                          */
2714                         if (!buffer_uptodate(bitmap_bh)) {
2715                                 brelse(bitmap_bh);
2716                                 goto make_io;
2717                         }
2718                         for (i = start; i < start + inodes_per_buffer; i++) {
2719                                 if (i == inode_offset)
2720                                         continue;
2721                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2722                                         break;
2723                         }
2724                         brelse(bitmap_bh);
2725                         if (i == start + inodes_per_buffer) {
2726                                 /* all other inodes are free, so skip I/O */
2727                                 memset(bh->b_data, 0, bh->b_size);
2728                                 set_buffer_uptodate(bh);
2729                                 unlock_buffer(bh);
2730                                 goto has_buffer;
2731                         }
2732                 }
2733
2734 make_io:
2735                 /*
2736                  * There are other valid inodes in the buffer, this inode
2737                  * has in-inode xattrs, or we don't have this inode in memory.
2738                  * Read the block from disk.
2739                  */
2740                 get_bh(bh);
2741                 bh->b_end_io = end_buffer_read_sync;
2742                 submit_bh(READ_META, bh);
2743                 wait_on_buffer(bh);
2744                 if (!buffer_uptodate(bh)) {
2745                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2746                                         "unable to read inode block - "
2747                                         "inode=%lu, block="E3FSBLK,
2748                                         inode->i_ino, block);
2749                         brelse(bh);
2750                         return -EIO;
2751                 }
2752         }
2753 has_buffer:
2754         iloc->bh = bh;
2755         return 0;
2756 }
2757
2758 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2759 {
2760         /* We have all inode data except xattrs in memory here. */
2761         return __ext3_get_inode_loc(inode, iloc,
2762                 !ext3_test_inode_state(inode, EXT3_STATE_XATTR));
2763 }
2764
2765 void ext3_set_inode_flags(struct inode *inode)
2766 {
2767         unsigned int flags = EXT3_I(inode)->i_flags;
2768
2769         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2770         if (flags & EXT3_SYNC_FL)
2771                 inode->i_flags |= S_SYNC;
2772         if (flags & EXT3_APPEND_FL)
2773                 inode->i_flags |= S_APPEND;
2774         if (flags & EXT3_IMMUTABLE_FL)
2775                 inode->i_flags |= S_IMMUTABLE;
2776         if (flags & EXT3_NOATIME_FL)
2777                 inode->i_flags |= S_NOATIME;
2778         if (flags & EXT3_DIRSYNC_FL)
2779                 inode->i_flags |= S_DIRSYNC;
2780 }
2781
2782 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2783 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2784 {
2785         unsigned int flags = ei->vfs_inode.i_flags;
2786
2787         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2788                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2789         if (flags & S_SYNC)
2790                 ei->i_flags |= EXT3_SYNC_FL;
2791         if (flags & S_APPEND)
2792                 ei->i_flags |= EXT3_APPEND_FL;
2793         if (flags & S_IMMUTABLE)
2794                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2795         if (flags & S_NOATIME)
2796                 ei->i_flags |= EXT3_NOATIME_FL;
2797         if (flags & S_DIRSYNC)
2798                 ei->i_flags |= EXT3_DIRSYNC_FL;
2799 }
2800
2801 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2802 {
2803         struct ext3_iloc iloc;
2804         struct ext3_inode *raw_inode;
2805         struct ext3_inode_info *ei;
2806         struct buffer_head *bh;
2807         struct inode *inode;
2808         journal_t *journal = EXT3_SB(sb)->s_journal;
2809         transaction_t *transaction;
2810         long ret;
2811         int block;
2812
2813         inode = iget_locked(sb, ino);
2814         if (!inode)
2815                 return ERR_PTR(-ENOMEM);
2816         if (!(inode->i_state & I_NEW))
2817                 return inode;
2818
2819         ei = EXT3_I(inode);
2820         ei->i_block_alloc_info = NULL;
2821
2822         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2823         if (ret < 0)
2824                 goto bad_inode;
2825         bh = iloc.bh;
2826         raw_inode = ext3_raw_inode(&iloc);
2827         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2828         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2829         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2830         if(!(test_opt (inode->i_sb, NO_UID32))) {
2831                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2832                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2833         }
2834         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2835         inode->i_size = le32_to_cpu(raw_inode->i_size);
2836         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2837         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2838         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2839         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2840
2841         ei->i_state_flags = 0;
2842         ei->i_dir_start_lookup = 0;
2843         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2844         /* We now have enough fields to check if the inode was active or not.
2845          * This is needed because nfsd might try to access dead inodes
2846          * the test is that same one that e2fsck uses
2847          * NeilBrown 1999oct15
2848          */
2849         if (inode->i_nlink == 0) {
2850                 if (inode->i_mode == 0 ||
2851                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2852                         /* this inode is deleted */
2853                         brelse (bh);
2854                         ret = -ESTALE;
2855                         goto bad_inode;
2856                 }
2857                 /* The only unlinked inodes we let through here have
2858                  * valid i_mode and are being read by the orphan
2859                  * recovery code: that's fine, we're about to complete
2860                  * the process of deleting those. */
2861         }
2862         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2863         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2864 #ifdef EXT3_FRAGMENTS
2865         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2866         ei->i_frag_no = raw_inode->i_frag;
2867         ei->i_frag_size = raw_inode->i_fsize;
2868 #endif
2869         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2870         if (!S_ISREG(inode->i_mode)) {
2871                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2872         } else {
2873                 inode->i_size |=
2874                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2875         }
2876         ei->i_disksize = inode->i_size;
2877         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2878         ei->i_block_group = iloc.block_group;
2879         /*
2880          * NOTE! The in-memory inode i_data array is in little-endian order
2881          * even on big-endian machines: we do NOT byteswap the block numbers!
2882          */
2883         for (block = 0; block < EXT3_N_BLOCKS; block++)
2884                 ei->i_data[block] = raw_inode->i_block[block];
2885         INIT_LIST_HEAD(&ei->i_orphan);
2886
2887         /*
2888          * Set transaction id's of transactions that have to be committed
2889          * to finish f[data]sync. We set them to currently running transaction
2890          * as we cannot be sure that the inode or some of its metadata isn't
2891          * part of the transaction - the inode could have been reclaimed and
2892          * now it is reread from disk.
2893          */
2894         if (journal) {
2895                 tid_t tid;
2896
2897                 spin_lock(&journal->j_state_lock);
2898                 if (journal->j_running_transaction)
2899                         transaction = journal->j_running_transaction;
2900                 else
2901                         transaction = journal->j_committing_transaction;
2902                 if (transaction)
2903                         tid = transaction->t_tid;
2904                 else
2905                         tid = journal->j_commit_sequence;
2906                 spin_unlock(&journal->j_state_lock);
2907                 atomic_set(&ei->i_sync_tid, tid);
2908                 atomic_set(&ei->i_datasync_tid, tid);
2909         }
2910
2911         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2912             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2913                 /*
2914                  * When mke2fs creates big inodes it does not zero out
2915                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2916                  * so ignore those first few inodes.
2917                  */
2918                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2919                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2920                     EXT3_INODE_SIZE(inode->i_sb)) {
2921                         brelse (bh);
2922                         ret = -EIO;
2923                         goto bad_inode;
2924                 }
2925                 if (ei->i_extra_isize == 0) {
2926                         /* The extra space is currently unused. Use it. */
2927                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2928                                             EXT3_GOOD_OLD_INODE_SIZE;
2929                 } else {
2930                         __le32 *magic = (void *)raw_inode +
2931                                         EXT3_GOOD_OLD_INODE_SIZE +
2932                                         ei->i_extra_isize;
2933                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2934                                  ext3_set_inode_state(inode, EXT3_STATE_XATTR);
2935                 }
2936         } else
2937                 ei->i_extra_isize = 0;
2938
2939         if (S_ISREG(inode->i_mode)) {
2940                 inode->i_op = &ext3_file_inode_operations;
2941                 inode->i_fop = &ext3_file_operations;
2942                 ext3_set_aops(inode);
2943         } else if (S_ISDIR(inode->i_mode)) {
2944                 inode->i_op = &ext3_dir_inode_operations;
2945                 inode->i_fop = &ext3_dir_operations;
2946         } else if (S_ISLNK(inode->i_mode)) {
2947                 if (ext3_inode_is_fast_symlink(inode)) {
2948                         inode->i_op = &ext3_fast_symlink_inode_operations;
2949                         nd_terminate_link(ei->i_data, inode->i_size,
2950                                 sizeof(ei->i_data) - 1);
2951                 } else {
2952                         inode->i_op = &ext3_symlink_inode_operations;
2953                         ext3_set_aops(inode);
2954                 }
2955         } else {
2956                 inode->i_op = &ext3_special_inode_operations;
2957                 if (raw_inode->i_block[0])
2958                         init_special_inode(inode, inode->i_mode,
2959                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2960                 else
2961                         init_special_inode(inode, inode->i_mode,
2962                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2963         }
2964         brelse (iloc.bh);
2965         ext3_set_inode_flags(inode);
2966         unlock_new_inode(inode);
2967         return inode;
2968
2969 bad_inode:
2970         iget_failed(inode);
2971         return ERR_PTR(ret);
2972 }
2973
2974 /*
2975  * Post the struct inode info into an on-disk inode location in the
2976  * buffer-cache.  This gobbles the caller's reference to the
2977  * buffer_head in the inode location struct.
2978  *
2979  * The caller must have write access to iloc->bh.
2980  */
2981 static int ext3_do_update_inode(handle_t *handle,
2982                                 struct inode *inode,
2983                                 struct ext3_iloc *iloc)
2984 {
2985         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2986         struct ext3_inode_info *ei = EXT3_I(inode);
2987         struct buffer_head *bh = iloc->bh;
2988         int err = 0, rc, block;
2989
2990 again:
2991         /* we can't allow multiple procs in here at once, its a bit racey */
2992         lock_buffer(bh);
2993
2994         /* For fields not not tracking in the in-memory inode,
2995          * initialise them to zero for new inodes. */
2996         if (ext3_test_inode_state(inode, EXT3_STATE_NEW))
2997                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2998
2999         ext3_get_inode_flags(ei);
3000         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3001         if(!(test_opt(inode->i_sb, NO_UID32))) {
3002                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3003                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3004 /*
3005  * Fix up interoperability with old kernels. Otherwise, old inodes get
3006  * re-used with the upper 16 bits of the uid/gid intact
3007  */
3008                 if(!ei->i_dtime) {
3009                         raw_inode->i_uid_high =
3010                                 cpu_to_le16(high_16_bits(inode->i_uid));
3011                         raw_inode->i_gid_high =
3012                                 cpu_to_le16(high_16_bits(inode->i_gid));
3013                 } else {
3014                         raw_inode->i_uid_high = 0;
3015                         raw_inode->i_gid_high = 0;
3016                 }
3017         } else {
3018                 raw_inode->i_uid_low =
3019                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3020                 raw_inode->i_gid_low =
3021                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3022                 raw_inode->i_uid_high = 0;
3023                 raw_inode->i_gid_high = 0;
3024         }
3025         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3026         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
3027         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
3028         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
3029         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
3030         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
3031         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3032         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
3033 #ifdef EXT3_FRAGMENTS
3034         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
3035         raw_inode->i_frag = ei->i_frag_no;
3036         raw_inode->i_fsize = ei->i_frag_size;
3037 #endif
3038         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
3039         if (!S_ISREG(inode->i_mode)) {
3040                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
3041         } else {
3042                 raw_inode->i_size_high =
3043                         cpu_to_le32(ei->i_disksize >> 32);
3044                 if (ei->i_disksize > 0x7fffffffULL) {
3045                         struct super_block *sb = inode->i_sb;
3046                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
3047                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
3048                             EXT3_SB(sb)->s_es->s_rev_level ==
3049                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
3050                                /* If this is the first large file
3051                                 * created, add a flag to the superblock.
3052                                 */
3053                                 unlock_buffer(bh);
3054                                 err = ext3_journal_get_write_access(handle,
3055                                                 EXT3_SB(sb)->s_sbh);
3056                                 if (err)
3057                                         goto out_brelse;
3058
3059                                 ext3_update_dynamic_rev(sb);
3060                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
3061                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
3062                                 handle->h_sync = 1;
3063                                 err = ext3_journal_dirty_metadata(handle,
3064                                                 EXT3_SB(sb)->s_sbh);
3065                                 /* get our lock and start over */
3066                                 goto again;
3067                         }
3068                 }
3069         }
3070         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3071         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3072                 if (old_valid_dev(inode->i_rdev)) {
3073                         raw_inode->i_block[0] =
3074                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3075                         raw_inode->i_block[1] = 0;
3076                 } else {
3077                         raw_inode->i_block[0] = 0;
3078                         raw_inode->i_block[1] =
3079                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3080                         raw_inode->i_block[2] = 0;
3081                 }
3082         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
3083                 raw_inode->i_block[block] = ei->i_data[block];
3084
3085         if (ei->i_extra_isize)
3086                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3087
3088         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
3089         unlock_buffer(bh);
3090         rc = ext3_journal_dirty_metadata(handle, bh);
3091         if (!err)
3092                 err = rc;
3093         ext3_clear_inode_state(inode, EXT3_STATE_NEW);
3094
3095         atomic_set(&ei->i_sync_tid, handle->h_transaction->t_tid);
3096 out_brelse:
3097         brelse (bh);
3098         ext3_std_error(inode->i_sb, err);
3099         return err;
3100 }
3101
3102 /*
3103  * ext3_write_inode()
3104  *
3105  * We are called from a few places:
3106  *
3107  * - Within generic_file_write() for O_SYNC files.
3108  *   Here, there will be no transaction running. We wait for any running
3109  *   trasnaction to commit.
3110  *
3111  * - Within sys_sync(), kupdate and such.
3112  *   We wait on commit, if tol to.
3113  *
3114  * - Within prune_icache() (PF_MEMALLOC == true)
3115  *   Here we simply return.  We can't afford to block kswapd on the
3116  *   journal commit.
3117  *
3118  * In all cases it is actually safe for us to return without doing anything,
3119  * because the inode has been copied into a raw inode buffer in
3120  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3121  * knfsd.
3122  *
3123  * Note that we are absolutely dependent upon all inode dirtiers doing the
3124  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3125  * which we are interested.
3126  *
3127  * It would be a bug for them to not do this.  The code:
3128  *
3129  *      mark_inode_dirty(inode)
3130  *      stuff();
3131  *      inode->i_size = expr;
3132  *
3133  * is in error because a kswapd-driven write_inode() could occur while
3134  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3135  * will no longer be on the superblock's dirty inode list.
3136  */
3137 int ext3_write_inode(struct inode *inode, struct writeback_control *wbc)
3138 {
3139         if (current->flags & PF_MEMALLOC)
3140                 return 0;
3141
3142         if (ext3_journal_current_handle()) {
3143                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3144                 dump_stack();
3145                 return -EIO;
3146         }
3147
3148         if (wbc->sync_mode != WB_SYNC_ALL)
3149                 return 0;
3150
3151         return ext3_force_commit(inode->i_sb);
3152 }
3153
3154 /*
3155  * ext3_setattr()
3156  *
3157  * Called from notify_change.
3158  *
3159  * We want to trap VFS attempts to truncate the file as soon as
3160  * possible.  In particular, we want to make sure that when the VFS
3161  * shrinks i_size, we put the inode on the orphan list and modify
3162  * i_disksize immediately, so that during the subsequent flushing of
3163  * dirty pages and freeing of disk blocks, we can guarantee that any
3164  * commit will leave the blocks being flushed in an unused state on
3165  * disk.  (On recovery, the inode will get truncated and the blocks will
3166  * be freed, so we have a strong guarantee that no future commit will
3167  * leave these blocks visible to the user.)
3168  *
3169  * Called with inode->sem down.
3170  */
3171 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3172 {
3173         struct inode *inode = dentry->d_inode;
3174         int error, rc = 0;
3175         const unsigned int ia_valid = attr->ia_valid;
3176
3177         error = inode_change_ok(inode, attr);
3178         if (error)
3179                 return error;
3180
3181         if (is_quota_modification(inode, attr))
3182                 dquot_initialize(inode);
3183         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3184                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3185                 handle_t *handle;
3186
3187                 /* (user+group)*(old+new) structure, inode write (sb,
3188                  * inode block, ? - but truncate inode update has it) */
3189                 handle = ext3_journal_start(inode, EXT3_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3190                                         EXT3_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)+3);
3191                 if (IS_ERR(handle)) {
3192                         error = PTR_ERR(handle);
3193                         goto err_out;
3194                 }
3195                 error = dquot_transfer(inode, attr);
3196                 if (error) {
3197                         ext3_journal_stop(handle);
3198                         return error;
3199                 }
3200                 /* Update corresponding info in inode so that everything is in
3201                  * one transaction */
3202                 if (attr->ia_valid & ATTR_UID)
3203                         inode->i_uid = attr->ia_uid;
3204                 if (attr->ia_valid & ATTR_GID)
3205                         inode->i_gid = attr->ia_gid;
3206                 error = ext3_mark_inode_dirty(handle, inode);
3207                 ext3_journal_stop(handle);
3208         }
3209
3210         if (S_ISREG(inode->i_mode) &&
3211             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3212                 handle_t *handle;
3213
3214                 handle = ext3_journal_start(inode, 3);
3215                 if (IS_ERR(handle)) {
3216                         error = PTR_ERR(handle);
3217                         goto err_out;
3218                 }
3219
3220                 error = ext3_orphan_add(handle, inode);
3221                 EXT3_I(inode)->i_disksize = attr->ia_size;
3222                 rc = ext3_mark_inode_dirty(handle, inode);
3223                 if (!error)
3224                         error = rc;
3225                 ext3_journal_stop(handle);
3226         }
3227
3228         if ((attr->ia_valid & ATTR_SIZE) &&
3229             attr->ia_size != i_size_read(inode)) {
3230                 rc = vmtruncate(inode, attr->ia_size);
3231                 if (rc)
3232                         goto err_out;
3233         }
3234
3235         setattr_copy(inode, attr);
3236         mark_inode_dirty(inode);
3237
3238         if (ia_valid & ATTR_MODE)
3239                 rc = ext3_acl_chmod(inode);
3240
3241 err_out:
3242         ext3_std_error(inode->i_sb, error);
3243         if (!error)
3244                 error = rc;
3245         return error;
3246 }
3247
3248
3249 /*
3250  * How many blocks doth make a writepage()?
3251  *
3252  * With N blocks per page, it may be:
3253  * N data blocks
3254  * 2 indirect block
3255  * 2 dindirect
3256  * 1 tindirect
3257  * N+5 bitmap blocks (from the above)
3258  * N+5 group descriptor summary blocks
3259  * 1 inode block
3260  * 1 superblock.
3261  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3262  *
3263  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3264  *
3265  * With ordered or writeback data it's the same, less the N data blocks.
3266  *
3267  * If the inode's direct blocks can hold an integral number of pages then a
3268  * page cannot straddle two indirect blocks, and we can only touch one indirect
3269  * and dindirect block, and the "5" above becomes "3".
3270  *
3271  * This still overestimates under most circumstances.  If we were to pass the
3272  * start and end offsets in here as well we could do block_to_path() on each
3273  * block and work out the exact number of indirects which are touched.  Pah.
3274  */
3275
3276 static int ext3_writepage_trans_blocks(struct inode *inode)
3277 {
3278         int bpp = ext3_journal_blocks_per_page(inode);
3279         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3280         int ret;
3281
3282         if (ext3_should_journal_data(inode))
3283                 ret = 3 * (bpp + indirects) + 2;
3284         else
3285                 ret = 2 * (bpp + indirects) + 2;
3286
3287 #ifdef CONFIG_QUOTA
3288         /* We know that structure was already allocated during dquot_initialize so
3289          * we will be updating only the data blocks + inodes */
3290         ret += EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
3291 #endif
3292
3293         return ret;
3294 }
3295
3296 /*
3297  * The caller must have previously called ext3_reserve_inode_write().
3298  * Give this, we know that the caller already has write access to iloc->bh.
3299  */
3300 int ext3_mark_iloc_dirty(handle_t *handle,
3301                 struct inode *inode, struct ext3_iloc *iloc)
3302 {
3303         int err = 0;
3304
3305         /* the do_update_inode consumes one bh->b_count */
3306         get_bh(iloc->bh);
3307
3308         /* ext3_do_update_inode() does journal_dirty_metadata */
3309         err = ext3_do_update_inode(handle, inode, iloc);
3310         put_bh(iloc->bh);
3311         return err;
3312 }
3313
3314 /*
3315  * On success, We end up with an outstanding reference count against
3316  * iloc->bh.  This _must_ be cleaned up later.
3317  */
3318
3319 int
3320 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3321                          struct ext3_iloc *iloc)
3322 {
3323         int err = 0;
3324         if (handle) {
3325                 err = ext3_get_inode_loc(inode, iloc);
3326                 if (!err) {
3327                         BUFFER_TRACE(iloc->bh, "get_write_access");
3328                         err = ext3_journal_get_write_access(handle, iloc->bh);
3329                         if (err) {
3330                                 brelse(iloc->bh);
3331                                 iloc->bh = NULL;
3332                         }
3333                 }
3334         }
3335         ext3_std_error(inode->i_sb, err);
3336         return err;
3337 }
3338
3339 /*
3340  * What we do here is to mark the in-core inode as clean with respect to inode
3341  * dirtiness (it may still be data-dirty).
3342  * This means that the in-core inode may be reaped by prune_icache
3343  * without having to perform any I/O.  This is a very good thing,
3344  * because *any* task may call prune_icache - even ones which
3345  * have a transaction open against a different journal.
3346  *
3347  * Is this cheating?  Not really.  Sure, we haven't written the
3348  * inode out, but prune_icache isn't a user-visible syncing function.
3349  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3350  * we start and wait on commits.
3351  *
3352  * Is this efficient/effective?  Well, we're being nice to the system
3353  * by cleaning up our inodes proactively so they can be reaped
3354  * without I/O.  But we are potentially leaving up to five seconds'
3355  * worth of inodes floating about which prune_icache wants us to
3356  * write out.  One way to fix that would be to get prune_icache()
3357  * to do a write_super() to free up some memory.  It has the desired
3358  * effect.
3359  */
3360 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3361 {
3362         struct ext3_iloc iloc;
3363         int err;
3364
3365         might_sleep();
3366         err = ext3_reserve_inode_write(handle, inode, &iloc);
3367         if (!err)
3368                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3369         return err;
3370 }
3371
3372 /*
3373  * ext3_dirty_inode() is called from __mark_inode_dirty()
3374  *
3375  * We're really interested in the case where a file is being extended.
3376  * i_size has been changed by generic_commit_write() and we thus need
3377  * to include the updated inode in the current transaction.
3378  *
3379  * Also, dquot_alloc_space() will always dirty the inode when blocks
3380  * are allocated to the file.
3381  *
3382  * If the inode is marked synchronous, we don't honour that here - doing
3383  * so would cause a commit on atime updates, which we don't bother doing.
3384  * We handle synchronous inodes at the highest possible level.
3385  */
3386 void ext3_dirty_inode(struct inode *inode)
3387 {
3388         handle_t *current_handle = ext3_journal_current_handle();
3389         handle_t *handle;
3390
3391         handle = ext3_journal_start(inode, 2);
3392         if (IS_ERR(handle))
3393                 goto out;
3394         if (current_handle &&
3395                 current_handle->h_transaction != handle->h_transaction) {
3396                 /* This task has a transaction open against a different fs */
3397                 printk(KERN_EMERG "%s: transactions do not match!\n",
3398                        __func__);
3399         } else {
3400                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3401                                 current_handle);
3402                 ext3_mark_inode_dirty(handle, inode);
3403         }
3404         ext3_journal_stop(handle);
3405 out:
3406         return;
3407 }
3408
3409 #if 0
3410 /*
3411  * Bind an inode's backing buffer_head into this transaction, to prevent
3412  * it from being flushed to disk early.  Unlike
3413  * ext3_reserve_inode_write, this leaves behind no bh reference and
3414  * returns no iloc structure, so the caller needs to repeat the iloc
3415  * lookup to mark the inode dirty later.
3416  */
3417 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3418 {
3419         struct ext3_iloc iloc;
3420
3421         int err = 0;
3422         if (handle) {
3423                 err = ext3_get_inode_loc(inode, &iloc);
3424                 if (!err) {
3425                         BUFFER_TRACE(iloc.bh, "get_write_access");
3426                         err = journal_get_write_access(handle, iloc.bh);
3427                         if (!err)
3428                                 err = ext3_journal_dirty_metadata(handle,
3429                                                                   iloc.bh);
3430                         brelse(iloc.bh);
3431                 }
3432         }
3433         ext3_std_error(inode->i_sb, err);
3434         return err;
3435 }
3436 #endif
3437
3438 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3439 {
3440         journal_t *journal;
3441         handle_t *handle;
3442         int err;
3443
3444         /*
3445          * We have to be very careful here: changing a data block's
3446          * journaling status dynamically is dangerous.  If we write a
3447          * data block to the journal, change the status and then delete
3448          * that block, we risk forgetting to revoke the old log record
3449          * from the journal and so a subsequent replay can corrupt data.
3450          * So, first we make sure that the journal is empty and that
3451          * nobody is changing anything.
3452          */
3453
3454         journal = EXT3_JOURNAL(inode);
3455         if (is_journal_aborted(journal))
3456                 return -EROFS;
3457
3458         journal_lock_updates(journal);
3459         journal_flush(journal);
3460
3461         /*
3462          * OK, there are no updates running now, and all cached data is
3463          * synced to disk.  We are now in a completely consistent state
3464          * which doesn't have anything in the journal, and we know that
3465          * no filesystem updates are running, so it is safe to modify
3466          * the inode's in-core data-journaling state flag now.
3467          */
3468
3469         if (val)
3470                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3471         else
3472                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3473         ext3_set_aops(inode);
3474
3475         journal_unlock_updates(journal);
3476
3477         /* Finally we can mark the inode as dirty. */
3478
3479         handle = ext3_journal_start(inode, 1);
3480         if (IS_ERR(handle))
3481                 return PTR_ERR(handle);
3482
3483         err = ext3_mark_inode_dirty(handle, inode);
3484         handle->h_sync = 1;
3485         ext3_journal_stop(handle);
3486         ext3_std_error(inode->i_sb, err);
3487
3488         return err;
3489 }