ext3: Fix data corruption in inodes with journalled data
[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 <trace/events/ext3.h>
42 #include "xattr.h"
43 #include "acl.h"
44
45 static int ext3_writepage_trans_blocks(struct inode *inode);
46 static int ext3_block_truncate_page(struct inode *inode, loff_t from);
47
48 /*
49  * Test whether an inode is a fast symlink.
50  */
51 static int ext3_inode_is_fast_symlink(struct inode *inode)
52 {
53         int ea_blocks = EXT3_I(inode)->i_file_acl ?
54                 (inode->i_sb->s_blocksize >> 9) : 0;
55
56         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
57 }
58
59 /*
60  * The ext3 forget function must perform a revoke if we are freeing data
61  * which has been journaled.  Metadata (eg. indirect blocks) must be
62  * revoked in all cases.
63  *
64  * "bh" may be NULL: a metadata block may have been freed from memory
65  * but there may still be a record of it in the journal, and that record
66  * still needs to be revoked.
67  */
68 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
69                         struct buffer_head *bh, ext3_fsblk_t blocknr)
70 {
71         int err;
72
73         might_sleep();
74
75         trace_ext3_forget(inode, is_metadata, blocknr);
76         BUFFER_TRACE(bh, "enter");
77
78         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
79                   "data mode %lx\n",
80                   bh, is_metadata, inode->i_mode,
81                   test_opt(inode->i_sb, DATA_FLAGS));
82
83         /* Never use the revoke function if we are doing full data
84          * journaling: there is no need to, and a V1 superblock won't
85          * support it.  Otherwise, only skip the revoke on un-journaled
86          * data blocks. */
87
88         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
89             (!is_metadata && !ext3_should_journal_data(inode))) {
90                 if (bh) {
91                         BUFFER_TRACE(bh, "call journal_forget");
92                         return ext3_journal_forget(handle, bh);
93                 }
94                 return 0;
95         }
96
97         /*
98          * data!=journal && (is_metadata || should_journal_data(inode))
99          */
100         BUFFER_TRACE(bh, "call ext3_journal_revoke");
101         err = ext3_journal_revoke(handle, blocknr, bh);
102         if (err)
103                 ext3_abort(inode->i_sb, __func__,
104                            "error %d when attempting revoke", err);
105         BUFFER_TRACE(bh, "exit");
106         return err;
107 }
108
109 /*
110  * Work out how many blocks we need to proceed with the next chunk of a
111  * truncate transaction.
112  */
113 static unsigned long blocks_for_truncate(struct inode *inode)
114 {
115         unsigned long needed;
116
117         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
118
119         /* Give ourselves just enough room to cope with inodes in which
120          * i_blocks is corrupt: we've seen disk corruptions in the past
121          * which resulted in random data in an inode which looked enough
122          * like a regular file for ext3 to try to delete it.  Things
123          * will go a bit crazy if that happens, but at least we should
124          * try not to panic the whole kernel. */
125         if (needed < 2)
126                 needed = 2;
127
128         /* But we need to bound the transaction so we don't overflow the
129          * journal. */
130         if (needed > EXT3_MAX_TRANS_DATA)
131                 needed = EXT3_MAX_TRANS_DATA;
132
133         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
134 }
135
136 /*
137  * Truncate transactions can be complex and absolutely huge.  So we need to
138  * be able to restart the transaction at a conventient checkpoint to make
139  * sure we don't overflow the journal.
140  *
141  * start_transaction gets us a new handle for a truncate transaction,
142  * and extend_transaction tries to extend the existing one a bit.  If
143  * extend fails, we need to propagate the failure up and restart the
144  * transaction in the top-level truncate loop. --sct
145  */
146 static handle_t *start_transaction(struct inode *inode)
147 {
148         handle_t *result;
149
150         result = ext3_journal_start(inode, blocks_for_truncate(inode));
151         if (!IS_ERR(result))
152                 return result;
153
154         ext3_std_error(inode->i_sb, PTR_ERR(result));
155         return result;
156 }
157
158 /*
159  * Try to extend this transaction for the purposes of truncation.
160  *
161  * Returns 0 if we managed to create more room.  If we can't create more
162  * room, and the transaction must be restarted we return 1.
163  */
164 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
165 {
166         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
167                 return 0;
168         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
169                 return 0;
170         return 1;
171 }
172
173 /*
174  * Restart the transaction associated with *handle.  This does a commit,
175  * so before we call here everything must be consistently dirtied against
176  * this transaction.
177  */
178 static int truncate_restart_transaction(handle_t *handle, struct inode *inode)
179 {
180         int ret;
181
182         jbd_debug(2, "restarting handle %p\n", handle);
183         /*
184          * Drop truncate_mutex to avoid deadlock with ext3_get_blocks_handle
185          * At this moment, get_block can be called only for blocks inside
186          * i_size since page cache has been already dropped and writes are
187          * blocked by i_mutex. So we can safely drop the truncate_mutex.
188          */
189         mutex_unlock(&EXT3_I(inode)->truncate_mutex);
190         ret = ext3_journal_restart(handle, blocks_for_truncate(inode));
191         mutex_lock(&EXT3_I(inode)->truncate_mutex);
192         return ret;
193 }
194
195 /*
196  * Called at inode eviction from icache
197  */
198 void ext3_evict_inode (struct inode *inode)
199 {
200         struct ext3_inode_info *ei = EXT3_I(inode);
201         struct ext3_block_alloc_info *rsv;
202         handle_t *handle;
203         int want_delete = 0;
204
205         trace_ext3_evict_inode(inode);
206         if (!inode->i_nlink && !is_bad_inode(inode)) {
207                 dquot_initialize(inode);
208                 want_delete = 1;
209         }
210
211         /*
212          * When journalling data dirty buffers are tracked only in the journal.
213          * So although mm thinks everything is clean and ready for reaping the
214          * inode might still have some pages to write in the running
215          * transaction or waiting to be checkpointed. Thus calling
216          * journal_invalidatepage() (via truncate_inode_pages()) to discard
217          * these buffers can cause data loss. Also even if we did not discard
218          * these buffers, we would have no way to find them after the inode
219          * is reaped and thus user could see stale data if he tries to read
220          * them before the transaction is checkpointed. So be careful and
221          * force everything to disk here... We use ei->i_datasync_tid to
222          * store the newest transaction containing inode's data.
223          *
224          * Note that directories do not have this problem because they don't
225          * use page cache.
226          */
227         if (inode->i_nlink && ext3_should_journal_data(inode) &&
228             (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
229                 tid_t commit_tid = atomic_read(&ei->i_datasync_tid);
230                 journal_t *journal = EXT3_SB(inode->i_sb)->s_journal;
231
232                 log_start_commit(journal, commit_tid);
233                 log_wait_commit(journal, commit_tid);
234                 filemap_write_and_wait(&inode->i_data);
235         }
236         truncate_inode_pages(&inode->i_data, 0);
237
238         ext3_discard_reservation(inode);
239         rsv = ei->i_block_alloc_info;
240         ei->i_block_alloc_info = NULL;
241         if (unlikely(rsv))
242                 kfree(rsv);
243
244         if (!want_delete)
245                 goto no_delete;
246
247         handle = start_transaction(inode);
248         if (IS_ERR(handle)) {
249                 /*
250                  * If we're going to skip the normal cleanup, we still need to
251                  * make sure that the in-core orphan linked list is properly
252                  * cleaned up.
253                  */
254                 ext3_orphan_del(NULL, inode);
255                 goto no_delete;
256         }
257
258         if (IS_SYNC(inode))
259                 handle->h_sync = 1;
260         inode->i_size = 0;
261         if (inode->i_blocks)
262                 ext3_truncate(inode);
263         /*
264          * Kill off the orphan record created when the inode lost the last
265          * link.  Note that ext3_orphan_del() has to be able to cope with the
266          * deletion of a non-existent orphan - ext3_truncate() could
267          * have removed the record.
268          */
269         ext3_orphan_del(handle, inode);
270         ei->i_dtime = get_seconds();
271
272         /*
273          * One subtle ordering requirement: if anything has gone wrong
274          * (transaction abort, IO errors, whatever), then we can still
275          * do these next steps (the fs will already have been marked as
276          * having errors), but we can't free the inode if the mark_dirty
277          * fails.
278          */
279         if (ext3_mark_inode_dirty(handle, inode)) {
280                 /* If that failed, just dquot_drop() and be done with that */
281                 dquot_drop(inode);
282                 end_writeback(inode);
283         } else {
284                 ext3_xattr_delete_inode(handle, inode);
285                 dquot_free_inode(inode);
286                 dquot_drop(inode);
287                 end_writeback(inode);
288                 ext3_free_inode(handle, inode);
289         }
290         ext3_journal_stop(handle);
291         return;
292 no_delete:
293         end_writeback(inode);
294         dquot_drop(inode);
295 }
296
297 typedef struct {
298         __le32  *p;
299         __le32  key;
300         struct buffer_head *bh;
301 } Indirect;
302
303 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
304 {
305         p->key = *(p->p = v);
306         p->bh = bh;
307 }
308
309 static int verify_chain(Indirect *from, Indirect *to)
310 {
311         while (from <= to && from->key == *from->p)
312                 from++;
313         return (from > to);
314 }
315
316 /**
317  *      ext3_block_to_path - parse the block number into array of offsets
318  *      @inode: inode in question (we are only interested in its superblock)
319  *      @i_block: block number to be parsed
320  *      @offsets: array to store the offsets in
321  *      @boundary: set this non-zero if the referred-to block is likely to be
322  *             followed (on disk) by an indirect block.
323  *
324  *      To store the locations of file's data ext3 uses a data structure common
325  *      for UNIX filesystems - tree of pointers anchored in the inode, with
326  *      data blocks at leaves and indirect blocks in intermediate nodes.
327  *      This function translates the block number into path in that tree -
328  *      return value is the path length and @offsets[n] is the offset of
329  *      pointer to (n+1)th node in the nth one. If @block is out of range
330  *      (negative or too large) warning is printed and zero returned.
331  *
332  *      Note: function doesn't find node addresses, so no IO is needed. All
333  *      we need to know is the capacity of indirect blocks (taken from the
334  *      inode->i_sb).
335  */
336
337 /*
338  * Portability note: the last comparison (check that we fit into triple
339  * indirect block) is spelled differently, because otherwise on an
340  * architecture with 32-bit longs and 8Kb pages we might get into trouble
341  * if our filesystem had 8Kb blocks. We might use long long, but that would
342  * kill us on x86. Oh, well, at least the sign propagation does not matter -
343  * i_block would have to be negative in the very beginning, so we would not
344  * get there at all.
345  */
346
347 static int ext3_block_to_path(struct inode *inode,
348                         long i_block, int offsets[4], int *boundary)
349 {
350         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
351         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
352         const long direct_blocks = EXT3_NDIR_BLOCKS,
353                 indirect_blocks = ptrs,
354                 double_blocks = (1 << (ptrs_bits * 2));
355         int n = 0;
356         int final = 0;
357
358         if (i_block < 0) {
359                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
360         } else if (i_block < direct_blocks) {
361                 offsets[n++] = i_block;
362                 final = direct_blocks;
363         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
364                 offsets[n++] = EXT3_IND_BLOCK;
365                 offsets[n++] = i_block;
366                 final = ptrs;
367         } else if ((i_block -= indirect_blocks) < double_blocks) {
368                 offsets[n++] = EXT3_DIND_BLOCK;
369                 offsets[n++] = i_block >> ptrs_bits;
370                 offsets[n++] = i_block & (ptrs - 1);
371                 final = ptrs;
372         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
373                 offsets[n++] = EXT3_TIND_BLOCK;
374                 offsets[n++] = i_block >> (ptrs_bits * 2);
375                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
376                 offsets[n++] = i_block & (ptrs - 1);
377                 final = ptrs;
378         } else {
379                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
380         }
381         if (boundary)
382                 *boundary = final - 1 - (i_block & (ptrs - 1));
383         return n;
384 }
385
386 /**
387  *      ext3_get_branch - read the chain of indirect blocks leading to data
388  *      @inode: inode in question
389  *      @depth: depth of the chain (1 - direct pointer, etc.)
390  *      @offsets: offsets of pointers in inode/indirect blocks
391  *      @chain: place to store the result
392  *      @err: here we store the error value
393  *
394  *      Function fills the array of triples <key, p, bh> and returns %NULL
395  *      if everything went OK or the pointer to the last filled triple
396  *      (incomplete one) otherwise. Upon the return chain[i].key contains
397  *      the number of (i+1)-th block in the chain (as it is stored in memory,
398  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
399  *      number (it points into struct inode for i==0 and into the bh->b_data
400  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
401  *      block for i>0 and NULL for i==0. In other words, it holds the block
402  *      numbers of the chain, addresses they were taken from (and where we can
403  *      verify that chain did not change) and buffer_heads hosting these
404  *      numbers.
405  *
406  *      Function stops when it stumbles upon zero pointer (absent block)
407  *              (pointer to last triple returned, *@err == 0)
408  *      or when it gets an IO error reading an indirect block
409  *              (ditto, *@err == -EIO)
410  *      or when it notices that chain had been changed while it was reading
411  *              (ditto, *@err == -EAGAIN)
412  *      or when it reads all @depth-1 indirect blocks successfully and finds
413  *      the whole chain, all way to the data (returns %NULL, *err == 0).
414  */
415 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
416                                  Indirect chain[4], int *err)
417 {
418         struct super_block *sb = inode->i_sb;
419         Indirect *p = chain;
420         struct buffer_head *bh;
421
422         *err = 0;
423         /* i_data is not going away, no lock needed */
424         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
425         if (!p->key)
426                 goto no_block;
427         while (--depth) {
428                 bh = sb_bread(sb, le32_to_cpu(p->key));
429                 if (!bh)
430                         goto failure;
431                 /* Reader: pointers */
432                 if (!verify_chain(chain, p))
433                         goto changed;
434                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
435                 /* Reader: end */
436                 if (!p->key)
437                         goto no_block;
438         }
439         return NULL;
440
441 changed:
442         brelse(bh);
443         *err = -EAGAIN;
444         goto no_block;
445 failure:
446         *err = -EIO;
447 no_block:
448         return p;
449 }
450
451 /**
452  *      ext3_find_near - find a place for allocation with sufficient locality
453  *      @inode: owner
454  *      @ind: descriptor of indirect block.
455  *
456  *      This function returns the preferred place for block allocation.
457  *      It is used when heuristic for sequential allocation fails.
458  *      Rules are:
459  *        + if there is a block to the left of our position - allocate near it.
460  *        + if pointer will live in indirect block - allocate near that block.
461  *        + if pointer will live in inode - allocate in the same
462  *          cylinder group.
463  *
464  * In the latter case we colour the starting block by the callers PID to
465  * prevent it from clashing with concurrent allocations for a different inode
466  * in the same block group.   The PID is used here so that functionally related
467  * files will be close-by on-disk.
468  *
469  *      Caller must make sure that @ind is valid and will stay that way.
470  */
471 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
472 {
473         struct ext3_inode_info *ei = EXT3_I(inode);
474         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
475         __le32 *p;
476         ext3_fsblk_t bg_start;
477         ext3_grpblk_t colour;
478
479         /* Try to find previous block */
480         for (p = ind->p - 1; p >= start; p--) {
481                 if (*p)
482                         return le32_to_cpu(*p);
483         }
484
485         /* No such thing, so let's try location of indirect block */
486         if (ind->bh)
487                 return ind->bh->b_blocknr;
488
489         /*
490          * It is going to be referred to from the inode itself? OK, just put it
491          * into the same cylinder group then.
492          */
493         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
494         colour = (current->pid % 16) *
495                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
496         return bg_start + colour;
497 }
498
499 /**
500  *      ext3_find_goal - find a preferred place for allocation.
501  *      @inode: owner
502  *      @block:  block we want
503  *      @partial: pointer to the last triple within a chain
504  *
505  *      Normally this function find the preferred place for block allocation,
506  *      returns it.
507  */
508
509 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
510                                    Indirect *partial)
511 {
512         struct ext3_block_alloc_info *block_i;
513
514         block_i =  EXT3_I(inode)->i_block_alloc_info;
515
516         /*
517          * try the heuristic for sequential allocation,
518          * failing that at least try to get decent locality.
519          */
520         if (block_i && (block == block_i->last_alloc_logical_block + 1)
521                 && (block_i->last_alloc_physical_block != 0)) {
522                 return block_i->last_alloc_physical_block + 1;
523         }
524
525         return ext3_find_near(inode, partial);
526 }
527
528 /**
529  *      ext3_blks_to_allocate - Look up the block map and count the number
530  *      of direct blocks need to be allocated for the given branch.
531  *
532  *      @branch: chain of indirect blocks
533  *      @k: number of blocks need for indirect blocks
534  *      @blks: number of data blocks to be mapped.
535  *      @blocks_to_boundary:  the offset in the indirect block
536  *
537  *      return the total number of blocks to be allocate, including the
538  *      direct and indirect blocks.
539  */
540 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
541                 int blocks_to_boundary)
542 {
543         unsigned long count = 0;
544
545         /*
546          * Simple case, [t,d]Indirect block(s) has not allocated yet
547          * then it's clear blocks on that path have not allocated
548          */
549         if (k > 0) {
550                 /* right now we don't handle cross boundary allocation */
551                 if (blks < blocks_to_boundary + 1)
552                         count += blks;
553                 else
554                         count += blocks_to_boundary + 1;
555                 return count;
556         }
557
558         count++;
559         while (count < blks && count <= blocks_to_boundary &&
560                 le32_to_cpu(*(branch[0].p + count)) == 0) {
561                 count++;
562         }
563         return count;
564 }
565
566 /**
567  *      ext3_alloc_blocks - multiple allocate blocks needed for a branch
568  *      @handle: handle for this transaction
569  *      @inode: owner
570  *      @goal: preferred place for allocation
571  *      @indirect_blks: the number of blocks need to allocate for indirect
572  *                      blocks
573  *      @blks:  number of blocks need to allocated for direct blocks
574  *      @new_blocks: on return it will store the new block numbers for
575  *      the indirect blocks(if needed) and the first direct block,
576  *      @err: here we store the error value
577  *
578  *      return the number of direct blocks allocated
579  */
580 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
581                         ext3_fsblk_t goal, int indirect_blks, int blks,
582                         ext3_fsblk_t new_blocks[4], int *err)
583 {
584         int target, i;
585         unsigned long count = 0;
586         int index = 0;
587         ext3_fsblk_t current_block = 0;
588         int ret = 0;
589
590         /*
591          * Here we try to allocate the requested multiple blocks at once,
592          * on a best-effort basis.
593          * To build a branch, we should allocate blocks for
594          * the indirect blocks(if not allocated yet), and at least
595          * the first direct block of this branch.  That's the
596          * minimum number of blocks need to allocate(required)
597          */
598         target = blks + indirect_blks;
599
600         while (1) {
601                 count = target;
602                 /* allocating blocks for indirect blocks and direct blocks */
603                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
604                 if (*err)
605                         goto failed_out;
606
607                 target -= count;
608                 /* allocate blocks for indirect blocks */
609                 while (index < indirect_blks && count) {
610                         new_blocks[index++] = current_block++;
611                         count--;
612                 }
613
614                 if (count > 0)
615                         break;
616         }
617
618         /* save the new block number for the first direct block */
619         new_blocks[index] = current_block;
620
621         /* total number of blocks allocated for direct blocks */
622         ret = count;
623         *err = 0;
624         return ret;
625 failed_out:
626         for (i = 0; i <index; i++)
627                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
628         return ret;
629 }
630
631 /**
632  *      ext3_alloc_branch - allocate and set up a chain of blocks.
633  *      @handle: handle for this transaction
634  *      @inode: owner
635  *      @indirect_blks: number of allocated indirect blocks
636  *      @blks: number of allocated direct blocks
637  *      @goal: preferred place for allocation
638  *      @offsets: offsets (in the blocks) to store the pointers to next.
639  *      @branch: place to store the chain in.
640  *
641  *      This function allocates blocks, zeroes out all but the last one,
642  *      links them into chain and (if we are synchronous) writes them to disk.
643  *      In other words, it prepares a branch that can be spliced onto the
644  *      inode. It stores the information about that chain in the branch[], in
645  *      the same format as ext3_get_branch() would do. We are calling it after
646  *      we had read the existing part of chain and partial points to the last
647  *      triple of that (one with zero ->key). Upon the exit we have the same
648  *      picture as after the successful ext3_get_block(), except that in one
649  *      place chain is disconnected - *branch->p is still zero (we did not
650  *      set the last link), but branch->key contains the number that should
651  *      be placed into *branch->p to fill that gap.
652  *
653  *      If allocation fails we free all blocks we've allocated (and forget
654  *      their buffer_heads) and return the error value the from failed
655  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
656  *      as described above and return 0.
657  */
658 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
659                         int indirect_blks, int *blks, ext3_fsblk_t goal,
660                         int *offsets, Indirect *branch)
661 {
662         int blocksize = inode->i_sb->s_blocksize;
663         int i, n = 0;
664         int err = 0;
665         struct buffer_head *bh;
666         int num;
667         ext3_fsblk_t new_blocks[4];
668         ext3_fsblk_t current_block;
669
670         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
671                                 *blks, new_blocks, &err);
672         if (err)
673                 return err;
674
675         branch[0].key = cpu_to_le32(new_blocks[0]);
676         /*
677          * metadata blocks and data blocks are allocated.
678          */
679         for (n = 1; n <= indirect_blks;  n++) {
680                 /*
681                  * Get buffer_head for parent block, zero it out
682                  * and set the pointer to new one, then send
683                  * parent to disk.
684                  */
685                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
686                 branch[n].bh = bh;
687                 lock_buffer(bh);
688                 BUFFER_TRACE(bh, "call get_create_access");
689                 err = ext3_journal_get_create_access(handle, bh);
690                 if (err) {
691                         unlock_buffer(bh);
692                         brelse(bh);
693                         goto failed;
694                 }
695
696                 memset(bh->b_data, 0, blocksize);
697                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
698                 branch[n].key = cpu_to_le32(new_blocks[n]);
699                 *branch[n].p = branch[n].key;
700                 if ( n == indirect_blks) {
701                         current_block = new_blocks[n];
702                         /*
703                          * End of chain, update the last new metablock of
704                          * the chain to point to the new allocated
705                          * data blocks numbers
706                          */
707                         for (i=1; i < num; i++)
708                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
709                 }
710                 BUFFER_TRACE(bh, "marking uptodate");
711                 set_buffer_uptodate(bh);
712                 unlock_buffer(bh);
713
714                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
715                 err = ext3_journal_dirty_metadata(handle, bh);
716                 if (err)
717                         goto failed;
718         }
719         *blks = num;
720         return err;
721 failed:
722         /* Allocation failed, free what we already allocated */
723         for (i = 1; i <= n ; i++) {
724                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
725                 ext3_journal_forget(handle, branch[i].bh);
726         }
727         for (i = 0; i <indirect_blks; i++)
728                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
729
730         ext3_free_blocks(handle, inode, new_blocks[i], num);
731
732         return err;
733 }
734
735 /**
736  * ext3_splice_branch - splice the allocated branch onto inode.
737  * @handle: handle for this transaction
738  * @inode: owner
739  * @block: (logical) number of block we are adding
740  * @where: location of missing link
741  * @num:   number of indirect blocks we are adding
742  * @blks:  number of direct blocks we are adding
743  *
744  * This function fills the missing link and does all housekeeping needed in
745  * inode (->i_blocks, etc.). In case of success we end up with the full
746  * chain to new block and return 0.
747  */
748 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
749                         long block, Indirect *where, int num, int blks)
750 {
751         int i;
752         int err = 0;
753         struct ext3_block_alloc_info *block_i;
754         ext3_fsblk_t current_block;
755         struct ext3_inode_info *ei = EXT3_I(inode);
756
757         block_i = ei->i_block_alloc_info;
758         /*
759          * If we're splicing into a [td]indirect block (as opposed to the
760          * inode) then we need to get write access to the [td]indirect block
761          * before the splice.
762          */
763         if (where->bh) {
764                 BUFFER_TRACE(where->bh, "get_write_access");
765                 err = ext3_journal_get_write_access(handle, where->bh);
766                 if (err)
767                         goto err_out;
768         }
769         /* That's it */
770
771         *where->p = where->key;
772
773         /*
774          * Update the host buffer_head or inode to point to more just allocated
775          * direct blocks blocks
776          */
777         if (num == 0 && blks > 1) {
778                 current_block = le32_to_cpu(where->key) + 1;
779                 for (i = 1; i < blks; i++)
780                         *(where->p + i ) = cpu_to_le32(current_block++);
781         }
782
783         /*
784          * update the most recently allocated logical & physical block
785          * in i_block_alloc_info, to assist find the proper goal block for next
786          * allocation
787          */
788         if (block_i) {
789                 block_i->last_alloc_logical_block = block + blks - 1;
790                 block_i->last_alloc_physical_block =
791                                 le32_to_cpu(where[num].key) + blks - 1;
792         }
793
794         /* We are done with atomic stuff, now do the rest of housekeeping */
795
796         inode->i_ctime = CURRENT_TIME_SEC;
797         ext3_mark_inode_dirty(handle, inode);
798         /* ext3_mark_inode_dirty already updated i_sync_tid */
799         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
800
801         /* had we spliced it onto indirect block? */
802         if (where->bh) {
803                 /*
804                  * If we spliced it onto an indirect block, we haven't
805                  * altered the inode.  Note however that if it is being spliced
806                  * onto an indirect block at the very end of the file (the
807                  * file is growing) then we *will* alter the inode to reflect
808                  * the new i_size.  But that is not done here - it is done in
809                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
810                  */
811                 jbd_debug(5, "splicing indirect only\n");
812                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
813                 err = ext3_journal_dirty_metadata(handle, where->bh);
814                 if (err)
815                         goto err_out;
816         } else {
817                 /*
818                  * OK, we spliced it into the inode itself on a direct block.
819                  * Inode was dirtied above.
820                  */
821                 jbd_debug(5, "splicing direct\n");
822         }
823         return err;
824
825 err_out:
826         for (i = 1; i <= num; i++) {
827                 BUFFER_TRACE(where[i].bh, "call journal_forget");
828                 ext3_journal_forget(handle, where[i].bh);
829                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
830         }
831         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
832
833         return err;
834 }
835
836 /*
837  * Allocation strategy is simple: if we have to allocate something, we will
838  * have to go the whole way to leaf. So let's do it before attaching anything
839  * to tree, set linkage between the newborn blocks, write them if sync is
840  * required, recheck the path, free and repeat if check fails, otherwise
841  * set the last missing link (that will protect us from any truncate-generated
842  * removals - all blocks on the path are immune now) and possibly force the
843  * write on the parent block.
844  * That has a nice additional property: no special recovery from the failed
845  * allocations is needed - we simply release blocks and do not touch anything
846  * reachable from inode.
847  *
848  * `handle' can be NULL if create == 0.
849  *
850  * The BKL may not be held on entry here.  Be sure to take it early.
851  * return > 0, # of blocks mapped or allocated.
852  * return = 0, if plain lookup failed.
853  * return < 0, error case.
854  */
855 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
856                 sector_t iblock, unsigned long maxblocks,
857                 struct buffer_head *bh_result,
858                 int create)
859 {
860         int err = -EIO;
861         int offsets[4];
862         Indirect chain[4];
863         Indirect *partial;
864         ext3_fsblk_t goal;
865         int indirect_blks;
866         int blocks_to_boundary = 0;
867         int depth;
868         struct ext3_inode_info *ei = EXT3_I(inode);
869         int count = 0;
870         ext3_fsblk_t first_block = 0;
871
872
873         trace_ext3_get_blocks_enter(inode, iblock, maxblocks, create);
874         J_ASSERT(handle != NULL || create == 0);
875         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
876
877         if (depth == 0)
878                 goto out;
879
880         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
881
882         /* Simplest case - block found, no allocation needed */
883         if (!partial) {
884                 first_block = le32_to_cpu(chain[depth - 1].key);
885                 clear_buffer_new(bh_result);
886                 count++;
887                 /*map more blocks*/
888                 while (count < maxblocks && count <= blocks_to_boundary) {
889                         ext3_fsblk_t blk;
890
891                         if (!verify_chain(chain, chain + depth - 1)) {
892                                 /*
893                                  * Indirect block might be removed by
894                                  * truncate while we were reading it.
895                                  * Handling of that case: forget what we've
896                                  * got now. Flag the err as EAGAIN, so it
897                                  * will reread.
898                                  */
899                                 err = -EAGAIN;
900                                 count = 0;
901                                 break;
902                         }
903                         blk = le32_to_cpu(*(chain[depth-1].p + count));
904
905                         if (blk == first_block + count)
906                                 count++;
907                         else
908                                 break;
909                 }
910                 if (err != -EAGAIN)
911                         goto got_it;
912         }
913
914         /* Next simple case - plain lookup or failed read of indirect block */
915         if (!create || err == -EIO)
916                 goto cleanup;
917
918         /*
919          * Block out ext3_truncate while we alter the tree
920          */
921         mutex_lock(&ei->truncate_mutex);
922
923         /*
924          * If the indirect block is missing while we are reading
925          * the chain(ext3_get_branch() returns -EAGAIN err), or
926          * if the chain has been changed after we grab the semaphore,
927          * (either because another process truncated this branch, or
928          * another get_block allocated this branch) re-grab the chain to see if
929          * the request block has been allocated or not.
930          *
931          * Since we already block the truncate/other get_block
932          * at this point, we will have the current copy of the chain when we
933          * splice the branch into the tree.
934          */
935         if (err == -EAGAIN || !verify_chain(chain, partial)) {
936                 while (partial > chain) {
937                         brelse(partial->bh);
938                         partial--;
939                 }
940                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
941                 if (!partial) {
942                         count++;
943                         mutex_unlock(&ei->truncate_mutex);
944                         if (err)
945                                 goto cleanup;
946                         clear_buffer_new(bh_result);
947                         goto got_it;
948                 }
949         }
950
951         /*
952          * Okay, we need to do block allocation.  Lazily initialize the block
953          * allocation info here if necessary
954         */
955         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
956                 ext3_init_block_alloc_info(inode);
957
958         goal = ext3_find_goal(inode, iblock, partial);
959
960         /* the number of blocks need to allocate for [d,t]indirect blocks */
961         indirect_blks = (chain + depth) - partial - 1;
962
963         /*
964          * Next look up the indirect map to count the totoal number of
965          * direct blocks to allocate for this branch.
966          */
967         count = ext3_blks_to_allocate(partial, indirect_blks,
968                                         maxblocks, blocks_to_boundary);
969         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
970                                 offsets + (partial - chain), partial);
971
972         /*
973          * The ext3_splice_branch call will free and forget any buffers
974          * on the new chain if there is a failure, but that risks using
975          * up transaction credits, especially for bitmaps where the
976          * credits cannot be returned.  Can we handle this somehow?  We
977          * may need to return -EAGAIN upwards in the worst case.  --sct
978          */
979         if (!err)
980                 err = ext3_splice_branch(handle, inode, iblock,
981                                         partial, indirect_blks, count);
982         mutex_unlock(&ei->truncate_mutex);
983         if (err)
984                 goto cleanup;
985
986         set_buffer_new(bh_result);
987 got_it:
988         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
989         if (count > blocks_to_boundary)
990                 set_buffer_boundary(bh_result);
991         err = count;
992         /* Clean up and exit */
993         partial = chain + depth - 1;    /* the whole chain */
994 cleanup:
995         while (partial > chain) {
996                 BUFFER_TRACE(partial->bh, "call brelse");
997                 brelse(partial->bh);
998                 partial--;
999         }
1000         BUFFER_TRACE(bh_result, "returned");
1001 out:
1002         trace_ext3_get_blocks_exit(inode, iblock,
1003                                    depth ? le32_to_cpu(chain[depth-1].key) : 0,
1004                                    count, err);
1005         return err;
1006 }
1007
1008 /* Maximum number of blocks we map for direct IO at once. */
1009 #define DIO_MAX_BLOCKS 4096
1010 /*
1011  * Number of credits we need for writing DIO_MAX_BLOCKS:
1012  * We need sb + group descriptor + bitmap + inode -> 4
1013  * For B blocks with A block pointers per block we need:
1014  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
1015  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
1016  */
1017 #define DIO_CREDITS 25
1018
1019 static int ext3_get_block(struct inode *inode, sector_t iblock,
1020                         struct buffer_head *bh_result, int create)
1021 {
1022         handle_t *handle = ext3_journal_current_handle();
1023         int ret = 0, started = 0;
1024         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1025
1026         if (create && !handle) {        /* Direct IO write... */
1027                 if (max_blocks > DIO_MAX_BLOCKS)
1028                         max_blocks = DIO_MAX_BLOCKS;
1029                 handle = ext3_journal_start(inode, DIO_CREDITS +
1030                                 EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb));
1031                 if (IS_ERR(handle)) {
1032                         ret = PTR_ERR(handle);
1033                         goto out;
1034                 }
1035                 started = 1;
1036         }
1037
1038         ret = ext3_get_blocks_handle(handle, inode, iblock,
1039                                         max_blocks, bh_result, create);
1040         if (ret > 0) {
1041                 bh_result->b_size = (ret << inode->i_blkbits);
1042                 ret = 0;
1043         }
1044         if (started)
1045                 ext3_journal_stop(handle);
1046 out:
1047         return ret;
1048 }
1049
1050 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
1051                 u64 start, u64 len)
1052 {
1053         return generic_block_fiemap(inode, fieinfo, start, len,
1054                                     ext3_get_block);
1055 }
1056
1057 /*
1058  * `handle' can be NULL if create is zero
1059  */
1060 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1061                                 long block, int create, int *errp)
1062 {
1063         struct buffer_head dummy;
1064         int fatal = 0, err;
1065
1066         J_ASSERT(handle != NULL || create == 0);
1067
1068         dummy.b_state = 0;
1069         dummy.b_blocknr = -1000;
1070         buffer_trace_init(&dummy.b_history);
1071         err = ext3_get_blocks_handle(handle, inode, block, 1,
1072                                         &dummy, create);
1073         /*
1074          * ext3_get_blocks_handle() returns number of blocks
1075          * mapped. 0 in case of a HOLE.
1076          */
1077         if (err > 0) {
1078                 if (err > 1)
1079                         WARN_ON(1);
1080                 err = 0;
1081         }
1082         *errp = err;
1083         if (!err && buffer_mapped(&dummy)) {
1084                 struct buffer_head *bh;
1085                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1086                 if (!bh) {
1087                         *errp = -EIO;
1088                         goto err;
1089                 }
1090                 if (buffer_new(&dummy)) {
1091                         J_ASSERT(create != 0);
1092                         J_ASSERT(handle != NULL);
1093
1094                         /*
1095                          * Now that we do not always journal data, we should
1096                          * keep in mind whether this should always journal the
1097                          * new buffer as metadata.  For now, regular file
1098                          * writes use ext3_get_block instead, so it's not a
1099                          * problem.
1100                          */
1101                         lock_buffer(bh);
1102                         BUFFER_TRACE(bh, "call get_create_access");
1103                         fatal = ext3_journal_get_create_access(handle, bh);
1104                         if (!fatal && !buffer_uptodate(bh)) {
1105                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1106                                 set_buffer_uptodate(bh);
1107                         }
1108                         unlock_buffer(bh);
1109                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1110                         err = ext3_journal_dirty_metadata(handle, bh);
1111                         if (!fatal)
1112                                 fatal = err;
1113                 } else {
1114                         BUFFER_TRACE(bh, "not a new buffer");
1115                 }
1116                 if (fatal) {
1117                         *errp = fatal;
1118                         brelse(bh);
1119                         bh = NULL;
1120                 }
1121                 return bh;
1122         }
1123 err:
1124         return NULL;
1125 }
1126
1127 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1128                                int block, int create, int *err)
1129 {
1130         struct buffer_head * bh;
1131
1132         bh = ext3_getblk(handle, inode, block, create, err);
1133         if (!bh)
1134                 return bh;
1135         if (buffer_uptodate(bh))
1136                 return bh;
1137         ll_rw_block(READ_META, 1, &bh);
1138         wait_on_buffer(bh);
1139         if (buffer_uptodate(bh))
1140                 return bh;
1141         put_bh(bh);
1142         *err = -EIO;
1143         return NULL;
1144 }
1145
1146 static int walk_page_buffers(   handle_t *handle,
1147                                 struct buffer_head *head,
1148                                 unsigned from,
1149                                 unsigned to,
1150                                 int *partial,
1151                                 int (*fn)(      handle_t *handle,
1152                                                 struct buffer_head *bh))
1153 {
1154         struct buffer_head *bh;
1155         unsigned block_start, block_end;
1156         unsigned blocksize = head->b_size;
1157         int err, ret = 0;
1158         struct buffer_head *next;
1159
1160         for (   bh = head, block_start = 0;
1161                 ret == 0 && (bh != head || !block_start);
1162                 block_start = block_end, bh = next)
1163         {
1164                 next = bh->b_this_page;
1165                 block_end = block_start + blocksize;
1166                 if (block_end <= from || block_start >= to) {
1167                         if (partial && !buffer_uptodate(bh))
1168                                 *partial = 1;
1169                         continue;
1170                 }
1171                 err = (*fn)(handle, bh);
1172                 if (!ret)
1173                         ret = err;
1174         }
1175         return ret;
1176 }
1177
1178 /*
1179  * To preserve ordering, it is essential that the hole instantiation and
1180  * the data write be encapsulated in a single transaction.  We cannot
1181  * close off a transaction and start a new one between the ext3_get_block()
1182  * and the commit_write().  So doing the journal_start at the start of
1183  * prepare_write() is the right place.
1184  *
1185  * Also, this function can nest inside ext3_writepage() ->
1186  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1187  * has generated enough buffer credits to do the whole page.  So we won't
1188  * block on the journal in that case, which is good, because the caller may
1189  * be PF_MEMALLOC.
1190  *
1191  * By accident, ext3 can be reentered when a transaction is open via
1192  * quota file writes.  If we were to commit the transaction while thus
1193  * reentered, there can be a deadlock - we would be holding a quota
1194  * lock, and the commit would never complete if another thread had a
1195  * transaction open and was blocking on the quota lock - a ranking
1196  * violation.
1197  *
1198  * So what we do is to rely on the fact that journal_stop/journal_start
1199  * will _not_ run commit under these circumstances because handle->h_ref
1200  * is elevated.  We'll still have enough credits for the tiny quotafile
1201  * write.
1202  */
1203 static int do_journal_get_write_access(handle_t *handle,
1204                                         struct buffer_head *bh)
1205 {
1206         int dirty = buffer_dirty(bh);
1207         int ret;
1208
1209         if (!buffer_mapped(bh) || buffer_freed(bh))
1210                 return 0;
1211         /*
1212          * __block_prepare_write() could have dirtied some buffers. Clean
1213          * the dirty bit as jbd2_journal_get_write_access() could complain
1214          * otherwise about fs integrity issues. Setting of the dirty bit
1215          * by __block_prepare_write() isn't a real problem here as we clear
1216          * the bit before releasing a page lock and thus writeback cannot
1217          * ever write the buffer.
1218          */
1219         if (dirty)
1220                 clear_buffer_dirty(bh);
1221         ret = ext3_journal_get_write_access(handle, bh);
1222         if (!ret && dirty)
1223                 ret = ext3_journal_dirty_metadata(handle, bh);
1224         return ret;
1225 }
1226
1227 /*
1228  * Truncate blocks that were not used by write. We have to truncate the
1229  * pagecache as well so that corresponding buffers get properly unmapped.
1230  */
1231 static void ext3_truncate_failed_write(struct inode *inode)
1232 {
1233         truncate_inode_pages(inode->i_mapping, inode->i_size);
1234         ext3_truncate(inode);
1235 }
1236
1237 /*
1238  * Truncate blocks that were not used by direct IO write. We have to zero out
1239  * the last file block as well because direct IO might have written to it.
1240  */
1241 static void ext3_truncate_failed_direct_write(struct inode *inode)
1242 {
1243         ext3_block_truncate_page(inode, inode->i_size);
1244         ext3_truncate(inode);
1245 }
1246
1247 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1248                                 loff_t pos, unsigned len, unsigned flags,
1249                                 struct page **pagep, void **fsdata)
1250 {
1251         struct inode *inode = mapping->host;
1252         int ret;
1253         handle_t *handle;
1254         int retries = 0;
1255         struct page *page;
1256         pgoff_t index;
1257         unsigned from, to;
1258         /* Reserve one block more for addition to orphan list in case
1259          * we allocate blocks but write fails for some reason */
1260         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1261
1262         trace_ext3_write_begin(inode, pos, len, flags);
1263
1264         index = pos >> PAGE_CACHE_SHIFT;
1265         from = pos & (PAGE_CACHE_SIZE - 1);
1266         to = from + len;
1267
1268 retry:
1269         page = grab_cache_page_write_begin(mapping, index, flags);
1270         if (!page)
1271                 return -ENOMEM;
1272         *pagep = page;
1273
1274         handle = ext3_journal_start(inode, needed_blocks);
1275         if (IS_ERR(handle)) {
1276                 unlock_page(page);
1277                 page_cache_release(page);
1278                 ret = PTR_ERR(handle);
1279                 goto out;
1280         }
1281         ret = __block_write_begin(page, pos, len, ext3_get_block);
1282         if (ret)
1283                 goto write_begin_failed;
1284
1285         if (ext3_should_journal_data(inode)) {
1286                 ret = walk_page_buffers(handle, page_buffers(page),
1287                                 from, to, NULL, do_journal_get_write_access);
1288         }
1289 write_begin_failed:
1290         if (ret) {
1291                 /*
1292                  * block_write_begin may have instantiated a few blocks
1293                  * outside i_size.  Trim these off again. Don't need
1294                  * i_size_read because we hold i_mutex.
1295                  *
1296                  * Add inode to orphan list in case we crash before truncate
1297                  * finishes. Do this only if ext3_can_truncate() agrees so
1298                  * that orphan processing code is happy.
1299                  */
1300                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1301                         ext3_orphan_add(handle, inode);
1302                 ext3_journal_stop(handle);
1303                 unlock_page(page);
1304                 page_cache_release(page);
1305                 if (pos + len > inode->i_size)
1306                         ext3_truncate_failed_write(inode);
1307         }
1308         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1309                 goto retry;
1310 out:
1311         return ret;
1312 }
1313
1314
1315 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1316 {
1317         int err = journal_dirty_data(handle, bh);
1318         if (err)
1319                 ext3_journal_abort_handle(__func__, __func__,
1320                                                 bh, handle, err);
1321         return err;
1322 }
1323
1324 /* For ordered writepage and write_end functions */
1325 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1326 {
1327         /*
1328          * Write could have mapped the buffer but it didn't copy the data in
1329          * yet. So avoid filing such buffer into a transaction.
1330          */
1331         if (buffer_mapped(bh) && buffer_uptodate(bh))
1332                 return ext3_journal_dirty_data(handle, bh);
1333         return 0;
1334 }
1335
1336 /* For write_end() in data=journal mode */
1337 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1338 {
1339         if (!buffer_mapped(bh) || buffer_freed(bh))
1340                 return 0;
1341         set_buffer_uptodate(bh);
1342         return ext3_journal_dirty_metadata(handle, bh);
1343 }
1344
1345 /*
1346  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1347  * for the whole page but later we failed to copy the data in. Update inode
1348  * size according to what we managed to copy. The rest is going to be
1349  * truncated in write_end function.
1350  */
1351 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1352 {
1353         /* What matters to us is i_disksize. We don't write i_size anywhere */
1354         if (pos + copied > inode->i_size)
1355                 i_size_write(inode, pos + copied);
1356         if (pos + copied > EXT3_I(inode)->i_disksize) {
1357                 EXT3_I(inode)->i_disksize = pos + copied;
1358                 mark_inode_dirty(inode);
1359         }
1360 }
1361
1362 /*
1363  * We need to pick up the new inode size which generic_commit_write gave us
1364  * `file' can be NULL - eg, when called from page_symlink().
1365  *
1366  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1367  * buffers are managed internally.
1368  */
1369 static int ext3_ordered_write_end(struct file *file,
1370                                 struct address_space *mapping,
1371                                 loff_t pos, unsigned len, unsigned copied,
1372                                 struct page *page, void *fsdata)
1373 {
1374         handle_t *handle = ext3_journal_current_handle();
1375         struct inode *inode = file->f_mapping->host;
1376         unsigned from, to;
1377         int ret = 0, ret2;
1378
1379         trace_ext3_ordered_write_end(inode, pos, len, copied);
1380         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1381
1382         from = pos & (PAGE_CACHE_SIZE - 1);
1383         to = from + copied;
1384         ret = walk_page_buffers(handle, page_buffers(page),
1385                 from, to, NULL, journal_dirty_data_fn);
1386
1387         if (ret == 0)
1388                 update_file_sizes(inode, pos, copied);
1389         /*
1390          * There may be allocated blocks outside of i_size because
1391          * we failed to copy some data. Prepare for truncate.
1392          */
1393         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1394                 ext3_orphan_add(handle, inode);
1395         ret2 = ext3_journal_stop(handle);
1396         if (!ret)
1397                 ret = ret2;
1398         unlock_page(page);
1399         page_cache_release(page);
1400
1401         if (pos + len > inode->i_size)
1402                 ext3_truncate_failed_write(inode);
1403         return ret ? ret : copied;
1404 }
1405
1406 static int ext3_writeback_write_end(struct file *file,
1407                                 struct address_space *mapping,
1408                                 loff_t pos, unsigned len, unsigned copied,
1409                                 struct page *page, void *fsdata)
1410 {
1411         handle_t *handle = ext3_journal_current_handle();
1412         struct inode *inode = file->f_mapping->host;
1413         int ret;
1414
1415         trace_ext3_writeback_write_end(inode, pos, len, copied);
1416         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1417         update_file_sizes(inode, pos, copied);
1418         /*
1419          * There may be allocated blocks outside of i_size because
1420          * we failed to copy some data. Prepare for truncate.
1421          */
1422         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1423                 ext3_orphan_add(handle, inode);
1424         ret = ext3_journal_stop(handle);
1425         unlock_page(page);
1426         page_cache_release(page);
1427
1428         if (pos + len > inode->i_size)
1429                 ext3_truncate_failed_write(inode);
1430         return ret ? ret : copied;
1431 }
1432
1433 static int ext3_journalled_write_end(struct file *file,
1434                                 struct address_space *mapping,
1435                                 loff_t pos, unsigned len, unsigned copied,
1436                                 struct page *page, void *fsdata)
1437 {
1438         handle_t *handle = ext3_journal_current_handle();
1439         struct inode *inode = mapping->host;
1440         struct ext3_inode_info *ei = EXT3_I(inode);
1441         int ret = 0, ret2;
1442         int partial = 0;
1443         unsigned from, to;
1444
1445         trace_ext3_journalled_write_end(inode, pos, len, copied);
1446         from = pos & (PAGE_CACHE_SIZE - 1);
1447         to = from + len;
1448
1449         if (copied < len) {
1450                 if (!PageUptodate(page))
1451                         copied = 0;
1452                 page_zero_new_buffers(page, from + copied, to);
1453                 to = from + copied;
1454         }
1455
1456         ret = walk_page_buffers(handle, page_buffers(page), from,
1457                                 to, &partial, write_end_fn);
1458         if (!partial)
1459                 SetPageUptodate(page);
1460
1461         if (pos + copied > inode->i_size)
1462                 i_size_write(inode, pos + copied);
1463         /*
1464          * There may be allocated blocks outside of i_size because
1465          * we failed to copy some data. Prepare for truncate.
1466          */
1467         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1468                 ext3_orphan_add(handle, inode);
1469         ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1470         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
1471         if (inode->i_size > ei->i_disksize) {
1472                 ei->i_disksize = inode->i_size;
1473                 ret2 = ext3_mark_inode_dirty(handle, inode);
1474                 if (!ret)
1475                         ret = ret2;
1476         }
1477
1478         ret2 = ext3_journal_stop(handle);
1479         if (!ret)
1480                 ret = ret2;
1481         unlock_page(page);
1482         page_cache_release(page);
1483
1484         if (pos + len > inode->i_size)
1485                 ext3_truncate_failed_write(inode);
1486         return ret ? ret : copied;
1487 }
1488
1489 /*
1490  * bmap() is special.  It gets used by applications such as lilo and by
1491  * the swapper to find the on-disk block of a specific piece of data.
1492  *
1493  * Naturally, this is dangerous if the block concerned is still in the
1494  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1495  * filesystem and enables swap, then they may get a nasty shock when the
1496  * data getting swapped to that swapfile suddenly gets overwritten by
1497  * the original zero's written out previously to the journal and
1498  * awaiting writeback in the kernel's buffer cache.
1499  *
1500  * So, if we see any bmap calls here on a modified, data-journaled file,
1501  * take extra steps to flush any blocks which might be in the cache.
1502  */
1503 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1504 {
1505         struct inode *inode = mapping->host;
1506         journal_t *journal;
1507         int err;
1508
1509         if (ext3_test_inode_state(inode, EXT3_STATE_JDATA)) {
1510                 /*
1511                  * This is a REALLY heavyweight approach, but the use of
1512                  * bmap on dirty files is expected to be extremely rare:
1513                  * only if we run lilo or swapon on a freshly made file
1514                  * do we expect this to happen.
1515                  *
1516                  * (bmap requires CAP_SYS_RAWIO so this does not
1517                  * represent an unprivileged user DOS attack --- we'd be
1518                  * in trouble if mortal users could trigger this path at
1519                  * will.)
1520                  *
1521                  * NB. EXT3_STATE_JDATA is not set on files other than
1522                  * regular files.  If somebody wants to bmap a directory
1523                  * or symlink and gets confused because the buffer
1524                  * hasn't yet been flushed to disk, they deserve
1525                  * everything they get.
1526                  */
1527
1528                 ext3_clear_inode_state(inode, EXT3_STATE_JDATA);
1529                 journal = EXT3_JOURNAL(inode);
1530                 journal_lock_updates(journal);
1531                 err = journal_flush(journal);
1532                 journal_unlock_updates(journal);
1533
1534                 if (err)
1535                         return 0;
1536         }
1537
1538         return generic_block_bmap(mapping,block,ext3_get_block);
1539 }
1540
1541 static int bget_one(handle_t *handle, struct buffer_head *bh)
1542 {
1543         get_bh(bh);
1544         return 0;
1545 }
1546
1547 static int bput_one(handle_t *handle, struct buffer_head *bh)
1548 {
1549         put_bh(bh);
1550         return 0;
1551 }
1552
1553 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1554 {
1555         return !buffer_mapped(bh);
1556 }
1557
1558 /*
1559  * Note that we always start a transaction even if we're not journalling
1560  * data.  This is to preserve ordering: any hole instantiation within
1561  * __block_write_full_page -> ext3_get_block() should be journalled
1562  * along with the data so we don't crash and then get metadata which
1563  * refers to old data.
1564  *
1565  * In all journalling modes block_write_full_page() will start the I/O.
1566  *
1567  * Problem:
1568  *
1569  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1570  *              ext3_writepage()
1571  *
1572  * Similar for:
1573  *
1574  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1575  *
1576  * Same applies to ext3_get_block().  We will deadlock on various things like
1577  * lock_journal and i_truncate_mutex.
1578  *
1579  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1580  * allocations fail.
1581  *
1582  * 16May01: If we're reentered then journal_current_handle() will be
1583  *          non-zero. We simply *return*.
1584  *
1585  * 1 July 2001: @@@ FIXME:
1586  *   In journalled data mode, a data buffer may be metadata against the
1587  *   current transaction.  But the same file is part of a shared mapping
1588  *   and someone does a writepage() on it.
1589  *
1590  *   We will move the buffer onto the async_data list, but *after* it has
1591  *   been dirtied. So there's a small window where we have dirty data on
1592  *   BJ_Metadata.
1593  *
1594  *   Note that this only applies to the last partial page in the file.  The
1595  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1596  *   broken code anyway: it's wrong for msync()).
1597  *
1598  *   It's a rare case: affects the final partial page, for journalled data
1599  *   where the file is subject to bith write() and writepage() in the same
1600  *   transction.  To fix it we'll need a custom block_write_full_page().
1601  *   We'll probably need that anyway for journalling writepage() output.
1602  *
1603  * We don't honour synchronous mounts for writepage().  That would be
1604  * disastrous.  Any write() or metadata operation will sync the fs for
1605  * us.
1606  *
1607  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1608  * we don't need to open a transaction here.
1609  */
1610 static int ext3_ordered_writepage(struct page *page,
1611                                 struct writeback_control *wbc)
1612 {
1613         struct inode *inode = page->mapping->host;
1614         struct buffer_head *page_bufs;
1615         handle_t *handle = NULL;
1616         int ret = 0;
1617         int err;
1618
1619         J_ASSERT(PageLocked(page));
1620         WARN_ON_ONCE(IS_RDONLY(inode));
1621
1622         /*
1623          * We give up here if we're reentered, because it might be for a
1624          * different filesystem.
1625          */
1626         if (ext3_journal_current_handle())
1627                 goto out_fail;
1628
1629         trace_ext3_ordered_writepage(page);
1630         if (!page_has_buffers(page)) {
1631                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1632                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1633                 page_bufs = page_buffers(page);
1634         } else {
1635                 page_bufs = page_buffers(page);
1636                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1637                                        NULL, buffer_unmapped)) {
1638                         /* Provide NULL get_block() to catch bugs if buffers
1639                          * weren't really mapped */
1640                         return block_write_full_page(page, NULL, wbc);
1641                 }
1642         }
1643         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1644
1645         if (IS_ERR(handle)) {
1646                 ret = PTR_ERR(handle);
1647                 goto out_fail;
1648         }
1649
1650         walk_page_buffers(handle, page_bufs, 0,
1651                         PAGE_CACHE_SIZE, NULL, bget_one);
1652
1653         ret = block_write_full_page(page, ext3_get_block, wbc);
1654
1655         /*
1656          * The page can become unlocked at any point now, and
1657          * truncate can then come in and change things.  So we
1658          * can't touch *page from now on.  But *page_bufs is
1659          * safe due to elevated refcount.
1660          */
1661
1662         /*
1663          * And attach them to the current transaction.  But only if
1664          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1665          * and generally junk.
1666          */
1667         if (ret == 0) {
1668                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1669                                         NULL, journal_dirty_data_fn);
1670                 if (!ret)
1671                         ret = err;
1672         }
1673         walk_page_buffers(handle, page_bufs, 0,
1674                         PAGE_CACHE_SIZE, NULL, bput_one);
1675         err = ext3_journal_stop(handle);
1676         if (!ret)
1677                 ret = err;
1678         return ret;
1679
1680 out_fail:
1681         redirty_page_for_writepage(wbc, page);
1682         unlock_page(page);
1683         return ret;
1684 }
1685
1686 static int ext3_writeback_writepage(struct page *page,
1687                                 struct writeback_control *wbc)
1688 {
1689         struct inode *inode = page->mapping->host;
1690         handle_t *handle = NULL;
1691         int ret = 0;
1692         int err;
1693
1694         J_ASSERT(PageLocked(page));
1695         WARN_ON_ONCE(IS_RDONLY(inode));
1696
1697         if (ext3_journal_current_handle())
1698                 goto out_fail;
1699
1700         trace_ext3_writeback_writepage(page);
1701         if (page_has_buffers(page)) {
1702                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1703                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1704                         /* Provide NULL get_block() to catch bugs if buffers
1705                          * weren't really mapped */
1706                         return block_write_full_page(page, NULL, wbc);
1707                 }
1708         }
1709
1710         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1711         if (IS_ERR(handle)) {
1712                 ret = PTR_ERR(handle);
1713                 goto out_fail;
1714         }
1715
1716         ret = block_write_full_page(page, ext3_get_block, wbc);
1717
1718         err = ext3_journal_stop(handle);
1719         if (!ret)
1720                 ret = err;
1721         return ret;
1722
1723 out_fail:
1724         redirty_page_for_writepage(wbc, page);
1725         unlock_page(page);
1726         return ret;
1727 }
1728
1729 static int ext3_journalled_writepage(struct page *page,
1730                                 struct writeback_control *wbc)
1731 {
1732         struct inode *inode = page->mapping->host;
1733         handle_t *handle = NULL;
1734         int ret = 0;
1735         int err;
1736
1737         J_ASSERT(PageLocked(page));
1738         WARN_ON_ONCE(IS_RDONLY(inode));
1739
1740         if (ext3_journal_current_handle())
1741                 goto no_write;
1742
1743         trace_ext3_journalled_writepage(page);
1744         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1745         if (IS_ERR(handle)) {
1746                 ret = PTR_ERR(handle);
1747                 goto no_write;
1748         }
1749
1750         if (!page_has_buffers(page) || PageChecked(page)) {
1751                 /*
1752                  * It's mmapped pagecache.  Add buffers and journal it.  There
1753                  * doesn't seem much point in redirtying the page here.
1754                  */
1755                 ClearPageChecked(page);
1756                 ret = __block_write_begin(page, 0, PAGE_CACHE_SIZE,
1757                                           ext3_get_block);
1758                 if (ret != 0) {
1759                         ext3_journal_stop(handle);
1760                         goto out_unlock;
1761                 }
1762                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1763                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1764
1765                 err = walk_page_buffers(handle, page_buffers(page), 0,
1766                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1767                 if (ret == 0)
1768                         ret = err;
1769                 ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1770                 atomic_set(&EXT3_I(inode)->i_datasync_tid,
1771                            handle->h_transaction->t_tid);
1772                 unlock_page(page);
1773         } else {
1774                 /*
1775                  * It may be a page full of checkpoint-mode buffers.  We don't
1776                  * really know unless we go poke around in the buffer_heads.
1777                  * But block_write_full_page will do the right thing.
1778                  */
1779                 ret = block_write_full_page(page, ext3_get_block, wbc);
1780         }
1781         err = ext3_journal_stop(handle);
1782         if (!ret)
1783                 ret = err;
1784 out:
1785         return ret;
1786
1787 no_write:
1788         redirty_page_for_writepage(wbc, page);
1789 out_unlock:
1790         unlock_page(page);
1791         goto out;
1792 }
1793
1794 static int ext3_readpage(struct file *file, struct page *page)
1795 {
1796         trace_ext3_readpage(page);
1797         return mpage_readpage(page, ext3_get_block);
1798 }
1799
1800 static int
1801 ext3_readpages(struct file *file, struct address_space *mapping,
1802                 struct list_head *pages, unsigned nr_pages)
1803 {
1804         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1805 }
1806
1807 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1808 {
1809         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1810
1811         trace_ext3_invalidatepage(page, offset);
1812
1813         /*
1814          * If it's a full truncate we just forget about the pending dirtying
1815          */
1816         if (offset == 0)
1817                 ClearPageChecked(page);
1818
1819         journal_invalidatepage(journal, page, offset);
1820 }
1821
1822 static int ext3_releasepage(struct page *page, gfp_t wait)
1823 {
1824         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1825
1826         trace_ext3_releasepage(page);
1827         WARN_ON(PageChecked(page));
1828         if (!page_has_buffers(page))
1829                 return 0;
1830         return journal_try_to_free_buffers(journal, page, wait);
1831 }
1832
1833 /*
1834  * If the O_DIRECT write will extend the file then add this inode to the
1835  * orphan list.  So recovery will truncate it back to the original size
1836  * if the machine crashes during the write.
1837  *
1838  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1839  * crashes then stale disk data _may_ be exposed inside the file. But current
1840  * VFS code falls back into buffered path in that case so we are safe.
1841  */
1842 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1843                         const struct iovec *iov, loff_t offset,
1844                         unsigned long nr_segs)
1845 {
1846         struct file *file = iocb->ki_filp;
1847         struct inode *inode = file->f_mapping->host;
1848         struct ext3_inode_info *ei = EXT3_I(inode);
1849         handle_t *handle;
1850         ssize_t ret;
1851         int orphan = 0;
1852         size_t count = iov_length(iov, nr_segs);
1853         int retries = 0;
1854
1855         trace_ext3_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
1856
1857         if (rw == WRITE) {
1858                 loff_t final_size = offset + count;
1859
1860                 if (final_size > inode->i_size) {
1861                         /* Credits for sb + inode write */
1862                         handle = ext3_journal_start(inode, 2);
1863                         if (IS_ERR(handle)) {
1864                                 ret = PTR_ERR(handle);
1865                                 goto out;
1866                         }
1867                         ret = ext3_orphan_add(handle, inode);
1868                         if (ret) {
1869                                 ext3_journal_stop(handle);
1870                                 goto out;
1871                         }
1872                         orphan = 1;
1873                         ei->i_disksize = inode->i_size;
1874                         ext3_journal_stop(handle);
1875                 }
1876         }
1877
1878 retry:
1879         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1880                                  offset, nr_segs,
1881                                  ext3_get_block, NULL);
1882         /*
1883          * In case of error extending write may have instantiated a few
1884          * blocks outside i_size. Trim these off again.
1885          */
1886         if (unlikely((rw & WRITE) && ret < 0)) {
1887                 loff_t isize = i_size_read(inode);
1888                 loff_t end = offset + iov_length(iov, nr_segs);
1889
1890                 if (end > isize)
1891                         ext3_truncate_failed_direct_write(inode);
1892         }
1893         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1894                 goto retry;
1895
1896         if (orphan) {
1897                 int err;
1898
1899                 /* Credits for sb + inode write */
1900                 handle = ext3_journal_start(inode, 2);
1901                 if (IS_ERR(handle)) {
1902                         /* This is really bad luck. We've written the data
1903                          * but cannot extend i_size. Truncate allocated blocks
1904                          * and pretend the write failed... */
1905                         ext3_truncate_failed_direct_write(inode);
1906                         ret = PTR_ERR(handle);
1907                         goto out;
1908                 }
1909                 if (inode->i_nlink)
1910                         ext3_orphan_del(handle, inode);
1911                 if (ret > 0) {
1912                         loff_t end = offset + ret;
1913                         if (end > inode->i_size) {
1914                                 ei->i_disksize = end;
1915                                 i_size_write(inode, end);
1916                                 /*
1917                                  * We're going to return a positive `ret'
1918                                  * here due to non-zero-length I/O, so there's
1919                                  * no way of reporting error returns from
1920                                  * ext3_mark_inode_dirty() to userspace.  So
1921                                  * ignore it.
1922                                  */
1923                                 ext3_mark_inode_dirty(handle, inode);
1924                         }
1925                 }
1926                 err = ext3_journal_stop(handle);
1927                 if (ret == 0)
1928                         ret = err;
1929         }
1930 out:
1931         trace_ext3_direct_IO_exit(inode, offset,
1932                                 iov_length(iov, nr_segs), rw, ret);
1933         return ret;
1934 }
1935
1936 /*
1937  * Pages can be marked dirty completely asynchronously from ext3's journalling
1938  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1939  * much here because ->set_page_dirty is called under VFS locks.  The page is
1940  * not necessarily locked.
1941  *
1942  * We cannot just dirty the page and leave attached buffers clean, because the
1943  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1944  * or jbddirty because all the journalling code will explode.
1945  *
1946  * So what we do is to mark the page "pending dirty" and next time writepage
1947  * is called, propagate that into the buffers appropriately.
1948  */
1949 static int ext3_journalled_set_page_dirty(struct page *page)
1950 {
1951         SetPageChecked(page);
1952         return __set_page_dirty_nobuffers(page);
1953 }
1954
1955 static const struct address_space_operations ext3_ordered_aops = {
1956         .readpage               = ext3_readpage,
1957         .readpages              = ext3_readpages,
1958         .writepage              = ext3_ordered_writepage,
1959         .write_begin            = ext3_write_begin,
1960         .write_end              = ext3_ordered_write_end,
1961         .bmap                   = ext3_bmap,
1962         .invalidatepage         = ext3_invalidatepage,
1963         .releasepage            = ext3_releasepage,
1964         .direct_IO              = ext3_direct_IO,
1965         .migratepage            = buffer_migrate_page,
1966         .is_partially_uptodate  = block_is_partially_uptodate,
1967         .error_remove_page      = generic_error_remove_page,
1968 };
1969
1970 static const struct address_space_operations ext3_writeback_aops = {
1971         .readpage               = ext3_readpage,
1972         .readpages              = ext3_readpages,
1973         .writepage              = ext3_writeback_writepage,
1974         .write_begin            = ext3_write_begin,
1975         .write_end              = ext3_writeback_write_end,
1976         .bmap                   = ext3_bmap,
1977         .invalidatepage         = ext3_invalidatepage,
1978         .releasepage            = ext3_releasepage,
1979         .direct_IO              = ext3_direct_IO,
1980         .migratepage            = buffer_migrate_page,
1981         .is_partially_uptodate  = block_is_partially_uptodate,
1982         .error_remove_page      = generic_error_remove_page,
1983 };
1984
1985 static const struct address_space_operations ext3_journalled_aops = {
1986         .readpage               = ext3_readpage,
1987         .readpages              = ext3_readpages,
1988         .writepage              = ext3_journalled_writepage,
1989         .write_begin            = ext3_write_begin,
1990         .write_end              = ext3_journalled_write_end,
1991         .set_page_dirty         = ext3_journalled_set_page_dirty,
1992         .bmap                   = ext3_bmap,
1993         .invalidatepage         = ext3_invalidatepage,
1994         .releasepage            = ext3_releasepage,
1995         .is_partially_uptodate  = block_is_partially_uptodate,
1996         .error_remove_page      = generic_error_remove_page,
1997 };
1998
1999 void ext3_set_aops(struct inode *inode)
2000 {
2001         if (ext3_should_order_data(inode))
2002                 inode->i_mapping->a_ops = &ext3_ordered_aops;
2003         else if (ext3_should_writeback_data(inode))
2004                 inode->i_mapping->a_ops = &ext3_writeback_aops;
2005         else
2006                 inode->i_mapping->a_ops = &ext3_journalled_aops;
2007 }
2008
2009 /*
2010  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
2011  * up to the end of the block which corresponds to `from'.
2012  * This required during truncate. We need to physically zero the tail end
2013  * of that block so it doesn't yield old data if the file is later grown.
2014  */
2015 static int ext3_block_truncate_page(struct inode *inode, loff_t from)
2016 {
2017         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
2018         unsigned offset = from & (PAGE_CACHE_SIZE - 1);
2019         unsigned blocksize, iblock, length, pos;
2020         struct page *page;
2021         handle_t *handle = NULL;
2022         struct buffer_head *bh;
2023         int err = 0;
2024
2025         /* Truncated on block boundary - nothing to do */
2026         blocksize = inode->i_sb->s_blocksize;
2027         if ((from & (blocksize - 1)) == 0)
2028                 return 0;
2029
2030         page = grab_cache_page(inode->i_mapping, index);
2031         if (!page)
2032                 return -ENOMEM;
2033         length = blocksize - (offset & (blocksize - 1));
2034         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
2035
2036         if (!page_has_buffers(page))
2037                 create_empty_buffers(page, blocksize, 0);
2038
2039         /* Find the buffer that contains "offset" */
2040         bh = page_buffers(page);
2041         pos = blocksize;
2042         while (offset >= pos) {
2043                 bh = bh->b_this_page;
2044                 iblock++;
2045                 pos += blocksize;
2046         }
2047
2048         err = 0;
2049         if (buffer_freed(bh)) {
2050                 BUFFER_TRACE(bh, "freed: skip");
2051                 goto unlock;
2052         }
2053
2054         if (!buffer_mapped(bh)) {
2055                 BUFFER_TRACE(bh, "unmapped");
2056                 ext3_get_block(inode, iblock, bh, 0);
2057                 /* unmapped? It's a hole - nothing to do */
2058                 if (!buffer_mapped(bh)) {
2059                         BUFFER_TRACE(bh, "still unmapped");
2060                         goto unlock;
2061                 }
2062         }
2063
2064         /* Ok, it's mapped. Make sure it's up-to-date */
2065         if (PageUptodate(page))
2066                 set_buffer_uptodate(bh);
2067
2068         if (!buffer_uptodate(bh)) {
2069                 err = -EIO;
2070                 ll_rw_block(READ, 1, &bh);
2071                 wait_on_buffer(bh);
2072                 /* Uhhuh. Read error. Complain and punt. */
2073                 if (!buffer_uptodate(bh))
2074                         goto unlock;
2075         }
2076
2077         /* data=writeback mode doesn't need transaction to zero-out data */
2078         if (!ext3_should_writeback_data(inode)) {
2079                 /* We journal at most one block */
2080                 handle = ext3_journal_start(inode, 1);
2081                 if (IS_ERR(handle)) {
2082                         clear_highpage(page);
2083                         flush_dcache_page(page);
2084                         err = PTR_ERR(handle);
2085                         goto unlock;
2086                 }
2087         }
2088
2089         if (ext3_should_journal_data(inode)) {
2090                 BUFFER_TRACE(bh, "get write access");
2091                 err = ext3_journal_get_write_access(handle, bh);
2092                 if (err)
2093                         goto stop;
2094         }
2095
2096         zero_user(page, offset, length);
2097         BUFFER_TRACE(bh, "zeroed end of block");
2098
2099         err = 0;
2100         if (ext3_should_journal_data(inode)) {
2101                 err = ext3_journal_dirty_metadata(handle, bh);
2102         } else {
2103                 if (ext3_should_order_data(inode))
2104                         err = ext3_journal_dirty_data(handle, bh);
2105                 mark_buffer_dirty(bh);
2106         }
2107 stop:
2108         if (handle)
2109                 ext3_journal_stop(handle);
2110
2111 unlock:
2112         unlock_page(page);
2113         page_cache_release(page);
2114         return err;
2115 }
2116
2117 /*
2118  * Probably it should be a library function... search for first non-zero word
2119  * or memcmp with zero_page, whatever is better for particular architecture.
2120  * Linus?
2121  */
2122 static inline int all_zeroes(__le32 *p, __le32 *q)
2123 {
2124         while (p < q)
2125                 if (*p++)
2126                         return 0;
2127         return 1;
2128 }
2129
2130 /**
2131  *      ext3_find_shared - find the indirect blocks for partial truncation.
2132  *      @inode:   inode in question
2133  *      @depth:   depth of the affected branch
2134  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
2135  *      @chain:   place to store the pointers to partial indirect blocks
2136  *      @top:     place to the (detached) top of branch
2137  *
2138  *      This is a helper function used by ext3_truncate().
2139  *
2140  *      When we do truncate() we may have to clean the ends of several
2141  *      indirect blocks but leave the blocks themselves alive. Block is
2142  *      partially truncated if some data below the new i_size is referred
2143  *      from it (and it is on the path to the first completely truncated
2144  *      data block, indeed).  We have to free the top of that path along
2145  *      with everything to the right of the path. Since no allocation
2146  *      past the truncation point is possible until ext3_truncate()
2147  *      finishes, we may safely do the latter, but top of branch may
2148  *      require special attention - pageout below the truncation point
2149  *      might try to populate it.
2150  *
2151  *      We atomically detach the top of branch from the tree, store the
2152  *      block number of its root in *@top, pointers to buffer_heads of
2153  *      partially truncated blocks - in @chain[].bh and pointers to
2154  *      their last elements that should not be removed - in
2155  *      @chain[].p. Return value is the pointer to last filled element
2156  *      of @chain.
2157  *
2158  *      The work left to caller to do the actual freeing of subtrees:
2159  *              a) free the subtree starting from *@top
2160  *              b) free the subtrees whose roots are stored in
2161  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2162  *              c) free the subtrees growing from the inode past the @chain[0].
2163  *                      (no partially truncated stuff there).  */
2164
2165 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2166                         int offsets[4], Indirect chain[4], __le32 *top)
2167 {
2168         Indirect *partial, *p;
2169         int k, err;
2170
2171         *top = 0;
2172         /* Make k index the deepest non-null offset + 1 */
2173         for (k = depth; k > 1 && !offsets[k-1]; k--)
2174                 ;
2175         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2176         /* Writer: pointers */
2177         if (!partial)
2178                 partial = chain + k-1;
2179         /*
2180          * If the branch acquired continuation since we've looked at it -
2181          * fine, it should all survive and (new) top doesn't belong to us.
2182          */
2183         if (!partial->key && *partial->p)
2184                 /* Writer: end */
2185                 goto no_top;
2186         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2187                 ;
2188         /*
2189          * OK, we've found the last block that must survive. The rest of our
2190          * branch should be detached before unlocking. However, if that rest
2191          * of branch is all ours and does not grow immediately from the inode
2192          * it's easier to cheat and just decrement partial->p.
2193          */
2194         if (p == chain + k - 1 && p > chain) {
2195                 p->p--;
2196         } else {
2197                 *top = *p->p;
2198                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2199 #if 0
2200                 *p->p = 0;
2201 #endif
2202         }
2203         /* Writer: end */
2204
2205         while(partial > p) {
2206                 brelse(partial->bh);
2207                 partial--;
2208         }
2209 no_top:
2210         return partial;
2211 }
2212
2213 /*
2214  * Zero a number of block pointers in either an inode or an indirect block.
2215  * If we restart the transaction we must again get write access to the
2216  * indirect block for further modification.
2217  *
2218  * We release `count' blocks on disk, but (last - first) may be greater
2219  * than `count' because there can be holes in there.
2220  */
2221 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2222                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2223                 unsigned long count, __le32 *first, __le32 *last)
2224 {
2225         __le32 *p;
2226         if (try_to_extend_transaction(handle, inode)) {
2227                 if (bh) {
2228                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2229                         if (ext3_journal_dirty_metadata(handle, bh))
2230                                 return;
2231                 }
2232                 ext3_mark_inode_dirty(handle, inode);
2233                 truncate_restart_transaction(handle, inode);
2234                 if (bh) {
2235                         BUFFER_TRACE(bh, "retaking write access");
2236                         if (ext3_journal_get_write_access(handle, bh))
2237                                 return;
2238                 }
2239         }
2240
2241         /*
2242          * Any buffers which are on the journal will be in memory. We find
2243          * them on the hash table so journal_revoke() will run journal_forget()
2244          * on them.  We've already detached each block from the file, so
2245          * bforget() in journal_forget() should be safe.
2246          *
2247          * AKPM: turn on bforget in journal_forget()!!!
2248          */
2249         for (p = first; p < last; p++) {
2250                 u32 nr = le32_to_cpu(*p);
2251                 if (nr) {
2252                         struct buffer_head *bh;
2253
2254                         *p = 0;
2255                         bh = sb_find_get_block(inode->i_sb, nr);
2256                         ext3_forget(handle, 0, inode, bh, nr);
2257                 }
2258         }
2259
2260         ext3_free_blocks(handle, inode, block_to_free, count);
2261 }
2262
2263 /**
2264  * ext3_free_data - free a list of data blocks
2265  * @handle:     handle for this transaction
2266  * @inode:      inode we are dealing with
2267  * @this_bh:    indirect buffer_head which contains *@first and *@last
2268  * @first:      array of block numbers
2269  * @last:       points immediately past the end of array
2270  *
2271  * We are freeing all blocks referred from that array (numbers are stored as
2272  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2273  *
2274  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2275  * blocks are contiguous then releasing them at one time will only affect one
2276  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2277  * actually use a lot of journal space.
2278  *
2279  * @this_bh will be %NULL if @first and @last point into the inode's direct
2280  * block pointers.
2281  */
2282 static void ext3_free_data(handle_t *handle, struct inode *inode,
2283                            struct buffer_head *this_bh,
2284                            __le32 *first, __le32 *last)
2285 {
2286         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2287         unsigned long count = 0;            /* Number of blocks in the run */
2288         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2289                                                corresponding to
2290                                                block_to_free */
2291         ext3_fsblk_t nr;                    /* Current block # */
2292         __le32 *p;                          /* Pointer into inode/ind
2293                                                for current block */
2294         int err;
2295
2296         if (this_bh) {                          /* For indirect block */
2297                 BUFFER_TRACE(this_bh, "get_write_access");
2298                 err = ext3_journal_get_write_access(handle, this_bh);
2299                 /* Important: if we can't update the indirect pointers
2300                  * to the blocks, we can't free them. */
2301                 if (err)
2302                         return;
2303         }
2304
2305         for (p = first; p < last; p++) {
2306                 nr = le32_to_cpu(*p);
2307                 if (nr) {
2308                         /* accumulate blocks to free if they're contiguous */
2309                         if (count == 0) {
2310                                 block_to_free = nr;
2311                                 block_to_free_p = p;
2312                                 count = 1;
2313                         } else if (nr == block_to_free + count) {
2314                                 count++;
2315                         } else {
2316                                 ext3_clear_blocks(handle, inode, this_bh,
2317                                                   block_to_free,
2318                                                   count, block_to_free_p, p);
2319                                 block_to_free = nr;
2320                                 block_to_free_p = p;
2321                                 count = 1;
2322                         }
2323                 }
2324         }
2325
2326         if (count > 0)
2327                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2328                                   count, block_to_free_p, p);
2329
2330         if (this_bh) {
2331                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2332
2333                 /*
2334                  * The buffer head should have an attached journal head at this
2335                  * point. However, if the data is corrupted and an indirect
2336                  * block pointed to itself, it would have been detached when
2337                  * the block was cleared. Check for this instead of OOPSing.
2338                  */
2339                 if (bh2jh(this_bh))
2340                         ext3_journal_dirty_metadata(handle, this_bh);
2341                 else
2342                         ext3_error(inode->i_sb, "ext3_free_data",
2343                                    "circular indirect block detected, "
2344                                    "inode=%lu, block=%llu",
2345                                    inode->i_ino,
2346                                    (unsigned long long)this_bh->b_blocknr);
2347         }
2348 }
2349
2350 /**
2351  *      ext3_free_branches - free an array of branches
2352  *      @handle: JBD handle for this transaction
2353  *      @inode: inode we are dealing with
2354  *      @parent_bh: the buffer_head which contains *@first and *@last
2355  *      @first: array of block numbers
2356  *      @last:  pointer immediately past the end of array
2357  *      @depth: depth of the branches to free
2358  *
2359  *      We are freeing all blocks referred from these branches (numbers are
2360  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2361  *      appropriately.
2362  */
2363 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2364                                struct buffer_head *parent_bh,
2365                                __le32 *first, __le32 *last, int depth)
2366 {
2367         ext3_fsblk_t nr;
2368         __le32 *p;
2369
2370         if (is_handle_aborted(handle))
2371                 return;
2372
2373         if (depth--) {
2374                 struct buffer_head *bh;
2375                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2376                 p = last;
2377                 while (--p >= first) {
2378                         nr = le32_to_cpu(*p);
2379                         if (!nr)
2380                                 continue;               /* A hole */
2381
2382                         /* Go read the buffer for the next level down */
2383                         bh = sb_bread(inode->i_sb, nr);
2384
2385                         /*
2386                          * A read failure? Report error and clear slot
2387                          * (should be rare).
2388                          */
2389                         if (!bh) {
2390                                 ext3_error(inode->i_sb, "ext3_free_branches",
2391                                            "Read failure, inode=%lu, block="E3FSBLK,
2392                                            inode->i_ino, nr);
2393                                 continue;
2394                         }
2395
2396                         /* This zaps the entire block.  Bottom up. */
2397                         BUFFER_TRACE(bh, "free child branches");
2398                         ext3_free_branches(handle, inode, bh,
2399                                            (__le32*)bh->b_data,
2400                                            (__le32*)bh->b_data + addr_per_block,
2401                                            depth);
2402
2403                         /*
2404                          * Everything below this this pointer has been
2405                          * released.  Now let this top-of-subtree go.
2406                          *
2407                          * We want the freeing of this indirect block to be
2408                          * atomic in the journal with the updating of the
2409                          * bitmap block which owns it.  So make some room in
2410                          * the journal.
2411                          *
2412                          * We zero the parent pointer *after* freeing its
2413                          * pointee in the bitmaps, so if extend_transaction()
2414                          * for some reason fails to put the bitmap changes and
2415                          * the release into the same transaction, recovery
2416                          * will merely complain about releasing a free block,
2417                          * rather than leaking blocks.
2418                          */
2419                         if (is_handle_aborted(handle))
2420                                 return;
2421                         if (try_to_extend_transaction(handle, inode)) {
2422                                 ext3_mark_inode_dirty(handle, inode);
2423                                 truncate_restart_transaction(handle, inode);
2424                         }
2425
2426                         /*
2427                          * We've probably journalled the indirect block several
2428                          * times during the truncate.  But it's no longer
2429                          * needed and we now drop it from the transaction via
2430                          * journal_revoke().
2431                          *
2432                          * That's easy if it's exclusively part of this
2433                          * transaction.  But if it's part of the committing
2434                          * transaction then journal_forget() will simply
2435                          * brelse() it.  That means that if the underlying
2436                          * block is reallocated in ext3_get_block(),
2437                          * unmap_underlying_metadata() will find this block
2438                          * and will try to get rid of it.  damn, damn. Thus
2439                          * we don't allow a block to be reallocated until
2440                          * a transaction freeing it has fully committed.
2441                          *
2442                          * We also have to make sure journal replay after a
2443                          * crash does not overwrite non-journaled data blocks
2444                          * with old metadata when the block got reallocated for
2445                          * data.  Thus we have to store a revoke record for a
2446                          * block in the same transaction in which we free the
2447                          * block.
2448                          */
2449                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2450
2451                         ext3_free_blocks(handle, inode, nr, 1);
2452
2453                         if (parent_bh) {
2454                                 /*
2455                                  * The block which we have just freed is
2456                                  * pointed to by an indirect block: journal it
2457                                  */
2458                                 BUFFER_TRACE(parent_bh, "get_write_access");
2459                                 if (!ext3_journal_get_write_access(handle,
2460                                                                    parent_bh)){
2461                                         *p = 0;
2462                                         BUFFER_TRACE(parent_bh,
2463                                         "call ext3_journal_dirty_metadata");
2464                                         ext3_journal_dirty_metadata(handle,
2465                                                                     parent_bh);
2466                                 }
2467                         }
2468                 }
2469         } else {
2470                 /* We have reached the bottom of the tree. */
2471                 BUFFER_TRACE(parent_bh, "free data blocks");
2472                 ext3_free_data(handle, inode, parent_bh, first, last);
2473         }
2474 }
2475
2476 int ext3_can_truncate(struct inode *inode)
2477 {
2478         if (S_ISREG(inode->i_mode))
2479                 return 1;
2480         if (S_ISDIR(inode->i_mode))
2481                 return 1;
2482         if (S_ISLNK(inode->i_mode))
2483                 return !ext3_inode_is_fast_symlink(inode);
2484         return 0;
2485 }
2486
2487 /*
2488  * ext3_truncate()
2489  *
2490  * We block out ext3_get_block() block instantiations across the entire
2491  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2492  * simultaneously on behalf of the same inode.
2493  *
2494  * As we work through the truncate and commmit bits of it to the journal there
2495  * is one core, guiding principle: the file's tree must always be consistent on
2496  * disk.  We must be able to restart the truncate after a crash.
2497  *
2498  * The file's tree may be transiently inconsistent in memory (although it
2499  * probably isn't), but whenever we close off and commit a journal transaction,
2500  * the contents of (the filesystem + the journal) must be consistent and
2501  * restartable.  It's pretty simple, really: bottom up, right to left (although
2502  * left-to-right works OK too).
2503  *
2504  * Note that at recovery time, journal replay occurs *before* the restart of
2505  * truncate against the orphan inode list.
2506  *
2507  * The committed inode has the new, desired i_size (which is the same as
2508  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2509  * that this inode's truncate did not complete and it will again call
2510  * ext3_truncate() to have another go.  So there will be instantiated blocks
2511  * to the right of the truncation point in a crashed ext3 filesystem.  But
2512  * that's fine - as long as they are linked from the inode, the post-crash
2513  * ext3_truncate() run will find them and release them.
2514  */
2515 void ext3_truncate(struct inode *inode)
2516 {
2517         handle_t *handle;
2518         struct ext3_inode_info *ei = EXT3_I(inode);
2519         __le32 *i_data = ei->i_data;
2520         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2521         int offsets[4];
2522         Indirect chain[4];
2523         Indirect *partial;
2524         __le32 nr = 0;
2525         int n;
2526         long last_block;
2527         unsigned blocksize = inode->i_sb->s_blocksize;
2528
2529         trace_ext3_truncate_enter(inode);
2530
2531         if (!ext3_can_truncate(inode))
2532                 goto out_notrans;
2533
2534         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2535                 ext3_set_inode_state(inode, EXT3_STATE_FLUSH_ON_CLOSE);
2536
2537         handle = start_transaction(inode);
2538         if (IS_ERR(handle))
2539                 goto out_notrans;
2540
2541         last_block = (inode->i_size + blocksize-1)
2542                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2543         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2544         if (n == 0)
2545                 goto out_stop;  /* error */
2546
2547         /*
2548          * OK.  This truncate is going to happen.  We add the inode to the
2549          * orphan list, so that if this truncate spans multiple transactions,
2550          * and we crash, we will resume the truncate when the filesystem
2551          * recovers.  It also marks the inode dirty, to catch the new size.
2552          *
2553          * Implication: the file must always be in a sane, consistent
2554          * truncatable state while each transaction commits.
2555          */
2556         if (ext3_orphan_add(handle, inode))
2557                 goto out_stop;
2558
2559         /*
2560          * The orphan list entry will now protect us from any crash which
2561          * occurs before the truncate completes, so it is now safe to propagate
2562          * the new, shorter inode size (held for now in i_size) into the
2563          * on-disk inode. We do this via i_disksize, which is the value which
2564          * ext3 *really* writes onto the disk inode.
2565          */
2566         ei->i_disksize = inode->i_size;
2567
2568         /*
2569          * From here we block out all ext3_get_block() callers who want to
2570          * modify the block allocation tree.
2571          */
2572         mutex_lock(&ei->truncate_mutex);
2573
2574         if (n == 1) {           /* direct blocks */
2575                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2576                                i_data + EXT3_NDIR_BLOCKS);
2577                 goto do_indirects;
2578         }
2579
2580         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2581         /* Kill the top of shared branch (not detached) */
2582         if (nr) {
2583                 if (partial == chain) {
2584                         /* Shared branch grows from the inode */
2585                         ext3_free_branches(handle, inode, NULL,
2586                                            &nr, &nr+1, (chain+n-1) - partial);
2587                         *partial->p = 0;
2588                         /*
2589                          * We mark the inode dirty prior to restart,
2590                          * and prior to stop.  No need for it here.
2591                          */
2592                 } else {
2593                         /* Shared branch grows from an indirect block */
2594                         ext3_free_branches(handle, inode, partial->bh,
2595                                         partial->p,
2596                                         partial->p+1, (chain+n-1) - partial);
2597                 }
2598         }
2599         /* Clear the ends of indirect blocks on the shared branch */
2600         while (partial > chain) {
2601                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2602                                    (__le32*)partial->bh->b_data+addr_per_block,
2603                                    (chain+n-1) - partial);
2604                 BUFFER_TRACE(partial->bh, "call brelse");
2605                 brelse (partial->bh);
2606                 partial--;
2607         }
2608 do_indirects:
2609         /* Kill the remaining (whole) subtrees */
2610         switch (offsets[0]) {
2611         default:
2612                 nr = i_data[EXT3_IND_BLOCK];
2613                 if (nr) {
2614                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2615                         i_data[EXT3_IND_BLOCK] = 0;
2616                 }
2617         case EXT3_IND_BLOCK:
2618                 nr = i_data[EXT3_DIND_BLOCK];
2619                 if (nr) {
2620                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2621                         i_data[EXT3_DIND_BLOCK] = 0;
2622                 }
2623         case EXT3_DIND_BLOCK:
2624                 nr = i_data[EXT3_TIND_BLOCK];
2625                 if (nr) {
2626                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2627                         i_data[EXT3_TIND_BLOCK] = 0;
2628                 }
2629         case EXT3_TIND_BLOCK:
2630                 ;
2631         }
2632
2633         ext3_discard_reservation(inode);
2634
2635         mutex_unlock(&ei->truncate_mutex);
2636         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2637         ext3_mark_inode_dirty(handle, inode);
2638
2639         /*
2640          * In a multi-transaction truncate, we only make the final transaction
2641          * synchronous
2642          */
2643         if (IS_SYNC(inode))
2644                 handle->h_sync = 1;
2645 out_stop:
2646         /*
2647          * If this was a simple ftruncate(), and the file will remain alive
2648          * then we need to clear up the orphan record which we created above.
2649          * However, if this was a real unlink then we were called by
2650          * ext3_evict_inode(), and we allow that function to clean up the
2651          * orphan info for us.
2652          */
2653         if (inode->i_nlink)
2654                 ext3_orphan_del(handle, inode);
2655
2656         ext3_journal_stop(handle);
2657         trace_ext3_truncate_exit(inode);
2658         return;
2659 out_notrans:
2660         /*
2661          * Delete the inode from orphan list so that it doesn't stay there
2662          * forever and trigger assertion on umount.
2663          */
2664         if (inode->i_nlink)
2665                 ext3_orphan_del(NULL, inode);
2666         trace_ext3_truncate_exit(inode);
2667 }
2668
2669 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2670                 unsigned long ino, struct ext3_iloc *iloc)
2671 {
2672         unsigned long block_group;
2673         unsigned long offset;
2674         ext3_fsblk_t block;
2675         struct ext3_group_desc *gdp;
2676
2677         if (!ext3_valid_inum(sb, ino)) {
2678                 /*
2679                  * This error is already checked for in namei.c unless we are
2680                  * looking at an NFS filehandle, in which case no error
2681                  * report is needed
2682                  */
2683                 return 0;
2684         }
2685
2686         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2687         gdp = ext3_get_group_desc(sb, block_group, NULL);
2688         if (!gdp)
2689                 return 0;
2690         /*
2691          * Figure out the offset within the block group inode table
2692          */
2693         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2694                 EXT3_INODE_SIZE(sb);
2695         block = le32_to_cpu(gdp->bg_inode_table) +
2696                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2697
2698         iloc->block_group = block_group;
2699         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2700         return block;
2701 }
2702
2703 /*
2704  * ext3_get_inode_loc returns with an extra refcount against the inode's
2705  * underlying buffer_head on success. If 'in_mem' is true, we have all
2706  * data in memory that is needed to recreate the on-disk version of this
2707  * inode.
2708  */
2709 static int __ext3_get_inode_loc(struct inode *inode,
2710                                 struct ext3_iloc *iloc, int in_mem)
2711 {
2712         ext3_fsblk_t block;
2713         struct buffer_head *bh;
2714
2715         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2716         if (!block)
2717                 return -EIO;
2718
2719         bh = sb_getblk(inode->i_sb, block);
2720         if (!bh) {
2721                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2722                                 "unable to read inode block - "
2723                                 "inode=%lu, block="E3FSBLK,
2724                                  inode->i_ino, block);
2725                 return -EIO;
2726         }
2727         if (!buffer_uptodate(bh)) {
2728                 lock_buffer(bh);
2729
2730                 /*
2731                  * If the buffer has the write error flag, we have failed
2732                  * to write out another inode in the same block.  In this
2733                  * case, we don't have to read the block because we may
2734                  * read the old inode data successfully.
2735                  */
2736                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2737                         set_buffer_uptodate(bh);
2738
2739                 if (buffer_uptodate(bh)) {
2740                         /* someone brought it uptodate while we waited */
2741                         unlock_buffer(bh);
2742                         goto has_buffer;
2743                 }
2744
2745                 /*
2746                  * If we have all information of the inode in memory and this
2747                  * is the only valid inode in the block, we need not read the
2748                  * block.
2749                  */
2750                 if (in_mem) {
2751                         struct buffer_head *bitmap_bh;
2752                         struct ext3_group_desc *desc;
2753                         int inodes_per_buffer;
2754                         int inode_offset, i;
2755                         int block_group;
2756                         int start;
2757
2758                         block_group = (inode->i_ino - 1) /
2759                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2760                         inodes_per_buffer = bh->b_size /
2761                                 EXT3_INODE_SIZE(inode->i_sb);
2762                         inode_offset = ((inode->i_ino - 1) %
2763                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2764                         start = inode_offset & ~(inodes_per_buffer - 1);
2765
2766                         /* Is the inode bitmap in cache? */
2767                         desc = ext3_get_group_desc(inode->i_sb,
2768                                                 block_group, NULL);
2769                         if (!desc)
2770                                 goto make_io;
2771
2772                         bitmap_bh = sb_getblk(inode->i_sb,
2773                                         le32_to_cpu(desc->bg_inode_bitmap));
2774                         if (!bitmap_bh)
2775                                 goto make_io;
2776
2777                         /*
2778                          * If the inode bitmap isn't in cache then the
2779                          * optimisation may end up performing two reads instead
2780                          * of one, so skip it.
2781                          */
2782                         if (!buffer_uptodate(bitmap_bh)) {
2783                                 brelse(bitmap_bh);
2784                                 goto make_io;
2785                         }
2786                         for (i = start; i < start + inodes_per_buffer; i++) {
2787                                 if (i == inode_offset)
2788                                         continue;
2789                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2790                                         break;
2791                         }
2792                         brelse(bitmap_bh);
2793                         if (i == start + inodes_per_buffer) {
2794                                 /* all other inodes are free, so skip I/O */
2795                                 memset(bh->b_data, 0, bh->b_size);
2796                                 set_buffer_uptodate(bh);
2797                                 unlock_buffer(bh);
2798                                 goto has_buffer;
2799                         }
2800                 }
2801
2802 make_io:
2803                 /*
2804                  * There are other valid inodes in the buffer, this inode
2805                  * has in-inode xattrs, or we don't have this inode in memory.
2806                  * Read the block from disk.
2807                  */
2808                 trace_ext3_load_inode(inode);
2809                 get_bh(bh);
2810                 bh->b_end_io = end_buffer_read_sync;
2811                 submit_bh(READ_META, bh);
2812                 wait_on_buffer(bh);
2813                 if (!buffer_uptodate(bh)) {
2814                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2815                                         "unable to read inode block - "
2816                                         "inode=%lu, block="E3FSBLK,
2817                                         inode->i_ino, block);
2818                         brelse(bh);
2819                         return -EIO;
2820                 }
2821         }
2822 has_buffer:
2823         iloc->bh = bh;
2824         return 0;
2825 }
2826
2827 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2828 {
2829         /* We have all inode data except xattrs in memory here. */
2830         return __ext3_get_inode_loc(inode, iloc,
2831                 !ext3_test_inode_state(inode, EXT3_STATE_XATTR));
2832 }
2833
2834 void ext3_set_inode_flags(struct inode *inode)
2835 {
2836         unsigned int flags = EXT3_I(inode)->i_flags;
2837
2838         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2839         if (flags & EXT3_SYNC_FL)
2840                 inode->i_flags |= S_SYNC;
2841         if (flags & EXT3_APPEND_FL)
2842                 inode->i_flags |= S_APPEND;
2843         if (flags & EXT3_IMMUTABLE_FL)
2844                 inode->i_flags |= S_IMMUTABLE;
2845         if (flags & EXT3_NOATIME_FL)
2846                 inode->i_flags |= S_NOATIME;
2847         if (flags & EXT3_DIRSYNC_FL)
2848                 inode->i_flags |= S_DIRSYNC;
2849 }
2850
2851 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2852 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2853 {
2854         unsigned int flags = ei->vfs_inode.i_flags;
2855
2856         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2857                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2858         if (flags & S_SYNC)
2859                 ei->i_flags |= EXT3_SYNC_FL;
2860         if (flags & S_APPEND)
2861                 ei->i_flags |= EXT3_APPEND_FL;
2862         if (flags & S_IMMUTABLE)
2863                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2864         if (flags & S_NOATIME)
2865                 ei->i_flags |= EXT3_NOATIME_FL;
2866         if (flags & S_DIRSYNC)
2867                 ei->i_flags |= EXT3_DIRSYNC_FL;
2868 }
2869
2870 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2871 {
2872         struct ext3_iloc iloc;
2873         struct ext3_inode *raw_inode;
2874         struct ext3_inode_info *ei;
2875         struct buffer_head *bh;
2876         struct inode *inode;
2877         journal_t *journal = EXT3_SB(sb)->s_journal;
2878         transaction_t *transaction;
2879         long ret;
2880         int block;
2881
2882         inode = iget_locked(sb, ino);
2883         if (!inode)
2884                 return ERR_PTR(-ENOMEM);
2885         if (!(inode->i_state & I_NEW))
2886                 return inode;
2887
2888         ei = EXT3_I(inode);
2889         ei->i_block_alloc_info = NULL;
2890
2891         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2892         if (ret < 0)
2893                 goto bad_inode;
2894         bh = iloc.bh;
2895         raw_inode = ext3_raw_inode(&iloc);
2896         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2897         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2898         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2899         if(!(test_opt (inode->i_sb, NO_UID32))) {
2900                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2901                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2902         }
2903         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2904         inode->i_size = le32_to_cpu(raw_inode->i_size);
2905         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2906         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2907         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2908         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2909
2910         ei->i_state_flags = 0;
2911         ei->i_dir_start_lookup = 0;
2912         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2913         /* We now have enough fields to check if the inode was active or not.
2914          * This is needed because nfsd might try to access dead inodes
2915          * the test is that same one that e2fsck uses
2916          * NeilBrown 1999oct15
2917          */
2918         if (inode->i_nlink == 0) {
2919                 if (inode->i_mode == 0 ||
2920                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2921                         /* this inode is deleted */
2922                         brelse (bh);
2923                         ret = -ESTALE;
2924                         goto bad_inode;
2925                 }
2926                 /* The only unlinked inodes we let through here have
2927                  * valid i_mode and are being read by the orphan
2928                  * recovery code: that's fine, we're about to complete
2929                  * the process of deleting those. */
2930         }
2931         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2932         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2933 #ifdef EXT3_FRAGMENTS
2934         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2935         ei->i_frag_no = raw_inode->i_frag;
2936         ei->i_frag_size = raw_inode->i_fsize;
2937 #endif
2938         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2939         if (!S_ISREG(inode->i_mode)) {
2940                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2941         } else {
2942                 inode->i_size |=
2943                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2944         }
2945         ei->i_disksize = inode->i_size;
2946         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2947         ei->i_block_group = iloc.block_group;
2948         /*
2949          * NOTE! The in-memory inode i_data array is in little-endian order
2950          * even on big-endian machines: we do NOT byteswap the block numbers!
2951          */
2952         for (block = 0; block < EXT3_N_BLOCKS; block++)
2953                 ei->i_data[block] = raw_inode->i_block[block];
2954         INIT_LIST_HEAD(&ei->i_orphan);
2955
2956         /*
2957          * Set transaction id's of transactions that have to be committed
2958          * to finish f[data]sync. We set them to currently running transaction
2959          * as we cannot be sure that the inode or some of its metadata isn't
2960          * part of the transaction - the inode could have been reclaimed and
2961          * now it is reread from disk.
2962          */
2963         if (journal) {
2964                 tid_t tid;
2965
2966                 spin_lock(&journal->j_state_lock);
2967                 if (journal->j_running_transaction)
2968                         transaction = journal->j_running_transaction;
2969                 else
2970                         transaction = journal->j_committing_transaction;
2971                 if (transaction)
2972                         tid = transaction->t_tid;
2973                 else
2974                         tid = journal->j_commit_sequence;
2975                 spin_unlock(&journal->j_state_lock);
2976                 atomic_set(&ei->i_sync_tid, tid);
2977                 atomic_set(&ei->i_datasync_tid, tid);
2978         }
2979
2980         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2981             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2982                 /*
2983                  * When mke2fs creates big inodes it does not zero out
2984                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2985                  * so ignore those first few inodes.
2986                  */
2987                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2988                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2989                     EXT3_INODE_SIZE(inode->i_sb)) {
2990                         brelse (bh);
2991                         ret = -EIO;
2992                         goto bad_inode;
2993                 }
2994                 if (ei->i_extra_isize == 0) {
2995                         /* The extra space is currently unused. Use it. */
2996                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2997                                             EXT3_GOOD_OLD_INODE_SIZE;
2998                 } else {
2999                         __le32 *magic = (void *)raw_inode +
3000                                         EXT3_GOOD_OLD_INODE_SIZE +
3001                                         ei->i_extra_isize;
3002                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
3003                                  ext3_set_inode_state(inode, EXT3_STATE_XATTR);
3004                 }
3005         } else
3006                 ei->i_extra_isize = 0;
3007
3008         if (S_ISREG(inode->i_mode)) {
3009                 inode->i_op = &ext3_file_inode_operations;
3010                 inode->i_fop = &ext3_file_operations;
3011                 ext3_set_aops(inode);
3012         } else if (S_ISDIR(inode->i_mode)) {
3013                 inode->i_op = &ext3_dir_inode_operations;
3014                 inode->i_fop = &ext3_dir_operations;
3015         } else if (S_ISLNK(inode->i_mode)) {
3016                 if (ext3_inode_is_fast_symlink(inode)) {
3017                         inode->i_op = &ext3_fast_symlink_inode_operations;
3018                         nd_terminate_link(ei->i_data, inode->i_size,
3019                                 sizeof(ei->i_data) - 1);
3020                 } else {
3021                         inode->i_op = &ext3_symlink_inode_operations;
3022                         ext3_set_aops(inode);
3023                 }
3024         } else {
3025                 inode->i_op = &ext3_special_inode_operations;
3026                 if (raw_inode->i_block[0])
3027                         init_special_inode(inode, inode->i_mode,
3028                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3029                 else
3030                         init_special_inode(inode, inode->i_mode,
3031                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3032         }
3033         brelse (iloc.bh);
3034         ext3_set_inode_flags(inode);
3035         unlock_new_inode(inode);
3036         return inode;
3037
3038 bad_inode:
3039         iget_failed(inode);
3040         return ERR_PTR(ret);
3041 }
3042
3043 /*
3044  * Post the struct inode info into an on-disk inode location in the
3045  * buffer-cache.  This gobbles the caller's reference to the
3046  * buffer_head in the inode location struct.
3047  *
3048  * The caller must have write access to iloc->bh.
3049  */
3050 static int ext3_do_update_inode(handle_t *handle,
3051                                 struct inode *inode,
3052                                 struct ext3_iloc *iloc)
3053 {
3054         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
3055         struct ext3_inode_info *ei = EXT3_I(inode);
3056         struct buffer_head *bh = iloc->bh;
3057         int err = 0, rc, block;
3058
3059 again:
3060         /* we can't allow multiple procs in here at once, its a bit racey */
3061         lock_buffer(bh);
3062
3063         /* For fields not not tracking in the in-memory inode,
3064          * initialise them to zero for new inodes. */
3065         if (ext3_test_inode_state(inode, EXT3_STATE_NEW))
3066                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
3067
3068         ext3_get_inode_flags(ei);
3069         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3070         if(!(test_opt(inode->i_sb, NO_UID32))) {
3071                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3072                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3073 /*
3074  * Fix up interoperability with old kernels. Otherwise, old inodes get
3075  * re-used with the upper 16 bits of the uid/gid intact
3076  */
3077                 if(!ei->i_dtime) {
3078                         raw_inode->i_uid_high =
3079                                 cpu_to_le16(high_16_bits(inode->i_uid));
3080                         raw_inode->i_gid_high =
3081                                 cpu_to_le16(high_16_bits(inode->i_gid));
3082                 } else {
3083                         raw_inode->i_uid_high = 0;
3084                         raw_inode->i_gid_high = 0;
3085                 }
3086         } else {
3087                 raw_inode->i_uid_low =
3088                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3089                 raw_inode->i_gid_low =
3090                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3091                 raw_inode->i_uid_high = 0;
3092                 raw_inode->i_gid_high = 0;
3093         }
3094         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3095         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
3096         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
3097         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
3098         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
3099         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
3100         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3101         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
3102 #ifdef EXT3_FRAGMENTS
3103         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
3104         raw_inode->i_frag = ei->i_frag_no;
3105         raw_inode->i_fsize = ei->i_frag_size;
3106 #endif
3107         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
3108         if (!S_ISREG(inode->i_mode)) {
3109                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
3110         } else {
3111                 raw_inode->i_size_high =
3112                         cpu_to_le32(ei->i_disksize >> 32);
3113                 if (ei->i_disksize > 0x7fffffffULL) {
3114                         struct super_block *sb = inode->i_sb;
3115                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
3116                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
3117                             EXT3_SB(sb)->s_es->s_rev_level ==
3118                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
3119                                /* If this is the first large file
3120                                 * created, add a flag to the superblock.
3121                                 */
3122                                 unlock_buffer(bh);
3123                                 err = ext3_journal_get_write_access(handle,
3124                                                 EXT3_SB(sb)->s_sbh);
3125                                 if (err)
3126                                         goto out_brelse;
3127
3128                                 ext3_update_dynamic_rev(sb);
3129                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
3130                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
3131                                 handle->h_sync = 1;
3132                                 err = ext3_journal_dirty_metadata(handle,
3133                                                 EXT3_SB(sb)->s_sbh);
3134                                 /* get our lock and start over */
3135                                 goto again;
3136                         }
3137                 }
3138         }
3139         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3140         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3141                 if (old_valid_dev(inode->i_rdev)) {
3142                         raw_inode->i_block[0] =
3143                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3144                         raw_inode->i_block[1] = 0;
3145                 } else {
3146                         raw_inode->i_block[0] = 0;
3147                         raw_inode->i_block[1] =
3148                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3149                         raw_inode->i_block[2] = 0;
3150                 }
3151         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
3152                 raw_inode->i_block[block] = ei->i_data[block];
3153
3154         if (ei->i_extra_isize)
3155                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3156
3157         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
3158         unlock_buffer(bh);
3159         rc = ext3_journal_dirty_metadata(handle, bh);
3160         if (!err)
3161                 err = rc;
3162         ext3_clear_inode_state(inode, EXT3_STATE_NEW);
3163
3164         atomic_set(&ei->i_sync_tid, handle->h_transaction->t_tid);
3165 out_brelse:
3166         brelse (bh);
3167         ext3_std_error(inode->i_sb, err);
3168         return err;
3169 }
3170
3171 /*
3172  * ext3_write_inode()
3173  *
3174  * We are called from a few places:
3175  *
3176  * - Within generic_file_write() for O_SYNC files.
3177  *   Here, there will be no transaction running. We wait for any running
3178  *   trasnaction to commit.
3179  *
3180  * - Within sys_sync(), kupdate and such.
3181  *   We wait on commit, if tol to.
3182  *
3183  * - Within prune_icache() (PF_MEMALLOC == true)
3184  *   Here we simply return.  We can't afford to block kswapd on the
3185  *   journal commit.
3186  *
3187  * In all cases it is actually safe for us to return without doing anything,
3188  * because the inode has been copied into a raw inode buffer in
3189  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3190  * knfsd.
3191  *
3192  * Note that we are absolutely dependent upon all inode dirtiers doing the
3193  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3194  * which we are interested.
3195  *
3196  * It would be a bug for them to not do this.  The code:
3197  *
3198  *      mark_inode_dirty(inode)
3199  *      stuff();
3200  *      inode->i_size = expr;
3201  *
3202  * is in error because a kswapd-driven write_inode() could occur while
3203  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3204  * will no longer be on the superblock's dirty inode list.
3205  */
3206 int ext3_write_inode(struct inode *inode, struct writeback_control *wbc)
3207 {
3208         if (current->flags & PF_MEMALLOC)
3209                 return 0;
3210
3211         if (ext3_journal_current_handle()) {
3212                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3213                 dump_stack();
3214                 return -EIO;
3215         }
3216
3217         if (wbc->sync_mode != WB_SYNC_ALL)
3218                 return 0;
3219
3220         return ext3_force_commit(inode->i_sb);
3221 }
3222
3223 /*
3224  * ext3_setattr()
3225  *
3226  * Called from notify_change.
3227  *
3228  * We want to trap VFS attempts to truncate the file as soon as
3229  * possible.  In particular, we want to make sure that when the VFS
3230  * shrinks i_size, we put the inode on the orphan list and modify
3231  * i_disksize immediately, so that during the subsequent flushing of
3232  * dirty pages and freeing of disk blocks, we can guarantee that any
3233  * commit will leave the blocks being flushed in an unused state on
3234  * disk.  (On recovery, the inode will get truncated and the blocks will
3235  * be freed, so we have a strong guarantee that no future commit will
3236  * leave these blocks visible to the user.)
3237  *
3238  * Called with inode->sem down.
3239  */
3240 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3241 {
3242         struct inode *inode = dentry->d_inode;
3243         int error, rc = 0;
3244         const unsigned int ia_valid = attr->ia_valid;
3245
3246         error = inode_change_ok(inode, attr);
3247         if (error)
3248                 return error;
3249
3250         if (is_quota_modification(inode, attr))
3251                 dquot_initialize(inode);
3252         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3253                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3254                 handle_t *handle;
3255
3256                 /* (user+group)*(old+new) structure, inode write (sb,
3257                  * inode block, ? - but truncate inode update has it) */
3258                 handle = ext3_journal_start(inode, EXT3_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3259                                         EXT3_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)+3);
3260                 if (IS_ERR(handle)) {
3261                         error = PTR_ERR(handle);
3262                         goto err_out;
3263                 }
3264                 error = dquot_transfer(inode, attr);
3265                 if (error) {
3266                         ext3_journal_stop(handle);
3267                         return error;
3268                 }
3269                 /* Update corresponding info in inode so that everything is in
3270                  * one transaction */
3271                 if (attr->ia_valid & ATTR_UID)
3272                         inode->i_uid = attr->ia_uid;
3273                 if (attr->ia_valid & ATTR_GID)
3274                         inode->i_gid = attr->ia_gid;
3275                 error = ext3_mark_inode_dirty(handle, inode);
3276                 ext3_journal_stop(handle);
3277         }
3278
3279         if (S_ISREG(inode->i_mode) &&
3280             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3281                 handle_t *handle;
3282
3283                 handle = ext3_journal_start(inode, 3);
3284                 if (IS_ERR(handle)) {
3285                         error = PTR_ERR(handle);
3286                         goto err_out;
3287                 }
3288
3289                 error = ext3_orphan_add(handle, inode);
3290                 if (error) {
3291                         ext3_journal_stop(handle);
3292                         goto err_out;
3293                 }
3294                 EXT3_I(inode)->i_disksize = attr->ia_size;
3295                 error = ext3_mark_inode_dirty(handle, inode);
3296                 ext3_journal_stop(handle);
3297                 if (error) {
3298                         /* Some hard fs error must have happened. Bail out. */
3299                         ext3_orphan_del(NULL, inode);
3300                         goto err_out;
3301                 }
3302                 rc = ext3_block_truncate_page(inode, attr->ia_size);
3303                 if (rc) {
3304                         /* Cleanup orphan list and exit */
3305                         handle = ext3_journal_start(inode, 3);
3306                         if (IS_ERR(handle)) {
3307                                 ext3_orphan_del(NULL, inode);
3308                                 goto err_out;
3309                         }
3310                         ext3_orphan_del(handle, inode);
3311                         ext3_journal_stop(handle);
3312                         goto err_out;
3313                 }
3314         }
3315
3316         if ((attr->ia_valid & ATTR_SIZE) &&
3317             attr->ia_size != i_size_read(inode)) {
3318                 truncate_setsize(inode, attr->ia_size);
3319                 ext3_truncate(inode);
3320         }
3321
3322         setattr_copy(inode, attr);
3323         mark_inode_dirty(inode);
3324
3325         if (ia_valid & ATTR_MODE)
3326                 rc = ext3_acl_chmod(inode);
3327
3328 err_out:
3329         ext3_std_error(inode->i_sb, error);
3330         if (!error)
3331                 error = rc;
3332         return error;
3333 }
3334
3335
3336 /*
3337  * How many blocks doth make a writepage()?
3338  *
3339  * With N blocks per page, it may be:
3340  * N data blocks
3341  * 2 indirect block
3342  * 2 dindirect
3343  * 1 tindirect
3344  * N+5 bitmap blocks (from the above)
3345  * N+5 group descriptor summary blocks
3346  * 1 inode block
3347  * 1 superblock.
3348  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3349  *
3350  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3351  *
3352  * With ordered or writeback data it's the same, less the N data blocks.
3353  *
3354  * If the inode's direct blocks can hold an integral number of pages then a
3355  * page cannot straddle two indirect blocks, and we can only touch one indirect
3356  * and dindirect block, and the "5" above becomes "3".
3357  *
3358  * This still overestimates under most circumstances.  If we were to pass the
3359  * start and end offsets in here as well we could do block_to_path() on each
3360  * block and work out the exact number of indirects which are touched.  Pah.
3361  */
3362
3363 static int ext3_writepage_trans_blocks(struct inode *inode)
3364 {
3365         int bpp = ext3_journal_blocks_per_page(inode);
3366         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3367         int ret;
3368
3369         if (ext3_should_journal_data(inode))
3370                 ret = 3 * (bpp + indirects) + 2;
3371         else
3372                 ret = 2 * (bpp + indirects) + indirects + 2;
3373
3374 #ifdef CONFIG_QUOTA
3375         /* We know that structure was already allocated during dquot_initialize so
3376          * we will be updating only the data blocks + inodes */
3377         ret += EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
3378 #endif
3379
3380         return ret;
3381 }
3382
3383 /*
3384  * The caller must have previously called ext3_reserve_inode_write().
3385  * Give this, we know that the caller already has write access to iloc->bh.
3386  */
3387 int ext3_mark_iloc_dirty(handle_t *handle,
3388                 struct inode *inode, struct ext3_iloc *iloc)
3389 {
3390         int err = 0;
3391
3392         /* the do_update_inode consumes one bh->b_count */
3393         get_bh(iloc->bh);
3394
3395         /* ext3_do_update_inode() does journal_dirty_metadata */
3396         err = ext3_do_update_inode(handle, inode, iloc);
3397         put_bh(iloc->bh);
3398         return err;
3399 }
3400
3401 /*
3402  * On success, We end up with an outstanding reference count against
3403  * iloc->bh.  This _must_ be cleaned up later.
3404  */
3405
3406 int
3407 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3408                          struct ext3_iloc *iloc)
3409 {
3410         int err = 0;
3411         if (handle) {
3412                 err = ext3_get_inode_loc(inode, iloc);
3413                 if (!err) {
3414                         BUFFER_TRACE(iloc->bh, "get_write_access");
3415                         err = ext3_journal_get_write_access(handle, iloc->bh);
3416                         if (err) {
3417                                 brelse(iloc->bh);
3418                                 iloc->bh = NULL;
3419                         }
3420                 }
3421         }
3422         ext3_std_error(inode->i_sb, err);
3423         return err;
3424 }
3425
3426 /*
3427  * What we do here is to mark the in-core inode as clean with respect to inode
3428  * dirtiness (it may still be data-dirty).
3429  * This means that the in-core inode may be reaped by prune_icache
3430  * without having to perform any I/O.  This is a very good thing,
3431  * because *any* task may call prune_icache - even ones which
3432  * have a transaction open against a different journal.
3433  *
3434  * Is this cheating?  Not really.  Sure, we haven't written the
3435  * inode out, but prune_icache isn't a user-visible syncing function.
3436  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3437  * we start and wait on commits.
3438  *
3439  * Is this efficient/effective?  Well, we're being nice to the system
3440  * by cleaning up our inodes proactively so they can be reaped
3441  * without I/O.  But we are potentially leaving up to five seconds'
3442  * worth of inodes floating about which prune_icache wants us to
3443  * write out.  One way to fix that would be to get prune_icache()
3444  * to do a write_super() to free up some memory.  It has the desired
3445  * effect.
3446  */
3447 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3448 {
3449         struct ext3_iloc iloc;
3450         int err;
3451
3452         might_sleep();
3453         trace_ext3_mark_inode_dirty(inode, _RET_IP_);
3454         err = ext3_reserve_inode_write(handle, inode, &iloc);
3455         if (!err)
3456                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3457         return err;
3458 }
3459
3460 /*
3461  * ext3_dirty_inode() is called from __mark_inode_dirty()
3462  *
3463  * We're really interested in the case where a file is being extended.
3464  * i_size has been changed by generic_commit_write() and we thus need
3465  * to include the updated inode in the current transaction.
3466  *
3467  * Also, dquot_alloc_space() will always dirty the inode when blocks
3468  * are allocated to the file.
3469  *
3470  * If the inode is marked synchronous, we don't honour that here - doing
3471  * so would cause a commit on atime updates, which we don't bother doing.
3472  * We handle synchronous inodes at the highest possible level.
3473  */
3474 void ext3_dirty_inode(struct inode *inode, int flags)
3475 {
3476         handle_t *current_handle = ext3_journal_current_handle();
3477         handle_t *handle;
3478
3479         handle = ext3_journal_start(inode, 2);
3480         if (IS_ERR(handle))
3481                 goto out;
3482         if (current_handle &&
3483                 current_handle->h_transaction != handle->h_transaction) {
3484                 /* This task has a transaction open against a different fs */
3485                 printk(KERN_EMERG "%s: transactions do not match!\n",
3486                        __func__);
3487         } else {
3488                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3489                                 current_handle);
3490                 ext3_mark_inode_dirty(handle, inode);
3491         }
3492         ext3_journal_stop(handle);
3493 out:
3494         return;
3495 }
3496
3497 #if 0
3498 /*
3499  * Bind an inode's backing buffer_head into this transaction, to prevent
3500  * it from being flushed to disk early.  Unlike
3501  * ext3_reserve_inode_write, this leaves behind no bh reference and
3502  * returns no iloc structure, so the caller needs to repeat the iloc
3503  * lookup to mark the inode dirty later.
3504  */
3505 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3506 {
3507         struct ext3_iloc iloc;
3508
3509         int err = 0;
3510         if (handle) {
3511                 err = ext3_get_inode_loc(inode, &iloc);
3512                 if (!err) {
3513                         BUFFER_TRACE(iloc.bh, "get_write_access");
3514                         err = journal_get_write_access(handle, iloc.bh);
3515                         if (!err)
3516                                 err = ext3_journal_dirty_metadata(handle,
3517                                                                   iloc.bh);
3518                         brelse(iloc.bh);
3519                 }
3520         }
3521         ext3_std_error(inode->i_sb, err);
3522         return err;
3523 }
3524 #endif
3525
3526 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3527 {
3528         journal_t *journal;
3529         handle_t *handle;
3530         int err;
3531
3532         /*
3533          * We have to be very careful here: changing a data block's
3534          * journaling status dynamically is dangerous.  If we write a
3535          * data block to the journal, change the status and then delete
3536          * that block, we risk forgetting to revoke the old log record
3537          * from the journal and so a subsequent replay can corrupt data.
3538          * So, first we make sure that the journal is empty and that
3539          * nobody is changing anything.
3540          */
3541
3542         journal = EXT3_JOURNAL(inode);
3543         if (is_journal_aborted(journal))
3544                 return -EROFS;
3545
3546         journal_lock_updates(journal);
3547         journal_flush(journal);
3548
3549         /*
3550          * OK, there are no updates running now, and all cached data is
3551          * synced to disk.  We are now in a completely consistent state
3552          * which doesn't have anything in the journal, and we know that
3553          * no filesystem updates are running, so it is safe to modify
3554          * the inode's in-core data-journaling state flag now.
3555          */
3556
3557         if (val)
3558                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3559         else
3560                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3561         ext3_set_aops(inode);
3562
3563         journal_unlock_updates(journal);
3564
3565         /* Finally we can mark the inode as dirty. */
3566
3567         handle = ext3_journal_start(inode, 1);
3568         if (IS_ERR(handle))
3569                 return PTR_ERR(handle);
3570
3571         err = ext3_mark_inode_dirty(handle, inode);
3572         handle->h_sync = 1;
3573         ext3_journal_stop(handle);
3574         ext3_std_error(inode->i_sb, err);
3575
3576         return err;
3577 }