d2e4547c7806340dc055ef82f15f4643a46411b8
[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_block_alloc_info *rsv;
201         handle_t *handle;
202         int want_delete = 0;
203
204         trace_ext3_evict_inode(inode);
205         if (!inode->i_nlink && !is_bad_inode(inode)) {
206                 dquot_initialize(inode);
207                 want_delete = 1;
208         }
209
210         truncate_inode_pages(&inode->i_data, 0);
211
212         ext3_discard_reservation(inode);
213         rsv = EXT3_I(inode)->i_block_alloc_info;
214         EXT3_I(inode)->i_block_alloc_info = NULL;
215         if (unlikely(rsv))
216                 kfree(rsv);
217
218         if (!want_delete)
219                 goto no_delete;
220
221         handle = start_transaction(inode);
222         if (IS_ERR(handle)) {
223                 /*
224                  * If we're going to skip the normal cleanup, we still need to
225                  * make sure that the in-core orphan linked list is properly
226                  * cleaned up.
227                  */
228                 ext3_orphan_del(NULL, inode);
229                 goto no_delete;
230         }
231
232         if (IS_SYNC(inode))
233                 handle->h_sync = 1;
234         inode->i_size = 0;
235         if (inode->i_blocks)
236                 ext3_truncate(inode);
237         /*
238          * Kill off the orphan record created when the inode lost the last
239          * link.  Note that ext3_orphan_del() has to be able to cope with the
240          * deletion of a non-existent orphan - ext3_truncate() could
241          * have removed the record.
242          */
243         ext3_orphan_del(handle, inode);
244         EXT3_I(inode)->i_dtime  = get_seconds();
245
246         /*
247          * One subtle ordering requirement: if anything has gone wrong
248          * (transaction abort, IO errors, whatever), then we can still
249          * do these next steps (the fs will already have been marked as
250          * having errors), but we can't free the inode if the mark_dirty
251          * fails.
252          */
253         if (ext3_mark_inode_dirty(handle, inode)) {
254                 /* If that failed, just dquot_drop() and be done with that */
255                 dquot_drop(inode);
256                 end_writeback(inode);
257         } else {
258                 ext3_xattr_delete_inode(handle, inode);
259                 dquot_free_inode(inode);
260                 dquot_drop(inode);
261                 end_writeback(inode);
262                 ext3_free_inode(handle, inode);
263         }
264         ext3_journal_stop(handle);
265         return;
266 no_delete:
267         end_writeback(inode);
268         dquot_drop(inode);
269 }
270
271 typedef struct {
272         __le32  *p;
273         __le32  key;
274         struct buffer_head *bh;
275 } Indirect;
276
277 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
278 {
279         p->key = *(p->p = v);
280         p->bh = bh;
281 }
282
283 static int verify_chain(Indirect *from, Indirect *to)
284 {
285         while (from <= to && from->key == *from->p)
286                 from++;
287         return (from > to);
288 }
289
290 /**
291  *      ext3_block_to_path - parse the block number into array of offsets
292  *      @inode: inode in question (we are only interested in its superblock)
293  *      @i_block: block number to be parsed
294  *      @offsets: array to store the offsets in
295  *      @boundary: set this non-zero if the referred-to block is likely to be
296  *             followed (on disk) by an indirect block.
297  *
298  *      To store the locations of file's data ext3 uses a data structure common
299  *      for UNIX filesystems - tree of pointers anchored in the inode, with
300  *      data blocks at leaves and indirect blocks in intermediate nodes.
301  *      This function translates the block number into path in that tree -
302  *      return value is the path length and @offsets[n] is the offset of
303  *      pointer to (n+1)th node in the nth one. If @block is out of range
304  *      (negative or too large) warning is printed and zero returned.
305  *
306  *      Note: function doesn't find node addresses, so no IO is needed. All
307  *      we need to know is the capacity of indirect blocks (taken from the
308  *      inode->i_sb).
309  */
310
311 /*
312  * Portability note: the last comparison (check that we fit into triple
313  * indirect block) is spelled differently, because otherwise on an
314  * architecture with 32-bit longs and 8Kb pages we might get into trouble
315  * if our filesystem had 8Kb blocks. We might use long long, but that would
316  * kill us on x86. Oh, well, at least the sign propagation does not matter -
317  * i_block would have to be negative in the very beginning, so we would not
318  * get there at all.
319  */
320
321 static int ext3_block_to_path(struct inode *inode,
322                         long i_block, int offsets[4], int *boundary)
323 {
324         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
325         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
326         const long direct_blocks = EXT3_NDIR_BLOCKS,
327                 indirect_blocks = ptrs,
328                 double_blocks = (1 << (ptrs_bits * 2));
329         int n = 0;
330         int final = 0;
331
332         if (i_block < 0) {
333                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
334         } else if (i_block < direct_blocks) {
335                 offsets[n++] = i_block;
336                 final = direct_blocks;
337         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
338                 offsets[n++] = EXT3_IND_BLOCK;
339                 offsets[n++] = i_block;
340                 final = ptrs;
341         } else if ((i_block -= indirect_blocks) < double_blocks) {
342                 offsets[n++] = EXT3_DIND_BLOCK;
343                 offsets[n++] = i_block >> ptrs_bits;
344                 offsets[n++] = i_block & (ptrs - 1);
345                 final = ptrs;
346         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
347                 offsets[n++] = EXT3_TIND_BLOCK;
348                 offsets[n++] = i_block >> (ptrs_bits * 2);
349                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
350                 offsets[n++] = i_block & (ptrs - 1);
351                 final = ptrs;
352         } else {
353                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
354         }
355         if (boundary)
356                 *boundary = final - 1 - (i_block & (ptrs - 1));
357         return n;
358 }
359
360 /**
361  *      ext3_get_branch - read the chain of indirect blocks leading to data
362  *      @inode: inode in question
363  *      @depth: depth of the chain (1 - direct pointer, etc.)
364  *      @offsets: offsets of pointers in inode/indirect blocks
365  *      @chain: place to store the result
366  *      @err: here we store the error value
367  *
368  *      Function fills the array of triples <key, p, bh> and returns %NULL
369  *      if everything went OK or the pointer to the last filled triple
370  *      (incomplete one) otherwise. Upon the return chain[i].key contains
371  *      the number of (i+1)-th block in the chain (as it is stored in memory,
372  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
373  *      number (it points into struct inode for i==0 and into the bh->b_data
374  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
375  *      block for i>0 and NULL for i==0. In other words, it holds the block
376  *      numbers of the chain, addresses they were taken from (and where we can
377  *      verify that chain did not change) and buffer_heads hosting these
378  *      numbers.
379  *
380  *      Function stops when it stumbles upon zero pointer (absent block)
381  *              (pointer to last triple returned, *@err == 0)
382  *      or when it gets an IO error reading an indirect block
383  *              (ditto, *@err == -EIO)
384  *      or when it notices that chain had been changed while it was reading
385  *              (ditto, *@err == -EAGAIN)
386  *      or when it reads all @depth-1 indirect blocks successfully and finds
387  *      the whole chain, all way to the data (returns %NULL, *err == 0).
388  */
389 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
390                                  Indirect chain[4], int *err)
391 {
392         struct super_block *sb = inode->i_sb;
393         Indirect *p = chain;
394         struct buffer_head *bh;
395
396         *err = 0;
397         /* i_data is not going away, no lock needed */
398         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
399         if (!p->key)
400                 goto no_block;
401         while (--depth) {
402                 bh = sb_bread(sb, le32_to_cpu(p->key));
403                 if (!bh)
404                         goto failure;
405                 /* Reader: pointers */
406                 if (!verify_chain(chain, p))
407                         goto changed;
408                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
409                 /* Reader: end */
410                 if (!p->key)
411                         goto no_block;
412         }
413         return NULL;
414
415 changed:
416         brelse(bh);
417         *err = -EAGAIN;
418         goto no_block;
419 failure:
420         *err = -EIO;
421 no_block:
422         return p;
423 }
424
425 /**
426  *      ext3_find_near - find a place for allocation with sufficient locality
427  *      @inode: owner
428  *      @ind: descriptor of indirect block.
429  *
430  *      This function returns the preferred place for block allocation.
431  *      It is used when heuristic for sequential allocation fails.
432  *      Rules are:
433  *        + if there is a block to the left of our position - allocate near it.
434  *        + if pointer will live in indirect block - allocate near that block.
435  *        + if pointer will live in inode - allocate in the same
436  *          cylinder group.
437  *
438  * In the latter case we colour the starting block by the callers PID to
439  * prevent it from clashing with concurrent allocations for a different inode
440  * in the same block group.   The PID is used here so that functionally related
441  * files will be close-by on-disk.
442  *
443  *      Caller must make sure that @ind is valid and will stay that way.
444  */
445 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
446 {
447         struct ext3_inode_info *ei = EXT3_I(inode);
448         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
449         __le32 *p;
450         ext3_fsblk_t bg_start;
451         ext3_grpblk_t colour;
452
453         /* Try to find previous block */
454         for (p = ind->p - 1; p >= start; p--) {
455                 if (*p)
456                         return le32_to_cpu(*p);
457         }
458
459         /* No such thing, so let's try location of indirect block */
460         if (ind->bh)
461                 return ind->bh->b_blocknr;
462
463         /*
464          * It is going to be referred to from the inode itself? OK, just put it
465          * into the same cylinder group then.
466          */
467         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
468         colour = (current->pid % 16) *
469                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
470         return bg_start + colour;
471 }
472
473 /**
474  *      ext3_find_goal - find a preferred place for allocation.
475  *      @inode: owner
476  *      @block:  block we want
477  *      @partial: pointer to the last triple within a chain
478  *
479  *      Normally this function find the preferred place for block allocation,
480  *      returns it.
481  */
482
483 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
484                                    Indirect *partial)
485 {
486         struct ext3_block_alloc_info *block_i;
487
488         block_i =  EXT3_I(inode)->i_block_alloc_info;
489
490         /*
491          * try the heuristic for sequential allocation,
492          * failing that at least try to get decent locality.
493          */
494         if (block_i && (block == block_i->last_alloc_logical_block + 1)
495                 && (block_i->last_alloc_physical_block != 0)) {
496                 return block_i->last_alloc_physical_block + 1;
497         }
498
499         return ext3_find_near(inode, partial);
500 }
501
502 /**
503  *      ext3_blks_to_allocate - Look up the block map and count the number
504  *      of direct blocks need to be allocated for the given branch.
505  *
506  *      @branch: chain of indirect blocks
507  *      @k: number of blocks need for indirect blocks
508  *      @blks: number of data blocks to be mapped.
509  *      @blocks_to_boundary:  the offset in the indirect block
510  *
511  *      return the total number of blocks to be allocate, including the
512  *      direct and indirect blocks.
513  */
514 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
515                 int blocks_to_boundary)
516 {
517         unsigned long count = 0;
518
519         /*
520          * Simple case, [t,d]Indirect block(s) has not allocated yet
521          * then it's clear blocks on that path have not allocated
522          */
523         if (k > 0) {
524                 /* right now we don't handle cross boundary allocation */
525                 if (blks < blocks_to_boundary + 1)
526                         count += blks;
527                 else
528                         count += blocks_to_boundary + 1;
529                 return count;
530         }
531
532         count++;
533         while (count < blks && count <= blocks_to_boundary &&
534                 le32_to_cpu(*(branch[0].p + count)) == 0) {
535                 count++;
536         }
537         return count;
538 }
539
540 /**
541  *      ext3_alloc_blocks - multiple allocate blocks needed for a branch
542  *      @handle: handle for this transaction
543  *      @inode: owner
544  *      @goal: preferred place for allocation
545  *      @indirect_blks: the number of blocks need to allocate for indirect
546  *                      blocks
547  *      @blks:  number of blocks need to allocated for direct blocks
548  *      @new_blocks: on return it will store the new block numbers for
549  *      the indirect blocks(if needed) and the first direct block,
550  *      @err: here we store the error value
551  *
552  *      return the number of direct blocks allocated
553  */
554 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
555                         ext3_fsblk_t goal, int indirect_blks, int blks,
556                         ext3_fsblk_t new_blocks[4], int *err)
557 {
558         int target, i;
559         unsigned long count = 0;
560         int index = 0;
561         ext3_fsblk_t current_block = 0;
562         int ret = 0;
563
564         /*
565          * Here we try to allocate the requested multiple blocks at once,
566          * on a best-effort basis.
567          * To build a branch, we should allocate blocks for
568          * the indirect blocks(if not allocated yet), and at least
569          * the first direct block of this branch.  That's the
570          * minimum number of blocks need to allocate(required)
571          */
572         target = blks + indirect_blks;
573
574         while (1) {
575                 count = target;
576                 /* allocating blocks for indirect blocks and direct blocks */
577                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
578                 if (*err)
579                         goto failed_out;
580
581                 target -= count;
582                 /* allocate blocks for indirect blocks */
583                 while (index < indirect_blks && count) {
584                         new_blocks[index++] = current_block++;
585                         count--;
586                 }
587
588                 if (count > 0)
589                         break;
590         }
591
592         /* save the new block number for the first direct block */
593         new_blocks[index] = current_block;
594
595         /* total number of blocks allocated for direct blocks */
596         ret = count;
597         *err = 0;
598         return ret;
599 failed_out:
600         for (i = 0; i <index; i++)
601                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
602         return ret;
603 }
604
605 /**
606  *      ext3_alloc_branch - allocate and set up a chain of blocks.
607  *      @handle: handle for this transaction
608  *      @inode: owner
609  *      @indirect_blks: number of allocated indirect blocks
610  *      @blks: number of allocated direct blocks
611  *      @goal: preferred place for allocation
612  *      @offsets: offsets (in the blocks) to store the pointers to next.
613  *      @branch: place to store the chain in.
614  *
615  *      This function allocates blocks, zeroes out all but the last one,
616  *      links them into chain and (if we are synchronous) writes them to disk.
617  *      In other words, it prepares a branch that can be spliced onto the
618  *      inode. It stores the information about that chain in the branch[], in
619  *      the same format as ext3_get_branch() would do. We are calling it after
620  *      we had read the existing part of chain and partial points to the last
621  *      triple of that (one with zero ->key). Upon the exit we have the same
622  *      picture as after the successful ext3_get_block(), except that in one
623  *      place chain is disconnected - *branch->p is still zero (we did not
624  *      set the last link), but branch->key contains the number that should
625  *      be placed into *branch->p to fill that gap.
626  *
627  *      If allocation fails we free all blocks we've allocated (and forget
628  *      their buffer_heads) and return the error value the from failed
629  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
630  *      as described above and return 0.
631  */
632 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
633                         int indirect_blks, int *blks, ext3_fsblk_t goal,
634                         int *offsets, Indirect *branch)
635 {
636         int blocksize = inode->i_sb->s_blocksize;
637         int i, n = 0;
638         int err = 0;
639         struct buffer_head *bh;
640         int num;
641         ext3_fsblk_t new_blocks[4];
642         ext3_fsblk_t current_block;
643
644         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
645                                 *blks, new_blocks, &err);
646         if (err)
647                 return err;
648
649         branch[0].key = cpu_to_le32(new_blocks[0]);
650         /*
651          * metadata blocks and data blocks are allocated.
652          */
653         for (n = 1; n <= indirect_blks;  n++) {
654                 /*
655                  * Get buffer_head for parent block, zero it out
656                  * and set the pointer to new one, then send
657                  * parent to disk.
658                  */
659                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
660                 branch[n].bh = bh;
661                 lock_buffer(bh);
662                 BUFFER_TRACE(bh, "call get_create_access");
663                 err = ext3_journal_get_create_access(handle, bh);
664                 if (err) {
665                         unlock_buffer(bh);
666                         brelse(bh);
667                         goto failed;
668                 }
669
670                 memset(bh->b_data, 0, blocksize);
671                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
672                 branch[n].key = cpu_to_le32(new_blocks[n]);
673                 *branch[n].p = branch[n].key;
674                 if ( n == indirect_blks) {
675                         current_block = new_blocks[n];
676                         /*
677                          * End of chain, update the last new metablock of
678                          * the chain to point to the new allocated
679                          * data blocks numbers
680                          */
681                         for (i=1; i < num; i++)
682                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
683                 }
684                 BUFFER_TRACE(bh, "marking uptodate");
685                 set_buffer_uptodate(bh);
686                 unlock_buffer(bh);
687
688                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
689                 err = ext3_journal_dirty_metadata(handle, bh);
690                 if (err)
691                         goto failed;
692         }
693         *blks = num;
694         return err;
695 failed:
696         /* Allocation failed, free what we already allocated */
697         for (i = 1; i <= n ; i++) {
698                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
699                 ext3_journal_forget(handle, branch[i].bh);
700         }
701         for (i = 0; i <indirect_blks; i++)
702                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
703
704         ext3_free_blocks(handle, inode, new_blocks[i], num);
705
706         return err;
707 }
708
709 /**
710  * ext3_splice_branch - splice the allocated branch onto inode.
711  * @handle: handle for this transaction
712  * @inode: owner
713  * @block: (logical) number of block we are adding
714  * @where: location of missing link
715  * @num:   number of indirect blocks we are adding
716  * @blks:  number of direct blocks we are adding
717  *
718  * This function fills the missing link and does all housekeeping needed in
719  * inode (->i_blocks, etc.). In case of success we end up with the full
720  * chain to new block and return 0.
721  */
722 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
723                         long block, Indirect *where, int num, int blks)
724 {
725         int i;
726         int err = 0;
727         struct ext3_block_alloc_info *block_i;
728         ext3_fsblk_t current_block;
729         struct ext3_inode_info *ei = EXT3_I(inode);
730
731         block_i = ei->i_block_alloc_info;
732         /*
733          * If we're splicing into a [td]indirect block (as opposed to the
734          * inode) then we need to get write access to the [td]indirect block
735          * before the splice.
736          */
737         if (where->bh) {
738                 BUFFER_TRACE(where->bh, "get_write_access");
739                 err = ext3_journal_get_write_access(handle, where->bh);
740                 if (err)
741                         goto err_out;
742         }
743         /* That's it */
744
745         *where->p = where->key;
746
747         /*
748          * Update the host buffer_head or inode to point to more just allocated
749          * direct blocks blocks
750          */
751         if (num == 0 && blks > 1) {
752                 current_block = le32_to_cpu(where->key) + 1;
753                 for (i = 1; i < blks; i++)
754                         *(where->p + i ) = cpu_to_le32(current_block++);
755         }
756
757         /*
758          * update the most recently allocated logical & physical block
759          * in i_block_alloc_info, to assist find the proper goal block for next
760          * allocation
761          */
762         if (block_i) {
763                 block_i->last_alloc_logical_block = block + blks - 1;
764                 block_i->last_alloc_physical_block =
765                                 le32_to_cpu(where[num].key) + blks - 1;
766         }
767
768         /* We are done with atomic stuff, now do the rest of housekeeping */
769
770         inode->i_ctime = CURRENT_TIME_SEC;
771         ext3_mark_inode_dirty(handle, inode);
772         /* ext3_mark_inode_dirty already updated i_sync_tid */
773         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
774
775         /* had we spliced it onto indirect block? */
776         if (where->bh) {
777                 /*
778                  * If we spliced it onto an indirect block, we haven't
779                  * altered the inode.  Note however that if it is being spliced
780                  * onto an indirect block at the very end of the file (the
781                  * file is growing) then we *will* alter the inode to reflect
782                  * the new i_size.  But that is not done here - it is done in
783                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
784                  */
785                 jbd_debug(5, "splicing indirect only\n");
786                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
787                 err = ext3_journal_dirty_metadata(handle, where->bh);
788                 if (err)
789                         goto err_out;
790         } else {
791                 /*
792                  * OK, we spliced it into the inode itself on a direct block.
793                  * Inode was dirtied above.
794                  */
795                 jbd_debug(5, "splicing direct\n");
796         }
797         return err;
798
799 err_out:
800         for (i = 1; i <= num; i++) {
801                 BUFFER_TRACE(where[i].bh, "call journal_forget");
802                 ext3_journal_forget(handle, where[i].bh);
803                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
804         }
805         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
806
807         return err;
808 }
809
810 /*
811  * Allocation strategy is simple: if we have to allocate something, we will
812  * have to go the whole way to leaf. So let's do it before attaching anything
813  * to tree, set linkage between the newborn blocks, write them if sync is
814  * required, recheck the path, free and repeat if check fails, otherwise
815  * set the last missing link (that will protect us from any truncate-generated
816  * removals - all blocks on the path are immune now) and possibly force the
817  * write on the parent block.
818  * That has a nice additional property: no special recovery from the failed
819  * allocations is needed - we simply release blocks and do not touch anything
820  * reachable from inode.
821  *
822  * `handle' can be NULL if create == 0.
823  *
824  * The BKL may not be held on entry here.  Be sure to take it early.
825  * return > 0, # of blocks mapped or allocated.
826  * return = 0, if plain lookup failed.
827  * return < 0, error case.
828  */
829 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
830                 sector_t iblock, unsigned long maxblocks,
831                 struct buffer_head *bh_result,
832                 int create)
833 {
834         int err = -EIO;
835         int offsets[4];
836         Indirect chain[4];
837         Indirect *partial;
838         ext3_fsblk_t goal;
839         int indirect_blks;
840         int blocks_to_boundary = 0;
841         int depth;
842         struct ext3_inode_info *ei = EXT3_I(inode);
843         int count = 0;
844         ext3_fsblk_t first_block = 0;
845
846
847         trace_ext3_get_blocks_enter(inode, iblock, maxblocks, create);
848         J_ASSERT(handle != NULL || create == 0);
849         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
850
851         if (depth == 0)
852                 goto out;
853
854         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
855
856         /* Simplest case - block found, no allocation needed */
857         if (!partial) {
858                 first_block = le32_to_cpu(chain[depth - 1].key);
859                 clear_buffer_new(bh_result);
860                 count++;
861                 /*map more blocks*/
862                 while (count < maxblocks && count <= blocks_to_boundary) {
863                         ext3_fsblk_t blk;
864
865                         if (!verify_chain(chain, chain + depth - 1)) {
866                                 /*
867                                  * Indirect block might be removed by
868                                  * truncate while we were reading it.
869                                  * Handling of that case: forget what we've
870                                  * got now. Flag the err as EAGAIN, so it
871                                  * will reread.
872                                  */
873                                 err = -EAGAIN;
874                                 count = 0;
875                                 break;
876                         }
877                         blk = le32_to_cpu(*(chain[depth-1].p + count));
878
879                         if (blk == first_block + count)
880                                 count++;
881                         else
882                                 break;
883                 }
884                 if (err != -EAGAIN)
885                         goto got_it;
886         }
887
888         /* Next simple case - plain lookup or failed read of indirect block */
889         if (!create || err == -EIO)
890                 goto cleanup;
891
892         /*
893          * Block out ext3_truncate while we alter the tree
894          */
895         mutex_lock(&ei->truncate_mutex);
896
897         /*
898          * If the indirect block is missing while we are reading
899          * the chain(ext3_get_branch() returns -EAGAIN err), or
900          * if the chain has been changed after we grab the semaphore,
901          * (either because another process truncated this branch, or
902          * another get_block allocated this branch) re-grab the chain to see if
903          * the request block has been allocated or not.
904          *
905          * Since we already block the truncate/other get_block
906          * at this point, we will have the current copy of the chain when we
907          * splice the branch into the tree.
908          */
909         if (err == -EAGAIN || !verify_chain(chain, partial)) {
910                 while (partial > chain) {
911                         brelse(partial->bh);
912                         partial--;
913                 }
914                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
915                 if (!partial) {
916                         count++;
917                         mutex_unlock(&ei->truncate_mutex);
918                         if (err)
919                                 goto cleanup;
920                         clear_buffer_new(bh_result);
921                         goto got_it;
922                 }
923         }
924
925         /*
926          * Okay, we need to do block allocation.  Lazily initialize the block
927          * allocation info here if necessary
928         */
929         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
930                 ext3_init_block_alloc_info(inode);
931
932         goal = ext3_find_goal(inode, iblock, partial);
933
934         /* the number of blocks need to allocate for [d,t]indirect blocks */
935         indirect_blks = (chain + depth) - partial - 1;
936
937         /*
938          * Next look up the indirect map to count the totoal number of
939          * direct blocks to allocate for this branch.
940          */
941         count = ext3_blks_to_allocate(partial, indirect_blks,
942                                         maxblocks, blocks_to_boundary);
943         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
944                                 offsets + (partial - chain), partial);
945
946         /*
947          * The ext3_splice_branch call will free and forget any buffers
948          * on the new chain if there is a failure, but that risks using
949          * up transaction credits, especially for bitmaps where the
950          * credits cannot be returned.  Can we handle this somehow?  We
951          * may need to return -EAGAIN upwards in the worst case.  --sct
952          */
953         if (!err)
954                 err = ext3_splice_branch(handle, inode, iblock,
955                                         partial, indirect_blks, count);
956         mutex_unlock(&ei->truncate_mutex);
957         if (err)
958                 goto cleanup;
959
960         set_buffer_new(bh_result);
961 got_it:
962         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
963         if (count > blocks_to_boundary)
964                 set_buffer_boundary(bh_result);
965         err = count;
966         /* Clean up and exit */
967         partial = chain + depth - 1;    /* the whole chain */
968 cleanup:
969         while (partial > chain) {
970                 BUFFER_TRACE(partial->bh, "call brelse");
971                 brelse(partial->bh);
972                 partial--;
973         }
974         BUFFER_TRACE(bh_result, "returned");
975 out:
976         trace_ext3_get_blocks_exit(inode, iblock,
977                                    depth ? le32_to_cpu(chain[depth-1].key) : 0,
978                                    count, err);
979         return err;
980 }
981
982 /* Maximum number of blocks we map for direct IO at once. */
983 #define DIO_MAX_BLOCKS 4096
984 /*
985  * Number of credits we need for writing DIO_MAX_BLOCKS:
986  * We need sb + group descriptor + bitmap + inode -> 4
987  * For B blocks with A block pointers per block we need:
988  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
989  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
990  */
991 #define DIO_CREDITS 25
992
993 static int ext3_get_block(struct inode *inode, sector_t iblock,
994                         struct buffer_head *bh_result, int create)
995 {
996         handle_t *handle = ext3_journal_current_handle();
997         int ret = 0, started = 0;
998         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
999
1000         if (create && !handle) {        /* Direct IO write... */
1001                 if (max_blocks > DIO_MAX_BLOCKS)
1002                         max_blocks = DIO_MAX_BLOCKS;
1003                 handle = ext3_journal_start(inode, DIO_CREDITS +
1004                                 EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb));
1005                 if (IS_ERR(handle)) {
1006                         ret = PTR_ERR(handle);
1007                         goto out;
1008                 }
1009                 started = 1;
1010         }
1011
1012         ret = ext3_get_blocks_handle(handle, inode, iblock,
1013                                         max_blocks, bh_result, create);
1014         if (ret > 0) {
1015                 bh_result->b_size = (ret << inode->i_blkbits);
1016                 ret = 0;
1017         }
1018         if (started)
1019                 ext3_journal_stop(handle);
1020 out:
1021         return ret;
1022 }
1023
1024 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
1025                 u64 start, u64 len)
1026 {
1027         return generic_block_fiemap(inode, fieinfo, start, len,
1028                                     ext3_get_block);
1029 }
1030
1031 /*
1032  * `handle' can be NULL if create is zero
1033  */
1034 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1035                                 long block, int create, int *errp)
1036 {
1037         struct buffer_head dummy;
1038         int fatal = 0, err;
1039
1040         J_ASSERT(handle != NULL || create == 0);
1041
1042         dummy.b_state = 0;
1043         dummy.b_blocknr = -1000;
1044         buffer_trace_init(&dummy.b_history);
1045         err = ext3_get_blocks_handle(handle, inode, block, 1,
1046                                         &dummy, create);
1047         /*
1048          * ext3_get_blocks_handle() returns number of blocks
1049          * mapped. 0 in case of a HOLE.
1050          */
1051         if (err > 0) {
1052                 if (err > 1)
1053                         WARN_ON(1);
1054                 err = 0;
1055         }
1056         *errp = err;
1057         if (!err && buffer_mapped(&dummy)) {
1058                 struct buffer_head *bh;
1059                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1060                 if (!bh) {
1061                         *errp = -EIO;
1062                         goto err;
1063                 }
1064                 if (buffer_new(&dummy)) {
1065                         J_ASSERT(create != 0);
1066                         J_ASSERT(handle != NULL);
1067
1068                         /*
1069                          * Now that we do not always journal data, we should
1070                          * keep in mind whether this should always journal the
1071                          * new buffer as metadata.  For now, regular file
1072                          * writes use ext3_get_block instead, so it's not a
1073                          * problem.
1074                          */
1075                         lock_buffer(bh);
1076                         BUFFER_TRACE(bh, "call get_create_access");
1077                         fatal = ext3_journal_get_create_access(handle, bh);
1078                         if (!fatal && !buffer_uptodate(bh)) {
1079                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1080                                 set_buffer_uptodate(bh);
1081                         }
1082                         unlock_buffer(bh);
1083                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1084                         err = ext3_journal_dirty_metadata(handle, bh);
1085                         if (!fatal)
1086                                 fatal = err;
1087                 } else {
1088                         BUFFER_TRACE(bh, "not a new buffer");
1089                 }
1090                 if (fatal) {
1091                         *errp = fatal;
1092                         brelse(bh);
1093                         bh = NULL;
1094                 }
1095                 return bh;
1096         }
1097 err:
1098         return NULL;
1099 }
1100
1101 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1102                                int block, int create, int *err)
1103 {
1104         struct buffer_head * bh;
1105
1106         bh = ext3_getblk(handle, inode, block, create, err);
1107         if (!bh)
1108                 return bh;
1109         if (buffer_uptodate(bh))
1110                 return bh;
1111         ll_rw_block(READ_META, 1, &bh);
1112         wait_on_buffer(bh);
1113         if (buffer_uptodate(bh))
1114                 return bh;
1115         put_bh(bh);
1116         *err = -EIO;
1117         return NULL;
1118 }
1119
1120 static int walk_page_buffers(   handle_t *handle,
1121                                 struct buffer_head *head,
1122                                 unsigned from,
1123                                 unsigned to,
1124                                 int *partial,
1125                                 int (*fn)(      handle_t *handle,
1126                                                 struct buffer_head *bh))
1127 {
1128         struct buffer_head *bh;
1129         unsigned block_start, block_end;
1130         unsigned blocksize = head->b_size;
1131         int err, ret = 0;
1132         struct buffer_head *next;
1133
1134         for (   bh = head, block_start = 0;
1135                 ret == 0 && (bh != head || !block_start);
1136                 block_start = block_end, bh = next)
1137         {
1138                 next = bh->b_this_page;
1139                 block_end = block_start + blocksize;
1140                 if (block_end <= from || block_start >= to) {
1141                         if (partial && !buffer_uptodate(bh))
1142                                 *partial = 1;
1143                         continue;
1144                 }
1145                 err = (*fn)(handle, bh);
1146                 if (!ret)
1147                         ret = err;
1148         }
1149         return ret;
1150 }
1151
1152 /*
1153  * To preserve ordering, it is essential that the hole instantiation and
1154  * the data write be encapsulated in a single transaction.  We cannot
1155  * close off a transaction and start a new one between the ext3_get_block()
1156  * and the commit_write().  So doing the journal_start at the start of
1157  * prepare_write() is the right place.
1158  *
1159  * Also, this function can nest inside ext3_writepage() ->
1160  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1161  * has generated enough buffer credits to do the whole page.  So we won't
1162  * block on the journal in that case, which is good, because the caller may
1163  * be PF_MEMALLOC.
1164  *
1165  * By accident, ext3 can be reentered when a transaction is open via
1166  * quota file writes.  If we were to commit the transaction while thus
1167  * reentered, there can be a deadlock - we would be holding a quota
1168  * lock, and the commit would never complete if another thread had a
1169  * transaction open and was blocking on the quota lock - a ranking
1170  * violation.
1171  *
1172  * So what we do is to rely on the fact that journal_stop/journal_start
1173  * will _not_ run commit under these circumstances because handle->h_ref
1174  * is elevated.  We'll still have enough credits for the tiny quotafile
1175  * write.
1176  */
1177 static int do_journal_get_write_access(handle_t *handle,
1178                                         struct buffer_head *bh)
1179 {
1180         int dirty = buffer_dirty(bh);
1181         int ret;
1182
1183         if (!buffer_mapped(bh) || buffer_freed(bh))
1184                 return 0;
1185         /*
1186          * __block_prepare_write() could have dirtied some buffers. Clean
1187          * the dirty bit as jbd2_journal_get_write_access() could complain
1188          * otherwise about fs integrity issues. Setting of the dirty bit
1189          * by __block_prepare_write() isn't a real problem here as we clear
1190          * the bit before releasing a page lock and thus writeback cannot
1191          * ever write the buffer.
1192          */
1193         if (dirty)
1194                 clear_buffer_dirty(bh);
1195         ret = ext3_journal_get_write_access(handle, bh);
1196         if (!ret && dirty)
1197                 ret = ext3_journal_dirty_metadata(handle, bh);
1198         return ret;
1199 }
1200
1201 /*
1202  * Truncate blocks that were not used by write. We have to truncate the
1203  * pagecache as well so that corresponding buffers get properly unmapped.
1204  */
1205 static void ext3_truncate_failed_write(struct inode *inode)
1206 {
1207         truncate_inode_pages(inode->i_mapping, inode->i_size);
1208         ext3_truncate(inode);
1209 }
1210
1211 /*
1212  * Truncate blocks that were not used by direct IO write. We have to zero out
1213  * the last file block as well because direct IO might have written to it.
1214  */
1215 static void ext3_truncate_failed_direct_write(struct inode *inode)
1216 {
1217         ext3_block_truncate_page(inode, inode->i_size);
1218         ext3_truncate(inode);
1219 }
1220
1221 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1222                                 loff_t pos, unsigned len, unsigned flags,
1223                                 struct page **pagep, void **fsdata)
1224 {
1225         struct inode *inode = mapping->host;
1226         int ret;
1227         handle_t *handle;
1228         int retries = 0;
1229         struct page *page;
1230         pgoff_t index;
1231         unsigned from, to;
1232         /* Reserve one block more for addition to orphan list in case
1233          * we allocate blocks but write fails for some reason */
1234         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1235
1236         trace_ext3_write_begin(inode, pos, len, flags);
1237
1238         index = pos >> PAGE_CACHE_SHIFT;
1239         from = pos & (PAGE_CACHE_SIZE - 1);
1240         to = from + len;
1241
1242 retry:
1243         page = grab_cache_page_write_begin(mapping, index, flags);
1244         if (!page)
1245                 return -ENOMEM;
1246         *pagep = page;
1247
1248         handle = ext3_journal_start(inode, needed_blocks);
1249         if (IS_ERR(handle)) {
1250                 unlock_page(page);
1251                 page_cache_release(page);
1252                 ret = PTR_ERR(handle);
1253                 goto out;
1254         }
1255         ret = __block_write_begin(page, pos, len, ext3_get_block);
1256         if (ret)
1257                 goto write_begin_failed;
1258
1259         if (ext3_should_journal_data(inode)) {
1260                 ret = walk_page_buffers(handle, page_buffers(page),
1261                                 from, to, NULL, do_journal_get_write_access);
1262         }
1263 write_begin_failed:
1264         if (ret) {
1265                 /*
1266                  * block_write_begin may have instantiated a few blocks
1267                  * outside i_size.  Trim these off again. Don't need
1268                  * i_size_read because we hold i_mutex.
1269                  *
1270                  * Add inode to orphan list in case we crash before truncate
1271                  * finishes. Do this only if ext3_can_truncate() agrees so
1272                  * that orphan processing code is happy.
1273                  */
1274                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1275                         ext3_orphan_add(handle, inode);
1276                 ext3_journal_stop(handle);
1277                 unlock_page(page);
1278                 page_cache_release(page);
1279                 if (pos + len > inode->i_size)
1280                         ext3_truncate_failed_write(inode);
1281         }
1282         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1283                 goto retry;
1284 out:
1285         return ret;
1286 }
1287
1288
1289 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1290 {
1291         int err = journal_dirty_data(handle, bh);
1292         if (err)
1293                 ext3_journal_abort_handle(__func__, __func__,
1294                                                 bh, handle, err);
1295         return err;
1296 }
1297
1298 /* For ordered writepage and write_end functions */
1299 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1300 {
1301         /*
1302          * Write could have mapped the buffer but it didn't copy the data in
1303          * yet. So avoid filing such buffer into a transaction.
1304          */
1305         if (buffer_mapped(bh) && buffer_uptodate(bh))
1306                 return ext3_journal_dirty_data(handle, bh);
1307         return 0;
1308 }
1309
1310 /* For write_end() in data=journal mode */
1311 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1312 {
1313         if (!buffer_mapped(bh) || buffer_freed(bh))
1314                 return 0;
1315         set_buffer_uptodate(bh);
1316         return ext3_journal_dirty_metadata(handle, bh);
1317 }
1318
1319 /*
1320  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1321  * for the whole page but later we failed to copy the data in. Update inode
1322  * size according to what we managed to copy. The rest is going to be
1323  * truncated in write_end function.
1324  */
1325 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1326 {
1327         /* What matters to us is i_disksize. We don't write i_size anywhere */
1328         if (pos + copied > inode->i_size)
1329                 i_size_write(inode, pos + copied);
1330         if (pos + copied > EXT3_I(inode)->i_disksize) {
1331                 EXT3_I(inode)->i_disksize = pos + copied;
1332                 mark_inode_dirty(inode);
1333         }
1334 }
1335
1336 /*
1337  * We need to pick up the new inode size which generic_commit_write gave us
1338  * `file' can be NULL - eg, when called from page_symlink().
1339  *
1340  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1341  * buffers are managed internally.
1342  */
1343 static int ext3_ordered_write_end(struct file *file,
1344                                 struct address_space *mapping,
1345                                 loff_t pos, unsigned len, unsigned copied,
1346                                 struct page *page, void *fsdata)
1347 {
1348         handle_t *handle = ext3_journal_current_handle();
1349         struct inode *inode = file->f_mapping->host;
1350         unsigned from, to;
1351         int ret = 0, ret2;
1352
1353         trace_ext3_ordered_write_end(inode, pos, len, copied);
1354         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1355
1356         from = pos & (PAGE_CACHE_SIZE - 1);
1357         to = from + copied;
1358         ret = walk_page_buffers(handle, page_buffers(page),
1359                 from, to, NULL, journal_dirty_data_fn);
1360
1361         if (ret == 0)
1362                 update_file_sizes(inode, pos, copied);
1363         /*
1364          * There may be allocated blocks outside of i_size because
1365          * we failed to copy some data. Prepare for truncate.
1366          */
1367         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1368                 ext3_orphan_add(handle, inode);
1369         ret2 = ext3_journal_stop(handle);
1370         if (!ret)
1371                 ret = ret2;
1372         unlock_page(page);
1373         page_cache_release(page);
1374
1375         if (pos + len > inode->i_size)
1376                 ext3_truncate_failed_write(inode);
1377         return ret ? ret : copied;
1378 }
1379
1380 static int ext3_writeback_write_end(struct file *file,
1381                                 struct address_space *mapping,
1382                                 loff_t pos, unsigned len, unsigned copied,
1383                                 struct page *page, void *fsdata)
1384 {
1385         handle_t *handle = ext3_journal_current_handle();
1386         struct inode *inode = file->f_mapping->host;
1387         int ret;
1388
1389         trace_ext3_writeback_write_end(inode, pos, len, copied);
1390         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1391         update_file_sizes(inode, pos, copied);
1392         /*
1393          * There may be allocated blocks outside of i_size because
1394          * we failed to copy some data. Prepare for truncate.
1395          */
1396         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1397                 ext3_orphan_add(handle, inode);
1398         ret = ext3_journal_stop(handle);
1399         unlock_page(page);
1400         page_cache_release(page);
1401
1402         if (pos + len > inode->i_size)
1403                 ext3_truncate_failed_write(inode);
1404         return ret ? ret : copied;
1405 }
1406
1407 static int ext3_journalled_write_end(struct file *file,
1408                                 struct address_space *mapping,
1409                                 loff_t pos, unsigned len, unsigned copied,
1410                                 struct page *page, void *fsdata)
1411 {
1412         handle_t *handle = ext3_journal_current_handle();
1413         struct inode *inode = mapping->host;
1414         int ret = 0, ret2;
1415         int partial = 0;
1416         unsigned from, to;
1417
1418         trace_ext3_journalled_write_end(inode, pos, len, copied);
1419         from = pos & (PAGE_CACHE_SIZE - 1);
1420         to = from + len;
1421
1422         if (copied < len) {
1423                 if (!PageUptodate(page))
1424                         copied = 0;
1425                 page_zero_new_buffers(page, from + copied, to);
1426                 to = from + copied;
1427         }
1428
1429         ret = walk_page_buffers(handle, page_buffers(page), from,
1430                                 to, &partial, write_end_fn);
1431         if (!partial)
1432                 SetPageUptodate(page);
1433
1434         if (pos + copied > inode->i_size)
1435                 i_size_write(inode, pos + copied);
1436         /*
1437          * There may be allocated blocks outside of i_size because
1438          * we failed to copy some data. Prepare for truncate.
1439          */
1440         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1441                 ext3_orphan_add(handle, inode);
1442         ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1443         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1444                 EXT3_I(inode)->i_disksize = inode->i_size;
1445                 ret2 = ext3_mark_inode_dirty(handle, inode);
1446                 if (!ret)
1447                         ret = ret2;
1448         }
1449
1450         ret2 = ext3_journal_stop(handle);
1451         if (!ret)
1452                 ret = ret2;
1453         unlock_page(page);
1454         page_cache_release(page);
1455
1456         if (pos + len > inode->i_size)
1457                 ext3_truncate_failed_write(inode);
1458         return ret ? ret : copied;
1459 }
1460
1461 /*
1462  * bmap() is special.  It gets used by applications such as lilo and by
1463  * the swapper to find the on-disk block of a specific piece of data.
1464  *
1465  * Naturally, this is dangerous if the block concerned is still in the
1466  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1467  * filesystem and enables swap, then they may get a nasty shock when the
1468  * data getting swapped to that swapfile suddenly gets overwritten by
1469  * the original zero's written out previously to the journal and
1470  * awaiting writeback in the kernel's buffer cache.
1471  *
1472  * So, if we see any bmap calls here on a modified, data-journaled file,
1473  * take extra steps to flush any blocks which might be in the cache.
1474  */
1475 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1476 {
1477         struct inode *inode = mapping->host;
1478         journal_t *journal;
1479         int err;
1480
1481         if (ext3_test_inode_state(inode, EXT3_STATE_JDATA)) {
1482                 /*
1483                  * This is a REALLY heavyweight approach, but the use of
1484                  * bmap on dirty files is expected to be extremely rare:
1485                  * only if we run lilo or swapon on a freshly made file
1486                  * do we expect this to happen.
1487                  *
1488                  * (bmap requires CAP_SYS_RAWIO so this does not
1489                  * represent an unprivileged user DOS attack --- we'd be
1490                  * in trouble if mortal users could trigger this path at
1491                  * will.)
1492                  *
1493                  * NB. EXT3_STATE_JDATA is not set on files other than
1494                  * regular files.  If somebody wants to bmap a directory
1495                  * or symlink and gets confused because the buffer
1496                  * hasn't yet been flushed to disk, they deserve
1497                  * everything they get.
1498                  */
1499
1500                 ext3_clear_inode_state(inode, EXT3_STATE_JDATA);
1501                 journal = EXT3_JOURNAL(inode);
1502                 journal_lock_updates(journal);
1503                 err = journal_flush(journal);
1504                 journal_unlock_updates(journal);
1505
1506                 if (err)
1507                         return 0;
1508         }
1509
1510         return generic_block_bmap(mapping,block,ext3_get_block);
1511 }
1512
1513 static int bget_one(handle_t *handle, struct buffer_head *bh)
1514 {
1515         get_bh(bh);
1516         return 0;
1517 }
1518
1519 static int bput_one(handle_t *handle, struct buffer_head *bh)
1520 {
1521         put_bh(bh);
1522         return 0;
1523 }
1524
1525 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1526 {
1527         return !buffer_mapped(bh);
1528 }
1529
1530 /*
1531  * Note that we always start a transaction even if we're not journalling
1532  * data.  This is to preserve ordering: any hole instantiation within
1533  * __block_write_full_page -> ext3_get_block() should be journalled
1534  * along with the data so we don't crash and then get metadata which
1535  * refers to old data.
1536  *
1537  * In all journalling modes block_write_full_page() will start the I/O.
1538  *
1539  * Problem:
1540  *
1541  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1542  *              ext3_writepage()
1543  *
1544  * Similar for:
1545  *
1546  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1547  *
1548  * Same applies to ext3_get_block().  We will deadlock on various things like
1549  * lock_journal and i_truncate_mutex.
1550  *
1551  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1552  * allocations fail.
1553  *
1554  * 16May01: If we're reentered then journal_current_handle() will be
1555  *          non-zero. We simply *return*.
1556  *
1557  * 1 July 2001: @@@ FIXME:
1558  *   In journalled data mode, a data buffer may be metadata against the
1559  *   current transaction.  But the same file is part of a shared mapping
1560  *   and someone does a writepage() on it.
1561  *
1562  *   We will move the buffer onto the async_data list, but *after* it has
1563  *   been dirtied. So there's a small window where we have dirty data on
1564  *   BJ_Metadata.
1565  *
1566  *   Note that this only applies to the last partial page in the file.  The
1567  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1568  *   broken code anyway: it's wrong for msync()).
1569  *
1570  *   It's a rare case: affects the final partial page, for journalled data
1571  *   where the file is subject to bith write() and writepage() in the same
1572  *   transction.  To fix it we'll need a custom block_write_full_page().
1573  *   We'll probably need that anyway for journalling writepage() output.
1574  *
1575  * We don't honour synchronous mounts for writepage().  That would be
1576  * disastrous.  Any write() or metadata operation will sync the fs for
1577  * us.
1578  *
1579  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1580  * we don't need to open a transaction here.
1581  */
1582 static int ext3_ordered_writepage(struct page *page,
1583                                 struct writeback_control *wbc)
1584 {
1585         struct inode *inode = page->mapping->host;
1586         struct buffer_head *page_bufs;
1587         handle_t *handle = NULL;
1588         int ret = 0;
1589         int err;
1590
1591         J_ASSERT(PageLocked(page));
1592         WARN_ON_ONCE(IS_RDONLY(inode));
1593
1594         /*
1595          * We give up here if we're reentered, because it might be for a
1596          * different filesystem.
1597          */
1598         if (ext3_journal_current_handle())
1599                 goto out_fail;
1600
1601         trace_ext3_ordered_writepage(page);
1602         if (!page_has_buffers(page)) {
1603                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1604                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1605                 page_bufs = page_buffers(page);
1606         } else {
1607                 page_bufs = page_buffers(page);
1608                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1609                                        NULL, buffer_unmapped)) {
1610                         /* Provide NULL get_block() to catch bugs if buffers
1611                          * weren't really mapped */
1612                         return block_write_full_page(page, NULL, wbc);
1613                 }
1614         }
1615         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1616
1617         if (IS_ERR(handle)) {
1618                 ret = PTR_ERR(handle);
1619                 goto out_fail;
1620         }
1621
1622         walk_page_buffers(handle, page_bufs, 0,
1623                         PAGE_CACHE_SIZE, NULL, bget_one);
1624
1625         ret = block_write_full_page(page, ext3_get_block, wbc);
1626
1627         /*
1628          * The page can become unlocked at any point now, and
1629          * truncate can then come in and change things.  So we
1630          * can't touch *page from now on.  But *page_bufs is
1631          * safe due to elevated refcount.
1632          */
1633
1634         /*
1635          * And attach them to the current transaction.  But only if
1636          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1637          * and generally junk.
1638          */
1639         if (ret == 0) {
1640                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1641                                         NULL, journal_dirty_data_fn);
1642                 if (!ret)
1643                         ret = err;
1644         }
1645         walk_page_buffers(handle, page_bufs, 0,
1646                         PAGE_CACHE_SIZE, NULL, bput_one);
1647         err = ext3_journal_stop(handle);
1648         if (!ret)
1649                 ret = err;
1650         return ret;
1651
1652 out_fail:
1653         redirty_page_for_writepage(wbc, page);
1654         unlock_page(page);
1655         return ret;
1656 }
1657
1658 static int ext3_writeback_writepage(struct page *page,
1659                                 struct writeback_control *wbc)
1660 {
1661         struct inode *inode = page->mapping->host;
1662         handle_t *handle = NULL;
1663         int ret = 0;
1664         int err;
1665
1666         J_ASSERT(PageLocked(page));
1667         WARN_ON_ONCE(IS_RDONLY(inode));
1668
1669         if (ext3_journal_current_handle())
1670                 goto out_fail;
1671
1672         trace_ext3_writeback_writepage(page);
1673         if (page_has_buffers(page)) {
1674                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1675                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1676                         /* Provide NULL get_block() to catch bugs if buffers
1677                          * weren't really mapped */
1678                         return block_write_full_page(page, NULL, wbc);
1679                 }
1680         }
1681
1682         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1683         if (IS_ERR(handle)) {
1684                 ret = PTR_ERR(handle);
1685                 goto out_fail;
1686         }
1687
1688         ret = block_write_full_page(page, ext3_get_block, wbc);
1689
1690         err = ext3_journal_stop(handle);
1691         if (!ret)
1692                 ret = err;
1693         return ret;
1694
1695 out_fail:
1696         redirty_page_for_writepage(wbc, page);
1697         unlock_page(page);
1698         return ret;
1699 }
1700
1701 static int ext3_journalled_writepage(struct page *page,
1702                                 struct writeback_control *wbc)
1703 {
1704         struct inode *inode = page->mapping->host;
1705         handle_t *handle = NULL;
1706         int ret = 0;
1707         int err;
1708
1709         J_ASSERT(PageLocked(page));
1710         WARN_ON_ONCE(IS_RDONLY(inode));
1711
1712         if (ext3_journal_current_handle())
1713                 goto no_write;
1714
1715         trace_ext3_journalled_writepage(page);
1716         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1717         if (IS_ERR(handle)) {
1718                 ret = PTR_ERR(handle);
1719                 goto no_write;
1720         }
1721
1722         if (!page_has_buffers(page) || PageChecked(page)) {
1723                 /*
1724                  * It's mmapped pagecache.  Add buffers and journal it.  There
1725                  * doesn't seem much point in redirtying the page here.
1726                  */
1727                 ClearPageChecked(page);
1728                 ret = __block_write_begin(page, 0, PAGE_CACHE_SIZE,
1729                                           ext3_get_block);
1730                 if (ret != 0) {
1731                         ext3_journal_stop(handle);
1732                         goto out_unlock;
1733                 }
1734                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1735                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1736
1737                 err = walk_page_buffers(handle, page_buffers(page), 0,
1738                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1739                 if (ret == 0)
1740                         ret = err;
1741                 ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1742                 unlock_page(page);
1743         } else {
1744                 /*
1745                  * It may be a page full of checkpoint-mode buffers.  We don't
1746                  * really know unless we go poke around in the buffer_heads.
1747                  * But block_write_full_page will do the right thing.
1748                  */
1749                 ret = block_write_full_page(page, ext3_get_block, wbc);
1750         }
1751         err = ext3_journal_stop(handle);
1752         if (!ret)
1753                 ret = err;
1754 out:
1755         return ret;
1756
1757 no_write:
1758         redirty_page_for_writepage(wbc, page);
1759 out_unlock:
1760         unlock_page(page);
1761         goto out;
1762 }
1763
1764 static int ext3_readpage(struct file *file, struct page *page)
1765 {
1766         trace_ext3_readpage(page);
1767         return mpage_readpage(page, ext3_get_block);
1768 }
1769
1770 static int
1771 ext3_readpages(struct file *file, struct address_space *mapping,
1772                 struct list_head *pages, unsigned nr_pages)
1773 {
1774         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1775 }
1776
1777 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1778 {
1779         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1780
1781         trace_ext3_invalidatepage(page, offset);
1782
1783         /*
1784          * If it's a full truncate we just forget about the pending dirtying
1785          */
1786         if (offset == 0)
1787                 ClearPageChecked(page);
1788
1789         journal_invalidatepage(journal, page, offset);
1790 }
1791
1792 static int ext3_releasepage(struct page *page, gfp_t wait)
1793 {
1794         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1795
1796         trace_ext3_releasepage(page);
1797         WARN_ON(PageChecked(page));
1798         if (!page_has_buffers(page))
1799                 return 0;
1800         return journal_try_to_free_buffers(journal, page, wait);
1801 }
1802
1803 /*
1804  * If the O_DIRECT write will extend the file then add this inode to the
1805  * orphan list.  So recovery will truncate it back to the original size
1806  * if the machine crashes during the write.
1807  *
1808  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1809  * crashes then stale disk data _may_ be exposed inside the file. But current
1810  * VFS code falls back into buffered path in that case so we are safe.
1811  */
1812 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1813                         const struct iovec *iov, loff_t offset,
1814                         unsigned long nr_segs)
1815 {
1816         struct file *file = iocb->ki_filp;
1817         struct inode *inode = file->f_mapping->host;
1818         struct ext3_inode_info *ei = EXT3_I(inode);
1819         handle_t *handle;
1820         ssize_t ret;
1821         int orphan = 0;
1822         size_t count = iov_length(iov, nr_segs);
1823         int retries = 0;
1824
1825         trace_ext3_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
1826
1827         if (rw == WRITE) {
1828                 loff_t final_size = offset + count;
1829
1830                 if (final_size > inode->i_size) {
1831                         /* Credits for sb + inode write */
1832                         handle = ext3_journal_start(inode, 2);
1833                         if (IS_ERR(handle)) {
1834                                 ret = PTR_ERR(handle);
1835                                 goto out;
1836                         }
1837                         ret = ext3_orphan_add(handle, inode);
1838                         if (ret) {
1839                                 ext3_journal_stop(handle);
1840                                 goto out;
1841                         }
1842                         orphan = 1;
1843                         ei->i_disksize = inode->i_size;
1844                         ext3_journal_stop(handle);
1845                 }
1846         }
1847
1848 retry:
1849         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1850                                  offset, nr_segs,
1851                                  ext3_get_block, NULL);
1852         /*
1853          * In case of error extending write may have instantiated a few
1854          * blocks outside i_size. Trim these off again.
1855          */
1856         if (unlikely((rw & WRITE) && ret < 0)) {
1857                 loff_t isize = i_size_read(inode);
1858                 loff_t end = offset + iov_length(iov, nr_segs);
1859
1860                 if (end > isize)
1861                         ext3_truncate_failed_direct_write(inode);
1862         }
1863         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1864                 goto retry;
1865
1866         if (orphan) {
1867                 int err;
1868
1869                 /* Credits for sb + inode write */
1870                 handle = ext3_journal_start(inode, 2);
1871                 if (IS_ERR(handle)) {
1872                         /* This is really bad luck. We've written the data
1873                          * but cannot extend i_size. Truncate allocated blocks
1874                          * and pretend the write failed... */
1875                         ext3_truncate_failed_direct_write(inode);
1876                         ret = PTR_ERR(handle);
1877                         goto out;
1878                 }
1879                 if (inode->i_nlink)
1880                         ext3_orphan_del(handle, inode);
1881                 if (ret > 0) {
1882                         loff_t end = offset + ret;
1883                         if (end > inode->i_size) {
1884                                 ei->i_disksize = end;
1885                                 i_size_write(inode, end);
1886                                 /*
1887                                  * We're going to return a positive `ret'
1888                                  * here due to non-zero-length I/O, so there's
1889                                  * no way of reporting error returns from
1890                                  * ext3_mark_inode_dirty() to userspace.  So
1891                                  * ignore it.
1892                                  */
1893                                 ext3_mark_inode_dirty(handle, inode);
1894                         }
1895                 }
1896                 err = ext3_journal_stop(handle);
1897                 if (ret == 0)
1898                         ret = err;
1899         }
1900 out:
1901         trace_ext3_direct_IO_exit(inode, offset,
1902                                 iov_length(iov, nr_segs), rw, ret);
1903         return ret;
1904 }
1905
1906 /*
1907  * Pages can be marked dirty completely asynchronously from ext3's journalling
1908  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1909  * much here because ->set_page_dirty is called under VFS locks.  The page is
1910  * not necessarily locked.
1911  *
1912  * We cannot just dirty the page and leave attached buffers clean, because the
1913  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1914  * or jbddirty because all the journalling code will explode.
1915  *
1916  * So what we do is to mark the page "pending dirty" and next time writepage
1917  * is called, propagate that into the buffers appropriately.
1918  */
1919 static int ext3_journalled_set_page_dirty(struct page *page)
1920 {
1921         SetPageChecked(page);
1922         return __set_page_dirty_nobuffers(page);
1923 }
1924
1925 static const struct address_space_operations ext3_ordered_aops = {
1926         .readpage               = ext3_readpage,
1927         .readpages              = ext3_readpages,
1928         .writepage              = ext3_ordered_writepage,
1929         .write_begin            = ext3_write_begin,
1930         .write_end              = ext3_ordered_write_end,
1931         .bmap                   = ext3_bmap,
1932         .invalidatepage         = ext3_invalidatepage,
1933         .releasepage            = ext3_releasepage,
1934         .direct_IO              = ext3_direct_IO,
1935         .migratepage            = buffer_migrate_page,
1936         .is_partially_uptodate  = block_is_partially_uptodate,
1937         .error_remove_page      = generic_error_remove_page,
1938 };
1939
1940 static const struct address_space_operations ext3_writeback_aops = {
1941         .readpage               = ext3_readpage,
1942         .readpages              = ext3_readpages,
1943         .writepage              = ext3_writeback_writepage,
1944         .write_begin            = ext3_write_begin,
1945         .write_end              = ext3_writeback_write_end,
1946         .bmap                   = ext3_bmap,
1947         .invalidatepage         = ext3_invalidatepage,
1948         .releasepage            = ext3_releasepage,
1949         .direct_IO              = ext3_direct_IO,
1950         .migratepage            = buffer_migrate_page,
1951         .is_partially_uptodate  = block_is_partially_uptodate,
1952         .error_remove_page      = generic_error_remove_page,
1953 };
1954
1955 static const struct address_space_operations ext3_journalled_aops = {
1956         .readpage               = ext3_readpage,
1957         .readpages              = ext3_readpages,
1958         .writepage              = ext3_journalled_writepage,
1959         .write_begin            = ext3_write_begin,
1960         .write_end              = ext3_journalled_write_end,
1961         .set_page_dirty         = ext3_journalled_set_page_dirty,
1962         .bmap                   = ext3_bmap,
1963         .invalidatepage         = ext3_invalidatepage,
1964         .releasepage            = ext3_releasepage,
1965         .is_partially_uptodate  = block_is_partially_uptodate,
1966         .error_remove_page      = generic_error_remove_page,
1967 };
1968
1969 void ext3_set_aops(struct inode *inode)
1970 {
1971         if (ext3_should_order_data(inode))
1972                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1973         else if (ext3_should_writeback_data(inode))
1974                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1975         else
1976                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1977 }
1978
1979 /*
1980  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1981  * up to the end of the block which corresponds to `from'.
1982  * This required during truncate. We need to physically zero the tail end
1983  * of that block so it doesn't yield old data if the file is later grown.
1984  */
1985 static int ext3_block_truncate_page(struct inode *inode, loff_t from)
1986 {
1987         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1988         unsigned offset = from & (PAGE_CACHE_SIZE - 1);
1989         unsigned blocksize, iblock, length, pos;
1990         struct page *page;
1991         handle_t *handle = NULL;
1992         struct buffer_head *bh;
1993         int err = 0;
1994
1995         /* Truncated on block boundary - nothing to do */
1996         blocksize = inode->i_sb->s_blocksize;
1997         if ((from & (blocksize - 1)) == 0)
1998                 return 0;
1999
2000         page = grab_cache_page(inode->i_mapping, index);
2001         if (!page)
2002                 return -ENOMEM;
2003         length = blocksize - (offset & (blocksize - 1));
2004         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
2005
2006         if (!page_has_buffers(page))
2007                 create_empty_buffers(page, blocksize, 0);
2008
2009         /* Find the buffer that contains "offset" */
2010         bh = page_buffers(page);
2011         pos = blocksize;
2012         while (offset >= pos) {
2013                 bh = bh->b_this_page;
2014                 iblock++;
2015                 pos += blocksize;
2016         }
2017
2018         err = 0;
2019         if (buffer_freed(bh)) {
2020                 BUFFER_TRACE(bh, "freed: skip");
2021                 goto unlock;
2022         }
2023
2024         if (!buffer_mapped(bh)) {
2025                 BUFFER_TRACE(bh, "unmapped");
2026                 ext3_get_block(inode, iblock, bh, 0);
2027                 /* unmapped? It's a hole - nothing to do */
2028                 if (!buffer_mapped(bh)) {
2029                         BUFFER_TRACE(bh, "still unmapped");
2030                         goto unlock;
2031                 }
2032         }
2033
2034         /* Ok, it's mapped. Make sure it's up-to-date */
2035         if (PageUptodate(page))
2036                 set_buffer_uptodate(bh);
2037
2038         if (!buffer_uptodate(bh)) {
2039                 err = -EIO;
2040                 ll_rw_block(READ, 1, &bh);
2041                 wait_on_buffer(bh);
2042                 /* Uhhuh. Read error. Complain and punt. */
2043                 if (!buffer_uptodate(bh))
2044                         goto unlock;
2045         }
2046
2047         /* data=writeback mode doesn't need transaction to zero-out data */
2048         if (!ext3_should_writeback_data(inode)) {
2049                 /* We journal at most one block */
2050                 handle = ext3_journal_start(inode, 1);
2051                 if (IS_ERR(handle)) {
2052                         clear_highpage(page);
2053                         flush_dcache_page(page);
2054                         err = PTR_ERR(handle);
2055                         goto unlock;
2056                 }
2057         }
2058
2059         if (ext3_should_journal_data(inode)) {
2060                 BUFFER_TRACE(bh, "get write access");
2061                 err = ext3_journal_get_write_access(handle, bh);
2062                 if (err)
2063                         goto stop;
2064         }
2065
2066         zero_user(page, offset, length);
2067         BUFFER_TRACE(bh, "zeroed end of block");
2068
2069         err = 0;
2070         if (ext3_should_journal_data(inode)) {
2071                 err = ext3_journal_dirty_metadata(handle, bh);
2072         } else {
2073                 if (ext3_should_order_data(inode))
2074                         err = ext3_journal_dirty_data(handle, bh);
2075                 mark_buffer_dirty(bh);
2076         }
2077 stop:
2078         if (handle)
2079                 ext3_journal_stop(handle);
2080
2081 unlock:
2082         unlock_page(page);
2083         page_cache_release(page);
2084         return err;
2085 }
2086
2087 /*
2088  * Probably it should be a library function... search for first non-zero word
2089  * or memcmp with zero_page, whatever is better for particular architecture.
2090  * Linus?
2091  */
2092 static inline int all_zeroes(__le32 *p, __le32 *q)
2093 {
2094         while (p < q)
2095                 if (*p++)
2096                         return 0;
2097         return 1;
2098 }
2099
2100 /**
2101  *      ext3_find_shared - find the indirect blocks for partial truncation.
2102  *      @inode:   inode in question
2103  *      @depth:   depth of the affected branch
2104  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
2105  *      @chain:   place to store the pointers to partial indirect blocks
2106  *      @top:     place to the (detached) top of branch
2107  *
2108  *      This is a helper function used by ext3_truncate().
2109  *
2110  *      When we do truncate() we may have to clean the ends of several
2111  *      indirect blocks but leave the blocks themselves alive. Block is
2112  *      partially truncated if some data below the new i_size is referred
2113  *      from it (and it is on the path to the first completely truncated
2114  *      data block, indeed).  We have to free the top of that path along
2115  *      with everything to the right of the path. Since no allocation
2116  *      past the truncation point is possible until ext3_truncate()
2117  *      finishes, we may safely do the latter, but top of branch may
2118  *      require special attention - pageout below the truncation point
2119  *      might try to populate it.
2120  *
2121  *      We atomically detach the top of branch from the tree, store the
2122  *      block number of its root in *@top, pointers to buffer_heads of
2123  *      partially truncated blocks - in @chain[].bh and pointers to
2124  *      their last elements that should not be removed - in
2125  *      @chain[].p. Return value is the pointer to last filled element
2126  *      of @chain.
2127  *
2128  *      The work left to caller to do the actual freeing of subtrees:
2129  *              a) free the subtree starting from *@top
2130  *              b) free the subtrees whose roots are stored in
2131  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2132  *              c) free the subtrees growing from the inode past the @chain[0].
2133  *                      (no partially truncated stuff there).  */
2134
2135 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2136                         int offsets[4], Indirect chain[4], __le32 *top)
2137 {
2138         Indirect *partial, *p;
2139         int k, err;
2140
2141         *top = 0;
2142         /* Make k index the deepest non-null offset + 1 */
2143         for (k = depth; k > 1 && !offsets[k-1]; k--)
2144                 ;
2145         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2146         /* Writer: pointers */
2147         if (!partial)
2148                 partial = chain + k-1;
2149         /*
2150          * If the branch acquired continuation since we've looked at it -
2151          * fine, it should all survive and (new) top doesn't belong to us.
2152          */
2153         if (!partial->key && *partial->p)
2154                 /* Writer: end */
2155                 goto no_top;
2156         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2157                 ;
2158         /*
2159          * OK, we've found the last block that must survive. The rest of our
2160          * branch should be detached before unlocking. However, if that rest
2161          * of branch is all ours and does not grow immediately from the inode
2162          * it's easier to cheat and just decrement partial->p.
2163          */
2164         if (p == chain + k - 1 && p > chain) {
2165                 p->p--;
2166         } else {
2167                 *top = *p->p;
2168                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2169 #if 0
2170                 *p->p = 0;
2171 #endif
2172         }
2173         /* Writer: end */
2174
2175         while(partial > p) {
2176                 brelse(partial->bh);
2177                 partial--;
2178         }
2179 no_top:
2180         return partial;
2181 }
2182
2183 /*
2184  * Zero a number of block pointers in either an inode or an indirect block.
2185  * If we restart the transaction we must again get write access to the
2186  * indirect block for further modification.
2187  *
2188  * We release `count' blocks on disk, but (last - first) may be greater
2189  * than `count' because there can be holes in there.
2190  */
2191 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2192                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2193                 unsigned long count, __le32 *first, __le32 *last)
2194 {
2195         __le32 *p;
2196         if (try_to_extend_transaction(handle, inode)) {
2197                 if (bh) {
2198                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2199                         if (ext3_journal_dirty_metadata(handle, bh))
2200                                 return;
2201                 }
2202                 ext3_mark_inode_dirty(handle, inode);
2203                 truncate_restart_transaction(handle, inode);
2204                 if (bh) {
2205                         BUFFER_TRACE(bh, "retaking write access");
2206                         if (ext3_journal_get_write_access(handle, bh))
2207                                 return;
2208                 }
2209         }
2210
2211         /*
2212          * Any buffers which are on the journal will be in memory. We find
2213          * them on the hash table so journal_revoke() will run journal_forget()
2214          * on them.  We've already detached each block from the file, so
2215          * bforget() in journal_forget() should be safe.
2216          *
2217          * AKPM: turn on bforget in journal_forget()!!!
2218          */
2219         for (p = first; p < last; p++) {
2220                 u32 nr = le32_to_cpu(*p);
2221                 if (nr) {
2222                         struct buffer_head *bh;
2223
2224                         *p = 0;
2225                         bh = sb_find_get_block(inode->i_sb, nr);
2226                         ext3_forget(handle, 0, inode, bh, nr);
2227                 }
2228         }
2229
2230         ext3_free_blocks(handle, inode, block_to_free, count);
2231 }
2232
2233 /**
2234  * ext3_free_data - free a list of data blocks
2235  * @handle:     handle for this transaction
2236  * @inode:      inode we are dealing with
2237  * @this_bh:    indirect buffer_head which contains *@first and *@last
2238  * @first:      array of block numbers
2239  * @last:       points immediately past the end of array
2240  *
2241  * We are freeing all blocks referred from that array (numbers are stored as
2242  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2243  *
2244  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2245  * blocks are contiguous then releasing them at one time will only affect one
2246  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2247  * actually use a lot of journal space.
2248  *
2249  * @this_bh will be %NULL if @first and @last point into the inode's direct
2250  * block pointers.
2251  */
2252 static void ext3_free_data(handle_t *handle, struct inode *inode,
2253                            struct buffer_head *this_bh,
2254                            __le32 *first, __le32 *last)
2255 {
2256         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2257         unsigned long count = 0;            /* Number of blocks in the run */
2258         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2259                                                corresponding to
2260                                                block_to_free */
2261         ext3_fsblk_t nr;                    /* Current block # */
2262         __le32 *p;                          /* Pointer into inode/ind
2263                                                for current block */
2264         int err;
2265
2266         if (this_bh) {                          /* For indirect block */
2267                 BUFFER_TRACE(this_bh, "get_write_access");
2268                 err = ext3_journal_get_write_access(handle, this_bh);
2269                 /* Important: if we can't update the indirect pointers
2270                  * to the blocks, we can't free them. */
2271                 if (err)
2272                         return;
2273         }
2274
2275         for (p = first; p < last; p++) {
2276                 nr = le32_to_cpu(*p);
2277                 if (nr) {
2278                         /* accumulate blocks to free if they're contiguous */
2279                         if (count == 0) {
2280                                 block_to_free = nr;
2281                                 block_to_free_p = p;
2282                                 count = 1;
2283                         } else if (nr == block_to_free + count) {
2284                                 count++;
2285                         } else {
2286                                 ext3_clear_blocks(handle, inode, this_bh,
2287                                                   block_to_free,
2288                                                   count, block_to_free_p, p);
2289                                 block_to_free = nr;
2290                                 block_to_free_p = p;
2291                                 count = 1;
2292                         }
2293                 }
2294         }
2295
2296         if (count > 0)
2297                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2298                                   count, block_to_free_p, p);
2299
2300         if (this_bh) {
2301                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2302
2303                 /*
2304                  * The buffer head should have an attached journal head at this
2305                  * point. However, if the data is corrupted and an indirect
2306                  * block pointed to itself, it would have been detached when
2307                  * the block was cleared. Check for this instead of OOPSing.
2308                  */
2309                 if (bh2jh(this_bh))
2310                         ext3_journal_dirty_metadata(handle, this_bh);
2311                 else
2312                         ext3_error(inode->i_sb, "ext3_free_data",
2313                                    "circular indirect block detected, "
2314                                    "inode=%lu, block=%llu",
2315                                    inode->i_ino,
2316                                    (unsigned long long)this_bh->b_blocknr);
2317         }
2318 }
2319
2320 /**
2321  *      ext3_free_branches - free an array of branches
2322  *      @handle: JBD handle for this transaction
2323  *      @inode: inode we are dealing with
2324  *      @parent_bh: the buffer_head which contains *@first and *@last
2325  *      @first: array of block numbers
2326  *      @last:  pointer immediately past the end of array
2327  *      @depth: depth of the branches to free
2328  *
2329  *      We are freeing all blocks referred from these branches (numbers are
2330  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2331  *      appropriately.
2332  */
2333 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2334                                struct buffer_head *parent_bh,
2335                                __le32 *first, __le32 *last, int depth)
2336 {
2337         ext3_fsblk_t nr;
2338         __le32 *p;
2339
2340         if (is_handle_aborted(handle))
2341                 return;
2342
2343         if (depth--) {
2344                 struct buffer_head *bh;
2345                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2346                 p = last;
2347                 while (--p >= first) {
2348                         nr = le32_to_cpu(*p);
2349                         if (!nr)
2350                                 continue;               /* A hole */
2351
2352                         /* Go read the buffer for the next level down */
2353                         bh = sb_bread(inode->i_sb, nr);
2354
2355                         /*
2356                          * A read failure? Report error and clear slot
2357                          * (should be rare).
2358                          */
2359                         if (!bh) {
2360                                 ext3_error(inode->i_sb, "ext3_free_branches",
2361                                            "Read failure, inode=%lu, block="E3FSBLK,
2362                                            inode->i_ino, nr);
2363                                 continue;
2364                         }
2365
2366                         /* This zaps the entire block.  Bottom up. */
2367                         BUFFER_TRACE(bh, "free child branches");
2368                         ext3_free_branches(handle, inode, bh,
2369                                            (__le32*)bh->b_data,
2370                                            (__le32*)bh->b_data + addr_per_block,
2371                                            depth);
2372
2373                         /*
2374                          * Everything below this this pointer has been
2375                          * released.  Now let this top-of-subtree go.
2376                          *
2377                          * We want the freeing of this indirect block to be
2378                          * atomic in the journal with the updating of the
2379                          * bitmap block which owns it.  So make some room in
2380                          * the journal.
2381                          *
2382                          * We zero the parent pointer *after* freeing its
2383                          * pointee in the bitmaps, so if extend_transaction()
2384                          * for some reason fails to put the bitmap changes and
2385                          * the release into the same transaction, recovery
2386                          * will merely complain about releasing a free block,
2387                          * rather than leaking blocks.
2388                          */
2389                         if (is_handle_aborted(handle))
2390                                 return;
2391                         if (try_to_extend_transaction(handle, inode)) {
2392                                 ext3_mark_inode_dirty(handle, inode);
2393                                 truncate_restart_transaction(handle, inode);
2394                         }
2395
2396                         /*
2397                          * We've probably journalled the indirect block several
2398                          * times during the truncate.  But it's no longer
2399                          * needed and we now drop it from the transaction via
2400                          * journal_revoke().
2401                          *
2402                          * That's easy if it's exclusively part of this
2403                          * transaction.  But if it's part of the committing
2404                          * transaction then journal_forget() will simply
2405                          * brelse() it.  That means that if the underlying
2406                          * block is reallocated in ext3_get_block(),
2407                          * unmap_underlying_metadata() will find this block
2408                          * and will try to get rid of it.  damn, damn. Thus
2409                          * we don't allow a block to be reallocated until
2410                          * a transaction freeing it has fully committed.
2411                          *
2412                          * We also have to make sure journal replay after a
2413                          * crash does not overwrite non-journaled data blocks
2414                          * with old metadata when the block got reallocated for
2415                          * data.  Thus we have to store a revoke record for a
2416                          * block in the same transaction in which we free the
2417                          * block.
2418                          */
2419                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2420
2421                         ext3_free_blocks(handle, inode, nr, 1);
2422
2423                         if (parent_bh) {
2424                                 /*
2425                                  * The block which we have just freed is
2426                                  * pointed to by an indirect block: journal it
2427                                  */
2428                                 BUFFER_TRACE(parent_bh, "get_write_access");
2429                                 if (!ext3_journal_get_write_access(handle,
2430                                                                    parent_bh)){
2431                                         *p = 0;
2432                                         BUFFER_TRACE(parent_bh,
2433                                         "call ext3_journal_dirty_metadata");
2434                                         ext3_journal_dirty_metadata(handle,
2435                                                                     parent_bh);
2436                                 }
2437                         }
2438                 }
2439         } else {
2440                 /* We have reached the bottom of the tree. */
2441                 BUFFER_TRACE(parent_bh, "free data blocks");
2442                 ext3_free_data(handle, inode, parent_bh, first, last);
2443         }
2444 }
2445
2446 int ext3_can_truncate(struct inode *inode)
2447 {
2448         if (S_ISREG(inode->i_mode))
2449                 return 1;
2450         if (S_ISDIR(inode->i_mode))
2451                 return 1;
2452         if (S_ISLNK(inode->i_mode))
2453                 return !ext3_inode_is_fast_symlink(inode);
2454         return 0;
2455 }
2456
2457 /*
2458  * ext3_truncate()
2459  *
2460  * We block out ext3_get_block() block instantiations across the entire
2461  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2462  * simultaneously on behalf of the same inode.
2463  *
2464  * As we work through the truncate and commmit bits of it to the journal there
2465  * is one core, guiding principle: the file's tree must always be consistent on
2466  * disk.  We must be able to restart the truncate after a crash.
2467  *
2468  * The file's tree may be transiently inconsistent in memory (although it
2469  * probably isn't), but whenever we close off and commit a journal transaction,
2470  * the contents of (the filesystem + the journal) must be consistent and
2471  * restartable.  It's pretty simple, really: bottom up, right to left (although
2472  * left-to-right works OK too).
2473  *
2474  * Note that at recovery time, journal replay occurs *before* the restart of
2475  * truncate against the orphan inode list.
2476  *
2477  * The committed inode has the new, desired i_size (which is the same as
2478  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2479  * that this inode's truncate did not complete and it will again call
2480  * ext3_truncate() to have another go.  So there will be instantiated blocks
2481  * to the right of the truncation point in a crashed ext3 filesystem.  But
2482  * that's fine - as long as they are linked from the inode, the post-crash
2483  * ext3_truncate() run will find them and release them.
2484  */
2485 void ext3_truncate(struct inode *inode)
2486 {
2487         handle_t *handle;
2488         struct ext3_inode_info *ei = EXT3_I(inode);
2489         __le32 *i_data = ei->i_data;
2490         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2491         int offsets[4];
2492         Indirect chain[4];
2493         Indirect *partial;
2494         __le32 nr = 0;
2495         int n;
2496         long last_block;
2497         unsigned blocksize = inode->i_sb->s_blocksize;
2498
2499         trace_ext3_truncate_enter(inode);
2500
2501         if (!ext3_can_truncate(inode))
2502                 goto out_notrans;
2503
2504         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2505                 ext3_set_inode_state(inode, EXT3_STATE_FLUSH_ON_CLOSE);
2506
2507         handle = start_transaction(inode);
2508         if (IS_ERR(handle))
2509                 goto out_notrans;
2510
2511         last_block = (inode->i_size + blocksize-1)
2512                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2513         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2514         if (n == 0)
2515                 goto out_stop;  /* error */
2516
2517         /*
2518          * OK.  This truncate is going to happen.  We add the inode to the
2519          * orphan list, so that if this truncate spans multiple transactions,
2520          * and we crash, we will resume the truncate when the filesystem
2521          * recovers.  It also marks the inode dirty, to catch the new size.
2522          *
2523          * Implication: the file must always be in a sane, consistent
2524          * truncatable state while each transaction commits.
2525          */
2526         if (ext3_orphan_add(handle, inode))
2527                 goto out_stop;
2528
2529         /*
2530          * The orphan list entry will now protect us from any crash which
2531          * occurs before the truncate completes, so it is now safe to propagate
2532          * the new, shorter inode size (held for now in i_size) into the
2533          * on-disk inode. We do this via i_disksize, which is the value which
2534          * ext3 *really* writes onto the disk inode.
2535          */
2536         ei->i_disksize = inode->i_size;
2537
2538         /*
2539          * From here we block out all ext3_get_block() callers who want to
2540          * modify the block allocation tree.
2541          */
2542         mutex_lock(&ei->truncate_mutex);
2543
2544         if (n == 1) {           /* direct blocks */
2545                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2546                                i_data + EXT3_NDIR_BLOCKS);
2547                 goto do_indirects;
2548         }
2549
2550         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2551         /* Kill the top of shared branch (not detached) */
2552         if (nr) {
2553                 if (partial == chain) {
2554                         /* Shared branch grows from the inode */
2555                         ext3_free_branches(handle, inode, NULL,
2556                                            &nr, &nr+1, (chain+n-1) - partial);
2557                         *partial->p = 0;
2558                         /*
2559                          * We mark the inode dirty prior to restart,
2560                          * and prior to stop.  No need for it here.
2561                          */
2562                 } else {
2563                         /* Shared branch grows from an indirect block */
2564                         ext3_free_branches(handle, inode, partial->bh,
2565                                         partial->p,
2566                                         partial->p+1, (chain+n-1) - partial);
2567                 }
2568         }
2569         /* Clear the ends of indirect blocks on the shared branch */
2570         while (partial > chain) {
2571                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2572                                    (__le32*)partial->bh->b_data+addr_per_block,
2573                                    (chain+n-1) - partial);
2574                 BUFFER_TRACE(partial->bh, "call brelse");
2575                 brelse (partial->bh);
2576                 partial--;
2577         }
2578 do_indirects:
2579         /* Kill the remaining (whole) subtrees */
2580         switch (offsets[0]) {
2581         default:
2582                 nr = i_data[EXT3_IND_BLOCK];
2583                 if (nr) {
2584                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2585                         i_data[EXT3_IND_BLOCK] = 0;
2586                 }
2587         case EXT3_IND_BLOCK:
2588                 nr = i_data[EXT3_DIND_BLOCK];
2589                 if (nr) {
2590                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2591                         i_data[EXT3_DIND_BLOCK] = 0;
2592                 }
2593         case EXT3_DIND_BLOCK:
2594                 nr = i_data[EXT3_TIND_BLOCK];
2595                 if (nr) {
2596                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2597                         i_data[EXT3_TIND_BLOCK] = 0;
2598                 }
2599         case EXT3_TIND_BLOCK:
2600                 ;
2601         }
2602
2603         ext3_discard_reservation(inode);
2604
2605         mutex_unlock(&ei->truncate_mutex);
2606         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2607         ext3_mark_inode_dirty(handle, inode);
2608
2609         /*
2610          * In a multi-transaction truncate, we only make the final transaction
2611          * synchronous
2612          */
2613         if (IS_SYNC(inode))
2614                 handle->h_sync = 1;
2615 out_stop:
2616         /*
2617          * If this was a simple ftruncate(), and the file will remain alive
2618          * then we need to clear up the orphan record which we created above.
2619          * However, if this was a real unlink then we were called by
2620          * ext3_evict_inode(), and we allow that function to clean up the
2621          * orphan info for us.
2622          */
2623         if (inode->i_nlink)
2624                 ext3_orphan_del(handle, inode);
2625
2626         ext3_journal_stop(handle);
2627         trace_ext3_truncate_exit(inode);
2628         return;
2629 out_notrans:
2630         /*
2631          * Delete the inode from orphan list so that it doesn't stay there
2632          * forever and trigger assertion on umount.
2633          */
2634         if (inode->i_nlink)
2635                 ext3_orphan_del(NULL, inode);
2636         trace_ext3_truncate_exit(inode);
2637 }
2638
2639 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2640                 unsigned long ino, struct ext3_iloc *iloc)
2641 {
2642         unsigned long block_group;
2643         unsigned long offset;
2644         ext3_fsblk_t block;
2645         struct ext3_group_desc *gdp;
2646
2647         if (!ext3_valid_inum(sb, ino)) {
2648                 /*
2649                  * This error is already checked for in namei.c unless we are
2650                  * looking at an NFS filehandle, in which case no error
2651                  * report is needed
2652                  */
2653                 return 0;
2654         }
2655
2656         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2657         gdp = ext3_get_group_desc(sb, block_group, NULL);
2658         if (!gdp)
2659                 return 0;
2660         /*
2661          * Figure out the offset within the block group inode table
2662          */
2663         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2664                 EXT3_INODE_SIZE(sb);
2665         block = le32_to_cpu(gdp->bg_inode_table) +
2666                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2667
2668         iloc->block_group = block_group;
2669         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2670         return block;
2671 }
2672
2673 /*
2674  * ext3_get_inode_loc returns with an extra refcount against the inode's
2675  * underlying buffer_head on success. If 'in_mem' is true, we have all
2676  * data in memory that is needed to recreate the on-disk version of this
2677  * inode.
2678  */
2679 static int __ext3_get_inode_loc(struct inode *inode,
2680                                 struct ext3_iloc *iloc, int in_mem)
2681 {
2682         ext3_fsblk_t block;
2683         struct buffer_head *bh;
2684
2685         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2686         if (!block)
2687                 return -EIO;
2688
2689         bh = sb_getblk(inode->i_sb, block);
2690         if (!bh) {
2691                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2692                                 "unable to read inode block - "
2693                                 "inode=%lu, block="E3FSBLK,
2694                                  inode->i_ino, block);
2695                 return -EIO;
2696         }
2697         if (!buffer_uptodate(bh)) {
2698                 lock_buffer(bh);
2699
2700                 /*
2701                  * If the buffer has the write error flag, we have failed
2702                  * to write out another inode in the same block.  In this
2703                  * case, we don't have to read the block because we may
2704                  * read the old inode data successfully.
2705                  */
2706                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2707                         set_buffer_uptodate(bh);
2708
2709                 if (buffer_uptodate(bh)) {
2710                         /* someone brought it uptodate while we waited */
2711                         unlock_buffer(bh);
2712                         goto has_buffer;
2713                 }
2714
2715                 /*
2716                  * If we have all information of the inode in memory and this
2717                  * is the only valid inode in the block, we need not read the
2718                  * block.
2719                  */
2720                 if (in_mem) {
2721                         struct buffer_head *bitmap_bh;
2722                         struct ext3_group_desc *desc;
2723                         int inodes_per_buffer;
2724                         int inode_offset, i;
2725                         int block_group;
2726                         int start;
2727
2728                         block_group = (inode->i_ino - 1) /
2729                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2730                         inodes_per_buffer = bh->b_size /
2731                                 EXT3_INODE_SIZE(inode->i_sb);
2732                         inode_offset = ((inode->i_ino - 1) %
2733                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2734                         start = inode_offset & ~(inodes_per_buffer - 1);
2735
2736                         /* Is the inode bitmap in cache? */
2737                         desc = ext3_get_group_desc(inode->i_sb,
2738                                                 block_group, NULL);
2739                         if (!desc)
2740                                 goto make_io;
2741
2742                         bitmap_bh = sb_getblk(inode->i_sb,
2743                                         le32_to_cpu(desc->bg_inode_bitmap));
2744                         if (!bitmap_bh)
2745                                 goto make_io;
2746
2747                         /*
2748                          * If the inode bitmap isn't in cache then the
2749                          * optimisation may end up performing two reads instead
2750                          * of one, so skip it.
2751                          */
2752                         if (!buffer_uptodate(bitmap_bh)) {
2753                                 brelse(bitmap_bh);
2754                                 goto make_io;
2755                         }
2756                         for (i = start; i < start + inodes_per_buffer; i++) {
2757                                 if (i == inode_offset)
2758                                         continue;
2759                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2760                                         break;
2761                         }
2762                         brelse(bitmap_bh);
2763                         if (i == start + inodes_per_buffer) {
2764                                 /* all other inodes are free, so skip I/O */
2765                                 memset(bh->b_data, 0, bh->b_size);
2766                                 set_buffer_uptodate(bh);
2767                                 unlock_buffer(bh);
2768                                 goto has_buffer;
2769                         }
2770                 }
2771
2772 make_io:
2773                 /*
2774                  * There are other valid inodes in the buffer, this inode
2775                  * has in-inode xattrs, or we don't have this inode in memory.
2776                  * Read the block from disk.
2777                  */
2778                 trace_ext3_load_inode(inode);
2779                 get_bh(bh);
2780                 bh->b_end_io = end_buffer_read_sync;
2781                 submit_bh(READ_META, bh);
2782                 wait_on_buffer(bh);
2783                 if (!buffer_uptodate(bh)) {
2784                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2785                                         "unable to read inode block - "
2786                                         "inode=%lu, block="E3FSBLK,
2787                                         inode->i_ino, block);
2788                         brelse(bh);
2789                         return -EIO;
2790                 }
2791         }
2792 has_buffer:
2793         iloc->bh = bh;
2794         return 0;
2795 }
2796
2797 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2798 {
2799         /* We have all inode data except xattrs in memory here. */
2800         return __ext3_get_inode_loc(inode, iloc,
2801                 !ext3_test_inode_state(inode, EXT3_STATE_XATTR));
2802 }
2803
2804 void ext3_set_inode_flags(struct inode *inode)
2805 {
2806         unsigned int flags = EXT3_I(inode)->i_flags;
2807
2808         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2809         if (flags & EXT3_SYNC_FL)
2810                 inode->i_flags |= S_SYNC;
2811         if (flags & EXT3_APPEND_FL)
2812                 inode->i_flags |= S_APPEND;
2813         if (flags & EXT3_IMMUTABLE_FL)
2814                 inode->i_flags |= S_IMMUTABLE;
2815         if (flags & EXT3_NOATIME_FL)
2816                 inode->i_flags |= S_NOATIME;
2817         if (flags & EXT3_DIRSYNC_FL)
2818                 inode->i_flags |= S_DIRSYNC;
2819 }
2820
2821 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2822 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2823 {
2824         unsigned int flags = ei->vfs_inode.i_flags;
2825
2826         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2827                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2828         if (flags & S_SYNC)
2829                 ei->i_flags |= EXT3_SYNC_FL;
2830         if (flags & S_APPEND)
2831                 ei->i_flags |= EXT3_APPEND_FL;
2832         if (flags & S_IMMUTABLE)
2833                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2834         if (flags & S_NOATIME)
2835                 ei->i_flags |= EXT3_NOATIME_FL;
2836         if (flags & S_DIRSYNC)
2837                 ei->i_flags |= EXT3_DIRSYNC_FL;
2838 }
2839
2840 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2841 {
2842         struct ext3_iloc iloc;
2843         struct ext3_inode *raw_inode;
2844         struct ext3_inode_info *ei;
2845         struct buffer_head *bh;
2846         struct inode *inode;
2847         journal_t *journal = EXT3_SB(sb)->s_journal;
2848         transaction_t *transaction;
2849         long ret;
2850         int block;
2851
2852         inode = iget_locked(sb, ino);
2853         if (!inode)
2854                 return ERR_PTR(-ENOMEM);
2855         if (!(inode->i_state & I_NEW))
2856                 return inode;
2857
2858         ei = EXT3_I(inode);
2859         ei->i_block_alloc_info = NULL;
2860
2861         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2862         if (ret < 0)
2863                 goto bad_inode;
2864         bh = iloc.bh;
2865         raw_inode = ext3_raw_inode(&iloc);
2866         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2867         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2868         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2869         if(!(test_opt (inode->i_sb, NO_UID32))) {
2870                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2871                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2872         }
2873         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2874         inode->i_size = le32_to_cpu(raw_inode->i_size);
2875         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2876         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2877         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2878         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2879
2880         ei->i_state_flags = 0;
2881         ei->i_dir_start_lookup = 0;
2882         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2883         /* We now have enough fields to check if the inode was active or not.
2884          * This is needed because nfsd might try to access dead inodes
2885          * the test is that same one that e2fsck uses
2886          * NeilBrown 1999oct15
2887          */
2888         if (inode->i_nlink == 0) {
2889                 if (inode->i_mode == 0 ||
2890                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2891                         /* this inode is deleted */
2892                         brelse (bh);
2893                         ret = -ESTALE;
2894                         goto bad_inode;
2895                 }
2896                 /* The only unlinked inodes we let through here have
2897                  * valid i_mode and are being read by the orphan
2898                  * recovery code: that's fine, we're about to complete
2899                  * the process of deleting those. */
2900         }
2901         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2902         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2903 #ifdef EXT3_FRAGMENTS
2904         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2905         ei->i_frag_no = raw_inode->i_frag;
2906         ei->i_frag_size = raw_inode->i_fsize;
2907 #endif
2908         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2909         if (!S_ISREG(inode->i_mode)) {
2910                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2911         } else {
2912                 inode->i_size |=
2913                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2914         }
2915         ei->i_disksize = inode->i_size;
2916         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2917         ei->i_block_group = iloc.block_group;
2918         /*
2919          * NOTE! The in-memory inode i_data array is in little-endian order
2920          * even on big-endian machines: we do NOT byteswap the block numbers!
2921          */
2922         for (block = 0; block < EXT3_N_BLOCKS; block++)
2923                 ei->i_data[block] = raw_inode->i_block[block];
2924         INIT_LIST_HEAD(&ei->i_orphan);
2925
2926         /*
2927          * Set transaction id's of transactions that have to be committed
2928          * to finish f[data]sync. We set them to currently running transaction
2929          * as we cannot be sure that the inode or some of its metadata isn't
2930          * part of the transaction - the inode could have been reclaimed and
2931          * now it is reread from disk.
2932          */
2933         if (journal) {
2934                 tid_t tid;
2935
2936                 spin_lock(&journal->j_state_lock);
2937                 if (journal->j_running_transaction)
2938                         transaction = journal->j_running_transaction;
2939                 else
2940                         transaction = journal->j_committing_transaction;
2941                 if (transaction)
2942                         tid = transaction->t_tid;
2943                 else
2944                         tid = journal->j_commit_sequence;
2945                 spin_unlock(&journal->j_state_lock);
2946                 atomic_set(&ei->i_sync_tid, tid);
2947                 atomic_set(&ei->i_datasync_tid, tid);
2948         }
2949
2950         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2951             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2952                 /*
2953                  * When mke2fs creates big inodes it does not zero out
2954                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2955                  * so ignore those first few inodes.
2956                  */
2957                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2958                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2959                     EXT3_INODE_SIZE(inode->i_sb)) {
2960                         brelse (bh);
2961                         ret = -EIO;
2962                         goto bad_inode;
2963                 }
2964                 if (ei->i_extra_isize == 0) {
2965                         /* The extra space is currently unused. Use it. */
2966                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2967                                             EXT3_GOOD_OLD_INODE_SIZE;
2968                 } else {
2969                         __le32 *magic = (void *)raw_inode +
2970                                         EXT3_GOOD_OLD_INODE_SIZE +
2971                                         ei->i_extra_isize;
2972                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2973                                  ext3_set_inode_state(inode, EXT3_STATE_XATTR);
2974                 }
2975         } else
2976                 ei->i_extra_isize = 0;
2977
2978         if (S_ISREG(inode->i_mode)) {
2979                 inode->i_op = &ext3_file_inode_operations;
2980                 inode->i_fop = &ext3_file_operations;
2981                 ext3_set_aops(inode);
2982         } else if (S_ISDIR(inode->i_mode)) {
2983                 inode->i_op = &ext3_dir_inode_operations;
2984                 inode->i_fop = &ext3_dir_operations;
2985         } else if (S_ISLNK(inode->i_mode)) {
2986                 if (ext3_inode_is_fast_symlink(inode)) {
2987                         inode->i_op = &ext3_fast_symlink_inode_operations;
2988                         nd_terminate_link(ei->i_data, inode->i_size,
2989                                 sizeof(ei->i_data) - 1);
2990                 } else {
2991                         inode->i_op = &ext3_symlink_inode_operations;
2992                         ext3_set_aops(inode);
2993                 }
2994         } else {
2995                 inode->i_op = &ext3_special_inode_operations;
2996                 if (raw_inode->i_block[0])
2997                         init_special_inode(inode, inode->i_mode,
2998                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2999                 else
3000                         init_special_inode(inode, inode->i_mode,
3001                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3002         }
3003         brelse (iloc.bh);
3004         ext3_set_inode_flags(inode);
3005         unlock_new_inode(inode);
3006         return inode;
3007
3008 bad_inode:
3009         iget_failed(inode);
3010         return ERR_PTR(ret);
3011 }
3012
3013 /*
3014  * Post the struct inode info into an on-disk inode location in the
3015  * buffer-cache.  This gobbles the caller's reference to the
3016  * buffer_head in the inode location struct.
3017  *
3018  * The caller must have write access to iloc->bh.
3019  */
3020 static int ext3_do_update_inode(handle_t *handle,
3021                                 struct inode *inode,
3022                                 struct ext3_iloc *iloc)
3023 {
3024         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
3025         struct ext3_inode_info *ei = EXT3_I(inode);
3026         struct buffer_head *bh = iloc->bh;
3027         int err = 0, rc, block;
3028
3029 again:
3030         /* we can't allow multiple procs in here at once, its a bit racey */
3031         lock_buffer(bh);
3032
3033         /* For fields not not tracking in the in-memory inode,
3034          * initialise them to zero for new inodes. */
3035         if (ext3_test_inode_state(inode, EXT3_STATE_NEW))
3036                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
3037
3038         ext3_get_inode_flags(ei);
3039         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3040         if(!(test_opt(inode->i_sb, NO_UID32))) {
3041                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3042                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3043 /*
3044  * Fix up interoperability with old kernels. Otherwise, old inodes get
3045  * re-used with the upper 16 bits of the uid/gid intact
3046  */
3047                 if(!ei->i_dtime) {
3048                         raw_inode->i_uid_high =
3049                                 cpu_to_le16(high_16_bits(inode->i_uid));
3050                         raw_inode->i_gid_high =
3051                                 cpu_to_le16(high_16_bits(inode->i_gid));
3052                 } else {
3053                         raw_inode->i_uid_high = 0;
3054                         raw_inode->i_gid_high = 0;
3055                 }
3056         } else {
3057                 raw_inode->i_uid_low =
3058                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3059                 raw_inode->i_gid_low =
3060                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3061                 raw_inode->i_uid_high = 0;
3062                 raw_inode->i_gid_high = 0;
3063         }
3064         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3065         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
3066         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
3067         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
3068         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
3069         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
3070         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3071         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
3072 #ifdef EXT3_FRAGMENTS
3073         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
3074         raw_inode->i_frag = ei->i_frag_no;
3075         raw_inode->i_fsize = ei->i_frag_size;
3076 #endif
3077         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
3078         if (!S_ISREG(inode->i_mode)) {
3079                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
3080         } else {
3081                 raw_inode->i_size_high =
3082                         cpu_to_le32(ei->i_disksize >> 32);
3083                 if (ei->i_disksize > 0x7fffffffULL) {
3084                         struct super_block *sb = inode->i_sb;
3085                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
3086                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
3087                             EXT3_SB(sb)->s_es->s_rev_level ==
3088                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
3089                                /* If this is the first large file
3090                                 * created, add a flag to the superblock.
3091                                 */
3092                                 unlock_buffer(bh);
3093                                 err = ext3_journal_get_write_access(handle,
3094                                                 EXT3_SB(sb)->s_sbh);
3095                                 if (err)
3096                                         goto out_brelse;
3097
3098                                 ext3_update_dynamic_rev(sb);
3099                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
3100                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
3101                                 handle->h_sync = 1;
3102                                 err = ext3_journal_dirty_metadata(handle,
3103                                                 EXT3_SB(sb)->s_sbh);
3104                                 /* get our lock and start over */
3105                                 goto again;
3106                         }
3107                 }
3108         }
3109         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3110         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3111                 if (old_valid_dev(inode->i_rdev)) {
3112                         raw_inode->i_block[0] =
3113                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3114                         raw_inode->i_block[1] = 0;
3115                 } else {
3116                         raw_inode->i_block[0] = 0;
3117                         raw_inode->i_block[1] =
3118                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3119                         raw_inode->i_block[2] = 0;
3120                 }
3121         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
3122                 raw_inode->i_block[block] = ei->i_data[block];
3123
3124         if (ei->i_extra_isize)
3125                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3126
3127         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
3128         unlock_buffer(bh);
3129         rc = ext3_journal_dirty_metadata(handle, bh);
3130         if (!err)
3131                 err = rc;
3132         ext3_clear_inode_state(inode, EXT3_STATE_NEW);
3133
3134         atomic_set(&ei->i_sync_tid, handle->h_transaction->t_tid);
3135 out_brelse:
3136         brelse (bh);
3137         ext3_std_error(inode->i_sb, err);
3138         return err;
3139 }
3140
3141 /*
3142  * ext3_write_inode()
3143  *
3144  * We are called from a few places:
3145  *
3146  * - Within generic_file_write() for O_SYNC files.
3147  *   Here, there will be no transaction running. We wait for any running
3148  *   trasnaction to commit.
3149  *
3150  * - Within sys_sync(), kupdate and such.
3151  *   We wait on commit, if tol to.
3152  *
3153  * - Within prune_icache() (PF_MEMALLOC == true)
3154  *   Here we simply return.  We can't afford to block kswapd on the
3155  *   journal commit.
3156  *
3157  * In all cases it is actually safe for us to return without doing anything,
3158  * because the inode has been copied into a raw inode buffer in
3159  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3160  * knfsd.
3161  *
3162  * Note that we are absolutely dependent upon all inode dirtiers doing the
3163  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3164  * which we are interested.
3165  *
3166  * It would be a bug for them to not do this.  The code:
3167  *
3168  *      mark_inode_dirty(inode)
3169  *      stuff();
3170  *      inode->i_size = expr;
3171  *
3172  * is in error because a kswapd-driven write_inode() could occur while
3173  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3174  * will no longer be on the superblock's dirty inode list.
3175  */
3176 int ext3_write_inode(struct inode *inode, struct writeback_control *wbc)
3177 {
3178         if (current->flags & PF_MEMALLOC)
3179                 return 0;
3180
3181         if (ext3_journal_current_handle()) {
3182                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3183                 dump_stack();
3184                 return -EIO;
3185         }
3186
3187         if (wbc->sync_mode != WB_SYNC_ALL)
3188                 return 0;
3189
3190         return ext3_force_commit(inode->i_sb);
3191 }
3192
3193 /*
3194  * ext3_setattr()
3195  *
3196  * Called from notify_change.
3197  *
3198  * We want to trap VFS attempts to truncate the file as soon as
3199  * possible.  In particular, we want to make sure that when the VFS
3200  * shrinks i_size, we put the inode on the orphan list and modify
3201  * i_disksize immediately, so that during the subsequent flushing of
3202  * dirty pages and freeing of disk blocks, we can guarantee that any
3203  * commit will leave the blocks being flushed in an unused state on
3204  * disk.  (On recovery, the inode will get truncated and the blocks will
3205  * be freed, so we have a strong guarantee that no future commit will
3206  * leave these blocks visible to the user.)
3207  *
3208  * Called with inode->sem down.
3209  */
3210 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3211 {
3212         struct inode *inode = dentry->d_inode;
3213         int error, rc = 0;
3214         const unsigned int ia_valid = attr->ia_valid;
3215
3216         error = inode_change_ok(inode, attr);
3217         if (error)
3218                 return error;
3219
3220         if (is_quota_modification(inode, attr))
3221                 dquot_initialize(inode);
3222         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3223                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3224                 handle_t *handle;
3225
3226                 /* (user+group)*(old+new) structure, inode write (sb,
3227                  * inode block, ? - but truncate inode update has it) */
3228                 handle = ext3_journal_start(inode, EXT3_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3229                                         EXT3_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)+3);
3230                 if (IS_ERR(handle)) {
3231                         error = PTR_ERR(handle);
3232                         goto err_out;
3233                 }
3234                 error = dquot_transfer(inode, attr);
3235                 if (error) {
3236                         ext3_journal_stop(handle);
3237                         return error;
3238                 }
3239                 /* Update corresponding info in inode so that everything is in
3240                  * one transaction */
3241                 if (attr->ia_valid & ATTR_UID)
3242                         inode->i_uid = attr->ia_uid;
3243                 if (attr->ia_valid & ATTR_GID)
3244                         inode->i_gid = attr->ia_gid;
3245                 error = ext3_mark_inode_dirty(handle, inode);
3246                 ext3_journal_stop(handle);
3247         }
3248
3249         if (S_ISREG(inode->i_mode) &&
3250             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3251                 handle_t *handle;
3252
3253                 handle = ext3_journal_start(inode, 3);
3254                 if (IS_ERR(handle)) {
3255                         error = PTR_ERR(handle);
3256                         goto err_out;
3257                 }
3258
3259                 error = ext3_orphan_add(handle, inode);
3260                 if (error) {
3261                         ext3_journal_stop(handle);
3262                         goto err_out;
3263                 }
3264                 EXT3_I(inode)->i_disksize = attr->ia_size;
3265                 error = ext3_mark_inode_dirty(handle, inode);
3266                 ext3_journal_stop(handle);
3267                 if (error) {
3268                         /* Some hard fs error must have happened. Bail out. */
3269                         ext3_orphan_del(NULL, inode);
3270                         goto err_out;
3271                 }
3272                 rc = ext3_block_truncate_page(inode, attr->ia_size);
3273                 if (rc) {
3274                         /* Cleanup orphan list and exit */
3275                         handle = ext3_journal_start(inode, 3);
3276                         if (IS_ERR(handle)) {
3277                                 ext3_orphan_del(NULL, inode);
3278                                 goto err_out;
3279                         }
3280                         ext3_orphan_del(handle, inode);
3281                         ext3_journal_stop(handle);
3282                         goto err_out;
3283                 }
3284         }
3285
3286         if ((attr->ia_valid & ATTR_SIZE) &&
3287             attr->ia_size != i_size_read(inode)) {
3288                 truncate_setsize(inode, attr->ia_size);
3289                 ext3_truncate(inode);
3290         }
3291
3292         setattr_copy(inode, attr);
3293         mark_inode_dirty(inode);
3294
3295         if (ia_valid & ATTR_MODE)
3296                 rc = ext3_acl_chmod(inode);
3297
3298 err_out:
3299         ext3_std_error(inode->i_sb, error);
3300         if (!error)
3301                 error = rc;
3302         return error;
3303 }
3304
3305
3306 /*
3307  * How many blocks doth make a writepage()?
3308  *
3309  * With N blocks per page, it may be:
3310  * N data blocks
3311  * 2 indirect block
3312  * 2 dindirect
3313  * 1 tindirect
3314  * N+5 bitmap blocks (from the above)
3315  * N+5 group descriptor summary blocks
3316  * 1 inode block
3317  * 1 superblock.
3318  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3319  *
3320  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3321  *
3322  * With ordered or writeback data it's the same, less the N data blocks.
3323  *
3324  * If the inode's direct blocks can hold an integral number of pages then a
3325  * page cannot straddle two indirect blocks, and we can only touch one indirect
3326  * and dindirect block, and the "5" above becomes "3".
3327  *
3328  * This still overestimates under most circumstances.  If we were to pass the
3329  * start and end offsets in here as well we could do block_to_path() on each
3330  * block and work out the exact number of indirects which are touched.  Pah.
3331  */
3332
3333 static int ext3_writepage_trans_blocks(struct inode *inode)
3334 {
3335         int bpp = ext3_journal_blocks_per_page(inode);
3336         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3337         int ret;
3338
3339         if (ext3_should_journal_data(inode))
3340                 ret = 3 * (bpp + indirects) + 2;
3341         else
3342                 ret = 2 * (bpp + indirects) + indirects + 2;
3343
3344 #ifdef CONFIG_QUOTA
3345         /* We know that structure was already allocated during dquot_initialize so
3346          * we will be updating only the data blocks + inodes */
3347         ret += EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
3348 #endif
3349
3350         return ret;
3351 }
3352
3353 /*
3354  * The caller must have previously called ext3_reserve_inode_write().
3355  * Give this, we know that the caller already has write access to iloc->bh.
3356  */
3357 int ext3_mark_iloc_dirty(handle_t *handle,
3358                 struct inode *inode, struct ext3_iloc *iloc)
3359 {
3360         int err = 0;
3361
3362         /* the do_update_inode consumes one bh->b_count */
3363         get_bh(iloc->bh);
3364
3365         /* ext3_do_update_inode() does journal_dirty_metadata */
3366         err = ext3_do_update_inode(handle, inode, iloc);
3367         put_bh(iloc->bh);
3368         return err;
3369 }
3370
3371 /*
3372  * On success, We end up with an outstanding reference count against
3373  * iloc->bh.  This _must_ be cleaned up later.
3374  */
3375
3376 int
3377 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3378                          struct ext3_iloc *iloc)
3379 {
3380         int err = 0;
3381         if (handle) {
3382                 err = ext3_get_inode_loc(inode, iloc);
3383                 if (!err) {
3384                         BUFFER_TRACE(iloc->bh, "get_write_access");
3385                         err = ext3_journal_get_write_access(handle, iloc->bh);
3386                         if (err) {
3387                                 brelse(iloc->bh);
3388                                 iloc->bh = NULL;
3389                         }
3390                 }
3391         }
3392         ext3_std_error(inode->i_sb, err);
3393         return err;
3394 }
3395
3396 /*
3397  * What we do here is to mark the in-core inode as clean with respect to inode
3398  * dirtiness (it may still be data-dirty).
3399  * This means that the in-core inode may be reaped by prune_icache
3400  * without having to perform any I/O.  This is a very good thing,
3401  * because *any* task may call prune_icache - even ones which
3402  * have a transaction open against a different journal.
3403  *
3404  * Is this cheating?  Not really.  Sure, we haven't written the
3405  * inode out, but prune_icache isn't a user-visible syncing function.
3406  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3407  * we start and wait on commits.
3408  *
3409  * Is this efficient/effective?  Well, we're being nice to the system
3410  * by cleaning up our inodes proactively so they can be reaped
3411  * without I/O.  But we are potentially leaving up to five seconds'
3412  * worth of inodes floating about which prune_icache wants us to
3413  * write out.  One way to fix that would be to get prune_icache()
3414  * to do a write_super() to free up some memory.  It has the desired
3415  * effect.
3416  */
3417 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3418 {
3419         struct ext3_iloc iloc;
3420         int err;
3421
3422         might_sleep();
3423         trace_ext3_mark_inode_dirty(inode, _RET_IP_);
3424         err = ext3_reserve_inode_write(handle, inode, &iloc);
3425         if (!err)
3426                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3427         return err;
3428 }
3429
3430 /*
3431  * ext3_dirty_inode() is called from __mark_inode_dirty()
3432  *
3433  * We're really interested in the case where a file is being extended.
3434  * i_size has been changed by generic_commit_write() and we thus need
3435  * to include the updated inode in the current transaction.
3436  *
3437  * Also, dquot_alloc_space() will always dirty the inode when blocks
3438  * are allocated to the file.
3439  *
3440  * If the inode is marked synchronous, we don't honour that here - doing
3441  * so would cause a commit on atime updates, which we don't bother doing.
3442  * We handle synchronous inodes at the highest possible level.
3443  */
3444 void ext3_dirty_inode(struct inode *inode, int flags)
3445 {
3446         handle_t *current_handle = ext3_journal_current_handle();
3447         handle_t *handle;
3448
3449         handle = ext3_journal_start(inode, 2);
3450         if (IS_ERR(handle))
3451                 goto out;
3452         if (current_handle &&
3453                 current_handle->h_transaction != handle->h_transaction) {
3454                 /* This task has a transaction open against a different fs */
3455                 printk(KERN_EMERG "%s: transactions do not match!\n",
3456                        __func__);
3457         } else {
3458                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3459                                 current_handle);
3460                 ext3_mark_inode_dirty(handle, inode);
3461         }
3462         ext3_journal_stop(handle);
3463 out:
3464         return;
3465 }
3466
3467 #if 0
3468 /*
3469  * Bind an inode's backing buffer_head into this transaction, to prevent
3470  * it from being flushed to disk early.  Unlike
3471  * ext3_reserve_inode_write, this leaves behind no bh reference and
3472  * returns no iloc structure, so the caller needs to repeat the iloc
3473  * lookup to mark the inode dirty later.
3474  */
3475 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3476 {
3477         struct ext3_iloc iloc;
3478
3479         int err = 0;
3480         if (handle) {
3481                 err = ext3_get_inode_loc(inode, &iloc);
3482                 if (!err) {
3483                         BUFFER_TRACE(iloc.bh, "get_write_access");
3484                         err = journal_get_write_access(handle, iloc.bh);
3485                         if (!err)
3486                                 err = ext3_journal_dirty_metadata(handle,
3487                                                                   iloc.bh);
3488                         brelse(iloc.bh);
3489                 }
3490         }
3491         ext3_std_error(inode->i_sb, err);
3492         return err;
3493 }
3494 #endif
3495
3496 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3497 {
3498         journal_t *journal;
3499         handle_t *handle;
3500         int err;
3501
3502         /*
3503          * We have to be very careful here: changing a data block's
3504          * journaling status dynamically is dangerous.  If we write a
3505          * data block to the journal, change the status and then delete
3506          * that block, we risk forgetting to revoke the old log record
3507          * from the journal and so a subsequent replay can corrupt data.
3508          * So, first we make sure that the journal is empty and that
3509          * nobody is changing anything.
3510          */
3511
3512         journal = EXT3_JOURNAL(inode);
3513         if (is_journal_aborted(journal))
3514                 return -EROFS;
3515
3516         journal_lock_updates(journal);
3517         journal_flush(journal);
3518
3519         /*
3520          * OK, there are no updates running now, and all cached data is
3521          * synced to disk.  We are now in a completely consistent state
3522          * which doesn't have anything in the journal, and we know that
3523          * no filesystem updates are running, so it is safe to modify
3524          * the inode's in-core data-journaling state flag now.
3525          */
3526
3527         if (val)
3528                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3529         else
3530                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3531         ext3_set_aops(inode);
3532
3533         journal_unlock_updates(journal);
3534
3535         /* Finally we can mark the inode as dirty. */
3536
3537         handle = ext3_journal_start(inode, 1);
3538         if (IS_ERR(handle))
3539                 return PTR_ERR(handle);
3540
3541         err = ext3_mark_inode_dirty(handle, inode);
3542         handle->h_sync = 1;
3543         ext3_journal_stop(handle);
3544         ext3_std_error(inode->i_sb, err);
3545
3546         return err;
3547 }