ext3: Adjust trim start with first_data_block.
[linux-2.6.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include <linux/fiemap.h>
40 #include <linux/namei.h>
41 #include "xattr.h"
42 #include "acl.h"
43
44 static int ext3_writepage_trans_blocks(struct inode *inode);
45
46 /*
47  * Test whether an inode is a fast symlink.
48  */
49 static int ext3_inode_is_fast_symlink(struct inode *inode)
50 {
51         int ea_blocks = EXT3_I(inode)->i_file_acl ?
52                 (inode->i_sb->s_blocksize >> 9) : 0;
53
54         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
55 }
56
57 /*
58  * The ext3 forget function must perform a revoke if we are freeing data
59  * which has been journaled.  Metadata (eg. indirect blocks) must be
60  * revoked in all cases.
61  *
62  * "bh" may be NULL: a metadata block may have been freed from memory
63  * but there may still be a record of it in the journal, and that record
64  * still needs to be revoked.
65  */
66 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
67                         struct buffer_head *bh, ext3_fsblk_t blocknr)
68 {
69         int err;
70
71         might_sleep();
72
73         BUFFER_TRACE(bh, "enter");
74
75         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
76                   "data mode %lx\n",
77                   bh, is_metadata, inode->i_mode,
78                   test_opt(inode->i_sb, DATA_FLAGS));
79
80         /* Never use the revoke function if we are doing full data
81          * journaling: there is no need to, and a V1 superblock won't
82          * support it.  Otherwise, only skip the revoke on un-journaled
83          * data blocks. */
84
85         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
86             (!is_metadata && !ext3_should_journal_data(inode))) {
87                 if (bh) {
88                         BUFFER_TRACE(bh, "call journal_forget");
89                         return ext3_journal_forget(handle, bh);
90                 }
91                 return 0;
92         }
93
94         /*
95          * data!=journal && (is_metadata || should_journal_data(inode))
96          */
97         BUFFER_TRACE(bh, "call ext3_journal_revoke");
98         err = ext3_journal_revoke(handle, blocknr, bh);
99         if (err)
100                 ext3_abort(inode->i_sb, __func__,
101                            "error %d when attempting revoke", err);
102         BUFFER_TRACE(bh, "exit");
103         return err;
104 }
105
106 /*
107  * Work out how many blocks we need to proceed with the next chunk of a
108  * truncate transaction.
109  */
110 static unsigned long blocks_for_truncate(struct inode *inode)
111 {
112         unsigned long needed;
113
114         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
115
116         /* Give ourselves just enough room to cope with inodes in which
117          * i_blocks is corrupt: we've seen disk corruptions in the past
118          * which resulted in random data in an inode which looked enough
119          * like a regular file for ext3 to try to delete it.  Things
120          * will go a bit crazy if that happens, but at least we should
121          * try not to panic the whole kernel. */
122         if (needed < 2)
123                 needed = 2;
124
125         /* But we need to bound the transaction so we don't overflow the
126          * journal. */
127         if (needed > EXT3_MAX_TRANS_DATA)
128                 needed = EXT3_MAX_TRANS_DATA;
129
130         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
131 }
132
133 /*
134  * Truncate transactions can be complex and absolutely huge.  So we need to
135  * be able to restart the transaction at a conventient checkpoint to make
136  * sure we don't overflow the journal.
137  *
138  * start_transaction gets us a new handle for a truncate transaction,
139  * and extend_transaction tries to extend the existing one a bit.  If
140  * extend fails, we need to propagate the failure up and restart the
141  * transaction in the top-level truncate loop. --sct
142  */
143 static handle_t *start_transaction(struct inode *inode)
144 {
145         handle_t *result;
146
147         result = ext3_journal_start(inode, blocks_for_truncate(inode));
148         if (!IS_ERR(result))
149                 return result;
150
151         ext3_std_error(inode->i_sb, PTR_ERR(result));
152         return result;
153 }
154
155 /*
156  * Try to extend this transaction for the purposes of truncation.
157  *
158  * Returns 0 if we managed to create more room.  If we can't create more
159  * room, and the transaction must be restarted we return 1.
160  */
161 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
162 {
163         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
164                 return 0;
165         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
166                 return 0;
167         return 1;
168 }
169
170 /*
171  * Restart the transaction associated with *handle.  This does a commit,
172  * so before we call here everything must be consistently dirtied against
173  * this transaction.
174  */
175 static int truncate_restart_transaction(handle_t *handle, struct inode *inode)
176 {
177         int ret;
178
179         jbd_debug(2, "restarting handle %p\n", handle);
180         /*
181          * Drop truncate_mutex to avoid deadlock with ext3_get_blocks_handle
182          * At this moment, get_block can be called only for blocks inside
183          * i_size since page cache has been already dropped and writes are
184          * blocked by i_mutex. So we can safely drop the truncate_mutex.
185          */
186         mutex_unlock(&EXT3_I(inode)->truncate_mutex);
187         ret = ext3_journal_restart(handle, blocks_for_truncate(inode));
188         mutex_lock(&EXT3_I(inode)->truncate_mutex);
189         return ret;
190 }
191
192 /*
193  * Called at inode eviction from icache
194  */
195 void ext3_evict_inode (struct inode *inode)
196 {
197         struct ext3_block_alloc_info *rsv;
198         handle_t *handle;
199         int want_delete = 0;
200
201         if (!inode->i_nlink && !is_bad_inode(inode)) {
202                 dquot_initialize(inode);
203                 want_delete = 1;
204         }
205
206         truncate_inode_pages(&inode->i_data, 0);
207
208         ext3_discard_reservation(inode);
209         rsv = EXT3_I(inode)->i_block_alloc_info;
210         EXT3_I(inode)->i_block_alloc_info = NULL;
211         if (unlikely(rsv))
212                 kfree(rsv);
213
214         if (!want_delete)
215                 goto no_delete;
216
217         handle = start_transaction(inode);
218         if (IS_ERR(handle)) {
219                 /*
220                  * If we're going to skip the normal cleanup, we still need to
221                  * make sure that the in-core orphan linked list is properly
222                  * cleaned up.
223                  */
224                 ext3_orphan_del(NULL, inode);
225                 goto no_delete;
226         }
227
228         if (IS_SYNC(inode))
229                 handle->h_sync = 1;
230         inode->i_size = 0;
231         if (inode->i_blocks)
232                 ext3_truncate(inode);
233         /*
234          * Kill off the orphan record which ext3_truncate created.
235          * AKPM: I think this can be inside the above `if'.
236          * Note that ext3_orphan_del() has to be able to cope with the
237          * deletion of a non-existent orphan - this is because we don't
238          * know if ext3_truncate() actually created an orphan record.
239          * (Well, we could do this if we need to, but heck - it works)
240          */
241         ext3_orphan_del(handle, inode);
242         EXT3_I(inode)->i_dtime  = get_seconds();
243
244         /*
245          * One subtle ordering requirement: if anything has gone wrong
246          * (transaction abort, IO errors, whatever), then we can still
247          * do these next steps (the fs will already have been marked as
248          * having errors), but we can't free the inode if the mark_dirty
249          * fails.
250          */
251         if (ext3_mark_inode_dirty(handle, inode)) {
252                 /* If that failed, just dquot_drop() and be done with that */
253                 dquot_drop(inode);
254                 end_writeback(inode);
255         } else {
256                 ext3_xattr_delete_inode(handle, inode);
257                 dquot_free_inode(inode);
258                 dquot_drop(inode);
259                 end_writeback(inode);
260                 ext3_free_inode(handle, inode);
261         }
262         ext3_journal_stop(handle);
263         return;
264 no_delete:
265         end_writeback(inode);
266         dquot_drop(inode);
267 }
268
269 typedef struct {
270         __le32  *p;
271         __le32  key;
272         struct buffer_head *bh;
273 } Indirect;
274
275 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
276 {
277         p->key = *(p->p = v);
278         p->bh = bh;
279 }
280
281 static int verify_chain(Indirect *from, Indirect *to)
282 {
283         while (from <= to && from->key == *from->p)
284                 from++;
285         return (from > to);
286 }
287
288 /**
289  *      ext3_block_to_path - parse the block number into array of offsets
290  *      @inode: inode in question (we are only interested in its superblock)
291  *      @i_block: block number to be parsed
292  *      @offsets: array to store the offsets in
293  *      @boundary: set this non-zero if the referred-to block is likely to be
294  *             followed (on disk) by an indirect block.
295  *
296  *      To store the locations of file's data ext3 uses a data structure common
297  *      for UNIX filesystems - tree of pointers anchored in the inode, with
298  *      data blocks at leaves and indirect blocks in intermediate nodes.
299  *      This function translates the block number into path in that tree -
300  *      return value is the path length and @offsets[n] is the offset of
301  *      pointer to (n+1)th node in the nth one. If @block is out of range
302  *      (negative or too large) warning is printed and zero returned.
303  *
304  *      Note: function doesn't find node addresses, so no IO is needed. All
305  *      we need to know is the capacity of indirect blocks (taken from the
306  *      inode->i_sb).
307  */
308
309 /*
310  * Portability note: the last comparison (check that we fit into triple
311  * indirect block) is spelled differently, because otherwise on an
312  * architecture with 32-bit longs and 8Kb pages we might get into trouble
313  * if our filesystem had 8Kb blocks. We might use long long, but that would
314  * kill us on x86. Oh, well, at least the sign propagation does not matter -
315  * i_block would have to be negative in the very beginning, so we would not
316  * get there at all.
317  */
318
319 static int ext3_block_to_path(struct inode *inode,
320                         long i_block, int offsets[4], int *boundary)
321 {
322         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
323         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
324         const long direct_blocks = EXT3_NDIR_BLOCKS,
325                 indirect_blocks = ptrs,
326                 double_blocks = (1 << (ptrs_bits * 2));
327         int n = 0;
328         int final = 0;
329
330         if (i_block < 0) {
331                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
332         } else if (i_block < direct_blocks) {
333                 offsets[n++] = i_block;
334                 final = direct_blocks;
335         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
336                 offsets[n++] = EXT3_IND_BLOCK;
337                 offsets[n++] = i_block;
338                 final = ptrs;
339         } else if ((i_block -= indirect_blocks) < double_blocks) {
340                 offsets[n++] = EXT3_DIND_BLOCK;
341                 offsets[n++] = i_block >> ptrs_bits;
342                 offsets[n++] = i_block & (ptrs - 1);
343                 final = ptrs;
344         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
345                 offsets[n++] = EXT3_TIND_BLOCK;
346                 offsets[n++] = i_block >> (ptrs_bits * 2);
347                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
348                 offsets[n++] = i_block & (ptrs - 1);
349                 final = ptrs;
350         } else {
351                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
352         }
353         if (boundary)
354                 *boundary = final - 1 - (i_block & (ptrs - 1));
355         return n;
356 }
357
358 /**
359  *      ext3_get_branch - read the chain of indirect blocks leading to data
360  *      @inode: inode in question
361  *      @depth: depth of the chain (1 - direct pointer, etc.)
362  *      @offsets: offsets of pointers in inode/indirect blocks
363  *      @chain: place to store the result
364  *      @err: here we store the error value
365  *
366  *      Function fills the array of triples <key, p, bh> and returns %NULL
367  *      if everything went OK or the pointer to the last filled triple
368  *      (incomplete one) otherwise. Upon the return chain[i].key contains
369  *      the number of (i+1)-th block in the chain (as it is stored in memory,
370  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
371  *      number (it points into struct inode for i==0 and into the bh->b_data
372  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
373  *      block for i>0 and NULL for i==0. In other words, it holds the block
374  *      numbers of the chain, addresses they were taken from (and where we can
375  *      verify that chain did not change) and buffer_heads hosting these
376  *      numbers.
377  *
378  *      Function stops when it stumbles upon zero pointer (absent block)
379  *              (pointer to last triple returned, *@err == 0)
380  *      or when it gets an IO error reading an indirect block
381  *              (ditto, *@err == -EIO)
382  *      or when it notices that chain had been changed while it was reading
383  *              (ditto, *@err == -EAGAIN)
384  *      or when it reads all @depth-1 indirect blocks successfully and finds
385  *      the whole chain, all way to the data (returns %NULL, *err == 0).
386  */
387 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
388                                  Indirect chain[4], int *err)
389 {
390         struct super_block *sb = inode->i_sb;
391         Indirect *p = chain;
392         struct buffer_head *bh;
393
394         *err = 0;
395         /* i_data is not going away, no lock needed */
396         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
397         if (!p->key)
398                 goto no_block;
399         while (--depth) {
400                 bh = sb_bread(sb, le32_to_cpu(p->key));
401                 if (!bh)
402                         goto failure;
403                 /* Reader: pointers */
404                 if (!verify_chain(chain, p))
405                         goto changed;
406                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
407                 /* Reader: end */
408                 if (!p->key)
409                         goto no_block;
410         }
411         return NULL;
412
413 changed:
414         brelse(bh);
415         *err = -EAGAIN;
416         goto no_block;
417 failure:
418         *err = -EIO;
419 no_block:
420         return p;
421 }
422
423 /**
424  *      ext3_find_near - find a place for allocation with sufficient locality
425  *      @inode: owner
426  *      @ind: descriptor of indirect block.
427  *
428  *      This function returns the preferred place for block allocation.
429  *      It is used when heuristic for sequential allocation fails.
430  *      Rules are:
431  *        + if there is a block to the left of our position - allocate near it.
432  *        + if pointer will live in indirect block - allocate near that block.
433  *        + if pointer will live in inode - allocate in the same
434  *          cylinder group.
435  *
436  * In the latter case we colour the starting block by the callers PID to
437  * prevent it from clashing with concurrent allocations for a different inode
438  * in the same block group.   The PID is used here so that functionally related
439  * files will be close-by on-disk.
440  *
441  *      Caller must make sure that @ind is valid and will stay that way.
442  */
443 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
444 {
445         struct ext3_inode_info *ei = EXT3_I(inode);
446         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
447         __le32 *p;
448         ext3_fsblk_t bg_start;
449         ext3_grpblk_t colour;
450
451         /* Try to find previous block */
452         for (p = ind->p - 1; p >= start; p--) {
453                 if (*p)
454                         return le32_to_cpu(*p);
455         }
456
457         /* No such thing, so let's try location of indirect block */
458         if (ind->bh)
459                 return ind->bh->b_blocknr;
460
461         /*
462          * It is going to be referred to from the inode itself? OK, just put it
463          * into the same cylinder group then.
464          */
465         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
466         colour = (current->pid % 16) *
467                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
468         return bg_start + colour;
469 }
470
471 /**
472  *      ext3_find_goal - find a preferred place for allocation.
473  *      @inode: owner
474  *      @block:  block we want
475  *      @partial: pointer to the last triple within a chain
476  *
477  *      Normally this function find the preferred place for block allocation,
478  *      returns it.
479  */
480
481 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
482                                    Indirect *partial)
483 {
484         struct ext3_block_alloc_info *block_i;
485
486         block_i =  EXT3_I(inode)->i_block_alloc_info;
487
488         /*
489          * try the heuristic for sequential allocation,
490          * failing that at least try to get decent locality.
491          */
492         if (block_i && (block == block_i->last_alloc_logical_block + 1)
493                 && (block_i->last_alloc_physical_block != 0)) {
494                 return block_i->last_alloc_physical_block + 1;
495         }
496
497         return ext3_find_near(inode, partial);
498 }
499
500 /**
501  *      ext3_blks_to_allocate - Look up the block map and count the number
502  *      of direct blocks need to be allocated for the given branch.
503  *
504  *      @branch: chain of indirect blocks
505  *      @k: number of blocks need for indirect blocks
506  *      @blks: number of data blocks to be mapped.
507  *      @blocks_to_boundary:  the offset in the indirect block
508  *
509  *      return the total number of blocks to be allocate, including the
510  *      direct and indirect blocks.
511  */
512 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
513                 int blocks_to_boundary)
514 {
515         unsigned long count = 0;
516
517         /*
518          * Simple case, [t,d]Indirect block(s) has not allocated yet
519          * then it's clear blocks on that path have not allocated
520          */
521         if (k > 0) {
522                 /* right now we don't handle cross boundary allocation */
523                 if (blks < blocks_to_boundary + 1)
524                         count += blks;
525                 else
526                         count += blocks_to_boundary + 1;
527                 return count;
528         }
529
530         count++;
531         while (count < blks && count <= blocks_to_boundary &&
532                 le32_to_cpu(*(branch[0].p + count)) == 0) {
533                 count++;
534         }
535         return count;
536 }
537
538 /**
539  *      ext3_alloc_blocks - multiple allocate blocks needed for a branch
540  *      @handle: handle for this transaction
541  *      @inode: owner
542  *      @goal: preferred place for allocation
543  *      @indirect_blks: the number of blocks need to allocate for indirect
544  *                      blocks
545  *      @blks:  number of blocks need to allocated for direct blocks
546  *      @new_blocks: on return it will store the new block numbers for
547  *      the indirect blocks(if needed) and the first direct block,
548  *      @err: here we store the error value
549  *
550  *      return the number of direct blocks allocated
551  */
552 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
553                         ext3_fsblk_t goal, int indirect_blks, int blks,
554                         ext3_fsblk_t new_blocks[4], int *err)
555 {
556         int target, i;
557         unsigned long count = 0;
558         int index = 0;
559         ext3_fsblk_t current_block = 0;
560         int ret = 0;
561
562         /*
563          * Here we try to allocate the requested multiple blocks at once,
564          * on a best-effort basis.
565          * To build a branch, we should allocate blocks for
566          * the indirect blocks(if not allocated yet), and at least
567          * the first direct block of this branch.  That's the
568          * minimum number of blocks need to allocate(required)
569          */
570         target = blks + indirect_blks;
571
572         while (1) {
573                 count = target;
574                 /* allocating blocks for indirect blocks and direct blocks */
575                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
576                 if (*err)
577                         goto failed_out;
578
579                 target -= count;
580                 /* allocate blocks for indirect blocks */
581                 while (index < indirect_blks && count) {
582                         new_blocks[index++] = current_block++;
583                         count--;
584                 }
585
586                 if (count > 0)
587                         break;
588         }
589
590         /* save the new block number for the first direct block */
591         new_blocks[index] = current_block;
592
593         /* total number of blocks allocated for direct blocks */
594         ret = count;
595         *err = 0;
596         return ret;
597 failed_out:
598         for (i = 0; i <index; i++)
599                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
600         return ret;
601 }
602
603 /**
604  *      ext3_alloc_branch - allocate and set up a chain of blocks.
605  *      @handle: handle for this transaction
606  *      @inode: owner
607  *      @indirect_blks: number of allocated indirect blocks
608  *      @blks: number of allocated direct blocks
609  *      @goal: preferred place for allocation
610  *      @offsets: offsets (in the blocks) to store the pointers to next.
611  *      @branch: place to store the chain in.
612  *
613  *      This function allocates blocks, zeroes out all but the last one,
614  *      links them into chain and (if we are synchronous) writes them to disk.
615  *      In other words, it prepares a branch that can be spliced onto the
616  *      inode. It stores the information about that chain in the branch[], in
617  *      the same format as ext3_get_branch() would do. We are calling it after
618  *      we had read the existing part of chain and partial points to the last
619  *      triple of that (one with zero ->key). Upon the exit we have the same
620  *      picture as after the successful ext3_get_block(), except that in one
621  *      place chain is disconnected - *branch->p is still zero (we did not
622  *      set the last link), but branch->key contains the number that should
623  *      be placed into *branch->p to fill that gap.
624  *
625  *      If allocation fails we free all blocks we've allocated (and forget
626  *      their buffer_heads) and return the error value the from failed
627  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
628  *      as described above and return 0.
629  */
630 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
631                         int indirect_blks, int *blks, ext3_fsblk_t goal,
632                         int *offsets, Indirect *branch)
633 {
634         int blocksize = inode->i_sb->s_blocksize;
635         int i, n = 0;
636         int err = 0;
637         struct buffer_head *bh;
638         int num;
639         ext3_fsblk_t new_blocks[4];
640         ext3_fsblk_t current_block;
641
642         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
643                                 *blks, new_blocks, &err);
644         if (err)
645                 return err;
646
647         branch[0].key = cpu_to_le32(new_blocks[0]);
648         /*
649          * metadata blocks and data blocks are allocated.
650          */
651         for (n = 1; n <= indirect_blks;  n++) {
652                 /*
653                  * Get buffer_head for parent block, zero it out
654                  * and set the pointer to new one, then send
655                  * parent to disk.
656                  */
657                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
658                 branch[n].bh = bh;
659                 lock_buffer(bh);
660                 BUFFER_TRACE(bh, "call get_create_access");
661                 err = ext3_journal_get_create_access(handle, bh);
662                 if (err) {
663                         unlock_buffer(bh);
664                         brelse(bh);
665                         goto failed;
666                 }
667
668                 memset(bh->b_data, 0, blocksize);
669                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
670                 branch[n].key = cpu_to_le32(new_blocks[n]);
671                 *branch[n].p = branch[n].key;
672                 if ( n == indirect_blks) {
673                         current_block = new_blocks[n];
674                         /*
675                          * End of chain, update the last new metablock of
676                          * the chain to point to the new allocated
677                          * data blocks numbers
678                          */
679                         for (i=1; i < num; i++)
680                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
681                 }
682                 BUFFER_TRACE(bh, "marking uptodate");
683                 set_buffer_uptodate(bh);
684                 unlock_buffer(bh);
685
686                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
687                 err = ext3_journal_dirty_metadata(handle, bh);
688                 if (err)
689                         goto failed;
690         }
691         *blks = num;
692         return err;
693 failed:
694         /* Allocation failed, free what we already allocated */
695         for (i = 1; i <= n ; i++) {
696                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
697                 ext3_journal_forget(handle, branch[i].bh);
698         }
699         for (i = 0; i <indirect_blks; i++)
700                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
701
702         ext3_free_blocks(handle, inode, new_blocks[i], num);
703
704         return err;
705 }
706
707 /**
708  * ext3_splice_branch - splice the allocated branch onto inode.
709  * @handle: handle for this transaction
710  * @inode: owner
711  * @block: (logical) number of block we are adding
712  * @where: location of missing link
713  * @num:   number of indirect blocks we are adding
714  * @blks:  number of direct blocks we are adding
715  *
716  * This function fills the missing link and does all housekeeping needed in
717  * inode (->i_blocks, etc.). In case of success we end up with the full
718  * chain to new block and return 0.
719  */
720 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
721                         long block, Indirect *where, int num, int blks)
722 {
723         int i;
724         int err = 0;
725         struct ext3_block_alloc_info *block_i;
726         ext3_fsblk_t current_block;
727         struct ext3_inode_info *ei = EXT3_I(inode);
728
729         block_i = ei->i_block_alloc_info;
730         /*
731          * If we're splicing into a [td]indirect block (as opposed to the
732          * inode) then we need to get write access to the [td]indirect block
733          * before the splice.
734          */
735         if (where->bh) {
736                 BUFFER_TRACE(where->bh, "get_write_access");
737                 err = ext3_journal_get_write_access(handle, where->bh);
738                 if (err)
739                         goto err_out;
740         }
741         /* That's it */
742
743         *where->p = where->key;
744
745         /*
746          * Update the host buffer_head or inode to point to more just allocated
747          * direct blocks blocks
748          */
749         if (num == 0 && blks > 1) {
750                 current_block = le32_to_cpu(where->key) + 1;
751                 for (i = 1; i < blks; i++)
752                         *(where->p + i ) = cpu_to_le32(current_block++);
753         }
754
755         /*
756          * update the most recently allocated logical & physical block
757          * in i_block_alloc_info, to assist find the proper goal block for next
758          * allocation
759          */
760         if (block_i) {
761                 block_i->last_alloc_logical_block = block + blks - 1;
762                 block_i->last_alloc_physical_block =
763                                 le32_to_cpu(where[num].key) + blks - 1;
764         }
765
766         /* We are done with atomic stuff, now do the rest of housekeeping */
767
768         inode->i_ctime = CURRENT_TIME_SEC;
769         ext3_mark_inode_dirty(handle, inode);
770         /* ext3_mark_inode_dirty already updated i_sync_tid */
771         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
772
773         /* had we spliced it onto indirect block? */
774         if (where->bh) {
775                 /*
776                  * If we spliced it onto an indirect block, we haven't
777                  * altered the inode.  Note however that if it is being spliced
778                  * onto an indirect block at the very end of the file (the
779                  * file is growing) then we *will* alter the inode to reflect
780                  * the new i_size.  But that is not done here - it is done in
781                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
782                  */
783                 jbd_debug(5, "splicing indirect only\n");
784                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
785                 err = ext3_journal_dirty_metadata(handle, where->bh);
786                 if (err)
787                         goto err_out;
788         } else {
789                 /*
790                  * OK, we spliced it into the inode itself on a direct block.
791                  * Inode was dirtied above.
792                  */
793                 jbd_debug(5, "splicing direct\n");
794         }
795         return err;
796
797 err_out:
798         for (i = 1; i <= num; i++) {
799                 BUFFER_TRACE(where[i].bh, "call journal_forget");
800                 ext3_journal_forget(handle, where[i].bh);
801                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
802         }
803         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
804
805         return err;
806 }
807
808 /*
809  * Allocation strategy is simple: if we have to allocate something, we will
810  * have to go the whole way to leaf. So let's do it before attaching anything
811  * to tree, set linkage between the newborn blocks, write them if sync is
812  * required, recheck the path, free and repeat if check fails, otherwise
813  * set the last missing link (that will protect us from any truncate-generated
814  * removals - all blocks on the path are immune now) and possibly force the
815  * write on the parent block.
816  * That has a nice additional property: no special recovery from the failed
817  * allocations is needed - we simply release blocks and do not touch anything
818  * reachable from inode.
819  *
820  * `handle' can be NULL if create == 0.
821  *
822  * The BKL may not be held on entry here.  Be sure to take it early.
823  * return > 0, # of blocks mapped or allocated.
824  * return = 0, if plain lookup failed.
825  * return < 0, error case.
826  */
827 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
828                 sector_t iblock, unsigned long maxblocks,
829                 struct buffer_head *bh_result,
830                 int create)
831 {
832         int err = -EIO;
833         int offsets[4];
834         Indirect chain[4];
835         Indirect *partial;
836         ext3_fsblk_t goal;
837         int indirect_blks;
838         int blocks_to_boundary = 0;
839         int depth;
840         struct ext3_inode_info *ei = EXT3_I(inode);
841         int count = 0;
842         ext3_fsblk_t first_block = 0;
843
844
845         J_ASSERT(handle != NULL || create == 0);
846         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
847
848         if (depth == 0)
849                 goto out;
850
851         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
852
853         /* Simplest case - block found, no allocation needed */
854         if (!partial) {
855                 first_block = le32_to_cpu(chain[depth - 1].key);
856                 clear_buffer_new(bh_result);
857                 count++;
858                 /*map more blocks*/
859                 while (count < maxblocks && count <= blocks_to_boundary) {
860                         ext3_fsblk_t blk;
861
862                         if (!verify_chain(chain, chain + depth - 1)) {
863                                 /*
864                                  * Indirect block might be removed by
865                                  * truncate while we were reading it.
866                                  * Handling of that case: forget what we've
867                                  * got now. Flag the err as EAGAIN, so it
868                                  * will reread.
869                                  */
870                                 err = -EAGAIN;
871                                 count = 0;
872                                 break;
873                         }
874                         blk = le32_to_cpu(*(chain[depth-1].p + count));
875
876                         if (blk == first_block + count)
877                                 count++;
878                         else
879                                 break;
880                 }
881                 if (err != -EAGAIN)
882                         goto got_it;
883         }
884
885         /* Next simple case - plain lookup or failed read of indirect block */
886         if (!create || err == -EIO)
887                 goto cleanup;
888
889         mutex_lock(&ei->truncate_mutex);
890
891         /*
892          * If the indirect block is missing while we are reading
893          * the chain(ext3_get_branch() returns -EAGAIN err), or
894          * if the chain has been changed after we grab the semaphore,
895          * (either because another process truncated this branch, or
896          * another get_block allocated this branch) re-grab the chain to see if
897          * the request block has been allocated or not.
898          *
899          * Since we already block the truncate/other get_block
900          * at this point, we will have the current copy of the chain when we
901          * splice the branch into the tree.
902          */
903         if (err == -EAGAIN || !verify_chain(chain, partial)) {
904                 while (partial > chain) {
905                         brelse(partial->bh);
906                         partial--;
907                 }
908                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
909                 if (!partial) {
910                         count++;
911                         mutex_unlock(&ei->truncate_mutex);
912                         if (err)
913                                 goto cleanup;
914                         clear_buffer_new(bh_result);
915                         goto got_it;
916                 }
917         }
918
919         /*
920          * Okay, we need to do block allocation.  Lazily initialize the block
921          * allocation info here if necessary
922         */
923         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
924                 ext3_init_block_alloc_info(inode);
925
926         goal = ext3_find_goal(inode, iblock, partial);
927
928         /* the number of blocks need to allocate for [d,t]indirect blocks */
929         indirect_blks = (chain + depth) - partial - 1;
930
931         /*
932          * Next look up the indirect map to count the totoal number of
933          * direct blocks to allocate for this branch.
934          */
935         count = ext3_blks_to_allocate(partial, indirect_blks,
936                                         maxblocks, blocks_to_boundary);
937         /*
938          * Block out ext3_truncate while we alter the tree
939          */
940         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
941                                 offsets + (partial - chain), partial);
942
943         /*
944          * The ext3_splice_branch call will free and forget any buffers
945          * on the new chain if there is a failure, but that risks using
946          * up transaction credits, especially for bitmaps where the
947          * credits cannot be returned.  Can we handle this somehow?  We
948          * may need to return -EAGAIN upwards in the worst case.  --sct
949          */
950         if (!err)
951                 err = ext3_splice_branch(handle, inode, iblock,
952                                         partial, indirect_blks, count);
953         mutex_unlock(&ei->truncate_mutex);
954         if (err)
955                 goto cleanup;
956
957         set_buffer_new(bh_result);
958 got_it:
959         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
960         if (count > blocks_to_boundary)
961                 set_buffer_boundary(bh_result);
962         err = count;
963         /* Clean up and exit */
964         partial = chain + depth - 1;    /* the whole chain */
965 cleanup:
966         while (partial > chain) {
967                 BUFFER_TRACE(partial->bh, "call brelse");
968                 brelse(partial->bh);
969                 partial--;
970         }
971         BUFFER_TRACE(bh_result, "returned");
972 out:
973         return err;
974 }
975
976 /* Maximum number of blocks we map for direct IO at once. */
977 #define DIO_MAX_BLOCKS 4096
978 /*
979  * Number of credits we need for writing DIO_MAX_BLOCKS:
980  * We need sb + group descriptor + bitmap + inode -> 4
981  * For B blocks with A block pointers per block we need:
982  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
983  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
984  */
985 #define DIO_CREDITS 25
986
987 static int ext3_get_block(struct inode *inode, sector_t iblock,
988                         struct buffer_head *bh_result, int create)
989 {
990         handle_t *handle = ext3_journal_current_handle();
991         int ret = 0, started = 0;
992         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
993
994         if (create && !handle) {        /* Direct IO write... */
995                 if (max_blocks > DIO_MAX_BLOCKS)
996                         max_blocks = DIO_MAX_BLOCKS;
997                 handle = ext3_journal_start(inode, DIO_CREDITS +
998                                 EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb));
999                 if (IS_ERR(handle)) {
1000                         ret = PTR_ERR(handle);
1001                         goto out;
1002                 }
1003                 started = 1;
1004         }
1005
1006         ret = ext3_get_blocks_handle(handle, inode, iblock,
1007                                         max_blocks, bh_result, create);
1008         if (ret > 0) {
1009                 bh_result->b_size = (ret << inode->i_blkbits);
1010                 ret = 0;
1011         }
1012         if (started)
1013                 ext3_journal_stop(handle);
1014 out:
1015         return ret;
1016 }
1017
1018 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
1019                 u64 start, u64 len)
1020 {
1021         return generic_block_fiemap(inode, fieinfo, start, len,
1022                                     ext3_get_block);
1023 }
1024
1025 /*
1026  * `handle' can be NULL if create is zero
1027  */
1028 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1029                                 long block, int create, int *errp)
1030 {
1031         struct buffer_head dummy;
1032         int fatal = 0, err;
1033
1034         J_ASSERT(handle != NULL || create == 0);
1035
1036         dummy.b_state = 0;
1037         dummy.b_blocknr = -1000;
1038         buffer_trace_init(&dummy.b_history);
1039         err = ext3_get_blocks_handle(handle, inode, block, 1,
1040                                         &dummy, create);
1041         /*
1042          * ext3_get_blocks_handle() returns number of blocks
1043          * mapped. 0 in case of a HOLE.
1044          */
1045         if (err > 0) {
1046                 if (err > 1)
1047                         WARN_ON(1);
1048                 err = 0;
1049         }
1050         *errp = err;
1051         if (!err && buffer_mapped(&dummy)) {
1052                 struct buffer_head *bh;
1053                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1054                 if (!bh) {
1055                         *errp = -EIO;
1056                         goto err;
1057                 }
1058                 if (buffer_new(&dummy)) {
1059                         J_ASSERT(create != 0);
1060                         J_ASSERT(handle != NULL);
1061
1062                         /*
1063                          * Now that we do not always journal data, we should
1064                          * keep in mind whether this should always journal the
1065                          * new buffer as metadata.  For now, regular file
1066                          * writes use ext3_get_block instead, so it's not a
1067                          * problem.
1068                          */
1069                         lock_buffer(bh);
1070                         BUFFER_TRACE(bh, "call get_create_access");
1071                         fatal = ext3_journal_get_create_access(handle, bh);
1072                         if (!fatal && !buffer_uptodate(bh)) {
1073                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1074                                 set_buffer_uptodate(bh);
1075                         }
1076                         unlock_buffer(bh);
1077                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1078                         err = ext3_journal_dirty_metadata(handle, bh);
1079                         if (!fatal)
1080                                 fatal = err;
1081                 } else {
1082                         BUFFER_TRACE(bh, "not a new buffer");
1083                 }
1084                 if (fatal) {
1085                         *errp = fatal;
1086                         brelse(bh);
1087                         bh = NULL;
1088                 }
1089                 return bh;
1090         }
1091 err:
1092         return NULL;
1093 }
1094
1095 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1096                                int block, int create, int *err)
1097 {
1098         struct buffer_head * bh;
1099
1100         bh = ext3_getblk(handle, inode, block, create, err);
1101         if (!bh)
1102                 return bh;
1103         if (buffer_uptodate(bh))
1104                 return bh;
1105         ll_rw_block(READ_META, 1, &bh);
1106         wait_on_buffer(bh);
1107         if (buffer_uptodate(bh))
1108                 return bh;
1109         put_bh(bh);
1110         *err = -EIO;
1111         return NULL;
1112 }
1113
1114 static int walk_page_buffers(   handle_t *handle,
1115                                 struct buffer_head *head,
1116                                 unsigned from,
1117                                 unsigned to,
1118                                 int *partial,
1119                                 int (*fn)(      handle_t *handle,
1120                                                 struct buffer_head *bh))
1121 {
1122         struct buffer_head *bh;
1123         unsigned block_start, block_end;
1124         unsigned blocksize = head->b_size;
1125         int err, ret = 0;
1126         struct buffer_head *next;
1127
1128         for (   bh = head, block_start = 0;
1129                 ret == 0 && (bh != head || !block_start);
1130                 block_start = block_end, bh = next)
1131         {
1132                 next = bh->b_this_page;
1133                 block_end = block_start + blocksize;
1134                 if (block_end <= from || block_start >= to) {
1135                         if (partial && !buffer_uptodate(bh))
1136                                 *partial = 1;
1137                         continue;
1138                 }
1139                 err = (*fn)(handle, bh);
1140                 if (!ret)
1141                         ret = err;
1142         }
1143         return ret;
1144 }
1145
1146 /*
1147  * To preserve ordering, it is essential that the hole instantiation and
1148  * the data write be encapsulated in a single transaction.  We cannot
1149  * close off a transaction and start a new one between the ext3_get_block()
1150  * and the commit_write().  So doing the journal_start at the start of
1151  * prepare_write() is the right place.
1152  *
1153  * Also, this function can nest inside ext3_writepage() ->
1154  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1155  * has generated enough buffer credits to do the whole page.  So we won't
1156  * block on the journal in that case, which is good, because the caller may
1157  * be PF_MEMALLOC.
1158  *
1159  * By accident, ext3 can be reentered when a transaction is open via
1160  * quota file writes.  If we were to commit the transaction while thus
1161  * reentered, there can be a deadlock - we would be holding a quota
1162  * lock, and the commit would never complete if another thread had a
1163  * transaction open and was blocking on the quota lock - a ranking
1164  * violation.
1165  *
1166  * So what we do is to rely on the fact that journal_stop/journal_start
1167  * will _not_ run commit under these circumstances because handle->h_ref
1168  * is elevated.  We'll still have enough credits for the tiny quotafile
1169  * write.
1170  */
1171 static int do_journal_get_write_access(handle_t *handle,
1172                                         struct buffer_head *bh)
1173 {
1174         int dirty = buffer_dirty(bh);
1175         int ret;
1176
1177         if (!buffer_mapped(bh) || buffer_freed(bh))
1178                 return 0;
1179         /*
1180          * __block_prepare_write() could have dirtied some buffers. Clean
1181          * the dirty bit as jbd2_journal_get_write_access() could complain
1182          * otherwise about fs integrity issues. Setting of the dirty bit
1183          * by __block_prepare_write() isn't a real problem here as we clear
1184          * the bit before releasing a page lock and thus writeback cannot
1185          * ever write the buffer.
1186          */
1187         if (dirty)
1188                 clear_buffer_dirty(bh);
1189         ret = ext3_journal_get_write_access(handle, bh);
1190         if (!ret && dirty)
1191                 ret = ext3_journal_dirty_metadata(handle, bh);
1192         return ret;
1193 }
1194
1195 /*
1196  * Truncate blocks that were not used by write. We have to truncate the
1197  * pagecache as well so that corresponding buffers get properly unmapped.
1198  */
1199 static void ext3_truncate_failed_write(struct inode *inode)
1200 {
1201         truncate_inode_pages(inode->i_mapping, inode->i_size);
1202         ext3_truncate(inode);
1203 }
1204
1205 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1206                                 loff_t pos, unsigned len, unsigned flags,
1207                                 struct page **pagep, void **fsdata)
1208 {
1209         struct inode *inode = mapping->host;
1210         int ret;
1211         handle_t *handle;
1212         int retries = 0;
1213         struct page *page;
1214         pgoff_t index;
1215         unsigned from, to;
1216         /* Reserve one block more for addition to orphan list in case
1217          * we allocate blocks but write fails for some reason */
1218         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1219
1220         index = pos >> PAGE_CACHE_SHIFT;
1221         from = pos & (PAGE_CACHE_SIZE - 1);
1222         to = from + len;
1223
1224 retry:
1225         page = grab_cache_page_write_begin(mapping, index, flags);
1226         if (!page)
1227                 return -ENOMEM;
1228         *pagep = page;
1229
1230         handle = ext3_journal_start(inode, needed_blocks);
1231         if (IS_ERR(handle)) {
1232                 unlock_page(page);
1233                 page_cache_release(page);
1234                 ret = PTR_ERR(handle);
1235                 goto out;
1236         }
1237         ret = __block_write_begin(page, pos, len, ext3_get_block);
1238         if (ret)
1239                 goto write_begin_failed;
1240
1241         if (ext3_should_journal_data(inode)) {
1242                 ret = walk_page_buffers(handle, page_buffers(page),
1243                                 from, to, NULL, do_journal_get_write_access);
1244         }
1245 write_begin_failed:
1246         if (ret) {
1247                 /*
1248                  * block_write_begin may have instantiated a few blocks
1249                  * outside i_size.  Trim these off again. Don't need
1250                  * i_size_read because we hold i_mutex.
1251                  *
1252                  * Add inode to orphan list in case we crash before truncate
1253                  * finishes. Do this only if ext3_can_truncate() agrees so
1254                  * that orphan processing code is happy.
1255                  */
1256                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1257                         ext3_orphan_add(handle, inode);
1258                 ext3_journal_stop(handle);
1259                 unlock_page(page);
1260                 page_cache_release(page);
1261                 if (pos + len > inode->i_size)
1262                         ext3_truncate_failed_write(inode);
1263         }
1264         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1265                 goto retry;
1266 out:
1267         return ret;
1268 }
1269
1270
1271 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1272 {
1273         int err = journal_dirty_data(handle, bh);
1274         if (err)
1275                 ext3_journal_abort_handle(__func__, __func__,
1276                                                 bh, handle, err);
1277         return err;
1278 }
1279
1280 /* For ordered writepage and write_end functions */
1281 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1282 {
1283         /*
1284          * Write could have mapped the buffer but it didn't copy the data in
1285          * yet. So avoid filing such buffer into a transaction.
1286          */
1287         if (buffer_mapped(bh) && buffer_uptodate(bh))
1288                 return ext3_journal_dirty_data(handle, bh);
1289         return 0;
1290 }
1291
1292 /* For write_end() in data=journal mode */
1293 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1294 {
1295         if (!buffer_mapped(bh) || buffer_freed(bh))
1296                 return 0;
1297         set_buffer_uptodate(bh);
1298         return ext3_journal_dirty_metadata(handle, bh);
1299 }
1300
1301 /*
1302  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1303  * for the whole page but later we failed to copy the data in. Update inode
1304  * size according to what we managed to copy. The rest is going to be
1305  * truncated in write_end function.
1306  */
1307 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1308 {
1309         /* What matters to us is i_disksize. We don't write i_size anywhere */
1310         if (pos + copied > inode->i_size)
1311                 i_size_write(inode, pos + copied);
1312         if (pos + copied > EXT3_I(inode)->i_disksize) {
1313                 EXT3_I(inode)->i_disksize = pos + copied;
1314                 mark_inode_dirty(inode);
1315         }
1316 }
1317
1318 /*
1319  * We need to pick up the new inode size which generic_commit_write gave us
1320  * `file' can be NULL - eg, when called from page_symlink().
1321  *
1322  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1323  * buffers are managed internally.
1324  */
1325 static int ext3_ordered_write_end(struct file *file,
1326                                 struct address_space *mapping,
1327                                 loff_t pos, unsigned len, unsigned copied,
1328                                 struct page *page, void *fsdata)
1329 {
1330         handle_t *handle = ext3_journal_current_handle();
1331         struct inode *inode = file->f_mapping->host;
1332         unsigned from, to;
1333         int ret = 0, ret2;
1334
1335         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1336
1337         from = pos & (PAGE_CACHE_SIZE - 1);
1338         to = from + copied;
1339         ret = walk_page_buffers(handle, page_buffers(page),
1340                 from, to, NULL, journal_dirty_data_fn);
1341
1342         if (ret == 0)
1343                 update_file_sizes(inode, pos, copied);
1344         /*
1345          * There may be allocated blocks outside of i_size because
1346          * we failed to copy some data. Prepare for truncate.
1347          */
1348         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1349                 ext3_orphan_add(handle, inode);
1350         ret2 = ext3_journal_stop(handle);
1351         if (!ret)
1352                 ret = ret2;
1353         unlock_page(page);
1354         page_cache_release(page);
1355
1356         if (pos + len > inode->i_size)
1357                 ext3_truncate_failed_write(inode);
1358         return ret ? ret : copied;
1359 }
1360
1361 static int ext3_writeback_write_end(struct file *file,
1362                                 struct address_space *mapping,
1363                                 loff_t pos, unsigned len, unsigned copied,
1364                                 struct page *page, void *fsdata)
1365 {
1366         handle_t *handle = ext3_journal_current_handle();
1367         struct inode *inode = file->f_mapping->host;
1368         int ret;
1369
1370         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1371         update_file_sizes(inode, pos, copied);
1372         /*
1373          * There may be allocated blocks outside of i_size because
1374          * we failed to copy some data. Prepare for truncate.
1375          */
1376         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1377                 ext3_orphan_add(handle, inode);
1378         ret = ext3_journal_stop(handle);
1379         unlock_page(page);
1380         page_cache_release(page);
1381
1382         if (pos + len > inode->i_size)
1383                 ext3_truncate_failed_write(inode);
1384         return ret ? ret : copied;
1385 }
1386
1387 static int ext3_journalled_write_end(struct file *file,
1388                                 struct address_space *mapping,
1389                                 loff_t pos, unsigned len, unsigned copied,
1390                                 struct page *page, void *fsdata)
1391 {
1392         handle_t *handle = ext3_journal_current_handle();
1393         struct inode *inode = mapping->host;
1394         int ret = 0, ret2;
1395         int partial = 0;
1396         unsigned from, to;
1397
1398         from = pos & (PAGE_CACHE_SIZE - 1);
1399         to = from + len;
1400
1401         if (copied < len) {
1402                 if (!PageUptodate(page))
1403                         copied = 0;
1404                 page_zero_new_buffers(page, from + copied, to);
1405                 to = from + copied;
1406         }
1407
1408         ret = walk_page_buffers(handle, page_buffers(page), from,
1409                                 to, &partial, write_end_fn);
1410         if (!partial)
1411                 SetPageUptodate(page);
1412
1413         if (pos + copied > inode->i_size)
1414                 i_size_write(inode, pos + copied);
1415         /*
1416          * There may be allocated blocks outside of i_size because
1417          * we failed to copy some data. Prepare for truncate.
1418          */
1419         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1420                 ext3_orphan_add(handle, inode);
1421         ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1422         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1423                 EXT3_I(inode)->i_disksize = inode->i_size;
1424                 ret2 = ext3_mark_inode_dirty(handle, inode);
1425                 if (!ret)
1426                         ret = ret2;
1427         }
1428
1429         ret2 = ext3_journal_stop(handle);
1430         if (!ret)
1431                 ret = ret2;
1432         unlock_page(page);
1433         page_cache_release(page);
1434
1435         if (pos + len > inode->i_size)
1436                 ext3_truncate_failed_write(inode);
1437         return ret ? ret : copied;
1438 }
1439
1440 /*
1441  * bmap() is special.  It gets used by applications such as lilo and by
1442  * the swapper to find the on-disk block of a specific piece of data.
1443  *
1444  * Naturally, this is dangerous if the block concerned is still in the
1445  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1446  * filesystem and enables swap, then they may get a nasty shock when the
1447  * data getting swapped to that swapfile suddenly gets overwritten by
1448  * the original zero's written out previously to the journal and
1449  * awaiting writeback in the kernel's buffer cache.
1450  *
1451  * So, if we see any bmap calls here on a modified, data-journaled file,
1452  * take extra steps to flush any blocks which might be in the cache.
1453  */
1454 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1455 {
1456         struct inode *inode = mapping->host;
1457         journal_t *journal;
1458         int err;
1459
1460         if (ext3_test_inode_state(inode, EXT3_STATE_JDATA)) {
1461                 /*
1462                  * This is a REALLY heavyweight approach, but the use of
1463                  * bmap on dirty files is expected to be extremely rare:
1464                  * only if we run lilo or swapon on a freshly made file
1465                  * do we expect this to happen.
1466                  *
1467                  * (bmap requires CAP_SYS_RAWIO so this does not
1468                  * represent an unprivileged user DOS attack --- we'd be
1469                  * in trouble if mortal users could trigger this path at
1470                  * will.)
1471                  *
1472                  * NB. EXT3_STATE_JDATA is not set on files other than
1473                  * regular files.  If somebody wants to bmap a directory
1474                  * or symlink and gets confused because the buffer
1475                  * hasn't yet been flushed to disk, they deserve
1476                  * everything they get.
1477                  */
1478
1479                 ext3_clear_inode_state(inode, EXT3_STATE_JDATA);
1480                 journal = EXT3_JOURNAL(inode);
1481                 journal_lock_updates(journal);
1482                 err = journal_flush(journal);
1483                 journal_unlock_updates(journal);
1484
1485                 if (err)
1486                         return 0;
1487         }
1488
1489         return generic_block_bmap(mapping,block,ext3_get_block);
1490 }
1491
1492 static int bget_one(handle_t *handle, struct buffer_head *bh)
1493 {
1494         get_bh(bh);
1495         return 0;
1496 }
1497
1498 static int bput_one(handle_t *handle, struct buffer_head *bh)
1499 {
1500         put_bh(bh);
1501         return 0;
1502 }
1503
1504 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1505 {
1506         return !buffer_mapped(bh);
1507 }
1508
1509 /*
1510  * Note that we always start a transaction even if we're not journalling
1511  * data.  This is to preserve ordering: any hole instantiation within
1512  * __block_write_full_page -> ext3_get_block() should be journalled
1513  * along with the data so we don't crash and then get metadata which
1514  * refers to old data.
1515  *
1516  * In all journalling modes block_write_full_page() will start the I/O.
1517  *
1518  * Problem:
1519  *
1520  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1521  *              ext3_writepage()
1522  *
1523  * Similar for:
1524  *
1525  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1526  *
1527  * Same applies to ext3_get_block().  We will deadlock on various things like
1528  * lock_journal and i_truncate_mutex.
1529  *
1530  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1531  * allocations fail.
1532  *
1533  * 16May01: If we're reentered then journal_current_handle() will be
1534  *          non-zero. We simply *return*.
1535  *
1536  * 1 July 2001: @@@ FIXME:
1537  *   In journalled data mode, a data buffer may be metadata against the
1538  *   current transaction.  But the same file is part of a shared mapping
1539  *   and someone does a writepage() on it.
1540  *
1541  *   We will move the buffer onto the async_data list, but *after* it has
1542  *   been dirtied. So there's a small window where we have dirty data on
1543  *   BJ_Metadata.
1544  *
1545  *   Note that this only applies to the last partial page in the file.  The
1546  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1547  *   broken code anyway: it's wrong for msync()).
1548  *
1549  *   It's a rare case: affects the final partial page, for journalled data
1550  *   where the file is subject to bith write() and writepage() in the same
1551  *   transction.  To fix it we'll need a custom block_write_full_page().
1552  *   We'll probably need that anyway for journalling writepage() output.
1553  *
1554  * We don't honour synchronous mounts for writepage().  That would be
1555  * disastrous.  Any write() or metadata operation will sync the fs for
1556  * us.
1557  *
1558  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1559  * we don't need to open a transaction here.
1560  */
1561 static int ext3_ordered_writepage(struct page *page,
1562                                 struct writeback_control *wbc)
1563 {
1564         struct inode *inode = page->mapping->host;
1565         struct buffer_head *page_bufs;
1566         handle_t *handle = NULL;
1567         int ret = 0;
1568         int err;
1569
1570         J_ASSERT(PageLocked(page));
1571         WARN_ON_ONCE(IS_RDONLY(inode));
1572
1573         /*
1574          * We give up here if we're reentered, because it might be for a
1575          * different filesystem.
1576          */
1577         if (ext3_journal_current_handle())
1578                 goto out_fail;
1579
1580         if (!page_has_buffers(page)) {
1581                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1582                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1583                 page_bufs = page_buffers(page);
1584         } else {
1585                 page_bufs = page_buffers(page);
1586                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1587                                        NULL, buffer_unmapped)) {
1588                         /* Provide NULL get_block() to catch bugs if buffers
1589                          * weren't really mapped */
1590                         return block_write_full_page(page, NULL, wbc);
1591                 }
1592         }
1593         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1594
1595         if (IS_ERR(handle)) {
1596                 ret = PTR_ERR(handle);
1597                 goto out_fail;
1598         }
1599
1600         walk_page_buffers(handle, page_bufs, 0,
1601                         PAGE_CACHE_SIZE, NULL, bget_one);
1602
1603         ret = block_write_full_page(page, ext3_get_block, wbc);
1604
1605         /*
1606          * The page can become unlocked at any point now, and
1607          * truncate can then come in and change things.  So we
1608          * can't touch *page from now on.  But *page_bufs is
1609          * safe due to elevated refcount.
1610          */
1611
1612         /*
1613          * And attach them to the current transaction.  But only if
1614          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1615          * and generally junk.
1616          */
1617         if (ret == 0) {
1618                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1619                                         NULL, journal_dirty_data_fn);
1620                 if (!ret)
1621                         ret = err;
1622         }
1623         walk_page_buffers(handle, page_bufs, 0,
1624                         PAGE_CACHE_SIZE, NULL, bput_one);
1625         err = ext3_journal_stop(handle);
1626         if (!ret)
1627                 ret = err;
1628         return ret;
1629
1630 out_fail:
1631         redirty_page_for_writepage(wbc, page);
1632         unlock_page(page);
1633         return ret;
1634 }
1635
1636 static int ext3_writeback_writepage(struct page *page,
1637                                 struct writeback_control *wbc)
1638 {
1639         struct inode *inode = page->mapping->host;
1640         handle_t *handle = NULL;
1641         int ret = 0;
1642         int err;
1643
1644         J_ASSERT(PageLocked(page));
1645         WARN_ON_ONCE(IS_RDONLY(inode));
1646
1647         if (ext3_journal_current_handle())
1648                 goto out_fail;
1649
1650         if (page_has_buffers(page)) {
1651                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1652                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1653                         /* Provide NULL get_block() to catch bugs if buffers
1654                          * weren't really mapped */
1655                         return block_write_full_page(page, NULL, wbc);
1656                 }
1657         }
1658
1659         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1660         if (IS_ERR(handle)) {
1661                 ret = PTR_ERR(handle);
1662                 goto out_fail;
1663         }
1664
1665         ret = block_write_full_page(page, ext3_get_block, wbc);
1666
1667         err = ext3_journal_stop(handle);
1668         if (!ret)
1669                 ret = err;
1670         return ret;
1671
1672 out_fail:
1673         redirty_page_for_writepage(wbc, page);
1674         unlock_page(page);
1675         return ret;
1676 }
1677
1678 static int ext3_journalled_writepage(struct page *page,
1679                                 struct writeback_control *wbc)
1680 {
1681         struct inode *inode = page->mapping->host;
1682         handle_t *handle = NULL;
1683         int ret = 0;
1684         int err;
1685
1686         J_ASSERT(PageLocked(page));
1687         WARN_ON_ONCE(IS_RDONLY(inode));
1688
1689         if (ext3_journal_current_handle())
1690                 goto no_write;
1691
1692         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1693         if (IS_ERR(handle)) {
1694                 ret = PTR_ERR(handle);
1695                 goto no_write;
1696         }
1697
1698         if (!page_has_buffers(page) || PageChecked(page)) {
1699                 /*
1700                  * It's mmapped pagecache.  Add buffers and journal it.  There
1701                  * doesn't seem much point in redirtying the page here.
1702                  */
1703                 ClearPageChecked(page);
1704                 ret = __block_write_begin(page, 0, PAGE_CACHE_SIZE,
1705                                           ext3_get_block);
1706                 if (ret != 0) {
1707                         ext3_journal_stop(handle);
1708                         goto out_unlock;
1709                 }
1710                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1711                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1712
1713                 err = walk_page_buffers(handle, page_buffers(page), 0,
1714                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1715                 if (ret == 0)
1716                         ret = err;
1717                 ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1718                 unlock_page(page);
1719         } else {
1720                 /*
1721                  * It may be a page full of checkpoint-mode buffers.  We don't
1722                  * really know unless we go poke around in the buffer_heads.
1723                  * But block_write_full_page will do the right thing.
1724                  */
1725                 ret = block_write_full_page(page, ext3_get_block, wbc);
1726         }
1727         err = ext3_journal_stop(handle);
1728         if (!ret)
1729                 ret = err;
1730 out:
1731         return ret;
1732
1733 no_write:
1734         redirty_page_for_writepage(wbc, page);
1735 out_unlock:
1736         unlock_page(page);
1737         goto out;
1738 }
1739
1740 static int ext3_readpage(struct file *file, struct page *page)
1741 {
1742         return mpage_readpage(page, ext3_get_block);
1743 }
1744
1745 static int
1746 ext3_readpages(struct file *file, struct address_space *mapping,
1747                 struct list_head *pages, unsigned nr_pages)
1748 {
1749         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1750 }
1751
1752 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1753 {
1754         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1755
1756         /*
1757          * If it's a full truncate we just forget about the pending dirtying
1758          */
1759         if (offset == 0)
1760                 ClearPageChecked(page);
1761
1762         journal_invalidatepage(journal, page, offset);
1763 }
1764
1765 static int ext3_releasepage(struct page *page, gfp_t wait)
1766 {
1767         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1768
1769         WARN_ON(PageChecked(page));
1770         if (!page_has_buffers(page))
1771                 return 0;
1772         return journal_try_to_free_buffers(journal, page, wait);
1773 }
1774
1775 /*
1776  * If the O_DIRECT write will extend the file then add this inode to the
1777  * orphan list.  So recovery will truncate it back to the original size
1778  * if the machine crashes during the write.
1779  *
1780  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1781  * crashes then stale disk data _may_ be exposed inside the file. But current
1782  * VFS code falls back into buffered path in that case so we are safe.
1783  */
1784 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1785                         const struct iovec *iov, loff_t offset,
1786                         unsigned long nr_segs)
1787 {
1788         struct file *file = iocb->ki_filp;
1789         struct inode *inode = file->f_mapping->host;
1790         struct ext3_inode_info *ei = EXT3_I(inode);
1791         handle_t *handle;
1792         ssize_t ret;
1793         int orphan = 0;
1794         size_t count = iov_length(iov, nr_segs);
1795         int retries = 0;
1796
1797         if (rw == WRITE) {
1798                 loff_t final_size = offset + count;
1799
1800                 if (final_size > inode->i_size) {
1801                         /* Credits for sb + inode write */
1802                         handle = ext3_journal_start(inode, 2);
1803                         if (IS_ERR(handle)) {
1804                                 ret = PTR_ERR(handle);
1805                                 goto out;
1806                         }
1807                         ret = ext3_orphan_add(handle, inode);
1808                         if (ret) {
1809                                 ext3_journal_stop(handle);
1810                                 goto out;
1811                         }
1812                         orphan = 1;
1813                         ei->i_disksize = inode->i_size;
1814                         ext3_journal_stop(handle);
1815                 }
1816         }
1817
1818 retry:
1819         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1820                                  offset, nr_segs,
1821                                  ext3_get_block, NULL);
1822         /*
1823          * In case of error extending write may have instantiated a few
1824          * blocks outside i_size. Trim these off again.
1825          */
1826         if (unlikely((rw & WRITE) && ret < 0)) {
1827                 loff_t isize = i_size_read(inode);
1828                 loff_t end = offset + iov_length(iov, nr_segs);
1829
1830                 if (end > isize)
1831                         vmtruncate(inode, isize);
1832         }
1833         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1834                 goto retry;
1835
1836         if (orphan) {
1837                 int err;
1838
1839                 /* Credits for sb + inode write */
1840                 handle = ext3_journal_start(inode, 2);
1841                 if (IS_ERR(handle)) {
1842                         /* This is really bad luck. We've written the data
1843                          * but cannot extend i_size. Truncate allocated blocks
1844                          * and pretend the write failed... */
1845                         ext3_truncate(inode);
1846                         ret = PTR_ERR(handle);
1847                         goto out;
1848                 }
1849                 if (inode->i_nlink)
1850                         ext3_orphan_del(handle, inode);
1851                 if (ret > 0) {
1852                         loff_t end = offset + ret;
1853                         if (end > inode->i_size) {
1854                                 ei->i_disksize = end;
1855                                 i_size_write(inode, end);
1856                                 /*
1857                                  * We're going to return a positive `ret'
1858                                  * here due to non-zero-length I/O, so there's
1859                                  * no way of reporting error returns from
1860                                  * ext3_mark_inode_dirty() to userspace.  So
1861                                  * ignore it.
1862                                  */
1863                                 ext3_mark_inode_dirty(handle, inode);
1864                         }
1865                 }
1866                 err = ext3_journal_stop(handle);
1867                 if (ret == 0)
1868                         ret = err;
1869         }
1870 out:
1871         return ret;
1872 }
1873
1874 /*
1875  * Pages can be marked dirty completely asynchronously from ext3's journalling
1876  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1877  * much here because ->set_page_dirty is called under VFS locks.  The page is
1878  * not necessarily locked.
1879  *
1880  * We cannot just dirty the page and leave attached buffers clean, because the
1881  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1882  * or jbddirty because all the journalling code will explode.
1883  *
1884  * So what we do is to mark the page "pending dirty" and next time writepage
1885  * is called, propagate that into the buffers appropriately.
1886  */
1887 static int ext3_journalled_set_page_dirty(struct page *page)
1888 {
1889         SetPageChecked(page);
1890         return __set_page_dirty_nobuffers(page);
1891 }
1892
1893 static const struct address_space_operations ext3_ordered_aops = {
1894         .readpage               = ext3_readpage,
1895         .readpages              = ext3_readpages,
1896         .writepage              = ext3_ordered_writepage,
1897         .sync_page              = block_sync_page,
1898         .write_begin            = ext3_write_begin,
1899         .write_end              = ext3_ordered_write_end,
1900         .bmap                   = ext3_bmap,
1901         .invalidatepage         = ext3_invalidatepage,
1902         .releasepage            = ext3_releasepage,
1903         .direct_IO              = ext3_direct_IO,
1904         .migratepage            = buffer_migrate_page,
1905         .is_partially_uptodate  = block_is_partially_uptodate,
1906         .error_remove_page      = generic_error_remove_page,
1907 };
1908
1909 static const struct address_space_operations ext3_writeback_aops = {
1910         .readpage               = ext3_readpage,
1911         .readpages              = ext3_readpages,
1912         .writepage              = ext3_writeback_writepage,
1913         .sync_page              = block_sync_page,
1914         .write_begin            = ext3_write_begin,
1915         .write_end              = ext3_writeback_write_end,
1916         .bmap                   = ext3_bmap,
1917         .invalidatepage         = ext3_invalidatepage,
1918         .releasepage            = ext3_releasepage,
1919         .direct_IO              = ext3_direct_IO,
1920         .migratepage            = buffer_migrate_page,
1921         .is_partially_uptodate  = block_is_partially_uptodate,
1922         .error_remove_page      = generic_error_remove_page,
1923 };
1924
1925 static const struct address_space_operations ext3_journalled_aops = {
1926         .readpage               = ext3_readpage,
1927         .readpages              = ext3_readpages,
1928         .writepage              = ext3_journalled_writepage,
1929         .sync_page              = block_sync_page,
1930         .write_begin            = ext3_write_begin,
1931         .write_end              = ext3_journalled_write_end,
1932         .set_page_dirty         = ext3_journalled_set_page_dirty,
1933         .bmap                   = ext3_bmap,
1934         .invalidatepage         = ext3_invalidatepage,
1935         .releasepage            = ext3_releasepage,
1936         .is_partially_uptodate  = block_is_partially_uptodate,
1937         .error_remove_page      = generic_error_remove_page,
1938 };
1939
1940 void ext3_set_aops(struct inode *inode)
1941 {
1942         if (ext3_should_order_data(inode))
1943                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1944         else if (ext3_should_writeback_data(inode))
1945                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1946         else
1947                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1948 }
1949
1950 /*
1951  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1952  * up to the end of the block which corresponds to `from'.
1953  * This required during truncate. We need to physically zero the tail end
1954  * of that block so it doesn't yield old data if the file is later grown.
1955  */
1956 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1957                 struct address_space *mapping, loff_t from)
1958 {
1959         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1960         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1961         unsigned blocksize, iblock, length, pos;
1962         struct inode *inode = mapping->host;
1963         struct buffer_head *bh;
1964         int err = 0;
1965
1966         blocksize = inode->i_sb->s_blocksize;
1967         length = blocksize - (offset & (blocksize - 1));
1968         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1969
1970         if (!page_has_buffers(page))
1971                 create_empty_buffers(page, blocksize, 0);
1972
1973         /* Find the buffer that contains "offset" */
1974         bh = page_buffers(page);
1975         pos = blocksize;
1976         while (offset >= pos) {
1977                 bh = bh->b_this_page;
1978                 iblock++;
1979                 pos += blocksize;
1980         }
1981
1982         err = 0;
1983         if (buffer_freed(bh)) {
1984                 BUFFER_TRACE(bh, "freed: skip");
1985                 goto unlock;
1986         }
1987
1988         if (!buffer_mapped(bh)) {
1989                 BUFFER_TRACE(bh, "unmapped");
1990                 ext3_get_block(inode, iblock, bh, 0);
1991                 /* unmapped? It's a hole - nothing to do */
1992                 if (!buffer_mapped(bh)) {
1993                         BUFFER_TRACE(bh, "still unmapped");
1994                         goto unlock;
1995                 }
1996         }
1997
1998         /* Ok, it's mapped. Make sure it's up-to-date */
1999         if (PageUptodate(page))
2000                 set_buffer_uptodate(bh);
2001
2002         if (!buffer_uptodate(bh)) {
2003                 err = -EIO;
2004                 ll_rw_block(READ, 1, &bh);
2005                 wait_on_buffer(bh);
2006                 /* Uhhuh. Read error. Complain and punt. */
2007                 if (!buffer_uptodate(bh))
2008                         goto unlock;
2009         }
2010
2011         if (ext3_should_journal_data(inode)) {
2012                 BUFFER_TRACE(bh, "get write access");
2013                 err = ext3_journal_get_write_access(handle, bh);
2014                 if (err)
2015                         goto unlock;
2016         }
2017
2018         zero_user(page, offset, length);
2019         BUFFER_TRACE(bh, "zeroed end of block");
2020
2021         err = 0;
2022         if (ext3_should_journal_data(inode)) {
2023                 err = ext3_journal_dirty_metadata(handle, bh);
2024         } else {
2025                 if (ext3_should_order_data(inode))
2026                         err = ext3_journal_dirty_data(handle, bh);
2027                 mark_buffer_dirty(bh);
2028         }
2029
2030 unlock:
2031         unlock_page(page);
2032         page_cache_release(page);
2033         return err;
2034 }
2035
2036 /*
2037  * Probably it should be a library function... search for first non-zero word
2038  * or memcmp with zero_page, whatever is better for particular architecture.
2039  * Linus?
2040  */
2041 static inline int all_zeroes(__le32 *p, __le32 *q)
2042 {
2043         while (p < q)
2044                 if (*p++)
2045                         return 0;
2046         return 1;
2047 }
2048
2049 /**
2050  *      ext3_find_shared - find the indirect blocks for partial truncation.
2051  *      @inode:   inode in question
2052  *      @depth:   depth of the affected branch
2053  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
2054  *      @chain:   place to store the pointers to partial indirect blocks
2055  *      @top:     place to the (detached) top of branch
2056  *
2057  *      This is a helper function used by ext3_truncate().
2058  *
2059  *      When we do truncate() we may have to clean the ends of several
2060  *      indirect blocks but leave the blocks themselves alive. Block is
2061  *      partially truncated if some data below the new i_size is refered
2062  *      from it (and it is on the path to the first completely truncated
2063  *      data block, indeed).  We have to free the top of that path along
2064  *      with everything to the right of the path. Since no allocation
2065  *      past the truncation point is possible until ext3_truncate()
2066  *      finishes, we may safely do the latter, but top of branch may
2067  *      require special attention - pageout below the truncation point
2068  *      might try to populate it.
2069  *
2070  *      We atomically detach the top of branch from the tree, store the
2071  *      block number of its root in *@top, pointers to buffer_heads of
2072  *      partially truncated blocks - in @chain[].bh and pointers to
2073  *      their last elements that should not be removed - in
2074  *      @chain[].p. Return value is the pointer to last filled element
2075  *      of @chain.
2076  *
2077  *      The work left to caller to do the actual freeing of subtrees:
2078  *              a) free the subtree starting from *@top
2079  *              b) free the subtrees whose roots are stored in
2080  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2081  *              c) free the subtrees growing from the inode past the @chain[0].
2082  *                      (no partially truncated stuff there).  */
2083
2084 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2085                         int offsets[4], Indirect chain[4], __le32 *top)
2086 {
2087         Indirect *partial, *p;
2088         int k, err;
2089
2090         *top = 0;
2091         /* Make k index the deepest non-null offset + 1 */
2092         for (k = depth; k > 1 && !offsets[k-1]; k--)
2093                 ;
2094         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2095         /* Writer: pointers */
2096         if (!partial)
2097                 partial = chain + k-1;
2098         /*
2099          * If the branch acquired continuation since we've looked at it -
2100          * fine, it should all survive and (new) top doesn't belong to us.
2101          */
2102         if (!partial->key && *partial->p)
2103                 /* Writer: end */
2104                 goto no_top;
2105         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2106                 ;
2107         /*
2108          * OK, we've found the last block that must survive. The rest of our
2109          * branch should be detached before unlocking. However, if that rest
2110          * of branch is all ours and does not grow immediately from the inode
2111          * it's easier to cheat and just decrement partial->p.
2112          */
2113         if (p == chain + k - 1 && p > chain) {
2114                 p->p--;
2115         } else {
2116                 *top = *p->p;
2117                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2118 #if 0
2119                 *p->p = 0;
2120 #endif
2121         }
2122         /* Writer: end */
2123
2124         while(partial > p) {
2125                 brelse(partial->bh);
2126                 partial--;
2127         }
2128 no_top:
2129         return partial;
2130 }
2131
2132 /*
2133  * Zero a number of block pointers in either an inode or an indirect block.
2134  * If we restart the transaction we must again get write access to the
2135  * indirect block for further modification.
2136  *
2137  * We release `count' blocks on disk, but (last - first) may be greater
2138  * than `count' because there can be holes in there.
2139  */
2140 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2141                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2142                 unsigned long count, __le32 *first, __le32 *last)
2143 {
2144         __le32 *p;
2145         if (try_to_extend_transaction(handle, inode)) {
2146                 if (bh) {
2147                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2148                         if (ext3_journal_dirty_metadata(handle, bh))
2149                                 return;
2150                 }
2151                 ext3_mark_inode_dirty(handle, inode);
2152                 truncate_restart_transaction(handle, inode);
2153                 if (bh) {
2154                         BUFFER_TRACE(bh, "retaking write access");
2155                         if (ext3_journal_get_write_access(handle, bh))
2156                                 return;
2157                 }
2158         }
2159
2160         /*
2161          * Any buffers which are on the journal will be in memory. We find
2162          * them on the hash table so journal_revoke() will run journal_forget()
2163          * on them.  We've already detached each block from the file, so
2164          * bforget() in journal_forget() should be safe.
2165          *
2166          * AKPM: turn on bforget in journal_forget()!!!
2167          */
2168         for (p = first; p < last; p++) {
2169                 u32 nr = le32_to_cpu(*p);
2170                 if (nr) {
2171                         struct buffer_head *bh;
2172
2173                         *p = 0;
2174                         bh = sb_find_get_block(inode->i_sb, nr);
2175                         ext3_forget(handle, 0, inode, bh, nr);
2176                 }
2177         }
2178
2179         ext3_free_blocks(handle, inode, block_to_free, count);
2180 }
2181
2182 /**
2183  * ext3_free_data - free a list of data blocks
2184  * @handle:     handle for this transaction
2185  * @inode:      inode we are dealing with
2186  * @this_bh:    indirect buffer_head which contains *@first and *@last
2187  * @first:      array of block numbers
2188  * @last:       points immediately past the end of array
2189  *
2190  * We are freeing all blocks refered from that array (numbers are stored as
2191  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2192  *
2193  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2194  * blocks are contiguous then releasing them at one time will only affect one
2195  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2196  * actually use a lot of journal space.
2197  *
2198  * @this_bh will be %NULL if @first and @last point into the inode's direct
2199  * block pointers.
2200  */
2201 static void ext3_free_data(handle_t *handle, struct inode *inode,
2202                            struct buffer_head *this_bh,
2203                            __le32 *first, __le32 *last)
2204 {
2205         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2206         unsigned long count = 0;            /* Number of blocks in the run */
2207         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2208                                                corresponding to
2209                                                block_to_free */
2210         ext3_fsblk_t nr;                    /* Current block # */
2211         __le32 *p;                          /* Pointer into inode/ind
2212                                                for current block */
2213         int err;
2214
2215         if (this_bh) {                          /* For indirect block */
2216                 BUFFER_TRACE(this_bh, "get_write_access");
2217                 err = ext3_journal_get_write_access(handle, this_bh);
2218                 /* Important: if we can't update the indirect pointers
2219                  * to the blocks, we can't free them. */
2220                 if (err)
2221                         return;
2222         }
2223
2224         for (p = first; p < last; p++) {
2225                 nr = le32_to_cpu(*p);
2226                 if (nr) {
2227                         /* accumulate blocks to free if they're contiguous */
2228                         if (count == 0) {
2229                                 block_to_free = nr;
2230                                 block_to_free_p = p;
2231                                 count = 1;
2232                         } else if (nr == block_to_free + count) {
2233                                 count++;
2234                         } else {
2235                                 ext3_clear_blocks(handle, inode, this_bh,
2236                                                   block_to_free,
2237                                                   count, block_to_free_p, p);
2238                                 block_to_free = nr;
2239                                 block_to_free_p = p;
2240                                 count = 1;
2241                         }
2242                 }
2243         }
2244
2245         if (count > 0)
2246                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2247                                   count, block_to_free_p, p);
2248
2249         if (this_bh) {
2250                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2251
2252                 /*
2253                  * The buffer head should have an attached journal head at this
2254                  * point. However, if the data is corrupted and an indirect
2255                  * block pointed to itself, it would have been detached when
2256                  * the block was cleared. Check for this instead of OOPSing.
2257                  */
2258                 if (bh2jh(this_bh))
2259                         ext3_journal_dirty_metadata(handle, this_bh);
2260                 else
2261                         ext3_error(inode->i_sb, "ext3_free_data",
2262                                    "circular indirect block detected, "
2263                                    "inode=%lu, block=%llu",
2264                                    inode->i_ino,
2265                                    (unsigned long long)this_bh->b_blocknr);
2266         }
2267 }
2268
2269 /**
2270  *      ext3_free_branches - free an array of branches
2271  *      @handle: JBD handle for this transaction
2272  *      @inode: inode we are dealing with
2273  *      @parent_bh: the buffer_head which contains *@first and *@last
2274  *      @first: array of block numbers
2275  *      @last:  pointer immediately past the end of array
2276  *      @depth: depth of the branches to free
2277  *
2278  *      We are freeing all blocks refered from these branches (numbers are
2279  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2280  *      appropriately.
2281  */
2282 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2283                                struct buffer_head *parent_bh,
2284                                __le32 *first, __le32 *last, int depth)
2285 {
2286         ext3_fsblk_t nr;
2287         __le32 *p;
2288
2289         if (is_handle_aborted(handle))
2290                 return;
2291
2292         if (depth--) {
2293                 struct buffer_head *bh;
2294                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2295                 p = last;
2296                 while (--p >= first) {
2297                         nr = le32_to_cpu(*p);
2298                         if (!nr)
2299                                 continue;               /* A hole */
2300
2301                         /* Go read the buffer for the next level down */
2302                         bh = sb_bread(inode->i_sb, nr);
2303
2304                         /*
2305                          * A read failure? Report error and clear slot
2306                          * (should be rare).
2307                          */
2308                         if (!bh) {
2309                                 ext3_error(inode->i_sb, "ext3_free_branches",
2310                                            "Read failure, inode=%lu, block="E3FSBLK,
2311                                            inode->i_ino, nr);
2312                                 continue;
2313                         }
2314
2315                         /* This zaps the entire block.  Bottom up. */
2316                         BUFFER_TRACE(bh, "free child branches");
2317                         ext3_free_branches(handle, inode, bh,
2318                                            (__le32*)bh->b_data,
2319                                            (__le32*)bh->b_data + addr_per_block,
2320                                            depth);
2321
2322                         /*
2323                          * Everything below this this pointer has been
2324                          * released.  Now let this top-of-subtree go.
2325                          *
2326                          * We want the freeing of this indirect block to be
2327                          * atomic in the journal with the updating of the
2328                          * bitmap block which owns it.  So make some room in
2329                          * the journal.
2330                          *
2331                          * We zero the parent pointer *after* freeing its
2332                          * pointee in the bitmaps, so if extend_transaction()
2333                          * for some reason fails to put the bitmap changes and
2334                          * the release into the same transaction, recovery
2335                          * will merely complain about releasing a free block,
2336                          * rather than leaking blocks.
2337                          */
2338                         if (is_handle_aborted(handle))
2339                                 return;
2340                         if (try_to_extend_transaction(handle, inode)) {
2341                                 ext3_mark_inode_dirty(handle, inode);
2342                                 truncate_restart_transaction(handle, inode);
2343                         }
2344
2345                         /*
2346                          * We've probably journalled the indirect block several
2347                          * times during the truncate.  But it's no longer
2348                          * needed and we now drop it from the transaction via
2349                          * journal_revoke().
2350                          *
2351                          * That's easy if it's exclusively part of this
2352                          * transaction.  But if it's part of the committing
2353                          * transaction then journal_forget() will simply
2354                          * brelse() it.  That means that if the underlying
2355                          * block is reallocated in ext3_get_block(),
2356                          * unmap_underlying_metadata() will find this block
2357                          * and will try to get rid of it.  damn, damn. Thus
2358                          * we don't allow a block to be reallocated until
2359                          * a transaction freeing it has fully committed.
2360                          *
2361                          * We also have to make sure journal replay after a
2362                          * crash does not overwrite non-journaled data blocks
2363                          * with old metadata when the block got reallocated for
2364                          * data.  Thus we have to store a revoke record for a
2365                          * block in the same transaction in which we free the
2366                          * block.
2367                          */
2368                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2369
2370                         ext3_free_blocks(handle, inode, nr, 1);
2371
2372                         if (parent_bh) {
2373                                 /*
2374                                  * The block which we have just freed is
2375                                  * pointed to by an indirect block: journal it
2376                                  */
2377                                 BUFFER_TRACE(parent_bh, "get_write_access");
2378                                 if (!ext3_journal_get_write_access(handle,
2379                                                                    parent_bh)){
2380                                         *p = 0;
2381                                         BUFFER_TRACE(parent_bh,
2382                                         "call ext3_journal_dirty_metadata");
2383                                         ext3_journal_dirty_metadata(handle,
2384                                                                     parent_bh);
2385                                 }
2386                         }
2387                 }
2388         } else {
2389                 /* We have reached the bottom of the tree. */
2390                 BUFFER_TRACE(parent_bh, "free data blocks");
2391                 ext3_free_data(handle, inode, parent_bh, first, last);
2392         }
2393 }
2394
2395 int ext3_can_truncate(struct inode *inode)
2396 {
2397         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2398                 return 0;
2399         if (S_ISREG(inode->i_mode))
2400                 return 1;
2401         if (S_ISDIR(inode->i_mode))
2402                 return 1;
2403         if (S_ISLNK(inode->i_mode))
2404                 return !ext3_inode_is_fast_symlink(inode);
2405         return 0;
2406 }
2407
2408 /*
2409  * ext3_truncate()
2410  *
2411  * We block out ext3_get_block() block instantiations across the entire
2412  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2413  * simultaneously on behalf of the same inode.
2414  *
2415  * As we work through the truncate and commmit bits of it to the journal there
2416  * is one core, guiding principle: the file's tree must always be consistent on
2417  * disk.  We must be able to restart the truncate after a crash.
2418  *
2419  * The file's tree may be transiently inconsistent in memory (although it
2420  * probably isn't), but whenever we close off and commit a journal transaction,
2421  * the contents of (the filesystem + the journal) must be consistent and
2422  * restartable.  It's pretty simple, really: bottom up, right to left (although
2423  * left-to-right works OK too).
2424  *
2425  * Note that at recovery time, journal replay occurs *before* the restart of
2426  * truncate against the orphan inode list.
2427  *
2428  * The committed inode has the new, desired i_size (which is the same as
2429  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2430  * that this inode's truncate did not complete and it will again call
2431  * ext3_truncate() to have another go.  So there will be instantiated blocks
2432  * to the right of the truncation point in a crashed ext3 filesystem.  But
2433  * that's fine - as long as they are linked from the inode, the post-crash
2434  * ext3_truncate() run will find them and release them.
2435  */
2436 void ext3_truncate(struct inode *inode)
2437 {
2438         handle_t *handle;
2439         struct ext3_inode_info *ei = EXT3_I(inode);
2440         __le32 *i_data = ei->i_data;
2441         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2442         struct address_space *mapping = inode->i_mapping;
2443         int offsets[4];
2444         Indirect chain[4];
2445         Indirect *partial;
2446         __le32 nr = 0;
2447         int n;
2448         long last_block;
2449         unsigned blocksize = inode->i_sb->s_blocksize;
2450         struct page *page;
2451
2452         if (!ext3_can_truncate(inode))
2453                 goto out_notrans;
2454
2455         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2456                 ext3_set_inode_state(inode, EXT3_STATE_FLUSH_ON_CLOSE);
2457
2458         /*
2459          * We have to lock the EOF page here, because lock_page() nests
2460          * outside journal_start().
2461          */
2462         if ((inode->i_size & (blocksize - 1)) == 0) {
2463                 /* Block boundary? Nothing to do */
2464                 page = NULL;
2465         } else {
2466                 page = grab_cache_page(mapping,
2467                                 inode->i_size >> PAGE_CACHE_SHIFT);
2468                 if (!page)
2469                         goto out_notrans;
2470         }
2471
2472         handle = start_transaction(inode);
2473         if (IS_ERR(handle)) {
2474                 if (page) {
2475                         clear_highpage(page);
2476                         flush_dcache_page(page);
2477                         unlock_page(page);
2478                         page_cache_release(page);
2479                 }
2480                 goto out_notrans;
2481         }
2482
2483         last_block = (inode->i_size + blocksize-1)
2484                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2485
2486         if (page)
2487                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2488
2489         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2490         if (n == 0)
2491                 goto out_stop;  /* error */
2492
2493         /*
2494          * OK.  This truncate is going to happen.  We add the inode to the
2495          * orphan list, so that if this truncate spans multiple transactions,
2496          * and we crash, we will resume the truncate when the filesystem
2497          * recovers.  It also marks the inode dirty, to catch the new size.
2498          *
2499          * Implication: the file must always be in a sane, consistent
2500          * truncatable state while each transaction commits.
2501          */
2502         if (ext3_orphan_add(handle, inode))
2503                 goto out_stop;
2504
2505         /*
2506          * The orphan list entry will now protect us from any crash which
2507          * occurs before the truncate completes, so it is now safe to propagate
2508          * the new, shorter inode size (held for now in i_size) into the
2509          * on-disk inode. We do this via i_disksize, which is the value which
2510          * ext3 *really* writes onto the disk inode.
2511          */
2512         ei->i_disksize = inode->i_size;
2513
2514         /*
2515          * From here we block out all ext3_get_block() callers who want to
2516          * modify the block allocation tree.
2517          */
2518         mutex_lock(&ei->truncate_mutex);
2519
2520         if (n == 1) {           /* direct blocks */
2521                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2522                                i_data + EXT3_NDIR_BLOCKS);
2523                 goto do_indirects;
2524         }
2525
2526         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2527         /* Kill the top of shared branch (not detached) */
2528         if (nr) {
2529                 if (partial == chain) {
2530                         /* Shared branch grows from the inode */
2531                         ext3_free_branches(handle, inode, NULL,
2532                                            &nr, &nr+1, (chain+n-1) - partial);
2533                         *partial->p = 0;
2534                         /*
2535                          * We mark the inode dirty prior to restart,
2536                          * and prior to stop.  No need for it here.
2537                          */
2538                 } else {
2539                         /* Shared branch grows from an indirect block */
2540                         ext3_free_branches(handle, inode, partial->bh,
2541                                         partial->p,
2542                                         partial->p+1, (chain+n-1) - partial);
2543                 }
2544         }
2545         /* Clear the ends of indirect blocks on the shared branch */
2546         while (partial > chain) {
2547                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2548                                    (__le32*)partial->bh->b_data+addr_per_block,
2549                                    (chain+n-1) - partial);
2550                 BUFFER_TRACE(partial->bh, "call brelse");
2551                 brelse (partial->bh);
2552                 partial--;
2553         }
2554 do_indirects:
2555         /* Kill the remaining (whole) subtrees */
2556         switch (offsets[0]) {
2557         default:
2558                 nr = i_data[EXT3_IND_BLOCK];
2559                 if (nr) {
2560                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2561                         i_data[EXT3_IND_BLOCK] = 0;
2562                 }
2563         case EXT3_IND_BLOCK:
2564                 nr = i_data[EXT3_DIND_BLOCK];
2565                 if (nr) {
2566                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2567                         i_data[EXT3_DIND_BLOCK] = 0;
2568                 }
2569         case EXT3_DIND_BLOCK:
2570                 nr = i_data[EXT3_TIND_BLOCK];
2571                 if (nr) {
2572                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2573                         i_data[EXT3_TIND_BLOCK] = 0;
2574                 }
2575         case EXT3_TIND_BLOCK:
2576                 ;
2577         }
2578
2579         ext3_discard_reservation(inode);
2580
2581         mutex_unlock(&ei->truncate_mutex);
2582         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2583         ext3_mark_inode_dirty(handle, inode);
2584
2585         /*
2586          * In a multi-transaction truncate, we only make the final transaction
2587          * synchronous
2588          */
2589         if (IS_SYNC(inode))
2590                 handle->h_sync = 1;
2591 out_stop:
2592         /*
2593          * If this was a simple ftruncate(), and the file will remain alive
2594          * then we need to clear up the orphan record which we created above.
2595          * However, if this was a real unlink then we were called by
2596          * ext3_evict_inode(), and we allow that function to clean up the
2597          * orphan info for us.
2598          */
2599         if (inode->i_nlink)
2600                 ext3_orphan_del(handle, inode);
2601
2602         ext3_journal_stop(handle);
2603         return;
2604 out_notrans:
2605         /*
2606          * Delete the inode from orphan list so that it doesn't stay there
2607          * forever and trigger assertion on umount.
2608          */
2609         if (inode->i_nlink)
2610                 ext3_orphan_del(NULL, inode);
2611 }
2612
2613 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2614                 unsigned long ino, struct ext3_iloc *iloc)
2615 {
2616         unsigned long block_group;
2617         unsigned long offset;
2618         ext3_fsblk_t block;
2619         struct ext3_group_desc *gdp;
2620
2621         if (!ext3_valid_inum(sb, ino)) {
2622                 /*
2623                  * This error is already checked for in namei.c unless we are
2624                  * looking at an NFS filehandle, in which case no error
2625                  * report is needed
2626                  */
2627                 return 0;
2628         }
2629
2630         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2631         gdp = ext3_get_group_desc(sb, block_group, NULL);
2632         if (!gdp)
2633                 return 0;
2634         /*
2635          * Figure out the offset within the block group inode table
2636          */
2637         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2638                 EXT3_INODE_SIZE(sb);
2639         block = le32_to_cpu(gdp->bg_inode_table) +
2640                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2641
2642         iloc->block_group = block_group;
2643         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2644         return block;
2645 }
2646
2647 /*
2648  * ext3_get_inode_loc returns with an extra refcount against the inode's
2649  * underlying buffer_head on success. If 'in_mem' is true, we have all
2650  * data in memory that is needed to recreate the on-disk version of this
2651  * inode.
2652  */
2653 static int __ext3_get_inode_loc(struct inode *inode,
2654                                 struct ext3_iloc *iloc, int in_mem)
2655 {
2656         ext3_fsblk_t block;
2657         struct buffer_head *bh;
2658
2659         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2660         if (!block)
2661                 return -EIO;
2662
2663         bh = sb_getblk(inode->i_sb, block);
2664         if (!bh) {
2665                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2666                                 "unable to read inode block - "
2667                                 "inode=%lu, block="E3FSBLK,
2668                                  inode->i_ino, block);
2669                 return -EIO;
2670         }
2671         if (!buffer_uptodate(bh)) {
2672                 lock_buffer(bh);
2673
2674                 /*
2675                  * If the buffer has the write error flag, we have failed
2676                  * to write out another inode in the same block.  In this
2677                  * case, we don't have to read the block because we may
2678                  * read the old inode data successfully.
2679                  */
2680                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2681                         set_buffer_uptodate(bh);
2682
2683                 if (buffer_uptodate(bh)) {
2684                         /* someone brought it uptodate while we waited */
2685                         unlock_buffer(bh);
2686                         goto has_buffer;
2687                 }
2688
2689                 /*
2690                  * If we have all information of the inode in memory and this
2691                  * is the only valid inode in the block, we need not read the
2692                  * block.
2693                  */
2694                 if (in_mem) {
2695                         struct buffer_head *bitmap_bh;
2696                         struct ext3_group_desc *desc;
2697                         int inodes_per_buffer;
2698                         int inode_offset, i;
2699                         int block_group;
2700                         int start;
2701
2702                         block_group = (inode->i_ino - 1) /
2703                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2704                         inodes_per_buffer = bh->b_size /
2705                                 EXT3_INODE_SIZE(inode->i_sb);
2706                         inode_offset = ((inode->i_ino - 1) %
2707                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2708                         start = inode_offset & ~(inodes_per_buffer - 1);
2709
2710                         /* Is the inode bitmap in cache? */
2711                         desc = ext3_get_group_desc(inode->i_sb,
2712                                                 block_group, NULL);
2713                         if (!desc)
2714                                 goto make_io;
2715
2716                         bitmap_bh = sb_getblk(inode->i_sb,
2717                                         le32_to_cpu(desc->bg_inode_bitmap));
2718                         if (!bitmap_bh)
2719                                 goto make_io;
2720
2721                         /*
2722                          * If the inode bitmap isn't in cache then the
2723                          * optimisation may end up performing two reads instead
2724                          * of one, so skip it.
2725                          */
2726                         if (!buffer_uptodate(bitmap_bh)) {
2727                                 brelse(bitmap_bh);
2728                                 goto make_io;
2729                         }
2730                         for (i = start; i < start + inodes_per_buffer; i++) {
2731                                 if (i == inode_offset)
2732                                         continue;
2733                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2734                                         break;
2735                         }
2736                         brelse(bitmap_bh);
2737                         if (i == start + inodes_per_buffer) {
2738                                 /* all other inodes are free, so skip I/O */
2739                                 memset(bh->b_data, 0, bh->b_size);
2740                                 set_buffer_uptodate(bh);
2741                                 unlock_buffer(bh);
2742                                 goto has_buffer;
2743                         }
2744                 }
2745
2746 make_io:
2747                 /*
2748                  * There are other valid inodes in the buffer, this inode
2749                  * has in-inode xattrs, or we don't have this inode in memory.
2750                  * Read the block from disk.
2751                  */
2752                 get_bh(bh);
2753                 bh->b_end_io = end_buffer_read_sync;
2754                 submit_bh(READ_META, bh);
2755                 wait_on_buffer(bh);
2756                 if (!buffer_uptodate(bh)) {
2757                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2758                                         "unable to read inode block - "
2759                                         "inode=%lu, block="E3FSBLK,
2760                                         inode->i_ino, block);
2761                         brelse(bh);
2762                         return -EIO;
2763                 }
2764         }
2765 has_buffer:
2766         iloc->bh = bh;
2767         return 0;
2768 }
2769
2770 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2771 {
2772         /* We have all inode data except xattrs in memory here. */
2773         return __ext3_get_inode_loc(inode, iloc,
2774                 !ext3_test_inode_state(inode, EXT3_STATE_XATTR));
2775 }
2776
2777 void ext3_set_inode_flags(struct inode *inode)
2778 {
2779         unsigned int flags = EXT3_I(inode)->i_flags;
2780
2781         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2782         if (flags & EXT3_SYNC_FL)
2783                 inode->i_flags |= S_SYNC;
2784         if (flags & EXT3_APPEND_FL)
2785                 inode->i_flags |= S_APPEND;
2786         if (flags & EXT3_IMMUTABLE_FL)
2787                 inode->i_flags |= S_IMMUTABLE;
2788         if (flags & EXT3_NOATIME_FL)
2789                 inode->i_flags |= S_NOATIME;
2790         if (flags & EXT3_DIRSYNC_FL)
2791                 inode->i_flags |= S_DIRSYNC;
2792 }
2793
2794 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2795 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2796 {
2797         unsigned int flags = ei->vfs_inode.i_flags;
2798
2799         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2800                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2801         if (flags & S_SYNC)
2802                 ei->i_flags |= EXT3_SYNC_FL;
2803         if (flags & S_APPEND)
2804                 ei->i_flags |= EXT3_APPEND_FL;
2805         if (flags & S_IMMUTABLE)
2806                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2807         if (flags & S_NOATIME)
2808                 ei->i_flags |= EXT3_NOATIME_FL;
2809         if (flags & S_DIRSYNC)
2810                 ei->i_flags |= EXT3_DIRSYNC_FL;
2811 }
2812
2813 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2814 {
2815         struct ext3_iloc iloc;
2816         struct ext3_inode *raw_inode;
2817         struct ext3_inode_info *ei;
2818         struct buffer_head *bh;
2819         struct inode *inode;
2820         journal_t *journal = EXT3_SB(sb)->s_journal;
2821         transaction_t *transaction;
2822         long ret;
2823         int block;
2824
2825         inode = iget_locked(sb, ino);
2826         if (!inode)
2827                 return ERR_PTR(-ENOMEM);
2828         if (!(inode->i_state & I_NEW))
2829                 return inode;
2830
2831         ei = EXT3_I(inode);
2832         ei->i_block_alloc_info = NULL;
2833
2834         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2835         if (ret < 0)
2836                 goto bad_inode;
2837         bh = iloc.bh;
2838         raw_inode = ext3_raw_inode(&iloc);
2839         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2840         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2841         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2842         if(!(test_opt (inode->i_sb, NO_UID32))) {
2843                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2844                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2845         }
2846         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2847         inode->i_size = le32_to_cpu(raw_inode->i_size);
2848         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2849         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2850         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2851         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2852
2853         ei->i_state_flags = 0;
2854         ei->i_dir_start_lookup = 0;
2855         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2856         /* We now have enough fields to check if the inode was active or not.
2857          * This is needed because nfsd might try to access dead inodes
2858          * the test is that same one that e2fsck uses
2859          * NeilBrown 1999oct15
2860          */
2861         if (inode->i_nlink == 0) {
2862                 if (inode->i_mode == 0 ||
2863                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2864                         /* this inode is deleted */
2865                         brelse (bh);
2866                         ret = -ESTALE;
2867                         goto bad_inode;
2868                 }
2869                 /* The only unlinked inodes we let through here have
2870                  * valid i_mode and are being read by the orphan
2871                  * recovery code: that's fine, we're about to complete
2872                  * the process of deleting those. */
2873         }
2874         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2875         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2876 #ifdef EXT3_FRAGMENTS
2877         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2878         ei->i_frag_no = raw_inode->i_frag;
2879         ei->i_frag_size = raw_inode->i_fsize;
2880 #endif
2881         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2882         if (!S_ISREG(inode->i_mode)) {
2883                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2884         } else {
2885                 inode->i_size |=
2886                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2887         }
2888         ei->i_disksize = inode->i_size;
2889         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2890         ei->i_block_group = iloc.block_group;
2891         /*
2892          * NOTE! The in-memory inode i_data array is in little-endian order
2893          * even on big-endian machines: we do NOT byteswap the block numbers!
2894          */
2895         for (block = 0; block < EXT3_N_BLOCKS; block++)
2896                 ei->i_data[block] = raw_inode->i_block[block];
2897         INIT_LIST_HEAD(&ei->i_orphan);
2898
2899         /*
2900          * Set transaction id's of transactions that have to be committed
2901          * to finish f[data]sync. We set them to currently running transaction
2902          * as we cannot be sure that the inode or some of its metadata isn't
2903          * part of the transaction - the inode could have been reclaimed and
2904          * now it is reread from disk.
2905          */
2906         if (journal) {
2907                 tid_t tid;
2908
2909                 spin_lock(&journal->j_state_lock);
2910                 if (journal->j_running_transaction)
2911                         transaction = journal->j_running_transaction;
2912                 else
2913                         transaction = journal->j_committing_transaction;
2914                 if (transaction)
2915                         tid = transaction->t_tid;
2916                 else
2917                         tid = journal->j_commit_sequence;
2918                 spin_unlock(&journal->j_state_lock);
2919                 atomic_set(&ei->i_sync_tid, tid);
2920                 atomic_set(&ei->i_datasync_tid, tid);
2921         }
2922
2923         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2924             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2925                 /*
2926                  * When mke2fs creates big inodes it does not zero out
2927                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2928                  * so ignore those first few inodes.
2929                  */
2930                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2931                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2932                     EXT3_INODE_SIZE(inode->i_sb)) {
2933                         brelse (bh);
2934                         ret = -EIO;
2935                         goto bad_inode;
2936                 }
2937                 if (ei->i_extra_isize == 0) {
2938                         /* The extra space is currently unused. Use it. */
2939                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2940                                             EXT3_GOOD_OLD_INODE_SIZE;
2941                 } else {
2942                         __le32 *magic = (void *)raw_inode +
2943                                         EXT3_GOOD_OLD_INODE_SIZE +
2944                                         ei->i_extra_isize;
2945                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2946                                  ext3_set_inode_state(inode, EXT3_STATE_XATTR);
2947                 }
2948         } else
2949                 ei->i_extra_isize = 0;
2950
2951         if (S_ISREG(inode->i_mode)) {
2952                 inode->i_op = &ext3_file_inode_operations;
2953                 inode->i_fop = &ext3_file_operations;
2954                 ext3_set_aops(inode);
2955         } else if (S_ISDIR(inode->i_mode)) {
2956                 inode->i_op = &ext3_dir_inode_operations;
2957                 inode->i_fop = &ext3_dir_operations;
2958         } else if (S_ISLNK(inode->i_mode)) {
2959                 if (ext3_inode_is_fast_symlink(inode)) {
2960                         inode->i_op = &ext3_fast_symlink_inode_operations;
2961                         nd_terminate_link(ei->i_data, inode->i_size,
2962                                 sizeof(ei->i_data) - 1);
2963                 } else {
2964                         inode->i_op = &ext3_symlink_inode_operations;
2965                         ext3_set_aops(inode);
2966                 }
2967         } else {
2968                 inode->i_op = &ext3_special_inode_operations;
2969                 if (raw_inode->i_block[0])
2970                         init_special_inode(inode, inode->i_mode,
2971                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2972                 else
2973                         init_special_inode(inode, inode->i_mode,
2974                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2975         }
2976         brelse (iloc.bh);
2977         ext3_set_inode_flags(inode);
2978         unlock_new_inode(inode);
2979         return inode;
2980
2981 bad_inode:
2982         iget_failed(inode);
2983         return ERR_PTR(ret);
2984 }
2985
2986 /*
2987  * Post the struct inode info into an on-disk inode location in the
2988  * buffer-cache.  This gobbles the caller's reference to the
2989  * buffer_head in the inode location struct.
2990  *
2991  * The caller must have write access to iloc->bh.
2992  */
2993 static int ext3_do_update_inode(handle_t *handle,
2994                                 struct inode *inode,
2995                                 struct ext3_iloc *iloc)
2996 {
2997         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2998         struct ext3_inode_info *ei = EXT3_I(inode);
2999         struct buffer_head *bh = iloc->bh;
3000         int err = 0, rc, block;
3001
3002 again:
3003         /* we can't allow multiple procs in here at once, its a bit racey */
3004         lock_buffer(bh);
3005
3006         /* For fields not not tracking in the in-memory inode,
3007          * initialise them to zero for new inodes. */
3008         if (ext3_test_inode_state(inode, EXT3_STATE_NEW))
3009                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
3010
3011         ext3_get_inode_flags(ei);
3012         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3013         if(!(test_opt(inode->i_sb, NO_UID32))) {
3014                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3015                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3016 /*
3017  * Fix up interoperability with old kernels. Otherwise, old inodes get
3018  * re-used with the upper 16 bits of the uid/gid intact
3019  */
3020                 if(!ei->i_dtime) {
3021                         raw_inode->i_uid_high =
3022                                 cpu_to_le16(high_16_bits(inode->i_uid));
3023                         raw_inode->i_gid_high =
3024                                 cpu_to_le16(high_16_bits(inode->i_gid));
3025                 } else {
3026                         raw_inode->i_uid_high = 0;
3027                         raw_inode->i_gid_high = 0;
3028                 }
3029         } else {
3030                 raw_inode->i_uid_low =
3031                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3032                 raw_inode->i_gid_low =
3033                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3034                 raw_inode->i_uid_high = 0;
3035                 raw_inode->i_gid_high = 0;
3036         }
3037         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3038         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
3039         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
3040         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
3041         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
3042         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
3043         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3044         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
3045 #ifdef EXT3_FRAGMENTS
3046         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
3047         raw_inode->i_frag = ei->i_frag_no;
3048         raw_inode->i_fsize = ei->i_frag_size;
3049 #endif
3050         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
3051         if (!S_ISREG(inode->i_mode)) {
3052                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
3053         } else {
3054                 raw_inode->i_size_high =
3055                         cpu_to_le32(ei->i_disksize >> 32);
3056                 if (ei->i_disksize > 0x7fffffffULL) {
3057                         struct super_block *sb = inode->i_sb;
3058                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
3059                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
3060                             EXT3_SB(sb)->s_es->s_rev_level ==
3061                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
3062                                /* If this is the first large file
3063                                 * created, add a flag to the superblock.
3064                                 */
3065                                 unlock_buffer(bh);
3066                                 err = ext3_journal_get_write_access(handle,
3067                                                 EXT3_SB(sb)->s_sbh);
3068                                 if (err)
3069                                         goto out_brelse;
3070
3071                                 ext3_update_dynamic_rev(sb);
3072                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
3073                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
3074                                 handle->h_sync = 1;
3075                                 err = ext3_journal_dirty_metadata(handle,
3076                                                 EXT3_SB(sb)->s_sbh);
3077                                 /* get our lock and start over */
3078                                 goto again;
3079                         }
3080                 }
3081         }
3082         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3083         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3084                 if (old_valid_dev(inode->i_rdev)) {
3085                         raw_inode->i_block[0] =
3086                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3087                         raw_inode->i_block[1] = 0;
3088                 } else {
3089                         raw_inode->i_block[0] = 0;
3090                         raw_inode->i_block[1] =
3091                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3092                         raw_inode->i_block[2] = 0;
3093                 }
3094         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
3095                 raw_inode->i_block[block] = ei->i_data[block];
3096
3097         if (ei->i_extra_isize)
3098                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3099
3100         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
3101         unlock_buffer(bh);
3102         rc = ext3_journal_dirty_metadata(handle, bh);
3103         if (!err)
3104                 err = rc;
3105         ext3_clear_inode_state(inode, EXT3_STATE_NEW);
3106
3107         atomic_set(&ei->i_sync_tid, handle->h_transaction->t_tid);
3108 out_brelse:
3109         brelse (bh);
3110         ext3_std_error(inode->i_sb, err);
3111         return err;
3112 }
3113
3114 /*
3115  * ext3_write_inode()
3116  *
3117  * We are called from a few places:
3118  *
3119  * - Within generic_file_write() for O_SYNC files.
3120  *   Here, there will be no transaction running. We wait for any running
3121  *   trasnaction to commit.
3122  *
3123  * - Within sys_sync(), kupdate and such.
3124  *   We wait on commit, if tol to.
3125  *
3126  * - Within prune_icache() (PF_MEMALLOC == true)
3127  *   Here we simply return.  We can't afford to block kswapd on the
3128  *   journal commit.
3129  *
3130  * In all cases it is actually safe for us to return without doing anything,
3131  * because the inode has been copied into a raw inode buffer in
3132  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3133  * knfsd.
3134  *
3135  * Note that we are absolutely dependent upon all inode dirtiers doing the
3136  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3137  * which we are interested.
3138  *
3139  * It would be a bug for them to not do this.  The code:
3140  *
3141  *      mark_inode_dirty(inode)
3142  *      stuff();
3143  *      inode->i_size = expr;
3144  *
3145  * is in error because a kswapd-driven write_inode() could occur while
3146  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3147  * will no longer be on the superblock's dirty inode list.
3148  */
3149 int ext3_write_inode(struct inode *inode, struct writeback_control *wbc)
3150 {
3151         if (current->flags & PF_MEMALLOC)
3152                 return 0;
3153
3154         if (ext3_journal_current_handle()) {
3155                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3156                 dump_stack();
3157                 return -EIO;
3158         }
3159
3160         if (wbc->sync_mode != WB_SYNC_ALL)
3161                 return 0;
3162
3163         return ext3_force_commit(inode->i_sb);
3164 }
3165
3166 /*
3167  * ext3_setattr()
3168  *
3169  * Called from notify_change.
3170  *
3171  * We want to trap VFS attempts to truncate the file as soon as
3172  * possible.  In particular, we want to make sure that when the VFS
3173  * shrinks i_size, we put the inode on the orphan list and modify
3174  * i_disksize immediately, so that during the subsequent flushing of
3175  * dirty pages and freeing of disk blocks, we can guarantee that any
3176  * commit will leave the blocks being flushed in an unused state on
3177  * disk.  (On recovery, the inode will get truncated and the blocks will
3178  * be freed, so we have a strong guarantee that no future commit will
3179  * leave these blocks visible to the user.)
3180  *
3181  * Called with inode->sem down.
3182  */
3183 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3184 {
3185         struct inode *inode = dentry->d_inode;
3186         int error, rc = 0;
3187         const unsigned int ia_valid = attr->ia_valid;
3188
3189         error = inode_change_ok(inode, attr);
3190         if (error)
3191                 return error;
3192
3193         if (is_quota_modification(inode, attr))
3194                 dquot_initialize(inode);
3195         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3196                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3197                 handle_t *handle;
3198
3199                 /* (user+group)*(old+new) structure, inode write (sb,
3200                  * inode block, ? - but truncate inode update has it) */
3201                 handle = ext3_journal_start(inode, EXT3_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3202                                         EXT3_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)+3);
3203                 if (IS_ERR(handle)) {
3204                         error = PTR_ERR(handle);
3205                         goto err_out;
3206                 }
3207                 error = dquot_transfer(inode, attr);
3208                 if (error) {
3209                         ext3_journal_stop(handle);
3210                         return error;
3211                 }
3212                 /* Update corresponding info in inode so that everything is in
3213                  * one transaction */
3214                 if (attr->ia_valid & ATTR_UID)
3215                         inode->i_uid = attr->ia_uid;
3216                 if (attr->ia_valid & ATTR_GID)
3217                         inode->i_gid = attr->ia_gid;
3218                 error = ext3_mark_inode_dirty(handle, inode);
3219                 ext3_journal_stop(handle);
3220         }
3221
3222         if (S_ISREG(inode->i_mode) &&
3223             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3224                 handle_t *handle;
3225
3226                 handle = ext3_journal_start(inode, 3);
3227                 if (IS_ERR(handle)) {
3228                         error = PTR_ERR(handle);
3229                         goto err_out;
3230                 }
3231
3232                 error = ext3_orphan_add(handle, inode);
3233                 EXT3_I(inode)->i_disksize = attr->ia_size;
3234                 rc = ext3_mark_inode_dirty(handle, inode);
3235                 if (!error)
3236                         error = rc;
3237                 ext3_journal_stop(handle);
3238         }
3239
3240         if ((attr->ia_valid & ATTR_SIZE) &&
3241             attr->ia_size != i_size_read(inode)) {
3242                 rc = vmtruncate(inode, attr->ia_size);
3243                 if (rc)
3244                         goto err_out;
3245         }
3246
3247         setattr_copy(inode, attr);
3248         mark_inode_dirty(inode);
3249
3250         if (ia_valid & ATTR_MODE)
3251                 rc = ext3_acl_chmod(inode);
3252
3253 err_out:
3254         ext3_std_error(inode->i_sb, error);
3255         if (!error)
3256                 error = rc;
3257         return error;
3258 }
3259
3260
3261 /*
3262  * How many blocks doth make a writepage()?
3263  *
3264  * With N blocks per page, it may be:
3265  * N data blocks
3266  * 2 indirect block
3267  * 2 dindirect
3268  * 1 tindirect
3269  * N+5 bitmap blocks (from the above)
3270  * N+5 group descriptor summary blocks
3271  * 1 inode block
3272  * 1 superblock.
3273  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3274  *
3275  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3276  *
3277  * With ordered or writeback data it's the same, less the N data blocks.
3278  *
3279  * If the inode's direct blocks can hold an integral number of pages then a
3280  * page cannot straddle two indirect blocks, and we can only touch one indirect
3281  * and dindirect block, and the "5" above becomes "3".
3282  *
3283  * This still overestimates under most circumstances.  If we were to pass the
3284  * start and end offsets in here as well we could do block_to_path() on each
3285  * block and work out the exact number of indirects which are touched.  Pah.
3286  */
3287
3288 static int ext3_writepage_trans_blocks(struct inode *inode)
3289 {
3290         int bpp = ext3_journal_blocks_per_page(inode);
3291         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3292         int ret;
3293
3294         if (ext3_should_journal_data(inode))
3295                 ret = 3 * (bpp + indirects) + 2;
3296         else
3297                 ret = 2 * (bpp + indirects) + 2;
3298
3299 #ifdef CONFIG_QUOTA
3300         /* We know that structure was already allocated during dquot_initialize so
3301          * we will be updating only the data blocks + inodes */
3302         ret += EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
3303 #endif
3304
3305         return ret;
3306 }
3307
3308 /*
3309  * The caller must have previously called ext3_reserve_inode_write().
3310  * Give this, we know that the caller already has write access to iloc->bh.
3311  */
3312 int ext3_mark_iloc_dirty(handle_t *handle,
3313                 struct inode *inode, struct ext3_iloc *iloc)
3314 {
3315         int err = 0;
3316
3317         /* the do_update_inode consumes one bh->b_count */
3318         get_bh(iloc->bh);
3319
3320         /* ext3_do_update_inode() does journal_dirty_metadata */
3321         err = ext3_do_update_inode(handle, inode, iloc);
3322         put_bh(iloc->bh);
3323         return err;
3324 }
3325
3326 /*
3327  * On success, We end up with an outstanding reference count against
3328  * iloc->bh.  This _must_ be cleaned up later.
3329  */
3330
3331 int
3332 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3333                          struct ext3_iloc *iloc)
3334 {
3335         int err = 0;
3336         if (handle) {
3337                 err = ext3_get_inode_loc(inode, iloc);
3338                 if (!err) {
3339                         BUFFER_TRACE(iloc->bh, "get_write_access");
3340                         err = ext3_journal_get_write_access(handle, iloc->bh);
3341                         if (err) {
3342                                 brelse(iloc->bh);
3343                                 iloc->bh = NULL;
3344                         }
3345                 }
3346         }
3347         ext3_std_error(inode->i_sb, err);
3348         return err;
3349 }
3350
3351 /*
3352  * What we do here is to mark the in-core inode as clean with respect to inode
3353  * dirtiness (it may still be data-dirty).
3354  * This means that the in-core inode may be reaped by prune_icache
3355  * without having to perform any I/O.  This is a very good thing,
3356  * because *any* task may call prune_icache - even ones which
3357  * have a transaction open against a different journal.
3358  *
3359  * Is this cheating?  Not really.  Sure, we haven't written the
3360  * inode out, but prune_icache isn't a user-visible syncing function.
3361  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3362  * we start and wait on commits.
3363  *
3364  * Is this efficient/effective?  Well, we're being nice to the system
3365  * by cleaning up our inodes proactively so they can be reaped
3366  * without I/O.  But we are potentially leaving up to five seconds'
3367  * worth of inodes floating about which prune_icache wants us to
3368  * write out.  One way to fix that would be to get prune_icache()
3369  * to do a write_super() to free up some memory.  It has the desired
3370  * effect.
3371  */
3372 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3373 {
3374         struct ext3_iloc iloc;
3375         int err;
3376
3377         might_sleep();
3378         err = ext3_reserve_inode_write(handle, inode, &iloc);
3379         if (!err)
3380                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3381         return err;
3382 }
3383
3384 /*
3385  * ext3_dirty_inode() is called from __mark_inode_dirty()
3386  *
3387  * We're really interested in the case where a file is being extended.
3388  * i_size has been changed by generic_commit_write() and we thus need
3389  * to include the updated inode in the current transaction.
3390  *
3391  * Also, dquot_alloc_space() will always dirty the inode when blocks
3392  * are allocated to the file.
3393  *
3394  * If the inode is marked synchronous, we don't honour that here - doing
3395  * so would cause a commit on atime updates, which we don't bother doing.
3396  * We handle synchronous inodes at the highest possible level.
3397  */
3398 void ext3_dirty_inode(struct inode *inode)
3399 {
3400         handle_t *current_handle = ext3_journal_current_handle();
3401         handle_t *handle;
3402
3403         handle = ext3_journal_start(inode, 2);
3404         if (IS_ERR(handle))
3405                 goto out;
3406         if (current_handle &&
3407                 current_handle->h_transaction != handle->h_transaction) {
3408                 /* This task has a transaction open against a different fs */
3409                 printk(KERN_EMERG "%s: transactions do not match!\n",
3410                        __func__);
3411         } else {
3412                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3413                                 current_handle);
3414                 ext3_mark_inode_dirty(handle, inode);
3415         }
3416         ext3_journal_stop(handle);
3417 out:
3418         return;
3419 }
3420
3421 #if 0
3422 /*
3423  * Bind an inode's backing buffer_head into this transaction, to prevent
3424  * it from being flushed to disk early.  Unlike
3425  * ext3_reserve_inode_write, this leaves behind no bh reference and
3426  * returns no iloc structure, so the caller needs to repeat the iloc
3427  * lookup to mark the inode dirty later.
3428  */
3429 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3430 {
3431         struct ext3_iloc iloc;
3432
3433         int err = 0;
3434         if (handle) {
3435                 err = ext3_get_inode_loc(inode, &iloc);
3436                 if (!err) {
3437                         BUFFER_TRACE(iloc.bh, "get_write_access");
3438                         err = journal_get_write_access(handle, iloc.bh);
3439                         if (!err)
3440                                 err = ext3_journal_dirty_metadata(handle,
3441                                                                   iloc.bh);
3442                         brelse(iloc.bh);
3443                 }
3444         }
3445         ext3_std_error(inode->i_sb, err);
3446         return err;
3447 }
3448 #endif
3449
3450 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3451 {
3452         journal_t *journal;
3453         handle_t *handle;
3454         int err;
3455
3456         /*
3457          * We have to be very careful here: changing a data block's
3458          * journaling status dynamically is dangerous.  If we write a
3459          * data block to the journal, change the status and then delete
3460          * that block, we risk forgetting to revoke the old log record
3461          * from the journal and so a subsequent replay can corrupt data.
3462          * So, first we make sure that the journal is empty and that
3463          * nobody is changing anything.
3464          */
3465
3466         journal = EXT3_JOURNAL(inode);
3467         if (is_journal_aborted(journal))
3468                 return -EROFS;
3469
3470         journal_lock_updates(journal);
3471         journal_flush(journal);
3472
3473         /*
3474          * OK, there are no updates running now, and all cached data is
3475          * synced to disk.  We are now in a completely consistent state
3476          * which doesn't have anything in the journal, and we know that
3477          * no filesystem updates are running, so it is safe to modify
3478          * the inode's in-core data-journaling state flag now.
3479          */
3480
3481         if (val)
3482                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3483         else
3484                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3485         ext3_set_aops(inode);
3486
3487         journal_unlock_updates(journal);
3488
3489         /* Finally we can mark the inode as dirty. */
3490
3491         handle = ext3_journal_start(inode, 1);
3492         if (IS_ERR(handle))
3493                 return PTR_ERR(handle);
3494
3495         err = ext3_mark_inode_dirty(handle, inode);
3496         handle->h_sync = 1;
3497         ext3_journal_stop(handle);
3498         ext3_std_error(inode->i_sb, err);
3499
3500         return err;
3501 }