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