vfs: pagecache usage optimization for pagesize!=blocksize
[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 "xattr.h"
40 #include "acl.h"
41
42 static int ext3_writepage_trans_blocks(struct inode *inode);
43
44 /*
45  * Test whether an inode is a fast symlink.
46  */
47 static int ext3_inode_is_fast_symlink(struct inode *inode)
48 {
49         int ea_blocks = EXT3_I(inode)->i_file_acl ?
50                 (inode->i_sb->s_blocksize >> 9) : 0;
51
52         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
53 }
54
55 /*
56  * The ext3 forget function must perform a revoke if we are freeing data
57  * which has been journaled.  Metadata (eg. indirect blocks) must be
58  * revoked in all cases.
59  *
60  * "bh" may be NULL: a metadata block may have been freed from memory
61  * but there may still be a record of it in the journal, and that record
62  * still needs to be revoked.
63  */
64 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
65                         struct buffer_head *bh, ext3_fsblk_t blocknr)
66 {
67         int err;
68
69         might_sleep();
70
71         BUFFER_TRACE(bh, "enter");
72
73         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
74                   "data mode %lx\n",
75                   bh, is_metadata, inode->i_mode,
76                   test_opt(inode->i_sb, DATA_FLAGS));
77
78         /* Never use the revoke function if we are doing full data
79          * journaling: there is no need to, and a V1 superblock won't
80          * support it.  Otherwise, only skip the revoke on un-journaled
81          * data blocks. */
82
83         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
84             (!is_metadata && !ext3_should_journal_data(inode))) {
85                 if (bh) {
86                         BUFFER_TRACE(bh, "call journal_forget");
87                         return ext3_journal_forget(handle, bh);
88                 }
89                 return 0;
90         }
91
92         /*
93          * data!=journal && (is_metadata || should_journal_data(inode))
94          */
95         BUFFER_TRACE(bh, "call ext3_journal_revoke");
96         err = ext3_journal_revoke(handle, blocknr, bh);
97         if (err)
98                 ext3_abort(inode->i_sb, __func__,
99                            "error %d when attempting revoke", err);
100         BUFFER_TRACE(bh, "exit");
101         return err;
102 }
103
104 /*
105  * Work out how many blocks we need to proceed with the next chunk of a
106  * truncate transaction.
107  */
108 static unsigned long blocks_for_truncate(struct inode *inode)
109 {
110         unsigned long needed;
111
112         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
113
114         /* Give ourselves just enough room to cope with inodes in which
115          * i_blocks is corrupt: we've seen disk corruptions in the past
116          * which resulted in random data in an inode which looked enough
117          * like a regular file for ext3 to try to delete it.  Things
118          * will go a bit crazy if that happens, but at least we should
119          * try not to panic the whole kernel. */
120         if (needed < 2)
121                 needed = 2;
122
123         /* But we need to bound the transaction so we don't overflow the
124          * journal. */
125         if (needed > EXT3_MAX_TRANS_DATA)
126                 needed = EXT3_MAX_TRANS_DATA;
127
128         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
129 }
130
131 /*
132  * Truncate transactions can be complex and absolutely huge.  So we need to
133  * be able to restart the transaction at a conventient checkpoint to make
134  * sure we don't overflow the journal.
135  *
136  * start_transaction gets us a new handle for a truncate transaction,
137  * and extend_transaction tries to extend the existing one a bit.  If
138  * extend fails, we need to propagate the failure up and restart the
139  * transaction in the top-level truncate loop. --sct
140  */
141 static handle_t *start_transaction(struct inode *inode)
142 {
143         handle_t *result;
144
145         result = ext3_journal_start(inode, blocks_for_truncate(inode));
146         if (!IS_ERR(result))
147                 return result;
148
149         ext3_std_error(inode->i_sb, PTR_ERR(result));
150         return result;
151 }
152
153 /*
154  * Try to extend this transaction for the purposes of truncation.
155  *
156  * Returns 0 if we managed to create more room.  If we can't create more
157  * room, and the transaction must be restarted we return 1.
158  */
159 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
160 {
161         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
162                 return 0;
163         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
164                 return 0;
165         return 1;
166 }
167
168 /*
169  * Restart the transaction associated with *handle.  This does a commit,
170  * so before we call here everything must be consistently dirtied against
171  * this transaction.
172  */
173 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
174 {
175         jbd_debug(2, "restarting handle %p\n", handle);
176         return ext3_journal_restart(handle, blocks_for_truncate(inode));
177 }
178
179 /*
180  * Called at the last iput() if i_nlink is zero.
181  */
182 void ext3_delete_inode (struct inode * inode)
183 {
184         handle_t *handle;
185
186         truncate_inode_pages(&inode->i_data, 0);
187
188         if (is_bad_inode(inode))
189                 goto no_delete;
190
191         handle = start_transaction(inode);
192         if (IS_ERR(handle)) {
193                 /*
194                  * If we're going to skip the normal cleanup, we still need to
195                  * make sure that the in-core orphan linked list is properly
196                  * cleaned up.
197                  */
198                 ext3_orphan_del(NULL, inode);
199                 goto no_delete;
200         }
201
202         if (IS_SYNC(inode))
203                 handle->h_sync = 1;
204         inode->i_size = 0;
205         if (inode->i_blocks)
206                 ext3_truncate(inode);
207         /*
208          * Kill off the orphan record which ext3_truncate created.
209          * AKPM: I think this can be inside the above `if'.
210          * Note that ext3_orphan_del() has to be able to cope with the
211          * deletion of a non-existent orphan - this is because we don't
212          * know if ext3_truncate() actually created an orphan record.
213          * (Well, we could do this if we need to, but heck - it works)
214          */
215         ext3_orphan_del(handle, inode);
216         EXT3_I(inode)->i_dtime  = get_seconds();
217
218         /*
219          * One subtle ordering requirement: if anything has gone wrong
220          * (transaction abort, IO errors, whatever), then we can still
221          * do these next steps (the fs will already have been marked as
222          * having errors), but we can't free the inode if the mark_dirty
223          * fails.
224          */
225         if (ext3_mark_inode_dirty(handle, inode))
226                 /* If that failed, just do the required in-core inode clear. */
227                 clear_inode(inode);
228         else
229                 ext3_free_inode(handle, inode);
230         ext3_journal_stop(handle);
231         return;
232 no_delete:
233         clear_inode(inode);     /* We must guarantee clearing of inode... */
234 }
235
236 typedef struct {
237         __le32  *p;
238         __le32  key;
239         struct buffer_head *bh;
240 } Indirect;
241
242 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
243 {
244         p->key = *(p->p = v);
245         p->bh = bh;
246 }
247
248 static int verify_chain(Indirect *from, Indirect *to)
249 {
250         while (from <= to && from->key == *from->p)
251                 from++;
252         return (from > to);
253 }
254
255 /**
256  *      ext3_block_to_path - parse the block number into array of offsets
257  *      @inode: inode in question (we are only interested in its superblock)
258  *      @i_block: block number to be parsed
259  *      @offsets: array to store the offsets in
260  *      @boundary: set this non-zero if the referred-to block is likely to be
261  *             followed (on disk) by an indirect block.
262  *
263  *      To store the locations of file's data ext3 uses a data structure common
264  *      for UNIX filesystems - tree of pointers anchored in the inode, with
265  *      data blocks at leaves and indirect blocks in intermediate nodes.
266  *      This function translates the block number into path in that tree -
267  *      return value is the path length and @offsets[n] is the offset of
268  *      pointer to (n+1)th node in the nth one. If @block is out of range
269  *      (negative or too large) warning is printed and zero returned.
270  *
271  *      Note: function doesn't find node addresses, so no IO is needed. All
272  *      we need to know is the capacity of indirect blocks (taken from the
273  *      inode->i_sb).
274  */
275
276 /*
277  * Portability note: the last comparison (check that we fit into triple
278  * indirect block) is spelled differently, because otherwise on an
279  * architecture with 32-bit longs and 8Kb pages we might get into trouble
280  * if our filesystem had 8Kb blocks. We might use long long, but that would
281  * kill us on x86. Oh, well, at least the sign propagation does not matter -
282  * i_block would have to be negative in the very beginning, so we would not
283  * get there at all.
284  */
285
286 static int ext3_block_to_path(struct inode *inode,
287                         long i_block, int offsets[4], int *boundary)
288 {
289         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
290         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
291         const long direct_blocks = EXT3_NDIR_BLOCKS,
292                 indirect_blocks = ptrs,
293                 double_blocks = (1 << (ptrs_bits * 2));
294         int n = 0;
295         int final = 0;
296
297         if (i_block < 0) {
298                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
299         } else if (i_block < direct_blocks) {
300                 offsets[n++] = i_block;
301                 final = direct_blocks;
302         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
303                 offsets[n++] = EXT3_IND_BLOCK;
304                 offsets[n++] = i_block;
305                 final = ptrs;
306         } else if ((i_block -= indirect_blocks) < double_blocks) {
307                 offsets[n++] = EXT3_DIND_BLOCK;
308                 offsets[n++] = i_block >> ptrs_bits;
309                 offsets[n++] = i_block & (ptrs - 1);
310                 final = ptrs;
311         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
312                 offsets[n++] = EXT3_TIND_BLOCK;
313                 offsets[n++] = i_block >> (ptrs_bits * 2);
314                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
315                 offsets[n++] = i_block & (ptrs - 1);
316                 final = ptrs;
317         } else {
318                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
319         }
320         if (boundary)
321                 *boundary = final - 1 - (i_block & (ptrs - 1));
322         return n;
323 }
324
325 /**
326  *      ext3_get_branch - read the chain of indirect blocks leading to data
327  *      @inode: inode in question
328  *      @depth: depth of the chain (1 - direct pointer, etc.)
329  *      @offsets: offsets of pointers in inode/indirect blocks
330  *      @chain: place to store the result
331  *      @err: here we store the error value
332  *
333  *      Function fills the array of triples <key, p, bh> and returns %NULL
334  *      if everything went OK or the pointer to the last filled triple
335  *      (incomplete one) otherwise. Upon the return chain[i].key contains
336  *      the number of (i+1)-th block in the chain (as it is stored in memory,
337  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
338  *      number (it points into struct inode for i==0 and into the bh->b_data
339  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
340  *      block for i>0 and NULL for i==0. In other words, it holds the block
341  *      numbers of the chain, addresses they were taken from (and where we can
342  *      verify that chain did not change) and buffer_heads hosting these
343  *      numbers.
344  *
345  *      Function stops when it stumbles upon zero pointer (absent block)
346  *              (pointer to last triple returned, *@err == 0)
347  *      or when it gets an IO error reading an indirect block
348  *              (ditto, *@err == -EIO)
349  *      or when it notices that chain had been changed while it was reading
350  *              (ditto, *@err == -EAGAIN)
351  *      or when it reads all @depth-1 indirect blocks successfully and finds
352  *      the whole chain, all way to the data (returns %NULL, *err == 0).
353  */
354 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
355                                  Indirect chain[4], int *err)
356 {
357         struct super_block *sb = inode->i_sb;
358         Indirect *p = chain;
359         struct buffer_head *bh;
360
361         *err = 0;
362         /* i_data is not going away, no lock needed */
363         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
364         if (!p->key)
365                 goto no_block;
366         while (--depth) {
367                 bh = sb_bread(sb, le32_to_cpu(p->key));
368                 if (!bh)
369                         goto failure;
370                 /* Reader: pointers */
371                 if (!verify_chain(chain, p))
372                         goto changed;
373                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
374                 /* Reader: end */
375                 if (!p->key)
376                         goto no_block;
377         }
378         return NULL;
379
380 changed:
381         brelse(bh);
382         *err = -EAGAIN;
383         goto no_block;
384 failure:
385         *err = -EIO;
386 no_block:
387         return p;
388 }
389
390 /**
391  *      ext3_find_near - find a place for allocation with sufficient locality
392  *      @inode: owner
393  *      @ind: descriptor of indirect block.
394  *
395  *      This function returns the preferred place for block allocation.
396  *      It is used when heuristic for sequential allocation fails.
397  *      Rules are:
398  *        + if there is a block to the left of our position - allocate near it.
399  *        + if pointer will live in indirect block - allocate near that block.
400  *        + if pointer will live in inode - allocate in the same
401  *          cylinder group.
402  *
403  * In the latter case we colour the starting block by the callers PID to
404  * prevent it from clashing with concurrent allocations for a different inode
405  * in the same block group.   The PID is used here so that functionally related
406  * files will be close-by on-disk.
407  *
408  *      Caller must make sure that @ind is valid and will stay that way.
409  */
410 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
411 {
412         struct ext3_inode_info *ei = EXT3_I(inode);
413         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
414         __le32 *p;
415         ext3_fsblk_t bg_start;
416         ext3_grpblk_t colour;
417
418         /* Try to find previous block */
419         for (p = ind->p - 1; p >= start; p--) {
420                 if (*p)
421                         return le32_to_cpu(*p);
422         }
423
424         /* No such thing, so let's try location of indirect block */
425         if (ind->bh)
426                 return ind->bh->b_blocknr;
427
428         /*
429          * It is going to be referred to from the inode itself? OK, just put it
430          * into the same cylinder group then.
431          */
432         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
433         colour = (current->pid % 16) *
434                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
435         return bg_start + colour;
436 }
437
438 /**
439  *      ext3_find_goal - find a preferred place for allocation.
440  *      @inode: owner
441  *      @block:  block we want
442  *      @partial: pointer to the last triple within a chain
443  *
444  *      Normally this function find the preferred place for block allocation,
445  *      returns it.
446  */
447
448 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
449                                    Indirect *partial)
450 {
451         struct ext3_block_alloc_info *block_i;
452
453         block_i =  EXT3_I(inode)->i_block_alloc_info;
454
455         /*
456          * try the heuristic for sequential allocation,
457          * failing that at least try to get decent locality.
458          */
459         if (block_i && (block == block_i->last_alloc_logical_block + 1)
460                 && (block_i->last_alloc_physical_block != 0)) {
461                 return block_i->last_alloc_physical_block + 1;
462         }
463
464         return ext3_find_near(inode, partial);
465 }
466
467 /**
468  *      ext3_blks_to_allocate: Look up the block map and count the number
469  *      of direct blocks need to be allocated for the given branch.
470  *
471  *      @branch: chain of indirect blocks
472  *      @k: number of blocks need for indirect blocks
473  *      @blks: number of data blocks to be mapped.
474  *      @blocks_to_boundary:  the offset in the indirect block
475  *
476  *      return the total number of blocks to be allocate, including the
477  *      direct and indirect blocks.
478  */
479 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
480                 int blocks_to_boundary)
481 {
482         unsigned long count = 0;
483
484         /*
485          * Simple case, [t,d]Indirect block(s) has not allocated yet
486          * then it's clear blocks on that path have not allocated
487          */
488         if (k > 0) {
489                 /* right now we don't handle cross boundary allocation */
490                 if (blks < blocks_to_boundary + 1)
491                         count += blks;
492                 else
493                         count += blocks_to_boundary + 1;
494                 return count;
495         }
496
497         count++;
498         while (count < blks && count <= blocks_to_boundary &&
499                 le32_to_cpu(*(branch[0].p + count)) == 0) {
500                 count++;
501         }
502         return count;
503 }
504
505 /**
506  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
507  *      @indirect_blks: the number of blocks need to allocate for indirect
508  *                      blocks
509  *
510  *      @new_blocks: on return it will store the new block numbers for
511  *      the indirect blocks(if needed) and the first direct block,
512  *      @blks:  on return it will store the total number of allocated
513  *              direct blocks
514  */
515 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
516                         ext3_fsblk_t goal, int indirect_blks, int blks,
517                         ext3_fsblk_t new_blocks[4], int *err)
518 {
519         int target, i;
520         unsigned long count = 0;
521         int index = 0;
522         ext3_fsblk_t current_block = 0;
523         int ret = 0;
524
525         /*
526          * Here we try to allocate the requested multiple blocks at once,
527          * on a best-effort basis.
528          * To build a branch, we should allocate blocks for
529          * the indirect blocks(if not allocated yet), and at least
530          * the first direct block of this branch.  That's the
531          * minimum number of blocks need to allocate(required)
532          */
533         target = blks + indirect_blks;
534
535         while (1) {
536                 count = target;
537                 /* allocating blocks for indirect blocks and direct blocks */
538                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
539                 if (*err)
540                         goto failed_out;
541
542                 target -= count;
543                 /* allocate blocks for indirect blocks */
544                 while (index < indirect_blks && count) {
545                         new_blocks[index++] = current_block++;
546                         count--;
547                 }
548
549                 if (count > 0)
550                         break;
551         }
552
553         /* save the new block number for the first direct block */
554         new_blocks[index] = current_block;
555
556         /* total number of blocks allocated for direct blocks */
557         ret = count;
558         *err = 0;
559         return ret;
560 failed_out:
561         for (i = 0; i <index; i++)
562                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
563         return ret;
564 }
565
566 /**
567  *      ext3_alloc_branch - allocate and set up a chain of blocks.
568  *      @inode: owner
569  *      @indirect_blks: number of allocated indirect blocks
570  *      @blks: number of allocated direct blocks
571  *      @offsets: offsets (in the blocks) to store the pointers to next.
572  *      @branch: place to store the chain in.
573  *
574  *      This function allocates blocks, zeroes out all but the last one,
575  *      links them into chain and (if we are synchronous) writes them to disk.
576  *      In other words, it prepares a branch that can be spliced onto the
577  *      inode. It stores the information about that chain in the branch[], in
578  *      the same format as ext3_get_branch() would do. We are calling it after
579  *      we had read the existing part of chain and partial points to the last
580  *      triple of that (one with zero ->key). Upon the exit we have the same
581  *      picture as after the successful ext3_get_block(), except that in one
582  *      place chain is disconnected - *branch->p is still zero (we did not
583  *      set the last link), but branch->key contains the number that should
584  *      be placed into *branch->p to fill that gap.
585  *
586  *      If allocation fails we free all blocks we've allocated (and forget
587  *      their buffer_heads) and return the error value the from failed
588  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
589  *      as described above and return 0.
590  */
591 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
592                         int indirect_blks, int *blks, ext3_fsblk_t goal,
593                         int *offsets, Indirect *branch)
594 {
595         int blocksize = inode->i_sb->s_blocksize;
596         int i, n = 0;
597         int err = 0;
598         struct buffer_head *bh;
599         int num;
600         ext3_fsblk_t new_blocks[4];
601         ext3_fsblk_t current_block;
602
603         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
604                                 *blks, new_blocks, &err);
605         if (err)
606                 return err;
607
608         branch[0].key = cpu_to_le32(new_blocks[0]);
609         /*
610          * metadata blocks and data blocks are allocated.
611          */
612         for (n = 1; n <= indirect_blks;  n++) {
613                 /*
614                  * Get buffer_head for parent block, zero it out
615                  * and set the pointer to new one, then send
616                  * parent to disk.
617                  */
618                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
619                 branch[n].bh = bh;
620                 lock_buffer(bh);
621                 BUFFER_TRACE(bh, "call get_create_access");
622                 err = ext3_journal_get_create_access(handle, bh);
623                 if (err) {
624                         unlock_buffer(bh);
625                         brelse(bh);
626                         goto failed;
627                 }
628
629                 memset(bh->b_data, 0, blocksize);
630                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
631                 branch[n].key = cpu_to_le32(new_blocks[n]);
632                 *branch[n].p = branch[n].key;
633                 if ( n == indirect_blks) {
634                         current_block = new_blocks[n];
635                         /*
636                          * End of chain, update the last new metablock of
637                          * the chain to point to the new allocated
638                          * data blocks numbers
639                          */
640                         for (i=1; i < num; i++)
641                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
642                 }
643                 BUFFER_TRACE(bh, "marking uptodate");
644                 set_buffer_uptodate(bh);
645                 unlock_buffer(bh);
646
647                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
648                 err = ext3_journal_dirty_metadata(handle, bh);
649                 if (err)
650                         goto failed;
651         }
652         *blks = num;
653         return err;
654 failed:
655         /* Allocation failed, free what we already allocated */
656         for (i = 1; i <= n ; i++) {
657                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
658                 ext3_journal_forget(handle, branch[i].bh);
659         }
660         for (i = 0; i <indirect_blks; i++)
661                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
662
663         ext3_free_blocks(handle, inode, new_blocks[i], num);
664
665         return err;
666 }
667
668 /**
669  * ext3_splice_branch - splice the allocated branch onto inode.
670  * @inode: owner
671  * @block: (logical) number of block we are adding
672  * @chain: chain of indirect blocks (with a missing link - see
673  *      ext3_alloc_branch)
674  * @where: location of missing link
675  * @num:   number of indirect blocks we are adding
676  * @blks:  number of direct blocks we are adding
677  *
678  * This function fills the missing link and does all housekeeping needed in
679  * inode (->i_blocks, etc.). In case of success we end up with the full
680  * chain to new block and return 0.
681  */
682 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
683                         long block, Indirect *where, int num, int blks)
684 {
685         int i;
686         int err = 0;
687         struct ext3_block_alloc_info *block_i;
688         ext3_fsblk_t current_block;
689
690         block_i = EXT3_I(inode)->i_block_alloc_info;
691         /*
692          * If we're splicing into a [td]indirect block (as opposed to the
693          * inode) then we need to get write access to the [td]indirect block
694          * before the splice.
695          */
696         if (where->bh) {
697                 BUFFER_TRACE(where->bh, "get_write_access");
698                 err = ext3_journal_get_write_access(handle, where->bh);
699                 if (err)
700                         goto err_out;
701         }
702         /* That's it */
703
704         *where->p = where->key;
705
706         /*
707          * Update the host buffer_head or inode to point to more just allocated
708          * direct blocks blocks
709          */
710         if (num == 0 && blks > 1) {
711                 current_block = le32_to_cpu(where->key) + 1;
712                 for (i = 1; i < blks; i++)
713                         *(where->p + i ) = cpu_to_le32(current_block++);
714         }
715
716         /*
717          * update the most recently allocated logical & physical block
718          * in i_block_alloc_info, to assist find the proper goal block for next
719          * allocation
720          */
721         if (block_i) {
722                 block_i->last_alloc_logical_block = block + blks - 1;
723                 block_i->last_alloc_physical_block =
724                                 le32_to_cpu(where[num].key) + blks - 1;
725         }
726
727         /* We are done with atomic stuff, now do the rest of housekeeping */
728
729         inode->i_ctime = CURRENT_TIME_SEC;
730         ext3_mark_inode_dirty(handle, inode);
731
732         /* had we spliced it onto indirect block? */
733         if (where->bh) {
734                 /*
735                  * If we spliced it onto an indirect block, we haven't
736                  * altered the inode.  Note however that if it is being spliced
737                  * onto an indirect block at the very end of the file (the
738                  * file is growing) then we *will* alter the inode to reflect
739                  * the new i_size.  But that is not done here - it is done in
740                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
741                  */
742                 jbd_debug(5, "splicing indirect only\n");
743                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
744                 err = ext3_journal_dirty_metadata(handle, where->bh);
745                 if (err)
746                         goto err_out;
747         } else {
748                 /*
749                  * OK, we spliced it into the inode itself on a direct block.
750                  * Inode was dirtied above.
751                  */
752                 jbd_debug(5, "splicing direct\n");
753         }
754         return err;
755
756 err_out:
757         for (i = 1; i <= num; i++) {
758                 BUFFER_TRACE(where[i].bh, "call journal_forget");
759                 ext3_journal_forget(handle, where[i].bh);
760                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
761         }
762         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
763
764         return err;
765 }
766
767 /*
768  * Allocation strategy is simple: if we have to allocate something, we will
769  * have to go the whole way to leaf. So let's do it before attaching anything
770  * to tree, set linkage between the newborn blocks, write them if sync is
771  * required, recheck the path, free and repeat if check fails, otherwise
772  * set the last missing link (that will protect us from any truncate-generated
773  * removals - all blocks on the path are immune now) and possibly force the
774  * write on the parent block.
775  * That has a nice additional property: no special recovery from the failed
776  * allocations is needed - we simply release blocks and do not touch anything
777  * reachable from inode.
778  *
779  * `handle' can be NULL if create == 0.
780  *
781  * The BKL may not be held on entry here.  Be sure to take it early.
782  * return > 0, # of blocks mapped or allocated.
783  * return = 0, if plain lookup failed.
784  * return < 0, error case.
785  */
786 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
787                 sector_t iblock, unsigned long maxblocks,
788                 struct buffer_head *bh_result,
789                 int create, int extend_disksize)
790 {
791         int err = -EIO;
792         int offsets[4];
793         Indirect chain[4];
794         Indirect *partial;
795         ext3_fsblk_t goal;
796         int indirect_blks;
797         int blocks_to_boundary = 0;
798         int depth;
799         struct ext3_inode_info *ei = EXT3_I(inode);
800         int count = 0;
801         ext3_fsblk_t first_block = 0;
802
803
804         J_ASSERT(handle != NULL || create == 0);
805         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
806
807         if (depth == 0)
808                 goto out;
809
810         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
811
812         /* Simplest case - block found, no allocation needed */
813         if (!partial) {
814                 first_block = le32_to_cpu(chain[depth - 1].key);
815                 clear_buffer_new(bh_result);
816                 count++;
817                 /*map more blocks*/
818                 while (count < maxblocks && count <= blocks_to_boundary) {
819                         ext3_fsblk_t blk;
820
821                         if (!verify_chain(chain, partial)) {
822                                 /*
823                                  * Indirect block might be removed by
824                                  * truncate while we were reading it.
825                                  * Handling of that case: forget what we've
826                                  * got now. Flag the err as EAGAIN, so it
827                                  * will reread.
828                                  */
829                                 err = -EAGAIN;
830                                 count = 0;
831                                 break;
832                         }
833                         blk = le32_to_cpu(*(chain[depth-1].p + count));
834
835                         if (blk == first_block + count)
836                                 count++;
837                         else
838                                 break;
839                 }
840                 if (err != -EAGAIN)
841                         goto got_it;
842         }
843
844         /* Next simple case - plain lookup or failed read of indirect block */
845         if (!create || err == -EIO)
846                 goto cleanup;
847
848         mutex_lock(&ei->truncate_mutex);
849
850         /*
851          * If the indirect block is missing while we are reading
852          * the chain(ext3_get_branch() returns -EAGAIN err), or
853          * if the chain has been changed after we grab the semaphore,
854          * (either because another process truncated this branch, or
855          * another get_block allocated this branch) re-grab the chain to see if
856          * the request block has been allocated or not.
857          *
858          * Since we already block the truncate/other get_block
859          * at this point, we will have the current copy of the chain when we
860          * splice the branch into the tree.
861          */
862         if (err == -EAGAIN || !verify_chain(chain, partial)) {
863                 while (partial > chain) {
864                         brelse(partial->bh);
865                         partial--;
866                 }
867                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
868                 if (!partial) {
869                         count++;
870                         mutex_unlock(&ei->truncate_mutex);
871                         if (err)
872                                 goto cleanup;
873                         clear_buffer_new(bh_result);
874                         goto got_it;
875                 }
876         }
877
878         /*
879          * Okay, we need to do block allocation.  Lazily initialize the block
880          * allocation info here if necessary
881         */
882         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
883                 ext3_init_block_alloc_info(inode);
884
885         goal = ext3_find_goal(inode, iblock, partial);
886
887         /* the number of blocks need to allocate for [d,t]indirect blocks */
888         indirect_blks = (chain + depth) - partial - 1;
889
890         /*
891          * Next look up the indirect map to count the totoal number of
892          * direct blocks to allocate for this branch.
893          */
894         count = ext3_blks_to_allocate(partial, indirect_blks,
895                                         maxblocks, blocks_to_boundary);
896         /*
897          * Block out ext3_truncate while we alter the tree
898          */
899         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
900                                 offsets + (partial - chain), partial);
901
902         /*
903          * The ext3_splice_branch call will free and forget any buffers
904          * on the new chain if there is a failure, but that risks using
905          * up transaction credits, especially for bitmaps where the
906          * credits cannot be returned.  Can we handle this somehow?  We
907          * may need to return -EAGAIN upwards in the worst case.  --sct
908          */
909         if (!err)
910                 err = ext3_splice_branch(handle, inode, iblock,
911                                         partial, indirect_blks, count);
912         /*
913          * i_disksize growing is protected by truncate_mutex.  Don't forget to
914          * protect it if you're about to implement concurrent
915          * ext3_get_block() -bzzz
916         */
917         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
918                 ei->i_disksize = inode->i_size;
919         mutex_unlock(&ei->truncate_mutex);
920         if (err)
921                 goto cleanup;
922
923         set_buffer_new(bh_result);
924 got_it:
925         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
926         if (count > blocks_to_boundary)
927                 set_buffer_boundary(bh_result);
928         err = count;
929         /* Clean up and exit */
930         partial = chain + depth - 1;    /* the whole chain */
931 cleanup:
932         while (partial > chain) {
933                 BUFFER_TRACE(partial->bh, "call brelse");
934                 brelse(partial->bh);
935                 partial--;
936         }
937         BUFFER_TRACE(bh_result, "returned");
938 out:
939         return err;
940 }
941
942 /* Maximum number of blocks we map for direct IO at once. */
943 #define DIO_MAX_BLOCKS 4096
944 /*
945  * Number of credits we need for writing DIO_MAX_BLOCKS:
946  * We need sb + group descriptor + bitmap + inode -> 4
947  * For B blocks with A block pointers per block we need:
948  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
949  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
950  */
951 #define DIO_CREDITS 25
952
953 static int ext3_get_block(struct inode *inode, sector_t iblock,
954                         struct buffer_head *bh_result, int create)
955 {
956         handle_t *handle = ext3_journal_current_handle();
957         int ret = 0, started = 0;
958         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
959
960         if (create && !handle) {        /* Direct IO write... */
961                 if (max_blocks > DIO_MAX_BLOCKS)
962                         max_blocks = DIO_MAX_BLOCKS;
963                 handle = ext3_journal_start(inode, DIO_CREDITS +
964                                 2 * EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb));
965                 if (IS_ERR(handle)) {
966                         ret = PTR_ERR(handle);
967                         goto out;
968                 }
969                 started = 1;
970         }
971
972         ret = ext3_get_blocks_handle(handle, inode, iblock,
973                                         max_blocks, bh_result, create, 0);
974         if (ret > 0) {
975                 bh_result->b_size = (ret << inode->i_blkbits);
976                 ret = 0;
977         }
978         if (started)
979                 ext3_journal_stop(handle);
980 out:
981         return ret;
982 }
983
984 /*
985  * `handle' can be NULL if create is zero
986  */
987 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
988                                 long block, int create, int *errp)
989 {
990         struct buffer_head dummy;
991         int fatal = 0, err;
992
993         J_ASSERT(handle != NULL || create == 0);
994
995         dummy.b_state = 0;
996         dummy.b_blocknr = -1000;
997         buffer_trace_init(&dummy.b_history);
998         err = ext3_get_blocks_handle(handle, inode, block, 1,
999                                         &dummy, create, 1);
1000         /*
1001          * ext3_get_blocks_handle() returns number of blocks
1002          * mapped. 0 in case of a HOLE.
1003          */
1004         if (err > 0) {
1005                 if (err > 1)
1006                         WARN_ON(1);
1007                 err = 0;
1008         }
1009         *errp = err;
1010         if (!err && buffer_mapped(&dummy)) {
1011                 struct buffer_head *bh;
1012                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1013                 if (!bh) {
1014                         *errp = -EIO;
1015                         goto err;
1016                 }
1017                 if (buffer_new(&dummy)) {
1018                         J_ASSERT(create != 0);
1019                         J_ASSERT(handle != NULL);
1020
1021                         /*
1022                          * Now that we do not always journal data, we should
1023                          * keep in mind whether this should always journal the
1024                          * new buffer as metadata.  For now, regular file
1025                          * writes use ext3_get_block instead, so it's not a
1026                          * problem.
1027                          */
1028                         lock_buffer(bh);
1029                         BUFFER_TRACE(bh, "call get_create_access");
1030                         fatal = ext3_journal_get_create_access(handle, bh);
1031                         if (!fatal && !buffer_uptodate(bh)) {
1032                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1033                                 set_buffer_uptodate(bh);
1034                         }
1035                         unlock_buffer(bh);
1036                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1037                         err = ext3_journal_dirty_metadata(handle, bh);
1038                         if (!fatal)
1039                                 fatal = err;
1040                 } else {
1041                         BUFFER_TRACE(bh, "not a new buffer");
1042                 }
1043                 if (fatal) {
1044                         *errp = fatal;
1045                         brelse(bh);
1046                         bh = NULL;
1047                 }
1048                 return bh;
1049         }
1050 err:
1051         return NULL;
1052 }
1053
1054 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1055                                int block, int create, int *err)
1056 {
1057         struct buffer_head * bh;
1058
1059         bh = ext3_getblk(handle, inode, block, create, err);
1060         if (!bh)
1061                 return bh;
1062         if (buffer_uptodate(bh))
1063                 return bh;
1064         ll_rw_block(READ_META, 1, &bh);
1065         wait_on_buffer(bh);
1066         if (buffer_uptodate(bh))
1067                 return bh;
1068         put_bh(bh);
1069         *err = -EIO;
1070         return NULL;
1071 }
1072
1073 static int walk_page_buffers(   handle_t *handle,
1074                                 struct buffer_head *head,
1075                                 unsigned from,
1076                                 unsigned to,
1077                                 int *partial,
1078                                 int (*fn)(      handle_t *handle,
1079                                                 struct buffer_head *bh))
1080 {
1081         struct buffer_head *bh;
1082         unsigned block_start, block_end;
1083         unsigned blocksize = head->b_size;
1084         int err, ret = 0;
1085         struct buffer_head *next;
1086
1087         for (   bh = head, block_start = 0;
1088                 ret == 0 && (bh != head || !block_start);
1089                 block_start = block_end, bh = next)
1090         {
1091                 next = bh->b_this_page;
1092                 block_end = block_start + blocksize;
1093                 if (block_end <= from || block_start >= to) {
1094                         if (partial && !buffer_uptodate(bh))
1095                                 *partial = 1;
1096                         continue;
1097                 }
1098                 err = (*fn)(handle, bh);
1099                 if (!ret)
1100                         ret = err;
1101         }
1102         return ret;
1103 }
1104
1105 /*
1106  * To preserve ordering, it is essential that the hole instantiation and
1107  * the data write be encapsulated in a single transaction.  We cannot
1108  * close off a transaction and start a new one between the ext3_get_block()
1109  * and the commit_write().  So doing the journal_start at the start of
1110  * prepare_write() is the right place.
1111  *
1112  * Also, this function can nest inside ext3_writepage() ->
1113  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1114  * has generated enough buffer credits to do the whole page.  So we won't
1115  * block on the journal in that case, which is good, because the caller may
1116  * be PF_MEMALLOC.
1117  *
1118  * By accident, ext3 can be reentered when a transaction is open via
1119  * quota file writes.  If we were to commit the transaction while thus
1120  * reentered, there can be a deadlock - we would be holding a quota
1121  * lock, and the commit would never complete if another thread had a
1122  * transaction open and was blocking on the quota lock - a ranking
1123  * violation.
1124  *
1125  * So what we do is to rely on the fact that journal_stop/journal_start
1126  * will _not_ run commit under these circumstances because handle->h_ref
1127  * is elevated.  We'll still have enough credits for the tiny quotafile
1128  * write.
1129  */
1130 static int do_journal_get_write_access(handle_t *handle,
1131                                         struct buffer_head *bh)
1132 {
1133         if (!buffer_mapped(bh) || buffer_freed(bh))
1134                 return 0;
1135         return ext3_journal_get_write_access(handle, bh);
1136 }
1137
1138 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1139                                 loff_t pos, unsigned len, unsigned flags,
1140                                 struct page **pagep, void **fsdata)
1141 {
1142         struct inode *inode = mapping->host;
1143         int ret, needed_blocks = ext3_writepage_trans_blocks(inode);
1144         handle_t *handle;
1145         int retries = 0;
1146         struct page *page;
1147         pgoff_t index;
1148         unsigned from, to;
1149
1150         index = pos >> PAGE_CACHE_SHIFT;
1151         from = pos & (PAGE_CACHE_SIZE - 1);
1152         to = from + len;
1153
1154 retry:
1155         page = __grab_cache_page(mapping, index);
1156         if (!page)
1157                 return -ENOMEM;
1158         *pagep = page;
1159
1160         handle = ext3_journal_start(inode, needed_blocks);
1161         if (IS_ERR(handle)) {
1162                 unlock_page(page);
1163                 page_cache_release(page);
1164                 ret = PTR_ERR(handle);
1165                 goto out;
1166         }
1167         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1168                                                         ext3_get_block);
1169         if (ret)
1170                 goto write_begin_failed;
1171
1172         if (ext3_should_journal_data(inode)) {
1173                 ret = walk_page_buffers(handle, page_buffers(page),
1174                                 from, to, NULL, do_journal_get_write_access);
1175         }
1176 write_begin_failed:
1177         if (ret) {
1178                 ext3_journal_stop(handle);
1179                 unlock_page(page);
1180                 page_cache_release(page);
1181         }
1182         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1183                 goto retry;
1184 out:
1185         return ret;
1186 }
1187
1188
1189 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1190 {
1191         int err = journal_dirty_data(handle, bh);
1192         if (err)
1193                 ext3_journal_abort_handle(__func__, __func__,
1194                                                 bh, handle, err);
1195         return err;
1196 }
1197
1198 /* For write_end() in data=journal mode */
1199 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1200 {
1201         if (!buffer_mapped(bh) || buffer_freed(bh))
1202                 return 0;
1203         set_buffer_uptodate(bh);
1204         return ext3_journal_dirty_metadata(handle, bh);
1205 }
1206
1207 /*
1208  * Generic write_end handler for ordered and writeback ext3 journal modes.
1209  * We can't use generic_write_end, because that unlocks the page and we need to
1210  * unlock the page after ext3_journal_stop, but ext3_journal_stop must run
1211  * after block_write_end.
1212  */
1213 static int ext3_generic_write_end(struct file *file,
1214                                 struct address_space *mapping,
1215                                 loff_t pos, unsigned len, unsigned copied,
1216                                 struct page *page, void *fsdata)
1217 {
1218         struct inode *inode = file->f_mapping->host;
1219
1220         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1221
1222         if (pos+copied > inode->i_size) {
1223                 i_size_write(inode, pos+copied);
1224                 mark_inode_dirty(inode);
1225         }
1226
1227         return copied;
1228 }
1229
1230 /*
1231  * We need to pick up the new inode size which generic_commit_write gave us
1232  * `file' can be NULL - eg, when called from page_symlink().
1233  *
1234  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1235  * buffers are managed internally.
1236  */
1237 static int ext3_ordered_write_end(struct file *file,
1238                                 struct address_space *mapping,
1239                                 loff_t pos, unsigned len, unsigned copied,
1240                                 struct page *page, void *fsdata)
1241 {
1242         handle_t *handle = ext3_journal_current_handle();
1243         struct inode *inode = file->f_mapping->host;
1244         unsigned from, to;
1245         int ret = 0, ret2;
1246
1247         from = pos & (PAGE_CACHE_SIZE - 1);
1248         to = from + len;
1249
1250         ret = walk_page_buffers(handle, page_buffers(page),
1251                 from, to, NULL, ext3_journal_dirty_data);
1252
1253         if (ret == 0) {
1254                 /*
1255                  * generic_write_end() will run mark_inode_dirty() if i_size
1256                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1257                  * into that.
1258                  */
1259                 loff_t new_i_size;
1260
1261                 new_i_size = pos + copied;
1262                 if (new_i_size > EXT3_I(inode)->i_disksize)
1263                         EXT3_I(inode)->i_disksize = new_i_size;
1264                 ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1265                                                         page, fsdata);
1266                 copied = ret2;
1267                 if (ret2 < 0)
1268                         ret = ret2;
1269         }
1270         ret2 = ext3_journal_stop(handle);
1271         if (!ret)
1272                 ret = ret2;
1273         unlock_page(page);
1274         page_cache_release(page);
1275
1276         return ret ? ret : copied;
1277 }
1278
1279 static int ext3_writeback_write_end(struct file *file,
1280                                 struct address_space *mapping,
1281                                 loff_t pos, unsigned len, unsigned copied,
1282                                 struct page *page, void *fsdata)
1283 {
1284         handle_t *handle = ext3_journal_current_handle();
1285         struct inode *inode = file->f_mapping->host;
1286         int ret = 0, ret2;
1287         loff_t new_i_size;
1288
1289         new_i_size = pos + copied;
1290         if (new_i_size > EXT3_I(inode)->i_disksize)
1291                 EXT3_I(inode)->i_disksize = new_i_size;
1292
1293         ret2 = ext3_generic_write_end(file, mapping, pos, len, copied,
1294                                                         page, fsdata);
1295         copied = ret2;
1296         if (ret2 < 0)
1297                 ret = ret2;
1298
1299         ret2 = ext3_journal_stop(handle);
1300         if (!ret)
1301                 ret = ret2;
1302         unlock_page(page);
1303         page_cache_release(page);
1304
1305         return ret ? ret : copied;
1306 }
1307
1308 static int ext3_journalled_write_end(struct file *file,
1309                                 struct address_space *mapping,
1310                                 loff_t pos, unsigned len, unsigned copied,
1311                                 struct page *page, void *fsdata)
1312 {
1313         handle_t *handle = ext3_journal_current_handle();
1314         struct inode *inode = mapping->host;
1315         int ret = 0, ret2;
1316         int partial = 0;
1317         unsigned from, to;
1318
1319         from = pos & (PAGE_CACHE_SIZE - 1);
1320         to = from + len;
1321
1322         if (copied < len) {
1323                 if (!PageUptodate(page))
1324                         copied = 0;
1325                 page_zero_new_buffers(page, from+copied, to);
1326         }
1327
1328         ret = walk_page_buffers(handle, page_buffers(page), from,
1329                                 to, &partial, write_end_fn);
1330         if (!partial)
1331                 SetPageUptodate(page);
1332         if (pos+copied > inode->i_size)
1333                 i_size_write(inode, pos+copied);
1334         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1335         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1336                 EXT3_I(inode)->i_disksize = inode->i_size;
1337                 ret2 = ext3_mark_inode_dirty(handle, inode);
1338                 if (!ret)
1339                         ret = ret2;
1340         }
1341
1342         ret2 = ext3_journal_stop(handle);
1343         if (!ret)
1344                 ret = ret2;
1345         unlock_page(page);
1346         page_cache_release(page);
1347
1348         return ret ? ret : copied;
1349 }
1350
1351 /*
1352  * bmap() is special.  It gets used by applications such as lilo and by
1353  * the swapper to find the on-disk block of a specific piece of data.
1354  *
1355  * Naturally, this is dangerous if the block concerned is still in the
1356  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1357  * filesystem and enables swap, then they may get a nasty shock when the
1358  * data getting swapped to that swapfile suddenly gets overwritten by
1359  * the original zero's written out previously to the journal and
1360  * awaiting writeback in the kernel's buffer cache.
1361  *
1362  * So, if we see any bmap calls here on a modified, data-journaled file,
1363  * take extra steps to flush any blocks which might be in the cache.
1364  */
1365 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1366 {
1367         struct inode *inode = mapping->host;
1368         journal_t *journal;
1369         int err;
1370
1371         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1372                 /*
1373                  * This is a REALLY heavyweight approach, but the use of
1374                  * bmap on dirty files is expected to be extremely rare:
1375                  * only if we run lilo or swapon on a freshly made file
1376                  * do we expect this to happen.
1377                  *
1378                  * (bmap requires CAP_SYS_RAWIO so this does not
1379                  * represent an unprivileged user DOS attack --- we'd be
1380                  * in trouble if mortal users could trigger this path at
1381                  * will.)
1382                  *
1383                  * NB. EXT3_STATE_JDATA is not set on files other than
1384                  * regular files.  If somebody wants to bmap a directory
1385                  * or symlink and gets confused because the buffer
1386                  * hasn't yet been flushed to disk, they deserve
1387                  * everything they get.
1388                  */
1389
1390                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1391                 journal = EXT3_JOURNAL(inode);
1392                 journal_lock_updates(journal);
1393                 err = journal_flush(journal);
1394                 journal_unlock_updates(journal);
1395
1396                 if (err)
1397                         return 0;
1398         }
1399
1400         return generic_block_bmap(mapping,block,ext3_get_block);
1401 }
1402
1403 static int bget_one(handle_t *handle, struct buffer_head *bh)
1404 {
1405         get_bh(bh);
1406         return 0;
1407 }
1408
1409 static int bput_one(handle_t *handle, struct buffer_head *bh)
1410 {
1411         put_bh(bh);
1412         return 0;
1413 }
1414
1415 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1416 {
1417         if (buffer_mapped(bh))
1418                 return ext3_journal_dirty_data(handle, bh);
1419         return 0;
1420 }
1421
1422 /*
1423  * Note that we always start a transaction even if we're not journalling
1424  * data.  This is to preserve ordering: any hole instantiation within
1425  * __block_write_full_page -> ext3_get_block() should be journalled
1426  * along with the data so we don't crash and then get metadata which
1427  * refers to old data.
1428  *
1429  * In all journalling modes block_write_full_page() will start the I/O.
1430  *
1431  * Problem:
1432  *
1433  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1434  *              ext3_writepage()
1435  *
1436  * Similar for:
1437  *
1438  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1439  *
1440  * Same applies to ext3_get_block().  We will deadlock on various things like
1441  * lock_journal and i_truncate_mutex.
1442  *
1443  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1444  * allocations fail.
1445  *
1446  * 16May01: If we're reentered then journal_current_handle() will be
1447  *          non-zero. We simply *return*.
1448  *
1449  * 1 July 2001: @@@ FIXME:
1450  *   In journalled data mode, a data buffer may be metadata against the
1451  *   current transaction.  But the same file is part of a shared mapping
1452  *   and someone does a writepage() on it.
1453  *
1454  *   We will move the buffer onto the async_data list, but *after* it has
1455  *   been dirtied. So there's a small window where we have dirty data on
1456  *   BJ_Metadata.
1457  *
1458  *   Note that this only applies to the last partial page in the file.  The
1459  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1460  *   broken code anyway: it's wrong for msync()).
1461  *
1462  *   It's a rare case: affects the final partial page, for journalled data
1463  *   where the file is subject to bith write() and writepage() in the same
1464  *   transction.  To fix it we'll need a custom block_write_full_page().
1465  *   We'll probably need that anyway for journalling writepage() output.
1466  *
1467  * We don't honour synchronous mounts for writepage().  That would be
1468  * disastrous.  Any write() or metadata operation will sync the fs for
1469  * us.
1470  *
1471  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1472  * we don't need to open a transaction here.
1473  */
1474 static int ext3_ordered_writepage(struct page *page,
1475                                 struct writeback_control *wbc)
1476 {
1477         struct inode *inode = page->mapping->host;
1478         struct buffer_head *page_bufs;
1479         handle_t *handle = NULL;
1480         int ret = 0;
1481         int err;
1482
1483         J_ASSERT(PageLocked(page));
1484
1485         /*
1486          * We give up here if we're reentered, because it might be for a
1487          * different filesystem.
1488          */
1489         if (ext3_journal_current_handle())
1490                 goto out_fail;
1491
1492         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1493
1494         if (IS_ERR(handle)) {
1495                 ret = PTR_ERR(handle);
1496                 goto out_fail;
1497         }
1498
1499         if (!page_has_buffers(page)) {
1500                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1501                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1502         }
1503         page_bufs = page_buffers(page);
1504         walk_page_buffers(handle, page_bufs, 0,
1505                         PAGE_CACHE_SIZE, NULL, bget_one);
1506
1507         ret = block_write_full_page(page, ext3_get_block, wbc);
1508
1509         /*
1510          * The page can become unlocked at any point now, and
1511          * truncate can then come in and change things.  So we
1512          * can't touch *page from now on.  But *page_bufs is
1513          * safe due to elevated refcount.
1514          */
1515
1516         /*
1517          * And attach them to the current transaction.  But only if
1518          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1519          * and generally junk.
1520          */
1521         if (ret == 0) {
1522                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1523                                         NULL, journal_dirty_data_fn);
1524                 if (!ret)
1525                         ret = err;
1526         }
1527         walk_page_buffers(handle, page_bufs, 0,
1528                         PAGE_CACHE_SIZE, NULL, bput_one);
1529         err = ext3_journal_stop(handle);
1530         if (!ret)
1531                 ret = err;
1532         return ret;
1533
1534 out_fail:
1535         redirty_page_for_writepage(wbc, page);
1536         unlock_page(page);
1537         return ret;
1538 }
1539
1540 static int ext3_writeback_writepage(struct page *page,
1541                                 struct writeback_control *wbc)
1542 {
1543         struct inode *inode = page->mapping->host;
1544         handle_t *handle = NULL;
1545         int ret = 0;
1546         int err;
1547
1548         if (ext3_journal_current_handle())
1549                 goto out_fail;
1550
1551         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1552         if (IS_ERR(handle)) {
1553                 ret = PTR_ERR(handle);
1554                 goto out_fail;
1555         }
1556
1557         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1558                 ret = nobh_writepage(page, ext3_get_block, wbc);
1559         else
1560                 ret = block_write_full_page(page, ext3_get_block, wbc);
1561
1562         err = ext3_journal_stop(handle);
1563         if (!ret)
1564                 ret = err;
1565         return ret;
1566
1567 out_fail:
1568         redirty_page_for_writepage(wbc, page);
1569         unlock_page(page);
1570         return ret;
1571 }
1572
1573 static int ext3_journalled_writepage(struct page *page,
1574                                 struct writeback_control *wbc)
1575 {
1576         struct inode *inode = page->mapping->host;
1577         handle_t *handle = NULL;
1578         int ret = 0;
1579         int err;
1580
1581         if (ext3_journal_current_handle())
1582                 goto no_write;
1583
1584         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1585         if (IS_ERR(handle)) {
1586                 ret = PTR_ERR(handle);
1587                 goto no_write;
1588         }
1589
1590         if (!page_has_buffers(page) || PageChecked(page)) {
1591                 /*
1592                  * It's mmapped pagecache.  Add buffers and journal it.  There
1593                  * doesn't seem much point in redirtying the page here.
1594                  */
1595                 ClearPageChecked(page);
1596                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1597                                         ext3_get_block);
1598                 if (ret != 0) {
1599                         ext3_journal_stop(handle);
1600                         goto out_unlock;
1601                 }
1602                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1603                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1604
1605                 err = walk_page_buffers(handle, page_buffers(page), 0,
1606                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1607                 if (ret == 0)
1608                         ret = err;
1609                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1610                 unlock_page(page);
1611         } else {
1612                 /*
1613                  * It may be a page full of checkpoint-mode buffers.  We don't
1614                  * really know unless we go poke around in the buffer_heads.
1615                  * But block_write_full_page will do the right thing.
1616                  */
1617                 ret = block_write_full_page(page, ext3_get_block, wbc);
1618         }
1619         err = ext3_journal_stop(handle);
1620         if (!ret)
1621                 ret = err;
1622 out:
1623         return ret;
1624
1625 no_write:
1626         redirty_page_for_writepage(wbc, page);
1627 out_unlock:
1628         unlock_page(page);
1629         goto out;
1630 }
1631
1632 static int ext3_readpage(struct file *file, struct page *page)
1633 {
1634         return mpage_readpage(page, ext3_get_block);
1635 }
1636
1637 static int
1638 ext3_readpages(struct file *file, struct address_space *mapping,
1639                 struct list_head *pages, unsigned nr_pages)
1640 {
1641         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1642 }
1643
1644 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1645 {
1646         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1647
1648         /*
1649          * If it's a full truncate we just forget about the pending dirtying
1650          */
1651         if (offset == 0)
1652                 ClearPageChecked(page);
1653
1654         journal_invalidatepage(journal, page, offset);
1655 }
1656
1657 static int ext3_releasepage(struct page *page, gfp_t wait)
1658 {
1659         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1660
1661         WARN_ON(PageChecked(page));
1662         if (!page_has_buffers(page))
1663                 return 0;
1664         return journal_try_to_free_buffers(journal, page, wait);
1665 }
1666
1667 /*
1668  * If the O_DIRECT write will extend the file then add this inode to the
1669  * orphan list.  So recovery will truncate it back to the original size
1670  * if the machine crashes during the write.
1671  *
1672  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1673  * crashes then stale disk data _may_ be exposed inside the file. But current
1674  * VFS code falls back into buffered path in that case so we are safe.
1675  */
1676 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1677                         const struct iovec *iov, loff_t offset,
1678                         unsigned long nr_segs)
1679 {
1680         struct file *file = iocb->ki_filp;
1681         struct inode *inode = file->f_mapping->host;
1682         struct ext3_inode_info *ei = EXT3_I(inode);
1683         handle_t *handle;
1684         ssize_t ret;
1685         int orphan = 0;
1686         size_t count = iov_length(iov, nr_segs);
1687
1688         if (rw == WRITE) {
1689                 loff_t final_size = offset + count;
1690
1691                 if (final_size > inode->i_size) {
1692                         /* Credits for sb + inode write */
1693                         handle = ext3_journal_start(inode, 2);
1694                         if (IS_ERR(handle)) {
1695                                 ret = PTR_ERR(handle);
1696                                 goto out;
1697                         }
1698                         ret = ext3_orphan_add(handle, inode);
1699                         if (ret) {
1700                                 ext3_journal_stop(handle);
1701                                 goto out;
1702                         }
1703                         orphan = 1;
1704                         ei->i_disksize = inode->i_size;
1705                         ext3_journal_stop(handle);
1706                 }
1707         }
1708
1709         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1710                                  offset, nr_segs,
1711                                  ext3_get_block, NULL);
1712
1713         if (orphan) {
1714                 int err;
1715
1716                 /* Credits for sb + inode write */
1717                 handle = ext3_journal_start(inode, 2);
1718                 if (IS_ERR(handle)) {
1719                         /* This is really bad luck. We've written the data
1720                          * but cannot extend i_size. Bail out and pretend
1721                          * the write failed... */
1722                         ret = PTR_ERR(handle);
1723                         goto out;
1724                 }
1725                 if (inode->i_nlink)
1726                         ext3_orphan_del(handle, inode);
1727                 if (ret > 0) {
1728                         loff_t end = offset + ret;
1729                         if (end > inode->i_size) {
1730                                 ei->i_disksize = end;
1731                                 i_size_write(inode, end);
1732                                 /*
1733                                  * We're going to return a positive `ret'
1734                                  * here due to non-zero-length I/O, so there's
1735                                  * no way of reporting error returns from
1736                                  * ext3_mark_inode_dirty() to userspace.  So
1737                                  * ignore it.
1738                                  */
1739                                 ext3_mark_inode_dirty(handle, inode);
1740                         }
1741                 }
1742                 err = ext3_journal_stop(handle);
1743                 if (ret == 0)
1744                         ret = err;
1745         }
1746 out:
1747         return ret;
1748 }
1749
1750 /*
1751  * Pages can be marked dirty completely asynchronously from ext3's journalling
1752  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1753  * much here because ->set_page_dirty is called under VFS locks.  The page is
1754  * not necessarily locked.
1755  *
1756  * We cannot just dirty the page and leave attached buffers clean, because the
1757  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1758  * or jbddirty because all the journalling code will explode.
1759  *
1760  * So what we do is to mark the page "pending dirty" and next time writepage
1761  * is called, propagate that into the buffers appropriately.
1762  */
1763 static int ext3_journalled_set_page_dirty(struct page *page)
1764 {
1765         SetPageChecked(page);
1766         return __set_page_dirty_nobuffers(page);
1767 }
1768
1769 static const struct address_space_operations ext3_ordered_aops = {
1770         .readpage               = ext3_readpage,
1771         .readpages              = ext3_readpages,
1772         .writepage              = ext3_ordered_writepage,
1773         .sync_page              = block_sync_page,
1774         .write_begin            = ext3_write_begin,
1775         .write_end              = ext3_ordered_write_end,
1776         .bmap                   = ext3_bmap,
1777         .invalidatepage         = ext3_invalidatepage,
1778         .releasepage            = ext3_releasepage,
1779         .direct_IO              = ext3_direct_IO,
1780         .migratepage            = buffer_migrate_page,
1781         .is_partially_uptodate  = block_is_partially_uptodate,
1782 };
1783
1784 static const struct address_space_operations ext3_writeback_aops = {
1785         .readpage               = ext3_readpage,
1786         .readpages              = ext3_readpages,
1787         .writepage              = ext3_writeback_writepage,
1788         .sync_page              = block_sync_page,
1789         .write_begin            = ext3_write_begin,
1790         .write_end              = ext3_writeback_write_end,
1791         .bmap                   = ext3_bmap,
1792         .invalidatepage         = ext3_invalidatepage,
1793         .releasepage            = ext3_releasepage,
1794         .direct_IO              = ext3_direct_IO,
1795         .migratepage            = buffer_migrate_page,
1796         .is_partially_uptodate  = block_is_partially_uptodate,
1797 };
1798
1799 static const struct address_space_operations ext3_journalled_aops = {
1800         .readpage               = ext3_readpage,
1801         .readpages              = ext3_readpages,
1802         .writepage              = ext3_journalled_writepage,
1803         .sync_page              = block_sync_page,
1804         .write_begin            = ext3_write_begin,
1805         .write_end              = ext3_journalled_write_end,
1806         .set_page_dirty         = ext3_journalled_set_page_dirty,
1807         .bmap                   = ext3_bmap,
1808         .invalidatepage         = ext3_invalidatepage,
1809         .releasepage            = ext3_releasepage,
1810         .is_partially_uptodate  = block_is_partially_uptodate,
1811 };
1812
1813 void ext3_set_aops(struct inode *inode)
1814 {
1815         if (ext3_should_order_data(inode))
1816                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1817         else if (ext3_should_writeback_data(inode))
1818                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1819         else
1820                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1821 }
1822
1823 /*
1824  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1825  * up to the end of the block which corresponds to `from'.
1826  * This required during truncate. We need to physically zero the tail end
1827  * of that block so it doesn't yield old data if the file is later grown.
1828  */
1829 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1830                 struct address_space *mapping, loff_t from)
1831 {
1832         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1833         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1834         unsigned blocksize, iblock, length, pos;
1835         struct inode *inode = mapping->host;
1836         struct buffer_head *bh;
1837         int err = 0;
1838
1839         blocksize = inode->i_sb->s_blocksize;
1840         length = blocksize - (offset & (blocksize - 1));
1841         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1842
1843         /*
1844          * For "nobh" option,  we can only work if we don't need to
1845          * read-in the page - otherwise we create buffers to do the IO.
1846          */
1847         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1848              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1849                 zero_user(page, offset, length);
1850                 set_page_dirty(page);
1851                 goto unlock;
1852         }
1853
1854         if (!page_has_buffers(page))
1855                 create_empty_buffers(page, blocksize, 0);
1856
1857         /* Find the buffer that contains "offset" */
1858         bh = page_buffers(page);
1859         pos = blocksize;
1860         while (offset >= pos) {
1861                 bh = bh->b_this_page;
1862                 iblock++;
1863                 pos += blocksize;
1864         }
1865
1866         err = 0;
1867         if (buffer_freed(bh)) {
1868                 BUFFER_TRACE(bh, "freed: skip");
1869                 goto unlock;
1870         }
1871
1872         if (!buffer_mapped(bh)) {
1873                 BUFFER_TRACE(bh, "unmapped");
1874                 ext3_get_block(inode, iblock, bh, 0);
1875                 /* unmapped? It's a hole - nothing to do */
1876                 if (!buffer_mapped(bh)) {
1877                         BUFFER_TRACE(bh, "still unmapped");
1878                         goto unlock;
1879                 }
1880         }
1881
1882         /* Ok, it's mapped. Make sure it's up-to-date */
1883         if (PageUptodate(page))
1884                 set_buffer_uptodate(bh);
1885
1886         if (!buffer_uptodate(bh)) {
1887                 err = -EIO;
1888                 ll_rw_block(READ, 1, &bh);
1889                 wait_on_buffer(bh);
1890                 /* Uhhuh. Read error. Complain and punt. */
1891                 if (!buffer_uptodate(bh))
1892                         goto unlock;
1893         }
1894
1895         if (ext3_should_journal_data(inode)) {
1896                 BUFFER_TRACE(bh, "get write access");
1897                 err = ext3_journal_get_write_access(handle, bh);
1898                 if (err)
1899                         goto unlock;
1900         }
1901
1902         zero_user(page, offset, length);
1903         BUFFER_TRACE(bh, "zeroed end of block");
1904
1905         err = 0;
1906         if (ext3_should_journal_data(inode)) {
1907                 err = ext3_journal_dirty_metadata(handle, bh);
1908         } else {
1909                 if (ext3_should_order_data(inode))
1910                         err = ext3_journal_dirty_data(handle, bh);
1911                 mark_buffer_dirty(bh);
1912         }
1913
1914 unlock:
1915         unlock_page(page);
1916         page_cache_release(page);
1917         return err;
1918 }
1919
1920 /*
1921  * Probably it should be a library function... search for first non-zero word
1922  * or memcmp with zero_page, whatever is better for particular architecture.
1923  * Linus?
1924  */
1925 static inline int all_zeroes(__le32 *p, __le32 *q)
1926 {
1927         while (p < q)
1928                 if (*p++)
1929                         return 0;
1930         return 1;
1931 }
1932
1933 /**
1934  *      ext3_find_shared - find the indirect blocks for partial truncation.
1935  *      @inode:   inode in question
1936  *      @depth:   depth of the affected branch
1937  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1938  *      @chain:   place to store the pointers to partial indirect blocks
1939  *      @top:     place to the (detached) top of branch
1940  *
1941  *      This is a helper function used by ext3_truncate().
1942  *
1943  *      When we do truncate() we may have to clean the ends of several
1944  *      indirect blocks but leave the blocks themselves alive. Block is
1945  *      partially truncated if some data below the new i_size is refered
1946  *      from it (and it is on the path to the first completely truncated
1947  *      data block, indeed).  We have to free the top of that path along
1948  *      with everything to the right of the path. Since no allocation
1949  *      past the truncation point is possible until ext3_truncate()
1950  *      finishes, we may safely do the latter, but top of branch may
1951  *      require special attention - pageout below the truncation point
1952  *      might try to populate it.
1953  *
1954  *      We atomically detach the top of branch from the tree, store the
1955  *      block number of its root in *@top, pointers to buffer_heads of
1956  *      partially truncated blocks - in @chain[].bh and pointers to
1957  *      their last elements that should not be removed - in
1958  *      @chain[].p. Return value is the pointer to last filled element
1959  *      of @chain.
1960  *
1961  *      The work left to caller to do the actual freeing of subtrees:
1962  *              a) free the subtree starting from *@top
1963  *              b) free the subtrees whose roots are stored in
1964  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1965  *              c) free the subtrees growing from the inode past the @chain[0].
1966  *                      (no partially truncated stuff there).  */
1967
1968 static Indirect *ext3_find_shared(struct inode *inode, int depth,
1969                         int offsets[4], Indirect chain[4], __le32 *top)
1970 {
1971         Indirect *partial, *p;
1972         int k, err;
1973
1974         *top = 0;
1975         /* Make k index the deepest non-null offest + 1 */
1976         for (k = depth; k > 1 && !offsets[k-1]; k--)
1977                 ;
1978         partial = ext3_get_branch(inode, k, offsets, chain, &err);
1979         /* Writer: pointers */
1980         if (!partial)
1981                 partial = chain + k-1;
1982         /*
1983          * If the branch acquired continuation since we've looked at it -
1984          * fine, it should all survive and (new) top doesn't belong to us.
1985          */
1986         if (!partial->key && *partial->p)
1987                 /* Writer: end */
1988                 goto no_top;
1989         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1990                 ;
1991         /*
1992          * OK, we've found the last block that must survive. The rest of our
1993          * branch should be detached before unlocking. However, if that rest
1994          * of branch is all ours and does not grow immediately from the inode
1995          * it's easier to cheat and just decrement partial->p.
1996          */
1997         if (p == chain + k - 1 && p > chain) {
1998                 p->p--;
1999         } else {
2000                 *top = *p->p;
2001                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2002 #if 0
2003                 *p->p = 0;
2004 #endif
2005         }
2006         /* Writer: end */
2007
2008         while(partial > p) {
2009                 brelse(partial->bh);
2010                 partial--;
2011         }
2012 no_top:
2013         return partial;
2014 }
2015
2016 /*
2017  * Zero a number of block pointers in either an inode or an indirect block.
2018  * If we restart the transaction we must again get write access to the
2019  * indirect block for further modification.
2020  *
2021  * We release `count' blocks on disk, but (last - first) may be greater
2022  * than `count' because there can be holes in there.
2023  */
2024 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2025                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2026                 unsigned long count, __le32 *first, __le32 *last)
2027 {
2028         __le32 *p;
2029         if (try_to_extend_transaction(handle, inode)) {
2030                 if (bh) {
2031                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2032                         ext3_journal_dirty_metadata(handle, bh);
2033                 }
2034                 ext3_mark_inode_dirty(handle, inode);
2035                 ext3_journal_test_restart(handle, inode);
2036                 if (bh) {
2037                         BUFFER_TRACE(bh, "retaking write access");
2038                         ext3_journal_get_write_access(handle, bh);
2039                 }
2040         }
2041
2042         /*
2043          * Any buffers which are on the journal will be in memory. We find
2044          * them on the hash table so journal_revoke() will run journal_forget()
2045          * on them.  We've already detached each block from the file, so
2046          * bforget() in journal_forget() should be safe.
2047          *
2048          * AKPM: turn on bforget in journal_forget()!!!
2049          */
2050         for (p = first; p < last; p++) {
2051                 u32 nr = le32_to_cpu(*p);
2052                 if (nr) {
2053                         struct buffer_head *bh;
2054
2055                         *p = 0;
2056                         bh = sb_find_get_block(inode->i_sb, nr);
2057                         ext3_forget(handle, 0, inode, bh, nr);
2058                 }
2059         }
2060
2061         ext3_free_blocks(handle, inode, block_to_free, count);
2062 }
2063
2064 /**
2065  * ext3_free_data - free a list of data blocks
2066  * @handle:     handle for this transaction
2067  * @inode:      inode we are dealing with
2068  * @this_bh:    indirect buffer_head which contains *@first and *@last
2069  * @first:      array of block numbers
2070  * @last:       points immediately past the end of array
2071  *
2072  * We are freeing all blocks refered from that array (numbers are stored as
2073  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2074  *
2075  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2076  * blocks are contiguous then releasing them at one time will only affect one
2077  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2078  * actually use a lot of journal space.
2079  *
2080  * @this_bh will be %NULL if @first and @last point into the inode's direct
2081  * block pointers.
2082  */
2083 static void ext3_free_data(handle_t *handle, struct inode *inode,
2084                            struct buffer_head *this_bh,
2085                            __le32 *first, __le32 *last)
2086 {
2087         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2088         unsigned long count = 0;            /* Number of blocks in the run */
2089         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2090                                                corresponding to
2091                                                block_to_free */
2092         ext3_fsblk_t nr;                    /* Current block # */
2093         __le32 *p;                          /* Pointer into inode/ind
2094                                                for current block */
2095         int err;
2096
2097         if (this_bh) {                          /* For indirect block */
2098                 BUFFER_TRACE(this_bh, "get_write_access");
2099                 err = ext3_journal_get_write_access(handle, this_bh);
2100                 /* Important: if we can't update the indirect pointers
2101                  * to the blocks, we can't free them. */
2102                 if (err)
2103                         return;
2104         }
2105
2106         for (p = first; p < last; p++) {
2107                 nr = le32_to_cpu(*p);
2108                 if (nr) {
2109                         /* accumulate blocks to free if they're contiguous */
2110                         if (count == 0) {
2111                                 block_to_free = nr;
2112                                 block_to_free_p = p;
2113                                 count = 1;
2114                         } else if (nr == block_to_free + count) {
2115                                 count++;
2116                         } else {
2117                                 ext3_clear_blocks(handle, inode, this_bh,
2118                                                   block_to_free,
2119                                                   count, block_to_free_p, p);
2120                                 block_to_free = nr;
2121                                 block_to_free_p = p;
2122                                 count = 1;
2123                         }
2124                 }
2125         }
2126
2127         if (count > 0)
2128                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2129                                   count, block_to_free_p, p);
2130
2131         if (this_bh) {
2132                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2133
2134                 /*
2135                  * The buffer head should have an attached journal head at this
2136                  * point. However, if the data is corrupted and an indirect
2137                  * block pointed to itself, it would have been detached when
2138                  * the block was cleared. Check for this instead of OOPSing.
2139                  */
2140                 if (bh2jh(this_bh))
2141                         ext3_journal_dirty_metadata(handle, this_bh);
2142                 else
2143                         ext3_error(inode->i_sb, "ext3_free_data",
2144                                    "circular indirect block detected, "
2145                                    "inode=%lu, block=%llu",
2146                                    inode->i_ino,
2147                                    (unsigned long long)this_bh->b_blocknr);
2148         }
2149 }
2150
2151 /**
2152  *      ext3_free_branches - free an array of branches
2153  *      @handle: JBD handle for this transaction
2154  *      @inode: inode we are dealing with
2155  *      @parent_bh: the buffer_head which contains *@first and *@last
2156  *      @first: array of block numbers
2157  *      @last:  pointer immediately past the end of array
2158  *      @depth: depth of the branches to free
2159  *
2160  *      We are freeing all blocks refered from these branches (numbers are
2161  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2162  *      appropriately.
2163  */
2164 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2165                                struct buffer_head *parent_bh,
2166                                __le32 *first, __le32 *last, int depth)
2167 {
2168         ext3_fsblk_t nr;
2169         __le32 *p;
2170
2171         if (is_handle_aborted(handle))
2172                 return;
2173
2174         if (depth--) {
2175                 struct buffer_head *bh;
2176                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2177                 p = last;
2178                 while (--p >= first) {
2179                         nr = le32_to_cpu(*p);
2180                         if (!nr)
2181                                 continue;               /* A hole */
2182
2183                         /* Go read the buffer for the next level down */
2184                         bh = sb_bread(inode->i_sb, nr);
2185
2186                         /*
2187                          * A read failure? Report error and clear slot
2188                          * (should be rare).
2189                          */
2190                         if (!bh) {
2191                                 ext3_error(inode->i_sb, "ext3_free_branches",
2192                                            "Read failure, inode=%lu, block="E3FSBLK,
2193                                            inode->i_ino, nr);
2194                                 continue;
2195                         }
2196
2197                         /* This zaps the entire block.  Bottom up. */
2198                         BUFFER_TRACE(bh, "free child branches");
2199                         ext3_free_branches(handle, inode, bh,
2200                                            (__le32*)bh->b_data,
2201                                            (__le32*)bh->b_data + addr_per_block,
2202                                            depth);
2203
2204                         /*
2205                          * We've probably journalled the indirect block several
2206                          * times during the truncate.  But it's no longer
2207                          * needed and we now drop it from the transaction via
2208                          * journal_revoke().
2209                          *
2210                          * That's easy if it's exclusively part of this
2211                          * transaction.  But if it's part of the committing
2212                          * transaction then journal_forget() will simply
2213                          * brelse() it.  That means that if the underlying
2214                          * block is reallocated in ext3_get_block(),
2215                          * unmap_underlying_metadata() will find this block
2216                          * and will try to get rid of it.  damn, damn.
2217                          *
2218                          * If this block has already been committed to the
2219                          * journal, a revoke record will be written.  And
2220                          * revoke records must be emitted *before* clearing
2221                          * this block's bit in the bitmaps.
2222                          */
2223                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2224
2225                         /*
2226                          * Everything below this this pointer has been
2227                          * released.  Now let this top-of-subtree go.
2228                          *
2229                          * We want the freeing of this indirect block to be
2230                          * atomic in the journal with the updating of the
2231                          * bitmap block which owns it.  So make some room in
2232                          * the journal.
2233                          *
2234                          * We zero the parent pointer *after* freeing its
2235                          * pointee in the bitmaps, so if extend_transaction()
2236                          * for some reason fails to put the bitmap changes and
2237                          * the release into the same transaction, recovery
2238                          * will merely complain about releasing a free block,
2239                          * rather than leaking blocks.
2240                          */
2241                         if (is_handle_aborted(handle))
2242                                 return;
2243                         if (try_to_extend_transaction(handle, inode)) {
2244                                 ext3_mark_inode_dirty(handle, inode);
2245                                 ext3_journal_test_restart(handle, inode);
2246                         }
2247
2248                         ext3_free_blocks(handle, inode, nr, 1);
2249
2250                         if (parent_bh) {
2251                                 /*
2252                                  * The block which we have just freed is
2253                                  * pointed to by an indirect block: journal it
2254                                  */
2255                                 BUFFER_TRACE(parent_bh, "get_write_access");
2256                                 if (!ext3_journal_get_write_access(handle,
2257                                                                    parent_bh)){
2258                                         *p = 0;
2259                                         BUFFER_TRACE(parent_bh,
2260                                         "call ext3_journal_dirty_metadata");
2261                                         ext3_journal_dirty_metadata(handle,
2262                                                                     parent_bh);
2263                                 }
2264                         }
2265                 }
2266         } else {
2267                 /* We have reached the bottom of the tree. */
2268                 BUFFER_TRACE(parent_bh, "free data blocks");
2269                 ext3_free_data(handle, inode, parent_bh, first, last);
2270         }
2271 }
2272
2273 int ext3_can_truncate(struct inode *inode)
2274 {
2275         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2276                 return 0;
2277         if (S_ISREG(inode->i_mode))
2278                 return 1;
2279         if (S_ISDIR(inode->i_mode))
2280                 return 1;
2281         if (S_ISLNK(inode->i_mode))
2282                 return !ext3_inode_is_fast_symlink(inode);
2283         return 0;
2284 }
2285
2286 /*
2287  * ext3_truncate()
2288  *
2289  * We block out ext3_get_block() block instantiations across the entire
2290  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2291  * simultaneously on behalf of the same inode.
2292  *
2293  * As we work through the truncate and commmit bits of it to the journal there
2294  * is one core, guiding principle: the file's tree must always be consistent on
2295  * disk.  We must be able to restart the truncate after a crash.
2296  *
2297  * The file's tree may be transiently inconsistent in memory (although it
2298  * probably isn't), but whenever we close off and commit a journal transaction,
2299  * the contents of (the filesystem + the journal) must be consistent and
2300  * restartable.  It's pretty simple, really: bottom up, right to left (although
2301  * left-to-right works OK too).
2302  *
2303  * Note that at recovery time, journal replay occurs *before* the restart of
2304  * truncate against the orphan inode list.
2305  *
2306  * The committed inode has the new, desired i_size (which is the same as
2307  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2308  * that this inode's truncate did not complete and it will again call
2309  * ext3_truncate() to have another go.  So there will be instantiated blocks
2310  * to the right of the truncation point in a crashed ext3 filesystem.  But
2311  * that's fine - as long as they are linked from the inode, the post-crash
2312  * ext3_truncate() run will find them and release them.
2313  */
2314 void ext3_truncate(struct inode *inode)
2315 {
2316         handle_t *handle;
2317         struct ext3_inode_info *ei = EXT3_I(inode);
2318         __le32 *i_data = ei->i_data;
2319         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2320         struct address_space *mapping = inode->i_mapping;
2321         int offsets[4];
2322         Indirect chain[4];
2323         Indirect *partial;
2324         __le32 nr = 0;
2325         int n;
2326         long last_block;
2327         unsigned blocksize = inode->i_sb->s_blocksize;
2328         struct page *page;
2329
2330         if (!ext3_can_truncate(inode))
2331                 return;
2332
2333         /*
2334          * We have to lock the EOF page here, because lock_page() nests
2335          * outside journal_start().
2336          */
2337         if ((inode->i_size & (blocksize - 1)) == 0) {
2338                 /* Block boundary? Nothing to do */
2339                 page = NULL;
2340         } else {
2341                 page = grab_cache_page(mapping,
2342                                 inode->i_size >> PAGE_CACHE_SHIFT);
2343                 if (!page)
2344                         return;
2345         }
2346
2347         handle = start_transaction(inode);
2348         if (IS_ERR(handle)) {
2349                 if (page) {
2350                         clear_highpage(page);
2351                         flush_dcache_page(page);
2352                         unlock_page(page);
2353                         page_cache_release(page);
2354                 }
2355                 return;         /* AKPM: return what? */
2356         }
2357
2358         last_block = (inode->i_size + blocksize-1)
2359                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2360
2361         if (page)
2362                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2363
2364         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2365         if (n == 0)
2366                 goto out_stop;  /* error */
2367
2368         /*
2369          * OK.  This truncate is going to happen.  We add the inode to the
2370          * orphan list, so that if this truncate spans multiple transactions,
2371          * and we crash, we will resume the truncate when the filesystem
2372          * recovers.  It also marks the inode dirty, to catch the new size.
2373          *
2374          * Implication: the file must always be in a sane, consistent
2375          * truncatable state while each transaction commits.
2376          */
2377         if (ext3_orphan_add(handle, inode))
2378                 goto out_stop;
2379
2380         /*
2381          * The orphan list entry will now protect us from any crash which
2382          * occurs before the truncate completes, so it is now safe to propagate
2383          * the new, shorter inode size (held for now in i_size) into the
2384          * on-disk inode. We do this via i_disksize, which is the value which
2385          * ext3 *really* writes onto the disk inode.
2386          */
2387         ei->i_disksize = inode->i_size;
2388
2389         /*
2390          * From here we block out all ext3_get_block() callers who want to
2391          * modify the block allocation tree.
2392          */
2393         mutex_lock(&ei->truncate_mutex);
2394
2395         if (n == 1) {           /* direct blocks */
2396                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2397                                i_data + EXT3_NDIR_BLOCKS);
2398                 goto do_indirects;
2399         }
2400
2401         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2402         /* Kill the top of shared branch (not detached) */
2403         if (nr) {
2404                 if (partial == chain) {
2405                         /* Shared branch grows from the inode */
2406                         ext3_free_branches(handle, inode, NULL,
2407                                            &nr, &nr+1, (chain+n-1) - partial);
2408                         *partial->p = 0;
2409                         /*
2410                          * We mark the inode dirty prior to restart,
2411                          * and prior to stop.  No need for it here.
2412                          */
2413                 } else {
2414                         /* Shared branch grows from an indirect block */
2415                         BUFFER_TRACE(partial->bh, "get_write_access");
2416                         ext3_free_branches(handle, inode, partial->bh,
2417                                         partial->p,
2418                                         partial->p+1, (chain+n-1) - partial);
2419                 }
2420         }
2421         /* Clear the ends of indirect blocks on the shared branch */
2422         while (partial > chain) {
2423                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2424                                    (__le32*)partial->bh->b_data+addr_per_block,
2425                                    (chain+n-1) - partial);
2426                 BUFFER_TRACE(partial->bh, "call brelse");
2427                 brelse (partial->bh);
2428                 partial--;
2429         }
2430 do_indirects:
2431         /* Kill the remaining (whole) subtrees */
2432         switch (offsets[0]) {
2433         default:
2434                 nr = i_data[EXT3_IND_BLOCK];
2435                 if (nr) {
2436                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2437                         i_data[EXT3_IND_BLOCK] = 0;
2438                 }
2439         case EXT3_IND_BLOCK:
2440                 nr = i_data[EXT3_DIND_BLOCK];
2441                 if (nr) {
2442                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2443                         i_data[EXT3_DIND_BLOCK] = 0;
2444                 }
2445         case EXT3_DIND_BLOCK:
2446                 nr = i_data[EXT3_TIND_BLOCK];
2447                 if (nr) {
2448                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2449                         i_data[EXT3_TIND_BLOCK] = 0;
2450                 }
2451         case EXT3_TIND_BLOCK:
2452                 ;
2453         }
2454
2455         ext3_discard_reservation(inode);
2456
2457         mutex_unlock(&ei->truncate_mutex);
2458         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2459         ext3_mark_inode_dirty(handle, inode);
2460
2461         /*
2462          * In a multi-transaction truncate, we only make the final transaction
2463          * synchronous
2464          */
2465         if (IS_SYNC(inode))
2466                 handle->h_sync = 1;
2467 out_stop:
2468         /*
2469          * If this was a simple ftruncate(), and the file will remain alive
2470          * then we need to clear up the orphan record which we created above.
2471          * However, if this was a real unlink then we were called by
2472          * ext3_delete_inode(), and we allow that function to clean up the
2473          * orphan info for us.
2474          */
2475         if (inode->i_nlink)
2476                 ext3_orphan_del(handle, inode);
2477
2478         ext3_journal_stop(handle);
2479 }
2480
2481 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2482                 unsigned long ino, struct ext3_iloc *iloc)
2483 {
2484         unsigned long block_group;
2485         unsigned long offset;
2486         ext3_fsblk_t block;
2487         struct ext3_group_desc *gdp;
2488
2489         if (!ext3_valid_inum(sb, ino)) {
2490                 /*
2491                  * This error is already checked for in namei.c unless we are
2492                  * looking at an NFS filehandle, in which case no error
2493                  * report is needed
2494                  */
2495                 return 0;
2496         }
2497
2498         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2499         gdp = ext3_get_group_desc(sb, block_group, NULL);
2500         if (!gdp)
2501                 return 0;
2502         /*
2503          * Figure out the offset within the block group inode table
2504          */
2505         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2506                 EXT3_INODE_SIZE(sb);
2507         block = le32_to_cpu(gdp->bg_inode_table) +
2508                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2509
2510         iloc->block_group = block_group;
2511         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2512         return block;
2513 }
2514
2515 /*
2516  * ext3_get_inode_loc returns with an extra refcount against the inode's
2517  * underlying buffer_head on success. If 'in_mem' is true, we have all
2518  * data in memory that is needed to recreate the on-disk version of this
2519  * inode.
2520  */
2521 static int __ext3_get_inode_loc(struct inode *inode,
2522                                 struct ext3_iloc *iloc, int in_mem)
2523 {
2524         ext3_fsblk_t block;
2525         struct buffer_head *bh;
2526
2527         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2528         if (!block)
2529                 return -EIO;
2530
2531         bh = sb_getblk(inode->i_sb, block);
2532         if (!bh) {
2533                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2534                                 "unable to read inode block - "
2535                                 "inode=%lu, block="E3FSBLK,
2536                                  inode->i_ino, block);
2537                 return -EIO;
2538         }
2539         if (!buffer_uptodate(bh)) {
2540                 lock_buffer(bh);
2541
2542                 /*
2543                  * If the buffer has the write error flag, we have failed
2544                  * to write out another inode in the same block.  In this
2545                  * case, we don't have to read the block because we may
2546                  * read the old inode data successfully.
2547                  */
2548                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2549                         set_buffer_uptodate(bh);
2550
2551                 if (buffer_uptodate(bh)) {
2552                         /* someone brought it uptodate while we waited */
2553                         unlock_buffer(bh);
2554                         goto has_buffer;
2555                 }
2556
2557                 /*
2558                  * If we have all information of the inode in memory and this
2559                  * is the only valid inode in the block, we need not read the
2560                  * block.
2561                  */
2562                 if (in_mem) {
2563                         struct buffer_head *bitmap_bh;
2564                         struct ext3_group_desc *desc;
2565                         int inodes_per_buffer;
2566                         int inode_offset, i;
2567                         int block_group;
2568                         int start;
2569
2570                         block_group = (inode->i_ino - 1) /
2571                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2572                         inodes_per_buffer = bh->b_size /
2573                                 EXT3_INODE_SIZE(inode->i_sb);
2574                         inode_offset = ((inode->i_ino - 1) %
2575                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2576                         start = inode_offset & ~(inodes_per_buffer - 1);
2577
2578                         /* Is the inode bitmap in cache? */
2579                         desc = ext3_get_group_desc(inode->i_sb,
2580                                                 block_group, NULL);
2581                         if (!desc)
2582                                 goto make_io;
2583
2584                         bitmap_bh = sb_getblk(inode->i_sb,
2585                                         le32_to_cpu(desc->bg_inode_bitmap));
2586                         if (!bitmap_bh)
2587                                 goto make_io;
2588
2589                         /*
2590                          * If the inode bitmap isn't in cache then the
2591                          * optimisation may end up performing two reads instead
2592                          * of one, so skip it.
2593                          */
2594                         if (!buffer_uptodate(bitmap_bh)) {
2595                                 brelse(bitmap_bh);
2596                                 goto make_io;
2597                         }
2598                         for (i = start; i < start + inodes_per_buffer; i++) {
2599                                 if (i == inode_offset)
2600                                         continue;
2601                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2602                                         break;
2603                         }
2604                         brelse(bitmap_bh);
2605                         if (i == start + inodes_per_buffer) {
2606                                 /* all other inodes are free, so skip I/O */
2607                                 memset(bh->b_data, 0, bh->b_size);
2608                                 set_buffer_uptodate(bh);
2609                                 unlock_buffer(bh);
2610                                 goto has_buffer;
2611                         }
2612                 }
2613
2614 make_io:
2615                 /*
2616                  * There are other valid inodes in the buffer, this inode
2617                  * has in-inode xattrs, or we don't have this inode in memory.
2618                  * Read the block from disk.
2619                  */
2620                 get_bh(bh);
2621                 bh->b_end_io = end_buffer_read_sync;
2622                 submit_bh(READ_META, bh);
2623                 wait_on_buffer(bh);
2624                 if (!buffer_uptodate(bh)) {
2625                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2626                                         "unable to read inode block - "
2627                                         "inode=%lu, block="E3FSBLK,
2628                                         inode->i_ino, block);
2629                         brelse(bh);
2630                         return -EIO;
2631                 }
2632         }
2633 has_buffer:
2634         iloc->bh = bh;
2635         return 0;
2636 }
2637
2638 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2639 {
2640         /* We have all inode data except xattrs in memory here. */
2641         return __ext3_get_inode_loc(inode, iloc,
2642                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2643 }
2644
2645 void ext3_set_inode_flags(struct inode *inode)
2646 {
2647         unsigned int flags = EXT3_I(inode)->i_flags;
2648
2649         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2650         if (flags & EXT3_SYNC_FL)
2651                 inode->i_flags |= S_SYNC;
2652         if (flags & EXT3_APPEND_FL)
2653                 inode->i_flags |= S_APPEND;
2654         if (flags & EXT3_IMMUTABLE_FL)
2655                 inode->i_flags |= S_IMMUTABLE;
2656         if (flags & EXT3_NOATIME_FL)
2657                 inode->i_flags |= S_NOATIME;
2658         if (flags & EXT3_DIRSYNC_FL)
2659                 inode->i_flags |= S_DIRSYNC;
2660 }
2661
2662 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2663 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2664 {
2665         unsigned int flags = ei->vfs_inode.i_flags;
2666
2667         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2668                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2669         if (flags & S_SYNC)
2670                 ei->i_flags |= EXT3_SYNC_FL;
2671         if (flags & S_APPEND)
2672                 ei->i_flags |= EXT3_APPEND_FL;
2673         if (flags & S_IMMUTABLE)
2674                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2675         if (flags & S_NOATIME)
2676                 ei->i_flags |= EXT3_NOATIME_FL;
2677         if (flags & S_DIRSYNC)
2678                 ei->i_flags |= EXT3_DIRSYNC_FL;
2679 }
2680
2681 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2682 {
2683         struct ext3_iloc iloc;
2684         struct ext3_inode *raw_inode;
2685         struct ext3_inode_info *ei;
2686         struct buffer_head *bh;
2687         struct inode *inode;
2688         long ret;
2689         int block;
2690
2691         inode = iget_locked(sb, ino);
2692         if (!inode)
2693                 return ERR_PTR(-ENOMEM);
2694         if (!(inode->i_state & I_NEW))
2695                 return inode;
2696
2697         ei = EXT3_I(inode);
2698 #ifdef CONFIG_EXT3_FS_POSIX_ACL
2699         ei->i_acl = EXT3_ACL_NOT_CACHED;
2700         ei->i_default_acl = EXT3_ACL_NOT_CACHED;
2701 #endif
2702         ei->i_block_alloc_info = NULL;
2703
2704         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2705         if (ret < 0)
2706                 goto bad_inode;
2707         bh = iloc.bh;
2708         raw_inode = ext3_raw_inode(&iloc);
2709         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2710         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2711         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2712         if(!(test_opt (inode->i_sb, NO_UID32))) {
2713                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2714                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2715         }
2716         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2717         inode->i_size = le32_to_cpu(raw_inode->i_size);
2718         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2719         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2720         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2721         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2722
2723         ei->i_state = 0;
2724         ei->i_dir_start_lookup = 0;
2725         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2726         /* We now have enough fields to check if the inode was active or not.
2727          * This is needed because nfsd might try to access dead inodes
2728          * the test is that same one that e2fsck uses
2729          * NeilBrown 1999oct15
2730          */
2731         if (inode->i_nlink == 0) {
2732                 if (inode->i_mode == 0 ||
2733                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2734                         /* this inode is deleted */
2735                         brelse (bh);
2736                         ret = -ESTALE;
2737                         goto bad_inode;
2738                 }
2739                 /* The only unlinked inodes we let through here have
2740                  * valid i_mode and are being read by the orphan
2741                  * recovery code: that's fine, we're about to complete
2742                  * the process of deleting those. */
2743         }
2744         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2745         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2746 #ifdef EXT3_FRAGMENTS
2747         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2748         ei->i_frag_no = raw_inode->i_frag;
2749         ei->i_frag_size = raw_inode->i_fsize;
2750 #endif
2751         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2752         if (!S_ISREG(inode->i_mode)) {
2753                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2754         } else {
2755                 inode->i_size |=
2756                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2757         }
2758         ei->i_disksize = inode->i_size;
2759         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2760         ei->i_block_group = iloc.block_group;
2761         /*
2762          * NOTE! The in-memory inode i_data array is in little-endian order
2763          * even on big-endian machines: we do NOT byteswap the block numbers!
2764          */
2765         for (block = 0; block < EXT3_N_BLOCKS; block++)
2766                 ei->i_data[block] = raw_inode->i_block[block];
2767         INIT_LIST_HEAD(&ei->i_orphan);
2768
2769         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2770             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2771                 /*
2772                  * When mke2fs creates big inodes it does not zero out
2773                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2774                  * so ignore those first few inodes.
2775                  */
2776                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2777                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2778                     EXT3_INODE_SIZE(inode->i_sb)) {
2779                         brelse (bh);
2780                         ret = -EIO;
2781                         goto bad_inode;
2782                 }
2783                 if (ei->i_extra_isize == 0) {
2784                         /* The extra space is currently unused. Use it. */
2785                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2786                                             EXT3_GOOD_OLD_INODE_SIZE;
2787                 } else {
2788                         __le32 *magic = (void *)raw_inode +
2789                                         EXT3_GOOD_OLD_INODE_SIZE +
2790                                         ei->i_extra_isize;
2791                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2792                                  ei->i_state |= EXT3_STATE_XATTR;
2793                 }
2794         } else
2795                 ei->i_extra_isize = 0;
2796
2797         if (S_ISREG(inode->i_mode)) {
2798                 inode->i_op = &ext3_file_inode_operations;
2799                 inode->i_fop = &ext3_file_operations;
2800                 ext3_set_aops(inode);
2801         } else if (S_ISDIR(inode->i_mode)) {
2802                 inode->i_op = &ext3_dir_inode_operations;
2803                 inode->i_fop = &ext3_dir_operations;
2804         } else if (S_ISLNK(inode->i_mode)) {
2805                 if (ext3_inode_is_fast_symlink(inode))
2806                         inode->i_op = &ext3_fast_symlink_inode_operations;
2807                 else {
2808                         inode->i_op = &ext3_symlink_inode_operations;
2809                         ext3_set_aops(inode);
2810                 }
2811         } else {
2812                 inode->i_op = &ext3_special_inode_operations;
2813                 if (raw_inode->i_block[0])
2814                         init_special_inode(inode, inode->i_mode,
2815                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2816                 else
2817                         init_special_inode(inode, inode->i_mode,
2818                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2819         }
2820         brelse (iloc.bh);
2821         ext3_set_inode_flags(inode);
2822         unlock_new_inode(inode);
2823         return inode;
2824
2825 bad_inode:
2826         iget_failed(inode);
2827         return ERR_PTR(ret);
2828 }
2829
2830 /*
2831  * Post the struct inode info into an on-disk inode location in the
2832  * buffer-cache.  This gobbles the caller's reference to the
2833  * buffer_head in the inode location struct.
2834  *
2835  * The caller must have write access to iloc->bh.
2836  */
2837 static int ext3_do_update_inode(handle_t *handle,
2838                                 struct inode *inode,
2839                                 struct ext3_iloc *iloc)
2840 {
2841         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2842         struct ext3_inode_info *ei = EXT3_I(inode);
2843         struct buffer_head *bh = iloc->bh;
2844         int err = 0, rc, block;
2845
2846         /* For fields not not tracking in the in-memory inode,
2847          * initialise them to zero for new inodes. */
2848         if (ei->i_state & EXT3_STATE_NEW)
2849                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2850
2851         ext3_get_inode_flags(ei);
2852         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2853         if(!(test_opt(inode->i_sb, NO_UID32))) {
2854                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2855                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2856 /*
2857  * Fix up interoperability with old kernels. Otherwise, old inodes get
2858  * re-used with the upper 16 bits of the uid/gid intact
2859  */
2860                 if(!ei->i_dtime) {
2861                         raw_inode->i_uid_high =
2862                                 cpu_to_le16(high_16_bits(inode->i_uid));
2863                         raw_inode->i_gid_high =
2864                                 cpu_to_le16(high_16_bits(inode->i_gid));
2865                 } else {
2866                         raw_inode->i_uid_high = 0;
2867                         raw_inode->i_gid_high = 0;
2868                 }
2869         } else {
2870                 raw_inode->i_uid_low =
2871                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2872                 raw_inode->i_gid_low =
2873                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2874                 raw_inode->i_uid_high = 0;
2875                 raw_inode->i_gid_high = 0;
2876         }
2877         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2878         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2879         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2880         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2881         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2882         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2883         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2884         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2885 #ifdef EXT3_FRAGMENTS
2886         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2887         raw_inode->i_frag = ei->i_frag_no;
2888         raw_inode->i_fsize = ei->i_frag_size;
2889 #endif
2890         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2891         if (!S_ISREG(inode->i_mode)) {
2892                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2893         } else {
2894                 raw_inode->i_size_high =
2895                         cpu_to_le32(ei->i_disksize >> 32);
2896                 if (ei->i_disksize > 0x7fffffffULL) {
2897                         struct super_block *sb = inode->i_sb;
2898                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2899                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2900                             EXT3_SB(sb)->s_es->s_rev_level ==
2901                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2902                                /* If this is the first large file
2903                                 * created, add a flag to the superblock.
2904                                 */
2905                                 err = ext3_journal_get_write_access(handle,
2906                                                 EXT3_SB(sb)->s_sbh);
2907                                 if (err)
2908                                         goto out_brelse;
2909                                 ext3_update_dynamic_rev(sb);
2910                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2911                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2912                                 sb->s_dirt = 1;
2913                                 handle->h_sync = 1;
2914                                 err = ext3_journal_dirty_metadata(handle,
2915                                                 EXT3_SB(sb)->s_sbh);
2916                         }
2917                 }
2918         }
2919         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2920         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2921                 if (old_valid_dev(inode->i_rdev)) {
2922                         raw_inode->i_block[0] =
2923                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2924                         raw_inode->i_block[1] = 0;
2925                 } else {
2926                         raw_inode->i_block[0] = 0;
2927                         raw_inode->i_block[1] =
2928                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2929                         raw_inode->i_block[2] = 0;
2930                 }
2931         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2932                 raw_inode->i_block[block] = ei->i_data[block];
2933
2934         if (ei->i_extra_isize)
2935                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2936
2937         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2938         rc = ext3_journal_dirty_metadata(handle, bh);
2939         if (!err)
2940                 err = rc;
2941         ei->i_state &= ~EXT3_STATE_NEW;
2942
2943 out_brelse:
2944         brelse (bh);
2945         ext3_std_error(inode->i_sb, err);
2946         return err;
2947 }
2948
2949 /*
2950  * ext3_write_inode()
2951  *
2952  * We are called from a few places:
2953  *
2954  * - Within generic_file_write() for O_SYNC files.
2955  *   Here, there will be no transaction running. We wait for any running
2956  *   trasnaction to commit.
2957  *
2958  * - Within sys_sync(), kupdate and such.
2959  *   We wait on commit, if tol to.
2960  *
2961  * - Within prune_icache() (PF_MEMALLOC == true)
2962  *   Here we simply return.  We can't afford to block kswapd on the
2963  *   journal commit.
2964  *
2965  * In all cases it is actually safe for us to return without doing anything,
2966  * because the inode has been copied into a raw inode buffer in
2967  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
2968  * knfsd.
2969  *
2970  * Note that we are absolutely dependent upon all inode dirtiers doing the
2971  * right thing: they *must* call mark_inode_dirty() after dirtying info in
2972  * which we are interested.
2973  *
2974  * It would be a bug for them to not do this.  The code:
2975  *
2976  *      mark_inode_dirty(inode)
2977  *      stuff();
2978  *      inode->i_size = expr;
2979  *
2980  * is in error because a kswapd-driven write_inode() could occur while
2981  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
2982  * will no longer be on the superblock's dirty inode list.
2983  */
2984 int ext3_write_inode(struct inode *inode, int wait)
2985 {
2986         if (current->flags & PF_MEMALLOC)
2987                 return 0;
2988
2989         if (ext3_journal_current_handle()) {
2990                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
2991                 dump_stack();
2992                 return -EIO;
2993         }
2994
2995         if (!wait)
2996                 return 0;
2997
2998         return ext3_force_commit(inode->i_sb);
2999 }
3000
3001 /*
3002  * ext3_setattr()
3003  *
3004  * Called from notify_change.
3005  *
3006  * We want to trap VFS attempts to truncate the file as soon as
3007  * possible.  In particular, we want to make sure that when the VFS
3008  * shrinks i_size, we put the inode on the orphan list and modify
3009  * i_disksize immediately, so that during the subsequent flushing of
3010  * dirty pages and freeing of disk blocks, we can guarantee that any
3011  * commit will leave the blocks being flushed in an unused state on
3012  * disk.  (On recovery, the inode will get truncated and the blocks will
3013  * be freed, so we have a strong guarantee that no future commit will
3014  * leave these blocks visible to the user.)
3015  *
3016  * Called with inode->sem down.
3017  */
3018 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3019 {
3020         struct inode *inode = dentry->d_inode;
3021         int error, rc = 0;
3022         const unsigned int ia_valid = attr->ia_valid;
3023
3024         error = inode_change_ok(inode, attr);
3025         if (error)
3026                 return error;
3027
3028         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3029                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3030                 handle_t *handle;
3031
3032                 /* (user+group)*(old+new) structure, inode write (sb,
3033                  * inode block, ? - but truncate inode update has it) */
3034                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3035                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3036                 if (IS_ERR(handle)) {
3037                         error = PTR_ERR(handle);
3038                         goto err_out;
3039                 }
3040                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
3041                 if (error) {
3042                         ext3_journal_stop(handle);
3043                         return error;
3044                 }
3045                 /* Update corresponding info in inode so that everything is in
3046                  * one transaction */
3047                 if (attr->ia_valid & ATTR_UID)
3048                         inode->i_uid = attr->ia_uid;
3049                 if (attr->ia_valid & ATTR_GID)
3050                         inode->i_gid = attr->ia_gid;
3051                 error = ext3_mark_inode_dirty(handle, inode);
3052                 ext3_journal_stop(handle);
3053         }
3054
3055         if (S_ISREG(inode->i_mode) &&
3056             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3057                 handle_t *handle;
3058
3059                 handle = ext3_journal_start(inode, 3);
3060                 if (IS_ERR(handle)) {
3061                         error = PTR_ERR(handle);
3062                         goto err_out;
3063                 }
3064
3065                 error = ext3_orphan_add(handle, inode);
3066                 EXT3_I(inode)->i_disksize = attr->ia_size;
3067                 rc = ext3_mark_inode_dirty(handle, inode);
3068                 if (!error)
3069                         error = rc;
3070                 ext3_journal_stop(handle);
3071         }
3072
3073         rc = inode_setattr(inode, attr);
3074
3075         /* If inode_setattr's call to ext3_truncate failed to get a
3076          * transaction handle at all, we need to clean up the in-core
3077          * orphan list manually. */
3078         if (inode->i_nlink)
3079                 ext3_orphan_del(NULL, inode);
3080
3081         if (!rc && (ia_valid & ATTR_MODE))
3082                 rc = ext3_acl_chmod(inode);
3083
3084 err_out:
3085         ext3_std_error(inode->i_sb, error);
3086         if (!error)
3087                 error = rc;
3088         return error;
3089 }
3090
3091
3092 /*
3093  * How many blocks doth make a writepage()?
3094  *
3095  * With N blocks per page, it may be:
3096  * N data blocks
3097  * 2 indirect block
3098  * 2 dindirect
3099  * 1 tindirect
3100  * N+5 bitmap blocks (from the above)
3101  * N+5 group descriptor summary blocks
3102  * 1 inode block
3103  * 1 superblock.
3104  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3105  *
3106  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3107  *
3108  * With ordered or writeback data it's the same, less the N data blocks.
3109  *
3110  * If the inode's direct blocks can hold an integral number of pages then a
3111  * page cannot straddle two indirect blocks, and we can only touch one indirect
3112  * and dindirect block, and the "5" above becomes "3".
3113  *
3114  * This still overestimates under most circumstances.  If we were to pass the
3115  * start and end offsets in here as well we could do block_to_path() on each
3116  * block and work out the exact number of indirects which are touched.  Pah.
3117  */
3118
3119 static int ext3_writepage_trans_blocks(struct inode *inode)
3120 {
3121         int bpp = ext3_journal_blocks_per_page(inode);
3122         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3123         int ret;
3124
3125         if (ext3_should_journal_data(inode))
3126                 ret = 3 * (bpp + indirects) + 2;
3127         else
3128                 ret = 2 * (bpp + indirects) + 2;
3129
3130 #ifdef CONFIG_QUOTA
3131         /* We know that structure was already allocated during DQUOT_INIT so
3132          * we will be updating only the data blocks + inodes */
3133         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3134 #endif
3135
3136         return ret;
3137 }
3138
3139 /*
3140  * The caller must have previously called ext3_reserve_inode_write().
3141  * Give this, we know that the caller already has write access to iloc->bh.
3142  */
3143 int ext3_mark_iloc_dirty(handle_t *handle,
3144                 struct inode *inode, struct ext3_iloc *iloc)
3145 {
3146         int err = 0;
3147
3148         /* the do_update_inode consumes one bh->b_count */
3149         get_bh(iloc->bh);
3150
3151         /* ext3_do_update_inode() does journal_dirty_metadata */
3152         err = ext3_do_update_inode(handle, inode, iloc);
3153         put_bh(iloc->bh);
3154         return err;
3155 }
3156
3157 /*
3158  * On success, We end up with an outstanding reference count against
3159  * iloc->bh.  This _must_ be cleaned up later.
3160  */
3161
3162 int
3163 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3164                          struct ext3_iloc *iloc)
3165 {
3166         int err = 0;
3167         if (handle) {
3168                 err = ext3_get_inode_loc(inode, iloc);
3169                 if (!err) {
3170                         BUFFER_TRACE(iloc->bh, "get_write_access");
3171                         err = ext3_journal_get_write_access(handle, iloc->bh);
3172                         if (err) {
3173                                 brelse(iloc->bh);
3174                                 iloc->bh = NULL;
3175                         }
3176                 }
3177         }
3178         ext3_std_error(inode->i_sb, err);
3179         return err;
3180 }
3181
3182 /*
3183  * What we do here is to mark the in-core inode as clean with respect to inode
3184  * dirtiness (it may still be data-dirty).
3185  * This means that the in-core inode may be reaped by prune_icache
3186  * without having to perform any I/O.  This is a very good thing,
3187  * because *any* task may call prune_icache - even ones which
3188  * have a transaction open against a different journal.
3189  *
3190  * Is this cheating?  Not really.  Sure, we haven't written the
3191  * inode out, but prune_icache isn't a user-visible syncing function.
3192  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3193  * we start and wait on commits.
3194  *
3195  * Is this efficient/effective?  Well, we're being nice to the system
3196  * by cleaning up our inodes proactively so they can be reaped
3197  * without I/O.  But we are potentially leaving up to five seconds'
3198  * worth of inodes floating about which prune_icache wants us to
3199  * write out.  One way to fix that would be to get prune_icache()
3200  * to do a write_super() to free up some memory.  It has the desired
3201  * effect.
3202  */
3203 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3204 {
3205         struct ext3_iloc iloc;
3206         int err;
3207
3208         might_sleep();
3209         err = ext3_reserve_inode_write(handle, inode, &iloc);
3210         if (!err)
3211                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3212         return err;
3213 }
3214
3215 /*
3216  * ext3_dirty_inode() is called from __mark_inode_dirty()
3217  *
3218  * We're really interested in the case where a file is being extended.
3219  * i_size has been changed by generic_commit_write() and we thus need
3220  * to include the updated inode in the current transaction.
3221  *
3222  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3223  * are allocated to the file.
3224  *
3225  * If the inode is marked synchronous, we don't honour that here - doing
3226  * so would cause a commit on atime updates, which we don't bother doing.
3227  * We handle synchronous inodes at the highest possible level.
3228  */
3229 void ext3_dirty_inode(struct inode *inode)
3230 {
3231         handle_t *current_handle = ext3_journal_current_handle();
3232         handle_t *handle;
3233
3234         handle = ext3_journal_start(inode, 2);
3235         if (IS_ERR(handle))
3236                 goto out;
3237         if (current_handle &&
3238                 current_handle->h_transaction != handle->h_transaction) {
3239                 /* This task has a transaction open against a different fs */
3240                 printk(KERN_EMERG "%s: transactions do not match!\n",
3241                        __func__);
3242         } else {
3243                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3244                                 current_handle);
3245                 ext3_mark_inode_dirty(handle, inode);
3246         }
3247         ext3_journal_stop(handle);
3248 out:
3249         return;
3250 }
3251
3252 #if 0
3253 /*
3254  * Bind an inode's backing buffer_head into this transaction, to prevent
3255  * it from being flushed to disk early.  Unlike
3256  * ext3_reserve_inode_write, this leaves behind no bh reference and
3257  * returns no iloc structure, so the caller needs to repeat the iloc
3258  * lookup to mark the inode dirty later.
3259  */
3260 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3261 {
3262         struct ext3_iloc iloc;
3263
3264         int err = 0;
3265         if (handle) {
3266                 err = ext3_get_inode_loc(inode, &iloc);
3267                 if (!err) {
3268                         BUFFER_TRACE(iloc.bh, "get_write_access");
3269                         err = journal_get_write_access(handle, iloc.bh);
3270                         if (!err)
3271                                 err = ext3_journal_dirty_metadata(handle,
3272                                                                   iloc.bh);
3273                         brelse(iloc.bh);
3274                 }
3275         }
3276         ext3_std_error(inode->i_sb, err);
3277         return err;
3278 }
3279 #endif
3280
3281 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3282 {
3283         journal_t *journal;
3284         handle_t *handle;
3285         int err;
3286
3287         /*
3288          * We have to be very careful here: changing a data block's
3289          * journaling status dynamically is dangerous.  If we write a
3290          * data block to the journal, change the status and then delete
3291          * that block, we risk forgetting to revoke the old log record
3292          * from the journal and so a subsequent replay can corrupt data.
3293          * So, first we make sure that the journal is empty and that
3294          * nobody is changing anything.
3295          */
3296
3297         journal = EXT3_JOURNAL(inode);
3298         if (is_journal_aborted(journal))
3299                 return -EROFS;
3300
3301         journal_lock_updates(journal);
3302         journal_flush(journal);
3303
3304         /*
3305          * OK, there are no updates running now, and all cached data is
3306          * synced to disk.  We are now in a completely consistent state
3307          * which doesn't have anything in the journal, and we know that
3308          * no filesystem updates are running, so it is safe to modify
3309          * the inode's in-core data-journaling state flag now.
3310          */
3311
3312         if (val)
3313                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3314         else
3315                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3316         ext3_set_aops(inode);
3317
3318         journal_unlock_updates(journal);
3319
3320         /* Finally we can mark the inode as dirty. */
3321
3322         handle = ext3_journal_start(inode, 1);
3323         if (IS_ERR(handle))
3324                 return PTR_ERR(handle);
3325
3326         err = ext3_mark_inode_dirty(handle, inode);
3327         handle->h_sync = 1;
3328         ext3_journal_stop(handle);
3329         ext3_std_error(inode->i_sb, err);
3330
3331         return err;
3332 }