Merge branch 'agp-patches' of git://git.kernel.org/pub/scm/linux/kernel/git/airlied...
[linux-2.6] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/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 ext4_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/ext4_jbd2.h>
29 #include <linux/jbd2.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 /*
43  * Test whether an inode is a fast symlink.
44  */
45 static int ext4_inode_is_fast_symlink(struct inode *inode)
46 {
47         int ea_blocks = EXT4_I(inode)->i_file_acl ?
48                 (inode->i_sb->s_blocksize >> 9) : 0;
49
50         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
51 }
52
53 /*
54  * The ext4 forget function must perform a revoke if we are freeing data
55  * which has been journaled.  Metadata (eg. indirect blocks) must be
56  * revoked in all cases.
57  *
58  * "bh" may be NULL: a metadata block may have been freed from memory
59  * but there may still be a record of it in the journal, and that record
60  * still needs to be revoked.
61  */
62 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
63                         struct buffer_head *bh, ext4_fsblk_t blocknr)
64 {
65         int err;
66
67         might_sleep();
68
69         BUFFER_TRACE(bh, "enter");
70
71         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
72                   "data mode %lx\n",
73                   bh, is_metadata, inode->i_mode,
74                   test_opt(inode->i_sb, DATA_FLAGS));
75
76         /* Never use the revoke function if we are doing full data
77          * journaling: there is no need to, and a V1 superblock won't
78          * support it.  Otherwise, only skip the revoke on un-journaled
79          * data blocks. */
80
81         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
82             (!is_metadata && !ext4_should_journal_data(inode))) {
83                 if (bh) {
84                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
85                         return ext4_journal_forget(handle, bh);
86                 }
87                 return 0;
88         }
89
90         /*
91          * data!=journal && (is_metadata || should_journal_data(inode))
92          */
93         BUFFER_TRACE(bh, "call ext4_journal_revoke");
94         err = ext4_journal_revoke(handle, blocknr, bh);
95         if (err)
96                 ext4_abort(inode->i_sb, __FUNCTION__,
97                            "error %d when attempting revoke", err);
98         BUFFER_TRACE(bh, "exit");
99         return err;
100 }
101
102 /*
103  * Work out how many blocks we need to proceed with the next chunk of a
104  * truncate transaction.
105  */
106 static unsigned long blocks_for_truncate(struct inode *inode)
107 {
108         ext4_lblk_t needed;
109
110         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
111
112         /* Give ourselves just enough room to cope with inodes in which
113          * i_blocks is corrupt: we've seen disk corruptions in the past
114          * which resulted in random data in an inode which looked enough
115          * like a regular file for ext4 to try to delete it.  Things
116          * will go a bit crazy if that happens, but at least we should
117          * try not to panic the whole kernel. */
118         if (needed < 2)
119                 needed = 2;
120
121         /* But we need to bound the transaction so we don't overflow the
122          * journal. */
123         if (needed > EXT4_MAX_TRANS_DATA)
124                 needed = EXT4_MAX_TRANS_DATA;
125
126         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
127 }
128
129 /*
130  * Truncate transactions can be complex and absolutely huge.  So we need to
131  * be able to restart the transaction at a conventient checkpoint to make
132  * sure we don't overflow the journal.
133  *
134  * start_transaction gets us a new handle for a truncate transaction,
135  * and extend_transaction tries to extend the existing one a bit.  If
136  * extend fails, we need to propagate the failure up and restart the
137  * transaction in the top-level truncate loop. --sct
138  */
139 static handle_t *start_transaction(struct inode *inode)
140 {
141         handle_t *result;
142
143         result = ext4_journal_start(inode, blocks_for_truncate(inode));
144         if (!IS_ERR(result))
145                 return result;
146
147         ext4_std_error(inode->i_sb, PTR_ERR(result));
148         return result;
149 }
150
151 /*
152  * Try to extend this transaction for the purposes of truncation.
153  *
154  * Returns 0 if we managed to create more room.  If we can't create more
155  * room, and the transaction must be restarted we return 1.
156  */
157 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
158 {
159         if (handle->h_buffer_credits > EXT4_RESERVE_TRANS_BLOCKS)
160                 return 0;
161         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
162                 return 0;
163         return 1;
164 }
165
166 /*
167  * Restart the transaction associated with *handle.  This does a commit,
168  * so before we call here everything must be consistently dirtied against
169  * this transaction.
170  */
171 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
172 {
173         jbd_debug(2, "restarting handle %p\n", handle);
174         return ext4_journal_restart(handle, blocks_for_truncate(inode));
175 }
176
177 /*
178  * Called at the last iput() if i_nlink is zero.
179  */
180 void ext4_delete_inode (struct inode * inode)
181 {
182         handle_t *handle;
183
184         truncate_inode_pages(&inode->i_data, 0);
185
186         if (is_bad_inode(inode))
187                 goto no_delete;
188
189         handle = start_transaction(inode);
190         if (IS_ERR(handle)) {
191                 /*
192                  * If we're going to skip the normal cleanup, we still need to
193                  * make sure that the in-core orphan linked list is properly
194                  * cleaned up.
195                  */
196                 ext4_orphan_del(NULL, inode);
197                 goto no_delete;
198         }
199
200         if (IS_SYNC(inode))
201                 handle->h_sync = 1;
202         inode->i_size = 0;
203         if (inode->i_blocks)
204                 ext4_truncate(inode);
205         /*
206          * Kill off the orphan record which ext4_truncate created.
207          * AKPM: I think this can be inside the above `if'.
208          * Note that ext4_orphan_del() has to be able to cope with the
209          * deletion of a non-existent orphan - this is because we don't
210          * know if ext4_truncate() actually created an orphan record.
211          * (Well, we could do this if we need to, but heck - it works)
212          */
213         ext4_orphan_del(handle, inode);
214         EXT4_I(inode)->i_dtime  = get_seconds();
215
216         /*
217          * One subtle ordering requirement: if anything has gone wrong
218          * (transaction abort, IO errors, whatever), then we can still
219          * do these next steps (the fs will already have been marked as
220          * having errors), but we can't free the inode if the mark_dirty
221          * fails.
222          */
223         if (ext4_mark_inode_dirty(handle, inode))
224                 /* If that failed, just do the required in-core inode clear. */
225                 clear_inode(inode);
226         else
227                 ext4_free_inode(handle, inode);
228         ext4_journal_stop(handle);
229         return;
230 no_delete:
231         clear_inode(inode);     /* We must guarantee clearing of inode... */
232 }
233
234 typedef struct {
235         __le32  *p;
236         __le32  key;
237         struct buffer_head *bh;
238 } Indirect;
239
240 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
241 {
242         p->key = *(p->p = v);
243         p->bh = bh;
244 }
245
246 /**
247  *      ext4_block_to_path - parse the block number into array of offsets
248  *      @inode: inode in question (we are only interested in its superblock)
249  *      @i_block: block number to be parsed
250  *      @offsets: array to store the offsets in
251  *      @boundary: set this non-zero if the referred-to block is likely to be
252  *             followed (on disk) by an indirect block.
253  *
254  *      To store the locations of file's data ext4 uses a data structure common
255  *      for UNIX filesystems - tree of pointers anchored in the inode, with
256  *      data blocks at leaves and indirect blocks in intermediate nodes.
257  *      This function translates the block number into path in that tree -
258  *      return value is the path length and @offsets[n] is the offset of
259  *      pointer to (n+1)th node in the nth one. If @block is out of range
260  *      (negative or too large) warning is printed and zero returned.
261  *
262  *      Note: function doesn't find node addresses, so no IO is needed. All
263  *      we need to know is the capacity of indirect blocks (taken from the
264  *      inode->i_sb).
265  */
266
267 /*
268  * Portability note: the last comparison (check that we fit into triple
269  * indirect block) is spelled differently, because otherwise on an
270  * architecture with 32-bit longs and 8Kb pages we might get into trouble
271  * if our filesystem had 8Kb blocks. We might use long long, but that would
272  * kill us on x86. Oh, well, at least the sign propagation does not matter -
273  * i_block would have to be negative in the very beginning, so we would not
274  * get there at all.
275  */
276
277 static int ext4_block_to_path(struct inode *inode,
278                         ext4_lblk_t i_block,
279                         ext4_lblk_t offsets[4], int *boundary)
280 {
281         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
282         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
283         const long direct_blocks = EXT4_NDIR_BLOCKS,
284                 indirect_blocks = ptrs,
285                 double_blocks = (1 << (ptrs_bits * 2));
286         int n = 0;
287         int final = 0;
288
289         if (i_block < 0) {
290                 ext4_warning (inode->i_sb, "ext4_block_to_path", "block < 0");
291         } else if (i_block < direct_blocks) {
292                 offsets[n++] = i_block;
293                 final = direct_blocks;
294         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
295                 offsets[n++] = EXT4_IND_BLOCK;
296                 offsets[n++] = i_block;
297                 final = ptrs;
298         } else if ((i_block -= indirect_blocks) < double_blocks) {
299                 offsets[n++] = EXT4_DIND_BLOCK;
300                 offsets[n++] = i_block >> ptrs_bits;
301                 offsets[n++] = i_block & (ptrs - 1);
302                 final = ptrs;
303         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
304                 offsets[n++] = EXT4_TIND_BLOCK;
305                 offsets[n++] = i_block >> (ptrs_bits * 2);
306                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
307                 offsets[n++] = i_block & (ptrs - 1);
308                 final = ptrs;
309         } else {
310                 ext4_warning(inode->i_sb, "ext4_block_to_path",
311                                 "block %lu > max",
312                                 i_block + direct_blocks +
313                                 indirect_blocks + double_blocks);
314         }
315         if (boundary)
316                 *boundary = final - 1 - (i_block & (ptrs - 1));
317         return n;
318 }
319
320 /**
321  *      ext4_get_branch - read the chain of indirect blocks leading to data
322  *      @inode: inode in question
323  *      @depth: depth of the chain (1 - direct pointer, etc.)
324  *      @offsets: offsets of pointers in inode/indirect blocks
325  *      @chain: place to store the result
326  *      @err: here we store the error value
327  *
328  *      Function fills the array of triples <key, p, bh> and returns %NULL
329  *      if everything went OK or the pointer to the last filled triple
330  *      (incomplete one) otherwise. Upon the return chain[i].key contains
331  *      the number of (i+1)-th block in the chain (as it is stored in memory,
332  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
333  *      number (it points into struct inode for i==0 and into the bh->b_data
334  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
335  *      block for i>0 and NULL for i==0. In other words, it holds the block
336  *      numbers of the chain, addresses they were taken from (and where we can
337  *      verify that chain did not change) and buffer_heads hosting these
338  *      numbers.
339  *
340  *      Function stops when it stumbles upon zero pointer (absent block)
341  *              (pointer to last triple returned, *@err == 0)
342  *      or when it gets an IO error reading an indirect block
343  *              (ditto, *@err == -EIO)
344  *      or when it reads all @depth-1 indirect blocks successfully and finds
345  *      the whole chain, all way to the data (returns %NULL, *err == 0).
346  *
347  *      Need to be called with
348  *      down_read(&EXT4_I(inode)->i_data_sem)
349  */
350 static Indirect *ext4_get_branch(struct inode *inode, int depth,
351                                  ext4_lblk_t  *offsets,
352                                  Indirect chain[4], int *err)
353 {
354         struct super_block *sb = inode->i_sb;
355         Indirect *p = chain;
356         struct buffer_head *bh;
357
358         *err = 0;
359         /* i_data is not going away, no lock needed */
360         add_chain (chain, NULL, EXT4_I(inode)->i_data + *offsets);
361         if (!p->key)
362                 goto no_block;
363         while (--depth) {
364                 bh = sb_bread(sb, le32_to_cpu(p->key));
365                 if (!bh)
366                         goto failure;
367                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
368                 /* Reader: end */
369                 if (!p->key)
370                         goto no_block;
371         }
372         return NULL;
373
374 failure:
375         *err = -EIO;
376 no_block:
377         return p;
378 }
379
380 /**
381  *      ext4_find_near - find a place for allocation with sufficient locality
382  *      @inode: owner
383  *      @ind: descriptor of indirect block.
384  *
385  *      This function returns the prefered place for block allocation.
386  *      It is used when heuristic for sequential allocation fails.
387  *      Rules are:
388  *        + if there is a block to the left of our position - allocate near it.
389  *        + if pointer will live in indirect block - allocate near that block.
390  *        + if pointer will live in inode - allocate in the same
391  *          cylinder group.
392  *
393  * In the latter case we colour the starting block by the callers PID to
394  * prevent it from clashing with concurrent allocations for a different inode
395  * in the same block group.   The PID is used here so that functionally related
396  * files will be close-by on-disk.
397  *
398  *      Caller must make sure that @ind is valid and will stay that way.
399  */
400 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
401 {
402         struct ext4_inode_info *ei = EXT4_I(inode);
403         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
404         __le32 *p;
405         ext4_fsblk_t bg_start;
406         ext4_grpblk_t colour;
407
408         /* Try to find previous block */
409         for (p = ind->p - 1; p >= start; p--) {
410                 if (*p)
411                         return le32_to_cpu(*p);
412         }
413
414         /* No such thing, so let's try location of indirect block */
415         if (ind->bh)
416                 return ind->bh->b_blocknr;
417
418         /*
419          * It is going to be referred to from the inode itself? OK, just put it
420          * into the same cylinder group then.
421          */
422         bg_start = ext4_group_first_block_no(inode->i_sb, ei->i_block_group);
423         colour = (current->pid % 16) *
424                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
425         return bg_start + colour;
426 }
427
428 /**
429  *      ext4_find_goal - find a prefered place for allocation.
430  *      @inode: owner
431  *      @block:  block we want
432  *      @chain:  chain of indirect blocks
433  *      @partial: pointer to the last triple within a chain
434  *      @goal:  place to store the result.
435  *
436  *      Normally this function find the prefered place for block allocation,
437  *      stores it in *@goal and returns zero.
438  */
439
440 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
441                 Indirect chain[4], Indirect *partial)
442 {
443         struct ext4_block_alloc_info *block_i;
444
445         block_i =  EXT4_I(inode)->i_block_alloc_info;
446
447         /*
448          * try the heuristic for sequential allocation,
449          * failing that at least try to get decent locality.
450          */
451         if (block_i && (block == block_i->last_alloc_logical_block + 1)
452                 && (block_i->last_alloc_physical_block != 0)) {
453                 return block_i->last_alloc_physical_block + 1;
454         }
455
456         return ext4_find_near(inode, partial);
457 }
458
459 /**
460  *      ext4_blks_to_allocate: Look up the block map and count the number
461  *      of direct blocks need to be allocated for the given branch.
462  *
463  *      @branch: chain of indirect blocks
464  *      @k: number of blocks need for indirect blocks
465  *      @blks: number of data blocks to be mapped.
466  *      @blocks_to_boundary:  the offset in the indirect block
467  *
468  *      return the total number of blocks to be allocate, including the
469  *      direct and indirect blocks.
470  */
471 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
472                 int blocks_to_boundary)
473 {
474         unsigned long count = 0;
475
476         /*
477          * Simple case, [t,d]Indirect block(s) has not allocated yet
478          * then it's clear blocks on that path have not allocated
479          */
480         if (k > 0) {
481                 /* right now we don't handle cross boundary allocation */
482                 if (blks < blocks_to_boundary + 1)
483                         count += blks;
484                 else
485                         count += blocks_to_boundary + 1;
486                 return count;
487         }
488
489         count++;
490         while (count < blks && count <= blocks_to_boundary &&
491                 le32_to_cpu(*(branch[0].p + count)) == 0) {
492                 count++;
493         }
494         return count;
495 }
496
497 /**
498  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
499  *      @indirect_blks: the number of blocks need to allocate for indirect
500  *                      blocks
501  *
502  *      @new_blocks: on return it will store the new block numbers for
503  *      the indirect blocks(if needed) and the first direct block,
504  *      @blks:  on return it will store the total number of allocated
505  *              direct blocks
506  */
507 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
508                         ext4_fsblk_t goal, int indirect_blks, int blks,
509                         ext4_fsblk_t new_blocks[4], int *err)
510 {
511         int target, i;
512         unsigned long count = 0;
513         int index = 0;
514         ext4_fsblk_t current_block = 0;
515         int ret = 0;
516
517         /*
518          * Here we try to allocate the requested multiple blocks at once,
519          * on a best-effort basis.
520          * To build a branch, we should allocate blocks for
521          * the indirect blocks(if not allocated yet), and at least
522          * the first direct block of this branch.  That's the
523          * minimum number of blocks need to allocate(required)
524          */
525         target = blks + indirect_blks;
526
527         while (1) {
528                 count = target;
529                 /* allocating blocks for indirect blocks and direct blocks */
530                 current_block = ext4_new_blocks(handle,inode,goal,&count,err);
531                 if (*err)
532                         goto failed_out;
533
534                 target -= count;
535                 /* allocate blocks for indirect blocks */
536                 while (index < indirect_blks && count) {
537                         new_blocks[index++] = current_block++;
538                         count--;
539                 }
540
541                 if (count > 0)
542                         break;
543         }
544
545         /* save the new block number for the first direct block */
546         new_blocks[index] = current_block;
547
548         /* total number of blocks allocated for direct blocks */
549         ret = count;
550         *err = 0;
551         return ret;
552 failed_out:
553         for (i = 0; i <index; i++)
554                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
555         return ret;
556 }
557
558 /**
559  *      ext4_alloc_branch - allocate and set up a chain of blocks.
560  *      @inode: owner
561  *      @indirect_blks: number of allocated indirect blocks
562  *      @blks: number of allocated direct blocks
563  *      @offsets: offsets (in the blocks) to store the pointers to next.
564  *      @branch: place to store the chain in.
565  *
566  *      This function allocates blocks, zeroes out all but the last one,
567  *      links them into chain and (if we are synchronous) writes them to disk.
568  *      In other words, it prepares a branch that can be spliced onto the
569  *      inode. It stores the information about that chain in the branch[], in
570  *      the same format as ext4_get_branch() would do. We are calling it after
571  *      we had read the existing part of chain and partial points to the last
572  *      triple of that (one with zero ->key). Upon the exit we have the same
573  *      picture as after the successful ext4_get_block(), except that in one
574  *      place chain is disconnected - *branch->p is still zero (we did not
575  *      set the last link), but branch->key contains the number that should
576  *      be placed into *branch->p to fill that gap.
577  *
578  *      If allocation fails we free all blocks we've allocated (and forget
579  *      their buffer_heads) and return the error value the from failed
580  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
581  *      as described above and return 0.
582  */
583 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
584                         int indirect_blks, int *blks, ext4_fsblk_t goal,
585                         ext4_lblk_t *offsets, Indirect *branch)
586 {
587         int blocksize = inode->i_sb->s_blocksize;
588         int i, n = 0;
589         int err = 0;
590         struct buffer_head *bh;
591         int num;
592         ext4_fsblk_t new_blocks[4];
593         ext4_fsblk_t current_block;
594
595         num = ext4_alloc_blocks(handle, inode, goal, indirect_blks,
596                                 *blks, new_blocks, &err);
597         if (err)
598                 return err;
599
600         branch[0].key = cpu_to_le32(new_blocks[0]);
601         /*
602          * metadata blocks and data blocks are allocated.
603          */
604         for (n = 1; n <= indirect_blks;  n++) {
605                 /*
606                  * Get buffer_head for parent block, zero it out
607                  * and set the pointer to new one, then send
608                  * parent to disk.
609                  */
610                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
611                 branch[n].bh = bh;
612                 lock_buffer(bh);
613                 BUFFER_TRACE(bh, "call get_create_access");
614                 err = ext4_journal_get_create_access(handle, bh);
615                 if (err) {
616                         unlock_buffer(bh);
617                         brelse(bh);
618                         goto failed;
619                 }
620
621                 memset(bh->b_data, 0, blocksize);
622                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
623                 branch[n].key = cpu_to_le32(new_blocks[n]);
624                 *branch[n].p = branch[n].key;
625                 if ( n == indirect_blks) {
626                         current_block = new_blocks[n];
627                         /*
628                          * End of chain, update the last new metablock of
629                          * the chain to point to the new allocated
630                          * data blocks numbers
631                          */
632                         for (i=1; i < num; i++)
633                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
634                 }
635                 BUFFER_TRACE(bh, "marking uptodate");
636                 set_buffer_uptodate(bh);
637                 unlock_buffer(bh);
638
639                 BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
640                 err = ext4_journal_dirty_metadata(handle, bh);
641                 if (err)
642                         goto failed;
643         }
644         *blks = num;
645         return err;
646 failed:
647         /* Allocation failed, free what we already allocated */
648         for (i = 1; i <= n ; i++) {
649                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
650                 ext4_journal_forget(handle, branch[i].bh);
651         }
652         for (i = 0; i <indirect_blks; i++)
653                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
654
655         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
656
657         return err;
658 }
659
660 /**
661  * ext4_splice_branch - splice the allocated branch onto inode.
662  * @inode: owner
663  * @block: (logical) number of block we are adding
664  * @chain: chain of indirect blocks (with a missing link - see
665  *      ext4_alloc_branch)
666  * @where: location of missing link
667  * @num:   number of indirect blocks we are adding
668  * @blks:  number of direct blocks we are adding
669  *
670  * This function fills the missing link and does all housekeeping needed in
671  * inode (->i_blocks, etc.). In case of success we end up with the full
672  * chain to new block and return 0.
673  */
674 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
675                         ext4_lblk_t block, Indirect *where, int num, int blks)
676 {
677         int i;
678         int err = 0;
679         struct ext4_block_alloc_info *block_i;
680         ext4_fsblk_t current_block;
681
682         block_i = EXT4_I(inode)->i_block_alloc_info;
683         /*
684          * If we're splicing into a [td]indirect block (as opposed to the
685          * inode) then we need to get write access to the [td]indirect block
686          * before the splice.
687          */
688         if (where->bh) {
689                 BUFFER_TRACE(where->bh, "get_write_access");
690                 err = ext4_journal_get_write_access(handle, where->bh);
691                 if (err)
692                         goto err_out;
693         }
694         /* That's it */
695
696         *where->p = where->key;
697
698         /*
699          * Update the host buffer_head or inode to point to more just allocated
700          * direct blocks blocks
701          */
702         if (num == 0 && blks > 1) {
703                 current_block = le32_to_cpu(where->key) + 1;
704                 for (i = 1; i < blks; i++)
705                         *(where->p + i ) = cpu_to_le32(current_block++);
706         }
707
708         /*
709          * update the most recently allocated logical & physical block
710          * in i_block_alloc_info, to assist find the proper goal block for next
711          * allocation
712          */
713         if (block_i) {
714                 block_i->last_alloc_logical_block = block + blks - 1;
715                 block_i->last_alloc_physical_block =
716                                 le32_to_cpu(where[num].key) + blks - 1;
717         }
718
719         /* We are done with atomic stuff, now do the rest of housekeeping */
720
721         inode->i_ctime = ext4_current_time(inode);
722         ext4_mark_inode_dirty(handle, inode);
723
724         /* had we spliced it onto indirect block? */
725         if (where->bh) {
726                 /*
727                  * If we spliced it onto an indirect block, we haven't
728                  * altered the inode.  Note however that if it is being spliced
729                  * onto an indirect block at the very end of the file (the
730                  * file is growing) then we *will* alter the inode to reflect
731                  * the new i_size.  But that is not done here - it is done in
732                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
733                  */
734                 jbd_debug(5, "splicing indirect only\n");
735                 BUFFER_TRACE(where->bh, "call ext4_journal_dirty_metadata");
736                 err = ext4_journal_dirty_metadata(handle, where->bh);
737                 if (err)
738                         goto err_out;
739         } else {
740                 /*
741                  * OK, we spliced it into the inode itself on a direct block.
742                  * Inode was dirtied above.
743                  */
744                 jbd_debug(5, "splicing direct\n");
745         }
746         return err;
747
748 err_out:
749         for (i = 1; i <= num; i++) {
750                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
751                 ext4_journal_forget(handle, where[i].bh);
752                 ext4_free_blocks(handle, inode,
753                                         le32_to_cpu(where[i-1].key), 1, 0);
754         }
755         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
756
757         return err;
758 }
759
760 /*
761  * Allocation strategy is simple: if we have to allocate something, we will
762  * have to go the whole way to leaf. So let's do it before attaching anything
763  * to tree, set linkage between the newborn blocks, write them if sync is
764  * required, recheck the path, free and repeat if check fails, otherwise
765  * set the last missing link (that will protect us from any truncate-generated
766  * removals - all blocks on the path are immune now) and possibly force the
767  * write on the parent block.
768  * That has a nice additional property: no special recovery from the failed
769  * allocations is needed - we simply release blocks and do not touch anything
770  * reachable from inode.
771  *
772  * `handle' can be NULL if create == 0.
773  *
774  * The BKL may not be held on entry here.  Be sure to take it early.
775  * return > 0, # of blocks mapped or allocated.
776  * return = 0, if plain lookup failed.
777  * return < 0, error case.
778  *
779  *
780  * Need to be called with
781  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
782  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
783  */
784 int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
785                 ext4_lblk_t iblock, unsigned long maxblocks,
786                 struct buffer_head *bh_result,
787                 int create, int extend_disksize)
788 {
789         int err = -EIO;
790         ext4_lblk_t offsets[4];
791         Indirect chain[4];
792         Indirect *partial;
793         ext4_fsblk_t goal;
794         int indirect_blks;
795         int blocks_to_boundary = 0;
796         int depth;
797         struct ext4_inode_info *ei = EXT4_I(inode);
798         int count = 0;
799         ext4_fsblk_t first_block = 0;
800
801
802         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
803         J_ASSERT(handle != NULL || create == 0);
804         depth = ext4_block_to_path(inode, iblock, offsets,
805                                         &blocks_to_boundary);
806
807         if (depth == 0)
808                 goto out;
809
810         partial = ext4_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                         ext4_fsblk_t blk;
820
821                         blk = le32_to_cpu(*(chain[depth-1].p + count));
822
823                         if (blk == first_block + count)
824                                 count++;
825                         else
826                                 break;
827                 }
828                 goto got_it;
829         }
830
831         /* Next simple case - plain lookup or failed read of indirect block */
832         if (!create || err == -EIO)
833                 goto cleanup;
834
835         /*
836          * Okay, we need to do block allocation.  Lazily initialize the block
837          * allocation info here if necessary
838         */
839         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
840                 ext4_init_block_alloc_info(inode);
841
842         goal = ext4_find_goal(inode, iblock, chain, partial);
843
844         /* the number of blocks need to allocate for [d,t]indirect blocks */
845         indirect_blks = (chain + depth) - partial - 1;
846
847         /*
848          * Next look up the indirect map to count the totoal number of
849          * direct blocks to allocate for this branch.
850          */
851         count = ext4_blks_to_allocate(partial, indirect_blks,
852                                         maxblocks, blocks_to_boundary);
853         /*
854          * Block out ext4_truncate while we alter the tree
855          */
856         err = ext4_alloc_branch(handle, inode, indirect_blks, &count, goal,
857                                 offsets + (partial - chain), partial);
858
859         /*
860          * The ext4_splice_branch call will free and forget any buffers
861          * on the new chain if there is a failure, but that risks using
862          * up transaction credits, especially for bitmaps where the
863          * credits cannot be returned.  Can we handle this somehow?  We
864          * may need to return -EAGAIN upwards in the worst case.  --sct
865          */
866         if (!err)
867                 err = ext4_splice_branch(handle, inode, iblock,
868                                         partial, indirect_blks, count);
869         /*
870          * i_disksize growing is protected by i_data_sem.  Don't forget to
871          * protect it if you're about to implement concurrent
872          * ext4_get_block() -bzzz
873         */
874         if (!err && extend_disksize && inode->i_size > ei->i_disksize)
875                 ei->i_disksize = inode->i_size;
876         if (err)
877                 goto cleanup;
878
879         set_buffer_new(bh_result);
880 got_it:
881         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
882         if (count > blocks_to_boundary)
883                 set_buffer_boundary(bh_result);
884         err = count;
885         /* Clean up and exit */
886         partial = chain + depth - 1;    /* the whole chain */
887 cleanup:
888         while (partial > chain) {
889                 BUFFER_TRACE(partial->bh, "call brelse");
890                 brelse(partial->bh);
891                 partial--;
892         }
893         BUFFER_TRACE(bh_result, "returned");
894 out:
895         return err;
896 }
897
898 #define DIO_CREDITS (EXT4_RESERVE_TRANS_BLOCKS + 32)
899
900 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
901                         unsigned long max_blocks, struct buffer_head *bh,
902                         int create, int extend_disksize)
903 {
904         int retval;
905         /*
906          * Try to see if we can get  the block without requesting
907          * for new file system block.
908          */
909         down_read((&EXT4_I(inode)->i_data_sem));
910         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
911                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
912                                 bh, 0, 0);
913         } else {
914                 retval = ext4_get_blocks_handle(handle,
915                                 inode, block, max_blocks, bh, 0, 0);
916         }
917         up_read((&EXT4_I(inode)->i_data_sem));
918         if (!create || (retval > 0))
919                 return retval;
920
921         /*
922          * We need to allocate new blocks which will result
923          * in i_data update
924          */
925         down_write((&EXT4_I(inode)->i_data_sem));
926         /*
927          * We need to check for EXT4 here because migrate
928          * could have changed the inode type in between
929          */
930         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
931                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
932                                 bh, create, extend_disksize);
933         } else {
934                 retval = ext4_get_blocks_handle(handle, inode, block,
935                                 max_blocks, bh, create, extend_disksize);
936         }
937         up_write((&EXT4_I(inode)->i_data_sem));
938         return retval;
939 }
940
941 static int ext4_get_block(struct inode *inode, sector_t iblock,
942                         struct buffer_head *bh_result, int create)
943 {
944         handle_t *handle = ext4_journal_current_handle();
945         int ret = 0;
946         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
947
948         if (!create)
949                 goto get_block;         /* A read */
950
951         if (max_blocks == 1)
952                 goto get_block;         /* A single block get */
953
954         if (handle->h_transaction->t_state == T_LOCKED) {
955                 /*
956                  * Huge direct-io writes can hold off commits for long
957                  * periods of time.  Let this commit run.
958                  */
959                 ext4_journal_stop(handle);
960                 handle = ext4_journal_start(inode, DIO_CREDITS);
961                 if (IS_ERR(handle))
962                         ret = PTR_ERR(handle);
963                 goto get_block;
964         }
965
966         if (handle->h_buffer_credits <= EXT4_RESERVE_TRANS_BLOCKS) {
967                 /*
968                  * Getting low on buffer credits...
969                  */
970                 ret = ext4_journal_extend(handle, DIO_CREDITS);
971                 if (ret > 0) {
972                         /*
973                          * Couldn't extend the transaction.  Start a new one.
974                          */
975                         ret = ext4_journal_restart(handle, DIO_CREDITS);
976                 }
977         }
978
979 get_block:
980         if (ret == 0) {
981                 ret = ext4_get_blocks_wrap(handle, inode, iblock,
982                                         max_blocks, bh_result, create, 0);
983                 if (ret > 0) {
984                         bh_result->b_size = (ret << inode->i_blkbits);
985                         ret = 0;
986                 }
987         }
988         return ret;
989 }
990
991 /*
992  * `handle' can be NULL if create is zero
993  */
994 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
995                                 ext4_lblk_t block, int create, int *errp)
996 {
997         struct buffer_head dummy;
998         int fatal = 0, err;
999
1000         J_ASSERT(handle != NULL || create == 0);
1001
1002         dummy.b_state = 0;
1003         dummy.b_blocknr = -1000;
1004         buffer_trace_init(&dummy.b_history);
1005         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1006                                         &dummy, create, 1);
1007         /*
1008          * ext4_get_blocks_handle() returns number of blocks
1009          * mapped. 0 in case of a HOLE.
1010          */
1011         if (err > 0) {
1012                 if (err > 1)
1013                         WARN_ON(1);
1014                 err = 0;
1015         }
1016         *errp = err;
1017         if (!err && buffer_mapped(&dummy)) {
1018                 struct buffer_head *bh;
1019                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1020                 if (!bh) {
1021                         *errp = -EIO;
1022                         goto err;
1023                 }
1024                 if (buffer_new(&dummy)) {
1025                         J_ASSERT(create != 0);
1026                         J_ASSERT(handle != NULL);
1027
1028                         /*
1029                          * Now that we do not always journal data, we should
1030                          * keep in mind whether this should always journal the
1031                          * new buffer as metadata.  For now, regular file
1032                          * writes use ext4_get_block instead, so it's not a
1033                          * problem.
1034                          */
1035                         lock_buffer(bh);
1036                         BUFFER_TRACE(bh, "call get_create_access");
1037                         fatal = ext4_journal_get_create_access(handle, bh);
1038                         if (!fatal && !buffer_uptodate(bh)) {
1039                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1040                                 set_buffer_uptodate(bh);
1041                         }
1042                         unlock_buffer(bh);
1043                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
1044                         err = ext4_journal_dirty_metadata(handle, bh);
1045                         if (!fatal)
1046                                 fatal = err;
1047                 } else {
1048                         BUFFER_TRACE(bh, "not a new buffer");
1049                 }
1050                 if (fatal) {
1051                         *errp = fatal;
1052                         brelse(bh);
1053                         bh = NULL;
1054                 }
1055                 return bh;
1056         }
1057 err:
1058         return NULL;
1059 }
1060
1061 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1062                                ext4_lblk_t block, int create, int *err)
1063 {
1064         struct buffer_head * bh;
1065
1066         bh = ext4_getblk(handle, inode, block, create, err);
1067         if (!bh)
1068                 return bh;
1069         if (buffer_uptodate(bh))
1070                 return bh;
1071         ll_rw_block(READ_META, 1, &bh);
1072         wait_on_buffer(bh);
1073         if (buffer_uptodate(bh))
1074                 return bh;
1075         put_bh(bh);
1076         *err = -EIO;
1077         return NULL;
1078 }
1079
1080 static int walk_page_buffers(   handle_t *handle,
1081                                 struct buffer_head *head,
1082                                 unsigned from,
1083                                 unsigned to,
1084                                 int *partial,
1085                                 int (*fn)(      handle_t *handle,
1086                                                 struct buffer_head *bh))
1087 {
1088         struct buffer_head *bh;
1089         unsigned block_start, block_end;
1090         unsigned blocksize = head->b_size;
1091         int err, ret = 0;
1092         struct buffer_head *next;
1093
1094         for (   bh = head, block_start = 0;
1095                 ret == 0 && (bh != head || !block_start);
1096                 block_start = block_end, bh = next)
1097         {
1098                 next = bh->b_this_page;
1099                 block_end = block_start + blocksize;
1100                 if (block_end <= from || block_start >= to) {
1101                         if (partial && !buffer_uptodate(bh))
1102                                 *partial = 1;
1103                         continue;
1104                 }
1105                 err = (*fn)(handle, bh);
1106                 if (!ret)
1107                         ret = err;
1108         }
1109         return ret;
1110 }
1111
1112 /*
1113  * To preserve ordering, it is essential that the hole instantiation and
1114  * the data write be encapsulated in a single transaction.  We cannot
1115  * close off a transaction and start a new one between the ext4_get_block()
1116  * and the commit_write().  So doing the jbd2_journal_start at the start of
1117  * prepare_write() is the right place.
1118  *
1119  * Also, this function can nest inside ext4_writepage() ->
1120  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1121  * has generated enough buffer credits to do the whole page.  So we won't
1122  * block on the journal in that case, which is good, because the caller may
1123  * be PF_MEMALLOC.
1124  *
1125  * By accident, ext4 can be reentered when a transaction is open via
1126  * quota file writes.  If we were to commit the transaction while thus
1127  * reentered, there can be a deadlock - we would be holding a quota
1128  * lock, and the commit would never complete if another thread had a
1129  * transaction open and was blocking on the quota lock - a ranking
1130  * violation.
1131  *
1132  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1133  * will _not_ run commit under these circumstances because handle->h_ref
1134  * is elevated.  We'll still have enough credits for the tiny quotafile
1135  * write.
1136  */
1137 static int do_journal_get_write_access(handle_t *handle,
1138                                         struct buffer_head *bh)
1139 {
1140         if (!buffer_mapped(bh) || buffer_freed(bh))
1141                 return 0;
1142         return ext4_journal_get_write_access(handle, bh);
1143 }
1144
1145 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1146                                 loff_t pos, unsigned len, unsigned flags,
1147                                 struct page **pagep, void **fsdata)
1148 {
1149         struct inode *inode = mapping->host;
1150         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1151         handle_t *handle;
1152         int retries = 0;
1153         struct page *page;
1154         pgoff_t index;
1155         unsigned from, to;
1156
1157         index = pos >> PAGE_CACHE_SHIFT;
1158         from = pos & (PAGE_CACHE_SIZE - 1);
1159         to = from + len;
1160
1161 retry:
1162         page = __grab_cache_page(mapping, index);
1163         if (!page)
1164                 return -ENOMEM;
1165         *pagep = page;
1166
1167         handle = ext4_journal_start(inode, needed_blocks);
1168         if (IS_ERR(handle)) {
1169                 unlock_page(page);
1170                 page_cache_release(page);
1171                 ret = PTR_ERR(handle);
1172                 goto out;
1173         }
1174
1175         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1176                                                         ext4_get_block);
1177
1178         if (!ret && ext4_should_journal_data(inode)) {
1179                 ret = walk_page_buffers(handle, page_buffers(page),
1180                                 from, to, NULL, do_journal_get_write_access);
1181         }
1182
1183         if (ret) {
1184                 ext4_journal_stop(handle);
1185                 unlock_page(page);
1186                 page_cache_release(page);
1187         }
1188
1189         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1190                 goto retry;
1191 out:
1192         return ret;
1193 }
1194
1195 int ext4_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1196 {
1197         int err = jbd2_journal_dirty_data(handle, bh);
1198         if (err)
1199                 ext4_journal_abort_handle(__FUNCTION__, __FUNCTION__,
1200                                                 bh, handle, err);
1201         return err;
1202 }
1203
1204 /* For write_end() in data=journal mode */
1205 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1206 {
1207         if (!buffer_mapped(bh) || buffer_freed(bh))
1208                 return 0;
1209         set_buffer_uptodate(bh);
1210         return ext4_journal_dirty_metadata(handle, bh);
1211 }
1212
1213 /*
1214  * Generic write_end handler for ordered and writeback ext4 journal modes.
1215  * We can't use generic_write_end, because that unlocks the page and we need to
1216  * unlock the page after ext4_journal_stop, but ext4_journal_stop must run
1217  * after block_write_end.
1218  */
1219 static int ext4_generic_write_end(struct file *file,
1220                                 struct address_space *mapping,
1221                                 loff_t pos, unsigned len, unsigned copied,
1222                                 struct page *page, void *fsdata)
1223 {
1224         struct inode *inode = file->f_mapping->host;
1225
1226         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1227
1228         if (pos+copied > inode->i_size) {
1229                 i_size_write(inode, pos+copied);
1230                 mark_inode_dirty(inode);
1231         }
1232
1233         return copied;
1234 }
1235
1236 /*
1237  * We need to pick up the new inode size which generic_commit_write gave us
1238  * `file' can be NULL - eg, when called from page_symlink().
1239  *
1240  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1241  * buffers are managed internally.
1242  */
1243 static int ext4_ordered_write_end(struct file *file,
1244                                 struct address_space *mapping,
1245                                 loff_t pos, unsigned len, unsigned copied,
1246                                 struct page *page, void *fsdata)
1247 {
1248         handle_t *handle = ext4_journal_current_handle();
1249         struct inode *inode = file->f_mapping->host;
1250         unsigned from, to;
1251         int ret = 0, ret2;
1252
1253         from = pos & (PAGE_CACHE_SIZE - 1);
1254         to = from + len;
1255
1256         ret = walk_page_buffers(handle, page_buffers(page),
1257                 from, to, NULL, ext4_journal_dirty_data);
1258
1259         if (ret == 0) {
1260                 /*
1261                  * generic_write_end() will run mark_inode_dirty() if i_size
1262                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1263                  * into that.
1264                  */
1265                 loff_t new_i_size;
1266
1267                 new_i_size = pos + copied;
1268                 if (new_i_size > EXT4_I(inode)->i_disksize)
1269                         EXT4_I(inode)->i_disksize = new_i_size;
1270                 copied = ext4_generic_write_end(file, mapping, pos, len, copied,
1271                                                         page, fsdata);
1272                 if (copied < 0)
1273                         ret = copied;
1274         }
1275         ret2 = ext4_journal_stop(handle);
1276         if (!ret)
1277                 ret = ret2;
1278         unlock_page(page);
1279         page_cache_release(page);
1280
1281         return ret ? ret : copied;
1282 }
1283
1284 static int ext4_writeback_write_end(struct file *file,
1285                                 struct address_space *mapping,
1286                                 loff_t pos, unsigned len, unsigned copied,
1287                                 struct page *page, void *fsdata)
1288 {
1289         handle_t *handle = ext4_journal_current_handle();
1290         struct inode *inode = file->f_mapping->host;
1291         int ret = 0, ret2;
1292         loff_t new_i_size;
1293
1294         new_i_size = pos + copied;
1295         if (new_i_size > EXT4_I(inode)->i_disksize)
1296                 EXT4_I(inode)->i_disksize = new_i_size;
1297
1298         copied = ext4_generic_write_end(file, mapping, pos, len, copied,
1299                                                         page, fsdata);
1300         if (copied < 0)
1301                 ret = copied;
1302
1303         ret2 = ext4_journal_stop(handle);
1304         if (!ret)
1305                 ret = ret2;
1306         unlock_page(page);
1307         page_cache_release(page);
1308
1309         return ret ? ret : copied;
1310 }
1311
1312 static int ext4_journalled_write_end(struct file *file,
1313                                 struct address_space *mapping,
1314                                 loff_t pos, unsigned len, unsigned copied,
1315                                 struct page *page, void *fsdata)
1316 {
1317         handle_t *handle = ext4_journal_current_handle();
1318         struct inode *inode = mapping->host;
1319         int ret = 0, ret2;
1320         int partial = 0;
1321         unsigned from, to;
1322
1323         from = pos & (PAGE_CACHE_SIZE - 1);
1324         to = from + len;
1325
1326         if (copied < len) {
1327                 if (!PageUptodate(page))
1328                         copied = 0;
1329                 page_zero_new_buffers(page, from+copied, to);
1330         }
1331
1332         ret = walk_page_buffers(handle, page_buffers(page), from,
1333                                 to, &partial, write_end_fn);
1334         if (!partial)
1335                 SetPageUptodate(page);
1336         if (pos+copied > inode->i_size)
1337                 i_size_write(inode, pos+copied);
1338         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1339         if (inode->i_size > EXT4_I(inode)->i_disksize) {
1340                 EXT4_I(inode)->i_disksize = inode->i_size;
1341                 ret2 = ext4_mark_inode_dirty(handle, inode);
1342                 if (!ret)
1343                         ret = ret2;
1344         }
1345
1346         ret2 = ext4_journal_stop(handle);
1347         if (!ret)
1348                 ret = ret2;
1349         unlock_page(page);
1350         page_cache_release(page);
1351
1352         return ret ? ret : copied;
1353 }
1354
1355 /*
1356  * bmap() is special.  It gets used by applications such as lilo and by
1357  * the swapper to find the on-disk block of a specific piece of data.
1358  *
1359  * Naturally, this is dangerous if the block concerned is still in the
1360  * journal.  If somebody makes a swapfile on an ext4 data-journaling
1361  * filesystem and enables swap, then they may get a nasty shock when the
1362  * data getting swapped to that swapfile suddenly gets overwritten by
1363  * the original zero's written out previously to the journal and
1364  * awaiting writeback in the kernel's buffer cache.
1365  *
1366  * So, if we see any bmap calls here on a modified, data-journaled file,
1367  * take extra steps to flush any blocks which might be in the cache.
1368  */
1369 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
1370 {
1371         struct inode *inode = mapping->host;
1372         journal_t *journal;
1373         int err;
1374
1375         if (EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
1376                 /*
1377                  * This is a REALLY heavyweight approach, but the use of
1378                  * bmap on dirty files is expected to be extremely rare:
1379                  * only if we run lilo or swapon on a freshly made file
1380                  * do we expect this to happen.
1381                  *
1382                  * (bmap requires CAP_SYS_RAWIO so this does not
1383                  * represent an unprivileged user DOS attack --- we'd be
1384                  * in trouble if mortal users could trigger this path at
1385                  * will.)
1386                  *
1387                  * NB. EXT4_STATE_JDATA is not set on files other than
1388                  * regular files.  If somebody wants to bmap a directory
1389                  * or symlink and gets confused because the buffer
1390                  * hasn't yet been flushed to disk, they deserve
1391                  * everything they get.
1392                  */
1393
1394                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
1395                 journal = EXT4_JOURNAL(inode);
1396                 jbd2_journal_lock_updates(journal);
1397                 err = jbd2_journal_flush(journal);
1398                 jbd2_journal_unlock_updates(journal);
1399
1400                 if (err)
1401                         return 0;
1402         }
1403
1404         return generic_block_bmap(mapping,block,ext4_get_block);
1405 }
1406
1407 static int bget_one(handle_t *handle, struct buffer_head *bh)
1408 {
1409         get_bh(bh);
1410         return 0;
1411 }
1412
1413 static int bput_one(handle_t *handle, struct buffer_head *bh)
1414 {
1415         put_bh(bh);
1416         return 0;
1417 }
1418
1419 static int jbd2_journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1420 {
1421         if (buffer_mapped(bh))
1422                 return ext4_journal_dirty_data(handle, bh);
1423         return 0;
1424 }
1425
1426 /*
1427  * Note that we always start a transaction even if we're not journalling
1428  * data.  This is to preserve ordering: any hole instantiation within
1429  * __block_write_full_page -> ext4_get_block() should be journalled
1430  * along with the data so we don't crash and then get metadata which
1431  * refers to old data.
1432  *
1433  * In all journalling modes block_write_full_page() will start the I/O.
1434  *
1435  * Problem:
1436  *
1437  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1438  *              ext4_writepage()
1439  *
1440  * Similar for:
1441  *
1442  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1443  *
1444  * Same applies to ext4_get_block().  We will deadlock on various things like
1445  * lock_journal and i_data_sem
1446  *
1447  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1448  * allocations fail.
1449  *
1450  * 16May01: If we're reentered then journal_current_handle() will be
1451  *          non-zero. We simply *return*.
1452  *
1453  * 1 July 2001: @@@ FIXME:
1454  *   In journalled data mode, a data buffer may be metadata against the
1455  *   current transaction.  But the same file is part of a shared mapping
1456  *   and someone does a writepage() on it.
1457  *
1458  *   We will move the buffer onto the async_data list, but *after* it has
1459  *   been dirtied. So there's a small window where we have dirty data on
1460  *   BJ_Metadata.
1461  *
1462  *   Note that this only applies to the last partial page in the file.  The
1463  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1464  *   broken code anyway: it's wrong for msync()).
1465  *
1466  *   It's a rare case: affects the final partial page, for journalled data
1467  *   where the file is subject to bith write() and writepage() in the same
1468  *   transction.  To fix it we'll need a custom block_write_full_page().
1469  *   We'll probably need that anyway for journalling writepage() output.
1470  *
1471  * We don't honour synchronous mounts for writepage().  That would be
1472  * disastrous.  Any write() or metadata operation will sync the fs for
1473  * us.
1474  *
1475  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1476  * we don't need to open a transaction here.
1477  */
1478 static int ext4_ordered_writepage(struct page *page,
1479                                 struct writeback_control *wbc)
1480 {
1481         struct inode *inode = page->mapping->host;
1482         struct buffer_head *page_bufs;
1483         handle_t *handle = NULL;
1484         int ret = 0;
1485         int err;
1486
1487         J_ASSERT(PageLocked(page));
1488
1489         /*
1490          * We give up here if we're reentered, because it might be for a
1491          * different filesystem.
1492          */
1493         if (ext4_journal_current_handle())
1494                 goto out_fail;
1495
1496         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1497
1498         if (IS_ERR(handle)) {
1499                 ret = PTR_ERR(handle);
1500                 goto out_fail;
1501         }
1502
1503         if (!page_has_buffers(page)) {
1504                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1505                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1506         }
1507         page_bufs = page_buffers(page);
1508         walk_page_buffers(handle, page_bufs, 0,
1509                         PAGE_CACHE_SIZE, NULL, bget_one);
1510
1511         ret = block_write_full_page(page, ext4_get_block, wbc);
1512
1513         /*
1514          * The page can become unlocked at any point now, and
1515          * truncate can then come in and change things.  So we
1516          * can't touch *page from now on.  But *page_bufs is
1517          * safe due to elevated refcount.
1518          */
1519
1520         /*
1521          * And attach them to the current transaction.  But only if
1522          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1523          * and generally junk.
1524          */
1525         if (ret == 0) {
1526                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1527                                         NULL, jbd2_journal_dirty_data_fn);
1528                 if (!ret)
1529                         ret = err;
1530         }
1531         walk_page_buffers(handle, page_bufs, 0,
1532                         PAGE_CACHE_SIZE, NULL, bput_one);
1533         err = ext4_journal_stop(handle);
1534         if (!ret)
1535                 ret = err;
1536         return ret;
1537
1538 out_fail:
1539         redirty_page_for_writepage(wbc, page);
1540         unlock_page(page);
1541         return ret;
1542 }
1543
1544 static int ext4_writeback_writepage(struct page *page,
1545                                 struct writeback_control *wbc)
1546 {
1547         struct inode *inode = page->mapping->host;
1548         handle_t *handle = NULL;
1549         int ret = 0;
1550         int err;
1551
1552         if (ext4_journal_current_handle())
1553                 goto out_fail;
1554
1555         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1556         if (IS_ERR(handle)) {
1557                 ret = PTR_ERR(handle);
1558                 goto out_fail;
1559         }
1560
1561         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
1562                 ret = nobh_writepage(page, ext4_get_block, wbc);
1563         else
1564                 ret = block_write_full_page(page, ext4_get_block, wbc);
1565
1566         err = ext4_journal_stop(handle);
1567         if (!ret)
1568                 ret = err;
1569         return ret;
1570
1571 out_fail:
1572         redirty_page_for_writepage(wbc, page);
1573         unlock_page(page);
1574         return ret;
1575 }
1576
1577 static int ext4_journalled_writepage(struct page *page,
1578                                 struct writeback_control *wbc)
1579 {
1580         struct inode *inode = page->mapping->host;
1581         handle_t *handle = NULL;
1582         int ret = 0;
1583         int err;
1584
1585         if (ext4_journal_current_handle())
1586                 goto no_write;
1587
1588         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1589         if (IS_ERR(handle)) {
1590                 ret = PTR_ERR(handle);
1591                 goto no_write;
1592         }
1593
1594         if (!page_has_buffers(page) || PageChecked(page)) {
1595                 /*
1596                  * It's mmapped pagecache.  Add buffers and journal it.  There
1597                  * doesn't seem much point in redirtying the page here.
1598                  */
1599                 ClearPageChecked(page);
1600                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1601                                         ext4_get_block);
1602                 if (ret != 0) {
1603                         ext4_journal_stop(handle);
1604                         goto out_unlock;
1605                 }
1606                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1607                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1608
1609                 err = walk_page_buffers(handle, page_buffers(page), 0,
1610                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1611                 if (ret == 0)
1612                         ret = err;
1613                 EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1614                 unlock_page(page);
1615         } else {
1616                 /*
1617                  * It may be a page full of checkpoint-mode buffers.  We don't
1618                  * really know unless we go poke around in the buffer_heads.
1619                  * But block_write_full_page will do the right thing.
1620                  */
1621                 ret = block_write_full_page(page, ext4_get_block, wbc);
1622         }
1623         err = ext4_journal_stop(handle);
1624         if (!ret)
1625                 ret = err;
1626 out:
1627         return ret;
1628
1629 no_write:
1630         redirty_page_for_writepage(wbc, page);
1631 out_unlock:
1632         unlock_page(page);
1633         goto out;
1634 }
1635
1636 static int ext4_readpage(struct file *file, struct page *page)
1637 {
1638         return mpage_readpage(page, ext4_get_block);
1639 }
1640
1641 static int
1642 ext4_readpages(struct file *file, struct address_space *mapping,
1643                 struct list_head *pages, unsigned nr_pages)
1644 {
1645         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
1646 }
1647
1648 static void ext4_invalidatepage(struct page *page, unsigned long offset)
1649 {
1650         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
1651
1652         /*
1653          * If it's a full truncate we just forget about the pending dirtying
1654          */
1655         if (offset == 0)
1656                 ClearPageChecked(page);
1657
1658         jbd2_journal_invalidatepage(journal, page, offset);
1659 }
1660
1661 static int ext4_releasepage(struct page *page, gfp_t wait)
1662 {
1663         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
1664
1665         WARN_ON(PageChecked(page));
1666         if (!page_has_buffers(page))
1667                 return 0;
1668         return jbd2_journal_try_to_free_buffers(journal, page, wait);
1669 }
1670
1671 /*
1672  * If the O_DIRECT write will extend the file then add this inode to the
1673  * orphan list.  So recovery will truncate it back to the original size
1674  * if the machine crashes during the write.
1675  *
1676  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1677  * crashes then stale disk data _may_ be exposed inside the file.
1678  */
1679 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
1680                         const struct iovec *iov, loff_t offset,
1681                         unsigned long nr_segs)
1682 {
1683         struct file *file = iocb->ki_filp;
1684         struct inode *inode = file->f_mapping->host;
1685         struct ext4_inode_info *ei = EXT4_I(inode);
1686         handle_t *handle = NULL;
1687         ssize_t ret;
1688         int orphan = 0;
1689         size_t count = iov_length(iov, nr_segs);
1690
1691         if (rw == WRITE) {
1692                 loff_t final_size = offset + count;
1693
1694                 handle = ext4_journal_start(inode, DIO_CREDITS);
1695                 if (IS_ERR(handle)) {
1696                         ret = PTR_ERR(handle);
1697                         goto out;
1698                 }
1699                 if (final_size > inode->i_size) {
1700                         ret = ext4_orphan_add(handle, inode);
1701                         if (ret)
1702                                 goto out_stop;
1703                         orphan = 1;
1704                         ei->i_disksize = inode->i_size;
1705                 }
1706         }
1707
1708         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1709                                  offset, nr_segs,
1710                                  ext4_get_block, NULL);
1711
1712         /*
1713          * Reacquire the handle: ext4_get_block() can restart the transaction
1714          */
1715         handle = ext4_journal_current_handle();
1716
1717 out_stop:
1718         if (handle) {
1719                 int err;
1720
1721                 if (orphan && inode->i_nlink)
1722                         ext4_orphan_del(handle, inode);
1723                 if (orphan && ret > 0) {
1724                         loff_t end = offset + ret;
1725                         if (end > inode->i_size) {
1726                                 ei->i_disksize = end;
1727                                 i_size_write(inode, end);
1728                                 /*
1729                                  * We're going to return a positive `ret'
1730                                  * here due to non-zero-length I/O, so there's
1731                                  * no way of reporting error returns from
1732                                  * ext4_mark_inode_dirty() to userspace.  So
1733                                  * ignore it.
1734                                  */
1735                                 ext4_mark_inode_dirty(handle, inode);
1736                         }
1737                 }
1738                 err = ext4_journal_stop(handle);
1739                 if (ret == 0)
1740                         ret = err;
1741         }
1742 out:
1743         return ret;
1744 }
1745
1746 /*
1747  * Pages can be marked dirty completely asynchronously from ext4's journalling
1748  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1749  * much here because ->set_page_dirty is called under VFS locks.  The page is
1750  * not necessarily locked.
1751  *
1752  * We cannot just dirty the page and leave attached buffers clean, because the
1753  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1754  * or jbddirty because all the journalling code will explode.
1755  *
1756  * So what we do is to mark the page "pending dirty" and next time writepage
1757  * is called, propagate that into the buffers appropriately.
1758  */
1759 static int ext4_journalled_set_page_dirty(struct page *page)
1760 {
1761         SetPageChecked(page);
1762         return __set_page_dirty_nobuffers(page);
1763 }
1764
1765 static const struct address_space_operations ext4_ordered_aops = {
1766         .readpage       = ext4_readpage,
1767         .readpages      = ext4_readpages,
1768         .writepage      = ext4_ordered_writepage,
1769         .sync_page      = block_sync_page,
1770         .write_begin    = ext4_write_begin,
1771         .write_end      = ext4_ordered_write_end,
1772         .bmap           = ext4_bmap,
1773         .invalidatepage = ext4_invalidatepage,
1774         .releasepage    = ext4_releasepage,
1775         .direct_IO      = ext4_direct_IO,
1776         .migratepage    = buffer_migrate_page,
1777 };
1778
1779 static const struct address_space_operations ext4_writeback_aops = {
1780         .readpage       = ext4_readpage,
1781         .readpages      = ext4_readpages,
1782         .writepage      = ext4_writeback_writepage,
1783         .sync_page      = block_sync_page,
1784         .write_begin    = ext4_write_begin,
1785         .write_end      = ext4_writeback_write_end,
1786         .bmap           = ext4_bmap,
1787         .invalidatepage = ext4_invalidatepage,
1788         .releasepage    = ext4_releasepage,
1789         .direct_IO      = ext4_direct_IO,
1790         .migratepage    = buffer_migrate_page,
1791 };
1792
1793 static const struct address_space_operations ext4_journalled_aops = {
1794         .readpage       = ext4_readpage,
1795         .readpages      = ext4_readpages,
1796         .writepage      = ext4_journalled_writepage,
1797         .sync_page      = block_sync_page,
1798         .write_begin    = ext4_write_begin,
1799         .write_end      = ext4_journalled_write_end,
1800         .set_page_dirty = ext4_journalled_set_page_dirty,
1801         .bmap           = ext4_bmap,
1802         .invalidatepage = ext4_invalidatepage,
1803         .releasepage    = ext4_releasepage,
1804 };
1805
1806 void ext4_set_aops(struct inode *inode)
1807 {
1808         if (ext4_should_order_data(inode))
1809                 inode->i_mapping->a_ops = &ext4_ordered_aops;
1810         else if (ext4_should_writeback_data(inode))
1811                 inode->i_mapping->a_ops = &ext4_writeback_aops;
1812         else
1813                 inode->i_mapping->a_ops = &ext4_journalled_aops;
1814 }
1815
1816 /*
1817  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
1818  * up to the end of the block which corresponds to `from'.
1819  * This required during truncate. We need to physically zero the tail end
1820  * of that block so it doesn't yield old data if the file is later grown.
1821  */
1822 int ext4_block_truncate_page(handle_t *handle, struct page *page,
1823                 struct address_space *mapping, loff_t from)
1824 {
1825         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1826         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1827         unsigned blocksize, length, pos;
1828         ext4_lblk_t iblock;
1829         struct inode *inode = mapping->host;
1830         struct buffer_head *bh;
1831         int err = 0;
1832
1833         blocksize = inode->i_sb->s_blocksize;
1834         length = blocksize - (offset & (blocksize - 1));
1835         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1836
1837         /*
1838          * For "nobh" option,  we can only work if we don't need to
1839          * read-in the page - otherwise we create buffers to do the IO.
1840          */
1841         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1842              ext4_should_writeback_data(inode) && PageUptodate(page)) {
1843                 zero_user(page, offset, length);
1844                 set_page_dirty(page);
1845                 goto unlock;
1846         }
1847
1848         if (!page_has_buffers(page))
1849                 create_empty_buffers(page, blocksize, 0);
1850
1851         /* Find the buffer that contains "offset" */
1852         bh = page_buffers(page);
1853         pos = blocksize;
1854         while (offset >= pos) {
1855                 bh = bh->b_this_page;
1856                 iblock++;
1857                 pos += blocksize;
1858         }
1859
1860         err = 0;
1861         if (buffer_freed(bh)) {
1862                 BUFFER_TRACE(bh, "freed: skip");
1863                 goto unlock;
1864         }
1865
1866         if (!buffer_mapped(bh)) {
1867                 BUFFER_TRACE(bh, "unmapped");
1868                 ext4_get_block(inode, iblock, bh, 0);
1869                 /* unmapped? It's a hole - nothing to do */
1870                 if (!buffer_mapped(bh)) {
1871                         BUFFER_TRACE(bh, "still unmapped");
1872                         goto unlock;
1873                 }
1874         }
1875
1876         /* Ok, it's mapped. Make sure it's up-to-date */
1877         if (PageUptodate(page))
1878                 set_buffer_uptodate(bh);
1879
1880         if (!buffer_uptodate(bh)) {
1881                 err = -EIO;
1882                 ll_rw_block(READ, 1, &bh);
1883                 wait_on_buffer(bh);
1884                 /* Uhhuh. Read error. Complain and punt. */
1885                 if (!buffer_uptodate(bh))
1886                         goto unlock;
1887         }
1888
1889         if (ext4_should_journal_data(inode)) {
1890                 BUFFER_TRACE(bh, "get write access");
1891                 err = ext4_journal_get_write_access(handle, bh);
1892                 if (err)
1893                         goto unlock;
1894         }
1895
1896         zero_user(page, offset, length);
1897
1898         BUFFER_TRACE(bh, "zeroed end of block");
1899
1900         err = 0;
1901         if (ext4_should_journal_data(inode)) {
1902                 err = ext4_journal_dirty_metadata(handle, bh);
1903         } else {
1904                 if (ext4_should_order_data(inode))
1905                         err = ext4_journal_dirty_data(handle, bh);
1906                 mark_buffer_dirty(bh);
1907         }
1908
1909 unlock:
1910         unlock_page(page);
1911         page_cache_release(page);
1912         return err;
1913 }
1914
1915 /*
1916  * Probably it should be a library function... search for first non-zero word
1917  * or memcmp with zero_page, whatever is better for particular architecture.
1918  * Linus?
1919  */
1920 static inline int all_zeroes(__le32 *p, __le32 *q)
1921 {
1922         while (p < q)
1923                 if (*p++)
1924                         return 0;
1925         return 1;
1926 }
1927
1928 /**
1929  *      ext4_find_shared - find the indirect blocks for partial truncation.
1930  *      @inode:   inode in question
1931  *      @depth:   depth of the affected branch
1932  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
1933  *      @chain:   place to store the pointers to partial indirect blocks
1934  *      @top:     place to the (detached) top of branch
1935  *
1936  *      This is a helper function used by ext4_truncate().
1937  *
1938  *      When we do truncate() we may have to clean the ends of several
1939  *      indirect blocks but leave the blocks themselves alive. Block is
1940  *      partially truncated if some data below the new i_size is refered
1941  *      from it (and it is on the path to the first completely truncated
1942  *      data block, indeed).  We have to free the top of that path along
1943  *      with everything to the right of the path. Since no allocation
1944  *      past the truncation point is possible until ext4_truncate()
1945  *      finishes, we may safely do the latter, but top of branch may
1946  *      require special attention - pageout below the truncation point
1947  *      might try to populate it.
1948  *
1949  *      We atomically detach the top of branch from the tree, store the
1950  *      block number of its root in *@top, pointers to buffer_heads of
1951  *      partially truncated blocks - in @chain[].bh and pointers to
1952  *      their last elements that should not be removed - in
1953  *      @chain[].p. Return value is the pointer to last filled element
1954  *      of @chain.
1955  *
1956  *      The work left to caller to do the actual freeing of subtrees:
1957  *              a) free the subtree starting from *@top
1958  *              b) free the subtrees whose roots are stored in
1959  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
1960  *              c) free the subtrees growing from the inode past the @chain[0].
1961  *                      (no partially truncated stuff there).  */
1962
1963 static Indirect *ext4_find_shared(struct inode *inode, int depth,
1964                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
1965 {
1966         Indirect *partial, *p;
1967         int k, err;
1968
1969         *top = 0;
1970         /* Make k index the deepest non-null offest + 1 */
1971         for (k = depth; k > 1 && !offsets[k-1]; k--)
1972                 ;
1973         partial = ext4_get_branch(inode, k, offsets, chain, &err);
1974         /* Writer: pointers */
1975         if (!partial)
1976                 partial = chain + k-1;
1977         /*
1978          * If the branch acquired continuation since we've looked at it -
1979          * fine, it should all survive and (new) top doesn't belong to us.
1980          */
1981         if (!partial->key && *partial->p)
1982                 /* Writer: end */
1983                 goto no_top;
1984         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
1985                 ;
1986         /*
1987          * OK, we've found the last block that must survive. The rest of our
1988          * branch should be detached before unlocking. However, if that rest
1989          * of branch is all ours and does not grow immediately from the inode
1990          * it's easier to cheat and just decrement partial->p.
1991          */
1992         if (p == chain + k - 1 && p > chain) {
1993                 p->p--;
1994         } else {
1995                 *top = *p->p;
1996                 /* Nope, don't do this in ext4.  Must leave the tree intact */
1997 #if 0
1998                 *p->p = 0;
1999 #endif
2000         }
2001         /* Writer: end */
2002
2003         while(partial > p) {
2004                 brelse(partial->bh);
2005                 partial--;
2006         }
2007 no_top:
2008         return partial;
2009 }
2010
2011 /*
2012  * Zero a number of block pointers in either an inode or an indirect block.
2013  * If we restart the transaction we must again get write access to the
2014  * indirect block for further modification.
2015  *
2016  * We release `count' blocks on disk, but (last - first) may be greater
2017  * than `count' because there can be holes in there.
2018  */
2019 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
2020                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
2021                 unsigned long count, __le32 *first, __le32 *last)
2022 {
2023         __le32 *p;
2024         if (try_to_extend_transaction(handle, inode)) {
2025                 if (bh) {
2026                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
2027                         ext4_journal_dirty_metadata(handle, bh);
2028                 }
2029                 ext4_mark_inode_dirty(handle, inode);
2030                 ext4_journal_test_restart(handle, inode);
2031                 if (bh) {
2032                         BUFFER_TRACE(bh, "retaking write access");
2033                         ext4_journal_get_write_access(handle, bh);
2034                 }
2035         }
2036
2037         /*
2038          * Any buffers which are on the journal will be in memory. We find
2039          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
2040          * on them.  We've already detached each block from the file, so
2041          * bforget() in jbd2_journal_forget() should be safe.
2042          *
2043          * AKPM: turn on bforget in jbd2_journal_forget()!!!
2044          */
2045         for (p = first; p < last; p++) {
2046                 u32 nr = le32_to_cpu(*p);
2047                 if (nr) {
2048                         struct buffer_head *tbh;
2049
2050                         *p = 0;
2051                         tbh = sb_find_get_block(inode->i_sb, nr);
2052                         ext4_forget(handle, 0, inode, tbh, nr);
2053                 }
2054         }
2055
2056         ext4_free_blocks(handle, inode, block_to_free, count, 0);
2057 }
2058
2059 /**
2060  * ext4_free_data - free a list of data blocks
2061  * @handle:     handle for this transaction
2062  * @inode:      inode we are dealing with
2063  * @this_bh:    indirect buffer_head which contains *@first and *@last
2064  * @first:      array of block numbers
2065  * @last:       points immediately past the end of array
2066  *
2067  * We are freeing all blocks refered from that array (numbers are stored as
2068  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2069  *
2070  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2071  * blocks are contiguous then releasing them at one time will only affect one
2072  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2073  * actually use a lot of journal space.
2074  *
2075  * @this_bh will be %NULL if @first and @last point into the inode's direct
2076  * block pointers.
2077  */
2078 static void ext4_free_data(handle_t *handle, struct inode *inode,
2079                            struct buffer_head *this_bh,
2080                            __le32 *first, __le32 *last)
2081 {
2082         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2083         unsigned long count = 0;            /* Number of blocks in the run */
2084         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2085                                                corresponding to
2086                                                block_to_free */
2087         ext4_fsblk_t nr;                    /* Current block # */
2088         __le32 *p;                          /* Pointer into inode/ind
2089                                                for current block */
2090         int err;
2091
2092         if (this_bh) {                          /* For indirect block */
2093                 BUFFER_TRACE(this_bh, "get_write_access");
2094                 err = ext4_journal_get_write_access(handle, this_bh);
2095                 /* Important: if we can't update the indirect pointers
2096                  * to the blocks, we can't free them. */
2097                 if (err)
2098                         return;
2099         }
2100
2101         for (p = first; p < last; p++) {
2102                 nr = le32_to_cpu(*p);
2103                 if (nr) {
2104                         /* accumulate blocks to free if they're contiguous */
2105                         if (count == 0) {
2106                                 block_to_free = nr;
2107                                 block_to_free_p = p;
2108                                 count = 1;
2109                         } else if (nr == block_to_free + count) {
2110                                 count++;
2111                         } else {
2112                                 ext4_clear_blocks(handle, inode, this_bh,
2113                                                   block_to_free,
2114                                                   count, block_to_free_p, p);
2115                                 block_to_free = nr;
2116                                 block_to_free_p = p;
2117                                 count = 1;
2118                         }
2119                 }
2120         }
2121
2122         if (count > 0)
2123                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
2124                                   count, block_to_free_p, p);
2125
2126         if (this_bh) {
2127                 BUFFER_TRACE(this_bh, "call ext4_journal_dirty_metadata");
2128                 ext4_journal_dirty_metadata(handle, this_bh);
2129         }
2130 }
2131
2132 /**
2133  *      ext4_free_branches - free an array of branches
2134  *      @handle: JBD handle for this transaction
2135  *      @inode: inode we are dealing with
2136  *      @parent_bh: the buffer_head which contains *@first and *@last
2137  *      @first: array of block numbers
2138  *      @last:  pointer immediately past the end of array
2139  *      @depth: depth of the branches to free
2140  *
2141  *      We are freeing all blocks refered from these branches (numbers are
2142  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2143  *      appropriately.
2144  */
2145 static void ext4_free_branches(handle_t *handle, struct inode *inode,
2146                                struct buffer_head *parent_bh,
2147                                __le32 *first, __le32 *last, int depth)
2148 {
2149         ext4_fsblk_t nr;
2150         __le32 *p;
2151
2152         if (is_handle_aborted(handle))
2153                 return;
2154
2155         if (depth--) {
2156                 struct buffer_head *bh;
2157                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
2158                 p = last;
2159                 while (--p >= first) {
2160                         nr = le32_to_cpu(*p);
2161                         if (!nr)
2162                                 continue;               /* A hole */
2163
2164                         /* Go read the buffer for the next level down */
2165                         bh = sb_bread(inode->i_sb, nr);
2166
2167                         /*
2168                          * A read failure? Report error and clear slot
2169                          * (should be rare).
2170                          */
2171                         if (!bh) {
2172                                 ext4_error(inode->i_sb, "ext4_free_branches",
2173                                            "Read failure, inode=%lu, block=%llu",
2174                                            inode->i_ino, nr);
2175                                 continue;
2176                         }
2177
2178                         /* This zaps the entire block.  Bottom up. */
2179                         BUFFER_TRACE(bh, "free child branches");
2180                         ext4_free_branches(handle, inode, bh,
2181                                            (__le32*)bh->b_data,
2182                                            (__le32*)bh->b_data + addr_per_block,
2183                                            depth);
2184
2185                         /*
2186                          * We've probably journalled the indirect block several
2187                          * times during the truncate.  But it's no longer
2188                          * needed and we now drop it from the transaction via
2189                          * jbd2_journal_revoke().
2190                          *
2191                          * That's easy if it's exclusively part of this
2192                          * transaction.  But if it's part of the committing
2193                          * transaction then jbd2_journal_forget() will simply
2194                          * brelse() it.  That means that if the underlying
2195                          * block is reallocated in ext4_get_block(),
2196                          * unmap_underlying_metadata() will find this block
2197                          * and will try to get rid of it.  damn, damn.
2198                          *
2199                          * If this block has already been committed to the
2200                          * journal, a revoke record will be written.  And
2201                          * revoke records must be emitted *before* clearing
2202                          * this block's bit in the bitmaps.
2203                          */
2204                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
2205
2206                         /*
2207                          * Everything below this this pointer has been
2208                          * released.  Now let this top-of-subtree go.
2209                          *
2210                          * We want the freeing of this indirect block to be
2211                          * atomic in the journal with the updating of the
2212                          * bitmap block which owns it.  So make some room in
2213                          * the journal.
2214                          *
2215                          * We zero the parent pointer *after* freeing its
2216                          * pointee in the bitmaps, so if extend_transaction()
2217                          * for some reason fails to put the bitmap changes and
2218                          * the release into the same transaction, recovery
2219                          * will merely complain about releasing a free block,
2220                          * rather than leaking blocks.
2221                          */
2222                         if (is_handle_aborted(handle))
2223                                 return;
2224                         if (try_to_extend_transaction(handle, inode)) {
2225                                 ext4_mark_inode_dirty(handle, inode);
2226                                 ext4_journal_test_restart(handle, inode);
2227                         }
2228
2229                         ext4_free_blocks(handle, inode, nr, 1, 1);
2230
2231                         if (parent_bh) {
2232                                 /*
2233                                  * The block which we have just freed is
2234                                  * pointed to by an indirect block: journal it
2235                                  */
2236                                 BUFFER_TRACE(parent_bh, "get_write_access");
2237                                 if (!ext4_journal_get_write_access(handle,
2238                                                                    parent_bh)){
2239                                         *p = 0;
2240                                         BUFFER_TRACE(parent_bh,
2241                                         "call ext4_journal_dirty_metadata");
2242                                         ext4_journal_dirty_metadata(handle,
2243                                                                     parent_bh);
2244                                 }
2245                         }
2246                 }
2247         } else {
2248                 /* We have reached the bottom of the tree. */
2249                 BUFFER_TRACE(parent_bh, "free data blocks");
2250                 ext4_free_data(handle, inode, parent_bh, first, last);
2251         }
2252 }
2253
2254 /*
2255  * ext4_truncate()
2256  *
2257  * We block out ext4_get_block() block instantiations across the entire
2258  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
2259  * simultaneously on behalf of the same inode.
2260  *
2261  * As we work through the truncate and commmit bits of it to the journal there
2262  * is one core, guiding principle: the file's tree must always be consistent on
2263  * disk.  We must be able to restart the truncate after a crash.
2264  *
2265  * The file's tree may be transiently inconsistent in memory (although it
2266  * probably isn't), but whenever we close off and commit a journal transaction,
2267  * the contents of (the filesystem + the journal) must be consistent and
2268  * restartable.  It's pretty simple, really: bottom up, right to left (although
2269  * left-to-right works OK too).
2270  *
2271  * Note that at recovery time, journal replay occurs *before* the restart of
2272  * truncate against the orphan inode list.
2273  *
2274  * The committed inode has the new, desired i_size (which is the same as
2275  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
2276  * that this inode's truncate did not complete and it will again call
2277  * ext4_truncate() to have another go.  So there will be instantiated blocks
2278  * to the right of the truncation point in a crashed ext4 filesystem.  But
2279  * that's fine - as long as they are linked from the inode, the post-crash
2280  * ext4_truncate() run will find them and release them.
2281  */
2282 void ext4_truncate(struct inode *inode)
2283 {
2284         handle_t *handle;
2285         struct ext4_inode_info *ei = EXT4_I(inode);
2286         __le32 *i_data = ei->i_data;
2287         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
2288         struct address_space *mapping = inode->i_mapping;
2289         ext4_lblk_t offsets[4];
2290         Indirect chain[4];
2291         Indirect *partial;
2292         __le32 nr = 0;
2293         int n;
2294         ext4_lblk_t last_block;
2295         unsigned blocksize = inode->i_sb->s_blocksize;
2296         struct page *page;
2297
2298         if (!(S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
2299             S_ISLNK(inode->i_mode)))
2300                 return;
2301         if (ext4_inode_is_fast_symlink(inode))
2302                 return;
2303         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2304                 return;
2305
2306         /*
2307          * We have to lock the EOF page here, because lock_page() nests
2308          * outside jbd2_journal_start().
2309          */
2310         if ((inode->i_size & (blocksize - 1)) == 0) {
2311                 /* Block boundary? Nothing to do */
2312                 page = NULL;
2313         } else {
2314                 page = grab_cache_page(mapping,
2315                                 inode->i_size >> PAGE_CACHE_SHIFT);
2316                 if (!page)
2317                         return;
2318         }
2319
2320         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
2321                 ext4_ext_truncate(inode, page);
2322                 return;
2323         }
2324
2325         handle = start_transaction(inode);
2326         if (IS_ERR(handle)) {
2327                 if (page) {
2328                         clear_highpage(page);
2329                         flush_dcache_page(page);
2330                         unlock_page(page);
2331                         page_cache_release(page);
2332                 }
2333                 return;         /* AKPM: return what? */
2334         }
2335
2336         last_block = (inode->i_size + blocksize-1)
2337                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
2338
2339         if (page)
2340                 ext4_block_truncate_page(handle, page, mapping, inode->i_size);
2341
2342         n = ext4_block_to_path(inode, last_block, offsets, NULL);
2343         if (n == 0)
2344                 goto out_stop;  /* error */
2345
2346         /*
2347          * OK.  This truncate is going to happen.  We add the inode to the
2348          * orphan list, so that if this truncate spans multiple transactions,
2349          * and we crash, we will resume the truncate when the filesystem
2350          * recovers.  It also marks the inode dirty, to catch the new size.
2351          *
2352          * Implication: the file must always be in a sane, consistent
2353          * truncatable state while each transaction commits.
2354          */
2355         if (ext4_orphan_add(handle, inode))
2356                 goto out_stop;
2357
2358         /*
2359          * The orphan list entry will now protect us from any crash which
2360          * occurs before the truncate completes, so it is now safe to propagate
2361          * the new, shorter inode size (held for now in i_size) into the
2362          * on-disk inode. We do this via i_disksize, which is the value which
2363          * ext4 *really* writes onto the disk inode.
2364          */
2365         ei->i_disksize = inode->i_size;
2366
2367         /*
2368          * From here we block out all ext4_get_block() callers who want to
2369          * modify the block allocation tree.
2370          */
2371         down_write(&ei->i_data_sem);
2372
2373         if (n == 1) {           /* direct blocks */
2374                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
2375                                i_data + EXT4_NDIR_BLOCKS);
2376                 goto do_indirects;
2377         }
2378
2379         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
2380         /* Kill the top of shared branch (not detached) */
2381         if (nr) {
2382                 if (partial == chain) {
2383                         /* Shared branch grows from the inode */
2384                         ext4_free_branches(handle, inode, NULL,
2385                                            &nr, &nr+1, (chain+n-1) - partial);
2386                         *partial->p = 0;
2387                         /*
2388                          * We mark the inode dirty prior to restart,
2389                          * and prior to stop.  No need for it here.
2390                          */
2391                 } else {
2392                         /* Shared branch grows from an indirect block */
2393                         BUFFER_TRACE(partial->bh, "get_write_access");
2394                         ext4_free_branches(handle, inode, partial->bh,
2395                                         partial->p,
2396                                         partial->p+1, (chain+n-1) - partial);
2397                 }
2398         }
2399         /* Clear the ends of indirect blocks on the shared branch */
2400         while (partial > chain) {
2401                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
2402                                    (__le32*)partial->bh->b_data+addr_per_block,
2403                                    (chain+n-1) - partial);
2404                 BUFFER_TRACE(partial->bh, "call brelse");
2405                 brelse (partial->bh);
2406                 partial--;
2407         }
2408 do_indirects:
2409         /* Kill the remaining (whole) subtrees */
2410         switch (offsets[0]) {
2411         default:
2412                 nr = i_data[EXT4_IND_BLOCK];
2413                 if (nr) {
2414                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2415                         i_data[EXT4_IND_BLOCK] = 0;
2416                 }
2417         case EXT4_IND_BLOCK:
2418                 nr = i_data[EXT4_DIND_BLOCK];
2419                 if (nr) {
2420                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2421                         i_data[EXT4_DIND_BLOCK] = 0;
2422                 }
2423         case EXT4_DIND_BLOCK:
2424                 nr = i_data[EXT4_TIND_BLOCK];
2425                 if (nr) {
2426                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2427                         i_data[EXT4_TIND_BLOCK] = 0;
2428                 }
2429         case EXT4_TIND_BLOCK:
2430                 ;
2431         }
2432
2433         ext4_discard_reservation(inode);
2434
2435         up_write(&ei->i_data_sem);
2436         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
2437         ext4_mark_inode_dirty(handle, inode);
2438
2439         /*
2440          * In a multi-transaction truncate, we only make the final transaction
2441          * synchronous
2442          */
2443         if (IS_SYNC(inode))
2444                 handle->h_sync = 1;
2445 out_stop:
2446         /*
2447          * If this was a simple ftruncate(), and the file will remain alive
2448          * then we need to clear up the orphan record which we created above.
2449          * However, if this was a real unlink then we were called by
2450          * ext4_delete_inode(), and we allow that function to clean up the
2451          * orphan info for us.
2452          */
2453         if (inode->i_nlink)
2454                 ext4_orphan_del(handle, inode);
2455
2456         ext4_journal_stop(handle);
2457 }
2458
2459 static ext4_fsblk_t ext4_get_inode_block(struct super_block *sb,
2460                 unsigned long ino, struct ext4_iloc *iloc)
2461 {
2462         unsigned long desc, group_desc;
2463         ext4_group_t block_group;
2464         unsigned long offset;
2465         ext4_fsblk_t block;
2466         struct buffer_head *bh;
2467         struct ext4_group_desc * gdp;
2468
2469         if (!ext4_valid_inum(sb, ino)) {
2470                 /*
2471                  * This error is already checked for in namei.c unless we are
2472                  * looking at an NFS filehandle, in which case no error
2473                  * report is needed
2474                  */
2475                 return 0;
2476         }
2477
2478         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
2479         if (block_group >= EXT4_SB(sb)->s_groups_count) {
2480                 ext4_error(sb,"ext4_get_inode_block","group >= groups count");
2481                 return 0;
2482         }
2483         smp_rmb();
2484         group_desc = block_group >> EXT4_DESC_PER_BLOCK_BITS(sb);
2485         desc = block_group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2486         bh = EXT4_SB(sb)->s_group_desc[group_desc];
2487         if (!bh) {
2488                 ext4_error (sb, "ext4_get_inode_block",
2489                             "Descriptor not loaded");
2490                 return 0;
2491         }
2492
2493         gdp = (struct ext4_group_desc *)((__u8 *)bh->b_data +
2494                 desc * EXT4_DESC_SIZE(sb));
2495         /*
2496          * Figure out the offset within the block group inode table
2497          */
2498         offset = ((ino - 1) % EXT4_INODES_PER_GROUP(sb)) *
2499                 EXT4_INODE_SIZE(sb);
2500         block = ext4_inode_table(sb, gdp) +
2501                 (offset >> EXT4_BLOCK_SIZE_BITS(sb));
2502
2503         iloc->block_group = block_group;
2504         iloc->offset = offset & (EXT4_BLOCK_SIZE(sb) - 1);
2505         return block;
2506 }
2507
2508 /*
2509  * ext4_get_inode_loc returns with an extra refcount against the inode's
2510  * underlying buffer_head on success. If 'in_mem' is true, we have all
2511  * data in memory that is needed to recreate the on-disk version of this
2512  * inode.
2513  */
2514 static int __ext4_get_inode_loc(struct inode *inode,
2515                                 struct ext4_iloc *iloc, int in_mem)
2516 {
2517         ext4_fsblk_t block;
2518         struct buffer_head *bh;
2519
2520         block = ext4_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2521         if (!block)
2522                 return -EIO;
2523
2524         bh = sb_getblk(inode->i_sb, block);
2525         if (!bh) {
2526                 ext4_error (inode->i_sb, "ext4_get_inode_loc",
2527                                 "unable to read inode block - "
2528                                 "inode=%lu, block=%llu",
2529                                  inode->i_ino, block);
2530                 return -EIO;
2531         }
2532         if (!buffer_uptodate(bh)) {
2533                 lock_buffer(bh);
2534                 if (buffer_uptodate(bh)) {
2535                         /* someone brought it uptodate while we waited */
2536                         unlock_buffer(bh);
2537                         goto has_buffer;
2538                 }
2539
2540                 /*
2541                  * If we have all information of the inode in memory and this
2542                  * is the only valid inode in the block, we need not read the
2543                  * block.
2544                  */
2545                 if (in_mem) {
2546                         struct buffer_head *bitmap_bh;
2547                         struct ext4_group_desc *desc;
2548                         int inodes_per_buffer;
2549                         int inode_offset, i;
2550                         ext4_group_t block_group;
2551                         int start;
2552
2553                         block_group = (inode->i_ino - 1) /
2554                                         EXT4_INODES_PER_GROUP(inode->i_sb);
2555                         inodes_per_buffer = bh->b_size /
2556                                 EXT4_INODE_SIZE(inode->i_sb);
2557                         inode_offset = ((inode->i_ino - 1) %
2558                                         EXT4_INODES_PER_GROUP(inode->i_sb));
2559                         start = inode_offset & ~(inodes_per_buffer - 1);
2560
2561                         /* Is the inode bitmap in cache? */
2562                         desc = ext4_get_group_desc(inode->i_sb,
2563                                                 block_group, NULL);
2564                         if (!desc)
2565                                 goto make_io;
2566
2567                         bitmap_bh = sb_getblk(inode->i_sb,
2568                                 ext4_inode_bitmap(inode->i_sb, desc));
2569                         if (!bitmap_bh)
2570                                 goto make_io;
2571
2572                         /*
2573                          * If the inode bitmap isn't in cache then the
2574                          * optimisation may end up performing two reads instead
2575                          * of one, so skip it.
2576                          */
2577                         if (!buffer_uptodate(bitmap_bh)) {
2578                                 brelse(bitmap_bh);
2579                                 goto make_io;
2580                         }
2581                         for (i = start; i < start + inodes_per_buffer; i++) {
2582                                 if (i == inode_offset)
2583                                         continue;
2584                                 if (ext4_test_bit(i, bitmap_bh->b_data))
2585                                         break;
2586                         }
2587                         brelse(bitmap_bh);
2588                         if (i == start + inodes_per_buffer) {
2589                                 /* all other inodes are free, so skip I/O */
2590                                 memset(bh->b_data, 0, bh->b_size);
2591                                 set_buffer_uptodate(bh);
2592                                 unlock_buffer(bh);
2593                                 goto has_buffer;
2594                         }
2595                 }
2596
2597 make_io:
2598                 /*
2599                  * There are other valid inodes in the buffer, this inode
2600                  * has in-inode xattrs, or we don't have this inode in memory.
2601                  * Read the block from disk.
2602                  */
2603                 get_bh(bh);
2604                 bh->b_end_io = end_buffer_read_sync;
2605                 submit_bh(READ_META, bh);
2606                 wait_on_buffer(bh);
2607                 if (!buffer_uptodate(bh)) {
2608                         ext4_error(inode->i_sb, "ext4_get_inode_loc",
2609                                         "unable to read inode block - "
2610                                         "inode=%lu, block=%llu",
2611                                         inode->i_ino, block);
2612                         brelse(bh);
2613                         return -EIO;
2614                 }
2615         }
2616 has_buffer:
2617         iloc->bh = bh;
2618         return 0;
2619 }
2620
2621 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
2622 {
2623         /* We have all inode data except xattrs in memory here. */
2624         return __ext4_get_inode_loc(inode, iloc,
2625                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
2626 }
2627
2628 void ext4_set_inode_flags(struct inode *inode)
2629 {
2630         unsigned int flags = EXT4_I(inode)->i_flags;
2631
2632         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2633         if (flags & EXT4_SYNC_FL)
2634                 inode->i_flags |= S_SYNC;
2635         if (flags & EXT4_APPEND_FL)
2636                 inode->i_flags |= S_APPEND;
2637         if (flags & EXT4_IMMUTABLE_FL)
2638                 inode->i_flags |= S_IMMUTABLE;
2639         if (flags & EXT4_NOATIME_FL)
2640                 inode->i_flags |= S_NOATIME;
2641         if (flags & EXT4_DIRSYNC_FL)
2642                 inode->i_flags |= S_DIRSYNC;
2643 }
2644
2645 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
2646 void ext4_get_inode_flags(struct ext4_inode_info *ei)
2647 {
2648         unsigned int flags = ei->vfs_inode.i_flags;
2649
2650         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
2651                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
2652         if (flags & S_SYNC)
2653                 ei->i_flags |= EXT4_SYNC_FL;
2654         if (flags & S_APPEND)
2655                 ei->i_flags |= EXT4_APPEND_FL;
2656         if (flags & S_IMMUTABLE)
2657                 ei->i_flags |= EXT4_IMMUTABLE_FL;
2658         if (flags & S_NOATIME)
2659                 ei->i_flags |= EXT4_NOATIME_FL;
2660         if (flags & S_DIRSYNC)
2661                 ei->i_flags |= EXT4_DIRSYNC_FL;
2662 }
2663 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
2664                                         struct ext4_inode_info *ei)
2665 {
2666         blkcnt_t i_blocks ;
2667         struct inode *inode = &(ei->vfs_inode);
2668         struct super_block *sb = inode->i_sb;
2669
2670         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
2671                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
2672                 /* we are using combined 48 bit field */
2673                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
2674                                         le32_to_cpu(raw_inode->i_blocks_lo);
2675                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
2676                         /* i_blocks represent file system block size */
2677                         return i_blocks  << (inode->i_blkbits - 9);
2678                 } else {
2679                         return i_blocks;
2680                 }
2681         } else {
2682                 return le32_to_cpu(raw_inode->i_blocks_lo);
2683         }
2684 }
2685
2686 void ext4_read_inode(struct inode * inode)
2687 {
2688         struct ext4_iloc iloc;
2689         struct ext4_inode *raw_inode;
2690         struct ext4_inode_info *ei = EXT4_I(inode);
2691         struct buffer_head *bh;
2692         int block;
2693
2694 #ifdef CONFIG_EXT4DEV_FS_POSIX_ACL
2695         ei->i_acl = EXT4_ACL_NOT_CACHED;
2696         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
2697 #endif
2698         ei->i_block_alloc_info = NULL;
2699
2700         if (__ext4_get_inode_loc(inode, &iloc, 0))
2701                 goto bad_inode;
2702         bh = iloc.bh;
2703         raw_inode = ext4_raw_inode(&iloc);
2704         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2705         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2706         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2707         if(!(test_opt (inode->i_sb, NO_UID32))) {
2708                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2709                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2710         }
2711         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2712
2713         ei->i_state = 0;
2714         ei->i_dir_start_lookup = 0;
2715         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2716         /* We now have enough fields to check if the inode was active or not.
2717          * This is needed because nfsd might try to access dead inodes
2718          * the test is that same one that e2fsck uses
2719          * NeilBrown 1999oct15
2720          */
2721         if (inode->i_nlink == 0) {
2722                 if (inode->i_mode == 0 ||
2723                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
2724                         /* this inode is deleted */
2725                         brelse (bh);
2726                         goto bad_inode;
2727                 }
2728                 /* The only unlinked inodes we let through here have
2729                  * valid i_mode and are being read by the orphan
2730                  * recovery code: that's fine, we're about to complete
2731                  * the process of deleting those. */
2732         }
2733         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2734         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
2735         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
2736         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
2737             cpu_to_le32(EXT4_OS_HURD)) {
2738                 ei->i_file_acl |=
2739                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
2740         }
2741         inode->i_size = ext4_isize(raw_inode);
2742         ei->i_disksize = inode->i_size;
2743         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2744         ei->i_block_group = iloc.block_group;
2745         /*
2746          * NOTE! The in-memory inode i_data array is in little-endian order
2747          * even on big-endian machines: we do NOT byteswap the block numbers!
2748          */
2749         for (block = 0; block < EXT4_N_BLOCKS; block++)
2750                 ei->i_data[block] = raw_inode->i_block[block];
2751         INIT_LIST_HEAD(&ei->i_orphan);
2752
2753         if (inode->i_ino >= EXT4_FIRST_INO(inode->i_sb) + 1 &&
2754             EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
2755                 /*
2756                  * When mke2fs creates big inodes it does not zero out
2757                  * the unused bytes above EXT4_GOOD_OLD_INODE_SIZE,
2758                  * so ignore those first few inodes.
2759                  */
2760                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2761                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2762                     EXT4_INODE_SIZE(inode->i_sb)) {
2763                         brelse (bh);
2764                         goto bad_inode;
2765                 }
2766                 if (ei->i_extra_isize == 0) {
2767                         /* The extra space is currently unused. Use it. */
2768                         ei->i_extra_isize = sizeof(struct ext4_inode) -
2769                                             EXT4_GOOD_OLD_INODE_SIZE;
2770                 } else {
2771                         __le32 *magic = (void *)raw_inode +
2772                                         EXT4_GOOD_OLD_INODE_SIZE +
2773                                         ei->i_extra_isize;
2774                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
2775                                  ei->i_state |= EXT4_STATE_XATTR;
2776                 }
2777         } else
2778                 ei->i_extra_isize = 0;
2779
2780         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
2781         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
2782         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
2783         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
2784
2785         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
2786         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
2787                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
2788                         inode->i_version |=
2789                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
2790         }
2791
2792         if (S_ISREG(inode->i_mode)) {
2793                 inode->i_op = &ext4_file_inode_operations;
2794                 inode->i_fop = &ext4_file_operations;
2795                 ext4_set_aops(inode);
2796         } else if (S_ISDIR(inode->i_mode)) {
2797                 inode->i_op = &ext4_dir_inode_operations;
2798                 inode->i_fop = &ext4_dir_operations;
2799         } else if (S_ISLNK(inode->i_mode)) {
2800                 if (ext4_inode_is_fast_symlink(inode))
2801                         inode->i_op = &ext4_fast_symlink_inode_operations;
2802                 else {
2803                         inode->i_op = &ext4_symlink_inode_operations;
2804                         ext4_set_aops(inode);
2805                 }
2806         } else {
2807                 inode->i_op = &ext4_special_inode_operations;
2808                 if (raw_inode->i_block[0])
2809                         init_special_inode(inode, inode->i_mode,
2810                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2811                 else
2812                         init_special_inode(inode, inode->i_mode,
2813                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2814         }
2815         brelse (iloc.bh);
2816         ext4_set_inode_flags(inode);
2817         return;
2818
2819 bad_inode:
2820         make_bad_inode(inode);
2821         return;
2822 }
2823
2824 static int ext4_inode_blocks_set(handle_t *handle,
2825                                 struct ext4_inode *raw_inode,
2826                                 struct ext4_inode_info *ei)
2827 {
2828         struct inode *inode = &(ei->vfs_inode);
2829         u64 i_blocks = inode->i_blocks;
2830         struct super_block *sb = inode->i_sb;
2831         int err = 0;
2832
2833         if (i_blocks <= ~0U) {
2834                 /*
2835                  * i_blocks can be represnted in a 32 bit variable
2836                  * as multiple of 512 bytes
2837                  */
2838                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
2839                 raw_inode->i_blocks_high = 0;
2840                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
2841         } else if (i_blocks <= 0xffffffffffffULL) {
2842                 /*
2843                  * i_blocks can be represented in a 48 bit variable
2844                  * as multiple of 512 bytes
2845                  */
2846                 err = ext4_update_rocompat_feature(handle, sb,
2847                                             EXT4_FEATURE_RO_COMPAT_HUGE_FILE);
2848                 if (err)
2849                         goto  err_out;
2850                 /* i_block is stored in the split  48 bit fields */
2851                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
2852                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
2853                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
2854         } else {
2855                 /*
2856                  * i_blocks should be represented in a 48 bit variable
2857                  * as multiple of  file system block size
2858                  */
2859                 err = ext4_update_rocompat_feature(handle, sb,
2860                                             EXT4_FEATURE_RO_COMPAT_HUGE_FILE);
2861                 if (err)
2862                         goto  err_out;
2863                 ei->i_flags |= EXT4_HUGE_FILE_FL;
2864                 /* i_block is stored in file system block size */
2865                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
2866                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
2867                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
2868         }
2869 err_out:
2870         return err;
2871 }
2872
2873 /*
2874  * Post the struct inode info into an on-disk inode location in the
2875  * buffer-cache.  This gobbles the caller's reference to the
2876  * buffer_head in the inode location struct.
2877  *
2878  * The caller must have write access to iloc->bh.
2879  */
2880 static int ext4_do_update_inode(handle_t *handle,
2881                                 struct inode *inode,
2882                                 struct ext4_iloc *iloc)
2883 {
2884         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
2885         struct ext4_inode_info *ei = EXT4_I(inode);
2886         struct buffer_head *bh = iloc->bh;
2887         int err = 0, rc, block;
2888
2889         /* For fields not not tracking in the in-memory inode,
2890          * initialise them to zero for new inodes. */
2891         if (ei->i_state & EXT4_STATE_NEW)
2892                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
2893
2894         ext4_get_inode_flags(ei);
2895         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2896         if(!(test_opt(inode->i_sb, NO_UID32))) {
2897                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2898                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2899 /*
2900  * Fix up interoperability with old kernels. Otherwise, old inodes get
2901  * re-used with the upper 16 bits of the uid/gid intact
2902  */
2903                 if(!ei->i_dtime) {
2904                         raw_inode->i_uid_high =
2905                                 cpu_to_le16(high_16_bits(inode->i_uid));
2906                         raw_inode->i_gid_high =
2907                                 cpu_to_le16(high_16_bits(inode->i_gid));
2908                 } else {
2909                         raw_inode->i_uid_high = 0;
2910                         raw_inode->i_gid_high = 0;
2911                 }
2912         } else {
2913                 raw_inode->i_uid_low =
2914                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2915                 raw_inode->i_gid_low =
2916                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2917                 raw_inode->i_uid_high = 0;
2918                 raw_inode->i_gid_high = 0;
2919         }
2920         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2921
2922         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
2923         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
2924         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
2925         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
2926
2927         if (ext4_inode_blocks_set(handle, raw_inode, ei))
2928                 goto out_brelse;
2929         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2930         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2931         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
2932             cpu_to_le32(EXT4_OS_HURD))
2933                 raw_inode->i_file_acl_high =
2934                         cpu_to_le16(ei->i_file_acl >> 32);
2935         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
2936         ext4_isize_set(raw_inode, ei->i_disksize);
2937         if (ei->i_disksize > 0x7fffffffULL) {
2938                 struct super_block *sb = inode->i_sb;
2939                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
2940                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
2941                                 EXT4_SB(sb)->s_es->s_rev_level ==
2942                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
2943                         /* If this is the first large file
2944                          * created, add a flag to the superblock.
2945                          */
2946                         err = ext4_journal_get_write_access(handle,
2947                                         EXT4_SB(sb)->s_sbh);
2948                         if (err)
2949                                 goto out_brelse;
2950                         ext4_update_dynamic_rev(sb);
2951                         EXT4_SET_RO_COMPAT_FEATURE(sb,
2952                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
2953                         sb->s_dirt = 1;
2954                         handle->h_sync = 1;
2955                         err = ext4_journal_dirty_metadata(handle,
2956                                         EXT4_SB(sb)->s_sbh);
2957                 }
2958         }
2959         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2960         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2961                 if (old_valid_dev(inode->i_rdev)) {
2962                         raw_inode->i_block[0] =
2963                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2964                         raw_inode->i_block[1] = 0;
2965                 } else {
2966                         raw_inode->i_block[0] = 0;
2967                         raw_inode->i_block[1] =
2968                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2969                         raw_inode->i_block[2] = 0;
2970                 }
2971         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
2972                 raw_inode->i_block[block] = ei->i_data[block];
2973
2974         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
2975         if (ei->i_extra_isize) {
2976                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
2977                         raw_inode->i_version_hi =
2978                         cpu_to_le32(inode->i_version >> 32);
2979                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2980         }
2981
2982
2983         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
2984         rc = ext4_journal_dirty_metadata(handle, bh);
2985         if (!err)
2986                 err = rc;
2987         ei->i_state &= ~EXT4_STATE_NEW;
2988
2989 out_brelse:
2990         brelse (bh);
2991         ext4_std_error(inode->i_sb, err);
2992         return err;
2993 }
2994
2995 /*
2996  * ext4_write_inode()
2997  *
2998  * We are called from a few places:
2999  *
3000  * - Within generic_file_write() for O_SYNC files.
3001  *   Here, there will be no transaction running. We wait for any running
3002  *   trasnaction to commit.
3003  *
3004  * - Within sys_sync(), kupdate and such.
3005  *   We wait on commit, if tol to.
3006  *
3007  * - Within prune_icache() (PF_MEMALLOC == true)
3008  *   Here we simply return.  We can't afford to block kswapd on the
3009  *   journal commit.
3010  *
3011  * In all cases it is actually safe for us to return without doing anything,
3012  * because the inode has been copied into a raw inode buffer in
3013  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3014  * knfsd.
3015  *
3016  * Note that we are absolutely dependent upon all inode dirtiers doing the
3017  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3018  * which we are interested.
3019  *
3020  * It would be a bug for them to not do this.  The code:
3021  *
3022  *      mark_inode_dirty(inode)
3023  *      stuff();
3024  *      inode->i_size = expr;
3025  *
3026  * is in error because a kswapd-driven write_inode() could occur while
3027  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3028  * will no longer be on the superblock's dirty inode list.
3029  */
3030 int ext4_write_inode(struct inode *inode, int wait)
3031 {
3032         if (current->flags & PF_MEMALLOC)
3033                 return 0;
3034
3035         if (ext4_journal_current_handle()) {
3036                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3037                 dump_stack();
3038                 return -EIO;
3039         }
3040
3041         if (!wait)
3042                 return 0;
3043
3044         return ext4_force_commit(inode->i_sb);
3045 }
3046
3047 /*
3048  * ext4_setattr()
3049  *
3050  * Called from notify_change.
3051  *
3052  * We want to trap VFS attempts to truncate the file as soon as
3053  * possible.  In particular, we want to make sure that when the VFS
3054  * shrinks i_size, we put the inode on the orphan list and modify
3055  * i_disksize immediately, so that during the subsequent flushing of
3056  * dirty pages and freeing of disk blocks, we can guarantee that any
3057  * commit will leave the blocks being flushed in an unused state on
3058  * disk.  (On recovery, the inode will get truncated and the blocks will
3059  * be freed, so we have a strong guarantee that no future commit will
3060  * leave these blocks visible to the user.)
3061  *
3062  * Called with inode->sem down.
3063  */
3064 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
3065 {
3066         struct inode *inode = dentry->d_inode;
3067         int error, rc = 0;
3068         const unsigned int ia_valid = attr->ia_valid;
3069
3070         error = inode_change_ok(inode, attr);
3071         if (error)
3072                 return error;
3073
3074         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3075                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3076                 handle_t *handle;
3077
3078                 /* (user+group)*(old+new) structure, inode write (sb,
3079                  * inode block, ? - but truncate inode update has it) */
3080                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
3081                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3082                 if (IS_ERR(handle)) {
3083                         error = PTR_ERR(handle);
3084                         goto err_out;
3085                 }
3086                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
3087                 if (error) {
3088                         ext4_journal_stop(handle);
3089                         return error;
3090                 }
3091                 /* Update corresponding info in inode so that everything is in
3092                  * one transaction */
3093                 if (attr->ia_valid & ATTR_UID)
3094                         inode->i_uid = attr->ia_uid;
3095                 if (attr->ia_valid & ATTR_GID)
3096                         inode->i_gid = attr->ia_gid;
3097                 error = ext4_mark_inode_dirty(handle, inode);
3098                 ext4_journal_stop(handle);
3099         }
3100
3101         if (attr->ia_valid & ATTR_SIZE) {
3102                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
3103                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3104
3105                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
3106                                 error = -EFBIG;
3107                                 goto err_out;
3108                         }
3109                 }
3110         }
3111
3112         if (S_ISREG(inode->i_mode) &&
3113             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3114                 handle_t *handle;
3115
3116                 handle = ext4_journal_start(inode, 3);
3117                 if (IS_ERR(handle)) {
3118                         error = PTR_ERR(handle);
3119                         goto err_out;
3120                 }
3121
3122                 error = ext4_orphan_add(handle, inode);
3123                 EXT4_I(inode)->i_disksize = attr->ia_size;
3124                 rc = ext4_mark_inode_dirty(handle, inode);
3125                 if (!error)
3126                         error = rc;
3127                 ext4_journal_stop(handle);
3128         }
3129
3130         rc = inode_setattr(inode, attr);
3131
3132         /* If inode_setattr's call to ext4_truncate failed to get a
3133          * transaction handle at all, we need to clean up the in-core
3134          * orphan list manually. */
3135         if (inode->i_nlink)
3136                 ext4_orphan_del(NULL, inode);
3137
3138         if (!rc && (ia_valid & ATTR_MODE))
3139                 rc = ext4_acl_chmod(inode);
3140
3141 err_out:
3142         ext4_std_error(inode->i_sb, error);
3143         if (!error)
3144                 error = rc;
3145         return error;
3146 }
3147
3148
3149 /*
3150  * How many blocks doth make a writepage()?
3151  *
3152  * With N blocks per page, it may be:
3153  * N data blocks
3154  * 2 indirect block
3155  * 2 dindirect
3156  * 1 tindirect
3157  * N+5 bitmap blocks (from the above)
3158  * N+5 group descriptor summary blocks
3159  * 1 inode block
3160  * 1 superblock.
3161  * 2 * EXT4_SINGLEDATA_TRANS_BLOCKS for the quote files
3162  *
3163  * 3 * (N + 5) + 2 + 2 * EXT4_SINGLEDATA_TRANS_BLOCKS
3164  *
3165  * With ordered or writeback data it's the same, less the N data blocks.
3166  *
3167  * If the inode's direct blocks can hold an integral number of pages then a
3168  * page cannot straddle two indirect blocks, and we can only touch one indirect
3169  * and dindirect block, and the "5" above becomes "3".
3170  *
3171  * This still overestimates under most circumstances.  If we were to pass the
3172  * start and end offsets in here as well we could do block_to_path() on each
3173  * block and work out the exact number of indirects which are touched.  Pah.
3174  */
3175
3176 int ext4_writepage_trans_blocks(struct inode *inode)
3177 {
3178         int bpp = ext4_journal_blocks_per_page(inode);
3179         int indirects = (EXT4_NDIR_BLOCKS % bpp) ? 5 : 3;
3180         int ret;
3181
3182         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3183                 return ext4_ext_writepage_trans_blocks(inode, bpp);
3184
3185         if (ext4_should_journal_data(inode))
3186                 ret = 3 * (bpp + indirects) + 2;
3187         else
3188                 ret = 2 * (bpp + indirects) + 2;
3189
3190 #ifdef CONFIG_QUOTA
3191         /* We know that structure was already allocated during DQUOT_INIT so
3192          * we will be updating only the data blocks + inodes */
3193         ret += 2*EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb);
3194 #endif
3195
3196         return ret;
3197 }
3198
3199 /*
3200  * The caller must have previously called ext4_reserve_inode_write().
3201  * Give this, we know that the caller already has write access to iloc->bh.
3202  */
3203 int ext4_mark_iloc_dirty(handle_t *handle,
3204                 struct inode *inode, struct ext4_iloc *iloc)
3205 {
3206         int err = 0;
3207
3208         if (test_opt(inode->i_sb, I_VERSION))
3209                 inode_inc_iversion(inode);
3210
3211         /* the do_update_inode consumes one bh->b_count */
3212         get_bh(iloc->bh);
3213
3214         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
3215         err = ext4_do_update_inode(handle, inode, iloc);
3216         put_bh(iloc->bh);
3217         return err;
3218 }
3219
3220 /*
3221  * On success, We end up with an outstanding reference count against
3222  * iloc->bh.  This _must_ be cleaned up later.
3223  */
3224
3225 int
3226 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
3227                          struct ext4_iloc *iloc)
3228 {
3229         int err = 0;
3230         if (handle) {
3231                 err = ext4_get_inode_loc(inode, iloc);
3232                 if (!err) {
3233                         BUFFER_TRACE(iloc->bh, "get_write_access");
3234                         err = ext4_journal_get_write_access(handle, iloc->bh);
3235                         if (err) {
3236                                 brelse(iloc->bh);
3237                                 iloc->bh = NULL;
3238                         }
3239                 }
3240         }
3241         ext4_std_error(inode->i_sb, err);
3242         return err;
3243 }
3244
3245 /*
3246  * Expand an inode by new_extra_isize bytes.
3247  * Returns 0 on success or negative error number on failure.
3248  */
3249 static int ext4_expand_extra_isize(struct inode *inode,
3250                                    unsigned int new_extra_isize,
3251                                    struct ext4_iloc iloc,
3252                                    handle_t *handle)
3253 {
3254         struct ext4_inode *raw_inode;
3255         struct ext4_xattr_ibody_header *header;
3256         struct ext4_xattr_entry *entry;
3257
3258         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
3259                 return 0;
3260
3261         raw_inode = ext4_raw_inode(&iloc);
3262
3263         header = IHDR(inode, raw_inode);
3264         entry = IFIRST(header);
3265
3266         /* No extended attributes present */
3267         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
3268                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
3269                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
3270                         new_extra_isize);
3271                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
3272                 return 0;
3273         }
3274
3275         /* try to expand with EAs present */
3276         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
3277                                           raw_inode, handle);
3278 }
3279
3280 /*
3281  * What we do here is to mark the in-core inode as clean with respect to inode
3282  * dirtiness (it may still be data-dirty).
3283  * This means that the in-core inode may be reaped by prune_icache
3284  * without having to perform any I/O.  This is a very good thing,
3285  * because *any* task may call prune_icache - even ones which
3286  * have a transaction open against a different journal.
3287  *
3288  * Is this cheating?  Not really.  Sure, we haven't written the
3289  * inode out, but prune_icache isn't a user-visible syncing function.
3290  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3291  * we start and wait on commits.
3292  *
3293  * Is this efficient/effective?  Well, we're being nice to the system
3294  * by cleaning up our inodes proactively so they can be reaped
3295  * without I/O.  But we are potentially leaving up to five seconds'
3296  * worth of inodes floating about which prune_icache wants us to
3297  * write out.  One way to fix that would be to get prune_icache()
3298  * to do a write_super() to free up some memory.  It has the desired
3299  * effect.
3300  */
3301 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
3302 {
3303         struct ext4_iloc iloc;
3304         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3305         static unsigned int mnt_count;
3306         int err, ret;
3307
3308         might_sleep();
3309         err = ext4_reserve_inode_write(handle, inode, &iloc);
3310         if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
3311             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
3312                 /*
3313                  * We need extra buffer credits since we may write into EA block
3314                  * with this same handle. If journal_extend fails, then it will
3315                  * only result in a minor loss of functionality for that inode.
3316                  * If this is felt to be critical, then e2fsck should be run to
3317                  * force a large enough s_min_extra_isize.
3318                  */
3319                 if ((jbd2_journal_extend(handle,
3320                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
3321                         ret = ext4_expand_extra_isize(inode,
3322                                                       sbi->s_want_extra_isize,
3323                                                       iloc, handle);
3324                         if (ret) {
3325                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
3326                                 if (mnt_count !=
3327                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
3328                                         ext4_warning(inode->i_sb, __FUNCTION__,
3329                                         "Unable to expand inode %lu. Delete"
3330                                         " some EAs or run e2fsck.",
3331                                         inode->i_ino);
3332                                         mnt_count =
3333                                           le16_to_cpu(sbi->s_es->s_mnt_count);
3334                                 }
3335                         }
3336                 }
3337         }
3338         if (!err)
3339                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
3340         return err;
3341 }
3342
3343 /*
3344  * ext4_dirty_inode() is called from __mark_inode_dirty()
3345  *
3346  * We're really interested in the case where a file is being extended.
3347  * i_size has been changed by generic_commit_write() and we thus need
3348  * to include the updated inode in the current transaction.
3349  *
3350  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
3351  * are allocated to the file.
3352  *
3353  * If the inode is marked synchronous, we don't honour that here - doing
3354  * so would cause a commit on atime updates, which we don't bother doing.
3355  * We handle synchronous inodes at the highest possible level.
3356  */
3357 void ext4_dirty_inode(struct inode *inode)
3358 {
3359         handle_t *current_handle = ext4_journal_current_handle();
3360         handle_t *handle;
3361
3362         handle = ext4_journal_start(inode, 2);
3363         if (IS_ERR(handle))
3364                 goto out;
3365         if (current_handle &&
3366                 current_handle->h_transaction != handle->h_transaction) {
3367                 /* This task has a transaction open against a different fs */
3368                 printk(KERN_EMERG "%s: transactions do not match!\n",
3369                        __FUNCTION__);
3370         } else {
3371                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3372                                 current_handle);
3373                 ext4_mark_inode_dirty(handle, inode);
3374         }
3375         ext4_journal_stop(handle);
3376 out:
3377         return;
3378 }
3379
3380 #if 0
3381 /*
3382  * Bind an inode's backing buffer_head into this transaction, to prevent
3383  * it from being flushed to disk early.  Unlike
3384  * ext4_reserve_inode_write, this leaves behind no bh reference and
3385  * returns no iloc structure, so the caller needs to repeat the iloc
3386  * lookup to mark the inode dirty later.
3387  */
3388 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
3389 {
3390         struct ext4_iloc iloc;
3391
3392         int err = 0;
3393         if (handle) {
3394                 err = ext4_get_inode_loc(inode, &iloc);
3395                 if (!err) {
3396                         BUFFER_TRACE(iloc.bh, "get_write_access");
3397                         err = jbd2_journal_get_write_access(handle, iloc.bh);
3398                         if (!err)
3399                                 err = ext4_journal_dirty_metadata(handle,
3400                                                                   iloc.bh);
3401                         brelse(iloc.bh);
3402                 }
3403         }
3404         ext4_std_error(inode->i_sb, err);
3405         return err;
3406 }
3407 #endif
3408
3409 int ext4_change_inode_journal_flag(struct inode *inode, int val)
3410 {
3411         journal_t *journal;
3412         handle_t *handle;
3413         int err;
3414
3415         /*
3416          * We have to be very careful here: changing a data block's
3417          * journaling status dynamically is dangerous.  If we write a
3418          * data block to the journal, change the status and then delete
3419          * that block, we risk forgetting to revoke the old log record
3420          * from the journal and so a subsequent replay can corrupt data.
3421          * So, first we make sure that the journal is empty and that
3422          * nobody is changing anything.
3423          */
3424
3425         journal = EXT4_JOURNAL(inode);
3426         if (is_journal_aborted(journal))
3427                 return -EROFS;
3428
3429         jbd2_journal_lock_updates(journal);
3430         jbd2_journal_flush(journal);
3431
3432         /*
3433          * OK, there are no updates running now, and all cached data is
3434          * synced to disk.  We are now in a completely consistent state
3435          * which doesn't have anything in the journal, and we know that
3436          * no filesystem updates are running, so it is safe to modify
3437          * the inode's in-core data-journaling state flag now.
3438          */
3439
3440         if (val)
3441                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
3442         else
3443                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
3444         ext4_set_aops(inode);
3445
3446         jbd2_journal_unlock_updates(journal);
3447
3448         /* Finally we can mark the inode as dirty. */
3449
3450         handle = ext4_journal_start(inode, 1);
3451         if (IS_ERR(handle))
3452                 return PTR_ERR(handle);
3453
3454         err = ext4_mark_inode_dirty(handle, inode);
3455         handle->h_sync = 1;
3456         ext4_journal_stop(handle);
3457         ext4_std_error(inode->i_sb, err);
3458
3459         return err;
3460 }