d12f961b8ae584b50172b809d1f90454c1ba4b1b
[pandora-kernel.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include <linux/fiemap.h>
40 #include <linux/namei.h>
41 #include <trace/events/ext3.h>
42 #include "xattr.h"
43 #include "acl.h"
44
45 static int ext3_writepage_trans_blocks(struct inode *inode);
46 static int ext3_block_truncate_page(struct inode *inode, loff_t from);
47
48 /*
49  * Test whether an inode is a fast symlink.
50  */
51 static int ext3_inode_is_fast_symlink(struct inode *inode)
52 {
53         int ea_blocks = EXT3_I(inode)->i_file_acl ?
54                 (inode->i_sb->s_blocksize >> 9) : 0;
55
56         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
57 }
58
59 /*
60  * The ext3 forget function must perform a revoke if we are freeing data
61  * which has been journaled.  Metadata (eg. indirect blocks) must be
62  * revoked in all cases.
63  *
64  * "bh" may be NULL: a metadata block may have been freed from memory
65  * but there may still be a record of it in the journal, and that record
66  * still needs to be revoked.
67  */
68 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
69                         struct buffer_head *bh, ext3_fsblk_t blocknr)
70 {
71         int err;
72
73         might_sleep();
74
75         trace_ext3_forget(inode, is_metadata, blocknr);
76         BUFFER_TRACE(bh, "enter");
77
78         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
79                   "data mode %lx\n",
80                   bh, is_metadata, inode->i_mode,
81                   test_opt(inode->i_sb, DATA_FLAGS));
82
83         /* Never use the revoke function if we are doing full data
84          * journaling: there is no need to, and a V1 superblock won't
85          * support it.  Otherwise, only skip the revoke on un-journaled
86          * data blocks. */
87
88         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
89             (!is_metadata && !ext3_should_journal_data(inode))) {
90                 if (bh) {
91                         BUFFER_TRACE(bh, "call journal_forget");
92                         return ext3_journal_forget(handle, bh);
93                 }
94                 return 0;
95         }
96
97         /*
98          * data!=journal && (is_metadata || should_journal_data(inode))
99          */
100         BUFFER_TRACE(bh, "call ext3_journal_revoke");
101         err = ext3_journal_revoke(handle, blocknr, bh);
102         if (err)
103                 ext3_abort(inode->i_sb, __func__,
104                            "error %d when attempting revoke", err);
105         BUFFER_TRACE(bh, "exit");
106         return err;
107 }
108
109 /*
110  * Work out how many blocks we need to proceed with the next chunk of a
111  * truncate transaction.
112  */
113 static unsigned long blocks_for_truncate(struct inode *inode)
114 {
115         unsigned long needed;
116
117         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
118
119         /* Give ourselves just enough room to cope with inodes in which
120          * i_blocks is corrupt: we've seen disk corruptions in the past
121          * which resulted in random data in an inode which looked enough
122          * like a regular file for ext3 to try to delete it.  Things
123          * will go a bit crazy if that happens, but at least we should
124          * try not to panic the whole kernel. */
125         if (needed < 2)
126                 needed = 2;
127
128         /* But we need to bound the transaction so we don't overflow the
129          * journal. */
130         if (needed > EXT3_MAX_TRANS_DATA)
131                 needed = EXT3_MAX_TRANS_DATA;
132
133         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
134 }
135
136 /*
137  * Truncate transactions can be complex and absolutely huge.  So we need to
138  * be able to restart the transaction at a conventient checkpoint to make
139  * sure we don't overflow the journal.
140  *
141  * start_transaction gets us a new handle for a truncate transaction,
142  * and extend_transaction tries to extend the existing one a bit.  If
143  * extend fails, we need to propagate the failure up and restart the
144  * transaction in the top-level truncate loop. --sct
145  */
146 static handle_t *start_transaction(struct inode *inode)
147 {
148         handle_t *result;
149
150         result = ext3_journal_start(inode, blocks_for_truncate(inode));
151         if (!IS_ERR(result))
152                 return result;
153
154         ext3_std_error(inode->i_sb, PTR_ERR(result));
155         return result;
156 }
157
158 /*
159  * Try to extend this transaction for the purposes of truncation.
160  *
161  * Returns 0 if we managed to create more room.  If we can't create more
162  * room, and the transaction must be restarted we return 1.
163  */
164 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
165 {
166         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
167                 return 0;
168         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
169                 return 0;
170         return 1;
171 }
172
173 /*
174  * Restart the transaction associated with *handle.  This does a commit,
175  * so before we call here everything must be consistently dirtied against
176  * this transaction.
177  */
178 static int truncate_restart_transaction(handle_t *handle, struct inode *inode)
179 {
180         int ret;
181
182         jbd_debug(2, "restarting handle %p\n", handle);
183         /*
184          * Drop truncate_mutex to avoid deadlock with ext3_get_blocks_handle
185          * At this moment, get_block can be called only for blocks inside
186          * i_size since page cache has been already dropped and writes are
187          * blocked by i_mutex. So we can safely drop the truncate_mutex.
188          */
189         mutex_unlock(&EXT3_I(inode)->truncate_mutex);
190         ret = ext3_journal_restart(handle, blocks_for_truncate(inode));
191         mutex_lock(&EXT3_I(inode)->truncate_mutex);
192         return ret;
193 }
194
195 /*
196  * Called at inode eviction from icache
197  */
198 void ext3_evict_inode (struct inode *inode)
199 {
200         struct ext3_inode_info *ei = EXT3_I(inode);
201         struct ext3_block_alloc_info *rsv;
202         handle_t *handle;
203         int want_delete = 0;
204
205         trace_ext3_evict_inode(inode);
206         if (!inode->i_nlink && !is_bad_inode(inode)) {
207                 dquot_initialize(inode);
208                 want_delete = 1;
209         }
210
211         /*
212          * When journalling data dirty buffers are tracked only in the journal.
213          * So although mm thinks everything is clean and ready for reaping the
214          * inode might still have some pages to write in the running
215          * transaction or waiting to be checkpointed. Thus calling
216          * journal_invalidatepage() (via truncate_inode_pages()) to discard
217          * these buffers can cause data loss. Also even if we did not discard
218          * these buffers, we would have no way to find them after the inode
219          * is reaped and thus user could see stale data if he tries to read
220          * them before the transaction is checkpointed. So be careful and
221          * force everything to disk here... We use ei->i_datasync_tid to
222          * store the newest transaction containing inode's data.
223          *
224          * Note that directories do not have this problem because they don't
225          * use page cache.
226          */
227         if (inode->i_nlink && ext3_should_journal_data(inode) &&
228             (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
229                 tid_t commit_tid = atomic_read(&ei->i_datasync_tid);
230                 journal_t *journal = EXT3_SB(inode->i_sb)->s_journal;
231
232                 log_start_commit(journal, commit_tid);
233                 log_wait_commit(journal, commit_tid);
234                 filemap_write_and_wait(&inode->i_data);
235         }
236         truncate_inode_pages(&inode->i_data, 0);
237
238         ext3_discard_reservation(inode);
239         rsv = ei->i_block_alloc_info;
240         ei->i_block_alloc_info = NULL;
241         if (unlikely(rsv))
242                 kfree(rsv);
243
244         if (!want_delete)
245                 goto no_delete;
246
247         handle = start_transaction(inode);
248         if (IS_ERR(handle)) {
249                 /*
250                  * If we're going to skip the normal cleanup, we still need to
251                  * make sure that the in-core orphan linked list is properly
252                  * cleaned up.
253                  */
254                 ext3_orphan_del(NULL, inode);
255                 goto no_delete;
256         }
257
258         if (IS_SYNC(inode))
259                 handle->h_sync = 1;
260         inode->i_size = 0;
261         if (inode->i_blocks)
262                 ext3_truncate(inode);
263         /*
264          * Kill off the orphan record created when the inode lost the last
265          * link.  Note that ext3_orphan_del() has to be able to cope with the
266          * deletion of a non-existent orphan - ext3_truncate() could
267          * have removed the record.
268          */
269         ext3_orphan_del(handle, inode);
270         ei->i_dtime = get_seconds();
271
272         /*
273          * One subtle ordering requirement: if anything has gone wrong
274          * (transaction abort, IO errors, whatever), then we can still
275          * do these next steps (the fs will already have been marked as
276          * having errors), but we can't free the inode if the mark_dirty
277          * fails.
278          */
279         if (ext3_mark_inode_dirty(handle, inode)) {
280                 /* If that failed, just dquot_drop() and be done with that */
281                 dquot_drop(inode);
282                 end_writeback(inode);
283         } else {
284                 ext3_xattr_delete_inode(handle, inode);
285                 dquot_free_inode(inode);
286                 dquot_drop(inode);
287                 end_writeback(inode);
288                 ext3_free_inode(handle, inode);
289         }
290         ext3_journal_stop(handle);
291         return;
292 no_delete:
293         end_writeback(inode);
294         dquot_drop(inode);
295 }
296
297 typedef struct {
298         __le32  *p;
299         __le32  key;
300         struct buffer_head *bh;
301 } Indirect;
302
303 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
304 {
305         p->key = *(p->p = v);
306         p->bh = bh;
307 }
308
309 static int verify_chain(Indirect *from, Indirect *to)
310 {
311         while (from <= to && from->key == *from->p)
312                 from++;
313         return (from > to);
314 }
315
316 /**
317  *      ext3_block_to_path - parse the block number into array of offsets
318  *      @inode: inode in question (we are only interested in its superblock)
319  *      @i_block: block number to be parsed
320  *      @offsets: array to store the offsets in
321  *      @boundary: set this non-zero if the referred-to block is likely to be
322  *             followed (on disk) by an indirect block.
323  *
324  *      To store the locations of file's data ext3 uses a data structure common
325  *      for UNIX filesystems - tree of pointers anchored in the inode, with
326  *      data blocks at leaves and indirect blocks in intermediate nodes.
327  *      This function translates the block number into path in that tree -
328  *      return value is the path length and @offsets[n] is the offset of
329  *      pointer to (n+1)th node in the nth one. If @block is out of range
330  *      (negative or too large) warning is printed and zero returned.
331  *
332  *      Note: function doesn't find node addresses, so no IO is needed. All
333  *      we need to know is the capacity of indirect blocks (taken from the
334  *      inode->i_sb).
335  */
336
337 /*
338  * Portability note: the last comparison (check that we fit into triple
339  * indirect block) is spelled differently, because otherwise on an
340  * architecture with 32-bit longs and 8Kb pages we might get into trouble
341  * if our filesystem had 8Kb blocks. We might use long long, but that would
342  * kill us on x86. Oh, well, at least the sign propagation does not matter -
343  * i_block would have to be negative in the very beginning, so we would not
344  * get there at all.
345  */
346
347 static int ext3_block_to_path(struct inode *inode,
348                         long i_block, int offsets[4], int *boundary)
349 {
350         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
351         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
352         const long direct_blocks = EXT3_NDIR_BLOCKS,
353                 indirect_blocks = ptrs,
354                 double_blocks = (1 << (ptrs_bits * 2));
355         int n = 0;
356         int final = 0;
357
358         if (i_block < 0) {
359                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
360         } else if (i_block < direct_blocks) {
361                 offsets[n++] = i_block;
362                 final = direct_blocks;
363         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
364                 offsets[n++] = EXT3_IND_BLOCK;
365                 offsets[n++] = i_block;
366                 final = ptrs;
367         } else if ((i_block -= indirect_blocks) < double_blocks) {
368                 offsets[n++] = EXT3_DIND_BLOCK;
369                 offsets[n++] = i_block >> ptrs_bits;
370                 offsets[n++] = i_block & (ptrs - 1);
371                 final = ptrs;
372         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
373                 offsets[n++] = EXT3_TIND_BLOCK;
374                 offsets[n++] = i_block >> (ptrs_bits * 2);
375                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
376                 offsets[n++] = i_block & (ptrs - 1);
377                 final = ptrs;
378         } else {
379                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
380         }
381         if (boundary)
382                 *boundary = final - 1 - (i_block & (ptrs - 1));
383         return n;
384 }
385
386 /**
387  *      ext3_get_branch - read the chain of indirect blocks leading to data
388  *      @inode: inode in question
389  *      @depth: depth of the chain (1 - direct pointer, etc.)
390  *      @offsets: offsets of pointers in inode/indirect blocks
391  *      @chain: place to store the result
392  *      @err: here we store the error value
393  *
394  *      Function fills the array of triples <key, p, bh> and returns %NULL
395  *      if everything went OK or the pointer to the last filled triple
396  *      (incomplete one) otherwise. Upon the return chain[i].key contains
397  *      the number of (i+1)-th block in the chain (as it is stored in memory,
398  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
399  *      number (it points into struct inode for i==0 and into the bh->b_data
400  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
401  *      block for i>0 and NULL for i==0. In other words, it holds the block
402  *      numbers of the chain, addresses they were taken from (and where we can
403  *      verify that chain did not change) and buffer_heads hosting these
404  *      numbers.
405  *
406  *      Function stops when it stumbles upon zero pointer (absent block)
407  *              (pointer to last triple returned, *@err == 0)
408  *      or when it gets an IO error reading an indirect block
409  *              (ditto, *@err == -EIO)
410  *      or when it notices that chain had been changed while it was reading
411  *              (ditto, *@err == -EAGAIN)
412  *      or when it reads all @depth-1 indirect blocks successfully and finds
413  *      the whole chain, all way to the data (returns %NULL, *err == 0).
414  */
415 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
416                                  Indirect chain[4], int *err)
417 {
418         struct super_block *sb = inode->i_sb;
419         Indirect *p = chain;
420         struct buffer_head *bh;
421
422         *err = 0;
423         /* i_data is not going away, no lock needed */
424         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
425         if (!p->key)
426                 goto no_block;
427         while (--depth) {
428                 bh = sb_bread(sb, le32_to_cpu(p->key));
429                 if (!bh)
430                         goto failure;
431                 /* Reader: pointers */
432                 if (!verify_chain(chain, p))
433                         goto changed;
434                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
435                 /* Reader: end */
436                 if (!p->key)
437                         goto no_block;
438         }
439         return NULL;
440
441 changed:
442         brelse(bh);
443         *err = -EAGAIN;
444         goto no_block;
445 failure:
446         *err = -EIO;
447 no_block:
448         return p;
449 }
450
451 /**
452  *      ext3_find_near - find a place for allocation with sufficient locality
453  *      @inode: owner
454  *      @ind: descriptor of indirect block.
455  *
456  *      This function returns the preferred place for block allocation.
457  *      It is used when heuristic for sequential allocation fails.
458  *      Rules are:
459  *        + if there is a block to the left of our position - allocate near it.
460  *        + if pointer will live in indirect block - allocate near that block.
461  *        + if pointer will live in inode - allocate in the same
462  *          cylinder group.
463  *
464  * In the latter case we colour the starting block by the callers PID to
465  * prevent it from clashing with concurrent allocations for a different inode
466  * in the same block group.   The PID is used here so that functionally related
467  * files will be close-by on-disk.
468  *
469  *      Caller must make sure that @ind is valid and will stay that way.
470  */
471 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
472 {
473         struct ext3_inode_info *ei = EXT3_I(inode);
474         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
475         __le32 *p;
476         ext3_fsblk_t bg_start;
477         ext3_grpblk_t colour;
478
479         /* Try to find previous block */
480         for (p = ind->p - 1; p >= start; p--) {
481                 if (*p)
482                         return le32_to_cpu(*p);
483         }
484
485         /* No such thing, so let's try location of indirect block */
486         if (ind->bh)
487                 return ind->bh->b_blocknr;
488
489         /*
490          * It is going to be referred to from the inode itself? OK, just put it
491          * into the same cylinder group then.
492          */
493         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
494         colour = (current->pid % 16) *
495                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
496         return bg_start + colour;
497 }
498
499 /**
500  *      ext3_find_goal - find a preferred place for allocation.
501  *      @inode: owner
502  *      @block:  block we want
503  *      @partial: pointer to the last triple within a chain
504  *
505  *      Normally this function find the preferred place for block allocation,
506  *      returns it.
507  */
508
509 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
510                                    Indirect *partial)
511 {
512         struct ext3_block_alloc_info *block_i;
513
514         block_i =  EXT3_I(inode)->i_block_alloc_info;
515
516         /*
517          * try the heuristic for sequential allocation,
518          * failing that at least try to get decent locality.
519          */
520         if (block_i && (block == block_i->last_alloc_logical_block + 1)
521                 && (block_i->last_alloc_physical_block != 0)) {
522                 return block_i->last_alloc_physical_block + 1;
523         }
524
525         return ext3_find_near(inode, partial);
526 }
527
528 /**
529  *      ext3_blks_to_allocate - Look up the block map and count the number
530  *      of direct blocks need to be allocated for the given branch.
531  *
532  *      @branch: chain of indirect blocks
533  *      @k: number of blocks need for indirect blocks
534  *      @blks: number of data blocks to be mapped.
535  *      @blocks_to_boundary:  the offset in the indirect block
536  *
537  *      return the total number of blocks to be allocate, including the
538  *      direct and indirect blocks.
539  */
540 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
541                 int blocks_to_boundary)
542 {
543         unsigned long count = 0;
544
545         /*
546          * Simple case, [t,d]Indirect block(s) has not allocated yet
547          * then it's clear blocks on that path have not allocated
548          */
549         if (k > 0) {
550                 /* right now we don't handle cross boundary allocation */
551                 if (blks < blocks_to_boundary + 1)
552                         count += blks;
553                 else
554                         count += blocks_to_boundary + 1;
555                 return count;
556         }
557
558         count++;
559         while (count < blks && count <= blocks_to_boundary &&
560                 le32_to_cpu(*(branch[0].p + count)) == 0) {
561                 count++;
562         }
563         return count;
564 }
565
566 /**
567  *      ext3_alloc_blocks - multiple allocate blocks needed for a branch
568  *      @handle: handle for this transaction
569  *      @inode: owner
570  *      @goal: preferred place for allocation
571  *      @indirect_blks: the number of blocks need to allocate for indirect
572  *                      blocks
573  *      @blks:  number of blocks need to allocated for direct blocks
574  *      @new_blocks: on return it will store the new block numbers for
575  *      the indirect blocks(if needed) and the first direct block,
576  *      @err: here we store the error value
577  *
578  *      return the number of direct blocks allocated
579  */
580 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
581                         ext3_fsblk_t goal, int indirect_blks, int blks,
582                         ext3_fsblk_t new_blocks[4], int *err)
583 {
584         int target, i;
585         unsigned long count = 0;
586         int index = 0;
587         ext3_fsblk_t current_block = 0;
588         int ret = 0;
589
590         /*
591          * Here we try to allocate the requested multiple blocks at once,
592          * on a best-effort basis.
593          * To build a branch, we should allocate blocks for
594          * the indirect blocks(if not allocated yet), and at least
595          * the first direct block of this branch.  That's the
596          * minimum number of blocks need to allocate(required)
597          */
598         target = blks + indirect_blks;
599
600         while (1) {
601                 count = target;
602                 /* allocating blocks for indirect blocks and direct blocks */
603                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
604                 if (*err)
605                         goto failed_out;
606
607                 target -= count;
608                 /* allocate blocks for indirect blocks */
609                 while (index < indirect_blks && count) {
610                         new_blocks[index++] = current_block++;
611                         count--;
612                 }
613
614                 if (count > 0)
615                         break;
616         }
617
618         /* save the new block number for the first direct block */
619         new_blocks[index] = current_block;
620
621         /* total number of blocks allocated for direct blocks */
622         ret = count;
623         *err = 0;
624         return ret;
625 failed_out:
626         for (i = 0; i <index; i++)
627                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
628         return ret;
629 }
630
631 /**
632  *      ext3_alloc_branch - allocate and set up a chain of blocks.
633  *      @handle: handle for this transaction
634  *      @inode: owner
635  *      @indirect_blks: number of allocated indirect blocks
636  *      @blks: number of allocated direct blocks
637  *      @goal: preferred place for allocation
638  *      @offsets: offsets (in the blocks) to store the pointers to next.
639  *      @branch: place to store the chain in.
640  *
641  *      This function allocates blocks, zeroes out all but the last one,
642  *      links them into chain and (if we are synchronous) writes them to disk.
643  *      In other words, it prepares a branch that can be spliced onto the
644  *      inode. It stores the information about that chain in the branch[], in
645  *      the same format as ext3_get_branch() would do. We are calling it after
646  *      we had read the existing part of chain and partial points to the last
647  *      triple of that (one with zero ->key). Upon the exit we have the same
648  *      picture as after the successful ext3_get_block(), except that in one
649  *      place chain is disconnected - *branch->p is still zero (we did not
650  *      set the last link), but branch->key contains the number that should
651  *      be placed into *branch->p to fill that gap.
652  *
653  *      If allocation fails we free all blocks we've allocated (and forget
654  *      their buffer_heads) and return the error value the from failed
655  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
656  *      as described above and return 0.
657  */
658 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
659                         int indirect_blks, int *blks, ext3_fsblk_t goal,
660                         int *offsets, Indirect *branch)
661 {
662         int blocksize = inode->i_sb->s_blocksize;
663         int i, n = 0;
664         int err = 0;
665         struct buffer_head *bh;
666         int num;
667         ext3_fsblk_t new_blocks[4];
668         ext3_fsblk_t current_block;
669
670         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
671                                 *blks, new_blocks, &err);
672         if (err)
673                 return err;
674
675         branch[0].key = cpu_to_le32(new_blocks[0]);
676         /*
677          * metadata blocks and data blocks are allocated.
678          */
679         for (n = 1; n <= indirect_blks;  n++) {
680                 /*
681                  * Get buffer_head for parent block, zero it out
682                  * and set the pointer to new one, then send
683                  * parent to disk.
684                  */
685                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
686                 branch[n].bh = bh;
687                 lock_buffer(bh);
688                 BUFFER_TRACE(bh, "call get_create_access");
689                 err = ext3_journal_get_create_access(handle, bh);
690                 if (err) {
691                         unlock_buffer(bh);
692                         brelse(bh);
693                         goto failed;
694                 }
695
696                 memset(bh->b_data, 0, blocksize);
697                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
698                 branch[n].key = cpu_to_le32(new_blocks[n]);
699                 *branch[n].p = branch[n].key;
700                 if ( n == indirect_blks) {
701                         current_block = new_blocks[n];
702                         /*
703                          * End of chain, update the last new metablock of
704                          * the chain to point to the new allocated
705                          * data blocks numbers
706                          */
707                         for (i=1; i < num; i++)
708                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
709                 }
710                 BUFFER_TRACE(bh, "marking uptodate");
711                 set_buffer_uptodate(bh);
712                 unlock_buffer(bh);
713
714                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
715                 err = ext3_journal_dirty_metadata(handle, bh);
716                 if (err)
717                         goto failed;
718         }
719         *blks = num;
720         return err;
721 failed:
722         /* Allocation failed, free what we already allocated */
723         for (i = 1; i <= n ; i++) {
724                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
725                 ext3_journal_forget(handle, branch[i].bh);
726         }
727         for (i = 0; i <indirect_blks; i++)
728                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
729
730         ext3_free_blocks(handle, inode, new_blocks[i], num);
731
732         return err;
733 }
734
735 /**
736  * ext3_splice_branch - splice the allocated branch onto inode.
737  * @handle: handle for this transaction
738  * @inode: owner
739  * @block: (logical) number of block we are adding
740  * @where: location of missing link
741  * @num:   number of indirect blocks we are adding
742  * @blks:  number of direct blocks we are adding
743  *
744  * This function fills the missing link and does all housekeeping needed in
745  * inode (->i_blocks, etc.). In case of success we end up with the full
746  * chain to new block and return 0.
747  */
748 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
749                         long block, Indirect *where, int num, int blks)
750 {
751         int i;
752         int err = 0;
753         struct ext3_block_alloc_info *block_i;
754         ext3_fsblk_t current_block;
755         struct ext3_inode_info *ei = EXT3_I(inode);
756
757         block_i = ei->i_block_alloc_info;
758         /*
759          * If we're splicing into a [td]indirect block (as opposed to the
760          * inode) then we need to get write access to the [td]indirect block
761          * before the splice.
762          */
763         if (where->bh) {
764                 BUFFER_TRACE(where->bh, "get_write_access");
765                 err = ext3_journal_get_write_access(handle, where->bh);
766                 if (err)
767                         goto err_out;
768         }
769         /* That's it */
770
771         *where->p = where->key;
772
773         /*
774          * Update the host buffer_head or inode to point to more just allocated
775          * direct blocks blocks
776          */
777         if (num == 0 && blks > 1) {
778                 current_block = le32_to_cpu(where->key) + 1;
779                 for (i = 1; i < blks; i++)
780                         *(where->p + i ) = cpu_to_le32(current_block++);
781         }
782
783         /*
784          * update the most recently allocated logical & physical block
785          * in i_block_alloc_info, to assist find the proper goal block for next
786          * allocation
787          */
788         if (block_i) {
789                 block_i->last_alloc_logical_block = block + blks - 1;
790                 block_i->last_alloc_physical_block =
791                                 le32_to_cpu(where[num].key) + blks - 1;
792         }
793
794         /* We are done with atomic stuff, now do the rest of housekeeping */
795
796         inode->i_ctime = CURRENT_TIME_SEC;
797         ext3_mark_inode_dirty(handle, inode);
798         /* ext3_mark_inode_dirty already updated i_sync_tid */
799         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
800
801         /* had we spliced it onto indirect block? */
802         if (where->bh) {
803                 /*
804                  * If we spliced it onto an indirect block, we haven't
805                  * altered the inode.  Note however that if it is being spliced
806                  * onto an indirect block at the very end of the file (the
807                  * file is growing) then we *will* alter the inode to reflect
808                  * the new i_size.  But that is not done here - it is done in
809                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
810                  */
811                 jbd_debug(5, "splicing indirect only\n");
812                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
813                 err = ext3_journal_dirty_metadata(handle, where->bh);
814                 if (err)
815                         goto err_out;
816         } else {
817                 /*
818                  * OK, we spliced it into the inode itself on a direct block.
819                  * Inode was dirtied above.
820                  */
821                 jbd_debug(5, "splicing direct\n");
822         }
823         return err;
824
825 err_out:
826         for (i = 1; i <= num; i++) {
827                 BUFFER_TRACE(where[i].bh, "call journal_forget");
828                 ext3_journal_forget(handle, where[i].bh);
829                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
830         }
831         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
832
833         return err;
834 }
835
836 /*
837  * Allocation strategy is simple: if we have to allocate something, we will
838  * have to go the whole way to leaf. So let's do it before attaching anything
839  * to tree, set linkage between the newborn blocks, write them if sync is
840  * required, recheck the path, free and repeat if check fails, otherwise
841  * set the last missing link (that will protect us from any truncate-generated
842  * removals - all blocks on the path are immune now) and possibly force the
843  * write on the parent block.
844  * That has a nice additional property: no special recovery from the failed
845  * allocations is needed - we simply release blocks and do not touch anything
846  * reachable from inode.
847  *
848  * `handle' can be NULL if create == 0.
849  *
850  * The BKL may not be held on entry here.  Be sure to take it early.
851  * return > 0, # of blocks mapped or allocated.
852  * return = 0, if plain lookup failed.
853  * return < 0, error case.
854  */
855 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
856                 sector_t iblock, unsigned long maxblocks,
857                 struct buffer_head *bh_result,
858                 int create)
859 {
860         int err = -EIO;
861         int offsets[4];
862         Indirect chain[4];
863         Indirect *partial;
864         ext3_fsblk_t goal;
865         int indirect_blks;
866         int blocks_to_boundary = 0;
867         int depth;
868         struct ext3_inode_info *ei = EXT3_I(inode);
869         int count = 0;
870         ext3_fsblk_t first_block = 0;
871
872
873         trace_ext3_get_blocks_enter(inode, iblock, maxblocks, create);
874         J_ASSERT(handle != NULL || create == 0);
875         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
876
877         if (depth == 0)
878                 goto out;
879
880         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
881
882         /* Simplest case - block found, no allocation needed */
883         if (!partial) {
884                 first_block = le32_to_cpu(chain[depth - 1].key);
885                 clear_buffer_new(bh_result);
886                 count++;
887                 /*map more blocks*/
888                 while (count < maxblocks && count <= blocks_to_boundary) {
889                         ext3_fsblk_t blk;
890
891                         if (!verify_chain(chain, chain + depth - 1)) {
892                                 /*
893                                  * Indirect block might be removed by
894                                  * truncate while we were reading it.
895                                  * Handling of that case: forget what we've
896                                  * got now. Flag the err as EAGAIN, so it
897                                  * will reread.
898                                  */
899                                 err = -EAGAIN;
900                                 count = 0;
901                                 break;
902                         }
903                         blk = le32_to_cpu(*(chain[depth-1].p + count));
904
905                         if (blk == first_block + count)
906                                 count++;
907                         else
908                                 break;
909                 }
910                 if (err != -EAGAIN)
911                         goto got_it;
912         }
913
914         /* Next simple case - plain lookup or failed read of indirect block */
915         if (!create || err == -EIO)
916                 goto cleanup;
917
918         /*
919          * Block out ext3_truncate while we alter the tree
920          */
921         mutex_lock(&ei->truncate_mutex);
922
923         /*
924          * If the indirect block is missing while we are reading
925          * the chain(ext3_get_branch() returns -EAGAIN err), or
926          * if the chain has been changed after we grab the semaphore,
927          * (either because another process truncated this branch, or
928          * another get_block allocated this branch) re-grab the chain to see if
929          * the request block has been allocated or not.
930          *
931          * Since we already block the truncate/other get_block
932          * at this point, we will have the current copy of the chain when we
933          * splice the branch into the tree.
934          */
935         if (err == -EAGAIN || !verify_chain(chain, partial)) {
936                 while (partial > chain) {
937                         brelse(partial->bh);
938                         partial--;
939                 }
940                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
941                 if (!partial) {
942                         count++;
943                         mutex_unlock(&ei->truncate_mutex);
944                         if (err)
945                                 goto cleanup;
946                         clear_buffer_new(bh_result);
947                         goto got_it;
948                 }
949         }
950
951         /*
952          * Okay, we need to do block allocation.  Lazily initialize the block
953          * allocation info here if necessary
954         */
955         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
956                 ext3_init_block_alloc_info(inode);
957
958         goal = ext3_find_goal(inode, iblock, partial);
959
960         /* the number of blocks need to allocate for [d,t]indirect blocks */
961         indirect_blks = (chain + depth) - partial - 1;
962
963         /*
964          * Next look up the indirect map to count the totoal number of
965          * direct blocks to allocate for this branch.
966          */
967         count = ext3_blks_to_allocate(partial, indirect_blks,
968                                         maxblocks, blocks_to_boundary);
969         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
970                                 offsets + (partial - chain), partial);
971
972         /*
973          * The ext3_splice_branch call will free and forget any buffers
974          * on the new chain if there is a failure, but that risks using
975          * up transaction credits, especially for bitmaps where the
976          * credits cannot be returned.  Can we handle this somehow?  We
977          * may need to return -EAGAIN upwards in the worst case.  --sct
978          */
979         if (!err)
980                 err = ext3_splice_branch(handle, inode, iblock,
981                                         partial, indirect_blks, count);
982         mutex_unlock(&ei->truncate_mutex);
983         if (err)
984                 goto cleanup;
985
986         set_buffer_new(bh_result);
987 got_it:
988         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
989         if (count > blocks_to_boundary)
990                 set_buffer_boundary(bh_result);
991         err = count;
992         /* Clean up and exit */
993         partial = chain + depth - 1;    /* the whole chain */
994 cleanup:
995         while (partial > chain) {
996                 BUFFER_TRACE(partial->bh, "call brelse");
997                 brelse(partial->bh);
998                 partial--;
999         }
1000         BUFFER_TRACE(bh_result, "returned");
1001 out:
1002         trace_ext3_get_blocks_exit(inode, iblock,
1003                                    depth ? le32_to_cpu(chain[depth-1].key) : 0,
1004                                    count, err);
1005         return err;
1006 }
1007
1008 /* Maximum number of blocks we map for direct IO at once. */
1009 #define DIO_MAX_BLOCKS 4096
1010 /*
1011  * Number of credits we need for writing DIO_MAX_BLOCKS:
1012  * We need sb + group descriptor + bitmap + inode -> 4
1013  * For B blocks with A block pointers per block we need:
1014  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
1015  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
1016  */
1017 #define DIO_CREDITS 25
1018
1019 static int ext3_get_block(struct inode *inode, sector_t iblock,
1020                         struct buffer_head *bh_result, int create)
1021 {
1022         handle_t *handle = ext3_journal_current_handle();
1023         int ret = 0, started = 0;
1024         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1025
1026         if (create && !handle) {        /* Direct IO write... */
1027                 if (max_blocks > DIO_MAX_BLOCKS)
1028                         max_blocks = DIO_MAX_BLOCKS;
1029                 handle = ext3_journal_start(inode, DIO_CREDITS +
1030                                 EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb));
1031                 if (IS_ERR(handle)) {
1032                         ret = PTR_ERR(handle);
1033                         goto out;
1034                 }
1035                 started = 1;
1036         }
1037
1038         ret = ext3_get_blocks_handle(handle, inode, iblock,
1039                                         max_blocks, bh_result, create);
1040         if (ret > 0) {
1041                 bh_result->b_size = (ret << inode->i_blkbits);
1042                 ret = 0;
1043         }
1044         if (started)
1045                 ext3_journal_stop(handle);
1046 out:
1047         return ret;
1048 }
1049
1050 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
1051                 u64 start, u64 len)
1052 {
1053         return generic_block_fiemap(inode, fieinfo, start, len,
1054                                     ext3_get_block);
1055 }
1056
1057 /*
1058  * `handle' can be NULL if create is zero
1059  */
1060 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
1061                                 long block, int create, int *errp)
1062 {
1063         struct buffer_head dummy;
1064         int fatal = 0, err;
1065
1066         J_ASSERT(handle != NULL || create == 0);
1067
1068         dummy.b_state = 0;
1069         dummy.b_blocknr = -1000;
1070         buffer_trace_init(&dummy.b_history);
1071         err = ext3_get_blocks_handle(handle, inode, block, 1,
1072                                         &dummy, create);
1073         /*
1074          * ext3_get_blocks_handle() returns number of blocks
1075          * mapped. 0 in case of a HOLE.
1076          */
1077         if (err > 0) {
1078                 if (err > 1)
1079                         WARN_ON(1);
1080                 err = 0;
1081         }
1082         *errp = err;
1083         if (!err && buffer_mapped(&dummy)) {
1084                 struct buffer_head *bh;
1085                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1086                 if (!bh) {
1087                         *errp = -EIO;
1088                         goto err;
1089                 }
1090                 if (buffer_new(&dummy)) {
1091                         J_ASSERT(create != 0);
1092                         J_ASSERT(handle != NULL);
1093
1094                         /*
1095                          * Now that we do not always journal data, we should
1096                          * keep in mind whether this should always journal the
1097                          * new buffer as metadata.  For now, regular file
1098                          * writes use ext3_get_block instead, so it's not a
1099                          * problem.
1100                          */
1101                         lock_buffer(bh);
1102                         BUFFER_TRACE(bh, "call get_create_access");
1103                         fatal = ext3_journal_get_create_access(handle, bh);
1104                         if (!fatal && !buffer_uptodate(bh)) {
1105                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1106                                 set_buffer_uptodate(bh);
1107                         }
1108                         unlock_buffer(bh);
1109                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1110                         err = ext3_journal_dirty_metadata(handle, bh);
1111                         if (!fatal)
1112                                 fatal = err;
1113                 } else {
1114                         BUFFER_TRACE(bh, "not a new buffer");
1115                 }
1116                 if (fatal) {
1117                         *errp = fatal;
1118                         brelse(bh);
1119                         bh = NULL;
1120                 }
1121                 return bh;
1122         }
1123 err:
1124         return NULL;
1125 }
1126
1127 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1128                                int block, int create, int *err)
1129 {
1130         struct buffer_head * bh;
1131
1132         bh = ext3_getblk(handle, inode, block, create, err);
1133         if (!bh)
1134                 return bh;
1135         if (buffer_uptodate(bh))
1136                 return bh;
1137         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
1138         wait_on_buffer(bh);
1139         if (buffer_uptodate(bh))
1140                 return bh;
1141         put_bh(bh);
1142         *err = -EIO;
1143         return NULL;
1144 }
1145
1146 static int walk_page_buffers(   handle_t *handle,
1147                                 struct buffer_head *head,
1148                                 unsigned from,
1149                                 unsigned to,
1150                                 int *partial,
1151                                 int (*fn)(      handle_t *handle,
1152                                                 struct buffer_head *bh))
1153 {
1154         struct buffer_head *bh;
1155         unsigned block_start, block_end;
1156         unsigned blocksize = head->b_size;
1157         int err, ret = 0;
1158         struct buffer_head *next;
1159
1160         for (   bh = head, block_start = 0;
1161                 ret == 0 && (bh != head || !block_start);
1162                 block_start = block_end, bh = next)
1163         {
1164                 next = bh->b_this_page;
1165                 block_end = block_start + blocksize;
1166                 if (block_end <= from || block_start >= to) {
1167                         if (partial && !buffer_uptodate(bh))
1168                                 *partial = 1;
1169                         continue;
1170                 }
1171                 err = (*fn)(handle, bh);
1172                 if (!ret)
1173                         ret = err;
1174         }
1175         return ret;
1176 }
1177
1178 /*
1179  * To preserve ordering, it is essential that the hole instantiation and
1180  * the data write be encapsulated in a single transaction.  We cannot
1181  * close off a transaction and start a new one between the ext3_get_block()
1182  * and the commit_write().  So doing the journal_start at the start of
1183  * prepare_write() is the right place.
1184  *
1185  * Also, this function can nest inside ext3_writepage() ->
1186  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1187  * has generated enough buffer credits to do the whole page.  So we won't
1188  * block on the journal in that case, which is good, because the caller may
1189  * be PF_MEMALLOC.
1190  *
1191  * By accident, ext3 can be reentered when a transaction is open via
1192  * quota file writes.  If we were to commit the transaction while thus
1193  * reentered, there can be a deadlock - we would be holding a quota
1194  * lock, and the commit would never complete if another thread had a
1195  * transaction open and was blocking on the quota lock - a ranking
1196  * violation.
1197  *
1198  * So what we do is to rely on the fact that journal_stop/journal_start
1199  * will _not_ run commit under these circumstances because handle->h_ref
1200  * is elevated.  We'll still have enough credits for the tiny quotafile
1201  * write.
1202  */
1203 static int do_journal_get_write_access(handle_t *handle,
1204                                         struct buffer_head *bh)
1205 {
1206         int dirty = buffer_dirty(bh);
1207         int ret;
1208
1209         if (!buffer_mapped(bh) || buffer_freed(bh))
1210                 return 0;
1211         /*
1212          * __block_prepare_write() could have dirtied some buffers. Clean
1213          * the dirty bit as jbd2_journal_get_write_access() could complain
1214          * otherwise about fs integrity issues. Setting of the dirty bit
1215          * by __block_prepare_write() isn't a real problem here as we clear
1216          * the bit before releasing a page lock and thus writeback cannot
1217          * ever write the buffer.
1218          */
1219         if (dirty)
1220                 clear_buffer_dirty(bh);
1221         ret = ext3_journal_get_write_access(handle, bh);
1222         if (!ret && dirty)
1223                 ret = ext3_journal_dirty_metadata(handle, bh);
1224         return ret;
1225 }
1226
1227 /*
1228  * Truncate blocks that were not used by write. We have to truncate the
1229  * pagecache as well so that corresponding buffers get properly unmapped.
1230  */
1231 static void ext3_truncate_failed_write(struct inode *inode)
1232 {
1233         truncate_inode_pages(inode->i_mapping, inode->i_size);
1234         ext3_truncate(inode);
1235 }
1236
1237 /*
1238  * Truncate blocks that were not used by direct IO write. We have to zero out
1239  * the last file block as well because direct IO might have written to it.
1240  */
1241 static void ext3_truncate_failed_direct_write(struct inode *inode)
1242 {
1243         ext3_block_truncate_page(inode, inode->i_size);
1244         ext3_truncate(inode);
1245 }
1246
1247 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1248                                 loff_t pos, unsigned len, unsigned flags,
1249                                 struct page **pagep, void **fsdata)
1250 {
1251         struct inode *inode = mapping->host;
1252         int ret;
1253         handle_t *handle;
1254         int retries = 0;
1255         struct page *page;
1256         pgoff_t index;
1257         unsigned from, to;
1258         /* Reserve one block more for addition to orphan list in case
1259          * we allocate blocks but write fails for some reason */
1260         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1261
1262         trace_ext3_write_begin(inode, pos, len, flags);
1263
1264         index = pos >> PAGE_CACHE_SHIFT;
1265         from = pos & (PAGE_CACHE_SIZE - 1);
1266         to = from + len;
1267
1268 retry:
1269         page = grab_cache_page_write_begin(mapping, index, flags);
1270         if (!page)
1271                 return -ENOMEM;
1272         *pagep = page;
1273
1274         handle = ext3_journal_start(inode, needed_blocks);
1275         if (IS_ERR(handle)) {
1276                 unlock_page(page);
1277                 page_cache_release(page);
1278                 ret = PTR_ERR(handle);
1279                 goto out;
1280         }
1281         ret = __block_write_begin(page, pos, len, ext3_get_block);
1282         if (ret)
1283                 goto write_begin_failed;
1284
1285         if (ext3_should_journal_data(inode)) {
1286                 ret = walk_page_buffers(handle, page_buffers(page),
1287                                 from, to, NULL, do_journal_get_write_access);
1288         }
1289 write_begin_failed:
1290         if (ret) {
1291                 /*
1292                  * block_write_begin may have instantiated a few blocks
1293                  * outside i_size.  Trim these off again. Don't need
1294                  * i_size_read because we hold i_mutex.
1295                  *
1296                  * Add inode to orphan list in case we crash before truncate
1297                  * finishes. Do this only if ext3_can_truncate() agrees so
1298                  * that orphan processing code is happy.
1299                  */
1300                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1301                         ext3_orphan_add(handle, inode);
1302                 ext3_journal_stop(handle);
1303                 unlock_page(page);
1304                 page_cache_release(page);
1305                 if (pos + len > inode->i_size)
1306                         ext3_truncate_failed_write(inode);
1307         }
1308         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1309                 goto retry;
1310 out:
1311         return ret;
1312 }
1313
1314
1315 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1316 {
1317         int err = journal_dirty_data(handle, bh);
1318         if (err)
1319                 ext3_journal_abort_handle(__func__, __func__,
1320                                                 bh, handle, err);
1321         return err;
1322 }
1323
1324 /* For ordered writepage and write_end functions */
1325 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1326 {
1327         /*
1328          * Write could have mapped the buffer but it didn't copy the data in
1329          * yet. So avoid filing such buffer into a transaction.
1330          */
1331         if (buffer_mapped(bh) && buffer_uptodate(bh))
1332                 return ext3_journal_dirty_data(handle, bh);
1333         return 0;
1334 }
1335
1336 /* For write_end() in data=journal mode */
1337 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1338 {
1339         if (!buffer_mapped(bh) || buffer_freed(bh))
1340                 return 0;
1341         set_buffer_uptodate(bh);
1342         return ext3_journal_dirty_metadata(handle, bh);
1343 }
1344
1345 /*
1346  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1347  * for the whole page but later we failed to copy the data in. Update inode
1348  * size according to what we managed to copy. The rest is going to be
1349  * truncated in write_end function.
1350  */
1351 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1352 {
1353         /* What matters to us is i_disksize. We don't write i_size anywhere */
1354         if (pos + copied > inode->i_size)
1355                 i_size_write(inode, pos + copied);
1356         if (pos + copied > EXT3_I(inode)->i_disksize) {
1357                 EXT3_I(inode)->i_disksize = pos + copied;
1358                 mark_inode_dirty(inode);
1359         }
1360 }
1361
1362 /*
1363  * We need to pick up the new inode size which generic_commit_write gave us
1364  * `file' can be NULL - eg, when called from page_symlink().
1365  *
1366  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1367  * buffers are managed internally.
1368  */
1369 static int ext3_ordered_write_end(struct file *file,
1370                                 struct address_space *mapping,
1371                                 loff_t pos, unsigned len, unsigned copied,
1372                                 struct page *page, void *fsdata)
1373 {
1374         handle_t *handle = ext3_journal_current_handle();
1375         struct inode *inode = file->f_mapping->host;
1376         unsigned from, to;
1377         int ret = 0, ret2;
1378
1379         trace_ext3_ordered_write_end(inode, pos, len, copied);
1380         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1381
1382         from = pos & (PAGE_CACHE_SIZE - 1);
1383         to = from + copied;
1384         ret = walk_page_buffers(handle, page_buffers(page),
1385                 from, to, NULL, journal_dirty_data_fn);
1386
1387         if (ret == 0)
1388                 update_file_sizes(inode, pos, copied);
1389         /*
1390          * There may be allocated blocks outside of i_size because
1391          * we failed to copy some data. Prepare for truncate.
1392          */
1393         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1394                 ext3_orphan_add(handle, inode);
1395         ret2 = ext3_journal_stop(handle);
1396         if (!ret)
1397                 ret = ret2;
1398         unlock_page(page);
1399         page_cache_release(page);
1400
1401         if (pos + len > inode->i_size)
1402                 ext3_truncate_failed_write(inode);
1403         return ret ? ret : copied;
1404 }
1405
1406 static int ext3_writeback_write_end(struct file *file,
1407                                 struct address_space *mapping,
1408                                 loff_t pos, unsigned len, unsigned copied,
1409                                 struct page *page, void *fsdata)
1410 {
1411         handle_t *handle = ext3_journal_current_handle();
1412         struct inode *inode = file->f_mapping->host;
1413         int ret;
1414
1415         trace_ext3_writeback_write_end(inode, pos, len, copied);
1416         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1417         update_file_sizes(inode, pos, copied);
1418         /*
1419          * There may be allocated blocks outside of i_size because
1420          * we failed to copy some data. Prepare for truncate.
1421          */
1422         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1423                 ext3_orphan_add(handle, inode);
1424         ret = ext3_journal_stop(handle);
1425         unlock_page(page);
1426         page_cache_release(page);
1427
1428         if (pos + len > inode->i_size)
1429                 ext3_truncate_failed_write(inode);
1430         return ret ? ret : copied;
1431 }
1432
1433 static int ext3_journalled_write_end(struct file *file,
1434                                 struct address_space *mapping,
1435                                 loff_t pos, unsigned len, unsigned copied,
1436                                 struct page *page, void *fsdata)
1437 {
1438         handle_t *handle = ext3_journal_current_handle();
1439         struct inode *inode = mapping->host;
1440         struct ext3_inode_info *ei = EXT3_I(inode);
1441         int ret = 0, ret2;
1442         int partial = 0;
1443         unsigned from, to;
1444
1445         trace_ext3_journalled_write_end(inode, pos, len, copied);
1446         from = pos & (PAGE_CACHE_SIZE - 1);
1447         to = from + len;
1448
1449         if (copied < len) {
1450                 if (!PageUptodate(page))
1451                         copied = 0;
1452                 page_zero_new_buffers(page, from + copied, to);
1453                 to = from + copied;
1454         }
1455
1456         ret = walk_page_buffers(handle, page_buffers(page), from,
1457                                 to, &partial, write_end_fn);
1458         if (!partial)
1459                 SetPageUptodate(page);
1460
1461         if (pos + copied > inode->i_size)
1462                 i_size_write(inode, pos + copied);
1463         /*
1464          * There may be allocated blocks outside of i_size because
1465          * we failed to copy some data. Prepare for truncate.
1466          */
1467         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1468                 ext3_orphan_add(handle, inode);
1469         ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1470         atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
1471         if (inode->i_size > ei->i_disksize) {
1472                 ei->i_disksize = inode->i_size;
1473                 ret2 = ext3_mark_inode_dirty(handle, inode);
1474                 if (!ret)
1475                         ret = ret2;
1476         }
1477
1478         ret2 = ext3_journal_stop(handle);
1479         if (!ret)
1480                 ret = ret2;
1481         unlock_page(page);
1482         page_cache_release(page);
1483
1484         if (pos + len > inode->i_size)
1485                 ext3_truncate_failed_write(inode);
1486         return ret ? ret : copied;
1487 }
1488
1489 /*
1490  * bmap() is special.  It gets used by applications such as lilo and by
1491  * the swapper to find the on-disk block of a specific piece of data.
1492  *
1493  * Naturally, this is dangerous if the block concerned is still in the
1494  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1495  * filesystem and enables swap, then they may get a nasty shock when the
1496  * data getting swapped to that swapfile suddenly gets overwritten by
1497  * the original zero's written out previously to the journal and
1498  * awaiting writeback in the kernel's buffer cache.
1499  *
1500  * So, if we see any bmap calls here on a modified, data-journaled file,
1501  * take extra steps to flush any blocks which might be in the cache.
1502  */
1503 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1504 {
1505         struct inode *inode = mapping->host;
1506         journal_t *journal;
1507         int err;
1508
1509         if (ext3_test_inode_state(inode, EXT3_STATE_JDATA)) {
1510                 /*
1511                  * This is a REALLY heavyweight approach, but the use of
1512                  * bmap on dirty files is expected to be extremely rare:
1513                  * only if we run lilo or swapon on a freshly made file
1514                  * do we expect this to happen.
1515                  *
1516                  * (bmap requires CAP_SYS_RAWIO so this does not
1517                  * represent an unprivileged user DOS attack --- we'd be
1518                  * in trouble if mortal users could trigger this path at
1519                  * will.)
1520                  *
1521                  * NB. EXT3_STATE_JDATA is not set on files other than
1522                  * regular files.  If somebody wants to bmap a directory
1523                  * or symlink and gets confused because the buffer
1524                  * hasn't yet been flushed to disk, they deserve
1525                  * everything they get.
1526                  */
1527
1528                 ext3_clear_inode_state(inode, EXT3_STATE_JDATA);
1529                 journal = EXT3_JOURNAL(inode);
1530                 journal_lock_updates(journal);
1531                 err = journal_flush(journal);
1532                 journal_unlock_updates(journal);
1533
1534                 if (err)
1535                         return 0;
1536         }
1537
1538         return generic_block_bmap(mapping,block,ext3_get_block);
1539 }
1540
1541 static int bget_one(handle_t *handle, struct buffer_head *bh)
1542 {
1543         get_bh(bh);
1544         return 0;
1545 }
1546
1547 static int bput_one(handle_t *handle, struct buffer_head *bh)
1548 {
1549         put_bh(bh);
1550         return 0;
1551 }
1552
1553 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1554 {
1555         return !buffer_mapped(bh);
1556 }
1557
1558 /*
1559  * Note that we always start a transaction even if we're not journalling
1560  * data.  This is to preserve ordering: any hole instantiation within
1561  * __block_write_full_page -> ext3_get_block() should be journalled
1562  * along with the data so we don't crash and then get metadata which
1563  * refers to old data.
1564  *
1565  * In all journalling modes block_write_full_page() will start the I/O.
1566  *
1567  * Problem:
1568  *
1569  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1570  *              ext3_writepage()
1571  *
1572  * Similar for:
1573  *
1574  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1575  *
1576  * Same applies to ext3_get_block().  We will deadlock on various things like
1577  * lock_journal and i_truncate_mutex.
1578  *
1579  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1580  * allocations fail.
1581  *
1582  * 16May01: If we're reentered then journal_current_handle() will be
1583  *          non-zero. We simply *return*.
1584  *
1585  * 1 July 2001: @@@ FIXME:
1586  *   In journalled data mode, a data buffer may be metadata against the
1587  *   current transaction.  But the same file is part of a shared mapping
1588  *   and someone does a writepage() on it.
1589  *
1590  *   We will move the buffer onto the async_data list, but *after* it has
1591  *   been dirtied. So there's a small window where we have dirty data on
1592  *   BJ_Metadata.
1593  *
1594  *   Note that this only applies to the last partial page in the file.  The
1595  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1596  *   broken code anyway: it's wrong for msync()).
1597  *
1598  *   It's a rare case: affects the final partial page, for journalled data
1599  *   where the file is subject to bith write() and writepage() in the same
1600  *   transction.  To fix it we'll need a custom block_write_full_page().
1601  *   We'll probably need that anyway for journalling writepage() output.
1602  *
1603  * We don't honour synchronous mounts for writepage().  That would be
1604  * disastrous.  Any write() or metadata operation will sync the fs for
1605  * us.
1606  *
1607  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1608  * we don't need to open a transaction here.
1609  */
1610 static int ext3_ordered_writepage(struct page *page,
1611                                 struct writeback_control *wbc)
1612 {
1613         struct inode *inode = page->mapping->host;
1614         struct buffer_head *page_bufs;
1615         handle_t *handle = NULL;
1616         int ret = 0;
1617         int err;
1618
1619         J_ASSERT(PageLocked(page));
1620         /*
1621          * We don't want to warn for emergency remount. The condition is
1622          * ordered to avoid dereferencing inode->i_sb in non-error case to
1623          * avoid slow-downs.
1624          */
1625         WARN_ON_ONCE(IS_RDONLY(inode) &&
1626                      !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ERROR_FS));
1627
1628         /*
1629          * We give up here if we're reentered, because it might be for a
1630          * different filesystem.
1631          */
1632         if (ext3_journal_current_handle())
1633                 goto out_fail;
1634
1635         trace_ext3_ordered_writepage(page);
1636         if (!page_has_buffers(page)) {
1637                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1638                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1639                 page_bufs = page_buffers(page);
1640         } else {
1641                 page_bufs = page_buffers(page);
1642                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1643                                        NULL, buffer_unmapped)) {
1644                         /* Provide NULL get_block() to catch bugs if buffers
1645                          * weren't really mapped */
1646                         return block_write_full_page(page, NULL, wbc);
1647                 }
1648         }
1649         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1650
1651         if (IS_ERR(handle)) {
1652                 ret = PTR_ERR(handle);
1653                 goto out_fail;
1654         }
1655
1656         walk_page_buffers(handle, page_bufs, 0,
1657                         PAGE_CACHE_SIZE, NULL, bget_one);
1658
1659         ret = block_write_full_page(page, ext3_get_block, wbc);
1660
1661         /*
1662          * The page can become unlocked at any point now, and
1663          * truncate can then come in and change things.  So we
1664          * can't touch *page from now on.  But *page_bufs is
1665          * safe due to elevated refcount.
1666          */
1667
1668         /*
1669          * And attach them to the current transaction.  But only if
1670          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1671          * and generally junk.
1672          */
1673         if (ret == 0) {
1674                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1675                                         NULL, journal_dirty_data_fn);
1676                 if (!ret)
1677                         ret = err;
1678         }
1679         walk_page_buffers(handle, page_bufs, 0,
1680                         PAGE_CACHE_SIZE, NULL, bput_one);
1681         err = ext3_journal_stop(handle);
1682         if (!ret)
1683                 ret = err;
1684         return ret;
1685
1686 out_fail:
1687         redirty_page_for_writepage(wbc, page);
1688         unlock_page(page);
1689         return ret;
1690 }
1691
1692 static int ext3_writeback_writepage(struct page *page,
1693                                 struct writeback_control *wbc)
1694 {
1695         struct inode *inode = page->mapping->host;
1696         handle_t *handle = NULL;
1697         int ret = 0;
1698         int err;
1699
1700         J_ASSERT(PageLocked(page));
1701         /*
1702          * We don't want to warn for emergency remount. The condition is
1703          * ordered to avoid dereferencing inode->i_sb in non-error case to
1704          * avoid slow-downs.
1705          */
1706         WARN_ON_ONCE(IS_RDONLY(inode) &&
1707                      !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ERROR_FS));
1708
1709         if (ext3_journal_current_handle())
1710                 goto out_fail;
1711
1712         trace_ext3_writeback_writepage(page);
1713         if (page_has_buffers(page)) {
1714                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1715                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1716                         /* Provide NULL get_block() to catch bugs if buffers
1717                          * weren't really mapped */
1718                         return block_write_full_page(page, NULL, wbc);
1719                 }
1720         }
1721
1722         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1723         if (IS_ERR(handle)) {
1724                 ret = PTR_ERR(handle);
1725                 goto out_fail;
1726         }
1727
1728         ret = block_write_full_page(page, ext3_get_block, wbc);
1729
1730         err = ext3_journal_stop(handle);
1731         if (!ret)
1732                 ret = err;
1733         return ret;
1734
1735 out_fail:
1736         redirty_page_for_writepage(wbc, page);
1737         unlock_page(page);
1738         return ret;
1739 }
1740
1741 static int ext3_journalled_writepage(struct page *page,
1742                                 struct writeback_control *wbc)
1743 {
1744         struct inode *inode = page->mapping->host;
1745         handle_t *handle = NULL;
1746         int ret = 0;
1747         int err;
1748
1749         J_ASSERT(PageLocked(page));
1750         /*
1751          * We don't want to warn for emergency remount. The condition is
1752          * ordered to avoid dereferencing inode->i_sb in non-error case to
1753          * avoid slow-downs.
1754          */
1755         WARN_ON_ONCE(IS_RDONLY(inode) &&
1756                      !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ERROR_FS));
1757
1758         if (ext3_journal_current_handle())
1759                 goto no_write;
1760
1761         trace_ext3_journalled_writepage(page);
1762         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1763         if (IS_ERR(handle)) {
1764                 ret = PTR_ERR(handle);
1765                 goto no_write;
1766         }
1767
1768         if (!page_has_buffers(page) || PageChecked(page)) {
1769                 /*
1770                  * It's mmapped pagecache.  Add buffers and journal it.  There
1771                  * doesn't seem much point in redirtying the page here.
1772                  */
1773                 ClearPageChecked(page);
1774                 ret = __block_write_begin(page, 0, PAGE_CACHE_SIZE,
1775                                           ext3_get_block);
1776                 if (ret != 0) {
1777                         ext3_journal_stop(handle);
1778                         goto out_unlock;
1779                 }
1780                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1781                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1782
1783                 err = walk_page_buffers(handle, page_buffers(page), 0,
1784                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1785                 if (ret == 0)
1786                         ret = err;
1787                 ext3_set_inode_state(inode, EXT3_STATE_JDATA);
1788                 atomic_set(&EXT3_I(inode)->i_datasync_tid,
1789                            handle->h_transaction->t_tid);
1790                 unlock_page(page);
1791         } else {
1792                 /*
1793                  * It may be a page full of checkpoint-mode buffers.  We don't
1794                  * really know unless we go poke around in the buffer_heads.
1795                  * But block_write_full_page will do the right thing.
1796                  */
1797                 ret = block_write_full_page(page, ext3_get_block, wbc);
1798         }
1799         err = ext3_journal_stop(handle);
1800         if (!ret)
1801                 ret = err;
1802 out:
1803         return ret;
1804
1805 no_write:
1806         redirty_page_for_writepage(wbc, page);
1807 out_unlock:
1808         unlock_page(page);
1809         goto out;
1810 }
1811
1812 static int ext3_readpage(struct file *file, struct page *page)
1813 {
1814         trace_ext3_readpage(page);
1815         return mpage_readpage(page, ext3_get_block);
1816 }
1817
1818 static int
1819 ext3_readpages(struct file *file, struct address_space *mapping,
1820                 struct list_head *pages, unsigned nr_pages)
1821 {
1822         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1823 }
1824
1825 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1826 {
1827         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1828
1829         trace_ext3_invalidatepage(page, offset);
1830
1831         /*
1832          * If it's a full truncate we just forget about the pending dirtying
1833          */
1834         if (offset == 0)
1835                 ClearPageChecked(page);
1836
1837         journal_invalidatepage(journal, page, offset);
1838 }
1839
1840 static int ext3_releasepage(struct page *page, gfp_t wait)
1841 {
1842         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1843
1844         trace_ext3_releasepage(page);
1845         WARN_ON(PageChecked(page));
1846         if (!page_has_buffers(page))
1847                 return 0;
1848         return journal_try_to_free_buffers(journal, page, wait);
1849 }
1850
1851 /*
1852  * If the O_DIRECT write will extend the file then add this inode to the
1853  * orphan list.  So recovery will truncate it back to the original size
1854  * if the machine crashes during the write.
1855  *
1856  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1857  * crashes then stale disk data _may_ be exposed inside the file. But current
1858  * VFS code falls back into buffered path in that case so we are safe.
1859  */
1860 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1861                         const struct iovec *iov, loff_t offset,
1862                         unsigned long nr_segs)
1863 {
1864         struct file *file = iocb->ki_filp;
1865         struct inode *inode = file->f_mapping->host;
1866         struct ext3_inode_info *ei = EXT3_I(inode);
1867         handle_t *handle;
1868         ssize_t ret;
1869         int orphan = 0;
1870         size_t count = iov_length(iov, nr_segs);
1871         int retries = 0;
1872
1873         trace_ext3_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
1874
1875         if (rw == WRITE) {
1876                 loff_t final_size = offset + count;
1877
1878                 if (final_size > inode->i_size) {
1879                         /* Credits for sb + inode write */
1880                         handle = ext3_journal_start(inode, 2);
1881                         if (IS_ERR(handle)) {
1882                                 ret = PTR_ERR(handle);
1883                                 goto out;
1884                         }
1885                         ret = ext3_orphan_add(handle, inode);
1886                         if (ret) {
1887                                 ext3_journal_stop(handle);
1888                                 goto out;
1889                         }
1890                         orphan = 1;
1891                         ei->i_disksize = inode->i_size;
1892                         ext3_journal_stop(handle);
1893                 }
1894         }
1895
1896 retry:
1897         ret = blockdev_direct_IO(rw, iocb, inode, iov, offset, nr_segs,
1898                                  ext3_get_block);
1899         /*
1900          * In case of error extending write may have instantiated a few
1901          * blocks outside i_size. Trim these off again.
1902          */
1903         if (unlikely((rw & WRITE) && ret < 0)) {
1904                 loff_t isize = i_size_read(inode);
1905                 loff_t end = offset + iov_length(iov, nr_segs);
1906
1907                 if (end > isize)
1908                         ext3_truncate_failed_direct_write(inode);
1909         }
1910         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1911                 goto retry;
1912
1913         if (orphan) {
1914                 int err;
1915
1916                 /* Credits for sb + inode write */
1917                 handle = ext3_journal_start(inode, 2);
1918                 if (IS_ERR(handle)) {
1919                         /* This is really bad luck. We've written the data
1920                          * but cannot extend i_size. Truncate allocated blocks
1921                          * and pretend the write failed... */
1922                         ext3_truncate_failed_direct_write(inode);
1923                         ret = PTR_ERR(handle);
1924                         goto out;
1925                 }
1926                 if (inode->i_nlink)
1927                         ext3_orphan_del(handle, inode);
1928                 if (ret > 0) {
1929                         loff_t end = offset + ret;
1930                         if (end > inode->i_size) {
1931                                 ei->i_disksize = end;
1932                                 i_size_write(inode, end);
1933                                 /*
1934                                  * We're going to return a positive `ret'
1935                                  * here due to non-zero-length I/O, so there's
1936                                  * no way of reporting error returns from
1937                                  * ext3_mark_inode_dirty() to userspace.  So
1938                                  * ignore it.
1939                                  */
1940                                 ext3_mark_inode_dirty(handle, inode);
1941                         }
1942                 }
1943                 err = ext3_journal_stop(handle);
1944                 if (ret == 0)
1945                         ret = err;
1946         }
1947 out:
1948         trace_ext3_direct_IO_exit(inode, offset,
1949                                 iov_length(iov, nr_segs), rw, ret);
1950         return ret;
1951 }
1952
1953 /*
1954  * Pages can be marked dirty completely asynchronously from ext3's journalling
1955  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1956  * much here because ->set_page_dirty is called under VFS locks.  The page is
1957  * not necessarily locked.
1958  *
1959  * We cannot just dirty the page and leave attached buffers clean, because the
1960  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1961  * or jbddirty because all the journalling code will explode.
1962  *
1963  * So what we do is to mark the page "pending dirty" and next time writepage
1964  * is called, propagate that into the buffers appropriately.
1965  */
1966 static int ext3_journalled_set_page_dirty(struct page *page)
1967 {
1968         SetPageChecked(page);
1969         return __set_page_dirty_nobuffers(page);
1970 }
1971
1972 static const struct address_space_operations ext3_ordered_aops = {
1973         .readpage               = ext3_readpage,
1974         .readpages              = ext3_readpages,
1975         .writepage              = ext3_ordered_writepage,
1976         .write_begin            = ext3_write_begin,
1977         .write_end              = ext3_ordered_write_end,
1978         .bmap                   = ext3_bmap,
1979         .invalidatepage         = ext3_invalidatepage,
1980         .releasepage            = ext3_releasepage,
1981         .direct_IO              = ext3_direct_IO,
1982         .migratepage            = buffer_migrate_page,
1983         .is_partially_uptodate  = block_is_partially_uptodate,
1984         .error_remove_page      = generic_error_remove_page,
1985 };
1986
1987 static const struct address_space_operations ext3_writeback_aops = {
1988         .readpage               = ext3_readpage,
1989         .readpages              = ext3_readpages,
1990         .writepage              = ext3_writeback_writepage,
1991         .write_begin            = ext3_write_begin,
1992         .write_end              = ext3_writeback_write_end,
1993         .bmap                   = ext3_bmap,
1994         .invalidatepage         = ext3_invalidatepage,
1995         .releasepage            = ext3_releasepage,
1996         .direct_IO              = ext3_direct_IO,
1997         .migratepage            = buffer_migrate_page,
1998         .is_partially_uptodate  = block_is_partially_uptodate,
1999         .error_remove_page      = generic_error_remove_page,
2000 };
2001
2002 static const struct address_space_operations ext3_journalled_aops = {
2003         .readpage               = ext3_readpage,
2004         .readpages              = ext3_readpages,
2005         .writepage              = ext3_journalled_writepage,
2006         .write_begin            = ext3_write_begin,
2007         .write_end              = ext3_journalled_write_end,
2008         .set_page_dirty         = ext3_journalled_set_page_dirty,
2009         .bmap                   = ext3_bmap,
2010         .invalidatepage         = ext3_invalidatepage,
2011         .releasepage            = ext3_releasepage,
2012         .is_partially_uptodate  = block_is_partially_uptodate,
2013         .error_remove_page      = generic_error_remove_page,
2014 };
2015
2016 void ext3_set_aops(struct inode *inode)
2017 {
2018         if (ext3_should_order_data(inode))
2019                 inode->i_mapping->a_ops = &ext3_ordered_aops;
2020         else if (ext3_should_writeback_data(inode))
2021                 inode->i_mapping->a_ops = &ext3_writeback_aops;
2022         else
2023                 inode->i_mapping->a_ops = &ext3_journalled_aops;
2024 }
2025
2026 /*
2027  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
2028  * up to the end of the block which corresponds to `from'.
2029  * This required during truncate. We need to physically zero the tail end
2030  * of that block so it doesn't yield old data if the file is later grown.
2031  */
2032 static int ext3_block_truncate_page(struct inode *inode, loff_t from)
2033 {
2034         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
2035         unsigned offset = from & (PAGE_CACHE_SIZE - 1);
2036         unsigned blocksize, iblock, length, pos;
2037         struct page *page;
2038         handle_t *handle = NULL;
2039         struct buffer_head *bh;
2040         int err = 0;
2041
2042         /* Truncated on block boundary - nothing to do */
2043         blocksize = inode->i_sb->s_blocksize;
2044         if ((from & (blocksize - 1)) == 0)
2045                 return 0;
2046
2047         page = grab_cache_page(inode->i_mapping, index);
2048         if (!page)
2049                 return -ENOMEM;
2050         length = blocksize - (offset & (blocksize - 1));
2051         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
2052
2053         if (!page_has_buffers(page))
2054                 create_empty_buffers(page, blocksize, 0);
2055
2056         /* Find the buffer that contains "offset" */
2057         bh = page_buffers(page);
2058         pos = blocksize;
2059         while (offset >= pos) {
2060                 bh = bh->b_this_page;
2061                 iblock++;
2062                 pos += blocksize;
2063         }
2064
2065         err = 0;
2066         if (buffer_freed(bh)) {
2067                 BUFFER_TRACE(bh, "freed: skip");
2068                 goto unlock;
2069         }
2070
2071         if (!buffer_mapped(bh)) {
2072                 BUFFER_TRACE(bh, "unmapped");
2073                 ext3_get_block(inode, iblock, bh, 0);
2074                 /* unmapped? It's a hole - nothing to do */
2075                 if (!buffer_mapped(bh)) {
2076                         BUFFER_TRACE(bh, "still unmapped");
2077                         goto unlock;
2078                 }
2079         }
2080
2081         /* Ok, it's mapped. Make sure it's up-to-date */
2082         if (PageUptodate(page))
2083                 set_buffer_uptodate(bh);
2084
2085         if (!buffer_uptodate(bh)) {
2086                 err = -EIO;
2087                 ll_rw_block(READ, 1, &bh);
2088                 wait_on_buffer(bh);
2089                 /* Uhhuh. Read error. Complain and punt. */
2090                 if (!buffer_uptodate(bh))
2091                         goto unlock;
2092         }
2093
2094         /* data=writeback mode doesn't need transaction to zero-out data */
2095         if (!ext3_should_writeback_data(inode)) {
2096                 /* We journal at most one block */
2097                 handle = ext3_journal_start(inode, 1);
2098                 if (IS_ERR(handle)) {
2099                         clear_highpage(page);
2100                         flush_dcache_page(page);
2101                         err = PTR_ERR(handle);
2102                         goto unlock;
2103                 }
2104         }
2105
2106         if (ext3_should_journal_data(inode)) {
2107                 BUFFER_TRACE(bh, "get write access");
2108                 err = ext3_journal_get_write_access(handle, bh);
2109                 if (err)
2110                         goto stop;
2111         }
2112
2113         zero_user(page, offset, length);
2114         BUFFER_TRACE(bh, "zeroed end of block");
2115
2116         err = 0;
2117         if (ext3_should_journal_data(inode)) {
2118                 err = ext3_journal_dirty_metadata(handle, bh);
2119         } else {
2120                 if (ext3_should_order_data(inode))
2121                         err = ext3_journal_dirty_data(handle, bh);
2122                 mark_buffer_dirty(bh);
2123         }
2124 stop:
2125         if (handle)
2126                 ext3_journal_stop(handle);
2127
2128 unlock:
2129         unlock_page(page);
2130         page_cache_release(page);
2131         return err;
2132 }
2133
2134 /*
2135  * Probably it should be a library function... search for first non-zero word
2136  * or memcmp with zero_page, whatever is better for particular architecture.
2137  * Linus?
2138  */
2139 static inline int all_zeroes(__le32 *p, __le32 *q)
2140 {
2141         while (p < q)
2142                 if (*p++)
2143                         return 0;
2144         return 1;
2145 }
2146
2147 /**
2148  *      ext3_find_shared - find the indirect blocks for partial truncation.
2149  *      @inode:   inode in question
2150  *      @depth:   depth of the affected branch
2151  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
2152  *      @chain:   place to store the pointers to partial indirect blocks
2153  *      @top:     place to the (detached) top of branch
2154  *
2155  *      This is a helper function used by ext3_truncate().
2156  *
2157  *      When we do truncate() we may have to clean the ends of several
2158  *      indirect blocks but leave the blocks themselves alive. Block is
2159  *      partially truncated if some data below the new i_size is referred
2160  *      from it (and it is on the path to the first completely truncated
2161  *      data block, indeed).  We have to free the top of that path along
2162  *      with everything to the right of the path. Since no allocation
2163  *      past the truncation point is possible until ext3_truncate()
2164  *      finishes, we may safely do the latter, but top of branch may
2165  *      require special attention - pageout below the truncation point
2166  *      might try to populate it.
2167  *
2168  *      We atomically detach the top of branch from the tree, store the
2169  *      block number of its root in *@top, pointers to buffer_heads of
2170  *      partially truncated blocks - in @chain[].bh and pointers to
2171  *      their last elements that should not be removed - in
2172  *      @chain[].p. Return value is the pointer to last filled element
2173  *      of @chain.
2174  *
2175  *      The work left to caller to do the actual freeing of subtrees:
2176  *              a) free the subtree starting from *@top
2177  *              b) free the subtrees whose roots are stored in
2178  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2179  *              c) free the subtrees growing from the inode past the @chain[0].
2180  *                      (no partially truncated stuff there).  */
2181
2182 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2183                         int offsets[4], Indirect chain[4], __le32 *top)
2184 {
2185         Indirect *partial, *p;
2186         int k, err;
2187
2188         *top = 0;
2189         /* Make k index the deepest non-null offset + 1 */
2190         for (k = depth; k > 1 && !offsets[k-1]; k--)
2191                 ;
2192         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2193         /* Writer: pointers */
2194         if (!partial)
2195                 partial = chain + k-1;
2196         /*
2197          * If the branch acquired continuation since we've looked at it -
2198          * fine, it should all survive and (new) top doesn't belong to us.
2199          */
2200         if (!partial->key && *partial->p)
2201                 /* Writer: end */
2202                 goto no_top;
2203         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2204                 ;
2205         /*
2206          * OK, we've found the last block that must survive. The rest of our
2207          * branch should be detached before unlocking. However, if that rest
2208          * of branch is all ours and does not grow immediately from the inode
2209          * it's easier to cheat and just decrement partial->p.
2210          */
2211         if (p == chain + k - 1 && p > chain) {
2212                 p->p--;
2213         } else {
2214                 *top = *p->p;
2215                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2216 #if 0
2217                 *p->p = 0;
2218 #endif
2219         }
2220         /* Writer: end */
2221
2222         while(partial > p) {
2223                 brelse(partial->bh);
2224                 partial--;
2225         }
2226 no_top:
2227         return partial;
2228 }
2229
2230 /*
2231  * Zero a number of block pointers in either an inode or an indirect block.
2232  * If we restart the transaction we must again get write access to the
2233  * indirect block for further modification.
2234  *
2235  * We release `count' blocks on disk, but (last - first) may be greater
2236  * than `count' because there can be holes in there.
2237  */
2238 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2239                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2240                 unsigned long count, __le32 *first, __le32 *last)
2241 {
2242         __le32 *p;
2243         if (try_to_extend_transaction(handle, inode)) {
2244                 if (bh) {
2245                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2246                         if (ext3_journal_dirty_metadata(handle, bh))
2247                                 return;
2248                 }
2249                 ext3_mark_inode_dirty(handle, inode);
2250                 truncate_restart_transaction(handle, inode);
2251                 if (bh) {
2252                         BUFFER_TRACE(bh, "retaking write access");
2253                         if (ext3_journal_get_write_access(handle, bh))
2254                                 return;
2255                 }
2256         }
2257
2258         /*
2259          * Any buffers which are on the journal will be in memory. We find
2260          * them on the hash table so journal_revoke() will run journal_forget()
2261          * on them.  We've already detached each block from the file, so
2262          * bforget() in journal_forget() should be safe.
2263          *
2264          * AKPM: turn on bforget in journal_forget()!!!
2265          */
2266         for (p = first; p < last; p++) {
2267                 u32 nr = le32_to_cpu(*p);
2268                 if (nr) {
2269                         struct buffer_head *bh;
2270
2271                         *p = 0;
2272                         bh = sb_find_get_block(inode->i_sb, nr);
2273                         ext3_forget(handle, 0, inode, bh, nr);
2274                 }
2275         }
2276
2277         ext3_free_blocks(handle, inode, block_to_free, count);
2278 }
2279
2280 /**
2281  * ext3_free_data - free a list of data blocks
2282  * @handle:     handle for this transaction
2283  * @inode:      inode we are dealing with
2284  * @this_bh:    indirect buffer_head which contains *@first and *@last
2285  * @first:      array of block numbers
2286  * @last:       points immediately past the end of array
2287  *
2288  * We are freeing all blocks referred from that array (numbers are stored as
2289  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2290  *
2291  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2292  * blocks are contiguous then releasing them at one time will only affect one
2293  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2294  * actually use a lot of journal space.
2295  *
2296  * @this_bh will be %NULL if @first and @last point into the inode's direct
2297  * block pointers.
2298  */
2299 static void ext3_free_data(handle_t *handle, struct inode *inode,
2300                            struct buffer_head *this_bh,
2301                            __le32 *first, __le32 *last)
2302 {
2303         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2304         unsigned long count = 0;            /* Number of blocks in the run */
2305         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2306                                                corresponding to
2307                                                block_to_free */
2308         ext3_fsblk_t nr;                    /* Current block # */
2309         __le32 *p;                          /* Pointer into inode/ind
2310                                                for current block */
2311         int err;
2312
2313         if (this_bh) {                          /* For indirect block */
2314                 BUFFER_TRACE(this_bh, "get_write_access");
2315                 err = ext3_journal_get_write_access(handle, this_bh);
2316                 /* Important: if we can't update the indirect pointers
2317                  * to the blocks, we can't free them. */
2318                 if (err)
2319                         return;
2320         }
2321
2322         for (p = first; p < last; p++) {
2323                 nr = le32_to_cpu(*p);
2324                 if (nr) {
2325                         /* accumulate blocks to free if they're contiguous */
2326                         if (count == 0) {
2327                                 block_to_free = nr;
2328                                 block_to_free_p = p;
2329                                 count = 1;
2330                         } else if (nr == block_to_free + count) {
2331                                 count++;
2332                         } else {
2333                                 ext3_clear_blocks(handle, inode, this_bh,
2334                                                   block_to_free,
2335                                                   count, block_to_free_p, p);
2336                                 block_to_free = nr;
2337                                 block_to_free_p = p;
2338                                 count = 1;
2339                         }
2340                 }
2341         }
2342
2343         if (count > 0)
2344                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2345                                   count, block_to_free_p, p);
2346
2347         if (this_bh) {
2348                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2349
2350                 /*
2351                  * The buffer head should have an attached journal head at this
2352                  * point. However, if the data is corrupted and an indirect
2353                  * block pointed to itself, it would have been detached when
2354                  * the block was cleared. Check for this instead of OOPSing.
2355                  */
2356                 if (bh2jh(this_bh))
2357                         ext3_journal_dirty_metadata(handle, this_bh);
2358                 else
2359                         ext3_error(inode->i_sb, "ext3_free_data",
2360                                    "circular indirect block detected, "
2361                                    "inode=%lu, block=%llu",
2362                                    inode->i_ino,
2363                                    (unsigned long long)this_bh->b_blocknr);
2364         }
2365 }
2366
2367 /**
2368  *      ext3_free_branches - free an array of branches
2369  *      @handle: JBD handle for this transaction
2370  *      @inode: inode we are dealing with
2371  *      @parent_bh: the buffer_head which contains *@first and *@last
2372  *      @first: array of block numbers
2373  *      @last:  pointer immediately past the end of array
2374  *      @depth: depth of the branches to free
2375  *
2376  *      We are freeing all blocks referred from these branches (numbers are
2377  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2378  *      appropriately.
2379  */
2380 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2381                                struct buffer_head *parent_bh,
2382                                __le32 *first, __le32 *last, int depth)
2383 {
2384         ext3_fsblk_t nr;
2385         __le32 *p;
2386
2387         if (is_handle_aborted(handle))
2388                 return;
2389
2390         if (depth--) {
2391                 struct buffer_head *bh;
2392                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2393                 p = last;
2394                 while (--p >= first) {
2395                         nr = le32_to_cpu(*p);
2396                         if (!nr)
2397                                 continue;               /* A hole */
2398
2399                         /* Go read the buffer for the next level down */
2400                         bh = sb_bread(inode->i_sb, nr);
2401
2402                         /*
2403                          * A read failure? Report error and clear slot
2404                          * (should be rare).
2405                          */
2406                         if (!bh) {
2407                                 ext3_error(inode->i_sb, "ext3_free_branches",
2408                                            "Read failure, inode=%lu, block="E3FSBLK,
2409                                            inode->i_ino, nr);
2410                                 continue;
2411                         }
2412
2413                         /* This zaps the entire block.  Bottom up. */
2414                         BUFFER_TRACE(bh, "free child branches");
2415                         ext3_free_branches(handle, inode, bh,
2416                                            (__le32*)bh->b_data,
2417                                            (__le32*)bh->b_data + addr_per_block,
2418                                            depth);
2419
2420                         /*
2421                          * Everything below this this pointer has been
2422                          * released.  Now let this top-of-subtree go.
2423                          *
2424                          * We want the freeing of this indirect block to be
2425                          * atomic in the journal with the updating of the
2426                          * bitmap block which owns it.  So make some room in
2427                          * the journal.
2428                          *
2429                          * We zero the parent pointer *after* freeing its
2430                          * pointee in the bitmaps, so if extend_transaction()
2431                          * for some reason fails to put the bitmap changes and
2432                          * the release into the same transaction, recovery
2433                          * will merely complain about releasing a free block,
2434                          * rather than leaking blocks.
2435                          */
2436                         if (is_handle_aborted(handle))
2437                                 return;
2438                         if (try_to_extend_transaction(handle, inode)) {
2439                                 ext3_mark_inode_dirty(handle, inode);
2440                                 truncate_restart_transaction(handle, inode);
2441                         }
2442
2443                         /*
2444                          * We've probably journalled the indirect block several
2445                          * times during the truncate.  But it's no longer
2446                          * needed and we now drop it from the transaction via
2447                          * journal_revoke().
2448                          *
2449                          * That's easy if it's exclusively part of this
2450                          * transaction.  But if it's part of the committing
2451                          * transaction then journal_forget() will simply
2452                          * brelse() it.  That means that if the underlying
2453                          * block is reallocated in ext3_get_block(),
2454                          * unmap_underlying_metadata() will find this block
2455                          * and will try to get rid of it.  damn, damn. Thus
2456                          * we don't allow a block to be reallocated until
2457                          * a transaction freeing it has fully committed.
2458                          *
2459                          * We also have to make sure journal replay after a
2460                          * crash does not overwrite non-journaled data blocks
2461                          * with old metadata when the block got reallocated for
2462                          * data.  Thus we have to store a revoke record for a
2463                          * block in the same transaction in which we free the
2464                          * block.
2465                          */
2466                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2467
2468                         ext3_free_blocks(handle, inode, nr, 1);
2469
2470                         if (parent_bh) {
2471                                 /*
2472                                  * The block which we have just freed is
2473                                  * pointed to by an indirect block: journal it
2474                                  */
2475                                 BUFFER_TRACE(parent_bh, "get_write_access");
2476                                 if (!ext3_journal_get_write_access(handle,
2477                                                                    parent_bh)){
2478                                         *p = 0;
2479                                         BUFFER_TRACE(parent_bh,
2480                                         "call ext3_journal_dirty_metadata");
2481                                         ext3_journal_dirty_metadata(handle,
2482                                                                     parent_bh);
2483                                 }
2484                         }
2485                 }
2486         } else {
2487                 /* We have reached the bottom of the tree. */
2488                 BUFFER_TRACE(parent_bh, "free data blocks");
2489                 ext3_free_data(handle, inode, parent_bh, first, last);
2490         }
2491 }
2492
2493 int ext3_can_truncate(struct inode *inode)
2494 {
2495         if (S_ISREG(inode->i_mode))
2496                 return 1;
2497         if (S_ISDIR(inode->i_mode))
2498                 return 1;
2499         if (S_ISLNK(inode->i_mode))
2500                 return !ext3_inode_is_fast_symlink(inode);
2501         return 0;
2502 }
2503
2504 /*
2505  * ext3_truncate()
2506  *
2507  * We block out ext3_get_block() block instantiations across the entire
2508  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2509  * simultaneously on behalf of the same inode.
2510  *
2511  * As we work through the truncate and commmit bits of it to the journal there
2512  * is one core, guiding principle: the file's tree must always be consistent on
2513  * disk.  We must be able to restart the truncate after a crash.
2514  *
2515  * The file's tree may be transiently inconsistent in memory (although it
2516  * probably isn't), but whenever we close off and commit a journal transaction,
2517  * the contents of (the filesystem + the journal) must be consistent and
2518  * restartable.  It's pretty simple, really: bottom up, right to left (although
2519  * left-to-right works OK too).
2520  *
2521  * Note that at recovery time, journal replay occurs *before* the restart of
2522  * truncate against the orphan inode list.
2523  *
2524  * The committed inode has the new, desired i_size (which is the same as
2525  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2526  * that this inode's truncate did not complete and it will again call
2527  * ext3_truncate() to have another go.  So there will be instantiated blocks
2528  * to the right of the truncation point in a crashed ext3 filesystem.  But
2529  * that's fine - as long as they are linked from the inode, the post-crash
2530  * ext3_truncate() run will find them and release them.
2531  */
2532 void ext3_truncate(struct inode *inode)
2533 {
2534         handle_t *handle;
2535         struct ext3_inode_info *ei = EXT3_I(inode);
2536         __le32 *i_data = ei->i_data;
2537         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2538         int offsets[4];
2539         Indirect chain[4];
2540         Indirect *partial;
2541         __le32 nr = 0;
2542         int n;
2543         long last_block;
2544         unsigned blocksize = inode->i_sb->s_blocksize;
2545
2546         trace_ext3_truncate_enter(inode);
2547
2548         if (!ext3_can_truncate(inode))
2549                 goto out_notrans;
2550
2551         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2552                 ext3_set_inode_state(inode, EXT3_STATE_FLUSH_ON_CLOSE);
2553
2554         handle = start_transaction(inode);
2555         if (IS_ERR(handle))
2556                 goto out_notrans;
2557
2558         last_block = (inode->i_size + blocksize-1)
2559                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2560         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2561         if (n == 0)
2562                 goto out_stop;  /* error */
2563
2564         /*
2565          * OK.  This truncate is going to happen.  We add the inode to the
2566          * orphan list, so that if this truncate spans multiple transactions,
2567          * and we crash, we will resume the truncate when the filesystem
2568          * recovers.  It also marks the inode dirty, to catch the new size.
2569          *
2570          * Implication: the file must always be in a sane, consistent
2571          * truncatable state while each transaction commits.
2572          */
2573         if (ext3_orphan_add(handle, inode))
2574                 goto out_stop;
2575
2576         /*
2577          * The orphan list entry will now protect us from any crash which
2578          * occurs before the truncate completes, so it is now safe to propagate
2579          * the new, shorter inode size (held for now in i_size) into the
2580          * on-disk inode. We do this via i_disksize, which is the value which
2581          * ext3 *really* writes onto the disk inode.
2582          */
2583         ei->i_disksize = inode->i_size;
2584
2585         /*
2586          * From here we block out all ext3_get_block() callers who want to
2587          * modify the block allocation tree.
2588          */
2589         mutex_lock(&ei->truncate_mutex);
2590
2591         if (n == 1) {           /* direct blocks */
2592                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2593                                i_data + EXT3_NDIR_BLOCKS);
2594                 goto do_indirects;
2595         }
2596
2597         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2598         /* Kill the top of shared branch (not detached) */
2599         if (nr) {
2600                 if (partial == chain) {
2601                         /* Shared branch grows from the inode */
2602                         ext3_free_branches(handle, inode, NULL,
2603                                            &nr, &nr+1, (chain+n-1) - partial);
2604                         *partial->p = 0;
2605                         /*
2606                          * We mark the inode dirty prior to restart,
2607                          * and prior to stop.  No need for it here.
2608                          */
2609                 } else {
2610                         /* Shared branch grows from an indirect block */
2611                         ext3_free_branches(handle, inode, partial->bh,
2612                                         partial->p,
2613                                         partial->p+1, (chain+n-1) - partial);
2614                 }
2615         }
2616         /* Clear the ends of indirect blocks on the shared branch */
2617         while (partial > chain) {
2618                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2619                                    (__le32*)partial->bh->b_data+addr_per_block,
2620                                    (chain+n-1) - partial);
2621                 BUFFER_TRACE(partial->bh, "call brelse");
2622                 brelse (partial->bh);
2623                 partial--;
2624         }
2625 do_indirects:
2626         /* Kill the remaining (whole) subtrees */
2627         switch (offsets[0]) {
2628         default:
2629                 nr = i_data[EXT3_IND_BLOCK];
2630                 if (nr) {
2631                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2632                         i_data[EXT3_IND_BLOCK] = 0;
2633                 }
2634         case EXT3_IND_BLOCK:
2635                 nr = i_data[EXT3_DIND_BLOCK];
2636                 if (nr) {
2637                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2638                         i_data[EXT3_DIND_BLOCK] = 0;
2639                 }
2640         case EXT3_DIND_BLOCK:
2641                 nr = i_data[EXT3_TIND_BLOCK];
2642                 if (nr) {
2643                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2644                         i_data[EXT3_TIND_BLOCK] = 0;
2645                 }
2646         case EXT3_TIND_BLOCK:
2647                 ;
2648         }
2649
2650         ext3_discard_reservation(inode);
2651
2652         mutex_unlock(&ei->truncate_mutex);
2653         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2654         ext3_mark_inode_dirty(handle, inode);
2655
2656         /*
2657          * In a multi-transaction truncate, we only make the final transaction
2658          * synchronous
2659          */
2660         if (IS_SYNC(inode))
2661                 handle->h_sync = 1;
2662 out_stop:
2663         /*
2664          * If this was a simple ftruncate(), and the file will remain alive
2665          * then we need to clear up the orphan record which we created above.
2666          * However, if this was a real unlink then we were called by
2667          * ext3_evict_inode(), and we allow that function to clean up the
2668          * orphan info for us.
2669          */
2670         if (inode->i_nlink)
2671                 ext3_orphan_del(handle, inode);
2672
2673         ext3_journal_stop(handle);
2674         trace_ext3_truncate_exit(inode);
2675         return;
2676 out_notrans:
2677         /*
2678          * Delete the inode from orphan list so that it doesn't stay there
2679          * forever and trigger assertion on umount.
2680          */
2681         if (inode->i_nlink)
2682                 ext3_orphan_del(NULL, inode);
2683         trace_ext3_truncate_exit(inode);
2684 }
2685
2686 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2687                 unsigned long ino, struct ext3_iloc *iloc)
2688 {
2689         unsigned long block_group;
2690         unsigned long offset;
2691         ext3_fsblk_t block;
2692         struct ext3_group_desc *gdp;
2693
2694         if (!ext3_valid_inum(sb, ino)) {
2695                 /*
2696                  * This error is already checked for in namei.c unless we are
2697                  * looking at an NFS filehandle, in which case no error
2698                  * report is needed
2699                  */
2700                 return 0;
2701         }
2702
2703         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2704         gdp = ext3_get_group_desc(sb, block_group, NULL);
2705         if (!gdp)
2706                 return 0;
2707         /*
2708          * Figure out the offset within the block group inode table
2709          */
2710         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2711                 EXT3_INODE_SIZE(sb);
2712         block = le32_to_cpu(gdp->bg_inode_table) +
2713                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2714
2715         iloc->block_group = block_group;
2716         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2717         return block;
2718 }
2719
2720 /*
2721  * ext3_get_inode_loc returns with an extra refcount against the inode's
2722  * underlying buffer_head on success. If 'in_mem' is true, we have all
2723  * data in memory that is needed to recreate the on-disk version of this
2724  * inode.
2725  */
2726 static int __ext3_get_inode_loc(struct inode *inode,
2727                                 struct ext3_iloc *iloc, int in_mem)
2728 {
2729         ext3_fsblk_t block;
2730         struct buffer_head *bh;
2731
2732         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2733         if (!block)
2734                 return -EIO;
2735
2736         bh = sb_getblk(inode->i_sb, block);
2737         if (!bh) {
2738                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2739                                 "unable to read inode block - "
2740                                 "inode=%lu, block="E3FSBLK,
2741                                  inode->i_ino, block);
2742                 return -EIO;
2743         }
2744         if (!buffer_uptodate(bh)) {
2745                 lock_buffer(bh);
2746
2747                 /*
2748                  * If the buffer has the write error flag, we have failed
2749                  * to write out another inode in the same block.  In this
2750                  * case, we don't have to read the block because we may
2751                  * read the old inode data successfully.
2752                  */
2753                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2754                         set_buffer_uptodate(bh);
2755
2756                 if (buffer_uptodate(bh)) {
2757                         /* someone brought it uptodate while we waited */
2758                         unlock_buffer(bh);
2759                         goto has_buffer;
2760                 }
2761
2762                 /*
2763                  * If we have all information of the inode in memory and this
2764                  * is the only valid inode in the block, we need not read the
2765                  * block.
2766                  */
2767                 if (in_mem) {
2768                         struct buffer_head *bitmap_bh;
2769                         struct ext3_group_desc *desc;
2770                         int inodes_per_buffer;
2771                         int inode_offset, i;
2772                         int block_group;
2773                         int start;
2774
2775                         block_group = (inode->i_ino - 1) /
2776                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2777                         inodes_per_buffer = bh->b_size /
2778                                 EXT3_INODE_SIZE(inode->i_sb);
2779                         inode_offset = ((inode->i_ino - 1) %
2780                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2781                         start = inode_offset & ~(inodes_per_buffer - 1);
2782
2783                         /* Is the inode bitmap in cache? */
2784                         desc = ext3_get_group_desc(inode->i_sb,
2785                                                 block_group, NULL);
2786                         if (!desc)
2787                                 goto make_io;
2788
2789                         bitmap_bh = sb_getblk(inode->i_sb,
2790                                         le32_to_cpu(desc->bg_inode_bitmap));
2791                         if (!bitmap_bh)
2792                                 goto make_io;
2793
2794                         /*
2795                          * If the inode bitmap isn't in cache then the
2796                          * optimisation may end up performing two reads instead
2797                          * of one, so skip it.
2798                          */
2799                         if (!buffer_uptodate(bitmap_bh)) {
2800                                 brelse(bitmap_bh);
2801                                 goto make_io;
2802                         }
2803                         for (i = start; i < start + inodes_per_buffer; i++) {
2804                                 if (i == inode_offset)
2805                                         continue;
2806                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2807                                         break;
2808                         }
2809                         brelse(bitmap_bh);
2810                         if (i == start + inodes_per_buffer) {
2811                                 /* all other inodes are free, so skip I/O */
2812                                 memset(bh->b_data, 0, bh->b_size);
2813                                 set_buffer_uptodate(bh);
2814                                 unlock_buffer(bh);
2815                                 goto has_buffer;
2816                         }
2817                 }
2818
2819 make_io:
2820                 /*
2821                  * There are other valid inodes in the buffer, this inode
2822                  * has in-inode xattrs, or we don't have this inode in memory.
2823                  * Read the block from disk.
2824                  */
2825                 trace_ext3_load_inode(inode);
2826                 get_bh(bh);
2827                 bh->b_end_io = end_buffer_read_sync;
2828                 submit_bh(READ | REQ_META | REQ_PRIO, bh);
2829                 wait_on_buffer(bh);
2830                 if (!buffer_uptodate(bh)) {
2831                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2832                                         "unable to read inode block - "
2833                                         "inode=%lu, block="E3FSBLK,
2834                                         inode->i_ino, block);
2835                         brelse(bh);
2836                         return -EIO;
2837                 }
2838         }
2839 has_buffer:
2840         iloc->bh = bh;
2841         return 0;
2842 }
2843
2844 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2845 {
2846         /* We have all inode data except xattrs in memory here. */
2847         return __ext3_get_inode_loc(inode, iloc,
2848                 !ext3_test_inode_state(inode, EXT3_STATE_XATTR));
2849 }
2850
2851 void ext3_set_inode_flags(struct inode *inode)
2852 {
2853         unsigned int flags = EXT3_I(inode)->i_flags;
2854
2855         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2856         if (flags & EXT3_SYNC_FL)
2857                 inode->i_flags |= S_SYNC;
2858         if (flags & EXT3_APPEND_FL)
2859                 inode->i_flags |= S_APPEND;
2860         if (flags & EXT3_IMMUTABLE_FL)
2861                 inode->i_flags |= S_IMMUTABLE;
2862         if (flags & EXT3_NOATIME_FL)
2863                 inode->i_flags |= S_NOATIME;
2864         if (flags & EXT3_DIRSYNC_FL)
2865                 inode->i_flags |= S_DIRSYNC;
2866 }
2867
2868 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2869 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2870 {
2871         unsigned int flags = ei->vfs_inode.i_flags;
2872
2873         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2874                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2875         if (flags & S_SYNC)
2876                 ei->i_flags |= EXT3_SYNC_FL;
2877         if (flags & S_APPEND)
2878                 ei->i_flags |= EXT3_APPEND_FL;
2879         if (flags & S_IMMUTABLE)
2880                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2881         if (flags & S_NOATIME)
2882                 ei->i_flags |= EXT3_NOATIME_FL;
2883         if (flags & S_DIRSYNC)
2884                 ei->i_flags |= EXT3_DIRSYNC_FL;
2885 }
2886
2887 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2888 {
2889         struct ext3_iloc iloc;
2890         struct ext3_inode *raw_inode;
2891         struct ext3_inode_info *ei;
2892         struct buffer_head *bh;
2893         struct inode *inode;
2894         journal_t *journal = EXT3_SB(sb)->s_journal;
2895         transaction_t *transaction;
2896         long ret;
2897         int block;
2898
2899         inode = iget_locked(sb, ino);
2900         if (!inode)
2901                 return ERR_PTR(-ENOMEM);
2902         if (!(inode->i_state & I_NEW))
2903                 return inode;
2904
2905         ei = EXT3_I(inode);
2906         ei->i_block_alloc_info = NULL;
2907
2908         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2909         if (ret < 0)
2910                 goto bad_inode;
2911         bh = iloc.bh;
2912         raw_inode = ext3_raw_inode(&iloc);
2913         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2914         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2915         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2916         if(!(test_opt (inode->i_sb, NO_UID32))) {
2917                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2918                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2919         }
2920         set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
2921         inode->i_size = le32_to_cpu(raw_inode->i_size);
2922         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2923         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2924         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2925         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2926
2927         ei->i_state_flags = 0;
2928         ei->i_dir_start_lookup = 0;
2929         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2930         /* We now have enough fields to check if the inode was active or not.
2931          * This is needed because nfsd might try to access dead inodes
2932          * the test is that same one that e2fsck uses
2933          * NeilBrown 1999oct15
2934          */
2935         if (inode->i_nlink == 0) {
2936                 if (inode->i_mode == 0 ||
2937                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2938                         /* this inode is deleted */
2939                         brelse (bh);
2940                         ret = -ESTALE;
2941                         goto bad_inode;
2942                 }
2943                 /* The only unlinked inodes we let through here have
2944                  * valid i_mode and are being read by the orphan
2945                  * recovery code: that's fine, we're about to complete
2946                  * the process of deleting those. */
2947         }
2948         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2949         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2950 #ifdef EXT3_FRAGMENTS
2951         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2952         ei->i_frag_no = raw_inode->i_frag;
2953         ei->i_frag_size = raw_inode->i_fsize;
2954 #endif
2955         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2956         if (!S_ISREG(inode->i_mode)) {
2957                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2958         } else {
2959                 inode->i_size |=
2960                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2961         }
2962         ei->i_disksize = inode->i_size;
2963         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2964         ei->i_block_group = iloc.block_group;
2965         /*
2966          * NOTE! The in-memory inode i_data array is in little-endian order
2967          * even on big-endian machines: we do NOT byteswap the block numbers!
2968          */
2969         for (block = 0; block < EXT3_N_BLOCKS; block++)
2970                 ei->i_data[block] = raw_inode->i_block[block];
2971         INIT_LIST_HEAD(&ei->i_orphan);
2972
2973         /*
2974          * Set transaction id's of transactions that have to be committed
2975          * to finish f[data]sync. We set them to currently running transaction
2976          * as we cannot be sure that the inode or some of its metadata isn't
2977          * part of the transaction - the inode could have been reclaimed and
2978          * now it is reread from disk.
2979          */
2980         if (journal) {
2981                 tid_t tid;
2982
2983                 spin_lock(&journal->j_state_lock);
2984                 if (journal->j_running_transaction)
2985                         transaction = journal->j_running_transaction;
2986                 else
2987                         transaction = journal->j_committing_transaction;
2988                 if (transaction)
2989                         tid = transaction->t_tid;
2990                 else
2991                         tid = journal->j_commit_sequence;
2992                 spin_unlock(&journal->j_state_lock);
2993                 atomic_set(&ei->i_sync_tid, tid);
2994                 atomic_set(&ei->i_datasync_tid, tid);
2995         }
2996
2997         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2998             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2999                 /*
3000                  * When mke2fs creates big inodes it does not zero out
3001                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
3002                  * so ignore those first few inodes.
3003                  */
3004                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3005                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3006                     EXT3_INODE_SIZE(inode->i_sb)) {
3007                         brelse (bh);
3008                         ret = -EIO;
3009                         goto bad_inode;
3010                 }
3011                 if (ei->i_extra_isize == 0) {
3012                         /* The extra space is currently unused. Use it. */
3013                         ei->i_extra_isize = sizeof(struct ext3_inode) -
3014                                             EXT3_GOOD_OLD_INODE_SIZE;
3015                 } else {
3016                         __le32 *magic = (void *)raw_inode +
3017                                         EXT3_GOOD_OLD_INODE_SIZE +
3018                                         ei->i_extra_isize;
3019                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
3020                                  ext3_set_inode_state(inode, EXT3_STATE_XATTR);
3021                 }
3022         } else
3023                 ei->i_extra_isize = 0;
3024
3025         if (S_ISREG(inode->i_mode)) {
3026                 inode->i_op = &ext3_file_inode_operations;
3027                 inode->i_fop = &ext3_file_operations;
3028                 ext3_set_aops(inode);
3029         } else if (S_ISDIR(inode->i_mode)) {
3030                 inode->i_op = &ext3_dir_inode_operations;
3031                 inode->i_fop = &ext3_dir_operations;
3032         } else if (S_ISLNK(inode->i_mode)) {
3033                 if (ext3_inode_is_fast_symlink(inode)) {
3034                         inode->i_op = &ext3_fast_symlink_inode_operations;
3035                         nd_terminate_link(ei->i_data, inode->i_size,
3036                                 sizeof(ei->i_data) - 1);
3037                 } else {
3038                         inode->i_op = &ext3_symlink_inode_operations;
3039                         ext3_set_aops(inode);
3040                 }
3041         } else {
3042                 inode->i_op = &ext3_special_inode_operations;
3043                 if (raw_inode->i_block[0])
3044                         init_special_inode(inode, inode->i_mode,
3045                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3046                 else
3047                         init_special_inode(inode, inode->i_mode,
3048                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3049         }
3050         brelse (iloc.bh);
3051         ext3_set_inode_flags(inode);
3052         unlock_new_inode(inode);
3053         return inode;
3054
3055 bad_inode:
3056         iget_failed(inode);
3057         return ERR_PTR(ret);
3058 }
3059
3060 /*
3061  * Post the struct inode info into an on-disk inode location in the
3062  * buffer-cache.  This gobbles the caller's reference to the
3063  * buffer_head in the inode location struct.
3064  *
3065  * The caller must have write access to iloc->bh.
3066  */
3067 static int ext3_do_update_inode(handle_t *handle,
3068                                 struct inode *inode,
3069                                 struct ext3_iloc *iloc)
3070 {
3071         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
3072         struct ext3_inode_info *ei = EXT3_I(inode);
3073         struct buffer_head *bh = iloc->bh;
3074         int err = 0, rc, block;
3075         int need_datasync = 0;
3076         __le32 disksize;
3077
3078 again:
3079         /* we can't allow multiple procs in here at once, its a bit racey */
3080         lock_buffer(bh);
3081
3082         /* For fields not not tracking in the in-memory inode,
3083          * initialise them to zero for new inodes. */
3084         if (ext3_test_inode_state(inode, EXT3_STATE_NEW))
3085                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
3086
3087         ext3_get_inode_flags(ei);
3088         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3089         if(!(test_opt(inode->i_sb, NO_UID32))) {
3090                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3091                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3092 /*
3093  * Fix up interoperability with old kernels. Otherwise, old inodes get
3094  * re-used with the upper 16 bits of the uid/gid intact
3095  */
3096                 if(!ei->i_dtime) {
3097                         raw_inode->i_uid_high =
3098                                 cpu_to_le16(high_16_bits(inode->i_uid));
3099                         raw_inode->i_gid_high =
3100                                 cpu_to_le16(high_16_bits(inode->i_gid));
3101                 } else {
3102                         raw_inode->i_uid_high = 0;
3103                         raw_inode->i_gid_high = 0;
3104                 }
3105         } else {
3106                 raw_inode->i_uid_low =
3107                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3108                 raw_inode->i_gid_low =
3109                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3110                 raw_inode->i_uid_high = 0;
3111                 raw_inode->i_gid_high = 0;
3112         }
3113         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3114         disksize = cpu_to_le32(ei->i_disksize);
3115         if (disksize != raw_inode->i_size) {
3116                 need_datasync = 1;
3117                 raw_inode->i_size = disksize;
3118         }
3119         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
3120         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
3121         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
3122         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
3123         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3124         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
3125 #ifdef EXT3_FRAGMENTS
3126         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
3127         raw_inode->i_frag = ei->i_frag_no;
3128         raw_inode->i_fsize = ei->i_frag_size;
3129 #endif
3130         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
3131         if (!S_ISREG(inode->i_mode)) {
3132                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
3133         } else {
3134                 disksize = cpu_to_le32(ei->i_disksize >> 32);
3135                 if (disksize != raw_inode->i_size_high) {
3136                         raw_inode->i_size_high = disksize;
3137                         need_datasync = 1;
3138                 }
3139                 if (ei->i_disksize > 0x7fffffffULL) {
3140                         struct super_block *sb = inode->i_sb;
3141                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
3142                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
3143                             EXT3_SB(sb)->s_es->s_rev_level ==
3144                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
3145                                /* If this is the first large file
3146                                 * created, add a flag to the superblock.
3147                                 */
3148                                 unlock_buffer(bh);
3149                                 err = ext3_journal_get_write_access(handle,
3150                                                 EXT3_SB(sb)->s_sbh);
3151                                 if (err)
3152                                         goto out_brelse;
3153
3154                                 ext3_update_dynamic_rev(sb);
3155                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
3156                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
3157                                 handle->h_sync = 1;
3158                                 err = ext3_journal_dirty_metadata(handle,
3159                                                 EXT3_SB(sb)->s_sbh);
3160                                 /* get our lock and start over */
3161                                 goto again;
3162                         }
3163                 }
3164         }
3165         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3166         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3167                 if (old_valid_dev(inode->i_rdev)) {
3168                         raw_inode->i_block[0] =
3169                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3170                         raw_inode->i_block[1] = 0;
3171                 } else {
3172                         raw_inode->i_block[0] = 0;
3173                         raw_inode->i_block[1] =
3174                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3175                         raw_inode->i_block[2] = 0;
3176                 }
3177         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
3178                 raw_inode->i_block[block] = ei->i_data[block];
3179
3180         if (ei->i_extra_isize)
3181                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3182
3183         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
3184         unlock_buffer(bh);
3185         rc = ext3_journal_dirty_metadata(handle, bh);
3186         if (!err)
3187                 err = rc;
3188         ext3_clear_inode_state(inode, EXT3_STATE_NEW);
3189
3190         atomic_set(&ei->i_sync_tid, handle->h_transaction->t_tid);
3191         if (need_datasync)
3192                 atomic_set(&ei->i_datasync_tid, handle->h_transaction->t_tid);
3193 out_brelse:
3194         brelse (bh);
3195         ext3_std_error(inode->i_sb, err);
3196         return err;
3197 }
3198
3199 /*
3200  * ext3_write_inode()
3201  *
3202  * We are called from a few places:
3203  *
3204  * - Within generic_file_write() for O_SYNC files.
3205  *   Here, there will be no transaction running. We wait for any running
3206  *   trasnaction to commit.
3207  *
3208  * - Within sys_sync(), kupdate and such.
3209  *   We wait on commit, if tol to.
3210  *
3211  * - Within prune_icache() (PF_MEMALLOC == true)
3212  *   Here we simply return.  We can't afford to block kswapd on the
3213  *   journal commit.
3214  *
3215  * In all cases it is actually safe for us to return without doing anything,
3216  * because the inode has been copied into a raw inode buffer in
3217  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3218  * knfsd.
3219  *
3220  * Note that we are absolutely dependent upon all inode dirtiers doing the
3221  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3222  * which we are interested.
3223  *
3224  * It would be a bug for them to not do this.  The code:
3225  *
3226  *      mark_inode_dirty(inode)
3227  *      stuff();
3228  *      inode->i_size = expr;
3229  *
3230  * is in error because a kswapd-driven write_inode() could occur while
3231  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3232  * will no longer be on the superblock's dirty inode list.
3233  */
3234 int ext3_write_inode(struct inode *inode, struct writeback_control *wbc)
3235 {
3236         if (current->flags & PF_MEMALLOC)
3237                 return 0;
3238
3239         if (ext3_journal_current_handle()) {
3240                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3241                 dump_stack();
3242                 return -EIO;
3243         }
3244
3245         if (wbc->sync_mode != WB_SYNC_ALL)
3246                 return 0;
3247
3248         return ext3_force_commit(inode->i_sb);
3249 }
3250
3251 /*
3252  * ext3_setattr()
3253  *
3254  * Called from notify_change.
3255  *
3256  * We want to trap VFS attempts to truncate the file as soon as
3257  * possible.  In particular, we want to make sure that when the VFS
3258  * shrinks i_size, we put the inode on the orphan list and modify
3259  * i_disksize immediately, so that during the subsequent flushing of
3260  * dirty pages and freeing of disk blocks, we can guarantee that any
3261  * commit will leave the blocks being flushed in an unused state on
3262  * disk.  (On recovery, the inode will get truncated and the blocks will
3263  * be freed, so we have a strong guarantee that no future commit will
3264  * leave these blocks visible to the user.)
3265  *
3266  * Called with inode->sem down.
3267  */
3268 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3269 {
3270         struct inode *inode = dentry->d_inode;
3271         int error, rc = 0;
3272         const unsigned int ia_valid = attr->ia_valid;
3273
3274         error = setattr_prepare(dentry, attr);
3275         if (error)
3276                 return error;
3277
3278         if (is_quota_modification(inode, attr))
3279                 dquot_initialize(inode);
3280         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3281                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3282                 handle_t *handle;
3283
3284                 /* (user+group)*(old+new) structure, inode write (sb,
3285                  * inode block, ? - but truncate inode update has it) */
3286                 handle = ext3_journal_start(inode, EXT3_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3287                                         EXT3_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)+3);
3288                 if (IS_ERR(handle)) {
3289                         error = PTR_ERR(handle);
3290                         goto err_out;
3291                 }
3292                 error = dquot_transfer(inode, attr);
3293                 if (error) {
3294                         ext3_journal_stop(handle);
3295                         return error;
3296                 }
3297                 /* Update corresponding info in inode so that everything is in
3298                  * one transaction */
3299                 if (attr->ia_valid & ATTR_UID)
3300                         inode->i_uid = attr->ia_uid;
3301                 if (attr->ia_valid & ATTR_GID)
3302                         inode->i_gid = attr->ia_gid;
3303                 error = ext3_mark_inode_dirty(handle, inode);
3304                 ext3_journal_stop(handle);
3305         }
3306
3307         if (attr->ia_valid & ATTR_SIZE)
3308                 inode_dio_wait(inode);
3309
3310         if (S_ISREG(inode->i_mode) &&
3311             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3312                 handle_t *handle;
3313
3314                 handle = ext3_journal_start(inode, 3);
3315                 if (IS_ERR(handle)) {
3316                         error = PTR_ERR(handle);
3317                         goto err_out;
3318                 }
3319
3320                 error = ext3_orphan_add(handle, inode);
3321                 if (error) {
3322                         ext3_journal_stop(handle);
3323                         goto err_out;
3324                 }
3325                 EXT3_I(inode)->i_disksize = attr->ia_size;
3326                 error = ext3_mark_inode_dirty(handle, inode);
3327                 ext3_journal_stop(handle);
3328                 if (error) {
3329                         /* Some hard fs error must have happened. Bail out. */
3330                         ext3_orphan_del(NULL, inode);
3331                         goto err_out;
3332                 }
3333                 rc = ext3_block_truncate_page(inode, attr->ia_size);
3334                 if (rc) {
3335                         /* Cleanup orphan list and exit */
3336                         handle = ext3_journal_start(inode, 3);
3337                         if (IS_ERR(handle)) {
3338                                 ext3_orphan_del(NULL, inode);
3339                                 goto err_out;
3340                         }
3341                         ext3_orphan_del(handle, inode);
3342                         ext3_journal_stop(handle);
3343                         goto err_out;
3344                 }
3345         }
3346
3347         if ((attr->ia_valid & ATTR_SIZE) &&
3348             attr->ia_size != i_size_read(inode)) {
3349                 truncate_setsize(inode, attr->ia_size);
3350                 ext3_truncate(inode);
3351         }
3352
3353         setattr_copy(inode, attr);
3354         mark_inode_dirty(inode);
3355
3356         if (ia_valid & ATTR_MODE)
3357                 rc = ext3_acl_chmod(inode);
3358
3359 err_out:
3360         ext3_std_error(inode->i_sb, error);
3361         if (!error)
3362                 error = rc;
3363         return error;
3364 }
3365
3366
3367 /*
3368  * How many blocks doth make a writepage()?
3369  *
3370  * With N blocks per page, it may be:
3371  * N data blocks
3372  * 2 indirect block
3373  * 2 dindirect
3374  * 1 tindirect
3375  * N+5 bitmap blocks (from the above)
3376  * N+5 group descriptor summary blocks
3377  * 1 inode block
3378  * 1 superblock.
3379  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3380  *
3381  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3382  *
3383  * With ordered or writeback data it's the same, less the N data blocks.
3384  *
3385  * If the inode's direct blocks can hold an integral number of pages then a
3386  * page cannot straddle two indirect blocks, and we can only touch one indirect
3387  * and dindirect block, and the "5" above becomes "3".
3388  *
3389  * This still overestimates under most circumstances.  If we were to pass the
3390  * start and end offsets in here as well we could do block_to_path() on each
3391  * block and work out the exact number of indirects which are touched.  Pah.
3392  */
3393
3394 static int ext3_writepage_trans_blocks(struct inode *inode)
3395 {
3396         int bpp = ext3_journal_blocks_per_page(inode);
3397         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3398         int ret;
3399
3400         if (ext3_should_journal_data(inode))
3401                 ret = 3 * (bpp + indirects) + 2;
3402         else
3403                 ret = 2 * (bpp + indirects) + indirects + 2;
3404
3405 #ifdef CONFIG_QUOTA
3406         /* We know that structure was already allocated during dquot_initialize so
3407          * we will be updating only the data blocks + inodes */
3408         ret += EXT3_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
3409 #endif
3410
3411         return ret;
3412 }
3413
3414 /*
3415  * The caller must have previously called ext3_reserve_inode_write().
3416  * Give this, we know that the caller already has write access to iloc->bh.
3417  */
3418 int ext3_mark_iloc_dirty(handle_t *handle,
3419                 struct inode *inode, struct ext3_iloc *iloc)
3420 {
3421         int err = 0;
3422
3423         /* the do_update_inode consumes one bh->b_count */
3424         get_bh(iloc->bh);
3425
3426         /* ext3_do_update_inode() does journal_dirty_metadata */
3427         err = ext3_do_update_inode(handle, inode, iloc);
3428         put_bh(iloc->bh);
3429         return err;
3430 }
3431
3432 /*
3433  * On success, We end up with an outstanding reference count against
3434  * iloc->bh.  This _must_ be cleaned up later.
3435  */
3436
3437 int
3438 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3439                          struct ext3_iloc *iloc)
3440 {
3441         int err = 0;
3442         if (handle) {
3443                 err = ext3_get_inode_loc(inode, iloc);
3444                 if (!err) {
3445                         BUFFER_TRACE(iloc->bh, "get_write_access");
3446                         err = ext3_journal_get_write_access(handle, iloc->bh);
3447                         if (err) {
3448                                 brelse(iloc->bh);
3449                                 iloc->bh = NULL;
3450                         }
3451                 }
3452         }
3453         ext3_std_error(inode->i_sb, err);
3454         return err;
3455 }
3456
3457 /*
3458  * What we do here is to mark the in-core inode as clean with respect to inode
3459  * dirtiness (it may still be data-dirty).
3460  * This means that the in-core inode may be reaped by prune_icache
3461  * without having to perform any I/O.  This is a very good thing,
3462  * because *any* task may call prune_icache - even ones which
3463  * have a transaction open against a different journal.
3464  *
3465  * Is this cheating?  Not really.  Sure, we haven't written the
3466  * inode out, but prune_icache isn't a user-visible syncing function.
3467  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3468  * we start and wait on commits.
3469  *
3470  * Is this efficient/effective?  Well, we're being nice to the system
3471  * by cleaning up our inodes proactively so they can be reaped
3472  * without I/O.  But we are potentially leaving up to five seconds'
3473  * worth of inodes floating about which prune_icache wants us to
3474  * write out.  One way to fix that would be to get prune_icache()
3475  * to do a write_super() to free up some memory.  It has the desired
3476  * effect.
3477  */
3478 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3479 {
3480         struct ext3_iloc iloc;
3481         int err;
3482
3483         might_sleep();
3484         trace_ext3_mark_inode_dirty(inode, _RET_IP_);
3485         err = ext3_reserve_inode_write(handle, inode, &iloc);
3486         if (!err)
3487                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3488         return err;
3489 }
3490
3491 /*
3492  * ext3_dirty_inode() is called from __mark_inode_dirty()
3493  *
3494  * We're really interested in the case where a file is being extended.
3495  * i_size has been changed by generic_commit_write() and we thus need
3496  * to include the updated inode in the current transaction.
3497  *
3498  * Also, dquot_alloc_space() will always dirty the inode when blocks
3499  * are allocated to the file.
3500  *
3501  * If the inode is marked synchronous, we don't honour that here - doing
3502  * so would cause a commit on atime updates, which we don't bother doing.
3503  * We handle synchronous inodes at the highest possible level.
3504  */
3505 void ext3_dirty_inode(struct inode *inode, int flags)
3506 {
3507         handle_t *current_handle = ext3_journal_current_handle();
3508         handle_t *handle;
3509
3510         handle = ext3_journal_start(inode, 2);
3511         if (IS_ERR(handle))
3512                 goto out;
3513         if (current_handle &&
3514                 current_handle->h_transaction != handle->h_transaction) {
3515                 /* This task has a transaction open against a different fs */
3516                 printk(KERN_EMERG "%s: transactions do not match!\n",
3517                        __func__);
3518         } else {
3519                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3520                                 current_handle);
3521                 ext3_mark_inode_dirty(handle, inode);
3522         }
3523         ext3_journal_stop(handle);
3524 out:
3525         return;
3526 }
3527
3528 #if 0
3529 /*
3530  * Bind an inode's backing buffer_head into this transaction, to prevent
3531  * it from being flushed to disk early.  Unlike
3532  * ext3_reserve_inode_write, this leaves behind no bh reference and
3533  * returns no iloc structure, so the caller needs to repeat the iloc
3534  * lookup to mark the inode dirty later.
3535  */
3536 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3537 {
3538         struct ext3_iloc iloc;
3539
3540         int err = 0;
3541         if (handle) {
3542                 err = ext3_get_inode_loc(inode, &iloc);
3543                 if (!err) {
3544                         BUFFER_TRACE(iloc.bh, "get_write_access");
3545                         err = journal_get_write_access(handle, iloc.bh);
3546                         if (!err)
3547                                 err = ext3_journal_dirty_metadata(handle,
3548                                                                   iloc.bh);
3549                         brelse(iloc.bh);
3550                 }
3551         }
3552         ext3_std_error(inode->i_sb, err);
3553         return err;
3554 }
3555 #endif
3556
3557 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3558 {
3559         journal_t *journal;
3560         handle_t *handle;
3561         int err;
3562
3563         /*
3564          * We have to be very careful here: changing a data block's
3565          * journaling status dynamically is dangerous.  If we write a
3566          * data block to the journal, change the status and then delete
3567          * that block, we risk forgetting to revoke the old log record
3568          * from the journal and so a subsequent replay can corrupt data.
3569          * So, first we make sure that the journal is empty and that
3570          * nobody is changing anything.
3571          */
3572
3573         journal = EXT3_JOURNAL(inode);
3574         if (is_journal_aborted(journal))
3575                 return -EROFS;
3576
3577         journal_lock_updates(journal);
3578         journal_flush(journal);
3579
3580         /*
3581          * OK, there are no updates running now, and all cached data is
3582          * synced to disk.  We are now in a completely consistent state
3583          * which doesn't have anything in the journal, and we know that
3584          * no filesystem updates are running, so it is safe to modify
3585          * the inode's in-core data-journaling state flag now.
3586          */
3587
3588         if (val)
3589                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3590         else
3591                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3592         ext3_set_aops(inode);
3593
3594         journal_unlock_updates(journal);
3595
3596         /* Finally we can mark the inode as dirty. */
3597
3598         handle = ext3_journal_start(inode, 1);
3599         if (IS_ERR(handle))
3600                 return PTR_ERR(handle);
3601
3602         err = ext3_mark_inode_dirty(handle, inode);
3603         handle->h_sync = 1;
3604         ext3_journal_stop(handle);
3605         ext3_std_error(inode->i_sb, err);
3606
3607         return err;
3608 }