ext3: Get rid of extenddisksize parameter of ext3_get_blocks_handle()
[linux-2.6.git] / fs / ext3 / inode.c
1 /*
2  *  linux/fs/ext3/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext3_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/ext3_jbd.h>
29 #include <linux/jbd.h>
30 #include <linux/highuid.h>
31 #include <linux/pagemap.h>
32 #include <linux/quotaops.h>
33 #include <linux/string.h>
34 #include <linux/buffer_head.h>
35 #include <linux/writeback.h>
36 #include <linux/mpage.h>
37 #include <linux/uio.h>
38 #include <linux/bio.h>
39 #include <linux/fiemap.h>
40 #include <linux/namei.h>
41 #include "xattr.h"
42 #include "acl.h"
43
44 static int ext3_writepage_trans_blocks(struct inode *inode);
45
46 /*
47  * Test whether an inode is a fast symlink.
48  */
49 static int ext3_inode_is_fast_symlink(struct inode *inode)
50 {
51         int ea_blocks = EXT3_I(inode)->i_file_acl ?
52                 (inode->i_sb->s_blocksize >> 9) : 0;
53
54         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
55 }
56
57 /*
58  * The ext3 forget function must perform a revoke if we are freeing data
59  * which has been journaled.  Metadata (eg. indirect blocks) must be
60  * revoked in all cases.
61  *
62  * "bh" may be NULL: a metadata block may have been freed from memory
63  * but there may still be a record of it in the journal, and that record
64  * still needs to be revoked.
65  */
66 int ext3_forget(handle_t *handle, int is_metadata, struct inode *inode,
67                         struct buffer_head *bh, ext3_fsblk_t blocknr)
68 {
69         int err;
70
71         might_sleep();
72
73         BUFFER_TRACE(bh, "enter");
74
75         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
76                   "data mode %lx\n",
77                   bh, is_metadata, inode->i_mode,
78                   test_opt(inode->i_sb, DATA_FLAGS));
79
80         /* Never use the revoke function if we are doing full data
81          * journaling: there is no need to, and a V1 superblock won't
82          * support it.  Otherwise, only skip the revoke on un-journaled
83          * data blocks. */
84
85         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT3_MOUNT_JOURNAL_DATA ||
86             (!is_metadata && !ext3_should_journal_data(inode))) {
87                 if (bh) {
88                         BUFFER_TRACE(bh, "call journal_forget");
89                         return ext3_journal_forget(handle, bh);
90                 }
91                 return 0;
92         }
93
94         /*
95          * data!=journal && (is_metadata || should_journal_data(inode))
96          */
97         BUFFER_TRACE(bh, "call ext3_journal_revoke");
98         err = ext3_journal_revoke(handle, blocknr, bh);
99         if (err)
100                 ext3_abort(inode->i_sb, __func__,
101                            "error %d when attempting revoke", err);
102         BUFFER_TRACE(bh, "exit");
103         return err;
104 }
105
106 /*
107  * Work out how many blocks we need to proceed with the next chunk of a
108  * truncate transaction.
109  */
110 static unsigned long blocks_for_truncate(struct inode *inode)
111 {
112         unsigned long needed;
113
114         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
115
116         /* Give ourselves just enough room to cope with inodes in which
117          * i_blocks is corrupt: we've seen disk corruptions in the past
118          * which resulted in random data in an inode which looked enough
119          * like a regular file for ext3 to try to delete it.  Things
120          * will go a bit crazy if that happens, but at least we should
121          * try not to panic the whole kernel. */
122         if (needed < 2)
123                 needed = 2;
124
125         /* But we need to bound the transaction so we don't overflow the
126          * journal. */
127         if (needed > EXT3_MAX_TRANS_DATA)
128                 needed = EXT3_MAX_TRANS_DATA;
129
130         return EXT3_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
131 }
132
133 /*
134  * Truncate transactions can be complex and absolutely huge.  So we need to
135  * be able to restart the transaction at a conventient checkpoint to make
136  * sure we don't overflow the journal.
137  *
138  * start_transaction gets us a new handle for a truncate transaction,
139  * and extend_transaction tries to extend the existing one a bit.  If
140  * extend fails, we need to propagate the failure up and restart the
141  * transaction in the top-level truncate loop. --sct
142  */
143 static handle_t *start_transaction(struct inode *inode)
144 {
145         handle_t *result;
146
147         result = ext3_journal_start(inode, blocks_for_truncate(inode));
148         if (!IS_ERR(result))
149                 return result;
150
151         ext3_std_error(inode->i_sb, PTR_ERR(result));
152         return result;
153 }
154
155 /*
156  * Try to extend this transaction for the purposes of truncation.
157  *
158  * Returns 0 if we managed to create more room.  If we can't create more
159  * room, and the transaction must be restarted we return 1.
160  */
161 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
162 {
163         if (handle->h_buffer_credits > EXT3_RESERVE_TRANS_BLOCKS)
164                 return 0;
165         if (!ext3_journal_extend(handle, blocks_for_truncate(inode)))
166                 return 0;
167         return 1;
168 }
169
170 /*
171  * Restart the transaction associated with *handle.  This does a commit,
172  * so before we call here everything must be consistently dirtied against
173  * this transaction.
174  */
175 static int ext3_journal_test_restart(handle_t *handle, struct inode *inode)
176 {
177         jbd_debug(2, "restarting handle %p\n", handle);
178         return ext3_journal_restart(handle, blocks_for_truncate(inode));
179 }
180
181 /*
182  * Called at the last iput() if i_nlink is zero.
183  */
184 void ext3_delete_inode (struct inode * inode)
185 {
186         handle_t *handle;
187
188         truncate_inode_pages(&inode->i_data, 0);
189
190         if (is_bad_inode(inode))
191                 goto no_delete;
192
193         handle = start_transaction(inode);
194         if (IS_ERR(handle)) {
195                 /*
196                  * If we're going to skip the normal cleanup, we still need to
197                  * make sure that the in-core orphan linked list is properly
198                  * cleaned up.
199                  */
200                 ext3_orphan_del(NULL, inode);
201                 goto no_delete;
202         }
203
204         if (IS_SYNC(inode))
205                 handle->h_sync = 1;
206         inode->i_size = 0;
207         if (inode->i_blocks)
208                 ext3_truncate(inode);
209         /*
210          * Kill off the orphan record which ext3_truncate created.
211          * AKPM: I think this can be inside the above `if'.
212          * Note that ext3_orphan_del() has to be able to cope with the
213          * deletion of a non-existent orphan - this is because we don't
214          * know if ext3_truncate() actually created an orphan record.
215          * (Well, we could do this if we need to, but heck - it works)
216          */
217         ext3_orphan_del(handle, inode);
218         EXT3_I(inode)->i_dtime  = get_seconds();
219
220         /*
221          * One subtle ordering requirement: if anything has gone wrong
222          * (transaction abort, IO errors, whatever), then we can still
223          * do these next steps (the fs will already have been marked as
224          * having errors), but we can't free the inode if the mark_dirty
225          * fails.
226          */
227         if (ext3_mark_inode_dirty(handle, inode))
228                 /* If that failed, just do the required in-core inode clear. */
229                 clear_inode(inode);
230         else
231                 ext3_free_inode(handle, inode);
232         ext3_journal_stop(handle);
233         return;
234 no_delete:
235         clear_inode(inode);     /* We must guarantee clearing of inode... */
236 }
237
238 typedef struct {
239         __le32  *p;
240         __le32  key;
241         struct buffer_head *bh;
242 } Indirect;
243
244 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
245 {
246         p->key = *(p->p = v);
247         p->bh = bh;
248 }
249
250 static int verify_chain(Indirect *from, Indirect *to)
251 {
252         while (from <= to && from->key == *from->p)
253                 from++;
254         return (from > to);
255 }
256
257 /**
258  *      ext3_block_to_path - parse the block number into array of offsets
259  *      @inode: inode in question (we are only interested in its superblock)
260  *      @i_block: block number to be parsed
261  *      @offsets: array to store the offsets in
262  *      @boundary: set this non-zero if the referred-to block is likely to be
263  *             followed (on disk) by an indirect block.
264  *
265  *      To store the locations of file's data ext3 uses a data structure common
266  *      for UNIX filesystems - tree of pointers anchored in the inode, with
267  *      data blocks at leaves and indirect blocks in intermediate nodes.
268  *      This function translates the block number into path in that tree -
269  *      return value is the path length and @offsets[n] is the offset of
270  *      pointer to (n+1)th node in the nth one. If @block is out of range
271  *      (negative or too large) warning is printed and zero returned.
272  *
273  *      Note: function doesn't find node addresses, so no IO is needed. All
274  *      we need to know is the capacity of indirect blocks (taken from the
275  *      inode->i_sb).
276  */
277
278 /*
279  * Portability note: the last comparison (check that we fit into triple
280  * indirect block) is spelled differently, because otherwise on an
281  * architecture with 32-bit longs and 8Kb pages we might get into trouble
282  * if our filesystem had 8Kb blocks. We might use long long, but that would
283  * kill us on x86. Oh, well, at least the sign propagation does not matter -
284  * i_block would have to be negative in the very beginning, so we would not
285  * get there at all.
286  */
287
288 static int ext3_block_to_path(struct inode *inode,
289                         long i_block, int offsets[4], int *boundary)
290 {
291         int ptrs = EXT3_ADDR_PER_BLOCK(inode->i_sb);
292         int ptrs_bits = EXT3_ADDR_PER_BLOCK_BITS(inode->i_sb);
293         const long direct_blocks = EXT3_NDIR_BLOCKS,
294                 indirect_blocks = ptrs,
295                 double_blocks = (1 << (ptrs_bits * 2));
296         int n = 0;
297         int final = 0;
298
299         if (i_block < 0) {
300                 ext3_warning (inode->i_sb, "ext3_block_to_path", "block < 0");
301         } else if (i_block < direct_blocks) {
302                 offsets[n++] = i_block;
303                 final = direct_blocks;
304         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
305                 offsets[n++] = EXT3_IND_BLOCK;
306                 offsets[n++] = i_block;
307                 final = ptrs;
308         } else if ((i_block -= indirect_blocks) < double_blocks) {
309                 offsets[n++] = EXT3_DIND_BLOCK;
310                 offsets[n++] = i_block >> ptrs_bits;
311                 offsets[n++] = i_block & (ptrs - 1);
312                 final = ptrs;
313         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
314                 offsets[n++] = EXT3_TIND_BLOCK;
315                 offsets[n++] = i_block >> (ptrs_bits * 2);
316                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
317                 offsets[n++] = i_block & (ptrs - 1);
318                 final = ptrs;
319         } else {
320                 ext3_warning(inode->i_sb, "ext3_block_to_path", "block > big");
321         }
322         if (boundary)
323                 *boundary = final - 1 - (i_block & (ptrs - 1));
324         return n;
325 }
326
327 /**
328  *      ext3_get_branch - read the chain of indirect blocks leading to data
329  *      @inode: inode in question
330  *      @depth: depth of the chain (1 - direct pointer, etc.)
331  *      @offsets: offsets of pointers in inode/indirect blocks
332  *      @chain: place to store the result
333  *      @err: here we store the error value
334  *
335  *      Function fills the array of triples <key, p, bh> and returns %NULL
336  *      if everything went OK or the pointer to the last filled triple
337  *      (incomplete one) otherwise. Upon the return chain[i].key contains
338  *      the number of (i+1)-th block in the chain (as it is stored in memory,
339  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
340  *      number (it points into struct inode for i==0 and into the bh->b_data
341  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
342  *      block for i>0 and NULL for i==0. In other words, it holds the block
343  *      numbers of the chain, addresses they were taken from (and where we can
344  *      verify that chain did not change) and buffer_heads hosting these
345  *      numbers.
346  *
347  *      Function stops when it stumbles upon zero pointer (absent block)
348  *              (pointer to last triple returned, *@err == 0)
349  *      or when it gets an IO error reading an indirect block
350  *              (ditto, *@err == -EIO)
351  *      or when it notices that chain had been changed while it was reading
352  *              (ditto, *@err == -EAGAIN)
353  *      or when it reads all @depth-1 indirect blocks successfully and finds
354  *      the whole chain, all way to the data (returns %NULL, *err == 0).
355  */
356 static Indirect *ext3_get_branch(struct inode *inode, int depth, int *offsets,
357                                  Indirect chain[4], int *err)
358 {
359         struct super_block *sb = inode->i_sb;
360         Indirect *p = chain;
361         struct buffer_head *bh;
362
363         *err = 0;
364         /* i_data is not going away, no lock needed */
365         add_chain (chain, NULL, EXT3_I(inode)->i_data + *offsets);
366         if (!p->key)
367                 goto no_block;
368         while (--depth) {
369                 bh = sb_bread(sb, le32_to_cpu(p->key));
370                 if (!bh)
371                         goto failure;
372                 /* Reader: pointers */
373                 if (!verify_chain(chain, p))
374                         goto changed;
375                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
376                 /* Reader: end */
377                 if (!p->key)
378                         goto no_block;
379         }
380         return NULL;
381
382 changed:
383         brelse(bh);
384         *err = -EAGAIN;
385         goto no_block;
386 failure:
387         *err = -EIO;
388 no_block:
389         return p;
390 }
391
392 /**
393  *      ext3_find_near - find a place for allocation with sufficient locality
394  *      @inode: owner
395  *      @ind: descriptor of indirect block.
396  *
397  *      This function returns the preferred place for block allocation.
398  *      It is used when heuristic for sequential allocation fails.
399  *      Rules are:
400  *        + if there is a block to the left of our position - allocate near it.
401  *        + if pointer will live in indirect block - allocate near that block.
402  *        + if pointer will live in inode - allocate in the same
403  *          cylinder group.
404  *
405  * In the latter case we colour the starting block by the callers PID to
406  * prevent it from clashing with concurrent allocations for a different inode
407  * in the same block group.   The PID is used here so that functionally related
408  * files will be close-by on-disk.
409  *
410  *      Caller must make sure that @ind is valid and will stay that way.
411  */
412 static ext3_fsblk_t ext3_find_near(struct inode *inode, Indirect *ind)
413 {
414         struct ext3_inode_info *ei = EXT3_I(inode);
415         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
416         __le32 *p;
417         ext3_fsblk_t bg_start;
418         ext3_grpblk_t colour;
419
420         /* Try to find previous block */
421         for (p = ind->p - 1; p >= start; p--) {
422                 if (*p)
423                         return le32_to_cpu(*p);
424         }
425
426         /* No such thing, so let's try location of indirect block */
427         if (ind->bh)
428                 return ind->bh->b_blocknr;
429
430         /*
431          * It is going to be referred to from the inode itself? OK, just put it
432          * into the same cylinder group then.
433          */
434         bg_start = ext3_group_first_block_no(inode->i_sb, ei->i_block_group);
435         colour = (current->pid % 16) *
436                         (EXT3_BLOCKS_PER_GROUP(inode->i_sb) / 16);
437         return bg_start + colour;
438 }
439
440 /**
441  *      ext3_find_goal - find a preferred place for allocation.
442  *      @inode: owner
443  *      @block:  block we want
444  *      @partial: pointer to the last triple within a chain
445  *
446  *      Normally this function find the preferred place for block allocation,
447  *      returns it.
448  */
449
450 static ext3_fsblk_t ext3_find_goal(struct inode *inode, long block,
451                                    Indirect *partial)
452 {
453         struct ext3_block_alloc_info *block_i;
454
455         block_i =  EXT3_I(inode)->i_block_alloc_info;
456
457         /*
458          * try the heuristic for sequential allocation,
459          * failing that at least try to get decent locality.
460          */
461         if (block_i && (block == block_i->last_alloc_logical_block + 1)
462                 && (block_i->last_alloc_physical_block != 0)) {
463                 return block_i->last_alloc_physical_block + 1;
464         }
465
466         return ext3_find_near(inode, partial);
467 }
468
469 /**
470  *      ext3_blks_to_allocate: Look up the block map and count the number
471  *      of direct blocks need to be allocated for the given branch.
472  *
473  *      @branch: chain of indirect blocks
474  *      @k: number of blocks need for indirect blocks
475  *      @blks: number of data blocks to be mapped.
476  *      @blocks_to_boundary:  the offset in the indirect block
477  *
478  *      return the total number of blocks to be allocate, including the
479  *      direct and indirect blocks.
480  */
481 static int ext3_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
482                 int blocks_to_boundary)
483 {
484         unsigned long count = 0;
485
486         /*
487          * Simple case, [t,d]Indirect block(s) has not allocated yet
488          * then it's clear blocks on that path have not allocated
489          */
490         if (k > 0) {
491                 /* right now we don't handle cross boundary allocation */
492                 if (blks < blocks_to_boundary + 1)
493                         count += blks;
494                 else
495                         count += blocks_to_boundary + 1;
496                 return count;
497         }
498
499         count++;
500         while (count < blks && count <= blocks_to_boundary &&
501                 le32_to_cpu(*(branch[0].p + count)) == 0) {
502                 count++;
503         }
504         return count;
505 }
506
507 /**
508  *      ext3_alloc_blocks: multiple allocate blocks needed for a branch
509  *      @indirect_blks: the number of blocks need to allocate for indirect
510  *                      blocks
511  *
512  *      @new_blocks: on return it will store the new block numbers for
513  *      the indirect blocks(if needed) and the first direct block,
514  *      @blks:  on return it will store the total number of allocated
515  *              direct blocks
516  */
517 static int ext3_alloc_blocks(handle_t *handle, struct inode *inode,
518                         ext3_fsblk_t goal, int indirect_blks, int blks,
519                         ext3_fsblk_t new_blocks[4], int *err)
520 {
521         int target, i;
522         unsigned long count = 0;
523         int index = 0;
524         ext3_fsblk_t current_block = 0;
525         int ret = 0;
526
527         /*
528          * Here we try to allocate the requested multiple blocks at once,
529          * on a best-effort basis.
530          * To build a branch, we should allocate blocks for
531          * the indirect blocks(if not allocated yet), and at least
532          * the first direct block of this branch.  That's the
533          * minimum number of blocks need to allocate(required)
534          */
535         target = blks + indirect_blks;
536
537         while (1) {
538                 count = target;
539                 /* allocating blocks for indirect blocks and direct blocks */
540                 current_block = ext3_new_blocks(handle,inode,goal,&count,err);
541                 if (*err)
542                         goto failed_out;
543
544                 target -= count;
545                 /* allocate blocks for indirect blocks */
546                 while (index < indirect_blks && count) {
547                         new_blocks[index++] = current_block++;
548                         count--;
549                 }
550
551                 if (count > 0)
552                         break;
553         }
554
555         /* save the new block number for the first direct block */
556         new_blocks[index] = current_block;
557
558         /* total number of blocks allocated for direct blocks */
559         ret = count;
560         *err = 0;
561         return ret;
562 failed_out:
563         for (i = 0; i <index; i++)
564                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
565         return ret;
566 }
567
568 /**
569  *      ext3_alloc_branch - allocate and set up a chain of blocks.
570  *      @inode: owner
571  *      @indirect_blks: number of allocated indirect blocks
572  *      @blks: number of allocated direct blocks
573  *      @offsets: offsets (in the blocks) to store the pointers to next.
574  *      @branch: place to store the chain in.
575  *
576  *      This function allocates blocks, zeroes out all but the last one,
577  *      links them into chain and (if we are synchronous) writes them to disk.
578  *      In other words, it prepares a branch that can be spliced onto the
579  *      inode. It stores the information about that chain in the branch[], in
580  *      the same format as ext3_get_branch() would do. We are calling it after
581  *      we had read the existing part of chain and partial points to the last
582  *      triple of that (one with zero ->key). Upon the exit we have the same
583  *      picture as after the successful ext3_get_block(), except that in one
584  *      place chain is disconnected - *branch->p is still zero (we did not
585  *      set the last link), but branch->key contains the number that should
586  *      be placed into *branch->p to fill that gap.
587  *
588  *      If allocation fails we free all blocks we've allocated (and forget
589  *      their buffer_heads) and return the error value the from failed
590  *      ext3_alloc_block() (normally -ENOSPC). Otherwise we set the chain
591  *      as described above and return 0.
592  */
593 static int ext3_alloc_branch(handle_t *handle, struct inode *inode,
594                         int indirect_blks, int *blks, ext3_fsblk_t goal,
595                         int *offsets, Indirect *branch)
596 {
597         int blocksize = inode->i_sb->s_blocksize;
598         int i, n = 0;
599         int err = 0;
600         struct buffer_head *bh;
601         int num;
602         ext3_fsblk_t new_blocks[4];
603         ext3_fsblk_t current_block;
604
605         num = ext3_alloc_blocks(handle, inode, goal, indirect_blks,
606                                 *blks, new_blocks, &err);
607         if (err)
608                 return err;
609
610         branch[0].key = cpu_to_le32(new_blocks[0]);
611         /*
612          * metadata blocks and data blocks are allocated.
613          */
614         for (n = 1; n <= indirect_blks;  n++) {
615                 /*
616                  * Get buffer_head for parent block, zero it out
617                  * and set the pointer to new one, then send
618                  * parent to disk.
619                  */
620                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
621                 branch[n].bh = bh;
622                 lock_buffer(bh);
623                 BUFFER_TRACE(bh, "call get_create_access");
624                 err = ext3_journal_get_create_access(handle, bh);
625                 if (err) {
626                         unlock_buffer(bh);
627                         brelse(bh);
628                         goto failed;
629                 }
630
631                 memset(bh->b_data, 0, blocksize);
632                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
633                 branch[n].key = cpu_to_le32(new_blocks[n]);
634                 *branch[n].p = branch[n].key;
635                 if ( n == indirect_blks) {
636                         current_block = new_blocks[n];
637                         /*
638                          * End of chain, update the last new metablock of
639                          * the chain to point to the new allocated
640                          * data blocks numbers
641                          */
642                         for (i=1; i < num; i++)
643                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
644                 }
645                 BUFFER_TRACE(bh, "marking uptodate");
646                 set_buffer_uptodate(bh);
647                 unlock_buffer(bh);
648
649                 BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
650                 err = ext3_journal_dirty_metadata(handle, bh);
651                 if (err)
652                         goto failed;
653         }
654         *blks = num;
655         return err;
656 failed:
657         /* Allocation failed, free what we already allocated */
658         for (i = 1; i <= n ; i++) {
659                 BUFFER_TRACE(branch[i].bh, "call journal_forget");
660                 ext3_journal_forget(handle, branch[i].bh);
661         }
662         for (i = 0; i <indirect_blks; i++)
663                 ext3_free_blocks(handle, inode, new_blocks[i], 1);
664
665         ext3_free_blocks(handle, inode, new_blocks[i], num);
666
667         return err;
668 }
669
670 /**
671  * ext3_splice_branch - splice the allocated branch onto inode.
672  * @inode: owner
673  * @block: (logical) number of block we are adding
674  * @chain: chain of indirect blocks (with a missing link - see
675  *      ext3_alloc_branch)
676  * @where: location of missing link
677  * @num:   number of indirect blocks we are adding
678  * @blks:  number of direct blocks we are adding
679  *
680  * This function fills the missing link and does all housekeeping needed in
681  * inode (->i_blocks, etc.). In case of success we end up with the full
682  * chain to new block and return 0.
683  */
684 static int ext3_splice_branch(handle_t *handle, struct inode *inode,
685                         long block, Indirect *where, int num, int blks)
686 {
687         int i;
688         int err = 0;
689         struct ext3_block_alloc_info *block_i;
690         ext3_fsblk_t current_block;
691
692         block_i = EXT3_I(inode)->i_block_alloc_info;
693         /*
694          * If we're splicing into a [td]indirect block (as opposed to the
695          * inode) then we need to get write access to the [td]indirect block
696          * before the splice.
697          */
698         if (where->bh) {
699                 BUFFER_TRACE(where->bh, "get_write_access");
700                 err = ext3_journal_get_write_access(handle, where->bh);
701                 if (err)
702                         goto err_out;
703         }
704         /* That's it */
705
706         *where->p = where->key;
707
708         /*
709          * Update the host buffer_head or inode to point to more just allocated
710          * direct blocks blocks
711          */
712         if (num == 0 && blks > 1) {
713                 current_block = le32_to_cpu(where->key) + 1;
714                 for (i = 1; i < blks; i++)
715                         *(where->p + i ) = cpu_to_le32(current_block++);
716         }
717
718         /*
719          * update the most recently allocated logical & physical block
720          * in i_block_alloc_info, to assist find the proper goal block for next
721          * allocation
722          */
723         if (block_i) {
724                 block_i->last_alloc_logical_block = block + blks - 1;
725                 block_i->last_alloc_physical_block =
726                                 le32_to_cpu(where[num].key) + blks - 1;
727         }
728
729         /* We are done with atomic stuff, now do the rest of housekeeping */
730
731         inode->i_ctime = CURRENT_TIME_SEC;
732         ext3_mark_inode_dirty(handle, inode);
733
734         /* had we spliced it onto indirect block? */
735         if (where->bh) {
736                 /*
737                  * If we spliced it onto an indirect block, we haven't
738                  * altered the inode.  Note however that if it is being spliced
739                  * onto an indirect block at the very end of the file (the
740                  * file is growing) then we *will* alter the inode to reflect
741                  * the new i_size.  But that is not done here - it is done in
742                  * generic_commit_write->__mark_inode_dirty->ext3_dirty_inode.
743                  */
744                 jbd_debug(5, "splicing indirect only\n");
745                 BUFFER_TRACE(where->bh, "call ext3_journal_dirty_metadata");
746                 err = ext3_journal_dirty_metadata(handle, where->bh);
747                 if (err)
748                         goto err_out;
749         } else {
750                 /*
751                  * OK, we spliced it into the inode itself on a direct block.
752                  * Inode was dirtied above.
753                  */
754                 jbd_debug(5, "splicing direct\n");
755         }
756         return err;
757
758 err_out:
759         for (i = 1; i <= num; i++) {
760                 BUFFER_TRACE(where[i].bh, "call journal_forget");
761                 ext3_journal_forget(handle, where[i].bh);
762                 ext3_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
763         }
764         ext3_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
765
766         return err;
767 }
768
769 /*
770  * Allocation strategy is simple: if we have to allocate something, we will
771  * have to go the whole way to leaf. So let's do it before attaching anything
772  * to tree, set linkage between the newborn blocks, write them if sync is
773  * required, recheck the path, free and repeat if check fails, otherwise
774  * set the last missing link (that will protect us from any truncate-generated
775  * removals - all blocks on the path are immune now) and possibly force the
776  * write on the parent block.
777  * That has a nice additional property: no special recovery from the failed
778  * allocations is needed - we simply release blocks and do not touch anything
779  * reachable from inode.
780  *
781  * `handle' can be NULL if create == 0.
782  *
783  * The BKL may not be held on entry here.  Be sure to take it early.
784  * return > 0, # of blocks mapped or allocated.
785  * return = 0, if plain lookup failed.
786  * return < 0, error case.
787  */
788 int ext3_get_blocks_handle(handle_t *handle, struct inode *inode,
789                 sector_t iblock, unsigned long maxblocks,
790                 struct buffer_head *bh_result,
791                 int create)
792 {
793         int err = -EIO;
794         int offsets[4];
795         Indirect chain[4];
796         Indirect *partial;
797         ext3_fsblk_t goal;
798         int indirect_blks;
799         int blocks_to_boundary = 0;
800         int depth;
801         struct ext3_inode_info *ei = EXT3_I(inode);
802         int count = 0;
803         ext3_fsblk_t first_block = 0;
804
805
806         J_ASSERT(handle != NULL || create == 0);
807         depth = ext3_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
808
809         if (depth == 0)
810                 goto out;
811
812         partial = ext3_get_branch(inode, depth, offsets, chain, &err);
813
814         /* Simplest case - block found, no allocation needed */
815         if (!partial) {
816                 first_block = le32_to_cpu(chain[depth - 1].key);
817                 clear_buffer_new(bh_result);
818                 count++;
819                 /*map more blocks*/
820                 while (count < maxblocks && count <= blocks_to_boundary) {
821                         ext3_fsblk_t blk;
822
823                         if (!verify_chain(chain, chain + depth - 1)) {
824                                 /*
825                                  * Indirect block might be removed by
826                                  * truncate while we were reading it.
827                                  * Handling of that case: forget what we've
828                                  * got now. Flag the err as EAGAIN, so it
829                                  * will reread.
830                                  */
831                                 err = -EAGAIN;
832                                 count = 0;
833                                 break;
834                         }
835                         blk = le32_to_cpu(*(chain[depth-1].p + count));
836
837                         if (blk == first_block + count)
838                                 count++;
839                         else
840                                 break;
841                 }
842                 if (err != -EAGAIN)
843                         goto got_it;
844         }
845
846         /* Next simple case - plain lookup or failed read of indirect block */
847         if (!create || err == -EIO)
848                 goto cleanup;
849
850         mutex_lock(&ei->truncate_mutex);
851
852         /*
853          * If the indirect block is missing while we are reading
854          * the chain(ext3_get_branch() returns -EAGAIN err), or
855          * if the chain has been changed after we grab the semaphore,
856          * (either because another process truncated this branch, or
857          * another get_block allocated this branch) re-grab the chain to see if
858          * the request block has been allocated or not.
859          *
860          * Since we already block the truncate/other get_block
861          * at this point, we will have the current copy of the chain when we
862          * splice the branch into the tree.
863          */
864         if (err == -EAGAIN || !verify_chain(chain, partial)) {
865                 while (partial > chain) {
866                         brelse(partial->bh);
867                         partial--;
868                 }
869                 partial = ext3_get_branch(inode, depth, offsets, chain, &err);
870                 if (!partial) {
871                         count++;
872                         mutex_unlock(&ei->truncate_mutex);
873                         if (err)
874                                 goto cleanup;
875                         clear_buffer_new(bh_result);
876                         goto got_it;
877                 }
878         }
879
880         /*
881          * Okay, we need to do block allocation.  Lazily initialize the block
882          * allocation info here if necessary
883         */
884         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
885                 ext3_init_block_alloc_info(inode);
886
887         goal = ext3_find_goal(inode, iblock, partial);
888
889         /* the number of blocks need to allocate for [d,t]indirect blocks */
890         indirect_blks = (chain + depth) - partial - 1;
891
892         /*
893          * Next look up the indirect map to count the totoal number of
894          * direct blocks to allocate for this branch.
895          */
896         count = ext3_blks_to_allocate(partial, indirect_blks,
897                                         maxblocks, blocks_to_boundary);
898         /*
899          * Block out ext3_truncate while we alter the tree
900          */
901         err = ext3_alloc_branch(handle, inode, indirect_blks, &count, goal,
902                                 offsets + (partial - chain), partial);
903
904         /*
905          * The ext3_splice_branch call will free and forget any buffers
906          * on the new chain if there is a failure, but that risks using
907          * up transaction credits, especially for bitmaps where the
908          * credits cannot be returned.  Can we handle this somehow?  We
909          * may need to return -EAGAIN upwards in the worst case.  --sct
910          */
911         if (!err)
912                 err = ext3_splice_branch(handle, inode, iblock,
913                                         partial, indirect_blks, count);
914         mutex_unlock(&ei->truncate_mutex);
915         if (err)
916                 goto cleanup;
917
918         set_buffer_new(bh_result);
919 got_it:
920         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
921         if (count > blocks_to_boundary)
922                 set_buffer_boundary(bh_result);
923         err = count;
924         /* Clean up and exit */
925         partial = chain + depth - 1;    /* the whole chain */
926 cleanup:
927         while (partial > chain) {
928                 BUFFER_TRACE(partial->bh, "call brelse");
929                 brelse(partial->bh);
930                 partial--;
931         }
932         BUFFER_TRACE(bh_result, "returned");
933 out:
934         return err;
935 }
936
937 /* Maximum number of blocks we map for direct IO at once. */
938 #define DIO_MAX_BLOCKS 4096
939 /*
940  * Number of credits we need for writing DIO_MAX_BLOCKS:
941  * We need sb + group descriptor + bitmap + inode -> 4
942  * For B blocks with A block pointers per block we need:
943  * 1 (triple ind.) + (B/A/A + 2) (doubly ind.) + (B/A + 2) (indirect).
944  * If we plug in 4096 for B and 256 for A (for 1KB block size), we get 25.
945  */
946 #define DIO_CREDITS 25
947
948 static int ext3_get_block(struct inode *inode, sector_t iblock,
949                         struct buffer_head *bh_result, int create)
950 {
951         handle_t *handle = ext3_journal_current_handle();
952         int ret = 0, started = 0;
953         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
954
955         if (create && !handle) {        /* Direct IO write... */
956                 if (max_blocks > DIO_MAX_BLOCKS)
957                         max_blocks = DIO_MAX_BLOCKS;
958                 handle = ext3_journal_start(inode, DIO_CREDITS +
959                                 2 * EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb));
960                 if (IS_ERR(handle)) {
961                         ret = PTR_ERR(handle);
962                         goto out;
963                 }
964                 started = 1;
965         }
966
967         ret = ext3_get_blocks_handle(handle, inode, iblock,
968                                         max_blocks, bh_result, create);
969         if (ret > 0) {
970                 bh_result->b_size = (ret << inode->i_blkbits);
971                 ret = 0;
972         }
973         if (started)
974                 ext3_journal_stop(handle);
975 out:
976         return ret;
977 }
978
979 int ext3_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
980                 u64 start, u64 len)
981 {
982         return generic_block_fiemap(inode, fieinfo, start, len,
983                                     ext3_get_block);
984 }
985
986 /*
987  * `handle' can be NULL if create is zero
988  */
989 struct buffer_head *ext3_getblk(handle_t *handle, struct inode *inode,
990                                 long block, int create, int *errp)
991 {
992         struct buffer_head dummy;
993         int fatal = 0, err;
994
995         J_ASSERT(handle != NULL || create == 0);
996
997         dummy.b_state = 0;
998         dummy.b_blocknr = -1000;
999         buffer_trace_init(&dummy.b_history);
1000         err = ext3_get_blocks_handle(handle, inode, block, 1,
1001                                         &dummy, create);
1002         /*
1003          * ext3_get_blocks_handle() returns number of blocks
1004          * mapped. 0 in case of a HOLE.
1005          */
1006         if (err > 0) {
1007                 if (err > 1)
1008                         WARN_ON(1);
1009                 err = 0;
1010         }
1011         *errp = err;
1012         if (!err && buffer_mapped(&dummy)) {
1013                 struct buffer_head *bh;
1014                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1015                 if (!bh) {
1016                         *errp = -EIO;
1017                         goto err;
1018                 }
1019                 if (buffer_new(&dummy)) {
1020                         J_ASSERT(create != 0);
1021                         J_ASSERT(handle != NULL);
1022
1023                         /*
1024                          * Now that we do not always journal data, we should
1025                          * keep in mind whether this should always journal the
1026                          * new buffer as metadata.  For now, regular file
1027                          * writes use ext3_get_block instead, so it's not a
1028                          * problem.
1029                          */
1030                         lock_buffer(bh);
1031                         BUFFER_TRACE(bh, "call get_create_access");
1032                         fatal = ext3_journal_get_create_access(handle, bh);
1033                         if (!fatal && !buffer_uptodate(bh)) {
1034                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1035                                 set_buffer_uptodate(bh);
1036                         }
1037                         unlock_buffer(bh);
1038                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
1039                         err = ext3_journal_dirty_metadata(handle, bh);
1040                         if (!fatal)
1041                                 fatal = err;
1042                 } else {
1043                         BUFFER_TRACE(bh, "not a new buffer");
1044                 }
1045                 if (fatal) {
1046                         *errp = fatal;
1047                         brelse(bh);
1048                         bh = NULL;
1049                 }
1050                 return bh;
1051         }
1052 err:
1053         return NULL;
1054 }
1055
1056 struct buffer_head *ext3_bread(handle_t *handle, struct inode *inode,
1057                                int block, int create, int *err)
1058 {
1059         struct buffer_head * bh;
1060
1061         bh = ext3_getblk(handle, inode, block, create, err);
1062         if (!bh)
1063                 return bh;
1064         if (buffer_uptodate(bh))
1065                 return bh;
1066         ll_rw_block(READ_META, 1, &bh);
1067         wait_on_buffer(bh);
1068         if (buffer_uptodate(bh))
1069                 return bh;
1070         put_bh(bh);
1071         *err = -EIO;
1072         return NULL;
1073 }
1074
1075 static int walk_page_buffers(   handle_t *handle,
1076                                 struct buffer_head *head,
1077                                 unsigned from,
1078                                 unsigned to,
1079                                 int *partial,
1080                                 int (*fn)(      handle_t *handle,
1081                                                 struct buffer_head *bh))
1082 {
1083         struct buffer_head *bh;
1084         unsigned block_start, block_end;
1085         unsigned blocksize = head->b_size;
1086         int err, ret = 0;
1087         struct buffer_head *next;
1088
1089         for (   bh = head, block_start = 0;
1090                 ret == 0 && (bh != head || !block_start);
1091                 block_start = block_end, bh = next)
1092         {
1093                 next = bh->b_this_page;
1094                 block_end = block_start + blocksize;
1095                 if (block_end <= from || block_start >= to) {
1096                         if (partial && !buffer_uptodate(bh))
1097                                 *partial = 1;
1098                         continue;
1099                 }
1100                 err = (*fn)(handle, bh);
1101                 if (!ret)
1102                         ret = err;
1103         }
1104         return ret;
1105 }
1106
1107 /*
1108  * To preserve ordering, it is essential that the hole instantiation and
1109  * the data write be encapsulated in a single transaction.  We cannot
1110  * close off a transaction and start a new one between the ext3_get_block()
1111  * and the commit_write().  So doing the journal_start at the start of
1112  * prepare_write() is the right place.
1113  *
1114  * Also, this function can nest inside ext3_writepage() ->
1115  * block_write_full_page(). In that case, we *know* that ext3_writepage()
1116  * has generated enough buffer credits to do the whole page.  So we won't
1117  * block on the journal in that case, which is good, because the caller may
1118  * be PF_MEMALLOC.
1119  *
1120  * By accident, ext3 can be reentered when a transaction is open via
1121  * quota file writes.  If we were to commit the transaction while thus
1122  * reentered, there can be a deadlock - we would be holding a quota
1123  * lock, and the commit would never complete if another thread had a
1124  * transaction open and was blocking on the quota lock - a ranking
1125  * violation.
1126  *
1127  * So what we do is to rely on the fact that journal_stop/journal_start
1128  * will _not_ run commit under these circumstances because handle->h_ref
1129  * is elevated.  We'll still have enough credits for the tiny quotafile
1130  * write.
1131  */
1132 static int do_journal_get_write_access(handle_t *handle,
1133                                         struct buffer_head *bh)
1134 {
1135         if (!buffer_mapped(bh) || buffer_freed(bh))
1136                 return 0;
1137         return ext3_journal_get_write_access(handle, bh);
1138 }
1139
1140 static int ext3_write_begin(struct file *file, struct address_space *mapping,
1141                                 loff_t pos, unsigned len, unsigned flags,
1142                                 struct page **pagep, void **fsdata)
1143 {
1144         struct inode *inode = mapping->host;
1145         int ret;
1146         handle_t *handle;
1147         int retries = 0;
1148         struct page *page;
1149         pgoff_t index;
1150         unsigned from, to;
1151         /* Reserve one block more for addition to orphan list in case
1152          * we allocate blocks but write fails for some reason */
1153         int needed_blocks = ext3_writepage_trans_blocks(inode) + 1;
1154
1155         index = pos >> PAGE_CACHE_SHIFT;
1156         from = pos & (PAGE_CACHE_SIZE - 1);
1157         to = from + len;
1158
1159 retry:
1160         page = grab_cache_page_write_begin(mapping, index, flags);
1161         if (!page)
1162                 return -ENOMEM;
1163         *pagep = page;
1164
1165         handle = ext3_journal_start(inode, needed_blocks);
1166         if (IS_ERR(handle)) {
1167                 unlock_page(page);
1168                 page_cache_release(page);
1169                 ret = PTR_ERR(handle);
1170                 goto out;
1171         }
1172         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1173                                                         ext3_get_block);
1174         if (ret)
1175                 goto write_begin_failed;
1176
1177         if (ext3_should_journal_data(inode)) {
1178                 ret = walk_page_buffers(handle, page_buffers(page),
1179                                 from, to, NULL, do_journal_get_write_access);
1180         }
1181 write_begin_failed:
1182         if (ret) {
1183                 /*
1184                  * block_write_begin may have instantiated a few blocks
1185                  * outside i_size.  Trim these off again. Don't need
1186                  * i_size_read because we hold i_mutex.
1187                  *
1188                  * Add inode to orphan list in case we crash before truncate
1189                  * finishes. Do this only if ext3_can_truncate() agrees so
1190                  * that orphan processing code is happy.
1191                  */
1192                 if (pos + len > inode->i_size && ext3_can_truncate(inode))
1193                         ext3_orphan_add(handle, inode);
1194                 ext3_journal_stop(handle);
1195                 unlock_page(page);
1196                 page_cache_release(page);
1197                 if (pos + len > inode->i_size)
1198                         ext3_truncate(inode);
1199         }
1200         if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
1201                 goto retry;
1202 out:
1203         return ret;
1204 }
1205
1206
1207 int ext3_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
1208 {
1209         int err = journal_dirty_data(handle, bh);
1210         if (err)
1211                 ext3_journal_abort_handle(__func__, __func__,
1212                                                 bh, handle, err);
1213         return err;
1214 }
1215
1216 /* For ordered writepage and write_end functions */
1217 static int journal_dirty_data_fn(handle_t *handle, struct buffer_head *bh)
1218 {
1219         /*
1220          * Write could have mapped the buffer but it didn't copy the data in
1221          * yet. So avoid filing such buffer into a transaction.
1222          */
1223         if (buffer_mapped(bh) && buffer_uptodate(bh))
1224                 return ext3_journal_dirty_data(handle, bh);
1225         return 0;
1226 }
1227
1228 /* For write_end() in data=journal mode */
1229 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1230 {
1231         if (!buffer_mapped(bh) || buffer_freed(bh))
1232                 return 0;
1233         set_buffer_uptodate(bh);
1234         return ext3_journal_dirty_metadata(handle, bh);
1235 }
1236
1237 /*
1238  * This is nasty and subtle: ext3_write_begin() could have allocated blocks
1239  * for the whole page but later we failed to copy the data in. Update inode
1240  * size according to what we managed to copy. The rest is going to be
1241  * truncated in write_end function.
1242  */
1243 static void update_file_sizes(struct inode *inode, loff_t pos, unsigned copied)
1244 {
1245         /* What matters to us is i_disksize. We don't write i_size anywhere */
1246         if (pos + copied > inode->i_size)
1247                 i_size_write(inode, pos + copied);
1248         if (pos + copied > EXT3_I(inode)->i_disksize) {
1249                 EXT3_I(inode)->i_disksize = pos + copied;
1250                 mark_inode_dirty(inode);
1251         }
1252 }
1253
1254 /*
1255  * We need to pick up the new inode size which generic_commit_write gave us
1256  * `file' can be NULL - eg, when called from page_symlink().
1257  *
1258  * ext3 never places buffers on inode->i_mapping->private_list.  metadata
1259  * buffers are managed internally.
1260  */
1261 static int ext3_ordered_write_end(struct file *file,
1262                                 struct address_space *mapping,
1263                                 loff_t pos, unsigned len, unsigned copied,
1264                                 struct page *page, void *fsdata)
1265 {
1266         handle_t *handle = ext3_journal_current_handle();
1267         struct inode *inode = file->f_mapping->host;
1268         unsigned from, to;
1269         int ret = 0, ret2;
1270
1271         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1272
1273         from = pos & (PAGE_CACHE_SIZE - 1);
1274         to = from + copied;
1275         ret = walk_page_buffers(handle, page_buffers(page),
1276                 from, to, NULL, journal_dirty_data_fn);
1277
1278         if (ret == 0)
1279                 update_file_sizes(inode, pos, copied);
1280         /*
1281          * There may be allocated blocks outside of i_size because
1282          * we failed to copy some data. Prepare for truncate.
1283          */
1284         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1285                 ext3_orphan_add(handle, inode);
1286         ret2 = ext3_journal_stop(handle);
1287         if (!ret)
1288                 ret = ret2;
1289         unlock_page(page);
1290         page_cache_release(page);
1291
1292         if (pos + len > inode->i_size)
1293                 ext3_truncate(inode);
1294         return ret ? ret : copied;
1295 }
1296
1297 static int ext3_writeback_write_end(struct file *file,
1298                                 struct address_space *mapping,
1299                                 loff_t pos, unsigned len, unsigned copied,
1300                                 struct page *page, void *fsdata)
1301 {
1302         handle_t *handle = ext3_journal_current_handle();
1303         struct inode *inode = file->f_mapping->host;
1304         int ret;
1305
1306         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1307         update_file_sizes(inode, pos, copied);
1308         /*
1309          * There may be allocated blocks outside of i_size because
1310          * we failed to copy some data. Prepare for truncate.
1311          */
1312         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1313                 ext3_orphan_add(handle, inode);
1314         ret = ext3_journal_stop(handle);
1315         unlock_page(page);
1316         page_cache_release(page);
1317
1318         if (pos + len > inode->i_size)
1319                 ext3_truncate(inode);
1320         return ret ? ret : copied;
1321 }
1322
1323 static int ext3_journalled_write_end(struct file *file,
1324                                 struct address_space *mapping,
1325                                 loff_t pos, unsigned len, unsigned copied,
1326                                 struct page *page, void *fsdata)
1327 {
1328         handle_t *handle = ext3_journal_current_handle();
1329         struct inode *inode = mapping->host;
1330         int ret = 0, ret2;
1331         int partial = 0;
1332         unsigned from, to;
1333
1334         from = pos & (PAGE_CACHE_SIZE - 1);
1335         to = from + len;
1336
1337         if (copied < len) {
1338                 if (!PageUptodate(page))
1339                         copied = 0;
1340                 page_zero_new_buffers(page, from + copied, to);
1341                 to = from + copied;
1342         }
1343
1344         ret = walk_page_buffers(handle, page_buffers(page), from,
1345                                 to, &partial, write_end_fn);
1346         if (!partial)
1347                 SetPageUptodate(page);
1348
1349         if (pos + copied > inode->i_size)
1350                 i_size_write(inode, pos + copied);
1351         /*
1352          * There may be allocated blocks outside of i_size because
1353          * we failed to copy some data. Prepare for truncate.
1354          */
1355         if (pos + len > inode->i_size && ext3_can_truncate(inode))
1356                 ext3_orphan_add(handle, inode);
1357         EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1358         if (inode->i_size > EXT3_I(inode)->i_disksize) {
1359                 EXT3_I(inode)->i_disksize = inode->i_size;
1360                 ret2 = ext3_mark_inode_dirty(handle, inode);
1361                 if (!ret)
1362                         ret = ret2;
1363         }
1364
1365         ret2 = ext3_journal_stop(handle);
1366         if (!ret)
1367                 ret = ret2;
1368         unlock_page(page);
1369         page_cache_release(page);
1370
1371         if (pos + len > inode->i_size)
1372                 ext3_truncate(inode);
1373         return ret ? ret : copied;
1374 }
1375
1376 /*
1377  * bmap() is special.  It gets used by applications such as lilo and by
1378  * the swapper to find the on-disk block of a specific piece of data.
1379  *
1380  * Naturally, this is dangerous if the block concerned is still in the
1381  * journal.  If somebody makes a swapfile on an ext3 data-journaling
1382  * filesystem and enables swap, then they may get a nasty shock when the
1383  * data getting swapped to that swapfile suddenly gets overwritten by
1384  * the original zero's written out previously to the journal and
1385  * awaiting writeback in the kernel's buffer cache.
1386  *
1387  * So, if we see any bmap calls here on a modified, data-journaled file,
1388  * take extra steps to flush any blocks which might be in the cache.
1389  */
1390 static sector_t ext3_bmap(struct address_space *mapping, sector_t block)
1391 {
1392         struct inode *inode = mapping->host;
1393         journal_t *journal;
1394         int err;
1395
1396         if (EXT3_I(inode)->i_state & EXT3_STATE_JDATA) {
1397                 /*
1398                  * This is a REALLY heavyweight approach, but the use of
1399                  * bmap on dirty files is expected to be extremely rare:
1400                  * only if we run lilo or swapon on a freshly made file
1401                  * do we expect this to happen.
1402                  *
1403                  * (bmap requires CAP_SYS_RAWIO so this does not
1404                  * represent an unprivileged user DOS attack --- we'd be
1405                  * in trouble if mortal users could trigger this path at
1406                  * will.)
1407                  *
1408                  * NB. EXT3_STATE_JDATA is not set on files other than
1409                  * regular files.  If somebody wants to bmap a directory
1410                  * or symlink and gets confused because the buffer
1411                  * hasn't yet been flushed to disk, they deserve
1412                  * everything they get.
1413                  */
1414
1415                 EXT3_I(inode)->i_state &= ~EXT3_STATE_JDATA;
1416                 journal = EXT3_JOURNAL(inode);
1417                 journal_lock_updates(journal);
1418                 err = journal_flush(journal);
1419                 journal_unlock_updates(journal);
1420
1421                 if (err)
1422                         return 0;
1423         }
1424
1425         return generic_block_bmap(mapping,block,ext3_get_block);
1426 }
1427
1428 static int bget_one(handle_t *handle, struct buffer_head *bh)
1429 {
1430         get_bh(bh);
1431         return 0;
1432 }
1433
1434 static int bput_one(handle_t *handle, struct buffer_head *bh)
1435 {
1436         put_bh(bh);
1437         return 0;
1438 }
1439
1440 static int buffer_unmapped(handle_t *handle, struct buffer_head *bh)
1441 {
1442         return !buffer_mapped(bh);
1443 }
1444
1445 /*
1446  * Note that we always start a transaction even if we're not journalling
1447  * data.  This is to preserve ordering: any hole instantiation within
1448  * __block_write_full_page -> ext3_get_block() should be journalled
1449  * along with the data so we don't crash and then get metadata which
1450  * refers to old data.
1451  *
1452  * In all journalling modes block_write_full_page() will start the I/O.
1453  *
1454  * Problem:
1455  *
1456  *      ext3_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1457  *              ext3_writepage()
1458  *
1459  * Similar for:
1460  *
1461  *      ext3_file_write() -> generic_file_write() -> __alloc_pages() -> ...
1462  *
1463  * Same applies to ext3_get_block().  We will deadlock on various things like
1464  * lock_journal and i_truncate_mutex.
1465  *
1466  * Setting PF_MEMALLOC here doesn't work - too many internal memory
1467  * allocations fail.
1468  *
1469  * 16May01: If we're reentered then journal_current_handle() will be
1470  *          non-zero. We simply *return*.
1471  *
1472  * 1 July 2001: @@@ FIXME:
1473  *   In journalled data mode, a data buffer may be metadata against the
1474  *   current transaction.  But the same file is part of a shared mapping
1475  *   and someone does a writepage() on it.
1476  *
1477  *   We will move the buffer onto the async_data list, but *after* it has
1478  *   been dirtied. So there's a small window where we have dirty data on
1479  *   BJ_Metadata.
1480  *
1481  *   Note that this only applies to the last partial page in the file.  The
1482  *   bit which block_write_full_page() uses prepare/commit for.  (That's
1483  *   broken code anyway: it's wrong for msync()).
1484  *
1485  *   It's a rare case: affects the final partial page, for journalled data
1486  *   where the file is subject to bith write() and writepage() in the same
1487  *   transction.  To fix it we'll need a custom block_write_full_page().
1488  *   We'll probably need that anyway for journalling writepage() output.
1489  *
1490  * We don't honour synchronous mounts for writepage().  That would be
1491  * disastrous.  Any write() or metadata operation will sync the fs for
1492  * us.
1493  *
1494  * AKPM2: if all the page's buffers are mapped to disk and !data=journal,
1495  * we don't need to open a transaction here.
1496  */
1497 static int ext3_ordered_writepage(struct page *page,
1498                                 struct writeback_control *wbc)
1499 {
1500         struct inode *inode = page->mapping->host;
1501         struct buffer_head *page_bufs;
1502         handle_t *handle = NULL;
1503         int ret = 0;
1504         int err;
1505
1506         J_ASSERT(PageLocked(page));
1507
1508         /*
1509          * We give up here if we're reentered, because it might be for a
1510          * different filesystem.
1511          */
1512         if (ext3_journal_current_handle())
1513                 goto out_fail;
1514
1515         if (!page_has_buffers(page)) {
1516                 create_empty_buffers(page, inode->i_sb->s_blocksize,
1517                                 (1 << BH_Dirty)|(1 << BH_Uptodate));
1518                 page_bufs = page_buffers(page);
1519         } else {
1520                 page_bufs = page_buffers(page);
1521                 if (!walk_page_buffers(NULL, page_bufs, 0, PAGE_CACHE_SIZE,
1522                                        NULL, buffer_unmapped)) {
1523                         /* Provide NULL get_block() to catch bugs if buffers
1524                          * weren't really mapped */
1525                         return block_write_full_page(page, NULL, wbc);
1526                 }
1527         }
1528         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1529
1530         if (IS_ERR(handle)) {
1531                 ret = PTR_ERR(handle);
1532                 goto out_fail;
1533         }
1534
1535         walk_page_buffers(handle, page_bufs, 0,
1536                         PAGE_CACHE_SIZE, NULL, bget_one);
1537
1538         ret = block_write_full_page(page, ext3_get_block, wbc);
1539
1540         /*
1541          * The page can become unlocked at any point now, and
1542          * truncate can then come in and change things.  So we
1543          * can't touch *page from now on.  But *page_bufs is
1544          * safe due to elevated refcount.
1545          */
1546
1547         /*
1548          * And attach them to the current transaction.  But only if
1549          * block_write_full_page() succeeded.  Otherwise they are unmapped,
1550          * and generally junk.
1551          */
1552         if (ret == 0) {
1553                 err = walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE,
1554                                         NULL, journal_dirty_data_fn);
1555                 if (!ret)
1556                         ret = err;
1557         }
1558         walk_page_buffers(handle, page_bufs, 0,
1559                         PAGE_CACHE_SIZE, NULL, bput_one);
1560         err = ext3_journal_stop(handle);
1561         if (!ret)
1562                 ret = err;
1563         return ret;
1564
1565 out_fail:
1566         redirty_page_for_writepage(wbc, page);
1567         unlock_page(page);
1568         return ret;
1569 }
1570
1571 static int ext3_writeback_writepage(struct page *page,
1572                                 struct writeback_control *wbc)
1573 {
1574         struct inode *inode = page->mapping->host;
1575         handle_t *handle = NULL;
1576         int ret = 0;
1577         int err;
1578
1579         if (ext3_journal_current_handle())
1580                 goto out_fail;
1581
1582         if (page_has_buffers(page)) {
1583                 if (!walk_page_buffers(NULL, page_buffers(page), 0,
1584                                       PAGE_CACHE_SIZE, NULL, buffer_unmapped)) {
1585                         /* Provide NULL get_block() to catch bugs if buffers
1586                          * weren't really mapped */
1587                         return block_write_full_page(page, NULL, wbc);
1588                 }
1589         }
1590
1591         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1592         if (IS_ERR(handle)) {
1593                 ret = PTR_ERR(handle);
1594                 goto out_fail;
1595         }
1596
1597         if (test_opt(inode->i_sb, NOBH) && ext3_should_writeback_data(inode))
1598                 ret = nobh_writepage(page, ext3_get_block, wbc);
1599         else
1600                 ret = block_write_full_page(page, ext3_get_block, wbc);
1601
1602         err = ext3_journal_stop(handle);
1603         if (!ret)
1604                 ret = err;
1605         return ret;
1606
1607 out_fail:
1608         redirty_page_for_writepage(wbc, page);
1609         unlock_page(page);
1610         return ret;
1611 }
1612
1613 static int ext3_journalled_writepage(struct page *page,
1614                                 struct writeback_control *wbc)
1615 {
1616         struct inode *inode = page->mapping->host;
1617         handle_t *handle = NULL;
1618         int ret = 0;
1619         int err;
1620
1621         if (ext3_journal_current_handle())
1622                 goto no_write;
1623
1624         handle = ext3_journal_start(inode, ext3_writepage_trans_blocks(inode));
1625         if (IS_ERR(handle)) {
1626                 ret = PTR_ERR(handle);
1627                 goto no_write;
1628         }
1629
1630         if (!page_has_buffers(page) || PageChecked(page)) {
1631                 /*
1632                  * It's mmapped pagecache.  Add buffers and journal it.  There
1633                  * doesn't seem much point in redirtying the page here.
1634                  */
1635                 ClearPageChecked(page);
1636                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
1637                                         ext3_get_block);
1638                 if (ret != 0) {
1639                         ext3_journal_stop(handle);
1640                         goto out_unlock;
1641                 }
1642                 ret = walk_page_buffers(handle, page_buffers(page), 0,
1643                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
1644
1645                 err = walk_page_buffers(handle, page_buffers(page), 0,
1646                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
1647                 if (ret == 0)
1648                         ret = err;
1649                 EXT3_I(inode)->i_state |= EXT3_STATE_JDATA;
1650                 unlock_page(page);
1651         } else {
1652                 /*
1653                  * It may be a page full of checkpoint-mode buffers.  We don't
1654                  * really know unless we go poke around in the buffer_heads.
1655                  * But block_write_full_page will do the right thing.
1656                  */
1657                 ret = block_write_full_page(page, ext3_get_block, wbc);
1658         }
1659         err = ext3_journal_stop(handle);
1660         if (!ret)
1661                 ret = err;
1662 out:
1663         return ret;
1664
1665 no_write:
1666         redirty_page_for_writepage(wbc, page);
1667 out_unlock:
1668         unlock_page(page);
1669         goto out;
1670 }
1671
1672 static int ext3_readpage(struct file *file, struct page *page)
1673 {
1674         return mpage_readpage(page, ext3_get_block);
1675 }
1676
1677 static int
1678 ext3_readpages(struct file *file, struct address_space *mapping,
1679                 struct list_head *pages, unsigned nr_pages)
1680 {
1681         return mpage_readpages(mapping, pages, nr_pages, ext3_get_block);
1682 }
1683
1684 static void ext3_invalidatepage(struct page *page, unsigned long offset)
1685 {
1686         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1687
1688         /*
1689          * If it's a full truncate we just forget about the pending dirtying
1690          */
1691         if (offset == 0)
1692                 ClearPageChecked(page);
1693
1694         journal_invalidatepage(journal, page, offset);
1695 }
1696
1697 static int ext3_releasepage(struct page *page, gfp_t wait)
1698 {
1699         journal_t *journal = EXT3_JOURNAL(page->mapping->host);
1700
1701         WARN_ON(PageChecked(page));
1702         if (!page_has_buffers(page))
1703                 return 0;
1704         return journal_try_to_free_buffers(journal, page, wait);
1705 }
1706
1707 /*
1708  * If the O_DIRECT write will extend the file then add this inode to the
1709  * orphan list.  So recovery will truncate it back to the original size
1710  * if the machine crashes during the write.
1711  *
1712  * If the O_DIRECT write is intantiating holes inside i_size and the machine
1713  * crashes then stale disk data _may_ be exposed inside the file. But current
1714  * VFS code falls back into buffered path in that case so we are safe.
1715  */
1716 static ssize_t ext3_direct_IO(int rw, struct kiocb *iocb,
1717                         const struct iovec *iov, loff_t offset,
1718                         unsigned long nr_segs)
1719 {
1720         struct file *file = iocb->ki_filp;
1721         struct inode *inode = file->f_mapping->host;
1722         struct ext3_inode_info *ei = EXT3_I(inode);
1723         handle_t *handle;
1724         ssize_t ret;
1725         int orphan = 0;
1726         size_t count = iov_length(iov, nr_segs);
1727
1728         if (rw == WRITE) {
1729                 loff_t final_size = offset + count;
1730
1731                 if (final_size > inode->i_size) {
1732                         /* Credits for sb + inode write */
1733                         handle = ext3_journal_start(inode, 2);
1734                         if (IS_ERR(handle)) {
1735                                 ret = PTR_ERR(handle);
1736                                 goto out;
1737                         }
1738                         ret = ext3_orphan_add(handle, inode);
1739                         if (ret) {
1740                                 ext3_journal_stop(handle);
1741                                 goto out;
1742                         }
1743                         orphan = 1;
1744                         ei->i_disksize = inode->i_size;
1745                         ext3_journal_stop(handle);
1746                 }
1747         }
1748
1749         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
1750                                  offset, nr_segs,
1751                                  ext3_get_block, NULL);
1752
1753         if (orphan) {
1754                 int err;
1755
1756                 /* Credits for sb + inode write */
1757                 handle = ext3_journal_start(inode, 2);
1758                 if (IS_ERR(handle)) {
1759                         /* This is really bad luck. We've written the data
1760                          * but cannot extend i_size. Bail out and pretend
1761                          * the write failed... */
1762                         ret = PTR_ERR(handle);
1763                         goto out;
1764                 }
1765                 if (inode->i_nlink)
1766                         ext3_orphan_del(handle, inode);
1767                 if (ret > 0) {
1768                         loff_t end = offset + ret;
1769                         if (end > inode->i_size) {
1770                                 ei->i_disksize = end;
1771                                 i_size_write(inode, end);
1772                                 /*
1773                                  * We're going to return a positive `ret'
1774                                  * here due to non-zero-length I/O, so there's
1775                                  * no way of reporting error returns from
1776                                  * ext3_mark_inode_dirty() to userspace.  So
1777                                  * ignore it.
1778                                  */
1779                                 ext3_mark_inode_dirty(handle, inode);
1780                         }
1781                 }
1782                 err = ext3_journal_stop(handle);
1783                 if (ret == 0)
1784                         ret = err;
1785         }
1786 out:
1787         return ret;
1788 }
1789
1790 /*
1791  * Pages can be marked dirty completely asynchronously from ext3's journalling
1792  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
1793  * much here because ->set_page_dirty is called under VFS locks.  The page is
1794  * not necessarily locked.
1795  *
1796  * We cannot just dirty the page and leave attached buffers clean, because the
1797  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
1798  * or jbddirty because all the journalling code will explode.
1799  *
1800  * So what we do is to mark the page "pending dirty" and next time writepage
1801  * is called, propagate that into the buffers appropriately.
1802  */
1803 static int ext3_journalled_set_page_dirty(struct page *page)
1804 {
1805         SetPageChecked(page);
1806         return __set_page_dirty_nobuffers(page);
1807 }
1808
1809 static const struct address_space_operations ext3_ordered_aops = {
1810         .readpage               = ext3_readpage,
1811         .readpages              = ext3_readpages,
1812         .writepage              = ext3_ordered_writepage,
1813         .sync_page              = block_sync_page,
1814         .write_begin            = ext3_write_begin,
1815         .write_end              = ext3_ordered_write_end,
1816         .bmap                   = ext3_bmap,
1817         .invalidatepage         = ext3_invalidatepage,
1818         .releasepage            = ext3_releasepage,
1819         .direct_IO              = ext3_direct_IO,
1820         .migratepage            = buffer_migrate_page,
1821         .is_partially_uptodate  = block_is_partially_uptodate,
1822 };
1823
1824 static const struct address_space_operations ext3_writeback_aops = {
1825         .readpage               = ext3_readpage,
1826         .readpages              = ext3_readpages,
1827         .writepage              = ext3_writeback_writepage,
1828         .sync_page              = block_sync_page,
1829         .write_begin            = ext3_write_begin,
1830         .write_end              = ext3_writeback_write_end,
1831         .bmap                   = ext3_bmap,
1832         .invalidatepage         = ext3_invalidatepage,
1833         .releasepage            = ext3_releasepage,
1834         .direct_IO              = ext3_direct_IO,
1835         .migratepage            = buffer_migrate_page,
1836         .is_partially_uptodate  = block_is_partially_uptodate,
1837 };
1838
1839 static const struct address_space_operations ext3_journalled_aops = {
1840         .readpage               = ext3_readpage,
1841         .readpages              = ext3_readpages,
1842         .writepage              = ext3_journalled_writepage,
1843         .sync_page              = block_sync_page,
1844         .write_begin            = ext3_write_begin,
1845         .write_end              = ext3_journalled_write_end,
1846         .set_page_dirty         = ext3_journalled_set_page_dirty,
1847         .bmap                   = ext3_bmap,
1848         .invalidatepage         = ext3_invalidatepage,
1849         .releasepage            = ext3_releasepage,
1850         .is_partially_uptodate  = block_is_partially_uptodate,
1851 };
1852
1853 void ext3_set_aops(struct inode *inode)
1854 {
1855         if (ext3_should_order_data(inode))
1856                 inode->i_mapping->a_ops = &ext3_ordered_aops;
1857         else if (ext3_should_writeback_data(inode))
1858                 inode->i_mapping->a_ops = &ext3_writeback_aops;
1859         else
1860                 inode->i_mapping->a_ops = &ext3_journalled_aops;
1861 }
1862
1863 /*
1864  * ext3_block_truncate_page() zeroes out a mapping from file offset `from'
1865  * up to the end of the block which corresponds to `from'.
1866  * This required during truncate. We need to physically zero the tail end
1867  * of that block so it doesn't yield old data if the file is later grown.
1868  */
1869 static int ext3_block_truncate_page(handle_t *handle, struct page *page,
1870                 struct address_space *mapping, loff_t from)
1871 {
1872         ext3_fsblk_t index = from >> PAGE_CACHE_SHIFT;
1873         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1874         unsigned blocksize, iblock, length, pos;
1875         struct inode *inode = mapping->host;
1876         struct buffer_head *bh;
1877         int err = 0;
1878
1879         blocksize = inode->i_sb->s_blocksize;
1880         length = blocksize - (offset & (blocksize - 1));
1881         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1882
1883         /*
1884          * For "nobh" option,  we can only work if we don't need to
1885          * read-in the page - otherwise we create buffers to do the IO.
1886          */
1887         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
1888              ext3_should_writeback_data(inode) && PageUptodate(page)) {
1889                 zero_user(page, offset, length);
1890                 set_page_dirty(page);
1891                 goto unlock;
1892         }
1893
1894         if (!page_has_buffers(page))
1895                 create_empty_buffers(page, blocksize, 0);
1896
1897         /* Find the buffer that contains "offset" */
1898         bh = page_buffers(page);
1899         pos = blocksize;
1900         while (offset >= pos) {
1901                 bh = bh->b_this_page;
1902                 iblock++;
1903                 pos += blocksize;
1904         }
1905
1906         err = 0;
1907         if (buffer_freed(bh)) {
1908                 BUFFER_TRACE(bh, "freed: skip");
1909                 goto unlock;
1910         }
1911
1912         if (!buffer_mapped(bh)) {
1913                 BUFFER_TRACE(bh, "unmapped");
1914                 ext3_get_block(inode, iblock, bh, 0);
1915                 /* unmapped? It's a hole - nothing to do */
1916                 if (!buffer_mapped(bh)) {
1917                         BUFFER_TRACE(bh, "still unmapped");
1918                         goto unlock;
1919                 }
1920         }
1921
1922         /* Ok, it's mapped. Make sure it's up-to-date */
1923         if (PageUptodate(page))
1924                 set_buffer_uptodate(bh);
1925
1926         if (!buffer_uptodate(bh)) {
1927                 err = -EIO;
1928                 ll_rw_block(READ, 1, &bh);
1929                 wait_on_buffer(bh);
1930                 /* Uhhuh. Read error. Complain and punt. */
1931                 if (!buffer_uptodate(bh))
1932                         goto unlock;
1933         }
1934
1935         if (ext3_should_journal_data(inode)) {
1936                 BUFFER_TRACE(bh, "get write access");
1937                 err = ext3_journal_get_write_access(handle, bh);
1938                 if (err)
1939                         goto unlock;
1940         }
1941
1942         zero_user(page, offset, length);
1943         BUFFER_TRACE(bh, "zeroed end of block");
1944
1945         err = 0;
1946         if (ext3_should_journal_data(inode)) {
1947                 err = ext3_journal_dirty_metadata(handle, bh);
1948         } else {
1949                 if (ext3_should_order_data(inode))
1950                         err = ext3_journal_dirty_data(handle, bh);
1951                 mark_buffer_dirty(bh);
1952         }
1953
1954 unlock:
1955         unlock_page(page);
1956         page_cache_release(page);
1957         return err;
1958 }
1959
1960 /*
1961  * Probably it should be a library function... search for first non-zero word
1962  * or memcmp with zero_page, whatever is better for particular architecture.
1963  * Linus?
1964  */
1965 static inline int all_zeroes(__le32 *p, __le32 *q)
1966 {
1967         while (p < q)
1968                 if (*p++)
1969                         return 0;
1970         return 1;
1971 }
1972
1973 /**
1974  *      ext3_find_shared - find the indirect blocks for partial truncation.
1975  *      @inode:   inode in question
1976  *      @depth:   depth of the affected branch
1977  *      @offsets: offsets of pointers in that branch (see ext3_block_to_path)
1978  *      @chain:   place to store the pointers to partial indirect blocks
1979  *      @top:     place to the (detached) top of branch
1980  *
1981  *      This is a helper function used by ext3_truncate().
1982  *
1983  *      When we do truncate() we may have to clean the ends of several
1984  *      indirect blocks but leave the blocks themselves alive. Block is
1985  *      partially truncated if some data below the new i_size is refered
1986  *      from it (and it is on the path to the first completely truncated
1987  *      data block, indeed).  We have to free the top of that path along
1988  *      with everything to the right of the path. Since no allocation
1989  *      past the truncation point is possible until ext3_truncate()
1990  *      finishes, we may safely do the latter, but top of branch may
1991  *      require special attention - pageout below the truncation point
1992  *      might try to populate it.
1993  *
1994  *      We atomically detach the top of branch from the tree, store the
1995  *      block number of its root in *@top, pointers to buffer_heads of
1996  *      partially truncated blocks - in @chain[].bh and pointers to
1997  *      their last elements that should not be removed - in
1998  *      @chain[].p. Return value is the pointer to last filled element
1999  *      of @chain.
2000  *
2001  *      The work left to caller to do the actual freeing of subtrees:
2002  *              a) free the subtree starting from *@top
2003  *              b) free the subtrees whose roots are stored in
2004  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
2005  *              c) free the subtrees growing from the inode past the @chain[0].
2006  *                      (no partially truncated stuff there).  */
2007
2008 static Indirect *ext3_find_shared(struct inode *inode, int depth,
2009                         int offsets[4], Indirect chain[4], __le32 *top)
2010 {
2011         Indirect *partial, *p;
2012         int k, err;
2013
2014         *top = 0;
2015         /* Make k index the deepest non-null offest + 1 */
2016         for (k = depth; k > 1 && !offsets[k-1]; k--)
2017                 ;
2018         partial = ext3_get_branch(inode, k, offsets, chain, &err);
2019         /* Writer: pointers */
2020         if (!partial)
2021                 partial = chain + k-1;
2022         /*
2023          * If the branch acquired continuation since we've looked at it -
2024          * fine, it should all survive and (new) top doesn't belong to us.
2025          */
2026         if (!partial->key && *partial->p)
2027                 /* Writer: end */
2028                 goto no_top;
2029         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
2030                 ;
2031         /*
2032          * OK, we've found the last block that must survive. The rest of our
2033          * branch should be detached before unlocking. However, if that rest
2034          * of branch is all ours and does not grow immediately from the inode
2035          * it's easier to cheat and just decrement partial->p.
2036          */
2037         if (p == chain + k - 1 && p > chain) {
2038                 p->p--;
2039         } else {
2040                 *top = *p->p;
2041                 /* Nope, don't do this in ext3.  Must leave the tree intact */
2042 #if 0
2043                 *p->p = 0;
2044 #endif
2045         }
2046         /* Writer: end */
2047
2048         while(partial > p) {
2049                 brelse(partial->bh);
2050                 partial--;
2051         }
2052 no_top:
2053         return partial;
2054 }
2055
2056 /*
2057  * Zero a number of block pointers in either an inode or an indirect block.
2058  * If we restart the transaction we must again get write access to the
2059  * indirect block for further modification.
2060  *
2061  * We release `count' blocks on disk, but (last - first) may be greater
2062  * than `count' because there can be holes in there.
2063  */
2064 static void ext3_clear_blocks(handle_t *handle, struct inode *inode,
2065                 struct buffer_head *bh, ext3_fsblk_t block_to_free,
2066                 unsigned long count, __le32 *first, __le32 *last)
2067 {
2068         __le32 *p;
2069         if (try_to_extend_transaction(handle, inode)) {
2070                 if (bh) {
2071                         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2072                         ext3_journal_dirty_metadata(handle, bh);
2073                 }
2074                 ext3_mark_inode_dirty(handle, inode);
2075                 ext3_journal_test_restart(handle, inode);
2076                 if (bh) {
2077                         BUFFER_TRACE(bh, "retaking write access");
2078                         ext3_journal_get_write_access(handle, bh);
2079                 }
2080         }
2081
2082         /*
2083          * Any buffers which are on the journal will be in memory. We find
2084          * them on the hash table so journal_revoke() will run journal_forget()
2085          * on them.  We've already detached each block from the file, so
2086          * bforget() in journal_forget() should be safe.
2087          *
2088          * AKPM: turn on bforget in journal_forget()!!!
2089          */
2090         for (p = first; p < last; p++) {
2091                 u32 nr = le32_to_cpu(*p);
2092                 if (nr) {
2093                         struct buffer_head *bh;
2094
2095                         *p = 0;
2096                         bh = sb_find_get_block(inode->i_sb, nr);
2097                         ext3_forget(handle, 0, inode, bh, nr);
2098                 }
2099         }
2100
2101         ext3_free_blocks(handle, inode, block_to_free, count);
2102 }
2103
2104 /**
2105  * ext3_free_data - free a list of data blocks
2106  * @handle:     handle for this transaction
2107  * @inode:      inode we are dealing with
2108  * @this_bh:    indirect buffer_head which contains *@first and *@last
2109  * @first:      array of block numbers
2110  * @last:       points immediately past the end of array
2111  *
2112  * We are freeing all blocks refered from that array (numbers are stored as
2113  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
2114  *
2115  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
2116  * blocks are contiguous then releasing them at one time will only affect one
2117  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
2118  * actually use a lot of journal space.
2119  *
2120  * @this_bh will be %NULL if @first and @last point into the inode's direct
2121  * block pointers.
2122  */
2123 static void ext3_free_data(handle_t *handle, struct inode *inode,
2124                            struct buffer_head *this_bh,
2125                            __le32 *first, __le32 *last)
2126 {
2127         ext3_fsblk_t block_to_free = 0;    /* Starting block # of a run */
2128         unsigned long count = 0;            /* Number of blocks in the run */
2129         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
2130                                                corresponding to
2131                                                block_to_free */
2132         ext3_fsblk_t nr;                    /* Current block # */
2133         __le32 *p;                          /* Pointer into inode/ind
2134                                                for current block */
2135         int err;
2136
2137         if (this_bh) {                          /* For indirect block */
2138                 BUFFER_TRACE(this_bh, "get_write_access");
2139                 err = ext3_journal_get_write_access(handle, this_bh);
2140                 /* Important: if we can't update the indirect pointers
2141                  * to the blocks, we can't free them. */
2142                 if (err)
2143                         return;
2144         }
2145
2146         for (p = first; p < last; p++) {
2147                 nr = le32_to_cpu(*p);
2148                 if (nr) {
2149                         /* accumulate blocks to free if they're contiguous */
2150                         if (count == 0) {
2151                                 block_to_free = nr;
2152                                 block_to_free_p = p;
2153                                 count = 1;
2154                         } else if (nr == block_to_free + count) {
2155                                 count++;
2156                         } else {
2157                                 ext3_clear_blocks(handle, inode, this_bh,
2158                                                   block_to_free,
2159                                                   count, block_to_free_p, p);
2160                                 block_to_free = nr;
2161                                 block_to_free_p = p;
2162                                 count = 1;
2163                         }
2164                 }
2165         }
2166
2167         if (count > 0)
2168                 ext3_clear_blocks(handle, inode, this_bh, block_to_free,
2169                                   count, block_to_free_p, p);
2170
2171         if (this_bh) {
2172                 BUFFER_TRACE(this_bh, "call ext3_journal_dirty_metadata");
2173
2174                 /*
2175                  * The buffer head should have an attached journal head at this
2176                  * point. However, if the data is corrupted and an indirect
2177                  * block pointed to itself, it would have been detached when
2178                  * the block was cleared. Check for this instead of OOPSing.
2179                  */
2180                 if (bh2jh(this_bh))
2181                         ext3_journal_dirty_metadata(handle, this_bh);
2182                 else
2183                         ext3_error(inode->i_sb, "ext3_free_data",
2184                                    "circular indirect block detected, "
2185                                    "inode=%lu, block=%llu",
2186                                    inode->i_ino,
2187                                    (unsigned long long)this_bh->b_blocknr);
2188         }
2189 }
2190
2191 /**
2192  *      ext3_free_branches - free an array of branches
2193  *      @handle: JBD handle for this transaction
2194  *      @inode: inode we are dealing with
2195  *      @parent_bh: the buffer_head which contains *@first and *@last
2196  *      @first: array of block numbers
2197  *      @last:  pointer immediately past the end of array
2198  *      @depth: depth of the branches to free
2199  *
2200  *      We are freeing all blocks refered from these branches (numbers are
2201  *      stored as little-endian 32-bit) and updating @inode->i_blocks
2202  *      appropriately.
2203  */
2204 static void ext3_free_branches(handle_t *handle, struct inode *inode,
2205                                struct buffer_head *parent_bh,
2206                                __le32 *first, __le32 *last, int depth)
2207 {
2208         ext3_fsblk_t nr;
2209         __le32 *p;
2210
2211         if (is_handle_aborted(handle))
2212                 return;
2213
2214         if (depth--) {
2215                 struct buffer_head *bh;
2216                 int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2217                 p = last;
2218                 while (--p >= first) {
2219                         nr = le32_to_cpu(*p);
2220                         if (!nr)
2221                                 continue;               /* A hole */
2222
2223                         /* Go read the buffer for the next level down */
2224                         bh = sb_bread(inode->i_sb, nr);
2225
2226                         /*
2227                          * A read failure? Report error and clear slot
2228                          * (should be rare).
2229                          */
2230                         if (!bh) {
2231                                 ext3_error(inode->i_sb, "ext3_free_branches",
2232                                            "Read failure, inode=%lu, block="E3FSBLK,
2233                                            inode->i_ino, nr);
2234                                 continue;
2235                         }
2236
2237                         /* This zaps the entire block.  Bottom up. */
2238                         BUFFER_TRACE(bh, "free child branches");
2239                         ext3_free_branches(handle, inode, bh,
2240                                            (__le32*)bh->b_data,
2241                                            (__le32*)bh->b_data + addr_per_block,
2242                                            depth);
2243
2244                         /*
2245                          * We've probably journalled the indirect block several
2246                          * times during the truncate.  But it's no longer
2247                          * needed and we now drop it from the transaction via
2248                          * journal_revoke().
2249                          *
2250                          * That's easy if it's exclusively part of this
2251                          * transaction.  But if it's part of the committing
2252                          * transaction then journal_forget() will simply
2253                          * brelse() it.  That means that if the underlying
2254                          * block is reallocated in ext3_get_block(),
2255                          * unmap_underlying_metadata() will find this block
2256                          * and will try to get rid of it.  damn, damn.
2257                          *
2258                          * If this block has already been committed to the
2259                          * journal, a revoke record will be written.  And
2260                          * revoke records must be emitted *before* clearing
2261                          * this block's bit in the bitmaps.
2262                          */
2263                         ext3_forget(handle, 1, inode, bh, bh->b_blocknr);
2264
2265                         /*
2266                          * Everything below this this pointer has been
2267                          * released.  Now let this top-of-subtree go.
2268                          *
2269                          * We want the freeing of this indirect block to be
2270                          * atomic in the journal with the updating of the
2271                          * bitmap block which owns it.  So make some room in
2272                          * the journal.
2273                          *
2274                          * We zero the parent pointer *after* freeing its
2275                          * pointee in the bitmaps, so if extend_transaction()
2276                          * for some reason fails to put the bitmap changes and
2277                          * the release into the same transaction, recovery
2278                          * will merely complain about releasing a free block,
2279                          * rather than leaking blocks.
2280                          */
2281                         if (is_handle_aborted(handle))
2282                                 return;
2283                         if (try_to_extend_transaction(handle, inode)) {
2284                                 ext3_mark_inode_dirty(handle, inode);
2285                                 ext3_journal_test_restart(handle, inode);
2286                         }
2287
2288                         ext3_free_blocks(handle, inode, nr, 1);
2289
2290                         if (parent_bh) {
2291                                 /*
2292                                  * The block which we have just freed is
2293                                  * pointed to by an indirect block: journal it
2294                                  */
2295                                 BUFFER_TRACE(parent_bh, "get_write_access");
2296                                 if (!ext3_journal_get_write_access(handle,
2297                                                                    parent_bh)){
2298                                         *p = 0;
2299                                         BUFFER_TRACE(parent_bh,
2300                                         "call ext3_journal_dirty_metadata");
2301                                         ext3_journal_dirty_metadata(handle,
2302                                                                     parent_bh);
2303                                 }
2304                         }
2305                 }
2306         } else {
2307                 /* We have reached the bottom of the tree. */
2308                 BUFFER_TRACE(parent_bh, "free data blocks");
2309                 ext3_free_data(handle, inode, parent_bh, first, last);
2310         }
2311 }
2312
2313 int ext3_can_truncate(struct inode *inode)
2314 {
2315         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
2316                 return 0;
2317         if (S_ISREG(inode->i_mode))
2318                 return 1;
2319         if (S_ISDIR(inode->i_mode))
2320                 return 1;
2321         if (S_ISLNK(inode->i_mode))
2322                 return !ext3_inode_is_fast_symlink(inode);
2323         return 0;
2324 }
2325
2326 /*
2327  * ext3_truncate()
2328  *
2329  * We block out ext3_get_block() block instantiations across the entire
2330  * transaction, and VFS/VM ensures that ext3_truncate() cannot run
2331  * simultaneously on behalf of the same inode.
2332  *
2333  * As we work through the truncate and commmit bits of it to the journal there
2334  * is one core, guiding principle: the file's tree must always be consistent on
2335  * disk.  We must be able to restart the truncate after a crash.
2336  *
2337  * The file's tree may be transiently inconsistent in memory (although it
2338  * probably isn't), but whenever we close off and commit a journal transaction,
2339  * the contents of (the filesystem + the journal) must be consistent and
2340  * restartable.  It's pretty simple, really: bottom up, right to left (although
2341  * left-to-right works OK too).
2342  *
2343  * Note that at recovery time, journal replay occurs *before* the restart of
2344  * truncate against the orphan inode list.
2345  *
2346  * The committed inode has the new, desired i_size (which is the same as
2347  * i_disksize in this case).  After a crash, ext3_orphan_cleanup() will see
2348  * that this inode's truncate did not complete and it will again call
2349  * ext3_truncate() to have another go.  So there will be instantiated blocks
2350  * to the right of the truncation point in a crashed ext3 filesystem.  But
2351  * that's fine - as long as they are linked from the inode, the post-crash
2352  * ext3_truncate() run will find them and release them.
2353  */
2354 void ext3_truncate(struct inode *inode)
2355 {
2356         handle_t *handle;
2357         struct ext3_inode_info *ei = EXT3_I(inode);
2358         __le32 *i_data = ei->i_data;
2359         int addr_per_block = EXT3_ADDR_PER_BLOCK(inode->i_sb);
2360         struct address_space *mapping = inode->i_mapping;
2361         int offsets[4];
2362         Indirect chain[4];
2363         Indirect *partial;
2364         __le32 nr = 0;
2365         int n;
2366         long last_block;
2367         unsigned blocksize = inode->i_sb->s_blocksize;
2368         struct page *page;
2369
2370         if (!ext3_can_truncate(inode))
2371                 goto out_notrans;
2372
2373         if (inode->i_size == 0 && ext3_should_writeback_data(inode))
2374                 ei->i_state |= EXT3_STATE_FLUSH_ON_CLOSE;
2375
2376         /*
2377          * We have to lock the EOF page here, because lock_page() nests
2378          * outside journal_start().
2379          */
2380         if ((inode->i_size & (blocksize - 1)) == 0) {
2381                 /* Block boundary? Nothing to do */
2382                 page = NULL;
2383         } else {
2384                 page = grab_cache_page(mapping,
2385                                 inode->i_size >> PAGE_CACHE_SHIFT);
2386                 if (!page)
2387                         goto out_notrans;
2388         }
2389
2390         handle = start_transaction(inode);
2391         if (IS_ERR(handle)) {
2392                 if (page) {
2393                         clear_highpage(page);
2394                         flush_dcache_page(page);
2395                         unlock_page(page);
2396                         page_cache_release(page);
2397                 }
2398                 goto out_notrans;
2399         }
2400
2401         last_block = (inode->i_size + blocksize-1)
2402                                         >> EXT3_BLOCK_SIZE_BITS(inode->i_sb);
2403
2404         if (page)
2405                 ext3_block_truncate_page(handle, page, mapping, inode->i_size);
2406
2407         n = ext3_block_to_path(inode, last_block, offsets, NULL);
2408         if (n == 0)
2409                 goto out_stop;  /* error */
2410
2411         /*
2412          * OK.  This truncate is going to happen.  We add the inode to the
2413          * orphan list, so that if this truncate spans multiple transactions,
2414          * and we crash, we will resume the truncate when the filesystem
2415          * recovers.  It also marks the inode dirty, to catch the new size.
2416          *
2417          * Implication: the file must always be in a sane, consistent
2418          * truncatable state while each transaction commits.
2419          */
2420         if (ext3_orphan_add(handle, inode))
2421                 goto out_stop;
2422
2423         /*
2424          * The orphan list entry will now protect us from any crash which
2425          * occurs before the truncate completes, so it is now safe to propagate
2426          * the new, shorter inode size (held for now in i_size) into the
2427          * on-disk inode. We do this via i_disksize, which is the value which
2428          * ext3 *really* writes onto the disk inode.
2429          */
2430         ei->i_disksize = inode->i_size;
2431
2432         /*
2433          * From here we block out all ext3_get_block() callers who want to
2434          * modify the block allocation tree.
2435          */
2436         mutex_lock(&ei->truncate_mutex);
2437
2438         if (n == 1) {           /* direct blocks */
2439                 ext3_free_data(handle, inode, NULL, i_data+offsets[0],
2440                                i_data + EXT3_NDIR_BLOCKS);
2441                 goto do_indirects;
2442         }
2443
2444         partial = ext3_find_shared(inode, n, offsets, chain, &nr);
2445         /* Kill the top of shared branch (not detached) */
2446         if (nr) {
2447                 if (partial == chain) {
2448                         /* Shared branch grows from the inode */
2449                         ext3_free_branches(handle, inode, NULL,
2450                                            &nr, &nr+1, (chain+n-1) - partial);
2451                         *partial->p = 0;
2452                         /*
2453                          * We mark the inode dirty prior to restart,
2454                          * and prior to stop.  No need for it here.
2455                          */
2456                 } else {
2457                         /* Shared branch grows from an indirect block */
2458                         BUFFER_TRACE(partial->bh, "get_write_access");
2459                         ext3_free_branches(handle, inode, partial->bh,
2460                                         partial->p,
2461                                         partial->p+1, (chain+n-1) - partial);
2462                 }
2463         }
2464         /* Clear the ends of indirect blocks on the shared branch */
2465         while (partial > chain) {
2466                 ext3_free_branches(handle, inode, partial->bh, partial->p + 1,
2467                                    (__le32*)partial->bh->b_data+addr_per_block,
2468                                    (chain+n-1) - partial);
2469                 BUFFER_TRACE(partial->bh, "call brelse");
2470                 brelse (partial->bh);
2471                 partial--;
2472         }
2473 do_indirects:
2474         /* Kill the remaining (whole) subtrees */
2475         switch (offsets[0]) {
2476         default:
2477                 nr = i_data[EXT3_IND_BLOCK];
2478                 if (nr) {
2479                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
2480                         i_data[EXT3_IND_BLOCK] = 0;
2481                 }
2482         case EXT3_IND_BLOCK:
2483                 nr = i_data[EXT3_DIND_BLOCK];
2484                 if (nr) {
2485                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
2486                         i_data[EXT3_DIND_BLOCK] = 0;
2487                 }
2488         case EXT3_DIND_BLOCK:
2489                 nr = i_data[EXT3_TIND_BLOCK];
2490                 if (nr) {
2491                         ext3_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
2492                         i_data[EXT3_TIND_BLOCK] = 0;
2493                 }
2494         case EXT3_TIND_BLOCK:
2495                 ;
2496         }
2497
2498         ext3_discard_reservation(inode);
2499
2500         mutex_unlock(&ei->truncate_mutex);
2501         inode->i_mtime = inode->i_ctime = CURRENT_TIME_SEC;
2502         ext3_mark_inode_dirty(handle, inode);
2503
2504         /*
2505          * In a multi-transaction truncate, we only make the final transaction
2506          * synchronous
2507          */
2508         if (IS_SYNC(inode))
2509                 handle->h_sync = 1;
2510 out_stop:
2511         /*
2512          * If this was a simple ftruncate(), and the file will remain alive
2513          * then we need to clear up the orphan record which we created above.
2514          * However, if this was a real unlink then we were called by
2515          * ext3_delete_inode(), and we allow that function to clean up the
2516          * orphan info for us.
2517          */
2518         if (inode->i_nlink)
2519                 ext3_orphan_del(handle, inode);
2520
2521         ext3_journal_stop(handle);
2522         return;
2523 out_notrans:
2524         /*
2525          * Delete the inode from orphan list so that it doesn't stay there
2526          * forever and trigger assertion on umount.
2527          */
2528         if (inode->i_nlink)
2529                 ext3_orphan_del(NULL, inode);
2530 }
2531
2532 static ext3_fsblk_t ext3_get_inode_block(struct super_block *sb,
2533                 unsigned long ino, struct ext3_iloc *iloc)
2534 {
2535         unsigned long block_group;
2536         unsigned long offset;
2537         ext3_fsblk_t block;
2538         struct ext3_group_desc *gdp;
2539
2540         if (!ext3_valid_inum(sb, ino)) {
2541                 /*
2542                  * This error is already checked for in namei.c unless we are
2543                  * looking at an NFS filehandle, in which case no error
2544                  * report is needed
2545                  */
2546                 return 0;
2547         }
2548
2549         block_group = (ino - 1) / EXT3_INODES_PER_GROUP(sb);
2550         gdp = ext3_get_group_desc(sb, block_group, NULL);
2551         if (!gdp)
2552                 return 0;
2553         /*
2554          * Figure out the offset within the block group inode table
2555          */
2556         offset = ((ino - 1) % EXT3_INODES_PER_GROUP(sb)) *
2557                 EXT3_INODE_SIZE(sb);
2558         block = le32_to_cpu(gdp->bg_inode_table) +
2559                 (offset >> EXT3_BLOCK_SIZE_BITS(sb));
2560
2561         iloc->block_group = block_group;
2562         iloc->offset = offset & (EXT3_BLOCK_SIZE(sb) - 1);
2563         return block;
2564 }
2565
2566 /*
2567  * ext3_get_inode_loc returns with an extra refcount against the inode's
2568  * underlying buffer_head on success. If 'in_mem' is true, we have all
2569  * data in memory that is needed to recreate the on-disk version of this
2570  * inode.
2571  */
2572 static int __ext3_get_inode_loc(struct inode *inode,
2573                                 struct ext3_iloc *iloc, int in_mem)
2574 {
2575         ext3_fsblk_t block;
2576         struct buffer_head *bh;
2577
2578         block = ext3_get_inode_block(inode->i_sb, inode->i_ino, iloc);
2579         if (!block)
2580                 return -EIO;
2581
2582         bh = sb_getblk(inode->i_sb, block);
2583         if (!bh) {
2584                 ext3_error (inode->i_sb, "ext3_get_inode_loc",
2585                                 "unable to read inode block - "
2586                                 "inode=%lu, block="E3FSBLK,
2587                                  inode->i_ino, block);
2588                 return -EIO;
2589         }
2590         if (!buffer_uptodate(bh)) {
2591                 lock_buffer(bh);
2592
2593                 /*
2594                  * If the buffer has the write error flag, we have failed
2595                  * to write out another inode in the same block.  In this
2596                  * case, we don't have to read the block because we may
2597                  * read the old inode data successfully.
2598                  */
2599                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
2600                         set_buffer_uptodate(bh);
2601
2602                 if (buffer_uptodate(bh)) {
2603                         /* someone brought it uptodate while we waited */
2604                         unlock_buffer(bh);
2605                         goto has_buffer;
2606                 }
2607
2608                 /*
2609                  * If we have all information of the inode in memory and this
2610                  * is the only valid inode in the block, we need not read the
2611                  * block.
2612                  */
2613                 if (in_mem) {
2614                         struct buffer_head *bitmap_bh;
2615                         struct ext3_group_desc *desc;
2616                         int inodes_per_buffer;
2617                         int inode_offset, i;
2618                         int block_group;
2619                         int start;
2620
2621                         block_group = (inode->i_ino - 1) /
2622                                         EXT3_INODES_PER_GROUP(inode->i_sb);
2623                         inodes_per_buffer = bh->b_size /
2624                                 EXT3_INODE_SIZE(inode->i_sb);
2625                         inode_offset = ((inode->i_ino - 1) %
2626                                         EXT3_INODES_PER_GROUP(inode->i_sb));
2627                         start = inode_offset & ~(inodes_per_buffer - 1);
2628
2629                         /* Is the inode bitmap in cache? */
2630                         desc = ext3_get_group_desc(inode->i_sb,
2631                                                 block_group, NULL);
2632                         if (!desc)
2633                                 goto make_io;
2634
2635                         bitmap_bh = sb_getblk(inode->i_sb,
2636                                         le32_to_cpu(desc->bg_inode_bitmap));
2637                         if (!bitmap_bh)
2638                                 goto make_io;
2639
2640                         /*
2641                          * If the inode bitmap isn't in cache then the
2642                          * optimisation may end up performing two reads instead
2643                          * of one, so skip it.
2644                          */
2645                         if (!buffer_uptodate(bitmap_bh)) {
2646                                 brelse(bitmap_bh);
2647                                 goto make_io;
2648                         }
2649                         for (i = start; i < start + inodes_per_buffer; i++) {
2650                                 if (i == inode_offset)
2651                                         continue;
2652                                 if (ext3_test_bit(i, bitmap_bh->b_data))
2653                                         break;
2654                         }
2655                         brelse(bitmap_bh);
2656                         if (i == start + inodes_per_buffer) {
2657                                 /* all other inodes are free, so skip I/O */
2658                                 memset(bh->b_data, 0, bh->b_size);
2659                                 set_buffer_uptodate(bh);
2660                                 unlock_buffer(bh);
2661                                 goto has_buffer;
2662                         }
2663                 }
2664
2665 make_io:
2666                 /*
2667                  * There are other valid inodes in the buffer, this inode
2668                  * has in-inode xattrs, or we don't have this inode in memory.
2669                  * Read the block from disk.
2670                  */
2671                 get_bh(bh);
2672                 bh->b_end_io = end_buffer_read_sync;
2673                 submit_bh(READ_META, bh);
2674                 wait_on_buffer(bh);
2675                 if (!buffer_uptodate(bh)) {
2676                         ext3_error(inode->i_sb, "ext3_get_inode_loc",
2677                                         "unable to read inode block - "
2678                                         "inode=%lu, block="E3FSBLK,
2679                                         inode->i_ino, block);
2680                         brelse(bh);
2681                         return -EIO;
2682                 }
2683         }
2684 has_buffer:
2685         iloc->bh = bh;
2686         return 0;
2687 }
2688
2689 int ext3_get_inode_loc(struct inode *inode, struct ext3_iloc *iloc)
2690 {
2691         /* We have all inode data except xattrs in memory here. */
2692         return __ext3_get_inode_loc(inode, iloc,
2693                 !(EXT3_I(inode)->i_state & EXT3_STATE_XATTR));
2694 }
2695
2696 void ext3_set_inode_flags(struct inode *inode)
2697 {
2698         unsigned int flags = EXT3_I(inode)->i_flags;
2699
2700         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
2701         if (flags & EXT3_SYNC_FL)
2702                 inode->i_flags |= S_SYNC;
2703         if (flags & EXT3_APPEND_FL)
2704                 inode->i_flags |= S_APPEND;
2705         if (flags & EXT3_IMMUTABLE_FL)
2706                 inode->i_flags |= S_IMMUTABLE;
2707         if (flags & EXT3_NOATIME_FL)
2708                 inode->i_flags |= S_NOATIME;
2709         if (flags & EXT3_DIRSYNC_FL)
2710                 inode->i_flags |= S_DIRSYNC;
2711 }
2712
2713 /* Propagate flags from i_flags to EXT3_I(inode)->i_flags */
2714 void ext3_get_inode_flags(struct ext3_inode_info *ei)
2715 {
2716         unsigned int flags = ei->vfs_inode.i_flags;
2717
2718         ei->i_flags &= ~(EXT3_SYNC_FL|EXT3_APPEND_FL|
2719                         EXT3_IMMUTABLE_FL|EXT3_NOATIME_FL|EXT3_DIRSYNC_FL);
2720         if (flags & S_SYNC)
2721                 ei->i_flags |= EXT3_SYNC_FL;
2722         if (flags & S_APPEND)
2723                 ei->i_flags |= EXT3_APPEND_FL;
2724         if (flags & S_IMMUTABLE)
2725                 ei->i_flags |= EXT3_IMMUTABLE_FL;
2726         if (flags & S_NOATIME)
2727                 ei->i_flags |= EXT3_NOATIME_FL;
2728         if (flags & S_DIRSYNC)
2729                 ei->i_flags |= EXT3_DIRSYNC_FL;
2730 }
2731
2732 struct inode *ext3_iget(struct super_block *sb, unsigned long ino)
2733 {
2734         struct ext3_iloc iloc;
2735         struct ext3_inode *raw_inode;
2736         struct ext3_inode_info *ei;
2737         struct buffer_head *bh;
2738         struct inode *inode;
2739         long ret;
2740         int block;
2741
2742         inode = iget_locked(sb, ino);
2743         if (!inode)
2744                 return ERR_PTR(-ENOMEM);
2745         if (!(inode->i_state & I_NEW))
2746                 return inode;
2747
2748         ei = EXT3_I(inode);
2749         ei->i_block_alloc_info = NULL;
2750
2751         ret = __ext3_get_inode_loc(inode, &iloc, 0);
2752         if (ret < 0)
2753                 goto bad_inode;
2754         bh = iloc.bh;
2755         raw_inode = ext3_raw_inode(&iloc);
2756         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
2757         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
2758         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
2759         if(!(test_opt (inode->i_sb, NO_UID32))) {
2760                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
2761                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
2762         }
2763         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
2764         inode->i_size = le32_to_cpu(raw_inode->i_size);
2765         inode->i_atime.tv_sec = (signed)le32_to_cpu(raw_inode->i_atime);
2766         inode->i_ctime.tv_sec = (signed)le32_to_cpu(raw_inode->i_ctime);
2767         inode->i_mtime.tv_sec = (signed)le32_to_cpu(raw_inode->i_mtime);
2768         inode->i_atime.tv_nsec = inode->i_ctime.tv_nsec = inode->i_mtime.tv_nsec = 0;
2769
2770         ei->i_state = 0;
2771         ei->i_dir_start_lookup = 0;
2772         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
2773         /* We now have enough fields to check if the inode was active or not.
2774          * This is needed because nfsd might try to access dead inodes
2775          * the test is that same one that e2fsck uses
2776          * NeilBrown 1999oct15
2777          */
2778         if (inode->i_nlink == 0) {
2779                 if (inode->i_mode == 0 ||
2780                     !(EXT3_SB(inode->i_sb)->s_mount_state & EXT3_ORPHAN_FS)) {
2781                         /* this inode is deleted */
2782                         brelse (bh);
2783                         ret = -ESTALE;
2784                         goto bad_inode;
2785                 }
2786                 /* The only unlinked inodes we let through here have
2787                  * valid i_mode and are being read by the orphan
2788                  * recovery code: that's fine, we're about to complete
2789                  * the process of deleting those. */
2790         }
2791         inode->i_blocks = le32_to_cpu(raw_inode->i_blocks);
2792         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
2793 #ifdef EXT3_FRAGMENTS
2794         ei->i_faddr = le32_to_cpu(raw_inode->i_faddr);
2795         ei->i_frag_no = raw_inode->i_frag;
2796         ei->i_frag_size = raw_inode->i_fsize;
2797 #endif
2798         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl);
2799         if (!S_ISREG(inode->i_mode)) {
2800                 ei->i_dir_acl = le32_to_cpu(raw_inode->i_dir_acl);
2801         } else {
2802                 inode->i_size |=
2803                         ((__u64)le32_to_cpu(raw_inode->i_size_high)) << 32;
2804         }
2805         ei->i_disksize = inode->i_size;
2806         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
2807         ei->i_block_group = iloc.block_group;
2808         /*
2809          * NOTE! The in-memory inode i_data array is in little-endian order
2810          * even on big-endian machines: we do NOT byteswap the block numbers!
2811          */
2812         for (block = 0; block < EXT3_N_BLOCKS; block++)
2813                 ei->i_data[block] = raw_inode->i_block[block];
2814         INIT_LIST_HEAD(&ei->i_orphan);
2815
2816         if (inode->i_ino >= EXT3_FIRST_INO(inode->i_sb) + 1 &&
2817             EXT3_INODE_SIZE(inode->i_sb) > EXT3_GOOD_OLD_INODE_SIZE) {
2818                 /*
2819                  * When mke2fs creates big inodes it does not zero out
2820                  * the unused bytes above EXT3_GOOD_OLD_INODE_SIZE,
2821                  * so ignore those first few inodes.
2822                  */
2823                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
2824                 if (EXT3_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
2825                     EXT3_INODE_SIZE(inode->i_sb)) {
2826                         brelse (bh);
2827                         ret = -EIO;
2828                         goto bad_inode;
2829                 }
2830                 if (ei->i_extra_isize == 0) {
2831                         /* The extra space is currently unused. Use it. */
2832                         ei->i_extra_isize = sizeof(struct ext3_inode) -
2833                                             EXT3_GOOD_OLD_INODE_SIZE;
2834                 } else {
2835                         __le32 *magic = (void *)raw_inode +
2836                                         EXT3_GOOD_OLD_INODE_SIZE +
2837                                         ei->i_extra_isize;
2838                         if (*magic == cpu_to_le32(EXT3_XATTR_MAGIC))
2839                                  ei->i_state |= EXT3_STATE_XATTR;
2840                 }
2841         } else
2842                 ei->i_extra_isize = 0;
2843
2844         if (S_ISREG(inode->i_mode)) {
2845                 inode->i_op = &ext3_file_inode_operations;
2846                 inode->i_fop = &ext3_file_operations;
2847                 ext3_set_aops(inode);
2848         } else if (S_ISDIR(inode->i_mode)) {
2849                 inode->i_op = &ext3_dir_inode_operations;
2850                 inode->i_fop = &ext3_dir_operations;
2851         } else if (S_ISLNK(inode->i_mode)) {
2852                 if (ext3_inode_is_fast_symlink(inode)) {
2853                         inode->i_op = &ext3_fast_symlink_inode_operations;
2854                         nd_terminate_link(ei->i_data, inode->i_size,
2855                                 sizeof(ei->i_data) - 1);
2856                 } else {
2857                         inode->i_op = &ext3_symlink_inode_operations;
2858                         ext3_set_aops(inode);
2859                 }
2860         } else {
2861                 inode->i_op = &ext3_special_inode_operations;
2862                 if (raw_inode->i_block[0])
2863                         init_special_inode(inode, inode->i_mode,
2864                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
2865                 else
2866                         init_special_inode(inode, inode->i_mode,
2867                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
2868         }
2869         brelse (iloc.bh);
2870         ext3_set_inode_flags(inode);
2871         unlock_new_inode(inode);
2872         return inode;
2873
2874 bad_inode:
2875         iget_failed(inode);
2876         return ERR_PTR(ret);
2877 }
2878
2879 /*
2880  * Post the struct inode info into an on-disk inode location in the
2881  * buffer-cache.  This gobbles the caller's reference to the
2882  * buffer_head in the inode location struct.
2883  *
2884  * The caller must have write access to iloc->bh.
2885  */
2886 static int ext3_do_update_inode(handle_t *handle,
2887                                 struct inode *inode,
2888                                 struct ext3_iloc *iloc)
2889 {
2890         struct ext3_inode *raw_inode = ext3_raw_inode(iloc);
2891         struct ext3_inode_info *ei = EXT3_I(inode);
2892         struct buffer_head *bh = iloc->bh;
2893         int err = 0, rc, block;
2894
2895         /* For fields not not tracking in the in-memory inode,
2896          * initialise them to zero for new inodes. */
2897         if (ei->i_state & EXT3_STATE_NEW)
2898                 memset(raw_inode, 0, EXT3_SB(inode->i_sb)->s_inode_size);
2899
2900         ext3_get_inode_flags(ei);
2901         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
2902         if(!(test_opt(inode->i_sb, NO_UID32))) {
2903                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
2904                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
2905 /*
2906  * Fix up interoperability with old kernels. Otherwise, old inodes get
2907  * re-used with the upper 16 bits of the uid/gid intact
2908  */
2909                 if(!ei->i_dtime) {
2910                         raw_inode->i_uid_high =
2911                                 cpu_to_le16(high_16_bits(inode->i_uid));
2912                         raw_inode->i_gid_high =
2913                                 cpu_to_le16(high_16_bits(inode->i_gid));
2914                 } else {
2915                         raw_inode->i_uid_high = 0;
2916                         raw_inode->i_gid_high = 0;
2917                 }
2918         } else {
2919                 raw_inode->i_uid_low =
2920                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
2921                 raw_inode->i_gid_low =
2922                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
2923                 raw_inode->i_uid_high = 0;
2924                 raw_inode->i_gid_high = 0;
2925         }
2926         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
2927         raw_inode->i_size = cpu_to_le32(ei->i_disksize);
2928         raw_inode->i_atime = cpu_to_le32(inode->i_atime.tv_sec);
2929         raw_inode->i_ctime = cpu_to_le32(inode->i_ctime.tv_sec);
2930         raw_inode->i_mtime = cpu_to_le32(inode->i_mtime.tv_sec);
2931         raw_inode->i_blocks = cpu_to_le32(inode->i_blocks);
2932         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
2933         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
2934 #ifdef EXT3_FRAGMENTS
2935         raw_inode->i_faddr = cpu_to_le32(ei->i_faddr);
2936         raw_inode->i_frag = ei->i_frag_no;
2937         raw_inode->i_fsize = ei->i_frag_size;
2938 #endif
2939         raw_inode->i_file_acl = cpu_to_le32(ei->i_file_acl);
2940         if (!S_ISREG(inode->i_mode)) {
2941                 raw_inode->i_dir_acl = cpu_to_le32(ei->i_dir_acl);
2942         } else {
2943                 raw_inode->i_size_high =
2944                         cpu_to_le32(ei->i_disksize >> 32);
2945                 if (ei->i_disksize > 0x7fffffffULL) {
2946                         struct super_block *sb = inode->i_sb;
2947                         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
2948                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE) ||
2949                             EXT3_SB(sb)->s_es->s_rev_level ==
2950                                         cpu_to_le32(EXT3_GOOD_OLD_REV)) {
2951                                /* If this is the first large file
2952                                 * created, add a flag to the superblock.
2953                                 */
2954                                 err = ext3_journal_get_write_access(handle,
2955                                                 EXT3_SB(sb)->s_sbh);
2956                                 if (err)
2957                                         goto out_brelse;
2958                                 ext3_update_dynamic_rev(sb);
2959                                 EXT3_SET_RO_COMPAT_FEATURE(sb,
2960                                         EXT3_FEATURE_RO_COMPAT_LARGE_FILE);
2961                                 handle->h_sync = 1;
2962                                 err = ext3_journal_dirty_metadata(handle,
2963                                                 EXT3_SB(sb)->s_sbh);
2964                         }
2965                 }
2966         }
2967         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
2968         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
2969                 if (old_valid_dev(inode->i_rdev)) {
2970                         raw_inode->i_block[0] =
2971                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
2972                         raw_inode->i_block[1] = 0;
2973                 } else {
2974                         raw_inode->i_block[0] = 0;
2975                         raw_inode->i_block[1] =
2976                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
2977                         raw_inode->i_block[2] = 0;
2978                 }
2979         } else for (block = 0; block < EXT3_N_BLOCKS; block++)
2980                 raw_inode->i_block[block] = ei->i_data[block];
2981
2982         if (ei->i_extra_isize)
2983                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
2984
2985         BUFFER_TRACE(bh, "call ext3_journal_dirty_metadata");
2986         rc = ext3_journal_dirty_metadata(handle, bh);
2987         if (!err)
2988                 err = rc;
2989         ei->i_state &= ~EXT3_STATE_NEW;
2990
2991 out_brelse:
2992         brelse (bh);
2993         ext3_std_error(inode->i_sb, err);
2994         return err;
2995 }
2996
2997 /*
2998  * ext3_write_inode()
2999  *
3000  * We are called from a few places:
3001  *
3002  * - Within generic_file_write() for O_SYNC files.
3003  *   Here, there will be no transaction running. We wait for any running
3004  *   trasnaction to commit.
3005  *
3006  * - Within sys_sync(), kupdate and such.
3007  *   We wait on commit, if tol to.
3008  *
3009  * - Within prune_icache() (PF_MEMALLOC == true)
3010  *   Here we simply return.  We can't afford to block kswapd on the
3011  *   journal commit.
3012  *
3013  * In all cases it is actually safe for us to return without doing anything,
3014  * because the inode has been copied into a raw inode buffer in
3015  * ext3_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3016  * knfsd.
3017  *
3018  * Note that we are absolutely dependent upon all inode dirtiers doing the
3019  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3020  * which we are interested.
3021  *
3022  * It would be a bug for them to not do this.  The code:
3023  *
3024  *      mark_inode_dirty(inode)
3025  *      stuff();
3026  *      inode->i_size = expr;
3027  *
3028  * is in error because a kswapd-driven write_inode() could occur while
3029  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3030  * will no longer be on the superblock's dirty inode list.
3031  */
3032 int ext3_write_inode(struct inode *inode, int wait)
3033 {
3034         if (current->flags & PF_MEMALLOC)
3035                 return 0;
3036
3037         if (ext3_journal_current_handle()) {
3038                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3039                 dump_stack();
3040                 return -EIO;
3041         }
3042
3043         if (!wait)
3044                 return 0;
3045
3046         return ext3_force_commit(inode->i_sb);
3047 }
3048
3049 /*
3050  * ext3_setattr()
3051  *
3052  * Called from notify_change.
3053  *
3054  * We want to trap VFS attempts to truncate the file as soon as
3055  * possible.  In particular, we want to make sure that when the VFS
3056  * shrinks i_size, we put the inode on the orphan list and modify
3057  * i_disksize immediately, so that during the subsequent flushing of
3058  * dirty pages and freeing of disk blocks, we can guarantee that any
3059  * commit will leave the blocks being flushed in an unused state on
3060  * disk.  (On recovery, the inode will get truncated and the blocks will
3061  * be freed, so we have a strong guarantee that no future commit will
3062  * leave these blocks visible to the user.)
3063  *
3064  * Called with inode->sem down.
3065  */
3066 int ext3_setattr(struct dentry *dentry, struct iattr *attr)
3067 {
3068         struct inode *inode = dentry->d_inode;
3069         int error, rc = 0;
3070         const unsigned int ia_valid = attr->ia_valid;
3071
3072         error = inode_change_ok(inode, attr);
3073         if (error)
3074                 return error;
3075
3076         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3077                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3078                 handle_t *handle;
3079
3080                 /* (user+group)*(old+new) structure, inode write (sb,
3081                  * inode block, ? - but truncate inode update has it) */
3082                 handle = ext3_journal_start(inode, 2*(EXT3_QUOTA_INIT_BLOCKS(inode->i_sb)+
3083                                         EXT3_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
3084                 if (IS_ERR(handle)) {
3085                         error = PTR_ERR(handle);
3086                         goto err_out;
3087                 }
3088                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
3089                 if (error) {
3090                         ext3_journal_stop(handle);
3091                         return error;
3092                 }
3093                 /* Update corresponding info in inode so that everything is in
3094                  * one transaction */
3095                 if (attr->ia_valid & ATTR_UID)
3096                         inode->i_uid = attr->ia_uid;
3097                 if (attr->ia_valid & ATTR_GID)
3098                         inode->i_gid = attr->ia_gid;
3099                 error = ext3_mark_inode_dirty(handle, inode);
3100                 ext3_journal_stop(handle);
3101         }
3102
3103         if (S_ISREG(inode->i_mode) &&
3104             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
3105                 handle_t *handle;
3106
3107                 handle = ext3_journal_start(inode, 3);
3108                 if (IS_ERR(handle)) {
3109                         error = PTR_ERR(handle);
3110                         goto err_out;
3111                 }
3112
3113                 error = ext3_orphan_add(handle, inode);
3114                 EXT3_I(inode)->i_disksize = attr->ia_size;
3115                 rc = ext3_mark_inode_dirty(handle, inode);
3116                 if (!error)
3117                         error = rc;
3118                 ext3_journal_stop(handle);
3119         }
3120
3121         rc = inode_setattr(inode, attr);
3122
3123         if (!rc && (ia_valid & ATTR_MODE))
3124                 rc = ext3_acl_chmod(inode);
3125
3126 err_out:
3127         ext3_std_error(inode->i_sb, error);
3128         if (!error)
3129                 error = rc;
3130         return error;
3131 }
3132
3133
3134 /*
3135  * How many blocks doth make a writepage()?
3136  *
3137  * With N blocks per page, it may be:
3138  * N data blocks
3139  * 2 indirect block
3140  * 2 dindirect
3141  * 1 tindirect
3142  * N+5 bitmap blocks (from the above)
3143  * N+5 group descriptor summary blocks
3144  * 1 inode block
3145  * 1 superblock.
3146  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quote files
3147  *
3148  * 3 * (N + 5) + 2 + 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
3149  *
3150  * With ordered or writeback data it's the same, less the N data blocks.
3151  *
3152  * If the inode's direct blocks can hold an integral number of pages then a
3153  * page cannot straddle two indirect blocks, and we can only touch one indirect
3154  * and dindirect block, and the "5" above becomes "3".
3155  *
3156  * This still overestimates under most circumstances.  If we were to pass the
3157  * start and end offsets in here as well we could do block_to_path() on each
3158  * block and work out the exact number of indirects which are touched.  Pah.
3159  */
3160
3161 static int ext3_writepage_trans_blocks(struct inode *inode)
3162 {
3163         int bpp = ext3_journal_blocks_per_page(inode);
3164         int indirects = (EXT3_NDIR_BLOCKS % bpp) ? 5 : 3;
3165         int ret;
3166
3167         if (ext3_should_journal_data(inode))
3168                 ret = 3 * (bpp + indirects) + 2;
3169         else
3170                 ret = 2 * (bpp + indirects) + 2;
3171
3172 #ifdef CONFIG_QUOTA
3173         /* We know that structure was already allocated during vfs_dq_init so
3174          * we will be updating only the data blocks + inodes */
3175         ret += 2*EXT3_QUOTA_TRANS_BLOCKS(inode->i_sb);
3176 #endif
3177
3178         return ret;
3179 }
3180
3181 /*
3182  * The caller must have previously called ext3_reserve_inode_write().
3183  * Give this, we know that the caller already has write access to iloc->bh.
3184  */
3185 int ext3_mark_iloc_dirty(handle_t *handle,
3186                 struct inode *inode, struct ext3_iloc *iloc)
3187 {
3188         int err = 0;
3189
3190         /* the do_update_inode consumes one bh->b_count */
3191         get_bh(iloc->bh);
3192
3193         /* ext3_do_update_inode() does journal_dirty_metadata */
3194         err = ext3_do_update_inode(handle, inode, iloc);
3195         put_bh(iloc->bh);
3196         return err;
3197 }
3198
3199 /*
3200  * On success, We end up with an outstanding reference count against
3201  * iloc->bh.  This _must_ be cleaned up later.
3202  */
3203
3204 int
3205 ext3_reserve_inode_write(handle_t *handle, struct inode *inode,
3206                          struct ext3_iloc *iloc)
3207 {
3208         int err = 0;
3209         if (handle) {
3210                 err = ext3_get_inode_loc(inode, iloc);
3211                 if (!err) {
3212                         BUFFER_TRACE(iloc->bh, "get_write_access");
3213                         err = ext3_journal_get_write_access(handle, iloc->bh);
3214                         if (err) {
3215                                 brelse(iloc->bh);
3216                                 iloc->bh = NULL;
3217                         }
3218                 }
3219         }
3220         ext3_std_error(inode->i_sb, err);
3221         return err;
3222 }
3223
3224 /*
3225  * What we do here is to mark the in-core inode as clean with respect to inode
3226  * dirtiness (it may still be data-dirty).
3227  * This means that the in-core inode may be reaped by prune_icache
3228  * without having to perform any I/O.  This is a very good thing,
3229  * because *any* task may call prune_icache - even ones which
3230  * have a transaction open against a different journal.
3231  *
3232  * Is this cheating?  Not really.  Sure, we haven't written the
3233  * inode out, but prune_icache isn't a user-visible syncing function.
3234  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
3235  * we start and wait on commits.
3236  *
3237  * Is this efficient/effective?  Well, we're being nice to the system
3238  * by cleaning up our inodes proactively so they can be reaped
3239  * without I/O.  But we are potentially leaving up to five seconds'
3240  * worth of inodes floating about which prune_icache wants us to
3241  * write out.  One way to fix that would be to get prune_icache()
3242  * to do a write_super() to free up some memory.  It has the desired
3243  * effect.
3244  */
3245 int ext3_mark_inode_dirty(handle_t *handle, struct inode *inode)
3246 {
3247         struct ext3_iloc iloc;
3248         int err;
3249
3250         might_sleep();
3251         err = ext3_reserve_inode_write(handle, inode, &iloc);
3252         if (!err)
3253                 err = ext3_mark_iloc_dirty(handle, inode, &iloc);
3254         return err;
3255 }
3256
3257 /*
3258  * ext3_dirty_inode() is called from __mark_inode_dirty()
3259  *
3260  * We're really interested in the case where a file is being extended.
3261  * i_size has been changed by generic_commit_write() and we thus need
3262  * to include the updated inode in the current transaction.
3263  *
3264  * Also, vfs_dq_alloc_space() will always dirty the inode when blocks
3265  * are allocated to the file.
3266  *
3267  * If the inode is marked synchronous, we don't honour that here - doing
3268  * so would cause a commit on atime updates, which we don't bother doing.
3269  * We handle synchronous inodes at the highest possible level.
3270  */
3271 void ext3_dirty_inode(struct inode *inode)
3272 {
3273         handle_t *current_handle = ext3_journal_current_handle();
3274         handle_t *handle;
3275
3276         handle = ext3_journal_start(inode, 2);
3277         if (IS_ERR(handle))
3278                 goto out;
3279         if (current_handle &&
3280                 current_handle->h_transaction != handle->h_transaction) {
3281                 /* This task has a transaction open against a different fs */
3282                 printk(KERN_EMERG "%s: transactions do not match!\n",
3283                        __func__);
3284         } else {
3285                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
3286                                 current_handle);
3287                 ext3_mark_inode_dirty(handle, inode);
3288         }
3289         ext3_journal_stop(handle);
3290 out:
3291         return;
3292 }
3293
3294 #if 0
3295 /*
3296  * Bind an inode's backing buffer_head into this transaction, to prevent
3297  * it from being flushed to disk early.  Unlike
3298  * ext3_reserve_inode_write, this leaves behind no bh reference and
3299  * returns no iloc structure, so the caller needs to repeat the iloc
3300  * lookup to mark the inode dirty later.
3301  */
3302 static int ext3_pin_inode(handle_t *handle, struct inode *inode)
3303 {
3304         struct ext3_iloc iloc;
3305
3306         int err = 0;
3307         if (handle) {
3308                 err = ext3_get_inode_loc(inode, &iloc);
3309                 if (!err) {
3310                         BUFFER_TRACE(iloc.bh, "get_write_access");
3311                         err = journal_get_write_access(handle, iloc.bh);
3312                         if (!err)
3313                                 err = ext3_journal_dirty_metadata(handle,
3314                                                                   iloc.bh);
3315                         brelse(iloc.bh);
3316                 }
3317         }
3318         ext3_std_error(inode->i_sb, err);
3319         return err;
3320 }
3321 #endif
3322
3323 int ext3_change_inode_journal_flag(struct inode *inode, int val)
3324 {
3325         journal_t *journal;
3326         handle_t *handle;
3327         int err;
3328
3329         /*
3330          * We have to be very careful here: changing a data block's
3331          * journaling status dynamically is dangerous.  If we write a
3332          * data block to the journal, change the status and then delete
3333          * that block, we risk forgetting to revoke the old log record
3334          * from the journal and so a subsequent replay can corrupt data.
3335          * So, first we make sure that the journal is empty and that
3336          * nobody is changing anything.
3337          */
3338
3339         journal = EXT3_JOURNAL(inode);
3340         if (is_journal_aborted(journal))
3341                 return -EROFS;
3342
3343         journal_lock_updates(journal);
3344         journal_flush(journal);
3345
3346         /*
3347          * OK, there are no updates running now, and all cached data is
3348          * synced to disk.  We are now in a completely consistent state
3349          * which doesn't have anything in the journal, and we know that
3350          * no filesystem updates are running, so it is safe to modify
3351          * the inode's in-core data-journaling state flag now.
3352          */
3353
3354         if (val)
3355                 EXT3_I(inode)->i_flags |= EXT3_JOURNAL_DATA_FL;
3356         else
3357                 EXT3_I(inode)->i_flags &= ~EXT3_JOURNAL_DATA_FL;
3358         ext3_set_aops(inode);
3359
3360         journal_unlock_updates(journal);
3361
3362         /* Finally we can mark the inode as dirty. */
3363
3364         handle = ext3_journal_start(inode, 1);
3365         if (IS_ERR(handle))
3366                 return PTR_ERR(handle);
3367
3368         err = ext3_mark_inode_dirty(handle, inode);
3369         handle->h_sync = 1;
3370         ext3_journal_stop(handle);
3371         ext3_std_error(inode->i_sb, err);
3372
3373         return err;
3374 }