ext4: Fix possible lost inode write in no journal mode
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42
43 #include "ext4_jbd2.h"
44 #include "xattr.h"
45 #include "acl.h"
46 #include "ext4_extents.h"
47
48 #include <trace/events/ext4.h>
49
50 #define MPAGE_DA_EXTENT_TAIL 0x01
51
52 static inline int ext4_begin_ordered_truncate(struct inode *inode,
53                                               loff_t new_size)
54 {
55         return jbd2_journal_begin_ordered_truncate(
56                                         EXT4_SB(inode->i_sb)->s_journal,
57                                         &EXT4_I(inode)->jinode,
58                                         new_size);
59 }
60
61 static void ext4_invalidatepage(struct page *page, unsigned long offset);
62
63 /*
64  * Test whether an inode is a fast symlink.
65  */
66 static int ext4_inode_is_fast_symlink(struct inode *inode)
67 {
68         int ea_blocks = EXT4_I(inode)->i_file_acl ?
69                 (inode->i_sb->s_blocksize >> 9) : 0;
70
71         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
72 }
73
74 /*
75  * Work out how many blocks we need to proceed with the next chunk of a
76  * truncate transaction.
77  */
78 static unsigned long blocks_for_truncate(struct inode *inode)
79 {
80         ext4_lblk_t needed;
81
82         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
83
84         /* Give ourselves just enough room to cope with inodes in which
85          * i_blocks is corrupt: we've seen disk corruptions in the past
86          * which resulted in random data in an inode which looked enough
87          * like a regular file for ext4 to try to delete it.  Things
88          * will go a bit crazy if that happens, but at least we should
89          * try not to panic the whole kernel. */
90         if (needed < 2)
91                 needed = 2;
92
93         /* But we need to bound the transaction so we don't overflow the
94          * journal. */
95         if (needed > EXT4_MAX_TRANS_DATA)
96                 needed = EXT4_MAX_TRANS_DATA;
97
98         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
99 }
100
101 /*
102  * Truncate transactions can be complex and absolutely huge.  So we need to
103  * be able to restart the transaction at a conventient checkpoint to make
104  * sure we don't overflow the journal.
105  *
106  * start_transaction gets us a new handle for a truncate transaction,
107  * and extend_transaction tries to extend the existing one a bit.  If
108  * extend fails, we need to propagate the failure up and restart the
109  * transaction in the top-level truncate loop. --sct
110  */
111 static handle_t *start_transaction(struct inode *inode)
112 {
113         handle_t *result;
114
115         result = ext4_journal_start(inode, blocks_for_truncate(inode));
116         if (!IS_ERR(result))
117                 return result;
118
119         ext4_std_error(inode->i_sb, PTR_ERR(result));
120         return result;
121 }
122
123 /*
124  * Try to extend this transaction for the purposes of truncation.
125  *
126  * Returns 0 if we managed to create more room.  If we can't create more
127  * room, and the transaction must be restarted we return 1.
128  */
129 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
130 {
131         if (!ext4_handle_valid(handle))
132                 return 0;
133         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
134                 return 0;
135         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
136                 return 0;
137         return 1;
138 }
139
140 /*
141  * Restart the transaction associated with *handle.  This does a commit,
142  * so before we call here everything must be consistently dirtied against
143  * this transaction.
144  */
145 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
146                                  int nblocks)
147 {
148         int ret;
149
150         /*
151          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
152          * moment, get_block can be called only for blocks inside i_size since
153          * page cache has been already dropped and writes are blocked by
154          * i_mutex. So we can safely drop the i_data_sem here.
155          */
156         BUG_ON(EXT4_JOURNAL(inode) == NULL);
157         jbd_debug(2, "restarting handle %p\n", handle);
158         up_write(&EXT4_I(inode)->i_data_sem);
159         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
160         down_write(&EXT4_I(inode)->i_data_sem);
161         ext4_discard_preallocations(inode);
162
163         return ret;
164 }
165
166 /*
167  * Called at the last iput() if i_nlink is zero.
168  */
169 void ext4_delete_inode(struct inode *inode)
170 {
171         handle_t *handle;
172         int err;
173
174         if (!is_bad_inode(inode))
175                 dquot_initialize(inode);
176
177         if (ext4_should_order_data(inode))
178                 ext4_begin_ordered_truncate(inode, 0);
179         truncate_inode_pages(&inode->i_data, 0);
180
181         if (is_bad_inode(inode))
182                 goto no_delete;
183
184         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
185         if (IS_ERR(handle)) {
186                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
187                 /*
188                  * If we're going to skip the normal cleanup, we still need to
189                  * make sure that the in-core orphan linked list is properly
190                  * cleaned up.
191                  */
192                 ext4_orphan_del(NULL, inode);
193                 goto no_delete;
194         }
195
196         if (IS_SYNC(inode))
197                 ext4_handle_sync(handle);
198         inode->i_size = 0;
199         err = ext4_mark_inode_dirty(handle, inode);
200         if (err) {
201                 ext4_warning(inode->i_sb,
202                              "couldn't mark inode dirty (err %d)", err);
203                 goto stop_handle;
204         }
205         if (inode->i_blocks)
206                 ext4_truncate(inode);
207
208         /*
209          * ext4_ext_truncate() doesn't reserve any slop when it
210          * restarts journal transactions; therefore there may not be
211          * enough credits left in the handle to remove the inode from
212          * the orphan list and set the dtime field.
213          */
214         if (!ext4_handle_has_enough_credits(handle, 3)) {
215                 err = ext4_journal_extend(handle, 3);
216                 if (err > 0)
217                         err = ext4_journal_restart(handle, 3);
218                 if (err != 0) {
219                         ext4_warning(inode->i_sb,
220                                      "couldn't extend journal (err %d)", err);
221                 stop_handle:
222                         ext4_journal_stop(handle);
223                         goto no_delete;
224                 }
225         }
226
227         /*
228          * Kill off the orphan record which ext4_truncate created.
229          * AKPM: I think this can be inside the above `if'.
230          * Note that ext4_orphan_del() has to be able to cope with the
231          * deletion of a non-existent orphan - this is because we don't
232          * know if ext4_truncate() actually created an orphan record.
233          * (Well, we could do this if we need to, but heck - it works)
234          */
235         ext4_orphan_del(handle, inode);
236         EXT4_I(inode)->i_dtime  = get_seconds();
237
238         /*
239          * One subtle ordering requirement: if anything has gone wrong
240          * (transaction abort, IO errors, whatever), then we can still
241          * do these next steps (the fs will already have been marked as
242          * having errors), but we can't free the inode if the mark_dirty
243          * fails.
244          */
245         if (ext4_mark_inode_dirty(handle, inode))
246                 /* If that failed, just do the required in-core inode clear. */
247                 clear_inode(inode);
248         else
249                 ext4_free_inode(handle, inode);
250         ext4_journal_stop(handle);
251         return;
252 no_delete:
253         clear_inode(inode);     /* We must guarantee clearing of inode... */
254 }
255
256 typedef struct {
257         __le32  *p;
258         __le32  key;
259         struct buffer_head *bh;
260 } Indirect;
261
262 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
263 {
264         p->key = *(p->p = v);
265         p->bh = bh;
266 }
267
268 /**
269  *      ext4_block_to_path - parse the block number into array of offsets
270  *      @inode: inode in question (we are only interested in its superblock)
271  *      @i_block: block number to be parsed
272  *      @offsets: array to store the offsets in
273  *      @boundary: set this non-zero if the referred-to block is likely to be
274  *             followed (on disk) by an indirect block.
275  *
276  *      To store the locations of file's data ext4 uses a data structure common
277  *      for UNIX filesystems - tree of pointers anchored in the inode, with
278  *      data blocks at leaves and indirect blocks in intermediate nodes.
279  *      This function translates the block number into path in that tree -
280  *      return value is the path length and @offsets[n] is the offset of
281  *      pointer to (n+1)th node in the nth one. If @block is out of range
282  *      (negative or too large) warning is printed and zero returned.
283  *
284  *      Note: function doesn't find node addresses, so no IO is needed. All
285  *      we need to know is the capacity of indirect blocks (taken from the
286  *      inode->i_sb).
287  */
288
289 /*
290  * Portability note: the last comparison (check that we fit into triple
291  * indirect block) is spelled differently, because otherwise on an
292  * architecture with 32-bit longs and 8Kb pages we might get into trouble
293  * if our filesystem had 8Kb blocks. We might use long long, but that would
294  * kill us on x86. Oh, well, at least the sign propagation does not matter -
295  * i_block would have to be negative in the very beginning, so we would not
296  * get there at all.
297  */
298
299 static int ext4_block_to_path(struct inode *inode,
300                               ext4_lblk_t i_block,
301                               ext4_lblk_t offsets[4], int *boundary)
302 {
303         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
304         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
305         const long direct_blocks = EXT4_NDIR_BLOCKS,
306                 indirect_blocks = ptrs,
307                 double_blocks = (1 << (ptrs_bits * 2));
308         int n = 0;
309         int final = 0;
310
311         if (i_block < direct_blocks) {
312                 offsets[n++] = i_block;
313                 final = direct_blocks;
314         } else if ((i_block -= direct_blocks) < indirect_blocks) {
315                 offsets[n++] = EXT4_IND_BLOCK;
316                 offsets[n++] = i_block;
317                 final = ptrs;
318         } else if ((i_block -= indirect_blocks) < double_blocks) {
319                 offsets[n++] = EXT4_DIND_BLOCK;
320                 offsets[n++] = i_block >> ptrs_bits;
321                 offsets[n++] = i_block & (ptrs - 1);
322                 final = ptrs;
323         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
324                 offsets[n++] = EXT4_TIND_BLOCK;
325                 offsets[n++] = i_block >> (ptrs_bits * 2);
326                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
327                 offsets[n++] = i_block & (ptrs - 1);
328                 final = ptrs;
329         } else {
330                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
331                              i_block + direct_blocks +
332                              indirect_blocks + double_blocks, inode->i_ino);
333         }
334         if (boundary)
335                 *boundary = final - 1 - (i_block & (ptrs - 1));
336         return n;
337 }
338
339 static int __ext4_check_blockref(const char *function, struct inode *inode,
340                                  __le32 *p, unsigned int max)
341 {
342         __le32 *bref = p;
343         unsigned int blk;
344
345         while (bref < p+max) {
346                 blk = le32_to_cpu(*bref++);
347                 if (blk &&
348                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
349                                                     blk, 1))) {
350                         __ext4_error(inode->i_sb, function,
351                                    "invalid block reference %u "
352                                    "in inode #%lu", blk, inode->i_ino);
353                         return -EIO;
354                 }
355         }
356         return 0;
357 }
358
359
360 #define ext4_check_indirect_blockref(inode, bh)                         \
361         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
362                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
363
364 #define ext4_check_inode_blockref(inode)                                \
365         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
366                               EXT4_NDIR_BLOCKS)
367
368 /**
369  *      ext4_get_branch - read the chain of indirect blocks leading to data
370  *      @inode: inode in question
371  *      @depth: depth of the chain (1 - direct pointer, etc.)
372  *      @offsets: offsets of pointers in inode/indirect blocks
373  *      @chain: place to store the result
374  *      @err: here we store the error value
375  *
376  *      Function fills the array of triples <key, p, bh> and returns %NULL
377  *      if everything went OK or the pointer to the last filled triple
378  *      (incomplete one) otherwise. Upon the return chain[i].key contains
379  *      the number of (i+1)-th block in the chain (as it is stored in memory,
380  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
381  *      number (it points into struct inode for i==0 and into the bh->b_data
382  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
383  *      block for i>0 and NULL for i==0. In other words, it holds the block
384  *      numbers of the chain, addresses they were taken from (and where we can
385  *      verify that chain did not change) and buffer_heads hosting these
386  *      numbers.
387  *
388  *      Function stops when it stumbles upon zero pointer (absent block)
389  *              (pointer to last triple returned, *@err == 0)
390  *      or when it gets an IO error reading an indirect block
391  *              (ditto, *@err == -EIO)
392  *      or when it reads all @depth-1 indirect blocks successfully and finds
393  *      the whole chain, all way to the data (returns %NULL, *err == 0).
394  *
395  *      Need to be called with
396  *      down_read(&EXT4_I(inode)->i_data_sem)
397  */
398 static Indirect *ext4_get_branch(struct inode *inode, int depth,
399                                  ext4_lblk_t  *offsets,
400                                  Indirect chain[4], int *err)
401 {
402         struct super_block *sb = inode->i_sb;
403         Indirect *p = chain;
404         struct buffer_head *bh;
405
406         *err = 0;
407         /* i_data is not going away, no lock needed */
408         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
409         if (!p->key)
410                 goto no_block;
411         while (--depth) {
412                 bh = sb_getblk(sb, le32_to_cpu(p->key));
413                 if (unlikely(!bh))
414                         goto failure;
415
416                 if (!bh_uptodate_or_lock(bh)) {
417                         if (bh_submit_read(bh) < 0) {
418                                 put_bh(bh);
419                                 goto failure;
420                         }
421                         /* validate block references */
422                         if (ext4_check_indirect_blockref(inode, bh)) {
423                                 put_bh(bh);
424                                 goto failure;
425                         }
426                 }
427
428                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
429                 /* Reader: end */
430                 if (!p->key)
431                         goto no_block;
432         }
433         return NULL;
434
435 failure:
436         *err = -EIO;
437 no_block:
438         return p;
439 }
440
441 /**
442  *      ext4_find_near - find a place for allocation with sufficient locality
443  *      @inode: owner
444  *      @ind: descriptor of indirect block.
445  *
446  *      This function returns the preferred place for block allocation.
447  *      It is used when heuristic for sequential allocation fails.
448  *      Rules are:
449  *        + if there is a block to the left of our position - allocate near it.
450  *        + if pointer will live in indirect block - allocate near that block.
451  *        + if pointer will live in inode - allocate in the same
452  *          cylinder group.
453  *
454  * In the latter case we colour the starting block by the callers PID to
455  * prevent it from clashing with concurrent allocations for a different inode
456  * in the same block group.   The PID is used here so that functionally related
457  * files will be close-by on-disk.
458  *
459  *      Caller must make sure that @ind is valid and will stay that way.
460  */
461 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
462 {
463         struct ext4_inode_info *ei = EXT4_I(inode);
464         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
465         __le32 *p;
466         ext4_fsblk_t bg_start;
467         ext4_fsblk_t last_block;
468         ext4_grpblk_t colour;
469         ext4_group_t block_group;
470         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
471
472         /* Try to find previous block */
473         for (p = ind->p - 1; p >= start; p--) {
474                 if (*p)
475                         return le32_to_cpu(*p);
476         }
477
478         /* No such thing, so let's try location of indirect block */
479         if (ind->bh)
480                 return ind->bh->b_blocknr;
481
482         /*
483          * It is going to be referred to from the inode itself? OK, just put it
484          * into the same cylinder group then.
485          */
486         block_group = ei->i_block_group;
487         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
488                 block_group &= ~(flex_size-1);
489                 if (S_ISREG(inode->i_mode))
490                         block_group++;
491         }
492         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
493         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
494
495         /*
496          * If we are doing delayed allocation, we don't need take
497          * colour into account.
498          */
499         if (test_opt(inode->i_sb, DELALLOC))
500                 return bg_start;
501
502         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
503                 colour = (current->pid % 16) *
504                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
505         else
506                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
507         return bg_start + colour;
508 }
509
510 /**
511  *      ext4_find_goal - find a preferred place for allocation.
512  *      @inode: owner
513  *      @block:  block we want
514  *      @partial: pointer to the last triple within a chain
515  *
516  *      Normally this function find the preferred place for block allocation,
517  *      returns it.
518  *      Because this is only used for non-extent files, we limit the block nr
519  *      to 32 bits.
520  */
521 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
522                                    Indirect *partial)
523 {
524         ext4_fsblk_t goal;
525
526         /*
527          * XXX need to get goal block from mballoc's data structures
528          */
529
530         goal = ext4_find_near(inode, partial);
531         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
532         return goal;
533 }
534
535 /**
536  *      ext4_blks_to_allocate: Look up the block map and count the number
537  *      of direct blocks need to be allocated for the given branch.
538  *
539  *      @branch: chain of indirect blocks
540  *      @k: number of blocks need for indirect blocks
541  *      @blks: number of data blocks to be mapped.
542  *      @blocks_to_boundary:  the offset in the indirect block
543  *
544  *      return the total number of blocks to be allocate, including the
545  *      direct and indirect blocks.
546  */
547 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
548                                  int blocks_to_boundary)
549 {
550         unsigned int count = 0;
551
552         /*
553          * Simple case, [t,d]Indirect block(s) has not allocated yet
554          * then it's clear blocks on that path have not allocated
555          */
556         if (k > 0) {
557                 /* right now we don't handle cross boundary allocation */
558                 if (blks < blocks_to_boundary + 1)
559                         count += blks;
560                 else
561                         count += blocks_to_boundary + 1;
562                 return count;
563         }
564
565         count++;
566         while (count < blks && count <= blocks_to_boundary &&
567                 le32_to_cpu(*(branch[0].p + count)) == 0) {
568                 count++;
569         }
570         return count;
571 }
572
573 /**
574  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
575  *      @indirect_blks: the number of blocks need to allocate for indirect
576  *                      blocks
577  *
578  *      @new_blocks: on return it will store the new block numbers for
579  *      the indirect blocks(if needed) and the first direct block,
580  *      @blks:  on return it will store the total number of allocated
581  *              direct blocks
582  */
583 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
584                              ext4_lblk_t iblock, ext4_fsblk_t goal,
585                              int indirect_blks, int blks,
586                              ext4_fsblk_t new_blocks[4], int *err)
587 {
588         struct ext4_allocation_request ar;
589         int target, i;
590         unsigned long count = 0, blk_allocated = 0;
591         int index = 0;
592         ext4_fsblk_t current_block = 0;
593         int ret = 0;
594
595         /*
596          * Here we try to allocate the requested multiple blocks at once,
597          * on a best-effort basis.
598          * To build a branch, we should allocate blocks for
599          * the indirect blocks(if not allocated yet), and at least
600          * the first direct block of this branch.  That's the
601          * minimum number of blocks need to allocate(required)
602          */
603         /* first we try to allocate the indirect blocks */
604         target = indirect_blks;
605         while (target > 0) {
606                 count = target;
607                 /* allocating blocks for indirect blocks and direct blocks */
608                 current_block = ext4_new_meta_blocks(handle, inode,
609                                                         goal, &count, err);
610                 if (*err)
611                         goto failed_out;
612
613                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
614                         EXT4_ERROR_INODE(inode,
615                                          "current_block %llu + count %lu > %d!",
616                                          current_block, count,
617                                          EXT4_MAX_BLOCK_FILE_PHYS);
618                         *err = -EIO;
619                         goto failed_out;
620                 }
621
622                 target -= count;
623                 /* allocate blocks for indirect blocks */
624                 while (index < indirect_blks && count) {
625                         new_blocks[index++] = current_block++;
626                         count--;
627                 }
628                 if (count > 0) {
629                         /*
630                          * save the new block number
631                          * for the first direct block
632                          */
633                         new_blocks[index] = current_block;
634                         printk(KERN_INFO "%s returned more blocks than "
635                                                 "requested\n", __func__);
636                         WARN_ON(1);
637                         break;
638                 }
639         }
640
641         target = blks - count ;
642         blk_allocated = count;
643         if (!target)
644                 goto allocated;
645         /* Now allocate data blocks */
646         memset(&ar, 0, sizeof(ar));
647         ar.inode = inode;
648         ar.goal = goal;
649         ar.len = target;
650         ar.logical = iblock;
651         if (S_ISREG(inode->i_mode))
652                 /* enable in-core preallocation only for regular files */
653                 ar.flags = EXT4_MB_HINT_DATA;
654
655         current_block = ext4_mb_new_blocks(handle, &ar, err);
656         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
657                 EXT4_ERROR_INODE(inode,
658                                  "current_block %llu + ar.len %d > %d!",
659                                  current_block, ar.len,
660                                  EXT4_MAX_BLOCK_FILE_PHYS);
661                 *err = -EIO;
662                 goto failed_out;
663         }
664
665         if (*err && (target == blks)) {
666                 /*
667                  * if the allocation failed and we didn't allocate
668                  * any blocks before
669                  */
670                 goto failed_out;
671         }
672         if (!*err) {
673                 if (target == blks) {
674                         /*
675                          * save the new block number
676                          * for the first direct block
677                          */
678                         new_blocks[index] = current_block;
679                 }
680                 blk_allocated += ar.len;
681         }
682 allocated:
683         /* total number of blocks allocated for direct blocks */
684         ret = blk_allocated;
685         *err = 0;
686         return ret;
687 failed_out:
688         for (i = 0; i < index; i++)
689                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
690         return ret;
691 }
692
693 /**
694  *      ext4_alloc_branch - allocate and set up a chain of blocks.
695  *      @inode: owner
696  *      @indirect_blks: number of allocated indirect blocks
697  *      @blks: number of allocated direct blocks
698  *      @offsets: offsets (in the blocks) to store the pointers to next.
699  *      @branch: place to store the chain in.
700  *
701  *      This function allocates blocks, zeroes out all but the last one,
702  *      links them into chain and (if we are synchronous) writes them to disk.
703  *      In other words, it prepares a branch that can be spliced onto the
704  *      inode. It stores the information about that chain in the branch[], in
705  *      the same format as ext4_get_branch() would do. We are calling it after
706  *      we had read the existing part of chain and partial points to the last
707  *      triple of that (one with zero ->key). Upon the exit we have the same
708  *      picture as after the successful ext4_get_block(), except that in one
709  *      place chain is disconnected - *branch->p is still zero (we did not
710  *      set the last link), but branch->key contains the number that should
711  *      be placed into *branch->p to fill that gap.
712  *
713  *      If allocation fails we free all blocks we've allocated (and forget
714  *      their buffer_heads) and return the error value the from failed
715  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
716  *      as described above and return 0.
717  */
718 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
719                              ext4_lblk_t iblock, int indirect_blks,
720                              int *blks, ext4_fsblk_t goal,
721                              ext4_lblk_t *offsets, Indirect *branch)
722 {
723         int blocksize = inode->i_sb->s_blocksize;
724         int i, n = 0;
725         int err = 0;
726         struct buffer_head *bh;
727         int num;
728         ext4_fsblk_t new_blocks[4];
729         ext4_fsblk_t current_block;
730
731         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
732                                 *blks, new_blocks, &err);
733         if (err)
734                 return err;
735
736         branch[0].key = cpu_to_le32(new_blocks[0]);
737         /*
738          * metadata blocks and data blocks are allocated.
739          */
740         for (n = 1; n <= indirect_blks;  n++) {
741                 /*
742                  * Get buffer_head for parent block, zero it out
743                  * and set the pointer to new one, then send
744                  * parent to disk.
745                  */
746                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
747                 branch[n].bh = bh;
748                 lock_buffer(bh);
749                 BUFFER_TRACE(bh, "call get_create_access");
750                 err = ext4_journal_get_create_access(handle, bh);
751                 if (err) {
752                         /* Don't brelse(bh) here; it's done in
753                          * ext4_journal_forget() below */
754                         unlock_buffer(bh);
755                         goto failed;
756                 }
757
758                 memset(bh->b_data, 0, blocksize);
759                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
760                 branch[n].key = cpu_to_le32(new_blocks[n]);
761                 *branch[n].p = branch[n].key;
762                 if (n == indirect_blks) {
763                         current_block = new_blocks[n];
764                         /*
765                          * End of chain, update the last new metablock of
766                          * the chain to point to the new allocated
767                          * data blocks numbers
768                          */
769                         for (i = 1; i < num; i++)
770                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
771                 }
772                 BUFFER_TRACE(bh, "marking uptodate");
773                 set_buffer_uptodate(bh);
774                 unlock_buffer(bh);
775
776                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
777                 err = ext4_handle_dirty_metadata(handle, inode, bh);
778                 if (err)
779                         goto failed;
780         }
781         *blks = num;
782         return err;
783 failed:
784         /* Allocation failed, free what we already allocated */
785         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
786         for (i = 1; i <= n ; i++) {
787                 /* 
788                  * branch[i].bh is newly allocated, so there is no
789                  * need to revoke the block, which is why we don't
790                  * need to set EXT4_FREE_BLOCKS_METADATA.
791                  */
792                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
793                                  EXT4_FREE_BLOCKS_FORGET);
794         }
795         for (i = n+1; i < indirect_blks; i++)
796                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
797
798         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
799
800         return err;
801 }
802
803 /**
804  * ext4_splice_branch - splice the allocated branch onto inode.
805  * @inode: owner
806  * @block: (logical) number of block we are adding
807  * @chain: chain of indirect blocks (with a missing link - see
808  *      ext4_alloc_branch)
809  * @where: location of missing link
810  * @num:   number of indirect blocks we are adding
811  * @blks:  number of direct blocks we are adding
812  *
813  * This function fills the missing link and does all housekeeping needed in
814  * inode (->i_blocks, etc.). In case of success we end up with the full
815  * chain to new block and return 0.
816  */
817 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
818                               ext4_lblk_t block, Indirect *where, int num,
819                               int blks)
820 {
821         int i;
822         int err = 0;
823         ext4_fsblk_t current_block;
824
825         /*
826          * If we're splicing into a [td]indirect block (as opposed to the
827          * inode) then we need to get write access to the [td]indirect block
828          * before the splice.
829          */
830         if (where->bh) {
831                 BUFFER_TRACE(where->bh, "get_write_access");
832                 err = ext4_journal_get_write_access(handle, where->bh);
833                 if (err)
834                         goto err_out;
835         }
836         /* That's it */
837
838         *where->p = where->key;
839
840         /*
841          * Update the host buffer_head or inode to point to more just allocated
842          * direct blocks blocks
843          */
844         if (num == 0 && blks > 1) {
845                 current_block = le32_to_cpu(where->key) + 1;
846                 for (i = 1; i < blks; i++)
847                         *(where->p + i) = cpu_to_le32(current_block++);
848         }
849
850         /* We are done with atomic stuff, now do the rest of housekeeping */
851         /* had we spliced it onto indirect block? */
852         if (where->bh) {
853                 /*
854                  * If we spliced it onto an indirect block, we haven't
855                  * altered the inode.  Note however that if it is being spliced
856                  * onto an indirect block at the very end of the file (the
857                  * file is growing) then we *will* alter the inode to reflect
858                  * the new i_size.  But that is not done here - it is done in
859                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
860                  */
861                 jbd_debug(5, "splicing indirect only\n");
862                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
863                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
864                 if (err)
865                         goto err_out;
866         } else {
867                 /*
868                  * OK, we spliced it into the inode itself on a direct block.
869                  */
870                 ext4_mark_inode_dirty(handle, inode);
871                 jbd_debug(5, "splicing direct\n");
872         }
873         return err;
874
875 err_out:
876         for (i = 1; i <= num; i++) {
877                 /* 
878                  * branch[i].bh is newly allocated, so there is no
879                  * need to revoke the block, which is why we don't
880                  * need to set EXT4_FREE_BLOCKS_METADATA.
881                  */
882                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
883                                  EXT4_FREE_BLOCKS_FORGET);
884         }
885         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
886                          blks, 0);
887
888         return err;
889 }
890
891 /*
892  * The ext4_ind_get_blocks() function handles non-extents inodes
893  * (i.e., using the traditional indirect/double-indirect i_blocks
894  * scheme) for ext4_get_blocks().
895  *
896  * Allocation strategy is simple: if we have to allocate something, we will
897  * have to go the whole way to leaf. So let's do it before attaching anything
898  * to tree, set linkage between the newborn blocks, write them if sync is
899  * required, recheck the path, free and repeat if check fails, otherwise
900  * set the last missing link (that will protect us from any truncate-generated
901  * removals - all blocks on the path are immune now) and possibly force the
902  * write on the parent block.
903  * That has a nice additional property: no special recovery from the failed
904  * allocations is needed - we simply release blocks and do not touch anything
905  * reachable from inode.
906  *
907  * `handle' can be NULL if create == 0.
908  *
909  * return > 0, # of blocks mapped or allocated.
910  * return = 0, if plain lookup failed.
911  * return < 0, error case.
912  *
913  * The ext4_ind_get_blocks() function should be called with
914  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
915  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
916  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
917  * blocks.
918  */
919 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
920                                ext4_lblk_t iblock, unsigned int maxblocks,
921                                struct buffer_head *bh_result,
922                                int flags)
923 {
924         int err = -EIO;
925         ext4_lblk_t offsets[4];
926         Indirect chain[4];
927         Indirect *partial;
928         ext4_fsblk_t goal;
929         int indirect_blks;
930         int blocks_to_boundary = 0;
931         int depth;
932         int count = 0;
933         ext4_fsblk_t first_block = 0;
934
935         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
936         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
937         depth = ext4_block_to_path(inode, iblock, offsets,
938                                    &blocks_to_boundary);
939
940         if (depth == 0)
941                 goto out;
942
943         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
944
945         /* Simplest case - block found, no allocation needed */
946         if (!partial) {
947                 first_block = le32_to_cpu(chain[depth - 1].key);
948                 clear_buffer_new(bh_result);
949                 count++;
950                 /*map more blocks*/
951                 while (count < maxblocks && count <= blocks_to_boundary) {
952                         ext4_fsblk_t blk;
953
954                         blk = le32_to_cpu(*(chain[depth-1].p + count));
955
956                         if (blk == first_block + count)
957                                 count++;
958                         else
959                                 break;
960                 }
961                 goto got_it;
962         }
963
964         /* Next simple case - plain lookup or failed read of indirect block */
965         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
966                 goto cleanup;
967
968         /*
969          * Okay, we need to do block allocation.
970         */
971         goal = ext4_find_goal(inode, iblock, partial);
972
973         /* the number of blocks need to allocate for [d,t]indirect blocks */
974         indirect_blks = (chain + depth) - partial - 1;
975
976         /*
977          * Next look up the indirect map to count the totoal number of
978          * direct blocks to allocate for this branch.
979          */
980         count = ext4_blks_to_allocate(partial, indirect_blks,
981                                         maxblocks, blocks_to_boundary);
982         /*
983          * Block out ext4_truncate while we alter the tree
984          */
985         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
986                                 &count, goal,
987                                 offsets + (partial - chain), partial);
988
989         /*
990          * The ext4_splice_branch call will free and forget any buffers
991          * on the new chain if there is a failure, but that risks using
992          * up transaction credits, especially for bitmaps where the
993          * credits cannot be returned.  Can we handle this somehow?  We
994          * may need to return -EAGAIN upwards in the worst case.  --sct
995          */
996         if (!err)
997                 err = ext4_splice_branch(handle, inode, iblock,
998                                          partial, indirect_blks, count);
999         if (err)
1000                 goto cleanup;
1001
1002         set_buffer_new(bh_result);
1003
1004         ext4_update_inode_fsync_trans(handle, inode, 1);
1005 got_it:
1006         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1007         if (count > blocks_to_boundary)
1008                 set_buffer_boundary(bh_result);
1009         err = count;
1010         /* Clean up and exit */
1011         partial = chain + depth - 1;    /* the whole chain */
1012 cleanup:
1013         while (partial > chain) {
1014                 BUFFER_TRACE(partial->bh, "call brelse");
1015                 brelse(partial->bh);
1016                 partial--;
1017         }
1018         BUFFER_TRACE(bh_result, "returned");
1019 out:
1020         return err;
1021 }
1022
1023 #ifdef CONFIG_QUOTA
1024 qsize_t *ext4_get_reserved_space(struct inode *inode)
1025 {
1026         return &EXT4_I(inode)->i_reserved_quota;
1027 }
1028 #endif
1029
1030 /*
1031  * Calculate the number of metadata blocks need to reserve
1032  * to allocate a new block at @lblocks for non extent file based file
1033  */
1034 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1035                                               sector_t lblock)
1036 {
1037         struct ext4_inode_info *ei = EXT4_I(inode);
1038         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1039         int blk_bits;
1040
1041         if (lblock < EXT4_NDIR_BLOCKS)
1042                 return 0;
1043
1044         lblock -= EXT4_NDIR_BLOCKS;
1045
1046         if (ei->i_da_metadata_calc_len &&
1047             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1048                 ei->i_da_metadata_calc_len++;
1049                 return 0;
1050         }
1051         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1052         ei->i_da_metadata_calc_len = 1;
1053         blk_bits = order_base_2(lblock);
1054         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1055 }
1056
1057 /*
1058  * Calculate the number of metadata blocks need to reserve
1059  * to allocate a block located at @lblock
1060  */
1061 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1062 {
1063         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1064                 return ext4_ext_calc_metadata_amount(inode, lblock);
1065
1066         return ext4_indirect_calc_metadata_amount(inode, lblock);
1067 }
1068
1069 /*
1070  * Called with i_data_sem down, which is important since we can call
1071  * ext4_discard_preallocations() from here.
1072  */
1073 void ext4_da_update_reserve_space(struct inode *inode,
1074                                         int used, int quota_claim)
1075 {
1076         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1077         struct ext4_inode_info *ei = EXT4_I(inode);
1078         int mdb_free = 0, allocated_meta_blocks = 0;
1079
1080         spin_lock(&ei->i_block_reservation_lock);
1081         trace_ext4_da_update_reserve_space(inode, used);
1082         if (unlikely(used > ei->i_reserved_data_blocks)) {
1083                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1084                          "with only %d reserved data blocks\n",
1085                          __func__, inode->i_ino, used,
1086                          ei->i_reserved_data_blocks);
1087                 WARN_ON(1);
1088                 used = ei->i_reserved_data_blocks;
1089         }
1090
1091         /* Update per-inode reservations */
1092         ei->i_reserved_data_blocks -= used;
1093         used += ei->i_allocated_meta_blocks;
1094         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1095         allocated_meta_blocks = ei->i_allocated_meta_blocks;
1096         ei->i_allocated_meta_blocks = 0;
1097         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
1098
1099         if (ei->i_reserved_data_blocks == 0) {
1100                 /*
1101                  * We can release all of the reserved metadata blocks
1102                  * only when we have written all of the delayed
1103                  * allocation blocks.
1104                  */
1105                 mdb_free = ei->i_reserved_meta_blocks;
1106                 ei->i_reserved_meta_blocks = 0;
1107                 ei->i_da_metadata_calc_len = 0;
1108                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1109         }
1110         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1111
1112         /* Update quota subsystem */
1113         if (quota_claim) {
1114                 dquot_claim_block(inode, used);
1115                 if (mdb_free)
1116                         dquot_release_reservation_block(inode, mdb_free);
1117         } else {
1118                 /*
1119                  * We did fallocate with an offset that is already delayed
1120                  * allocated. So on delayed allocated writeback we should
1121                  * not update the quota for allocated blocks. But then
1122                  * converting an fallocate region to initialized region would
1123                  * have caused a metadata allocation. So claim quota for
1124                  * that
1125                  */
1126                 if (allocated_meta_blocks)
1127                         dquot_claim_block(inode, allocated_meta_blocks);
1128                 dquot_release_reservation_block(inode, mdb_free + used);
1129         }
1130
1131         /*
1132          * If we have done all the pending block allocations and if
1133          * there aren't any writers on the inode, we can discard the
1134          * inode's preallocations.
1135          */
1136         if ((ei->i_reserved_data_blocks == 0) &&
1137             (atomic_read(&inode->i_writecount) == 0))
1138                 ext4_discard_preallocations(inode);
1139 }
1140
1141 static int check_block_validity(struct inode *inode, const char *msg,
1142                                 sector_t logical, sector_t phys, int len)
1143 {
1144         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1145                 __ext4_error(inode->i_sb, msg,
1146                            "inode #%lu logical block %llu mapped to %llu "
1147                            "(size %d)", inode->i_ino,
1148                            (unsigned long long) logical,
1149                            (unsigned long long) phys, len);
1150                 return -EIO;
1151         }
1152         return 0;
1153 }
1154
1155 /*
1156  * Return the number of contiguous dirty pages in a given inode
1157  * starting at page frame idx.
1158  */
1159 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1160                                     unsigned int max_pages)
1161 {
1162         struct address_space *mapping = inode->i_mapping;
1163         pgoff_t index;
1164         struct pagevec pvec;
1165         pgoff_t num = 0;
1166         int i, nr_pages, done = 0;
1167
1168         if (max_pages == 0)
1169                 return 0;
1170         pagevec_init(&pvec, 0);
1171         while (!done) {
1172                 index = idx;
1173                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1174                                               PAGECACHE_TAG_DIRTY,
1175                                               (pgoff_t)PAGEVEC_SIZE);
1176                 if (nr_pages == 0)
1177                         break;
1178                 for (i = 0; i < nr_pages; i++) {
1179                         struct page *page = pvec.pages[i];
1180                         struct buffer_head *bh, *head;
1181
1182                         lock_page(page);
1183                         if (unlikely(page->mapping != mapping) ||
1184                             !PageDirty(page) ||
1185                             PageWriteback(page) ||
1186                             page->index != idx) {
1187                                 done = 1;
1188                                 unlock_page(page);
1189                                 break;
1190                         }
1191                         if (page_has_buffers(page)) {
1192                                 bh = head = page_buffers(page);
1193                                 do {
1194                                         if (!buffer_delay(bh) &&
1195                                             !buffer_unwritten(bh))
1196                                                 done = 1;
1197                                         bh = bh->b_this_page;
1198                                 } while (!done && (bh != head));
1199                         }
1200                         unlock_page(page);
1201                         if (done)
1202                                 break;
1203                         idx++;
1204                         num++;
1205                         if (num >= max_pages)
1206                                 break;
1207                 }
1208                 pagevec_release(&pvec);
1209         }
1210         return num;
1211 }
1212
1213 /*
1214  * The ext4_get_blocks() function tries to look up the requested blocks,
1215  * and returns if the blocks are already mapped.
1216  *
1217  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1218  * and store the allocated blocks in the result buffer head and mark it
1219  * mapped.
1220  *
1221  * If file type is extents based, it will call ext4_ext_get_blocks(),
1222  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1223  * based files
1224  *
1225  * On success, it returns the number of blocks being mapped or allocate.
1226  * if create==0 and the blocks are pre-allocated and uninitialized block,
1227  * the result buffer head is unmapped. If the create ==1, it will make sure
1228  * the buffer head is mapped.
1229  *
1230  * It returns 0 if plain look up failed (blocks have not been allocated), in
1231  * that casem, buffer head is unmapped
1232  *
1233  * It returns the error in case of allocation failure.
1234  */
1235 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1236                     unsigned int max_blocks, struct buffer_head *bh,
1237                     int flags)
1238 {
1239         int retval;
1240
1241         clear_buffer_mapped(bh);
1242         clear_buffer_unwritten(bh);
1243
1244         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1245                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1246                   (unsigned long)block);
1247         /*
1248          * Try to see if we can get the block without requesting a new
1249          * file system block.
1250          */
1251         down_read((&EXT4_I(inode)->i_data_sem));
1252         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1253                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1254                                 bh, 0);
1255         } else {
1256                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1257                                              bh, 0);
1258         }
1259         up_read((&EXT4_I(inode)->i_data_sem));
1260
1261         if (retval > 0 && buffer_mapped(bh)) {
1262                 int ret = check_block_validity(inode, "file system corruption",
1263                                                block, bh->b_blocknr, retval);
1264                 if (ret != 0)
1265                         return ret;
1266         }
1267
1268         /* If it is only a block(s) look up */
1269         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1270                 return retval;
1271
1272         /*
1273          * Returns if the blocks have already allocated
1274          *
1275          * Note that if blocks have been preallocated
1276          * ext4_ext_get_block() returns th create = 0
1277          * with buffer head unmapped.
1278          */
1279         if (retval > 0 && buffer_mapped(bh))
1280                 return retval;
1281
1282         /*
1283          * When we call get_blocks without the create flag, the
1284          * BH_Unwritten flag could have gotten set if the blocks
1285          * requested were part of a uninitialized extent.  We need to
1286          * clear this flag now that we are committed to convert all or
1287          * part of the uninitialized extent to be an initialized
1288          * extent.  This is because we need to avoid the combination
1289          * of BH_Unwritten and BH_Mapped flags being simultaneously
1290          * set on the buffer_head.
1291          */
1292         clear_buffer_unwritten(bh);
1293
1294         /*
1295          * New blocks allocate and/or writing to uninitialized extent
1296          * will possibly result in updating i_data, so we take
1297          * the write lock of i_data_sem, and call get_blocks()
1298          * with create == 1 flag.
1299          */
1300         down_write((&EXT4_I(inode)->i_data_sem));
1301
1302         /*
1303          * if the caller is from delayed allocation writeout path
1304          * we have already reserved fs blocks for allocation
1305          * let the underlying get_block() function know to
1306          * avoid double accounting
1307          */
1308         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1309                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1310         /*
1311          * We need to check for EXT4 here because migrate
1312          * could have changed the inode type in between
1313          */
1314         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1315                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1316                                               bh, flags);
1317         } else {
1318                 retval = ext4_ind_get_blocks(handle, inode, block,
1319                                              max_blocks, bh, flags);
1320
1321                 if (retval > 0 && buffer_new(bh)) {
1322                         /*
1323                          * We allocated new blocks which will result in
1324                          * i_data's format changing.  Force the migrate
1325                          * to fail by clearing migrate flags
1326                          */
1327                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1328                 }
1329
1330                 /*
1331                  * Update reserved blocks/metadata blocks after successful
1332                  * block allocation which had been deferred till now. We don't
1333                  * support fallocate for non extent files. So we can update
1334                  * reserve space here.
1335                  */
1336                 if ((retval > 0) &&
1337                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1338                         ext4_da_update_reserve_space(inode, retval, 1);
1339         }
1340         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1341                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1342
1343         up_write((&EXT4_I(inode)->i_data_sem));
1344         if (retval > 0 && buffer_mapped(bh)) {
1345                 int ret = check_block_validity(inode, "file system "
1346                                                "corruption after allocation",
1347                                                block, bh->b_blocknr, retval);
1348                 if (ret != 0)
1349                         return ret;
1350         }
1351         return retval;
1352 }
1353
1354 /* Maximum number of blocks we map for direct IO at once. */
1355 #define DIO_MAX_BLOCKS 4096
1356
1357 int ext4_get_block(struct inode *inode, sector_t iblock,
1358                    struct buffer_head *bh_result, int create)
1359 {
1360         handle_t *handle = ext4_journal_current_handle();
1361         int ret = 0, started = 0;
1362         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1363         int dio_credits;
1364
1365         if (create && !handle) {
1366                 /* Direct IO write... */
1367                 if (max_blocks > DIO_MAX_BLOCKS)
1368                         max_blocks = DIO_MAX_BLOCKS;
1369                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1370                 handle = ext4_journal_start(inode, dio_credits);
1371                 if (IS_ERR(handle)) {
1372                         ret = PTR_ERR(handle);
1373                         goto out;
1374                 }
1375                 started = 1;
1376         }
1377
1378         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1379                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1380         if (ret > 0) {
1381                 bh_result->b_size = (ret << inode->i_blkbits);
1382                 ret = 0;
1383         }
1384         if (started)
1385                 ext4_journal_stop(handle);
1386 out:
1387         return ret;
1388 }
1389
1390 /*
1391  * `handle' can be NULL if create is zero
1392  */
1393 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1394                                 ext4_lblk_t block, int create, int *errp)
1395 {
1396         struct buffer_head dummy;
1397         int fatal = 0, err;
1398         int flags = 0;
1399
1400         J_ASSERT(handle != NULL || create == 0);
1401
1402         dummy.b_state = 0;
1403         dummy.b_blocknr = -1000;
1404         buffer_trace_init(&dummy.b_history);
1405         if (create)
1406                 flags |= EXT4_GET_BLOCKS_CREATE;
1407         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1408         /*
1409          * ext4_get_blocks() returns number of blocks mapped. 0 in
1410          * case of a HOLE.
1411          */
1412         if (err > 0) {
1413                 if (err > 1)
1414                         WARN_ON(1);
1415                 err = 0;
1416         }
1417         *errp = err;
1418         if (!err && buffer_mapped(&dummy)) {
1419                 struct buffer_head *bh;
1420                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1421                 if (!bh) {
1422                         *errp = -EIO;
1423                         goto err;
1424                 }
1425                 if (buffer_new(&dummy)) {
1426                         J_ASSERT(create != 0);
1427                         J_ASSERT(handle != NULL);
1428
1429                         /*
1430                          * Now that we do not always journal data, we should
1431                          * keep in mind whether this should always journal the
1432                          * new buffer as metadata.  For now, regular file
1433                          * writes use ext4_get_block instead, so it's not a
1434                          * problem.
1435                          */
1436                         lock_buffer(bh);
1437                         BUFFER_TRACE(bh, "call get_create_access");
1438                         fatal = ext4_journal_get_create_access(handle, bh);
1439                         if (!fatal && !buffer_uptodate(bh)) {
1440                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1441                                 set_buffer_uptodate(bh);
1442                         }
1443                         unlock_buffer(bh);
1444                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1445                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1446                         if (!fatal)
1447                                 fatal = err;
1448                 } else {
1449                         BUFFER_TRACE(bh, "not a new buffer");
1450                 }
1451                 if (fatal) {
1452                         *errp = fatal;
1453                         brelse(bh);
1454                         bh = NULL;
1455                 }
1456                 return bh;
1457         }
1458 err:
1459         return NULL;
1460 }
1461
1462 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1463                                ext4_lblk_t block, int create, int *err)
1464 {
1465         struct buffer_head *bh;
1466
1467         bh = ext4_getblk(handle, inode, block, create, err);
1468         if (!bh)
1469                 return bh;
1470         if (buffer_uptodate(bh))
1471                 return bh;
1472         ll_rw_block(READ_META, 1, &bh);
1473         wait_on_buffer(bh);
1474         if (buffer_uptodate(bh))
1475                 return bh;
1476         put_bh(bh);
1477         *err = -EIO;
1478         return NULL;
1479 }
1480
1481 static int walk_page_buffers(handle_t *handle,
1482                              struct buffer_head *head,
1483                              unsigned from,
1484                              unsigned to,
1485                              int *partial,
1486                              int (*fn)(handle_t *handle,
1487                                        struct buffer_head *bh))
1488 {
1489         struct buffer_head *bh;
1490         unsigned block_start, block_end;
1491         unsigned blocksize = head->b_size;
1492         int err, ret = 0;
1493         struct buffer_head *next;
1494
1495         for (bh = head, block_start = 0;
1496              ret == 0 && (bh != head || !block_start);
1497              block_start = block_end, bh = next) {
1498                 next = bh->b_this_page;
1499                 block_end = block_start + blocksize;
1500                 if (block_end <= from || block_start >= to) {
1501                         if (partial && !buffer_uptodate(bh))
1502                                 *partial = 1;
1503                         continue;
1504                 }
1505                 err = (*fn)(handle, bh);
1506                 if (!ret)
1507                         ret = err;
1508         }
1509         return ret;
1510 }
1511
1512 /*
1513  * To preserve ordering, it is essential that the hole instantiation and
1514  * the data write be encapsulated in a single transaction.  We cannot
1515  * close off a transaction and start a new one between the ext4_get_block()
1516  * and the commit_write().  So doing the jbd2_journal_start at the start of
1517  * prepare_write() is the right place.
1518  *
1519  * Also, this function can nest inside ext4_writepage() ->
1520  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1521  * has generated enough buffer credits to do the whole page.  So we won't
1522  * block on the journal in that case, which is good, because the caller may
1523  * be PF_MEMALLOC.
1524  *
1525  * By accident, ext4 can be reentered when a transaction is open via
1526  * quota file writes.  If we were to commit the transaction while thus
1527  * reentered, there can be a deadlock - we would be holding a quota
1528  * lock, and the commit would never complete if another thread had a
1529  * transaction open and was blocking on the quota lock - a ranking
1530  * violation.
1531  *
1532  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1533  * will _not_ run commit under these circumstances because handle->h_ref
1534  * is elevated.  We'll still have enough credits for the tiny quotafile
1535  * write.
1536  */
1537 static int do_journal_get_write_access(handle_t *handle,
1538                                        struct buffer_head *bh)
1539 {
1540         if (!buffer_mapped(bh) || buffer_freed(bh))
1541                 return 0;
1542         return ext4_journal_get_write_access(handle, bh);
1543 }
1544
1545 /*
1546  * Truncate blocks that were not used by write. We have to truncate the
1547  * pagecache as well so that corresponding buffers get properly unmapped.
1548  */
1549 static void ext4_truncate_failed_write(struct inode *inode)
1550 {
1551         truncate_inode_pages(inode->i_mapping, inode->i_size);
1552         ext4_truncate(inode);
1553 }
1554
1555 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1556                    struct buffer_head *bh_result, int create);
1557 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1558                             loff_t pos, unsigned len, unsigned flags,
1559                             struct page **pagep, void **fsdata)
1560 {
1561         struct inode *inode = mapping->host;
1562         int ret, needed_blocks;
1563         handle_t *handle;
1564         int retries = 0;
1565         struct page *page;
1566         pgoff_t index;
1567         unsigned from, to;
1568
1569         trace_ext4_write_begin(inode, pos, len, flags);
1570         /*
1571          * Reserve one block more for addition to orphan list in case
1572          * we allocate blocks but write fails for some reason
1573          */
1574         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1575         index = pos >> PAGE_CACHE_SHIFT;
1576         from = pos & (PAGE_CACHE_SIZE - 1);
1577         to = from + len;
1578
1579 retry:
1580         handle = ext4_journal_start(inode, needed_blocks);
1581         if (IS_ERR(handle)) {
1582                 ret = PTR_ERR(handle);
1583                 goto out;
1584         }
1585
1586         /* We cannot recurse into the filesystem as the transaction is already
1587          * started */
1588         flags |= AOP_FLAG_NOFS;
1589
1590         page = grab_cache_page_write_begin(mapping, index, flags);
1591         if (!page) {
1592                 ext4_journal_stop(handle);
1593                 ret = -ENOMEM;
1594                 goto out;
1595         }
1596         *pagep = page;
1597
1598         if (ext4_should_dioread_nolock(inode))
1599                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1600                                 fsdata, ext4_get_block_write);
1601         else
1602                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1603                                 fsdata, ext4_get_block);
1604
1605         if (!ret && ext4_should_journal_data(inode)) {
1606                 ret = walk_page_buffers(handle, page_buffers(page),
1607                                 from, to, NULL, do_journal_get_write_access);
1608         }
1609
1610         if (ret) {
1611                 unlock_page(page);
1612                 page_cache_release(page);
1613                 /*
1614                  * block_write_begin may have instantiated a few blocks
1615                  * outside i_size.  Trim these off again. Don't need
1616                  * i_size_read because we hold i_mutex.
1617                  *
1618                  * Add inode to orphan list in case we crash before
1619                  * truncate finishes
1620                  */
1621                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1622                         ext4_orphan_add(handle, inode);
1623
1624                 ext4_journal_stop(handle);
1625                 if (pos + len > inode->i_size) {
1626                         ext4_truncate_failed_write(inode);
1627                         /*
1628                          * If truncate failed early the inode might
1629                          * still be on the orphan list; we need to
1630                          * make sure the inode is removed from the
1631                          * orphan list in that case.
1632                          */
1633                         if (inode->i_nlink)
1634                                 ext4_orphan_del(NULL, inode);
1635                 }
1636         }
1637
1638         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1639                 goto retry;
1640 out:
1641         return ret;
1642 }
1643
1644 /* For write_end() in data=journal mode */
1645 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1646 {
1647         if (!buffer_mapped(bh) || buffer_freed(bh))
1648                 return 0;
1649         set_buffer_uptodate(bh);
1650         return ext4_handle_dirty_metadata(handle, NULL, bh);
1651 }
1652
1653 static int ext4_generic_write_end(struct file *file,
1654                                   struct address_space *mapping,
1655                                   loff_t pos, unsigned len, unsigned copied,
1656                                   struct page *page, void *fsdata)
1657 {
1658         int i_size_changed = 0;
1659         struct inode *inode = mapping->host;
1660         handle_t *handle = ext4_journal_current_handle();
1661
1662         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1663
1664         /*
1665          * No need to use i_size_read() here, the i_size
1666          * cannot change under us because we hold i_mutex.
1667          *
1668          * But it's important to update i_size while still holding page lock:
1669          * page writeout could otherwise come in and zero beyond i_size.
1670          */
1671         if (pos + copied > inode->i_size) {
1672                 i_size_write(inode, pos + copied);
1673                 i_size_changed = 1;
1674         }
1675
1676         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1677                 /* We need to mark inode dirty even if
1678                  * new_i_size is less that inode->i_size
1679                  * bu greater than i_disksize.(hint delalloc)
1680                  */
1681                 ext4_update_i_disksize(inode, (pos + copied));
1682                 i_size_changed = 1;
1683         }
1684         unlock_page(page);
1685         page_cache_release(page);
1686
1687         /*
1688          * Don't mark the inode dirty under page lock. First, it unnecessarily
1689          * makes the holding time of page lock longer. Second, it forces lock
1690          * ordering of page lock and transaction start for journaling
1691          * filesystems.
1692          */
1693         if (i_size_changed)
1694                 ext4_mark_inode_dirty(handle, inode);
1695
1696         return copied;
1697 }
1698
1699 /*
1700  * We need to pick up the new inode size which generic_commit_write gave us
1701  * `file' can be NULL - eg, when called from page_symlink().
1702  *
1703  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1704  * buffers are managed internally.
1705  */
1706 static int ext4_ordered_write_end(struct file *file,
1707                                   struct address_space *mapping,
1708                                   loff_t pos, unsigned len, unsigned copied,
1709                                   struct page *page, void *fsdata)
1710 {
1711         handle_t *handle = ext4_journal_current_handle();
1712         struct inode *inode = mapping->host;
1713         int ret = 0, ret2;
1714
1715         trace_ext4_ordered_write_end(inode, pos, len, copied);
1716         ret = ext4_jbd2_file_inode(handle, inode);
1717
1718         if (ret == 0) {
1719                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1720                                                         page, fsdata);
1721                 copied = ret2;
1722                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1723                         /* if we have allocated more blocks and copied
1724                          * less. We will have blocks allocated outside
1725                          * inode->i_size. So truncate them
1726                          */
1727                         ext4_orphan_add(handle, inode);
1728                 if (ret2 < 0)
1729                         ret = ret2;
1730         }
1731         ret2 = ext4_journal_stop(handle);
1732         if (!ret)
1733                 ret = ret2;
1734
1735         if (pos + len > inode->i_size) {
1736                 ext4_truncate_failed_write(inode);
1737                 /*
1738                  * If truncate failed early the inode might still be
1739                  * on the orphan list; we need to make sure the inode
1740                  * is removed from the orphan list in that case.
1741                  */
1742                 if (inode->i_nlink)
1743                         ext4_orphan_del(NULL, inode);
1744         }
1745
1746
1747         return ret ? ret : copied;
1748 }
1749
1750 static int ext4_writeback_write_end(struct file *file,
1751                                     struct address_space *mapping,
1752                                     loff_t pos, unsigned len, unsigned copied,
1753                                     struct page *page, void *fsdata)
1754 {
1755         handle_t *handle = ext4_journal_current_handle();
1756         struct inode *inode = mapping->host;
1757         int ret = 0, ret2;
1758
1759         trace_ext4_writeback_write_end(inode, pos, len, copied);
1760         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1761                                                         page, fsdata);
1762         copied = ret2;
1763         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1764                 /* if we have allocated more blocks and copied
1765                  * less. We will have blocks allocated outside
1766                  * inode->i_size. So truncate them
1767                  */
1768                 ext4_orphan_add(handle, inode);
1769
1770         if (ret2 < 0)
1771                 ret = ret2;
1772
1773         ret2 = ext4_journal_stop(handle);
1774         if (!ret)
1775                 ret = ret2;
1776
1777         if (pos + len > inode->i_size) {
1778                 ext4_truncate_failed_write(inode);
1779                 /*
1780                  * If truncate failed early the inode might still be
1781                  * on the orphan list; we need to make sure the inode
1782                  * is removed from the orphan list in that case.
1783                  */
1784                 if (inode->i_nlink)
1785                         ext4_orphan_del(NULL, inode);
1786         }
1787
1788         return ret ? ret : copied;
1789 }
1790
1791 static int ext4_journalled_write_end(struct file *file,
1792                                      struct address_space *mapping,
1793                                      loff_t pos, unsigned len, unsigned copied,
1794                                      struct page *page, void *fsdata)
1795 {
1796         handle_t *handle = ext4_journal_current_handle();
1797         struct inode *inode = mapping->host;
1798         int ret = 0, ret2;
1799         int partial = 0;
1800         unsigned from, to;
1801         loff_t new_i_size;
1802
1803         trace_ext4_journalled_write_end(inode, pos, len, copied);
1804         from = pos & (PAGE_CACHE_SIZE - 1);
1805         to = from + len;
1806
1807         if (copied < len) {
1808                 if (!PageUptodate(page))
1809                         copied = 0;
1810                 page_zero_new_buffers(page, from+copied, to);
1811         }
1812
1813         ret = walk_page_buffers(handle, page_buffers(page), from,
1814                                 to, &partial, write_end_fn);
1815         if (!partial)
1816                 SetPageUptodate(page);
1817         new_i_size = pos + copied;
1818         if (new_i_size > inode->i_size)
1819                 i_size_write(inode, pos+copied);
1820         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1821         if (new_i_size > EXT4_I(inode)->i_disksize) {
1822                 ext4_update_i_disksize(inode, new_i_size);
1823                 ret2 = ext4_mark_inode_dirty(handle, inode);
1824                 if (!ret)
1825                         ret = ret2;
1826         }
1827
1828         unlock_page(page);
1829         page_cache_release(page);
1830         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1831                 /* if we have allocated more blocks and copied
1832                  * less. We will have blocks allocated outside
1833                  * inode->i_size. So truncate them
1834                  */
1835                 ext4_orphan_add(handle, inode);
1836
1837         ret2 = ext4_journal_stop(handle);
1838         if (!ret)
1839                 ret = ret2;
1840         if (pos + len > inode->i_size) {
1841                 ext4_truncate_failed_write(inode);
1842                 /*
1843                  * If truncate failed early the inode might still be
1844                  * on the orphan list; we need to make sure the inode
1845                  * is removed from the orphan list in that case.
1846                  */
1847                 if (inode->i_nlink)
1848                         ext4_orphan_del(NULL, inode);
1849         }
1850
1851         return ret ? ret : copied;
1852 }
1853
1854 /*
1855  * Reserve a single block located at lblock
1856  */
1857 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1858 {
1859         int retries = 0;
1860         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1861         struct ext4_inode_info *ei = EXT4_I(inode);
1862         unsigned long md_needed, md_reserved;
1863         int ret;
1864
1865         /*
1866          * recalculate the amount of metadata blocks to reserve
1867          * in order to allocate nrblocks
1868          * worse case is one extent per block
1869          */
1870 repeat:
1871         spin_lock(&ei->i_block_reservation_lock);
1872         md_reserved = ei->i_reserved_meta_blocks;
1873         md_needed = ext4_calc_metadata_amount(inode, lblock);
1874         trace_ext4_da_reserve_space(inode, md_needed);
1875         spin_unlock(&ei->i_block_reservation_lock);
1876
1877         /*
1878          * Make quota reservation here to prevent quota overflow
1879          * later. Real quota accounting is done at pages writeout
1880          * time.
1881          */
1882         ret = dquot_reserve_block(inode, md_needed + 1);
1883         if (ret)
1884                 return ret;
1885
1886         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1887                 dquot_release_reservation_block(inode, md_needed + 1);
1888                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1889                         yield();
1890                         goto repeat;
1891                 }
1892                 return -ENOSPC;
1893         }
1894         spin_lock(&ei->i_block_reservation_lock);
1895         ei->i_reserved_data_blocks++;
1896         ei->i_reserved_meta_blocks += md_needed;
1897         spin_unlock(&ei->i_block_reservation_lock);
1898
1899         return 0;       /* success */
1900 }
1901
1902 static void ext4_da_release_space(struct inode *inode, int to_free)
1903 {
1904         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1905         struct ext4_inode_info *ei = EXT4_I(inode);
1906
1907         if (!to_free)
1908                 return;         /* Nothing to release, exit */
1909
1910         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1911
1912         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1913                 /*
1914                  * if there aren't enough reserved blocks, then the
1915                  * counter is messed up somewhere.  Since this
1916                  * function is called from invalidate page, it's
1917                  * harmless to return without any action.
1918                  */
1919                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1920                          "ino %lu, to_free %d with only %d reserved "
1921                          "data blocks\n", inode->i_ino, to_free,
1922                          ei->i_reserved_data_blocks);
1923                 WARN_ON(1);
1924                 to_free = ei->i_reserved_data_blocks;
1925         }
1926         ei->i_reserved_data_blocks -= to_free;
1927
1928         if (ei->i_reserved_data_blocks == 0) {
1929                 /*
1930                  * We can release all of the reserved metadata blocks
1931                  * only when we have written all of the delayed
1932                  * allocation blocks.
1933                  */
1934                 to_free += ei->i_reserved_meta_blocks;
1935                 ei->i_reserved_meta_blocks = 0;
1936                 ei->i_da_metadata_calc_len = 0;
1937         }
1938
1939         /* update fs dirty blocks counter */
1940         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1941
1942         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1943
1944         dquot_release_reservation_block(inode, to_free);
1945 }
1946
1947 static void ext4_da_page_release_reservation(struct page *page,
1948                                              unsigned long offset)
1949 {
1950         int to_release = 0;
1951         struct buffer_head *head, *bh;
1952         unsigned int curr_off = 0;
1953
1954         head = page_buffers(page);
1955         bh = head;
1956         do {
1957                 unsigned int next_off = curr_off + bh->b_size;
1958
1959                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1960                         to_release++;
1961                         clear_buffer_delay(bh);
1962                 }
1963                 curr_off = next_off;
1964         } while ((bh = bh->b_this_page) != head);
1965         ext4_da_release_space(page->mapping->host, to_release);
1966 }
1967
1968 /*
1969  * Delayed allocation stuff
1970  */
1971
1972 /*
1973  * mpage_da_submit_io - walks through extent of pages and try to write
1974  * them with writepage() call back
1975  *
1976  * @mpd->inode: inode
1977  * @mpd->first_page: first page of the extent
1978  * @mpd->next_page: page after the last page of the extent
1979  *
1980  * By the time mpage_da_submit_io() is called we expect all blocks
1981  * to be allocated. this may be wrong if allocation failed.
1982  *
1983  * As pages are already locked by write_cache_pages(), we can't use it
1984  */
1985 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1986 {
1987         long pages_skipped;
1988         struct pagevec pvec;
1989         unsigned long index, end;
1990         int ret = 0, err, nr_pages, i;
1991         struct inode *inode = mpd->inode;
1992         struct address_space *mapping = inode->i_mapping;
1993
1994         BUG_ON(mpd->next_page <= mpd->first_page);
1995         /*
1996          * We need to start from the first_page to the next_page - 1
1997          * to make sure we also write the mapped dirty buffer_heads.
1998          * If we look at mpd->b_blocknr we would only be looking
1999          * at the currently mapped buffer_heads.
2000          */
2001         index = mpd->first_page;
2002         end = mpd->next_page - 1;
2003
2004         pagevec_init(&pvec, 0);
2005         while (index <= end) {
2006                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2007                 if (nr_pages == 0)
2008                         break;
2009                 for (i = 0; i < nr_pages; i++) {
2010                         struct page *page = pvec.pages[i];
2011
2012                         index = page->index;
2013                         if (index > end)
2014                                 break;
2015                         index++;
2016
2017                         BUG_ON(!PageLocked(page));
2018                         BUG_ON(PageWriteback(page));
2019
2020                         pages_skipped = mpd->wbc->pages_skipped;
2021                         err = mapping->a_ops->writepage(page, mpd->wbc);
2022                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2023                                 /*
2024                                  * have successfully written the page
2025                                  * without skipping the same
2026                                  */
2027                                 mpd->pages_written++;
2028                         /*
2029                          * In error case, we have to continue because
2030                          * remaining pages are still locked
2031                          * XXX: unlock and re-dirty them?
2032                          */
2033                         if (ret == 0)
2034                                 ret = err;
2035                 }
2036                 pagevec_release(&pvec);
2037         }
2038         return ret;
2039 }
2040
2041 /*
2042  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2043  *
2044  * @mpd->inode - inode to walk through
2045  * @exbh->b_blocknr - first block on a disk
2046  * @exbh->b_size - amount of space in bytes
2047  * @logical - first logical block to start assignment with
2048  *
2049  * the function goes through all passed space and put actual disk
2050  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2051  */
2052 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2053                                  struct buffer_head *exbh)
2054 {
2055         struct inode *inode = mpd->inode;
2056         struct address_space *mapping = inode->i_mapping;
2057         int blocks = exbh->b_size >> inode->i_blkbits;
2058         sector_t pblock = exbh->b_blocknr, cur_logical;
2059         struct buffer_head *head, *bh;
2060         pgoff_t index, end;
2061         struct pagevec pvec;
2062         int nr_pages, i;
2063
2064         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2065         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2066         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2067
2068         pagevec_init(&pvec, 0);
2069
2070         while (index <= end) {
2071                 /* XXX: optimize tail */
2072                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2073                 if (nr_pages == 0)
2074                         break;
2075                 for (i = 0; i < nr_pages; i++) {
2076                         struct page *page = pvec.pages[i];
2077
2078                         index = page->index;
2079                         if (index > end)
2080                                 break;
2081                         index++;
2082
2083                         BUG_ON(!PageLocked(page));
2084                         BUG_ON(PageWriteback(page));
2085                         BUG_ON(!page_has_buffers(page));
2086
2087                         bh = page_buffers(page);
2088                         head = bh;
2089
2090                         /* skip blocks out of the range */
2091                         do {
2092                                 if (cur_logical >= logical)
2093                                         break;
2094                                 cur_logical++;
2095                         } while ((bh = bh->b_this_page) != head);
2096
2097                         do {
2098                                 if (cur_logical >= logical + blocks)
2099                                         break;
2100
2101                                 if (buffer_delay(bh) ||
2102                                                 buffer_unwritten(bh)) {
2103
2104                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2105
2106                                         if (buffer_delay(bh)) {
2107                                                 clear_buffer_delay(bh);
2108                                                 bh->b_blocknr = pblock;
2109                                         } else {
2110                                                 /*
2111                                                  * unwritten already should have
2112                                                  * blocknr assigned. Verify that
2113                                                  */
2114                                                 clear_buffer_unwritten(bh);
2115                                                 BUG_ON(bh->b_blocknr != pblock);
2116                                         }
2117
2118                                 } else if (buffer_mapped(bh))
2119                                         BUG_ON(bh->b_blocknr != pblock);
2120
2121                                 if (buffer_uninit(exbh))
2122                                         set_buffer_uninit(bh);
2123                                 cur_logical++;
2124                                 pblock++;
2125                         } while ((bh = bh->b_this_page) != head);
2126                 }
2127                 pagevec_release(&pvec);
2128         }
2129 }
2130
2131
2132 /*
2133  * __unmap_underlying_blocks - just a helper function to unmap
2134  * set of blocks described by @bh
2135  */
2136 static inline void __unmap_underlying_blocks(struct inode *inode,
2137                                              struct buffer_head *bh)
2138 {
2139         struct block_device *bdev = inode->i_sb->s_bdev;
2140         int blocks, i;
2141
2142         blocks = bh->b_size >> inode->i_blkbits;
2143         for (i = 0; i < blocks; i++)
2144                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2145 }
2146
2147 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2148                                         sector_t logical, long blk_cnt)
2149 {
2150         int nr_pages, i;
2151         pgoff_t index, end;
2152         struct pagevec pvec;
2153         struct inode *inode = mpd->inode;
2154         struct address_space *mapping = inode->i_mapping;
2155
2156         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2157         end   = (logical + blk_cnt - 1) >>
2158                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2159         while (index <= end) {
2160                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2161                 if (nr_pages == 0)
2162                         break;
2163                 for (i = 0; i < nr_pages; i++) {
2164                         struct page *page = pvec.pages[i];
2165                         if (page->index > end)
2166                                 break;
2167                         BUG_ON(!PageLocked(page));
2168                         BUG_ON(PageWriteback(page));
2169                         block_invalidatepage(page, 0);
2170                         ClearPageUptodate(page);
2171                         unlock_page(page);
2172                 }
2173                 index = pvec.pages[nr_pages - 1]->index + 1;
2174                 pagevec_release(&pvec);
2175         }
2176         return;
2177 }
2178
2179 static void ext4_print_free_blocks(struct inode *inode)
2180 {
2181         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2182         printk(KERN_CRIT "Total free blocks count %lld\n",
2183                ext4_count_free_blocks(inode->i_sb));
2184         printk(KERN_CRIT "Free/Dirty block details\n");
2185         printk(KERN_CRIT "free_blocks=%lld\n",
2186                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2187         printk(KERN_CRIT "dirty_blocks=%lld\n",
2188                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2189         printk(KERN_CRIT "Block reservation details\n");
2190         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2191                EXT4_I(inode)->i_reserved_data_blocks);
2192         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2193                EXT4_I(inode)->i_reserved_meta_blocks);
2194         return;
2195 }
2196
2197 /*
2198  * mpage_da_map_blocks - go through given space
2199  *
2200  * @mpd - bh describing space
2201  *
2202  * The function skips space we know is already mapped to disk blocks.
2203  *
2204  */
2205 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2206 {
2207         int err, blks, get_blocks_flags;
2208         struct buffer_head new;
2209         sector_t next = mpd->b_blocknr;
2210         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2211         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2212         handle_t *handle = NULL;
2213
2214         /*
2215          * We consider only non-mapped and non-allocated blocks
2216          */
2217         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2218                 !(mpd->b_state & (1 << BH_Delay)) &&
2219                 !(mpd->b_state & (1 << BH_Unwritten)))
2220                 return 0;
2221
2222         /*
2223          * If we didn't accumulate anything to write simply return
2224          */
2225         if (!mpd->b_size)
2226                 return 0;
2227
2228         handle = ext4_journal_current_handle();
2229         BUG_ON(!handle);
2230
2231         /*
2232          * Call ext4_get_blocks() to allocate any delayed allocation
2233          * blocks, or to convert an uninitialized extent to be
2234          * initialized (in the case where we have written into
2235          * one or more preallocated blocks).
2236          *
2237          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2238          * indicate that we are on the delayed allocation path.  This
2239          * affects functions in many different parts of the allocation
2240          * call path.  This flag exists primarily because we don't
2241          * want to change *many* call functions, so ext4_get_blocks()
2242          * will set the magic i_delalloc_reserved_flag once the
2243          * inode's allocation semaphore is taken.
2244          *
2245          * If the blocks in questions were delalloc blocks, set
2246          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2247          * variables are updated after the blocks have been allocated.
2248          */
2249         new.b_state = 0;
2250         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2251         if (ext4_should_dioread_nolock(mpd->inode))
2252                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2253         if (mpd->b_state & (1 << BH_Delay))
2254                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2255
2256         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2257                                &new, get_blocks_flags);
2258         if (blks < 0) {
2259                 err = blks;
2260                 /*
2261                  * If get block returns with error we simply
2262                  * return. Later writepage will redirty the page and
2263                  * writepages will find the dirty page again
2264                  */
2265                 if (err == -EAGAIN)
2266                         return 0;
2267
2268                 if (err == -ENOSPC &&
2269                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2270                         mpd->retval = err;
2271                         return 0;
2272                 }
2273
2274                 /*
2275                  * get block failure will cause us to loop in
2276                  * writepages, because a_ops->writepage won't be able
2277                  * to make progress. The page will be redirtied by
2278                  * writepage and writepages will again try to write
2279                  * the same.
2280                  */
2281                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2282                          "delayed block allocation failed for inode %lu at "
2283                          "logical offset %llu with max blocks %zd with "
2284                          "error %d\n", mpd->inode->i_ino,
2285                          (unsigned long long) next,
2286                          mpd->b_size >> mpd->inode->i_blkbits, err);
2287                 printk(KERN_CRIT "This should not happen!!  "
2288                        "Data will be lost\n");
2289                 if (err == -ENOSPC) {
2290                         ext4_print_free_blocks(mpd->inode);
2291                 }
2292                 /* invalidate all the pages */
2293                 ext4_da_block_invalidatepages(mpd, next,
2294                                 mpd->b_size >> mpd->inode->i_blkbits);
2295                 return err;
2296         }
2297         BUG_ON(blks == 0);
2298
2299         new.b_size = (blks << mpd->inode->i_blkbits);
2300
2301         if (buffer_new(&new))
2302                 __unmap_underlying_blocks(mpd->inode, &new);
2303
2304         /*
2305          * If blocks are delayed marked, we need to
2306          * put actual blocknr and drop delayed bit
2307          */
2308         if ((mpd->b_state & (1 << BH_Delay)) ||
2309             (mpd->b_state & (1 << BH_Unwritten)))
2310                 mpage_put_bnr_to_bhs(mpd, next, &new);
2311
2312         if (ext4_should_order_data(mpd->inode)) {
2313                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2314                 if (err)
2315                         return err;
2316         }
2317
2318         /*
2319          * Update on-disk size along with block allocation.
2320          */
2321         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2322         if (disksize > i_size_read(mpd->inode))
2323                 disksize = i_size_read(mpd->inode);
2324         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2325                 ext4_update_i_disksize(mpd->inode, disksize);
2326                 return ext4_mark_inode_dirty(handle, mpd->inode);
2327         }
2328
2329         return 0;
2330 }
2331
2332 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2333                 (1 << BH_Delay) | (1 << BH_Unwritten))
2334
2335 /*
2336  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2337  *
2338  * @mpd->lbh - extent of blocks
2339  * @logical - logical number of the block in the file
2340  * @bh - bh of the block (used to access block's state)
2341  *
2342  * the function is used to collect contig. blocks in same state
2343  */
2344 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2345                                    sector_t logical, size_t b_size,
2346                                    unsigned long b_state)
2347 {
2348         sector_t next;
2349         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2350
2351         /* check if thereserved journal credits might overflow */
2352         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2353                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2354                         /*
2355                          * With non-extent format we are limited by the journal
2356                          * credit available.  Total credit needed to insert
2357                          * nrblocks contiguous blocks is dependent on the
2358                          * nrblocks.  So limit nrblocks.
2359                          */
2360                         goto flush_it;
2361                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2362                                 EXT4_MAX_TRANS_DATA) {
2363                         /*
2364                          * Adding the new buffer_head would make it cross the
2365                          * allowed limit for which we have journal credit
2366                          * reserved. So limit the new bh->b_size
2367                          */
2368                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2369                                                 mpd->inode->i_blkbits;
2370                         /* we will do mpage_da_submit_io in the next loop */
2371                 }
2372         }
2373         /*
2374          * First block in the extent
2375          */
2376         if (mpd->b_size == 0) {
2377                 mpd->b_blocknr = logical;
2378                 mpd->b_size = b_size;
2379                 mpd->b_state = b_state & BH_FLAGS;
2380                 return;
2381         }
2382
2383         next = mpd->b_blocknr + nrblocks;
2384         /*
2385          * Can we merge the block to our big extent?
2386          */
2387         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2388                 mpd->b_size += b_size;
2389                 return;
2390         }
2391
2392 flush_it:
2393         /*
2394          * We couldn't merge the block to our extent, so we
2395          * need to flush current  extent and start new one
2396          */
2397         if (mpage_da_map_blocks(mpd) == 0)
2398                 mpage_da_submit_io(mpd);
2399         mpd->io_done = 1;
2400         return;
2401 }
2402
2403 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2404 {
2405         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2406 }
2407
2408 /*
2409  * __mpage_da_writepage - finds extent of pages and blocks
2410  *
2411  * @page: page to consider
2412  * @wbc: not used, we just follow rules
2413  * @data: context
2414  *
2415  * The function finds extents of pages and scan them for all blocks.
2416  */
2417 static int __mpage_da_writepage(struct page *page,
2418                                 struct writeback_control *wbc, void *data)
2419 {
2420         struct mpage_da_data *mpd = data;
2421         struct inode *inode = mpd->inode;
2422         struct buffer_head *bh, *head;
2423         sector_t logical;
2424
2425         if (mpd->io_done) {
2426                 /*
2427                  * Rest of the page in the page_vec
2428                  * redirty then and skip then. We will
2429                  * try to write them again after
2430                  * starting a new transaction
2431                  */
2432                 redirty_page_for_writepage(wbc, page);
2433                 unlock_page(page);
2434                 return MPAGE_DA_EXTENT_TAIL;
2435         }
2436         /*
2437          * Can we merge this page to current extent?
2438          */
2439         if (mpd->next_page != page->index) {
2440                 /*
2441                  * Nope, we can't. So, we map non-allocated blocks
2442                  * and start IO on them using writepage()
2443                  */
2444                 if (mpd->next_page != mpd->first_page) {
2445                         if (mpage_da_map_blocks(mpd) == 0)
2446                                 mpage_da_submit_io(mpd);
2447                         /*
2448                          * skip rest of the page in the page_vec
2449                          */
2450                         mpd->io_done = 1;
2451                         redirty_page_for_writepage(wbc, page);
2452                         unlock_page(page);
2453                         return MPAGE_DA_EXTENT_TAIL;
2454                 }
2455
2456                 /*
2457                  * Start next extent of pages ...
2458                  */
2459                 mpd->first_page = page->index;
2460
2461                 /*
2462                  * ... and blocks
2463                  */
2464                 mpd->b_size = 0;
2465                 mpd->b_state = 0;
2466                 mpd->b_blocknr = 0;
2467         }
2468
2469         mpd->next_page = page->index + 1;
2470         logical = (sector_t) page->index <<
2471                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2472
2473         if (!page_has_buffers(page)) {
2474                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2475                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2476                 if (mpd->io_done)
2477                         return MPAGE_DA_EXTENT_TAIL;
2478         } else {
2479                 /*
2480                  * Page with regular buffer heads, just add all dirty ones
2481                  */
2482                 head = page_buffers(page);
2483                 bh = head;
2484                 do {
2485                         BUG_ON(buffer_locked(bh));
2486                         /*
2487                          * We need to try to allocate
2488                          * unmapped blocks in the same page.
2489                          * Otherwise we won't make progress
2490                          * with the page in ext4_writepage
2491                          */
2492                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2493                                 mpage_add_bh_to_extent(mpd, logical,
2494                                                        bh->b_size,
2495                                                        bh->b_state);
2496                                 if (mpd->io_done)
2497                                         return MPAGE_DA_EXTENT_TAIL;
2498                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2499                                 /*
2500                                  * mapped dirty buffer. We need to update
2501                                  * the b_state because we look at
2502                                  * b_state in mpage_da_map_blocks. We don't
2503                                  * update b_size because if we find an
2504                                  * unmapped buffer_head later we need to
2505                                  * use the b_state flag of that buffer_head.
2506                                  */
2507                                 if (mpd->b_size == 0)
2508                                         mpd->b_state = bh->b_state & BH_FLAGS;
2509                         }
2510                         logical++;
2511                 } while ((bh = bh->b_this_page) != head);
2512         }
2513
2514         return 0;
2515 }
2516
2517 /*
2518  * This is a special get_blocks_t callback which is used by
2519  * ext4_da_write_begin().  It will either return mapped block or
2520  * reserve space for a single block.
2521  *
2522  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2523  * We also have b_blocknr = -1 and b_bdev initialized properly
2524  *
2525  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2526  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2527  * initialized properly.
2528  */
2529 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2530                                   struct buffer_head *bh_result, int create)
2531 {
2532         int ret = 0;
2533         sector_t invalid_block = ~((sector_t) 0xffff);
2534
2535         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2536                 invalid_block = ~0;
2537
2538         BUG_ON(create == 0);
2539         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2540
2541         /*
2542          * first, we need to know whether the block is allocated already
2543          * preallocated blocks are unmapped but should treated
2544          * the same as allocated blocks.
2545          */
2546         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2547         if ((ret == 0) && !buffer_delay(bh_result)) {
2548                 /* the block isn't (pre)allocated yet, let's reserve space */
2549                 /*
2550                  * XXX: __block_prepare_write() unmaps passed block,
2551                  * is it OK?
2552                  */
2553                 ret = ext4_da_reserve_space(inode, iblock);
2554                 if (ret)
2555                         /* not enough space to reserve */
2556                         return ret;
2557
2558                 map_bh(bh_result, inode->i_sb, invalid_block);
2559                 set_buffer_new(bh_result);
2560                 set_buffer_delay(bh_result);
2561         } else if (ret > 0) {
2562                 bh_result->b_size = (ret << inode->i_blkbits);
2563                 if (buffer_unwritten(bh_result)) {
2564                         /* A delayed write to unwritten bh should
2565                          * be marked new and mapped.  Mapped ensures
2566                          * that we don't do get_block multiple times
2567                          * when we write to the same offset and new
2568                          * ensures that we do proper zero out for
2569                          * partial write.
2570                          */
2571                         set_buffer_new(bh_result);
2572                         set_buffer_mapped(bh_result);
2573                 }
2574                 ret = 0;
2575         }
2576
2577         return ret;
2578 }
2579
2580 /*
2581  * This function is used as a standard get_block_t calback function
2582  * when there is no desire to allocate any blocks.  It is used as a
2583  * callback function for block_prepare_write(), nobh_writepage(), and
2584  * block_write_full_page().  These functions should only try to map a
2585  * single block at a time.
2586  *
2587  * Since this function doesn't do block allocations even if the caller
2588  * requests it by passing in create=1, it is critically important that
2589  * any caller checks to make sure that any buffer heads are returned
2590  * by this function are either all already mapped or marked for
2591  * delayed allocation before calling nobh_writepage() or
2592  * block_write_full_page().  Otherwise, b_blocknr could be left
2593  * unitialized, and the page write functions will be taken by
2594  * surprise.
2595  */
2596 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2597                                    struct buffer_head *bh_result, int create)
2598 {
2599         int ret = 0;
2600         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2601
2602         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2603
2604         /*
2605          * we don't want to do block allocation in writepage
2606          * so call get_block_wrap with create = 0
2607          */
2608         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2609         if (ret > 0) {
2610                 bh_result->b_size = (ret << inode->i_blkbits);
2611                 ret = 0;
2612         }
2613         return ret;
2614 }
2615
2616 static int bget_one(handle_t *handle, struct buffer_head *bh)
2617 {
2618         get_bh(bh);
2619         return 0;
2620 }
2621
2622 static int bput_one(handle_t *handle, struct buffer_head *bh)
2623 {
2624         put_bh(bh);
2625         return 0;
2626 }
2627
2628 static int __ext4_journalled_writepage(struct page *page,
2629                                        unsigned int len)
2630 {
2631         struct address_space *mapping = page->mapping;
2632         struct inode *inode = mapping->host;
2633         struct buffer_head *page_bufs;
2634         handle_t *handle = NULL;
2635         int ret = 0;
2636         int err;
2637
2638         page_bufs = page_buffers(page);
2639         BUG_ON(!page_bufs);
2640         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2641         /* As soon as we unlock the page, it can go away, but we have
2642          * references to buffers so we are safe */
2643         unlock_page(page);
2644
2645         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2646         if (IS_ERR(handle)) {
2647                 ret = PTR_ERR(handle);
2648                 goto out;
2649         }
2650
2651         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2652                                 do_journal_get_write_access);
2653
2654         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2655                                 write_end_fn);
2656         if (ret == 0)
2657                 ret = err;
2658         err = ext4_journal_stop(handle);
2659         if (!ret)
2660                 ret = err;
2661
2662         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2663         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2664 out:
2665         return ret;
2666 }
2667
2668 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2669 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2670
2671 /*
2672  * Note that we don't need to start a transaction unless we're journaling data
2673  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2674  * need to file the inode to the transaction's list in ordered mode because if
2675  * we are writing back data added by write(), the inode is already there and if
2676  * we are writing back data modified via mmap(), noone guarantees in which
2677  * transaction the data will hit the disk. In case we are journaling data, we
2678  * cannot start transaction directly because transaction start ranks above page
2679  * lock so we have to do some magic.
2680  *
2681  * This function can get called via...
2682  *   - ext4_da_writepages after taking page lock (have journal handle)
2683  *   - journal_submit_inode_data_buffers (no journal handle)
2684  *   - shrink_page_list via pdflush (no journal handle)
2685  *   - grab_page_cache when doing write_begin (have journal handle)
2686  *
2687  * We don't do any block allocation in this function. If we have page with
2688  * multiple blocks we need to write those buffer_heads that are mapped. This
2689  * is important for mmaped based write. So if we do with blocksize 1K
2690  * truncate(f, 1024);
2691  * a = mmap(f, 0, 4096);
2692  * a[0] = 'a';
2693  * truncate(f, 4096);
2694  * we have in the page first buffer_head mapped via page_mkwrite call back
2695  * but other bufer_heads would be unmapped but dirty(dirty done via the
2696  * do_wp_page). So writepage should write the first block. If we modify
2697  * the mmap area beyond 1024 we will again get a page_fault and the
2698  * page_mkwrite callback will do the block allocation and mark the
2699  * buffer_heads mapped.
2700  *
2701  * We redirty the page if we have any buffer_heads that is either delay or
2702  * unwritten in the page.
2703  *
2704  * We can get recursively called as show below.
2705  *
2706  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2707  *              ext4_writepage()
2708  *
2709  * But since we don't do any block allocation we should not deadlock.
2710  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2711  */
2712 static int ext4_writepage(struct page *page,
2713                           struct writeback_control *wbc)
2714 {
2715         int ret = 0;
2716         loff_t size;
2717         unsigned int len;
2718         struct buffer_head *page_bufs = NULL;
2719         struct inode *inode = page->mapping->host;
2720
2721         trace_ext4_writepage(inode, page);
2722         size = i_size_read(inode);
2723         if (page->index == size >> PAGE_CACHE_SHIFT)
2724                 len = size & ~PAGE_CACHE_MASK;
2725         else
2726                 len = PAGE_CACHE_SIZE;
2727
2728         if (page_has_buffers(page)) {
2729                 page_bufs = page_buffers(page);
2730                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2731                                         ext4_bh_delay_or_unwritten)) {
2732                         /*
2733                          * We don't want to do  block allocation
2734                          * So redirty the page and return
2735                          * We may reach here when we do a journal commit
2736                          * via journal_submit_inode_data_buffers.
2737                          * If we don't have mapping block we just ignore
2738                          * them. We can also reach here via shrink_page_list
2739                          */
2740                         redirty_page_for_writepage(wbc, page);
2741                         unlock_page(page);
2742                         return 0;
2743                 }
2744         } else {
2745                 /*
2746                  * The test for page_has_buffers() is subtle:
2747                  * We know the page is dirty but it lost buffers. That means
2748                  * that at some moment in time after write_begin()/write_end()
2749                  * has been called all buffers have been clean and thus they
2750                  * must have been written at least once. So they are all
2751                  * mapped and we can happily proceed with mapping them
2752                  * and writing the page.
2753                  *
2754                  * Try to initialize the buffer_heads and check whether
2755                  * all are mapped and non delay. We don't want to
2756                  * do block allocation here.
2757                  */
2758                 ret = block_prepare_write(page, 0, len,
2759                                           noalloc_get_block_write);
2760                 if (!ret) {
2761                         page_bufs = page_buffers(page);
2762                         /* check whether all are mapped and non delay */
2763                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2764                                                 ext4_bh_delay_or_unwritten)) {
2765                                 redirty_page_for_writepage(wbc, page);
2766                                 unlock_page(page);
2767                                 return 0;
2768                         }
2769                 } else {
2770                         /*
2771                          * We can't do block allocation here
2772                          * so just redity the page and unlock
2773                          * and return
2774                          */
2775                         redirty_page_for_writepage(wbc, page);
2776                         unlock_page(page);
2777                         return 0;
2778                 }
2779                 /* now mark the buffer_heads as dirty and uptodate */
2780                 block_commit_write(page, 0, len);
2781         }
2782
2783         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2784                 /*
2785                  * It's mmapped pagecache.  Add buffers and journal it.  There
2786                  * doesn't seem much point in redirtying the page here.
2787                  */
2788                 ClearPageChecked(page);
2789                 return __ext4_journalled_writepage(page, len);
2790         }
2791
2792         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2793                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2794         else if (page_bufs && buffer_uninit(page_bufs)) {
2795                 ext4_set_bh_endio(page_bufs, inode);
2796                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2797                                             wbc, ext4_end_io_buffer_write);
2798         } else
2799                 ret = block_write_full_page(page, noalloc_get_block_write,
2800                                             wbc);
2801
2802         return ret;
2803 }
2804
2805 /*
2806  * This is called via ext4_da_writepages() to
2807  * calulate the total number of credits to reserve to fit
2808  * a single extent allocation into a single transaction,
2809  * ext4_da_writpeages() will loop calling this before
2810  * the block allocation.
2811  */
2812
2813 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2814 {
2815         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2816
2817         /*
2818          * With non-extent format the journal credit needed to
2819          * insert nrblocks contiguous block is dependent on
2820          * number of contiguous block. So we will limit
2821          * number of contiguous block to a sane value
2822          */
2823         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2824             (max_blocks > EXT4_MAX_TRANS_DATA))
2825                 max_blocks = EXT4_MAX_TRANS_DATA;
2826
2827         return ext4_chunk_trans_blocks(inode, max_blocks);
2828 }
2829
2830 static int ext4_da_writepages(struct address_space *mapping,
2831                               struct writeback_control *wbc)
2832 {
2833         pgoff_t index;
2834         int range_whole = 0;
2835         handle_t *handle = NULL;
2836         struct mpage_da_data mpd;
2837         struct inode *inode = mapping->host;
2838         int no_nrwrite_index_update;
2839         int pages_written = 0;
2840         long pages_skipped;
2841         unsigned int max_pages;
2842         int range_cyclic, cycled = 1, io_done = 0;
2843         int needed_blocks, ret = 0;
2844         long desired_nr_to_write, nr_to_writebump = 0;
2845         loff_t range_start = wbc->range_start;
2846         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2847
2848         trace_ext4_da_writepages(inode, wbc);
2849
2850         /*
2851          * No pages to write? This is mainly a kludge to avoid starting
2852          * a transaction for special inodes like journal inode on last iput()
2853          * because that could violate lock ordering on umount
2854          */
2855         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2856                 return 0;
2857
2858         /*
2859          * If the filesystem has aborted, it is read-only, so return
2860          * right away instead of dumping stack traces later on that
2861          * will obscure the real source of the problem.  We test
2862          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2863          * the latter could be true if the filesystem is mounted
2864          * read-only, and in that case, ext4_da_writepages should
2865          * *never* be called, so if that ever happens, we would want
2866          * the stack trace.
2867          */
2868         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2869                 return -EROFS;
2870
2871         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2872                 range_whole = 1;
2873
2874         range_cyclic = wbc->range_cyclic;
2875         if (wbc->range_cyclic) {
2876                 index = mapping->writeback_index;
2877                 if (index)
2878                         cycled = 0;
2879                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2880                 wbc->range_end  = LLONG_MAX;
2881                 wbc->range_cyclic = 0;
2882         } else
2883                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2884
2885         /*
2886          * This works around two forms of stupidity.  The first is in
2887          * the writeback code, which caps the maximum number of pages
2888          * written to be 1024 pages.  This is wrong on multiple
2889          * levels; different architectues have a different page size,
2890          * which changes the maximum amount of data which gets
2891          * written.  Secondly, 4 megabytes is way too small.  XFS
2892          * forces this value to be 16 megabytes by multiplying
2893          * nr_to_write parameter by four, and then relies on its
2894          * allocator to allocate larger extents to make them
2895          * contiguous.  Unfortunately this brings us to the second
2896          * stupidity, which is that ext4's mballoc code only allocates
2897          * at most 2048 blocks.  So we force contiguous writes up to
2898          * the number of dirty blocks in the inode, or
2899          * sbi->max_writeback_mb_bump whichever is smaller.
2900          */
2901         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2902         if (!range_cyclic && range_whole)
2903                 desired_nr_to_write = wbc->nr_to_write * 8;
2904         else
2905                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2906                                                            max_pages);
2907         if (desired_nr_to_write > max_pages)
2908                 desired_nr_to_write = max_pages;
2909
2910         if (wbc->nr_to_write < desired_nr_to_write) {
2911                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2912                 wbc->nr_to_write = desired_nr_to_write;
2913         }
2914
2915         mpd.wbc = wbc;
2916         mpd.inode = mapping->host;
2917
2918         /*
2919          * we don't want write_cache_pages to update
2920          * nr_to_write and writeback_index
2921          */
2922         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2923         wbc->no_nrwrite_index_update = 1;
2924         pages_skipped = wbc->pages_skipped;
2925
2926 retry:
2927         while (!ret && wbc->nr_to_write > 0) {
2928
2929                 /*
2930                  * we  insert one extent at a time. So we need
2931                  * credit needed for single extent allocation.
2932                  * journalled mode is currently not supported
2933                  * by delalloc
2934                  */
2935                 BUG_ON(ext4_should_journal_data(inode));
2936                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2937
2938                 /* start a new transaction*/
2939                 handle = ext4_journal_start(inode, needed_blocks);
2940                 if (IS_ERR(handle)) {
2941                         ret = PTR_ERR(handle);
2942                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2943                                "%ld pages, ino %lu; err %d\n", __func__,
2944                                 wbc->nr_to_write, inode->i_ino, ret);
2945                         goto out_writepages;
2946                 }
2947
2948                 /*
2949                  * Now call __mpage_da_writepage to find the next
2950                  * contiguous region of logical blocks that need
2951                  * blocks to be allocated by ext4.  We don't actually
2952                  * submit the blocks for I/O here, even though
2953                  * write_cache_pages thinks it will, and will set the
2954                  * pages as clean for write before calling
2955                  * __mpage_da_writepage().
2956                  */
2957                 mpd.b_size = 0;
2958                 mpd.b_state = 0;
2959                 mpd.b_blocknr = 0;
2960                 mpd.first_page = 0;
2961                 mpd.next_page = 0;
2962                 mpd.io_done = 0;
2963                 mpd.pages_written = 0;
2964                 mpd.retval = 0;
2965                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2966                                         &mpd);
2967                 /*
2968                  * If we have a contiguous extent of pages and we
2969                  * haven't done the I/O yet, map the blocks and submit
2970                  * them for I/O.
2971                  */
2972                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2973                         if (mpage_da_map_blocks(&mpd) == 0)
2974                                 mpage_da_submit_io(&mpd);
2975                         mpd.io_done = 1;
2976                         ret = MPAGE_DA_EXTENT_TAIL;
2977                 }
2978                 trace_ext4_da_write_pages(inode, &mpd);
2979                 wbc->nr_to_write -= mpd.pages_written;
2980
2981                 ext4_journal_stop(handle);
2982
2983                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2984                         /* commit the transaction which would
2985                          * free blocks released in the transaction
2986                          * and try again
2987                          */
2988                         jbd2_journal_force_commit_nested(sbi->s_journal);
2989                         wbc->pages_skipped = pages_skipped;
2990                         ret = 0;
2991                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2992                         /*
2993                          * got one extent now try with
2994                          * rest of the pages
2995                          */
2996                         pages_written += mpd.pages_written;
2997                         wbc->pages_skipped = pages_skipped;
2998                         ret = 0;
2999                         io_done = 1;
3000                 } else if (wbc->nr_to_write)
3001                         /*
3002                          * There is no more writeout needed
3003                          * or we requested for a noblocking writeout
3004                          * and we found the device congested
3005                          */
3006                         break;
3007         }
3008         if (!io_done && !cycled) {
3009                 cycled = 1;
3010                 index = 0;
3011                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3012                 wbc->range_end  = mapping->writeback_index - 1;
3013                 goto retry;
3014         }
3015         if (pages_skipped != wbc->pages_skipped)
3016                 ext4_msg(inode->i_sb, KERN_CRIT,
3017                          "This should not happen leaving %s "
3018                          "with nr_to_write = %ld ret = %d\n",
3019                          __func__, wbc->nr_to_write, ret);
3020
3021         /* Update index */
3022         index += pages_written;
3023         wbc->range_cyclic = range_cyclic;
3024         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3025                 /*
3026                  * set the writeback_index so that range_cyclic
3027                  * mode will write it back later
3028                  */
3029                 mapping->writeback_index = index;
3030
3031 out_writepages:
3032         if (!no_nrwrite_index_update)
3033                 wbc->no_nrwrite_index_update = 0;
3034         wbc->nr_to_write -= nr_to_writebump;
3035         wbc->range_start = range_start;
3036         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3037         return ret;
3038 }
3039
3040 #define FALL_BACK_TO_NONDELALLOC 1
3041 static int ext4_nonda_switch(struct super_block *sb)
3042 {
3043         s64 free_blocks, dirty_blocks;
3044         struct ext4_sb_info *sbi = EXT4_SB(sb);
3045
3046         /*
3047          * switch to non delalloc mode if we are running low
3048          * on free block. The free block accounting via percpu
3049          * counters can get slightly wrong with percpu_counter_batch getting
3050          * accumulated on each CPU without updating global counters
3051          * Delalloc need an accurate free block accounting. So switch
3052          * to non delalloc when we are near to error range.
3053          */
3054         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3055         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3056         if (2 * free_blocks < 3 * dirty_blocks ||
3057                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3058                 /*
3059                  * free block count is less than 150% of dirty blocks
3060                  * or free blocks is less than watermark
3061                  */
3062                 return 1;
3063         }
3064         /*
3065          * Even if we don't switch but are nearing capacity,
3066          * start pushing delalloc when 1/2 of free blocks are dirty.
3067          */
3068         if (free_blocks < 2 * dirty_blocks)
3069                 writeback_inodes_sb_if_idle(sb);
3070
3071         return 0;
3072 }
3073
3074 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3075                                loff_t pos, unsigned len, unsigned flags,
3076                                struct page **pagep, void **fsdata)
3077 {
3078         int ret, retries = 0, quota_retries = 0;
3079         struct page *page;
3080         pgoff_t index;
3081         unsigned from, to;
3082         struct inode *inode = mapping->host;
3083         handle_t *handle;
3084
3085         index = pos >> PAGE_CACHE_SHIFT;
3086         from = pos & (PAGE_CACHE_SIZE - 1);
3087         to = from + len;
3088
3089         if (ext4_nonda_switch(inode->i_sb)) {
3090                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3091                 return ext4_write_begin(file, mapping, pos,
3092                                         len, flags, pagep, fsdata);
3093         }
3094         *fsdata = (void *)0;
3095         trace_ext4_da_write_begin(inode, pos, len, flags);
3096 retry:
3097         /*
3098          * With delayed allocation, we don't log the i_disksize update
3099          * if there is delayed block allocation. But we still need
3100          * to journalling the i_disksize update if writes to the end
3101          * of file which has an already mapped buffer.
3102          */
3103         handle = ext4_journal_start(inode, 1);
3104         if (IS_ERR(handle)) {
3105                 ret = PTR_ERR(handle);
3106                 goto out;
3107         }
3108         /* We cannot recurse into the filesystem as the transaction is already
3109          * started */
3110         flags |= AOP_FLAG_NOFS;
3111
3112         page = grab_cache_page_write_begin(mapping, index, flags);
3113         if (!page) {
3114                 ext4_journal_stop(handle);
3115                 ret = -ENOMEM;
3116                 goto out;
3117         }
3118         *pagep = page;
3119
3120         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3121                                 ext4_da_get_block_prep);
3122         if (ret < 0) {
3123                 unlock_page(page);
3124                 ext4_journal_stop(handle);
3125                 page_cache_release(page);
3126                 /*
3127                  * block_write_begin may have instantiated a few blocks
3128                  * outside i_size.  Trim these off again. Don't need
3129                  * i_size_read because we hold i_mutex.
3130                  */
3131                 if (pos + len > inode->i_size)
3132                         ext4_truncate_failed_write(inode);
3133         }
3134
3135         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3136                 goto retry;
3137
3138         if ((ret == -EDQUOT) &&
3139             EXT4_I(inode)->i_reserved_meta_blocks &&
3140             (quota_retries++ < 3)) {
3141                 /*
3142                  * Since we often over-estimate the number of meta
3143                  * data blocks required, we may sometimes get a
3144                  * spurios out of quota error even though there would
3145                  * be enough space once we write the data blocks and
3146                  * find out how many meta data blocks were _really_
3147                  * required.  So try forcing the inode write to see if
3148                  * that helps.
3149                  */
3150                 write_inode_now(inode, (quota_retries == 3));
3151                 goto retry;
3152         }
3153 out:
3154         return ret;
3155 }
3156
3157 /*
3158  * Check if we should update i_disksize
3159  * when write to the end of file but not require block allocation
3160  */
3161 static int ext4_da_should_update_i_disksize(struct page *page,
3162                                             unsigned long offset)
3163 {
3164         struct buffer_head *bh;
3165         struct inode *inode = page->mapping->host;
3166         unsigned int idx;
3167         int i;
3168
3169         bh = page_buffers(page);
3170         idx = offset >> inode->i_blkbits;
3171
3172         for (i = 0; i < idx; i++)
3173                 bh = bh->b_this_page;
3174
3175         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3176                 return 0;
3177         return 1;
3178 }
3179
3180 static int ext4_da_write_end(struct file *file,
3181                              struct address_space *mapping,
3182                              loff_t pos, unsigned len, unsigned copied,
3183                              struct page *page, void *fsdata)
3184 {
3185         struct inode *inode = mapping->host;
3186         int ret = 0, ret2;
3187         handle_t *handle = ext4_journal_current_handle();
3188         loff_t new_i_size;
3189         unsigned long start, end;
3190         int write_mode = (int)(unsigned long)fsdata;
3191
3192         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3193                 if (ext4_should_order_data(inode)) {
3194                         return ext4_ordered_write_end(file, mapping, pos,
3195                                         len, copied, page, fsdata);
3196                 } else if (ext4_should_writeback_data(inode)) {
3197                         return ext4_writeback_write_end(file, mapping, pos,
3198                                         len, copied, page, fsdata);
3199                 } else {
3200                         BUG();
3201                 }
3202         }
3203
3204         trace_ext4_da_write_end(inode, pos, len, copied);
3205         start = pos & (PAGE_CACHE_SIZE - 1);
3206         end = start + copied - 1;
3207
3208         /*
3209          * generic_write_end() will run mark_inode_dirty() if i_size
3210          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3211          * into that.
3212          */
3213
3214         new_i_size = pos + copied;
3215         if (new_i_size > EXT4_I(inode)->i_disksize) {
3216                 if (ext4_da_should_update_i_disksize(page, end)) {
3217                         down_write(&EXT4_I(inode)->i_data_sem);
3218                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3219                                 /*
3220                                  * Updating i_disksize when extending file
3221                                  * without needing block allocation
3222                                  */
3223                                 if (ext4_should_order_data(inode))
3224                                         ret = ext4_jbd2_file_inode(handle,
3225                                                                    inode);
3226
3227                                 EXT4_I(inode)->i_disksize = new_i_size;
3228                         }
3229                         up_write(&EXT4_I(inode)->i_data_sem);
3230                         /* We need to mark inode dirty even if
3231                          * new_i_size is less that inode->i_size
3232                          * bu greater than i_disksize.(hint delalloc)
3233                          */
3234                         ext4_mark_inode_dirty(handle, inode);
3235                 }
3236         }
3237         ret2 = generic_write_end(file, mapping, pos, len, copied,
3238                                                         page, fsdata);
3239         copied = ret2;
3240         if (ret2 < 0)
3241                 ret = ret2;
3242         ret2 = ext4_journal_stop(handle);
3243         if (!ret)
3244                 ret = ret2;
3245
3246         return ret ? ret : copied;
3247 }
3248
3249 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3250 {
3251         /*
3252          * Drop reserved blocks
3253          */
3254         BUG_ON(!PageLocked(page));
3255         if (!page_has_buffers(page))
3256                 goto out;
3257
3258         ext4_da_page_release_reservation(page, offset);
3259
3260 out:
3261         ext4_invalidatepage(page, offset);
3262
3263         return;
3264 }
3265
3266 /*
3267  * Force all delayed allocation blocks to be allocated for a given inode.
3268  */
3269 int ext4_alloc_da_blocks(struct inode *inode)
3270 {
3271         trace_ext4_alloc_da_blocks(inode);
3272
3273         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3274             !EXT4_I(inode)->i_reserved_meta_blocks)
3275                 return 0;
3276
3277         /*
3278          * We do something simple for now.  The filemap_flush() will
3279          * also start triggering a write of the data blocks, which is
3280          * not strictly speaking necessary (and for users of
3281          * laptop_mode, not even desirable).  However, to do otherwise
3282          * would require replicating code paths in:
3283          *
3284          * ext4_da_writepages() ->
3285          *    write_cache_pages() ---> (via passed in callback function)
3286          *        __mpage_da_writepage() -->
3287          *           mpage_add_bh_to_extent()
3288          *           mpage_da_map_blocks()
3289          *
3290          * The problem is that write_cache_pages(), located in
3291          * mm/page-writeback.c, marks pages clean in preparation for
3292          * doing I/O, which is not desirable if we're not planning on
3293          * doing I/O at all.
3294          *
3295          * We could call write_cache_pages(), and then redirty all of
3296          * the pages by calling redirty_page_for_writeback() but that
3297          * would be ugly in the extreme.  So instead we would need to
3298          * replicate parts of the code in the above functions,
3299          * simplifying them becuase we wouldn't actually intend to
3300          * write out the pages, but rather only collect contiguous
3301          * logical block extents, call the multi-block allocator, and
3302          * then update the buffer heads with the block allocations.
3303          *
3304          * For now, though, we'll cheat by calling filemap_flush(),
3305          * which will map the blocks, and start the I/O, but not
3306          * actually wait for the I/O to complete.
3307          */
3308         return filemap_flush(inode->i_mapping);
3309 }
3310
3311 /*
3312  * bmap() is special.  It gets used by applications such as lilo and by
3313  * the swapper to find the on-disk block of a specific piece of data.
3314  *
3315  * Naturally, this is dangerous if the block concerned is still in the
3316  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3317  * filesystem and enables swap, then they may get a nasty shock when the
3318  * data getting swapped to that swapfile suddenly gets overwritten by
3319  * the original zero's written out previously to the journal and
3320  * awaiting writeback in the kernel's buffer cache.
3321  *
3322  * So, if we see any bmap calls here on a modified, data-journaled file,
3323  * take extra steps to flush any blocks which might be in the cache.
3324  */
3325 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3326 {
3327         struct inode *inode = mapping->host;
3328         journal_t *journal;
3329         int err;
3330
3331         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3332                         test_opt(inode->i_sb, DELALLOC)) {
3333                 /*
3334                  * With delalloc we want to sync the file
3335                  * so that we can make sure we allocate
3336                  * blocks for file
3337                  */
3338                 filemap_write_and_wait(mapping);
3339         }
3340
3341         if (EXT4_JOURNAL(inode) &&
3342             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3343                 /*
3344                  * This is a REALLY heavyweight approach, but the use of
3345                  * bmap on dirty files is expected to be extremely rare:
3346                  * only if we run lilo or swapon on a freshly made file
3347                  * do we expect this to happen.
3348                  *
3349                  * (bmap requires CAP_SYS_RAWIO so this does not
3350                  * represent an unprivileged user DOS attack --- we'd be
3351                  * in trouble if mortal users could trigger this path at
3352                  * will.)
3353                  *
3354                  * NB. EXT4_STATE_JDATA is not set on files other than
3355                  * regular files.  If somebody wants to bmap a directory
3356                  * or symlink and gets confused because the buffer
3357                  * hasn't yet been flushed to disk, they deserve
3358                  * everything they get.
3359                  */
3360
3361                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3362                 journal = EXT4_JOURNAL(inode);
3363                 jbd2_journal_lock_updates(journal);
3364                 err = jbd2_journal_flush(journal);
3365                 jbd2_journal_unlock_updates(journal);
3366
3367                 if (err)
3368                         return 0;
3369         }
3370
3371         return generic_block_bmap(mapping, block, ext4_get_block);
3372 }
3373
3374 static int ext4_readpage(struct file *file, struct page *page)
3375 {
3376         return mpage_readpage(page, ext4_get_block);
3377 }
3378
3379 static int
3380 ext4_readpages(struct file *file, struct address_space *mapping,
3381                 struct list_head *pages, unsigned nr_pages)
3382 {
3383         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3384 }
3385
3386 static void ext4_free_io_end(ext4_io_end_t *io)
3387 {
3388         BUG_ON(!io);
3389         if (io->page)
3390                 put_page(io->page);
3391         iput(io->inode);
3392         kfree(io);
3393 }
3394
3395 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3396 {
3397         struct buffer_head *head, *bh;
3398         unsigned int curr_off = 0;
3399
3400         if (!page_has_buffers(page))
3401                 return;
3402         head = bh = page_buffers(page);
3403         do {
3404                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3405                                         && bh->b_private) {
3406                         ext4_free_io_end(bh->b_private);
3407                         bh->b_private = NULL;
3408                         bh->b_end_io = NULL;
3409                 }
3410                 curr_off = curr_off + bh->b_size;
3411                 bh = bh->b_this_page;
3412         } while (bh != head);
3413 }
3414
3415 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3416 {
3417         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3418
3419         /*
3420          * free any io_end structure allocated for buffers to be discarded
3421          */
3422         if (ext4_should_dioread_nolock(page->mapping->host))
3423                 ext4_invalidatepage_free_endio(page, offset);
3424         /*
3425          * If it's a full truncate we just forget about the pending dirtying
3426          */
3427         if (offset == 0)
3428                 ClearPageChecked(page);
3429
3430         if (journal)
3431                 jbd2_journal_invalidatepage(journal, page, offset);
3432         else
3433                 block_invalidatepage(page, offset);
3434 }
3435
3436 static int ext4_releasepage(struct page *page, gfp_t wait)
3437 {
3438         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3439
3440         WARN_ON(PageChecked(page));
3441         if (!page_has_buffers(page))
3442                 return 0;
3443         if (journal)
3444                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3445         else
3446                 return try_to_free_buffers(page);
3447 }
3448
3449 /*
3450  * O_DIRECT for ext3 (or indirect map) based files
3451  *
3452  * If the O_DIRECT write will extend the file then add this inode to the
3453  * orphan list.  So recovery will truncate it back to the original size
3454  * if the machine crashes during the write.
3455  *
3456  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3457  * crashes then stale disk data _may_ be exposed inside the file. But current
3458  * VFS code falls back into buffered path in that case so we are safe.
3459  */
3460 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3461                               const struct iovec *iov, loff_t offset,
3462                               unsigned long nr_segs)
3463 {
3464         struct file *file = iocb->ki_filp;
3465         struct inode *inode = file->f_mapping->host;
3466         struct ext4_inode_info *ei = EXT4_I(inode);
3467         handle_t *handle;
3468         ssize_t ret;
3469         int orphan = 0;
3470         size_t count = iov_length(iov, nr_segs);
3471         int retries = 0;
3472
3473         if (rw == WRITE) {
3474                 loff_t final_size = offset + count;
3475
3476                 if (final_size > inode->i_size) {
3477                         /* Credits for sb + inode write */
3478                         handle = ext4_journal_start(inode, 2);
3479                         if (IS_ERR(handle)) {
3480                                 ret = PTR_ERR(handle);
3481                                 goto out;
3482                         }
3483                         ret = ext4_orphan_add(handle, inode);
3484                         if (ret) {
3485                                 ext4_journal_stop(handle);
3486                                 goto out;
3487                         }
3488                         orphan = 1;
3489                         ei->i_disksize = inode->i_size;
3490                         ext4_journal_stop(handle);
3491                 }
3492         }
3493
3494 retry:
3495         if (rw == READ && ext4_should_dioread_nolock(inode))
3496                 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
3497                                  inode->i_sb->s_bdev, iov,
3498                                  offset, nr_segs,
3499                                  ext4_get_block, NULL);
3500         else
3501                 ret = blockdev_direct_IO(rw, iocb, inode,
3502                                  inode->i_sb->s_bdev, iov,
3503                                  offset, nr_segs,
3504                                  ext4_get_block, NULL);
3505         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3506                 goto retry;
3507
3508         if (orphan) {
3509                 int err;
3510
3511                 /* Credits for sb + inode write */
3512                 handle = ext4_journal_start(inode, 2);
3513                 if (IS_ERR(handle)) {
3514                         /* This is really bad luck. We've written the data
3515                          * but cannot extend i_size. Bail out and pretend
3516                          * the write failed... */
3517                         ret = PTR_ERR(handle);
3518                         if (inode->i_nlink)
3519                                 ext4_orphan_del(NULL, inode);
3520
3521                         goto out;
3522                 }
3523                 if (inode->i_nlink)
3524                         ext4_orphan_del(handle, inode);
3525                 if (ret > 0) {
3526                         loff_t end = offset + ret;
3527                         if (end > inode->i_size) {
3528                                 ei->i_disksize = end;
3529                                 i_size_write(inode, end);
3530                                 /*
3531                                  * We're going to return a positive `ret'
3532                                  * here due to non-zero-length I/O, so there's
3533                                  * no way of reporting error returns from
3534                                  * ext4_mark_inode_dirty() to userspace.  So
3535                                  * ignore it.
3536                                  */
3537                                 ext4_mark_inode_dirty(handle, inode);
3538                         }
3539                 }
3540                 err = ext4_journal_stop(handle);
3541                 if (ret == 0)
3542                         ret = err;
3543         }
3544 out:
3545         return ret;
3546 }
3547
3548 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3549                    struct buffer_head *bh_result, int create)
3550 {
3551         handle_t *handle = ext4_journal_current_handle();
3552         int ret = 0;
3553         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3554         int dio_credits;
3555         int started = 0;
3556
3557         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3558                    inode->i_ino, create);
3559         /*
3560          * ext4_get_block in prepare for a DIO write or buffer write.
3561          * We allocate an uinitialized extent if blocks haven't been allocated.
3562          * The extent will be converted to initialized after IO complete.
3563          */
3564         create = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3565
3566         if (!handle) {
3567                 if (max_blocks > DIO_MAX_BLOCKS)
3568                         max_blocks = DIO_MAX_BLOCKS;
3569                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3570                 handle = ext4_journal_start(inode, dio_credits);
3571                 if (IS_ERR(handle)) {
3572                         ret = PTR_ERR(handle);
3573                         goto out;
3574                 }
3575                 started = 1;
3576         }
3577
3578         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3579                               create);
3580         if (ret > 0) {
3581                 bh_result->b_size = (ret << inode->i_blkbits);
3582                 ret = 0;
3583         }
3584         if (started)
3585                 ext4_journal_stop(handle);
3586 out:
3587         return ret;
3588 }
3589
3590 static void dump_completed_IO(struct inode * inode)
3591 {
3592 #ifdef  EXT4_DEBUG
3593         struct list_head *cur, *before, *after;
3594         ext4_io_end_t *io, *io0, *io1;
3595         unsigned long flags;
3596
3597         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3598                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3599                 return;
3600         }
3601
3602         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3603         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3604         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3605                 cur = &io->list;
3606                 before = cur->prev;
3607                 io0 = container_of(before, ext4_io_end_t, list);
3608                 after = cur->next;
3609                 io1 = container_of(after, ext4_io_end_t, list);
3610
3611                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3612                             io, inode->i_ino, io0, io1);
3613         }
3614         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3615 #endif
3616 }
3617
3618 /*
3619  * check a range of space and convert unwritten extents to written.
3620  */
3621 static int ext4_end_io_nolock(ext4_io_end_t *io)
3622 {
3623         struct inode *inode = io->inode;
3624         loff_t offset = io->offset;
3625         ssize_t size = io->size;
3626         int ret = 0;
3627
3628         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3629                    "list->prev 0x%p\n",
3630                    io, inode->i_ino, io->list.next, io->list.prev);
3631
3632         if (list_empty(&io->list))
3633                 return ret;
3634
3635         if (io->flag != EXT4_IO_UNWRITTEN)
3636                 return ret;
3637
3638         ret = ext4_convert_unwritten_extents(inode, offset, size);
3639         if (ret < 0) {
3640                 printk(KERN_EMERG "%s: failed to convert unwritten"
3641                         "extents to written extents, error is %d"
3642                         " io is still on inode %lu aio dio list\n",
3643                        __func__, ret, inode->i_ino);
3644                 return ret;
3645         }
3646
3647         /* clear the DIO AIO unwritten flag */
3648         io->flag = 0;
3649         return ret;
3650 }
3651
3652 /*
3653  * work on completed aio dio IO, to convert unwritten extents to extents
3654  */
3655 static void ext4_end_io_work(struct work_struct *work)
3656 {
3657         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3658         struct inode            *inode = io->inode;
3659         struct ext4_inode_info  *ei = EXT4_I(inode);
3660         unsigned long           flags;
3661         int                     ret;
3662
3663         mutex_lock(&inode->i_mutex);
3664         ret = ext4_end_io_nolock(io);
3665         if (ret < 0) {
3666                 mutex_unlock(&inode->i_mutex);
3667                 return;
3668         }
3669
3670         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3671         if (!list_empty(&io->list))
3672                 list_del_init(&io->list);
3673         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3674         mutex_unlock(&inode->i_mutex);
3675         ext4_free_io_end(io);
3676 }
3677
3678 /*
3679  * This function is called from ext4_sync_file().
3680  *
3681  * When IO is completed, the work to convert unwritten extents to
3682  * written is queued on workqueue but may not get immediately
3683  * scheduled. When fsync is called, we need to ensure the
3684  * conversion is complete before fsync returns.
3685  * The inode keeps track of a list of pending/completed IO that
3686  * might needs to do the conversion. This function walks through
3687  * the list and convert the related unwritten extents for completed IO
3688  * to written.
3689  * The function return the number of pending IOs on success.
3690  */
3691 int flush_completed_IO(struct inode *inode)
3692 {
3693         ext4_io_end_t *io;
3694         struct ext4_inode_info *ei = EXT4_I(inode);
3695         unsigned long flags;
3696         int ret = 0;
3697         int ret2 = 0;
3698
3699         if (list_empty(&ei->i_completed_io_list))
3700                 return ret;
3701
3702         dump_completed_IO(inode);
3703         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3704         while (!list_empty(&ei->i_completed_io_list)){
3705                 io = list_entry(ei->i_completed_io_list.next,
3706                                 ext4_io_end_t, list);
3707                 /*
3708                  * Calling ext4_end_io_nolock() to convert completed
3709                  * IO to written.
3710                  *
3711                  * When ext4_sync_file() is called, run_queue() may already
3712                  * about to flush the work corresponding to this io structure.
3713                  * It will be upset if it founds the io structure related
3714                  * to the work-to-be schedule is freed.
3715                  *
3716                  * Thus we need to keep the io structure still valid here after
3717                  * convertion finished. The io structure has a flag to
3718                  * avoid double converting from both fsync and background work
3719                  * queue work.
3720                  */
3721                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3722                 ret = ext4_end_io_nolock(io);
3723                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3724                 if (ret < 0)
3725                         ret2 = ret;
3726                 else
3727                         list_del_init(&io->list);
3728         }
3729         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3730         return (ret2 < 0) ? ret2 : 0;
3731 }
3732
3733 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3734 {
3735         ext4_io_end_t *io = NULL;
3736
3737         io = kmalloc(sizeof(*io), flags);
3738
3739         if (io) {
3740                 igrab(inode);
3741                 io->inode = inode;
3742                 io->flag = 0;
3743                 io->offset = 0;
3744                 io->size = 0;
3745                 io->page = NULL;
3746                 INIT_WORK(&io->work, ext4_end_io_work);
3747                 INIT_LIST_HEAD(&io->list);
3748         }
3749
3750         return io;
3751 }
3752
3753 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3754                             ssize_t size, void *private)
3755 {
3756         ext4_io_end_t *io_end = iocb->private;
3757         struct workqueue_struct *wq;
3758         unsigned long flags;
3759         struct ext4_inode_info *ei;
3760
3761         /* if not async direct IO or dio with 0 bytes write, just return */
3762         if (!io_end || !size)
3763                 return;
3764
3765         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3766                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3767                   iocb->private, io_end->inode->i_ino, iocb, offset,
3768                   size);
3769
3770         /* if not aio dio with unwritten extents, just free io and return */
3771         if (io_end->flag != EXT4_IO_UNWRITTEN){
3772                 ext4_free_io_end(io_end);
3773                 iocb->private = NULL;
3774                 return;
3775         }
3776
3777         io_end->offset = offset;
3778         io_end->size = size;
3779         io_end->flag = EXT4_IO_UNWRITTEN;
3780         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3781
3782         /* queue the work to convert unwritten extents to written */
3783         queue_work(wq, &io_end->work);
3784
3785         /* Add the io_end to per-inode completed aio dio list*/
3786         ei = EXT4_I(io_end->inode);
3787         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3788         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3789         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3790         iocb->private = NULL;
3791 }
3792
3793 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3794 {
3795         ext4_io_end_t *io_end = bh->b_private;
3796         struct workqueue_struct *wq;
3797         struct inode *inode;
3798         unsigned long flags;
3799
3800         if (!test_clear_buffer_uninit(bh) || !io_end)
3801                 goto out;
3802
3803         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3804                 printk("sb umounted, discard end_io request for inode %lu\n",
3805                         io_end->inode->i_ino);
3806                 ext4_free_io_end(io_end);
3807                 goto out;
3808         }
3809
3810         io_end->flag = EXT4_IO_UNWRITTEN;
3811         inode = io_end->inode;
3812
3813         /* Add the io_end to per-inode completed io list*/
3814         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3815         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3816         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3817
3818         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3819         /* queue the work to convert unwritten extents to written */
3820         queue_work(wq, &io_end->work);
3821 out:
3822         bh->b_private = NULL;
3823         bh->b_end_io = NULL;
3824         clear_buffer_uninit(bh);
3825         end_buffer_async_write(bh, uptodate);
3826 }
3827
3828 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3829 {
3830         ext4_io_end_t *io_end;
3831         struct page *page = bh->b_page;
3832         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3833         size_t size = bh->b_size;
3834
3835 retry:
3836         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3837         if (!io_end) {
3838                 if (printk_ratelimit())
3839                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3840                 schedule();
3841                 goto retry;
3842         }
3843         io_end->offset = offset;
3844         io_end->size = size;
3845         /*
3846          * We need to hold a reference to the page to make sure it
3847          * doesn't get evicted before ext4_end_io_work() has a chance
3848          * to convert the extent from written to unwritten.
3849          */
3850         io_end->page = page;
3851         get_page(io_end->page);
3852
3853         bh->b_private = io_end;
3854         bh->b_end_io = ext4_end_io_buffer_write;
3855         return 0;
3856 }
3857
3858 /*
3859  * For ext4 extent files, ext4 will do direct-io write to holes,
3860  * preallocated extents, and those write extend the file, no need to
3861  * fall back to buffered IO.
3862  *
3863  * For holes, we fallocate those blocks, mark them as unintialized
3864  * If those blocks were preallocated, we mark sure they are splited, but
3865  * still keep the range to write as unintialized.
3866  *
3867  * The unwrritten extents will be converted to written when DIO is completed.
3868  * For async direct IO, since the IO may still pending when return, we
3869  * set up an end_io call back function, which will do the convertion
3870  * when async direct IO completed.
3871  *
3872  * If the O_DIRECT write will extend the file then add this inode to the
3873  * orphan list.  So recovery will truncate it back to the original size
3874  * if the machine crashes during the write.
3875  *
3876  */
3877 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3878                               const struct iovec *iov, loff_t offset,
3879                               unsigned long nr_segs)
3880 {
3881         struct file *file = iocb->ki_filp;
3882         struct inode *inode = file->f_mapping->host;
3883         ssize_t ret;
3884         size_t count = iov_length(iov, nr_segs);
3885
3886         loff_t final_size = offset + count;
3887         if (rw == WRITE && final_size <= inode->i_size) {
3888                 /*
3889                  * We could direct write to holes and fallocate.
3890                  *
3891                  * Allocated blocks to fill the hole are marked as uninitialized
3892                  * to prevent paralel buffered read to expose the stale data
3893                  * before DIO complete the data IO.
3894                  *
3895                  * As to previously fallocated extents, ext4 get_block
3896                  * will just simply mark the buffer mapped but still
3897                  * keep the extents uninitialized.
3898                  *
3899                  * for non AIO case, we will convert those unwritten extents
3900                  * to written after return back from blockdev_direct_IO.
3901                  *
3902                  * for async DIO, the conversion needs to be defered when
3903                  * the IO is completed. The ext4 end_io callback function
3904                  * will be called to take care of the conversion work.
3905                  * Here for async case, we allocate an io_end structure to
3906                  * hook to the iocb.
3907                  */
3908                 iocb->private = NULL;
3909                 EXT4_I(inode)->cur_aio_dio = NULL;
3910                 if (!is_sync_kiocb(iocb)) {
3911                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3912                         if (!iocb->private)
3913                                 return -ENOMEM;
3914                         /*
3915                          * we save the io structure for current async
3916                          * direct IO, so that later ext4_get_blocks()
3917                          * could flag the io structure whether there
3918                          * is a unwritten extents needs to be converted
3919                          * when IO is completed.
3920                          */
3921                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3922                 }
3923
3924                 ret = blockdev_direct_IO(rw, iocb, inode,
3925                                          inode->i_sb->s_bdev, iov,
3926                                          offset, nr_segs,
3927                                          ext4_get_block_write,
3928                                          ext4_end_io_dio);
3929                 if (iocb->private)
3930                         EXT4_I(inode)->cur_aio_dio = NULL;
3931                 /*
3932                  * The io_end structure takes a reference to the inode,
3933                  * that structure needs to be destroyed and the
3934                  * reference to the inode need to be dropped, when IO is
3935                  * complete, even with 0 byte write, or failed.
3936                  *
3937                  * In the successful AIO DIO case, the io_end structure will be
3938                  * desctroyed and the reference to the inode will be dropped
3939                  * after the end_io call back function is called.
3940                  *
3941                  * In the case there is 0 byte write, or error case, since
3942                  * VFS direct IO won't invoke the end_io call back function,
3943                  * we need to free the end_io structure here.
3944                  */
3945                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3946                         ext4_free_io_end(iocb->private);
3947                         iocb->private = NULL;
3948                 } else if (ret > 0 && ext4_test_inode_state(inode,
3949                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3950                         int err;
3951                         /*
3952                          * for non AIO case, since the IO is already
3953                          * completed, we could do the convertion right here
3954                          */
3955                         err = ext4_convert_unwritten_extents(inode,
3956                                                              offset, ret);
3957                         if (err < 0)
3958                                 ret = err;
3959                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3960                 }
3961                 return ret;
3962         }
3963
3964         /* for write the the end of file case, we fall back to old way */
3965         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3966 }
3967
3968 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3969                               const struct iovec *iov, loff_t offset,
3970                               unsigned long nr_segs)
3971 {
3972         struct file *file = iocb->ki_filp;
3973         struct inode *inode = file->f_mapping->host;
3974
3975         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3976                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3977
3978         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3979 }
3980
3981 /*
3982  * Pages can be marked dirty completely asynchronously from ext4's journalling
3983  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3984  * much here because ->set_page_dirty is called under VFS locks.  The page is
3985  * not necessarily locked.
3986  *
3987  * We cannot just dirty the page and leave attached buffers clean, because the
3988  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3989  * or jbddirty because all the journalling code will explode.
3990  *
3991  * So what we do is to mark the page "pending dirty" and next time writepage
3992  * is called, propagate that into the buffers appropriately.
3993  */
3994 static int ext4_journalled_set_page_dirty(struct page *page)
3995 {
3996         SetPageChecked(page);
3997         return __set_page_dirty_nobuffers(page);
3998 }
3999
4000 static const struct address_space_operations ext4_ordered_aops = {
4001         .readpage               = ext4_readpage,
4002         .readpages              = ext4_readpages,
4003         .writepage              = ext4_writepage,
4004         .sync_page              = block_sync_page,
4005         .write_begin            = ext4_write_begin,
4006         .write_end              = ext4_ordered_write_end,
4007         .bmap                   = ext4_bmap,
4008         .invalidatepage         = ext4_invalidatepage,
4009         .releasepage            = ext4_releasepage,
4010         .direct_IO              = ext4_direct_IO,
4011         .migratepage            = buffer_migrate_page,
4012         .is_partially_uptodate  = block_is_partially_uptodate,
4013         .error_remove_page      = generic_error_remove_page,
4014 };
4015
4016 static const struct address_space_operations ext4_writeback_aops = {
4017         .readpage               = ext4_readpage,
4018         .readpages              = ext4_readpages,
4019         .writepage              = ext4_writepage,
4020         .sync_page              = block_sync_page,
4021         .write_begin            = ext4_write_begin,
4022         .write_end              = ext4_writeback_write_end,
4023         .bmap                   = ext4_bmap,
4024         .invalidatepage         = ext4_invalidatepage,
4025         .releasepage            = ext4_releasepage,
4026         .direct_IO              = ext4_direct_IO,
4027         .migratepage            = buffer_migrate_page,
4028         .is_partially_uptodate  = block_is_partially_uptodate,
4029         .error_remove_page      = generic_error_remove_page,
4030 };
4031
4032 static const struct address_space_operations ext4_journalled_aops = {
4033         .readpage               = ext4_readpage,
4034         .readpages              = ext4_readpages,
4035         .writepage              = ext4_writepage,
4036         .sync_page              = block_sync_page,
4037         .write_begin            = ext4_write_begin,
4038         .write_end              = ext4_journalled_write_end,
4039         .set_page_dirty         = ext4_journalled_set_page_dirty,
4040         .bmap                   = ext4_bmap,
4041         .invalidatepage         = ext4_invalidatepage,
4042         .releasepage            = ext4_releasepage,
4043         .is_partially_uptodate  = block_is_partially_uptodate,
4044         .error_remove_page      = generic_error_remove_page,
4045 };
4046
4047 static const struct address_space_operations ext4_da_aops = {
4048         .readpage               = ext4_readpage,
4049         .readpages              = ext4_readpages,
4050         .writepage              = ext4_writepage,
4051         .writepages             = ext4_da_writepages,
4052         .sync_page              = block_sync_page,
4053         .write_begin            = ext4_da_write_begin,
4054         .write_end              = ext4_da_write_end,
4055         .bmap                   = ext4_bmap,
4056         .invalidatepage         = ext4_da_invalidatepage,
4057         .releasepage            = ext4_releasepage,
4058         .direct_IO              = ext4_direct_IO,
4059         .migratepage            = buffer_migrate_page,
4060         .is_partially_uptodate  = block_is_partially_uptodate,
4061         .error_remove_page      = generic_error_remove_page,
4062 };
4063
4064 void ext4_set_aops(struct inode *inode)
4065 {
4066         if (ext4_should_order_data(inode) &&
4067                 test_opt(inode->i_sb, DELALLOC))
4068                 inode->i_mapping->a_ops = &ext4_da_aops;
4069         else if (ext4_should_order_data(inode))
4070                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4071         else if (ext4_should_writeback_data(inode) &&
4072                  test_opt(inode->i_sb, DELALLOC))
4073                 inode->i_mapping->a_ops = &ext4_da_aops;
4074         else if (ext4_should_writeback_data(inode))
4075                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4076         else
4077                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4078 }
4079
4080 /*
4081  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4082  * up to the end of the block which corresponds to `from'.
4083  * This required during truncate. We need to physically zero the tail end
4084  * of that block so it doesn't yield old data if the file is later grown.
4085  */
4086 int ext4_block_truncate_page(handle_t *handle,
4087                 struct address_space *mapping, loff_t from)
4088 {
4089         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4090         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4091         unsigned blocksize, length, pos;
4092         ext4_lblk_t iblock;
4093         struct inode *inode = mapping->host;
4094         struct buffer_head *bh;
4095         struct page *page;
4096         int err = 0;
4097
4098         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4099                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4100         if (!page)
4101                 return -EINVAL;
4102
4103         blocksize = inode->i_sb->s_blocksize;
4104         length = blocksize - (offset & (blocksize - 1));
4105         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4106
4107         /*
4108          * For "nobh" option,  we can only work if we don't need to
4109          * read-in the page - otherwise we create buffers to do the IO.
4110          */
4111         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
4112              ext4_should_writeback_data(inode) && PageUptodate(page)) {
4113                 zero_user(page, offset, length);
4114                 set_page_dirty(page);
4115                 goto unlock;
4116         }
4117
4118         if (!page_has_buffers(page))
4119                 create_empty_buffers(page, blocksize, 0);
4120
4121         /* Find the buffer that contains "offset" */
4122         bh = page_buffers(page);
4123         pos = blocksize;
4124         while (offset >= pos) {
4125                 bh = bh->b_this_page;
4126                 iblock++;
4127                 pos += blocksize;
4128         }
4129
4130         err = 0;
4131         if (buffer_freed(bh)) {
4132                 BUFFER_TRACE(bh, "freed: skip");
4133                 goto unlock;
4134         }
4135
4136         if (!buffer_mapped(bh)) {
4137                 BUFFER_TRACE(bh, "unmapped");
4138                 ext4_get_block(inode, iblock, bh, 0);
4139                 /* unmapped? It's a hole - nothing to do */
4140                 if (!buffer_mapped(bh)) {
4141                         BUFFER_TRACE(bh, "still unmapped");
4142                         goto unlock;
4143                 }
4144         }
4145
4146         /* Ok, it's mapped. Make sure it's up-to-date */
4147         if (PageUptodate(page))
4148                 set_buffer_uptodate(bh);
4149
4150         if (!buffer_uptodate(bh)) {
4151                 err = -EIO;
4152                 ll_rw_block(READ, 1, &bh);
4153                 wait_on_buffer(bh);
4154                 /* Uhhuh. Read error. Complain and punt. */
4155                 if (!buffer_uptodate(bh))
4156                         goto unlock;
4157         }
4158
4159         if (ext4_should_journal_data(inode)) {
4160                 BUFFER_TRACE(bh, "get write access");
4161                 err = ext4_journal_get_write_access(handle, bh);
4162                 if (err)
4163                         goto unlock;
4164         }
4165
4166         zero_user(page, offset, length);
4167
4168         BUFFER_TRACE(bh, "zeroed end of block");
4169
4170         err = 0;
4171         if (ext4_should_journal_data(inode)) {
4172                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4173         } else {
4174                 if (ext4_should_order_data(inode))
4175                         err = ext4_jbd2_file_inode(handle, inode);
4176                 mark_buffer_dirty(bh);
4177         }
4178
4179 unlock:
4180         unlock_page(page);
4181         page_cache_release(page);
4182         return err;
4183 }
4184
4185 /*
4186  * Probably it should be a library function... search for first non-zero word
4187  * or memcmp with zero_page, whatever is better for particular architecture.
4188  * Linus?
4189  */
4190 static inline int all_zeroes(__le32 *p, __le32 *q)
4191 {
4192         while (p < q)
4193                 if (*p++)
4194                         return 0;
4195         return 1;
4196 }
4197
4198 /**
4199  *      ext4_find_shared - find the indirect blocks for partial truncation.
4200  *      @inode:   inode in question
4201  *      @depth:   depth of the affected branch
4202  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4203  *      @chain:   place to store the pointers to partial indirect blocks
4204  *      @top:     place to the (detached) top of branch
4205  *
4206  *      This is a helper function used by ext4_truncate().
4207  *
4208  *      When we do truncate() we may have to clean the ends of several
4209  *      indirect blocks but leave the blocks themselves alive. Block is
4210  *      partially truncated if some data below the new i_size is refered
4211  *      from it (and it is on the path to the first completely truncated
4212  *      data block, indeed).  We have to free the top of that path along
4213  *      with everything to the right of the path. Since no allocation
4214  *      past the truncation point is possible until ext4_truncate()
4215  *      finishes, we may safely do the latter, but top of branch may
4216  *      require special attention - pageout below the truncation point
4217  *      might try to populate it.
4218  *
4219  *      We atomically detach the top of branch from the tree, store the
4220  *      block number of its root in *@top, pointers to buffer_heads of
4221  *      partially truncated blocks - in @chain[].bh and pointers to
4222  *      their last elements that should not be removed - in
4223  *      @chain[].p. Return value is the pointer to last filled element
4224  *      of @chain.
4225  *
4226  *      The work left to caller to do the actual freeing of subtrees:
4227  *              a) free the subtree starting from *@top
4228  *              b) free the subtrees whose roots are stored in
4229  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4230  *              c) free the subtrees growing from the inode past the @chain[0].
4231  *                      (no partially truncated stuff there).  */
4232
4233 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4234                                   ext4_lblk_t offsets[4], Indirect chain[4],
4235                                   __le32 *top)
4236 {
4237         Indirect *partial, *p;
4238         int k, err;
4239
4240         *top = 0;
4241         /* Make k index the deepest non-null offset + 1 */
4242         for (k = depth; k > 1 && !offsets[k-1]; k--)
4243                 ;
4244         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4245         /* Writer: pointers */
4246         if (!partial)
4247                 partial = chain + k-1;
4248         /*
4249          * If the branch acquired continuation since we've looked at it -
4250          * fine, it should all survive and (new) top doesn't belong to us.
4251          */
4252         if (!partial->key && *partial->p)
4253                 /* Writer: end */
4254                 goto no_top;
4255         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4256                 ;
4257         /*
4258          * OK, we've found the last block that must survive. The rest of our
4259          * branch should be detached before unlocking. However, if that rest
4260          * of branch is all ours and does not grow immediately from the inode
4261          * it's easier to cheat and just decrement partial->p.
4262          */
4263         if (p == chain + k - 1 && p > chain) {
4264                 p->p--;
4265         } else {
4266                 *top = *p->p;
4267                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4268 #if 0
4269                 *p->p = 0;
4270 #endif
4271         }
4272         /* Writer: end */
4273
4274         while (partial > p) {
4275                 brelse(partial->bh);
4276                 partial--;
4277         }
4278 no_top:
4279         return partial;
4280 }
4281
4282 /*
4283  * Zero a number of block pointers in either an inode or an indirect block.
4284  * If we restart the transaction we must again get write access to the
4285  * indirect block for further modification.
4286  *
4287  * We release `count' blocks on disk, but (last - first) may be greater
4288  * than `count' because there can be holes in there.
4289  */
4290 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4291                              struct buffer_head *bh,
4292                              ext4_fsblk_t block_to_free,
4293                              unsigned long count, __le32 *first,
4294                              __le32 *last)
4295 {
4296         __le32 *p;
4297         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4298
4299         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4300                 flags |= EXT4_FREE_BLOCKS_METADATA;
4301
4302         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4303                                    count)) {
4304                 ext4_error(inode->i_sb, "inode #%lu: "
4305                            "attempt to clear blocks %llu len %lu, invalid",
4306                            inode->i_ino, (unsigned long long) block_to_free,
4307                            count);
4308                 return 1;
4309         }
4310
4311         if (try_to_extend_transaction(handle, inode)) {
4312                 if (bh) {
4313                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4314                         ext4_handle_dirty_metadata(handle, inode, bh);
4315                 }
4316                 ext4_mark_inode_dirty(handle, inode);
4317                 ext4_truncate_restart_trans(handle, inode,
4318                                             blocks_for_truncate(inode));
4319                 if (bh) {
4320                         BUFFER_TRACE(bh, "retaking write access");
4321                         ext4_journal_get_write_access(handle, bh);
4322                 }
4323         }
4324
4325         for (p = first; p < last; p++)
4326                 *p = 0;
4327
4328         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4329         return 0;
4330 }
4331
4332 /**
4333  * ext4_free_data - free a list of data blocks
4334  * @handle:     handle for this transaction
4335  * @inode:      inode we are dealing with
4336  * @this_bh:    indirect buffer_head which contains *@first and *@last
4337  * @first:      array of block numbers
4338  * @last:       points immediately past the end of array
4339  *
4340  * We are freeing all blocks refered from that array (numbers are stored as
4341  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4342  *
4343  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4344  * blocks are contiguous then releasing them at one time will only affect one
4345  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4346  * actually use a lot of journal space.
4347  *
4348  * @this_bh will be %NULL if @first and @last point into the inode's direct
4349  * block pointers.
4350  */
4351 static void ext4_free_data(handle_t *handle, struct inode *inode,
4352                            struct buffer_head *this_bh,
4353                            __le32 *first, __le32 *last)
4354 {
4355         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4356         unsigned long count = 0;            /* Number of blocks in the run */
4357         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4358                                                corresponding to
4359                                                block_to_free */
4360         ext4_fsblk_t nr;                    /* Current block # */
4361         __le32 *p;                          /* Pointer into inode/ind
4362                                                for current block */
4363         int err;
4364
4365         if (this_bh) {                          /* For indirect block */
4366                 BUFFER_TRACE(this_bh, "get_write_access");
4367                 err = ext4_journal_get_write_access(handle, this_bh);
4368                 /* Important: if we can't update the indirect pointers
4369                  * to the blocks, we can't free them. */
4370                 if (err)
4371                         return;
4372         }
4373
4374         for (p = first; p < last; p++) {
4375                 nr = le32_to_cpu(*p);
4376                 if (nr) {
4377                         /* accumulate blocks to free if they're contiguous */
4378                         if (count == 0) {
4379                                 block_to_free = nr;
4380                                 block_to_free_p = p;
4381                                 count = 1;
4382                         } else if (nr == block_to_free + count) {
4383                                 count++;
4384                         } else {
4385                                 if (ext4_clear_blocks(handle, inode, this_bh,
4386                                                       block_to_free, count,
4387                                                       block_to_free_p, p))
4388                                         break;
4389                                 block_to_free = nr;
4390                                 block_to_free_p = p;
4391                                 count = 1;
4392                         }
4393                 }
4394         }
4395
4396         if (count > 0)
4397                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4398                                   count, block_to_free_p, p);
4399
4400         if (this_bh) {
4401                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4402
4403                 /*
4404                  * The buffer head should have an attached journal head at this
4405                  * point. However, if the data is corrupted and an indirect
4406                  * block pointed to itself, it would have been detached when
4407                  * the block was cleared. Check for this instead of OOPSing.
4408                  */
4409                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4410                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4411                 else
4412                         ext4_error(inode->i_sb,
4413                                    "circular indirect block detected, "
4414                                    "inode=%lu, block=%llu",
4415                                    inode->i_ino,
4416                                    (unsigned long long) this_bh->b_blocknr);
4417         }
4418 }
4419
4420 /**
4421  *      ext4_free_branches - free an array of branches
4422  *      @handle: JBD handle for this transaction
4423  *      @inode: inode we are dealing with
4424  *      @parent_bh: the buffer_head which contains *@first and *@last
4425  *      @first: array of block numbers
4426  *      @last:  pointer immediately past the end of array
4427  *      @depth: depth of the branches to free
4428  *
4429  *      We are freeing all blocks refered from these branches (numbers are
4430  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4431  *      appropriately.
4432  */
4433 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4434                                struct buffer_head *parent_bh,
4435                                __le32 *first, __le32 *last, int depth)
4436 {
4437         ext4_fsblk_t nr;
4438         __le32 *p;
4439
4440         if (ext4_handle_is_aborted(handle))
4441                 return;
4442
4443         if (depth--) {
4444                 struct buffer_head *bh;
4445                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4446                 p = last;
4447                 while (--p >= first) {
4448                         nr = le32_to_cpu(*p);
4449                         if (!nr)
4450                                 continue;               /* A hole */
4451
4452                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4453                                                    nr, 1)) {
4454                                 ext4_error(inode->i_sb,
4455                                            "indirect mapped block in inode "
4456                                            "#%lu invalid (level %d, blk #%lu)",
4457                                            inode->i_ino, depth,
4458                                            (unsigned long) nr);
4459                                 break;
4460                         }
4461
4462                         /* Go read the buffer for the next level down */
4463                         bh = sb_bread(inode->i_sb, nr);
4464
4465                         /*
4466                          * A read failure? Report error and clear slot
4467                          * (should be rare).
4468                          */
4469                         if (!bh) {
4470                                 ext4_error(inode->i_sb,
4471                                            "Read failure, inode=%lu, block=%llu",
4472                                            inode->i_ino, nr);
4473                                 continue;
4474                         }
4475
4476                         /* This zaps the entire block.  Bottom up. */
4477                         BUFFER_TRACE(bh, "free child branches");
4478                         ext4_free_branches(handle, inode, bh,
4479                                         (__le32 *) bh->b_data,
4480                                         (__le32 *) bh->b_data + addr_per_block,
4481                                         depth);
4482
4483                         /*
4484                          * We've probably journalled the indirect block several
4485                          * times during the truncate.  But it's no longer
4486                          * needed and we now drop it from the transaction via
4487                          * jbd2_journal_revoke().
4488                          *
4489                          * That's easy if it's exclusively part of this
4490                          * transaction.  But if it's part of the committing
4491                          * transaction then jbd2_journal_forget() will simply
4492                          * brelse() it.  That means that if the underlying
4493                          * block is reallocated in ext4_get_block(),
4494                          * unmap_underlying_metadata() will find this block
4495                          * and will try to get rid of it.  damn, damn.
4496                          *
4497                          * If this block has already been committed to the
4498                          * journal, a revoke record will be written.  And
4499                          * revoke records must be emitted *before* clearing
4500                          * this block's bit in the bitmaps.
4501                          */
4502                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4503
4504                         /*
4505                          * Everything below this this pointer has been
4506                          * released.  Now let this top-of-subtree go.
4507                          *
4508                          * We want the freeing of this indirect block to be
4509                          * atomic in the journal with the updating of the
4510                          * bitmap block which owns it.  So make some room in
4511                          * the journal.
4512                          *
4513                          * We zero the parent pointer *after* freeing its
4514                          * pointee in the bitmaps, so if extend_transaction()
4515                          * for some reason fails to put the bitmap changes and
4516                          * the release into the same transaction, recovery
4517                          * will merely complain about releasing a free block,
4518                          * rather than leaking blocks.
4519                          */
4520                         if (ext4_handle_is_aborted(handle))
4521                                 return;
4522                         if (try_to_extend_transaction(handle, inode)) {
4523                                 ext4_mark_inode_dirty(handle, inode);
4524                                 ext4_truncate_restart_trans(handle, inode,
4525                                             blocks_for_truncate(inode));
4526                         }
4527
4528                         ext4_free_blocks(handle, inode, 0, nr, 1,
4529                                          EXT4_FREE_BLOCKS_METADATA);
4530
4531                         if (parent_bh) {
4532                                 /*
4533                                  * The block which we have just freed is
4534                                  * pointed to by an indirect block: journal it
4535                                  */
4536                                 BUFFER_TRACE(parent_bh, "get_write_access");
4537                                 if (!ext4_journal_get_write_access(handle,
4538                                                                    parent_bh)){
4539                                         *p = 0;
4540                                         BUFFER_TRACE(parent_bh,
4541                                         "call ext4_handle_dirty_metadata");
4542                                         ext4_handle_dirty_metadata(handle,
4543                                                                    inode,
4544                                                                    parent_bh);
4545                                 }
4546                         }
4547                 }
4548         } else {
4549                 /* We have reached the bottom of the tree. */
4550                 BUFFER_TRACE(parent_bh, "free data blocks");
4551                 ext4_free_data(handle, inode, parent_bh, first, last);
4552         }
4553 }
4554
4555 int ext4_can_truncate(struct inode *inode)
4556 {
4557         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4558                 return 0;
4559         if (S_ISREG(inode->i_mode))
4560                 return 1;
4561         if (S_ISDIR(inode->i_mode))
4562                 return 1;
4563         if (S_ISLNK(inode->i_mode))
4564                 return !ext4_inode_is_fast_symlink(inode);
4565         return 0;
4566 }
4567
4568 /*
4569  * ext4_truncate()
4570  *
4571  * We block out ext4_get_block() block instantiations across the entire
4572  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4573  * simultaneously on behalf of the same inode.
4574  *
4575  * As we work through the truncate and commmit bits of it to the journal there
4576  * is one core, guiding principle: the file's tree must always be consistent on
4577  * disk.  We must be able to restart the truncate after a crash.
4578  *
4579  * The file's tree may be transiently inconsistent in memory (although it
4580  * probably isn't), but whenever we close off and commit a journal transaction,
4581  * the contents of (the filesystem + the journal) must be consistent and
4582  * restartable.  It's pretty simple, really: bottom up, right to left (although
4583  * left-to-right works OK too).
4584  *
4585  * Note that at recovery time, journal replay occurs *before* the restart of
4586  * truncate against the orphan inode list.
4587  *
4588  * The committed inode has the new, desired i_size (which is the same as
4589  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4590  * that this inode's truncate did not complete and it will again call
4591  * ext4_truncate() to have another go.  So there will be instantiated blocks
4592  * to the right of the truncation point in a crashed ext4 filesystem.  But
4593  * that's fine - as long as they are linked from the inode, the post-crash
4594  * ext4_truncate() run will find them and release them.
4595  */
4596 void ext4_truncate(struct inode *inode)
4597 {
4598         handle_t *handle;
4599         struct ext4_inode_info *ei = EXT4_I(inode);
4600         __le32 *i_data = ei->i_data;
4601         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4602         struct address_space *mapping = inode->i_mapping;
4603         ext4_lblk_t offsets[4];
4604         Indirect chain[4];
4605         Indirect *partial;
4606         __le32 nr = 0;
4607         int n;
4608         ext4_lblk_t last_block;
4609         unsigned blocksize = inode->i_sb->s_blocksize;
4610
4611         if (!ext4_can_truncate(inode))
4612                 return;
4613
4614         EXT4_I(inode)->i_flags &= ~EXT4_EOFBLOCKS_FL;
4615
4616         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4617                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4618
4619         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4620                 ext4_ext_truncate(inode);
4621                 return;
4622         }
4623
4624         handle = start_transaction(inode);
4625         if (IS_ERR(handle))
4626                 return;         /* AKPM: return what? */
4627
4628         last_block = (inode->i_size + blocksize-1)
4629                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4630
4631         if (inode->i_size & (blocksize - 1))
4632                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4633                         goto out_stop;
4634
4635         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4636         if (n == 0)
4637                 goto out_stop;  /* error */
4638
4639         /*
4640          * OK.  This truncate is going to happen.  We add the inode to the
4641          * orphan list, so that if this truncate spans multiple transactions,
4642          * and we crash, we will resume the truncate when the filesystem
4643          * recovers.  It also marks the inode dirty, to catch the new size.
4644          *
4645          * Implication: the file must always be in a sane, consistent
4646          * truncatable state while each transaction commits.
4647          */
4648         if (ext4_orphan_add(handle, inode))
4649                 goto out_stop;
4650
4651         /*
4652          * From here we block out all ext4_get_block() callers who want to
4653          * modify the block allocation tree.
4654          */
4655         down_write(&ei->i_data_sem);
4656
4657         ext4_discard_preallocations(inode);
4658
4659         /*
4660          * The orphan list entry will now protect us from any crash which
4661          * occurs before the truncate completes, so it is now safe to propagate
4662          * the new, shorter inode size (held for now in i_size) into the
4663          * on-disk inode. We do this via i_disksize, which is the value which
4664          * ext4 *really* writes onto the disk inode.
4665          */
4666         ei->i_disksize = inode->i_size;
4667
4668         if (n == 1) {           /* direct blocks */
4669                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4670                                i_data + EXT4_NDIR_BLOCKS);
4671                 goto do_indirects;
4672         }
4673
4674         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4675         /* Kill the top of shared branch (not detached) */
4676         if (nr) {
4677                 if (partial == chain) {
4678                         /* Shared branch grows from the inode */
4679                         ext4_free_branches(handle, inode, NULL,
4680                                            &nr, &nr+1, (chain+n-1) - partial);
4681                         *partial->p = 0;
4682                         /*
4683                          * We mark the inode dirty prior to restart,
4684                          * and prior to stop.  No need for it here.
4685                          */
4686                 } else {
4687                         /* Shared branch grows from an indirect block */
4688                         BUFFER_TRACE(partial->bh, "get_write_access");
4689                         ext4_free_branches(handle, inode, partial->bh,
4690                                         partial->p,
4691                                         partial->p+1, (chain+n-1) - partial);
4692                 }
4693         }
4694         /* Clear the ends of indirect blocks on the shared branch */
4695         while (partial > chain) {
4696                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4697                                    (__le32*)partial->bh->b_data+addr_per_block,
4698                                    (chain+n-1) - partial);
4699                 BUFFER_TRACE(partial->bh, "call brelse");
4700                 brelse(partial->bh);
4701                 partial--;
4702         }
4703 do_indirects:
4704         /* Kill the remaining (whole) subtrees */
4705         switch (offsets[0]) {
4706         default:
4707                 nr = i_data[EXT4_IND_BLOCK];
4708                 if (nr) {
4709                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4710                         i_data[EXT4_IND_BLOCK] = 0;
4711                 }
4712         case EXT4_IND_BLOCK:
4713                 nr = i_data[EXT4_DIND_BLOCK];
4714                 if (nr) {
4715                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4716                         i_data[EXT4_DIND_BLOCK] = 0;
4717                 }
4718         case EXT4_DIND_BLOCK:
4719                 nr = i_data[EXT4_TIND_BLOCK];
4720                 if (nr) {
4721                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4722                         i_data[EXT4_TIND_BLOCK] = 0;
4723                 }
4724         case EXT4_TIND_BLOCK:
4725                 ;
4726         }
4727
4728         up_write(&ei->i_data_sem);
4729         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4730         ext4_mark_inode_dirty(handle, inode);
4731
4732         /*
4733          * In a multi-transaction truncate, we only make the final transaction
4734          * synchronous
4735          */
4736         if (IS_SYNC(inode))
4737                 ext4_handle_sync(handle);
4738 out_stop:
4739         /*
4740          * If this was a simple ftruncate(), and the file will remain alive
4741          * then we need to clear up the orphan record which we created above.
4742          * However, if this was a real unlink then we were called by
4743          * ext4_delete_inode(), and we allow that function to clean up the
4744          * orphan info for us.
4745          */
4746         if (inode->i_nlink)
4747                 ext4_orphan_del(handle, inode);
4748
4749         ext4_journal_stop(handle);
4750 }
4751
4752 /*
4753  * ext4_get_inode_loc returns with an extra refcount against the inode's
4754  * underlying buffer_head on success. If 'in_mem' is true, we have all
4755  * data in memory that is needed to recreate the on-disk version of this
4756  * inode.
4757  */
4758 static int __ext4_get_inode_loc(struct inode *inode,
4759                                 struct ext4_iloc *iloc, int in_mem)
4760 {
4761         struct ext4_group_desc  *gdp;
4762         struct buffer_head      *bh;
4763         struct super_block      *sb = inode->i_sb;
4764         ext4_fsblk_t            block;
4765         int                     inodes_per_block, inode_offset;
4766
4767         iloc->bh = NULL;
4768         if (!ext4_valid_inum(sb, inode->i_ino))
4769                 return -EIO;
4770
4771         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4772         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4773         if (!gdp)
4774                 return -EIO;
4775
4776         /*
4777          * Figure out the offset within the block group inode table
4778          */
4779         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4780         inode_offset = ((inode->i_ino - 1) %
4781                         EXT4_INODES_PER_GROUP(sb));
4782         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4783         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4784
4785         bh = sb_getblk(sb, block);
4786         if (!bh) {
4787                 ext4_error(sb, "unable to read inode block - "
4788                            "inode=%lu, block=%llu", inode->i_ino, block);
4789                 return -EIO;
4790         }
4791         if (!buffer_uptodate(bh)) {
4792                 lock_buffer(bh);
4793
4794                 /*
4795                  * If the buffer has the write error flag, we have failed
4796                  * to write out another inode in the same block.  In this
4797                  * case, we don't have to read the block because we may
4798                  * read the old inode data successfully.
4799                  */
4800                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4801                         set_buffer_uptodate(bh);
4802
4803                 if (buffer_uptodate(bh)) {
4804                         /* someone brought it uptodate while we waited */
4805                         unlock_buffer(bh);
4806                         goto has_buffer;
4807                 }
4808
4809                 /*
4810                  * If we have all information of the inode in memory and this
4811                  * is the only valid inode in the block, we need not read the
4812                  * block.
4813                  */
4814                 if (in_mem) {
4815                         struct buffer_head *bitmap_bh;
4816                         int i, start;
4817
4818                         start = inode_offset & ~(inodes_per_block - 1);
4819
4820                         /* Is the inode bitmap in cache? */
4821                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4822                         if (!bitmap_bh)
4823                                 goto make_io;
4824
4825                         /*
4826                          * If the inode bitmap isn't in cache then the
4827                          * optimisation may end up performing two reads instead
4828                          * of one, so skip it.
4829                          */
4830                         if (!buffer_uptodate(bitmap_bh)) {
4831                                 brelse(bitmap_bh);
4832                                 goto make_io;
4833                         }
4834                         for (i = start; i < start + inodes_per_block; i++) {
4835                                 if (i == inode_offset)
4836                                         continue;
4837                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4838                                         break;
4839                         }
4840                         brelse(bitmap_bh);
4841                         if (i == start + inodes_per_block) {
4842                                 /* all other inodes are free, so skip I/O */
4843                                 memset(bh->b_data, 0, bh->b_size);
4844                                 set_buffer_uptodate(bh);
4845                                 unlock_buffer(bh);
4846                                 goto has_buffer;
4847                         }
4848                 }
4849
4850 make_io:
4851                 /*
4852                  * If we need to do any I/O, try to pre-readahead extra
4853                  * blocks from the inode table.
4854                  */
4855                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4856                         ext4_fsblk_t b, end, table;
4857                         unsigned num;
4858
4859                         table = ext4_inode_table(sb, gdp);
4860                         /* s_inode_readahead_blks is always a power of 2 */
4861                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4862                         if (table > b)
4863                                 b = table;
4864                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4865                         num = EXT4_INODES_PER_GROUP(sb);
4866                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4867                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4868                                 num -= ext4_itable_unused_count(sb, gdp);
4869                         table += num / inodes_per_block;
4870                         if (end > table)
4871                                 end = table;
4872                         while (b <= end)
4873                                 sb_breadahead(sb, b++);
4874                 }
4875
4876                 /*
4877                  * There are other valid inodes in the buffer, this inode
4878                  * has in-inode xattrs, or we don't have this inode in memory.
4879                  * Read the block from disk.
4880                  */
4881                 get_bh(bh);
4882                 bh->b_end_io = end_buffer_read_sync;
4883                 submit_bh(READ_META, bh);
4884                 wait_on_buffer(bh);
4885                 if (!buffer_uptodate(bh)) {
4886                         ext4_error(sb, "unable to read inode block - inode=%lu,"
4887                                    " block=%llu", inode->i_ino, block);
4888                         brelse(bh);
4889                         return -EIO;
4890                 }
4891         }
4892 has_buffer:
4893         iloc->bh = bh;
4894         return 0;
4895 }
4896
4897 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4898 {
4899         /* We have all inode data except xattrs in memory here. */
4900         return __ext4_get_inode_loc(inode, iloc,
4901                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4902 }
4903
4904 void ext4_set_inode_flags(struct inode *inode)
4905 {
4906         unsigned int flags = EXT4_I(inode)->i_flags;
4907
4908         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4909         if (flags & EXT4_SYNC_FL)
4910                 inode->i_flags |= S_SYNC;
4911         if (flags & EXT4_APPEND_FL)
4912                 inode->i_flags |= S_APPEND;
4913         if (flags & EXT4_IMMUTABLE_FL)
4914                 inode->i_flags |= S_IMMUTABLE;
4915         if (flags & EXT4_NOATIME_FL)
4916                 inode->i_flags |= S_NOATIME;
4917         if (flags & EXT4_DIRSYNC_FL)
4918                 inode->i_flags |= S_DIRSYNC;
4919 }
4920
4921 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4922 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4923 {
4924         unsigned int flags = ei->vfs_inode.i_flags;
4925
4926         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4927                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4928         if (flags & S_SYNC)
4929                 ei->i_flags |= EXT4_SYNC_FL;
4930         if (flags & S_APPEND)
4931                 ei->i_flags |= EXT4_APPEND_FL;
4932         if (flags & S_IMMUTABLE)
4933                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4934         if (flags & S_NOATIME)
4935                 ei->i_flags |= EXT4_NOATIME_FL;
4936         if (flags & S_DIRSYNC)
4937                 ei->i_flags |= EXT4_DIRSYNC_FL;
4938 }
4939
4940 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4941                                   struct ext4_inode_info *ei)
4942 {
4943         blkcnt_t i_blocks ;
4944         struct inode *inode = &(ei->vfs_inode);
4945         struct super_block *sb = inode->i_sb;
4946
4947         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4948                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4949                 /* we are using combined 48 bit field */
4950                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4951                                         le32_to_cpu(raw_inode->i_blocks_lo);
4952                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4953                         /* i_blocks represent file system block size */
4954                         return i_blocks  << (inode->i_blkbits - 9);
4955                 } else {
4956                         return i_blocks;
4957                 }
4958         } else {
4959                 return le32_to_cpu(raw_inode->i_blocks_lo);
4960         }
4961 }
4962
4963 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4964 {
4965         struct ext4_iloc iloc;
4966         struct ext4_inode *raw_inode;
4967         struct ext4_inode_info *ei;
4968         struct inode *inode;
4969         journal_t *journal = EXT4_SB(sb)->s_journal;
4970         long ret;
4971         int block;
4972
4973         inode = iget_locked(sb, ino);
4974         if (!inode)
4975                 return ERR_PTR(-ENOMEM);
4976         if (!(inode->i_state & I_NEW))
4977                 return inode;
4978
4979         ei = EXT4_I(inode);
4980         iloc.bh = 0;
4981
4982         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4983         if (ret < 0)
4984                 goto bad_inode;
4985         raw_inode = ext4_raw_inode(&iloc);
4986         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4987         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4988         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4989         if (!(test_opt(inode->i_sb, NO_UID32))) {
4990                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4991                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4992         }
4993         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4994
4995         ei->i_state_flags = 0;
4996         ei->i_dir_start_lookup = 0;
4997         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4998         /* We now have enough fields to check if the inode was active or not.
4999          * This is needed because nfsd might try to access dead inodes
5000          * the test is that same one that e2fsck uses
5001          * NeilBrown 1999oct15
5002          */
5003         if (inode->i_nlink == 0) {
5004                 if (inode->i_mode == 0 ||
5005                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5006                         /* this inode is deleted */
5007                         ret = -ESTALE;
5008                         goto bad_inode;
5009                 }
5010                 /* The only unlinked inodes we let through here have
5011                  * valid i_mode and are being read by the orphan
5012                  * recovery code: that's fine, we're about to complete
5013                  * the process of deleting those. */
5014         }
5015         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5016         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5017         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5018         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5019                 ei->i_file_acl |=
5020                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5021         inode->i_size = ext4_isize(raw_inode);
5022         ei->i_disksize = inode->i_size;
5023 #ifdef CONFIG_QUOTA
5024         ei->i_reserved_quota = 0;
5025 #endif
5026         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5027         ei->i_block_group = iloc.block_group;
5028         ei->i_last_alloc_group = ~0;
5029         /*
5030          * NOTE! The in-memory inode i_data array is in little-endian order
5031          * even on big-endian machines: we do NOT byteswap the block numbers!
5032          */
5033         for (block = 0; block < EXT4_N_BLOCKS; block++)
5034                 ei->i_data[block] = raw_inode->i_block[block];
5035         INIT_LIST_HEAD(&ei->i_orphan);
5036
5037         /*
5038          * Set transaction id's of transactions that have to be committed
5039          * to finish f[data]sync. We set them to currently running transaction
5040          * as we cannot be sure that the inode or some of its metadata isn't
5041          * part of the transaction - the inode could have been reclaimed and
5042          * now it is reread from disk.
5043          */
5044         if (journal) {
5045                 transaction_t *transaction;
5046                 tid_t tid;
5047
5048                 spin_lock(&journal->j_state_lock);
5049                 if (journal->j_running_transaction)
5050                         transaction = journal->j_running_transaction;
5051                 else
5052                         transaction = journal->j_committing_transaction;
5053                 if (transaction)
5054                         tid = transaction->t_tid;
5055                 else
5056                         tid = journal->j_commit_sequence;
5057                 spin_unlock(&journal->j_state_lock);
5058                 ei->i_sync_tid = tid;
5059                 ei->i_datasync_tid = tid;
5060         }
5061
5062         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5063                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5064                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5065                     EXT4_INODE_SIZE(inode->i_sb)) {
5066                         ret = -EIO;
5067                         goto bad_inode;
5068                 }
5069                 if (ei->i_extra_isize == 0) {
5070                         /* The extra space is currently unused. Use it. */
5071                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5072                                             EXT4_GOOD_OLD_INODE_SIZE;
5073                 } else {
5074                         __le32 *magic = (void *)raw_inode +
5075                                         EXT4_GOOD_OLD_INODE_SIZE +
5076                                         ei->i_extra_isize;
5077                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5078                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5079                 }
5080         } else
5081                 ei->i_extra_isize = 0;
5082
5083         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5084         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5085         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5086         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5087
5088         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5089         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5090                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5091                         inode->i_version |=
5092                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5093         }
5094
5095         ret = 0;
5096         if (ei->i_file_acl &&
5097             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5098                 ext4_error(sb, "bad extended attribute block %llu inode #%lu",
5099                            ei->i_file_acl, inode->i_ino);
5100                 ret = -EIO;
5101                 goto bad_inode;
5102         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
5103                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5104                     (S_ISLNK(inode->i_mode) &&
5105                      !ext4_inode_is_fast_symlink(inode)))
5106                         /* Validate extent which is part of inode */
5107                         ret = ext4_ext_check_inode(inode);
5108         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5109                    (S_ISLNK(inode->i_mode) &&
5110                     !ext4_inode_is_fast_symlink(inode))) {
5111                 /* Validate block references which are part of inode */
5112                 ret = ext4_check_inode_blockref(inode);
5113         }
5114         if (ret)
5115                 goto bad_inode;
5116
5117         if (S_ISREG(inode->i_mode)) {
5118                 inode->i_op = &ext4_file_inode_operations;
5119                 inode->i_fop = &ext4_file_operations;
5120                 ext4_set_aops(inode);
5121         } else if (S_ISDIR(inode->i_mode)) {
5122                 inode->i_op = &ext4_dir_inode_operations;
5123                 inode->i_fop = &ext4_dir_operations;
5124         } else if (S_ISLNK(inode->i_mode)) {
5125                 if (ext4_inode_is_fast_symlink(inode)) {
5126                         inode->i_op = &ext4_fast_symlink_inode_operations;
5127                         nd_terminate_link(ei->i_data, inode->i_size,
5128                                 sizeof(ei->i_data) - 1);
5129                 } else {
5130                         inode->i_op = &ext4_symlink_inode_operations;
5131                         ext4_set_aops(inode);
5132                 }
5133         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5134               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5135                 inode->i_op = &ext4_special_inode_operations;
5136                 if (raw_inode->i_block[0])
5137                         init_special_inode(inode, inode->i_mode,
5138                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5139                 else
5140                         init_special_inode(inode, inode->i_mode,
5141                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5142         } else {
5143                 ret = -EIO;
5144                 ext4_error(inode->i_sb, "bogus i_mode (%o) for inode=%lu",
5145                            inode->i_mode, inode->i_ino);
5146                 goto bad_inode;
5147         }
5148         brelse(iloc.bh);
5149         ext4_set_inode_flags(inode);
5150         unlock_new_inode(inode);
5151         return inode;
5152
5153 bad_inode:
5154         brelse(iloc.bh);
5155         iget_failed(inode);
5156         return ERR_PTR(ret);
5157 }
5158
5159 static int ext4_inode_blocks_set(handle_t *handle,
5160                                 struct ext4_inode *raw_inode,
5161                                 struct ext4_inode_info *ei)
5162 {
5163         struct inode *inode = &(ei->vfs_inode);
5164         u64 i_blocks = inode->i_blocks;
5165         struct super_block *sb = inode->i_sb;
5166
5167         if (i_blocks <= ~0U) {
5168                 /*
5169                  * i_blocks can be represnted in a 32 bit variable
5170                  * as multiple of 512 bytes
5171                  */
5172                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5173                 raw_inode->i_blocks_high = 0;
5174                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5175                 return 0;
5176         }
5177         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5178                 return -EFBIG;
5179
5180         if (i_blocks <= 0xffffffffffffULL) {
5181                 /*
5182                  * i_blocks can be represented in a 48 bit variable
5183                  * as multiple of 512 bytes
5184                  */
5185                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5186                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5187                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5188         } else {
5189                 ei->i_flags |= EXT4_HUGE_FILE_FL;
5190                 /* i_block is stored in file system block size */
5191                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5192                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5193                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5194         }
5195         return 0;
5196 }
5197
5198 /*
5199  * Post the struct inode info into an on-disk inode location in the
5200  * buffer-cache.  This gobbles the caller's reference to the
5201  * buffer_head in the inode location struct.
5202  *
5203  * The caller must have write access to iloc->bh.
5204  */
5205 static int ext4_do_update_inode(handle_t *handle,
5206                                 struct inode *inode,
5207                                 struct ext4_iloc *iloc)
5208 {
5209         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5210         struct ext4_inode_info *ei = EXT4_I(inode);
5211         struct buffer_head *bh = iloc->bh;
5212         int err = 0, rc, block;
5213
5214         /* For fields not not tracking in the in-memory inode,
5215          * initialise them to zero for new inodes. */
5216         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5217                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5218
5219         ext4_get_inode_flags(ei);
5220         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5221         if (!(test_opt(inode->i_sb, NO_UID32))) {
5222                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5223                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5224 /*
5225  * Fix up interoperability with old kernels. Otherwise, old inodes get
5226  * re-used with the upper 16 bits of the uid/gid intact
5227  */
5228                 if (!ei->i_dtime) {
5229                         raw_inode->i_uid_high =
5230                                 cpu_to_le16(high_16_bits(inode->i_uid));
5231                         raw_inode->i_gid_high =
5232                                 cpu_to_le16(high_16_bits(inode->i_gid));
5233                 } else {
5234                         raw_inode->i_uid_high = 0;
5235                         raw_inode->i_gid_high = 0;
5236                 }
5237         } else {
5238                 raw_inode->i_uid_low =
5239                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5240                 raw_inode->i_gid_low =
5241                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5242                 raw_inode->i_uid_high = 0;
5243                 raw_inode->i_gid_high = 0;
5244         }
5245         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5246
5247         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5248         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5249         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5250         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5251
5252         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5253                 goto out_brelse;
5254         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5255         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5256         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5257             cpu_to_le32(EXT4_OS_HURD))
5258                 raw_inode->i_file_acl_high =
5259                         cpu_to_le16(ei->i_file_acl >> 32);
5260         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5261         ext4_isize_set(raw_inode, ei->i_disksize);
5262         if (ei->i_disksize > 0x7fffffffULL) {
5263                 struct super_block *sb = inode->i_sb;
5264                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5265                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5266                                 EXT4_SB(sb)->s_es->s_rev_level ==
5267                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5268                         /* If this is the first large file
5269                          * created, add a flag to the superblock.
5270                          */
5271                         err = ext4_journal_get_write_access(handle,
5272                                         EXT4_SB(sb)->s_sbh);
5273                         if (err)
5274                                 goto out_brelse;
5275                         ext4_update_dynamic_rev(sb);
5276                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5277                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5278                         sb->s_dirt = 1;
5279                         ext4_handle_sync(handle);
5280                         err = ext4_handle_dirty_metadata(handle, NULL,
5281                                         EXT4_SB(sb)->s_sbh);
5282                 }
5283         }
5284         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5285         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5286                 if (old_valid_dev(inode->i_rdev)) {
5287                         raw_inode->i_block[0] =
5288                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5289                         raw_inode->i_block[1] = 0;
5290                 } else {
5291                         raw_inode->i_block[0] = 0;
5292                         raw_inode->i_block[1] =
5293                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5294                         raw_inode->i_block[2] = 0;
5295                 }
5296         } else
5297                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5298                         raw_inode->i_block[block] = ei->i_data[block];
5299
5300         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5301         if (ei->i_extra_isize) {
5302                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5303                         raw_inode->i_version_hi =
5304                         cpu_to_le32(inode->i_version >> 32);
5305                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5306         }
5307
5308         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5309         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5310         if (!err)
5311                 err = rc;
5312         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5313
5314         ext4_update_inode_fsync_trans(handle, inode, 0);
5315 out_brelse:
5316         brelse(bh);
5317         ext4_std_error(inode->i_sb, err);
5318         return err;
5319 }
5320
5321 /*
5322  * ext4_write_inode()
5323  *
5324  * We are called from a few places:
5325  *
5326  * - Within generic_file_write() for O_SYNC files.
5327  *   Here, there will be no transaction running. We wait for any running
5328  *   trasnaction to commit.
5329  *
5330  * - Within sys_sync(), kupdate and such.
5331  *   We wait on commit, if tol to.
5332  *
5333  * - Within prune_icache() (PF_MEMALLOC == true)
5334  *   Here we simply return.  We can't afford to block kswapd on the
5335  *   journal commit.
5336  *
5337  * In all cases it is actually safe for us to return without doing anything,
5338  * because the inode has been copied into a raw inode buffer in
5339  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5340  * knfsd.
5341  *
5342  * Note that we are absolutely dependent upon all inode dirtiers doing the
5343  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5344  * which we are interested.
5345  *
5346  * It would be a bug for them to not do this.  The code:
5347  *
5348  *      mark_inode_dirty(inode)
5349  *      stuff();
5350  *      inode->i_size = expr;
5351  *
5352  * is in error because a kswapd-driven write_inode() could occur while
5353  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5354  * will no longer be on the superblock's dirty inode list.
5355  */
5356 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5357 {
5358         int err;
5359
5360         if (current->flags & PF_MEMALLOC)
5361                 return 0;
5362
5363         if (EXT4_SB(inode->i_sb)->s_journal) {
5364                 if (ext4_journal_current_handle()) {
5365                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5366                         dump_stack();
5367                         return -EIO;
5368                 }
5369
5370                 if (wbc->sync_mode != WB_SYNC_ALL)
5371                         return 0;
5372
5373                 err = ext4_force_commit(inode->i_sb);
5374         } else {
5375                 struct ext4_iloc iloc;
5376
5377                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5378                 if (err)
5379                         return err;
5380                 if (wbc->sync_mode == WB_SYNC_ALL)
5381                         sync_dirty_buffer(iloc.bh);
5382                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5383                         ext4_error(inode->i_sb, "IO error syncing inode, "
5384                                    "inode=%lu, block=%llu", inode->i_ino,
5385                                    (unsigned long long)iloc.bh->b_blocknr);
5386                         err = -EIO;
5387                 }
5388         }
5389         return err;
5390 }
5391
5392 /*
5393  * ext4_setattr()
5394  *
5395  * Called from notify_change.
5396  *
5397  * We want to trap VFS attempts to truncate the file as soon as
5398  * possible.  In particular, we want to make sure that when the VFS
5399  * shrinks i_size, we put the inode on the orphan list and modify
5400  * i_disksize immediately, so that during the subsequent flushing of
5401  * dirty pages and freeing of disk blocks, we can guarantee that any
5402  * commit will leave the blocks being flushed in an unused state on
5403  * disk.  (On recovery, the inode will get truncated and the blocks will
5404  * be freed, so we have a strong guarantee that no future commit will
5405  * leave these blocks visible to the user.)
5406  *
5407  * Another thing we have to assure is that if we are in ordered mode
5408  * and inode is still attached to the committing transaction, we must
5409  * we start writeout of all the dirty pages which are being truncated.
5410  * This way we are sure that all the data written in the previous
5411  * transaction are already on disk (truncate waits for pages under
5412  * writeback).
5413  *
5414  * Called with inode->i_mutex down.
5415  */
5416 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5417 {
5418         struct inode *inode = dentry->d_inode;
5419         int error, rc = 0;
5420         const unsigned int ia_valid = attr->ia_valid;
5421
5422         error = inode_change_ok(inode, attr);
5423         if (error)
5424                 return error;
5425
5426         if (ia_valid & ATTR_SIZE)
5427                 dquot_initialize(inode);
5428         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5429                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5430                 handle_t *handle;
5431
5432                 /* (user+group)*(old+new) structure, inode write (sb,
5433                  * inode block, ? - but truncate inode update has it) */
5434                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5435                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5436                 if (IS_ERR(handle)) {
5437                         error = PTR_ERR(handle);
5438                         goto err_out;
5439                 }
5440                 error = dquot_transfer(inode, attr);
5441                 if (error) {
5442                         ext4_journal_stop(handle);
5443                         return error;
5444                 }
5445                 /* Update corresponding info in inode so that everything is in
5446                  * one transaction */
5447                 if (attr->ia_valid & ATTR_UID)
5448                         inode->i_uid = attr->ia_uid;
5449                 if (attr->ia_valid & ATTR_GID)
5450                         inode->i_gid = attr->ia_gid;
5451                 error = ext4_mark_inode_dirty(handle, inode);
5452                 ext4_journal_stop(handle);
5453         }
5454
5455         if (attr->ia_valid & ATTR_SIZE) {
5456                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5457                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5458
5459                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5460                                 error = -EFBIG;
5461                                 goto err_out;
5462                         }
5463                 }
5464         }
5465
5466         if (S_ISREG(inode->i_mode) &&
5467             attr->ia_valid & ATTR_SIZE &&
5468             (attr->ia_size < inode->i_size ||
5469              (EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))) {
5470                 handle_t *handle;
5471
5472                 handle = ext4_journal_start(inode, 3);
5473                 if (IS_ERR(handle)) {
5474                         error = PTR_ERR(handle);
5475                         goto err_out;
5476                 }
5477
5478                 error = ext4_orphan_add(handle, inode);
5479                 EXT4_I(inode)->i_disksize = attr->ia_size;
5480                 rc = ext4_mark_inode_dirty(handle, inode);
5481                 if (!error)
5482                         error = rc;
5483                 ext4_journal_stop(handle);
5484
5485                 if (ext4_should_order_data(inode)) {
5486                         error = ext4_begin_ordered_truncate(inode,
5487                                                             attr->ia_size);
5488                         if (error) {
5489                                 /* Do as much error cleanup as possible */
5490                                 handle = ext4_journal_start(inode, 3);
5491                                 if (IS_ERR(handle)) {
5492                                         ext4_orphan_del(NULL, inode);
5493                                         goto err_out;
5494                                 }
5495                                 ext4_orphan_del(handle, inode);
5496                                 ext4_journal_stop(handle);
5497                                 goto err_out;
5498                         }
5499                 }
5500                 /* ext4_truncate will clear the flag */
5501                 if ((EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))
5502                         ext4_truncate(inode);
5503         }
5504
5505         rc = inode_setattr(inode, attr);
5506
5507         /* If inode_setattr's call to ext4_truncate failed to get a
5508          * transaction handle at all, we need to clean up the in-core
5509          * orphan list manually. */
5510         if (inode->i_nlink)
5511                 ext4_orphan_del(NULL, inode);
5512
5513         if (!rc && (ia_valid & ATTR_MODE))
5514                 rc = ext4_acl_chmod(inode);
5515
5516 err_out:
5517         ext4_std_error(inode->i_sb, error);
5518         if (!error)
5519                 error = rc;
5520         return error;
5521 }
5522
5523 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5524                  struct kstat *stat)
5525 {
5526         struct inode *inode;
5527         unsigned long delalloc_blocks;
5528
5529         inode = dentry->d_inode;
5530         generic_fillattr(inode, stat);
5531
5532         /*
5533          * We can't update i_blocks if the block allocation is delayed
5534          * otherwise in the case of system crash before the real block
5535          * allocation is done, we will have i_blocks inconsistent with
5536          * on-disk file blocks.
5537          * We always keep i_blocks updated together with real
5538          * allocation. But to not confuse with user, stat
5539          * will return the blocks that include the delayed allocation
5540          * blocks for this file.
5541          */
5542         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5543         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5544         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5545
5546         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5547         return 0;
5548 }
5549
5550 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5551                                       int chunk)
5552 {
5553