ext4: remove initialized but not read variables
[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 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63
64 /*
65  * Test whether an inode is a fast symlink.
66  */
67 static int ext4_inode_is_fast_symlink(struct inode *inode)
68 {
69         int ea_blocks = EXT4_I(inode)->i_file_acl ?
70                 (inode->i_sb->s_blocksize >> 9) : 0;
71
72         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
73 }
74
75 /*
76  * Work out how many blocks we need to proceed with the next chunk of a
77  * truncate transaction.
78  */
79 static unsigned long blocks_for_truncate(struct inode *inode)
80 {
81         ext4_lblk_t needed;
82
83         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
84
85         /* Give ourselves just enough room to cope with inodes in which
86          * i_blocks is corrupt: we've seen disk corruptions in the past
87          * which resulted in random data in an inode which looked enough
88          * like a regular file for ext4 to try to delete it.  Things
89          * will go a bit crazy if that happens, but at least we should
90          * try not to panic the whole kernel. */
91         if (needed < 2)
92                 needed = 2;
93
94         /* But we need to bound the transaction so we don't overflow the
95          * journal. */
96         if (needed > EXT4_MAX_TRANS_DATA)
97                 needed = EXT4_MAX_TRANS_DATA;
98
99         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
100 }
101
102 /*
103  * Truncate transactions can be complex and absolutely huge.  So we need to
104  * be able to restart the transaction at a conventient checkpoint to make
105  * sure we don't overflow the journal.
106  *
107  * start_transaction gets us a new handle for a truncate transaction,
108  * and extend_transaction tries to extend the existing one a bit.  If
109  * extend fails, we need to propagate the failure up and restart the
110  * transaction in the top-level truncate loop. --sct
111  */
112 static handle_t *start_transaction(struct inode *inode)
113 {
114         handle_t *result;
115
116         result = ext4_journal_start(inode, blocks_for_truncate(inode));
117         if (!IS_ERR(result))
118                 return result;
119
120         ext4_std_error(inode->i_sb, PTR_ERR(result));
121         return result;
122 }
123
124 /*
125  * Try to extend this transaction for the purposes of truncation.
126  *
127  * Returns 0 if we managed to create more room.  If we can't create more
128  * room, and the transaction must be restarted we return 1.
129  */
130 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
131 {
132         if (!ext4_handle_valid(handle))
133                 return 0;
134         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
135                 return 0;
136         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
137                 return 0;
138         return 1;
139 }
140
141 /*
142  * Restart the transaction associated with *handle.  This does a commit,
143  * so before we call here everything must be consistently dirtied against
144  * this transaction.
145  */
146 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
147                                  int nblocks)
148 {
149         int ret;
150
151         /*
152          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
153          * moment, get_block can be called only for blocks inside i_size since
154          * page cache has been already dropped and writes are blocked by
155          * i_mutex. So we can safely drop the i_data_sem here.
156          */
157         BUG_ON(EXT4_JOURNAL(inode) == NULL);
158         jbd_debug(2, "restarting handle %p\n", handle);
159         up_write(&EXT4_I(inode)->i_data_sem);
160         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
161         down_write(&EXT4_I(inode)->i_data_sem);
162         ext4_discard_preallocations(inode);
163
164         return ret;
165 }
166
167 /*
168  * Called at the last iput() if i_nlink is zero.
169  */
170 void ext4_delete_inode(struct inode *inode)
171 {
172         handle_t *handle;
173         int err;
174
175         if (!is_bad_inode(inode))
176                 dquot_initialize(inode);
177
178         if (ext4_should_order_data(inode))
179                 ext4_begin_ordered_truncate(inode, 0);
180         truncate_inode_pages(&inode->i_data, 0);
181
182         if (is_bad_inode(inode))
183                 goto no_delete;
184
185         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
186         if (IS_ERR(handle)) {
187                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
188                 /*
189                  * If we're going to skip the normal cleanup, we still need to
190                  * make sure that the in-core orphan linked list is properly
191                  * cleaned up.
192                  */
193                 ext4_orphan_del(NULL, inode);
194                 goto no_delete;
195         }
196
197         if (IS_SYNC(inode))
198                 ext4_handle_sync(handle);
199         inode->i_size = 0;
200         err = ext4_mark_inode_dirty(handle, inode);
201         if (err) {
202                 ext4_warning(inode->i_sb,
203                              "couldn't mark inode dirty (err %d)", err);
204                 goto stop_handle;
205         }
206         if (inode->i_blocks)
207                 ext4_truncate(inode);
208
209         /*
210          * ext4_ext_truncate() doesn't reserve any slop when it
211          * restarts journal transactions; therefore there may not be
212          * enough credits left in the handle to remove the inode from
213          * the orphan list and set the dtime field.
214          */
215         if (!ext4_handle_has_enough_credits(handle, 3)) {
216                 err = ext4_journal_extend(handle, 3);
217                 if (err > 0)
218                         err = ext4_journal_restart(handle, 3);
219                 if (err != 0) {
220                         ext4_warning(inode->i_sb,
221                                      "couldn't extend journal (err %d)", err);
222                 stop_handle:
223                         ext4_journal_stop(handle);
224                         goto no_delete;
225                 }
226         }
227
228         /*
229          * Kill off the orphan record which ext4_truncate created.
230          * AKPM: I think this can be inside the above `if'.
231          * Note that ext4_orphan_del() has to be able to cope with the
232          * deletion of a non-existent orphan - this is because we don't
233          * know if ext4_truncate() actually created an orphan record.
234          * (Well, we could do this if we need to, but heck - it works)
235          */
236         ext4_orphan_del(handle, inode);
237         EXT4_I(inode)->i_dtime  = get_seconds();
238
239         /*
240          * One subtle ordering requirement: if anything has gone wrong
241          * (transaction abort, IO errors, whatever), then we can still
242          * do these next steps (the fs will already have been marked as
243          * having errors), but we can't free the inode if the mark_dirty
244          * fails.
245          */
246         if (ext4_mark_inode_dirty(handle, inode))
247                 /* If that failed, just do the required in-core inode clear. */
248                 clear_inode(inode);
249         else
250                 ext4_free_inode(handle, inode);
251         ext4_journal_stop(handle);
252         return;
253 no_delete:
254         clear_inode(inode);     /* We must guarantee clearing of inode... */
255 }
256
257 typedef struct {
258         __le32  *p;
259         __le32  key;
260         struct buffer_head *bh;
261 } Indirect;
262
263 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
264 {
265         p->key = *(p->p = v);
266         p->bh = bh;
267 }
268
269 /**
270  *      ext4_block_to_path - parse the block number into array of offsets
271  *      @inode: inode in question (we are only interested in its superblock)
272  *      @i_block: block number to be parsed
273  *      @offsets: array to store the offsets in
274  *      @boundary: set this non-zero if the referred-to block is likely to be
275  *             followed (on disk) by an indirect block.
276  *
277  *      To store the locations of file's data ext4 uses a data structure common
278  *      for UNIX filesystems - tree of pointers anchored in the inode, with
279  *      data blocks at leaves and indirect blocks in intermediate nodes.
280  *      This function translates the block number into path in that tree -
281  *      return value is the path length and @offsets[n] is the offset of
282  *      pointer to (n+1)th node in the nth one. If @block is out of range
283  *      (negative or too large) warning is printed and zero returned.
284  *
285  *      Note: function doesn't find node addresses, so no IO is needed. All
286  *      we need to know is the capacity of indirect blocks (taken from the
287  *      inode->i_sb).
288  */
289
290 /*
291  * Portability note: the last comparison (check that we fit into triple
292  * indirect block) is spelled differently, because otherwise on an
293  * architecture with 32-bit longs and 8Kb pages we might get into trouble
294  * if our filesystem had 8Kb blocks. We might use long long, but that would
295  * kill us on x86. Oh, well, at least the sign propagation does not matter -
296  * i_block would have to be negative in the very beginning, so we would not
297  * get there at all.
298  */
299
300 static int ext4_block_to_path(struct inode *inode,
301                               ext4_lblk_t i_block,
302                               ext4_lblk_t offsets[4], int *boundary)
303 {
304         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
305         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
306         const long direct_blocks = EXT4_NDIR_BLOCKS,
307                 indirect_blocks = ptrs,
308                 double_blocks = (1 << (ptrs_bits * 2));
309         int n = 0;
310         int final = 0;
311
312         if (i_block < direct_blocks) {
313                 offsets[n++] = i_block;
314                 final = direct_blocks;
315         } else if ((i_block -= direct_blocks) < indirect_blocks) {
316                 offsets[n++] = EXT4_IND_BLOCK;
317                 offsets[n++] = i_block;
318                 final = ptrs;
319         } else if ((i_block -= indirect_blocks) < double_blocks) {
320                 offsets[n++] = EXT4_DIND_BLOCK;
321                 offsets[n++] = i_block >> ptrs_bits;
322                 offsets[n++] = i_block & (ptrs - 1);
323                 final = ptrs;
324         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
325                 offsets[n++] = EXT4_TIND_BLOCK;
326                 offsets[n++] = i_block >> (ptrs_bits * 2);
327                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
328                 offsets[n++] = i_block & (ptrs - 1);
329                 final = ptrs;
330         } else {
331                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
332                              i_block + direct_blocks +
333                              indirect_blocks + double_blocks, inode->i_ino);
334         }
335         if (boundary)
336                 *boundary = final - 1 - (i_block & (ptrs - 1));
337         return n;
338 }
339
340 static int __ext4_check_blockref(const char *function, struct inode *inode,
341                                  __le32 *p, unsigned int max)
342 {
343         __le32 *bref = p;
344         unsigned int blk;
345
346         while (bref < p+max) {
347                 blk = le32_to_cpu(*bref++);
348                 if (blk &&
349                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
350                                                     blk, 1))) {
351                         ext4_error_inode(function, inode,
352                                          "invalid block reference %u", blk);
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_map_blocks() function handles non-extents inodes
893  * (i.e., using the traditional indirect/double-indirect i_blocks
894  * scheme) for ext4_map_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_map_blocks(handle_t *handle, struct inode *inode,
920                                struct ext4_map_blocks *map,
921                                int flags)
922 {
923         int err = -EIO;
924         ext4_lblk_t offsets[4];
925         Indirect chain[4];
926         Indirect *partial;
927         ext4_fsblk_t goal;
928         int indirect_blks;
929         int blocks_to_boundary = 0;
930         int depth;
931         int count = 0;
932         ext4_fsblk_t first_block = 0;
933
934         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
935         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
936         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
937                                    &blocks_to_boundary);
938
939         if (depth == 0)
940                 goto out;
941
942         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
943
944         /* Simplest case - block found, no allocation needed */
945         if (!partial) {
946                 first_block = le32_to_cpu(chain[depth - 1].key);
947                 count++;
948                 /*map more blocks*/
949                 while (count < map->m_len && count <= blocks_to_boundary) {
950                         ext4_fsblk_t blk;
951
952                         blk = le32_to_cpu(*(chain[depth-1].p + count));
953
954                         if (blk == first_block + count)
955                                 count++;
956                         else
957                                 break;
958                 }
959                 goto got_it;
960         }
961
962         /* Next simple case - plain lookup or failed read of indirect block */
963         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
964                 goto cleanup;
965
966         /*
967          * Okay, we need to do block allocation.
968         */
969         goal = ext4_find_goal(inode, map->m_lblk, partial);
970
971         /* the number of blocks need to allocate for [d,t]indirect blocks */
972         indirect_blks = (chain + depth) - partial - 1;
973
974         /*
975          * Next look up the indirect map to count the totoal number of
976          * direct blocks to allocate for this branch.
977          */
978         count = ext4_blks_to_allocate(partial, indirect_blks,
979                                       map->m_len, blocks_to_boundary);
980         /*
981          * Block out ext4_truncate while we alter the tree
982          */
983         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
984                                 &count, goal,
985                                 offsets + (partial - chain), partial);
986
987         /*
988          * The ext4_splice_branch call will free and forget any buffers
989          * on the new chain if there is a failure, but that risks using
990          * up transaction credits, especially for bitmaps where the
991          * credits cannot be returned.  Can we handle this somehow?  We
992          * may need to return -EAGAIN upwards in the worst case.  --sct
993          */
994         if (!err)
995                 err = ext4_splice_branch(handle, inode, map->m_lblk,
996                                          partial, indirect_blks, count);
997         if (err)
998                 goto cleanup;
999
1000         map->m_flags |= EXT4_MAP_NEW;
1001
1002         ext4_update_inode_fsync_trans(handle, inode, 1);
1003 got_it:
1004         map->m_flags |= EXT4_MAP_MAPPED;
1005         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1006         map->m_len = count;
1007         if (count > blocks_to_boundary)
1008                 map->m_flags |= EXT4_MAP_BOUNDARY;
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 out:
1019         return err;
1020 }
1021
1022 #ifdef CONFIG_QUOTA
1023 qsize_t *ext4_get_reserved_space(struct inode *inode)
1024 {
1025         return &EXT4_I(inode)->i_reserved_quota;
1026 }
1027 #endif
1028
1029 /*
1030  * Calculate the number of metadata blocks need to reserve
1031  * to allocate a new block at @lblocks for non extent file based file
1032  */
1033 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1034                                               sector_t lblock)
1035 {
1036         struct ext4_inode_info *ei = EXT4_I(inode);
1037         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1038         int blk_bits;
1039
1040         if (lblock < EXT4_NDIR_BLOCKS)
1041                 return 0;
1042
1043         lblock -= EXT4_NDIR_BLOCKS;
1044
1045         if (ei->i_da_metadata_calc_len &&
1046             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1047                 ei->i_da_metadata_calc_len++;
1048                 return 0;
1049         }
1050         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1051         ei->i_da_metadata_calc_len = 1;
1052         blk_bits = order_base_2(lblock);
1053         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1054 }
1055
1056 /*
1057  * Calculate the number of metadata blocks need to reserve
1058  * to allocate a block located at @lblock
1059  */
1060 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1061 {
1062         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1063                 return ext4_ext_calc_metadata_amount(inode, lblock);
1064
1065         return ext4_indirect_calc_metadata_amount(inode, lblock);
1066 }
1067
1068 /*
1069  * Called with i_data_sem down, which is important since we can call
1070  * ext4_discard_preallocations() from here.
1071  */
1072 void ext4_da_update_reserve_space(struct inode *inode,
1073                                         int used, int quota_claim)
1074 {
1075         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1076         struct ext4_inode_info *ei = EXT4_I(inode);
1077
1078         spin_lock(&ei->i_block_reservation_lock);
1079         trace_ext4_da_update_reserve_space(inode, used);
1080         if (unlikely(used > ei->i_reserved_data_blocks)) {
1081                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1082                          "with only %d reserved data blocks\n",
1083                          __func__, inode->i_ino, used,
1084                          ei->i_reserved_data_blocks);
1085                 WARN_ON(1);
1086                 used = ei->i_reserved_data_blocks;
1087         }
1088
1089         /* Update per-inode reservations */
1090         ei->i_reserved_data_blocks -= used;
1091         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1092         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1093                            used + ei->i_allocated_meta_blocks);
1094         ei->i_allocated_meta_blocks = 0;
1095
1096         if (ei->i_reserved_data_blocks == 0) {
1097                 /*
1098                  * We can release all of the reserved metadata blocks
1099                  * only when we have written all of the delayed
1100                  * allocation blocks.
1101                  */
1102                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1103                                    ei->i_reserved_meta_blocks);
1104                 ei->i_reserved_meta_blocks = 0;
1105                 ei->i_da_metadata_calc_len = 0;
1106         }
1107         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1108
1109         /* Update quota subsystem for data blocks */
1110         if (quota_claim)
1111                 dquot_claim_block(inode, used);
1112         else {
1113                 /*
1114                  * We did fallocate with an offset that is already delayed
1115                  * allocated. So on delayed allocated writeback we should
1116                  * not re-claim the quota for fallocated blocks.
1117                  */
1118                 dquot_release_reservation_block(inode, used);
1119         }
1120
1121         /*
1122          * If we have done all the pending block allocations and if
1123          * there aren't any writers on the inode, we can discard the
1124          * inode's preallocations.
1125          */
1126         if ((ei->i_reserved_data_blocks == 0) &&
1127             (atomic_read(&inode->i_writecount) == 0))
1128                 ext4_discard_preallocations(inode);
1129 }
1130
1131 static int check_block_validity(struct inode *inode, const char *func,
1132                                 struct ext4_map_blocks *map)
1133 {
1134         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1135                                    map->m_len)) {
1136                 ext4_error_inode(func, inode,
1137                            "lblock %lu mapped to illegal pblock %llu "
1138                            "(length %d)", (unsigned long) map->m_lblk,
1139                                  map->m_pblk, map->m_len);
1140                 return -EIO;
1141         }
1142         return 0;
1143 }
1144
1145 /*
1146  * Return the number of contiguous dirty pages in a given inode
1147  * starting at page frame idx.
1148  */
1149 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1150                                     unsigned int max_pages)
1151 {
1152         struct address_space *mapping = inode->i_mapping;
1153         pgoff_t index;
1154         struct pagevec pvec;
1155         pgoff_t num = 0;
1156         int i, nr_pages, done = 0;
1157
1158         if (max_pages == 0)
1159                 return 0;
1160         pagevec_init(&pvec, 0);
1161         while (!done) {
1162                 index = idx;
1163                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1164                                               PAGECACHE_TAG_DIRTY,
1165                                               (pgoff_t)PAGEVEC_SIZE);
1166                 if (nr_pages == 0)
1167                         break;
1168                 for (i = 0; i < nr_pages; i++) {
1169                         struct page *page = pvec.pages[i];
1170                         struct buffer_head *bh, *head;
1171
1172                         lock_page(page);
1173                         if (unlikely(page->mapping != mapping) ||
1174                             !PageDirty(page) ||
1175                             PageWriteback(page) ||
1176                             page->index != idx) {
1177                                 done = 1;
1178                                 unlock_page(page);
1179                                 break;
1180                         }
1181                         if (page_has_buffers(page)) {
1182                                 bh = head = page_buffers(page);
1183                                 do {
1184                                         if (!buffer_delay(bh) &&
1185                                             !buffer_unwritten(bh))
1186                                                 done = 1;
1187                                         bh = bh->b_this_page;
1188                                 } while (!done && (bh != head));
1189                         }
1190                         unlock_page(page);
1191                         if (done)
1192                                 break;
1193                         idx++;
1194                         num++;
1195                         if (num >= max_pages)
1196                                 break;
1197                 }
1198                 pagevec_release(&pvec);
1199         }
1200         return num;
1201 }
1202
1203 /*
1204  * The ext4_map_blocks() function tries to look up the requested blocks,
1205  * and returns if the blocks are already mapped.
1206  *
1207  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1208  * and store the allocated blocks in the result buffer head and mark it
1209  * mapped.
1210  *
1211  * If file type is extents based, it will call ext4_ext_map_blocks(),
1212  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1213  * based files
1214  *
1215  * On success, it returns the number of blocks being mapped or allocate.
1216  * if create==0 and the blocks are pre-allocated and uninitialized block,
1217  * the result buffer head is unmapped. If the create ==1, it will make sure
1218  * the buffer head is mapped.
1219  *
1220  * It returns 0 if plain look up failed (blocks have not been allocated), in
1221  * that casem, buffer head is unmapped
1222  *
1223  * It returns the error in case of allocation failure.
1224  */
1225 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1226                     struct ext4_map_blocks *map, int flags)
1227 {
1228         int retval;
1229
1230         map->m_flags = 0;
1231         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1232                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1233                   (unsigned long) map->m_lblk);
1234         /*
1235          * Try to see if we can get the block without requesting a new
1236          * file system block.
1237          */
1238         down_read((&EXT4_I(inode)->i_data_sem));
1239         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1240                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1241         } else {
1242                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1243         }
1244         up_read((&EXT4_I(inode)->i_data_sem));
1245
1246         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1247                 int ret = check_block_validity(inode, __func__, map);
1248                 if (ret != 0)
1249                         return ret;
1250         }
1251
1252         /* If it is only a block(s) look up */
1253         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1254                 return retval;
1255
1256         /*
1257          * Returns if the blocks have already allocated
1258          *
1259          * Note that if blocks have been preallocated
1260          * ext4_ext_get_block() returns th create = 0
1261          * with buffer head unmapped.
1262          */
1263         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1264                 return retval;
1265
1266         /*
1267          * When we call get_blocks without the create flag, the
1268          * BH_Unwritten flag could have gotten set if the blocks
1269          * requested were part of a uninitialized extent.  We need to
1270          * clear this flag now that we are committed to convert all or
1271          * part of the uninitialized extent to be an initialized
1272          * extent.  This is because we need to avoid the combination
1273          * of BH_Unwritten and BH_Mapped flags being simultaneously
1274          * set on the buffer_head.
1275          */
1276         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1277
1278         /*
1279          * New blocks allocate and/or writing to uninitialized extent
1280          * will possibly result in updating i_data, so we take
1281          * the write lock of i_data_sem, and call get_blocks()
1282          * with create == 1 flag.
1283          */
1284         down_write((&EXT4_I(inode)->i_data_sem));
1285
1286         /*
1287          * if the caller is from delayed allocation writeout path
1288          * we have already reserved fs blocks for allocation
1289          * let the underlying get_block() function know to
1290          * avoid double accounting
1291          */
1292         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1293                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1294         /*
1295          * We need to check for EXT4 here because migrate
1296          * could have changed the inode type in between
1297          */
1298         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1299                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1300         } else {
1301                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1302
1303                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1304                         /*
1305                          * We allocated new blocks which will result in
1306                          * i_data's format changing.  Force the migrate
1307                          * to fail by clearing migrate flags
1308                          */
1309                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1310                 }
1311
1312                 /*
1313                  * Update reserved blocks/metadata blocks after successful
1314                  * block allocation which had been deferred till now. We don't
1315                  * support fallocate for non extent files. So we can update
1316                  * reserve space here.
1317                  */
1318                 if ((retval > 0) &&
1319                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1320                         ext4_da_update_reserve_space(inode, retval, 1);
1321         }
1322         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1323                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1324
1325         up_write((&EXT4_I(inode)->i_data_sem));
1326         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1327                 int ret = check_block_validity(inode,
1328                                                "ext4_map_blocks_after_alloc",
1329                                                map);
1330                 if (ret != 0)
1331                         return ret;
1332         }
1333         return retval;
1334 }
1335
1336 /* Maximum number of blocks we map for direct IO at once. */
1337 #define DIO_MAX_BLOCKS 4096
1338
1339 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1340                            struct buffer_head *bh, int flags)
1341 {
1342         handle_t *handle = ext4_journal_current_handle();
1343         struct ext4_map_blocks map;
1344         int ret = 0, started = 0;
1345         int dio_credits;
1346
1347         map.m_lblk = iblock;
1348         map.m_len = bh->b_size >> inode->i_blkbits;
1349
1350         if (flags && !handle) {
1351                 /* Direct IO write... */
1352                 if (map.m_len > DIO_MAX_BLOCKS)
1353                         map.m_len = DIO_MAX_BLOCKS;
1354                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1355                 handle = ext4_journal_start(inode, dio_credits);
1356                 if (IS_ERR(handle)) {
1357                         ret = PTR_ERR(handle);
1358                         return ret;
1359                 }
1360                 started = 1;
1361         }
1362
1363         ret = ext4_map_blocks(handle, inode, &map, flags);
1364         if (ret > 0) {
1365                 map_bh(bh, inode->i_sb, map.m_pblk);
1366                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1367                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1368                 ret = 0;
1369         }
1370         if (started)
1371                 ext4_journal_stop(handle);
1372         return ret;
1373 }
1374
1375 int ext4_get_block(struct inode *inode, sector_t iblock,
1376                    struct buffer_head *bh, int create)
1377 {
1378         return _ext4_get_block(inode, iblock, bh,
1379                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1380 }
1381
1382 /*
1383  * `handle' can be NULL if create is zero
1384  */
1385 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1386                                 ext4_lblk_t block, int create, int *errp)
1387 {
1388         struct ext4_map_blocks map;
1389         struct buffer_head *bh;
1390         int fatal = 0, err;
1391
1392         J_ASSERT(handle != NULL || create == 0);
1393
1394         map.m_lblk = block;
1395         map.m_len = 1;
1396         err = ext4_map_blocks(handle, inode, &map,
1397                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1398
1399         if (err < 0)
1400                 *errp = err;
1401         if (err <= 0)
1402                 return NULL;
1403         *errp = 0;
1404
1405         bh = sb_getblk(inode->i_sb, map.m_pblk);
1406         if (!bh) {
1407                 *errp = -EIO;
1408                 return NULL;
1409         }
1410         if (map.m_flags & EXT4_MAP_NEW) {
1411                 J_ASSERT(create != 0);
1412                 J_ASSERT(handle != NULL);
1413
1414                 /*
1415                  * Now that we do not always journal data, we should
1416                  * keep in mind whether this should always journal the
1417                  * new buffer as metadata.  For now, regular file
1418                  * writes use ext4_get_block instead, so it's not a
1419                  * problem.
1420                  */
1421                 lock_buffer(bh);
1422                 BUFFER_TRACE(bh, "call get_create_access");
1423                 fatal = ext4_journal_get_create_access(handle, bh);
1424                 if (!fatal && !buffer_uptodate(bh)) {
1425                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1426                         set_buffer_uptodate(bh);
1427                 }
1428                 unlock_buffer(bh);
1429                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1430                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1431                 if (!fatal)
1432                         fatal = err;
1433         } else {
1434                 BUFFER_TRACE(bh, "not a new buffer");
1435         }
1436         if (fatal) {
1437                 *errp = fatal;
1438                 brelse(bh);
1439                 bh = NULL;
1440         }
1441         return bh;
1442 }
1443
1444 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1445                                ext4_lblk_t block, int create, int *err)
1446 {
1447         struct buffer_head *bh;
1448
1449         bh = ext4_getblk(handle, inode, block, create, err);
1450         if (!bh)
1451                 return bh;
1452         if (buffer_uptodate(bh))
1453                 return bh;
1454         ll_rw_block(READ_META, 1, &bh);
1455         wait_on_buffer(bh);
1456         if (buffer_uptodate(bh))
1457                 return bh;
1458         put_bh(bh);
1459         *err = -EIO;
1460         return NULL;
1461 }
1462
1463 static int walk_page_buffers(handle_t *handle,
1464                              struct buffer_head *head,
1465                              unsigned from,
1466                              unsigned to,
1467                              int *partial,
1468                              int (*fn)(handle_t *handle,
1469                                        struct buffer_head *bh))
1470 {
1471         struct buffer_head *bh;
1472         unsigned block_start, block_end;
1473         unsigned blocksize = head->b_size;
1474         int err, ret = 0;
1475         struct buffer_head *next;
1476
1477         for (bh = head, block_start = 0;
1478              ret == 0 && (bh != head || !block_start);
1479              block_start = block_end, bh = next) {
1480                 next = bh->b_this_page;
1481                 block_end = block_start + blocksize;
1482                 if (block_end <= from || block_start >= to) {
1483                         if (partial && !buffer_uptodate(bh))
1484                                 *partial = 1;
1485                         continue;
1486                 }
1487                 err = (*fn)(handle, bh);
1488                 if (!ret)
1489                         ret = err;
1490         }
1491         return ret;
1492 }
1493
1494 /*
1495  * To preserve ordering, it is essential that the hole instantiation and
1496  * the data write be encapsulated in a single transaction.  We cannot
1497  * close off a transaction and start a new one between the ext4_get_block()
1498  * and the commit_write().  So doing the jbd2_journal_start at the start of
1499  * prepare_write() is the right place.
1500  *
1501  * Also, this function can nest inside ext4_writepage() ->
1502  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1503  * has generated enough buffer credits to do the whole page.  So we won't
1504  * block on the journal in that case, which is good, because the caller may
1505  * be PF_MEMALLOC.
1506  *
1507  * By accident, ext4 can be reentered when a transaction is open via
1508  * quota file writes.  If we were to commit the transaction while thus
1509  * reentered, there can be a deadlock - we would be holding a quota
1510  * lock, and the commit would never complete if another thread had a
1511  * transaction open and was blocking on the quota lock - a ranking
1512  * violation.
1513  *
1514  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1515  * will _not_ run commit under these circumstances because handle->h_ref
1516  * is elevated.  We'll still have enough credits for the tiny quotafile
1517  * write.
1518  */
1519 static int do_journal_get_write_access(handle_t *handle,
1520                                        struct buffer_head *bh)
1521 {
1522         if (!buffer_mapped(bh) || buffer_freed(bh))
1523                 return 0;
1524         return ext4_journal_get_write_access(handle, bh);
1525 }
1526
1527 /*
1528  * Truncate blocks that were not used by write. We have to truncate the
1529  * pagecache as well so that corresponding buffers get properly unmapped.
1530  */
1531 static void ext4_truncate_failed_write(struct inode *inode)
1532 {
1533         truncate_inode_pages(inode->i_mapping, inode->i_size);
1534         ext4_truncate(inode);
1535 }
1536
1537 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1538                    struct buffer_head *bh_result, int create);
1539 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1540                             loff_t pos, unsigned len, unsigned flags,
1541                             struct page **pagep, void **fsdata)
1542 {
1543         struct inode *inode = mapping->host;
1544         int ret, needed_blocks;
1545         handle_t *handle;
1546         int retries = 0;
1547         struct page *page;
1548         pgoff_t index;
1549         unsigned from, to;
1550
1551         trace_ext4_write_begin(inode, pos, len, flags);
1552         /*
1553          * Reserve one block more for addition to orphan list in case
1554          * we allocate blocks but write fails for some reason
1555          */
1556         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1557         index = pos >> PAGE_CACHE_SHIFT;
1558         from = pos & (PAGE_CACHE_SIZE - 1);
1559         to = from + len;
1560
1561 retry:
1562         handle = ext4_journal_start(inode, needed_blocks);
1563         if (IS_ERR(handle)) {
1564                 ret = PTR_ERR(handle);
1565                 goto out;
1566         }
1567
1568         /* We cannot recurse into the filesystem as the transaction is already
1569          * started */
1570         flags |= AOP_FLAG_NOFS;
1571
1572         page = grab_cache_page_write_begin(mapping, index, flags);
1573         if (!page) {
1574                 ext4_journal_stop(handle);
1575                 ret = -ENOMEM;
1576                 goto out;
1577         }
1578         *pagep = page;
1579
1580         if (ext4_should_dioread_nolock(inode))
1581                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1582                                 fsdata, ext4_get_block_write);
1583         else
1584                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1585                                 fsdata, ext4_get_block);
1586
1587         if (!ret && ext4_should_journal_data(inode)) {
1588                 ret = walk_page_buffers(handle, page_buffers(page),
1589                                 from, to, NULL, do_journal_get_write_access);
1590         }
1591
1592         if (ret) {
1593                 unlock_page(page);
1594                 page_cache_release(page);
1595                 /*
1596                  * block_write_begin may have instantiated a few blocks
1597                  * outside i_size.  Trim these off again. Don't need
1598                  * i_size_read because we hold i_mutex.
1599                  *
1600                  * Add inode to orphan list in case we crash before
1601                  * truncate finishes
1602                  */
1603                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1604                         ext4_orphan_add(handle, inode);
1605
1606                 ext4_journal_stop(handle);
1607                 if (pos + len > inode->i_size) {
1608                         ext4_truncate_failed_write(inode);
1609                         /*
1610                          * If truncate failed early the inode might
1611                          * still be on the orphan list; we need to
1612                          * make sure the inode is removed from the
1613                          * orphan list in that case.
1614                          */
1615                         if (inode->i_nlink)
1616                                 ext4_orphan_del(NULL, inode);
1617                 }
1618         }
1619
1620         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1621                 goto retry;
1622 out:
1623         return ret;
1624 }
1625
1626 /* For write_end() in data=journal mode */
1627 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1628 {
1629         if (!buffer_mapped(bh) || buffer_freed(bh))
1630                 return 0;
1631         set_buffer_uptodate(bh);
1632         return ext4_handle_dirty_metadata(handle, NULL, bh);
1633 }
1634
1635 static int ext4_generic_write_end(struct file *file,
1636                                   struct address_space *mapping,
1637                                   loff_t pos, unsigned len, unsigned copied,
1638                                   struct page *page, void *fsdata)
1639 {
1640         int i_size_changed = 0;
1641         struct inode *inode = mapping->host;
1642         handle_t *handle = ext4_journal_current_handle();
1643
1644         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1645
1646         /*
1647          * No need to use i_size_read() here, the i_size
1648          * cannot change under us because we hold i_mutex.
1649          *
1650          * But it's important to update i_size while still holding page lock:
1651          * page writeout could otherwise come in and zero beyond i_size.
1652          */
1653         if (pos + copied > inode->i_size) {
1654                 i_size_write(inode, pos + copied);
1655                 i_size_changed = 1;
1656         }
1657
1658         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1659                 /* We need to mark inode dirty even if
1660                  * new_i_size is less that inode->i_size
1661                  * bu greater than i_disksize.(hint delalloc)
1662                  */
1663                 ext4_update_i_disksize(inode, (pos + copied));
1664                 i_size_changed = 1;
1665         }
1666         unlock_page(page);
1667         page_cache_release(page);
1668
1669         /*
1670          * Don't mark the inode dirty under page lock. First, it unnecessarily
1671          * makes the holding time of page lock longer. Second, it forces lock
1672          * ordering of page lock and transaction start for journaling
1673          * filesystems.
1674          */
1675         if (i_size_changed)
1676                 ext4_mark_inode_dirty(handle, inode);
1677
1678         return copied;
1679 }
1680
1681 /*
1682  * We need to pick up the new inode size which generic_commit_write gave us
1683  * `file' can be NULL - eg, when called from page_symlink().
1684  *
1685  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1686  * buffers are managed internally.
1687  */
1688 static int ext4_ordered_write_end(struct file *file,
1689                                   struct address_space *mapping,
1690                                   loff_t pos, unsigned len, unsigned copied,
1691                                   struct page *page, void *fsdata)
1692 {
1693         handle_t *handle = ext4_journal_current_handle();
1694         struct inode *inode = mapping->host;
1695         int ret = 0, ret2;
1696
1697         trace_ext4_ordered_write_end(inode, pos, len, copied);
1698         ret = ext4_jbd2_file_inode(handle, inode);
1699
1700         if (ret == 0) {
1701                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1702                                                         page, fsdata);
1703                 copied = ret2;
1704                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1705                         /* if we have allocated more blocks and copied
1706                          * less. We will have blocks allocated outside
1707                          * inode->i_size. So truncate them
1708                          */
1709                         ext4_orphan_add(handle, inode);
1710                 if (ret2 < 0)
1711                         ret = ret2;
1712         }
1713         ret2 = ext4_journal_stop(handle);
1714         if (!ret)
1715                 ret = ret2;
1716
1717         if (pos + len > inode->i_size) {
1718                 ext4_truncate_failed_write(inode);
1719                 /*
1720                  * If truncate failed early the inode might still be
1721                  * on the orphan list; we need to make sure the inode
1722                  * is removed from the orphan list in that case.
1723                  */
1724                 if (inode->i_nlink)
1725                         ext4_orphan_del(NULL, inode);
1726         }
1727
1728
1729         return ret ? ret : copied;
1730 }
1731
1732 static int ext4_writeback_write_end(struct file *file,
1733                                     struct address_space *mapping,
1734                                     loff_t pos, unsigned len, unsigned copied,
1735                                     struct page *page, void *fsdata)
1736 {
1737         handle_t *handle = ext4_journal_current_handle();
1738         struct inode *inode = mapping->host;
1739         int ret = 0, ret2;
1740
1741         trace_ext4_writeback_write_end(inode, pos, len, copied);
1742         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1743                                                         page, fsdata);
1744         copied = ret2;
1745         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1746                 /* if we have allocated more blocks and copied
1747                  * less. We will have blocks allocated outside
1748                  * inode->i_size. So truncate them
1749                  */
1750                 ext4_orphan_add(handle, inode);
1751
1752         if (ret2 < 0)
1753                 ret = ret2;
1754
1755         ret2 = ext4_journal_stop(handle);
1756         if (!ret)
1757                 ret = ret2;
1758
1759         if (pos + len > inode->i_size) {
1760                 ext4_truncate_failed_write(inode);
1761                 /*
1762                  * If truncate failed early the inode might still be
1763                  * on the orphan list; we need to make sure the inode
1764                  * is removed from the orphan list in that case.
1765                  */
1766                 if (inode->i_nlink)
1767                         ext4_orphan_del(NULL, inode);
1768         }
1769
1770         return ret ? ret : copied;
1771 }
1772
1773 static int ext4_journalled_write_end(struct file *file,
1774                                      struct address_space *mapping,
1775                                      loff_t pos, unsigned len, unsigned copied,
1776                                      struct page *page, void *fsdata)
1777 {
1778         handle_t *handle = ext4_journal_current_handle();
1779         struct inode *inode = mapping->host;
1780         int ret = 0, ret2;
1781         int partial = 0;
1782         unsigned from, to;
1783         loff_t new_i_size;
1784
1785         trace_ext4_journalled_write_end(inode, pos, len, copied);
1786         from = pos & (PAGE_CACHE_SIZE - 1);
1787         to = from + len;
1788
1789         if (copied < len) {
1790                 if (!PageUptodate(page))
1791                         copied = 0;
1792                 page_zero_new_buffers(page, from+copied, to);
1793         }
1794
1795         ret = walk_page_buffers(handle, page_buffers(page), from,
1796                                 to, &partial, write_end_fn);
1797         if (!partial)
1798                 SetPageUptodate(page);
1799         new_i_size = pos + copied;
1800         if (new_i_size > inode->i_size)
1801                 i_size_write(inode, pos+copied);
1802         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1803         if (new_i_size > EXT4_I(inode)->i_disksize) {
1804                 ext4_update_i_disksize(inode, new_i_size);
1805                 ret2 = ext4_mark_inode_dirty(handle, inode);
1806                 if (!ret)
1807                         ret = ret2;
1808         }
1809
1810         unlock_page(page);
1811         page_cache_release(page);
1812         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1813                 /* if we have allocated more blocks and copied
1814                  * less. We will have blocks allocated outside
1815                  * inode->i_size. So truncate them
1816                  */
1817                 ext4_orphan_add(handle, inode);
1818
1819         ret2 = ext4_journal_stop(handle);
1820         if (!ret)
1821                 ret = ret2;
1822         if (pos + len > inode->i_size) {
1823                 ext4_truncate_failed_write(inode);
1824                 /*
1825                  * If truncate failed early the inode might still be
1826                  * on the orphan list; we need to make sure the inode
1827                  * is removed from the orphan list in that case.
1828                  */
1829                 if (inode->i_nlink)
1830                         ext4_orphan_del(NULL, inode);
1831         }
1832
1833         return ret ? ret : copied;
1834 }
1835
1836 /*
1837  * Reserve a single block located at lblock
1838  */
1839 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1840 {
1841         int retries = 0;
1842         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1843         struct ext4_inode_info *ei = EXT4_I(inode);
1844         unsigned long md_needed;
1845         int ret;
1846
1847         /*
1848          * recalculate the amount of metadata blocks to reserve
1849          * in order to allocate nrblocks
1850          * worse case is one extent per block
1851          */
1852 repeat:
1853         spin_lock(&ei->i_block_reservation_lock);
1854         md_needed = ext4_calc_metadata_amount(inode, lblock);
1855         trace_ext4_da_reserve_space(inode, md_needed);
1856         spin_unlock(&ei->i_block_reservation_lock);
1857
1858         /*
1859          * We will charge metadata quota at writeout time; this saves
1860          * us from metadata over-estimation, though we may go over by
1861          * a small amount in the end.  Here we just reserve for data.
1862          */
1863         ret = dquot_reserve_block(inode, 1);
1864         if (ret)
1865                 return ret;
1866         /*
1867          * We do still charge estimated metadata to the sb though;
1868          * we cannot afford to run out of free blocks.
1869          */
1870         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1871                 dquot_release_reservation_block(inode, 1);
1872                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1873                         yield();
1874                         goto repeat;
1875                 }
1876                 return -ENOSPC;
1877         }
1878         spin_lock(&ei->i_block_reservation_lock);
1879         ei->i_reserved_data_blocks++;
1880         ei->i_reserved_meta_blocks += md_needed;
1881         spin_unlock(&ei->i_block_reservation_lock);
1882
1883         return 0;       /* success */
1884 }
1885
1886 static void ext4_da_release_space(struct inode *inode, int to_free)
1887 {
1888         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1889         struct ext4_inode_info *ei = EXT4_I(inode);
1890
1891         if (!to_free)
1892                 return;         /* Nothing to release, exit */
1893
1894         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1895
1896         trace_ext4_da_release_space(inode, to_free);
1897         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1898                 /*
1899                  * if there aren't enough reserved blocks, then the
1900                  * counter is messed up somewhere.  Since this
1901                  * function is called from invalidate page, it's
1902                  * harmless to return without any action.
1903                  */
1904                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1905                          "ino %lu, to_free %d with only %d reserved "
1906                          "data blocks\n", inode->i_ino, to_free,
1907                          ei->i_reserved_data_blocks);
1908                 WARN_ON(1);
1909                 to_free = ei->i_reserved_data_blocks;
1910         }
1911         ei->i_reserved_data_blocks -= to_free;
1912
1913         if (ei->i_reserved_data_blocks == 0) {
1914                 /*
1915                  * We can release all of the reserved metadata blocks
1916                  * only when we have written all of the delayed
1917                  * allocation blocks.
1918                  */
1919                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1920                                    ei->i_reserved_meta_blocks);
1921                 ei->i_reserved_meta_blocks = 0;
1922                 ei->i_da_metadata_calc_len = 0;
1923         }
1924
1925         /* update fs dirty data blocks counter */
1926         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1927
1928         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1929
1930         dquot_release_reservation_block(inode, to_free);
1931 }
1932
1933 static void ext4_da_page_release_reservation(struct page *page,
1934                                              unsigned long offset)
1935 {
1936         int to_release = 0;
1937         struct buffer_head *head, *bh;
1938         unsigned int curr_off = 0;
1939
1940         head = page_buffers(page);
1941         bh = head;
1942         do {
1943                 unsigned int next_off = curr_off + bh->b_size;
1944
1945                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1946                         to_release++;
1947                         clear_buffer_delay(bh);
1948                 }
1949                 curr_off = next_off;
1950         } while ((bh = bh->b_this_page) != head);
1951         ext4_da_release_space(page->mapping->host, to_release);
1952 }
1953
1954 /*
1955  * Delayed allocation stuff
1956  */
1957
1958 /*
1959  * mpage_da_submit_io - walks through extent of pages and try to write
1960  * them with writepage() call back
1961  *
1962  * @mpd->inode: inode
1963  * @mpd->first_page: first page of the extent
1964  * @mpd->next_page: page after the last page of the extent
1965  *
1966  * By the time mpage_da_submit_io() is called we expect all blocks
1967  * to be allocated. this may be wrong if allocation failed.
1968  *
1969  * As pages are already locked by write_cache_pages(), we can't use it
1970  */
1971 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1972 {
1973         long pages_skipped;
1974         struct pagevec pvec;
1975         unsigned long index, end;
1976         int ret = 0, err, nr_pages, i;
1977         struct inode *inode = mpd->inode;
1978         struct address_space *mapping = inode->i_mapping;
1979
1980         BUG_ON(mpd->next_page <= mpd->first_page);
1981         /*
1982          * We need to start from the first_page to the next_page - 1
1983          * to make sure we also write the mapped dirty buffer_heads.
1984          * If we look at mpd->b_blocknr we would only be looking
1985          * at the currently mapped buffer_heads.
1986          */
1987         index = mpd->first_page;
1988         end = mpd->next_page - 1;
1989
1990         pagevec_init(&pvec, 0);
1991         while (index <= end) {
1992                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1993                 if (nr_pages == 0)
1994                         break;
1995                 for (i = 0; i < nr_pages; i++) {
1996                         struct page *page = pvec.pages[i];
1997
1998                         index = page->index;
1999                         if (index > end)
2000                                 break;
2001                         index++;
2002
2003                         BUG_ON(!PageLocked(page));
2004                         BUG_ON(PageWriteback(page));
2005
2006                         pages_skipped = mpd->wbc->pages_skipped;
2007                         err = mapping->a_ops->writepage(page, mpd->wbc);
2008                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2009                                 /*
2010                                  * have successfully written the page
2011                                  * without skipping the same
2012                                  */
2013                                 mpd->pages_written++;
2014                         /*
2015                          * In error case, we have to continue because
2016                          * remaining pages are still locked
2017                          * XXX: unlock and re-dirty them?
2018                          */
2019                         if (ret == 0)
2020                                 ret = err;
2021                 }
2022                 pagevec_release(&pvec);
2023         }
2024         return ret;
2025 }
2026
2027 /*
2028  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2029  *
2030  * the function goes through all passed space and put actual disk
2031  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2032  */
2033 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2034                                  struct ext4_map_blocks *map)
2035 {
2036         struct inode *inode = mpd->inode;
2037         struct address_space *mapping = inode->i_mapping;
2038         int blocks = map->m_len;
2039         sector_t pblock = map->m_pblk, cur_logical;
2040         struct buffer_head *head, *bh;
2041         pgoff_t index, end;
2042         struct pagevec pvec;
2043         int nr_pages, i;
2044
2045         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2046         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2047         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2048
2049         pagevec_init(&pvec, 0);
2050
2051         while (index <= end) {
2052                 /* XXX: optimize tail */
2053                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2054                 if (nr_pages == 0)
2055                         break;
2056                 for (i = 0; i < nr_pages; i++) {
2057                         struct page *page = pvec.pages[i];
2058
2059                         index = page->index;
2060                         if (index > end)
2061                                 break;
2062                         index++;
2063
2064                         BUG_ON(!PageLocked(page));
2065                         BUG_ON(PageWriteback(page));
2066                         BUG_ON(!page_has_buffers(page));
2067
2068                         bh = page_buffers(page);
2069                         head = bh;
2070
2071                         /* skip blocks out of the range */
2072                         do {
2073                                 if (cur_logical >= map->m_lblk)
2074                                         break;
2075                                 cur_logical++;
2076                         } while ((bh = bh->b_this_page) != head);
2077
2078                         do {
2079                                 if (cur_logical >= map->m_lblk + blocks)
2080                                         break;
2081
2082                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2083
2084                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2085
2086                                         if (buffer_delay(bh)) {
2087                                                 clear_buffer_delay(bh);
2088                                                 bh->b_blocknr = pblock;
2089                                         } else {
2090                                                 /*
2091                                                  * unwritten already should have
2092                                                  * blocknr assigned. Verify that
2093                                                  */
2094                                                 clear_buffer_unwritten(bh);
2095                                                 BUG_ON(bh->b_blocknr != pblock);
2096                                         }
2097
2098                                 } else if (buffer_mapped(bh))
2099                                         BUG_ON(bh->b_blocknr != pblock);
2100
2101                                 if (map->m_flags & EXT4_MAP_UNINIT)
2102                                         set_buffer_uninit(bh);
2103                                 cur_logical++;
2104                                 pblock++;
2105                         } while ((bh = bh->b_this_page) != head);
2106                 }
2107                 pagevec_release(&pvec);
2108         }
2109 }
2110
2111
2112 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2113                                         sector_t logical, long blk_cnt)
2114 {
2115         int nr_pages, i;
2116         pgoff_t index, end;
2117         struct pagevec pvec;
2118         struct inode *inode = mpd->inode;
2119         struct address_space *mapping = inode->i_mapping;
2120
2121         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2122         end   = (logical + blk_cnt - 1) >>
2123                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2124         while (index <= end) {
2125                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2126                 if (nr_pages == 0)
2127                         break;
2128                 for (i = 0; i < nr_pages; i++) {
2129                         struct page *page = pvec.pages[i];
2130                         if (page->index > end)
2131                                 break;
2132                         BUG_ON(!PageLocked(page));
2133                         BUG_ON(PageWriteback(page));
2134                         block_invalidatepage(page, 0);
2135                         ClearPageUptodate(page);
2136                         unlock_page(page);
2137                 }
2138                 index = pvec.pages[nr_pages - 1]->index + 1;
2139                 pagevec_release(&pvec);
2140         }
2141         return;
2142 }
2143
2144 static void ext4_print_free_blocks(struct inode *inode)
2145 {
2146         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2147         printk(KERN_CRIT "Total free blocks count %lld\n",
2148                ext4_count_free_blocks(inode->i_sb));
2149         printk(KERN_CRIT "Free/Dirty block details\n");
2150         printk(KERN_CRIT "free_blocks=%lld\n",
2151                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2152         printk(KERN_CRIT "dirty_blocks=%lld\n",
2153                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2154         printk(KERN_CRIT "Block reservation details\n");
2155         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2156                EXT4_I(inode)->i_reserved_data_blocks);
2157         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2158                EXT4_I(inode)->i_reserved_meta_blocks);
2159         return;
2160 }
2161
2162 /*
2163  * mpage_da_map_blocks - go through given space
2164  *
2165  * @mpd - bh describing space
2166  *
2167  * The function skips space we know is already mapped to disk blocks.
2168  *
2169  */
2170 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2171 {
2172         int err, blks, get_blocks_flags;
2173         struct ext4_map_blocks map;
2174         sector_t next = mpd->b_blocknr;
2175         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2176         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2177         handle_t *handle = NULL;
2178
2179         /*
2180          * We consider only non-mapped and non-allocated blocks
2181          */
2182         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2183                 !(mpd->b_state & (1 << BH_Delay)) &&
2184                 !(mpd->b_state & (1 << BH_Unwritten)))
2185                 return 0;
2186
2187         /*
2188          * If we didn't accumulate anything to write simply return
2189          */
2190         if (!mpd->b_size)
2191                 return 0;
2192
2193         handle = ext4_journal_current_handle();
2194         BUG_ON(!handle);
2195
2196         /*
2197          * Call ext4_get_blocks() to allocate any delayed allocation
2198          * blocks, or to convert an uninitialized extent to be
2199          * initialized (in the case where we have written into
2200          * one or more preallocated blocks).
2201          *
2202          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2203          * indicate that we are on the delayed allocation path.  This
2204          * affects functions in many different parts of the allocation
2205          * call path.  This flag exists primarily because we don't
2206          * want to change *many* call functions, so ext4_get_blocks()
2207          * will set the magic i_delalloc_reserved_flag once the
2208          * inode's allocation semaphore is taken.
2209          *
2210          * If the blocks in questions were delalloc blocks, set
2211          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2212          * variables are updated after the blocks have been allocated.
2213          */
2214         map.m_lblk = next;
2215         map.m_len = max_blocks;
2216         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2217         if (ext4_should_dioread_nolock(mpd->inode))
2218                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2219         if (mpd->b_state & (1 << BH_Delay))
2220                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2221
2222         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2223         if (blks < 0) {
2224                 err = blks;
2225                 /*
2226                  * If get block returns with error we simply
2227                  * return. Later writepage will redirty the page and
2228                  * writepages will find the dirty page again
2229                  */
2230                 if (err == -EAGAIN)
2231                         return 0;
2232
2233                 if (err == -ENOSPC &&
2234                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2235                         mpd->retval = err;
2236                         return 0;
2237                 }
2238
2239                 /*
2240                  * get block failure will cause us to loop in
2241                  * writepages, because a_ops->writepage won't be able
2242                  * to make progress. The page will be redirtied by
2243                  * writepage and writepages will again try to write
2244                  * the same.
2245                  */
2246                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2247                          "delayed block allocation failed for inode %lu at "
2248                          "logical offset %llu with max blocks %zd with "
2249                          "error %d", mpd->inode->i_ino,
2250                          (unsigned long long) next,
2251                          mpd->b_size >> mpd->inode->i_blkbits, err);
2252                 printk(KERN_CRIT "This should not happen!!  "
2253                        "Data will be lost\n");
2254                 if (err == -ENOSPC) {
2255                         ext4_print_free_blocks(mpd->inode);
2256                 }
2257                 /* invalidate all the pages */
2258                 ext4_da_block_invalidatepages(mpd, next,
2259                                 mpd->b_size >> mpd->inode->i_blkbits);
2260                 return err;
2261         }
2262         BUG_ON(blks == 0);
2263
2264         if (map.m_flags & EXT4_MAP_NEW) {
2265                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2266                 int i;
2267
2268                 for (i = 0; i < map.m_len; i++)
2269                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2270         }
2271
2272         /*
2273          * If blocks are delayed marked, we need to
2274          * put actual blocknr and drop delayed bit
2275          */
2276         if ((mpd->b_state & (1 << BH_Delay)) ||
2277             (mpd->b_state & (1 << BH_Unwritten)))
2278                 mpage_put_bnr_to_bhs(mpd, &map);
2279
2280         if (ext4_should_order_data(mpd->inode)) {
2281                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2282                 if (err)
2283                         return err;
2284         }
2285
2286         /*
2287          * Update on-disk size along with block allocation.
2288          */
2289         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2290         if (disksize > i_size_read(mpd->inode))
2291                 disksize = i_size_read(mpd->inode);
2292         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2293                 ext4_update_i_disksize(mpd->inode, disksize);
2294                 return ext4_mark_inode_dirty(handle, mpd->inode);
2295         }
2296
2297         return 0;
2298 }
2299
2300 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2301                 (1 << BH_Delay) | (1 << BH_Unwritten))
2302
2303 /*
2304  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2305  *
2306  * @mpd->lbh - extent of blocks
2307  * @logical - logical number of the block in the file
2308  * @bh - bh of the block (used to access block's state)
2309  *
2310  * the function is used to collect contig. blocks in same state
2311  */
2312 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2313                                    sector_t logical, size_t b_size,
2314                                    unsigned long b_state)
2315 {
2316         sector_t next;
2317         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2318
2319         /*
2320          * XXX Don't go larger than mballoc is willing to allocate
2321          * This is a stopgap solution.  We eventually need to fold
2322          * mpage_da_submit_io() into this function and then call
2323          * ext4_get_blocks() multiple times in a loop
2324          */
2325         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2326                 goto flush_it;
2327
2328         /* check if thereserved journal credits might overflow */
2329         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2330                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2331                         /*
2332                          * With non-extent format we are limited by the journal
2333                          * credit available.  Total credit needed to insert
2334                          * nrblocks contiguous blocks is dependent on the
2335                          * nrblocks.  So limit nrblocks.
2336                          */
2337                         goto flush_it;
2338                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2339                                 EXT4_MAX_TRANS_DATA) {
2340                         /*
2341                          * Adding the new buffer_head would make it cross the
2342                          * allowed limit for which we have journal credit
2343                          * reserved. So limit the new bh->b_size
2344                          */
2345                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2346                                                 mpd->inode->i_blkbits;
2347                         /* we will do mpage_da_submit_io in the next loop */
2348                 }
2349         }
2350         /*
2351          * First block in the extent
2352          */
2353         if (mpd->b_size == 0) {
2354                 mpd->b_blocknr = logical;
2355                 mpd->b_size = b_size;
2356                 mpd->b_state = b_state & BH_FLAGS;
2357                 return;
2358         }
2359
2360         next = mpd->b_blocknr + nrblocks;
2361         /*
2362          * Can we merge the block to our big extent?
2363          */
2364         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2365                 mpd->b_size += b_size;
2366                 return;
2367         }
2368
2369 flush_it:
2370         /*
2371          * We couldn't merge the block to our extent, so we
2372          * need to flush current  extent and start new one
2373          */
2374         if (mpage_da_map_blocks(mpd) == 0)
2375                 mpage_da_submit_io(mpd);
2376         mpd->io_done = 1;
2377         return;
2378 }
2379
2380 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2381 {
2382         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2383 }
2384
2385 /*
2386  * __mpage_da_writepage - finds extent of pages and blocks
2387  *
2388  * @page: page to consider
2389  * @wbc: not used, we just follow rules
2390  * @data: context
2391  *
2392  * The function finds extents of pages and scan them for all blocks.
2393  */
2394 static int __mpage_da_writepage(struct page *page,
2395                                 struct writeback_control *wbc, void *data)
2396 {
2397         struct mpage_da_data *mpd = data;
2398         struct inode *inode = mpd->inode;
2399         struct buffer_head *bh, *head;
2400         sector_t logical;
2401
2402         /*
2403          * Can we merge this page to current extent?
2404          */
2405         if (mpd->next_page != page->index) {
2406                 /*
2407                  * Nope, we can't. So, we map non-allocated blocks
2408                  * and start IO on them using writepage()
2409                  */
2410                 if (mpd->next_page != mpd->first_page) {
2411                         if (mpage_da_map_blocks(mpd) == 0)
2412                                 mpage_da_submit_io(mpd);
2413                         /*
2414                          * skip rest of the page in the page_vec
2415                          */
2416                         mpd->io_done = 1;
2417                         redirty_page_for_writepage(wbc, page);
2418                         unlock_page(page);
2419                         return MPAGE_DA_EXTENT_TAIL;
2420                 }
2421
2422                 /*
2423                  * Start next extent of pages ...
2424                  */
2425                 mpd->first_page = page->index;
2426
2427                 /*
2428                  * ... and blocks
2429                  */
2430                 mpd->b_size = 0;
2431                 mpd->b_state = 0;
2432                 mpd->b_blocknr = 0;
2433         }
2434
2435         mpd->next_page = page->index + 1;
2436         logical = (sector_t) page->index <<
2437                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2438
2439         if (!page_has_buffers(page)) {
2440                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2441                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2442                 if (mpd->io_done)
2443                         return MPAGE_DA_EXTENT_TAIL;
2444         } else {
2445                 /*
2446                  * Page with regular buffer heads, just add all dirty ones
2447                  */
2448                 head = page_buffers(page);
2449                 bh = head;
2450                 do {
2451                         BUG_ON(buffer_locked(bh));
2452                         /*
2453                          * We need to try to allocate
2454                          * unmapped blocks in the same page.
2455                          * Otherwise we won't make progress
2456                          * with the page in ext4_writepage
2457                          */
2458                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2459                                 mpage_add_bh_to_extent(mpd, logical,
2460                                                        bh->b_size,
2461                                                        bh->b_state);
2462                                 if (mpd->io_done)
2463                                         return MPAGE_DA_EXTENT_TAIL;
2464                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2465                                 /*
2466                                  * mapped dirty buffer. We need to update
2467                                  * the b_state because we look at
2468                                  * b_state in mpage_da_map_blocks. We don't
2469                                  * update b_size because if we find an
2470                                  * unmapped buffer_head later we need to
2471                                  * use the b_state flag of that buffer_head.
2472                                  */
2473                                 if (mpd->b_size == 0)
2474                                         mpd->b_state = bh->b_state & BH_FLAGS;
2475                         }
2476                         logical++;
2477                 } while ((bh = bh->b_this_page) != head);
2478         }
2479
2480         return 0;
2481 }
2482
2483 /*
2484  * This is a special get_blocks_t callback which is used by
2485  * ext4_da_write_begin().  It will either return mapped block or
2486  * reserve space for a single block.
2487  *
2488  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2489  * We also have b_blocknr = -1 and b_bdev initialized properly
2490  *
2491  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2492  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2493  * initialized properly.
2494  */
2495 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2496                                   struct buffer_head *bh, int create)
2497 {
2498         struct ext4_map_blocks map;
2499         int ret = 0;
2500         sector_t invalid_block = ~((sector_t) 0xffff);
2501
2502         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2503                 invalid_block = ~0;
2504
2505         BUG_ON(create == 0);
2506         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2507
2508         map.m_lblk = iblock;
2509         map.m_len = 1;
2510
2511         /*
2512          * first, we need to know whether the block is allocated already
2513          * preallocated blocks are unmapped but should treated
2514          * the same as allocated blocks.
2515          */
2516         ret = ext4_map_blocks(NULL, inode, &map, 0);
2517         if (ret < 0)
2518                 return ret;
2519         if (ret == 0) {
2520                 if (buffer_delay(bh))
2521                         return 0; /* Not sure this could or should happen */
2522                 /*
2523                  * XXX: __block_prepare_write() unmaps passed block,
2524                  * is it OK?
2525                  */
2526                 ret = ext4_da_reserve_space(inode, iblock);
2527                 if (ret)
2528                         /* not enough space to reserve */
2529                         return ret;
2530
2531                 map_bh(bh, inode->i_sb, invalid_block);
2532                 set_buffer_new(bh);
2533                 set_buffer_delay(bh);
2534                 return 0;
2535         }
2536
2537         map_bh(bh, inode->i_sb, map.m_pblk);
2538         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2539
2540         if (buffer_unwritten(bh)) {
2541                 /* A delayed write to unwritten bh should be marked
2542                  * new and mapped.  Mapped ensures that we don't do
2543                  * get_block multiple times when we write to the same
2544                  * offset and new ensures that we do proper zero out
2545                  * for partial write.
2546                  */
2547                 set_buffer_new(bh);
2548                 set_buffer_mapped(bh);
2549         }
2550         return 0;
2551 }
2552
2553 /*
2554  * This function is used as a standard get_block_t calback function
2555  * when there is no desire to allocate any blocks.  It is used as a
2556  * callback function for block_prepare_write(), nobh_writepage(), and
2557  * block_write_full_page().  These functions should only try to map a
2558  * single block at a time.
2559  *
2560  * Since this function doesn't do block allocations even if the caller
2561  * requests it by passing in create=1, it is critically important that
2562  * any caller checks to make sure that any buffer heads are returned
2563  * by this function are either all already mapped or marked for
2564  * delayed allocation before calling nobh_writepage() or
2565  * block_write_full_page().  Otherwise, b_blocknr could be left
2566  * unitialized, and the page write functions will be taken by
2567  * surprise.
2568  */
2569 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2570                                    struct buffer_head *bh_result, int create)
2571 {
2572         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2573         return _ext4_get_block(inode, iblock, bh_result, 0);
2574 }
2575
2576 static int bget_one(handle_t *handle, struct buffer_head *bh)
2577 {
2578         get_bh(bh);
2579         return 0;
2580 }
2581
2582 static int bput_one(handle_t *handle, struct buffer_head *bh)
2583 {
2584         put_bh(bh);
2585         return 0;
2586 }
2587
2588 static int __ext4_journalled_writepage(struct page *page,
2589                                        unsigned int len)
2590 {
2591         struct address_space *mapping = page->mapping;
2592         struct inode *inode = mapping->host;
2593         struct buffer_head *page_bufs;
2594         handle_t *handle = NULL;
2595         int ret = 0;
2596         int err;
2597
2598         page_bufs = page_buffers(page);
2599         BUG_ON(!page_bufs);
2600         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2601         /* As soon as we unlock the page, it can go away, but we have
2602          * references to buffers so we are safe */
2603         unlock_page(page);
2604
2605         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2606         if (IS_ERR(handle)) {
2607                 ret = PTR_ERR(handle);
2608                 goto out;
2609         }
2610
2611         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2612                                 do_journal_get_write_access);
2613
2614         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2615                                 write_end_fn);
2616         if (ret == 0)
2617                 ret = err;
2618         err = ext4_journal_stop(handle);
2619         if (!ret)
2620                 ret = err;
2621
2622         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2623         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2624 out:
2625         return ret;
2626 }
2627
2628 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2629 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2630
2631 /*
2632  * Note that we don't need to start a transaction unless we're journaling data
2633  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2634  * need to file the inode to the transaction's list in ordered mode because if
2635  * we are writing back data added by write(), the inode is already there and if
2636  * we are writing back data modified via mmap(), noone guarantees in which
2637  * transaction the data will hit the disk. In case we are journaling data, we
2638  * cannot start transaction directly because transaction start ranks above page
2639  * lock so we have to do some magic.
2640  *
2641  * This function can get called via...
2642  *   - ext4_da_writepages after taking page lock (have journal handle)
2643  *   - journal_submit_inode_data_buffers (no journal handle)
2644  *   - shrink_page_list via pdflush (no journal handle)
2645  *   - grab_page_cache when doing write_begin (have journal handle)
2646  *
2647  * We don't do any block allocation in this function. If we have page with
2648  * multiple blocks we need to write those buffer_heads that are mapped. This
2649  * is important for mmaped based write. So if we do with blocksize 1K
2650  * truncate(f, 1024);
2651  * a = mmap(f, 0, 4096);
2652  * a[0] = 'a';
2653  * truncate(f, 4096);
2654  * we have in the page first buffer_head mapped via page_mkwrite call back
2655  * but other bufer_heads would be unmapped but dirty(dirty done via the
2656  * do_wp_page). So writepage should write the first block. If we modify
2657  * the mmap area beyond 1024 we will again get a page_fault and the
2658  * page_mkwrite callback will do the block allocation and mark the
2659  * buffer_heads mapped.
2660  *
2661  * We redirty the page if we have any buffer_heads that is either delay or
2662  * unwritten in the page.
2663  *
2664  * We can get recursively called as show below.
2665  *
2666  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2667  *              ext4_writepage()
2668  *
2669  * But since we don't do any block allocation we should not deadlock.
2670  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2671  */
2672 static int ext4_writepage(struct page *page,
2673                           struct writeback_control *wbc)
2674 {
2675         int ret = 0;
2676         loff_t size;
2677         unsigned int len;
2678         struct buffer_head *page_bufs = NULL;
2679         struct inode *inode = page->mapping->host;
2680
2681         trace_ext4_writepage(inode, page);
2682         size = i_size_read(inode);
2683         if (page->index == size >> PAGE_CACHE_SHIFT)
2684                 len = size & ~PAGE_CACHE_MASK;
2685         else
2686                 len = PAGE_CACHE_SIZE;
2687
2688         if (page_has_buffers(page)) {
2689                 page_bufs = page_buffers(page);
2690                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2691                                         ext4_bh_delay_or_unwritten)) {
2692                         /*
2693                          * We don't want to do  block allocation
2694                          * So redirty the page and return
2695                          * We may reach here when we do a journal commit
2696                          * via journal_submit_inode_data_buffers.
2697                          * If we don't have mapping block we just ignore
2698                          * them. We can also reach here via shrink_page_list
2699                          */
2700                         redirty_page_for_writepage(wbc, page);
2701                         unlock_page(page);
2702                         return 0;
2703                 }
2704         } else {
2705                 /*
2706                  * The test for page_has_buffers() is subtle:
2707                  * We know the page is dirty but it lost buffers. That means
2708                  * that at some moment in time after write_begin()/write_end()
2709                  * has been called all buffers have been clean and thus they
2710                  * must have been written at least once. So they are all
2711                  * mapped and we can happily proceed with mapping them
2712                  * and writing the page.
2713                  *
2714                  * Try to initialize the buffer_heads and check whether
2715                  * all are mapped and non delay. We don't want to
2716                  * do block allocation here.
2717                  */
2718                 ret = block_prepare_write(page, 0, len,
2719                                           noalloc_get_block_write);
2720                 if (!ret) {
2721                         page_bufs = page_buffers(page);
2722                         /* check whether all are mapped and non delay */
2723                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2724                                                 ext4_bh_delay_or_unwritten)) {
2725                                 redirty_page_for_writepage(wbc, page);
2726                                 unlock_page(page);
2727                                 return 0;
2728                         }
2729                 } else {
2730                         /*
2731                          * We can't do block allocation here
2732                          * so just redity the page and unlock
2733                          * and return
2734                          */
2735                         redirty_page_for_writepage(wbc, page);
2736                         unlock_page(page);
2737                         return 0;
2738                 }
2739                 /* now mark the buffer_heads as dirty and uptodate */
2740                 block_commit_write(page, 0, len);
2741         }
2742
2743         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2744                 /*
2745                  * It's mmapped pagecache.  Add buffers and journal it.  There
2746                  * doesn't seem much point in redirtying the page here.
2747                  */
2748                 ClearPageChecked(page);
2749                 return __ext4_journalled_writepage(page, len);
2750         }
2751
2752         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2753                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2754         else if (page_bufs && buffer_uninit(page_bufs)) {
2755                 ext4_set_bh_endio(page_bufs, inode);
2756                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2757                                             wbc, ext4_end_io_buffer_write);
2758         } else
2759                 ret = block_write_full_page(page, noalloc_get_block_write,
2760                                             wbc);
2761
2762         return ret;
2763 }
2764
2765 /*
2766  * This is called via ext4_da_writepages() to
2767  * calulate the total number of credits to reserve to fit
2768  * a single extent allocation into a single transaction,
2769  * ext4_da_writpeages() will loop calling this before
2770  * the block allocation.
2771  */
2772
2773 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2774 {
2775         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2776
2777         /*
2778          * With non-extent format the journal credit needed to
2779          * insert nrblocks contiguous block is dependent on
2780          * number of contiguous block. So we will limit
2781          * number of contiguous block to a sane value
2782          */
2783         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2784             (max_blocks > EXT4_MAX_TRANS_DATA))
2785                 max_blocks = EXT4_MAX_TRANS_DATA;
2786
2787         return ext4_chunk_trans_blocks(inode, max_blocks);
2788 }
2789
2790 /*
2791  * write_cache_pages_da - walk the list of dirty pages of the given
2792  * address space and call the callback function (which usually writes
2793  * the pages).
2794  *
2795  * This is a forked version of write_cache_pages().  Differences:
2796  *      Range cyclic is ignored.
2797  *      no_nrwrite_index_update is always presumed true
2798  */
2799 static int write_cache_pages_da(struct address_space *mapping,
2800                                 struct writeback_control *wbc,
2801                                 struct mpage_da_data *mpd)
2802 {
2803         int ret = 0;
2804         int done = 0;
2805         struct pagevec pvec;
2806         int nr_pages;
2807         pgoff_t index;
2808         pgoff_t end;            /* Inclusive */
2809         long nr_to_write = wbc->nr_to_write;
2810
2811         pagevec_init(&pvec, 0);
2812         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2813         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2814
2815         while (!done && (index <= end)) {
2816                 int i;
2817
2818                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2819                               PAGECACHE_TAG_DIRTY,
2820                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2821                 if (nr_pages == 0)
2822                         break;
2823
2824                 for (i = 0; i < nr_pages; i++) {
2825                         struct page *page = pvec.pages[i];
2826
2827                         /*
2828                          * At this point, the page may be truncated or
2829                          * invalidated (changing page->mapping to NULL), or
2830                          * even swizzled back from swapper_space to tmpfs file
2831                          * mapping. However, page->index will not change
2832                          * because we have a reference on the page.
2833                          */
2834                         if (page->index > end) {
2835                                 done = 1;
2836                                 break;
2837                         }
2838
2839                         lock_page(page);
2840
2841                         /*
2842                          * Page truncated or invalidated. We can freely skip it
2843                          * then, even for data integrity operations: the page
2844                          * has disappeared concurrently, so there could be no
2845                          * real expectation of this data interity operation
2846                          * even if there is now a new, dirty page at the same
2847                          * pagecache address.
2848                          */
2849                         if (unlikely(page->mapping != mapping)) {
2850 continue_unlock:
2851                                 unlock_page(page);
2852                                 continue;
2853                         }
2854
2855                         if (!PageDirty(page)) {
2856                                 /* someone wrote it for us */
2857                                 goto continue_unlock;
2858                         }
2859
2860                         if (PageWriteback(page)) {
2861                                 if (wbc->sync_mode != WB_SYNC_NONE)
2862                                         wait_on_page_writeback(page);
2863                                 else
2864                                         goto continue_unlock;
2865                         }
2866
2867                         BUG_ON(PageWriteback(page));
2868                         if (!clear_page_dirty_for_io(page))
2869                                 goto continue_unlock;
2870
2871                         ret = __mpage_da_writepage(page, wbc, mpd);
2872                         if (unlikely(ret)) {
2873                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2874                                         unlock_page(page);
2875                                         ret = 0;
2876                                 } else {
2877                                         done = 1;
2878                                         break;
2879                                 }
2880                         }
2881
2882                         if (nr_to_write > 0) {
2883                                 nr_to_write--;
2884                                 if (nr_to_write == 0 &&
2885                                     wbc->sync_mode == WB_SYNC_NONE) {
2886                                         /*
2887                                          * We stop writing back only if we are
2888                                          * not doing integrity sync. In case of
2889                                          * integrity sync we have to keep going
2890                                          * because someone may be concurrently
2891                                          * dirtying pages, and we might have
2892                                          * synced a lot of newly appeared dirty
2893                                          * pages, but have not synced all of the
2894                                          * old dirty pages.
2895                                          */
2896                                         done = 1;
2897                                         break;
2898                                 }
2899                         }
2900                 }
2901                 pagevec_release(&pvec);
2902                 cond_resched();
2903         }
2904         return ret;
2905 }
2906
2907
2908 static int ext4_da_writepages(struct address_space *mapping,
2909                               struct writeback_control *wbc)
2910 {
2911         pgoff_t index;
2912         int range_whole = 0;
2913         handle_t *handle = NULL;
2914         struct mpage_da_data mpd;
2915         struct inode *inode = mapping->host;
2916         int pages_written = 0;
2917         long pages_skipped;
2918         unsigned int max_pages;
2919         int range_cyclic, cycled = 1, io_done = 0;
2920         int needed_blocks, ret = 0;
2921         long desired_nr_to_write, nr_to_writebump = 0;
2922         loff_t range_start = wbc->range_start;
2923         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2924
2925         trace_ext4_da_writepages(inode, wbc);
2926
2927         /*
2928          * No pages to write? This is mainly a kludge to avoid starting
2929          * a transaction for special inodes like journal inode on last iput()
2930          * because that could violate lock ordering on umount
2931          */
2932         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2933                 return 0;
2934
2935         /*
2936          * If the filesystem has aborted, it is read-only, so return
2937          * right away instead of dumping stack traces later on that
2938          * will obscure the real source of the problem.  We test
2939          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2940          * the latter could be true if the filesystem is mounted
2941          * read-only, and in that case, ext4_da_writepages should
2942          * *never* be called, so if that ever happens, we would want
2943          * the stack trace.
2944          */
2945         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2946                 return -EROFS;
2947
2948         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2949                 range_whole = 1;
2950
2951         range_cyclic = wbc->range_cyclic;
2952         if (wbc->range_cyclic) {
2953                 index = mapping->writeback_index;
2954                 if (index)
2955                         cycled = 0;
2956                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2957                 wbc->range_end  = LLONG_MAX;
2958                 wbc->range_cyclic = 0;
2959         } else
2960                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2961
2962         /*
2963          * This works around two forms of stupidity.  The first is in
2964          * the writeback code, which caps the maximum number of pages
2965          * written to be 1024 pages.  This is wrong on multiple
2966          * levels; different architectues have a different page size,
2967          * which changes the maximum amount of data which gets
2968          * written.  Secondly, 4 megabytes is way too small.  XFS
2969          * forces this value to be 16 megabytes by multiplying
2970          * nr_to_write parameter by four, and then relies on its
2971          * allocator to allocate larger extents to make them
2972          * contiguous.  Unfortunately this brings us to the second
2973          * stupidity, which is that ext4's mballoc code only allocates
2974          * at most 2048 blocks.  So we force contiguous writes up to
2975          * the number of dirty blocks in the inode, or
2976          * sbi->max_writeback_mb_bump whichever is smaller.
2977          */
2978         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2979         if (!range_cyclic && range_whole)
2980                 desired_nr_to_write = wbc->nr_to_write * 8;
2981         else
2982                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2983                                                            max_pages);
2984         if (desired_nr_to_write > max_pages)
2985                 desired_nr_to_write = max_pages;
2986
2987         if (wbc->nr_to_write < desired_nr_to_write) {
2988                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2989                 wbc->nr_to_write = desired_nr_to_write;
2990         }
2991
2992         mpd.wbc = wbc;
2993         mpd.inode = mapping->host;
2994
2995         pages_skipped = wbc->pages_skipped;
2996
2997 retry:
2998         while (!ret && wbc->nr_to_write > 0) {
2999
3000                 /*
3001                  * we  insert one extent at a time. So we need
3002                  * credit needed for single extent allocation.
3003                  * journalled mode is currently not supported
3004                  * by delalloc
3005                  */
3006                 BUG_ON(ext4_should_journal_data(inode));
3007                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3008
3009                 /* start a new transaction*/
3010                 handle = ext4_journal_start(inode, needed_blocks);
3011                 if (IS_ERR(handle)) {
3012                         ret = PTR_ERR(handle);
3013                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3014                                "%ld pages, ino %lu; err %d", __func__,
3015                                 wbc->nr_to_write, inode->i_ino, ret);
3016                         goto out_writepages;
3017                 }
3018
3019                 /*
3020                  * Now call __mpage_da_writepage to find the next
3021                  * contiguous region of logical blocks that need
3022                  * blocks to be allocated by ext4.  We don't actually
3023                  * submit the blocks for I/O here, even though
3024                  * write_cache_pages thinks it will, and will set the
3025                  * pages as clean for write before calling
3026                  * __mpage_da_writepage().
3027                  */
3028                 mpd.b_size = 0;
3029                 mpd.b_state = 0;
3030                 mpd.b_blocknr = 0;
3031                 mpd.first_page = 0;
3032                 mpd.next_page = 0;
3033                 mpd.io_done = 0;
3034                 mpd.pages_written = 0;
3035                 mpd.retval = 0;
3036                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3037                 /*
3038                  * If we have a contiguous extent of pages and we
3039                  * haven't done the I/O yet, map the blocks and submit
3040                  * them for I/O.
3041                  */
3042                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3043                         if (mpage_da_map_blocks(&mpd) == 0)
3044                                 mpage_da_submit_io(&mpd);
3045                         mpd.io_done = 1;
3046                         ret = MPAGE_DA_EXTENT_TAIL;
3047                 }
3048                 trace_ext4_da_write_pages(inode, &mpd);
3049                 wbc->nr_to_write -= mpd.pages_written;
3050
3051                 ext4_journal_stop(handle);
3052
3053                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3054                         /* commit the transaction which would
3055                          * free blocks released in the transaction
3056                          * and try again
3057                          */
3058                         jbd2_journal_force_commit_nested(sbi->s_journal);
3059                         wbc->pages_skipped = pages_skipped;
3060                         ret = 0;
3061                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3062                         /*
3063                          * got one extent now try with
3064                          * rest of the pages
3065                          */
3066                         pages_written += mpd.pages_written;
3067                         wbc->pages_skipped = pages_skipped;
3068                         ret = 0;
3069                         io_done = 1;
3070                 } else if (wbc->nr_to_write)
3071                         /*
3072                          * There is no more writeout needed
3073                          * or we requested for a noblocking writeout
3074                          * and we found the device congested
3075                          */
3076                         break;
3077         }
3078         if (!io_done && !cycled) {
3079                 cycled = 1;
3080                 index = 0;
3081                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3082                 wbc->range_end  = mapping->writeback_index - 1;
3083                 goto retry;
3084         }
3085         if (pages_skipped != wbc->pages_skipped)
3086                 ext4_msg(inode->i_sb, KERN_CRIT,
3087                          "This should not happen leaving %s "
3088                          "with nr_to_write = %ld ret = %d",
3089                          __func__, wbc->nr_to_write, ret);
3090
3091         /* Update index */
3092         index += pages_written;
3093         wbc->range_cyclic = range_cyclic;
3094         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3095                 /*
3096                  * set the writeback_index so that range_cyclic
3097                  * mode will write it back later
3098                  */
3099                 mapping->writeback_index = index;
3100
3101 out_writepages:
3102         wbc->nr_to_write -= nr_to_writebump;
3103         wbc->range_start = range_start;
3104         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3105         return ret;
3106 }
3107
3108 #define FALL_BACK_TO_NONDELALLOC 1
3109 static int ext4_nonda_switch(struct super_block *sb)
3110 {
3111         s64 free_blocks, dirty_blocks;
3112         struct ext4_sb_info *sbi = EXT4_SB(sb);
3113
3114         /*
3115          * switch to non delalloc mode if we are running low
3116          * on free block. The free block accounting via percpu
3117          * counters can get slightly wrong with percpu_counter_batch getting
3118          * accumulated on each CPU without updating global counters
3119          * Delalloc need an accurate free block accounting. So switch
3120          * to non delalloc when we are near to error range.
3121          */
3122         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3123         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3124         if (2 * free_blocks < 3 * dirty_blocks ||
3125                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3126                 /*
3127                  * free block count is less than 150% of dirty blocks
3128                  * or free blocks is less than watermark
3129                  */
3130                 return 1;
3131         }
3132         /*
3133          * Even if we don't switch but are nearing capacity,
3134          * start pushing delalloc when 1/2 of free blocks are dirty.
3135          */
3136         if (free_blocks < 2 * dirty_blocks)
3137                 writeback_inodes_sb_if_idle(sb);
3138
3139         return 0;
3140 }
3141
3142 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3143                                loff_t pos, unsigned len, unsigned flags,
3144                                struct page **pagep, void **fsdata)
3145 {
3146         int ret, retries = 0;
3147         struct page *page;
3148         pgoff_t index;
3149         struct inode *inode = mapping->host;
3150         handle_t *handle;
3151
3152         index = pos >> PAGE_CACHE_SHIFT;
3153
3154         if (ext4_nonda_switch(inode->i_sb)) {
3155                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3156                 return ext4_write_begin(file, mapping, pos,
3157                                         len, flags, pagep, fsdata);
3158         }
3159         *fsdata = (void *)0;
3160         trace_ext4_da_write_begin(inode, pos, len, flags);
3161 retry:
3162         /*
3163          * With delayed allocation, we don't log the i_disksize update
3164          * if there is delayed block allocation. But we still need
3165          * to journalling the i_disksize update if writes to the end
3166          * of file which has an already mapped buffer.
3167          */
3168         handle = ext4_journal_start(inode, 1);
3169         if (IS_ERR(handle)) {
3170                 ret = PTR_ERR(handle);
3171                 goto out;
3172         }
3173         /* We cannot recurse into the filesystem as the transaction is already
3174          * started */
3175         flags |= AOP_FLAG_NOFS;
3176
3177         page = grab_cache_page_write_begin(mapping, index, flags);
3178         if (!page) {
3179                 ext4_journal_stop(handle);
3180                 ret = -ENOMEM;
3181                 goto out;
3182         }
3183         *pagep = page;
3184
3185         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3186                                 ext4_da_get_block_prep);
3187         if (ret < 0) {
3188                 unlock_page(page);
3189                 ext4_journal_stop(handle);
3190                 page_cache_release(page);
3191                 /*
3192                  * block_write_begin may have instantiated a few blocks
3193                  * outside i_size.  Trim these off again. Don't need
3194                  * i_size_read because we hold i_mutex.
3195                  */
3196                 if (pos + len > inode->i_size)
3197                         ext4_truncate_failed_write(inode);
3198         }
3199
3200         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3201                 goto retry;
3202 out:
3203         return ret;
3204 }
3205
3206 /*
3207  * Check if we should update i_disksize
3208  * when write to the end of file but not require block allocation
3209  */
3210 static int ext4_da_should_update_i_disksize(struct page *page,
3211                                             unsigned long offset)
3212 {
3213         struct buffer_head *bh;
3214         struct inode *inode = page->mapping->host;
3215         unsigned int idx;
3216         int i;
3217
3218         bh = page_buffers(page);
3219         idx = offset >> inode->i_blkbits;
3220
3221         for (i = 0; i < idx; i++)
3222                 bh = bh->b_this_page;
3223
3224         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3225                 return 0;
3226         return 1;
3227 }
3228
3229 static int ext4_da_write_end(struct file *file,
3230                              struct address_space *mapping,
3231                              loff_t pos, unsigned len, unsigned copied,
3232                              struct page *page, void *fsdata)
3233 {
3234         struct inode *inode = mapping->host;
3235         int ret = 0, ret2;
3236         handle_t *handle = ext4_journal_current_handle();
3237         loff_t new_i_size;
3238         unsigned long start, end;
3239         int write_mode = (int)(unsigned long)fsdata;
3240
3241         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3242                 if (ext4_should_order_data(inode)) {
3243                         return ext4_ordered_write_end(file, mapping, pos,
3244                                         len, copied, page, fsdata);
3245                 } else if (ext4_should_writeback_data(inode)) {
3246                         return ext4_writeback_write_end(file, mapping, pos,
3247                                         len, copied, page, fsdata);
3248                 } else {
3249                         BUG();
3250                 }
3251         }
3252
3253         trace_ext4_da_write_end(inode, pos, len, copied);
3254         start = pos & (PAGE_CACHE_SIZE - 1);
3255         end = start + copied - 1;
3256
3257         /*
3258          * generic_write_end() will run mark_inode_dirty() if i_size
3259          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3260          * into that.
3261          */
3262
3263         new_i_size = pos + copied;
3264         if (new_i_size > EXT4_I(inode)->i_disksize) {
3265                 if (ext4_da_should_update_i_disksize(page, end)) {
3266                         down_write(&EXT4_I(inode)->i_data_sem);
3267                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3268                                 /*
3269                                  * Updating i_disksize when extending file
3270                                  * without needing block allocation
3271                                  */
3272                                 if (ext4_should_order_data(inode))
3273                                         ret = ext4_jbd2_file_inode(handle,
3274                                                                    inode);
3275
3276                                 EXT4_I(inode)->i_disksize = new_i_size;
3277                         }
3278                         up_write(&EXT4_I(inode)->i_data_sem);
3279                         /* We need to mark inode dirty even if
3280                          * new_i_size is less that inode->i_size
3281                          * bu greater than i_disksize.(hint delalloc)
3282                          */
3283                         ext4_mark_inode_dirty(handle, inode);
3284                 }
3285         }
3286         ret2 = generic_write_end(file, mapping, pos, len, copied,
3287                                                         page, fsdata);
3288         copied = ret2;
3289         if (ret2 < 0)
3290                 ret = ret2;
3291         ret2 = ext4_journal_stop(handle);
3292         if (!ret)
3293                 ret = ret2;
3294
3295         return ret ? ret : copied;
3296 }
3297
3298 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3299 {
3300         /*
3301          * Drop reserved blocks
3302          */
3303         BUG_ON(!PageLocked(page));
3304         if (!page_has_buffers(page))
3305                 goto out;
3306
3307         ext4_da_page_release_reservation(page, offset);
3308
3309 out:
3310         ext4_invalidatepage(page, offset);
3311
3312         return;
3313 }
3314
3315 /*
3316  * Force all delayed allocation blocks to be allocated for a given inode.
3317  */
3318 int ext4_alloc_da_blocks(struct inode *inode)
3319 {
3320         trace_ext4_alloc_da_blocks(inode);
3321
3322         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3323             !EXT4_I(inode)->i_reserved_meta_blocks)
3324                 return 0;
3325
3326         /*
3327          * We do something simple for now.  The filemap_flush() will
3328          * also start triggering a write of the data blocks, which is
3329          * not strictly speaking necessary (and for users of
3330          * laptop_mode, not even desirable).  However, to do otherwise
3331          * would require replicating code paths in:
3332          *
3333          * ext4_da_writepages() ->
3334          *    write_cache_pages() ---> (via passed in callback function)
3335          *        __mpage_da_writepage() -->
3336          *           mpage_add_bh_to_extent()
3337          *           mpage_da_map_blocks()
3338          *
3339          * The problem is that write_cache_pages(), located in
3340          * mm/page-writeback.c, marks pages clean in preparation for
3341          * doing I/O, which is not desirable if we're not planning on
3342          * doing I/O at all.
3343          *
3344          * We could call write_cache_pages(), and then redirty all of
3345          * the pages by calling redirty_page_for_writeback() but that
3346          * would be ugly in the extreme.  So instead we would need to
3347          * replicate parts of the code in the above functions,
3348          * simplifying them becuase we wouldn't actually intend to
3349          * write out the pages, but rather only collect contiguous
3350          * logical block extents, call the multi-block allocator, and
3351          * then update the buffer heads with the block allocations.
3352          *
3353          * For now, though, we'll cheat by calling filemap_flush(),
3354          * which will map the blocks, and start the I/O, but not
3355          * actually wait for the I/O to complete.
3356          */
3357         return filemap_flush(inode->i_mapping);
3358 }
3359
3360 /*
3361  * bmap() is special.  It gets used by applications such as lilo and by
3362  * the swapper to find the on-disk block of a specific piece of data.
3363  *
3364  * Naturally, this is dangerous if the block concerned is still in the
3365  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3366  * filesystem and enables swap, then they may get a nasty shock when the
3367  * data getting swapped to that swapfile suddenly gets overwritten by
3368  * the original zero's written out previously to the journal and
3369  * awaiting writeback in the kernel's buffer cache.
3370  *
3371  * So, if we see any bmap calls here on a modified, data-journaled file,
3372  * take extra steps to flush any blocks which might be in the cache.
3373  */
3374 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3375 {
3376         struct inode *inode = mapping->host;
3377         journal_t *journal;
3378         int err;
3379
3380         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3381                         test_opt(inode->i_sb, DELALLOC)) {
3382                 /*
3383                  * With delalloc we want to sync the file
3384                  * so that we can make sure we allocate
3385                  * blocks for file
3386                  */
3387                 filemap_write_and_wait(mapping);
3388         }
3389
3390         if (EXT4_JOURNAL(inode) &&
3391             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3392                 /*
3393                  * This is a REALLY heavyweight approach, but the use of
3394                  * bmap on dirty files is expected to be extremely rare:
3395                  * only if we run lilo or swapon on a freshly made file
3396                  * do we expect this to happen.
3397                  *
3398                  * (bmap requires CAP_SYS_RAWIO so this does not
3399                  * represent an unprivileged user DOS attack --- we'd be
3400                  * in trouble if mortal users could trigger this path at
3401                  * will.)
3402                  *
3403                  * NB. EXT4_STATE_JDATA is not set on files other than
3404                  * regular files.  If somebody wants to bmap a directory
3405                  * or symlink and gets confused because the buffer
3406                  * hasn't yet been flushed to disk, they deserve
3407                  * everything they get.
3408                  */
3409
3410                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3411                 journal = EXT4_JOURNAL(inode);
3412                 jbd2_journal_lock_updates(journal);
3413                 err = jbd2_journal_flush(journal);
3414                 jbd2_journal_unlock_updates(journal);
3415
3416                 if (err)
3417                         return 0;
3418         }
3419
3420         return generic_block_bmap(mapping, block, ext4_get_block);
3421 }
3422
3423 static int ext4_readpage(struct file *file, struct page *page)
3424 {
3425         return mpage_readpage(page, ext4_get_block);
3426 }
3427
3428 static int
3429 ext4_readpages(struct file *file, struct address_space *mapping,
3430                 struct list_head *pages, unsigned nr_pages)
3431 {
3432         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3433 }
3434
3435 static void ext4_free_io_end(ext4_io_end_t *io)
3436 {
3437         BUG_ON(!io);
3438         if (io->page)
3439                 put_page(io->page);
3440         iput(io->inode);
3441         kfree(io);
3442 }
3443
3444 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3445 {
3446         struct buffer_head *head, *bh;
3447         unsigned int curr_off = 0;
3448
3449         if (!page_has_buffers(page))
3450                 return;
3451         head = bh = page_buffers(page);
3452         do {
3453                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3454                                         && bh->b_private) {
3455                         ext4_free_io_end(bh->b_private);
3456                         bh->b_private = NULL;
3457                         bh->b_end_io = NULL;
3458                 }
3459                 curr_off = curr_off + bh->b_size;
3460                 bh = bh->b_this_page;
3461         } while (bh != head);
3462 }
3463
3464 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3465 {
3466         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3467
3468         /*
3469          * free any io_end structure allocated for buffers to be discarded
3470          */
3471         if (ext4_should_dioread_nolock(page->mapping->host))
3472                 ext4_invalidatepage_free_endio(page, offset);
3473         /*
3474          * If it's a full truncate we just forget about the pending dirtying
3475          */
3476         if (offset == 0)
3477                 ClearPageChecked(page);
3478
3479         if (journal)
3480                 jbd2_journal_invalidatepage(journal, page, offset);
3481         else
3482                 block_invalidatepage(page, offset);
3483 }
3484
3485 static int ext4_releasepage(struct page *page, gfp_t wait)
3486 {
3487         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3488
3489         WARN_ON(PageChecked(page));
3490         if (!page_has_buffers(page))
3491                 return 0;
3492         if (journal)
3493                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3494         else
3495                 return try_to_free_buffers(page);
3496 }
3497
3498 /*
3499  * O_DIRECT for ext3 (or indirect map) based files
3500  *
3501  * If the O_DIRECT write will extend the file then add this inode to the
3502  * orphan list.  So recovery will truncate it back to the original size
3503  * if the machine crashes during the write.
3504  *
3505  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3506  * crashes then stale disk data _may_ be exposed inside the file. But current
3507  * VFS code falls back into buffered path in that case so we are safe.
3508  */
3509 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3510                               const struct iovec *iov, loff_t offset,
3511                               unsigned long nr_segs)
3512 {
3513         struct file *file = iocb->ki_filp;
3514         struct inode *inode = file->f_mapping->host;
3515         struct ext4_inode_info *ei = EXT4_I(inode);
3516         handle_t *handle;
3517         ssize_t ret;
3518         int orphan = 0;
3519         size_t count = iov_length(iov, nr_segs);
3520         int retries = 0;
3521
3522         if (rw == WRITE) {
3523                 loff_t final_size = offset + count;
3524
3525                 if (final_size > inode->i_size) {
3526                         /* Credits for sb + inode write */
3527                         handle = ext4_journal_start(inode, 2);
3528                         if (IS_ERR(handle)) {
3529                                 ret = PTR_ERR(handle);
3530                                 goto out;
3531                         }
3532                         ret = ext4_orphan_add(handle, inode);
3533                         if (ret) {
3534                                 ext4_journal_stop(handle);
3535                                 goto out;
3536                         }
3537                         orphan = 1;
3538                         ei->i_disksize = inode->i_size;
3539                         ext4_journal_stop(handle);
3540                 }
3541         }
3542
3543 retry:
3544         if (rw == READ && ext4_should_dioread_nolock(inode))
3545                 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
3546                                  inode->i_sb->s_bdev, iov,
3547                                  offset, nr_segs,
3548                                  ext4_get_block, NULL);
3549         else
3550                 ret = blockdev_direct_IO(rw, iocb, inode,
3551                                  inode->i_sb->s_bdev, iov,
3552                                  offset, nr_segs,
3553                                  ext4_get_block, NULL);
3554         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3555                 goto retry;
3556
3557         if (orphan) {
3558                 int err;
3559
3560                 /* Credits for sb + inode write */
3561                 handle = ext4_journal_start(inode, 2);
3562                 if (IS_ERR(handle)) {
3563                         /* This is really bad luck. We've written the data
3564                          * but cannot extend i_size. Bail out and pretend
3565                          * the write failed... */
3566                         ret = PTR_ERR(handle);
3567                         if (inode->i_nlink)
3568                                 ext4_orphan_del(NULL, inode);
3569
3570                         goto out;
3571                 }
3572                 if (inode->i_nlink)
3573                         ext4_orphan_del(handle, inode);
3574                 if (ret > 0) {
3575                         loff_t end = offset + ret;
3576                         if (end > inode->i_size) {
3577                                 ei->i_disksize = end;
3578                                 i_size_write(inode, end);
3579                                 /*
3580                                  * We're going to return a positive `ret'
3581                                  * here due to non-zero-length I/O, so there's
3582                                  * no way of reporting error returns from
3583                                  * ext4_mark_inode_dirty() to userspace.  So
3584                                  * ignore it.
3585                                  */
3586                                 ext4_mark_inode_dirty(handle, inode);
3587                         }
3588                 }
3589                 err = ext4_journal_stop(handle);
3590                 if (ret == 0)
3591                         ret = err;
3592         }
3593 out:
3594         return ret;
3595 }
3596
3597 /*
3598  * ext4_get_block used when preparing for a DIO write or buffer write.
3599  * We allocate an uinitialized extent if blocks haven't been allocated.
3600  * The extent will be converted to initialized after the IO is complete.
3601  */
3602 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3603                    struct buffer_head *bh_result, int create)
3604 {
3605         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3606                    inode->i_ino, create);
3607         return _ext4_get_block(inode, iblock, bh_result,
3608                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3609 }
3610
3611 static void dump_completed_IO(struct inode * inode)
3612 {
3613 #ifdef  EXT4_DEBUG
3614         struct list_head *cur, *before, *after;
3615         ext4_io_end_t *io, *io0, *io1;
3616         unsigned long flags;
3617
3618         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3619                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3620                 return;
3621         }
3622
3623         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3624         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3625         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3626                 cur = &io->list;
3627                 before = cur->prev;
3628                 io0 = container_of(before, ext4_io_end_t, list);
3629                 after = cur->next;
3630                 io1 = container_of(after, ext4_io_end_t, list);
3631
3632                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3633                             io, inode->i_ino, io0, io1);
3634         }
3635         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3636 #endif
3637 }
3638
3639 /*
3640  * check a range of space and convert unwritten extents to written.
3641  */
3642 static int ext4_end_io_nolock(ext4_io_end_t *io)
3643 {
3644         struct inode *inode = io->inode;
3645         loff_t offset = io->offset;
3646         ssize_t size = io->size;
3647         int ret = 0;
3648
3649         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3650                    "list->prev 0x%p\n",
3651                    io, inode->i_ino, io->list.next, io->list.prev);
3652
3653         if (list_empty(&io->list))
3654                 return ret;
3655
3656         if (io->flag != EXT4_IO_UNWRITTEN)
3657                 return ret;
3658
3659         ret = ext4_convert_unwritten_extents(inode, offset, size);
3660         if (ret < 0) {
3661                 printk(KERN_EMERG "%s: failed to convert unwritten"
3662                         "extents to written extents, error is %d"
3663                         " io is still on inode %lu aio dio list\n",
3664                        __func__, ret, inode->i_ino);
3665                 return ret;
3666         }
3667
3668         /* clear the DIO AIO unwritten flag */
3669         io->flag = 0;
3670         return ret;
3671 }
3672
3673 /*
3674  * work on completed aio dio IO, to convert unwritten extents to extents
3675  */
3676 static void ext4_end_io_work(struct work_struct *work)
3677 {
3678         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3679         struct inode            *inode = io->inode;
3680         struct ext4_inode_info  *ei = EXT4_I(inode);
3681         unsigned long           flags;
3682         int                     ret;
3683
3684         mutex_lock(&inode->i_mutex);
3685         ret = ext4_end_io_nolock(io);
3686         if (ret < 0) {
3687                 mutex_unlock(&inode->i_mutex);
3688                 return;
3689         }
3690
3691         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3692         if (!list_empty(&io->list))
3693                 list_del_init(&io->list);
3694         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3695         mutex_unlock(&inode->i_mutex);
3696         ext4_free_io_end(io);
3697 }
3698
3699 /*
3700  * This function is called from ext4_sync_file().
3701  *
3702  * When IO is completed, the work to convert unwritten extents to
3703  * written is queued on workqueue but may not get immediately
3704  * scheduled. When fsync is called, we need to ensure the
3705  * conversion is complete before fsync returns.
3706  * The inode keeps track of a list of pending/completed IO that
3707  * might needs to do the conversion. This function walks through
3708  * the list and convert the related unwritten extents for completed IO
3709  * to written.
3710  * The function return the number of pending IOs on success.
3711  */
3712 int flush_completed_IO(struct inode *inode)
3713 {
3714         ext4_io_end_t *io;
3715         struct ext4_inode_info *ei = EXT4_I(inode);
3716         unsigned long flags;
3717         int ret = 0;
3718         int ret2 = 0;
3719
3720         if (list_empty(&ei->i_completed_io_list))
3721                 return ret;
3722
3723         dump_completed_IO(inode);
3724         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3725         while (!list_empty(&ei->i_completed_io_list)){
3726                 io = list_entry(ei->i_completed_io_list.next,
3727                                 ext4_io_end_t, list);
3728                 /*
3729                  * Calling ext4_end_io_nolock() to convert completed
3730                  * IO to written.
3731                  *
3732                  * When ext4_sync_file() is called, run_queue() may already
3733                  * about to flush the work corresponding to this io structure.
3734                  * It will be upset if it founds the io structure related
3735                  * to the work-to-be schedule is freed.
3736                  *
3737                  * Thus we need to keep the io structure still valid here after
3738                  * convertion finished. The io structure has a flag to
3739                  * avoid double converting from both fsync and background work
3740                  * queue work.
3741                  */
3742                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3743                 ret = ext4_end_io_nolock(io);
3744                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3745                 if (ret < 0)
3746                         ret2 = ret;
3747                 else
3748                         list_del_init(&io->list);
3749         }
3750         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3751         return (ret2 < 0) ? ret2 : 0;
3752 }
3753
3754 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3755 {
3756         ext4_io_end_t *io = NULL;
3757
3758         io = kmalloc(sizeof(*io), flags);
3759
3760         if (io) {
3761                 igrab(inode);
3762                 io->inode = inode;
3763                 io->flag = 0;
3764                 io->offset = 0;
3765                 io->size = 0;
3766                 io->page = NULL;
3767                 INIT_WORK(&io->work, ext4_end_io_work);
3768                 INIT_LIST_HEAD(&io->list);
3769         }
3770
3771         return io;
3772 }
3773
3774 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3775                             ssize_t size, void *private)
3776 {
3777         ext4_io_end_t *io_end = iocb->private;
3778         struct workqueue_struct *wq;
3779         unsigned long flags;
3780         struct ext4_inode_info *ei;
3781
3782         /* if not async direct IO or dio with 0 bytes write, just return */
3783         if (!io_end || !size)
3784                 return;
3785
3786         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3787                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3788                   iocb->private, io_end->inode->i_ino, iocb, offset,
3789                   size);
3790
3791         /* if not aio dio with unwritten extents, just free io and return */
3792         if (io_end->flag != EXT4_IO_UNWRITTEN){
3793                 ext4_free_io_end(io_end);
3794                 iocb->private = NULL;
3795                 return;
3796         }
3797
3798         io_end->offset = offset;
3799         io_end->size = size;
3800         io_end->flag = EXT4_IO_UNWRITTEN;
3801         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3802
3803         /* queue the work to convert unwritten extents to written */
3804         queue_work(wq, &io_end->work);
3805
3806         /* Add the io_end to per-inode completed aio dio list*/
3807         ei = EXT4_I(io_end->inode);
3808         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3809         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3810         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3811         iocb->private = NULL;
3812 }
3813
3814 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3815 {
3816         ext4_io_end_t *io_end = bh->b_private;
3817         struct workqueue_struct *wq;
3818         struct inode *inode;
3819         unsigned long flags;
3820
3821         if (!test_clear_buffer_uninit(bh) || !io_end)
3822                 goto out;
3823
3824         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3825                 printk("sb umounted, discard end_io request for inode %lu\n",
3826                         io_end->inode->i_ino);
3827                 ext4_free_io_end(io_end);
3828                 goto out;
3829         }
3830
3831         io_end->flag = EXT4_IO_UNWRITTEN;
3832         inode = io_end->inode;
3833
3834         /* Add the io_end to per-inode completed io list*/
3835         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3836         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3837         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3838
3839         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3840         /* queue the work to convert unwritten extents to written */
3841         queue_work(wq, &io_end->work);
3842 out:
3843         bh->b_private = NULL;
3844         bh->b_end_io = NULL;
3845         clear_buffer_uninit(bh);
3846         end_buffer_async_write(bh, uptodate);
3847 }
3848
3849 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3850 {
3851         ext4_io_end_t *io_end;
3852         struct page *page = bh->b_page;
3853         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3854         size_t size = bh->b_size;
3855
3856 retry:
3857         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3858         if (!io_end) {
3859                 if (printk_ratelimit())
3860                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3861                 schedule();
3862                 goto retry;
3863         }
3864         io_end->offset = offset;
3865         io_end->size = size;
3866         /*
3867          * We need to hold a reference to the page to make sure it
3868          * doesn't get evicted before ext4_end_io_work() has a chance
3869          * to convert the extent from written to unwritten.
3870          */
3871         io_end->page = page;
3872         get_page(io_end->page);
3873
3874         bh->b_private = io_end;
3875         bh->b_end_io = ext4_end_io_buffer_write;
3876         return 0;
3877 }
3878
3879 /*
3880  * For ext4 extent files, ext4 will do direct-io write to holes,
3881  * preallocated extents, and those write extend the file, no need to
3882  * fall back to buffered IO.
3883  *
3884  * For holes, we fallocate those blocks, mark them as unintialized
3885  * If those blocks were preallocated, we mark sure they are splited, but
3886  * still keep the range to write as unintialized.
3887  *
3888  * The unwrritten extents will be converted to written when DIO is completed.
3889  * For async direct IO, since the IO may still pending when return, we
3890  * set up an end_io call back function, which will do the convertion
3891  * when async direct IO completed.
3892  *
3893  * If the O_DIRECT write will extend the file then add this inode to the
3894  * orphan list.  So recovery will truncate it back to the original size
3895  * if the machine crashes during the write.
3896  *
3897  */
3898 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3899                               const struct iovec *iov, loff_t offset,
3900                               unsigned long nr_segs)
3901 {
3902         struct file *file = iocb->ki_filp;
3903         struct inode *inode = file->f_mapping->host;
3904         ssize_t ret;
3905         size_t count = iov_length(iov, nr_segs);
3906
3907         loff_t final_size = offset + count;
3908         if (rw == WRITE && final_size <= inode->i_size) {
3909                 /*
3910                  * We could direct write to holes and fallocate.
3911                  *
3912                  * Allocated blocks to fill the hole are marked as uninitialized
3913                  * to prevent paralel buffered read to expose the stale data
3914                  * before DIO complete the data IO.
3915                  *
3916                  * As to previously fallocated extents, ext4 get_block
3917                  * will just simply mark the buffer mapped but still
3918                  * keep the extents uninitialized.
3919                  *
3920                  * for non AIO case, we will convert those unwritten extents
3921                  * to written after return back from blockdev_direct_IO.
3922                  *
3923                  * for async DIO, the conversion needs to be defered when
3924                  * the IO is completed. The ext4 end_io callback function
3925                  * will be called to take care of the conversion work.
3926                  * Here for async case, we allocate an io_end structure to
3927                  * hook to the iocb.
3928                  */
3929                 iocb->private = NULL;
3930                 EXT4_I(inode)->cur_aio_dio = NULL;
3931                 if (!is_sync_kiocb(iocb)) {
3932                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3933                         if (!iocb->private)
3934                                 return -ENOMEM;
3935                         /*
3936                          * we save the io structure for current async
3937                          * direct IO, so that later ext4_get_blocks()
3938                          * could flag the io structure whether there
3939                          * is a unwritten extents needs to be converted
3940                          * when IO is completed.
3941                          */
3942                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3943                 }
3944
3945                 ret = blockdev_direct_IO(rw, iocb, inode,
3946                                          inode->i_sb->s_bdev, iov,
3947                                          offset, nr_segs,
3948                                          ext4_get_block_write,
3949                                          ext4_end_io_dio);
3950                 if (iocb->private)
3951                         EXT4_I(inode)->cur_aio_dio = NULL;
3952                 /*
3953                  * The io_end structure takes a reference to the inode,
3954                  * that structure needs to be destroyed and the
3955                  * reference to the inode need to be dropped, when IO is
3956                  * complete, even with 0 byte write, or failed.
3957                  *
3958                  * In the successful AIO DIO case, the io_end structure will be
3959                  * desctroyed and the reference to the inode will be dropped
3960                  * after the end_io call back function is called.
3961                  *
3962                  * In the case there is 0 byte write, or error case, since
3963                  * VFS direct IO won't invoke the end_io call back function,
3964                  * we need to free the end_io structure here.
3965                  */
3966                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3967                         ext4_free_io_end(iocb->private);
3968                         iocb->private = NULL;
3969                 } else if (ret > 0 && ext4_test_inode_state(inode,
3970                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3971                         int err;
3972                         /*
3973                          * for non AIO case, since the IO is already
3974                          * completed, we could do the convertion right here
3975                          */
3976                         err = ext4_convert_unwritten_extents(inode,
3977                                                              offset, ret);
3978                         if (err < 0)
3979                                 ret = err;
3980                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3981                 }
3982                 return ret;
3983         }
3984
3985         /* for write the the end of file case, we fall back to old way */
3986         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3987 }
3988
3989 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3990                               const struct iovec *iov, loff_t offset,
3991                               unsigned long nr_segs)
3992 {
3993         struct file *file = iocb->ki_filp;
3994         struct inode *inode = file->f_mapping->host;
3995
3996         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3997                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3998
3999         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4000 }
4001
4002 /*
4003  * Pages can be marked dirty completely asynchronously from ext4's journalling
4004  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
4005  * much here because ->set_page_dirty is called under VFS locks.  The page is
4006  * not necessarily locked.
4007  *
4008  * We cannot just dirty the page and leave attached buffers clean, because the
4009  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4010  * or jbddirty because all the journalling code will explode.
4011  *
4012  * So what we do is to mark the page "pending dirty" and next time writepage
4013  * is called, propagate that into the buffers appropriately.
4014  */
4015 static int ext4_journalled_set_page_dirty(struct page *page)
4016 {
4017         SetPageChecked(page);
4018         return __set_page_dirty_nobuffers(page);
4019 }
4020
4021 static const struct address_space_operations ext4_ordered_aops = {
4022         .readpage               = ext4_readpage,
4023         .readpages              = ext4_readpages,
4024         .writepage              = ext4_writepage,
4025         .sync_page              = block_sync_page,
4026         .write_begin            = ext4_write_begin,
4027         .write_end              = ext4_ordered_write_end,
4028         .bmap                   = ext4_bmap,
4029         .invalidatepage         = ext4_invalidatepage,
4030         .releasepage            = ext4_releasepage,
4031         .direct_IO              = ext4_direct_IO,
4032         .migratepage            = buffer_migrate_page,
4033         .is_partially_uptodate  = block_is_partially_uptodate,
4034         .error_remove_page      = generic_error_remove_page,
4035 };
4036
4037 static const struct address_space_operations ext4_writeback_aops = {
4038         .readpage               = ext4_readpage,
4039         .readpages              = ext4_readpages,
4040         .writepage              = ext4_writepage,
4041         .sync_page              = block_sync_page,
4042         .write_begin            = ext4_write_begin,
4043         .write_end              = ext4_writeback_write_end,
4044         .bmap                   = ext4_bmap,
4045         .invalidatepage         = ext4_invalidatepage,
4046         .releasepage            = ext4_releasepage,
4047         .direct_IO              = ext4_direct_IO,
4048         .migratepage            = buffer_migrate_page,
4049         .is_partially_uptodate  = block_is_partially_uptodate,
4050         .error_remove_page      = generic_error_remove_page,
4051 };
4052
4053 static const struct address_space_operations ext4_journalled_aops = {
4054         .readpage               = ext4_readpage,
4055         .readpages              = ext4_readpages,
4056         .writepage              = ext4_writepage,
4057         .sync_page              = block_sync_page,
4058         .write_begin            = ext4_write_begin,
4059         .write_end              = ext4_journalled_write_end,
4060         .set_page_dirty         = ext4_journalled_set_page_dirty,
4061         .bmap                   = ext4_bmap,
4062         .invalidatepage         = ext4_invalidatepage,
4063         .releasepage            = ext4_releasepage,
4064         .is_partially_uptodate  = block_is_partially_uptodate,
4065         .error_remove_page      = generic_error_remove_page,
4066 };
4067
4068 static const struct address_space_operations ext4_da_aops = {
4069         .readpage               = ext4_readpage,
4070         .readpages              = ext4_readpages,
4071         .writepage              = ext4_writepage,
4072         .writepages             = ext4_da_writepages,
4073         .sync_page              = block_sync_page,
4074         .write_begin            = ext4_da_write_begin,
4075         .write_end              = ext4_da_write_end,
4076         .bmap                   = ext4_bmap,
4077         .invalidatepage         = ext4_da_invalidatepage,
4078         .releasepage            = ext4_releasepage,
4079         .direct_IO              = ext4_direct_IO,
4080         .migratepage            = buffer_migrate_page,
4081         .is_partially_uptodate  = block_is_partially_uptodate,
4082         .error_remove_page      = generic_error_remove_page,
4083 };
4084
4085 void ext4_set_aops(struct inode *inode)
4086 {
4087         if (ext4_should_order_data(inode) &&
4088                 test_opt(inode->i_sb, DELALLOC))
4089                 inode->i_mapping->a_ops = &ext4_da_aops;
4090         else if (ext4_should_order_data(inode))
4091                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4092         else if (ext4_should_writeback_data(inode) &&
4093                  test_opt(inode->i_sb, DELALLOC))
4094                 inode->i_mapping->a_ops = &ext4_da_aops;
4095         else if (ext4_should_writeback_data(inode))
4096                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4097         else
4098                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4099 }
4100
4101 /*
4102  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4103  * up to the end of the block which corresponds to `from'.
4104  * This required during truncate. We need to physically zero the tail end
4105  * of that block so it doesn't yield old data if the file is later grown.
4106  */
4107 int ext4_block_truncate_page(handle_t *handle,
4108                 struct address_space *mapping, loff_t from)
4109 {
4110         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4111         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4112         unsigned blocksize, length, pos;
4113         ext4_lblk_t iblock;
4114         struct inode *inode = mapping->host;
4115         struct buffer_head *bh;
4116         struct page *page;
4117         int err = 0;
4118
4119         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4120                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4121         if (!page)
4122                 return -EINVAL;
4123
4124         blocksize = inode->i_sb->s_blocksize;
4125         length = blocksize - (offset & (blocksize - 1));
4126         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4127
4128         /*
4129          * For "nobh" option,  we can only work if we don't need to
4130          * read-in the page - otherwise we create buffers to do the IO.
4131          */
4132         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
4133              ext4_should_writeback_data(inode) && PageUptodate(page)) {
4134                 zero_user(page, offset, length);
4135                 set_page_dirty(page);
4136                 goto unlock;
4137         }
4138
4139         if (!page_has_buffers(page))
4140                 create_empty_buffers(page, blocksize, 0);
4141
4142         /* Find the buffer that contains "offset" */
4143         bh = page_buffers(page);
4144         pos = blocksize;
4145         while (offset >= pos) {
4146                 bh = bh->b_this_page;
4147                 iblock++;
4148                 pos += blocksize;
4149         }
4150
4151         err = 0;
4152         if (buffer_freed(bh)) {
4153                 BUFFER_TRACE(bh, "freed: skip");
4154                 goto unlock;
4155         }
4156
4157         if (!buffer_mapped(bh)) {
4158                 BUFFER_TRACE(bh, "unmapped");
4159                 ext4_get_block(inode, iblock, bh, 0);
4160                 /* unmapped? It's a hole - nothing to do */
4161                 if (!buffer_mapped(bh)) {
4162                         BUFFER_TRACE(bh, "still unmapped");
4163                         goto unlock;
4164                 }
4165         }
4166
4167         /* Ok, it's mapped. Make sure it's up-to-date */
4168         if (PageUptodate(page))
4169                 set_buffer_uptodate(bh);
4170
4171         if (!buffer_uptodate(bh)) {
4172                 err = -EIO;
4173                 ll_rw_block(READ, 1, &bh);
4174                 wait_on_buffer(bh);
4175                 /* Uhhuh. Read error. Complain and punt. */
4176                 if (!buffer_uptodate(bh))
4177                         goto unlock;
4178         }
4179
4180         if (ext4_should_journal_data(inode)) {
4181                 BUFFER_TRACE(bh, "get write access");
4182                 err = ext4_journal_get_write_access(handle, bh);
4183                 if (err)
4184                         goto unlock;
4185         }
4186
4187         zero_user(page, offset, length);
4188
4189         BUFFER_TRACE(bh, "zeroed end of block");
4190
4191         err = 0;
4192         if (ext4_should_journal_data(inode)) {
4193                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4194         } else {
4195                 if (ext4_should_order_data(inode))
4196                         err = ext4_jbd2_file_inode(handle, inode);
4197                 mark_buffer_dirty(bh);
4198         }
4199
4200 unlock:
4201         unlock_page(page);
4202         page_cache_release(page);
4203         return err;
4204 }
4205
4206 /*
4207  * Probably it should be a library function... search for first non-zero word
4208  * or memcmp with zero_page, whatever is better for particular architecture.
4209  * Linus?
4210  */
4211 static inline int all_zeroes(__le32 *p, __le32 *q)
4212 {
4213         while (p < q)
4214                 if (*p++)
4215                         return 0;
4216         return 1;
4217 }
4218
4219 /**
4220  *      ext4_find_shared - find the indirect blocks for partial truncation.
4221  *      @inode:   inode in question
4222  *      @depth:   depth of the affected branch
4223  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4224  *      @chain:   place to store the pointers to partial indirect blocks
4225  *      @top:     place to the (detached) top of branch
4226  *
4227  *      This is a helper function used by ext4_truncate().
4228  *
4229  *      When we do truncate() we may have to clean the ends of several
4230  *      indirect blocks but leave the blocks themselves alive. Block is
4231  *      partially truncated if some data below the new i_size is refered
4232  *      from it (and it is on the path to the first completely truncated
4233  *      data block, indeed).  We have to free the top of that path along
4234  *      with everything to the right of the path. Since no allocation
4235  *      past the truncation point is possible until ext4_truncate()
4236  *      finishes, we may safely do the latter, but top of branch may
4237  *      require special attention - pageout below the truncation point
4238  *      might try to populate it.
4239  *
4240  *      We atomically detach the top of branch from the tree, store the
4241  *      block number of its root in *@top, pointers to buffer_heads of
4242  *      partially truncated blocks - in @chain[].bh and pointers to
4243  *      their last elements that should not be removed - in
4244  *      @chain[].p. Return value is the pointer to last filled element
4245  *      of @chain.
4246  *
4247  *      The work left to caller to do the actual freeing of subtrees:
4248  *              a) free the subtree starting from *@top
4249  *              b) free the subtrees whose roots are stored in
4250  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4251  *              c) free the subtrees growing from the inode past the @chain[0].
4252  *                      (no partially truncated stuff there).  */
4253
4254 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4255                                   ext4_lblk_t offsets[4], Indirect chain[4],
4256                                   __le32 *top)
4257 {
4258         Indirect *partial, *p;
4259         int k, err;
4260
4261         *top = 0;
4262         /* Make k index the deepest non-null offset + 1 */
4263         for (k = depth; k > 1 && !offsets[k-1]; k--)
4264                 ;
4265         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4266         /* Writer: pointers */
4267         if (!partial)
4268                 partial = chain + k-1;
4269         /*
4270          * If the branch acquired continuation since we've looked at it -
4271          * fine, it should all survive and (new) top doesn't belong to us.
4272          */
4273         if (!partial->key && *partial->p)
4274                 /* Writer: end */
4275                 goto no_top;
4276         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4277                 ;
4278         /*
4279          * OK, we've found the last block that must survive. The rest of our
4280          * branch should be detached before unlocking. However, if that rest
4281          * of branch is all ours and does not grow immediately from the inode
4282          * it's easier to cheat and just decrement partial->p.
4283          */
4284         if (p == chain + k - 1 && p > chain) {
4285                 p->p--;
4286         } else {
4287                 *top = *p->p;
4288                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4289 #if 0
4290                 *p->p = 0;
4291 #endif
4292         }
4293         /* Writer: end */
4294
4295         while (partial > p) {
4296                 brelse(partial->bh);
4297                 partial--;
4298         }
4299 no_top:
4300         return partial;
4301 }
4302
4303 /*
4304  * Zero a number of block pointers in either an inode or an indirect block.
4305  * If we restart the transaction we must again get write access to the
4306  * indirect block for further modification.
4307  *
4308  * We release `count' blocks on disk, but (last - first) may be greater
4309  * than `count' because there can be holes in there.
4310  */
4311 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4312                              struct buffer_head *bh,
4313                              ext4_fsblk_t block_to_free,
4314                              unsigned long count, __le32 *first,
4315                              __le32 *last)
4316 {
4317         __le32 *p;
4318         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4319
4320         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4321                 flags |= EXT4_FREE_BLOCKS_METADATA;
4322
4323         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4324                                    count)) {
4325                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4326                                  "blocks %llu len %lu",
4327                                  (unsigned long long) block_to_free, count);
4328                 return 1;
4329         }
4330
4331         if (try_to_extend_transaction(handle, inode)) {
4332                 if (bh) {
4333                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4334                         ext4_handle_dirty_metadata(handle, inode, bh);
4335                 }
4336                 ext4_mark_inode_dirty(handle, inode);
4337                 ext4_truncate_restart_trans(handle, inode,
4338                                             blocks_for_truncate(inode));
4339                 if (bh) {
4340                         BUFFER_TRACE(bh, "retaking write access");
4341                         ext4_journal_get_write_access(handle, bh);
4342                 }
4343         }
4344
4345         for (p = first; p < last; p++)
4346                 *p = 0;
4347
4348         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4349         return 0;
4350 }
4351
4352 /**
4353  * ext4_free_data - free a list of data blocks
4354  * @handle:     handle for this transaction
4355  * @inode:      inode we are dealing with
4356  * @this_bh:    indirect buffer_head which contains *@first and *@last
4357  * @first:      array of block numbers
4358  * @last:       points immediately past the end of array
4359  *
4360  * We are freeing all blocks refered from that array (numbers are stored as
4361  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4362  *
4363  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4364  * blocks are contiguous then releasing them at one time will only affect one
4365  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4366  * actually use a lot of journal space.
4367  *
4368  * @this_bh will be %NULL if @first and @last point into the inode's direct
4369  * block pointers.
4370  */
4371 static void ext4_free_data(handle_t *handle, struct inode *inode,
4372                            struct buffer_head *this_bh,
4373                            __le32 *first, __le32 *last)
4374 {
4375         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4376         unsigned long count = 0;            /* Number of blocks in the run */
4377         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4378                                                corresponding to
4379                                                block_to_free */
4380         ext4_fsblk_t nr;                    /* Current block # */
4381         __le32 *p;                          /* Pointer into inode/ind
4382                                                for current block */
4383         int err;
4384
4385         if (this_bh) {                          /* For indirect block */
4386                 BUFFER_TRACE(this_bh, "get_write_access");
4387                 err = ext4_journal_get_write_access(handle, this_bh);
4388                 /* Important: if we can't update the indirect pointers
4389                  * to the blocks, we can't free them. */
4390                 if (err)
4391                         return;
4392         }
4393
4394         for (p = first; p < last; p++) {
4395                 nr = le32_to_cpu(*p);
4396                 if (nr) {
4397                         /* accumulate blocks to free if they're contiguous */
4398                         if (count == 0) {
4399                                 block_to_free = nr;
4400                                 block_to_free_p = p;
4401                                 count = 1;
4402                         } else if (nr == block_to_free + count) {
4403                                 count++;
4404                         } else {
4405                                 if (ext4_clear_blocks(handle, inode, this_bh,
4406                                                       block_to_free, count,
4407                                                       block_to_free_p, p))
4408                                         break;
4409                                 block_to_free = nr;
4410                                 block_to_free_p = p;
4411                                 count = 1;
4412                         }
4413                 }
4414         }
4415
4416         if (count > 0)
4417                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4418                                   count, block_to_free_p, p);
4419
4420         if (this_bh) {
4421                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4422
4423                 /*
4424                  * The buffer head should have an attached journal head at this
4425                  * point. However, if the data is corrupted and an indirect
4426                  * block pointed to itself, it would have been detached when
4427                  * the block was cleared. Check for this instead of OOPSing.
4428                  */
4429                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4430                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4431                 else
4432                         EXT4_ERROR_INODE(inode,
4433                                          "circular indirect block detected at "
4434                                          "block %llu",
4435                                 (unsigned long long) this_bh->b_blocknr);
4436         }
4437 }
4438
4439 /**
4440  *      ext4_free_branches - free an array of branches
4441  *      @handle: JBD handle for this transaction
4442  *      @inode: inode we are dealing with
4443  *      @parent_bh: the buffer_head which contains *@first and *@last
4444  *      @first: array of block numbers
4445  *      @last:  pointer immediately past the end of array
4446  *      @depth: depth of the branches to free
4447  *
4448  *      We are freeing all blocks refered from these branches (numbers are
4449  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4450  *      appropriately.
4451  */
4452 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4453                                struct buffer_head *parent_bh,
4454                                __le32 *first, __le32 *last, int depth)
4455 {
4456         ext4_fsblk_t nr;
4457         __le32 *p;
4458
4459         if (ext4_handle_is_aborted(handle))
4460                 return;
4461
4462         if (depth--) {
4463                 struct buffer_head *bh;
4464                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4465                 p = last;
4466                 while (--p >= first) {
4467                         nr = le32_to_cpu(*p);
4468                         if (!nr)
4469                                 continue;               /* A hole */
4470
4471                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4472                                                    nr, 1)) {
4473                                 EXT4_ERROR_INODE(inode,
4474                                                  "invalid indirect mapped "
4475                                                  "block %lu (level %d)",
4476                                                  (unsigned long) nr, depth);
4477                                 break;
4478                         }
4479
4480                         /* Go read the buffer for the next level down */
4481                         bh = sb_bread(inode->i_sb, nr);
4482
4483                         /*
4484                          * A read failure? Report error and clear slot
4485                          * (should be rare).
4486                          */
4487                         if (!bh) {
4488                                 EXT4_ERROR_INODE(inode,
4489                                                  "Read failure block=%llu",
4490                                                  (unsigned long long) nr);
4491                                 continue;
4492                         }
4493
4494                         /* This zaps the entire block.  Bottom up. */
4495                         BUFFER_TRACE(bh, "free child branches");
4496                         ext4_free_branches(handle, inode, bh,
4497                                         (__le32 *) bh->b_data,
4498                                         (__le32 *) bh->b_data + addr_per_block,
4499                                         depth);
4500
4501                         /*
4502                          * We've probably journalled the indirect block several
4503                          * times during the truncate.  But it's no longer
4504                          * needed and we now drop it from the transaction via
4505                          * jbd2_journal_revoke().
4506                          *
4507                          * That's easy if it's exclusively part of this
4508                          * transaction.  But if it's part of the committing
4509                          * transaction then jbd2_journal_forget() will simply
4510                          * brelse() it.  That means that if the underlying
4511                          * block is reallocated in ext4_get_block(),
4512                          * unmap_underlying_metadata() will find this block
4513                          * and will try to get rid of it.  damn, damn.
4514                          *
4515                          * If this block has already been committed to the
4516                          * journal, a revoke record will be written.  And
4517                          * revoke records must be emitted *before* clearing
4518                          * this block's bit in the bitmaps.
4519                          */
4520                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4521
4522                         /*
4523                          * Everything below this this pointer has been
4524                          * released.  Now let this top-of-subtree go.
4525                          *
4526                          * We want the freeing of this indirect block to be
4527                          * atomic in the journal with the updating of the
4528                          * bitmap block which owns it.  So make some room in
4529                          * the journal.
4530                          *
4531                          * We zero the parent pointer *after* freeing its
4532                          * pointee in the bitmaps, so if extend_transaction()
4533                          * for some reason fails to put the bitmap changes and
4534                          * the release into the same transaction, recovery
4535                          * will merely complain about releasing a free block,
4536                          * rather than leaking blocks.
4537                          */
4538                         if (ext4_handle_is_aborted(handle))
4539                                 return;
4540                         if (try_to_extend_transaction(handle, inode)) {
4541                                 ext4_mark_inode_dirty(handle, inode);
4542                                 ext4_truncate_restart_trans(handle, inode,
4543                                             blocks_for_truncate(inode));
4544                         }
4545
4546                         ext4_free_blocks(handle, inode, 0, nr, 1,
4547                                          EXT4_FREE_BLOCKS_METADATA);
4548
4549                         if (parent_bh) {
4550                                 /*
4551                                  * The block which we have just freed is
4552                                  * pointed to by an indirect block: journal it
4553                                  */
4554                                 BUFFER_TRACE(parent_bh, "get_write_access");
4555                                 if (!ext4_journal_get_write_access(handle,
4556                                                                    parent_bh)){
4557                                         *p = 0;
4558                                         BUFFER_TRACE(parent_bh,
4559                                         "call ext4_handle_dirty_metadata");
4560                                         ext4_handle_dirty_metadata(handle,
4561                                                                    inode,
4562                                                                    parent_bh);
4563                                 }
4564                         }
4565                 }
4566         } else {
4567                 /* We have reached the bottom of the tree. */
4568                 BUFFER_TRACE(parent_bh, "free data blocks");
4569                 ext4_free_data(handle, inode, parent_bh, first, last);
4570         }
4571 }
4572
4573 int ext4_can_truncate(struct inode *inode)
4574 {
4575         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4576                 return 0;
4577         if (S_ISREG(inode->i_mode))
4578                 return 1;
4579         if (S_ISDIR(inode->i_mode))
4580                 return 1;
4581         if (S_ISLNK(inode->i_mode))
4582                 return !ext4_inode_is_fast_symlink(inode);
4583         return 0;
4584 }
4585
4586 /*
4587  * ext4_truncate()
4588  *
4589  * We block out ext4_get_block() block instantiations across the entire
4590  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4591  * simultaneously on behalf of the same inode.
4592  *
4593  * As we work through the truncate and commmit bits of it to the journal there
4594  * is one core, guiding principle: the file's tree must always be consistent on
4595  * disk.  We must be able to restart the truncate after a crash.
4596  *
4597  * The file's tree may be transiently inconsistent in memory (although it
4598  * probably isn't), but whenever we close off and commit a journal transaction,
4599  * the contents of (the filesystem + the journal) must be consistent and
4600  * restartable.  It's pretty simple, really: bottom up, right to left (although
4601  * left-to-right works OK too).
4602  *
4603  * Note that at recovery time, journal replay occurs *before* the restart of
4604  * truncate against the orphan inode list.
4605  *
4606  * The committed inode has the new, desired i_size (which is the same as
4607  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4608  * that this inode's truncate did not complete and it will again call
4609  * ext4_truncate() to have another go.  So there will be instantiated blocks
4610  * to the right of the truncation point in a crashed ext4 filesystem.  But
4611  * that's fine - as long as they are linked from the inode, the post-crash
4612  * ext4_truncate() run will find them and release them.
4613  */
4614 void ext4_truncate(struct inode *inode)
4615 {
4616         handle_t *handle;
4617         struct ext4_inode_info *ei = EXT4_I(inode);
4618         __le32 *i_data = ei->i_data;
4619         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4620         struct address_space *mapping = inode->i_mapping;
4621         ext4_lblk_t offsets[4];
4622         Indirect chain[4];
4623         Indirect *partial;
4624         __le32 nr = 0;
4625         int n;
4626         ext4_lblk_t last_block;
4627         unsigned blocksize = inode->i_sb->s_blocksize;
4628
4629         if (!ext4_can_truncate(inode))
4630                 return;
4631
4632         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4633
4634         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4635                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4636
4637         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4638                 ext4_ext_truncate(inode);
4639                 return;
4640         }
4641
4642         handle = start_transaction(inode);
4643         if (IS_ERR(handle))
4644                 return;         /* AKPM: return what? */
4645
4646         last_block = (inode->i_size + blocksize-1)
4647                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4648
4649         if (inode->i_size & (blocksize - 1))
4650                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4651                         goto out_stop;
4652
4653         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4654         if (n == 0)
4655                 goto out_stop;  /* error */
4656
4657         /*
4658          * OK.  This truncate is going to happen.  We add the inode to the
4659          * orphan list, so that if this truncate spans multiple transactions,
4660          * and we crash, we will resume the truncate when the filesystem
4661          * recovers.  It also marks the inode dirty, to catch the new size.
4662          *
4663          * Implication: the file must always be in a sane, consistent
4664          * truncatable state while each transaction commits.
4665          */
4666         if (ext4_orphan_add(handle, inode))
4667                 goto out_stop;
4668
4669         /*
4670          * From here we block out all ext4_get_block() callers who want to
4671          * modify the block allocation tree.
4672          */
4673         down_write(&ei->i_data_sem);
4674
4675         ext4_discard_preallocations(inode);
4676
4677         /*
4678          * The orphan list entry will now protect us from any crash which
4679          * occurs before the truncate completes, so it is now safe to propagate
4680          * the new, shorter inode size (held for now in i_size) into the
4681          * on-disk inode. We do this via i_disksize, which is the value which
4682          * ext4 *really* writes onto the disk inode.
4683          */
4684         ei->i_disksize = inode->i_size;
4685
4686         if (n == 1) {           /* direct blocks */
4687                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4688                                i_data + EXT4_NDIR_BLOCKS);
4689                 goto do_indirects;
4690         }
4691
4692         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4693         /* Kill the top of shared branch (not detached) */
4694         if (nr) {
4695                 if (partial == chain) {
4696                         /* Shared branch grows from the inode */
4697                         ext4_free_branches(handle, inode, NULL,
4698                                            &nr, &nr+1, (chain+n-1) - partial);
4699                         *partial->p = 0;
4700                         /*
4701                          * We mark the inode dirty prior to restart,
4702                          * and prior to stop.  No need for it here.
4703                          */
4704                 } else {
4705                         /* Shared branch grows from an indirect block */
4706                         BUFFER_TRACE(partial->bh, "get_write_access");
4707                         ext4_free_branches(handle, inode, partial->bh,
4708                                         partial->p,
4709                                         partial->p+1, (chain+n-1) - partial);
4710                 }
4711         }
4712         /* Clear the ends of indirect blocks on the shared branch */
4713         while (partial > chain) {
4714                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4715                                    (__le32*)partial->bh->b_data+addr_per_block,
4716                                    (chain+n-1) - partial);
4717                 BUFFER_TRACE(partial->bh, "call brelse");
4718                 brelse(partial->bh);
4719                 partial--;
4720         }
4721 do_indirects:
4722         /* Kill the remaining (whole) subtrees */
4723         switch (offsets[0]) {
4724         default:
4725                 nr = i_data[EXT4_IND_BLOCK];
4726                 if (nr) {
4727                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4728                         i_data[EXT4_IND_BLOCK] = 0;
4729                 }
4730         case EXT4_IND_BLOCK:
4731                 nr = i_data[EXT4_DIND_BLOCK];
4732                 if (nr) {
4733                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4734                         i_data[EXT4_DIND_BLOCK] = 0;
4735                 }
4736         case EXT4_DIND_BLOCK:
4737                 nr = i_data[EXT4_TIND_BLOCK];
4738                 if (nr) {
4739                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4740                         i_data[EXT4_TIND_BLOCK] = 0;
4741                 }
4742         case EXT4_TIND_BLOCK:
4743                 ;
4744         }
4745
4746         up_write(&ei->i_data_sem);
4747         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4748         ext4_mark_inode_dirty(handle, inode);
4749
4750         /*
4751          * In a multi-transaction truncate, we only make the final transaction
4752          * synchronous
4753          */
4754         if (IS_SYNC(inode))
4755                 ext4_handle_sync(handle);
4756 out_stop:
4757         /*
4758          * If this was a simple ftruncate(), and the file will remain alive
4759          * then we need to clear up the orphan record which we created above.
4760          * However, if this was a real unlink then we were called by
4761          * ext4_delete_inode(), and we allow that function to clean up the
4762          * orphan info for us.
4763          */
4764         if (inode->i_nlink)
4765                 ext4_orphan_del(handle, inode);
4766
4767         ext4_journal_stop(handle);
4768 }
4769
4770 /*
4771  * ext4_get_inode_loc returns with an extra refcount against the inode's
4772  * underlying buffer_head on success. If 'in_mem' is true, we have all
4773  * data in memory that is needed to recreate the on-disk version of this
4774  * inode.
4775  */
4776 static int __ext4_get_inode_loc(struct inode *inode,
4777                                 struct ext4_iloc *iloc, int in_mem)
4778 {
4779         struct ext4_group_desc  *gdp;
4780         struct buffer_head      *bh;
4781         struct super_block      *sb = inode->i_sb;
4782         ext4_fsblk_t            block;
4783         int                     inodes_per_block, inode_offset;
4784
4785         iloc->bh = NULL;
4786         if (!ext4_valid_inum(sb, inode->i_ino))
4787                 return -EIO;
4788
4789         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4790         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4791         if (!gdp)
4792                 return -EIO;
4793
4794         /*
4795          * Figure out the offset within the block group inode table
4796          */
4797         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4798         inode_offset = ((inode->i_ino - 1) %
4799                         EXT4_INODES_PER_GROUP(sb));
4800         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4801         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4802
4803         bh = sb_getblk(sb, block);
4804         if (!bh) {
4805                 EXT4_ERROR_INODE(inode, "unable to read inode block - "
4806                                  "block %llu", block);
4807                 return -EIO;
4808         }
4809         if (!buffer_uptodate(bh)) {
4810                 lock_buffer(bh);
4811
4812                 /*
4813                  * If the buffer has the write error flag, we have failed
4814                  * to write out another inode in the same block.  In this
4815                  * case, we don't have to read the block because we may
4816                  * read the old inode data successfully.
4817                  */
4818                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4819                         set_buffer_uptodate(bh);
4820
4821                 if (buffer_uptodate(bh)) {
4822                         /* someone brought it uptodate while we waited */
4823                         unlock_buffer(bh);
4824                         goto has_buffer;
4825                 }
4826
4827                 /*
4828                  * If we have all information of the inode in memory and this
4829                  * is the only valid inode in the block, we need not read the
4830                  * block.
4831                  */
4832                 if (in_mem) {
4833                         struct buffer_head *bitmap_bh;
4834                         int i, start;
4835
4836                         start = inode_offset & ~(inodes_per_block - 1);
4837
4838                         /* Is the inode bitmap in cache? */
4839                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4840                         if (!bitmap_bh)
4841                                 goto make_io;
4842
4843                         /*
4844                          * If the inode bitmap isn't in cache then the
4845                          * optimisation may end up performing two reads instead
4846                          * of one, so skip it.
4847                          */
4848                         if (!buffer_uptodate(bitmap_bh)) {
4849                                 brelse(bitmap_bh);
4850                                 goto make_io;
4851                         }
4852                         for (i = start; i < start + inodes_per_block; i++) {
4853                                 if (i == inode_offset)
4854                                         continue;
4855                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4856                                         break;
4857                         }
4858                         brelse(bitmap_bh);
4859                         if (i == start + inodes_per_block) {
4860                                 /* all other inodes are free, so skip I/O */
4861                                 memset(bh->b_data, 0, bh->b_size);
4862                                 set_buffer_uptodate(bh);
4863                                 unlock_buffer(bh);
4864                                 goto has_buffer;
4865                         }
4866                 }
4867
4868 make_io:
4869                 /*
4870                  * If we need to do any I/O, try to pre-readahead extra
4871                  * blocks from the inode table.
4872                  */
4873                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4874                         ext4_fsblk_t b, end, table;
4875                         unsigned num;
4876
4877                         table = ext4_inode_table(sb, gdp);
4878                         /* s_inode_readahead_blks is always a power of 2 */
4879                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4880                         if (table > b)
4881                                 b = table;
4882                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4883                         num = EXT4_INODES_PER_GROUP(sb);
4884                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4885                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4886                                 num -= ext4_itable_unused_count(sb, gdp);
4887                         table += num / inodes_per_block;
4888                         if (end > table)
4889                                 end = table;
4890                         while (b <= end)
4891                                 sb_breadahead(sb, b++);
4892                 }
4893
4894                 /*
4895                  * There are other valid inodes in the buffer, this inode
4896                  * has in-inode xattrs, or we don't have this inode in memory.
4897                  * Read the block from disk.
4898                  */
4899                 get_bh(bh);
4900                 bh->b_end_io = end_buffer_read_sync;
4901                 submit_bh(READ_META, bh);
4902                 wait_on_buffer(bh);
4903                 if (!buffer_uptodate(bh)) {
4904                         EXT4_ERROR_INODE(inode, "unable to read inode "
4905                                          "block %llu", block);
4906                         brelse(bh);
4907                         return -EIO;
4908                 }
4909         }
4910 has_buffer:
4911         iloc->bh = bh;
4912         return 0;
4913 }
4914
4915 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4916 {
4917         /* We have all inode data except xattrs in memory here. */
4918         return __ext4_get_inode_loc(inode, iloc,
4919                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4920 }
4921
4922 void ext4_set_inode_flags(struct inode *inode)
4923 {
4924         unsigned int flags = EXT4_I(inode)->i_flags;
4925
4926         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4927         if (flags & EXT4_SYNC_FL)
4928                 inode->i_flags |= S_SYNC;
4929         if (flags & EXT4_APPEND_FL)
4930                 inode->i_flags |= S_APPEND;
4931         if (flags & EXT4_IMMUTABLE_FL)
4932                 inode->i_flags |= S_IMMUTABLE;
4933         if (flags & EXT4_NOATIME_FL)
4934                 inode->i_flags |= S_NOATIME;
4935         if (flags & EXT4_DIRSYNC_FL)
4936                 inode->i_flags |= S_DIRSYNC;
4937 }
4938
4939 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4940 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4941 {
4942         unsigned int vfs_fl;
4943         unsigned long old_fl, new_fl;
4944
4945         do {
4946                 vfs_fl = ei->vfs_inode.i_flags;
4947                 old_fl = ei->i_flags;
4948                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4949                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4950                                 EXT4_DIRSYNC_FL);
4951                 if (vfs_fl & S_SYNC)
4952                         new_fl |= EXT4_SYNC_FL;
4953                 if (vfs_fl & S_APPEND)
4954                         new_fl |= EXT4_APPEND_FL;
4955                 if (vfs_fl & S_IMMUTABLE)
4956                         new_fl |= EXT4_IMMUTABLE_FL;
4957                 if (vfs_fl & S_NOATIME)
4958                         new_fl |= EXT4_NOATIME_FL;
4959                 if (vfs_fl & S_DIRSYNC)
4960                         new_fl |= EXT4_DIRSYNC_FL;
4961         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4962 }
4963
4964 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4965                                   struct ext4_inode_info *ei)
4966 {
4967         blkcnt_t i_blocks ;
4968         struct inode *inode = &(ei->vfs_inode);
4969         struct super_block *sb = inode->i_sb;
4970
4971         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4972                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4973                 /* we are using combined 48 bit field */
4974                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4975                                         le32_to_cpu(raw_inode->i_blocks_lo);
4976                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4977                         /* i_blocks represent file system block size */
4978                         return i_blocks  << (inode->i_blkbits - 9);
4979                 } else {
4980                         return i_blocks;
4981                 }
4982         } else {
4983                 return le32_to_cpu(raw_inode->i_blocks_lo);
4984         }
4985 }
4986
4987 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4988 {
4989         struct ext4_iloc iloc;
4990         struct ext4_inode *raw_inode;
4991         struct ext4_inode_info *ei;
4992         struct inode *inode;
4993         journal_t *journal = EXT4_SB(sb)->s_journal;
4994         long ret;
4995         int block;
4996
4997         inode = iget_locked(sb, ino);
4998         if (!inode)
4999                 return ERR_PTR(-ENOMEM);
5000         if (!(inode->i_state & I_NEW))
5001                 return inode;
5002
5003         ei = EXT4_I(inode);
5004         iloc.bh = 0;
5005
5006         ret = __ext4_get_inode_loc(inode, &iloc, 0);
5007         if (ret < 0)
5008                 goto bad_inode;
5009         raw_inode = ext4_raw_inode(&iloc);
5010         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
5011         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
5012         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
5013         if (!(test_opt(inode->i_sb, NO_UID32))) {
5014                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5015                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5016         }
5017         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5018
5019         ei->i_state_flags = 0;
5020         ei->i_dir_start_lookup = 0;
5021         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5022         /* We now have enough fields to check if the inode was active or not.
5023          * This is needed because nfsd might try to access dead inodes
5024          * the test is that same one that e2fsck uses
5025          * NeilBrown 1999oct15
5026          */
5027         if (inode->i_nlink == 0) {
5028                 if (inode->i_mode == 0 ||
5029                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5030                         /* this inode is deleted */
5031                         ret = -ESTALE;
5032                         goto bad_inode;
5033                 }
5034                 /* The only unlinked inodes we let through here have
5035                  * valid i_mode and are being read by the orphan
5036                  * recovery code: that's fine, we're about to complete
5037                  * the process of deleting those. */
5038         }
5039         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5040         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5041         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5042         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5043                 ei->i_file_acl |=
5044                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5045         inode->i_size = ext4_isize(raw_inode);
5046         ei->i_disksize = inode->i_size;
5047 #ifdef CONFIG_QUOTA
5048         ei->i_reserved_quota = 0;
5049 #endif
5050         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5051         ei->i_block_group = iloc.block_group;
5052         ei->i_last_alloc_group = ~0;
5053         /*
5054          * NOTE! The in-memory inode i_data array is in little-endian order
5055          * even on big-endian machines: we do NOT byteswap the block numbers!
5056          */
5057         for (block = 0; block < EXT4_N_BLOCKS; block++)
5058                 ei->i_data[block] = raw_inode->i_block[block];
5059         INIT_LIST_HEAD(&ei->i_orphan);
5060
5061         /*
5062          * Set transaction id's of transactions that have to be committed
5063          * to finish f[data]sync. We set them to currently running transaction
5064          * as we cannot be sure that the inode or some of its metadata isn't
5065          * part of the transaction - the inode could have been reclaimed and
5066          * now it is reread from disk.
5067          */
5068         if (journal) {
5069                 transaction_t *transaction;
5070                 tid_t tid;
5071
5072                 spin_lock(&journal->j_state_lock);
5073                 if (journal->j_running_transaction)
5074                         transaction = journal->j_running_transaction;
5075                 else
5076                         transaction = journal->j_committing_transaction;
5077                 if (transaction)
5078                         tid = transaction->t_tid;
5079                 else
5080                         tid = journal->j_commit_sequence;
5081                 spin_unlock(&journal->j_state_lock);
5082                 ei->i_sync_tid = tid;
5083                 ei->i_datasync_tid = tid;
5084         }
5085
5086         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5087                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5088                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5089                     EXT4_INODE_SIZE(inode->i_sb)) {
5090                         ret = -EIO;
5091                         goto bad_inode;
5092                 }
5093                 if (ei->i_extra_isize == 0) {
5094                         /* The extra space is currently unused. Use it. */
5095                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5096                                             EXT4_GOOD_OLD_INODE_SIZE;
5097                 } else {
5098                         __le32 *magic = (void *)raw_inode +
5099                                         EXT4_GOOD_OLD_INODE_SIZE +
5100                                         ei->i_extra_isize;
5101                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5102                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5103                 }
5104         } else
5105                 ei->i_extra_isize = 0;
5106
5107         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5108         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5109         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5110         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5111
5112         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5113         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5114                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5115                         inode->i_version |=
5116                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5117         }
5118
5119         ret = 0;
5120         if (ei->i_file_acl &&
5121             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5122                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
5123                                  ei->i_file_acl);
5124                 ret = -EIO;
5125                 goto bad_inode;
5126         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5127                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5128                     (S_ISLNK(inode->i_mode) &&
5129                      !ext4_inode_is_fast_symlink(inode)))
5130                         /* Validate extent which is part of inode */
5131                         ret = ext4_ext_check_inode(inode);
5132         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5133                    (S_ISLNK(inode->i_mode) &&
5134                     !ext4_inode_is_fast_symlink(inode))) {
5135                 /* Validate block references which are part of inode */
5136                 ret = ext4_check_inode_blockref(inode);
5137         }
5138         if (ret)
5139                 goto bad_inode;
5140
5141         if (S_ISREG(inode->i_mode)) {
5142                 inode->i_op = &ext4_file_inode_operations;
5143                 inode->i_fop = &ext4_file_operations;
5144                 ext4_set_aops(inode);
5145         } else if (S_ISDIR(inode->i_mode)) {
5146                 inode->i_op = &ext4_dir_inode_operations;
5147                 inode->i_fop = &ext4_dir_operations;
5148         } else if (S_ISLNK(inode->i_mode)) {
5149                 if (ext4_inode_is_fast_symlink(inode)) {
5150                         inode->i_op = &ext4_fast_symlink_inode_operations;
5151                         nd_terminate_link(ei->i_data, inode->i_size,
5152                                 sizeof(ei->i_data) - 1);
5153                 } else {
5154                         inode->i_op = &ext4_symlink_inode_operations;
5155                         ext4_set_aops(inode);
5156                 }
5157         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5158               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5159                 inode->i_op = &ext4_special_inode_operations;
5160                 if (raw_inode->i_block[0])
5161                         init_special_inode(inode, inode->i_mode,
5162                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5163                 else
5164                         init_special_inode(inode, inode->i_mode,
5165                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5166         } else {
5167                 ret = -EIO;
5168                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5169                 goto bad_inode;
5170         }
5171         brelse(iloc.bh);
5172         ext4_set_inode_flags(inode);
5173         unlock_new_inode(inode);
5174         return inode;
5175
5176 bad_inode:
5177         brelse(iloc.bh);
5178         iget_failed(inode);
5179         return ERR_PTR(ret);
5180 }
5181
5182 static int ext4_inode_blocks_set(handle_t *handle,
5183                                 struct ext4_inode *raw_inode,
5184                                 struct ext4_inode_info *ei)
5185 {
5186         struct inode *inode = &(ei->vfs_inode);
5187         u64 i_blocks = inode->i_blocks;
5188         struct super_block *sb = inode->i_sb;
5189
5190         if (i_blocks <= ~0U) {
5191                 /*
5192                  * i_blocks can be represnted in a 32 bit variable
5193                  * as multiple of 512 bytes
5194                  */
5195                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5196                 raw_inode->i_blocks_high = 0;
5197                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5198                 return 0;
5199         }
5200         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5201                 return -EFBIG;
5202
5203         if (i_blocks <= 0xffffffffffffULL) {
5204                 /*
5205                  * i_blocks can be represented in a 48 bit variable
5206                  * as multiple of 512 bytes
5207                  */
5208                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5209                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5210                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5211         } else {
5212                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5213                 /* i_block is stored in file system block size */
5214                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5215                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5216                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5217         }
5218         return 0;
5219 }
5220
5221 /*
5222  * Post the struct inode info into an on-disk inode location in the
5223  * buffer-cache.  This gobbles the caller's reference to the
5224  * buffer_head in the inode location struct.
5225  *
5226  * The caller must have write access to iloc->bh.
5227  */
5228 static int ext4_do_update_inode(handle_t *handle,
5229                                 struct inode *inode,
5230                                 struct ext4_iloc *iloc)
5231 {
5232         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5233         struct ext4_inode_info *ei = EXT4_I(inode);
5234         struct buffer_head *bh = iloc->bh;
5235         int err = 0, rc, block;
5236
5237         /* For fields not not tracking in the in-memory inode,
5238          * initialise them to zero for new inodes. */
5239         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5240                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5241
5242         ext4_get_inode_flags(ei);
5243         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5244         if (!(test_opt(inode->i_sb, NO_UID32))) {
5245                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5246                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5247 /*
5248  * Fix up interoperability with old kernels. Otherwise, old inodes get
5249  * re-used with the upper 16 bits of the uid/gid intact
5250  */
5251                 if (!ei->i_dtime) {
5252                         raw_inode->i_uid_high =
5253                                 cpu_to_le16(high_16_bits(inode->i_uid));
5254                         raw_inode->i_gid_high =
5255                                 cpu_to_le16(high_16_bits(inode->i_gid));
5256                 } else {
5257                         raw_inode->i_uid_high = 0;
5258                         raw_inode->i_gid_high = 0;
5259                 }
5260         } else {
5261                 raw_inode->i_uid_low =
5262                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5263                 raw_inode->i_gid_low =
5264                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5265                 raw_inode->i_uid_high = 0;
5266                 raw_inode->i_gid_high = 0;
5267         }
5268         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5269
5270         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5271         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5272         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5273         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5274
5275         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5276                 goto out_brelse;
5277         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5278         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5279         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5280             cpu_to_le32(EXT4_OS_HURD))
5281                 raw_inode->i_file_acl_high =
5282                         cpu_to_le16(ei->i_file_acl >> 32);
5283         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5284         ext4_isize_set(raw_inode, ei->i_disksize);
5285         if (ei->i_disksize > 0x7fffffffULL) {
5286                 struct super_block *sb = inode->i_sb;
5287                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5288                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5289                                 EXT4_SB(sb)->s_es->s_rev_level ==
5290                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5291                         /* If this is the first large file
5292                          * created, add a flag to the superblock.
5293                          */
5294                         err = ext4_journal_get_write_access(handle,
5295                                         EXT4_SB(sb)->s_sbh);
5296                         if (err)
5297                                 goto out_brelse;
5298                         ext4_update_dynamic_rev(sb);
5299                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5300                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5301                         sb->s_dirt = 1;
5302                         ext4_handle_sync(handle);
5303                         err = ext4_handle_dirty_metadata(handle, NULL,
5304                                         EXT4_SB(sb)->s_sbh);
5305                 }
5306         }
5307         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5308         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5309                 if (old_valid_dev(inode->i_rdev)) {
5310                         raw_inode->i_block[0] =
5311                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5312                         raw_inode->i_block[1] = 0;
5313                 } else {
5314                         raw_inode->i_block[0] = 0;
5315                         raw_inode->i_block[1] =
5316                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5317                         raw_inode->i_block[2] = 0;
5318                 }
5319         } else
5320                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5321                         raw_inode->i_block[block] = ei->i_data[block];
5322
5323         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5324         if (ei->i_extra_isize) {
5325                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5326                         raw_inode->i_version_hi =
5327                         cpu_to_le32(inode->i_version >> 32);
5328                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5329         }
5330
5331         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5332         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5333         if (!err)
5334                 err = rc;
5335         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5336
5337         ext4_update_inode_fsync_trans(handle, inode, 0);
5338 out_brelse:
5339         brelse(bh);
5340         ext4_std_error(inode->i_sb, err);
5341         return err;
5342 }
5343
5344 /*
5345  * ext4_write_inode()
5346  *
5347  * We are called from a few places:
5348  *
5349  * - Within generic_file_write() for O_SYNC files.
5350  *   Here, there will be no transaction running. We wait for any running
5351  *   trasnaction to commit.
5352  *
5353  * - Within sys_sync(), kupdate and such.
5354  *   We wait on commit, if tol to.
5355  *
5356  * - Within prune_icache() (PF_MEMALLOC == true)
5357  *   Here we simply return.  We can't afford to block kswapd on the
5358  *   journal commit.
5359  *
5360  * In all cases it is actually safe for us to return without doing anything,
5361  * because the inode has been copied into a raw inode buffer in
5362  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5363  * knfsd.
5364  *
5365  * Note that we are absolutely dependent upon all inode dirtiers doing the
5366  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5367  * which we are interested.
5368  *
5369  * It would be a bug for them to not do this.  The code:
5370  *
5371  *      mark_inode_dirty(inode)
5372  *      stuff();
5373  *      inode->i_size = expr;
5374  *
5375  * is in error because a kswapd-driven write_inode() could occur while
5376  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5377  * will no longer be on the superblock's dirty inode list.
5378  */
5379 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5380 {
5381         int err;
5382
5383         if (current->flags & PF_MEMALLOC)
5384                 return 0;
5385
5386         if (EXT4_SB(inode->i_sb)->s_journal) {
5387                 if (ext4_journal_current_handle()) {
5388                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5389                         dump_stack();
5390                         return -EIO;
5391                 }
5392
5393                 if (wbc->sync_mode != WB_SYNC_ALL)
5394                         return 0;
5395
5396                 err = ext4_force_commit(inode->i_sb);
5397         } else {
5398                 struct ext4_iloc iloc;
5399
5400                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5401                 if (err)
5402                         return err;
5403                 if (wbc->sync_mode == WB_SYNC_ALL)
5404                         sync_dirty_buffer(iloc.bh);
5405                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5406                         EXT4_ERROR_INODE(inode,
5407                                 "IO error syncing inode (block=%llu)",
5408                                 (unsigned long long) iloc.bh->b_blocknr);
5409                         err = -EIO;
5410                 }
5411                 brelse(iloc.bh);
5412         }
5413         return err;
5414 }
5415
5416 /*
5417  * ext4_setattr()
5418  *
5419  * Called from notify_change.
5420  *
5421  * We want to trap VFS attempts to truncate the file as soon as
5422  * possible.  In particular, we want to make sure that when the VFS
5423  * shrinks i_size, we put the inode on the orphan list and modify
5424  * i_disksize immediately, so that during the subsequent flushing of
5425  * dirty pages and freeing of disk blocks, we can guarantee that any
5426  * commit will leave the blocks being flushed in an unused state on
5427  * disk.  (On recovery, the inode will get truncated and the blocks will
5428  * be freed, so we have a strong guarantee that no future commit will
5429  * leave these blocks visible to the user.)
5430  *
5431  * Another thing we have to assure is that if we are in ordered mode
5432  * and inode is still attached to the committing transaction, we must
5433  * we start writeout of all the dirty pages which are being truncated.
5434  * This way we are sure that all the data written in the previous
5435  * transaction are already on disk (truncate waits for pages under
5436  * writeback).
5437  *
5438  * Called with inode->i_mutex down.
5439  */
5440 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5441 {
5442         struct inode *inode = dentry->d_inode;
5443         int error, rc = 0;
5444         const unsigned int ia_valid = attr->ia_valid;
5445
5446         error = inode_change_ok(inode, attr);
5447         if (error)
5448                 return error;
5449
5450         if (is_quota_modification(inode, attr))
5451                 dquot_initialize(inode);
5452         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5453                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5454                 handle_t *handle;
5455
5456                 /* (user+group)*(old+new) structure, inode write (sb,
5457                  * inode block, ? - but truncate inode update has it) */
5458                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5459                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5460                 if (IS_ERR(handle)) {
5461                         error = PTR_ERR(handle);
5462                         goto err_out;
5463                 }
5464                 error = dquot_transfer(inode, attr);
5465                 if (error) {
5466                         ext4_journal_stop(handle);
5467                         return error;
5468                 }
5469                 /* Update corresponding info in inode so that everything is in
5470                  * one transaction */
5471                 if (attr->ia_valid & ATTR_UID)
5472                         inode->i_uid = attr->ia_uid;
5473                 if (attr->ia_valid & ATTR_GID)
5474                         inode->i_gid = attr->ia_gid;
5475                 error = ext4_mark_inode_dirty(handle, inode);
5476                 ext4_journal_stop(handle);
5477         }
5478
5479         if (attr->ia_valid & ATTR_SIZE) {
5480                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5481                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5482
5483                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5484                                 error = -EFBIG;
5485                                 goto err_out;
5486                         }
5487                 }
5488         }
5489
5490         if (S_ISREG(inode->i_mode) &&
5491             attr->ia_valid & ATTR_SIZE &&
5492             (attr->ia_size < inode->i_size ||
5493              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5494                 handle_t *handle;
5495
5496                 handle = ext4_journal_start(inode, 3);
5497                 if (IS_ERR(handle)) {
5498                         error = PTR_ERR(handle);
5499                         goto err_out;
5500                 }
5501
5502                 error = ext4_orphan_add(handle, inode);
5503                 EXT4_I(inode)->i_disksize = attr->ia_size;
5504                 rc = ext4_mark_inode_dirty(handle, inode);
5505                 if (!error)
5506                         error = rc;
5507                 ext4_journal_stop(handle);
5508
5509                 if (ext4_should_order_data(inode)) {
5510                         error = ext4_begin_ordered_truncate(inode,
5511                                                             attr->ia_size);
5512                         if (error) {
5513                                 /* Do as much error cleanup as possible */
5514                                 handle = ext4_journal_start(inode, 3);
5515                                 if (IS_ERR(handle)) {
5516                                         ext4_orphan_del(NULL, inode);
5517                                         goto err_out;
5518                                 }
5519                                 ext4_orphan_del(handle, inode);
5520                                 ext4_journal_stop(handle);
5521                                 goto err_out;
5522                         }
5523                 }
5524                 /* ext4_truncate will clear the flag */
5525                 if ((ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))
5526                         ext4_truncate(inode);
5527         }
5528
5529         rc = inode_setattr(inode, attr);
5530
5531         /* If inode_setattr's call to ext4_truncate failed to get a
5532          * transaction handle at all, we need to clean up the in-core
5533          * orphan list manually. */