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